1 /*
   2  * Copyright (c) 1998, 2011, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/assembler.inline.hpp"
  27 #include "code/debugInfo.hpp"
  28 #include "code/debugInfoRec.hpp"
  29 #include "compiler/compileBroker.hpp"
  30 #include "compiler/oopMap.hpp"
  31 #include "memory/allocation.inline.hpp"
  32 #include "opto/callnode.hpp"
  33 #include "opto/cfgnode.hpp"
  34 #include "opto/locknode.hpp"
  35 #include "opto/machnode.hpp"
  36 #include "opto/output.hpp"
  37 #include "opto/regalloc.hpp"
  38 #include "opto/runtime.hpp"
  39 #include "opto/subnode.hpp"
  40 #include "opto/type.hpp"
  41 #include "runtime/handles.inline.hpp"
  42 #include "utilities/xmlstream.hpp"
  43 
  44 extern uint size_java_to_interp();
  45 extern uint reloc_java_to_interp();
  46 extern uint size_exception_handler();
  47 extern uint size_deopt_handler();
  48 
  49 #ifndef PRODUCT
  50 #define DEBUG_ARG(x) , x
  51 #else
  52 #define DEBUG_ARG(x)
  53 #endif
  54 
  55 extern int emit_exception_handler(CodeBuffer &cbuf);
  56 extern int emit_deopt_handler(CodeBuffer &cbuf);
  57 
  58 //------------------------------Output-----------------------------------------
  59 // Convert Nodes to instruction bits and pass off to the VM
  60 void Compile::Output() {
  61   // RootNode goes
  62   assert( _cfg->_broot->_nodes.size() == 0, "" );
  63 
  64   // The number of new nodes (mostly MachNop) is proportional to
  65   // the number of java calls and inner loops which are aligned.
  66   if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
  67                             C->inner_loops()*(OptoLoopAlignment-1)),
  68                            "out of nodes before code generation" ) ) {
  69     return;
  70   }
  71   // Make sure I can find the Start Node
  72   Block_Array& bbs = _cfg->_bbs;
  73   Block *entry = _cfg->_blocks[1];
  74   Block *broot = _cfg->_broot;
  75 
  76   const StartNode *start = entry->_nodes[0]->as_Start();
  77 
  78   // Replace StartNode with prolog
  79   MachPrologNode *prolog = new (this) MachPrologNode();
  80   entry->_nodes.map( 0, prolog );
  81   bbs.map( prolog->_idx, entry );
  82   bbs.map( start->_idx, NULL ); // start is no longer in any block
  83 
  84   // Virtual methods need an unverified entry point
  85 
  86   if( is_osr_compilation() ) {
  87     if( PoisonOSREntry ) {
  88       // TODO: Should use a ShouldNotReachHereNode...
  89       _cfg->insert( broot, 0, new (this) MachBreakpointNode() );
  90     }
  91   } else {
  92     if( _method && !_method->flags().is_static() ) {
  93       // Insert unvalidated entry point
  94       _cfg->insert( broot, 0, new (this) MachUEPNode() );
  95     }
  96 
  97   }
  98 
  99 
 100   // Break before main entry point
 101   if( (_method && _method->break_at_execute())
 102 #ifndef PRODUCT
 103     ||(OptoBreakpoint && is_method_compilation())
 104     ||(OptoBreakpointOSR && is_osr_compilation())
 105     ||(OptoBreakpointC2R && !_method)
 106 #endif
 107     ) {
 108     // checking for _method means that OptoBreakpoint does not apply to
 109     // runtime stubs or frame converters
 110     _cfg->insert( entry, 1, new (this) MachBreakpointNode() );
 111   }
 112 
 113   // Insert epilogs before every return
 114   for( uint i=0; i<_cfg->_num_blocks; i++ ) {
 115     Block *b = _cfg->_blocks[i];
 116     if( !b->is_connector() && b->non_connector_successor(0) == _cfg->_broot ) { // Found a program exit point?
 117       Node *m = b->end();
 118       if( m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt ) {
 119         MachEpilogNode *epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
 120         b->add_inst( epilog );
 121         bbs.map(epilog->_idx, b);
 122         //_regalloc->set_bad(epilog->_idx); // Already initialized this way.
 123       }
 124     }
 125   }
 126 
 127 # ifdef ENABLE_ZAP_DEAD_LOCALS
 128   if ( ZapDeadCompiledLocals )  Insert_zap_nodes();
 129 # endif
 130 
 131   uint* blk_starts = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks+1);
 132   blk_starts[0]    = 0;
 133 
 134   // Initialize code buffer and process short branches.
 135   CodeBuffer* cb = init_buffer(blk_starts);
 136 
 137   if (cb == NULL || failing())  return;
 138 
 139   ScheduleAndBundle();
 140 
 141 #ifndef PRODUCT
 142   if (trace_opto_output()) {
 143     tty->print("\n---- After ScheduleAndBundle ----\n");
 144     for (uint i = 0; i < _cfg->_num_blocks; i++) {
 145       tty->print("\nBB#%03d:\n", i);
 146       Block *bb = _cfg->_blocks[i];
 147       for (uint j = 0; j < bb->_nodes.size(); j++) {
 148         Node *n = bb->_nodes[j];
 149         OptoReg::Name reg = _regalloc->get_reg_first(n);
 150         tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
 151         n->dump();
 152       }
 153     }
 154   }
 155 #endif
 156 
 157   if (failing())  return;
 158 
 159   finalize_offsets_and_shorten(blk_starts);
 160 
 161   BuildOopMaps();
 162 
 163   if (failing())  return;
 164 
 165   fill_buffer(cb, blk_starts);
 166 }
 167 
 168 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
 169   // Determine if we need to generate a stack overflow check.
 170   // Do it if the method is not a stub function and
 171   // has java calls or has frame size > vm_page_size/8.
 172   return (stub_function() == NULL &&
 173           (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3));
 174 }
 175 
 176 bool Compile::need_register_stack_bang() const {
 177   // Determine if we need to generate a register stack overflow check.
 178   // This is only used on architectures which have split register
 179   // and memory stacks (ie. IA64).
 180   // Bang if the method is not a stub function and has java calls
 181   return (stub_function() == NULL && has_java_calls());
 182 }
 183 
 184 # ifdef ENABLE_ZAP_DEAD_LOCALS
 185 
 186 
 187 // In order to catch compiler oop-map bugs, we have implemented
 188 // a debugging mode called ZapDeadCompilerLocals.
 189 // This mode causes the compiler to insert a call to a runtime routine,
 190 // "zap_dead_locals", right before each place in compiled code
 191 // that could potentially be a gc-point (i.e., a safepoint or oop map point).
 192 // The runtime routine checks that locations mapped as oops are really
 193 // oops, that locations mapped as values do not look like oops,
 194 // and that locations mapped as dead are not used later
 195 // (by zapping them to an invalid address).
 196 
 197 int Compile::_CompiledZap_count = 0;
 198 
 199 void Compile::Insert_zap_nodes() {
 200   bool skip = false;
 201 
 202 
 203   // Dink with static counts because code code without the extra
 204   // runtime calls is MUCH faster for debugging purposes
 205 
 206        if ( CompileZapFirst  ==  0  ) ; // nothing special
 207   else if ( CompileZapFirst  >  CompiledZap_count() )  skip = true;
 208   else if ( CompileZapFirst  == CompiledZap_count() )
 209     warning("starting zap compilation after skipping");
 210 
 211        if ( CompileZapLast  ==  -1  ) ; // nothing special
 212   else if ( CompileZapLast  <   CompiledZap_count() )  skip = true;
 213   else if ( CompileZapLast  ==  CompiledZap_count() )
 214     warning("about to compile last zap");
 215 
 216   ++_CompiledZap_count; // counts skipped zaps, too
 217 
 218   if ( skip )  return;
 219 
 220 
 221   if ( _method == NULL )
 222     return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care
 223 
 224   // Insert call to zap runtime stub before every node with an oop map
 225   for( uint i=0; i<_cfg->_num_blocks; i++ ) {
 226     Block *b = _cfg->_blocks[i];
 227     for ( uint j = 0;  j < b->_nodes.size();  ++j ) {
 228       Node *n = b->_nodes[j];
 229 
 230       // Determining if we should insert a zap-a-lot node in output.
 231       // We do that for all nodes that has oopmap info, except for calls
 232       // to allocation.  Calls to allocation passes in the old top-of-eden pointer
 233       // and expect the C code to reset it.  Hence, there can be no safepoints between
 234       // the inlined-allocation and the call to new_Java, etc.
 235       // We also cannot zap monitor calls, as they must hold the microlock
 236       // during the call to Zap, which also wants to grab the microlock.
 237       bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL);
 238       if ( insert ) { // it is MachSafePoint
 239         if ( !n->is_MachCall() ) {
 240           insert = false;
 241         } else if ( n->is_MachCall() ) {
 242           MachCallNode* call = n->as_MachCall();
 243           if (call->entry_point() == OptoRuntime::new_instance_Java() ||
 244               call->entry_point() == OptoRuntime::new_array_Java() ||
 245               call->entry_point() == OptoRuntime::multianewarray2_Java() ||
 246               call->entry_point() == OptoRuntime::multianewarray3_Java() ||
 247               call->entry_point() == OptoRuntime::multianewarray4_Java() ||
 248               call->entry_point() == OptoRuntime::multianewarray5_Java() ||
 249               call->entry_point() == OptoRuntime::slow_arraycopy_Java() ||
 250               call->entry_point() == OptoRuntime::complete_monitor_locking_Java()
 251               ) {
 252             insert = false;
 253           }
 254         }
 255         if (insert) {
 256           Node *zap = call_zap_node(n->as_MachSafePoint(), i);
 257           b->_nodes.insert( j, zap );
 258           _cfg->_bbs.map( zap->_idx, b );
 259           ++j;
 260         }
 261       }
 262     }
 263   }
 264 }
 265 
 266 
 267 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) {
 268   const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type();
 269   CallStaticJavaNode* ideal_node =
 270     new (this, tf->domain()->cnt()) CallStaticJavaNode( tf,
 271          OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()),
 272                             "call zap dead locals stub", 0, TypePtr::BOTTOM);
 273   // We need to copy the OopMap from the site we're zapping at.
 274   // We have to make a copy, because the zap site might not be
 275   // a call site, and zap_dead is a call site.
 276   OopMap* clone = node_to_check->oop_map()->deep_copy();
 277 
 278   // Add the cloned OopMap to the zap node
 279   ideal_node->set_oop_map(clone);
 280   return _matcher->match_sfpt(ideal_node);
 281 }
 282 
 283 //------------------------------is_node_getting_a_safepoint--------------------
 284 bool Compile::is_node_getting_a_safepoint( Node* n) {
 285   // This code duplicates the logic prior to the call of add_safepoint
 286   // below in this file.
 287   if( n->is_MachSafePoint() ) return true;
 288   return false;
 289 }
 290 
 291 # endif // ENABLE_ZAP_DEAD_LOCALS
 292 
 293 //------------------------------compute_loop_first_inst_sizes------------------
 294 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
 295 // of a loop. When aligning a loop we need to provide enough instructions
 296 // in cpu's fetch buffer to feed decoders. The loop alignment could be
 297 // avoided if we have enough instructions in fetch buffer at the head of a loop.
 298 // By default, the size is set to 999999 by Block's constructor so that
 299 // a loop will be aligned if the size is not reset here.
 300 //
 301 // Note: Mach instructions could contain several HW instructions
 302 // so the size is estimated only.
 303 //
 304 void Compile::compute_loop_first_inst_sizes() {
 305   // The next condition is used to gate the loop alignment optimization.
 306   // Don't aligned a loop if there are enough instructions at the head of a loop
 307   // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
 308   // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
 309   // equal to 11 bytes which is the largest address NOP instruction.
 310   if( MaxLoopPad < OptoLoopAlignment-1 ) {
 311     uint last_block = _cfg->_num_blocks-1;
 312     for( uint i=1; i <= last_block; i++ ) {
 313       Block *b = _cfg->_blocks[i];
 314       // Check the first loop's block which requires an alignment.
 315       if( b->loop_alignment() > (uint)relocInfo::addr_unit() ) {
 316         uint sum_size = 0;
 317         uint inst_cnt = NumberOfLoopInstrToAlign;
 318         inst_cnt = b->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
 319 
 320         // Check subsequent fallthrough blocks if the loop's first
 321         // block(s) does not have enough instructions.
 322         Block *nb = b;
 323         while( inst_cnt > 0 &&
 324                i < last_block &&
 325                !_cfg->_blocks[i+1]->has_loop_alignment() &&
 326                !nb->has_successor(b) ) {
 327           i++;
 328           nb = _cfg->_blocks[i];
 329           inst_cnt  = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
 330         } // while( inst_cnt > 0 && i < last_block  )
 331 
 332         b->set_first_inst_size(sum_size);
 333       } // f( b->head()->is_Loop() )
 334     } // for( i <= last_block )
 335   } // if( MaxLoopPad < OptoLoopAlignment-1 )
 336 }
 337 
 338 //----------------------shorten_branches---------------------------------------
 339 // The architecture description provides short branch variants for some long
 340 // branch instructions. Replace eligible long branches with short branches.
 341 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
 342 
 343   // ------------------
 344   // Compute size of each block, method size, and relocation information size
 345   uint nblocks  = _cfg->_num_blocks;
 346 
 347   uint*      jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
 348   uint*      jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
 349   int*       jmp_nidx   = NEW_RESOURCE_ARRAY(int ,nblocks);
 350   DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
 351   DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
 352 
 353   bool has_short_branch_candidate = false;
 354 
 355   // Initialize the sizes to 0
 356   code_size  = 0;          // Size in bytes of generated code
 357   stub_size  = 0;          // Size in bytes of all stub entries
 358   // Size in bytes of all relocation entries, including those in local stubs.
 359   // Start with 2-bytes of reloc info for the unvalidated entry point
 360   reloc_size = 1;          // Number of relocation entries
 361 
 362   // Make three passes.  The first computes pessimistic blk_starts,
 363   // relative jmp_offset and reloc_size information.  The second performs
 364   // short branch substitution using the pessimistic sizing.  The
 365   // third inserts nops where needed.
 366 
 367   // Step one, perform a pessimistic sizing pass.
