1 /*
   2  * Copyright (c) 1998, 2024, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/assembler.inline.hpp"
  27 #include "asm/macroAssembler.inline.hpp"
  28 #include "code/compiledIC.hpp"
  29 #include "code/debugInfo.hpp"
  30 #include "code/debugInfoRec.hpp"
  31 #include "compiler/compileBroker.hpp"
  32 #include "compiler/compilerDirectives.hpp"
  33 #include "compiler/disassembler.hpp"
  34 #include "compiler/oopMap.hpp"
  35 #include "gc/shared/barrierSet.hpp"
  36 #include "gc/shared/c2/barrierSetC2.hpp"
  37 #include "memory/allocation.inline.hpp"
  38 #include "memory/allocation.hpp"
  39 #include "opto/ad.hpp"
  40 #include "opto/block.hpp"
  41 #include "opto/c2compiler.hpp"
  42 #include "opto/c2_MacroAssembler.hpp"
  43 #include "opto/callnode.hpp"
  44 #include "opto/cfgnode.hpp"
  45 #include "opto/locknode.hpp"
  46 #include "opto/machnode.hpp"
  47 #include "opto/node.hpp"
  48 #include "opto/optoreg.hpp"
  49 #include "opto/output.hpp"
  50 #include "opto/regalloc.hpp"
  51 #include "opto/runtime.hpp"
  52 #include "opto/subnode.hpp"
  53 #include "opto/type.hpp"
  54 #include "runtime/handles.inline.hpp"
  55 #include "runtime/sharedRuntime.hpp"
  56 #include "utilities/macros.hpp"
  57 #include "utilities/powerOfTwo.hpp"
  58 #include "utilities/xmlstream.hpp"
  59 
  60 #ifndef PRODUCT
  61 #define DEBUG_ARG(x) , x
  62 #else
  63 #define DEBUG_ARG(x)
  64 #endif
  65 
  66 //------------------------------Scheduling----------------------------------
  67 // This class contains all the information necessary to implement instruction
  68 // scheduling and bundling.
  69 class Scheduling {
  70 
  71 private:
  72   // Arena to use
  73   Arena *_arena;
  74 
  75   // Control-Flow Graph info
  76   PhaseCFG *_cfg;
  77 
  78   // Register Allocation info
  79   PhaseRegAlloc *_regalloc;
  80 
  81   // Number of nodes in the method
  82   uint _node_bundling_limit;
  83 
  84   // List of scheduled nodes. Generated in reverse order
  85   Node_List _scheduled;
  86 
  87   // List of nodes currently available for choosing for scheduling
  88   Node_List _available;
  89 
  90   // For each instruction beginning a bundle, the number of following
  91   // nodes to be bundled with it.
  92   Bundle *_node_bundling_base;
  93 
  94   // Mapping from register to Node
  95   Node_List _reg_node;
  96 
  97   // Free list for pinch nodes.
  98   Node_List _pinch_free_list;
  99 
 100   // Number of uses of this node within the containing basic block.
 101   short *_uses;
 102 
 103   // Schedulable portion of current block.  Skips Region/Phi/CreateEx up
 104   // front, branch+proj at end.  Also skips Catch/CProj (same as
 105   // branch-at-end), plus just-prior exception-throwing call.
 106   uint _bb_start, _bb_end;
 107 
 108   // Latency from the end of the basic block as scheduled
 109   unsigned short *_current_latency;
 110 
 111   // Remember the next node
 112   Node *_next_node;
 113 
 114   // Use this for an unconditional branch delay slot
 115   Node *_unconditional_delay_slot;
 116 
 117   // Pointer to a Nop
 118   MachNopNode *_nop;
 119 
 120   // Length of the current bundle, in instructions
 121   uint _bundle_instr_count;
 122 
 123   // Current Cycle number, for computing latencies and bundling
 124   uint _bundle_cycle_number;
 125 
 126   // Bundle information
 127   Pipeline_Use_Element _bundle_use_elements[resource_count];
 128   Pipeline_Use         _bundle_use;
 129 
 130   // Dump the available list
 131   void dump_available() const;
 132 
 133 public:
 134   Scheduling(Arena *arena, Compile &compile);
 135 
 136   // Destructor
 137   NOT_PRODUCT( ~Scheduling(); )
 138 
 139   // Step ahead "i" cycles
 140   void step(uint i);
 141 
 142   // Step ahead 1 cycle, and clear the bundle state (for example,
 143   // at a branch target)
 144   void step_and_clear();
 145 
 146   Bundle* node_bundling(const Node *n) {
 147     assert(valid_bundle_info(n), "oob");
 148     return (&_node_bundling_base[n->_idx]);
 149   }
 150 
 151   bool valid_bundle_info(const Node *n) const {
 152     return (_node_bundling_limit > n->_idx);
 153   }
 154 
 155   bool starts_bundle(const Node *n) const {
 156     return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle());
 157   }
 158 
 159   // Do the scheduling
 160   void DoScheduling();
 161 
 162   // Compute the register antidependencies within a basic block
 163   void ComputeRegisterAntidependencies(Block *bb);
 164   void verify_do_def( Node *n, OptoReg::Name def, const char *msg );
 165   void verify_good_schedule( Block *b, const char *msg );
 166   void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def );
 167   void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg );
 168 
 169   // Add a node to the current bundle
 170   void AddNodeToBundle(Node *n, const Block *bb);
 171 
 172   // Return an integer less than, equal to, or greater than zero
 173   // if the stack offset of the first argument is respectively
 174   // less than, equal to, or greater than the second.
 175   int compare_two_spill_nodes(Node* first, Node* second);
 176 
 177   // Add a node to the list of available nodes
 178   void AddNodeToAvailableList(Node *n);
 179 
 180   // Compute the local use count for the nodes in a block, and compute
 181   // the list of instructions with no uses in the block as available
 182   void ComputeUseCount(const Block *bb);
 183 
 184   // Choose an instruction from the available list to add to the bundle
 185   Node * ChooseNodeToBundle();
 186 
 187   // See if this Node fits into the currently accumulating bundle
 188   bool NodeFitsInBundle(Node *n);
 189 
 190   // Decrement the use count for a node
 191  void DecrementUseCounts(Node *n, const Block *bb);
 192 
 193   // Garbage collect pinch nodes for reuse by other blocks.
