1 /*
   2  * Copyright (c) 1998, 2025, 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 "asm/assembler.inline.hpp"
  26 #include "code/aotCodeCache.hpp"
  27 #include "code/compiledIC.hpp"
  28 #include "code/debugInfo.hpp"
  29 #include "code/debugInfoRec.hpp"
  30 #include "compiler/compileBroker.hpp"
  31 #include "compiler/compilerDirectives.hpp"
  32 #include "compiler/disassembler.hpp"
  33 #include "compiler/oopMap.hpp"
  34 #include "gc/shared/barrierSet.hpp"
  35 #include "gc/shared/c2/barrierSetC2.hpp"
  36 #include "memory/allocation.hpp"
  37 #include "opto/ad.hpp"
  38 #include "opto/block.hpp"
  39 #include "opto/c2_MacroAssembler.hpp"
  40 #include "opto/c2compiler.hpp"
  41 #include "opto/callnode.hpp"
  42 #include "opto/cfgnode.hpp"
  43 #include "opto/locknode.hpp"
  44 #include "opto/machnode.hpp"
  45 #include "opto/node.hpp"
  46 #include "opto/optoreg.hpp"
  47 #include "opto/output.hpp"
  48 #include "opto/regalloc.hpp"
  49 #include "opto/type.hpp"
  50 #include "runtime/sharedRuntime.hpp"
  51 #include "utilities/macros.hpp"
  52 #include "utilities/powerOfTwo.hpp"
  53 #include "utilities/xmlstream.hpp"
  54 
  55 #ifndef PRODUCT
  56 #define DEBUG_ARG(x) , x
  57 #else
  58 #define DEBUG_ARG(x)
  59 #endif
  60 
  61 //------------------------------Scheduling----------------------------------
  62 // This class contains all the information necessary to implement instruction
  63 // scheduling and bundling.
  64 class Scheduling {
  65 
  66 private:
  67   // Arena to use
  68   Arena *_arena;
  69 
  70   // Control-Flow Graph info
  71   PhaseCFG *_cfg;
  72 
  73   // Register Allocation info
  74   PhaseRegAlloc *_regalloc;
  75 
  76   // Number of nodes in the method
  77   uint _node_bundling_limit;
  78 
  79   // List of scheduled nodes. Generated in reverse order
  80   Node_List _scheduled;
  81 
  82   // List of nodes currently available for choosing for scheduling
  83   Node_List _available;
  84 
  85   // For each instruction beginning a bundle, the number of following
  86   // nodes to be bundled with it.
  87   Bundle *_node_bundling_base;
  88 
  89   // Mapping from register to Node
  90   Node_List _reg_node;
  91 
  92   // Free list for pinch nodes.
  93   Node_List _pinch_free_list;
  94 
  95   // Number of uses of this node within the containing basic block.
  96   short *_uses;
  97 
  98   // Schedulable portion of current block.  Skips Region/Phi/CreateEx up
  99   // front, branch+proj at end.  Also skips Catch/CProj (same as
 100   // branch-at-end), plus just-prior exception-throwing call.
 101   uint _bb_start, _bb_end;
 102 
 103   // Latency from the end of the basic block as scheduled
 104   unsigned short *_current_latency;
 105 
 106   // Remember the next node
 107   Node *_next_node;
 108 
 109   // Length of the current bundle, in instructions
 110   uint _bundle_instr_count;
 111 
 112   // Current Cycle number, for computing latencies and bundling
 113   uint _bundle_cycle_number;
 114 
 115   // Bundle information
 116   Pipeline_Use_Element _bundle_use_elements[resource_count];
 117   Pipeline_Use         _bundle_use;
 118 
 119   // Dump the available list
 120   void dump_available() const;
 121 
 122 public:
 123   Scheduling(Arena *arena, Compile &compile);
 124 
 125   // Step ahead "i" cycles
 126   void step(uint i);
 127 
 128   // Step ahead 1 cycle, and clear the bundle state (for example,
 129   // at a branch target)
 130   void step_and_clear();
 131 
 132   Bundle* node_bundling(const Node *n) {
 133     assert(valid_bundle_info(n), "oob");
 134     return (&_node_bundling_base[n->_idx]);
 135   }
 136 
 137   bool valid_bundle_info(const Node *n) const {
 138     return (_node_bundling_limit > n->_idx);
 139   }
 140 
 141   bool starts_bundle(const Node *n) const {
 142     return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle());
 143   }
 144 
 145   // Do the scheduling
 146   void DoScheduling();
 147 
 148   // Compute the register antidependencies within a basic block
 149   void ComputeRegisterAntidependencies(Block *bb);
 150   void verify_do_def( Node *n, OptoReg::Name def, const char *msg );
 151   void verify_good_schedule( Block *b, const char *msg );
 152   void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def );
 153   void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg );
 154 
 155   // Add a node to the current bundle
 156   void AddNodeToBundle(Node *n, const Block *bb);
 157 
 158   // Return an integer less than, equal to, or greater than zero
 159   // if the stack offset of the first argument is respectively
 160   // less than, equal to, or greater than the second.
 161   int compare_two_spill_nodes(Node* first, Node* second);
 162 
 163   // Add a node to the list of available nodes
 164   void AddNodeToAvailableList(Node *n);
 165 
 166   // Compute the local use count for the nodes in a block, and compute
 167   // the list of instructions with no uses in the block as available
 168   void ComputeUseCount(const Block *bb);
 169 
 170   // Choose an instruction from the available list to add to the bundle
 171   Node * ChooseNodeToBundle();
 172 
 173   // See if this Node fits into the currently accumulating bundle
 174   bool NodeFitsInBundle(Node *n);
 175 
 176   // Decrement the use count for a node
 177  void DecrementUseCounts(Node *n, const Block *bb);
 178 
 179   // Garbage collect pinch nodes for reuse by other blocks.
 180   void garbage_collect_pinch_nodes();
 181   // Clean up a pinch node for reuse (helper for above).
 182   void cleanup_pinch( Node *pinch );
 183 
 184   // Information for statistics gathering
 185 #ifndef PRODUCT
 186 private:
 187   // Gather information on size of nops relative to total
 188   static uint _total_nop_size, _total_method_size;
 189   static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
 190 
 191 public:
 192   static void print_statistics();
 193 
 194   static void increment_instructions_per_bundle(uint i) {
 195     _total_instructions_per_bundle[i]++;
 196   }
 197 
 198   static void increment_nop_size(uint s) {
 199     _total_nop_size += s;
 200   }
 201 
 202   static void increment_method_size(uint s) {
 203     _total_method_size += s;
 204   }
 205 #endif
 206 
 207 };
 208 
 209 PhaseOutput::PhaseOutput()
 210   : Phase(Phase::Output),
 211     _code_buffer("Compile::Fill_buffer"),
 212     _first_block_size(0),
 213     _handler_table(),
 214     _inc_table(),
 215     _stub_list(),
 216     _oop_map_set(nullptr),
 217     _scratch_buffer_blob(nullptr),
 218     _scratch_locs_memory(nullptr),
 219     _scratch_const_size(-1),
 220     _in_scratch_emit_size(false),
 221     _frame_slots(0),
 222     _code_offsets(),
 223     _node_bundling_limit(0),
 224     _node_bundling_base(nullptr),
 225     _orig_pc_slot(0),
 226     _orig_pc_slot_offset_in_bytes(0),
 227     _buf_sizes(),
 228     _block(nullptr),
 229     _index(0) {
 230   C->set_output(this);
 231   if (C->stub_name() == nullptr) {
 232     _orig_pc_slot = C->fixed_slots() - (sizeof(address) / VMRegImpl::stack_slot_size);
 233   }
 234 }
 235 
 236 PhaseOutput::~PhaseOutput() {
 237   C->set_output(nullptr);
 238   if (_scratch_buffer_blob != nullptr) {
 239     BufferBlob::free(_scratch_buffer_blob);
 240   }
 241 }
 242 
 243 void PhaseOutput::perform_mach_node_analysis() {
 244   // Late barrier analysis must be done after schedule and bundle
 245   // Otherwise liveness based spilling will fail
 246   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 247   bs->late_barrier_analysis();
 248 
 249   pd_perform_mach_node_analysis();
 250 
 251   C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
 252 }
 253 
 254 // Convert Nodes to instruction bits and pass off to the VM
 255 void PhaseOutput::Output() {
 256   // RootNode goes
 257   assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
 258 
 259   // The number of new nodes (mostly MachNop) is proportional to
 260   // the number of java calls and inner loops which are aligned.
 261   if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
 262                             C->inner_loops()*(OptoLoopAlignment-1)),
 263                            "out of nodes before code generation" ) ) {
 264     return;
 265   }
 266   // Make sure I can find the Start Node
 267   Block *entry = C->cfg()->get_block(1);
 268   Block *broot = C->cfg()->get_root_block();
 269 
 270   const StartNode *start = entry->head()->as_Start();
 271 
 272   // Replace StartNode with prolog
 273   MachPrologNode *prolog = new MachPrologNode();
 274   entry->map_node(prolog, 0);
 275   C->cfg()->map_node_to_block(prolog, entry);
 276   C->cfg()->unmap_node_from_block(start); // start is no longer in any block
 277 
 278   // Virtual methods need an unverified entry point
 279 
 280   if( C->is_osr_compilation() ) {
 281     if( PoisonOSREntry ) {
 282       // TODO: Should use a ShouldNotReachHereNode...
 283       C->cfg()->insert( broot, 0, new MachBreakpointNode() );
 284     }
 285   } else {
 286     if( C->method() && !C->method()->flags().is_static() ) {
 287       // Insert unvalidated entry point
 288       C->cfg()->insert( broot, 0, new MachUEPNode() );
 289     }
 290 
 291   }
 292 
 293   // Break before main entry point
 294   if ((C->method() && C->directive()->BreakAtExecuteOption) ||
 295       (OptoBreakpoint && C->is_method_compilation())       ||
 296       (OptoBreakpointOSR && C->is_osr_compilation())       ||
 297       (OptoBreakpointC2R && !C->method())                   ) {
 298     // checking for C->method() means that OptoBreakpoint does not apply to
 299     // runtime stubs or frame converters
 300     C->cfg()->insert( entry, 1, new MachBreakpointNode() );
 301   }
 302 
 303   // Insert epilogs before every return
 304   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
 305     Block* block = C->cfg()->get_block(i);
 306     if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
 307       Node* m = block->end();
 308       if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
 309         MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
 310         block->add_inst(epilog);
 311         C->cfg()->map_node_to_block(epilog, block);
 312       }
 313     }
 314   }
 315 
 316   // Keeper of sizing aspects
 317   _buf_sizes = BufferSizingData();
 318 
 319   // Initialize code buffer
 320   estimate_buffer_size(_buf_sizes._const);
 321   if (C->failing()) return;
 322 
 323   // Pre-compute the length of blocks and replace
 324   // long branches with short if machine supports it.
 325   // Must be done before ScheduleAndBundle due to SPARC delay slots
 326   uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
 327   blk_starts[0] = 0;
 328   shorten_branches(blk_starts);
 329 
 330   ScheduleAndBundle();
 331   if (C->failing()) {
 332     return;
 333   }
 334 
 335   perform_mach_node_analysis();
 336 
 337   // Complete sizing of codebuffer
 338   CodeBuffer* cb = init_buffer();
 339   if (cb == nullptr || C->failing()) {
 340     return;
 341   }
 342 
 343   BuildOopMaps();
 344 
 345   if (C->failing())  {
 346     return;
 347   }
 348 
 349   C2_MacroAssembler masm(cb);
 350   fill_buffer(&masm, blk_starts);
 351 }
 352 
 353 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
 354   // Determine if we need to generate a stack overflow check.
 355   // Do it if the method is not a stub function and
 356   // has java calls or has frame size > vm_page_size/8.
 357   // The debug VM checks that deoptimization doesn't trigger an
 358   // unexpected stack overflow (compiled method stack banging should
 359   // guarantee it doesn't happen) so we always need the stack bang in
 360   // a debug VM.
 361   return (C->stub_function() == nullptr &&
 362           (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
 363            DEBUG_ONLY(|| true)));
 364 }
 365 
 366 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
 367 // of a loop. When aligning a loop we need to provide enough instructions
 368 // in cpu's fetch buffer to feed decoders. The loop alignment could be
 369 // avoided if we have enough instructions in fetch buffer at the head of a loop.
 370 // By default, the size is set to 999999 by Block's constructor so that
 371 // a loop will be aligned if the size is not reset here.
 372 //
 373 // Note: Mach instructions could contain several HW instructions
 374 // so the size is estimated only.
 375 //
 376 void PhaseOutput::compute_loop_first_inst_sizes() {
 377   // The next condition is used to gate the loop alignment optimization.
 378   // Don't aligned a loop if there are enough instructions at the head of a loop
 379   // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
 380   // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
 381   // equal to 11 bytes which is the largest address NOP instruction.
 382   if (MaxLoopPad < OptoLoopAlignment - 1) {
 383     uint last_block = C->cfg()->number_of_blocks() - 1;
 384     for (uint i = 1; i <= last_block; i++) {
 385       Block* block = C->cfg()->get_block(i);
 386       // Check the first loop's block which requires an alignment.
 387       if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
 388         uint sum_size = 0;
 389         uint inst_cnt = NumberOfLoopInstrToAlign;
 390         inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
 391 
 392         // Check subsequent fallthrough blocks if the loop's first
 393         // block(s) does not have enough instructions.
 394         Block *nb = block;
 395         while(inst_cnt > 0 &&
 396               i < last_block &&
 397               !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
 398               !nb->has_successor(block)) {
 399           i++;
 400           nb = C->cfg()->get_block(i);
 401           inst_cnt  = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
 402         } // while( inst_cnt > 0 && i < last_block  )
 403 
 404         block->set_first_inst_size(sum_size);
 405       } // f( b->head()->is_Loop() )
 406     } // for( i <= last_block )
 407   } // if( MaxLoopPad < OptoLoopAlignment-1 )
 408 }
 409 
 410 // The architecture description provides short branch variants for some long
 411 // branch instructions. Replace eligible long branches with short branches.
 412 void PhaseOutput::shorten_branches(uint* blk_starts) {
 413 
 414   Compile::TracePhase tp(_t_shortenBranches);
 415 
 416   // Compute size of each block, method size, and relocation information size
 417   uint nblocks  = C->cfg()->number_of_blocks();
 418 
 419   uint*      jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
 420   uint*      jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
 421   int*       jmp_nidx   = NEW_RESOURCE_ARRAY(int ,nblocks);
 422 
 423   // Collect worst case block paddings
 424   int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
 425   memset(block_worst_case_pad, 0, nblocks * sizeof(int));
 426 
 427   DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
 428   DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
 429 
 430   bool has_short_branch_candidate = false;
 431 
 432   // Initialize the sizes to 0
 433   int code_size  = 0;          // Size in bytes of generated code
 434   int stub_size  = 0;          // Size in bytes of all stub entries
 435   // Size in bytes of all relocation entries, including those in local stubs.
