Mercurial > jdk8-mips64-public > hotspot / file revision

 /*

  * Copyright (c) 1998, 2011, Oracle and/or its affiliates. All rights reserved.

  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.

  *

  * This code is free software; you can redistribute it and/or modify it

  * under the terms of the GNU General Public License version 2 only, as

  * published by the Free Software Foundation.

  *

  * This code is distributed in the hope that it will be useful, but WITHOUT

  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or

  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License

  * version 2 for more details (a copy is included in the LICENSE file that

  * accompanied this code).

  *

  * You should have received a copy of the GNU General Public License version

  * 2 along with this work; if not, write to the Free Software Foundation,

  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.

  *

  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA

  * or visit www.oracle.com if you need additional information or have any

  * questions.

  *

  */

 #include "precompiled.hpp"

 #include "asm/assembler.inline.hpp"

 #include "code/debugInfo.hpp"

 #include "code/debugInfoRec.hpp"

 #include "compiler/compileBroker.hpp"

 #include "compiler/oopMap.hpp"

 #include "memory/allocation.inline.hpp"

 #include "opto/callnode.hpp"

 #include "opto/cfgnode.hpp"

 #include "opto/locknode.hpp"

 #include "opto/machnode.hpp"

 #include "opto/output.hpp"

 #include "opto/regalloc.hpp"

 #include "opto/runtime.hpp"

 #include "opto/subnode.hpp"

 #include "opto/type.hpp"

 #include "runtime/handles.inline.hpp"

 #include "utilities/xmlstream.hpp"

 extern uint size_java_to_interp();

 extern uint reloc_java_to_interp();

 extern uint size_exception_handler();

 extern uint size_deopt_handler();

 #ifndef PRODUCT

 #define DEBUG_ARG(x) , x

 #else

 #define DEBUG_ARG(x)

 #endif

 extern int emit_exception_handler(CodeBuffer &cbuf);

 extern int emit_deopt_handler(CodeBuffer &cbuf);

 //------------------------------Output-----------------------------------------

 // Convert Nodes to instruction bits and pass off to the VM

 void Compile::Output() {

   // RootNode goes

   assert( _cfg->_broot->_nodes.size() == 0, "" );

   // The number of new nodes (mostly MachNop) is proportional to

   // the number of java calls and inner loops which are aligned.

   if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +

                             C->inner_loops()*(OptoLoopAlignment-1)),

                            "out of nodes before code generation" ) ) {

     return;

   }

   // Make sure I can find the Start Node

   Block_Array& bbs = _cfg->_bbs;

   Block *entry = _cfg->_blocks[1];

   Block *broot = _cfg->_broot;

   const StartNode *start = entry->_nodes[0]->as_Start();

   // Replace StartNode with prolog

   MachPrologNode *prolog = new (this) MachPrologNode();

   entry->_nodes.map( 0, prolog );

   bbs.map( prolog->_idx, entry );

   bbs.map( start->_idx, NULL ); // start is no longer in any block

   // Virtual methods need an unverified entry point

   if( is_osr_compilation() ) {

     if( PoisonOSREntry ) {

       // TODO: Should use a ShouldNotReachHereNode...

       _cfg->insert( broot, 0, new (this) MachBreakpointNode() );

     }

   } else {

     if( _method && !_method->flags().is_static() ) {

       // Insert unvalidated entry point

       _cfg->insert( broot, 0, new (this) MachUEPNode() );

     }

   }

   // Break before main entry point

   if( (_method && _method->break_at_execute())

 #ifndef PRODUCT

     ||(OptoBreakpoint && is_method_compilation())

     ||(OptoBreakpointOSR && is_osr_compilation())

     ||(OptoBreakpointC2R && !_method)

 #endif

     ) {

     // checking for _method means that OptoBreakpoint does not apply to

     // runtime stubs or frame converters

     _cfg->insert( entry, 1, new (this) MachBreakpointNode() );

   }

   // Insert epilogs before every return

   for( uint i=0; i<_cfg->_num_blocks; i++ ) {

     Block *b = _cfg->_blocks[i];

     if( !b->is_connector() && b->non_connector_successor(0) == _cfg->_broot ) { // Found a program exit point?

       Node *m = b->end();

       if( m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt ) {

         MachEpilogNode *epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);

         b->add_inst( epilog );

         bbs.map(epilog->_idx, b);

         //_regalloc->set_bad(epilog->_idx); // Already initialized this way.

       }

     }

   }

 # ifdef ENABLE_ZAP_DEAD_LOCALS

   if ( ZapDeadCompiledLocals )  Insert_zap_nodes();

 # endif

   uint* blk_starts = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks+1);

   blk_starts[0]    = 0;

   // Initialize code buffer and process short branches.

   CodeBuffer* cb = init_buffer(blk_starts);

   if (cb == NULL || failing())  return;

   ScheduleAndBundle();

 #ifndef PRODUCT

   if (trace_opto_output()) {

     tty->print("\n---- After ScheduleAndBundle ----\n");

     for (uint i = 0; i < _cfg->_num_blocks; i++) {

       tty->print("\nBB#%03d:\n", i);

       Block *bb = _cfg->_blocks[i];

       for (uint j = 0; j < bb->_nodes.size(); j++) {

         Node *n = bb->_nodes[j];

         OptoReg::Name reg = _regalloc->get_reg_first(n);

         tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");

         n->dump();

       }

     }

   }

 #endif

   if (failing())  return;

   BuildOopMaps();

   if (failing())  return;

   fill_buffer(cb, blk_starts);

 }

 bool Compile::need_stack_bang(int frame_size_in_bytes) const {

   // Determine if we need to generate a stack overflow check.

   // Do it if the method is not a stub function and

   // has java calls or has frame size > vm_page_size/8.

   return (stub_function() == NULL &&

           (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3));

 }

 bool Compile::need_register_stack_bang() const {

   // Determine if we need to generate a register stack overflow check.

   // This is only used on architectures which have split register

   // and memory stacks (ie. IA64).

   // Bang if the method is not a stub function and has java calls

   return (stub_function() == NULL && has_java_calls());

 }

 # ifdef ENABLE_ZAP_DEAD_LOCALS

 // In order to catch compiler oop-map bugs, we have implemented

 // a debugging mode called ZapDeadCompilerLocals.

 // This mode causes the compiler to insert a call to a runtime routine,

 // "zap_dead_locals", right before each place in compiled code

 // that could potentially be a gc-point (i.e., a safepoint or oop map point).

 // The runtime routine checks that locations mapped as oops are really

 // oops, that locations mapped as values do not look like oops,

 // and that locations mapped as dead are not used later

 // (by zapping them to an invalid address).

 int Compile::_CompiledZap_count = 0;

 void Compile::Insert_zap_nodes() {

   bool skip = false;

   // Dink with static counts because code code without the extra

   // runtime calls is MUCH faster for debugging purposes

        if ( CompileZapFirst  ==  0  ) ; // nothing special

   else if ( CompileZapFirst  >  CompiledZap_count() )  skip = true;

   else if ( CompileZapFirst  == CompiledZap_count() )

     warning("starting zap compilation after skipping");

        if ( CompileZapLast  ==  -1  ) ; // nothing special

   else if ( CompileZapLast  <   CompiledZap_count() )  skip = true;

   else if ( CompileZapLast  ==  CompiledZap_count() )

     warning("about to compile last zap");

   ++_CompiledZap_count; // counts skipped zaps, too

   if ( skip )  return;

   if ( _method == NULL )

     return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care

   // Insert call to zap runtime stub before every node with an oop map

   for( uint i=0; i<_cfg->_num_blocks; i++ ) {

     Block *b = _cfg->_blocks[i];

     for ( uint j = 0;  j < b->_nodes.size();  ++j ) {

       Node *n = b->_nodes[j];

       // Determining if we should insert a zap-a-lot node in output.

       // We do that for all nodes that has oopmap info, except for calls

       // to allocation.  Calls to allocation passes in the old top-of-eden pointer

       // and expect the C code to reset it.  Hence, there can be no safepoints between

       // the inlined-allocation and the call to new_Java, etc.

       // We also cannot zap monitor calls, as they must hold the microlock

       // during the call to Zap, which also wants to grab the microlock.

       bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL);

       if ( insert ) { // it is MachSafePoint

         if ( !n->is_MachCall() ) {

           insert = false;

         } else if ( n->is_MachCall() ) {

           MachCallNode* call = n->as_MachCall();

           if (call->entry_point() == OptoRuntime::new_instance_Java() ||

               call->entry_point() == OptoRuntime::new_array_Java() ||

               call->entry_point() == OptoRuntime::multianewarray2_Java() ||

               call->entry_point() == OptoRuntime::multianewarray3_Java() ||

               call->entry_point() == OptoRuntime::multianewarray4_Java() ||

               call->entry_point() == OptoRuntime::multianewarray5_Java() ||

               call->entry_point() == OptoRuntime::slow_arraycopy_Java() ||

               call->entry_point() == OptoRuntime::complete_monitor_locking_Java()

               ) {

             insert = false;

           }

         }

         if (insert) {

           Node *zap = call_zap_node(n->as_MachSafePoint(), i);

           b->_nodes.insert( j, zap );

           _cfg->_bbs.map( zap->_idx, b );

           ++j;

         }

       }

     }

   }

 }

 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) {

   const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type();

   CallStaticJavaNode* ideal_node =

     new (this, tf->domain()->cnt()) CallStaticJavaNode( tf,

          OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()),

                             "call zap dead locals stub", 0, TypePtr::BOTTOM);

   // We need to copy the OopMap from the site we're zapping at.

   // We have to make a copy, because the zap site might not be

   // a call site, and zap_dead is a call site.

   OopMap* clone = node_to_check->oop_map()->deep_copy();

   // Add the cloned OopMap to the zap node

   ideal_node->set_oop_map(clone);

   return _matcher->match_sfpt(ideal_node);

 }

 //------------------------------is_node_getting_a_safepoint--------------------

 bool Compile::is_node_getting_a_safepoint( Node* n) {

   // This code duplicates the logic prior to the call of add_safepoint

   // below in this file.

   if( n->is_MachSafePoint() ) return true;

   return false;

 }

 # endif // ENABLE_ZAP_DEAD_LOCALS

 //------------------------------compute_loop_first_inst_sizes------------------

 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top

 // of a loop. When aligning a loop we need to provide enough instructions

 // in cpu's fetch buffer to feed decoders. The loop alignment could be

 // avoided if we have enough instructions in fetch buffer at the head of a loop.

 // By default, the size is set to 999999 by Block's constructor so that

 // a loop will be aligned if the size is not reset here.

 //

 // Note: Mach instructions could contain several HW instructions

 // so the size is estimated only.

 //

 void Compile::compute_loop_first_inst_sizes() {

   // The next condition is used to gate the loop alignment optimization.