 194   void garbage_collect_pinch_nodes();
 195   // Clean up a pinch node for reuse (helper for above).
 196   void cleanup_pinch( Node *pinch );
 197 
 198   // Information for statistics gathering
 199 #ifndef PRODUCT
 200 private:
 201   // Gather information on size of nops relative to total
 202   uint _branches, _unconditional_delays;
 203 
 204   static uint _total_nop_size, _total_method_size;
 205   static uint _total_branches, _total_unconditional_delays;
 206   static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
 207 
 208 public:
 209   static void print_statistics();
 210 
 211   static void increment_instructions_per_bundle(uint i) {
 212     _total_instructions_per_bundle[i]++;
 213   }
 214 
 215   static void increment_nop_size(uint s) {
 216     _total_nop_size += s;
 217   }
 218 
 219   static void increment_method_size(uint s) {
 220     _total_method_size += s;
 221   }
 222 #endif
 223 
 224 };
 225 
 226 PhaseOutput::PhaseOutput()
 227   : Phase(Phase::Output),
 228     _code_buffer("Compile::Fill_buffer"),
 229     _first_block_size(0),
 230     _handler_table(),
 231     _inc_table(),
 232     _stub_list(),
 233     _oop_map_set(nullptr),
 234     _scratch_buffer_blob(nullptr),
 235     _scratch_locs_memory(nullptr),
 236     _scratch_const_size(-1),
 237     _in_scratch_emit_size(false),
 238     _frame_slots(0),
 239     _code_offsets(),
 240     _node_bundling_limit(0),
 241     _node_bundling_base(nullptr),
 242     _orig_pc_slot(0),
 243     _orig_pc_slot_offset_in_bytes(0),
 244     _buf_sizes(),
 245     _block(nullptr),
 246     _index(0) {
 247   C->set_output(this);
 248   if (C->stub_name() == nullptr) {
 249     _orig_pc_slot = C->fixed_slots() - (sizeof(address) / VMRegImpl::stack_slot_size);
 250   }
 251 }
 252 
 253 PhaseOutput::~PhaseOutput() {
 254   C->set_output(nullptr);
 255   if (_scratch_buffer_blob != nullptr) {
 256     BufferBlob::free(_scratch_buffer_blob);
 257   }
 258 }
 259 
 260 void PhaseOutput::perform_mach_node_analysis() {
 261   // Late barrier analysis must be done after schedule and bundle
 262   // Otherwise liveness based spilling will fail
 263   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 264   bs->late_barrier_analysis();
 265 
 266   pd_perform_mach_node_analysis();
 267 
 268   C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
 269 }
 270 
 271 // Convert Nodes to instruction bits and pass off to the VM
 272 void PhaseOutput::Output() {
 273   // RootNode goes
 274   assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
 275 
 276   // The number of new nodes (mostly MachNop) is proportional to
 277   // the number of java calls and inner loops which are aligned.
 278   if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
 279                             C->inner_loops()*(OptoLoopAlignment-1)),
 280                            "out of nodes before code generation" ) ) {
 281     return;
 282   }
 283   // Make sure I can find the Start Node
 284   Block *entry = C->cfg()->get_block(1);
 285   Block *broot = C->cfg()->get_root_block();
 286 
 287   const StartNode *start = entry->head()->as_Start();
 288 
 289   // Replace StartNode with prolog
 290   MachPrologNode *prolog = new MachPrologNode();
 291   entry->map_node(prolog, 0);
 292   C->cfg()->map_node_to_block(prolog, entry);
 293   C->cfg()->unmap_node_from_block(start); // start is no longer in any block
 294 
 295   // Virtual methods need an unverified entry point
 296 
 297   if( C->is_osr_compilation() ) {
 298     if( PoisonOSREntry ) {
 299       // TODO: Should use a ShouldNotReachHereNode...
 300       C->cfg()->insert( broot, 0, new MachBreakpointNode() );
 301     }
 302   } else {
 303     if( C->method() && !C->method()->flags().is_static() ) {
 304       // Insert unvalidated entry point
 305       C->cfg()->insert( broot, 0, new MachUEPNode() );
 306     }
 307 
 308   }
 309 
 310   // Break before main entry point
 311   if ((C->method() && C->directive()->BreakAtExecuteOption) ||
 312       (OptoBreakpoint && C->is_method_compilation())       ||
 313       (OptoBreakpointOSR && C->is_osr_compilation())       ||
 314       (OptoBreakpointC2R && !C->method())                   ) {
 315     // checking for C->method() means that OptoBreakpoint does not apply to
 316     // runtime stubs or frame converters
 317     C->cfg()->insert( entry, 1, new MachBreakpointNode() );
 318   }
 319 
 320   // Insert epilogs before every return
 321   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
 322     Block* block = C->cfg()->get_block(i);
 323     if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
 324       Node* m = block->end();
 325       if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
 326         MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
 327         block->add_inst(epilog);
 328         C->cfg()->map_node_to_block(epilog, block);
 329       }
 330     }
 331   }
 332 
 333   // Keeper of sizing aspects
 334   _buf_sizes = BufferSizingData();
 335 
 336   // Initialize code buffer
 337   estimate_buffer_size(_buf_sizes._const);
 338   if (C->failing()) return;
 339 
 340   // Pre-compute the length of blocks and replace
 341   // long branches with short if machine supports it.
 342   // Must be done before ScheduleAndBundle due to SPARC delay slots
 343   uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
 344   blk_starts[0] = 0;
 345   shorten_branches(blk_starts);
 346 
 347   ScheduleAndBundle();
 348   if (C->failing()) {
 349     return;
 350   }
 351 
 352   perform_mach_node_analysis();
 353 
 354   // Complete sizing of codebuffer
 355   CodeBuffer* cb = init_buffer();
 356   if (cb == nullptr || C->failing()) {
 357     return;
 358   }
 359 
 360   BuildOopMaps();
 361 
 362   if (C->failing())  {
 363     return;
 364   }
 365 
 366   C2_MacroAssembler masm(cb);
 367   fill_buffer(&masm, blk_starts);
 368 }
 369 
 370 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
 371   // Determine if we need to generate a stack overflow check.
