src/share/vm/opto/memnode.cpp

Fri, 11 Jul 2008 01:14:44 -0700

author
trims
date
Fri, 11 Jul 2008 01:14:44 -0700
changeset 670
9c2ecc2ffb12
parent 631
d1605aabd0a1
parent 658
1dd146f17531
child 682
02a35ad4adf8
permissions
-rw-r--r--

Merge

     1 /*
     2  * Copyright 1997-2008 Sun Microsystems, Inc.  All Rights Reserved.
     3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
     4  *
     5  * This code is free software; you can redistribute it and/or modify it
     6  * under the terms of the GNU General Public License version 2 only, as
     7  * published by the Free Software Foundation.
     8  *
     9  * This code is distributed in the hope that it will be useful, but WITHOUT
    10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
    11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
    12  * version 2 for more details (a copy is included in the LICENSE file that
    13  * accompanied this code).
    14  *
    15  * You should have received a copy of the GNU General Public License version
    16  * 2 along with this work; if not, write to the Free Software Foundation,
    17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
    18  *
    19  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
    20  * CA 95054 USA or visit www.sun.com if you need additional information or
    21  * have any questions.
    22  *
    23  */
    25 // Portions of code courtesy of Clifford Click
    27 // Optimization - Graph Style
    29 #include "incls/_precompiled.incl"
    30 #include "incls/_memnode.cpp.incl"
    32 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
    34 //=============================================================================
    35 uint MemNode::size_of() const { return sizeof(*this); }
    37 const TypePtr *MemNode::adr_type() const {
    38   Node* adr = in(Address);
    39   const TypePtr* cross_check = NULL;
    40   DEBUG_ONLY(cross_check = _adr_type);
    41   return calculate_adr_type(adr->bottom_type(), cross_check);
    42 }
    44 #ifndef PRODUCT
    45 void MemNode::dump_spec(outputStream *st) const {
    46   if (in(Address) == NULL)  return; // node is dead
    47 #ifndef ASSERT
    48   // fake the missing field
    49   const TypePtr* _adr_type = NULL;
    50   if (in(Address) != NULL)
    51     _adr_type = in(Address)->bottom_type()->isa_ptr();
    52 #endif
    53   dump_adr_type(this, _adr_type, st);
    55   Compile* C = Compile::current();
    56   if( C->alias_type(_adr_type)->is_volatile() )
    57     st->print(" Volatile!");
    58 }
    60 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
    61   st->print(" @");
    62   if (adr_type == NULL) {
    63     st->print("NULL");
    64   } else {
    65     adr_type->dump_on(st);
    66     Compile* C = Compile::current();
    67     Compile::AliasType* atp = NULL;
    68     if (C->have_alias_type(adr_type))  atp = C->alias_type(adr_type);
    69     if (atp == NULL)
    70       st->print(", idx=?\?;");
    71     else if (atp->index() == Compile::AliasIdxBot)
    72       st->print(", idx=Bot;");
    73     else if (atp->index() == Compile::AliasIdxTop)
    74       st->print(", idx=Top;");
    75     else if (atp->index() == Compile::AliasIdxRaw)
    76       st->print(", idx=Raw;");
    77     else {
    78       ciField* field = atp->field();
    79       if (field) {
    80         st->print(", name=");
    81         field->print_name_on(st);
    82       }
    83       st->print(", idx=%d;", atp->index());
    84     }
    85   }
    86 }
    88 extern void print_alias_types();
    90 #endif
    92 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
    93   const TypeOopPtr *tinst = t_adr->isa_oopptr();
    94   if (tinst == NULL || !tinst->is_known_instance_field())
    95     return mchain;  // don't try to optimize non-instance types
    96   uint instance_id = tinst->instance_id();
    97   Node *prev = NULL;
    98   Node *result = mchain;
    99   while (prev != result) {
   100     prev = result;
   101     // skip over a call which does not affect this memory slice
   102     if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
   103       Node *proj_in = result->in(0);
   104       if (proj_in->is_Call()) {
   105         CallNode *call = proj_in->as_Call();
   106         if (!call->may_modify(t_adr, phase)) {
   107           result = call->in(TypeFunc::Memory);
   108         }
   109       } else if (proj_in->is_Initialize()) {
   110         AllocateNode* alloc = proj_in->as_Initialize()->allocation();
   111         // Stop if this is the initialization for the object instance which
   112         // which contains this memory slice, otherwise skip over it.
   113         if (alloc != NULL && alloc->_idx != instance_id) {
   114           result = proj_in->in(TypeFunc::Memory);
   115         }
   116       } else if (proj_in->is_MemBar()) {
   117         result = proj_in->in(TypeFunc::Memory);
   118       }
   119     } else if (result->is_MergeMem()) {
   120       result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
   121     }
   122   }
   123   return result;
   124 }
   126 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
   127   const TypeOopPtr *t_oop = t_adr->isa_oopptr();
   128   bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field();
   129   PhaseIterGVN *igvn = phase->is_IterGVN();
   130   Node *result = mchain;
   131   result = optimize_simple_memory_chain(result, t_adr, phase);
   132   if (is_instance && igvn != NULL  && result->is_Phi()) {
   133     PhiNode *mphi = result->as_Phi();
   134     assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
   135     const TypePtr *t = mphi->adr_type();
   136     if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
   137         t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
   138         t->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
   139       // clone the Phi with our address type
   140       result = mphi->split_out_instance(t_adr, igvn);
   141     } else {
   142       assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
   143     }
   144   }
   145   return result;
   146 }
   148 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
   149   uint alias_idx = phase->C->get_alias_index(tp);
   150   Node *mem = mmem;
   151 #ifdef ASSERT
   152   {
   153     // Check that current type is consistent with the alias index used during graph construction
   154     assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
   155     bool consistent =  adr_check == NULL || adr_check->empty() ||
   156                        phase->C->must_alias(adr_check, alias_idx );
   157     // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
   158     if( !consistent && adr_check != NULL && !adr_check->empty() &&
   159                tp->isa_aryptr() &&        tp->offset() == Type::OffsetBot &&
   160         adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
   161         ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
   162           adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
   163           adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
   164       // don't assert if it is dead code.
   165       consistent = true;
   166     }
   167     if( !consistent ) {
   168       st->print("alias_idx==%d, adr_check==", alias_idx);
   169       if( adr_check == NULL ) {
   170         st->print("NULL");
   171       } else {
   172         adr_check->dump();
   173       }
   174       st->cr();
   175       print_alias_types();
   176       assert(consistent, "adr_check must match alias idx");
   177     }
   178   }
   179 #endif
   180   // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
   181   // means an array I have not precisely typed yet.  Do not do any
   182   // alias stuff with it any time soon.
   183   const TypeOopPtr *tinst = tp->isa_oopptr();
   184   if( tp->base() != Type::AnyPtr &&
   185       !(tinst &&
   186         tinst->klass()->is_java_lang_Object() &&
   187         tinst->offset() == Type::OffsetBot) ) {
   188     // compress paths and change unreachable cycles to TOP
   189     // If not, we can update the input infinitely along a MergeMem cycle
   190     // Equivalent code in PhiNode::Ideal
   191     Node* m  = phase->transform(mmem);
   192     // If tranformed to a MergeMem, get the desired slice
   193     // Otherwise the returned node represents memory for every slice
   194     mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
   195     // Update input if it is progress over what we have now
   196   }
   197   return mem;
   198 }
   200 //--------------------------Ideal_common---------------------------------------
   201 // Look for degenerate control and memory inputs.  Bypass MergeMem inputs.
   202 // Unhook non-raw memories from complete (macro-expanded) initializations.
   203 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
   204   // If our control input is a dead region, kill all below the region
   205   Node *ctl = in(MemNode::Control);
   206   if (ctl && remove_dead_region(phase, can_reshape))
   207     return this;
   209   // Ignore if memory is dead, or self-loop
   210   Node *mem = in(MemNode::Memory);
   211   if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
   212   assert( mem != this, "dead loop in MemNode::Ideal" );
   214   Node *address = in(MemNode::Address);
   215   const Type *t_adr = phase->type( address );
   216   if( t_adr == Type::TOP )              return NodeSentinel; // caller will return NULL
   218   // Avoid independent memory operations
   219   Node* old_mem = mem;
   221   // The code which unhooks non-raw memories from complete (macro-expanded)
   222   // initializations was removed. After macro-expansion all stores catched
   223   // by Initialize node became raw stores and there is no information
   224   // which memory slices they modify. So it is unsafe to move any memory
   225   // operation above these stores. Also in most cases hooked non-raw memories
   226   // were already unhooked by using information from detect_ptr_independence()
   227   // and find_previous_store().
   229   if (mem->is_MergeMem()) {
   230     MergeMemNode* mmem = mem->as_MergeMem();
   231     const TypePtr *tp = t_adr->is_ptr();
   233     mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
   234   }
   236   if (mem != old_mem) {
   237     set_req(MemNode::Memory, mem);
   238     return this;
   239   }
   241   // let the subclass continue analyzing...
   242   return NULL;
   243 }
   245 // Helper function for proving some simple control dominations.
   246 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
   247 // Already assumes that 'dom' is available at 'sub', and that 'sub'
   248 // is not a constant (dominated by the method's StartNode).
   249 // Used by MemNode::find_previous_store to prove that the
   250 // control input of a memory operation predates (dominates)
   251 // an allocation it wants to look past.
   252 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
   253   if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
   254     return false; // Conservative answer for dead code
   256   // Check 'dom'. Skip Proj and CatchProj nodes.
   257   dom = dom->find_exact_control(dom);
   258   if (dom == NULL || dom->is_top())
   259     return false; // Conservative answer for dead code
   261   if (dom == sub) {
   262     // For the case when, for example, 'sub' is Initialize and the original
   263     // 'dom' is Proj node of the 'sub'.
   264     return false;
   265   }
   267   if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
   268     return true;
   270   // 'dom' dominates 'sub' if its control edge and control edges
   271   // of all its inputs dominate or equal to sub's control edge.
   273   // Currently 'sub' is either Allocate, Initialize or Start nodes.
   274   // Or Region for the check in LoadNode::Ideal();
   275   // 'sub' should have sub->in(0) != NULL.
   276   assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
   277          sub->is_Region(), "expecting only these nodes");
   279   // Get control edge of 'sub'.
   280   Node* orig_sub = sub;
   281   sub = sub->find_exact_control(sub->in(0));
   282   if (sub == NULL || sub->is_top())
   283     return false; // Conservative answer for dead code
   285   assert(sub->is_CFG(), "expecting control");
   287   if (sub == dom)
   288     return true;
   290   if (sub->is_Start() || sub->is_Root())
   291     return false;
   293   {
   294     // Check all control edges of 'dom'.
   296     ResourceMark rm;
   297     Arena* arena = Thread::current()->resource_area();
   298     Node_List nlist(arena);
   299     Unique_Node_List dom_list(arena);
   301     dom_list.push(dom);
   302     bool only_dominating_controls = false;
   304     for (uint next = 0; next < dom_list.size(); next++) {
   305       Node* n = dom_list.at(next);
   306       if (n == orig_sub)
   307         return false; // One of dom's inputs dominated by sub.
   308       if (!n->is_CFG() && n->pinned()) {
   309         // Check only own control edge for pinned non-control nodes.
   310         n = n->find_exact_control(n->in(0));
   311         if (n == NULL || n->is_top())
   312           return false; // Conservative answer for dead code
   313         assert(n->is_CFG(), "expecting control");
   314         dom_list.push(n);
   315       } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
   316         only_dominating_controls = true;
   317       } else if (n->is_CFG()) {
   318         if (n->dominates(sub, nlist))
   319           only_dominating_controls = true;
   320         else
   321           return false;
   322       } else {
   323         // First, own control edge.
   324         Node* m = n->find_exact_control(n->in(0));
   325         if (m != NULL) {
   326           if (m->is_top())
   327             return false; // Conservative answer for dead code
   328           dom_list.push(m);
   329         }
   330         // Now, the rest of edges.
   331         uint cnt = n->req();
   332         for (uint i = 1; i < cnt; i++) {
   333           m = n->find_exact_control(n->in(i));
   334           if (m == NULL || m->is_top())
   335             continue;
   336           dom_list.push(m);
   337         }
   338       }
   339     }
   340     return only_dominating_controls;
   341   }
   342 }
   344 //---------------------detect_ptr_independence---------------------------------
   345 // Used by MemNode::find_previous_store to prove that two base
   346 // pointers are never equal.
   347 // The pointers are accompanied by their associated allocations,
   348 // if any, which have been previously discovered by the caller.
   349 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
   350                                       Node* p2, AllocateNode* a2,
   351                                       PhaseTransform* phase) {
   352   // Attempt to prove that these two pointers cannot be aliased.
   353   // They may both manifestly be allocations, and they should differ.
   354   // Or, if they are not both allocations, they can be distinct constants.
   355   // Otherwise, one is an allocation and the other a pre-existing value.
   356   if (a1 == NULL && a2 == NULL) {           // neither an allocation
   357     return (p1 != p2) && p1->is_Con() && p2->is_Con();
   358   } else if (a1 != NULL && a2 != NULL) {    // both allocations
   359     return (a1 != a2);
   360   } else if (a1 != NULL) {                  // one allocation a1
   361     // (Note:  p2->is_Con implies p2->in(0)->is_Root, which dominates.)
   362     return all_controls_dominate(p2, a1);
   363   } else { //(a2 != NULL)                   // one allocation a2
   364     return all_controls_dominate(p1, a2);
   365   }
   366   return false;
   367 }
   370 // The logic for reordering loads and stores uses four steps:
   371 // (a) Walk carefully past stores and initializations which we
   372 //     can prove are independent of this load.
   373 // (b) Observe that the next memory state makes an exact match
   374 //     with self (load or store), and locate the relevant store.
   375 // (c) Ensure that, if we were to wire self directly to the store,
   376 //     the optimizer would fold it up somehow.
   377 // (d) Do the rewiring, and return, depending on some other part of
   378 //     the optimizer to fold up the load.
   379 // This routine handles steps (a) and (b).  Steps (c) and (d) are
   380 // specific to loads and stores, so they are handled by the callers.
   381 // (Currently, only LoadNode::Ideal has steps (c), (d).  More later.)
   382 //
   383 Node* MemNode::find_previous_store(PhaseTransform* phase) {
   384   Node*         ctrl   = in(MemNode::Control);
   385   Node*         adr    = in(MemNode::Address);
   386   intptr_t      offset = 0;
   387   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
   388   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
   390   if (offset == Type::OffsetBot)
   391     return NULL;            // cannot unalias unless there are precise offsets
   393   const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
   395   intptr_t size_in_bytes = memory_size();
   397   Node* mem = in(MemNode::Memory);   // start searching here...
   399   int cnt = 50;             // Cycle limiter
   400   for (;;) {                // While we can dance past unrelated stores...
   401     if (--cnt < 0)  break;  // Caught in cycle or a complicated dance?
   403     if (mem->is_Store()) {
   404       Node* st_adr = mem->in(MemNode::Address);
   405       intptr_t st_offset = 0;
   406       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
   407       if (st_base == NULL)
   408         break;              // inscrutable pointer
   409       if (st_offset != offset && st_offset != Type::OffsetBot) {
   410         const int MAX_STORE = BytesPerLong;
   411         if (st_offset >= offset + size_in_bytes ||
   412             st_offset <= offset - MAX_STORE ||
   413             st_offset <= offset - mem->as_Store()->memory_size()) {
   414           // Success:  The offsets are provably independent.
   415           // (You may ask, why not just test st_offset != offset and be done?
   416           // The answer is that stores of different sizes can co-exist
   417           // in the same sequence of RawMem effects.  We sometimes initialize
   418           // a whole 'tile' of array elements with a single jint or jlong.)
   419           mem = mem->in(MemNode::Memory);
   420           continue;           // (a) advance through independent store memory
   421         }
   422       }
   423       if (st_base != base &&
   424           detect_ptr_independence(base, alloc,
   425                                   st_base,
   426                                   AllocateNode::Ideal_allocation(st_base, phase),
   427                                   phase)) {
   428         // Success:  The bases are provably independent.
   429         mem = mem->in(MemNode::Memory);
   430         continue;           // (a) advance through independent store memory
   431       }
   433       // (b) At this point, if the bases or offsets do not agree, we lose,
   434       // since we have not managed to prove 'this' and 'mem' independent.
   435       if (st_base == base && st_offset == offset) {
   436         return mem;         // let caller handle steps (c), (d)
   437       }
   439     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
   440       InitializeNode* st_init = mem->in(0)->as_Initialize();
   441       AllocateNode*  st_alloc = st_init->allocation();
   442       if (st_alloc == NULL)
   443         break;              // something degenerated
   444       bool known_identical = false;
   445       bool known_independent = false;
   446       if (alloc == st_alloc)
   447         known_identical = true;
   448       else if (alloc != NULL)
   449         known_independent = true;
   450       else if (all_controls_dominate(this, st_alloc))
   451         known_independent = true;
   453       if (known_independent) {
   454         // The bases are provably independent: Either they are
   455         // manifestly distinct allocations, or else the control
   456         // of this load dominates the store's allocation.
   457         int alias_idx = phase->C->get_alias_index(adr_type());
   458         if (alias_idx == Compile::AliasIdxRaw) {
   459           mem = st_alloc->in(TypeFunc::Memory);
   460         } else {
   461           mem = st_init->memory(alias_idx);
   462         }
   463         continue;           // (a) advance through independent store memory
   464       }
   466       // (b) at this point, if we are not looking at a store initializing
   467       // the same allocation we are loading from, we lose.
   468       if (known_identical) {
   469         // From caller, can_see_stored_value will consult find_captured_store.
   470         return mem;         // let caller handle steps (c), (d)
   471       }
   473     } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
   474       // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
   475       if (mem->is_Proj() && mem->in(0)->is_Call()) {
   476         CallNode *call = mem->in(0)->as_Call();
   477         if (!call->may_modify(addr_t, phase)) {
   478           mem = call->in(TypeFunc::Memory);
   479           continue;         // (a) advance through independent call memory
   480         }
   481       } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
   482         mem = mem->in(0)->in(TypeFunc::Memory);
   483         continue;           // (a) advance through independent MemBar memory
   484       } else if (mem->is_MergeMem()) {
   485         int alias_idx = phase->C->get_alias_index(adr_type());
   486         mem = mem->as_MergeMem()->memory_at(alias_idx);
   487         continue;           // (a) advance through independent MergeMem memory
   488       }
   489     }
   491     // Unless there is an explicit 'continue', we must bail out here,
   492     // because 'mem' is an inscrutable memory state (e.g., a call).