   // Don't aligned a loop if there are enough instructions at the head of a loop

   // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad

   // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is

   // equal to 11 bytes which is the largest address NOP instruction.

   if( MaxLoopPad < OptoLoopAlignment-1 ) {

     uint last_block = _cfg->_num_blocks-1;

     for( uint i=1; i <= last_block; i++ ) {

       Block *b = _cfg->_blocks[i];

       // Check the first loop's block which requires an alignment.

       if( b->loop_alignment() > (uint)relocInfo::addr_unit() ) {

         uint sum_size = 0;

         uint inst_cnt = NumberOfLoopInstrToAlign;

         inst_cnt = b->compute_first_inst_size(sum_size, inst_cnt, _regalloc);

         // Check subsequent fallthrough blocks if the loop's first

         // block(s) does not have enough instructions.

         Block *nb = b;

         while( inst_cnt > 0 &&

                i < last_block &&

                !_cfg->_blocks[i+1]->has_loop_alignment() &&

                !nb->has_successor(b) ) {

           i++;

           nb = _cfg->_blocks[i];

           inst_cnt  = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);

         } // while( inst_cnt > 0 && i < last_block  )

         b->set_first_inst_size(sum_size);

       } // f( b->head()->is_Loop() )

     } // for( i <= last_block )

   } // if( MaxLoopPad < OptoLoopAlignment-1 )

 }

 //----------------------shorten_branches---------------------------------------

 // The architecture description provides short branch variants for some long

 // branch instructions. Replace eligible long branches with short branches.

 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {

   // ------------------

   // Compute size of each block, method size, and relocation information size

   uint nblocks  = _cfg->_num_blocks;

   uint*      jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);

   uint*      jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);

   int*       jmp_nidx   = NEW_RESOURCE_ARRAY(int ,nblocks);

   DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )

   DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )

   bool has_short_branch_candidate = false;

   // Initialize the sizes to 0

   code_size  = 0;          // Size in bytes of generated code

   stub_size  = 0;          // Size in bytes of all stub entries

   // Size in bytes of all relocation entries, including those in local stubs.

   // Start with 2-bytes of reloc info for the unvalidated entry point

   reloc_size = 1;          // Number of relocation entries

   // Make three passes.  The first computes pessimistic blk_starts,

   // relative jmp_offset and reloc_size information.  The second performs

   // short branch substitution using the pessimistic sizing.  The

   // third inserts nops where needed.

   // Step one, perform a pessimistic sizing pass.

   uint last_call_adr = max_uint;

   uint last_avoid_back_to_back_adr = max_uint;

   uint nop_size = (new (this) MachNopNode())->size(_regalloc);

   for (uint i = 0; i < nblocks; i++) { // For all blocks

     Block *b = _cfg->_blocks[i];

     // During short branch replacement, we store the relative (to blk_starts)

     // offset of jump in jmp_offset, rather than the absolute offset of jump.

     // This is so that we do not need to recompute sizes of all nodes when

     // we compute correct blk_starts in our next sizing pass.

     jmp_offset[i] = 0;

     jmp_size[i]   = 0;

     jmp_nidx[i]   = -1;

     DEBUG_ONLY( jmp_target[i] = 0; )

     DEBUG_ONLY( jmp_rule[i]   = 0; )

     // Sum all instruction sizes to compute block size

     uint last_inst = b->_nodes.size();

     uint blk_size = 0;

     for (uint j = 0; j < last_inst; j++) {

       Node* nj = b->_nodes[j];

       // Handle machine instruction nodes

       if (nj->is_Mach()) {

         MachNode *mach = nj->as_Mach();

         blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding

         reloc_size += mach->reloc();

         if( mach->is_MachCall() ) {

           MachCallNode *mcall = mach->as_MachCall();

           // This destination address is NOT PC-relative

           mcall->method_set((intptr_t)mcall->entry_point());

           if( mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method ) {

             stub_size  += size_java_to_interp();

             reloc_size += reloc_java_to_interp();

           }

         } else if (mach->is_MachSafePoint()) {

           // If call/safepoint are adjacent, account for possible

           // nop to disambiguate the two safepoints.

           // ScheduleAndBundle() can rearrange nodes in a block,

           // check for all offsets inside this block.

           if (last_call_adr >= blk_starts[i]) {

             blk_size += nop_size;

           }

         }

         if (mach->avoid_back_to_back()) {

           // Nop is inserted between "avoid back to back" instructions.

           // ScheduleAndBundle() can rearrange nodes in a block,

           // check for all offsets inside this block.

           if (last_avoid_back_to_back_adr >= blk_starts[i]) {

             blk_size += nop_size;

           }

         }

         if (mach->may_be_short_branch()) {

           if (!nj->is_MachBranch()) {

 #ifndef PRODUCT

             nj->dump(3);

 #endif

             Unimplemented();

           }

           assert(jmp_nidx[i] == -1, "block should have only one branch");

           jmp_offset[i] = blk_size;

           jmp_size[i]   = nj->size(_regalloc);

           jmp_nidx[i]   = j;

           has_short_branch_candidate = true;

         }

       }

       blk_size += nj->size(_regalloc);

       // Remember end of call offset

       if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {

         last_call_adr = blk_starts[i]+blk_size;

       }

       // Remember end of avoid_back_to_back offset

       if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) {

         last_avoid_back_to_back_adr = blk_starts[i]+blk_size;

       }

     }

     // When the next block starts a loop, we may insert pad NOP

     // instructions.  Since we cannot know our future alignment,

     // assume the worst.

     if (i< nblocks-1) {

       Block *nb = _cfg->_blocks[i+1];

       int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();

       if (max_loop_pad > 0) {

         assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");

         blk_size += max_loop_pad;

       }

     }

     // Save block size; update total method size

     blk_starts[i+1] = blk_starts[i]+blk_size;

   }

   // Step two, replace eligible long jumps.

   bool progress = true;

   uint last_may_be_short_branch_adr = max_uint;

   while (has_short_branch_candidate && progress) {

     progress = false;

     has_short_branch_candidate = false;

     int adjust_block_start = 0;

     for (uint i = 0; i < nblocks; i++) {

       Block *b = _cfg->_blocks[i];

       int idx = jmp_nidx[i];

       MachNode* mach = (idx == -1) ? NULL: b->_nodes[idx]->as_Mach();

       if (mach != NULL && mach->may_be_short_branch()) {

 #ifdef ASSERT

         assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");

         int j;

         // Find the branch; ignore trailing NOPs.

         for (j = b->_nodes.size()-1; j>=0; j--) {

           Node* n = b->_nodes[j];

           if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)

             break;

         }

         assert(j >= 0 && j == idx && b->_nodes[j] == (Node*)mach, "sanity");

 #endif

         int br_size = jmp_size[i];

         int br_offs = blk_starts[i] + jmp_offset[i];

         // This requires the TRUE branch target be in succs[0]

         uint bnum = b->non_connector_successor(0)->_pre_order;

         int offset = blk_starts[bnum] - br_offs;

         if (bnum > i) { // adjust following block's offset

           offset -= adjust_block_start;

         }

         // In the following code a nop could be inserted before

         // the branch which will increase the backward distance.

         bool needs_padding = ((uint)br_offs == last_may_be_short_branch_adr);

         if (needs_padding && offset <= 0)

           offset -= nop_size;

         if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {

           // We've got a winner.  Replace this branch.

           MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);

           // Update the jmp_size.

           int new_size = replacement->size(_regalloc);

           int diff     = br_size - new_size;

           assert(diff >= (int)nop_size, "short_branch size should be smaller");

           // Conservatively take into accound padding between

           // avoid_back_to_back branches. Previous branch could be

           // converted into avoid_back_to_back branch during next

           // rounds.

           if (needs_padding && replacement->avoid_back_to_back()) {

             jmp_offset[i] += nop_size;

             diff -= nop_size;

           }

           adjust_block_start += diff;

           b->_nodes.map(idx, replacement);

           mach->subsume_by(replacement);

           mach = replacement;

           progress = true;

           jmp_size[i] = new_size;

           DEBUG_ONLY( jmp_target[i] = bnum; );

           DEBUG_ONLY( jmp_rule[i] = mach->rule(); );

         } else {

           // The jump distance is not short, try again during next iteration.

           has_short_branch_candidate = true;

         }

       } // (mach->may_be_short_branch())

       if (mach != NULL && (mach->may_be_short_branch() ||

                            mach->avoid_back_to_back())) {

         last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];

       }

       blk_starts[i+1] -= adjust_block_start;

     }

   }

 #ifdef ASSERT

   for (uint i = 0; i < nblocks; i++) { // For all blocks

     if (jmp_target[i] != 0) {

       int br_size = jmp_size[i];

       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);

       if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {

         tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);

       }

       assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");

     }

   }

 #endif

   // Step 3, compute the offsets of all blocks, will be done in fill_buffer()

   // after ScheduleAndBundle().

   // ------------------

   // Compute size for code buffer

   code_size = blk_starts[nblocks];

   // Relocation records

   reloc_size += 1;              // Relo entry for exception handler

   // Adjust reloc_size to number of record of relocation info

   // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for

   // a relocation index.

   // The CodeBuffer will expand the locs array if this estimate is too low.

   reloc_size *= 10 / sizeof(relocInfo);

 }

 //------------------------------FillLocArray-----------------------------------

 // Create a bit of debug info and append it to the array.  The mapping is from

 // Java local or expression stack to constant, register or stack-slot.  For

 // doubles, insert 2 mappings and return 1 (to tell the caller that the next

 // entry has been taken care of and caller should skip it).

 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {

   // This should never have accepted Bad before

   assert(OptoReg::is_valid(regnum), "location must be valid");

   return (OptoReg::is_reg(regnum))

     ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )

     : new LocationValue(Location::new_stk_loc(l_type,  ra->reg2offset(regnum)));

 }

 ObjectValue*

 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {

   for (int i = 0; i < objs->length(); i++) {

     assert(objs->at(i)->is_object(), "corrupt object cache");

     ObjectValue* sv = (ObjectValue*) objs->at(i);

     if (sv->id() == id) {

       return sv;

     }

   }

   // Otherwise..

   return NULL;

 }

 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,

                                      ObjectValue* sv ) {

   assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");

   objs->append(sv);

 }

 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,

                             GrowableArray<ScopeValue*> *array,

                             GrowableArray<ScopeValue*> *objs ) {

   assert( local, "use _top instead of null" );

   if (array->length() != idx) {

     assert(array->length() == idx + 1, "Unexpected array count");

     // Old functionality:

     //   return

     // New functionality:

     //   Assert if the local is not top. In product mode let the new node

     //   override the old entry.

     assert(local == top(), "LocArray collision");

     if (local == top()) {

       return;

     }

     array->pop();

   }

   const Type *t = local->bottom_type();

   // Is it a safepoint scalar object node?

   if (local->is_SafePointScalarObject()) {

     SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();

     ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);

     if (sv == NULL) {

       ciKlass* cik = t->is_oopptr()->klass();

       assert(cik->is_instance_klass() ||

              cik->is_array_klass(), "Not supported allocation.");

       sv = new ObjectValue(spobj->_idx,

                            new ConstantOopWriteValue(cik->constant_encoding()));

       Compile::set_sv_for_object_node(objs, sv);

       uint first_ind = spobj->first_index();

       for (uint i = 0; i < spobj->n_fields(); i++) {

         Node* fld_node = sfpt->in(first_ind+i);

         (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);

       }

     }

     array->append(sv);

     return;

   }

   // Grab the register number for the local

   OptoReg::Name regnum = _regalloc->get_reg_first(local);

   if( OptoReg::is_valid(regnum) ) {// Got a register/stack?