 372   // Do it if the method is not a stub function and
 373   // has java calls or has frame size > vm_page_size/8.
 374   // The debug VM checks that deoptimization doesn't trigger an
 375   // unexpected stack overflow (compiled method stack banging should
 376   // guarantee it doesn't happen) so we always need the stack bang in
 377   // a debug VM.
 378   return (C->stub_function() == nullptr &&
 379           (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
 380            DEBUG_ONLY(|| true)));
 381 }
 382 
 383 bool PhaseOutput::need_register_stack_bang() const {
 384   // Determine if we need to generate a register stack overflow check.
 385   // This is only used on architectures which have split register
 386   // and memory stacks (ie. IA64).
 387   // Bang if the method is not a stub function and has java calls
 388   return (C->stub_function() == nullptr && C->has_java_calls());
 389 }
 390 
 391 
 392 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
 393 // of a loop. When aligning a loop we need to provide enough instructions
 394 // in cpu's fetch buffer to feed decoders. The loop alignment could be
 395 // avoided if we have enough instructions in fetch buffer at the head of a loop.
 396 // By default, the size is set to 999999 by Block's constructor so that
 397 // a loop will be aligned if the size is not reset here.
 398 //
 399 // Note: Mach instructions could contain several HW instructions
 400 // so the size is estimated only.
 401 //
 402 void PhaseOutput::compute_loop_first_inst_sizes() {
 403   // The next condition is used to gate the loop alignment optimization.
 404   // Don't aligned a loop if there are enough instructions at the head of a loop
 405   // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
 406   // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
 407   // equal to 11 bytes which is the largest address NOP instruction.
 408   if (MaxLoopPad < OptoLoopAlignment - 1) {
 409     uint last_block = C->cfg()->number_of_blocks() - 1;
 410     for (uint i = 1; i <= last_block; i++) {
 411       Block* block = C->cfg()->get_block(i);
 412       // Check the first loop's block which requires an alignment.
 413       if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
 414         uint sum_size = 0;
 415         uint inst_cnt = NumberOfLoopInstrToAlign;
 416         inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
 417 
 418         // Check subsequent fallthrough blocks if the loop's first
 419         // block(s) does not have enough instructions.
 420         Block *nb = block;
 421         while(inst_cnt > 0 &&
 422               i < last_block &&
 423               !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
 424               !nb->has_successor(block)) {
 425           i++;
 426           nb = C->cfg()->get_block(i);
 427           inst_cnt  = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
 428         } // while( inst_cnt > 0 && i < last_block  )
 429 
 430         block->set_first_inst_size(sum_size);
 431       } // f( b->head()->is_Loop() )
 432     } // for( i <= last_block )
 433   } // if( MaxLoopPad < OptoLoopAlignment-1 )
 434 }
 435 
 436 // The architecture description provides short branch variants for some long
 437 // branch instructions. Replace eligible long branches with short branches.
 438 void PhaseOutput::shorten_branches(uint* blk_starts) {
 439 
 440   Compile::TracePhase tp("shorten branches", &timers[_t_shortenBranches]);
 441 
 442   // Compute size of each block, method size, and relocation information size
 443   uint nblocks  = C->cfg()->number_of_blocks();
 444 
 445   uint*      jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
 446   uint*      jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
 447   int*       jmp_nidx   = NEW_RESOURCE_ARRAY(int ,nblocks);
 448 
 449   // Collect worst case block paddings
 450   int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
 451   memset(block_worst_case_pad, 0, nblocks * sizeof(int));
 452 
 453   DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
 454   DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
 455 
 456   bool has_short_branch_candidate = false;
 457 
 458   // Initialize the sizes to 0
 459   int code_size  = 0;          // Size in bytes of generated code
 460   int stub_size  = 0;          // Size in bytes of all stub entries
 461   // Size in bytes of all relocation entries, including those in local stubs.
 462   // Start with 2-bytes of reloc info for the unvalidated entry point
 463   int reloc_size = 1;          // Number of relocation entries
 464 
 465   // Make three passes.  The first computes pessimistic blk_starts,
 466   // relative jmp_offset and reloc_size information.  The second performs
 467   // short branch substitution using the pessimistic sizing.  The
 468   // third inserts nops where needed.
 469 
 470   // Step one, perform a pessimistic sizing pass.
 471   uint last_call_adr = max_juint;
 472   uint last_avoid_back_to_back_adr = max_juint;
 473   uint nop_size = (new MachNopNode())->size(C->regalloc());
 474   for (uint i = 0; i < nblocks; i++) { // For all blocks
 475     Block* block = C->cfg()->get_block(i);
 476     _block = block;
 477 
 478     // During short branch replacement, we store the relative (to blk_starts)
 479     // offset of jump in jmp_offset, rather than the absolute offset of jump.
 480     // This is so that we do not need to recompute sizes of all nodes when
 481     // we compute correct blk_starts in our next sizing pass.
 482     jmp_offset[i] = 0;
 483     jmp_size[i]   = 0;
 484     jmp_nidx[i]   = -1;
 485     DEBUG_ONLY( jmp_target[i] = 0; )
 486     DEBUG_ONLY( jmp_rule[i]   = 0; )
 487 
 488     // Sum all instruction sizes to compute block size
 489     uint last_inst = block->number_of_nodes();
 490     uint blk_size = 0;
 491     for (uint j = 0; j < last_inst; j++) {
 492       _index = j;
 493       Node* nj = block->get_node(_index);
 494       // Handle machine instruction nodes
 495       if (nj->is_Mach()) {
 496         MachNode* mach = nj->as_Mach();
 497         blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
 498         reloc_size += mach->reloc();
 499         if (mach->is_MachCall()) {
 500           // add size information for trampoline stub
 501           // class CallStubImpl is platform-specific and defined in the *.ad files.