 368   uint last_call_adr = max_uint;
 369   uint last_avoid_back_to_back_adr = max_uint;
 370   uint nop_size = (new (this) MachNopNode())->size(_regalloc);
 371   for (uint i = 0; i < nblocks; i++) { // For all blocks
 372     Block *b = _cfg->_blocks[i];
 373 
 374     // During short branch replacement, we store the relative (to blk_starts)
 375     // offset of jump in jmp_offset, rather than the absolute offset of jump.
 376     // This is so that we do not need to recompute sizes of all nodes when
 377     // we compute correct blk_starts in our next sizing pass.
 378     jmp_offset[i] = 0;
 379     jmp_size[i]   = 0;
 380     jmp_nidx[i]   = -1;
 381     DEBUG_ONLY( jmp_target[i] = 0; )
 382     DEBUG_ONLY( jmp_rule[i]   = 0; )
 383 
 384     // Sum all instruction sizes to compute block size
 385     uint last_inst = b->_nodes.size();
 386     uint blk_size = 0;
 387     for (uint j = 0; j < last_inst; j++) {
 388       Node* nj = b->_nodes[j];
 389       uint inst_size = nj->size(_regalloc);
 390       // Handle machine instruction nodes
 391       if (nj->is_Mach()) {
 392         MachNode *mach = nj->as_Mach();
 393         blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
 394         reloc_size += mach->reloc();
 395         if( mach->is_MachCall() ) {
 396           MachCallNode *mcall = mach->as_MachCall();
 397           // This destination address is NOT PC-relative
 398 
 399           mcall->method_set((intptr_t)mcall->entry_point());
 400 
 401           if( mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method ) {
 402             stub_size  += size_java_to_interp();
 403             reloc_size += reloc_java_to_interp();
 404           }
 405         } else if (mach->is_MachSafePoint()) {
 406           // If call/safepoint are adjacent, account for possible
 407           // nop to disambiguate the two safepoints.
 408           // ScheduleAndBundle() can rearrange nodes in a block,
 409           // check for all offsets inside this block.
 410           if (last_call_adr >= blk_starts[i]) {
 411             blk_size += nop_size;
 412           }
 413         }
 414         if (mach->avoid_back_to_back()) {
 415           // Nop is inserted between "avoid back to back" instructions.
 416           // ScheduleAndBundle() can rearrange nodes in a block,
 417           // check for all offsets inside this block.
 418           if (last_avoid_back_to_back_adr >= blk_starts[i]) {
 419             blk_size += nop_size;
 420           }
 421         }
 422         if (mach->may_be_short_branch()) {
 423           if (!nj->is_Branch()) {
 424 #ifndef PRODUCT
 425             nj->dump(3);
 426 #endif
 427             Unimplemented();
 428           }
 429           assert(jmp_nidx[i] == -1, "block should have only one branch");
 430           jmp_offset[i] = blk_size;
 431           jmp_size[i]   = inst_size;
 432           jmp_nidx[i]   = j;
 433           has_short_branch_candidate = true;
 434         }
 435       }
 436       blk_size += inst_size;
 437       // Remember end of call offset
 438       if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
 439         last_call_adr = blk_starts[i]+blk_size;
 440       }
 441       // Remember end of avoid_back_to_back offset
 442       if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) {
 443         last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
 444       }
 445     }
 446 
 447     // When the next block starts a loop, we may insert pad NOP
 448     // instructions.  Since we cannot know our future alignment,
 449     // assume the worst.
 450     if (i< nblocks-1) {
 451       Block *nb = _cfg->_blocks[i+1];
 452       int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
 453       if (max_loop_pad > 0) {
 454         assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
 455         blk_size += max_loop_pad;
 456       }
 457     }
 458 
 459     // Save block size; update total method size
 460     blk_starts[i+1] = blk_starts[i]+blk_size;
 461   }
 462 
 463   // Step two, replace eligible long jumps.
 464   bool progress = true;
 465   uint last_may_be_short_branch_adr = max_uint;
 466   while (has_short_branch_candidate && progress) {
 467     progress = false;
 468     has_short_branch_candidate = false;
 469     int adjust_block_start = 0;
 470     for (uint i = 0; i < nblocks; i++) {
 471       Block *b = _cfg->_blocks[i];
 472       int idx = jmp_nidx[i];
 473       MachNode* mach = (idx == -1) ? NULL: b->_nodes[idx]->as_Mach();
 474       if (mach != NULL && mach->may_be_short_branch()) {
 475 #ifdef ASSERT
 476         assert(jmp_size[i] > 0 && mach->is_Branch(), "sanity");
 477         int j;
 478         // Find the branch; ignore trailing NOPs.
 479         for (j = b->_nodes.size()-1; j>=0; j--) {
 480           Node* n = b->_nodes[j];
 481           if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
 482             break;
 483         }
 484         assert(j >= 0 && j == idx && b->_nodes[j] == (Node*)mach, "sanity");
 485 #endif
 486         int br_size = jmp_size[i];
 487         int br_offs = blk_starts[i] + jmp_offset[i];
 488 
 489         // This requires the TRUE branch target be in succs[0]
 490         uint bnum = b->non_connector_successor(0)->_pre_order;
 491         int offset = blk_starts[bnum] - br_offs;
 492         if (bnum > i) { // adjust following block's offset
 493           offset -= adjust_block_start;
 494         }
 495         // In the following code a nop could be inserted before
 496         // the branch which will increase the backward distance.
 497         bool needs_padding = ((uint)br_offs == last_may_be_short_branch_adr);
 498         if (needs_padding && offset <= 0)
 499           offset -= nop_size;
 500 
 501         if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
 502           // We've got a winner.  Replace this branch.
 503           MachNode* replacement = mach->short_branch_version(this);
 504 
 505           // Update the jmp_size.
 506           int new_size = replacement->size(_regalloc);
 507           int diff     = br_size - new_size;
 508           assert(diff >= (int)nop_size, "short_branch size should be smaller");
 509           // Conservatively take into accound padding between
 510           // avoid_back_to_back branches. Previous branch could be
 511           // converted into avoid_back_to_back branch during next
 512           // rounds.
 513           if (needs_padding && replacement->avoid_back_to_back()) {
 514             jmp_offset[i] += nop_size;
 515             diff -= nop_size;
 516           }
 517           adjust_block_start += diff;
 518           b->_nodes.map(idx, replacement);
 519           mach->subsume_by(replacement);
 520           mach = replacement;
 521           progress = true;
 522 
 523           jmp_size[i] = new_size;
 524           DEBUG_ONLY( jmp_target[i] = bnum; );
 525           DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
 526         } else {
 527           // The jump distance is not short, try again during next iteration.
 528           has_short_branch_candidate = true;
 529         }
 530       } // (mach->may_be_short_branch())
 531       if (mach != NULL && (mach->may_be_short_branch() ||
 532                            mach->avoid_back_to_back())) {
 533         last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
 534       }
 535       blk_starts[i+1] -= adjust_block_start;
 536     }
 537   }
 538 
 539 #ifdef ASSERT
 540   for (uint i = 0; i < nblocks; i++) { // For all blocks
 541     if (jmp_target[i] != 0) {
 542       int br_size = jmp_size[i];
 543       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
 544       if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
 545         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]);
 546       }
 547       assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
 548     }
 549   }
 550 #endif
 551 
 552   // Step 3, compute the offsets of all blocks, will be done in finalize_offsets_and_shorten()
 553   // after ScheduleAndBundle().
 554 
 555   // ------------------
 556   // Compute size for code buffer
 557   code_size = blk_starts[nblocks];
 558 
 559   // Relocation records
 560   reloc_size += 1;              // Relo entry for exception handler
 561 
 562   // Adjust reloc_size to number of record of relocation info
 563   // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
 564   // a relocation index.
 565   // The CodeBuffer will expand the locs array if this estimate is too low.
 566   reloc_size *= 10 / sizeof(relocInfo);
 567 }
 568 
 569 //----------------------finalize_offsets_and_shorten-------------------------
 570 void Compile::finalize_offsets_and_shorten(uint* blk_starts) {
 571   // blk_starts[] contains offsets calculated during short branches processing,
 572   // offsets should not be increased during following steps.
 573 
 574   // Compute the size of first NumberOfLoopInstrToAlign instructions at head
 575   // of a loop. It is used to determine the padding for loop alignment.
 576   compute_loop_first_inst_sizes();
 577 
 578   uint nblocks  = _cfg->_num_blocks;
 579 #ifdef ASSERT
 580   uint*      jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
 581   uint*      jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
 582   uint*      jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
 583   uint*      jmp_rule   = NEW_RESOURCE_ARRAY(uint,nblocks);
 584 #endif
 585 
 586   // Inserts nops where needed and do final short branches replacement.
 587   uint nop_size = (new (this) MachNopNode())->size(_regalloc);
 588   uint last_call_adr = max_uint;
 589   uint last_avoid_back_to_back_adr = max_uint;
 590 
 591   assert(blk_starts[0] == 0, "sanity");
 592   uint current_offset = 0;
 593   uint block_alignment_padding = 0;
 594 
 595   for (uint i=0; i < nblocks; i++) { // For all blocks
 596     Block *b = _cfg->_blocks[i];
 597 
 598 #ifdef ASSERT
 599     jmp_target[i] = 0;
 600     jmp_offset[i] = 0;
 601     jmp_size[i]   = 0;
 602     jmp_rule[i]   = 0;
 603 #endif
 604 
 605     // Maximum alignment was added before loop block during
 606     // Step One, as result padding for nodes was not added.
 607     // Take this into account for block's size change check
 608     // and allow increase block's size by the difference
 609     // of maximum and actual alignment paddings.
 610     DEBUG_ONLY( uint orig_blk_size = blk_starts[i+1] - blk_starts[i] + block_alignment_padding; )
 611     uint blk_offset = current_offset;
 612 
 613     uint last_inst = b->_nodes.size();
 614     for (uint j = 0; j<last_inst; j++) {
 615       Node* nj = b->_nodes[j];
 616 
 617       if (valid_bundle_info(nj) &&
 618           node_bundling(nj)->used_in_unconditional_delay()) {
 619         continue; // Skip instruction in delay slot
 620       }
 621 
 622       uint inst_size = nj->size(_regalloc);
 623       if (nj->is_Mach()) {
 624         MachNode *mach = nj->as_Mach();
 625         int padding = mach->compute_padding(current_offset);
 626 
 627         // If call/safepoint are adjacent insert a nop (5010568)
 628         if (padding == 0 && nj->is_MachSafePoint() && !nj->is_MachCall() &&
 629             current_offset == last_call_adr) {
 630           padding = nop_size;
 631         }
 632 
 633         // Inserted a nop between "avoid back to back" instructions.
 634         if (padding == 0 && mach->avoid_back_to_back() &&
 635             current_offset == last_avoid_back_to_back_adr) {
 636           padding = nop_size;
 637         }
 638 
 639         if (padding > 0) {
 640           assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
 641           int nops_cnt = padding / nop_size;
 642           MachNode *nop = new (this) MachNopNode(nops_cnt);
 643           b->_nodes.insert(j++, nop);
 644           _cfg->_bbs.map(nop->_idx, b);
 645           last_inst++;
 646           current_offset += padding;
 647         }
 648 
 649         // Try to replace long branch if delay slot is not used,
 650         // it is mostly for back branches since forward branch's
 651         // distance is not updated yet.
 652         bool delay_slot_is_used = valid_bundle_info(nj) &&
 653                                   node_bundling(nj)->use_unconditional_delay();
 654         if (!delay_slot_is_used && mach->may_be_short_branch()) {
 655           int br_size = inst_size;
 656 
 657           // This requires the TRUE branch target be in succs[0]
 658           uint bnum = b->non_connector_successor(0)->_pre_order;
 659           int offset = blk_starts[bnum] - current_offset;
 660           if (bnum >= i) {
 661             // Current and following block's offset are not
 662             // finilized yet, adjust distance.
 663             offset -= (blk_starts[i] - blk_offset);
 664           }
 665           // In the following code a nop could be inserted before
 666           // the branch which will increase the backward distance.
 667           bool needs_padding = (current_offset == last_avoid_back_to_back_adr);
 668           if (needs_padding && offset <= 0)
 669             offset -= nop_size;
 670 
 671           if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
 672             // We've got a winner.  Replace this branch.
 673             MachNode* replacement = mach->short_branch_version(this);
 674 
 675             // Update the jmp_size.
 676             int new_size = replacement->size(_regalloc);
 677             assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
 678             // Conservatively take into accound padding between
 679             // avoid_back_to_back branches. Previous branch could be
 680             // converted into avoid_back_to_back branch during next
 681             // rounds.
 682             if (needs_padding && replacement->avoid_back_to_back()) {
 683               MachNode *nop = new (this) MachNopNode();
 684               b->_nodes.insert(j++, nop);
 685               _cfg->_bbs.map(nop->_idx, b);
 686               last_inst++;
 687               current_offset += nop_size;
 688             }
 689             inst_size = new_size;
 690             b->_nodes.map(j, replacement);
 691             mach->subsume_by(replacement);
 692             nj = replacement;
 693 #ifdef ASSERT
 694             jmp_target[i] = bnum;
 695             jmp_offset[i] = current_offset - blk_offset;
 696             jmp_size[i]   = new_size;
 697             jmp_rule[i]   = mach->rule();
 698 #endif
 699           }
 700         }
 701       }
 702       current_offset += inst_size;
 703 
 704       // Remember end of call offset
 705       if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
 706         last_call_adr = current_offset;
 707       }
 708       // Remember end of avoid_back_to_back offset
 709       if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) {
 710         last_avoid_back_to_back_adr = current_offset;
 711       }
 712     }
 713     assert(blk_offset <= blk_starts[i], "shouldn't increase distance");
 714     blk_starts[i] = blk_offset;
 715 
 716     // When the next block is the top of a loop, we may insert pad NOP
 717     // instructions.