   493     break;
   494   }
   496   return NULL;              // bail out
   497 }
   499 //----------------------calculate_adr_type-------------------------------------
   500 // Helper function.  Notices when the given type of address hits top or bottom.
   501 // Also, asserts a cross-check of the type against the expected address type.
   502 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
   503   if (t == Type::TOP)  return NULL; // does not touch memory any more?
   504   #ifdef PRODUCT
   505   cross_check = NULL;
   506   #else
   507   if (!VerifyAliases || is_error_reported() || Node::in_dump())  cross_check = NULL;
   508   #endif
   509   const TypePtr* tp = t->isa_ptr();
   510   if (tp == NULL) {
   511     assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
   512     return TypePtr::BOTTOM;           // touches lots of memory
   513   } else {
   514     #ifdef ASSERT
   515     // %%%% [phh] We don't check the alias index if cross_check is
   516     //            TypeRawPtr::BOTTOM.  Needs to be investigated.
   517     if (cross_check != NULL &&
   518         cross_check != TypePtr::BOTTOM &&
   519         cross_check != TypeRawPtr::BOTTOM) {
   520       // Recheck the alias index, to see if it has changed (due to a bug).
   521       Compile* C = Compile::current();
   522       assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
   523              "must stay in the original alias category");
   524       // The type of the address must be contained in the adr_type,
   525       // disregarding "null"-ness.
   526       // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
   527       const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
   528       assert(cross_check->meet(tp_notnull) == cross_check,
   529              "real address must not escape from expected memory type");
   530     }
   531     #endif
   532     return tp;
   533   }
   534 }
   536 //------------------------adr_phi_is_loop_invariant----------------------------
   537 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
   538 // loop is loop invariant. Make a quick traversal of Phi and associated
   539 // CastPP nodes, looking to see if they are a closed group within the loop.
   540 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
   541   // The idea is that the phi-nest must boil down to only CastPP nodes
   542   // with the same data. This implies that any path into the loop already
   543   // includes such a CastPP, and so the original cast, whatever its input,
   544   // must be covered by an equivalent cast, with an earlier control input.
   545   ResourceMark rm;
   547   // The loop entry input of the phi should be the unique dominating
   548   // node for every Phi/CastPP in the loop.
   549   Unique_Node_List closure;
   550   closure.push(adr_phi->in(LoopNode::EntryControl));
   552   // Add the phi node and the cast to the worklist.
   553   Unique_Node_List worklist;
   554   worklist.push(adr_phi);
   555   if( cast != NULL ){
   556     if( !cast->is_ConstraintCast() ) return false;
   557     worklist.push(cast);
   558   }
   560   // Begin recursive walk of phi nodes.
   561   while( worklist.size() ){
   562     // Take a node off the worklist
   563     Node *n = worklist.pop();
   564     if( !closure.member(n) ){
   565       // Add it to the closure.
   566       closure.push(n);
   567       // Make a sanity check to ensure we don't waste too much time here.
   568       if( closure.size() > 20) return false;
   569       // This node is OK if:
   570       //  - it is a cast of an identical value
   571       //  - or it is a phi node (then we add its inputs to the worklist)
   572       // Otherwise, the node is not OK, and we presume the cast is not invariant
   573       if( n->is_ConstraintCast() ){
   574         worklist.push(n->in(1));
   575       } else if( n->is_Phi() ) {
   576         for( uint i = 1; i < n->req(); i++ ) {
   577           worklist.push(n->in(i));
   578         }
   579       } else {
   580         return false;
   581       }
   582     }
   583   }
   585   // Quit when the worklist is empty, and we've found no offending nodes.
   586   return true;
   587 }
   589 //------------------------------Ideal_DU_postCCP-------------------------------
   590 // Find any cast-away of null-ness and keep its control.  Null cast-aways are
   591 // going away in this pass and we need to make this memory op depend on the
   592 // gating null check.
   593 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
   594   return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
   595 }
   597 // I tried to leave the CastPP's in.  This makes the graph more accurate in
   598 // some sense; we get to keep around the knowledge that an oop is not-null
   599 // after some test.  Alas, the CastPP's interfere with GVN (some values are
   600 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
   601 // cannot collapse, etc).  This cost us 10% on SpecJVM, even when I removed
   602 // some of the more trivial cases in the optimizer.  Removing more useless
   603 // Phi's started allowing Loads to illegally float above null checks.  I gave
   604 // up on this approach.  CNC 10/20/2000
   605 // This static method may be called not from MemNode (EncodePNode calls it).
   606 // Only the control edge of the node 'n' might be updated.
   607 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
   608   Node *skipped_cast = NULL;
   609   // Need a null check?  Regular static accesses do not because they are
   610   // from constant addresses.  Array ops are gated by the range check (which
   611   // always includes a NULL check).  Just check field ops.
   612   if( n->in(MemNode::Control) == NULL ) {
   613     // Scan upwards for the highest location we can place this memory op.
   614     while( true ) {
   615       switch( adr->Opcode() ) {
   617       case Op_AddP:             // No change to NULL-ness, so peek thru AddP's
   618         adr = adr->in(AddPNode::Base);
   619         continue;
   621       case Op_DecodeN:         // No change to NULL-ness, so peek thru
   622         adr = adr->in(1);
   623         continue;
   625       case Op_CastPP:
   626         // If the CastPP is useless, just peek on through it.
   627         if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
   628           // Remember the cast that we've peeked though. If we peek
   629           // through more than one, then we end up remembering the highest
   630           // one, that is, if in a loop, the one closest to the top.
   631           skipped_cast = adr;
   632           adr = adr->in(1);
   633           continue;
   634         }
   635         // CastPP is going away in this pass!  We need this memory op to be
   636         // control-dependent on the test that is guarding the CastPP.
   637         ccp->hash_delete(n);
   638         n->set_req(MemNode::Control, adr->in(0));
   639         ccp->hash_insert(n);
   640         return n;
   642       case Op_Phi:
   643         // Attempt to float above a Phi to some dominating point.
   644         if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
   645           // If we've already peeked through a Cast (which could have set the
   646           // control), we can't float above a Phi, because the skipped Cast
   647           // may not be loop invariant.
   648           if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
   649             adr = adr->in(1);
   650             continue;
   651           }
   652         }
   654         // Intentional fallthrough!
   656         // No obvious dominating point.  The mem op is pinned below the Phi
   657         // by the Phi itself.  If the Phi goes away (no true value is merged)
   658         // then the mem op can float, but not indefinitely.  It must be pinned
   659         // behind the controls leading to the Phi.
   660       case Op_CheckCastPP:
   661         // These usually stick around to change address type, however a
   662         // useless one can be elided and we still need to pick up a control edge
   663         if (adr->in(0) == NULL) {
   664           // This CheckCastPP node has NO control and is likely useless. But we
   665           // need check further up the ancestor chain for a control input to keep
   666           // the node in place. 4959717.
   667           skipped_cast = adr;
   668           adr = adr->in(1);
   669           continue;
   670         }
   671         ccp->hash_delete(n);
   672         n->set_req(MemNode::Control, adr->in(0));
   673         ccp->hash_insert(n);
   674         return n;
   676         // List of "safe" opcodes; those that implicitly block the memory
   677         // op below any null check.
   678       case Op_CastX2P:          // no null checks on native pointers
   679       case Op_Parm:             // 'this' pointer is not null
   680       case Op_LoadP:            // Loading from within a klass
   681       case Op_LoadN:            // Loading from within a klass
   682       case Op_LoadKlass:        // Loading from within a klass
   683       case Op_LoadNKlass:       // Loading from within a klass
   684       case Op_ConP:             // Loading from a klass
   685       case Op_ConN:             // Loading from a klass
   686       case Op_CreateEx:         // Sucking up the guts of an exception oop
   687       case Op_Con:              // Reading from TLS
   688       case Op_CMoveP:           // CMoveP is pinned
   689       case Op_CMoveN:           // CMoveN is pinned
   690         break;                  // No progress
   692       case Op_Proj:             // Direct call to an allocation routine
   693       case Op_SCMemProj:        // Memory state from store conditional ops
   694 #ifdef ASSERT
   695         {
   696           assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
   697           const Node* call = adr->in(0);
   698           if (call->is_CallJava()) {
   699             const CallJavaNode* call_java = call->as_CallJava();
   700             const TypeTuple *r = call_java->tf()->range();
   701             assert(r->cnt() > TypeFunc::Parms, "must return value");
   702             const Type* ret_type = r->field_at(TypeFunc::Parms);
   703             assert(ret_type && ret_type->isa_ptr(), "must return pointer");
   704             // We further presume that this is one of
   705             // new_instance_Java, new_array_Java, or
   706             // the like, but do not assert for this.
   707           } else if (call->is_Allocate()) {
   708             // similar case to new_instance_Java, etc.
   709           } else if (!call->is_CallLeaf()) {
   710             // Projections from fetch_oop (OSR) are allowed as well.
   711             ShouldNotReachHere();
   712           }
   713         }
   714 #endif
   715         break;
   716       default:
   717         ShouldNotReachHere();
   718       }
   719       break;
   720     }
   721   }
   723   return  NULL;               // No progress
   724 }
   727 //=============================================================================
   728 uint LoadNode::size_of() const { return sizeof(*this); }
   729 uint LoadNode::cmp( const Node &n ) const
   730 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
   731 const Type *LoadNode::bottom_type() const { return _type; }
   732 uint LoadNode::ideal_reg() const {
   733   return Matcher::base2reg[_type->base()];
   734 }
   736 #ifndef PRODUCT
   737 void LoadNode::dump_spec(outputStream *st) const {
   738   MemNode::dump_spec(st);
   739   if( !Verbose && !WizardMode ) {
   740     // standard dump does this in Verbose and WizardMode
   741     st->print(" #"); _type->dump_on(st);
   742   }
   743 }
   744 #endif
   747 //----------------------------LoadNode::make-----------------------------------
   748 // Polymorphic factory method:
   749 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
   750   Compile* C = gvn.C;
   752   // sanity check the alias category against the created node type
   753   assert(!(adr_type->isa_oopptr() &&
   754            adr_type->offset() == oopDesc::klass_offset_in_bytes()),
   755          "use LoadKlassNode instead");
   756   assert(!(adr_type->isa_aryptr() &&
   757            adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
   758          "use LoadRangeNode instead");
   759   switch (bt) {
   760   case T_BOOLEAN:
   761   case T_BYTE:    return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int()    );
   762   case T_INT:     return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int()    );
   763   case T_CHAR:    return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int()    );
   764   case T_SHORT:   return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int()    );
   765   case T_LONG:    return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long()   );
   766   case T_FLOAT:   return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt              );
   767   case T_DOUBLE:  return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt              );
   768   case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr()    );
   769   case T_OBJECT:
   770 #ifdef _LP64
   771     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
   772       Node* load  = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
   773       return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr());
   774     } else
   775 #endif
   776     {
   777       assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
   778       return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
   779     }
   780   }
   781   ShouldNotReachHere();
   782   return (LoadNode*)NULL;
   783 }
   785 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
   786   bool require_atomic = true;
   787   return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
   788 }
   793 //------------------------------hash-------------------------------------------
   794 uint LoadNode::hash() const {
   795   // unroll addition of interesting fields
   796   return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
   797 }
   799 //---------------------------can_see_stored_value------------------------------
   800 // This routine exists to make sure this set of tests is done the same
   801 // everywhere.  We need to make a coordinated change: first LoadNode::Ideal
   802 // will change the graph shape in a way which makes memory alive twice at the
   803 // same time (uses the Oracle model of aliasing), then some
   804 // LoadXNode::Identity will fold things back to the equivalence-class model
   805 // of aliasing.
   806 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
   807   Node* ld_adr = in(MemNode::Address);
   809   const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
   810   Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
   811   if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
   812       atp->field() != NULL && !atp->field()->is_volatile()) {
   813     uint alias_idx = atp->index();
   814     bool final = atp->field()->is_final();
   815     Node* result = NULL;
   816     Node* current = st;
   817     // Skip through chains of MemBarNodes checking the MergeMems for
   818     // new states for the slice of this load.  Stop once any other
   819     // kind of node is encountered.  Loads from final memory can skip
   820     // through any kind of MemBar but normal loads shouldn't skip
   821     // through MemBarAcquire since the could allow them to move out of
   822     // a synchronized region.
   823     while (current->is_Proj()) {
   824       int opc = current->in(0)->Opcode();
   825       if ((final && opc == Op_MemBarAcquire) ||
   826           opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
   827         Node* mem = current->in(0)->in(TypeFunc::Memory);
   828         if (mem->is_MergeMem()) {
   829           MergeMemNode* merge = mem->as_MergeMem();
   830           Node* new_st = merge->memory_at(alias_idx);
   831           if (new_st == merge->base_memory()) {
   832             // Keep searching
   833             current = merge->base_memory();
   834             continue;
   835           }
   836           // Save the new memory state for the slice and fall through
   837           // to exit.
   838           result = new_st;
   839         }
   840       }
   841       break;
   842     }
   843     if (result != NULL) {
   844       st = result;
   845     }
   846   }
   849   // Loop around twice in the case Load -> Initialize -> Store.
   850   // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
   851   for (int trip = 0; trip <= 1; trip++) {
   853     if (st->is_Store()) {
   854       Node* st_adr = st->in(MemNode::Address);
   855       if (!phase->eqv(st_adr, ld_adr)) {
   856         // Try harder before giving up...  Match raw and non-raw pointers.
   857         intptr_t st_off = 0;
   858         AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
   859         if (alloc == NULL)       return NULL;
   860         intptr_t ld_off = 0;
   861         AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
   862         if (alloc != allo2)      return NULL;
   863         if (ld_off != st_off)    return NULL;
   864         // At this point we have proven something like this setup:
   865         //  A = Allocate(...)
   866         //  L = LoadQ(,  AddP(CastPP(, A.Parm),, #Off))
   867         //  S = StoreQ(, AddP(,        A.Parm  , #Off), V)
   868         // (Actually, we haven't yet proven the Q's are the same.)
   869         // In other words, we are loading from a casted version of
   870         // the same pointer-and-offset that we stored to.
   871         // Thus, we are able to replace L by V.
   872       }
   873       // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
   874       if (store_Opcode() != st->Opcode())
   875         return NULL;
   876       return st->in(MemNode::ValueIn);
   877     }
   879     intptr_t offset = 0;  // scratch
   881     // A load from a freshly-created object always returns zero.
   882     // (This can happen after LoadNode::Ideal resets the load's memory input
   883     // to find_captured_store, which returned InitializeNode::zero_memory.)
   884     if (st->is_Proj() && st->in(0)->is_Allocate() &&
   885         st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
   886         offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
   887       // return a zero value for the load's basic type
   888       // (This is one of the few places where a generic PhaseTransform
   889       // can create new nodes.  Think of it as lazily manifesting
   890       // virtually pre-existing constants.)
   891       return phase->zerocon(memory_type());
   892     }
   894     // A load from an initialization barrier can match a captured store.
   895     if (st->is_Proj() && st->in(0)->is_Initialize()) {
   896       InitializeNode* init = st->in(0)->as_Initialize();
   897       AllocateNode* alloc = init->allocation();
   898       if (alloc != NULL &&
   899           alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
   900         // examine a captured store value
   901         st = init->find_captured_store(offset, memory_size(), phase);
   902         if (st != NULL)
   903           continue;             // take one more trip around
   904       }
   905     }
   907     break;
   908   }
   910   return NULL;
   911 }
   913 //----------------------is_instance_field_load_with_local_phi------------------
   914 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
   915   if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
   916       in(MemNode::Address)->is_AddP() ) {
   917     const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
   918     // Only instances.
   919     if( t_oop != NULL && t_oop->is_known_instance_field() &&
   920         t_oop->offset() != Type::OffsetBot &&
   921         t_oop->offset() != Type::OffsetTop) {
   922       return true;
   923     }
   924   }
   925   return false;
   926 }
   928 //------------------------------Identity---------------------------------------
   929 // Loads are identity if previous store is to same address
   930 Node *LoadNode::Identity( PhaseTransform *phase ) {
   931   // If the previous store-maker is the right kind of Store, and the store is
   932   // to the same address, then we are equal to the value stored.
   933   Node* mem = in(MemNode::Memory);
   934   Node* value = can_see_stored_value(mem, phase);
   935   if( value ) {
   936     // byte, short & char stores truncate naturally.
   937     // A load has to load the truncated value which requires
   938     // some sort of masking operation and that requires an
   939     // Ideal call instead of an Identity call.
   940     if (memory_size() < BytesPerInt) {
   941       // If the input to the store does not fit with the load's result type,
   942       // it must be truncated via an Ideal call.
   943       if (!phase->type(value)->higher_equal(phase->type(this)))
   944         return this;
   945     }
   946     // (This works even when value is a Con, but LoadNode::Value
   947     // usually runs first, producing the singleton type of the Con.)
   948     return value;
   949   }
   951   // Search for an existing data phi which was generated before for the same
   952   // instance's field to avoid infinite genertion of phis in a loop.
   953   Node *region = mem->in(0);
   954   if (is_instance_field_load_with_local_phi(region)) {
   955     const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
   956     int this_index  = phase->C->get_alias_index(addr_t);
   957     int this_offset = addr_t->offset();
   958     int this_id    = addr_t->is_oopptr()->instance_id();
   959     const Type* this_type = bottom_type();
   960     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
   961       Node* phi = region->fast_out(i);
   962       if (phi->is_Phi() && phi != mem &&
   963           phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
   964         return phi;
   965       }
   966     }
   967   }
   969   return this;
   970 }
   973 // Returns true if the AliasType refers to the field that holds the
   974 // cached box array.  Currently only handles the IntegerCache case.