     // Record the double as two float registers.

     // The register mask for such a value always specifies two adjacent

     // float registers, with the lower register number even.

     // Normally, the allocation of high and low words to these registers

     // is irrelevant, because nearly all operations on register pairs

     // (e.g., StoreD) treat them as a single unit.

     // Here, we assume in addition that the words in these two registers

     // stored "naturally" (by operations like StoreD and double stores

     // within the interpreter) such that the lower-numbered register

     // is written to the lower memory address.  This may seem like

     // a machine dependency, but it is not--it is a requirement on

     // the author of the <arch>.ad file to ensure that, for every

     // even/odd double-register pair to which a double may be allocated,

     // the word in the even single-register is stored to the first

     // memory word.  (Note that register numbers are completely

     // arbitrary, and are not tied to any machine-level encodings.)

 #ifdef _LP64

     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {

       array->append(new ConstantIntValue(0));

       array->append(new_loc_value( _regalloc, regnum, Location::dbl ));

     } else if ( t->base() == Type::Long ) {

       array->append(new ConstantIntValue(0));

       array->append(new_loc_value( _regalloc, regnum, Location::lng ));

     } else if ( t->base() == Type::RawPtr ) {

       // jsr/ret return address which must be restored into a the full

       // width 64-bit stack slot.

       array->append(new_loc_value( _regalloc, regnum, Location::lng ));

     }

 #else //_LP64

 #ifdef SPARC

     if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {

       // For SPARC we have to swap high and low words for

       // long values stored in a single-register (g0-g7).

       array->append(new_loc_value( _regalloc,              regnum   , Location::normal ));

       array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));

     } else

 #endif //SPARC

     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {

       // Repack the double/long as two jints.

       // The convention the interpreter uses is that the second local

       // holds the first raw word of the native double representation.

       // This is actually reasonable, since locals and stack arrays

       // grow downwards in all implementations.

       // (If, on some machine, the interpreter's Java locals or stack

       // were to grow upwards, the embedded doubles would be word-swapped.)

       array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));

       array->append(new_loc_value( _regalloc,              regnum   , Location::normal ));

     }

 #endif //_LP64

     else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&

                OptoReg::is_reg(regnum) ) {

       array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()

                                    ? Location::float_in_dbl : Location::normal ));

     } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {

       array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long

                                    ? Location::int_in_long : Location::normal ));

     } else if( t->base() == Type::NarrowOop ) {

       array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));

     } else {

       array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));

     }

     return;

   }

   // No register.  It must be constant data.

   switch (t->base()) {

   case Type::Half:              // Second half of a double

     ShouldNotReachHere();       // Caller should skip 2nd halves

     break;

   case Type::AnyPtr:

     array->append(new ConstantOopWriteValue(NULL));

     break;

   case Type::AryPtr:

   case Type::InstPtr:

   case Type::KlassPtr:          // fall through

     array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));

     break;

   case Type::NarrowOop:

     if (t == TypeNarrowOop::NULL_PTR) {

       array->append(new ConstantOopWriteValue(NULL));

     } else {

       array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));

     }

     break;

   case Type::Int:

     array->append(new ConstantIntValue(t->is_int()->get_con()));

     break;

   case Type::RawPtr:

     // A return address (T_ADDRESS).

     assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");

 #ifdef _LP64

     // Must be restored to the full-width 64-bit stack slot.

     array->append(new ConstantLongValue(t->is_ptr()->get_con()));

 #else

     array->append(new ConstantIntValue(t->is_ptr()->get_con()));

 #endif

     break;

   case Type::FloatCon: {

     float f = t->is_float_constant()->getf();

     array->append(new ConstantIntValue(jint_cast(f)));

     break;

   }

   case Type::DoubleCon: {

     jdouble d = t->is_double_constant()->getd();

 #ifdef _LP64

     array->append(new ConstantIntValue(0));

     array->append(new ConstantDoubleValue(d));

 #else

     // Repack the double as two jints.

     // The convention the interpreter uses is that the second local

     // holds the first raw word of the native double representation.

     // This is actually reasonable, since locals and stack arrays

     // grow downwards in all implementations.

     // (If, on some machine, the interpreter's Java locals or stack

     // were to grow upwards, the embedded doubles would be word-swapped.)

     jint   *dp = (jint*)&d;

     array->append(new ConstantIntValue(dp[1]));

     array->append(new ConstantIntValue(dp[0]));

 #endif

     break;

   }

   case Type::Long: {

     jlong d = t->is_long()->get_con();

 #ifdef _LP64

     array->append(new ConstantIntValue(0));

     array->append(new ConstantLongValue(d));

 #else

     // Repack the long as two jints.

     // The convention the interpreter uses is that the second local

     // holds the first raw word of the native double representation.

     // This is actually reasonable, since locals and stack arrays

     // grow downwards in all implementations.

     // (If, on some machine, the interpreter's Java locals or stack

     // were to grow upwards, the embedded doubles would be word-swapped.)

     jint *dp = (jint*)&d;

     array->append(new ConstantIntValue(dp[1]));

     array->append(new ConstantIntValue(dp[0]));

 #endif

     break;

   }

   case Type::Top:               // Add an illegal value here

     array->append(new LocationValue(Location()));

     break;

   default:

     ShouldNotReachHere();

     break;

   }

 }

 // Determine if this node starts a bundle

 bool Compile::starts_bundle(const Node *n) const {

   return (_node_bundling_limit > n->_idx &&

           _node_bundling_base[n->_idx].starts_bundle());

 }

 //--------------------------Process_OopMap_Node--------------------------------

 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {

   // Handle special safepoint nodes for synchronization

   MachSafePointNode *sfn   = mach->as_MachSafePoint();

   MachCallNode      *mcall;

 #ifdef ENABLE_ZAP_DEAD_LOCALS

   assert( is_node_getting_a_safepoint(mach),  "logic does not match; false negative");

 #endif

   int safepoint_pc_offset = current_offset;

   bool is_method_handle_invoke = false;

   bool return_oop = false;

   // Add the safepoint in the DebugInfoRecorder

   if( !mach->is_MachCall() ) {

     mcall = NULL;

     debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);

   } else {

     mcall = mach->as_MachCall();

     // Is the call a MethodHandle call?

     if (mcall->is_MachCallJava()) {

       if (mcall->as_MachCallJava()->_method_handle_invoke) {

         assert(has_method_handle_invokes(), "must have been set during call generation");

         is_method_handle_invoke = true;

       }

     }

     // Check if a call returns an object.

     if (mcall->return_value_is_used() &&

         mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) {

       return_oop = true;

     }

     safepoint_pc_offset += mcall->ret_addr_offset();

     debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);

   }

   // Loop over the JVMState list to add scope information

   // Do not skip safepoints with a NULL method, they need monitor info

   JVMState* youngest_jvms = sfn->jvms();

   int max_depth = youngest_jvms->depth();

   // Allocate the object pool for scalar-replaced objects -- the map from

   // small-integer keys (which can be recorded in the local and ostack

   // arrays) to descriptions of the object state.

   GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();

   // Visit scopes from oldest to youngest.

   for (int depth = 1; depth <= max_depth; depth++) {

     JVMState* jvms = youngest_jvms->of_depth(depth);

     int idx;

     ciMethod* method = jvms->has_method() ? jvms->method() : NULL;

     // Safepoints that do not have method() set only provide oop-map and monitor info

     // to support GC; these do not support deoptimization.

     int num_locs = (method == NULL) ? 0 : jvms->loc_size();

     int num_exps = (method == NULL) ? 0 : jvms->stk_size();

     int num_mon  = jvms->nof_monitors();

     assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),

            "JVMS local count must match that of the method");

     // Add Local and Expression Stack Information

     // Insert locals into the locarray

     GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);

     for( idx = 0; idx < num_locs; idx++ ) {

       FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );

     }

     // Insert expression stack entries into the exparray

     GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);

     for( idx = 0; idx < num_exps; idx++ ) {

       FillLocArray( idx,  sfn, sfn->stack(jvms, idx), exparray, objs );

     }

     // Add in mappings of the monitors

     assert( !method ||

             !method->is_synchronized() ||

             method->is_native() ||

             num_mon > 0 ||

             !GenerateSynchronizationCode,

             "monitors must always exist for synchronized methods");

     // Build the growable array of ScopeValues for exp stack

     GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);

     // Loop over monitors and insert into array

     for(idx = 0; idx < num_mon; idx++) {

       // Grab the node that defines this monitor

       Node* box_node = sfn->monitor_box(jvms, idx);

       Node* obj_node = sfn->monitor_obj(jvms, idx);

       // Create ScopeValue for object

       ScopeValue *scval = NULL;

       if( obj_node->is_SafePointScalarObject() ) {

         SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();

         scval = Compile::sv_for_node_id(objs, spobj->_idx);

         if (scval == NULL) {

           const Type *t = obj_node->bottom_type();

           ciKlass* cik = t->is_oopptr()->klass();

           assert(cik->is_instance_klass() ||

                  cik->is_array_klass(), "Not supported allocation.");

           ObjectValue* sv = new ObjectValue(spobj->_idx,

                                 new ConstantOopWriteValue(cik->constant_encoding()));

           Compile::set_sv_for_object_node(objs, sv);

           uint first_ind = spobj->first_index();

           for (uint i = 0; i < spobj->n_fields(); i++) {

             Node* fld_node = sfn->in(first_ind+i);

             (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);

           }

           scval = sv;

         }

       } else if( !obj_node->is_Con() ) {

         OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);

         if( obj_node->bottom_type()->base() == Type::NarrowOop ) {

           scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );

         } else {

           scval = new_loc_value( _regalloc, obj_reg, Location::oop );

         }

       } else {

         const TypePtr *tp = obj_node->bottom_type()->make_ptr();

         scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());

       }

       OptoReg::Name box_reg = BoxLockNode::stack_slot(box_node);

       Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));

       while( !box_node->is_BoxLock() )  box_node = box_node->in(1);

       monarray->append(new MonitorValue(scval, basic_lock, box_node->as_BoxLock()->is_eliminated()));

     }

     // We dump the object pool first, since deoptimization reads it in first.

     debug_info()->dump_object_pool(objs);

     // Build first class objects to pass to scope

     DebugToken *locvals = debug_info()->create_scope_values(locarray);

     DebugToken *expvals = debug_info()->create_scope_values(exparray);

     DebugToken *monvals = debug_info()->create_monitor_values(monarray);

     // Make method available for all Safepoints

     ciMethod* scope_method = method ? method : _method;

     // Describe the scope here

     assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");

     assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");

     // Now we can describe the scope.

     debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);

   } // End jvms loop

   // Mark the end of the scope set.

   debug_info()->end_safepoint(safepoint_pc_offset);

 }

 // A simplified version of Process_OopMap_Node, to handle non-safepoints.

 class NonSafepointEmitter {

   Compile*  C;

   JVMState* _pending_jvms;

   int       _pending_offset;

   void emit_non_safepoint();

  public:

   NonSafepointEmitter(Compile* compile) {

     this->C = compile;

     _pending_jvms = NULL;

     _pending_offset = 0;

   }

   void observe_instruction(Node* n, int pc_offset) {

     if (!C->debug_info()->recording_non_safepoints())  return;

     Node_Notes* nn = C->node_notes_at(n->_idx);

     if (nn == NULL || nn->jvms() == NULL)  return;

     if (_pending_jvms != NULL &&

         _pending_jvms->same_calls_as(nn->jvms())) {

       // Repeated JVMS?  Stretch it up here.