 718     if (i < nblocks-1) {
 719       Block *nb = _cfg->_blocks[i+1];
 720       int padding = nb->alignment_padding(current_offset);
 721       if (padding > 0) {
 722         assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
 723         int nops_cnt = padding / nop_size;
 724         MachNode *nop = new (this) MachNopNode(nops_cnt);
 725         b->_nodes.insert(b->_nodes.size(), nop);
 726         _cfg->_bbs.map(nop->_idx, b);
 727         current_offset += padding;
 728       }
 729       int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
 730       assert(max_loop_pad >= padding, "sanity");
 731       block_alignment_padding = max_loop_pad - padding;
 732     }
 733     assert(orig_blk_size >= (current_offset - blk_offset), "shouldn't increase block size");
 734   }
 735   blk_starts[nblocks] = current_offset;
 736 
 737 #ifdef ASSERT
 738   for (uint i = 0; i < nblocks; i++) { // For all blocks
 739     if (jmp_target[i] != 0) {
 740       int br_size = jmp_size[i];
 741       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
 742       if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
 743         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]);
 744       }
 745       assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
 746     }
 747   }
 748 #endif
 749 }
 750 
 751 //------------------------------FillLocArray-----------------------------------
 752 // Create a bit of debug info and append it to the array.  The mapping is from
 753 // Java local or expression stack to constant, register or stack-slot.  For
 754 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
 755 // entry has been taken care of and caller should skip it).
 756 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
 757   // This should never have accepted Bad before
 758   assert(OptoReg::is_valid(regnum), "location must be valid");
 759   return (OptoReg::is_reg(regnum))
 760     ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
 761     : new LocationValue(Location::new_stk_loc(l_type,  ra->reg2offset(regnum)));
 762 }
 763 
 764 
 765 ObjectValue*
 766 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
 767   for (int i = 0; i < objs->length(); i++) {
 768     assert(objs->at(i)->is_object(), "corrupt object cache");
 769     ObjectValue* sv = (ObjectValue*) objs->at(i);
 770     if (sv->id() == id) {
 771       return sv;
 772     }
 773   }
 774   // Otherwise..
 775   return NULL;
 776 }
 777 
 778 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
 779                                      ObjectValue* sv ) {
 780   assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
 781   objs->append(sv);
 782 }
 783 
 784 
 785 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
 786                             GrowableArray<ScopeValue*> *array,
 787                             GrowableArray<ScopeValue*> *objs ) {
 788   assert( local, "use _top instead of null" );
 789   if (array->length() != idx) {
 790     assert(array->length() == idx + 1, "Unexpected array count");
 791     // Old functionality:
 792     //   return
 793     // New functionality:
 794     //   Assert if the local is not top. In product mode let the new node
 795     //   override the old entry.
 796     assert(local == top(), "LocArray collision");
 797     if (local == top()) {
 798       return;
 799     }
 800     array->pop();
 801   }
 802   const Type *t = local->bottom_type();
 803 
 804   // Is it a safepoint scalar object node?
 805   if (local->is_SafePointScalarObject()) {
 806     SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
 807 
 808     ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
 809     if (sv == NULL) {
 810       ciKlass* cik = t->is_oopptr()->klass();
 811       assert(cik->is_instance_klass() ||
 812              cik->is_array_klass(), "Not supported allocation.");
 813       sv = new ObjectValue(spobj->_idx,
 814                            new ConstantOopWriteValue(cik->constant_encoding()));
 815       Compile::set_sv_for_object_node(objs, sv);
 816 
 817       uint first_ind = spobj->first_index();
 818       for (uint i = 0; i < spobj->n_fields(); i++) {
 819         Node* fld_node = sfpt->in(first_ind+i);
 820         (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
 821       }
 822     }
 823     array->append(sv);
 824     return;
 825   }
 826 
 827   // Grab the register number for the local
 828   OptoReg::Name regnum = _regalloc->get_reg_first(local);
 829   if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
 830     // Record the double as two float registers.
 831     // The register mask for such a value always specifies two adjacent
 832     // float registers, with the lower register number even.
 833     // Normally, the allocation of high and low words to these registers
 834     // is irrelevant, because nearly all operations on register pairs
 835     // (e.g., StoreD) treat them as a single unit.
 836     // Here, we assume in addition that the words in these two registers
 837     // stored "naturally" (by operations like StoreD and double stores
 838     // within the interpreter) such that the lower-numbered register
 839     // is written to the lower memory address.  This may seem like
 840     // a machine dependency, but it is not--it is a requirement on
 841     // the author of the <arch>.ad file to ensure that, for every
 842     // even/odd double-register pair to which a double may be allocated,
 843     // the word in the even single-register is stored to the first
 844     // memory word.  (Note that register numbers are completely
 845     // arbitrary, and are not tied to any machine-level encodings.)
 846 #ifdef _LP64
 847     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
 848       array->append(new ConstantIntValue(0));
 849       array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
 850     } else if ( t->base() == Type::Long ) {
 851       array->append(new ConstantIntValue(0));
 852       array->append(new_loc_value( _regalloc, regnum, Location::lng ));
 853     } else if ( t->base() == Type::RawPtr ) {
 854       // jsr/ret return address which must be restored into a the full
 855       // width 64-bit stack slot.
 856       array->append(new_loc_value( _regalloc, regnum, Location::lng ));
 857     }
 858 #else //_LP64
 859 #ifdef SPARC
 860     if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
 861       // For SPARC we have to swap high and low words for
 862       // long values stored in a single-register (g0-g7).
 863       array->append(new_loc_value( _regalloc,              regnum   , Location::normal ));
 864       array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
 865     } else
 866 #endif //SPARC
 867     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
 868       // Repack the double/long as two jints.
 869       // The convention the interpreter uses is that the second local
 870       // holds the first raw word of the native double representation.
 871       // This is actually reasonable, since locals and stack arrays
 872       // grow downwards in all implementations.
 873       // (If, on some machine, the interpreter's Java locals or stack
 874       // were to grow upwards, the embedded doubles would be word-swapped.)
 875       array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
 876       array->append(new_loc_value( _regalloc,              regnum   , Location::normal ));
 877     }
 878 #endif //_LP64
 879     else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
 880                OptoReg::is_reg(regnum) ) {
 881       array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
 882                                    ? Location::float_in_dbl : Location::normal ));
 883     } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
 884       array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
 885                                    ? Location::int_in_long : Location::normal ));
 886     } else if( t->base() == Type::NarrowOop ) {
 887       array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
 888     } else {
 889       array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
 890     }
 891     return;
 892   }
 893 
 894   // No register.  It must be constant data.
 895   switch (t->base()) {
 896   case Type::Half:              // Second half of a double
 897     ShouldNotReachHere();       // Caller should skip 2nd halves
 898     break;
 899   case Type::AnyPtr:
 900     array->append(new ConstantOopWriteValue(NULL));
 901     break;
 902   case Type::AryPtr:
 903   case Type::InstPtr:
 904   case Type::KlassPtr:          // fall through
 905     array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
 906     break;
 907   case Type::NarrowOop:
 908     if (t == TypeNarrowOop::NULL_PTR) {
 909       array->append(new ConstantOopWriteValue(NULL));
 910     } else {
 911       array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
 912     }
 913     break;
 914   case Type::Int:
 915     array->append(new ConstantIntValue(t->is_int()->get_con()));
 916     break;
 917   case Type::RawPtr:
 918     // A return address (T_ADDRESS).
 919     assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
 920 #ifdef _LP64
 921     // Must be restored to the full-width 64-bit stack slot.
 922     array->append(new ConstantLongValue(t->is_ptr()->get_con()));
 923 #else
 924     array->append(new ConstantIntValue(t->is_ptr()->get_con()));
 925 #endif
 926     break;
 927   case Type::FloatCon: {
 928     float f = t->is_float_constant()->getf();
 929     array->append(new ConstantIntValue(jint_cast(f)));
 930     break;
 931   }
 932   case Type::DoubleCon: {
 933     jdouble d = t->is_double_constant()->getd();
 934 #ifdef _LP64
 935     array->append(new ConstantIntValue(0));
 936     array->append(new ConstantDoubleValue(d));
 937 #else
 938     // Repack the double as two jints.
 939     // The convention the interpreter uses is that the second local
 940     // holds the first raw word of the native double representation.
 941     // This is actually reasonable, since locals and stack arrays
 942     // grow downwards in all implementations.
 943     // (If, on some machine, the interpreter's Java locals or stack
 944     // were to grow upwards, the embedded doubles would be word-swapped.)
 945     jint   *dp = (jint*)&d;
 946     array->append(new ConstantIntValue(dp[1]));
 947     array->append(new ConstantIntValue(dp[0]));
 948 #endif
 949     break;
 950   }
 951   case Type::Long: {
 952     jlong d = t->is_long()->get_con();
 953 #ifdef _LP64
 954     array->append(new ConstantIntValue(0));
 955     array->append(new ConstantLongValue(d));
 956 #else
 957     // Repack the long as two jints.
 958     // The convention the interpreter uses is that the second local
 959     // holds the first raw word of the native double representation.
 960     // This is actually reasonable, since locals and stack arrays
 961     // grow downwards in all implementations.
 962     // (If, on some machine, the interpreter's Java locals or stack
 963     // were to grow upwards, the embedded doubles would be word-swapped.)
 964     jint *dp = (jint*)&d;
 965     array->append(new ConstantIntValue(dp[1]));
 966     array->append(new ConstantIntValue(dp[0]));
 967 #endif
 968     break;
 969   }
 970   case Type::Top:               // Add an illegal value here
 971     array->append(new LocationValue(Location()));
 972     break;
 973   default:
 974     ShouldNotReachHere();
 975     break;
 976   }
 977 }
 978 
 979 // Determine if this node starts a bundle
 980 bool Compile::starts_bundle(const Node *n) const {
 981   return (_node_bundling_limit > n->_idx &&
 982           _node_bundling_base[n->_idx].starts_bundle());
 983 }
 984 
 985 //--------------------------Process_OopMap_Node--------------------------------
 986 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
 987 
 988   // Handle special safepoint nodes for synchronization
 989   MachSafePointNode *sfn   = mach->as_MachSafePoint();
 990   MachCallNode      *mcall;
 991 
 992 #ifdef ENABLE_ZAP_DEAD_LOCALS
 993   assert( is_node_getting_a_safepoint(mach),  "logic does not match; false negative");
 994 #endif
 995 
 996   int safepoint_pc_offset = current_offset;
 997   bool is_method_handle_invoke = false;
 998   bool return_oop = false;
 999 
1000   // Add the safepoint in the DebugInfoRecorder
1001   if( !mach->is_MachCall() ) {
1002     mcall = NULL;
1003     debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1004   } else {
1005     mcall = mach->as_MachCall();
1006 
1007     // Is the call a MethodHandle call?
1008     if (mcall->is_MachCallJava()) {
1009       if (mcall->as_MachCallJava()->_method_handle_invoke) {
1010         assert(has_method_handle_invokes(), "must have been set during call generation");
1011         is_method_handle_invoke = true;
1012       }
1013     }
1014 
1015     // Check if a call returns an object.
1016     if (mcall->return_value_is_used() &&
1017         mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) {
1018       return_oop = true;
1019     }
1020     safepoint_pc_offset += mcall->ret_addr_offset();
1021     debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1022   }
1023 
1024   // Loop over the JVMState list to add scope information
1025   // Do not skip safepoints with a NULL method, they need monitor info
1026   JVMState* youngest_jvms = sfn->jvms();
1027   int max_depth = youngest_jvms->depth();
1028 
1029   // Allocate the object pool for scalar-replaced objects -- the map from
1030   // small-integer keys (which can be recorded in the local and ostack
1031   // arrays) to descriptions of the object state.
1032   GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1033 
1034   // Visit scopes from oldest to youngest.
1035   for (int depth = 1; depth <= max_depth; depth++) {
1036     JVMState* jvms = youngest_jvms->of_depth(depth);
1037     int idx;
1038     ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1039     // Safepoints that do not have method() set only provide oop-map and monitor info
1040     // to support GC; these do not support deoptimization.
1041     int num_locs = (method == NULL) ? 0 : jvms->loc_size();
1042     int num_exps = (method == NULL) ? 0 : jvms->stk_size();
1043     int num_mon  = jvms->nof_monitors();
1044     assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
1045            "JVMS local count must match that of the method");
1046 
1047     // Add Local and Expression Stack Information
1048 
1049     // Insert locals into the locarray
1050     GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1051     for( idx = 0; idx < num_locs; idx++ ) {
1052       FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1053     }
1054 
1055     // Insert expression stack entries into the exparray
1056     GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1057     for( idx = 0; idx < num_exps; idx++ ) {
1058       FillLocArray( idx,  sfn, sfn->stack(jvms, idx), exparray, objs );
1059     }
1060 
1061     // Add in mappings of the monitors
1062     assert( !method ||
1063             !method->is_synchronized() ||
1064             method->is_native() ||
1065             num_mon > 0 ||
1066             !GenerateSynchronizationCode,
1067             "monitors must always exist for synchronized methods");
1068 
1069     // Build the growable array of ScopeValues for exp stack
1070     GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1071 
1072     // Loop over monitors and insert into array
1073     for(idx = 0; idx < num_mon; idx++) {
1074       // Grab the node that defines this monitor
1075       Node* box_node = sfn->monitor_box(jvms, idx);
1076       Node* obj_node = sfn->monitor_obj(jvms, idx);
1077 
1078       // Create ScopeValue for object
1079       ScopeValue *scval = NULL;
1080 
1081       if( obj_node->is_SafePointScalarObject() ) {
1082         SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1083         scval = Compile::sv_for_node_id(objs, spobj->_idx);
1084         if (scval == NULL) {
1085           const Type *t = obj_node->bottom_type();
1086           ciKlass* cik = t->is_oopptr()->klass();
1087           assert(cik->is_instance_klass() ||
1088                  cik->is_array_klass(), "Not supported allocation.");
1089           ObjectValue* sv = new ObjectValue(spobj->_idx,
1090                                 new ConstantOopWriteValue(cik->constant_encoding()));
1091           Compile::set_sv_for_object_node(objs, sv);
1092 
1093           uint first_ind = spobj->first_index();
1094           for (uint i = 0; i < spobj->n_fields(); i++) {
1095             Node* fld_node = sfn->in(first_ind+i);
1096             (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1097           }
1098           scval = sv;
1099         }
1100       } else if( !obj_node->is_Con() ) {
1101         OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
1102         if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1103           scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
1104         } else {
1105           scval = new_loc_value( _regalloc, obj_reg, Location::oop );
1106         }
1107       } else {
1108         const TypePtr *tp = obj_node->bottom_type()->make_ptr();
1109         scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1110       }
1111 
1112       OptoReg::Name box_reg = BoxLockNode::stack_slot(box_node);
1113       Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
1114       while( !box_node->is_BoxLock() )  box_node = box_node->in(1);
1115       monarray->append(new MonitorValue(scval, basic_lock, box_node->as_BoxLock()->is_eliminated()));
1116     }
1117 
1118     // We dump the object pool first, since deoptimization reads it in first.