 502           stub_size  += CallStubImpl::size_call_trampoline();
 503           reloc_size += CallStubImpl::reloc_call_trampoline();
 504 
 505           MachCallNode *mcall = mach->as_MachCall();
 506           // This destination address is NOT PC-relative
 507 
 508           mcall->method_set((intptr_t)mcall->entry_point());
 509 
 510           if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
 511             stub_size  += CompiledDirectCall::to_interp_stub_size();
 512             reloc_size += CompiledDirectCall::reloc_to_interp_stub();
 513           }
 514         } else if (mach->is_MachSafePoint()) {
 515           // If call/safepoint are adjacent, account for possible
 516           // nop to disambiguate the two safepoints.
 517           // ScheduleAndBundle() can rearrange nodes in a block,
 518           // check for all offsets inside this block.
 519           if (last_call_adr >= blk_starts[i]) {
 520             blk_size += nop_size;
 521           }
 522         }
 523         if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
 524           // Nop is inserted between "avoid back to back" instructions.
 525           // ScheduleAndBundle() can rearrange nodes in a block,
 526           // check for all offsets inside this block.
 527           if (last_avoid_back_to_back_adr >= blk_starts[i]) {
 528             blk_size += nop_size;
 529           }
 530         }
 531         if (mach->may_be_short_branch()) {
 532           if (!nj->is_MachBranch()) {
 533 #ifndef PRODUCT
 534             nj->dump(3);
 535 #endif
 536             Unimplemented();
 537           }
 538           assert(jmp_nidx[i] == -1, "block should have only one branch");
 539           jmp_offset[i] = blk_size;
 540           jmp_size[i]   = nj->size(C->regalloc());
 541           jmp_nidx[i]   = j;
 542           has_short_branch_candidate = true;
 543         }
 544       }
 545       blk_size += nj->size(C->regalloc());
 546       // Remember end of call offset
 547       if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
 548         last_call_adr = blk_starts[i]+blk_size;
 549       }
 550       // Remember end of avoid_back_to_back offset
 551       if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
 552         last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
 553       }
 554     }
 555 
 556     // When the next block starts a loop, we may insert pad NOP
 557     // instructions.  Since we cannot know our future alignment,
 558     // assume the worst.
 559     if (i < nblocks - 1) {
 560       Block* nb = C->cfg()->get_block(i + 1);
 561       int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
 562       if (max_loop_pad > 0) {
 563         assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
 564         // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
 565         // If either is the last instruction in this block, bump by
 566         // max_loop_pad in lock-step with blk_size, so sizing
 567         // calculations in subsequent blocks still can conservatively
 568         // detect that it may the last instruction in this block.
 569         if (last_call_adr == blk_starts[i]+blk_size) {
 570           last_call_adr += max_loop_pad;
 571         }
 572         if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
 573           last_avoid_back_to_back_adr += max_loop_pad;
 574         }
 575         blk_size += max_loop_pad;
 576         block_worst_case_pad[i + 1] = max_loop_pad;
 577       }
 578     }
 579 
 580     // Save block size; update total method size
 581     blk_starts[i+1] = blk_starts[i]+blk_size;
 582   }
 583 
 584   // Step two, replace eligible long jumps.
 585   bool progress = true;
 586   uint last_may_be_short_branch_adr = max_juint;
 587   while (has_short_branch_candidate && progress) {
 588     progress = false;
 589     has_short_branch_candidate = false;
 590     int adjust_block_start = 0;
 591     for (uint i = 0; i < nblocks; i++) {
 592       Block* block = C->cfg()->get_block(i);
 593       int idx = jmp_nidx[i];
 594       MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach();
 595       if (mach != nullptr && mach->may_be_short_branch()) {
 596 #ifdef ASSERT
 597         assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
 598         int j;
 599         // Find the branch; ignore trailing NOPs.
 600         for (j = block->number_of_nodes()-1; j>=0; j--) {
 601           Node* n = block->get_node(j);
 602           if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
 603             break;
 604         }
 605         assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
 606 #endif
 607         int br_size = jmp_size[i];
 608         int br_offs = blk_starts[i] + jmp_offset[i];
 609 
 610         // This requires the TRUE branch target be in succs[0]
 611         uint bnum = block->non_connector_successor(0)->_pre_order;
 612         int offset = blk_starts[bnum] - br_offs;
 613         if (bnum > i) { // adjust following block's offset
 614           offset -= adjust_block_start;
 615         }
 616 
 617         // This block can be a loop header, account for the padding
 618         // in the previous block.
 619         int block_padding = block_worst_case_pad[i];
 620         assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
 621         // In the following code a nop could be inserted before
 622         // the branch which will increase the backward distance.
 623         bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
 624         assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
 625 
 626         if (needs_padding && offset <= 0)
 627           offset -= nop_size;
 628 
 629         if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
 630           // We've got a winner.  Replace this branch.
 631           MachNode* replacement = mach->as_MachBranch()->short_branch_version();
 632 
 633           // Update the jmp_size.
 634           int new_size = replacement->size(C->regalloc());
 635           int diff     = br_size - new_size;
 636           assert(diff >= (int)nop_size, "short_branch size should be smaller");
 637           // Conservatively take into account padding between
 638           // avoid_back_to_back branches. Previous branch could be
 639           // converted into avoid_back_to_back branch during next
 640           // rounds.
 641           if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
 642             jmp_offset[i] += nop_size;
 643             diff -= nop_size;
 644           }
 645           adjust_block_start += diff;
 646           block->map_node(replacement, idx);
 647           mach->subsume_by(replacement, C);
 648           mach = replacement;
 649           progress = true;
 650 
 651           jmp_size[i] = new_size;
 652           DEBUG_ONLY( jmp_target[i] = bnum; );
 653           DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
 654         } else {
 655           // The jump distance is not short, try again during next iteration.
 656           has_short_branch_candidate = true;
 657         }
 658       } // (mach->may_be_short_branch())
 659       if (mach != nullptr && (mach->may_be_short_branch() ||
 660                            mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
 661         last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
 662       }
 663       blk_starts[i+1] -= adjust_block_start;
 664     }
 665   }
 666 
 667 #ifdef ASSERT
 668   for (uint i = 0; i < nblocks; i++) { // For all blocks
 669     if (jmp_target[i] != 0) {
 670       int br_size = jmp_size[i];
 671       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
 672       if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
 673         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]);
 674       }
 675       assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
 676     }
 677   }
 678 #endif
 679 
 680   // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
 681   // after ScheduleAndBundle().