   975 static bool is_autobox_cache(Compile::AliasType* atp) {
   976   if (atp != NULL && atp->field() != NULL) {
   977     ciField* field = atp->field();
   978     ciSymbol* klass = field->holder()->name();
   979     if (field->name() == ciSymbol::cache_field_name() &&
   980         field->holder()->uses_default_loader() &&
   981         klass == ciSymbol::java_lang_Integer_IntegerCache()) {
   982       return true;
   983     }
   984   }
   985   return false;
   986 }
   988 // Fetch the base value in the autobox array
   989 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
   990   if (atp != NULL && atp->field() != NULL) {
   991     ciField* field = atp->field();
   992     ciSymbol* klass = field->holder()->name();
   993     if (field->name() == ciSymbol::cache_field_name() &&
   994         field->holder()->uses_default_loader() &&
   995         klass == ciSymbol::java_lang_Integer_IntegerCache()) {
   996       assert(field->is_constant(), "what?");
   997       ciObjArray* array = field->constant_value().as_object()->as_obj_array();
   998       // Fetch the box object at the base of the array and get its value
   999       ciInstance* box = array->obj_at(0)->as_instance();
  1000       ciInstanceKlass* ik = box->klass()->as_instance_klass();
  1001       if (ik->nof_nonstatic_fields() == 1) {
  1002         // This should be true nonstatic_field_at requires calling
  1003         // nof_nonstatic_fields so check it anyway
  1004         ciConstant c = box->field_value(ik->nonstatic_field_at(0));
  1005         cache_offset = c.as_int();
  1007       return true;
  1010   return false;
  1013 // Returns true if the AliasType refers to the value field of an
  1014 // autobox object.  Currently only handles Integer.
  1015 static bool is_autobox_object(Compile::AliasType* atp) {
  1016   if (atp != NULL && atp->field() != NULL) {
  1017     ciField* field = atp->field();
  1018     ciSymbol* klass = field->holder()->name();
  1019     if (field->name() == ciSymbol::value_name() &&
  1020         field->holder()->uses_default_loader() &&
  1021         klass == ciSymbol::java_lang_Integer()) {
  1022       return true;
  1025   return false;
  1029 // We're loading from an object which has autobox behaviour.
  1030 // If this object is result of a valueOf call we'll have a phi
  1031 // merging a newly allocated object and a load from the cache.
  1032 // We want to replace this load with the original incoming
  1033 // argument to the valueOf call.
  1034 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
  1035   Node* base = in(Address)->in(AddPNode::Base);
  1036   if (base->is_Phi() && base->req() == 3) {
  1037     AllocateNode* allocation = NULL;
  1038     int allocation_index = -1;
  1039     int load_index = -1;
  1040     for (uint i = 1; i < base->req(); i++) {
  1041       allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
  1042       if (allocation != NULL) {
  1043         allocation_index = i;
  1044         load_index = 3 - allocation_index;
  1045         break;
  1048     LoadNode* load = NULL;
  1049     if (allocation != NULL && base->in(load_index)->is_Load()) {
  1050       load = base->in(load_index)->as_Load();
  1052     if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
  1053       // Push the loads from the phi that comes from valueOf up
  1054       // through it to allow elimination of the loads and the recovery
  1055       // of the original value.
  1056       Node* mem_phi = in(Memory);
  1057       Node* offset = in(Address)->in(AddPNode::Offset);
  1059       Node* in1 = clone();
  1060       Node* in1_addr = in1->in(Address)->clone();
  1061       in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
  1062       in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
  1063       in1_addr->set_req(AddPNode::Offset, offset);
  1064       in1->set_req(0, base->in(allocation_index));
  1065       in1->set_req(Address, in1_addr);
  1066       in1->set_req(Memory, mem_phi->in(allocation_index));
  1068       Node* in2 = clone();
  1069       Node* in2_addr = in2->in(Address)->clone();
  1070       in2_addr->set_req(AddPNode::Base, base->in(load_index));
  1071       in2_addr->set_req(AddPNode::Address, base->in(load_index));
  1072       in2_addr->set_req(AddPNode::Offset, offset);
  1073       in2->set_req(0, base->in(load_index));
  1074       in2->set_req(Address, in2_addr);
  1075       in2->set_req(Memory, mem_phi->in(load_index));
  1077       in1_addr = phase->transform(in1_addr);
  1078       in1 =      phase->transform(in1);
  1079       in2_addr = phase->transform(in2_addr);
  1080       in2 =      phase->transform(in2);
  1082       PhiNode* result = PhiNode::make_blank(base->in(0), this);
  1083       result->set_req(allocation_index, in1);
  1084       result->set_req(load_index, in2);
  1085       return result;
  1087   } else if (base->is_Load()) {
  1088     // Eliminate the load of Integer.value for integers from the cache
  1089     // array by deriving the value from the index into the array.
  1090     // Capture the offset of the load and then reverse the computation.
  1091     Node* load_base = base->in(Address)->in(AddPNode::Base);
  1092     if (load_base != NULL) {
  1093       Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
  1094       intptr_t cache_offset;
  1095       int shift = -1;
  1096       Node* cache = NULL;
  1097       if (is_autobox_cache(atp)) {
  1098         shift  = exact_log2(type2aelembytes(T_OBJECT));
  1099         cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
  1101       if (cache != NULL && base->in(Address)->is_AddP()) {
  1102         Node* elements[4];
  1103         int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
  1104         int cache_low;
  1105         if (count > 0 && fetch_autobox_base(atp, cache_low)) {
  1106           int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
  1107           // Add up all the offsets making of the address of the load
  1108           Node* result = elements[0];
  1109           for (int i = 1; i < count; i++) {
  1110             result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
  1112           // Remove the constant offset from the address and then
  1113           // remove the scaling of the offset to recover the original index.
  1114           result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
  1115           if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
  1116             // Peel the shift off directly but wrap it in a dummy node
  1117             // since Ideal can't return existing nodes
  1118             result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
  1119           } else {
  1120             result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
  1122 #ifdef _LP64
  1123           result = new (phase->C, 2) ConvL2INode(phase->transform(result));
  1124 #endif
  1125           return result;
  1130   return NULL;
  1133 //------------------------------split_through_phi------------------------------
  1134 // Split instance field load through Phi.
  1135 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
  1136   Node* mem     = in(MemNode::Memory);
  1137   Node* address = in(MemNode::Address);
  1138   const TypePtr *addr_t = phase->type(address)->isa_ptr();
  1139   const TypeOopPtr *t_oop = addr_t->isa_oopptr();
  1141   assert(mem->is_Phi() && (t_oop != NULL) &&
  1142          t_oop->is_known_instance_field(), "invalide conditions");
  1144   Node *region = mem->in(0);
  1145   if (region == NULL) {
  1146     return NULL; // Wait stable graph
  1148   uint cnt = mem->req();
  1149   for( uint i = 1; i < cnt; i++ ) {
  1150     Node *in = mem->in(i);
  1151     if( in == NULL ) {
  1152       return NULL; // Wait stable graph
  1155   // Check for loop invariant.
  1156   if (cnt == 3) {
  1157     for( uint i = 1; i < cnt; i++ ) {
  1158       Node *in = mem->in(i);
  1159       Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
  1160       if (m == mem) {
  1161         set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
  1162         return this;
  1166   // Split through Phi (see original code in loopopts.cpp).
  1167   assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
  1169   // Do nothing here if Identity will find a value
  1170   // (to avoid infinite chain of value phis generation).
  1171   if ( !phase->eqv(this, this->Identity(phase)) )
  1172     return NULL;
  1174   // Skip the split if the region dominates some control edge of the address.
  1175   if (cnt == 3 && !MemNode::all_controls_dominate(address, region))
  1176     return NULL;
  1178   const Type* this_type = this->bottom_type();
  1179   int this_index  = phase->C->get_alias_index(addr_t);
  1180   int this_offset = addr_t->offset();
  1181   int this_iid    = addr_t->is_oopptr()->instance_id();
  1182   int wins = 0;
  1183   PhaseIterGVN *igvn = phase->is_IterGVN();
  1184   Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
  1185   for( uint i = 1; i < region->req(); i++ ) {
  1186     Node *x;
  1187     Node* the_clone = NULL;
  1188     if( region->in(i) == phase->C->top() ) {
  1189       x = phase->C->top();      // Dead path?  Use a dead data op
  1190     } else {
  1191       x = this->clone();        // Else clone up the data op
  1192       the_clone = x;            // Remember for possible deletion.
  1193       // Alter data node to use pre-phi inputs
  1194       if( this->in(0) == region ) {
  1195         x->set_req( 0, region->in(i) );
  1196       } else {
  1197         x->set_req( 0, NULL );
  1199       for( uint j = 1; j < this->req(); j++ ) {
  1200         Node *in = this->in(j);
  1201         if( in->is_Phi() && in->in(0) == region )
  1202           x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
  1205     // Check for a 'win' on some paths
  1206     const Type *t = x->Value(igvn);
  1208     bool singleton = t->singleton();
  1210     // See comments in PhaseIdealLoop::split_thru_phi().
  1211     if( singleton && t == Type::TOP ) {
  1212       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
  1215     if( singleton ) {
  1216       wins++;
  1217       x = igvn->makecon(t);
  1218     } else {
  1219       // We now call Identity to try to simplify the cloned node.
  1220       // Note that some Identity methods call phase->type(this).
  1221       // Make sure that the type array is big enough for
  1222       // our new node, even though we may throw the node away.
  1223       // (This tweaking with igvn only works because x is a new node.)
  1224       igvn->set_type(x, t);
  1225       Node *y = x->Identity(igvn);
  1226       if( y != x ) {
  1227         wins++;
  1228         x = y;
  1229       } else {
  1230         y = igvn->hash_find(x);
  1231         if( y ) {
  1232           wins++;
  1233           x = y;
  1234         } else {
  1235           // Else x is a new node we are keeping
  1236           // We do not need register_new_node_with_optimizer
  1237           // because set_type has already been called.
  1238           igvn->_worklist.push(x);
  1242     if (x != the_clone && the_clone != NULL)
  1243       igvn->remove_dead_node(the_clone);
  1244     phi->set_req(i, x);
  1246   if( wins > 0 ) {
  1247     // Record Phi
  1248     igvn->register_new_node_with_optimizer(phi);
  1249     return phi;
  1251   igvn->remove_dead_node(phi);
  1252   return NULL;
  1255 //------------------------------Ideal------------------------------------------
  1256 // If the load is from Field memory and the pointer is non-null, we can
  1257 // zero out the control input.
  1258 // If the offset is constant and the base is an object allocation,
  1259 // try to hook me up to the exact initializing store.
  1260 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  1261   Node* p = MemNode::Ideal_common(phase, can_reshape);
  1262   if (p)  return (p == NodeSentinel) ? NULL : p;
  1264   Node* ctrl    = in(MemNode::Control);
  1265   Node* address = in(MemNode::Address);
  1267   // Skip up past a SafePoint control.  Cannot do this for Stores because
  1268   // pointer stores & cardmarks must stay on the same side of a SafePoint.
  1269   if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
  1270       phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
  1271     ctrl = ctrl->in(0);
  1272     set_req(MemNode::Control,ctrl);
  1275   // Check for useless control edge in some common special cases
  1276   if (in(MemNode::Control) != NULL) {
  1277     intptr_t ignore = 0;
  1278     Node*    base   = AddPNode::Ideal_base_and_offset(address, phase, ignore);
  1279     if (base != NULL
  1280         && phase->type(base)->higher_equal(TypePtr::NOTNULL)
  1281         && all_controls_dominate(base, phase->C->start())) {
  1282       // A method-invariant, non-null address (constant or 'this' argument).
  1283       set_req(MemNode::Control, NULL);
  1287   if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
  1288     Node* base = in(Address)->in(AddPNode::Base);
  1289     if (base != NULL) {
  1290       Compile::AliasType* atp = phase->C->alias_type(adr_type());
  1291       if (is_autobox_object(atp)) {
  1292         Node* result = eliminate_autobox(phase);
  1293         if (result != NULL) return result;
  1298   Node* mem = in(MemNode::Memory);
  1299   const TypePtr *addr_t = phase->type(address)->isa_ptr();
  1301   if (addr_t != NULL) {
  1302     // try to optimize our memory input
  1303     Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
  1304     if (opt_mem != mem) {
  1305       set_req(MemNode::Memory, opt_mem);
  1306       return this;
  1308     const TypeOopPtr *t_oop = addr_t->isa_oopptr();
  1309     if (can_reshape && opt_mem->is_Phi() &&
  1310         (t_oop != NULL) && t_oop->is_known_instance_field()) {
  1311       // Split instance field load through Phi.
  1312       Node* result = split_through_phi(phase);
  1313       if (result != NULL) return result;
  1317   // Check for prior store with a different base or offset; make Load
  1318   // independent.  Skip through any number of them.  Bail out if the stores
  1319   // are in an endless dead cycle and report no progress.  This is a key
  1320   // transform for Reflection.  However, if after skipping through the Stores
  1321   // we can't then fold up against a prior store do NOT do the transform as
  1322   // this amounts to using the 'Oracle' model of aliasing.  It leaves the same
  1323   // array memory alive twice: once for the hoisted Load and again after the
  1324   // bypassed Store.  This situation only works if EVERYBODY who does
  1325   // anti-dependence work knows how to bypass.  I.e. we need all
  1326   // anti-dependence checks to ask the same Oracle.  Right now, that Oracle is
  1327   // the alias index stuff.  So instead, peek through Stores and IFF we can
  1328   // fold up, do so.
  1329   Node* prev_mem = find_previous_store(phase);
  1330   // Steps (a), (b):  Walk past independent stores to find an exact match.
  1331   if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
  1332     // (c) See if we can fold up on the spot, but don't fold up here.
  1333     // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
  1334     // just return a prior value, which is done by Identity calls.
  1335     if (can_see_stored_value(prev_mem, phase)) {
  1336       // Make ready for step (d):
  1337       set_req(MemNode::Memory, prev_mem);
  1338       return this;
  1342   return NULL;                  // No further progress
  1345 // Helper to recognize certain Klass fields which are invariant across
  1346 // some group of array types (e.g., int[] or all T[] where T < Object).
  1347 const Type*
  1348 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
  1349                                  ciKlass* klass) const {
  1350   if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
  1351     // The field is Klass::_modifier_flags.  Return its (constant) value.
  1352     // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
  1353     assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
  1354     return TypeInt::make(klass->modifier_flags());
  1356   if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
  1357     // The field is Klass::_access_flags.  Return its (constant) value.
  1358     // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
  1359     assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
  1360     return TypeInt::make(klass->access_flags());
  1362   if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
  1363     // The field is Klass::_layout_helper.  Return its constant value if known.
  1364     assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
  1365     return TypeInt::make(klass->layout_helper());
  1368   // No match.
  1369   return NULL;
  1372 //------------------------------Value-----------------------------------------
  1373 const Type *LoadNode::Value( PhaseTransform *phase ) const {
  1374   // Either input is TOP ==> the result is TOP
  1375   Node* mem = in(MemNode::Memory);
  1376   const Type *t1 = phase->type(mem);
  1377   if (t1 == Type::TOP)  return Type::TOP;
  1378   Node* adr = in(MemNode::Address);
  1379   const TypePtr* tp = phase->type(adr)->isa_ptr();
  1380   if (tp == NULL || tp->empty())  return Type::TOP;
  1381   int off = tp->offset();
  1382   assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
  1384   // Try to guess loaded type from pointer type
  1385   if (tp->base() == Type::AryPtr) {
  1386     const Type *t = tp->is_aryptr()->elem();
  1387     // Don't do this for integer types. There is only potential profit if
  1388     // the element type t is lower than _type; that is, for int types, if _type is
  1389     // more restrictive than t.  This only happens here if one is short and the other
  1390     // char (both 16 bits), and in those cases we've made an intentional decision
  1391     // to use one kind of load over the other. See AndINode::Ideal and 4965907.
  1392     // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
  1393     //
  1394     // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
  1395     // where the _gvn.type of the AddP is wider than 8.  This occurs when an earlier
  1396     // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
  1397     // subsumed by p1.  If p1 is on the worklist but has not yet been re-transformed,
  1398     // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
  1399     // In fact, that could have been the original type of p1, and p1 could have
  1400     // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
  1401     // expression (LShiftL quux 3) independently optimized to the constant 8.
  1402     if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
  1403         && Opcode() != Op_LoadKlass) {
  1404       // t might actually be lower than _type, if _type is a unique
  1405       // concrete subclass of abstract class t.
  1406       // Make sure the reference is not into the header, by comparing
  1407       // the offset against the offset of the start of the array's data.
  1408       // Different array types begin at slightly different offsets (12 vs. 16).
  1409       // We choose T_BYTE as an example base type that is least restrictive
  1410       // as to alignment, which will therefore produce the smallest
  1411       // possible base offset.
  1412       const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
  1413       if ((uint)off >= (uint)min_base_off) {  // is the offset beyond the header?
  1414         const Type* jt = t->join(_type);
  1415         // In any case, do not allow the join, per se, to empty out the type.
  1416         if (jt->empty() && !t->empty()) {
  1417           // This can happen if a interface-typed array narrows to a class type.
  1418           jt = _type;
  1421         if (EliminateAutoBox) {
  1422           // The pointers in the autobox arrays are always non-null
  1423           Node* base = in(Address)->in(AddPNode::Base);
  1424           if (base != NULL) {
  1425             Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
  1426             if (is_autobox_cache(atp)) {
  1427               return jt->join(TypePtr::NOTNULL)->is_ptr();
  1431         return jt;
  1434   } else if (tp->base() == Type::InstPtr) {
  1435     assert( off != Type::OffsetBot ||
  1436             // arrays can be cast to Objects
  1437             tp->is_oopptr()->klass()->is_java_lang_Object() ||
  1438             // unsafe field access may not have a constant offset
  1439             phase->C->has_unsafe_access(),
  1440             "Field accesses must be precise" );
  1441     // For oop loads, we expect the _type to be precise
  1442   } else if (tp->base() == Type::KlassPtr) {
  1443     assert( off != Type::OffsetBot ||
  1444             // arrays can be cast to Objects
  1445             tp->is_klassptr()->klass()->is_java_lang_Object() ||
  1446             // also allow array-loading from the primary supertype
  1447             // array during subtype checks
  1448             Opcode() == Op_LoadKlass,
  1449             "Field accesses must be precise" );
  1450     // For klass/static loads, we expect the _type to be precise
  1453   const TypeKlassPtr *tkls = tp->isa_klassptr();
  1454   if (tkls != NULL && !StressReflectiveCode) {
  1455     ciKlass* klass = tkls->klass();
  1456     if (klass->is_loaded() && tkls->klass_is_exact()) {
  1457       // We are loading a field from a Klass metaobject whose identity
  1458       // is known at compile time (the type is "exact" or "precise").
  1459       // Check for fields we know are maintained as constants by the VM.
  1460       if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
  1461         // The field is Klass::_super_check_offset.  Return its (constant) value.
  1462         // (Folds up type checking code.)
  1463         assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
  1464         return TypeInt::make(klass->super_check_offset());
  1466       // Compute index into primary_supers array
  1467       juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
  1468       // Check for overflowing; use unsigned compare to handle the negative case.