 436   // Start with 2-bytes of reloc info for the unvalidated entry point
 437   int reloc_size = 1;          // Number of relocation entries
 438 
 439   // Make three passes.  The first computes pessimistic blk_starts,
 440   // relative jmp_offset and reloc_size information.  The second performs
 441   // short branch substitution using the pessimistic sizing.  The
 442   // third inserts nops where needed.
 443 
 444   // Step one, perform a pessimistic sizing pass.
 445   uint last_call_adr = max_juint;
 446   uint last_avoid_back_to_back_adr = max_juint;
 447   uint nop_size = (new MachNopNode())->size(C->regalloc());
 448   for (uint i = 0; i < nblocks; i++) { // For all blocks
 449     Block* block = C->cfg()->get_block(i);
 450     _block = block;
 451 
 452     // During short branch replacement, we store the relative (to blk_starts)
 453     // offset of jump in jmp_offset, rather than the absolute offset of jump.
 454     // This is so that we do not need to recompute sizes of all nodes when
 455     // we compute correct blk_starts in our next sizing pass.
 456     jmp_offset[i] = 0;
 457     jmp_size[i]   = 0;
 458     jmp_nidx[i]   = -1;
 459     DEBUG_ONLY( jmp_target[i] = 0; )
 460     DEBUG_ONLY( jmp_rule[i]   = 0; )
 461 
 462     // Sum all instruction sizes to compute block size
 463     uint last_inst = block->number_of_nodes();
 464     uint blk_size = 0;
 465     for (uint j = 0; j < last_inst; j++) {
 466       _index = j;
 467       Node* nj = block->get_node(_index);
 468       // Handle machine instruction nodes
 469       if (nj->is_Mach()) {
 470         MachNode* mach = nj->as_Mach();
 471         blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
 472         reloc_size += mach->reloc();
 473         if (mach->is_MachCall()) {
 474           // add size information for trampoline stub
 475           // class CallStubImpl is platform-specific and defined in the *.ad files.
 476           stub_size  += CallStubImpl::size_call_trampoline();
 477           reloc_size += CallStubImpl::reloc_call_trampoline();
 478 
 479           MachCallNode *mcall = mach->as_MachCall();
 480           // This destination address is NOT PC-relative
 481 
 482           mcall->method_set((intptr_t)mcall->entry_point());
 483 
 484           if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
 485             stub_size  += CompiledDirectCall::to_interp_stub_size();
 486             reloc_size += CompiledDirectCall::reloc_to_interp_stub();
 487           }
 488         } else if (mach->is_MachSafePoint()) {
 489           // If call/safepoint are adjacent, account for possible
 490           // nop to disambiguate the two safepoints.
 491           // ScheduleAndBundle() can rearrange nodes in a block,
 492           // check for all offsets inside this block.
 493           if (last_call_adr >= blk_starts[i]) {
 494             blk_size += nop_size;
 495           }
 496         }
 497         if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
 498           // Nop is inserted between "avoid back to back" instructions.
 499           // ScheduleAndBundle() can rearrange nodes in a block,
 500           // check for all offsets inside this block.
 501           if (last_avoid_back_to_back_adr >= blk_starts[i]) {
 502             blk_size += nop_size;
 503           }
 504         }
 505         if (mach->may_be_short_branch()) {
 506           if (!nj->is_MachBranch()) {
 507 #ifndef PRODUCT
 508             nj->dump(3);
 509 #endif
 510             Unimplemented();
 511           }
 512           assert(jmp_nidx[i] == -1, "block should have only one branch");
 513           jmp_offset[i] = blk_size;
 514           jmp_size[i]   = nj->size(C->regalloc());
 515           jmp_nidx[i]   = j;
 516           has_short_branch_candidate = true;
 517         }
 518       }
 519       blk_size += nj->size(C->regalloc());
 520       // Remember end of call offset
 521       if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
 522         last_call_adr = blk_starts[i]+blk_size;
 523       }
 524       // Remember end of avoid_back_to_back offset
 525       if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
 526         last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
 527       }
 528     }
 529 
 530     // When the next block starts a loop, we may insert pad NOP
 531     // instructions.  Since we cannot know our future alignment,
 532     // assume the worst.
 533     if (i < nblocks - 1) {
 534       Block* nb = C->cfg()->get_block(i + 1);
 535       int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
 536       if (max_loop_pad > 0) {
 537         assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
 538         // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
 539         // If either is the last instruction in this block, bump by
 540         // max_loop_pad in lock-step with blk_size, so sizing
 541         // calculations in subsequent blocks still can conservatively
 542         // detect that it may the last instruction in this block.
 543         if (last_call_adr == blk_starts[i]+blk_size) {
 544           last_call_adr += max_loop_pad;
 545         }
 546         if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
 547           last_avoid_back_to_back_adr += max_loop_pad;
 548         }
 549         blk_size += max_loop_pad;
 550         block_worst_case_pad[i + 1] = max_loop_pad;
 551       }
 552     }
 553 
 554     // Save block size; update total method size
 555     blk_starts[i+1] = blk_starts[i]+blk_size;
 556   }
 557 
 558   // Step two, replace eligible long jumps.
 559   bool progress = true;
 560   uint last_may_be_short_branch_adr = max_juint;
 561   while (has_short_branch_candidate && progress) {
 562     progress = false;
 563     has_short_branch_candidate = false;
 564     int adjust_block_start = 0;
 565     for (uint i = 0; i < nblocks; i++) {
 566       Block* block = C->cfg()->get_block(i);
 567       int idx = jmp_nidx[i];
 568       MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach();
 569       if (mach != nullptr && mach->may_be_short_branch()) {
 570 #ifdef ASSERT
 571         assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
 572         int j;
 573         // Find the branch; ignore trailing NOPs.
 574         for (j = block->number_of_nodes()-1; j>=0; j--) {
 575           Node* n = block->get_node(j);
 576           if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
 577             break;
 578         }
 579         assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
 580 #endif
 581         int br_size = jmp_size[i];
 582         int br_offs = blk_starts[i] + jmp_offset[i];
 583 
 584         // This requires the TRUE branch target be in succs[0]
 585         uint bnum = block->non_connector_successor(0)->_pre_order;
 586         int offset = blk_starts[bnum] - br_offs;
 587         if (bnum > i) { // adjust following block's offset
 588           offset -= adjust_block_start;
 589         }
 590 
 591         // This block can be a loop header, account for the padding
 592         // in the previous block.
 593         int block_padding = block_worst_case_pad[i];
 594         assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
 595         // In the following code a nop could be inserted before
 596         // the branch which will increase the backward distance.
 597         bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
 598         assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
 599 
 600         if (needs_padding && offset <= 0)
 601           offset -= nop_size;
 602 
 603         if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
 604           // We've got a winner.  Replace this branch.
 605           MachNode* replacement = mach->as_MachBranch()->short_branch_version();
 606 
 607           // Update the jmp_size.
 608           int new_size = replacement->size(C->regalloc());
 609           int diff     = br_size - new_size;
 610           assert(diff >= (int)nop_size, "short_branch size should be smaller");
 611           // Conservatively take into account padding between
 612           // avoid_back_to_back branches. Previous branch could be
 613           // converted into avoid_back_to_back branch during next
 614           // rounds.
 615           if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
 616             jmp_offset[i] += nop_size;
 617             diff -= nop_size;
 618           }
 619           adjust_block_start += diff;
 620           block->map_node(replacement, idx);
 621           mach->subsume_by(replacement, C);
 622           mach = replacement;
 623           progress = true;
 624 
 625           jmp_size[i] = new_size;
 626           DEBUG_ONLY( jmp_target[i] = bnum; );
 627           DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
 628         } else {
 629           // The jump distance is not short, try again during next iteration.
 630           has_short_branch_candidate = true;
 631         }
 632       } // (mach->may_be_short_branch())
 633       if (mach != nullptr && (mach->may_be_short_branch() ||
 634                            mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
 635         last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
 636       }
 637       blk_starts[i+1] -= adjust_block_start;
 638     }
 639   }
 640 
 641 #ifdef ASSERT
 642   for (uint i = 0; i < nblocks; i++) { // For all blocks
 643     if (jmp_target[i] != 0) {
 644       int br_size = jmp_size[i];
 645       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
 646       if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
 647         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]);
 648       }
 649       assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
 650     }
 651   }
 652 #endif
 653 
 654   // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
 655   // after ScheduleAndBundle().
 656 
 657   // ------------------
 658   // Compute size for code buffer
 659   code_size = blk_starts[nblocks];
 660 
 661   // Relocation records
 662   reloc_size += 1;              // Relo entry for exception handler
 663 
 664   // Adjust reloc_size to number of record of relocation info
 665   // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
 666   // a relocation index.
 667   // The CodeBuffer will expand the locs array if this estimate is too low.
 668   reloc_size *= 10 / sizeof(relocInfo);
 669 
 670   _buf_sizes._reloc = reloc_size;
 671   _buf_sizes._code  = code_size;
 672   _buf_sizes._stub  = stub_size;
 673 }
 674 
 675 //------------------------------FillLocArray-----------------------------------
 676 // Create a bit of debug info and append it to the array.  The mapping is from
 677 // Java local or expression stack to constant, register or stack-slot.  For
 678 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
 679 // entry has been taken care of and caller should skip it).
 680 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
 681   // This should never have accepted Bad before
 682   assert(OptoReg::is_valid(regnum), "location must be valid");
 683   return (OptoReg::is_reg(regnum))
 684          ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
 685          : new LocationValue(Location::new_stk_loc(l_type,  ra->reg2offset(regnum)));
 686 }
 687 
 688 
 689 ObjectValue*
 690 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
 691   for (int i = 0; i < objs->length(); i++) {
 692     assert(objs->at(i)->is_object(), "corrupt object cache");
 693     ObjectValue* sv = objs->at(i)->as_ObjectValue();
 694     if (sv->id() == id) {
 695       return sv;
 696     }
 697   }
 698   // Otherwise..
 699   return nullptr;
 700 }
 701 
 702 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
 703                                      ObjectValue* sv ) {
 704   assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition");
 705   objs->append(sv);
 706 }
 707 
 708 
 709 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
 710                             GrowableArray<ScopeValue*> *array,
 711                             GrowableArray<ScopeValue*> *objs ) {
 712   assert( local, "use _top instead of null" );
 713   if (array->length() != idx) {
 714     assert(array->length() == idx + 1, "Unexpected array count");
 715     // Old functionality:
 716     //   return
 717     // New functionality:
 718     //   Assert if the local is not top. In product mode let the new node
 719     //   override the old entry.
 720     assert(local == C->top(), "LocArray collision");
 721     if (local == C->top()) {
 722       return;
 723     }
 724     array->pop();
 725   }
 726   const Type *t = local->bottom_type();
 727 
 728   // Is it a safepoint scalar object node?
 729   if (local->is_SafePointScalarObject()) {
 730     SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
 731 
 732     ObjectValue* sv = sv_for_node_id(objs, spobj->_idx);
 733     if (sv == nullptr) {
 734       ciKlass* cik = t->is_oopptr()->exact_klass();
 735       assert(cik->is_instance_klass() ||
 736              cik->is_array_klass(), "Not supported allocation.");
 737       sv = new ObjectValue(spobj->_idx,
 738                            new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
 739       set_sv_for_object_node(objs, sv);
 740 
 741       uint first_ind = spobj->first_index(sfpt->jvms());
 742       for (uint i = 0; i < spobj->n_fields(); i++) {
 743         Node* fld_node = sfpt->in(first_ind+i);
 744         (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
 745       }
 746     }
 747     array->append(sv);
 748     return;
 749   } else if (local->is_SafePointScalarMerge()) {
 750     SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
 751     ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
 752 
 753     if (mv == nullptr) {
 754       GrowableArray<ScopeValue*> deps;
 755 
 756       int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
 757       (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
 758       assert(deps.length() == 1, "missing value");
 759 
 760       int selector_idx = smerge->selector_idx(sfpt->jvms());
 761       (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
 762       assert(deps.length() == 2, "missing value");
 763 
 764       mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
 765       set_sv_for_object_node(objs, mv);
 766 
 767       for (uint i = 1; i < smerge->req(); i++) {
 768         Node* obj_node = smerge->in(i);
 769         int idx = mv->possible_objects()->length();
 770         (void)FillLocArray(idx, sfpt, obj_node, mv->possible_objects(), objs);
 771 
 772         // By default ObjectValues that are in 'possible_objects' are not root objects.
 773         // They will be marked as root later if they are directly referenced in a JVMS.
 774         assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
 775         assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
 776         mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
 777       }
 778     }
 779     array->append(mv);
 780     return;
 781   }
 782 
 783   // Grab the register number for the local
 784   OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
 785   if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
 786     // Record the double as two float registers.
 787     // The register mask for such a value always specifies two adjacent
 788     // float registers, with the lower register number even.
 789     // Normally, the allocation of high and low words to these registers
 790     // is irrelevant, because nearly all operations on register pairs
 791     // (e.g., StoreD) treat them as a single unit.
 792     // Here, we assume in addition that the words in these two registers
 793     // stored "naturally" (by operations like StoreD and double stores
 794     // within the interpreter) such that the lower-numbered register
 795     // is written to the lower memory address.  This may seem like
 796     // a machine dependency, but it is not--it is a requirement on
 797     // the author of the <arch>.ad file to ensure that, for every
 798     // even/odd double-register pair to which a double may be allocated,
 799     // the word in the even single-register is stored to the first
 800     // memory word.  (Note that register numbers are completely
 801     // arbitrary, and are not tied to any machine-level encodings.)
 802 #ifdef _LP64
 803     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
 804       array->append(new ConstantIntValue((jint)0));
 805       array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
 806     } else if ( t->base() == Type::Long ) {
 807       array->append(new ConstantIntValue((jint)0));
 808       array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
 809     } else if ( t->base() == Type::RawPtr ) {
 810       // jsr/ret return address which must be restored into the full
 811       // width 64-bit stack slot.
 812       array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
 813     }
 814 #else //_LP64
 815     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
 816       // Repack the double/long as two jints.
 817       // The convention the interpreter uses is that the second local
 818       // holds the first raw word of the native double representation.
 819       // This is actually reasonable, since locals and stack arrays
 820       // grow downwards in all implementations.
 821       // (If, on some machine, the interpreter's Java locals or stack
 822       // were to grow upwards, the embedded doubles would be word-swapped.)
 823       array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
 824       array->append(new_loc_value( C->regalloc(),              regnum   , Location::normal ));
 825     }
 826 #endif //_LP64
 827     else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
 828              OptoReg::is_reg(regnum) ) {
 829       array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
 830                                                       ? Location::float_in_dbl : Location::normal ));
 831     } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
 832       array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
 833                                                       ? Location::int_in_long : Location::normal ));
 834     } else if( t->base() == Type::NarrowOop ) {
 835       array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
 836     } else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
 837                t->base() == Type::VectorD || t->base() == Type::VectorX ||
 838                t->base() == Type::VectorY || t->base() == Type::VectorZ) {
 839       array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
 840     } else if (C->regalloc()->is_oop(local)) {
 841       assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
 842              t->base() == Type::AryPtr,
 843              "Unexpected type: %s", t->msg());
 844       array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
 845     } else {
 846       assert(t->base() == Type::Int || t->base() == Type::Half ||
 847              t->base() == Type::FloatCon || t->base() == Type::FloatBot,
 848              "Unexpected type: %s", t->msg());
 849       array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
 850     }
 851     return;
 852   }
 853 
 854   // No register.  It must be constant data.
