src/share/vm/opto/memnode.cpp

*** 131,141 ****
    result = optimize_simple_memory_chain(result, t_adr, phase);
    if (is_instance && igvn != NULL  && result->is_Phi()) {
      PhiNode *mphi = result->as_Phi();
      assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
      const TypePtr *t = mphi->adr_type();
!     if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM) {
        // clone the Phi with our address type
        result = mphi->split_out_instance(t_adr, igvn);
      } else {
        assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
      }
--- 131,143 ----
    result = optimize_simple_memory_chain(result, t_adr, phase);
    if (is_instance && igvn != NULL  && result->is_Phi()) {
      PhiNode *mphi = result->as_Phi();
      assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
      const TypePtr *t = mphi->adr_type();
!     if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
!         t->isa_oopptr() && !t->is_oopptr()->is_instance() &&
!         t->is_oopptr()->cast_to_instance(t_oop->instance_id()) == t_oop) {
        // clone the Phi with our address type
        result = mphi->split_out_instance(t_adr, igvn);
      } else {
        assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
      }
*** 261,271 ****
  
    // 'dom' dominates 'sub' if its control edge and control edges
    // of all its inputs dominate or equal to sub's control edge.
  
    // Currently 'sub' is either Allocate, Initialize or Start nodes.
!   assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start(), "expecting only these nodes");
  
    // Get control edge of 'sub'.
    sub = sub->find_exact_control(sub->in(0));
    if (sub == NULL || sub->is_top())
      return false; // Conservative answer for dead code
--- 263,276 ----
  
    // 'dom' dominates 'sub' if its control edge and control edges
    // of all its inputs dominate or equal to sub's control edge.
  
    // Currently 'sub' is either Allocate, Initialize or Start nodes.
!   // Or Region for the check in LoadNode::Ideal();
!   // 'sub' should have sub->in(0) != NULL.
!   assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
!          sub->is_Region(), "expecting only these nodes");
  
    // Get control edge of 'sub'.
    sub = sub->find_exact_control(sub->in(0));
    if (sub == NULL || sub->is_top())
      return false; // Conservative answer for dead code
*** 574,601 ****
  
  //------------------------------Ideal_DU_postCCP-------------------------------
  // Find any cast-away of null-ness and keep its control.  Null cast-aways are
  // going away in this pass and we need to make this memory op depend on the
  // gating null check.
  
  // I tried to leave the CastPP's in.  This makes the graph more accurate in
  // some sense; we get to keep around the knowledge that an oop is not-null
  // after some test.  Alas, the CastPP's interfere with GVN (some values are
  // the regular oop, some are the CastPP of the oop, all merge at Phi's which
  // cannot collapse, etc).  This cost us 10% on SpecJVM, even when I removed
  // some of the more trivial cases in the optimizer.  Removing more useless
  // Phi's started allowing Loads to illegally float above null checks.  I gave
  // up on this approach.  CNC 10/20/2000
! Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
!   Node *ctr = in(MemNode::Control);
!   Node *mem = in(MemNode::Memory);
!   Node *adr = in(MemNode::Address);
    Node *skipped_cast = NULL;
    // Need a null check?  Regular static accesses do not because they are
    // from constant addresses.  Array ops are gated by the range check (which
    // always includes a NULL check).  Just check field ops.
!   if( !ctr ) {
      // Scan upwards for the highest location we can place this memory op.
      while( true ) {
        switch( adr->Opcode() ) {
  
        case Op_AddP:             // No change to NULL-ness, so peek thru AddP's
--- 579,606 ----
  
  //------------------------------Ideal_DU_postCCP-------------------------------
  // Find any cast-away of null-ness and keep its control.  Null cast-aways are
  // going away in this pass and we need to make this memory op depend on the
  // gating null check.
+ Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
+   return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
+ }
  