       _pending_offset = pc_offset;

     } else {

       if (_pending_jvms != NULL &&

           _pending_offset < pc_offset) {

         emit_non_safepoint();

       }

       _pending_jvms = NULL;

       if (pc_offset > C->debug_info()->last_pc_offset()) {

         // This is the only way _pending_jvms can become non-NULL:

         _pending_jvms = nn->jvms();

         _pending_offset = pc_offset;

       }

     }

   }

   // Stay out of the way of real safepoints:

   void observe_safepoint(JVMState* jvms, int pc_offset) {

     if (_pending_jvms != NULL &&

         !_pending_jvms->same_calls_as(jvms) &&

         _pending_offset < pc_offset) {

       emit_non_safepoint();

     }

     _pending_jvms = NULL;

   }

   void flush_at_end() {

     if (_pending_jvms != NULL) {

       emit_non_safepoint();

     }

     _pending_jvms = NULL;

   }

 };

 void NonSafepointEmitter::emit_non_safepoint() {

   JVMState* youngest_jvms = _pending_jvms;

   int       pc_offset     = _pending_offset;

   // Clear it now:

   _pending_jvms = NULL;

   DebugInformationRecorder* debug_info = C->debug_info();

   assert(debug_info->recording_non_safepoints(), "sanity");

   debug_info->add_non_safepoint(pc_offset);

   int max_depth = youngest_jvms->depth();

   // Visit scopes from oldest to youngest.

   for (int depth = 1; depth <= max_depth; depth++) {

     JVMState* jvms = youngest_jvms->of_depth(depth);

     ciMethod* method = jvms->has_method() ? jvms->method() : NULL;

     assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");

     debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());

   }

   // Mark the end of the scope set.

   debug_info->end_non_safepoint(pc_offset);

 }

 // helper for fill_buffer bailout logic

 static void turn_off_compiler(Compile* C) {

   if (CodeCache::largest_free_block() >= CodeCacheMinimumFreeSpace*10) {

     // Do not turn off compilation if a single giant method has

     // blown the code cache size.

     C->record_failure("excessive request to CodeCache");

   } else {

     // Let CompilerBroker disable further compilations.

     C->record_failure("CodeCache is full");

   }

 }

 //------------------------------init_buffer------------------------------------

 CodeBuffer* Compile::init_buffer(uint* blk_starts) {

   // Set the initially allocated size

   int  code_req   = initial_code_capacity;

   int  locs_req   = initial_locs_capacity;

   int  stub_req   = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;

   int  const_req  = initial_const_capacity;

   int  pad_req    = NativeCall::instruction_size;

   // The extra spacing after the code is necessary on some platforms.

   // Sometimes we need to patch in a jump after the last instruction,

   // if the nmethod has been deoptimized.  (See 4932387, 4894843.)

   // Compute the byte offset where we can store the deopt pc.

   if (fixed_slots() != 0) {

     _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));

   }

   // Compute prolog code size

   _method_size = 0;

   _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;

 #ifdef IA64

   if (save_argument_registers()) {

     // 4815101: this is a stub with implicit and unknown precision fp args.

     // The usual spill mechanism can only generate stfd's in this case, which

     // doesn't work if the fp reg to spill contains a single-precision denorm.

     // Instead, we hack around the normal spill mechanism using stfspill's and

     // ldffill's in the MachProlog and MachEpilog emit methods.  We allocate

     // space here for the fp arg regs (f8-f15) we're going to thusly spill.

     //

     // If we ever implement 16-byte 'registers' == stack slots, we can

     // get rid of this hack and have SpillCopy generate stfspill/ldffill

     // instead of stfd/stfs/ldfd/ldfs.

     _frame_slots += 8*(16/BytesPerInt);

   }

 #endif

   assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");

   if (has_mach_constant_base_node()) {

     // Fill the constant table.

     // Note:  This must happen before shorten_branches.

     for (uint i = 0; i < _cfg->_num_blocks; i++) {

       Block* b = _cfg->_blocks[i];

       for (uint j = 0; j < b->_nodes.size(); j++) {

         Node* n = b->_nodes[j];

         // If the node is a MachConstantNode evaluate the constant

         // value section.

         if (n->is_MachConstant()) {

           MachConstantNode* machcon = n->as_MachConstant();

           machcon->eval_constant(C);

         }

       }

     }

     // Calculate the offsets of the constants and the size of the

     // constant table (including the padding to the next section).

     constant_table().calculate_offsets_and_size();

     const_req = constant_table().size();

   }

   // Initialize the space for the BufferBlob used to find and verify

   // instruction size in MachNode::emit_size()

   init_scratch_buffer_blob(const_req);

   if (failing())  return NULL; // Out of memory

   // Pre-compute the length of blocks and replace

   // long branches with short if machine supports it.

   shorten_branches(blk_starts, code_req, locs_req, stub_req);

   // nmethod and CodeBuffer count stubs & constants as part of method's code.

   int exception_handler_req = size_exception_handler();

   int deopt_handler_req = size_deopt_handler();

   exception_handler_req += MAX_stubs_size; // add marginal slop for handler

   deopt_handler_req += MAX_stubs_size; // add marginal slop for handler

   stub_req += MAX_stubs_size;   // ensure per-stub margin

   code_req += MAX_inst_size;    // ensure per-instruction margin

   if (StressCodeBuffers)

     code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10;  // force expansion

   int total_req =

     const_req +

     code_req +

     pad_req +

     stub_req +

     exception_handler_req +

     deopt_handler_req;               // deopt handler

   if (has_method_handle_invokes())

     total_req += deopt_handler_req;  // deopt MH handler

   CodeBuffer* cb = code_buffer();

   cb->initialize(total_req, locs_req);

   // Have we run out of code space?

   if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {

     turn_off_compiler(this);

     return NULL;

   }

   // Configure the code buffer.

   cb->initialize_consts_size(const_req);

   cb->initialize_stubs_size(stub_req);

   cb->initialize_oop_recorder(env()->oop_recorder());

   // fill in the nop array for bundling computations

   MachNode *_nop_list[Bundle::_nop_count];

   Bundle::initialize_nops(_nop_list, this);

   return cb;

 }

 //------------------------------fill_buffer------------------------------------

 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {

   // blk_starts[] contains offsets calculated during short branches processing,

   // offsets should not be increased during following steps.

   // Compute the size of first NumberOfLoopInstrToAlign instructions at head

   // of a loop. It is used to determine the padding for loop alignment.

   compute_loop_first_inst_sizes();

   // Create oopmap set.

   _oop_map_set = new OopMapSet();

   // !!!!! This preserves old handling of oopmaps for now

   debug_info()->set_oopmaps(_oop_map_set);

   uint nblocks  = _cfg->_num_blocks;

   // Count and start of implicit null check instructions

   uint inct_cnt = 0;

   uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);

   // Count and start of calls

   uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);

   uint  return_offset = 0;

   int nop_size = (new (this) MachNopNode())->size(_regalloc);

   int previous_offset = 0;

   int current_offset  = 0;

   int last_call_offset = -1;

   int last_avoid_back_to_back_offset = -1;

 #ifdef ASSERT

   int block_alignment_padding = 0;

   uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);

   uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);

   uint* jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);

   uint* jmp_rule   = NEW_RESOURCE_ARRAY(uint,nblocks);

 #endif

   // Create an array of unused labels, one for each basic block, if printing is enabled

 #ifndef PRODUCT

   int *node_offsets      = NULL;

   uint node_offset_limit = unique();

   if (print_assembly())

     node_offsets         = NEW_RESOURCE_ARRAY(int, node_offset_limit);

 #endif

   NonSafepointEmitter non_safepoints(this);  // emit non-safepoints lazily

   // Emit the constant table.

   if (has_mach_constant_base_node()) {

     constant_table().emit(*cb);

   }

   // Create an array of labels, one for each basic block

   Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);

   for (uint i=0; i <= nblocks; i++) {

     blk_labels[i].init();

   }

   // ------------------

   // Now fill in the code buffer

   Node *delay_slot = NULL;

   for (uint i=0; i < nblocks; i++) {

     guarantee(blk_starts[i] >= (uint)cb->insts_size(),"should not increase size");

     Block *b = _cfg->_blocks[i];

     Node *head = b->head();

     // If this block needs to start aligned (i.e, can be reached other

     // than by falling-thru from the previous block), then force the

     // start of a new bundle.

     if (Pipeline::requires_bundling() && starts_bundle(head))

       cb->flush_bundle(true);

 #ifdef ASSERT

     if (!b->is_connector()) {

       stringStream st;

       b->dump_head(&_cfg->_bbs, &st);

       MacroAssembler(cb).block_comment(st.as_string());

     }

     jmp_target[i] = 0;

     jmp_offset[i] = 0;

     jmp_size[i]   = 0;

     jmp_rule[i]   = 0;

     // Maximum alignment padding for loop block was used

     // during first round of branches shortening, as result

     // padding for nodes (sfpt after call) was not added.

     // Take this into account for block's size change check

     // and allow increase block's size by the difference

     // of maximum and actual alignment paddings.

     int orig_blk_size = blk_starts[i+1] - blk_starts[i] + block_alignment_padding;

 #endif

     int blk_offset = current_offset;

     // Define the label at the beginning of the basic block

     MacroAssembler(cb).bind(blk_labels[b->_pre_order]);

     uint last_inst = b->_nodes.size();

     // Emit block normally, except for last instruction.