 855   switch (t->base()) {
 856     case Type::Half:              // Second half of a double
 857       ShouldNotReachHere();       // Caller should skip 2nd halves
 858       break;
 859     case Type::AnyPtr:
 860       array->append(new ConstantOopWriteValue(nullptr));
 861       break;
 862     case Type::AryPtr:
 863     case Type::InstPtr:          // fall through
 864       array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
 865       break;
 866     case Type::NarrowOop:
 867       if (t == TypeNarrowOop::NULL_PTR) {
 868         array->append(new ConstantOopWriteValue(nullptr));
 869       } else {
 870         array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
 871       }
 872       break;
 873     case Type::Int:
 874       array->append(new ConstantIntValue(t->is_int()->get_con()));
 875       break;
 876     case Type::RawPtr:
 877       // A return address (T_ADDRESS).
 878       assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
 879 #ifdef _LP64
 880       // Must be restored to the full-width 64-bit stack slot.
 881       array->append(new ConstantLongValue(t->is_ptr()->get_con()));
 882 #else
 883       array->append(new ConstantIntValue(t->is_ptr()->get_con()));
 884 #endif
 885       break;
 886     case Type::FloatCon: {
 887       float f = t->is_float_constant()->getf();
 888       array->append(new ConstantIntValue(jint_cast(f)));
 889       break;
 890     }
 891     case Type::DoubleCon: {
 892       jdouble d = t->is_double_constant()->getd();
 893 #ifdef _LP64
 894       array->append(new ConstantIntValue((jint)0));
 895       array->append(new ConstantDoubleValue(d));
 896 #else
 897       // Repack the double as two jints.
 898     // The convention the interpreter uses is that the second local
 899     // holds the first raw word of the native double representation.
 900     // This is actually reasonable, since locals and stack arrays
 901     // grow downwards in all implementations.
 902     // (If, on some machine, the interpreter's Java locals or stack
 903     // were to grow upwards, the embedded doubles would be word-swapped.)
 904     jlong_accessor acc;
 905     acc.long_value = jlong_cast(d);
 906     array->append(new ConstantIntValue(acc.words[1]));
 907     array->append(new ConstantIntValue(acc.words[0]));
 908 #endif
 909       break;
 910     }
 911     case Type::Long: {
 912       jlong d = t->is_long()->get_con();
 913 #ifdef _LP64
 914       array->append(new ConstantIntValue((jint)0));
 915       array->append(new ConstantLongValue(d));
 916 #else
 917       // Repack the long as two jints.
 918     // The convention the interpreter uses is that the second local
 919     // holds the first raw word of the native double representation.
 920     // This is actually reasonable, since locals and stack arrays
 921     // grow downwards in all implementations.
 922     // (If, on some machine, the interpreter's Java locals or stack
 923     // were to grow upwards, the embedded doubles would be word-swapped.)
 924     jlong_accessor acc;
 925     acc.long_value = d;
 926     array->append(new ConstantIntValue(acc.words[1]));
 927     array->append(new ConstantIntValue(acc.words[0]));
 928 #endif
 929       break;
 930     }
 931     case Type::Top:               // Add an illegal value here
 932       array->append(new LocationValue(Location()));
 933       break;
 934     default:
 935       ShouldNotReachHere();
 936       break;
 937   }
 938 }
 939 
 940 // Determine if this node starts a bundle
 941 bool PhaseOutput::starts_bundle(const Node *n) const {
 942   return (_node_bundling_limit > n->_idx &&
 943           _node_bundling_base[n->_idx].starts_bundle());
 944 }
 945 
 946 // Determine if there is a monitor that has 'ov' as its owner.
 947 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const {
 948   for (int k = 0; k < monarray->length(); k++) {
 949     MonitorValue* mv = monarray->at(k);
 950     if (mv->owner() == ov) {
 951       return true;
 952     }
 953   }
 954 
 955   return false;
 956 }
 957 
 958 // Determine if there is a scalar replaced object description represented by 'ov'.
 959 bool PhaseOutput::contains_as_scalarized_obj(JVMState* jvms, MachSafePointNode* sfn,
 960                                              GrowableArray<ScopeValue*>* objs,
 961                                              ObjectValue* ov) const {
 962   for (int i = 0; i < jvms->scl_size(); i++) {
 963     Node* n = sfn->scalarized_obj(jvms, i);
 964     // Other kinds of nodes that we may encounter here, for instance constants
 965     // representing values of fields of objects scalarized, aren't relevant for
 966     // us, since they don't map to ObjectValue.
 967     if (!n->is_SafePointScalarObject()) {
 968       continue;
 969     }
 970 
 971     ObjectValue* other = sv_for_node_id(objs, n->_idx);
 972     if (ov == other) {
 973       return true;
 974     }
 975   }
 976   return false;
 977 }
 978 
 979 //--------------------------Process_OopMap_Node--------------------------------
 980 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
 981   // Handle special safepoint nodes for synchronization
 982   MachSafePointNode *sfn   = mach->as_MachSafePoint();
 983   MachCallNode      *mcall;
 984 
 985   int safepoint_pc_offset = current_offset;
 986   bool return_oop = false;
 987   bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
 988   bool arg_escape = false;
 989 
 990   // Add the safepoint in the DebugInfoRecorder
 991   if( !mach->is_MachCall() ) {
 992     mcall = nullptr;
 993     C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
 994   } else {
 995     mcall = mach->as_MachCall();
 996 
 997     if (mcall->is_MachCallJava()) {
 998       arg_escape = mcall->as_MachCallJava()->_arg_escape;
 999     }
1000 
1001     // Check if a call returns an object.
1002     if (mcall->returns_pointer()) {
1003       return_oop = true;
1004     }
1005     safepoint_pc_offset += mcall->ret_addr_offset();
1006     C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1007   }
1008 
1009   // Loop over the JVMState list to add scope information
1010   // Do not skip safepoints with a null method, they need monitor info
1011   JVMState* youngest_jvms = sfn->jvms();
1012   int max_depth = youngest_jvms->depth();
1013 
1014   // Allocate the object pool for scalar-replaced objects -- the map from
1015   // small-integer keys (which can be recorded in the local and ostack
1016   // arrays) to descriptions of the object state.
1017   GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1018 
1019   // Visit scopes from oldest to youngest.
1020   for (int depth = 1; depth <= max_depth; depth++) {
1021     JVMState* jvms = youngest_jvms->of_depth(depth);
1022     int idx;
1023     ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1024     // Safepoints that do not have method() set only provide oop-map and monitor info
1025     // to support GC; these do not support deoptimization.
1026     int num_locs = (method == nullptr) ? 0 : jvms->loc_size();
1027     int num_exps = (method == nullptr) ? 0 : jvms->stk_size();
1028     int num_mon  = jvms->nof_monitors();
1029     assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(),
1030            "JVMS local count must match that of the method");
1031 
1032     // Add Local and Expression Stack Information
1033 
1034     // Insert locals into the locarray
1035     GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1036     for( idx = 0; idx < num_locs; idx++ ) {
1037       FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1038     }
1039 
1040     // Insert expression stack entries into the exparray
1041     GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1042     for( idx = 0; idx < num_exps; idx++ ) {
1043       FillLocArray( idx,  sfn, sfn->stack(jvms, idx), exparray, objs );
1044     }
1045 
1046     // Add in mappings of the monitors
1047     assert( !method ||
1048             !method->is_synchronized() ||
1049             method->is_native() ||
1050             num_mon > 0,
1051             "monitors must always exist for synchronized methods");
1052 
1053     // Build the growable array of ScopeValues for exp stack
1054     GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1055 
1056     // Loop over monitors and insert into array
1057     for (idx = 0; idx < num_mon; idx++) {
1058       // Grab the node that defines this monitor
1059       Node* box_node = sfn->monitor_box(jvms, idx);
1060       Node* obj_node = sfn->monitor_obj(jvms, idx);
1061 
1062       // Create ScopeValue for object
1063       ScopeValue *scval = nullptr;
1064 
1065       if (obj_node->is_SafePointScalarObject()) {
1066         SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1067         scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1068         if (scval == nullptr) {
1069           const Type *t = spobj->bottom_type();
1070           ciKlass* cik = t->is_oopptr()->exact_klass();
1071           assert(cik->is_instance_klass() ||
1072                  cik->is_array_klass(), "Not supported allocation.");
1073           ObjectValue* sv = new ObjectValue(spobj->_idx,
1074                                             new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1075           PhaseOutput::set_sv_for_object_node(objs, sv);
1076 
1077           uint first_ind = spobj->first_index(youngest_jvms);
1078           for (uint i = 0; i < spobj->n_fields(); i++) {
1079             Node* fld_node = sfn->in(first_ind+i);
1080             (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1081           }
1082           scval = sv;
1083         }
1084       } else if (obj_node->is_SafePointScalarMerge()) {
1085         SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1086         ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1087 
1088         if (mv == nullptr) {
1089           GrowableArray<ScopeValue*> deps;
1090 
1091           int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1092           FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1093           assert(deps.length() == 1, "missing value");
1094 
1095           int selector_idx = smerge->selector_idx(youngest_jvms);
1096           FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr);
1097           assert(deps.length() == 2, "missing value");
1098 
1099           mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
1100           set_sv_for_object_node(objs, mv);
1101 
1102           for (uint i = 1; i < smerge->req(); i++) {
1103             Node* obj_node = smerge->in(i);
1104             int idx = mv->possible_objects()->length();
1105             (void)FillLocArray(idx, sfn, obj_node, mv->possible_objects(), objs);
1106 
1107             // By default ObjectValues that are in 'possible_objects' are not root objects.
1108             // They will be marked as root later if they are directly referenced in a JVMS.
1109             assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
1110             assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
1111             mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
1112           }
1113         }
1114         scval = mv;
1115       } else if (!obj_node->is_Con()) {
1116         OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1117         if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1118           scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1119         } else {
1120           scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1121         }
1122       } else {
1123         const TypePtr *tp = obj_node->get_ptr_type();
1124         scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1125       }
1126 
1127       OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1128       Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1129       bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1130       monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1131     }
1132 
1133     // Mark ObjectValue nodes as root nodes if they are directly
1134     // referenced in the JVMS.
1135     for (int i = 0; i < objs->length(); i++) {
1136       ScopeValue* sv = objs->at(i);
1137       if (sv->is_object_merge()) {
1138         ObjectMergeValue* merge = sv->as_ObjectMergeValue();
1139 
1140         for (int j = 0; j< merge->possible_objects()->length(); j++) {
1141           ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue();
1142           if (ov->is_root()) {
1143             // Already flagged as 'root' by something else. We shouldn't change it
1144             // to non-root in a younger JVMS because it may need to be alive in
1145             // a younger JVMS.
1146           } else {
1147             bool is_root = locarray->contains(ov) ||
1148                            exparray->contains(ov) ||
1149                            contains_as_owner(monarray, ov) ||
1150                            contains_as_scalarized_obj(jvms, sfn, objs, ov);
1151             ov->set_root(is_root);
1152           }
1153         }
1154       }
1155     }
1156 
1157     // We dump the object pool first, since deoptimization reads it in first.
1158     C->debug_info()->dump_object_pool(objs);
1159 
1160     // Build first class objects to pass to scope
1161     DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1162     DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1163     DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1164 
1165     // Make method available for all Safepoints
1166     ciMethod* scope_method = method ? method : C->method();
1167     // Describe the scope here
1168     assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1169     assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1170     // Now we can describe the scope.
1171     methodHandle null_mh;
1172     bool rethrow_exception = false;
1173     C->debug_info()->describe_scope(
1174       safepoint_pc_offset,
1175       null_mh,
1176       scope_method,
1177       jvms->bci(),
1178       jvms->should_reexecute(),
1179       rethrow_exception,
1180       return_oop,
1181       has_ea_local_in_scope,
1182       arg_escape,
1183       locvals,
1184       expvals,
1185       monvals
1186     );
1187   } // End jvms loop
1188 
1189   // Mark the end of the scope set.
1190   C->debug_info()->end_safepoint(safepoint_pc_offset);
1191 }
1192 
1193 
1194 
1195 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1196 class NonSafepointEmitter {
1197     Compile*  C;
1198     JVMState* _pending_jvms;
1199     int       _pending_offset;
1200 
1201     void emit_non_safepoint();
1202 
1203  public:
1204     NonSafepointEmitter(Compile* compile) {
1205       this->C = compile;
1206       _pending_jvms = nullptr;
1207       _pending_offset = 0;
1208     }
1209 
1210     void observe_instruction(Node* n, int pc_offset) {
1211       if (!C->debug_info()->recording_non_safepoints())  return;
1212 
1213       Node_Notes* nn = C->node_notes_at(n->_idx);
1214       if (nn == nullptr || nn->jvms() == nullptr)  return;
1215       if (_pending_jvms != nullptr &&
1216           _pending_jvms->same_calls_as(nn->jvms())) {
1217         // Repeated JVMS?  Stretch it up here.
1218         _pending_offset = pc_offset;
1219       } else {
1220         if (_pending_jvms != nullptr &&
1221             _pending_offset < pc_offset) {
1222           emit_non_safepoint();
1223         }
1224         _pending_jvms = nullptr;
1225         if (pc_offset > C->debug_info()->last_pc_offset()) {
1226           // This is the only way _pending_jvms can become non-null:
1227           _pending_jvms = nn->jvms();
1228           _pending_offset = pc_offset;
1229         }
1230       }
1231     }
1232 
1233     // Stay out of the way of real safepoints:
1234     void observe_safepoint(JVMState* jvms, int pc_offset) {
1235       if (_pending_jvms != nullptr &&
1236           !_pending_jvms->same_calls_as(jvms) &&
1237           _pending_offset < pc_offset) {
1238         emit_non_safepoint();
1239       }
1240       _pending_jvms = nullptr;
1241     }
1242 
1243     void flush_at_end() {
1244       if (_pending_jvms != nullptr) {
1245         emit_non_safepoint();
1246       }
1247       _pending_jvms = nullptr;
1248     }
1249 };
1250 
1251 void NonSafepointEmitter::emit_non_safepoint() {
1252   JVMState* youngest_jvms = _pending_jvms;
1253   int       pc_offset     = _pending_offset;
1254 
1255   // Clear it now:
1256   _pending_jvms = nullptr;
1257 
1258   DebugInformationRecorder* debug_info = C->debug_info();
1259   assert(debug_info->recording_non_safepoints(), "sanity");
1260 
1261   debug_info->add_non_safepoint(pc_offset);
1262   int max_depth = youngest_jvms->depth();
1263 
1264   // Visit scopes from oldest to youngest.