  1469       if( depth < ciKlass::primary_super_limit() ) {
  1470         // The field is an element of Klass::_primary_supers.  Return its (constant) value.
  1471         // (Folds up type checking code.)
  1472         assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
  1473         ciKlass *ss = klass->super_of_depth(depth);
  1474         return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
  1476       const Type* aift = load_array_final_field(tkls, klass);
  1477       if (aift != NULL)  return aift;
  1478       if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
  1479           && klass->is_array_klass()) {
  1480         // The field is arrayKlass::_component_mirror.  Return its (constant) value.
  1481         // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
  1482         assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
  1483         return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
  1485       if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
  1486         // The field is Klass::_java_mirror.  Return its (constant) value.
  1487         // (Folds up the 2nd indirection in anObjConstant.getClass().)
  1488         assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
  1489         return TypeInstPtr::make(klass->java_mirror());
  1493     // We can still check if we are loading from the primary_supers array at a
  1494     // shallow enough depth.  Even though the klass is not exact, entries less
  1495     // than or equal to its super depth are correct.
  1496     if (klass->is_loaded() ) {
  1497       ciType *inner = klass->klass();
  1498       while( inner->is_obj_array_klass() )
  1499         inner = inner->as_obj_array_klass()->base_element_type();
  1500       if( inner->is_instance_klass() &&
  1501           !inner->as_instance_klass()->flags().is_interface() ) {
  1502         // Compute index into primary_supers array
  1503         juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
  1504         // Check for overflowing; use unsigned compare to handle the negative case.
  1505         if( depth < ciKlass::primary_super_limit() &&
  1506             depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
  1507           // The field is an element of Klass::_primary_supers.  Return its (constant) value.
  1508           // (Folds up type checking code.)
  1509           assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
  1510           ciKlass *ss = klass->super_of_depth(depth);
  1511           return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
  1516     // If the type is enough to determine that the thing is not an array,
  1517     // we can give the layout_helper a positive interval type.
  1518     // This will help short-circuit some reflective code.
  1519     if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
  1520         && !klass->is_array_klass() // not directly typed as an array
  1521         && !klass->is_interface()  // specifically not Serializable & Cloneable
  1522         && !klass->is_java_lang_Object()   // not the supertype of all T[]
  1523         ) {
  1524       // Note:  When interfaces are reliable, we can narrow the interface
  1525       // test to (klass != Serializable && klass != Cloneable).
  1526       assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
  1527       jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
  1528       // The key property of this type is that it folds up tests
  1529       // for array-ness, since it proves that the layout_helper is positive.
  1530       // Thus, a generic value like the basic object layout helper works fine.
  1531       return TypeInt::make(min_size, max_jint, Type::WidenMin);
  1535   // If we are loading from a freshly-allocated object, produce a zero,
  1536   // if the load is provably beyond the header of the object.
  1537   // (Also allow a variable load from a fresh array to produce zero.)
  1538   if (ReduceFieldZeroing) {
  1539     Node* value = can_see_stored_value(mem,phase);
  1540     if (value != NULL && value->is_Con())
  1541       return value->bottom_type();
  1544   const TypeOopPtr *tinst = tp->isa_oopptr();
  1545   if (tinst != NULL && tinst->is_known_instance_field()) {
  1546     // If we have an instance type and our memory input is the
  1547     // programs's initial memory state, there is no matching store,
  1548     // so just return a zero of the appropriate type
  1549     Node *mem = in(MemNode::Memory);
  1550     if (mem->is_Parm() && mem->in(0)->is_Start()) {
  1551       assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
  1552       return Type::get_zero_type(_type->basic_type());
  1555   return _type;
  1558 //------------------------------match_edge-------------------------------------
  1559 // Do we Match on this edge index or not?  Match only the address.
  1560 uint LoadNode::match_edge(uint idx) const {
  1561   return idx == MemNode::Address;
  1564 //--------------------------LoadBNode::Ideal--------------------------------------
  1565 //
  1566 //  If the previous store is to the same address as this load,
  1567 //  and the value stored was larger than a byte, replace this load
  1568 //  with the value stored truncated to a byte.  If no truncation is
  1569 //  needed, the replacement is done in LoadNode::Identity().
  1570 //
  1571 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  1572   Node* mem = in(MemNode::Memory);
  1573   Node* value = can_see_stored_value(mem,phase);
  1574   if( value && !phase->type(value)->higher_equal( _type ) ) {
  1575     Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
  1576     return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
  1578   // Identity call will handle the case where truncation is not needed.
  1579   return LoadNode::Ideal(phase, can_reshape);
  1582 //--------------------------LoadCNode::Ideal--------------------------------------
  1583 //
  1584 //  If the previous store is to the same address as this load,
  1585 //  and the value stored was larger than a char, replace this load
  1586 //  with the value stored truncated to a char.  If no truncation is
  1587 //  needed, the replacement is done in LoadNode::Identity().
  1588 //
  1589 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  1590   Node* mem = in(MemNode::Memory);
  1591   Node* value = can_see_stored_value(mem,phase);
  1592   if( value && !phase->type(value)->higher_equal( _type ) )
  1593     return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
  1594   // Identity call will handle the case where truncation is not needed.
  1595   return LoadNode::Ideal(phase, can_reshape);
  1598 //--------------------------LoadSNode::Ideal--------------------------------------
  1599 //
  1600 //  If the previous store is to the same address as this load,
  1601 //  and the value stored was larger than a short, replace this load
  1602 //  with the value stored truncated to a short.  If no truncation is
  1603 //  needed, the replacement is done in LoadNode::Identity().
  1604 //
  1605 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  1606   Node* mem = in(MemNode::Memory);
  1607   Node* value = can_see_stored_value(mem,phase);
  1608   if( value && !phase->type(value)->higher_equal( _type ) ) {
  1609     Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
  1610     return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
  1612   // Identity call will handle the case where truncation is not needed.
  1613   return LoadNode::Ideal(phase, can_reshape);
  1616 //=============================================================================
  1617 //----------------------------LoadKlassNode::make------------------------------
  1618 // Polymorphic factory method:
  1619 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
  1620   Compile* C = gvn.C;
  1621   Node *ctl = NULL;
  1622   // sanity check the alias category against the created node type
  1623   const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr();
  1624   assert(adr_type != NULL, "expecting TypeOopPtr");
  1625 #ifdef _LP64
  1626   if (adr_type->is_ptr_to_narrowoop()) {
  1627     Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop()));
  1628     return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr());
  1630 #endif
  1631   assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
  1632   return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk);
  1635 //------------------------------Value------------------------------------------
  1636 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
  1637   return klass_value_common(phase);
  1640 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
  1641   // Either input is TOP ==> the result is TOP
  1642   const Type *t1 = phase->type( in(MemNode::Memory) );
  1643   if (t1 == Type::TOP)  return Type::TOP;
  1644   Node *adr = in(MemNode::Address);
  1645   const Type *t2 = phase->type( adr );
  1646   if (t2 == Type::TOP)  return Type::TOP;
  1647   const TypePtr *tp = t2->is_ptr();
  1648   if (TypePtr::above_centerline(tp->ptr()) ||
  1649       tp->ptr() == TypePtr::Null)  return Type::TOP;
  1651   // Return a more precise klass, if possible
  1652   const TypeInstPtr *tinst = tp->isa_instptr();
  1653   if (tinst != NULL) {
  1654     ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
  1655     int offset = tinst->offset();
  1656     if (ik == phase->C->env()->Class_klass()
  1657         && (offset == java_lang_Class::klass_offset_in_bytes() ||
  1658             offset == java_lang_Class::array_klass_offset_in_bytes())) {
  1659       // We are loading a special hidden field from a Class mirror object,
  1660       // the field which points to the VM's Klass metaobject.
  1661       ciType* t = tinst->java_mirror_type();
  1662       // java_mirror_type returns non-null for compile-time Class constants.
  1663       if (t != NULL) {
  1664         // constant oop => constant klass
  1665         if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
  1666           return TypeKlassPtr::make(ciArrayKlass::make(t));
  1668         if (!t->is_klass()) {
  1669           // a primitive Class (e.g., int.class) has NULL for a klass field
  1670           return TypePtr::NULL_PTR;
  1672         // (Folds up the 1st indirection in aClassConstant.getModifiers().)
  1673         return TypeKlassPtr::make(t->as_klass());
  1675       // non-constant mirror, so we can't tell what's going on
  1677     if( !ik->is_loaded() )
  1678       return _type;             // Bail out if not loaded
  1679     if (offset == oopDesc::klass_offset_in_bytes()) {
  1680       if (tinst->klass_is_exact()) {
  1681         return TypeKlassPtr::make(ik);
  1683       // See if we can become precise: no subklasses and no interface
  1684       // (Note:  We need to support verified interfaces.)
  1685       if (!ik->is_interface() && !ik->has_subklass()) {
  1686         //assert(!UseExactTypes, "this code should be useless with exact types");
  1687         // Add a dependence; if any subclass added we need to recompile
  1688         if (!ik->is_final()) {
  1689           // %%% should use stronger assert_unique_concrete_subtype instead
  1690           phase->C->dependencies()->assert_leaf_type(ik);
  1692         // Return precise klass
  1693         return TypeKlassPtr::make(ik);
  1696       // Return root of possible klass
  1697       return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
  1701   // Check for loading klass from an array
  1702   const TypeAryPtr *tary = tp->isa_aryptr();
  1703   if( tary != NULL ) {
  1704     ciKlass *tary_klass = tary->klass();
  1705     if (tary_klass != NULL   // can be NULL when at BOTTOM or TOP
  1706         && tary->offset() == oopDesc::klass_offset_in_bytes()) {
  1707       if (tary->klass_is_exact()) {
  1708         return TypeKlassPtr::make(tary_klass);
  1710       ciArrayKlass *ak = tary->klass()->as_array_klass();
  1711       // If the klass is an object array, we defer the question to the
  1712       // array component klass.
  1713       if( ak->is_obj_array_klass() ) {
  1714         assert( ak->is_loaded(), "" );
  1715         ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
  1716         if( base_k->is_loaded() && base_k->is_instance_klass() ) {
  1717           ciInstanceKlass* ik = base_k->as_instance_klass();
  1718           // See if we can become precise: no subklasses and no interface
  1719           if (!ik->is_interface() && !ik->has_subklass()) {
  1720             //assert(!UseExactTypes, "this code should be useless with exact types");
  1721             // Add a dependence; if any subclass added we need to recompile
  1722             if (!ik->is_final()) {
  1723               phase->C->dependencies()->assert_leaf_type(ik);
  1725             // Return precise array klass
  1726             return TypeKlassPtr::make(ak);
  1729         return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
  1730       } else {                  // Found a type-array?
  1731         //assert(!UseExactTypes, "this code should be useless with exact types");
  1732         assert( ak->is_type_array_klass(), "" );
  1733         return TypeKlassPtr::make(ak); // These are always precise
  1738   // Check for loading klass from an array klass
  1739   const TypeKlassPtr *tkls = tp->isa_klassptr();
  1740   if (tkls != NULL && !StressReflectiveCode) {
  1741     ciKlass* klass = tkls->klass();
  1742     if( !klass->is_loaded() )
  1743       return _type;             // Bail out if not loaded
  1744     if( klass->is_obj_array_klass() &&
  1745         (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
  1746       ciKlass* elem = klass->as_obj_array_klass()->element_klass();
  1747       // // Always returning precise element type is incorrect,
  1748       // // e.g., element type could be object and array may contain strings
  1749       // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
  1751       // The array's TypeKlassPtr was declared 'precise' or 'not precise'
  1752       // according to the element type's subclassing.
  1753       return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
  1755     if( klass->is_instance_klass() && tkls->klass_is_exact() &&
  1756         (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
  1757       ciKlass* sup = klass->as_instance_klass()->super();
  1758       // The field is Klass::_super.  Return its (constant) value.
  1759       // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
  1760       return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
  1764   // Bailout case
  1765   return LoadNode::Value(phase);
  1768 //------------------------------Identity---------------------------------------
  1769 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
  1770 // Also feed through the klass in Allocate(...klass...)._klass.
  1771 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
  1772   return klass_identity_common(phase);
  1775 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
  1776   Node* x = LoadNode::Identity(phase);
  1777   if (x != this)  return x;
  1779   // Take apart the address into an oop and and offset.
  1780   // Return 'this' if we cannot.
  1781   Node*    adr    = in(MemNode::Address);
  1782   intptr_t offset = 0;
  1783   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
  1784   if (base == NULL)     return this;
  1785   const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
  1786   if (toop == NULL)     return this;
  1788   // We can fetch the klass directly through an AllocateNode.
  1789   // This works even if the klass is not constant (clone or newArray).
  1790   if (offset == oopDesc::klass_offset_in_bytes()) {
  1791     Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
  1792     if (allocated_klass != NULL) {
  1793       return allocated_klass;
  1797   // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
  1798   // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
  1799   // See inline_native_Class_query for occurrences of these patterns.
  1800   // Java Example:  x.getClass().isAssignableFrom(y)
  1801   // Java Example:  Array.newInstance(x.getClass().getComponentType(), n)
  1802   //
  1803   // This improves reflective code, often making the Class
  1804   // mirror go completely dead.  (Current exception:  Class
  1805   // mirrors may appear in debug info, but we could clean them out by
  1806   // introducing a new debug info operator for klassOop.java_mirror).
  1807   if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
  1808       && (offset == java_lang_Class::klass_offset_in_bytes() ||
  1809           offset == java_lang_Class::array_klass_offset_in_bytes())) {
  1810     // We are loading a special hidden field from a Class mirror,
  1811     // the field which points to its Klass or arrayKlass metaobject.
  1812     if (base->is_Load()) {
  1813       Node* adr2 = base->in(MemNode::Address);
  1814       const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
  1815       if (tkls != NULL && !tkls->empty()
  1816           && (tkls->klass()->is_instance_klass() ||
  1817               tkls->klass()->is_array_klass())
  1818           && adr2->is_AddP()
  1819           ) {
  1820         int mirror_field = Klass::java_mirror_offset_in_bytes();
  1821         if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
  1822           mirror_field = in_bytes(arrayKlass::component_mirror_offset());
  1824         if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
  1825           return adr2->in(AddPNode::Base);
  1831   return this;
  1835 //------------------------------Value------------------------------------------
  1836 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
  1837   const Type *t = klass_value_common(phase);
  1838   if (t == Type::TOP)
  1839     return t;
  1841   return t->make_narrowoop();
  1844 //------------------------------Identity---------------------------------------
  1845 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
  1846 // Also feed through the klass in Allocate(...klass...)._klass.
  1847 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
  1848   Node *x = klass_identity_common(phase);
  1850   const Type *t = phase->type( x );
  1851   if( t == Type::TOP ) return x;
  1852   if( t->isa_narrowoop()) return x;
  1854   return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop()));
  1857 //------------------------------Value-----------------------------------------
  1858 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
  1859   // Either input is TOP ==> the result is TOP
  1860   const Type *t1 = phase->type( in(MemNode::Memory) );
  1861   if( t1 == Type::TOP ) return Type::TOP;
  1862   Node *adr = in(MemNode::Address);
  1863   const Type *t2 = phase->type( adr );
  1864   if( t2 == Type::TOP ) return Type::TOP;
  1865   const TypePtr *tp = t2->is_ptr();
  1866   if (TypePtr::above_centerline(tp->ptr()))  return Type::TOP;
  1867   const TypeAryPtr *tap = tp->isa_aryptr();
  1868   if( !tap ) return _type;
  1869   return tap->size();
  1872 //------------------------------Identity---------------------------------------
  1873 // Feed through the length in AllocateArray(...length...)._length.
  1874 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
  1875   Node* x = LoadINode::Identity(phase);
  1876   if (x != this)  return x;
  1878   // Take apart the address into an oop and and offset.
  1879   // Return 'this' if we cannot.
  1880   Node*    adr    = in(MemNode::Address);
  1881   intptr_t offset = 0;
  1882   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
  1883   if (base == NULL)     return this;
  1884   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
  1885   if (tary == NULL)     return this;
  1887   // We can fetch the length directly through an AllocateArrayNode.
  1888   // This works even if the length is not constant (clone or newArray).
  1889   if (offset == arrayOopDesc::length_offset_in_bytes()) {
  1890     Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase);
  1891     if (allocated_length != NULL) {
  1892       return allocated_length;
  1896   return this;
  1899 //=============================================================================
  1900 //---------------------------StoreNode::make-----------------------------------
  1901 // Polymorphic factory method:
  1902 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
  1903   Compile* C = gvn.C;
  1905   switch (bt) {
  1906   case T_BOOLEAN:
  1907   case T_BYTE:    return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
  1908   case T_INT:     return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
  1909   case T_CHAR:
  1910   case T_SHORT:   return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
  1911   case T_LONG:    return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
  1912   case T_FLOAT:   return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
  1913   case T_DOUBLE:  return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
  1914   case T_ADDRESS:
  1915   case T_OBJECT:
  1916 #ifdef _LP64
  1917     if (adr->bottom_type()->is_ptr_to_narrowoop() ||
  1918         (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
  1919          adr->bottom_type()->isa_rawptr())) {
  1920       val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop()));
  1921       return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val);
  1922     } else
  1923 #endif
  1925       return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
  1928   ShouldNotReachHere();
  1929   return (StoreNode*)NULL;
  1932 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
  1933   bool require_atomic = true;
  1934   return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
  1938 //--------------------------bottom_type----------------------------------------
  1939 const Type *StoreNode::bottom_type() const {
  1940   return Type::MEMORY;
  1943 //------------------------------hash-------------------------------------------
  1944 uint StoreNode::hash() const {
  1945   // unroll addition of interesting fields
  1946   //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
  1948   // Since they are not commoned, do not hash them:
  1949   return NO_HASH;
  1952 //------------------------------Ideal------------------------------------------
  1953 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
  1954 // When a store immediately follows a relevant allocation/initialization,
  1955 // try to capture it into the initialization, or hoist it above.
  1956 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  1957   Node* p = MemNode::Ideal_common(phase, can_reshape);
  1958   if (p)  return (p == NodeSentinel) ? NULL : p;
  1960   Node* mem     = in(MemNode::Memory);
  1961   Node* address = in(MemNode::Address);
  1963   // Back-to-back stores to same address?  Fold em up.
  1964   // Generally unsafe if I have intervening uses...
  1965   if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
  1966     // Looking at a dead closed cycle of memory?