1119     debug_info()->dump_object_pool(objs);
1120 
1121     // Build first class objects to pass to scope
1122     DebugToken *locvals = debug_info()->create_scope_values(locarray);
1123     DebugToken *expvals = debug_info()->create_scope_values(exparray);
1124     DebugToken *monvals = debug_info()->create_monitor_values(monarray);
1125 
1126     // Make method available for all Safepoints
1127     ciMethod* scope_method = method ? method : _method;
1128     // Describe the scope here
1129     assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1130     assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1131     // Now we can describe the scope.
1132     debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);
1133   } // End jvms loop
1134 
1135   // Mark the end of the scope set.
1136   debug_info()->end_safepoint(safepoint_pc_offset);
1137 }
1138 
1139 
1140 
1141 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1142 class NonSafepointEmitter {
1143   Compile*  C;
1144   JVMState* _pending_jvms;
1145   int       _pending_offset;
1146 
1147   void emit_non_safepoint();
1148 
1149  public:
1150   NonSafepointEmitter(Compile* compile) {
1151     this->C = compile;
1152     _pending_jvms = NULL;
1153     _pending_offset = 0;
1154   }
1155 
1156   void observe_instruction(Node* n, int pc_offset) {
1157     if (!C->debug_info()->recording_non_safepoints())  return;
1158 
1159     Node_Notes* nn = C->node_notes_at(n->_idx);
1160     if (nn == NULL || nn->jvms() == NULL)  return;
1161     if (_pending_jvms != NULL &&
1162         _pending_jvms->same_calls_as(nn->jvms())) {
1163       // Repeated JVMS?  Stretch it up here.
1164       _pending_offset = pc_offset;
1165     } else {
1166       if (_pending_jvms != NULL &&
1167           _pending_offset < pc_offset) {
1168         emit_non_safepoint();
1169       }
1170       _pending_jvms = NULL;
1171       if (pc_offset > C->debug_info()->last_pc_offset()) {
1172         // This is the only way _pending_jvms can become non-NULL:
1173         _pending_jvms = nn->jvms();
1174         _pending_offset = pc_offset;
1175       }
1176     }
1177   }
1178 
1179   // Stay out of the way of real safepoints:
1180   void observe_safepoint(JVMState* jvms, int pc_offset) {
1181     if (_pending_jvms != NULL &&
1182         !_pending_jvms->same_calls_as(jvms) &&
1183         _pending_offset < pc_offset) {
1184       emit_non_safepoint();
1185     }
1186     _pending_jvms = NULL;
1187   }
1188 
1189   void flush_at_end() {
1190     if (_pending_jvms != NULL) {
1191       emit_non_safepoint();
1192     }
1193     _pending_jvms = NULL;
1194   }
1195 };
1196 
1197 void NonSafepointEmitter::emit_non_safepoint() {
1198   JVMState* youngest_jvms = _pending_jvms;
1199   int       pc_offset     = _pending_offset;
1200 
1201   // Clear it now:
1202   _pending_jvms = NULL;
1203 
1204   DebugInformationRecorder* debug_info = C->debug_info();
1205   assert(debug_info->recording_non_safepoints(), "sanity");
1206 
1207   debug_info->add_non_safepoint(pc_offset);
1208   int max_depth = youngest_jvms->depth();
1209 
1210   // Visit scopes from oldest to youngest.
1211   for (int depth = 1; depth <= max_depth; depth++) {
1212     JVMState* jvms = youngest_jvms->of_depth(depth);
1213     ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1214     assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1215     debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());
1216   }
1217 
1218   // Mark the end of the scope set.
1219   debug_info->end_non_safepoint(pc_offset);
1220 }
1221 
1222 
1223 
1224 // helper for fill_buffer bailout logic
1225 static void turn_off_compiler(Compile* C) {
1226   if (CodeCache::largest_free_block() >= CodeCacheMinimumFreeSpace*10) {
1227     // Do not turn off compilation if a single giant method has
1228     // blown the code cache size.
1229     C->record_failure("excessive request to CodeCache");
1230   } else {
1231     // Let CompilerBroker disable further compilations.
1232     C->record_failure("CodeCache is full");
1233   }
1234 }
1235 
1236 
1237 //------------------------------init_buffer------------------------------------
1238 CodeBuffer* Compile::init_buffer(uint* blk_starts) {
1239 
1240   // Set the initially allocated size
1241   int  code_req   = initial_code_capacity;
1242   int  locs_req   = initial_locs_capacity;
1243   int  stub_req   = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
1244   int  const_req  = initial_const_capacity;
1245 
1246   int  pad_req    = NativeCall::instruction_size;
1247   // The extra spacing after the code is necessary on some platforms.
1248   // Sometimes we need to patch in a jump after the last instruction,
1249   // if the nmethod has been deoptimized.  (See 4932387, 4894843.)
1250 
1251   // Compute the byte offset where we can store the deopt pc.
1252   if (fixed_slots() != 0) {
1253     _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1254   }
1255 
1256   // Compute prolog code size
1257   _method_size = 0;
1258   _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
1259 #ifdef IA64
1260   if (save_argument_registers()) {
1261     // 4815101: this is a stub with implicit and unknown precision fp args.
1262     // The usual spill mechanism can only generate stfd's in this case, which
1263     // doesn't work if the fp reg to spill contains a single-precision denorm.
1264     // Instead, we hack around the normal spill mechanism using stfspill's and
1265     // ldffill's in the MachProlog and MachEpilog emit methods.  We allocate
1266     // space here for the fp arg regs (f8-f15) we're going to thusly spill.
1267     //
1268     // If we ever implement 16-byte 'registers' == stack slots, we can
1269     // get rid of this hack and have SpillCopy generate stfspill/ldffill
1270     // instead of stfd/stfs/ldfd/ldfs.
1271     _frame_slots += 8*(16/BytesPerInt);
1272   }
1273 #endif
1274   assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1275 
1276   if (has_mach_constant_base_node()) {
1277     // Fill the constant table.
1278     // Note:  This must happen before shorten_branches.
1279     for (uint i = 0; i < _cfg->_num_blocks; i++) {
1280       Block* b = _cfg->_blocks[i];
1281 
1282       for (uint j = 0; j < b->_nodes.size(); j++) {
1283         Node* n = b->_nodes[j];
1284 
1285         // If the node is a MachConstantNode evaluate the constant
1286         // value section.
1287         if (n->is_MachConstant()) {
1288           MachConstantNode* machcon = n->as_MachConstant();
1289           machcon->eval_constant(C);
1290         }
1291       }
1292     }
1293 
1294     // Calculate the offsets of the constants and the size of the
1295     // constant table (including the padding to the next section).
1296     constant_table().calculate_offsets_and_size();
1297     const_req = constant_table().size();
1298   }
1299 
1300   // Initialize the space for the BufferBlob used to find and verify
1301   // instruction size in MachNode::emit_size()
1302   init_scratch_buffer_blob(const_req);
1303   if (failing())  return NULL; // Out of memory
1304 
1305   // Pre-compute the length of blocks and replace
1306   // long branches with short if machine supports it.
1307   shorten_branches(blk_starts, code_req, locs_req, stub_req);
1308 
1309   // nmethod and CodeBuffer count stubs & constants as part of method's code.
1310   int exception_handler_req = size_exception_handler();
1311   int deopt_handler_req = size_deopt_handler();
1312   exception_handler_req += MAX_stubs_size; // add marginal slop for handler
1313   deopt_handler_req += MAX_stubs_size; // add marginal slop for handler
1314   stub_req += MAX_stubs_size;   // ensure per-stub margin
1315   code_req += MAX_inst_size;    // ensure per-instruction margin
1316 
1317   if (StressCodeBuffers)
1318     code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10;  // force expansion
1319 
1320   int total_req =
1321     const_req +
1322     code_req +
1323     pad_req +
1324     stub_req +
1325     exception_handler_req +
1326     deopt_handler_req;               // deopt handler
1327 
1328   if (has_method_handle_invokes())
1329     total_req += deopt_handler_req;  // deopt MH handler
1330 
1331   CodeBuffer* cb = code_buffer();
1332   cb->initialize(total_req, locs_req);
1333 
1334   // Have we run out of code space?
1335   if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1336     turn_off_compiler(this);
1337     return NULL;
1338   }
1339   // Configure the code buffer.
1340   cb->initialize_consts_size(const_req);
1341   cb->initialize_stubs_size(stub_req);
1342   cb->initialize_oop_recorder(env()->oop_recorder());
1343 
1344   // fill in the nop array for bundling computations
1345   MachNode *_nop_list[Bundle::_nop_count];
1346   Bundle::initialize_nops(_nop_list, this);
1347 
1348   return cb;
1349 }
1350 
1351 //------------------------------fill_buffer------------------------------------
1352 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1353 
1354   // Create oopmap set.
1355   _oop_map_set = new OopMapSet();
1356 
1357   // !!!!! This preserves old handling of oopmaps for now
1358   debug_info()->set_oopmaps(_oop_map_set);
1359 
1360   // Count and start of implicit null check instructions
1361   uint inct_cnt = 0;
1362   uint *inct_starts = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1);
1363 
1364   // Count and start of calls
1365   uint *call_returns = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1);
1366 
1367   uint  return_offset = 0;
1368   int nop_size = (new (this) MachNopNode())->size(_regalloc);
1369 
1370   int previous_offset = 0;
1371   int current_offset  = 0;
1372 #ifdef ASSERT
1373   int last_call_offset = -1;
1374   int last_avoid_back_to_back_offset = -1;
1375 #endif
1376   // Create an array of unused labels, one for each basic block, if printing is enabled
1377 #ifndef PRODUCT
1378   int *node_offsets      = NULL;
1379   uint node_offset_limit = unique();
1380 
1381   if (print_assembly())
1382     node_offsets         = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1383 #endif
1384 
1385   NonSafepointEmitter non_safepoints(this);  // emit non-safepoints lazily
1386 
1387   // Emit the constant table.
1388   if (has_mach_constant_base_node()) {
1389     constant_table().emit(*cb);
1390   }
1391 
1392   // Create an array of labels, one for each basic block
1393   Label *blk_labels = NEW_RESOURCE_ARRAY(Label, _cfg->_num_blocks+1);
1394   for (uint i=0; i <= _cfg->_num_blocks; i++) {
1395     blk_labels[i].init();
1396   }
1397 
1398   // ------------------
1399   // Now fill in the code buffer
1400   Node *delay_slot = NULL;
1401 
1402   for (uint i=0; i < _cfg->_num_blocks; i++) {
1403     guarantee(blk_starts[i] == (uint)cb->insts_size(),"should not change size");
1404 
1405     Block *b = _cfg->_blocks[i];
1406 
1407     Node *head = b->head();
1408 
1409     // If this block needs to start aligned (i.e, can be reached other
1410     // than by falling-thru from the previous block), then force the
1411     // start of a new bundle.
1412     if (Pipeline::requires_bundling() && starts_bundle(head))
1413       cb->flush_bundle(true);
1414 
1415 #ifdef ASSERT
1416     if (!b->is_connector()) {
1417       stringStream st;
1418       b->dump_head(&_cfg->_bbs, &st);
1419       MacroAssembler(cb).block_comment(st.as_string());
1420     }
1421 #endif
1422 
1423     // Define the label at the beginning of the basic block
1424     MacroAssembler(cb).bind(blk_labels[b->_pre_order]);
1425 
1426     uint last_inst = b->_nodes.size();
1427 
1428     // Emit block normally, except for last instruction.
1429     // Emit means "dump code bits into code buffer".
1430     for (uint j = 0; j<last_inst; j++) {
1431 
1432       // Get the node
1433       Node* n = b->_nodes[j];
1434 
1435       // See if delay slots are supported
1436       if (valid_bundle_info(n) &&
1437           node_bundling(n)->used_in_unconditional_delay()) {
1438         assert(delay_slot == NULL, "no use of delay slot node");
1439         assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1440 
1441         delay_slot = n;
1442         continue;
1443       }
1444 
1445       // If this starts a new instruction group, then flush the current one
1446       // (but allow split bundles)
1447       if (Pipeline::requires_bundling() && starts_bundle(n))
1448         cb->flush_bundle(false);
1449 
1450       // The following logic is duplicated in the code ifdeffed for
1451       // ENABLE_ZAP_DEAD_LOCALS which appears above in this file.  It
1452       // should be factored out.  Or maybe dispersed to the nodes?
1453 
1454       // Special handling for SafePoint/Call Nodes
1455       bool is_mcall = false;
1456       if (n->is_Mach()) {
1457         MachNode *mach = n->as_Mach();
1458         is_mcall = n->is_MachCall();
1459         bool is_sfn = n->is_MachSafePoint();
1460 
1461         // If this requires all previous instructions be flushed, then do so
1462         if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1463           cb->flush_bundle(true);
1464           current_offset = cb->insts_size();
1465         }
1466 
1467 #ifdef ASSERT
1468         // A padding may be needed again since a previous instruction
1469         // could be moved to delay slot.
1470 
1471         // align the instruction if necessary
1472         int padding = mach->compute_padding(current_offset);
1473         // Make sure safepoint node for polling is distinct from a call's
1474         // return by adding a nop if needed.
1475         if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1476           padding = nop_size;
1477         }
1478         if (padding == 0 && mach->avoid_back_to_back() &&
1479             current_offset == last_avoid_back_to_back_offset) {
1480           // Avoid back to back some instructions.