1265   for (int depth = 1; depth <= max_depth; depth++) {
1266     JVMState* jvms = youngest_jvms->of_depth(depth);
1267     ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1268     assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1269     methodHandle null_mh;
1270     debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1271   }
1272 
1273   // Mark the end of the scope set.
1274   debug_info->end_non_safepoint(pc_offset);
1275 }
1276 
1277 //------------------------------init_buffer------------------------------------
1278 void PhaseOutput::estimate_buffer_size(int& const_req) {
1279 
1280   // Set the initially allocated size
1281   const_req = initial_const_capacity;
1282 
1283   // The extra spacing after the code is necessary on some platforms.
1284   // Sometimes we need to patch in a jump after the last instruction,
1285   // if the nmethod has been deoptimized.  (See 4932387, 4894843.)
1286 
1287   // Compute the byte offset where we can store the deopt pc.
1288   if (C->fixed_slots() != 0) {
1289     _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1290   }
1291 
1292   // Compute prolog code size
1293   _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1294   assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1295 
1296   if (C->has_mach_constant_base_node()) {
1297     uint add_size = 0;
1298     // Fill the constant table.
1299     // Note:  This must happen before shorten_branches.
1300     for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1301       Block* b = C->cfg()->get_block(i);
1302 
1303       for (uint j = 0; j < b->number_of_nodes(); j++) {
1304         Node* n = b->get_node(j);
1305 
1306         // If the node is a MachConstantNode evaluate the constant
1307         // value section.
1308         if (n->is_MachConstant()) {
1309           MachConstantNode* machcon = n->as_MachConstant();
1310           machcon->eval_constant(C);
1311         } else if (n->is_Mach()) {
1312           // On Power there are more nodes that issue constants.
1313           add_size += (n->as_Mach()->ins_num_consts() * 8);
1314         }
1315       }
1316     }
1317 
1318     // Calculate the offsets of the constants and the size of the
1319     // constant table (including the padding to the next section).
1320     constant_table().calculate_offsets_and_size();
1321     const_req = constant_table().alignment() + constant_table().size() + add_size;
1322   }
1323 
1324   // Initialize the space for the BufferBlob used to find and verify
1325   // instruction size in MachNode::emit_size()
1326   init_scratch_buffer_blob(const_req);
1327 }
1328 
1329 CodeBuffer* PhaseOutput::init_buffer() {
1330   int stub_req  = _buf_sizes._stub;
1331   int code_req  = _buf_sizes._code;
1332   int const_req = _buf_sizes._const;
1333 
1334   int pad_req   = NativeCall::byte_size();
1335 
1336   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1337   stub_req += bs->estimate_stub_size();
1338 
1339   // nmethod and CodeBuffer count stubs & constants as part of method's code.
1340   // class HandlerImpl is platform-specific and defined in the *.ad files.
1341   int deopt_handler_req     = HandlerImpl::size_deopt_handler()     + MAX_stubs_size; // add marginal slop for handler
1342   stub_req += MAX_stubs_size;   // ensure per-stub margin
1343   code_req += MAX_inst_size;    // ensure per-instruction margin
1344 
1345   if (StressCodeBuffers)
1346     code_req = const_req = stub_req = deopt_handler_req = 0x10;  // force expansion
1347 
1348   int total_req =
1349           const_req +
1350           code_req +
1351           pad_req +
1352           stub_req +
1353           deopt_handler_req;               // deopt handler
1354 
1355   CodeBuffer* cb = code_buffer();
1356   cb->set_const_section_alignment(constant_table().alignment());
1357   cb->initialize(total_req, _buf_sizes._reloc);
1358 
1359   // Have we run out of code space?
1360   if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1361     C->record_failure("CodeCache is full");
1362     return nullptr;
1363   }
1364   // Configure the code buffer.
1365   cb->initialize_consts_size(const_req);
1366   cb->initialize_stubs_size(stub_req);
1367   cb->initialize_oop_recorder(C->env()->oop_recorder());
1368 
1369   return cb;
1370 }
1371 
1372 //------------------------------fill_buffer------------------------------------
1373 void PhaseOutput::fill_buffer(C2_MacroAssembler* masm, uint* blk_starts) {
1374   // blk_starts[] contains offsets calculated during short branches processing,
1375   // offsets should not be increased during following steps.
1376 
1377   // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1378   // of a loop. It is used to determine the padding for loop alignment.
1379   Compile::TracePhase tp(_t_fillBuffer);
1380 
1381   compute_loop_first_inst_sizes();
1382 
1383   // Create oopmap set.
1384   _oop_map_set = new OopMapSet();
1385 
1386   // !!!!! This preserves old handling of oopmaps for now
1387   C->debug_info()->set_oopmaps(_oop_map_set);
1388 
1389   uint nblocks  = C->cfg()->number_of_blocks();
1390   // Count and start of implicit null check instructions
1391   uint inct_cnt = 0;
1392   uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1393 
1394   // Count and start of calls
1395   uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1396 
1397   uint  return_offset = 0;
1398   int nop_size = (new MachNopNode())->size(C->regalloc());
1399 
1400   int previous_offset = 0;
1401   int current_offset  = 0;
1402   int last_call_offset = -1;
1403   int last_avoid_back_to_back_offset = -1;
1404 #ifdef ASSERT
1405   uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1406   uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1407   uint* jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
1408   uint* jmp_rule   = NEW_RESOURCE_ARRAY(uint,nblocks);
1409 #endif
1410 
1411   // Create an array of unused labels, one for each basic block, if printing is enabled
1412 #if defined(SUPPORT_OPTO_ASSEMBLY)
1413   int* node_offsets      = nullptr;
1414   uint node_offset_limit = C->unique();
1415 
1416   if (C->print_assembly()) {
1417     node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1418   }
1419   if (node_offsets != nullptr) {
1420     // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1421     memset(node_offsets, 0, node_offset_limit*sizeof(int));
1422   }
1423 #endif
1424 
1425   NonSafepointEmitter non_safepoints(C);  // emit non-safepoints lazily
1426 
1427   // Emit the constant table.
1428   if (C->has_mach_constant_base_node()) {
1429     if (!constant_table().emit(masm)) {
1430       C->record_failure("consts section overflow");
1431       return;
1432     }
1433   }
1434 
1435   // Create an array of labels, one for each basic block
1436   Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1437   for (uint i = 0; i <= nblocks; i++) {
1438     blk_labels[i].init();
1439   }
1440 
1441   // Now fill in the code buffer
1442   for (uint i = 0; i < nblocks; i++) {
1443     Block* block = C->cfg()->get_block(i);
1444     _block = block;
1445     Node* head = block->head();
1446 
1447     // If this block needs to start aligned (i.e, can be reached other
1448     // than by falling-thru from the previous block), then force the
1449     // start of a new bundle.
1450     if (Pipeline::requires_bundling() && starts_bundle(head)) {
1451       masm->code()->flush_bundle(true);
1452     }
1453 
1454 #ifdef ASSERT
1455     if (!block->is_connector()) {
1456       stringStream st;
1457       block->dump_head(C->cfg(), &st);
1458       masm->block_comment(st.freeze());
1459     }
1460     jmp_target[i] = 0;
1461     jmp_offset[i] = 0;
1462     jmp_size[i]   = 0;
1463     jmp_rule[i]   = 0;
1464 #endif
1465     int blk_offset = current_offset;
1466 
1467     // Define the label at the beginning of the basic block
1468     masm->bind(blk_labels[block->_pre_order]);
1469 
1470     uint last_inst = block->number_of_nodes();
1471 
1472     // Emit block normally, except for last instruction.
1473     // Emit means "dump code bits into code buffer".
1474     for (uint j = 0; j<last_inst; j++) {
1475       _index = j;
1476 
1477       // Get the node
1478       Node* n = block->get_node(j);
1479 
1480       // If this starts a new instruction group, then flush the current one
1481       // (but allow split bundles)
1482       if (Pipeline::requires_bundling() && starts_bundle(n))
1483         masm->code()->flush_bundle(false);
1484 
1485       // Special handling for SafePoint/Call Nodes
1486       bool is_mcall = false;
1487       if (n->is_Mach()) {
1488         MachNode *mach = n->as_Mach();
1489         is_mcall = n->is_MachCall();
1490         bool is_sfn = n->is_MachSafePoint();
1491 
1492         // If this requires all previous instructions be flushed, then do so
1493         if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1494           masm->code()->flush_bundle(true);
1495           current_offset = masm->offset();
1496         }
1497 
1498         // align the instruction if necessary
1499         int padding = mach->compute_padding(current_offset);
1500         // Make sure safepoint node for polling is distinct from a call's
1501         // return by adding a nop if needed.
1502         if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1503           padding = nop_size;
1504         }
1505         if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1506             current_offset == last_avoid_back_to_back_offset) {
1507           // Avoid back to back some instructions.
1508           padding = nop_size;
1509         }
1510 
1511         if (padding > 0) {
1512           assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1513           int nops_cnt = padding / nop_size;
1514           MachNode *nop = new MachNopNode(nops_cnt);
1515           block->insert_node(nop, j++);
1516           last_inst++;
1517           C->cfg()->map_node_to_block(nop, block);
1518           // Ensure enough space.
1519           masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1520           if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1521             C->record_failure("CodeCache is full");
1522             return;
1523           }
1524           nop->emit(masm, C->regalloc());
1525           masm->code()->flush_bundle(true);
1526           current_offset = masm->offset();
1527         }
1528 
1529         bool observe_safepoint = is_sfn;
1530         // Remember the start of the last call in a basic block
1531         if (is_mcall) {
1532           MachCallNode *mcall = mach->as_MachCall();
1533 
1534           // This destination address is NOT PC-relative
1535           mcall->method_set((intptr_t)mcall->entry_point());
1536 
1537           // Save the return address
1538           call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1539 
1540           observe_safepoint = mcall->guaranteed_safepoint();
1541         }
1542 
1543         // sfn will be valid whenever mcall is valid now because of inheritance
1544         if (observe_safepoint) {
1545           // Handle special safepoint nodes for synchronization
1546           if (!is_mcall) {
1547             MachSafePointNode *sfn = mach->as_MachSafePoint();
1548             // !!!!! Stubs only need an oopmap right now, so bail out
1549             if (sfn->jvms()->method() == nullptr) {
1550               // Write the oopmap directly to the code blob??!!
1551               continue;
1552             }
1553           } // End synchronization
1554 
1555           non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1556                                            current_offset);
1557           Process_OopMap_Node(mach, current_offset);
1558         } // End if safepoint
1559 
1560           // If this is a null check, then add the start of the previous instruction to the list
1561         else if( mach->is_MachNullCheck() ) {
1562           inct_starts[inct_cnt++] = previous_offset;
1563         }
1564 
1565           // If this is a branch, then fill in the label with the target BB's label
1566         else if (mach->is_MachBranch()) {
1567           // This requires the TRUE branch target be in succs[0]
1568           uint block_num = block->non_connector_successor(0)->_pre_order;
1569 
1570           // Try to replace long branch,
1571           // it is mostly for back branches since forward branch's
1572           // distance is not updated yet.
1573           if (mach->may_be_short_branch()) {
1574             int br_size = n->size(C->regalloc());
1575             int offset = blk_starts[block_num] - current_offset;
1576             if (block_num >= i) {
1577               // Current and following block's offset are not
1578               // finalized yet, adjust distance by the difference
1579               // between calculated and final offsets of current block.
1580               offset -= (blk_starts[i] - blk_offset);
1581             }
1582             // In the following code a nop could be inserted before
1583             // the branch which will increase the backward distance.
1584             bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1585             if (needs_padding && offset <= 0)
1586               offset -= nop_size;
1587 
1588             if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1589               // We've got a winner.  Replace this branch.
1590               MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1591 
1592               // Update the jmp_size.
1593               int new_size = replacement->size(C->regalloc());
1594               assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1595               // Insert padding between avoid_back_to_back branches.
1596               if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1597                 MachNode *nop = new MachNopNode();
1598                 block->insert_node(nop, j++);
1599                 C->cfg()->map_node_to_block(nop, block);
1600                 last_inst++;
1601                 nop->emit(masm, C->regalloc());
1602                 masm->code()->flush_bundle(true);
1603                 current_offset = masm->offset();
1604               }
1605 #ifdef ASSERT
1606               jmp_target[i] = block_num;
1607               jmp_offset[i] = current_offset - blk_offset;
1608               jmp_size[i]   = new_size;
1609               jmp_rule[i]   = mach->rule();
1610 #endif
1611               block->map_node(replacement, j);
1612               mach->subsume_by(replacement, C);
1613               n    = replacement;
1614               mach = replacement;
1615             }
1616           }
1617           mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1618         } else if (mach->ideal_Opcode() == Op_Jump) {
1619           for (uint h = 0; h < block->_num_succs; h++) {
1620             Block* succs_block = block->_succs[h];
1621             for (uint j = 1; j < succs_block->num_preds(); j++) {
1622               Node* jpn = succs_block->pred(j);
1623               if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1624                 uint block_num = succs_block->non_connector()->_pre_order;
1625                 Label *blkLabel = &blk_labels[block_num];
1626                 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1627               }
1628             }
1629           }
1630         } else if (!n->is_Proj()) {
1631           // Remember the beginning of the previous instruction, in case
1632           // it's followed by a flag-kill and a null-check.  Happens on
1633           // Intel all the time, with add-to-memory kind of opcodes.
1634           previous_offset = current_offset;
1635         }
1636 
1637         // Not an else-if!
1638         // If this is a trap based cmp then add its offset to the list.
1639         if (mach->is_TrapBasedCheckNode()) {
1640           inct_starts[inct_cnt++] = current_offset;
1641         }
1642       }
1643 
1644       // Verify that there is sufficient space remaining
1645       masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1646       if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1647         C->record_failure("CodeCache is full");
1648         return;
1649       }
1650 
1651       // Save the offset for the listing
1652 #if defined(SUPPORT_OPTO_ASSEMBLY)
1653       if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1654         node_offsets[n->_idx] = masm->offset();
1655       }
1656 #endif
1657       assert(!C->failing_internal() || C->failure_is_artificial(), "Should not reach here if failing.");
1658 
1659       // "Normal" instruction case
1660       DEBUG_ONLY(uint instr_offset = masm->offset());
1661       n->emit(masm, C->regalloc());
1662       current_offset = masm->offset();
1663 
1664       // Above we only verified that there is enough space in the instruction section.
1665       // However, the instruction may emit stubs that cause code buffer expansion.
1666       // Bail out here if expansion failed due to a lack of code cache space.
1667       if (C->failing()) {
1668         return;
1669       }
1670 
1671       assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1672              "ret_addr_offset() not within emitted code");
1673 
1674 #ifdef ASSERT
1675       uint n_size = n->size(C->regalloc());
1676       if (n_size < (current_offset-instr_offset)) {
1677         MachNode* mach = n->as_Mach();
1678         n->dump();
1679         mach->dump_format(C->regalloc(), tty);
1680         tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1681         Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty);
1682         tty->print_cr(" ------------------- ");
1683         BufferBlob* blob = this->scratch_buffer_blob();
1684         address blob_begin = blob->content_begin();
1685         Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1686         assert(false, "wrong size of mach node");
1687       }
1688 #endif
1689       non_safepoints.observe_instruction(n, current_offset);
1690 
1691       // mcall is last "call" that can be a safepoint
1692       // record it so we can see if a poll will directly follow it
1693       // in which case we'll need a pad to make the PcDesc sites unique
1694       // see  5010568. This can be slightly inaccurate but conservative
1695       // in the case that return address is not actually at current_offset.