     // Emit means "dump code bits into code buffer".

     for (uint j = 0; j<last_inst; j++) {

       // Get the node

       Node* n = b->_nodes[j];

       // See if delay slots are supported

       if (valid_bundle_info(n) &&

           node_bundling(n)->used_in_unconditional_delay()) {

         assert(delay_slot == NULL, "no use of delay slot node");

         assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");

         delay_slot = n;

         continue;

       }

       // If this starts a new instruction group, then flush the current one

       // (but allow split bundles)

       if (Pipeline::requires_bundling() && starts_bundle(n))

         cb->flush_bundle(false);

       // The following logic is duplicated in the code ifdeffed for

       // ENABLE_ZAP_DEAD_LOCALS which appears above in this file.  It

       // should be factored out.  Or maybe dispersed to the nodes?

       // Special handling for SafePoint/Call Nodes

       bool is_mcall = false;

       if (n->is_Mach()) {

         MachNode *mach = n->as_Mach();

         is_mcall = n->is_MachCall();

         bool is_sfn = n->is_MachSafePoint();

         // If this requires all previous instructions be flushed, then do so

         if (is_sfn || is_mcall || mach->alignment_required() != 1) {

           cb->flush_bundle(true);

           current_offset = cb->insts_size();

         }

         // A padding may be needed again since a previous instruction

         // could be moved to delay slot.

         // align the instruction if necessary

         int padding = mach->compute_padding(current_offset);

         // Make sure safepoint node for polling is distinct from a call's

         // return by adding a nop if needed.

         if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {

           padding = nop_size;

         }

         if (padding == 0 && mach->avoid_back_to_back() &&

             current_offset == last_avoid_back_to_back_offset) {

           // Avoid back to back some instructions.

           padding = nop_size;

         }

         if(padding > 0) {

           assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");

           int nops_cnt = padding / nop_size;

           MachNode *nop = new (this) MachNopNode(nops_cnt);

           b->_nodes.insert(j++, nop);

           last_inst++;

           _cfg->_bbs.map( nop->_idx, b );

           nop->emit(*cb, _regalloc);

           cb->flush_bundle(true);

           current_offset = cb->insts_size();

         }

         // Remember the start of the last call in a basic block

         if (is_mcall) {

           MachCallNode *mcall = mach->as_MachCall();

           // This destination address is NOT PC-relative

           mcall->method_set((intptr_t)mcall->entry_point());

           // Save the return address

           call_returns[b->_pre_order] = current_offset + mcall->ret_addr_offset();

           if (mcall->is_MachCallLeaf()) {

             is_mcall = false;

             is_sfn = false;

           }

         }

         // sfn will be valid whenever mcall is valid now because of inheritance

         if (is_sfn || is_mcall) {

           // Handle special safepoint nodes for synchronization

           if (!is_mcall) {

             MachSafePointNode *sfn = mach->as_MachSafePoint();

             // !!!!! Stubs only need an oopmap right now, so bail out

             if (sfn->jvms()->method() == NULL) {

               // Write the oopmap directly to the code blob??!!

 #             ifdef ENABLE_ZAP_DEAD_LOCALS

               assert( !is_node_getting_a_safepoint(sfn),  "logic does not match; false positive");

 #             endif

               continue;

             }

           } // End synchronization

           non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),

                                            current_offset);

           Process_OopMap_Node(mach, current_offset);

         } // End if safepoint

         // If this is a null check, then add the start of the previous instruction to the list

         else if( mach->is_MachNullCheck() ) {

           inct_starts[inct_cnt++] = previous_offset;

         }

         // If this is a branch, then fill in the label with the target BB's label

         else if (mach->is_MachBranch()) {

           // This requires the TRUE branch target be in succs[0]

           uint block_num = b->non_connector_successor(0)->_pre_order;

           // Try to replace long branch if delay slot is not used,

           // it is mostly for back branches since forward branch's

           // distance is not updated yet.

           bool delay_slot_is_used = valid_bundle_info(n) &&

                                     node_bundling(n)->use_unconditional_delay();

           if (!delay_slot_is_used && mach->may_be_short_branch()) {

            assert(delay_slot == NULL, "not expecting delay slot node");

            int br_size = n->size(_regalloc);

             int offset = blk_starts[block_num] - current_offset;

             if (block_num >= i) {

               // Current and following block's offset are not

               // finilized yet, adjust distance by the difference

               // between calculated and final offsets of current block.

               offset -= (blk_starts[i] - blk_offset);

             }

             // In the following code a nop could be inserted before

             // the branch which will increase the backward distance.

             bool needs_padding = (current_offset == last_avoid_back_to_back_offset);

             if (needs_padding && offset <= 0)

               offset -= nop_size;

             if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {

               // We've got a winner.  Replace this branch.

               MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);

               // Update the jmp_size.

               int new_size = replacement->size(_regalloc);

               assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");

               // Insert padding between avoid_back_to_back branches.

               if (needs_padding && replacement->avoid_back_to_back()) {

                 MachNode *nop = new (this) MachNopNode();

                 b->_nodes.insert(j++, nop);

                 _cfg->_bbs.map(nop->_idx, b);

                 last_inst++;

                 nop->emit(*cb, _regalloc);

                 cb->flush_bundle(true);

                 current_offset = cb->insts_size();

               }

 #ifdef ASSERT

               jmp_target[i] = block_num;

               jmp_offset[i] = current_offset - blk_offset;

               jmp_size[i]   = new_size;

               jmp_rule[i]   = mach->rule();

 #endif

               b->_nodes.map(j, replacement);

               mach->subsume_by(replacement);

               n    = replacement;

               mach = replacement;

             }

           }

           mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );

         } else if (mach->ideal_Opcode() == Op_Jump) {

           for (uint h = 0; h < b->_num_succs; h++) {

             Block* succs_block = b->_succs[h];

             for (uint j = 1; j < succs_block->num_preds(); j++) {

               Node* jpn = succs_block->pred(j);

               if (jpn->is_JumpProj() && jpn->in(0) == mach) {

                 uint block_num = succs_block->non_connector()->_pre_order;

                 Label *blkLabel = &blk_labels[block_num];

                 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);

               }

             }

           }

         }

 #ifdef ASSERT

         // Check that oop-store precedes the card-mark

         else if (mach->ideal_Opcode() == Op_StoreCM) {

           uint storeCM_idx = j;

           int count = 0;

           for (uint prec = mach->req(); prec < mach->len(); prec++) {

             Node *oop_store = mach->in(prec);  // Precedence edge

             if (oop_store == NULL) continue;

             count++;

             uint i4;

             for( i4 = 0; i4 < last_inst; ++i4 ) {

               if( b->_nodes[i4] == oop_store ) break;

             }

             // Note: This test can provide a false failure if other precedence

             // edges have been added to the storeCMNode.

             assert( i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");

           }

           assert(count > 0, "storeCM expects at least one precedence edge");

         }

 #endif

         else if (!n->is_Proj()) {

           // Remember the beginning of the previous instruction, in case

           // it's followed by a flag-kill and a null-check.  Happens on

           // Intel all the time, with add-to-memory kind of opcodes.

           previous_offset = current_offset;

         }

       }

       // Verify that there is sufficient space remaining

       cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);

       if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {

         turn_off_compiler(this);

         return;

       }

       // Save the offset for the listing

 #ifndef PRODUCT

       if (node_offsets && n->_idx < node_offset_limit)

         node_offsets[n->_idx] = cb->insts_size();

 #endif

       // "Normal" instruction case

       DEBUG_ONLY( uint instr_offset = cb->insts_size(); )

       n->emit(*cb, _regalloc);

       current_offset  = cb->insts_size();

 #ifdef ASSERT

       if (n->size(_regalloc) < (current_offset-instr_offset)) {

         n->dump();

         assert(false, "wrong size of mach node");

       }

 #endif

       non_safepoints.observe_instruction(n, current_offset);

       // mcall is last "call" that can be a safepoint

       // record it so we can see if a poll will directly follow it

       // in which case we'll need a pad to make the PcDesc sites unique

       // see  5010568. This can be slightly inaccurate but conservative

       // in the case that return address is not actually at current_offset.

       // This is a small price to pay.

       if (is_mcall) {

         last_call_offset = current_offset;

       }

       if (n->is_Mach() && n->as_Mach()->avoid_back_to_back()) {

         // Avoid back to back some instructions.

         last_avoid_back_to_back_offset = current_offset;

       }

       // See if this instruction has a delay slot

       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {

         assert(delay_slot != NULL, "expecting delay slot node");

         // Back up 1 instruction

         cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());

         // Save the offset for the listing

 #ifndef PRODUCT

         if (node_offsets && delay_slot->_idx < node_offset_limit)

           node_offsets[delay_slot->_idx] = cb->insts_size();

 #endif

         // Support a SafePoint in the delay slot

         if (delay_slot->is_MachSafePoint()) {

           MachNode *mach = delay_slot->as_Mach();

           // !!!!! Stubs only need an oopmap right now, so bail out

           if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {

             // Write the oopmap directly to the code blob??!!

 #           ifdef ENABLE_ZAP_DEAD_LOCALS

             assert( !is_node_getting_a_safepoint(mach),  "logic does not match; false positive");

 #           endif

             delay_slot = NULL;

             continue;

           }

           int adjusted_offset = current_offset - Pipeline::instr_unit_size();

           non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),

                                            adjusted_offset);

           // Generate an OopMap entry

           Process_OopMap_Node(mach, adjusted_offset);

         }

         // Insert the delay slot instruction

         delay_slot->emit(*cb, _regalloc);

         // Don't reuse it

         delay_slot = NULL;

       }

     } // End for all instructions in block

     assert((uint)blk_offset <= blk_starts[i], "shouldn't increase distance");

     blk_starts[i] = blk_offset;

     // If the next block is the top of a loop, pad this block out to align

     // the loop top a little. Helps prevent pipe stalls at loop back branches.

     if (i < nblocks-1) {

       Block *nb = _cfg->_blocks[i+1];

       int padding = nb->alignment_padding(current_offset);

       if( padding > 0 ) {

         MachNode *nop = new (this) MachNopNode(padding / nop_size);

         b->_nodes.insert( b->_nodes.size(), nop );

         _cfg->_bbs.map( nop->_idx, b );

         nop->emit(*cb, _regalloc);

         current_offset = cb->insts_size();

       }

 #ifdef ASSERT

       int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();

       block_alignment_padding = (max_loop_pad - padding);

       assert(block_alignment_padding >= 0, "sanity");

 #endif

     }

     // Verify that the distance for generated before forward

     // short branches is still valid.

     assert(orig_blk_size >= (current_offset - blk_offset), "shouldn't increase block size");

   } // End of for all blocks

   blk_starts[nblocks] = current_offset;

   non_safepoints.flush_at_end();

   // Offset too large?

   if (failing())  return;

   // Define a pseudo-label at the end of the code

   MacroAssembler(cb).bind( blk_labels[nblocks] );

   // Compute the size of the first block

   _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();

   assert(cb->insts_size() < 500000, "method is unreasonably large");

 #ifdef ASSERT

   for (uint i = 0; i < nblocks; i++) { // For all blocks

     if (jmp_target[i] != 0) {

       int br_size = jmp_size[i];

       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);

       if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {

         tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);

         assert(false, "Displacement too large for short jmp");

       }

     }

   }

 #endif

   // ------------------

 #ifndef PRODUCT

   // Information on the size of the method, without the extraneous code

   Scheduling::increment_method_size(cb->insts_size());

 #endif

   // ------------------

   // Fill in exception table entries.

   FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);

   // Only java methods have exception handlers and deopt handlers

   if (_method) {

     // Emit the exception handler code.

     _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb));

     // Emit the deopt handler code.

     _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb));

     // Emit the MethodHandle deopt handler code (if required).

     if (has_method_handle_invokes()) {

       // We can use the same code as for the normal deopt handler, we

       // just need a different entry point address.

       _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb));

     }

   }

   // One last check for failed CodeBuffer::expand:

   if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {

     turn_off_compiler(this);

     return;

   }

 #ifndef PRODUCT

   // Dump the assembly code, including basic-block numbers

   if (print_assembly()) {

     ttyLocker ttyl;  // keep the following output all in one block

     if (!VMThread::should_terminate()) {  // test this under the tty lock

       // This output goes directly to the tty, not the compiler log.

       // To enable tools to match it up with the compilation activity,

       // be sure to tag this tty output with the compile ID.

       if (xtty != NULL) {

         xtty->head("opto_assembly compile_id='%d'%s", compile_id(),

                    is_osr_compilation()    ? " compile_kind='osr'" :

                    "");

       }

       if (method() != NULL) {

         method()->print_oop();

         print_codes();

       }

       dump_asm(node_offsets, node_offset_limit);

       if (xtty != NULL) {

         xtty->tail("opto_assembly");

       }

     }

   }

 #endif

 }

 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {

   _inc_table.set_size(cnt);

   uint inct_cnt = 0;

   for( uint i=0; i<_cfg->_num_blocks; i++ ) {

     Block *b = _cfg->_blocks[i];

     Node *n = NULL;

     int j;

     // Find the branch; ignore trailing NOPs.

     for( j = b->_nodes.size()-1; j>=0; j-- ) {

       n = b->_nodes[j];

       if( !n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con )

         break;

     }

     // If we didn't find anything, continue

     if( j < 0 ) continue;

     // Compute ExceptionHandlerTable subtable entry and add it

     // (skip empty blocks)

     if( n->is_Catch() ) {

       // Get the offset of the return from the call

       uint call_return = call_returns[b->_pre_order];

 #ifdef ASSERT

       assert( call_return > 0, "no call seen for this basic block" );

       while( b->_nodes[--j]->is_MachProj() ) ;

       assert( b->_nodes[j]->is_MachCall(), "CatchProj must follow call" );

 #endif

       // last instruction is a CatchNode, find it's CatchProjNodes

       int nof_succs = b->_num_succs;

       // allocate space

       GrowableArray<intptr_t> handler_bcis(nof_succs);

       GrowableArray<intptr_t> handler_pcos(nof_succs);

       // iterate through all successors

       for (int j = 0; j < nof_succs; j++) {

         Block* s = b->_succs[j];

         bool found_p = false;

         for( uint k = 1; k < s->num_preds(); k++ ) {

           Node *pk = s->pred(k);

           if( pk->is_CatchProj() && pk->in(0) == n ) {

             const CatchProjNode* p = pk->as_CatchProj();

             found_p = true;

             // add the corresponding handler bci & pco information

             if( p->_con != CatchProjNode::fall_through_index ) {

               // p leads to an exception handler (and is not fall through)

               assert(s == _cfg->_blocks[s->_pre_order],"bad numbering");

               // no duplicates, please

               if( !handler_bcis.contains(p->handler_bci()) ) {

                 uint block_num = s->non_connector()->_pre_order;

                 handler_bcis.append(p->handler_bci());

                 handler_pcos.append(blk_labels[block_num].loc_pos());

               }

             }

           }

         }

         assert(found_p, "no matching predecessor found");

         // Note:  Due to empty block removal, one block may have

         // several CatchProj inputs, from the same Catch.

       }

       // Set the offset of the return from the call

       _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);

       continue;

     }

     // Handle implicit null exception table updates

     if( n->is_MachNullCheck() ) {

       uint block_num = b->non_connector_successor(0)->_pre_order;

       _inc_table.append( inct_starts[inct_cnt++], blk_labels[block_num].loc_pos() );

       continue;

     }

   } // End of for all blocks fill in exception table entries

 }

 // Static Variables

 #ifndef PRODUCT

 uint Scheduling::_total_nop_size = 0;

 uint Scheduling::_total_method_size = 0;

 uint Scheduling::_total_branches = 0;

 uint Scheduling::_total_unconditional_delays = 0;

 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];

 #endif

 // Initializer for class Scheduling

 Scheduling::Scheduling(Arena *arena, Compile &compile)

   : _arena(arena),

     _cfg(compile.cfg()),

     _bbs(compile.cfg()->_bbs),

     _regalloc(compile.regalloc()),

     _reg_node(arena),

     _bundle_instr_count(0),

     _bundle_cycle_number(0),

     _scheduled(arena),

     _available(arena),

     _next_node(NULL),

     _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),

     _pinch_free_list(arena)

 #ifndef PRODUCT

   , _branches(0)

   , _unconditional_delays(0)

 #endif

 {

   // Create a MachNopNode

   _nop = new (&compile) MachNopNode();

   // Now that the nops are in the array, save the count

   // (but allow entries for the nops)

   _node_bundling_limit = compile.unique();

   uint node_max = _regalloc->node_regs_max_index();

   compile.set_node_bundling_limit(_node_bundling_limit);

   // This one is persistent within the Compile class

   _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);

   // Allocate space for fixed-size arrays

   _node_latency    = NEW_ARENA_ARRAY(arena, unsigned short, node_max);

   _uses            = NEW_ARENA_ARRAY(arena, short,          node_max);

   _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);

   // Clear the arrays

   memset(_node_bundling_base, 0, node_max * sizeof(Bundle));

   memset(_node_latency,       0, node_max * sizeof(unsigned short));

   memset(_uses,               0, node_max * sizeof(short));

   memset(_current_latency,    0, node_max * sizeof(unsigned short));

   // Clear the bundling information

   memcpy(_bundle_use_elements,

     Pipeline_Use::elaborated_elements,

     sizeof(Pipeline_Use::elaborated_elements));

   // Get the last node

   Block *bb = _cfg->_blocks[_cfg->_blocks.size()-1];

   _next_node = bb->_nodes[bb->_nodes.size()-1];

 }

 #ifndef PRODUCT

 // Scheduling destructor

 Scheduling::~Scheduling() {

   _total_branches             += _branches;

   _total_unconditional_delays += _unconditional_delays;

 }

 #endif

 // Step ahead "i" cycles

 void Scheduling::step(uint i) {

   Bundle *bundle = node_bundling(_next_node);

   bundle->set_starts_bundle();

   // Update the bundle record, but leave the flags information alone

   if (_bundle_instr_count > 0) {

     bundle->set_instr_count(_bundle_instr_count);

     bundle->set_resources_used(_bundle_use.resourcesUsed());

   }

   // Update the state information

   _bundle_instr_count = 0;

   _bundle_cycle_number += i;

   _bundle_use.step(i);

 }

 void Scheduling::step_and_clear() {

   Bundle *bundle = node_bundling(_next_node);

   bundle->set_starts_bundle();

   // Update the bundle record

   if (_bundle_instr_count > 0) {

     bundle->set_instr_count(_bundle_instr_count);

     bundle->set_resources_used(_bundle_use.resourcesUsed());

     _bundle_cycle_number += 1;

   }

   // Clear the bundling information

   _bundle_instr_count = 0;

   _bundle_use.reset();

   memcpy(_bundle_use_elements,

     Pipeline_Use::elaborated_elements,

     sizeof(Pipeline_Use::elaborated_elements));

 }

 //------------------------------ScheduleAndBundle------------------------------

 // Perform instruction scheduling and bundling over the sequence of

 // instructions in backwards order.

 void Compile::ScheduleAndBundle() {

   // Don't optimize this if it isn't a method

   if (!_method)

     return;

   // Don't optimize this if scheduling is disabled

   if (!do_scheduling())

     return;

   NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); )

   // Create a data structure for all the scheduling information

   Scheduling scheduling(Thread::current()->resource_area(), *this);

   // Walk backwards over each basic block, computing the needed alignment

   // Walk over all the basic blocks

   scheduling.DoScheduling();

 }

 //------------------------------ComputeLocalLatenciesForward-------------------

 // Compute the latency of all the instructions.  This is fairly simple,

 // because we already have a legal ordering.  Walk over the instructions

 // from first to last, and compute the latency of the instruction based

 // on the latency of the preceding instruction(s).

 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {

 #ifndef PRODUCT

   if (_cfg->C->trace_opto_output())

     tty->print("# -> ComputeLocalLatenciesForward\n");

 #endif

   // Walk over all the schedulable instructions

   for( uint j=_bb_start; j < _bb_end; j++ ) {

     // This is a kludge, forcing all latency calculations to start at 1.

     // Used to allow latency 0 to force an instruction to the beginning

     // of the bb

     uint latency = 1;

     Node *use = bb->_nodes[j];

     uint nlen = use->len();

     // Walk over all the inputs

     for ( uint k=0; k < nlen; k++ ) {

       Node *def = use->in(k);

       if (!def)

         continue;

       uint l = _node_latency[def->_idx] + use->latency(k);

       if (latency < l)

         latency = l;

     }

     _node_latency[use->_idx] = latency;

 #ifndef PRODUCT

     if (_cfg->C->trace_opto_output()) {

       tty->print("# latency %4d: ", latency);

       use->dump();

     }

 #endif

   }

 #ifndef PRODUCT

   if (_cfg->C->trace_opto_output())

     tty->print("# <- ComputeLocalLatenciesForward\n");

 #endif

 } // end ComputeLocalLatenciesForward

 // See if this node fits into the present instruction bundle

 bool Scheduling::NodeFitsInBundle(Node *n) {

   uint n_idx = n->_idx;

   // If this is the unconditional delay instruction, then it fits

   if (n == _unconditional_delay_slot) {

 #ifndef PRODUCT

     if (_cfg->C->trace_opto_output())

       tty->print("#     NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);

 #endif

     return (true);

   }

   // If the node cannot be scheduled this cycle, skip it

   if (_current_latency[n_idx] > _bundle_cycle_number) {

 #ifndef PRODUCT

     if (_cfg->C->trace_opto_output())

       tty->print("#     NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",

         n->_idx, _current_latency[n_idx], _bundle_cycle_number);

 #endif

     return (false);

   }

   const Pipeline *node_pipeline = n->pipeline();

   uint instruction_count = node_pipeline->instructionCount();