  1967     assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
  1969     assert(Opcode() == mem->Opcode() ||
  1970            phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
  1971            "no mismatched stores, except on raw memory");
  1973     if (mem->outcnt() == 1 &&           // check for intervening uses
  1974         mem->as_Store()->memory_size() <= this->memory_size()) {
  1975       // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
  1976       // For example, 'mem' might be the final state at a conditional return.
  1977       // Or, 'mem' might be used by some node which is live at the same time
  1978       // 'this' is live, which might be unschedulable.  So, require exactly
  1979       // ONE user, the 'this' store, until such time as we clone 'mem' for
  1980       // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
  1981       if (can_reshape) {  // (%%% is this an anachronism?)
  1982         set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
  1983                   phase->is_IterGVN());
  1984       } else {
  1985         // It's OK to do this in the parser, since DU info is always accurate,
  1986         // and the parser always refers to nodes via SafePointNode maps.
  1987         set_req(MemNode::Memory, mem->in(MemNode::Memory));
  1989       return this;
  1993   // Capture an unaliased, unconditional, simple store into an initializer.
  1994   // Or, if it is independent of the allocation, hoist it above the allocation.
  1995   if (ReduceFieldZeroing && /*can_reshape &&*/
  1996       mem->is_Proj() && mem->in(0)->is_Initialize()) {
  1997     InitializeNode* init = mem->in(0)->as_Initialize();
  1998     intptr_t offset = init->can_capture_store(this, phase);
  1999     if (offset > 0) {
  2000       Node* moved = init->capture_store(this, offset, phase);
  2001       // If the InitializeNode captured me, it made a raw copy of me,
  2002       // and I need to disappear.
  2003       if (moved != NULL) {
  2004         // %%% hack to ensure that Ideal returns a new node:
  2005         mem = MergeMemNode::make(phase->C, mem);
  2006         return mem;             // fold me away
  2011   return NULL;                  // No further progress
  2014 //------------------------------Value-----------------------------------------
  2015 const Type *StoreNode::Value( PhaseTransform *phase ) const {
  2016   // Either input is TOP ==> the result is TOP
  2017   const Type *t1 = phase->type( in(MemNode::Memory) );
  2018   if( t1 == Type::TOP ) return Type::TOP;
  2019   const Type *t2 = phase->type( in(MemNode::Address) );
  2020   if( t2 == Type::TOP ) return Type::TOP;
  2021   const Type *t3 = phase->type( in(MemNode::ValueIn) );
  2022   if( t3 == Type::TOP ) return Type::TOP;
  2023   return Type::MEMORY;
  2026 //------------------------------Identity---------------------------------------
  2027 // Remove redundant stores:
  2028 //   Store(m, p, Load(m, p)) changes to m.
  2029 //   Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
  2030 Node *StoreNode::Identity( PhaseTransform *phase ) {
  2031   Node* mem = in(MemNode::Memory);
  2032   Node* adr = in(MemNode::Address);
  2033   Node* val = in(MemNode::ValueIn);
  2035   // Load then Store?  Then the Store is useless
  2036   if (val->is_Load() &&
  2037       phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
  2038       phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
  2039       val->as_Load()->store_Opcode() == Opcode()) {
  2040     return mem;
  2043   // Two stores in a row of the same value?
  2044   if (mem->is_Store() &&
  2045       phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
  2046       phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
  2047       mem->Opcode() == Opcode()) {
  2048     return mem;
  2051   // Store of zero anywhere into a freshly-allocated object?
  2052   // Then the store is useless.
  2053   // (It must already have been captured by the InitializeNode.)
  2054   if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
  2055     // a newly allocated object is already all-zeroes everywhere
  2056     if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
  2057       return mem;
  2060     // the store may also apply to zero-bits in an earlier object
  2061     Node* prev_mem = find_previous_store(phase);
  2062     // Steps (a), (b):  Walk past independent stores to find an exact match.
  2063     if (prev_mem != NULL) {
  2064       Node* prev_val = can_see_stored_value(prev_mem, phase);
  2065       if (prev_val != NULL && phase->eqv(prev_val, val)) {
  2066         // prev_val and val might differ by a cast; it would be good
  2067         // to keep the more informative of the two.
  2068         return mem;
  2073   return this;
  2076 //------------------------------match_edge-------------------------------------
  2077 // Do we Match on this edge index or not?  Match only memory & value
  2078 uint StoreNode::match_edge(uint idx) const {
  2079   return idx == MemNode::Address || idx == MemNode::ValueIn;
  2082 //------------------------------cmp--------------------------------------------
  2083 // Do not common stores up together.  They generally have to be split
  2084 // back up anyways, so do not bother.
  2085 uint StoreNode::cmp( const Node &n ) const {
  2086   return (&n == this);          // Always fail except on self
  2089 //------------------------------Ideal_masked_input-----------------------------
  2090 // Check for a useless mask before a partial-word store
  2091 // (StoreB ... (AndI valIn conIa) )
  2092 // If (conIa & mask == mask) this simplifies to
  2093 // (StoreB ... (valIn) )
  2094 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
  2095   Node *val = in(MemNode::ValueIn);
  2096   if( val->Opcode() == Op_AndI ) {
  2097     const TypeInt *t = phase->type( val->in(2) )->isa_int();
  2098     if( t && t->is_con() && (t->get_con() & mask) == mask ) {
  2099       set_req(MemNode::ValueIn, val->in(1));
  2100       return this;
  2103   return NULL;
  2107 //------------------------------Ideal_sign_extended_input----------------------
  2108 // Check for useless sign-extension before a partial-word store
  2109 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
  2110 // If (conIL == conIR && conIR <= num_bits)  this simplifies to
  2111 // (StoreB ... (valIn) )
  2112 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
  2113   Node *val = in(MemNode::ValueIn);
  2114   if( val->Opcode() == Op_RShiftI ) {
  2115     const TypeInt *t = phase->type( val->in(2) )->isa_int();
  2116     if( t && t->is_con() && (t->get_con() <= num_bits) ) {
  2117       Node *shl = val->in(1);
  2118       if( shl->Opcode() == Op_LShiftI ) {
  2119         const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
  2120         if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
  2121           set_req(MemNode::ValueIn, shl->in(1));
  2122           return this;
  2127   return NULL;
  2130 //------------------------------value_never_loaded-----------------------------------
  2131 // Determine whether there are any possible loads of the value stored.
  2132 // For simplicity, we actually check if there are any loads from the
  2133 // address stored to, not just for loads of the value stored by this node.
  2134 //
  2135 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
  2136   Node *adr = in(Address);
  2137   const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
  2138   if (adr_oop == NULL)
  2139     return false;
  2140   if (!adr_oop->is_known_instance_field())
  2141     return false; // if not a distinct instance, there may be aliases of the address
  2142   for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
  2143     Node *use = adr->fast_out(i);
  2144     int opc = use->Opcode();
  2145     if (use->is_Load() || use->is_LoadStore()) {
  2146       return false;
  2149   return true;
  2152 //=============================================================================
  2153 //------------------------------Ideal------------------------------------------
  2154 // If the store is from an AND mask that leaves the low bits untouched, then
  2155 // we can skip the AND operation.  If the store is from a sign-extension
  2156 // (a left shift, then right shift) we can skip both.
  2157 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
  2158   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
  2159   if( progress != NULL ) return progress;
  2161   progress = StoreNode::Ideal_sign_extended_input(phase, 24);
  2162   if( progress != NULL ) return progress;
  2164   // Finally check the default case
  2165   return StoreNode::Ideal(phase, can_reshape);
  2168 //=============================================================================
  2169 //------------------------------Ideal------------------------------------------
  2170 // If the store is from an AND mask that leaves the low bits untouched, then
  2171 // we can skip the AND operation
  2172 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
  2173   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
  2174   if( progress != NULL ) return progress;
  2176   progress = StoreNode::Ideal_sign_extended_input(phase, 16);
  2177   if( progress != NULL ) return progress;
  2179   // Finally check the default case
  2180   return StoreNode::Ideal(phase, can_reshape);
  2183 //=============================================================================
  2184 //------------------------------Identity---------------------------------------
  2185 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
  2186   // No need to card mark when storing a null ptr
  2187   Node* my_store = in(MemNode::OopStore);
  2188   if (my_store->is_Store()) {
  2189     const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
  2190     if( t1 == TypePtr::NULL_PTR ) {
  2191       return in(MemNode::Memory);
  2194   return this;
  2197 //------------------------------Value-----------------------------------------
  2198 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
  2199   // Either input is TOP ==> the result is TOP
  2200   const Type *t = phase->type( in(MemNode::Memory) );
  2201   if( t == Type::TOP ) return Type::TOP;
  2202   t = phase->type( in(MemNode::Address) );
  2203   if( t == Type::TOP ) return Type::TOP;
  2204   t = phase->type( in(MemNode::ValueIn) );
  2205   if( t == Type::TOP ) return Type::TOP;
  2206   // If extra input is TOP ==> the result is TOP
  2207   t = phase->type( in(MemNode::OopStore) );
  2208   if( t == Type::TOP ) return Type::TOP;
  2210   return StoreNode::Value( phase );
  2214 //=============================================================================
  2215 //----------------------------------SCMemProjNode------------------------------
  2216 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
  2218   return bottom_type();
  2221 //=============================================================================
  2222 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
  2223   init_req(MemNode::Control, c  );
  2224   init_req(MemNode::Memory , mem);
  2225   init_req(MemNode::Address, adr);
  2226   init_req(MemNode::ValueIn, val);
  2227   init_req(         ExpectedIn, ex );
  2228   init_class_id(Class_LoadStore);
  2232 //=============================================================================
  2233 //-------------------------------adr_type--------------------------------------
  2234 // Do we Match on this edge index or not?  Do not match memory
  2235 const TypePtr* ClearArrayNode::adr_type() const {
  2236   Node *adr = in(3);
  2237   return MemNode::calculate_adr_type(adr->bottom_type());
  2240 //------------------------------match_edge-------------------------------------
  2241 // Do we Match on this edge index or not?  Do not match memory
  2242 uint ClearArrayNode::match_edge(uint idx) const {
  2243   return idx > 1;
  2246 //------------------------------Identity---------------------------------------
  2247 // Clearing a zero length array does nothing
  2248 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
  2249   return phase->type(in(2))->higher_equal(TypeX::ZERO)  ? in(1) : this;
  2252 //------------------------------Idealize---------------------------------------
  2253 // Clearing a short array is faster with stores
  2254 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
  2255   const int unit = BytesPerLong;
  2256   const TypeX* t = phase->type(in(2))->isa_intptr_t();
  2257   if (!t)  return NULL;
  2258   if (!t->is_con())  return NULL;
  2259   intptr_t raw_count = t->get_con();
  2260   intptr_t size = raw_count;
  2261   if (!Matcher::init_array_count_is_in_bytes) size *= unit;
  2262   // Clearing nothing uses the Identity call.
  2263   // Negative clears are possible on dead ClearArrays
  2264   // (see jck test stmt114.stmt11402.val).
  2265   if (size <= 0 || size % unit != 0)  return NULL;
  2266   intptr_t count = size / unit;
  2267   // Length too long; use fast hardware clear
  2268   if (size > Matcher::init_array_short_size)  return NULL;
  2269   Node *mem = in(1);
  2270   if( phase->type(mem)==Type::TOP ) return NULL;
  2271   Node *adr = in(3);
  2272   const Type* at = phase->type(adr);
  2273   if( at==Type::TOP ) return NULL;
  2274   const TypePtr* atp = at->isa_ptr();
  2275   // adjust atp to be the correct array element address type
  2276   if (atp == NULL)  atp = TypePtr::BOTTOM;
  2277   else              atp = atp->add_offset(Type::OffsetBot);
  2278   // Get base for derived pointer purposes
  2279   if( adr->Opcode() != Op_AddP ) Unimplemented();
  2280   Node *base = adr->in(1);
  2282   Node *zero = phase->makecon(TypeLong::ZERO);
  2283   Node *off  = phase->MakeConX(BytesPerLong);
  2284   mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
  2285   count--;
  2286   while( count-- ) {
  2287     mem = phase->transform(mem);
  2288     adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
  2289     mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
  2291   return mem;
  2294 //----------------------------clear_memory-------------------------------------
  2295 // Generate code to initialize object storage to zero.
  2296 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
  2297                                    intptr_t start_offset,
  2298                                    Node* end_offset,
  2299                                    PhaseGVN* phase) {
  2300   Compile* C = phase->C;
  2301   intptr_t offset = start_offset;
  2303   int unit = BytesPerLong;
  2304   if ((offset % unit) != 0) {
  2305     Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
  2306     adr = phase->transform(adr);
  2307     const TypePtr* atp = TypeRawPtr::BOTTOM;
  2308     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
  2309     mem = phase->transform(mem);
  2310     offset += BytesPerInt;
  2312   assert((offset % unit) == 0, "");
  2314   // Initialize the remaining stuff, if any, with a ClearArray.
  2315   return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
  2318 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
  2319                                    Node* start_offset,
  2320                                    Node* end_offset,
  2321                                    PhaseGVN* phase) {
  2322   if (start_offset == end_offset) {
  2323     // nothing to do
  2324     return mem;
  2327   Compile* C = phase->C;
  2328   int unit = BytesPerLong;
  2329   Node* zbase = start_offset;
  2330   Node* zend  = end_offset;
  2332   // Scale to the unit required by the CPU:
  2333   if (!Matcher::init_array_count_is_in_bytes) {
  2334     Node* shift = phase->intcon(exact_log2(unit));
  2335     zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
  2336     zend  = phase->transform( new(C,3) URShiftXNode(zend,  shift) );
  2339   Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
  2340   Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
  2342   // Bulk clear double-words
  2343   Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
  2344   mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
  2345   return phase->transform(mem);
  2348 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
  2349                                    intptr_t start_offset,
  2350                                    intptr_t end_offset,
  2351                                    PhaseGVN* phase) {
  2352   if (start_offset == end_offset) {
  2353     // nothing to do
  2354     return mem;
  2357   Compile* C = phase->C;
  2358   assert((end_offset % BytesPerInt) == 0, "odd end offset");
  2359   intptr_t done_offset = end_offset;
  2360   if ((done_offset % BytesPerLong) != 0) {
  2361     done_offset -= BytesPerInt;
  2363   if (done_offset > start_offset) {
  2364     mem = clear_memory(ctl, mem, dest,
  2365                        start_offset, phase->MakeConX(done_offset), phase);
  2367   if (done_offset < end_offset) { // emit the final 32-bit store
  2368     Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
  2369     adr = phase->transform(adr);
  2370     const TypePtr* atp = TypeRawPtr::BOTTOM;
  2371     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
  2372     mem = phase->transform(mem);
  2373     done_offset += BytesPerInt;
  2375   assert(done_offset == end_offset, "");
  2376   return mem;
  2379 //=============================================================================
  2380 // Do we match on this edge? No memory edges
  2381 uint StrCompNode::match_edge(uint idx) const {
  2382   return idx == 5 || idx == 6;
  2385 //------------------------------Ideal------------------------------------------
  2386 // Return a node which is more "ideal" than the current node.  Strip out
  2387 // control copies
  2388 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
  2389   return remove_dead_region(phase, can_reshape) ? this : NULL;
  2392 //------------------------------Ideal------------------------------------------
  2393 // Return a node which is more "ideal" than the current node.  Strip out
  2394 // control copies
  2395 Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){
  2396   return remove_dead_region(phase, can_reshape) ? this : NULL;
  2400 //=============================================================================
  2401 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
  2402   : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
  2403     _adr_type(C->get_adr_type(alias_idx))
  2405   init_class_id(Class_MemBar);
  2406   Node* top = C->top();
  2407   init_req(TypeFunc::I_O,top);
  2408   init_req(TypeFunc::FramePtr,top);
  2409   init_req(TypeFunc::ReturnAdr,top);
  2410   if (precedent != NULL)
  2411     init_req(TypeFunc::Parms, precedent);
  2414 //------------------------------cmp--------------------------------------------
  2415 uint MemBarNode::hash() const { return NO_HASH; }
  2416 uint MemBarNode::cmp( const Node &n ) const {
  2417   return (&n == this);          // Always fail except on self
  2420 //------------------------------make-------------------------------------------
  2421 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
  2422   int len = Precedent + (pn == NULL? 0: 1);
  2423   switch (opcode) {
  2424   case Op_MemBarAcquire:   return new(C, len) MemBarAcquireNode(C,  atp, pn);
  2425   case Op_MemBarRelease:   return new(C, len) MemBarReleaseNode(C,  atp, pn);
  2426   case Op_MemBarVolatile:  return new(C, len) MemBarVolatileNode(C, atp, pn);
  2427   case Op_MemBarCPUOrder:  return new(C, len) MemBarCPUOrderNode(C, atp, pn);
  2428   case Op_Initialize:      return new(C, len) InitializeNode(C,     atp, pn);
  2429   default:                 ShouldNotReachHere(); return NULL;
  2433 //------------------------------Ideal------------------------------------------
  2434 // Return a node which is more "ideal" than the current node.  Strip out
  2435 // control copies
  2436 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  2437   if (remove_dead_region(phase, can_reshape))  return this;
  2438   return NULL;
  2441 //------------------------------Value------------------------------------------
  2442 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
  2443   if( !in(0) ) return Type::TOP;
  2444   if( phase->type(in(0)) == Type::TOP )
  2445     return Type::TOP;
  2446   return TypeTuple::MEMBAR;
  2449 //------------------------------match------------------------------------------
  2450 // Construct projections for memory.
  2451 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
  2452   switch (proj->_con) {
  2453   case TypeFunc::Control:
  2454   case TypeFunc::Memory:
  2455     return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
  2457   ShouldNotReachHere();
  2458   return NULL;
  2461 //===========================InitializeNode====================================
  2462 // SUMMARY:
  2463 // This node acts as a memory barrier on raw memory, after some raw stores.
  2464 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
  2465 // The Initialize can 'capture' suitably constrained stores as raw inits.
  2466 // It can coalesce related raw stores into larger units (called 'tiles').
  2467 // It can avoid zeroing new storage for memory units which have raw inits.
  2468 // At macro-expansion, it is marked 'complete', and does not optimize further.
  2469 //
  2470 // EXAMPLE:
  2471 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
  2472 //   ctl = incoming control; mem* = incoming memory
  2473 // (Note:  A star * on a memory edge denotes I/O and other standard edges.)
  2474 // First allocate uninitialized memory and fill in the header:
  2475 //   alloc = (Allocate ctl mem* 16 #short[].klass ...)