1481           padding = nop_size;
1482         }
1483         assert(padding == 0, "padding should be added already");
1484 #endif
1485         // Remember the start of the last call in a basic block
1486         if (is_mcall) {
1487           MachCallNode *mcall = mach->as_MachCall();
1488 
1489           // This destination address is NOT PC-relative
1490           mcall->method_set((intptr_t)mcall->entry_point());
1491 
1492           // Save the return address
1493           call_returns[b->_pre_order] = current_offset + mcall->ret_addr_offset();
1494 
1495           if (mcall->is_MachCallLeaf()) {
1496             is_mcall = false;
1497             is_sfn = false;
1498           }
1499         }
1500 
1501         // sfn will be valid whenever mcall is valid now because of inheritance
1502         if (is_sfn || is_mcall) {
1503 
1504           // Handle special safepoint nodes for synchronization
1505           if (!is_mcall) {
1506             MachSafePointNode *sfn = mach->as_MachSafePoint();
1507             // !!!!! Stubs only need an oopmap right now, so bail out
1508             if (sfn->jvms()->method() == NULL) {
1509               // Write the oopmap directly to the code blob??!!
1510 #             ifdef ENABLE_ZAP_DEAD_LOCALS
1511               assert( !is_node_getting_a_safepoint(sfn),  "logic does not match; false positive");
1512 #             endif
1513               continue;
1514             }
1515           } // End synchronization
1516 
1517           non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1518                                            current_offset);
1519           Process_OopMap_Node(mach, current_offset);
1520         } // End if safepoint
1521 
1522         // If this is a null check, then add the start of the previous instruction to the list
1523         else if( mach->is_MachNullCheck() ) {
1524           inct_starts[inct_cnt++] = previous_offset;
1525         }
1526 
1527         // If this is a branch, then fill in the label with the target BB's label
1528         else if (mach->is_Branch()) {
1529 
1530           if (mach->ideal_Opcode() == Op_Jump) {
1531             for (uint h = 0; h < b->_num_succs; h++) {
1532               Block* succs_block = b->_succs[h];
1533               for (uint j = 1; j < succs_block->num_preds(); j++) {
1534                 Node* jpn = succs_block->pred(j);
1535                 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1536                   uint block_num = succs_block->non_connector()->_pre_order;
1537                   Label *blkLabel = &blk_labels[block_num];
1538                   mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1539                 }
1540               }
1541             }
1542           } else {
1543             // For Branchs
1544             // This requires the TRUE branch target be in succs[0]
1545             uint block_num = b->non_connector_successor(0)->_pre_order;
1546             mach->label_set( &blk_labels[block_num], block_num );
1547           }
1548         }
1549 
1550 #ifdef ASSERT
1551         // Check that oop-store precedes the card-mark
1552         else if (mach->ideal_Opcode() == Op_StoreCM) {
1553           uint storeCM_idx = j;
1554           int count = 0;
1555           for (uint prec = mach->req(); prec < mach->len(); prec++) {
1556             Node *oop_store = mach->in(prec);  // Precedence edge
1557             if (oop_store == NULL) continue;
1558             count++;
1559             uint i4;
1560             for( i4 = 0; i4 < last_inst; ++i4 ) {
1561               if( b->_nodes[i4] == oop_store ) break;
1562             }
1563             // Note: This test can provide a false failure if other precedence
1564             // edges have been added to the storeCMNode.
1565             assert( i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1566           }
1567           assert(count > 0, "storeCM expects at least one precedence edge");
1568         }
1569 #endif
1570 
1571         else if (!n->is_Proj()) {
1572           // Remember the beginning of the previous instruction, in case
1573           // it's followed by a flag-kill and a null-check.  Happens on
1574           // Intel all the time, with add-to-memory kind of opcodes.
1575           previous_offset = current_offset;
1576         }
1577       }
1578 
1579       // Verify that there is sufficient space remaining
1580       cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1581       if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1582         turn_off_compiler(this);
1583         return;
1584       }
1585 
1586       // Save the offset for the listing
1587 #ifndef PRODUCT
1588       if (node_offsets && n->_idx < node_offset_limit)
1589         node_offsets[n->_idx] = cb->insts_size();
1590 #endif
1591 
1592       // "Normal" instruction case
1593       DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1594       n->emit(*cb, _regalloc);
1595       current_offset  = cb->insts_size();
1596 
1597 #ifdef ASSERT
1598       if (n->size(_regalloc) != (current_offset-instr_offset)) {
1599         n->dump();
1600         assert(n->size(_regalloc) == (current_offset-instr_offset), "wrong size of mach node");
1601       }
1602 #endif
1603       non_safepoints.observe_instruction(n, current_offset);
1604 
1605 #ifdef ASSERT
1606       // mcall is last "call" that can be a safepoint
1607       // record it so we can see if a poll will directly follow it
1608       // in which case we'll need a pad to make the PcDesc sites unique
1609       // see  5010568. This can be slightly inaccurate but conservative
1610       // in the case that return address is not actually at current_offset.
1611       // This is a small price to pay.
1612 
1613       if (is_mcall) {
1614         last_call_offset = current_offset;
1615       }
1616 
1617       if (n->is_Mach() && n->as_Mach()->avoid_back_to_back()) {
1618         // Avoid back to back some instructions.
1619         last_avoid_back_to_back_offset = current_offset;
1620       }
1621 #endif
1622 
1623       // See if this instruction has a delay slot
1624       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1625         assert(delay_slot != NULL, "expecting delay slot node");
1626 
1627         // Back up 1 instruction
1628         cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1629 
1630         // Save the offset for the listing
1631 #ifndef PRODUCT
1632         if (node_offsets && delay_slot->_idx < node_offset_limit)
1633           node_offsets[delay_slot->_idx] = cb->insts_size();
1634 #endif
1635 
1636         // Support a SafePoint in the delay slot
1637         if (delay_slot->is_MachSafePoint()) {
1638           MachNode *mach = delay_slot->as_Mach();
1639           // !!!!! Stubs only need an oopmap right now, so bail out
1640           if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1641             // Write the oopmap directly to the code blob??!!
1642 #           ifdef ENABLE_ZAP_DEAD_LOCALS
1643             assert( !is_node_getting_a_safepoint(mach),  "logic does not match; false positive");
1644 #           endif
1645             delay_slot = NULL;
1646             continue;
1647           }
1648 
1649           int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1650           non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1651                                            adjusted_offset);
1652           // Generate an OopMap entry
1653           Process_OopMap_Node(mach, adjusted_offset);
1654         }
1655 
1656         // Insert the delay slot instruction
1657         delay_slot->emit(*cb, _regalloc);
1658 
1659         // Don't reuse it
1660         delay_slot = NULL;
1661       }
1662 
1663     } // End for all instructions in block
1664 #ifdef ASSERT
1665     // If the next block is the top of a loop, pad this block out to align
1666     // the loop top a little. Helps prevent pipe stalls at loop back branches.
1667     if (i < _cfg->_num_blocks-1) {
1668       Block *nb = _cfg->_blocks[i+1];
1669       uint padding = nb->alignment_padding(current_offset);
1670       assert(padding == 0, "alignment should be added already");
1671     }
1672 #endif
1673   } // End of for all blocks
1674 
1675   non_safepoints.flush_at_end();
1676 
1677   // Offset too large?
1678   if (failing())  return;
1679 
1680   // Define a pseudo-label at the end of the code
1681   MacroAssembler(cb).bind( blk_labels[_cfg->_num_blocks] );
1682 
1683   // Compute the size of the first block
1684   _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1685 
1686   assert(cb->insts_size() < 500000, "method is unreasonably large");
1687 
1688   // ------------------
1689 
1690 #ifndef PRODUCT
1691   // Information on the size of the method, without the extraneous code
1692   Scheduling::increment_method_size(cb->insts_size());
1693 #endif
1694 
1695   // ------------------
1696   // Fill in exception table entries.
1697   FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1698 
1699   // Only java methods have exception handlers and deopt handlers
1700   if (_method) {
1701     // Emit the exception handler code.
1702     _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb));
1703     // Emit the deopt handler code.
1704     _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb));
1705 
1706     // Emit the MethodHandle deopt handler code (if required).
1707     if (has_method_handle_invokes()) {
1708       // We can use the same code as for the normal deopt handler, we
1709       // just need a different entry point address.
1710       _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb));
1711     }
1712   }
1713 
1714   // One last check for failed CodeBuffer::expand:
1715   if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1716     turn_off_compiler(this);
1717     return;
1718   }
1719 
1720 #ifndef PRODUCT
1721   // Dump the assembly code, including basic-block numbers
1722   if (print_assembly()) {
1723     ttyLocker ttyl;  // keep the following output all in one block
1724     if (!VMThread::should_terminate()) {  // test this under the tty lock
1725       // This output goes directly to the tty, not the compiler log.
1726       // To enable tools to match it up with the compilation activity,
1727       // be sure to tag this tty output with the compile ID.
1728       if (xtty != NULL) {
1729         xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1730                    is_osr_compilation()    ? " compile_kind='osr'" :
1731                    "");
1732       }
1733       if (method() != NULL) {
1734         method()->print_oop();
1735         print_codes();
1736       }
1737       dump_asm(node_offsets, node_offset_limit);
1738       if (xtty != NULL) {
1739         xtty->tail("opto_assembly");
1740       }
1741     }
1742   }
1743 #endif
1744 
1745 }
1746 
1747 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1748   _inc_table.set_size(cnt);
1749 
1750   uint inct_cnt = 0;
1751   for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1752     Block *b = _cfg->_blocks[i];
1753     Node *n = NULL;
1754     int j;
1755 
1756     // Find the branch; ignore trailing NOPs.
1757     for( j = b->_nodes.size()-1; j>=0; j-- ) {
1758       n = b->_nodes[j];
1759       if( !n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con )
1760         break;
1761     }
1762 
1763     // If we didn't find anything, continue
1764     if( j < 0 ) continue;
1765 
1766     // Compute ExceptionHandlerTable subtable entry and add it
1767     // (skip empty blocks)
1768     if( n->is_Catch() ) {
1769 
1770       // Get the offset of the return from the call
1771       uint call_return = call_returns[b->_pre_order];
1772 #ifdef ASSERT
1773       assert( call_return > 0, "no call seen for this basic block" );
1774       while( b->_nodes[--j]->is_MachProj() ) ;
1775       assert( b->_nodes[j]->is_MachCall(), "CatchProj must follow call" );
1776 #endif
1777       // last instruction is a CatchNode, find it's CatchProjNodes
1778       int nof_succs = b->_num_succs;
1779       // allocate space
1780       GrowableArray<intptr_t> handler_bcis(nof_succs);
1781       GrowableArray<intptr_t> handler_pcos(nof_succs);
1782       // iterate through all successors
1783       for (int j = 0; j < nof_succs; j++) {
1784         Block* s = b->_succs[j];
1785         bool found_p = false;
1786         for( uint k = 1; k < s->num_preds(); k++ ) {
1787           Node *pk = s->pred(k);
1788           if( pk->is_CatchProj() && pk->in(0) == n ) {
1789             const CatchProjNode* p = pk->as_CatchProj();
1790             found_p = true;
1791             // add the corresponding handler bci & pco information
1792             if( p->_con != CatchProjNode::fall_through_index ) {
1793               // p leads to an exception handler (and is not fall through)
1794               assert(s == _cfg->_blocks[s->_pre_order],"bad numbering");
1795               // no duplicates, please
1796               if( !handler_bcis.contains(p->handler_bci()) ) {
1797                 uint block_num = s->non_connector()->_pre_order;
1798                 handler_bcis.append(p->handler_bci());
1799                 handler_pcos.append(blk_labels[block_num].loc_pos());
1800               }
1801             }
1802           }
1803         }
1804         assert(found_p, "no matching predecessor found");
1805         // Note:  Due to empty block removal, one block may have
1806         // several CatchProj inputs, from the same Catch.
1807       }
1808 
1809       // Set the offset of the return from the call
1810       _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1811       continue;
1812     }
1813 
1814     // Handle implicit null exception table updates
1815     if( n->is_MachNullCheck() ) {
1816       uint block_num = b->non_connector_successor(0)->_pre_order;
1817       _inc_table.append( inct_starts[inct_cnt++], blk_labels[block_num].loc_pos() );
1818       continue;
1819     }
1820   } // End of for all blocks fill in exception table entries
1821 }
1822 
1823 // Static Variables
1824 #ifndef PRODUCT
1825 uint Scheduling::_total_nop_size = 0;
1826 uint Scheduling::_total_method_size = 0;
1827 uint Scheduling::_total_branches = 0;
1828 uint Scheduling::_total_unconditional_delays = 0;
1829 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1830 #endif
1831 
1832 // Initializer for class Scheduling
1833 
1834 Scheduling::Scheduling(Arena *arena, Compile &compile)
1835   : _arena(arena),
1836     _cfg(compile.cfg()),
1837     _bbs(compile.cfg()->_bbs),
1838     _regalloc(compile.regalloc()),
1839     _reg_node(arena),
1840     _bundle_instr_count(0),
1841     _bundle_cycle_number(0),
1842     _scheduled(arena),
1843     _available(arena),
1844     _next_node(NULL),
1845     _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),
1846     _pinch_free_list(arena)
1847 #ifndef PRODUCT
1848   , _branches(0)
1849   , _unconditional_delays(0)
1850 #endif
1851 {
1852   // Create a MachNopNode
1853   _nop = new (&compile) MachNopNode();
1854 
1855   // Now that the nops are in the array, save the count
1856   // (but allow entries for the nops)
1857   _node_bundling_limit = compile.unique();
1858   uint node_max = _regalloc->node_regs_max_index();
1859 
1860   compile.set_node_bundling_limit(_node_bundling_limit);
1861 
1862   // This one is persistent within the Compile class
1863   _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1864 
1865   // Allocate space for fixed-size arrays
1866   _node_latency    = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1867   _uses            = NEW_ARENA_ARRAY(arena, short,          node_max);
1868   _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1869 
1870   // Clear the arrays
1871   memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
1872   memset(_node_latency,       0, node_max * sizeof(unsigned short));
1873   memset(_uses,               0, node_max * sizeof(short));
1874   memset(_current_latency,    0, node_max * sizeof(unsigned short));
1875 
1876   // Clear the bundling information
1877   memcpy(_bundle_use_elements,
1878     Pipeline_Use::elaborated_elements,
1879     sizeof(Pipeline_Use::elaborated_elements));
1880 
1881   // Get the last node
1882   Block *bb = _cfg->_blocks[_cfg->_blocks.size()-1];
1883 
1884   _next_node = bb->_nodes[bb->_nodes.size()-1];
1885 }
1886 
1887 #ifndef PRODUCT
1888 // Scheduling destructor
1889 Scheduling::~Scheduling() {
1890   _total_branches             += _branches;
1891   _total_unconditional_delays += _unconditional_delays;
1892 }
1893 #endif
1894 
1895 // Step ahead "i" cycles
1896 void Scheduling::step(uint i) {
1897 
1898   Bundle *bundle = node_bundling(_next_node);
1899   bundle->set_starts_bundle();
1900 
1901   // Update the bundle record, but leave the flags information alone
1902   if (_bundle_instr_count > 0) {
1903     bundle->set_instr_count(_bundle_instr_count);
1904     bundle->set_resources_used(_bundle_use.resourcesUsed());
1905   }
1906 
1907   // Update the state information
1908   _bundle_instr_count = 0;
1909   _bundle_cycle_number += i;
1910   _bundle_use.step(i);
1911 }
1912 
1913 void Scheduling::step_and_clear() {
1914   Bundle *bundle = node_bundling(_next_node);
1915   bundle->set_starts_bundle();
1916 
1917   // Update the bundle record
1918   if (_bundle_instr_count > 0) {
1919     bundle->set_instr_count(_bundle_instr_count);
1920     bundle->set_resources_used(_bundle_use.resourcesUsed());
1921 
1922     _bundle_cycle_number += 1;
1923   }
1924 
1925   // Clear the bundling information
1926   _bundle_instr_count = 0;
1927   _bundle_use.reset();
1928 
1929   memcpy(_bundle_use_elements,
1930     Pipeline_Use::elaborated_elements,
1931     sizeof(Pipeline_Use::elaborated_elements));
1932 }
1933 
1934 //------------------------------ScheduleAndBundle------------------------------
1935 // Perform instruction scheduling and bundling over the sequence of
1936 // instructions in backwards order.