1696       // This is a small price to pay.
1697 
1698       if (is_mcall) {
1699         last_call_offset = current_offset;
1700       }
1701 
1702       if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1703         // Avoid back to back some instructions.
1704         last_avoid_back_to_back_offset = current_offset;
1705       }
1706 
1707     } // End for all instructions in block
1708 
1709     // If the next block is the top of a loop, pad this block out to align
1710     // the loop top a little. Helps prevent pipe stalls at loop back branches.
1711     if (i < nblocks-1) {
1712       Block *nb = C->cfg()->get_block(i + 1);
1713       int padding = nb->alignment_padding(current_offset);
1714       if( padding > 0 ) {
1715         MachNode *nop = new MachNopNode(padding / nop_size);
1716         block->insert_node(nop, block->number_of_nodes());
1717         C->cfg()->map_node_to_block(nop, block);
1718         nop->emit(masm, C->regalloc());
1719         current_offset = masm->offset();
1720       }
1721     }
1722     // Verify that the distance for generated before forward
1723     // short branches is still valid.
1724     guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1725 
1726     // Save new block start offset
1727     blk_starts[i] = blk_offset;
1728   } // End of for all blocks
1729   blk_starts[nblocks] = current_offset;
1730 
1731   non_safepoints.flush_at_end();
1732 
1733   // Offset too large?
1734   if (C->failing())  return;
1735 
1736   // Define a pseudo-label at the end of the code
1737   masm->bind( blk_labels[nblocks] );
1738 
1739   // Compute the size of the first block
1740   _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1741 
1742 #ifdef ASSERT
1743   for (uint i = 0; i < nblocks; i++) { // For all blocks
1744     if (jmp_target[i] != 0) {
1745       int br_size = jmp_size[i];
1746       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1747       if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1748         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]);
1749         assert(false, "Displacement too large for short jmp");
1750       }
1751     }
1752   }
1753 #endif
1754 
1755   if (!masm->code()->finalize_stubs()) {
1756     C->record_failure("CodeCache is full");
1757     return;
1758   }
1759 
1760   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1761   bs->emit_stubs(*masm->code());
1762   if (C->failing())  return;
1763 
1764   // Fill in stubs.
1765   assert(masm->inst_mark() == nullptr, "should be.");
1766   _stub_list.emit(*masm);
1767   if (C->failing())  return;
1768 
1769 #ifndef PRODUCT
1770   // Information on the size of the method, without the extraneous code
1771   Scheduling::increment_method_size(masm->offset());
1772 #endif
1773 
1774   // ------------------
1775   // Fill in exception table entries.
1776   FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1777 
1778   // Only java methods have exception handlers and deopt handlers
1779   // class HandlerImpl is platform-specific and defined in the *.ad files.
1780   if (C->method()) {
1781     if (C->failing()) {
1782       return; // CodeBuffer::expand failed
1783     }
1784     // Emit the deopt handler code.
1785     _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(masm));
1786   }
1787 
1788   // One last check for failed CodeBuffer::expand:
1789   if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1790     C->record_failure("CodeCache is full");
1791     return;
1792   }
1793 
1794 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1795   if (C->print_assembly()) {
1796     tty->cr();
1797     tty->print_cr("============================= C2-compiled nmethod ==============================");
1798   }
1799 #endif
1800 
1801 #if defined(SUPPORT_OPTO_ASSEMBLY)
1802   // Dump the assembly code, including basic-block numbers
1803   if (C->print_assembly()) {
1804     if (!VMThread::should_terminate()) {
1805       ResourceMark rm;
1806       stringStream method_metadata_str;
1807       if (C->method() != nullptr) {
1808         C->method()->print_metadata(&method_metadata_str);
1809       }
1810       stringStream dump_asm_str;
1811       dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
1812 
1813       // Make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
1814       NoSafepointVerifier nsv;
1815       ttyLocker ttyl;
1816       // This output goes directly to the tty, not the compiler log.
1817       // To enable tools to match it up with the compilation activity,
1818       // be sure to tag this tty output with the compile ID.
1819       if (xtty != nullptr) {
1820         xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
1821                    C->is_osr_compilation() ? " compile_kind='osr'" :
1822                    (C->for_preload() ? " compile_kind='AP'" : ""));
1823       }
1824       const char* is_aot = C->env()->is_aot_compile() ? (C->for_preload() ? "(AP) " : "(A) -") : "-----";
1825       if (C->method() != nullptr) {
1826         tty->print_cr("----------------------- MetaData before Compile_id = %d %s-------------------", C->compile_id(), is_aot);
1827         tty->print_raw(method_metadata_str.freeze());
1828       } else if (C->stub_name() != nullptr) {
1829         tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
1830       }
1831       tty->cr();
1832       tty->print_cr("------------------------ OptoAssembly for Compile_id = %d %s------------------", C->compile_id(), is_aot);
1833       tty->print_raw(dump_asm_str.freeze());
1834       tty->print_cr("--------------------------------------------------------------------------------");
1835       if (xtty != nullptr) {
1836         xtty->tail("opto_assembly");
1837       }
1838     }
1839   }
1840 #endif
1841 }
1842 
1843 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1844   _inc_table.set_size(cnt);
1845 
1846   uint inct_cnt = 0;
1847   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1848     Block* block = C->cfg()->get_block(i);
1849     Node *n = nullptr;
1850     int j;
1851 
1852     // Find the branch; ignore trailing NOPs.
1853     for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1854       n = block->get_node(j);
1855       if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1856         break;
1857       }
1858     }
1859 
1860     // If we didn't find anything, continue
1861     if (j < 0) {
1862       continue;
1863     }
1864 
1865     // Compute ExceptionHandlerTable subtable entry and add it
1866     // (skip empty blocks)
1867     if (n->is_Catch()) {
1868 
1869       // Get the offset of the return from the call
1870       uint call_return = call_returns[block->_pre_order];
1871 #ifdef ASSERT
1872       assert( call_return > 0, "no call seen for this basic block" );
1873       while (block->get_node(--j)->is_MachProj()) ;
1874       assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1875 #endif
1876       // last instruction is a CatchNode, find it's CatchProjNodes
1877       int nof_succs = block->_num_succs;
1878       // allocate space
1879       GrowableArray<intptr_t> handler_bcis(nof_succs);
1880       GrowableArray<intptr_t> handler_pcos(nof_succs);
1881       // iterate through all successors
1882       for (int j = 0; j < nof_succs; j++) {
1883         Block* s = block->_succs[j];
1884         bool found_p = false;
1885         for (uint k = 1; k < s->num_preds(); k++) {
1886           Node* pk = s->pred(k);
1887           if (pk->is_CatchProj() && pk->in(0) == n) {
1888             const CatchProjNode* p = pk->as_CatchProj();
1889             found_p = true;
1890             // add the corresponding handler bci & pco information
1891             if (p->_con != CatchProjNode::fall_through_index) {
1892               // p leads to an exception handler (and is not fall through)
1893               assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
1894               // no duplicates, please
1895               if (!handler_bcis.contains(p->handler_bci())) {
1896                 uint block_num = s->non_connector()->_pre_order;
1897                 handler_bcis.append(p->handler_bci());
1898                 handler_pcos.append(blk_labels[block_num].loc_pos());
1899               }
1900             }
1901           }
1902         }
1903         assert(found_p, "no matching predecessor found");
1904         // Note:  Due to empty block removal, one block may have
1905         // several CatchProj inputs, from the same Catch.
1906       }
1907 
1908       // Set the offset of the return from the call
1909       assert(handler_bcis.find(-1) != -1, "must have default handler");
1910       _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos);
1911       continue;
1912     }
1913 
1914     // Handle implicit null exception table updates
1915     if (n->is_MachNullCheck()) {
1916       MachNode* access = n->in(1)->as_Mach();
1917       assert(access->barrier_data() == 0 ||
1918              access->is_late_expanded_null_check_candidate(),
1919              "Implicit null checks on memory accesses with barriers are only supported on nodes explicitly marked as null-check candidates");
1920       uint block_num = block->non_connector_successor(0)->_pre_order;
1921       _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1922       continue;
1923     }
1924     // Handle implicit exception table updates: trap instructions.
1925     if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
1926       uint block_num = block->non_connector_successor(0)->_pre_order;
1927       _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1928       continue;
1929     }
1930   } // End of for all blocks fill in exception table entries
1931 }
1932 
1933 // Static Variables
1934 #ifndef PRODUCT
1935 uint Scheduling::_total_nop_size = 0;
1936 uint Scheduling::_total_method_size = 0;
1937 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1938 #endif
1939 
1940 // Initializer for class Scheduling
1941 
1942 Scheduling::Scheduling(Arena *arena, Compile &compile)
1943         : _arena(arena),
1944           _cfg(compile.cfg()),
1945           _regalloc(compile.regalloc()),
1946           _scheduled(arena),
1947           _available(arena),
1948           _reg_node(arena),
1949           _pinch_free_list(arena),
1950           _next_node(nullptr),
1951           _bundle_instr_count(0),
1952           _bundle_cycle_number(0),
1953           _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
1954 {
1955   // Save the count
1956   _node_bundling_limit = compile.unique();
1957   uint node_max = _regalloc->node_regs_max_index();
1958 
1959   compile.output()->set_node_bundling_limit(_node_bundling_limit);
1960 
1961   // This one is persistent within the Compile class
1962   _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1963 
1964   // Allocate space for fixed-size arrays
1965   _uses            = NEW_ARENA_ARRAY(arena, short,          node_max);
1966   _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1967 
1968   // Clear the arrays
1969   for (uint i = 0; i < node_max; i++) {
1970     ::new (&_node_bundling_base[i]) Bundle();
1971   }
1972   memset(_uses,               0, node_max * sizeof(short));
1973   memset(_current_latency,    0, node_max * sizeof(unsigned short));
1974 
1975   // Clear the bundling information
1976   memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
1977 
1978   // Get the last node
1979   Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
1980 
1981   _next_node = block->get_node(block->number_of_nodes() - 1);
1982 }
1983 
1984 // Step ahead "i" cycles
1985 void Scheduling::step(uint i) {
1986 
1987   Bundle *bundle = node_bundling(_next_node);
1988   bundle->set_starts_bundle();
1989 
1990   // Update the bundle record, but leave the flags information alone
1991   if (_bundle_instr_count > 0) {
1992     bundle->set_instr_count(_bundle_instr_count);
1993     bundle->set_resources_used(_bundle_use.resourcesUsed());
1994   }
1995 
1996   // Update the state information
1997   _bundle_instr_count = 0;
1998   _bundle_cycle_number += i;
1999   _bundle_use.step(i);
2000 }
2001 
2002 void Scheduling::step_and_clear() {
2003   Bundle *bundle = node_bundling(_next_node);
2004   bundle->set_starts_bundle();
2005 
2006   // Update the bundle record
2007   if (_bundle_instr_count > 0) {
2008     bundle->set_instr_count(_bundle_instr_count);
2009     bundle->set_resources_used(_bundle_use.resourcesUsed());
2010 
2011     _bundle_cycle_number += 1;
2012   }
2013 
2014   // Clear the bundling information
2015   _bundle_instr_count = 0;
2016   _bundle_use.reset();
2017 
2018   memcpy(_bundle_use_elements,
2019          Pipeline_Use::elaborated_elements,
2020          sizeof(Pipeline_Use::elaborated_elements));
2021 }
2022 
2023 // Perform instruction scheduling and bundling over the sequence of
2024 // instructions in backwards order.
2025 void PhaseOutput::ScheduleAndBundle() {
2026 
2027   // Don't optimize this if it isn't a method
2028   if (!C->method())
2029     return;
2030 
2031   // Don't optimize this if scheduling is disabled
2032   if (!C->do_scheduling())
2033     return;
2034 
2035   // Scheduling code works only with pairs (8 bytes) maximum.
2036   // And when the scalable vector register is used, we may spill/unspill
2037   // the whole reg regardless of the max vector size.