 682 
 683   // ------------------
 684   // Compute size for code buffer
 685   code_size = blk_starts[nblocks];
 686 
 687   // Relocation records
 688   reloc_size += 1;              // Relo entry for exception handler
 689 
 690   // Adjust reloc_size to number of record of relocation info
 691   // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
 692   // a relocation index.
 693   // The CodeBuffer will expand the locs array if this estimate is too low.
 694   reloc_size *= 10 / sizeof(relocInfo);
 695 
 696   _buf_sizes._reloc = reloc_size;
 697   _buf_sizes._code  = code_size;
 698   _buf_sizes._stub  = stub_size;
 699 }
 700 
 701 //------------------------------FillLocArray-----------------------------------
 702 // Create a bit of debug info and append it to the array.  The mapping is from
 703 // Java local or expression stack to constant, register or stack-slot.  For
 704 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
 705 // entry has been taken care of and caller should skip it).
 706 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
 707   // This should never have accepted Bad before
 708   assert(OptoReg::is_valid(regnum), "location must be valid");
 709   return (OptoReg::is_reg(regnum))
 710          ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
 711          : new LocationValue(Location::new_stk_loc(l_type,  ra->reg2offset(regnum)));
 712 }
 713 
 714 
 715 ObjectValue*
 716 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
 717   for (int i = 0; i < objs->length(); i++) {
 718     assert(objs->at(i)->is_object(), "corrupt object cache");
 719     ObjectValue* sv = (ObjectValue*) objs->at(i);
 720     if (sv->id() == id) {
 721       return sv;
 722     }
 723   }
 724   // Otherwise..
 725   return nullptr;
 726 }
 727 
 728 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
 729                                      ObjectValue* sv ) {
 730   assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition");
 731   objs->append(sv);
 732 }
 733 
 734 
 735 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
 736                             GrowableArray<ScopeValue*> *array,
 737                             GrowableArray<ScopeValue*> *objs ) {
 738   assert( local, "use _top instead of null" );
 739   if (array->length() != idx) {
 740     assert(array->length() == idx + 1, "Unexpected array count");
 741     // Old functionality:
 742     //   return
 743     // New functionality:
 744     //   Assert if the local is not top. In product mode let the new node
 745     //   override the old entry.
 746     assert(local == C->top(), "LocArray collision");
 747     if (local == C->top()) {
 748       return;
 749     }
 750     array->pop();
 751   }
 752   const Type *t = local->bottom_type();
 753 
 754   // Is it a safepoint scalar object node?
 755   if (local->is_SafePointScalarObject()) {
 756     SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
 757 
 758     ObjectValue* sv = (ObjectValue*) sv_for_node_id(objs, spobj->_idx);
 759     if (sv == nullptr) {
 760       ciKlass* cik = t->is_oopptr()->exact_klass();
 761       assert(cik->is_instance_klass() ||
 762              cik->is_array_klass(), "Not supported allocation.");
 763       sv = new ObjectValue(spobj->_idx,
 764                            new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
 765       set_sv_for_object_node(objs, sv);
 766 
 767       uint first_ind = spobj->first_index(sfpt->jvms());
 768       for (uint i = 0; i < spobj->n_fields(); i++) {
 769         Node* fld_node = sfpt->in(first_ind+i);
 770         (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
 771       }
 772     }
 773     array->append(sv);
 774     return;
 775   } else if (local->is_SafePointScalarMerge()) {
 776     SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
 777     ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
 778 
 779     if (mv == nullptr) {
 780       GrowableArray<ScopeValue*> deps;
 781 
 782       int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
 783       (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
 784       assert(deps.length() == 1, "missing value");
 785 
 786       int selector_idx = smerge->selector_idx(sfpt->jvms());
 787       (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
 788       assert(deps.length() == 2, "missing value");
 789 
 790       mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
 791       set_sv_for_object_node(objs, mv);
 792 
 793       for (uint i = 1; i < smerge->req(); i++) {
 794         Node* obj_node = smerge->in(i);
 795         (void)FillLocArray(mv->possible_objects()->length(), sfpt, obj_node, mv->possible_objects(), objs);
 796       }
 797     }
 798     array->append(mv);
 799     return;
 800   }
 801 
 802   // Grab the register number for the local
 803   OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
 804   if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
 805     // Record the double as two float registers.
 806     // The register mask for such a value always specifies two adjacent
 807     // float registers, with the lower register number even.
 808     // Normally, the allocation of high and low words to these registers
 809     // is irrelevant, because nearly all operations on register pairs
 810     // (e.g., StoreD) treat them as a single unit.
 811     // Here, we assume in addition that the words in these two registers
 812     // stored "naturally" (by operations like StoreD and double stores
 813     // within the interpreter) such that the lower-numbered register
 814     // is written to the lower memory address.  This may seem like
 815     // a machine dependency, but it is not--it is a requirement on
 816     // the author of the <arch>.ad file to ensure that, for every
 817     // even/odd double-register pair to which a double may be allocated,
 818     // the word in the even single-register is stored to the first
 819     // memory word.  (Note that register numbers are completely
 820     // arbitrary, and are not tied to any machine-level encodings.)
 821 #ifdef _LP64
 822     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
 823       array->append(new ConstantIntValue((jint)0));
 824       array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
 825     } else if ( t->base() == Type::Long ) {
 826       array->append(new ConstantIntValue((jint)0));
 827       array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
 828     } else if ( t->base() == Type::RawPtr ) {
 829       // jsr/ret return address which must be restored into the full
 830       // width 64-bit stack slot.
 831       array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
 832     }
 833 #else //_LP64
 834     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
 835       // Repack the double/long as two jints.
 836       // The convention the interpreter uses is that the second local
 837       // holds the first raw word of the native double representation.
 838       // This is actually reasonable, since locals and stack arrays
 839       // grow downwards in all implementations.
 840       // (If, on some machine, the interpreter's Java locals or stack
 841       // were to grow upwards, the embedded doubles would be word-swapped.)