1937 void Compile::ScheduleAndBundle() {
1938 
1939   // Don't optimize this if it isn't a method
1940   if (!_method)
1941     return;
1942 
1943   // Don't optimize this if scheduling is disabled
1944   if (!do_scheduling())
1945     return;
1946 
1947   NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); )
1948 
1949   // Create a data structure for all the scheduling information
1950   Scheduling scheduling(Thread::current()->resource_area(), *this);
1951 
1952   // Walk backwards over each basic block, computing the needed alignment
1953   // Walk over all the basic blocks
1954   scheduling.DoScheduling();
1955 }
1956 
1957 //------------------------------ComputeLocalLatenciesForward-------------------
1958 // Compute the latency of all the instructions.  This is fairly simple,
1959 // because we already have a legal ordering.  Walk over the instructions
1960 // from first to last, and compute the latency of the instruction based
1961 // on the latency of the preceding instruction(s).
1962 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1963 #ifndef PRODUCT
1964   if (_cfg->C->trace_opto_output())
1965     tty->print("# -> ComputeLocalLatenciesForward\n");
1966 #endif
1967 
1968   // Walk over all the schedulable instructions
1969   for( uint j=_bb_start; j < _bb_end; j++ ) {
1970 
1971     // This is a kludge, forcing all latency calculations to start at 1.
1972     // Used to allow latency 0 to force an instruction to the beginning
1973     // of the bb
1974     uint latency = 1;
1975     Node *use = bb->_nodes[j];
1976     uint nlen = use->len();
1977 
1978     // Walk over all the inputs
1979     for ( uint k=0; k < nlen; k++ ) {
1980       Node *def = use->in(k);
1981       if (!def)
1982         continue;
1983 
1984       uint l = _node_latency[def->_idx] + use->latency(k);
1985       if (latency < l)
1986         latency = l;
1987     }
1988 
1989     _node_latency[use->_idx] = latency;
1990 
1991 #ifndef PRODUCT
1992     if (_cfg->C->trace_opto_output()) {
1993       tty->print("# latency %4d: ", latency);
1994       use->dump();
1995     }
1996 #endif
1997   }
1998 
1999 #ifndef PRODUCT
2000   if (_cfg->C->trace_opto_output())
2001     tty->print("# <- ComputeLocalLatenciesForward\n");
2002 #endif
2003 
2004 } // end ComputeLocalLatenciesForward
2005 
2006 // See if this node fits into the present instruction bundle
2007 bool Scheduling::NodeFitsInBundle(Node *n) {
2008   uint n_idx = n->_idx;
2009 
2010   // If this is the unconditional delay instruction, then it fits
2011   if (n == _unconditional_delay_slot) {
2012 #ifndef PRODUCT
2013     if (_cfg->C->trace_opto_output())
2014       tty->print("#     NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2015 #endif
2016     return (true);
2017   }
2018 
2019   // If the node cannot be scheduled this cycle, skip it
2020   if (_current_latency[n_idx] > _bundle_cycle_number) {
2021 #ifndef PRODUCT
2022     if (_cfg->C->trace_opto_output())
2023       tty->print("#     NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2024         n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2025 #endif
2026     return (false);
2027   }
2028 
2029   const Pipeline *node_pipeline = n->pipeline();
2030 
2031   uint instruction_count = node_pipeline->instructionCount();
2032   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2033     instruction_count = 0;
2034   else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2035     instruction_count++;
2036 
2037   if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2038 #ifndef PRODUCT
2039     if (_cfg->C->trace_opto_output())
2040       tty->print("#     NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2041         n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2042 #endif
2043     return (false);
2044   }
2045 
2046   // Don't allow non-machine nodes to be handled this way
2047   if (!n->is_Mach() && instruction_count == 0)
2048     return (false);
2049 
2050   // See if there is any overlap
2051   uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2052 
2053   if (delay > 0) {
2054 #ifndef PRODUCT
2055     if (_cfg->C->trace_opto_output())
2056       tty->print("#     NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2057 #endif
2058     return false;
2059   }
2060 
2061 #ifndef PRODUCT
2062   if (_cfg->C->trace_opto_output())
2063     tty->print("#     NodeFitsInBundle [%4d]:  TRUE\n", n_idx);
2064 #endif
2065 
2066   return true;
2067 }
2068 
2069 Node * Scheduling::ChooseNodeToBundle() {
2070   uint siz = _available.size();
2071 
2072   if (siz == 0) {
2073 
2074 #ifndef PRODUCT
2075     if (_cfg->C->trace_opto_output())
2076       tty->print("#   ChooseNodeToBundle: NULL\n");
2077 #endif
2078     return (NULL);
2079   }
2080 
2081   // Fast path, if only 1 instruction in the bundle
2082   if (siz == 1) {
2083 #ifndef PRODUCT
2084     if (_cfg->C->trace_opto_output()) {
2085       tty->print("#   ChooseNodeToBundle (only 1): ");
2086       _available[0]->dump();
2087     }
2088 #endif
2089     return (_available[0]);
2090   }
2091 
2092   // Don't bother, if the bundle is already full
2093   if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2094     for ( uint i = 0; i < siz; i++ ) {
2095       Node *n = _available[i];
2096 
2097       // Skip projections, we'll handle them another way
2098       if (n->is_Proj())
2099         continue;
2100 
2101       // This presupposed that instructions are inserted into the
2102       // available list in a legality order; i.e. instructions that
2103       // must be inserted first are at the head of the list
2104       if (NodeFitsInBundle(n)) {
2105 #ifndef PRODUCT
2106         if (_cfg->C->trace_opto_output()) {
2107           tty->print("#   ChooseNodeToBundle: ");
2108           n->dump();
2109         }
2110 #endif
2111         return (n);
2112       }
2113     }
2114   }
2115 
2116   // Nothing fits in this bundle, choose the highest priority
2117 #ifndef PRODUCT
2118   if (_cfg->C->trace_opto_output()) {
2119     tty->print("#   ChooseNodeToBundle: ");
2120     _available[0]->dump();
2121   }
2122 #endif
2123 
2124   return _available[0];
2125 }
2126 
2127 //------------------------------AddNodeToAvailableList-------------------------
2128 void Scheduling::AddNodeToAvailableList(Node *n) {
2129   assert( !n->is_Proj(), "projections never directly made available" );
2130 #ifndef PRODUCT
2131   if (_cfg->C->trace_opto_output()) {
2132     tty->print("#   AddNodeToAvailableList: ");
2133     n->dump();
2134   }
2135 #endif
2136 
2137   int latency = _current_latency[n->_idx];
2138 
2139   // Insert in latency order (insertion sort)
2140   uint i;
2141   for ( i=0; i < _available.size(); i++ )
2142     if (_current_latency[_available[i]->_idx] > latency)
2143       break;
2144 
2145   // Special Check for compares following branches
2146   if( n->is_Mach() && _scheduled.size() > 0 ) {
2147     int op = n->as_Mach()->ideal_Opcode();
2148     Node *last = _scheduled[0];
2149     if( last->is_MachIf() && last->in(1) == n &&
2150         ( op == Op_CmpI ||
2151           op == Op_CmpU ||
2152           op == Op_CmpP ||
2153           op == Op_CmpF ||
2154           op == Op_CmpD ||
2155           op == Op_CmpL ) ) {
2156 
2157       // Recalculate position, moving to front of same latency
2158       for ( i=0 ; i < _available.size(); i++ )
2159         if (_current_latency[_available[i]->_idx] >= latency)
2160           break;
2161     }
2162   }
2163 
2164   // Insert the node in the available list
2165   _available.insert(i, n);
2166 
2167 #ifndef PRODUCT
2168   if (_cfg->C->trace_opto_output())
2169     dump_available();
2170 #endif
2171 }
2172 
2173 //------------------------------DecrementUseCounts-----------------------------
2174 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2175   for ( uint i=0; i < n->len(); i++ ) {
2176     Node *def = n->in(i);
2177     if (!def) continue;
2178     if( def->is_Proj() )        // If this is a machine projection, then
2179       def = def->in(0);         // propagate usage thru to the base instruction
2180 
2181     if( _bbs[def->_idx] != bb ) // Ignore if not block-local
2182       continue;
2183 
2184     // Compute the latency
2185     uint l = _bundle_cycle_number + n->latency(i);
2186     if (_current_latency[def->_idx] < l)
2187       _current_latency[def->_idx] = l;
2188 
2189     // If this does not have uses then schedule it
2190     if ((--_uses[def->_idx]) == 0)
2191       AddNodeToAvailableList(def);
2192   }
2193 }
2194 
2195 //------------------------------AddNodeToBundle--------------------------------
2196 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2197 #ifndef PRODUCT
2198   if (_cfg->C->trace_opto_output()) {
2199     tty->print("#   AddNodeToBundle: ");
2200     n->dump();
2201   }
2202 #endif
2203 
2204   // Remove this from the available list
2205   uint i;
2206   for (i = 0; i < _available.size(); i++)
2207     if (_available[i] == n)
2208       break;
2209   assert(i < _available.size(), "entry in _available list not found");
2210   _available.remove(i);
2211 
2212   // See if this fits in the current bundle
2213   const Pipeline *node_pipeline = n->pipeline();
2214   const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2215 
2216   // Check for instructions to be placed in the delay slot. We
2217   // do this before we actually schedule the current instruction,
2218   // because the delay slot follows the current instruction.
2219   if (Pipeline::_branch_has_delay_slot &&
2220       node_pipeline->hasBranchDelay() &&
2221       !_unconditional_delay_slot) {
2222 
2223     uint siz = _available.size();
2224 
2225     // Conditional branches can support an instruction that
2226     // is unconditionally executed and not dependent by the
2227     // branch, OR a conditionally executed instruction if
2228     // the branch is taken.  In practice, this means that
2229     // the first instruction at the branch target is
2230     // copied to the delay slot, and the branch goes to
2231     // the instruction after that at the branch target
2232     if ( n->is_Mach() && n->is_Branch() ) {
2233 
2234       assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2235       assert( !n->is_Catch(),         "should not look for delay slot for Catch" );
2236 
2237 #ifndef PRODUCT
2238       _branches++;
2239 #endif
2240 
2241       // At least 1 instruction is on the available list
2242       // that is not dependent on the branch
2243       for (uint i = 0; i < siz; i++) {
2244         Node *d = _available[i];
2245         const Pipeline *avail_pipeline = d->pipeline();
2246 
2247         // Don't allow safepoints in the branch shadow, that will
2248         // cause a number of difficulties
2249         if ( avail_pipeline->instructionCount() == 1 &&
2250             !avail_pipeline->hasMultipleBundles() &&
2251             !avail_pipeline->hasBranchDelay() &&
2252             Pipeline::instr_has_unit_size() &&
2253             d->size(_regalloc) == Pipeline::instr_unit_size() &&
2254             NodeFitsInBundle(d) &&
2255             !node_bundling(d)->used_in_delay()) {
2256 
2257           if (d->is_Mach() && !d->is_MachSafePoint()) {
2258             // A node that fits in the delay slot was found, so we need to
2259             // set the appropriate bits in the bundle pipeline information so
2260             // that it correctly indicates resource usage.  Later, when we
2261             // attempt to add this instruction to the bundle, we will skip
2262             // setting the resource usage.
2263             _unconditional_delay_slot = d;
2264             node_bundling(n)->set_use_unconditional_delay();
2265             node_bundling(d)->set_used_in_unconditional_delay();
2266             _bundle_use.add_usage(avail_pipeline->resourceUse());
2267             _current_latency[d->_idx] = _bundle_cycle_number;
2268             _next_node = d;
2269             ++_bundle_instr_count;
2270 #ifndef PRODUCT
2271             _unconditional_delays++;
2272 #endif
2273             break;
2274           }
2275         }
2276       }
2277     }
2278 
2279     // No delay slot, add a nop to the usage
2280     if (!_unconditional_delay_slot) {
2281       // See if adding an instruction in the delay slot will overflow
2282       // the bundle.