2038   if (C->max_vector_size() > 8 ||
2039       (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2040     return;
2041   }
2042 
2043   Compile::TracePhase tp(_t_instrSched);
2044 
2045   // Create a data structure for all the scheduling information
2046   Scheduling scheduling(Thread::current()->resource_area(), *C);
2047 
2048   // Walk backwards over each basic block, computing the needed alignment
2049   // Walk over all the basic blocks
2050   scheduling.DoScheduling();
2051 
2052 #ifndef PRODUCT
2053   if (C->trace_opto_output()) {
2054     // Buffer and print all at once
2055     ResourceMark rm;
2056     stringStream ss;
2057     ss.print("\n---- After ScheduleAndBundle ----\n");
2058     print_scheduling(&ss);
2059     tty->print("%s", ss.as_string());
2060   }
2061 #endif
2062 }
2063 
2064 #ifndef PRODUCT
2065 // Separated out so that it can be called directly from debugger
2066 void PhaseOutput::print_scheduling() {
2067   print_scheduling(tty);
2068 }
2069 
2070 void PhaseOutput::print_scheduling(outputStream* output_stream) {
2071   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2072     output_stream->print("\nBB#%03d:\n", i);
2073     Block* block = C->cfg()->get_block(i);
2074     for (uint j = 0; j < block->number_of_nodes(); j++) {
2075       Node* n = block->get_node(j);
2076       OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2077       output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2078       n->dump("\n", false, output_stream);
2079     }
2080   }
2081 }
2082 #endif
2083 
2084 // See if this node fits into the present instruction bundle
2085 bool Scheduling::NodeFitsInBundle(Node *n) {
2086   uint n_idx = n->_idx;
2087 
2088   // If the node cannot be scheduled this cycle, skip it
2089   if (_current_latency[n_idx] > _bundle_cycle_number) {
2090 #ifndef PRODUCT
2091     if (_cfg->C->trace_opto_output())
2092       tty->print("#     NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2093                  n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2094 #endif
2095     return (false);
2096   }
2097 
2098   const Pipeline *node_pipeline = n->pipeline();
2099 
2100   uint instruction_count = node_pipeline->instructionCount();
2101   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2102     instruction_count = 0;
2103 
2104   if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2105 #ifndef PRODUCT
2106     if (_cfg->C->trace_opto_output())
2107       tty->print("#     NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2108                  n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2109 #endif
2110     return (false);
2111   }
2112 
2113   // Don't allow non-machine nodes to be handled this way
2114   if (!n->is_Mach() && instruction_count == 0)
2115     return (false);
2116 
2117   // See if there is any overlap
2118   uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2119 
2120   if (delay > 0) {
2121 #ifndef PRODUCT
2122     if (_cfg->C->trace_opto_output())
2123       tty->print("#     NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2124 #endif
2125     return false;
2126   }
2127 
2128 #ifndef PRODUCT
2129   if (_cfg->C->trace_opto_output())
2130     tty->print("#     NodeFitsInBundle [%4d]:  TRUE\n", n_idx);
2131 #endif
2132 
2133   return true;
2134 }
2135 
2136 Node * Scheduling::ChooseNodeToBundle() {
2137   uint siz = _available.size();
2138 
2139   if (siz == 0) {
2140 
2141 #ifndef PRODUCT
2142     if (_cfg->C->trace_opto_output())
2143       tty->print("#   ChooseNodeToBundle: null\n");
2144 #endif
2145     return (nullptr);
2146   }
2147 
2148   // Fast path, if only 1 instruction in the bundle
2149   if (siz == 1) {
2150 #ifndef PRODUCT
2151     if (_cfg->C->trace_opto_output()) {
2152       tty->print("#   ChooseNodeToBundle (only 1): ");
2153       _available[0]->dump();
2154     }
2155 #endif
2156     return (_available[0]);
2157   }
2158 
2159   // Don't bother, if the bundle is already full
2160   if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2161     for ( uint i = 0; i < siz; i++ ) {
2162       Node *n = _available[i];
2163 
2164       // Skip projections, we'll handle them another way
2165       if (n->is_Proj())
2166         continue;
2167 
2168       // This presupposed that instructions are inserted into the
2169       // available list in a legality order; i.e. instructions that
2170       // must be inserted first are at the head of the list
2171       if (NodeFitsInBundle(n)) {
2172 #ifndef PRODUCT
2173         if (_cfg->C->trace_opto_output()) {
2174           tty->print("#   ChooseNodeToBundle: ");
2175           n->dump();
2176         }
2177 #endif
2178         return (n);
2179       }
2180     }
2181   }
2182 
2183   // Nothing fits in this bundle, choose the highest priority
2184 #ifndef PRODUCT
2185   if (_cfg->C->trace_opto_output()) {
2186     tty->print("#   ChooseNodeToBundle: ");
2187     _available[0]->dump();
2188   }
2189 #endif
2190 
2191   return _available[0];
2192 }
2193 
2194 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) {
2195   assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), "");
2196 
2197   OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1));
2198   OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first);
2199   OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1));
2200   OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second);
2201 
2202   // Comparison between stack -> reg and stack -> reg
2203   if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) &&
2204       OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) {
2205     return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo);
2206   }
2207 
2208   // Comparison between reg -> stack and reg -> stack
2209   if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) &&
2210       OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) {
2211     return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo);
2212   }
2213 
2214   return 0; // Not comparable
2215 }
2216 
2217 void Scheduling::AddNodeToAvailableList(Node *n) {
2218   assert( !n->is_Proj(), "projections never directly made available" );
2219 #ifndef PRODUCT
2220   if (_cfg->C->trace_opto_output()) {
2221     tty->print("#   AddNodeToAvailableList: ");
2222     n->dump();
2223   }
2224 #endif
2225 
2226   int latency = _current_latency[n->_idx];
2227 
2228   // Insert in latency order (insertion sort). If two MachSpillCopyNodes
2229   // for stack spilling or unspilling have the same latency, we sort
2230   // them in the order of stack offset. Some ports (e.g. aarch64) may also
2231   // have more opportunities to do ld/st merging
2232   uint i;
2233   for (i = 0; i < _available.size(); i++) {
2234     if (_current_latency[_available[i]->_idx] > latency) {
2235       break;
2236     } else if (_current_latency[_available[i]->_idx] == latency &&
2237                n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() &&
2238                compare_two_spill_nodes(n, _available[i]) > 0) {
2239       break;
2240     }
2241   }
2242 
2243   // Special Check for compares following branches
2244   if( n->is_Mach() && _scheduled.size() > 0 ) {
2245     int op = n->as_Mach()->ideal_Opcode();
2246     Node *last = _scheduled[0];
2247     if( last->is_MachIf() && last->in(1) == n &&
2248         ( op == Op_CmpI ||
2249           op == Op_CmpU ||
2250           op == Op_CmpUL ||
2251           op == Op_CmpP ||
2252           op == Op_CmpF ||
2253           op == Op_CmpD ||
2254           op == Op_CmpL ) ) {
2255 
2256       // Recalculate position, moving to front of same latency
2257       for ( i=0 ; i < _available.size(); i++ )
2258         if (_current_latency[_available[i]->_idx] >= latency)
2259           break;
2260     }
2261   }
2262 
2263   // Insert the node in the available list
2264   _available.insert(i, n);
2265 
2266 #ifndef PRODUCT
2267   if (_cfg->C->trace_opto_output())
2268     dump_available();
2269 #endif
2270 }
2271 
2272 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2273   for ( uint i=0; i < n->len(); i++ ) {
2274     Node *def = n->in(i);
2275     if (!def) continue;
2276     if( def->is_Proj() )        // If this is a machine projection, then
2277       def = def->in(0);         // propagate usage thru to the base instruction
2278 
2279     if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2280       continue;
2281     }
2282 
2283     // Compute the latency
2284     uint l = _bundle_cycle_number + n->latency(i);
2285     if (_current_latency[def->_idx] < l)
2286       _current_latency[def->_idx] = l;
2287 
2288     // If this does not have uses then schedule it
2289     if ((--_uses[def->_idx]) == 0)
2290       AddNodeToAvailableList(def);
2291   }
2292 }
2293 
2294 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2295 #ifndef PRODUCT
2296   if (_cfg->C->trace_opto_output()) {
2297     tty->print("#   AddNodeToBundle: ");
2298     n->dump();
2299   }
2300 #endif
2301 
2302   // Remove this from the available list
2303   uint i;
2304   for (i = 0; i < _available.size(); i++)
2305     if (_available[i] == n)
2306       break;
2307   assert(i < _available.size(), "entry in _available list not found");
2308   _available.remove(i);
2309 
2310   // See if this fits in the current bundle
2311   const Pipeline *node_pipeline = n->pipeline();
2312   const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2313 
2314 
2315   // Get the number of instructions
2316   uint instruction_count = node_pipeline->instructionCount();
2317   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2318     instruction_count = 0;
2319 
2320   // Compute the latency information
2321   uint delay = 0;
2322 
2323   if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2324     int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2325     if (relative_latency < 0)
2326       relative_latency = 0;
2327 
2328     delay = _bundle_use.full_latency(relative_latency, node_usage);
2329 
2330     // Does not fit in this bundle, start a new one
2331     if (delay > 0) {
2332       step(delay);
2333 
2334 #ifndef PRODUCT
2335       if (_cfg->C->trace_opto_output())
2336         tty->print("#  *** STEP(%d) ***\n", delay);
2337 #endif
2338     }
2339   }
2340 
2341   if (delay == 0) {
2342     if (node_pipeline->hasMultipleBundles()) {
2343 #ifndef PRODUCT
2344       if (_cfg->C->trace_opto_output())
2345         tty->print("#  *** STEP(multiple instructions) ***\n");
2346 #endif
2347       step(1);
2348     }
2349 
2350     else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2351 #ifndef PRODUCT
2352       if (_cfg->C->trace_opto_output())
2353         tty->print("#  *** STEP(%d >= %d instructions) ***\n",
2354                    instruction_count + _bundle_instr_count,
2355                    Pipeline::_max_instrs_per_cycle);
2356 #endif
2357       step(1);
2358     }
2359   }
2360 
2361   // Set the node's latency
2362   _current_latency[n->_idx] = _bundle_cycle_number;
2363 
2364   // Now merge the functional unit information
2365   if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2366     _bundle_use.add_usage(node_usage);
2367 
2368   // Increment the number of instructions in this bundle
2369   _bundle_instr_count += instruction_count;
2370 
2371   // Remember this node for later
2372   if (n->is_Mach())
2373     _next_node = n;
2374 
2375   // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2376   // not in the bb->_nodes array.  This happens for debug-info-only BoxLocks.
2377   // 'Schedule' them (basically ignore in the schedule) but do not insert them
2378   // into the block.  All other scheduled nodes get put in the schedule here.
2379   int op = n->Opcode();
2380   if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2381       (op != Op_Node &&         // Not an unused antidepedence node and
2382        // not an unallocated boxlock
2383        (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2384 
2385     // Push any trailing projections
2386     if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2387       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2388         Node *foi = n->fast_out(i);
2389         if( foi->is_Proj() )
2390           _scheduled.push(foi);
2391       }
2392     }
2393 
2394     // Put the instruction in the schedule list
2395     _scheduled.push(n);
2396   }
2397 
2398 #ifndef PRODUCT
2399   if (_cfg->C->trace_opto_output())
2400     dump_available();
2401 #endif
2402 
2403   // Walk all the definitions, decrementing use counts, and
2404   // if a definition has a 0 use count, place it in the available list.
2405   DecrementUseCounts(n,bb);
2406 }
2407 
2408 // This method sets the use count within a basic block.  We will ignore all
2409 // uses outside the current basic block.  As we are doing a backwards walk,
2410 // any node we reach that has a use count of 0 may be scheduled.  This also
2411 // avoids the problem of cyclic references from phi nodes, as long as phi
2412 // nodes are at the front of the basic block.  This method also initializes
2413 // the available list to the set of instructions that have no uses within this
2414 // basic block.
2415 void Scheduling::ComputeUseCount(const Block *bb) {
2416 #ifndef PRODUCT
2417   if (_cfg->C->trace_opto_output())
2418     tty->print("# -> ComputeUseCount\n");
2419 #endif
2420 
2421   // Clear the list of available and scheduled instructions, just in case
2422   _available.clear();
2423   _scheduled.clear();
2424 
2425 #ifdef ASSERT
2426   for( uint i=0; i < bb->number_of_nodes(); i++ )
2427     assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2428 #endif
2429 
2430   // Force the _uses count to never go to zero for unscheduable pieces
2431   // of the block
2432   for( uint k = 0; k < _bb_start; k++ )
2433     _uses[bb->get_node(k)->_idx] = 1;
2434   for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2435     _uses[bb->get_node(l)->_idx] = 1;
2436 
2437   // Iterate backwards over the instructions in the block.  Don't count the
2438   // branch projections at end or the block header instructions.
2439   for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2440     Node *n = bb->get_node(j);
2441     if( n->is_Proj() ) continue; // Projections handled another way
2442 
2443     // Account for all uses
2444     for ( uint k = 0; k < n->len(); k++ ) {
2445       Node *inp = n->in(k);
2446       if (!inp) continue;
2447       assert(inp != n, "no cycles allowed" );
2448       if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2449         if (inp->is_Proj()) { // Skip through Proj's
2450           inp = inp->in(0);
2451         }
2452         ++_uses[inp->_idx];     // Count 1 block-local use
2453       }
2454     }
2455 
2456     // If this instruction has a 0 use count, then it is available
2457     if (!_uses[n->_idx]) {
2458       _current_latency[n->_idx] = _bundle_cycle_number;
2459       AddNodeToAvailableList(n);
2460     }
2461 
2462 #ifndef PRODUCT
2463     if (_cfg->C->trace_opto_output()) {
2464       tty->print("#   uses: %3d: ", _uses[n->_idx]);
2465       n->dump();
2466     }
2467 #endif
2468   }
2469 
2470 #ifndef PRODUCT
2471   if (_cfg->C->trace_opto_output())
2472     tty->print("# <- ComputeUseCount\n");
2473 #endif
2474 }
2475 
2476 // This routine performs scheduling on each basic block in reverse order,
2477 // using instruction latencies and taking into account function unit
2478 // availability.
2479 void Scheduling::DoScheduling() {
2480 #ifndef PRODUCT
2481   if (_cfg->C->trace_opto_output())
2482     tty->print("# -> DoScheduling\n");
2483 #endif
2484 
2485   Block *succ_bb = nullptr;
2486   Block *bb;
2487   Compile* C = Compile::current();
2488 
2489   // Walk over all the basic blocks in reverse order
2490   for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2491     bb = _cfg->get_block(i);
2492 
2493 #ifndef PRODUCT
2494     if (_cfg->C->trace_opto_output()) {
2495       tty->print("#  Schedule BB#%03d (initial)\n", i);
2496       for (uint j = 0; j < bb->number_of_nodes(); j++) {
2497         bb->get_node(j)->dump();
2498       }
2499     }
2500 #endif
2501 
2502     // On the head node, skip processing
2503     if (bb == _cfg->get_root_block()) {
2504       continue;
2505     }
2506 
2507     // Skip empty, connector blocks
2508     if (bb->is_connector())
2509       continue;
2510 
2511     // If the following block is not the sole successor of
2512     // this one, then reset the pipeline information
2513     if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2514 #ifndef PRODUCT
2515       if (_cfg->C->trace_opto_output()) {
2516         tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2517                    _next_node->_idx, _bundle_instr_count);
2518       }
2519 #endif
2520       step_and_clear();
2521     }
2522 
2523     // Leave untouched the starting instruction, any Phis, a CreateEx node
2524     // or Top.  bb->get_node(_bb_start) is the first schedulable instruction.
2525     _bb_end = bb->number_of_nodes()-1;
2526     for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2527       Node *n = bb->get_node(_bb_start);
2528       // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2529       // Also, MachIdealNodes do not get scheduled
2530       if( !n->is_Mach() ) continue;     // Skip non-machine nodes
2531       MachNode *mach = n->as_Mach();
2532       int iop = mach->ideal_Opcode();
2533       if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2534       if( iop == Op_Con ) continue;      // Do not schedule Top
2535       if( iop == Op_Node &&     // Do not schedule PhiNodes, ProjNodes
2536           mach->pipeline() == MachNode::pipeline_class() &&
2537           !n->is_SpillCopy() && !n->is_MachMerge() )  // Breakpoints, Prolog, etc
2538         continue;
2539       break;                    // Funny loop structure to be sure...
2540     }
2541     // Compute last "interesting" instruction in block - last instruction we
2542     // might schedule.  _bb_end points just after last schedulable inst.
2543     Node *last = bb->get_node(_bb_end);
2544     // Ignore trailing NOPs.
2545     while (_bb_end > 0 && last->is_Mach() &&
2546            last->as_Mach()->ideal_Opcode() == Op_Con) {
2547       last = bb->get_node(--_bb_end);
2548     }
2549     assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2550     if( last->is_Catch() ||
2551         (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2552       // There might be a prior call.  Skip it.
2553       while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2554     } else if( last->is_MachNullCheck() ) {
2555       // Backup so the last null-checked memory instruction is
2556       // outside the schedulable range. Skip over the nullcheck,
2557       // projection, and the memory nodes.
2558       Node *mem = last->in(1);
2559       do {
2560         _bb_end--;
2561       } while (mem != bb->get_node(_bb_end));
2562     } else {
2563       // Set _bb_end to point after last schedulable inst.
2564       _bb_end++;
2565     }
2566 
2567     assert( _bb_start <= _bb_end, "inverted block ends" );
2568 
2569     // Compute the register antidependencies for the basic block
2570     ComputeRegisterAntidependencies(bb);
2571     if (C->failing())  return;  // too many D-U pinch points
2572 
2573     // Compute the usage within the block, and set the list of all nodes
2574     // in the block that have no uses within the block.