 842       array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
 843       array->append(new_loc_value( C->regalloc(),              regnum   , Location::normal ));
 844     }
 845 #endif //_LP64
 846     else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
 847              OptoReg::is_reg(regnum) ) {
 848       array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
 849                                                       ? Location::float_in_dbl : Location::normal ));
 850     } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
 851       array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
 852                                                       ? Location::int_in_long : Location::normal ));
 853     } else if( t->base() == Type::NarrowOop ) {
 854       array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
 855     } else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
 856                t->base() == Type::VectorD || t->base() == Type::VectorX ||
 857                t->base() == Type::VectorY || t->base() == Type::VectorZ) {
 858       array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
 859     } else if (C->regalloc()->is_oop(local)) {
 860       assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
 861              t->base() == Type::AryPtr,
 862              "Unexpected type: %s", t->msg());
 863       array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
 864     } else {
 865       assert(t->base() == Type::Int || t->base() == Type::Half ||
 866              t->base() == Type::FloatCon || t->base() == Type::FloatBot,
 867              "Unexpected type: %s", t->msg());
 868       array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
 869     }
 870     return;
 871   }
 872 
 873   // No register.  It must be constant data.
 874   switch (t->base()) {
 875     case Type::Half:              // Second half of a double
 876       ShouldNotReachHere();       // Caller should skip 2nd halves
 877       break;
 878     case Type::AnyPtr:
 879       array->append(new ConstantOopWriteValue(nullptr));
 880       break;
 881     case Type::AryPtr:
 882     case Type::InstPtr:          // fall through
 883       array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
 884       break;
 885     case Type::NarrowOop:
 886       if (t == TypeNarrowOop::NULL_PTR) {
 887         array->append(new ConstantOopWriteValue(nullptr));
 888       } else {
 889         array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
 890       }
 891       break;
 892     case Type::Int:
 893       array->append(new ConstantIntValue(t->is_int()->get_con()));
 894       break;
 895     case Type::RawPtr:
 896       // A return address (T_ADDRESS).
 897       assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
 898 #ifdef _LP64
 899       // Must be restored to the full-width 64-bit stack slot.
 900       array->append(new ConstantLongValue(t->is_ptr()->get_con()));
 901 #else
 902       array->append(new ConstantIntValue(t->is_ptr()->get_con()));
 903 #endif
 904       break;
 905     case Type::FloatCon: {
 906       float f = t->is_float_constant()->getf();
 907       array->append(new ConstantIntValue(jint_cast(f)));
 908       break;
 909     }
 910     case Type::DoubleCon: {
 911       jdouble d = t->is_double_constant()->getd();
 912 #ifdef _LP64
 913       array->append(new ConstantIntValue((jint)0));
 914       array->append(new ConstantDoubleValue(d));
 915 #else
 916       // Repack the double as two jints.
 917     // The convention the interpreter uses is that the second local
 918     // holds the first raw word of the native double representation.
 919     // This is actually reasonable, since locals and stack arrays
 920     // grow downwards in all implementations.
 921     // (If, on some machine, the interpreter's Java locals or stack
 922     // were to grow upwards, the embedded doubles would be word-swapped.)
 923     jlong_accessor acc;
 924     acc.long_value = jlong_cast(d);
 925     array->append(new ConstantIntValue(acc.words[1]));
 926     array->append(new ConstantIntValue(acc.words[0]));
 927 #endif
 928       break;
 929     }
 930     case Type::Long: {
 931       jlong d = t->is_long()->get_con();
 932 #ifdef _LP64
 933       array->append(new ConstantIntValue((jint)0));
 934       array->append(new ConstantLongValue(d));
 935 #else
 936       // Repack the long as two jints.
 937     // The convention the interpreter uses is that the second local
 938     // holds the first raw word of the native double representation.
 939     // This is actually reasonable, since locals and stack arrays
 940     // grow downwards in all implementations.
 941     // (If, on some machine, the interpreter's Java locals or stack
 942     // were to grow upwards, the embedded doubles would be word-swapped.)
 943     jlong_accessor acc;
 944     acc.long_value = d;
 945     array->append(new ConstantIntValue(acc.words[1]));
 946     array->append(new ConstantIntValue(acc.words[0]));
 947 #endif
 948       break;
 949     }
 950     case Type::Top:               // Add an illegal value here
 951       array->append(new LocationValue(Location()));
 952       break;
 953     default:
 954       ShouldNotReachHere();
 955       break;
 956   }
 957 }
 958 
 959 // Determine if this node starts a bundle
 960 bool PhaseOutput::starts_bundle(const Node *n) const {
 961   return (_node_bundling_limit > n->_idx &&
 962           _node_bundling_base[n->_idx].starts_bundle());
 963 }
 964 
 965 // Determine if there is a monitor that has 'ov' as its owner.
 966 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const {
 967   for (int k = 0; k < monarray->length(); k++) {
 968     MonitorValue* mv = monarray->at(k);
 969     if (mv->owner() == ov) {
 970       return true;
 971     }
 972   }
 973 
 974   return false;
 975 }
 976 
 977 //--------------------------Process_OopMap_Node--------------------------------
 978 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
 979   // Handle special safepoint nodes for synchronization
 980   MachSafePointNode *sfn   = mach->as_MachSafePoint();
 981   MachCallNode      *mcall;
 982 
 983   int safepoint_pc_offset = current_offset;
 984   bool is_method_handle_invoke = false;
 985   bool return_oop = false;
 986   bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
 987   bool arg_escape = false;
 988 
 989   // Add the safepoint in the DebugInfoRecorder
 990   if( !mach->is_MachCall() ) {
 991     mcall = nullptr;
 992     C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
 993   } else {
 994     mcall = mach->as_MachCall();
 995 
 996     // Is the call a MethodHandle call?
 997     if (mcall->is_MachCallJava()) {
 998       if (mcall->as_MachCallJava()->_method_handle_invoke) {
 999         assert(C->has_method_handle_invokes(), "must have been set during call generation");
1000         is_method_handle_invoke = true;
1001       }
1002       arg_escape = mcall->as_MachCallJava()->_arg_escape;
1003     }
1004 
1005     // Check if a call returns an object.