   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)

     instruction_count = 0;

   else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)

     instruction_count++;

   if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {

 #ifndef PRODUCT

     if (_cfg->C->trace_opto_output())

       tty->print("#     NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",

         n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);

 #endif

     return (false);

   }

   // Don't allow non-machine nodes to be handled this way

   if (!n->is_Mach() && instruction_count == 0)

     return (false);

   // See if there is any overlap

   uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());

   if (delay > 0) {

 #ifndef PRODUCT

     if (_cfg->C->trace_opto_output())

       tty->print("#     NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);

 #endif

     return false;

   }

 #ifndef PRODUCT

   if (_cfg->C->trace_opto_output())

     tty->print("#     NodeFitsInBundle [%4d]:  TRUE\n", n_idx);

 #endif

   return true;

 }

 Node * Scheduling::ChooseNodeToBundle() {

   uint siz = _available.size();

   if (siz == 0) {

 #ifndef PRODUCT

     if (_cfg->C->trace_opto_output())

       tty->print("#   ChooseNodeToBundle: NULL\n");

 #endif

     return (NULL);

   }

   // Fast path, if only 1 instruction in the bundle

   if (siz == 1) {

 #ifndef PRODUCT

     if (_cfg->C->trace_opto_output()) {

       tty->print("#   ChooseNodeToBundle (only 1): ");

       _available[0]->dump();

     }

 #endif

     return (_available[0]);

   }

   // Don't bother, if the bundle is already full

   if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {

     for ( uint i = 0; i < siz; i++ ) {

       Node *n = _available[i];

       // Skip projections, we'll handle them another way

       if (n->is_Proj())

         continue;

       // This presupposed that instructions are inserted into the

       // available list in a legality order; i.e. instructions that

       // must be inserted first are at the head of the list

       if (NodeFitsInBundle(n)) {

 #ifndef PRODUCT

         if (_cfg->C->trace_opto_output()) {

           tty->print("#   ChooseNodeToBundle: ");

           n->dump();

         }

 #endif

         return (n);

       }

     }

   }

   // Nothing fits in this bundle, choose the highest priority

 #ifndef PRODUCT

   if (_cfg->C->trace_opto_output()) {

     tty->print("#   ChooseNodeToBundle: ");

     _available[0]->dump();

   }

 #endif

   return _available[0];

 }

 //------------------------------AddNodeToAvailableList-------------------------

 void Scheduling::AddNodeToAvailableList(Node *n) {

   assert( !n->is_Proj(), "projections never directly made available" );

 #ifndef PRODUCT

   if (_cfg->C->trace_opto_output()) {

     tty->print("#   AddNodeToAvailableList: ");

     n->dump();

   }

 #endif

   int latency = _current_latency[n->_idx];

   // Insert in latency order (insertion sort)

   uint i;

   for ( i=0; i < _available.size(); i++ )

     if (_current_latency[_available[i]->_idx] > latency)

       break;

   // Special Check for compares following branches

   if( n->is_Mach() && _scheduled.size() > 0 ) {

     int op = n->as_Mach()->ideal_Opcode();

     Node *last = _scheduled[0];

     if( last->is_MachIf() && last->in(1) == n &&

         ( op == Op_CmpI ||

           op == Op_CmpU ||

           op == Op_CmpP ||

           op == Op_CmpF ||

           op == Op_CmpD ||

           op == Op_CmpL ) ) {

       // Recalculate position, moving to front of same latency

       for ( i=0 ; i < _available.size(); i++ )

         if (_current_latency[_available[i]->_idx] >= latency)

           break;

     }

   }

   // Insert the node in the available list

   _available.insert(i, n);

 #ifndef PRODUCT

   if (_cfg->C->trace_opto_output())

     dump_available();

 #endif

 }

 //------------------------------DecrementUseCounts-----------------------------

 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {

   for ( uint i=0; i < n->len(); i++ ) {

     Node *def = n->in(i);

     if (!def) continue;

     if( def->is_Proj() )        // If this is a machine projection, then

       def = def->in(0);         // propagate usage thru to the base instruction

     if( _bbs[def->_idx] != bb ) // Ignore if not block-local

       continue;

     // Compute the latency

     uint l = _bundle_cycle_number + n->latency(i);

     if (_current_latency[def->_idx] < l)

       _current_latency[def->_idx] = l;

     // If this does not have uses then schedule it

     if ((--_uses[def->_idx]) == 0)

       AddNodeToAvailableList(def);

   }

 }

 //------------------------------AddNodeToBundle--------------------------------

 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {

 #ifndef PRODUCT

   if (_cfg->C->trace_opto_output()) {

     tty->print("#   AddNodeToBundle: ");

     n->dump();

   }

 #endif

   // Remove this from the available list

   uint i;

   for (i = 0; i < _available.size(); i++)

     if (_available[i] == n)

       break;

   assert(i < _available.size(), "entry in _available list not found");

   _available.remove(i);

   // See if this fits in the current bundle

   const Pipeline *node_pipeline = n->pipeline();

   const Pipeline_Use& node_usage = node_pipeline->resourceUse();

   // Check for instructions to be placed in the delay slot. We

   // do this before we actually schedule the current instruction,

   // because the delay slot follows the current instruction.

   if (Pipeline::_branch_has_delay_slot &&

       node_pipeline->hasBranchDelay() &&

       !_unconditional_delay_slot) {

     uint siz = _available.size();

     // Conditional branches can support an instruction that

     // is unconditionally executed and not dependent by the

     // branch, OR a conditionally executed instruction if

     // the branch is taken.  In practice, this means that

     // the first instruction at the branch target is

     // copied to the delay slot, and the branch goes to

     // the instruction after that at the branch target

     if ( n->is_MachBranch() ) {

       assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );

       assert( !n->is_Catch(),         "should not look for delay slot for Catch" );

 #ifndef PRODUCT

       _branches++;

 #endif

       // At least 1 instruction is on the available list

       // that is not dependent on the branch

       for (uint i = 0; i < siz; i++) {

         Node *d = _available[i];

         const Pipeline *avail_pipeline = d->pipeline();

         // Don't allow safepoints in the branch shadow, that will

         // cause a number of difficulties

         if ( avail_pipeline->instructionCount() == 1 &&

             !avail_pipeline->hasMultipleBundles() &&

             !avail_pipeline->hasBranchDelay() &&

             Pipeline::instr_has_unit_size() &&

             d->size(_regalloc) == Pipeline::instr_unit_size() &&

             NodeFitsInBundle(d) &&

             !node_bundling(d)->used_in_delay()) {

           if (d->is_Mach() && !d->is_MachSafePoint()) {

             // A node that fits in the delay slot was found, so we need to

             // set the appropriate bits in the bundle pipeline information so

             // that it correctly indicates resource usage.  Later, when we

             // attempt to add this instruction to the bundle, we will skip

             // setting the resource usage.

             _unconditional_delay_slot = d;

             node_bundling(n)->set_use_unconditional_delay();

             node_bundling(d)->set_used_in_unconditional_delay();

             _bundle_use.add_usage(avail_pipeline->resourceUse());

             _current_latency[d->_idx] = _bundle_cycle_number;

             _next_node = d;

             ++_bundle_instr_count;

 #ifndef PRODUCT

             _unconditional_delays++;

 #endif

             break;

           }

         }

       }

     }

     // No delay slot, add a nop to the usage

     if (!_unconditional_delay_slot) {

       // See if adding an instruction in the delay slot will overflow

       // the bundle.

       if (!NodeFitsInBundle(_nop)) {

 #ifndef PRODUCT

         if (_cfg->C->trace_opto_output())

           tty->print("#  *** STEP(1 instruction for delay slot) ***\n");

 #endif

         step(1);

       }

       _bundle_use.add_usage(_nop->pipeline()->resourceUse());

       _next_node = _nop;

       ++_bundle_instr_count;

     }

     // See if the instruction in the delay slot requires a

     // step of the bundles

     if (!NodeFitsInBundle(n)) {

 #ifndef PRODUCT

         if (_cfg->C->trace_opto_output())

           tty->print("#  *** STEP(branch won't fit) ***\n");

 #endif

         // Update the state information

         _bundle_instr_count = 0;

         _bundle_cycle_number += 1;

         _bundle_use.step(1);

     }

   }

   // Get the number of instructions

   uint instruction_count = node_pipeline->instructionCount();

   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)

     instruction_count = 0;

   // Compute the latency information

   uint delay = 0;

   if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {

     int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;

     if (relative_latency < 0)

       relative_latency = 0;

     delay = _bundle_use.full_latency(relative_latency, node_usage);

     // Does not fit in this bundle, start a new one

     if (delay > 0) {

       step(delay);

 #ifndef PRODUCT

       if (_cfg->C->trace_opto_output())

         tty->print("#  *** STEP(%d) ***\n", delay);

 #endif

     }

   }

   // If this was placed in the delay slot, ignore it

   if (n != _unconditional_delay_slot) {

     if (delay == 0) {

       if (node_pipeline->hasMultipleBundles()) {

 #ifndef PRODUCT

         if (_cfg->C->trace_opto_output())

           tty->print("#  *** STEP(multiple instructions) ***\n");

 #endif

         step(1);

       }

       else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {

 #ifndef PRODUCT

         if (_cfg->C->trace_opto_output())

           tty->print("#  *** STEP(%d >= %d instructions) ***\n",

             instruction_count + _bundle_instr_count,

             Pipeline::_max_instrs_per_cycle);

 #endif

         step(1);

       }

     }

     if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)

       _bundle_instr_count++;

     // Set the node's latency

     _current_latency[n->_idx] = _bundle_cycle_number;

     // Now merge the functional unit information

     if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())

       _bundle_use.add_usage(node_usage);

     // Increment the number of instructions in this bundle

     _bundle_instr_count += instruction_count;

     // Remember this node for later

     if (n->is_Mach())

       _next_node = n;

   }

   // It's possible to have a BoxLock in the graph and in the _bbs mapping but

   // not in the bb->_nodes array.  This happens for debug-info-only BoxLocks.

   // 'Schedule' them (basically ignore in the schedule) but do not insert them

   // into the block.  All other scheduled nodes get put in the schedule here.

   int op = n->Opcode();

   if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR

       (op != Op_Node &&         // Not an unused antidepedence node and

        // not an unallocated boxlock

        (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {

     // Push any trailing projections

     if( bb->_nodes[bb->_nodes.size()-1] != n ) {

       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {

         Node *foi = n->fast_out(i);

         if( foi->is_Proj() )

           _scheduled.push(foi);

       }

     }

     // Put the instruction in the schedule list

     _scheduled.push(n);

   }

 #ifndef PRODUCT

   if (_cfg->C->trace_opto_output())

     dump_available();

 #endif

   // Walk all the definitions, decrementing use counts, and

   // if a definition has a 0 use count, place it in the available list.