2575     ComputeUseCount(bb);
2576 
2577     // Schedule the remaining instructions in the block
2578     while ( _available.size() > 0 ) {
2579       Node *n = ChooseNodeToBundle();
2580       guarantee(n != nullptr, "no nodes available");
2581       AddNodeToBundle(n,bb);
2582     }
2583 
2584     assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2585 #ifdef ASSERT
2586     for( uint l = _bb_start; l < _bb_end; l++ ) {
2587       Node *n = bb->get_node(l);
2588       uint m;
2589       for( m = 0; m < _bb_end-_bb_start; m++ )
2590         if( _scheduled[m] == n )
2591           break;
2592       assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2593     }
2594 #endif
2595 
2596     // Now copy the instructions (in reverse order) back to the block
2597     for ( uint k = _bb_start; k < _bb_end; k++ )
2598       bb->map_node(_scheduled[_bb_end-k-1], k);
2599 
2600 #ifndef PRODUCT
2601     if (_cfg->C->trace_opto_output()) {
2602       tty->print("#  Schedule BB#%03d (final)\n", i);
2603       uint current = 0;
2604       for (uint j = 0; j < bb->number_of_nodes(); j++) {
2605         Node *n = bb->get_node(j);
2606         if( valid_bundle_info(n) ) {
2607           Bundle *bundle = node_bundling(n);
2608           if (bundle->instr_count() > 0) {
2609             tty->print("*** Bundle: ");
2610             bundle->dump();
2611           }
2612           n->dump();
2613         }
2614       }
2615     }
2616 #endif
2617 #ifdef ASSERT
2618     verify_good_schedule(bb,"after block local scheduling");
2619 #endif
2620   }
2621 
2622 #ifndef PRODUCT
2623   if (_cfg->C->trace_opto_output())
2624     tty->print("# <- DoScheduling\n");
2625 #endif
2626 
2627   // Record final node-bundling array location
2628   _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2629 
2630 } // end DoScheduling
2631 
2632 // Verify that no live-range used in the block is killed in the block by a
2633 // wrong DEF.  This doesn't verify live-ranges that span blocks.
2634 
2635 // Check for edge existence.  Used to avoid adding redundant precedence edges.
2636 static bool edge_from_to( Node *from, Node *to ) {
2637   for( uint i=0; i<from->len(); i++ )
2638     if( from->in(i) == to )
2639       return true;
2640   return false;
2641 }
2642 
2643 #ifdef ASSERT
2644 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2645   // Check for bad kills
2646   if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2647     Node *prior_use = _reg_node[def];
2648     if( prior_use && !edge_from_to(prior_use,n) ) {
2649       tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2650       n->dump();
2651       tty->print_cr("...");
2652       prior_use->dump();
2653       assert(edge_from_to(prior_use,n), "%s", msg);
2654     }
2655     _reg_node.map(def,nullptr); // Kill live USEs
2656   }
2657 }
2658 
2659 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2660 
2661   // Zap to something reasonable for the verify code
2662   _reg_node.clear();
2663 
2664   // Walk over the block backwards.  Check to make sure each DEF doesn't
2665   // kill a live value (other than the one it's supposed to).  Add each
2666   // USE to the live set.
2667   for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2668     Node *n = b->get_node(i);
2669     int n_op = n->Opcode();
2670     if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2671       // Fat-proj kills a slew of registers
2672       RegMaskIterator rmi(n->out_RegMask());
2673       while (rmi.has_next()) {
2674         OptoReg::Name kill = rmi.next();
2675         verify_do_def(n, kill, msg);
2676       }
2677     } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2678       // Get DEF'd registers the normal way
2679       verify_do_def( n, _regalloc->get_reg_first(n), msg );
2680       verify_do_def( n, _regalloc->get_reg_second(n), msg );
2681     }
2682 
2683     // Now make all USEs live
2684     for( uint i=1; i<n->req(); i++ ) {
2685       Node *def = n->in(i);
2686       assert(def != nullptr, "input edge required");
2687       OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2688       OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2689       if( OptoReg::is_valid(reg_lo) ) {
2690         assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2691         _reg_node.map(reg_lo,n);
2692       }
2693       if( OptoReg::is_valid(reg_hi) ) {
2694         assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2695         _reg_node.map(reg_hi,n);
2696       }
2697     }
2698 
2699   }
2700 
2701   // Zap to something reasonable for the Antidependence code
2702   _reg_node.clear();
2703 }
2704 #endif
2705 
2706 // Conditionally add precedence edges.  Avoid putting edges on Projs.
2707 static void add_prec_edge_from_to( Node *from, Node *to ) {
2708   if( from->is_Proj() ) {       // Put precedence edge on Proj's input
2709     assert( from->req() == 1 && (from->len() == 1 || from->in(1) == nullptr), "no precedence edges on projections" );
2710     from = from->in(0);
2711   }
2712   if( from != to &&             // No cycles (for things like LD L0,[L0+4] )
2713       !edge_from_to( from, to ) ) // Avoid duplicate edge
2714     from->add_prec(to);
2715 }
2716 
2717 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2718   if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2719     return;
2720 
2721   if (OptoReg::is_reg(def_reg)) {
2722     VMReg vmreg = OptoReg::as_VMReg(def_reg);
2723     if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
2724       // This is one of the high slots of a vector register.
2725       // ScheduleAndBundle already checked there are no live wide
2726       // vectors in this method so it can be safely ignored.
2727       return;
2728     }
2729   }
2730 
2731   Node *pinch = _reg_node[def_reg]; // Get pinch point
2732   if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2733       is_def ) {    // Check for a true def (not a kill)
2734     _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2735     return;
2736   }
2737 
2738   Node *kill = def;             // Rename 'def' to more descriptive 'kill'
2739   DEBUG_ONLY( def = (Node*)((intptr_t)0xdeadbeef); )
2740 
2741   // After some number of kills there _may_ be a later def
2742   Node *later_def = nullptr;
2743 
2744   Compile* C = Compile::current();
2745 
2746   // Finding a kill requires a real pinch-point.
2747   // Check for not already having a pinch-point.
2748   // Pinch points are Op_Node's.
2749   if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2750     later_def = pinch;            // Must be def/kill as optimistic pinch-point
2751     if ( _pinch_free_list.size() > 0) {
2752       pinch = _pinch_free_list.pop();
2753     } else {
2754       pinch = new Node(1); // Pinch point to-be
2755     }
2756     if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2757       DEBUG_ONLY( pinch->dump(); );
2758       assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index());
2759       _cfg->C->record_method_not_compilable("too many D-U pinch points");
2760       return;
2761     }
2762     _cfg->map_node_to_block(pinch, b);      // Pretend it's valid in this block (lazy init)
2763     _reg_node.map(def_reg,pinch); // Record pinch-point
2764     //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
2765     if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2766       pinch->init_req(0, C->top());     // set not null for the next call
2767       add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2768       later_def = nullptr;           // and no later def
2769     }
2770     pinch->set_req(0,later_def);  // Hook later def so we can find it
2771   } else {                        // Else have valid pinch point
2772     if( pinch->in(0) )            // If there is a later-def
2773       later_def = pinch->in(0);   // Get it
2774   }
2775 
2776   // Add output-dependence edge from later def to kill
2777   if( later_def )               // If there is some original def
2778     add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2779 
2780   // See if current kill is also a use, and so is forced to be the pinch-point.
2781   if( pinch->Opcode() == Op_Node ) {
2782     Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2783     for( uint i=1; i<uses->req(); i++ ) {
2784       if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2785           _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2786         // Yes, found a use/kill pinch-point
2787         pinch->set_req(0,nullptr);  //
2788         pinch->replace_by(kill); // Move anti-dep edges up
2789         pinch = kill;
2790         _reg_node.map(def_reg,pinch);
2791         return;
2792       }
2793     }
2794   }
2795 
2796   // Add edge from kill to pinch-point
2797   add_prec_edge_from_to(kill,pinch);
2798 }
2799 
2800 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2801   if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2802     return;
2803   Node *pinch = _reg_node[use_reg]; // Get pinch point
2804   // Check for no later def_reg/kill in block
2805   if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b &&
2806       // Use has to be block-local as well
2807       _cfg->get_block_for_node(use) == b) {
2808     if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2809         pinch->req() == 1 ) {   // pinch not yet in block?
2810       pinch->del_req(0);        // yank pointer to later-def, also set flag
2811       // Insert the pinch-point in the block just after the last use
2812       b->insert_node(pinch, b->find_node(use) + 1);
2813       _bb_end++;                // Increase size scheduled region in block
2814     }
2815 
2816     add_prec_edge_from_to(pinch,use);
2817   }
2818 }
2819 
2820 // We insert antidependences between the reads and following write of
2821 // allocated registers to prevent illegal code motion. Hopefully, the
2822 // number of added references should be fairly small, especially as we
2823 // are only adding references within the current basic block.
2824 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2825 
2826 #ifdef ASSERT
2827   verify_good_schedule(b,"before block local scheduling");
2828 #endif
2829 
2830   // A valid schedule, for each register independently, is an endless cycle
2831   // of: a def, then some uses (connected to the def by true dependencies),
2832   // then some kills (defs with no uses), finally the cycle repeats with a new
2833   // def.  The uses are allowed to float relative to each other, as are the
2834   // kills.  No use is allowed to slide past a kill (or def).  This requires
2835   // antidependencies between all uses of a single def and all kills that
2836   // follow, up to the next def.  More edges are redundant, because later defs
2837   // & kills are already serialized with true or antidependencies.  To keep
2838   // the edge count down, we add a 'pinch point' node if there's more than
2839   // one use or more than one kill/def.
2840 
2841   // We add dependencies in one bottom-up pass.
2842 
2843   // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2844 
2845   // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2846   // register.  If not, we record the DEF/KILL in _reg_node, the
2847   // register-to-def mapping.  If there is a prior DEF/KILL, we insert a
2848   // "pinch point", a new Node that's in the graph but not in the block.
2849   // We put edges from the prior and current DEF/KILLs to the pinch point.
2850   // We put the pinch point in _reg_node.  If there's already a pinch point
2851   // we merely add an edge from the current DEF/KILL to the pinch point.
2852 
2853   // After doing the DEF/KILLs, we handle USEs.  For each used register, we
2854   // put an edge from the pinch point to the USE.
2855 
2856   // To be expedient, the _reg_node array is pre-allocated for the whole
2857   // compilation.  _reg_node is lazily initialized; it either contains a null,
2858   // or a valid def/kill/pinch-point, or a leftover node from some prior
2859   // block.  Leftover node from some prior block is treated like a null (no
2860   // prior def, so no anti-dependence needed).  Valid def is distinguished by
2861   // it being in the current block.
2862   bool fat_proj_seen = false;
2863   uint last_safept = _bb_end-1;
2864   Node* end_node         = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr;
2865   Node* last_safept_node = end_node;
2866   for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2867     Node *n = b->get_node(i);
2868     int is_def = n->outcnt();   // def if some uses prior to adding precedence edges
2869     if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2870       // Fat-proj kills a slew of registers
2871       // This can add edges to 'n' and obscure whether or not it was a def,
2872       // hence the is_def flag.
2873       fat_proj_seen = true;
2874       RegMaskIterator rmi(n->out_RegMask());
2875       while (rmi.has_next()) {
2876         OptoReg::Name kill = rmi.next();
2877         anti_do_def(b, n, kill, is_def);
2878       }
2879     } else {
2880       // Get DEF'd registers the normal way
2881       anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2882       anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2883     }
2884 
2885     // Kill projections on a branch should appear to occur on the
2886     // branch, not afterwards, so grab the masks from the projections
2887     // and process them.
2888     if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
2889       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2890         Node* use = n->fast_out(i);
2891         if (use->is_Proj()) {
2892           RegMaskIterator rmi(use->out_RegMask());
2893           while (rmi.has_next()) {
2894             OptoReg::Name kill = rmi.next();
2895             anti_do_def(b, n, kill, false);
2896           }
2897         }
2898       }
2899     }
2900 
2901     // Check each register used by this instruction for a following DEF/KILL
2902     // that must occur afterward and requires an anti-dependence edge.
2903     for( uint j=0; j<n->req(); j++ ) {
2904       Node *def = n->in(j);
2905       if( def ) {
2906         assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2907         anti_do_use( b, n, _regalloc->get_reg_first(def) );
2908         anti_do_use( b, n, _regalloc->get_reg_second(def) );
2909       }
2910     }
2911     // Do not allow defs of new derived values to float above GC
2912     // points unless the base is definitely available at the GC point.
2913 
2914     Node *m = b->get_node(i);
2915 
2916     // Add precedence edge from following safepoint to use of derived pointer
2917     if( last_safept_node != end_node &&
2918         m != last_safept_node) {
2919       for (uint k = 1; k < m->req(); k++) {
2920         const Type *t = m->in(k)->bottom_type();
2921         if( t->isa_oop_ptr() &&
2922             t->is_ptr()->offset() != 0 ) {
2923           last_safept_node->add_prec( m );
2924           break;
2925         }
2926       }
2927     }
2928 
2929     if( n->jvms() ) {           // Precedence edge from derived to safept
2930       // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2931       if( b->get_node(last_safept) != last_safept_node ) {
2932         last_safept = b->find_node(last_safept_node);
2933       }
2934       for( uint j=last_safept; j > i; j-- ) {
2935         Node *mach = b->get_node(j);
2936         if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2937           mach->add_prec( n );
2938       }
2939       last_safept = i;
2940       last_safept_node = m;
2941     }
2942   }
2943 
2944   if (fat_proj_seen) {
2945     // Garbage collect pinch nodes that were not consumed.
2946     // They are usually created by a fat kill MachProj for a call.
2947     garbage_collect_pinch_nodes();
2948   }
2949 }
2950 
2951 // Garbage collect pinch nodes for reuse by other blocks.
2952 //
2953 // The block scheduler's insertion of anti-dependence
2954 // edges creates many pinch nodes when the block contains
2955 // 2 or more Calls.  A pinch node is used to prevent a
2956 // combinatorial explosion of edges.  If a set of kills for a
2957 // register is anti-dependent on a set of uses (or defs), rather
2958 // than adding an edge in the graph between each pair of kill
2959 // and use (or def), a pinch is inserted between them:
2960 //
2961 //            use1   use2  use3
2962 //                \   |   /
2963 //                 \  |  /
2964 //                  pinch
2965 //                 /  |  \
2966 //                /   |   \
2967 //            kill1 kill2 kill3
2968 //
2969 // One pinch node is created per register killed when
2970 // the second call is encountered during a backwards pass
2971 // over the block.  Most of these pinch nodes are never
2972 // wired into the graph because the register is never
2973 // used or def'ed in the block.