  2476 //   ctl := alloc.Control; mem* := alloc.Memory*
  2477 //   rawmem = alloc.Memory; rawoop = alloc.RawAddress
  2478 // Then initialize to zero the non-header parts of the raw memory block:
  2479 //   init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
  2480 //   ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
  2481 // After the initialize node executes, the object is ready for service:
  2482 //   oop := (CheckCastPP init.Control alloc.RawAddress #short[])
  2483 // Suppose its body is immediately initialized as {1,2}:
  2484 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
  2485 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
  2486 //   mem.SLICE(#short[*]) := store2
  2487 //
  2488 // DETAILS:
  2489 // An InitializeNode collects and isolates object initialization after
  2490 // an AllocateNode and before the next possible safepoint.  As a
  2491 // memory barrier (MemBarNode), it keeps critical stores from drifting
  2492 // down past any safepoint or any publication of the allocation.
  2493 // Before this barrier, a newly-allocated object may have uninitialized bits.
  2494 // After this barrier, it may be treated as a real oop, and GC is allowed.
  2495 //
  2496 // The semantics of the InitializeNode include an implicit zeroing of
  2497 // the new object from object header to the end of the object.
  2498 // (The object header and end are determined by the AllocateNode.)
  2499 //
  2500 // Certain stores may be added as direct inputs to the InitializeNode.
  2501 // These stores must update raw memory, and they must be to addresses
  2502 // derived from the raw address produced by AllocateNode, and with
  2503 // a constant offset.  They must be ordered by increasing offset.
  2504 // The first one is at in(RawStores), the last at in(req()-1).
  2505 // Unlike most memory operations, they are not linked in a chain,
  2506 // but are displayed in parallel as users of the rawmem output of
  2507 // the allocation.
  2508 //
  2509 // (See comments in InitializeNode::capture_store, which continue
  2510 // the example given above.)
  2511 //
  2512 // When the associated Allocate is macro-expanded, the InitializeNode
  2513 // may be rewritten to optimize collected stores.  A ClearArrayNode
  2514 // may also be created at that point to represent any required zeroing.
  2515 // The InitializeNode is then marked 'complete', prohibiting further
  2516 // capturing of nearby memory operations.
  2517 //
  2518 // During macro-expansion, all captured initializations which store
  2519 // constant values of 32 bits or smaller are coalesced (if advantagous)
  2520 // into larger 'tiles' 32 or 64 bits.  This allows an object to be
  2521 // initialized in fewer memory operations.  Memory words which are
  2522 // covered by neither tiles nor non-constant stores are pre-zeroed
  2523 // by explicit stores of zero.  (The code shape happens to do all
  2524 // zeroing first, then all other stores, with both sequences occurring
  2525 // in order of ascending offsets.)
  2526 //
  2527 // Alternatively, code may be inserted between an AllocateNode and its
  2528 // InitializeNode, to perform arbitrary initialization of the new object.
  2529 // E.g., the object copying intrinsics insert complex data transfers here.
  2530 // The initialization must then be marked as 'complete' disable the
  2531 // built-in zeroing semantics and the collection of initializing stores.
  2532 //
  2533 // While an InitializeNode is incomplete, reads from the memory state
  2534 // produced by it are optimizable if they match the control edge and
  2535 // new oop address associated with the allocation/initialization.
  2536 // They return a stored value (if the offset matches) or else zero.
  2537 // A write to the memory state, if it matches control and address,
  2538 // and if it is to a constant offset, may be 'captured' by the
  2539 // InitializeNode.  It is cloned as a raw memory operation and rewired
  2540 // inside the initialization, to the raw oop produced by the allocation.
  2541 // Operations on addresses which are provably distinct (e.g., to
  2542 // other AllocateNodes) are allowed to bypass the initialization.
  2543 //
  2544 // The effect of all this is to consolidate object initialization
  2545 // (both arrays and non-arrays, both piecewise and bulk) into a
  2546 // single location, where it can be optimized as a unit.
  2547 //
  2548 // Only stores with an offset less than TrackedInitializationLimit words
  2549 // will be considered for capture by an InitializeNode.  This puts a
  2550 // reasonable limit on the complexity of optimized initializations.
  2552 //---------------------------InitializeNode------------------------------------
  2553 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
  2554   : _is_complete(false),
  2555     MemBarNode(C, adr_type, rawoop)
  2557   init_class_id(Class_Initialize);
  2559   assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
  2560   assert(in(RawAddress) == rawoop, "proper init");
  2561   // Note:  allocation() can be NULL, for secondary initialization barriers
  2564 // Since this node is not matched, it will be processed by the
  2565 // register allocator.  Declare that there are no constraints
  2566 // on the allocation of the RawAddress edge.
  2567 const RegMask &InitializeNode::in_RegMask(uint idx) const {
  2568   // This edge should be set to top, by the set_complete.  But be conservative.
  2569   if (idx == InitializeNode::RawAddress)
  2570     return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
  2571   return RegMask::Empty;
  2574 Node* InitializeNode::memory(uint alias_idx) {
  2575   Node* mem = in(Memory);
  2576   if (mem->is_MergeMem()) {
  2577     return mem->as_MergeMem()->memory_at(alias_idx);
  2578   } else {
  2579     // incoming raw memory is not split
  2580     return mem;
  2584 bool InitializeNode::is_non_zero() {
  2585   if (is_complete())  return false;
  2586   remove_extra_zeroes();
  2587   return (req() > RawStores);
  2590 void InitializeNode::set_complete(PhaseGVN* phase) {
  2591   assert(!is_complete(), "caller responsibility");
  2592   _is_complete = true;
  2594   // After this node is complete, it contains a bunch of
  2595   // raw-memory initializations.  There is no need for
  2596   // it to have anything to do with non-raw memory effects.
  2597   // Therefore, tell all non-raw users to re-optimize themselves,
  2598   // after skipping the memory effects of this initialization.
  2599   PhaseIterGVN* igvn = phase->is_IterGVN();
  2600   if (igvn)  igvn->add_users_to_worklist(this);
  2603 // convenience function
  2604 // return false if the init contains any stores already
  2605 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
  2606   InitializeNode* init = initialization();
  2607   if (init == NULL || init->is_complete())  return false;
  2608   init->remove_extra_zeroes();
  2609   // for now, if this allocation has already collected any inits, bail:
  2610   if (init->is_non_zero())  return false;
  2611   init->set_complete(phase);
  2612   return true;
  2615 void InitializeNode::remove_extra_zeroes() {
  2616   if (req() == RawStores)  return;
  2617   Node* zmem = zero_memory();
  2618   uint fill = RawStores;
  2619   for (uint i = fill; i < req(); i++) {
  2620     Node* n = in(i);
  2621     if (n->is_top() || n == zmem)  continue;  // skip
  2622     if (fill < i)  set_req(fill, n);          // compact
  2623     ++fill;
  2625   // delete any empty spaces created:
  2626   while (fill < req()) {
  2627     del_req(fill);
  2631 // Helper for remembering which stores go with which offsets.
  2632 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
  2633   if (!st->is_Store())  return -1;  // can happen to dead code via subsume_node
  2634   intptr_t offset = -1;
  2635   Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
  2636                                                phase, offset);
  2637   if (base == NULL)     return -1;  // something is dead,
  2638   if (offset < 0)       return -1;  //        dead, dead
  2639   return offset;
  2642 // Helper for proving that an initialization expression is
  2643 // "simple enough" to be folded into an object initialization.
  2644 // Attempts to prove that a store's initial value 'n' can be captured
  2645 // within the initialization without creating a vicious cycle, such as:
  2646 //     { Foo p = new Foo(); p.next = p; }
  2647 // True for constants and parameters and small combinations thereof.
  2648 bool InitializeNode::detect_init_independence(Node* n,
  2649                                               bool st_is_pinned,
  2650                                               int& count) {
  2651   if (n == NULL)      return true;   // (can this really happen?)
  2652   if (n->is_Proj())   n = n->in(0);
  2653   if (n == this)      return false;  // found a cycle
  2654   if (n->is_Con())    return true;
  2655   if (n->is_Start())  return true;   // params, etc., are OK
  2656   if (n->is_Root())   return true;   // even better
  2658   Node* ctl = n->in(0);
  2659   if (ctl != NULL && !ctl->is_top()) {
  2660     if (ctl->is_Proj())  ctl = ctl->in(0);
  2661     if (ctl == this)  return false;
  2663     // If we already know that the enclosing memory op is pinned right after
  2664     // the init, then any control flow that the store has picked up
  2665     // must have preceded the init, or else be equal to the init.
  2666     // Even after loop optimizations (which might change control edges)
  2667     // a store is never pinned *before* the availability of its inputs.
  2668     if (!MemNode::all_controls_dominate(n, this))
  2669       return false;                  // failed to prove a good control
  2673   // Check data edges for possible dependencies on 'this'.
  2674   if ((count += 1) > 20)  return false;  // complexity limit
  2675   for (uint i = 1; i < n->req(); i++) {
  2676     Node* m = n->in(i);
  2677     if (m == NULL || m == n || m->is_top())  continue;
  2678     uint first_i = n->find_edge(m);
  2679     if (i != first_i)  continue;  // process duplicate edge just once
  2680     if (!detect_init_independence(m, st_is_pinned, count)) {
  2681       return false;
  2685   return true;
  2688 // Here are all the checks a Store must pass before it can be moved into
  2689 // an initialization.  Returns zero if a check fails.
  2690 // On success, returns the (constant) offset to which the store applies,
  2691 // within the initialized memory.
  2692 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
  2693   const int FAIL = 0;
  2694   if (st->req() != MemNode::ValueIn + 1)
  2695     return FAIL;                // an inscrutable StoreNode (card mark?)
  2696   Node* ctl = st->in(MemNode::Control);
  2697   if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
  2698     return FAIL;                // must be unconditional after the initialization
  2699   Node* mem = st->in(MemNode::Memory);
  2700   if (!(mem->is_Proj() && mem->in(0) == this))
  2701     return FAIL;                // must not be preceded by other stores
  2702   Node* adr = st->in(MemNode::Address);
  2703   intptr_t offset;
  2704   AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
  2705   if (alloc == NULL)
  2706     return FAIL;                // inscrutable address
  2707   if (alloc != allocation())
  2708     return FAIL;                // wrong allocation!  (store needs to float up)
  2709   Node* val = st->in(MemNode::ValueIn);
  2710   int complexity_count = 0;
  2711   if (!detect_init_independence(val, true, complexity_count))
  2712     return FAIL;                // stored value must be 'simple enough'
  2714   return offset;                // success
  2717 // Find the captured store in(i) which corresponds to the range
  2718 // [start..start+size) in the initialized object.
  2719 // If there is one, return its index i.  If there isn't, return the
  2720 // negative of the index where it should be inserted.
  2721 // Return 0 if the queried range overlaps an initialization boundary
  2722 // or if dead code is encountered.
  2723 // If size_in_bytes is zero, do not bother with overlap checks.
  2724 int InitializeNode::captured_store_insertion_point(intptr_t start,
  2725                                                    int size_in_bytes,
  2726                                                    PhaseTransform* phase) {
  2727   const int FAIL = 0, MAX_STORE = BytesPerLong;
  2729   if (is_complete())
  2730     return FAIL;                // arraycopy got here first; punt
  2732   assert(allocation() != NULL, "must be present");
  2734   // no negatives, no header fields:
  2735   if (start < (intptr_t) allocation()->minimum_header_size())  return FAIL;
  2737   // after a certain size, we bail out on tracking all the stores:
  2738   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
  2739   if (start >= ti_limit)  return FAIL;
  2741   for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
  2742     if (i >= limit)  return -(int)i; // not found; here is where to put it
  2744     Node*    st     = in(i);
  2745     intptr_t st_off = get_store_offset(st, phase);
  2746     if (st_off < 0) {
  2747       if (st != zero_memory()) {
  2748         return FAIL;            // bail out if there is dead garbage
  2750     } else if (st_off > start) {
  2751       // ...we are done, since stores are ordered
  2752       if (st_off < start + size_in_bytes) {
  2753         return FAIL;            // the next store overlaps
  2755       return -(int)i;           // not found; here is where to put it
  2756     } else if (st_off < start) {
  2757       if (size_in_bytes != 0 &&
  2758           start < st_off + MAX_STORE &&
  2759           start < st_off + st->as_Store()->memory_size()) {
  2760         return FAIL;            // the previous store overlaps
  2762     } else {
  2763       if (size_in_bytes != 0 &&
  2764           st->as_Store()->memory_size() != size_in_bytes) {
  2765         return FAIL;            // mismatched store size
  2767       return i;
  2770     ++i;
  2774 // Look for a captured store which initializes at the offset 'start'
  2775 // with the given size.  If there is no such store, and no other
  2776 // initialization interferes, then return zero_memory (the memory
  2777 // projection of the AllocateNode).
  2778 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
  2779                                           PhaseTransform* phase) {
  2780   assert(stores_are_sane(phase), "");
  2781   int i = captured_store_insertion_point(start, size_in_bytes, phase);
  2782   if (i == 0) {
  2783     return NULL;                // something is dead
  2784   } else if (i < 0) {
  2785     return zero_memory();       // just primordial zero bits here
  2786   } else {
  2787     Node* st = in(i);           // here is the store at this position
  2788     assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
  2789     return st;
  2793 // Create, as a raw pointer, an address within my new object at 'offset'.
  2794 Node* InitializeNode::make_raw_address(intptr_t offset,
  2795                                        PhaseTransform* phase) {
  2796   Node* addr = in(RawAddress);
  2797   if (offset != 0) {
  2798     Compile* C = phase->C;
  2799     addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
  2800                                                  phase->MakeConX(offset)) );
  2802   return addr;
  2805 // Clone the given store, converting it into a raw store
  2806 // initializing a field or element of my new object.
  2807 // Caller is responsible for retiring the original store,
  2808 // with subsume_node or the like.
  2809 //
  2810 // From the example above InitializeNode::InitializeNode,
  2811 // here are the old stores to be captured:
  2812 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
  2813 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
  2814 //
  2815 // Here is the changed code; note the extra edges on init:
  2816 //   alloc = (Allocate ...)
  2817 //   rawoop = alloc.RawAddress
  2818 //   rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
  2819 //   rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
  2820 //   init = (Initialize alloc.Control alloc.Memory rawoop
  2821 //                      rawstore1 rawstore2)
  2822 //
  2823 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
  2824                                     PhaseTransform* phase) {
  2825   assert(stores_are_sane(phase), "");
  2827   if (start < 0)  return NULL;
  2828   assert(can_capture_store(st, phase) == start, "sanity");
  2830   Compile* C = phase->C;
  2831   int size_in_bytes = st->memory_size();
  2832   int i = captured_store_insertion_point(start, size_in_bytes, phase);
  2833   if (i == 0)  return NULL;     // bail out
  2834   Node* prev_mem = NULL;        // raw memory for the captured store
  2835   if (i > 0) {
  2836     prev_mem = in(i);           // there is a pre-existing store under this one
  2837     set_req(i, C->top());       // temporarily disconnect it
  2838     // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
  2839   } else {
  2840     i = -i;                     // no pre-existing store
  2841     prev_mem = zero_memory();   // a slice of the newly allocated object
  2842     if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
  2843       set_req(--i, C->top());   // reuse this edge; it has been folded away
  2844     else
  2845       ins_req(i, C->top());     // build a new edge
  2847   Node* new_st = st->clone();
  2848   new_st->set_req(MemNode::Control, in(Control));
  2849   new_st->set_req(MemNode::Memory,  prev_mem);
  2850   new_st->set_req(MemNode::Address, make_raw_address(start, phase));
  2851   new_st = phase->transform(new_st);
  2853   // At this point, new_st might have swallowed a pre-existing store
  2854   // at the same offset, or perhaps new_st might have disappeared,
  2855   // if it redundantly stored the same value (or zero to fresh memory).
  2857   // In any case, wire it in:
  2858   set_req(i, new_st);
  2860   // The caller may now kill the old guy.
  2861   DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
  2862   assert(check_st == new_st || check_st == NULL, "must be findable");
  2863   assert(!is_complete(), "");
  2864   return new_st;
  2867 static bool store_constant(jlong* tiles, int num_tiles,
  2868                            intptr_t st_off, int st_size,
  2869                            jlong con) {
  2870   if ((st_off & (st_size-1)) != 0)
  2871     return false;               // strange store offset (assume size==2**N)
  2872   address addr = (address)tiles + st_off;
  2873   assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
  2874   switch (st_size) {
  2875   case sizeof(jbyte):  *(jbyte*) addr = (jbyte) con; break;
  2876   case sizeof(jchar):  *(jchar*) addr = (jchar) con; break;
  2877   case sizeof(jint):   *(jint*)  addr = (jint)  con; break;
  2878   case sizeof(jlong):  *(jlong*) addr = (jlong) con; break;
  2879   default: return false;        // strange store size (detect size!=2**N here)
  2881   return true;                  // return success to caller
  2884 // Coalesce subword constants into int constants and possibly
  2885 // into long constants.  The goal, if the CPU permits,
  2886 // is to initialize the object with a small number of 64-bit tiles.
  2887 // Also, convert floating-point constants to bit patterns.
  2888 // Non-constants are not relevant to this pass.
  2889 //
  2890 // In terms of the running example on InitializeNode::InitializeNode
  2891 // and InitializeNode::capture_store, here is the transformation
  2892 // of rawstore1 and rawstore2 into rawstore12:
  2893 //   alloc = (Allocate ...)
  2894 //   rawoop = alloc.RawAddress
  2895 //   tile12 = 0x00010002
  2896 //   rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
  2897 //   init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
  2898 //
  2899 void
  2900 InitializeNode::coalesce_subword_stores(intptr_t header_size,
  2901                                         Node* size_in_bytes,
  2902                                         PhaseGVN* phase) {
  2903   Compile* C = phase->C;
  2905   assert(stores_are_sane(phase), "");
  2906   // Note:  After this pass, they are not completely sane,
  2907   // since there may be some overlaps.
  2909   int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
  2911   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
  2912   intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
  2913   size_limit = MIN2(size_limit, ti_limit);
  2914   size_limit = align_size_up(size_limit, BytesPerLong);
  2915   int num_tiles = size_limit / BytesPerLong;
  2917   // allocate space for the tile map:
  2918   const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
  2919   jlong  tiles_buf[small_len];
  2920   Node*  nodes_buf[small_len];
  2921   jlong  inits_buf[small_len];
  2922   jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
  2923                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
  2924   Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
  2925                   : NEW_RESOURCE_ARRAY(Node*, num_tiles));
  2926   jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
  2927                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
  2928   // tiles: exact bitwise model of all primitive constants
  2929   // nodes: last constant-storing node subsumed into the tiles model
  2930   // inits: which bytes (in each tile) are touched by any initializations
  2932   //// Pass A: Fill in the tile model with any relevant stores.