  // I tried to leave the CastPP's in.  This makes the graph more accurate in
  // some sense; we get to keep around the knowledge that an oop is not-null
  // after some test.  Alas, the CastPP's interfere with GVN (some values are
  // the regular oop, some are the CastPP of the oop, all merge at Phi's which
  // cannot collapse, etc).  This cost us 10% on SpecJVM, even when I removed
  // some of the more trivial cases in the optimizer.  Removing more useless
  // Phi's started allowing Loads to illegally float above null checks.  I gave
  // up on this approach.  CNC 10/20/2000
! Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
    Node *skipped_cast = NULL;
    // Need a null check?  Regular static accesses do not because they are
    // from constant addresses.  Array ops are gated by the range check (which
    // always includes a NULL check).  Just check field ops.
!   if( n->in(MemNode::Control) == NULL ) {
      // Scan upwards for the highest location we can place this memory op.
      while( true ) {
        switch( adr->Opcode() ) {
  
        case Op_AddP:             // No change to NULL-ness, so peek thru AddP's
*** 616,629 ****
            adr = adr->in(1);
            continue;
          }
          // CastPP is going away in this pass!  We need this memory op to be
          // control-dependent on the test that is guarding the CastPP.
!         ccp->hash_delete(this);
!         set_req(MemNode::Control, adr->in(0));
!         ccp->hash_insert(this);
!         return this;
  
        case Op_Phi:
          // Attempt to float above a Phi to some dominating point.
          if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
            // If we've already peeked through a Cast (which could have set the
--- 621,634 ----
            adr = adr->in(1);
            continue;
          }
          // CastPP is going away in this pass!  We need this memory op to be
          // control-dependent on the test that is guarding the CastPP.
!         ccp->hash_delete(n);
!         n->set_req(MemNode::Control, adr->in(0));
!         ccp->hash_insert(n);
!         return n;
  
        case Op_Phi:
          // Attempt to float above a Phi to some dominating point.
          if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
            // If we've already peeked through a Cast (which could have set the
*** 650,672 ****
            // the node in place. 4959717.
            skipped_cast = adr;
            adr = adr->in(1);
            continue;
          }
!         ccp->hash_delete(this);
!         set_req(MemNode::Control, adr->in(0));
!         ccp->hash_insert(this);
!         return this;
  
          // List of "safe" opcodes; those that implicitly block the memory
          // op below any null check.
        case Op_CastX2P:          // no null checks on native pointers
        case Op_Parm:             // 'this' pointer is not null
        case Op_LoadP:            // Loading from within a klass
        case Op_LoadN:            // Loading from within a klass
        case Op_LoadKlass:        // Loading from within a klass
        case Op_ConP:             // Loading from a klass
        case Op_CreateEx:         // Sucking up the guts of an exception oop
        case Op_Con:              // Reading from TLS
        case Op_CMoveP:           // CMoveP is pinned
          break;                  // No progress
  
--- 655,678 ----
            // the node in place. 4959717.
            skipped_cast = adr;
            adr = adr->in(1);
            continue;
          }
!         ccp->hash_delete(n);
!         n->set_req(MemNode::Control, adr->in(0));
!         ccp->hash_insert(n);
!         return n;
  
          // List of "safe" opcodes; those that implicitly block the memory
          // op below any null check.
        case Op_CastX2P:          // no null checks on native pointers
        case Op_Parm:             // 'this' pointer is not null
        case Op_LoadP:            // Loading from within a klass
        case Op_LoadN:            // Loading from within a klass
        case Op_LoadKlass:        // Loading from within a klass
        case Op_ConP:             // Loading from a klass
+       case Op_ConN:             // Loading from a klass
        case Op_CreateEx:         // Sucking up the guts of an exception oop
        case Op_Con:              // Reading from TLS
        case Op_CMoveP:           // CMoveP is pinned
          break;                  // No progress
  