2283       if (!NodeFitsInBundle(_nop)) {
2284 #ifndef PRODUCT
2285         if (_cfg->C->trace_opto_output())
2286           tty->print("#  *** STEP(1 instruction for delay slot) ***\n");
2287 #endif
2288         step(1);
2289       }
2290 
2291       _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2292       _next_node = _nop;
2293       ++_bundle_instr_count;
2294     }
2295 
2296     // See if the instruction in the delay slot requires a
2297     // step of the bundles
2298     if (!NodeFitsInBundle(n)) {
2299 #ifndef PRODUCT
2300         if (_cfg->C->trace_opto_output())
2301           tty->print("#  *** STEP(branch won't fit) ***\n");
2302 #endif
2303         // Update the state information
2304         _bundle_instr_count = 0;
2305         _bundle_cycle_number += 1;
2306         _bundle_use.step(1);
2307     }
2308   }
2309 
2310   // Get the number of instructions
2311   uint instruction_count = node_pipeline->instructionCount();
2312   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2313     instruction_count = 0;
2314 
2315   // Compute the latency information
2316   uint delay = 0;
2317 
2318   if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2319     int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2320     if (relative_latency < 0)
2321       relative_latency = 0;
2322 
2323     delay = _bundle_use.full_latency(relative_latency, node_usage);
2324 
2325     // Does not fit in this bundle, start a new one
2326     if (delay > 0) {
2327       step(delay);
2328 
2329 #ifndef PRODUCT
2330       if (_cfg->C->trace_opto_output())
2331         tty->print("#  *** STEP(%d) ***\n", delay);
2332 #endif
2333     }
2334   }
2335 
2336   // If this was placed in the delay slot, ignore it
2337   if (n != _unconditional_delay_slot) {
2338 
2339     if (delay == 0) {
2340       if (node_pipeline->hasMultipleBundles()) {
2341 #ifndef PRODUCT
2342         if (_cfg->C->trace_opto_output())
2343           tty->print("#  *** STEP(multiple instructions) ***\n");
2344 #endif
2345         step(1);
2346       }
2347 
2348       else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2349 #ifndef PRODUCT
2350         if (_cfg->C->trace_opto_output())
2351           tty->print("#  *** STEP(%d >= %d instructions) ***\n",
2352             instruction_count + _bundle_instr_count,
2353             Pipeline::_max_instrs_per_cycle);
2354 #endif
2355         step(1);
2356       }
2357     }
2358 
2359     if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2360       _bundle_instr_count++;
2361 
2362     // Set the node's latency
2363     _current_latency[n->_idx] = _bundle_cycle_number;
2364 
2365     // Now merge the functional unit information
2366     if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2367       _bundle_use.add_usage(node_usage);
2368 
2369     // Increment the number of instructions in this bundle
2370     _bundle_instr_count += instruction_count;
2371 
2372     // Remember this node for later
2373     if (n->is_Mach())
2374       _next_node = n;
2375   }
2376 
2377   // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2378   // not in the bb->_nodes array.  This happens for debug-info-only BoxLocks.
2379   // 'Schedule' them (basically ignore in the schedule) but do not insert them
2380   // into the block.  All other scheduled nodes get put in the schedule here.
2381   int op = n->Opcode();
2382   if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2383       (op != Op_Node &&         // Not an unused antidepedence node and
2384        // not an unallocated boxlock
2385        (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2386 
2387     // Push any trailing projections
2388     if( bb->_nodes[bb->_nodes.size()-1] != n ) {
2389       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2390         Node *foi = n->fast_out(i);
2391         if( foi->is_Proj() )
2392           _scheduled.push(foi);
2393       }
2394     }
2395 
2396     // Put the instruction in the schedule list
2397     _scheduled.push(n);
2398   }
2399 
2400 #ifndef PRODUCT
2401   if (_cfg->C->trace_opto_output())
2402     dump_available();
2403 #endif
2404 
2405   // Walk all the definitions, decrementing use counts, and
2406   // if a definition has a 0 use count, place it in the available list.
2407   DecrementUseCounts(n,bb);
2408 }
2409 
2410 //------------------------------ComputeUseCount--------------------------------
2411 // This method sets the use count within a basic block.  We will ignore all
2412 // uses outside the current basic block.  As we are doing a backwards walk,
2413 // any node we reach that has a use count of 0 may be scheduled.  This also
2414 // avoids the problem of cyclic references from phi nodes, as long as phi
2415 // nodes are at the front of the basic block.  This method also initializes
2416 // the available list to the set of instructions that have no uses within this
2417 // basic block.
2418 void Scheduling::ComputeUseCount(const Block *bb) {
2419 #ifndef PRODUCT
2420   if (_cfg->C->trace_opto_output())
2421     tty->print("# -> ComputeUseCount\n");
2422 #endif
2423 
2424   // Clear the list of available and scheduled instructions, just in case
2425   _available.clear();
2426   _scheduled.clear();
2427 
2428   // No delay slot specified
2429   _unconditional_delay_slot = NULL;
2430 
2431 #ifdef ASSERT
2432   for( uint i=0; i < bb->_nodes.size(); i++ )
2433     assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" );
2434 #endif
2435 
2436   // Force the _uses count to never go to zero for unscheduable pieces
2437   // of the block
2438   for( uint k = 0; k < _bb_start; k++ )
2439     _uses[bb->_nodes[k]->_idx] = 1;
2440   for( uint l = _bb_end; l < bb->_nodes.size(); l++ )
2441     _uses[bb->_nodes[l]->_idx] = 1;
2442 
2443   // Iterate backwards over the instructions in the block.  Don't count the
2444   // branch projections at end or the block header instructions.
2445   for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2446     Node *n = bb->_nodes[j];
2447     if( n->is_Proj() ) continue; // Projections handled another way
2448 
2449     // Account for all uses
2450     for ( uint k = 0; k < n->len(); k++ ) {
2451       Node *inp = n->in(k);
2452       if (!inp) continue;
2453       assert(inp != n, "no cycles allowed" );
2454       if( _bbs[inp->_idx] == bb ) { // Block-local use?
2455         if( inp->is_Proj() )    // Skip through Proj's
2456           inp = inp->in(0);
2457         ++_uses[inp->_idx];     // Count 1 block-local use
2458       }
2459     }
2460 
2461     // If this instruction has a 0 use count, then it is available
2462     if (!_uses[n->_idx]) {
2463       _current_latency[n->_idx] = _bundle_cycle_number;
2464       AddNodeToAvailableList(n);
2465     }
2466 
2467 #ifndef PRODUCT
2468     if (_cfg->C->trace_opto_output()) {
2469       tty->print("#   uses: %3d: ", _uses[n->_idx]);
2470       n->dump();
2471     }
2472 #endif
2473   }
2474 
2475 #ifndef PRODUCT
2476   if (_cfg->C->trace_opto_output())
2477     tty->print("# <- ComputeUseCount\n");
2478 #endif
2479 }
2480 
2481 // This routine performs scheduling on each basic block in reverse order,
2482 // using instruction latencies and taking into account function unit
2483 // availability.
2484 void Scheduling::DoScheduling() {
2485 #ifndef PRODUCT
2486   if (_cfg->C->trace_opto_output())
2487     tty->print("# -> DoScheduling\n");
2488 #endif
2489 
2490   Block *succ_bb = NULL;
2491   Block *bb;
2492 
2493   // Walk over all the basic blocks in reverse order
2494   for( int i=_cfg->_num_blocks-1; i >= 0; succ_bb = bb, i-- ) {
2495     bb = _cfg->_blocks[i];
2496 
2497 #ifndef PRODUCT
2498     if (_cfg->C->trace_opto_output()) {
2499       tty->print("#  Schedule BB#%03d (initial)\n", i);
2500       for (uint j = 0; j < bb->_nodes.size(); j++)
2501         bb->_nodes[j]->dump();
2502     }
2503 #endif
2504 
2505     // On the head node, skip processing
2506     if( bb == _cfg->_broot )
2507       continue;
2508 
2509     // Skip empty, connector blocks
2510     if (bb->is_connector())
2511       continue;
2512 
2513     // If the following block is not the sole successor of
2514     // this one, then reset the pipeline information
2515     if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2516 #ifndef PRODUCT
2517       if (_cfg->C->trace_opto_output()) {
2518         tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2519                    _next_node->_idx, _bundle_instr_count);
2520       }
2521 #endif
2522       step_and_clear();
2523     }
2524 
2525     // Leave untouched the starting instruction, any Phis, a CreateEx node
2526     // or Top.  bb->_nodes[_bb_start] is the first schedulable instruction.
2527     _bb_end = bb->_nodes.size()-1;
2528     for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2529       Node *n = bb->_nodes[_bb_start];
2530       // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2531       // Also, MachIdealNodes do not get scheduled
2532       if( !n->is_Mach() ) continue;     // Skip non-machine nodes
2533       MachNode *mach = n->as_Mach();
2534       int iop = mach->ideal_Opcode();
2535       if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2536       if( iop == Op_Con ) continue;      // Do not schedule Top
2537       if( iop == Op_Node &&     // Do not schedule PhiNodes, ProjNodes
2538           mach->pipeline() == MachNode::pipeline_class() &&
2539           !n->is_SpillCopy() )  // Breakpoints, Prolog, etc
2540         continue;
2541       break;                    // Funny loop structure to be sure...
2542     }
2543     // Compute last "interesting" instruction in block - last instruction we
2544     // might schedule.  _bb_end points just after last schedulable inst.  We
2545     // normally schedule conditional branches (despite them being forced last
2546     // in the block), because they have delay slots we can fill.  Calls all
2547     // have their delay slots filled in the template expansions, so we don't
2548     // bother scheduling them.
2549     Node *last = bb->_nodes[_bb_end];
2550     // Ignore trailing NOPs.
2551     while (_bb_end > 0 && last->is_Mach() &&
2552            last->as_Mach()->ideal_Opcode() == Op_Con) {
2553       last = bb->_nodes[--_bb_end];
2554     }
2555     assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2556     if( last->is_Catch() ||
2557        // Exclude unreachable path case when Halt node is in a separate block.
2558        (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2559       // There must be a prior call.  Skip it.
2560       while( !bb->_nodes[--_bb_end]->is_MachCall() ) {
2561         assert( bb->_nodes[_bb_end]->is_MachProj(), "skipping projections after expected call" );
2562       }
2563     } else if( last->is_MachNullCheck() ) {
2564       // Backup so the last null-checked memory instruction is
2565       // outside the schedulable range. Skip over the nullcheck,
2566       // projection, and the memory nodes.
2567       Node *mem = last->in(1);
2568       do {
2569         _bb_end--;
2570       } while (mem != bb->_nodes[_bb_end]);
2571     } else {
2572       // Set _bb_end to point after last schedulable inst.
2573       _bb_end++;
2574     }
2575 
2576     assert( _bb_start <= _bb_end, "inverted block ends" );
2577 
2578     // Compute the register antidependencies for the basic block
2579     ComputeRegisterAntidependencies(bb);
2580     if (_cfg->C->failing())  return;  // too many D-U pinch points
2581 
2582     // Compute intra-bb latencies for the nodes
2583     ComputeLocalLatenciesForward(bb);
2584 
2585     // Compute the usage within the block, and set the list of all nodes
2586     // in the block that have no uses within the block.
2587     ComputeUseCount(bb);
2588 
2589     // Schedule the remaining instructions in the block
2590     while ( _available.size() > 0 ) {
2591       Node *n = ChooseNodeToBundle();
2592       AddNodeToBundle(n,bb);
2593     }
2594 
2595     assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2596 #ifdef ASSERT
2597     for( uint l = _bb_start; l < _bb_end; l++ ) {
2598       Node *n = bb->_nodes[l];
2599       uint m;
2600       for( m = 0; m < _bb_end-_bb_start; m++ )
2601         if( _scheduled[m] == n )
2602           break;
2603       assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2604     }
2605 #endif
2606 
2607     // Now copy the instructions (in reverse order) back to the block
2608     for ( uint k = _bb_start; k < _bb_end; k++ )
2609       bb->_nodes.map(k, _scheduled[_bb_end-k-1]);
2610 
2611 #ifndef PRODUCT
2612     if (_cfg->C->trace_opto_output()) {
2613       tty->print("#  Schedule BB#%03d (final)\n", i);
2614       uint current = 0;
2615       for (uint j = 0; j < bb->_nodes.size(); j++) {
2616         Node *n = bb->_nodes[j];
2617         if( valid_bundle_info(n) ) {
2618           Bundle *bundle = node_bundling(n);
2619           if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2620             tty->print("*** Bundle: ");
2621             bundle->dump();
2622           }
2623           n->dump();
2624         }
2625       }
2626     }
2627 #endif
2628 #ifdef ASSERT
2629   verify_good_schedule(bb,"after block local scheduling");
2630 #endif
2631   }
2632 
2633 #ifndef PRODUCT
2634   if (_cfg->C->trace_opto_output())
2635     tty->print("# <- DoScheduling\n");
2636 #endif
2637 
2638   // Record final node-bundling array location
2639   _regalloc->C->set_node_bundling_base(_node_bundling_base);
2640 
2641 } // end DoScheduling
2642 
2643 //------------------------------verify_good_schedule---------------------------
2644 // Verify that no live-range used in the block is killed in the block by a
2645 // wrong DEF.  This doesn't verify live-ranges that span blocks.
2646 
2647 // Check for edge existence.  Used to avoid adding redundant precedence edges.
2648 static bool edge_from_to( Node *from, Node *to ) {
2649   for( uint i=0; i<from->len(); i++ )
2650     if( from->in(i) == to )
2651       return true;
2652   return false;
2653 }
2654 
2655 #ifdef ASSERT
2656 //------------------------------verify_do_def----------------------------------
2657 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2658   // Check for bad kills
2659   if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2660     Node *prior_use = _reg_node[def];
2661     if( prior_use && !edge_from_to(prior_use,n) ) {
2662       tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2663       n->dump();
2664       tty->print_cr("...");
2665       prior_use->dump();
2666       assert(edge_from_to(prior_use,n),msg);
2667     }
2668     _reg_node.map(def,NULL); // Kill live USEs
2669   }
2670 }
2671 
2672 //------------------------------verify_good_schedule---------------------------
2673 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2674 
2675   // Zap to something reasonable for the verify code
2676   _reg_node.clear();
2677 
2678   // Walk over the block backwards.  Check to make sure each DEF doesn't
2679   // kill a live value (other than the one it's supposed to).  Add each
2680   // USE to the live set.