  2934   Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
  2935   Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
  2936   Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
  2937   Node* zmem = zero_memory(); // initially zero memory state
  2938   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
  2939     Node* st = in(i);
  2940     intptr_t st_off = get_store_offset(st, phase);
  2942     // Figure out the store's offset and constant value:
  2943     if (st_off < header_size)             continue; //skip (ignore header)
  2944     if (st->in(MemNode::Memory) != zmem)  continue; //skip (odd store chain)
  2945     int st_size = st->as_Store()->memory_size();
  2946     if (st_off + st_size > size_limit)    break;
  2948     // Record which bytes are touched, whether by constant or not.
  2949     if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
  2950       continue;                 // skip (strange store size)
  2952     const Type* val = phase->type(st->in(MemNode::ValueIn));
  2953     if (!val->singleton())                continue; //skip (non-con store)
  2954     BasicType type = val->basic_type();
  2956     jlong con = 0;
  2957     switch (type) {
  2958     case T_INT:    con = val->is_int()->get_con();  break;
  2959     case T_LONG:   con = val->is_long()->get_con(); break;
  2960     case T_FLOAT:  con = jint_cast(val->getf());    break;
  2961     case T_DOUBLE: con = jlong_cast(val->getd());   break;
  2962     default:                              continue; //skip (odd store type)
  2965     if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
  2966         st->Opcode() == Op_StoreL) {
  2967       continue;                 // This StoreL is already optimal.
  2970     // Store down the constant.
  2971     store_constant(tiles, num_tiles, st_off, st_size, con);
  2973     intptr_t j = st_off >> LogBytesPerLong;
  2975     if (type == T_INT && st_size == BytesPerInt
  2976         && (st_off & BytesPerInt) == BytesPerInt) {
  2977       jlong lcon = tiles[j];
  2978       if (!Matcher::isSimpleConstant64(lcon) &&
  2979           st->Opcode() == Op_StoreI) {
  2980         // This StoreI is already optimal by itself.
  2981         jint* intcon = (jint*) &tiles[j];
  2982         intcon[1] = 0;  // undo the store_constant()
  2984         // If the previous store is also optimal by itself, back up and
  2985         // undo the action of the previous loop iteration... if we can.
  2986         // But if we can't, just let the previous half take care of itself.
  2987         st = nodes[j];
  2988         st_off -= BytesPerInt;
  2989         con = intcon[0];
  2990         if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
  2991           assert(st_off >= header_size, "still ignoring header");
  2992           assert(get_store_offset(st, phase) == st_off, "must be");
  2993           assert(in(i-1) == zmem, "must be");
  2994           DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
  2995           assert(con == tcon->is_int()->get_con(), "must be");
  2996           // Undo the effects of the previous loop trip, which swallowed st:
  2997           intcon[0] = 0;        // undo store_constant()
  2998           set_req(i-1, st);     // undo set_req(i, zmem)
  2999           nodes[j] = NULL;      // undo nodes[j] = st
  3000           --old_subword;        // undo ++old_subword
  3002         continue;               // This StoreI is already optimal.
  3006     // This store is not needed.
  3007     set_req(i, zmem);
  3008     nodes[j] = st;              // record for the moment
  3009     if (st_size < BytesPerLong) // something has changed
  3010           ++old_subword;        // includes int/float, but who's counting...
  3011     else  ++old_long;
  3014   if ((old_subword + old_long) == 0)
  3015     return;                     // nothing more to do
  3017   //// Pass B: Convert any non-zero tiles into optimal constant stores.
  3018   // Be sure to insert them before overlapping non-constant stores.
  3019   // (E.g., byte[] x = { 1,2,y,4 }  =>  x[int 0] = 0x01020004, x[2]=y.)
  3020   for (int j = 0; j < num_tiles; j++) {
  3021     jlong con  = tiles[j];
  3022     jlong init = inits[j];
  3023     if (con == 0)  continue;
  3024     jint con0,  con1;           // split the constant, address-wise
  3025     jint init0, init1;          // split the init map, address-wise
  3026     { union { jlong con; jint intcon[2]; } u;
  3027       u.con = con;
  3028       con0  = u.intcon[0];
  3029       con1  = u.intcon[1];
  3030       u.con = init;
  3031       init0 = u.intcon[0];
  3032       init1 = u.intcon[1];
  3035     Node* old = nodes[j];
  3036     assert(old != NULL, "need the prior store");
  3037     intptr_t offset = (j * BytesPerLong);
  3039     bool split = !Matcher::isSimpleConstant64(con);
  3041     if (offset < header_size) {
  3042       assert(offset + BytesPerInt >= header_size, "second int counts");
  3043       assert(*(jint*)&tiles[j] == 0, "junk in header");
  3044       split = true;             // only the second word counts
  3045       // Example:  int a[] = { 42 ... }
  3046     } else if (con0 == 0 && init0 == -1) {
  3047       split = true;             // first word is covered by full inits
  3048       // Example:  int a[] = { ... foo(), 42 ... }
  3049     } else if (con1 == 0 && init1 == -1) {
  3050       split = true;             // second word is covered by full inits
  3051       // Example:  int a[] = { ... 42, foo() ... }
  3054     // Here's a case where init0 is neither 0 nor -1:
  3055     //   byte a[] = { ... 0,0,foo(),0,  0,0,0,42 ... }
  3056     // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
  3057     // In this case the tile is not split; it is (jlong)42.
  3058     // The big tile is stored down, and then the foo() value is inserted.
  3059     // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
  3061     Node* ctl = old->in(MemNode::Control);
  3062     Node* adr = make_raw_address(offset, phase);
  3063     const TypePtr* atp = TypeRawPtr::BOTTOM;
  3065     // One or two coalesced stores to plop down.
  3066     Node*    st[2];
  3067     intptr_t off[2];
  3068     int  nst = 0;
  3069     if (!split) {
  3070       ++new_long;
  3071       off[nst] = offset;
  3072       st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
  3073                                   phase->longcon(con), T_LONG);
  3074     } else {
  3075       // Omit either if it is a zero.
  3076       if (con0 != 0) {
  3077         ++new_int;
  3078         off[nst]  = offset;
  3079         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
  3080                                     phase->intcon(con0), T_INT);
  3082       if (con1 != 0) {
  3083         ++new_int;
  3084         offset += BytesPerInt;
  3085         adr = make_raw_address(offset, phase);
  3086         off[nst]  = offset;
  3087         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
  3088                                     phase->intcon(con1), T_INT);
  3092     // Insert second store first, then the first before the second.
  3093     // Insert each one just before any overlapping non-constant stores.
  3094     while (nst > 0) {
  3095       Node* st1 = st[--nst];
  3096       C->copy_node_notes_to(st1, old);
  3097       st1 = phase->transform(st1);
  3098       offset = off[nst];
  3099       assert(offset >= header_size, "do not smash header");
  3100       int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
  3101       guarantee(ins_idx != 0, "must re-insert constant store");
  3102       if (ins_idx < 0)  ins_idx = -ins_idx;  // never overlap
  3103       if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
  3104         set_req(--ins_idx, st1);
  3105       else
  3106         ins_req(ins_idx, st1);
  3110   if (PrintCompilation && WizardMode)
  3111     tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
  3112                   old_subword, old_long, new_int, new_long);
  3113   if (C->log() != NULL)
  3114     C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
  3115                    old_subword, old_long, new_int, new_long);
  3117   // Clean up any remaining occurrences of zmem:
  3118   remove_extra_zeroes();
  3121 // Explore forward from in(start) to find the first fully initialized
  3122 // word, and return its offset.  Skip groups of subword stores which
  3123 // together initialize full words.  If in(start) is itself part of a
  3124 // fully initialized word, return the offset of in(start).  If there
  3125 // are no following full-word stores, or if something is fishy, return
  3126 // a negative value.
  3127 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
  3128   int       int_map = 0;
  3129   intptr_t  int_map_off = 0;
  3130   const int FULL_MAP = right_n_bits(BytesPerInt);  // the int_map we hope for
  3132   for (uint i = start, limit = req(); i < limit; i++) {
  3133     Node* st = in(i);
  3135     intptr_t st_off = get_store_offset(st, phase);
  3136     if (st_off < 0)  break;  // return conservative answer
  3138     int st_size = st->as_Store()->memory_size();
  3139     if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
  3140       return st_off;            // we found a complete word init
  3143     // update the map:
  3145     intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
  3146     if (this_int_off != int_map_off) {
  3147       // reset the map:
  3148       int_map = 0;
  3149       int_map_off = this_int_off;
  3152     int subword_off = st_off - this_int_off;
  3153     int_map |= right_n_bits(st_size) << subword_off;
  3154     if ((int_map & FULL_MAP) == FULL_MAP) {
  3155       return this_int_off;      // we found a complete word init
  3158     // Did this store hit or cross the word boundary?
  3159     intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
  3160     if (next_int_off == this_int_off + BytesPerInt) {
  3161       // We passed the current int, without fully initializing it.
  3162       int_map_off = next_int_off;
  3163       int_map >>= BytesPerInt;
  3164     } else if (next_int_off > this_int_off + BytesPerInt) {
  3165       // We passed the current and next int.
  3166       return this_int_off + BytesPerInt;
  3170   return -1;
  3174 // Called when the associated AllocateNode is expanded into CFG.
  3175 // At this point, we may perform additional optimizations.
  3176 // Linearize the stores by ascending offset, to make memory
  3177 // activity as coherent as possible.
  3178 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
  3179                                       intptr_t header_size,
  3180                                       Node* size_in_bytes,
  3181                                       PhaseGVN* phase) {
  3182   assert(!is_complete(), "not already complete");
  3183   assert(stores_are_sane(phase), "");
  3184   assert(allocation() != NULL, "must be present");
  3186   remove_extra_zeroes();
  3188   if (ReduceFieldZeroing || ReduceBulkZeroing)
  3189     // reduce instruction count for common initialization patterns
  3190     coalesce_subword_stores(header_size, size_in_bytes, phase);
  3192   Node* zmem = zero_memory();   // initially zero memory state
  3193   Node* inits = zmem;           // accumulating a linearized chain of inits
  3194   #ifdef ASSERT
  3195   intptr_t first_offset = allocation()->minimum_header_size();
  3196   intptr_t last_init_off = first_offset;  // previous init offset
  3197   intptr_t last_init_end = first_offset;  // previous init offset+size
  3198   intptr_t last_tile_end = first_offset;  // previous tile offset+size
  3199   #endif
  3200   intptr_t zeroes_done = header_size;
  3202   bool do_zeroing = true;       // we might give up if inits are very sparse
  3203   int  big_init_gaps = 0;       // how many large gaps have we seen?
  3205   if (ZeroTLAB)  do_zeroing = false;
  3206   if (!ReduceFieldZeroing && !ReduceBulkZeroing)  do_zeroing = false;
  3208   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
  3209     Node* st = in(i);
  3210     intptr_t st_off = get_store_offset(st, phase);
  3211     if (st_off < 0)
  3212       break;                    // unknown junk in the inits
  3213     if (st->in(MemNode::Memory) != zmem)
  3214       break;                    // complicated store chains somehow in list
  3216     int st_size = st->as_Store()->memory_size();
  3217     intptr_t next_init_off = st_off + st_size;
  3219     if (do_zeroing && zeroes_done < next_init_off) {
  3220       // See if this store needs a zero before it or under it.
  3221       intptr_t zeroes_needed = st_off;
  3223       if (st_size < BytesPerInt) {
  3224         // Look for subword stores which only partially initialize words.
  3225         // If we find some, we must lay down some word-level zeroes first,
  3226         // underneath the subword stores.
  3227         //
  3228         // Examples:
  3229         //   byte[] a = { p,q,r,s }  =>  a[0]=p,a[1]=q,a[2]=r,a[3]=s
  3230         //   byte[] a = { x,y,0,0 }  =>  a[0..3] = 0, a[0]=x,a[1]=y
  3231         //   byte[] a = { 0,0,z,0 }  =>  a[0..3] = 0, a[2]=z
  3232         //
  3233         // Note:  coalesce_subword_stores may have already done this,
  3234         // if it was prompted by constant non-zero subword initializers.
  3235         // But this case can still arise with non-constant stores.
  3237         intptr_t next_full_store = find_next_fullword_store(i, phase);
  3239         // In the examples above:
  3240         //   in(i)          p   q   r   s     x   y     z
  3241         //   st_off        12  13  14  15    12  13    14
  3242         //   st_size        1   1   1   1     1   1     1
  3243         //   next_full_s.  12  16  16  16    16  16    16
  3244         //   z's_done      12  16  16  16    12  16    12
  3245         //   z's_needed    12  16  16  16    16  16    16
  3246         //   zsize          0   0   0   0     4   0     4
  3247         if (next_full_store < 0) {
  3248           // Conservative tack:  Zero to end of current word.
  3249           zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
  3250         } else {
  3251           // Zero to beginning of next fully initialized word.
  3252           // Or, don't zero at all, if we are already in that word.
  3253           assert(next_full_store >= zeroes_needed, "must go forward");
  3254           assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
  3255           zeroes_needed = next_full_store;
  3259       if (zeroes_needed > zeroes_done) {
  3260         intptr_t zsize = zeroes_needed - zeroes_done;
  3261         // Do some incremental zeroing on rawmem, in parallel with inits.
  3262         zeroes_done = align_size_down(zeroes_done, BytesPerInt);
  3263         rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
  3264                                               zeroes_done, zeroes_needed,
  3265                                               phase);
  3266         zeroes_done = zeroes_needed;
  3267         if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
  3268           do_zeroing = false;   // leave the hole, next time
  3272     // Collect the store and move on:
  3273     st->set_req(MemNode::Memory, inits);
  3274     inits = st;                 // put it on the linearized chain
  3275     set_req(i, zmem);           // unhook from previous position
  3277     if (zeroes_done == st_off)
  3278       zeroes_done = next_init_off;
  3280     assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
  3282     #ifdef ASSERT
  3283     // Various order invariants.  Weaker than stores_are_sane because
  3284     // a large constant tile can be filled in by smaller non-constant stores.
  3285     assert(st_off >= last_init_off, "inits do not reverse");
  3286     last_init_off = st_off;
  3287     const Type* val = NULL;
  3288     if (st_size >= BytesPerInt &&
  3289         (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
  3290         (int)val->basic_type() < (int)T_OBJECT) {
  3291       assert(st_off >= last_tile_end, "tiles do not overlap");
  3292       assert(st_off >= last_init_end, "tiles do not overwrite inits");
  3293       last_tile_end = MAX2(last_tile_end, next_init_off);
  3294     } else {
  3295       intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
  3296       assert(st_tile_end >= last_tile_end, "inits stay with tiles");
  3297       assert(st_off      >= last_init_end, "inits do not overlap");
  3298       last_init_end = next_init_off;  // it's a non-tile
  3300     #endif //ASSERT
  3303   remove_extra_zeroes();        // clear out all the zmems left over
  3304   add_req(inits);
  3306   if (!ZeroTLAB) {
  3307     // If anything remains to be zeroed, zero it all now.
  3308     zeroes_done = align_size_down(zeroes_done, BytesPerInt);
  3309     // if it is the last unused 4 bytes of an instance, forget about it
  3310     intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
  3311     if (zeroes_done + BytesPerLong >= size_limit) {
  3312       assert(allocation() != NULL, "");
  3313       Node* klass_node = allocation()->in(AllocateNode::KlassNode);
  3314       ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
  3315       if (zeroes_done == k->layout_helper())
  3316         zeroes_done = size_limit;
  3318     if (zeroes_done < size_limit) {
  3319       rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
  3320                                             zeroes_done, size_in_bytes, phase);
  3324   set_complete(phase);
  3325   return rawmem;
  3329 #ifdef ASSERT
  3330 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
  3331   if (is_complete())
  3332     return true;                // stores could be anything at this point
  3333   assert(allocation() != NULL, "must be present");
  3334   intptr_t last_off = allocation()->minimum_header_size();
  3335   for (uint i = InitializeNode::RawStores; i < req(); i++) {
  3336     Node* st = in(i);
  3337     intptr_t st_off = get_store_offset(st, phase);
  3338     if (st_off < 0)  continue;  // ignore dead garbage
  3339     if (last_off > st_off) {
  3340       tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
  3341       this->dump(2);
  3342       assert(false, "ascending store offsets");
  3343       return false;
  3345     last_off = st_off + st->as_Store()->memory_size();
  3347   return true;
  3349 #endif //ASSERT
  3354 //============================MergeMemNode=====================================
  3355 //
  3356 // SEMANTICS OF MEMORY MERGES:  A MergeMem is a memory state assembled from several
  3357 // contributing store or call operations.  Each contributor provides the memory
  3358 // state for a particular "alias type" (see Compile::alias_type).  For example,
  3359 // if a MergeMem has an input X for alias category #6, then any memory reference
  3360 // to alias category #6 may use X as its memory state input, as an exact equivalent
  3361 // to using the MergeMem as a whole.
  3362 //   Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
  3363 //
  3364 // (Here, the <N> notation gives the index of the relevant adr_type.)
  3365 //
  3366 // In one special case (and more cases in the future), alias categories overlap.
  3367 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
  3368 // states.  Therefore, if a MergeMem has only one contributing input W for Bot,
  3369 // it is exactly equivalent to that state W:
  3370 //   MergeMem(<Bot>: W) <==> W
  3371 //
  3372 // Usually, the merge has more than one input.  In that case, where inputs
  3373 // overlap (i.e., one is Bot), the narrower alias type determines the memory
  3374 // state for that type, and the wider alias type (Bot) fills in everywhere else:
  3375 //   Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
  3376 //   Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
  3377 //
  3378 // A merge can take a "wide" memory state as one of its narrow inputs.
  3379 // This simply means that the merge observes out only the relevant parts of
  3380 // the wide input.  That is, wide memory states arriving at narrow merge inputs
  3381 // are implicitly "filtered" or "sliced" as necessary.  (This is rare.)