*** 674,685 ****
        case Op_SCMemProj:        // Memory state from store conditional ops
  #ifdef ASSERT
          {
            assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
            const Node* call = adr->in(0);
!           if (call->is_CallStaticJava()) {
!             const CallStaticJavaNode* call_java = call->as_CallStaticJava();
              const TypeTuple *r = call_java->tf()->range();
              assert(r->cnt() > TypeFunc::Parms, "must return value");
              const Type* ret_type = r->field_at(TypeFunc::Parms);
              assert(ret_type && ret_type->isa_ptr(), "must return pointer");
              // We further presume that this is one of
--- 680,691 ----
        case Op_SCMemProj:        // Memory state from store conditional ops
  #ifdef ASSERT
          {
            assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
            const Node* call = adr->in(0);
!           if (call->is_CallJava()) {
!             const CallJavaNode* call_java = call->as_CallJava();
              const TypeTuple *r = call_java->tf()->range();
              assert(r->cnt() > TypeFunc::Parms, "must return value");
              const Type* ret_type = r->field_at(TypeFunc::Parms);
              assert(ret_type && ret_type->isa_ptr(), "must return pointer");
              // We further presume that this is one of
*** 747,757 ****
    case T_FLOAT:   return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt              );
    case T_DOUBLE:  return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt              );
    case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr()    );
    case T_OBJECT:
  #ifdef _LP64
!     if (adr->bottom_type()->is_narrow()) {
        const TypeNarrowOop* narrowtype;
        if (rt->isa_narrowoop()) {
          narrowtype = rt->is_narrowoop();
        } else {
          narrowtype = rt->is_oopptr()->make_narrowoop();
--- 753,763 ----
    case T_FLOAT:   return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt              );
    case T_DOUBLE:  return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt              );
    case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr()    );
    case T_OBJECT:
  #ifdef _LP64
!     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
        const TypeNarrowOop* narrowtype;
        if (rt->isa_narrowoop()) {
          narrowtype = rt->is_narrowoop();
        } else {
          narrowtype = rt->is_oopptr()->make_narrowoop();
*** 760,770 ****
  
        return DecodeNNode::decode(&gvn, load);
      } else
  #endif
        {
!         assert(!adr->bottom_type()->is_narrow(), "should have got back a narrow oop");
          return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
        }
    }
    ShouldNotReachHere();
    return (LoadNode*)NULL;
--- 766,776 ----
  
        return DecodeNNode::decode(&gvn, load);
      } else
  #endif
      {
!       assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
        return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
      }
    }
    ShouldNotReachHere();
    return (LoadNode*)NULL;
*** 1203,1212 ****
--- 1209,1223 ----
          // Do nothing here if Identity will find a value
          // (to avoid infinite chain of value phis generation).
          if ( !phase->eqv(this, this->Identity(phase)) )
            return NULL;
  
+         // Skip the split if the loop head dominates some control edges
+         // of the address.
+         if (cnt == 3 && !MemNode::all_controls_dominate(address, region))
+           goto find_store;
+ 
          const Type* this_type = this->bottom_type();
          int this_index  = phase->C->get_alias_index(addr_t);
          int this_offset = addr_t->offset();
          int this_iid    = addr_t->is_oopptr()->instance_id();
          int wins = 0;
*** 1281,1291 ****
            igvn->remove_dead_node(phi);
          }
        }
      }
    }
! 
    // Check for prior store with a different base or offset; make Load
    // independent.  Skip through any number of them.  Bail out if the stores
    // are in an endless dead cycle and report no progress.  This is a key
    // transform for Reflection.  However, if after skipping through the Stores
    // we can't then fold up against a prior store do NOT do the transform as
--- 1292,1302 ----
            igvn->remove_dead_node(phi);
          }
        }
      }
    }
! find_store:
    // Check for prior store with a different base or offset; make Load
    // independent.  Skip through any number of them.  Bail out if the stores
    // are in an endless dead cycle and report no progress.  This is a key
    // transform for Reflection.  However, if after skipping through the Stores
    // we can't then fold up against a prior store do NOT do the transform as
*** 1833,1843 ****
    case T_FLOAT:   return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
    case T_DOUBLE:  return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
    case T_ADDRESS:
    case T_OBJECT:
  #ifdef _LP64
!     if (adr->bottom_type()->is_narrow() ||
          (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
           adr->bottom_type()->isa_rawptr())) {
        const TypePtr* type = val->bottom_type()->is_ptr();
        Node* cp = EncodePNode::encode(&gvn, val);
        return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, cp);
--- 1844,1854 ----
    case T_FLOAT:   return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
    case T_DOUBLE:  return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
    case T_ADDRESS:
    case T_OBJECT:
  #ifdef _LP64
!     if (adr->bottom_type()->is_ptr_to_narrowoop() ||
          (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
           adr->bottom_type()->isa_rawptr())) {
        const TypePtr* type = val->bottom_type()->is_ptr();
        Node* cp = EncodePNode::encode(&gvn, val);
        return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, cp);