2681   for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) {
2682     Node *n = b->_nodes[i];
2683     int n_op = n->Opcode();
2684     if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2685       // Fat-proj kills a slew of registers
2686       RegMask rm = n->out_RegMask();// Make local copy
2687       while( rm.is_NotEmpty() ) {
2688         OptoReg::Name kill = rm.find_first_elem();
2689         rm.Remove(kill);
2690         verify_do_def( n, kill, msg );
2691       }
2692     } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2693       // Get DEF'd registers the normal way
2694       verify_do_def( n, _regalloc->get_reg_first(n), msg );
2695       verify_do_def( n, _regalloc->get_reg_second(n), msg );
2696     }
2697 
2698     // Now make all USEs live
2699     for( uint i=1; i<n->req(); i++ ) {
2700       Node *def = n->in(i);
2701       assert(def != 0, "input edge required");
2702       OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2703       OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2704       if( OptoReg::is_valid(reg_lo) ) {
2705         assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);
2706         _reg_node.map(reg_lo,n);
2707       }
2708       if( OptoReg::is_valid(reg_hi) ) {
2709         assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);
2710         _reg_node.map(reg_hi,n);
2711       }
2712     }
2713 
2714   }
2715 
2716   // Zap to something reasonable for the Antidependence code
2717   _reg_node.clear();
2718 }
2719 #endif
2720 
2721 // Conditionally add precedence edges.  Avoid putting edges on Projs.
2722 static void add_prec_edge_from_to( Node *from, Node *to ) {
2723   if( from->is_Proj() ) {       // Put precedence edge on Proj's input
2724     assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2725     from = from->in(0);
2726   }
2727   if( from != to &&             // No cycles (for things like LD L0,[L0+4] )
2728       !edge_from_to( from, to ) ) // Avoid duplicate edge
2729     from->add_prec(to);
2730 }
2731 
2732 //------------------------------anti_do_def------------------------------------
2733 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2734   if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2735     return;
2736 
2737   Node *pinch = _reg_node[def_reg]; // Get pinch point
2738   if( !pinch || _bbs[pinch->_idx] != b || // No pinch-point yet?
2739       is_def ) {    // Check for a true def (not a kill)
2740     _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2741     return;
2742   }
2743 
2744   Node *kill = def;             // Rename 'def' to more descriptive 'kill'
2745   debug_only( def = (Node*)0xdeadbeef; )
2746 
2747   // After some number of kills there _may_ be a later def
2748   Node *later_def = NULL;
2749 
2750   // Finding a kill requires a real pinch-point.
2751   // Check for not already having a pinch-point.
2752   // Pinch points are Op_Node's.
2753   if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2754     later_def = pinch;            // Must be def/kill as optimistic pinch-point
2755     if ( _pinch_free_list.size() > 0) {
2756       pinch = _pinch_free_list.pop();
2757     } else {
2758       pinch = new (_cfg->C, 1) Node(1); // Pinch point to-be
2759     }
2760     if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2761       _cfg->C->record_method_not_compilable("too many D-U pinch points");
2762       return;
2763     }
2764     _bbs.map(pinch->_idx,b);      // Pretend it's valid in this block (lazy init)
2765     _reg_node.map(def_reg,pinch); // Record pinch-point
2766     //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2767     if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2768       pinch->init_req(0, _cfg->C->top());     // set not NULL for the next call
2769       add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2770       later_def = NULL;           // and no later def
2771     }
2772     pinch->set_req(0,later_def);  // Hook later def so we can find it
2773   } else {                        // Else have valid pinch point
2774     if( pinch->in(0) )            // If there is a later-def
2775       later_def = pinch->in(0);   // Get it
2776   }
2777 
2778   // Add output-dependence edge from later def to kill
2779   if( later_def )               // If there is some original def
2780     add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2781 
2782   // See if current kill is also a use, and so is forced to be the pinch-point.
2783   if( pinch->Opcode() == Op_Node ) {
2784     Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2785     for( uint i=1; i<uses->req(); i++ ) {
2786       if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2787           _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2788         // Yes, found a use/kill pinch-point
2789         pinch->set_req(0,NULL);  //
2790         pinch->replace_by(kill); // Move anti-dep edges up
2791         pinch = kill;
2792         _reg_node.map(def_reg,pinch);
2793         return;
2794       }
2795     }
2796   }
2797 
2798   // Add edge from kill to pinch-point
2799   add_prec_edge_from_to(kill,pinch);
2800 }
2801 
2802 //------------------------------anti_do_use------------------------------------
2803 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2804   if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2805     return;
2806   Node *pinch = _reg_node[use_reg]; // Get pinch point
2807   // Check for no later def_reg/kill in block
2808   if( pinch && _bbs[pinch->_idx] == b &&
2809       // Use has to be block-local as well
2810       _bbs[use->_idx] == b ) {
2811     if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2812         pinch->req() == 1 ) {   // pinch not yet in block?
2813       pinch->del_req(0);        // yank pointer to later-def, also set flag
2814       // Insert the pinch-point in the block just after the last use
2815       b->_nodes.insert(b->find_node(use)+1,pinch);
2816       _bb_end++;                // Increase size scheduled region in block
2817     }
2818 
2819     add_prec_edge_from_to(pinch,use);
2820   }
2821 }
2822 
2823 //------------------------------ComputeRegisterAntidependences-----------------
2824 // We insert antidependences between the reads and following write of
2825 // allocated registers to prevent illegal code motion. Hopefully, the
2826 // number of added references should be fairly small, especially as we
2827 // are only adding references within the current basic block.
2828 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2829 
2830 #ifdef ASSERT
2831   verify_good_schedule(b,"before block local scheduling");
2832 #endif
2833 
2834   // A valid schedule, for each register independently, is an endless cycle
2835   // of: a def, then some uses (connected to the def by true dependencies),
2836   // then some kills (defs with no uses), finally the cycle repeats with a new
2837   // def.  The uses are allowed to float relative to each other, as are the
2838   // kills.  No use is allowed to slide past a kill (or def).  This requires
2839   // antidependencies between all uses of a single def and all kills that
2840   // follow, up to the next def.  More edges are redundant, because later defs
2841   // & kills are already serialized with true or antidependencies.  To keep
2842   // the edge count down, we add a 'pinch point' node if there's more than
2843   // one use or more than one kill/def.
2844 
2845   // We add dependencies in one bottom-up pass.
2846 
2847   // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2848 
2849   // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2850   // register.  If not, we record the DEF/KILL in _reg_node, the
2851   // register-to-def mapping.  If there is a prior DEF/KILL, we insert a
2852   // "pinch point", a new Node that's in the graph but not in the block.
2853   // We put edges from the prior and current DEF/KILLs to the pinch point.
2854   // We put the pinch point in _reg_node.  If there's already a pinch point
2855   // we merely add an edge from the current DEF/KILL to the pinch point.
2856 
2857   // After doing the DEF/KILLs, we handle USEs.  For each used register, we
2858   // put an edge from the pinch point to the USE.
2859 
2860   // To be expedient, the _reg_node array is pre-allocated for the whole
2861   // compilation.  _reg_node is lazily initialized; it either contains a NULL,
2862   // or a valid def/kill/pinch-point, or a leftover node from some prior
2863   // block.  Leftover node from some prior block is treated like a NULL (no
2864   // prior def, so no anti-dependence needed).  Valid def is distinguished by
2865   // it being in the current block.
2866   bool fat_proj_seen = false;
2867   uint last_safept = _bb_end-1;
2868   Node* end_node         = (_bb_end-1 >= _bb_start) ? b->_nodes[last_safept] : NULL;
2869   Node* last_safept_node = end_node;
2870   for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2871     Node *n = b->_nodes[i];
2872     int is_def = n->outcnt();   // def if some uses prior to adding precedence edges
2873     if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2874       // Fat-proj kills a slew of registers
2875       // This can add edges to 'n' and obscure whether or not it was a def,
2876       // hence the is_def flag.
2877       fat_proj_seen = true;
2878       RegMask rm = n->out_RegMask();// Make local copy
2879       while( rm.is_NotEmpty() ) {
2880         OptoReg::Name kill = rm.find_first_elem();
2881         rm.Remove(kill);
2882         anti_do_def( b, n, kill, is_def );
2883       }
2884     } else {
2885       // Get DEF'd registers the normal way
2886       anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2887       anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2888     }
2889 
2890     // Kill projections on a branch should appear to occur on the
2891     // branch, not afterwards, so grab the masks from the projections
2892     // and process them.
2893     if (n->is_Branch()) {
2894       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2895         Node* use = n->fast_out(i);
2896         if (use->is_Proj()) {
2897           RegMask rm = use->out_RegMask();// Make local copy
2898           while( rm.is_NotEmpty() ) {
2899             OptoReg::Name kill = rm.find_first_elem();
2900             rm.Remove(kill);
2901             anti_do_def( b, n, kill, false );
2902           }
2903         }
2904       }
2905     }
2906 
2907     // Check each register used by this instruction for a following DEF/KILL
2908     // that must occur afterward and requires an anti-dependence edge.
2909     for( uint j=0; j<n->req(); j++ ) {
2910       Node *def = n->in(j);
2911       if( def ) {
2912         assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2913         anti_do_use( b, n, _regalloc->get_reg_first(def) );
2914         anti_do_use( b, n, _regalloc->get_reg_second(def) );
2915       }
2916     }
2917     // Do not allow defs of new derived values to float above GC
2918     // points unless the base is definitely available at the GC point.
2919 
2920     Node *m = b->_nodes[i];
2921 
2922     // Add precedence edge from following safepoint to use of derived pointer
2923     if( last_safept_node != end_node &&
2924         m != last_safept_node) {
2925       for (uint k = 1; k < m->req(); k++) {
2926         const Type *t = m->in(k)->bottom_type();
2927         if( t->isa_oop_ptr() &&
2928             t->is_ptr()->offset() != 0 ) {
2929           last_safept_node->add_prec( m );
2930           break;
2931         }
2932       }
2933     }
2934 
2935     if( n->jvms() ) {           // Precedence edge from derived to safept
2936       // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2937       if( b->_nodes[last_safept] != last_safept_node ) {
2938         last_safept = b->find_node(last_safept_node);
2939       }
2940       for( uint j=last_safept; j > i; j-- ) {
2941         Node *mach = b->_nodes[j];
2942         if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2943           mach->add_prec( n );
2944       }
2945       last_safept = i;
2946       last_safept_node = m;
2947     }
2948   }
2949 
2950   if (fat_proj_seen) {
2951     // Garbage collect pinch nodes that were not consumed.
2952     // They are usually created by a fat kill MachProj for a call.
2953     garbage_collect_pinch_nodes();
2954   }
2955 }
2956 
2957 //------------------------------garbage_collect_pinch_nodes-------------------------------
2958 
2959 // Garbage collect pinch nodes for reuse by other blocks.
2960 //
2961 // The block scheduler's insertion of anti-dependence
2962 // edges creates many pinch nodes when the block contains
2963 // 2 or more Calls.  A pinch node is used to prevent a
2964 // combinatorial explosion of edges.  If a set of kills for a
2965 // register is anti-dependent on a set of uses (or defs), rather
2966 // than adding an edge in the graph between each pair of kill
2967 // and use (or def), a pinch is inserted between them:
2968 //
2969 //            use1   use2  use3
2970 //                \   |   /
2971 //                 \  |  /
2972 //                  pinch
2973 //                 /  |  \
2974 //                /   |   \
2975 //            kill1 kill2 kill3
2976 //
2977 // One pinch node is created per register killed when
2978 // the second call is encountered during a backwards pass
2979 // over the block.  Most of these pinch nodes are never
2980 // wired into the graph because the register is never
2981 // used or def'ed in the block.
2982 //
2983 void Scheduling::garbage_collect_pinch_nodes() {
2984 #ifndef PRODUCT
2985     if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2986 #endif
2987     int trace_cnt = 0;
2988     for (uint k = 0; k < _reg_node.Size(); k++) {
2989       Node* pinch = _reg_node[k];
2990       if (pinch != NULL && pinch->Opcode() == Op_Node &&
2991           // no predecence input edges
2992           (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
2993         cleanup_pinch(pinch);
2994         _pinch_free_list.push(pinch);
2995         _reg_node.map(k, NULL);
2996 #ifndef PRODUCT
2997         if (_cfg->C->trace_opto_output()) {
2998           trace_cnt++;
2999           if (trace_cnt > 40) {
3000             tty->print("\n");
3001             trace_cnt = 0;
3002           }
3003           tty->print(" %d", pinch->_idx);
3004         }
3005 #endif
3006       }
3007     }
3008 #ifndef PRODUCT
3009     if (_cfg->C->trace_opto_output()) tty->print("\n");
3010 #endif
3011 }
3012 
3013 // Clean up a pinch node for reuse.
3014 void Scheduling::cleanup_pinch( Node *pinch ) {
3015   assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3016 
3017   for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3018     Node* use = pinch->last_out(i);
3019     uint uses_found = 0;
3020     for (uint j = use->req(); j < use->len(); j++) {
3021       if (use->in(j) == pinch) {
3022         use->rm_prec(j);
3023         uses_found++;
3024       }
3025     }
3026     assert(uses_found > 0, "must be a precedence edge");
3027     i -= uses_found;    // we deleted 1 or more copies of this edge
3028   }
3029   // May have a later_def entry
3030   pinch->set_req(0, NULL);
3031 }
3032 
3033 //------------------------------print_statistics-------------------------------
3034 #ifndef PRODUCT
3035 
3036 void Scheduling::dump_available() const {
3037   tty->print("#Availist  ");
3038   for (uint i = 0; i < _available.size(); i++)
3039     tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3040   tty->cr();
3041 }
3042 
3043 // Print Scheduling Statistics
3044 void Scheduling::print_statistics() {
3045   // Print the size added by nops for bundling
3046   tty->print("Nops added %d bytes to total of %d bytes",
3047     _total_nop_size, _total_method_size);
3048   if (_total_method_size > 0)
3049     tty->print(", for %.2f%%",
3050       ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3051   tty->print("\n");
3052 
3053   // Print the number of branch shadows filled
3054   if (Pipeline::_branch_has_delay_slot) {
3055     tty->print("Of %d branches, %d had unconditional delay slots filled",
3056       _total_branches, _total_unconditional_delays);
3057     if (_total_branches > 0)
3058       tty->print(", for %.2f%%",
3059         ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3060     tty->print("\n");
3061   }
3062 
3063   uint total_instructions = 0, total_bundles = 0;
3064 
3065   for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3066     uint bundle_count   = _total_instructions_per_bundle[i];
3067     total_instructions += bundle_count * i;
3068     total_bundles      += bundle_count;
3069   }
3070 
3071   if (total_bundles > 0)
3072     tty->print("Average ILP (excluding nops) is %.2f\n",
3073       ((double)total_instructions) / ((double)total_bundles));
3074 }
3075 #endif