  3382 //
  3383 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
  3384 // and that memory slices "leak through":
  3385 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
  3386 //
  3387 // But, in such a cascade, repeated memory slices can "block the leak":
  3388 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
  3389 //
  3390 // In the last example, Y is not part of the combined memory state of the
  3391 // outermost MergeMem.  The system must, of course, prevent unschedulable
  3392 // memory states from arising, so you can be sure that the state Y is somehow
  3393 // a precursor to state Y'.
  3394 //
  3395 //
  3396 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
  3397 // of each MergeMemNode array are exactly the numerical alias indexes, including
  3398 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw.  The functions
  3399 // Compile::alias_type (and kin) produce and manage these indexes.
  3400 //
  3401 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
  3402 // (Note that this provides quick access to the top node inside MergeMem methods,
  3403 // without the need to reach out via TLS to Compile::current.)
  3404 //
  3405 // As a consequence of what was just described, a MergeMem that represents a full
  3406 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
  3407 // containing all alias categories.
  3408 //
  3409 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
  3410 //
  3411 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
  3412 // a memory state for the alias type <N>, or else the top node, meaning that
  3413 // there is no particular input for that alias type.  Note that the length of
  3414 // a MergeMem is variable, and may be extended at any time to accommodate new
  3415 // memory states at larger alias indexes.  When merges grow, they are of course
  3416 // filled with "top" in the unused in() positions.
  3417 //
  3418 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
  3419 // (Top was chosen because it works smoothly with passes like GCM.)
  3420 //
  3421 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM.  (It is
  3422 // the type of random VM bits like TLS references.)  Since it is always the
  3423 // first non-Bot memory slice, some low-level loops use it to initialize an
  3424 // index variable:  for (i = AliasIdxRaw; i < req(); i++).
  3425 //
  3426 //
  3427 // ACCESSORS:  There is a special accessor MergeMemNode::base_memory which returns
  3428 // the distinguished "wide" state.  The accessor MergeMemNode::memory_at(N) returns
  3429 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
  3430 // it returns the base memory.  To prevent bugs, memory_at does not accept <Top>
  3431 // or <Bot> indexes.  The iterator MergeMemStream provides robust iteration over
  3432 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
  3433 //
  3434 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
  3435 // really that different from the other memory inputs.  An abbreviation called
  3436 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
  3437 //
  3438 //
  3439 // PARTIAL MEMORY STATES:  During optimization, MergeMem nodes may arise that represent
  3440 // partial memory states.  When a Phi splits through a MergeMem, the copy of the Phi
  3441 // that "emerges though" the base memory will be marked as excluding the alias types
  3442 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
  3443 //
  3444 //   Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
  3445 //     ==Ideal=>  MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
  3446 //
  3447 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
  3448 // (It is currently unimplemented.)  As you can see, the resulting merge is
  3449 // actually a disjoint union of memory states, rather than an overlay.
  3450 //
  3452 //------------------------------MergeMemNode-----------------------------------
  3453 Node* MergeMemNode::make_empty_memory() {
  3454   Node* empty_memory = (Node*) Compile::current()->top();
  3455   assert(empty_memory->is_top(), "correct sentinel identity");
  3456   return empty_memory;
  3459 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
  3460   init_class_id(Class_MergeMem);
  3461   // all inputs are nullified in Node::Node(int)
  3462   // set_input(0, NULL);  // no control input
  3464   // Initialize the edges uniformly to top, for starters.
  3465   Node* empty_mem = make_empty_memory();
  3466   for (uint i = Compile::AliasIdxTop; i < req(); i++) {
  3467     init_req(i,empty_mem);
  3469   assert(empty_memory() == empty_mem, "");
  3471   if( new_base != NULL && new_base->is_MergeMem() ) {
  3472     MergeMemNode* mdef = new_base->as_MergeMem();
  3473     assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
  3474     for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
  3475       mms.set_memory(mms.memory2());
  3477     assert(base_memory() == mdef->base_memory(), "");
  3478   } else {
  3479     set_base_memory(new_base);
  3483 // Make a new, untransformed MergeMem with the same base as 'mem'.
  3484 // If mem is itself a MergeMem, populate the result with the same edges.
  3485 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
  3486   return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
  3489 //------------------------------cmp--------------------------------------------
  3490 uint MergeMemNode::hash() const { return NO_HASH; }
  3491 uint MergeMemNode::cmp( const Node &n ) const {
  3492   return (&n == this);          // Always fail except on self
  3495 //------------------------------Identity---------------------------------------
  3496 Node* MergeMemNode::Identity(PhaseTransform *phase) {
  3497   // Identity if this merge point does not record any interesting memory
  3498   // disambiguations.
  3499   Node* base_mem = base_memory();
  3500   Node* empty_mem = empty_memory();
  3501   if (base_mem != empty_mem) {  // Memory path is not dead?
  3502     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
  3503       Node* mem = in(i);
  3504       if (mem != empty_mem && mem != base_mem) {
  3505         return this;            // Many memory splits; no change
  3509   return base_mem;              // No memory splits; ID on the one true input
  3512 //------------------------------Ideal------------------------------------------
  3513 // This method is invoked recursively on chains of MergeMem nodes
  3514 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  3515   // Remove chain'd MergeMems
  3516   //
  3517   // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
  3518   // relative to the "in(Bot)".  Since we are patching both at the same time,
  3519   // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
  3520   // but rewrite each "in(i)" relative to the new "in(Bot)".
  3521   Node *progress = NULL;
  3524   Node* old_base = base_memory();
  3525   Node* empty_mem = empty_memory();
  3526   if (old_base == empty_mem)
  3527     return NULL; // Dead memory path.
  3529   MergeMemNode* old_mbase;
  3530   if (old_base != NULL && old_base->is_MergeMem())
  3531     old_mbase = old_base->as_MergeMem();
  3532   else
  3533     old_mbase = NULL;
  3534   Node* new_base = old_base;
  3536   // simplify stacked MergeMems in base memory
  3537   if (old_mbase)  new_base = old_mbase->base_memory();
  3539   // the base memory might contribute new slices beyond my req()
  3540   if (old_mbase)  grow_to_match(old_mbase);
  3542   // Look carefully at the base node if it is a phi.
  3543   PhiNode* phi_base;
  3544   if (new_base != NULL && new_base->is_Phi())
  3545     phi_base = new_base->as_Phi();
  3546   else
  3547     phi_base = NULL;
  3549   Node*    phi_reg = NULL;
  3550   uint     phi_len = (uint)-1;
  3551   if (phi_base != NULL && !phi_base->is_copy()) {
  3552     // do not examine phi if degraded to a copy
  3553     phi_reg = phi_base->region();
  3554     phi_len = phi_base->req();
  3555     // see if the phi is unfinished
  3556     for (uint i = 1; i < phi_len; i++) {
  3557       if (phi_base->in(i) == NULL) {
  3558         // incomplete phi; do not look at it yet!
  3559         phi_reg = NULL;
  3560         phi_len = (uint)-1;
  3561         break;
  3566   // Note:  We do not call verify_sparse on entry, because inputs
  3567   // can normalize to the base_memory via subsume_node or similar
  3568   // mechanisms.  This method repairs that damage.
  3570   assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
  3572   // Look at each slice.
  3573   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
  3574     Node* old_in = in(i);
  3575     // calculate the old memory value
  3576     Node* old_mem = old_in;
  3577     if (old_mem == empty_mem)  old_mem = old_base;
  3578     assert(old_mem == memory_at(i), "");
  3580     // maybe update (reslice) the old memory value
  3582     // simplify stacked MergeMems
  3583     Node* new_mem = old_mem;
  3584     MergeMemNode* old_mmem;
  3585     if (old_mem != NULL && old_mem->is_MergeMem())
  3586       old_mmem = old_mem->as_MergeMem();
  3587     else
  3588       old_mmem = NULL;
  3589     if (old_mmem == this) {
  3590       // This can happen if loops break up and safepoints disappear.
  3591       // A merge of BotPtr (default) with a RawPtr memory derived from a
  3592       // safepoint can be rewritten to a merge of the same BotPtr with
  3593       // the BotPtr phi coming into the loop.  If that phi disappears
  3594       // also, we can end up with a self-loop of the mergemem.
  3595       // In general, if loops degenerate and memory effects disappear,
  3596       // a mergemem can be left looking at itself.  This simply means
  3597       // that the mergemem's default should be used, since there is
  3598       // no longer any apparent effect on this slice.
  3599       // Note: If a memory slice is a MergeMem cycle, it is unreachable
  3600       //       from start.  Update the input to TOP.
  3601       new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
  3603     else if (old_mmem != NULL) {
  3604       new_mem = old_mmem->memory_at(i);
  3606     // else preceeding memory was not a MergeMem
  3608     // replace equivalent phis (unfortunately, they do not GVN together)
  3609     if (new_mem != NULL && new_mem != new_base &&
  3610         new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
  3611       if (new_mem->is_Phi()) {
  3612         PhiNode* phi_mem = new_mem->as_Phi();
  3613         for (uint i = 1; i < phi_len; i++) {
  3614           if (phi_base->in(i) != phi_mem->in(i)) {
  3615             phi_mem = NULL;
  3616             break;
  3619         if (phi_mem != NULL) {
  3620           // equivalent phi nodes; revert to the def
  3621           new_mem = new_base;
  3626     // maybe store down a new value
  3627     Node* new_in = new_mem;
  3628     if (new_in == new_base)  new_in = empty_mem;
  3630     if (new_in != old_in) {
  3631       // Warning:  Do not combine this "if" with the previous "if"
  3632       // A memory slice might have be be rewritten even if it is semantically
  3633       // unchanged, if the base_memory value has changed.
  3634       set_req(i, new_in);
  3635       progress = this;          // Report progress
  3639   if (new_base != old_base) {
  3640     set_req(Compile::AliasIdxBot, new_base);
  3641     // Don't use set_base_memory(new_base), because we need to update du.
  3642     assert(base_memory() == new_base, "");
  3643     progress = this;
  3646   if( base_memory() == this ) {
  3647     // a self cycle indicates this memory path is dead
  3648     set_req(Compile::AliasIdxBot, empty_mem);
  3651   // Resolve external cycles by calling Ideal on a MergeMem base_memory
  3652   // Recursion must occur after the self cycle check above
  3653   if( base_memory()->is_MergeMem() ) {
  3654     MergeMemNode *new_mbase = base_memory()->as_MergeMem();
  3655     Node *m = phase->transform(new_mbase);  // Rollup any cycles
  3656     if( m != NULL && (m->is_top() ||
  3657         m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
  3658       // propagate rollup of dead cycle to self
  3659       set_req(Compile::AliasIdxBot, empty_mem);
  3663   if( base_memory() == empty_mem ) {
  3664     progress = this;
  3665     // Cut inputs during Parse phase only.
  3666     // During Optimize phase a dead MergeMem node will be subsumed by Top.
  3667     if( !can_reshape ) {
  3668       for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
  3669         if( in(i) != empty_mem ) { set_req(i, empty_mem); }
  3674   if( !progress && base_memory()->is_Phi() && can_reshape ) {
  3675     // Check if PhiNode::Ideal's "Split phis through memory merges"
  3676     // transform should be attempted. Look for this->phi->this cycle.
  3677     uint merge_width = req();
  3678     if (merge_width > Compile::AliasIdxRaw) {
  3679       PhiNode* phi = base_memory()->as_Phi();
  3680       for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
  3681         if (phi->in(i) == this) {
  3682           phase->is_IterGVN()->_worklist.push(phi);
  3683           break;
  3689   assert(progress || verify_sparse(), "please, no dups of base");
  3690   return progress;
  3693 //-------------------------set_base_memory-------------------------------------
  3694 void MergeMemNode::set_base_memory(Node *new_base) {
  3695   Node* empty_mem = empty_memory();
  3696   set_req(Compile::AliasIdxBot, new_base);
  3697   assert(memory_at(req()) == new_base, "must set default memory");
  3698   // Clear out other occurrences of new_base:
  3699   if (new_base != empty_mem) {
  3700     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
  3701       if (in(i) == new_base)  set_req(i, empty_mem);
  3706 //------------------------------out_RegMask------------------------------------
  3707 const RegMask &MergeMemNode::out_RegMask() const {
  3708   return RegMask::Empty;
  3711 //------------------------------dump_spec--------------------------------------
  3712 #ifndef PRODUCT
  3713 void MergeMemNode::dump_spec(outputStream *st) const {
  3714   st->print(" {");
  3715   Node* base_mem = base_memory();
  3716   for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
  3717     Node* mem = memory_at(i);
  3718     if (mem == base_mem) { st->print(" -"); continue; }
  3719     st->print( " N%d:", mem->_idx );
  3720     Compile::current()->get_adr_type(i)->dump_on(st);
  3722   st->print(" }");
  3724 #endif // !PRODUCT
  3727 #ifdef ASSERT
  3728 static bool might_be_same(Node* a, Node* b) {
  3729   if (a == b)  return true;
  3730   if (!(a->is_Phi() || b->is_Phi()))  return false;
  3731   // phis shift around during optimization
  3732   return true;  // pretty stupid...
  3735 // verify a narrow slice (either incoming or outgoing)
  3736 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
  3737   if (!VerifyAliases)       return;  // don't bother to verify unless requested
  3738   if (is_error_reported())  return;  // muzzle asserts when debugging an error
  3739   if (Node::in_dump())      return;  // muzzle asserts when printing
  3740   assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
  3741   assert(n != NULL, "");
  3742   // Elide intervening MergeMem's
  3743   while (n->is_MergeMem()) {
  3744     n = n->as_MergeMem()->memory_at(alias_idx);
  3746   Compile* C = Compile::current();
  3747   const TypePtr* n_adr_type = n->adr_type();
  3748   if (n == m->empty_memory()) {
  3749     // Implicit copy of base_memory()
  3750   } else if (n_adr_type != TypePtr::BOTTOM) {
  3751     assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
  3752     assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
  3753   } else {
  3754     // A few places like make_runtime_call "know" that VM calls are narrow,
  3755     // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
  3756     bool expected_wide_mem = false;
  3757     if (n == m->base_memory()) {
  3758       expected_wide_mem = true;
  3759     } else if (alias_idx == Compile::AliasIdxRaw ||
  3760                n == m->memory_at(Compile::AliasIdxRaw)) {
  3761       expected_wide_mem = true;
  3762     } else if (!C->alias_type(alias_idx)->is_rewritable()) {
  3763       // memory can "leak through" calls on channels that
  3764       // are write-once.  Allow this also.
  3765       expected_wide_mem = true;
  3767     assert(expected_wide_mem, "expected narrow slice replacement");
  3770 #else // !ASSERT
  3771 #define verify_memory_slice(m,i,n) (0)  // PRODUCT version is no-op
  3772 #endif
  3775 //-----------------------------memory_at---------------------------------------
  3776 Node* MergeMemNode::memory_at(uint alias_idx) const {
  3777   assert(alias_idx >= Compile::AliasIdxRaw ||
  3778          alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
  3779          "must avoid base_memory and AliasIdxTop");
  3781   // Otherwise, it is a narrow slice.
  3782   Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
  3783   Compile *C = Compile::current();
  3784   if (is_empty_memory(n)) {
  3785     // the array is sparse; empty slots are the "top" node
  3786     n = base_memory();
  3787     assert(Node::in_dump()
  3788            || n == NULL || n->bottom_type() == Type::TOP
  3789            || n->adr_type() == TypePtr::BOTTOM
  3790            || n->adr_type() == TypeRawPtr::BOTTOM
  3791            || Compile::current()->AliasLevel() == 0,
  3792            "must be a wide memory");
  3793     // AliasLevel == 0 if we are organizing the memory states manually.
  3794     // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
  3795   } else {
  3796     // make sure the stored slice is sane
  3797     #ifdef ASSERT
  3798     if (is_error_reported() || Node::in_dump()) {
  3799     } else if (might_be_same(n, base_memory())) {
  3800       // Give it a pass:  It is a mostly harmless repetition of the base.
  3801       // This can arise normally from node subsumption during optimization.
  3802     } else {
  3803       verify_memory_slice(this, alias_idx, n);
  3805     #endif
  3807   return n;
  3810 //---------------------------set_memory_at-------------------------------------
  3811 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
  3812   verify_memory_slice(this, alias_idx, n);
  3813   Node* empty_mem = empty_memory();
  3814   if (n == base_memory())  n = empty_mem;  // collapse default
  3815   uint need_req = alias_idx+1;
  3816   if (req() < need_req) {
  3817     if (n == empty_mem)  return;  // already the default, so do not grow me
  3818     // grow the sparse array
  3819     do {
  3820       add_req(empty_mem);
  3821     } while (req() < need_req);
  3823   set_req( alias_idx, n );
  3828 //--------------------------iteration_setup------------------------------------
  3829 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
  3830   if (other != NULL) {
  3831     grow_to_match(other);
  3832     // invariant:  the finite support of mm2 is within mm->req()
  3833     #ifdef ASSERT
  3834     for (uint i = req(); i < other->req(); i++) {
  3835       assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
  3837     #endif
  3839   // Replace spurious copies of base_memory by top.
  3840   Node* base_mem = base_memory();
  3841   if (base_mem != NULL && !base_mem->is_top()) {
  3842     for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
  3843       if (in(i) == base_mem)
  3844         set_req(i, empty_memory());
  3849 //---------------------------grow_to_match-------------------------------------
  3850 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
  3851   Node* empty_mem = empty_memory();
  3852   assert(other->is_empty_memory(empty_mem), "consistent sentinels");
  3853   // look for the finite support of the other memory
  3854   for (uint i = other->req(); --i >= req(); ) {
  3855     if (other->in(i) != empty_mem) {
  3856       uint new_len = i+1;
  3857       while (req() < new_len)  add_req(empty_mem);
  3858       break;
  3863 //---------------------------verify_sparse-------------------------------------
  3864 #ifndef PRODUCT
  3865 bool MergeMemNode::verify_sparse() const {
  3866   assert(is_empty_memory(make_empty_memory()), "sane sentinel");
  3867   Node* base_mem = base_memory();
  3868   // The following can happen in degenerate cases, since empty==top.
  3869   if (is_empty_memory(base_mem))  return true;
  3870   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
  3871     assert(in(i) != NULL, "sane slice");
  3872     if (in(i) == base_mem)  return false;  // should have been the sentinel value!
  3874   return true;
  3877 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
  3878   Node* n;
  3879   n = mm->in(idx);
  3880   if (mem == n)  return true;  // might be empty_memory()
  3881   n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
  3882   if (mem == n)  return true;
  3883   while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
  3884     if (mem == n)  return true;
  3885     if (n == NULL)  break;
  3887   return false;
  3889 #endif // !PRODUCT

mercurial