2974 //
2975 void Scheduling::garbage_collect_pinch_nodes() {
2976 #ifndef PRODUCT
2977   if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2978 #endif
2979   int trace_cnt = 0;
2980   for (uint k = 0; k < _reg_node.max(); k++) {
2981     Node* pinch = _reg_node[k];
2982     if ((pinch != nullptr) && pinch->Opcode() == Op_Node &&
2983         // no predecence input edges
2984         (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) {
2985       cleanup_pinch(pinch);
2986       _pinch_free_list.push(pinch);
2987       _reg_node.map(k, nullptr);
2988 #ifndef PRODUCT
2989       if (_cfg->C->trace_opto_output()) {
2990         trace_cnt++;
2991         if (trace_cnt > 40) {
2992           tty->print("\n");
2993           trace_cnt = 0;
2994         }
2995         tty->print(" %d", pinch->_idx);
2996       }
2997 #endif
2998     }
2999   }
3000 #ifndef PRODUCT
3001   if (_cfg->C->trace_opto_output()) tty->print("\n");
3002 #endif
3003 }
3004 
3005 // Clean up a pinch node for reuse.
3006 void Scheduling::cleanup_pinch( Node *pinch ) {
3007   assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3008 
3009   for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3010     Node* use = pinch->last_out(i);
3011     uint uses_found = 0;
3012     for (uint j = use->req(); j < use->len(); j++) {
3013       if (use->in(j) == pinch) {
3014         use->rm_prec(j);
3015         uses_found++;
3016       }
3017     }
3018     assert(uses_found > 0, "must be a precedence edge");
3019     i -= uses_found;    // we deleted 1 or more copies of this edge
3020   }
3021   // May have a later_def entry
3022   pinch->set_req(0, nullptr);
3023 }
3024 
3025 #ifndef PRODUCT
3026 
3027 void Scheduling::dump_available() const {
3028   tty->print("#Availist  ");
3029   for (uint i = 0; i < _available.size(); i++)
3030     tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3031   tty->cr();
3032 }
3033 
3034 // Print Scheduling Statistics
3035 void Scheduling::print_statistics() {
3036   // Print the size added by nops for bundling
3037   tty->print("Nops added %d bytes to total of %d bytes",
3038              _total_nop_size, _total_method_size);
3039   if (_total_method_size > 0)
3040     tty->print(", for %.2f%%",
3041                ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3042   tty->print("\n");
3043 
3044   uint total_instructions = 0, total_bundles = 0;
3045 
3046   for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3047     uint bundle_count   = _total_instructions_per_bundle[i];
3048     total_instructions += bundle_count * i;
3049     total_bundles      += bundle_count;
3050   }
3051 
3052   if (total_bundles > 0)
3053     tty->print("Average ILP (excluding nops) is %.2f\n",
3054                ((double)total_instructions) / ((double)total_bundles));
3055 }
3056 #endif
3057 
3058 //-----------------------init_scratch_buffer_blob------------------------------
3059 // Construct a temporary BufferBlob and cache it for this compile.
3060 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3061   // If there is already a scratch buffer blob allocated and the
3062   // constant section is big enough, use it.  Otherwise free the
3063   // current and allocate a new one.
3064   BufferBlob* blob = scratch_buffer_blob();
3065   if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3066     // Use the current blob.
3067   } else {
3068     if (blob != nullptr) {
3069       BufferBlob::free(blob);
3070     }
3071 
3072     ResourceMark rm;
3073     _scratch_const_size = const_size;
3074     int size = C2Compiler::initial_code_buffer_size(const_size);
3075     blob = BufferBlob::create("Compile::scratch_buffer", size);
3076     // Record the buffer blob for next time.
3077     set_scratch_buffer_blob(blob);
3078     // Have we run out of code space?
3079     if (scratch_buffer_blob() == nullptr) {
3080       // Let CompilerBroker disable further compilations.
3081       C->record_failure("Not enough space for scratch buffer in CodeCache");
3082       return;
3083     }
3084   }
3085 
3086   // Initialize the relocation buffers
3087   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3088   set_scratch_locs_memory(locs_buf);
3089 }
3090 
3091 
3092 //-----------------------scratch_emit_size-------------------------------------
3093 // Helper function that computes size by emitting code
3094 uint PhaseOutput::scratch_emit_size(const Node* n) {
3095   // Start scratch_emit_size section.
3096   set_in_scratch_emit_size(true);
3097 
3098   // Emit into a trash buffer and count bytes emitted.
3099   // This is a pretty expensive way to compute a size,
3100   // but it works well enough if seldom used.
3101   // All common fixed-size instructions are given a size
3102   // method by the AD file.
3103   // Note that the scratch buffer blob and locs memory are
3104   // allocated at the beginning of the compile task, and
3105   // may be shared by several calls to scratch_emit_size.
3106   // The allocation of the scratch buffer blob is particularly
3107   // expensive, since it has to grab the code cache lock.
3108   BufferBlob* blob = this->scratch_buffer_blob();
3109   assert(blob != nullptr, "Initialize BufferBlob at start");
3110   assert(blob->size() > MAX_inst_size, "sanity");
3111   relocInfo* locs_buf = scratch_locs_memory();
3112   address blob_begin = blob->content_begin();
3113   address blob_end   = (address)locs_buf;
3114   assert(blob->contains(blob_end), "sanity");
3115   CodeBuffer buf(blob_begin, blob_end - blob_begin);
3116   buf.initialize_consts_size(_scratch_const_size);
3117   buf.initialize_stubs_size(MAX_stubs_size);
3118   assert(locs_buf != nullptr, "sanity");
3119   int lsize = MAX_locs_size / 3;
3120   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3121   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3122   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3123   // Mark as scratch buffer.
3124   buf.consts()->set_scratch_emit();
3125   buf.insts()->set_scratch_emit();
3126   buf.stubs()->set_scratch_emit();
3127 
3128   // Do the emission.
3129 
3130   Label fakeL; // Fake label for branch instructions.
3131   Label*   saveL = nullptr;
3132   uint save_bnum = 0;
3133   bool is_branch = n->is_MachBranch();
3134   C2_MacroAssembler masm(&buf);
3135   masm.bind(fakeL);
3136   if (is_branch) {
3137     n->as_MachBranch()->save_label(&saveL, &save_bnum);
3138     n->as_MachBranch()->label_set(&fakeL, 0);
3139   }
3140   n->emit(&masm, C->regalloc());
3141 
3142   // Emitting into the scratch buffer should not fail
3143   assert(!C->failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3144 
3145   if (is_branch) // Restore label.
3146     n->as_MachBranch()->label_set(saveL, save_bnum);
3147 
3148   // End scratch_emit_size section.
3149   set_in_scratch_emit_size(false);
3150 
3151   return buf.insts_size();
3152 }
3153 
3154 void PhaseOutput::install() {
3155   if (C->should_install_code() && C->stub_function() != nullptr) {
3156     install_stub(C->stub_name());
3157   } else {
3158     install_code(C->method(),
3159                  C->entry_bci(),
3160                  CompilerThread::current()->compiler(),
3161                  C->has_unsafe_access(),
3162                  SharedRuntime::is_wide_vector(C->max_vector_size()));
3163   }
3164 }
3165 
3166 void PhaseOutput::install_code(ciMethod*         target,
3167                                int               entry_bci,
3168                                AbstractCompiler* compiler,
3169                                bool              has_unsafe_access,
3170                                bool              has_wide_vectors) {
3171   // Check if we want to skip execution of all compiled code.
3172   {
3173 #ifndef PRODUCT
3174     if (OptoNoExecute) {
3175       C->record_method_not_compilable("+OptoNoExecute");  // Flag as failed
3176       return;
3177     }
3178 #endif
3179     Compile::TracePhase tp(_t_registerMethod);
3180 
3181     if (C->is_osr_compilation()) {
3182       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3183       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3184     } else {
3185       if (!target->is_static()) {
3186         // The UEP of an nmethod ensures that the VEP is padded. However, the padding of the UEP is placed
3187         // before the inline cache check, so we don't have to execute any nop instructions when dispatching
3188         // through the UEP, yet we can ensure that the VEP is aligned appropriately.
3189         _code_offsets.set_value(CodeOffsets::Entry, _first_block_size - MacroAssembler::ic_check_size());
3190       }
3191       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3192       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3193     }
3194 
3195     C->env()->register_method(target,
3196                                      entry_bci,
3197                                      &_code_offsets,
3198                                      _orig_pc_slot_offset_in_bytes,
3199                                      code_buffer(),
3200                                      frame_size_in_words(),
3201                                      oop_map_set(),
3202                                      &_handler_table,
3203                                      inc_table(),
3204                                      compiler,
3205                                      C->has_clinit_barriers(),
3206                                      C->for_preload(),
3207                                      has_unsafe_access,
3208                                      SharedRuntime::is_wide_vector(C->max_vector_size()),
3209                                      C->has_monitors(),
3210                                      C->has_scoped_access(),
3211                                      0,
3212                                      C->should_install_code());
3213 
3214     if (C->log() != nullptr) { // Print code cache state into compiler log
3215       C->log()->code_cache_state();
3216     }
3217     assert(!C->has_clinit_barriers() || C->for_preload(), "class init barriers should be only in preload code");
3218   }
3219 }
3220 void PhaseOutput::install_stub(const char* stub_name) {
3221   // Entry point will be accessed using stub_entry_point();
3222   if (code_buffer() == nullptr) {
3223     Matcher::soft_match_failure();
3224   } else {
3225     if (PrintAssembly && (WizardMode || Verbose))
3226       tty->print_cr("### Stub::%s", stub_name);
3227 
3228     if (!C->failing()) {
3229       assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3230 
3231       // Make the NMethod
3232       // For now we mark the frame as never safe for profile stackwalking
3233       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3234                                                       code_buffer(),
3235                                                       CodeOffsets::frame_never_safe,
3236                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
3237                                                       frame_size_in_words(),
3238                                                       oop_map_set(),
3239                                                       false,
3240                                                       false);
3241 
3242       if (rs == nullptr) {
3243         C->record_failure("CodeCache is full");
3244       } else {
3245         assert(rs->is_runtime_stub(), "sanity check");
3246         C->set_stub_entry_point(rs->entry_point());
3247         BlobId blob_id = StubInfo::blob(C->stub_id());
3248         AOTCodeCache::store_code_blob(*rs, AOTCodeEntry::C2Blob, blob_id);
3249       }
3250     }
3251   }
3252 }
3253 
3254 // Support for bundling info
3255 Bundle* PhaseOutput::node_bundling(const Node *n) {
3256   assert(valid_bundle_info(n), "oob");
3257   return &_node_bundling_base[n->_idx];
3258 }
3259 
3260 bool PhaseOutput::valid_bundle_info(const Node *n) {
3261   return (_node_bundling_limit > n->_idx);
3262 }
3263 
3264 //------------------------------frame_size_in_words-----------------------------
3265 // frame_slots in units of words
3266 int PhaseOutput::frame_size_in_words() const {
3267   // shift is 0 in LP32 and 1 in LP64
3268   const int shift = (LogBytesPerWord - LogBytesPerInt);
3269   int words = _frame_slots >> shift;
3270   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3271   return words;
3272 }
3273 
3274 // To bang the stack of this compiled method we use the stack size
3275 // that the interpreter would need in case of a deoptimization. This
3276 // removes the need to bang the stack in the deoptimization blob which
3277 // in turn simplifies stack overflow handling.
3278 int PhaseOutput::bang_size_in_bytes() const {
3279   return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3280 }
3281 
3282 //------------------------------dump_asm---------------------------------------
3283 // Dump formatted assembly
3284 #if defined(SUPPORT_OPTO_ASSEMBLY)
3285 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3286 
3287   int pc_digits = 3; // #chars required for pc
3288   int sb_chars  = 3; // #chars for "start bundle" indicator
3289   int tab_size  = 8;
3290   if (pcs != nullptr) {
3291     int max_pc = 0;
3292     for (uint i = 0; i < pc_limit; i++) {
3293       max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3294     }
3295     pc_digits  = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3296   }
3297   int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3298 
3299   bool cut_short = false;
3300   st->print_cr("#");
3301   st->print("#  ");  C->tf()->dump_on(st);  st->cr();
3302   st->print_cr("#");
3303 
3304   // For all blocks
3305   int pc = 0x0;                 // Program counter
3306   char starts_bundle = ' ';
3307   C->regalloc()->dump_frame();
3308 
3309   Node *n = nullptr;
3310   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3311     if (VMThread::should_terminate()) {
3312       cut_short = true;
3313       break;
3314     }
3315     Block* block = C->cfg()->get_block(i);
3316     if (block->is_connector() && !Verbose) {
3317       continue;
3318     }
3319     n = block->head();
3320     if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3321       pc = pcs[n->_idx];
3322       st->print("%*.*x", pc_digits, pc_digits, pc);
3323     }
3324     st->fill_to(prefix_len);
3325     block->dump_head(C->cfg(), st);
3326     if (block->is_connector()) {
3327       st->fill_to(prefix_len);
3328       st->print_cr("# Empty connector block");
3329     } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3330       st->fill_to(prefix_len);
3331       st->print_cr("# Block is sole successor of call");
3332     }
3333 
3334     // For all instructions
3335     for (uint j = 0; j < block->number_of_nodes(); j++) {
3336       if (VMThread::should_terminate()) {
3337         cut_short = true;
3338         break;
3339       }
3340       n = block->get_node(j);
3341       if (valid_bundle_info(n)) {
3342         Bundle* bundle = node_bundling(n);
3343         if (bundle->starts_bundle()) {
3344           starts_bundle = '+';
3345         }
3346       }
3347 
3348       if (WizardMode) {
3349         n->dump();
3350       }
3351 
3352       if( !n->is_Region() &&    // Dont print in the Assembly
3353           !n->is_Phi() &&       // a few noisely useless nodes
3354           !n->is_Proj() &&
3355           !n->is_MachTemp() &&
3356           !n->is_SafePointScalarObject() &&
3357           !n->is_Catch() &&     // Would be nice to print exception table targets
3358           !n->is_MergeMem() &&  // Not very interesting
3359           !n->is_top() &&       // Debug info table constants
3360           !(n->is_Con() && !n->is_Mach())// Debug info table constants
3361           ) {
3362         if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3363           pc = pcs[n->_idx];
3364           st->print("%*.*x", pc_digits, pc_digits, pc);
3365         } else {
3366           st->fill_to(pc_digits);
3367         }
3368         st->print(" %c ", starts_bundle);
3369         starts_bundle = ' ';
3370         st->fill_to(prefix_len);
3371         n->format(C->regalloc(), st);
3372         st->cr();
3373       }
3374 
3375       // Dump the exception table as well
3376       if( n->is_Catch() && (Verbose || WizardMode) ) {
3377         // Print the exception table for this offset
3378         _handler_table.print_subtable_for(pc);
3379       }
3380       st->bol(); // Make sure we start on a new line
3381     }
3382     st->cr(); // one empty line between blocks
3383   } // End of per-block dump
3384 
3385   if (cut_short)  st->print_cr("*** disassembly is cut short ***");
3386 }
3387 #endif
3388 
3389 #ifndef PRODUCT
3390 void PhaseOutput::print_statistics() {
3391   Scheduling::print_statistics();
3392 }
3393 #endif