1006     if (mcall->returns_pointer()) {
1007       return_oop = true;
1008     }
1009     safepoint_pc_offset += mcall->ret_addr_offset();
1010     C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1011   }
1012 
1013   // Loop over the JVMState list to add scope information
1014   // Do not skip safepoints with a null method, they need monitor info
1015   JVMState* youngest_jvms = sfn->jvms();
1016   int max_depth = youngest_jvms->depth();
1017 
1018   // Allocate the object pool for scalar-replaced objects -- the map from
1019   // small-integer keys (which can be recorded in the local and ostack
1020   // arrays) to descriptions of the object state.
1021   GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1022 
1023   // Visit scopes from oldest to youngest.
1024   for (int depth = 1; depth <= max_depth; depth++) {
1025     JVMState* jvms = youngest_jvms->of_depth(depth);
1026     int idx;
1027     ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1028     // Safepoints that do not have method() set only provide oop-map and monitor info
1029     // to support GC; these do not support deoptimization.
1030     int num_locs = (method == nullptr) ? 0 : jvms->loc_size();
1031     int num_exps = (method == nullptr) ? 0 : jvms->stk_size();
1032     int num_mon  = jvms->nof_monitors();
1033     assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(),
1034            "JVMS local count must match that of the method");
1035 
1036     // Add Local and Expression Stack Information
1037 
1038     // Insert locals into the locarray
1039     GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1040     for( idx = 0; idx < num_locs; idx++ ) {
1041       FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1042     }
1043 
1044     // Insert expression stack entries into the exparray
1045     GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1046     for( idx = 0; idx < num_exps; idx++ ) {
1047       FillLocArray( idx,  sfn, sfn->stack(jvms, idx), exparray, objs );
1048     }
1049 
1050     // Add in mappings of the monitors
1051     assert( !method ||
1052             !method->is_synchronized() ||
1053             method->is_native() ||
1054             num_mon > 0 ||
1055             !GenerateSynchronizationCode,
1056             "monitors must always exist for synchronized methods");
1057 
1058     // Build the growable array of ScopeValues for exp stack
1059     GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1060 
1061     // Loop over monitors and insert into array
1062     for (idx = 0; idx < num_mon; idx++) {
1063       // Grab the node that defines this monitor
1064       Node* box_node = sfn->monitor_box(jvms, idx);
1065       Node* obj_node = sfn->monitor_obj(jvms, idx);
1066 
1067       // Create ScopeValue for object
1068       ScopeValue *scval = nullptr;
1069 
1070       if (obj_node->is_SafePointScalarObject()) {
1071         SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1072         scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1073         if (scval == nullptr) {
1074           const Type *t = spobj->bottom_type();
1075           ciKlass* cik = t->is_oopptr()->exact_klass();
1076           assert(cik->is_instance_klass() ||
1077                  cik->is_array_klass(), "Not supported allocation.");
1078           ObjectValue* sv = new ObjectValue(spobj->_idx,
1079                                             new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1080           PhaseOutput::set_sv_for_object_node(objs, sv);
1081 
1082           uint first_ind = spobj->first_index(youngest_jvms);
1083           for (uint i = 0; i < spobj->n_fields(); i++) {
1084             Node* fld_node = sfn->in(first_ind+i);
1085             (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1086           }
1087           scval = sv;
1088         }
1089       } else if (obj_node->is_SafePointScalarMerge()) {
1090         SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1091         ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1092 
1093         if (mv == nullptr) {
1094           GrowableArray<ScopeValue*> deps;
1095 
1096           int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1097           FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1098           assert(deps.length() == 1, "missing value");
1099 
1100           int selector_idx = smerge->selector_idx(youngest_jvms);
1101           FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr);
1102           assert(deps.length() == 2, "missing value");
1103 
1104           mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
1105           set_sv_for_object_node(objs, mv);
1106 
1107           for (uint i = 1; i < smerge->req(); i++) {
1108             Node* obj_node = smerge->in(i);
1109             FillLocArray(mv->possible_objects()->length(), sfn, obj_node, mv->possible_objects(), objs);
1110           }
1111         }
1112         scval = mv;
1113       } else if (!obj_node->is_Con()) {
1114         OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1115         if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1116           scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1117         } else {
1118           scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1119         }
1120       } else {
1121         const TypePtr *tp = obj_node->get_ptr_type();
1122         scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1123       }
1124 
1125       OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1126       Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1127       bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1128       monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1129     }
1130 
1131     // Mark ObjectValue nodes as root nodes if they are directly
1132     // referenced in the JVMS.
1133     for (int i = 0; i < objs->length(); i++) {
1134       ScopeValue* sv = objs->at(i);
1135       if (sv->is_object_merge()) {
1136         ObjectMergeValue* merge = sv->as_ObjectMergeValue();
1137 
1138         for (int j = 0; j< merge->possible_objects()->length(); j++) {
1139           ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue();
1140           bool is_root = locarray->contains(ov) || exparray->contains(ov) || contains_as_owner(monarray, ov);
1141           ov->set_root(is_root);
1142         }
1143       }
1144     }
1145 
1146     // We dump the object pool first, since deoptimization reads it in first.
1147     C->debug_info()->dump_object_pool(objs);
1148 
1149     // Build first class objects to pass to scope
1150     DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1151     DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1152     DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1153 
1154     // Make method available for all Safepoints
1155     ciMethod* scope_method = method ? method : C->method();
1156     // Describe the scope here
1157     assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1158     assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1159     // Now we can describe the scope.
1160     methodHandle null_mh;
1161     bool rethrow_exception = false;
1162     C->debug_info()->describe_scope(
1163       safepoint_pc_offset,
1164       null_mh,
1165       scope_method,
1166       jvms->bci(),
1167       jvms->should_reexecute(),
1168       rethrow_exception,
1169       is_method_handle_invoke,
1170       return_oop,
1171       has_ea_local_in_scope,
1172       arg_escape,
1173       locvals,
1174       expvals,
1175       monvals
1176     );
1177   } // End jvms loop
1178 
1179   // Mark the end of the scope set.