   DecrementUseCounts(n,bb);

 }

 //------------------------------ComputeUseCount--------------------------------

 // This method sets the use count within a basic block.  We will ignore all

 // uses outside the current basic block.  As we are doing a backwards walk,

 // any node we reach that has a use count of 0 may be scheduled.  This also

 // avoids the problem of cyclic references from phi nodes, as long as phi

 // nodes are at the front of the basic block.  This method also initializes

 // the available list to the set of instructions that have no uses within this

 // basic block.

 void Scheduling::ComputeUseCount(const Block *bb) {

 #ifndef PRODUCT

   if (_cfg->C->trace_opto_output())

     tty->print("# -> ComputeUseCount\n");

 #endif

   // Clear the list of available and scheduled instructions, just in case

   _available.clear();

   _scheduled.clear();

   // No delay slot specified

   _unconditional_delay_slot = NULL;

 #ifdef ASSERT

   for( uint i=0; i < bb->_nodes.size(); i++ )

     assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" );

 #endif

   // Force the _uses count to never go to zero for unscheduable pieces

   // of the block

   for( uint k = 0; k < _bb_start; k++ )

     _uses[bb->_nodes[k]->_idx] = 1;

   for( uint l = _bb_end; l < bb->_nodes.size(); l++ )

     _uses[bb->_nodes[l]->_idx] = 1;

   // Iterate backwards over the instructions in the block.  Don't count the

   // branch projections at end or the block header instructions.

   for( uint j = _bb_end-1; j >= _bb_start; j-- ) {

     Node *n = bb->_nodes[j];

     if( n->is_Proj() ) continue; // Projections handled another way

     // Account for all uses

     for ( uint k = 0; k < n->len(); k++ ) {

       Node *inp = n->in(k);

       if (!inp) continue;

       assert(inp != n, "no cycles allowed" );

       if( _bbs[inp->_idx] == bb ) { // Block-local use?

         if( inp->is_Proj() )    // Skip through Proj's

           inp = inp->in(0);

         ++_uses[inp->_idx];     // Count 1 block-local use

       }

     }

     // If this instruction has a 0 use count, then it is available

     if (!_uses[n->_idx]) {

       _current_latency[n->_idx] = _bundle_cycle_number;

       AddNodeToAvailableList(n);

     }

 #ifndef PRODUCT

     if (_cfg->C->trace_opto_output()) {

       tty->print("#   uses: %3d: ", _uses[n->_idx]);

       n->dump();

     }

 #endif

   }

 #ifndef PRODUCT

   if (_cfg->C->trace_opto_output())

     tty->print("# <- ComputeUseCount\n");

 #endif

 }

 // This routine performs scheduling on each basic block in reverse order,

 // using instruction latencies and taking into account function unit

 // availability.

 void Scheduling::DoScheduling() {

 #ifndef PRODUCT

   if (_cfg->C->trace_opto_output())

     tty->print("# -> DoScheduling\n");

 #endif

   Block *succ_bb = NULL;

   Block *bb;

   // Walk over all the basic blocks in reverse order

   for( int i=_cfg->_num_blocks-1; i >= 0; succ_bb = bb, i-- ) {

     bb = _cfg->_blocks[i];

 #ifndef PRODUCT

     if (_cfg->C->trace_opto_output()) {

       tty->print("#  Schedule BB#%03d (initial)\n", i);

       for (uint j = 0; j < bb->_nodes.size(); j++)

         bb->_nodes[j]->dump();

     }

 #endif

     // On the head node, skip processing

     if( bb == _cfg->_broot )

       continue;

     // Skip empty, connector blocks

     if (bb->is_connector())

       continue;

     // If the following block is not the sole successor of

     // this one, then reset the pipeline information

     if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {

 #ifndef PRODUCT

       if (_cfg->C->trace_opto_output()) {

         tty->print("*** bundle start of next BB, node %d, for %d instructions\n",

                    _next_node->_idx, _bundle_instr_count);

       }

 #endif

       step_and_clear();

     }

     // Leave untouched the starting instruction, any Phis, a CreateEx node

     // or Top.  bb->_nodes[_bb_start] is the first schedulable instruction.

     _bb_end = bb->_nodes.size()-1;

     for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {

       Node *n = bb->_nodes[_bb_start];

       // Things not matched, like Phinodes and ProjNodes don't get scheduled.

       // Also, MachIdealNodes do not get scheduled

       if( !n->is_Mach() ) continue;     // Skip non-machine nodes

       MachNode *mach = n->as_Mach();

       int iop = mach->ideal_Opcode();

       if( iop == Op_CreateEx ) continue; // CreateEx is pinned

       if( iop == Op_Con ) continue;      // Do not schedule Top

       if( iop == Op_Node &&     // Do not schedule PhiNodes, ProjNodes

           mach->pipeline() == MachNode::pipeline_class() &&

           !n->is_SpillCopy() )  // Breakpoints, Prolog, etc

         continue;

       break;                    // Funny loop structure to be sure...

     }

     // Compute last "interesting" instruction in block - last instruction we

     // might schedule.  _bb_end points just after last schedulable inst.  We

     // normally schedule conditional branches (despite them being forced last

     // in the block), because they have delay slots we can fill.  Calls all

     // have their delay slots filled in the template expansions, so we don't

     // bother scheduling them.

     Node *last = bb->_nodes[_bb_end];

     // Ignore trailing NOPs.

     while (_bb_end > 0 && last->is_Mach() &&

            last->as_Mach()->ideal_Opcode() == Op_Con) {

       last = bb->_nodes[--_bb_end];

     }

     assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");

     if( last->is_Catch() ||

        // Exclude unreachable path case when Halt node is in a separate block.

        (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {

       // There must be a prior call.  Skip it.

       while( !bb->_nodes[--_bb_end]->is_MachCall() ) {

         assert( bb->_nodes[_bb_end]->is_MachProj(), "skipping projections after expected call" );

       }

     } else if( last->is_MachNullCheck() ) {

       // Backup so the last null-checked memory instruction is

       // outside the schedulable range. Skip over the nullcheck,

       // projection, and the memory nodes.

       Node *mem = last->in(1);

       do {

         _bb_end--;

       } while (mem != bb->_nodes[_bb_end]);

     } else {

       // Set _bb_end to point after last schedulable inst.

       _bb_end++;

     }

     assert( _bb_start <= _bb_end, "inverted block ends" );

     // Compute the register antidependencies for the basic block

     ComputeRegisterAntidependencies(bb);

     if (_cfg->C->failing())  return;  // too many D-U pinch points

     // Compute intra-bb latencies for the nodes

     ComputeLocalLatenciesForward(bb);

     // Compute the usage within the block, and set the list of all nodes

     // in the block that have no uses within the block.

     ComputeUseCount(bb);

     // Schedule the remaining instructions in the block

     while ( _available.size() > 0 ) {

       Node *n = ChooseNodeToBundle();

       AddNodeToBundle(n,bb);

     }

     assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );

 #ifdef ASSERT

     for( uint l = _bb_start; l < _bb_end; l++ ) {

       Node *n = bb->_nodes[l];

       uint m;

       for( m = 0; m < _bb_end-_bb_start; m++ )

         if( _scheduled[m] == n )

           break;

       assert( m < _bb_end-_bb_start, "instruction missing in schedule" );

     }

 #endif

     // Now copy the instructions (in reverse order) back to the block

     for ( uint k = _bb_start; k < _bb_end; k++ )

       bb->_nodes.map(k, _scheduled[_bb_end-k-1]);

 #ifndef PRODUCT

     if (_cfg->C->trace_opto_output()) {

       tty->print("#  Schedule BB#%03d (final)\n", i);

       uint current = 0;

       for (uint j = 0; j < bb->_nodes.size(); j++) {

         Node *n = bb->_nodes[j];

         if( valid_bundle_info(n) ) {

           Bundle *bundle = node_bundling(n);

           if (bundle->instr_count() > 0 || bundle->flags() > 0) {

             tty->print("*** Bundle: ");

             bundle->dump();

           }

           n->dump();

         }

       }

     }

 #endif

 #ifdef ASSERT

   verify_good_schedule(bb,"after block local scheduling");

 #endif

   }

 #ifndef PRODUCT

   if (_cfg->C->trace_opto_output())

     tty->print("# <- DoScheduling\n");

 #endif

   // Record final node-bundling array location

   _regalloc->C->set_node_bundling_base(_node_bundling_base);

 } // end DoScheduling

 //------------------------------verify_good_schedule---------------------------

 // Verify that no live-range used in the block is killed in the block by a

 // wrong DEF.  This doesn't verify live-ranges that span blocks.

 // Check for edge existence.  Used to avoid adding redundant precedence edges.

 static bool edge_from_to( Node *from, Node *to ) {

   for( uint i=0; i<from->len(); i++ )

     if( from->in(i) == to )

       return true;

   return false;

 }

 #ifdef ASSERT

 //------------------------------verify_do_def----------------------------------

 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {

   // Check for bad kills

   if( OptoReg::is_valid(def) ) { // Ignore stores & control flow

     Node *prior_use = _reg_node[def];

     if( prior_use && !edge_from_to(prior_use,n) ) {

       tty->print("%s = ",OptoReg::as_VMReg(def)->name());

       n->dump();

       tty->print_cr("...");

       prior_use->dump();

       assert(edge_from_to(prior_use,n),msg);

     }

     _reg_node.map(def,NULL); // Kill live USEs

   }

 }

 //------------------------------verify_good_schedule---------------------------

 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {

   // Zap to something reasonable for the verify code

   _reg_node.clear();

   // Walk over the block backwards.  Check to make sure each DEF doesn't

   // kill a live value (other than the one it's supposed to).  Add each

   // USE to the live set.

   for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) {

     Node *n = b->_nodes[i];

     int n_op = n->Opcode();

     if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {

       // Fat-proj kills a slew of registers

       RegMask rm = n->out_RegMask();// Make local copy

       while( rm.is_NotEmpty() ) {

         OptoReg::Name kill = rm.find_first_elem();

         rm.Remove(kill);

         verify_do_def( n, kill, msg );

       }

     } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes

       // Get DEF'd registers the normal way

       verify_do_def( n, _regalloc->get_reg_first(n), msg );

       verify_do_def( n, _regalloc->get_reg_second(n), msg );

     }

     // Now make all USEs live

     for( uint i=1; i<n->req(); i++ ) {

       Node *def = n->in(i);

       assert(def != 0, "input edge required");

       OptoReg::Name reg_lo = _regalloc->get_reg_first(def);

       OptoReg::Name reg_hi = _regalloc->get_reg_second(def);

       if( OptoReg::is_valid(reg_lo) ) {

         assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);

         _reg_node.map(reg_lo,n);

       }

       if( OptoReg::is_valid(reg_hi) ) {

         assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);

         _reg_node.map(reg_hi,n);

       }

     }

   }

   // Zap to something reasonable for the Antidependence code

   _reg_node.clear();