1180   C->debug_info()->end_safepoint(safepoint_pc_offset);
1181 }
1182 
1183 
1184 
1185 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1186 class NonSafepointEmitter {
1187     Compile*  C;
1188     JVMState* _pending_jvms;
1189     int       _pending_offset;
1190 
1191     void emit_non_safepoint();
1192 
1193  public:
1194     NonSafepointEmitter(Compile* compile) {
1195       this->C = compile;
1196       _pending_jvms = nullptr;
1197       _pending_offset = 0;
1198     }
1199 
1200     void observe_instruction(Node* n, int pc_offset) {
1201       if (!C->debug_info()->recording_non_safepoints())  return;
1202 
1203       Node_Notes* nn = C->node_notes_at(n->_idx);
1204       if (nn == nullptr || nn->jvms() == nullptr)  return;
1205       if (_pending_jvms != nullptr &&
1206           _pending_jvms->same_calls_as(nn->jvms())) {
1207         // Repeated JVMS?  Stretch it up here.
1208         _pending_offset = pc_offset;
1209       } else {
1210         if (_pending_jvms != nullptr &&
1211             _pending_offset < pc_offset) {
1212           emit_non_safepoint();
1213         }
1214         _pending_jvms = nullptr;
1215         if (pc_offset > C->debug_info()->last_pc_offset()) {
1216           // This is the only way _pending_jvms can become non-null:
1217           _pending_jvms = nn->jvms();
1218           _pending_offset = pc_offset;
1219         }
1220       }
1221     }
1222 
1223     // Stay out of the way of real safepoints:
1224     void observe_safepoint(JVMState* jvms, int pc_offset) {
1225       if (_pending_jvms != nullptr &&
1226           !_pending_jvms->same_calls_as(jvms) &&
1227           _pending_offset < pc_offset) {
1228         emit_non_safepoint();
1229       }
1230       _pending_jvms = nullptr;
1231     }
1232 
1233     void flush_at_end() {
1234       if (_pending_jvms != nullptr) {
1235         emit_non_safepoint();
1236       }
1237       _pending_jvms = nullptr;
1238     }
1239 };
1240 
1241 void NonSafepointEmitter::emit_non_safepoint() {
1242   JVMState* youngest_jvms = _pending_jvms;
1243   int       pc_offset     = _pending_offset;
1244 
1245   // Clear it now:
1246   _pending_jvms = nullptr;
1247 
1248   DebugInformationRecorder* debug_info = C->debug_info();
1249   assert(debug_info->recording_non_safepoints(), "sanity");
1250 
1251   debug_info->add_non_safepoint(pc_offset);
1252   int max_depth = youngest_jvms->depth();
1253 
1254   // Visit scopes from oldest to youngest.
1255   for (int depth = 1; depth <= max_depth; depth++) {
1256     JVMState* jvms = youngest_jvms->of_depth(depth);
1257     ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1258     assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1259     methodHandle null_mh;
1260     debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1261   }
1262 
1263   // Mark the end of the scope set.
1264   debug_info->end_non_safepoint(pc_offset);
1265 }
1266 
1267 //------------------------------init_buffer------------------------------------
1268 void PhaseOutput::estimate_buffer_size(int& const_req) {
1269 
1270   // Set the initially allocated size
1271   const_req = initial_const_capacity;
1272 
1273   // The extra spacing after the code is necessary on some platforms.
1274   // Sometimes we need to patch in a jump after the last instruction,
1275   // if the nmethod has been deoptimized.  (See 4932387, 4894843.)
1276 
1277   // Compute the byte offset where we can store the deopt pc.
1278   if (C->fixed_slots() != 0) {
1279     _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1280   }
1281 
1282   // Compute prolog code size
1283   _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1284   assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1285 
1286   if (C->has_mach_constant_base_node()) {
1287     uint add_size = 0;
1288     // Fill the constant table.
1289     // Note:  This must happen before shorten_branches.
1290     for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1291       Block* b = C->cfg()->get_block(i);
1292 
1293       for (uint j = 0; j < b->number_of_nodes(); j++) {
1294         Node* n = b->get_node(j);
1295 
1296         // If the node is a MachConstantNode evaluate the constant
1297         // value section.
1298         if (n->is_MachConstant()) {
1299           MachConstantNode* machcon = n->as_MachConstant();
1300           machcon->eval_constant(C);
1301         } else if (n->is_Mach()) {
1302           // On Power there are more nodes that issue constants.
1303           add_size += (n->as_Mach()->ins_num_consts() * 8);
1304         }
1305       }
1306     }
1307 
1308     // Calculate the offsets of the constants and the size of the
1309     // constant table (including the padding to the next section).
1310     constant_table().calculate_offsets_and_size();
1311     const_req = constant_table().size() + add_size;
1312   }
1313 
1314   // Initialize the space for the BufferBlob used to find and verify
1315   // instruction size in MachNode::emit_size()
1316   init_scratch_buffer_blob(const_req);
1317 }
1318 
1319 CodeBuffer* PhaseOutput::init_buffer() {
1320   int stub_req  = _buf_sizes._stub;
1321   int code_req  = _buf_sizes._code;
1322   int const_req = _buf_sizes._const;
1323 
1324   int pad_req   = NativeCall::instruction_size;
1325 
1326   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1327   stub_req += bs->estimate_stub_size();
1328 
1329   // nmethod and CodeBuffer count stubs & constants as part of method's code.
1330   // class HandlerImpl is platform-specific and defined in the *.ad files.
1331   int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1332   int deopt_handler_req     = HandlerImpl::size_deopt_handler()     + MAX_stubs_size; // add marginal slop for handler
1333   stub_req += MAX_stubs_size;   // ensure per-stub margin
1334   code_req += MAX_inst_size;    // ensure per-instruction margin
1335 
1336   if (StressCodeBuffers)
1337     code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10;  // force expansion
1338 
1339   int total_req =
1340           const_req +
1341           code_req +
1342           pad_req +
1343           stub_req +
1344           exception_handler_req +
1345           deopt_handler_req;               // deopt handler
1346 
1347   if (C->has_method_handle_invokes())
1348     total_req += deopt_handler_req;  // deopt MH handler
1349 
1350   CodeBuffer* cb = code_buffer();
1351   cb->initialize(total_req, _buf_sizes._reloc);
1352 
1353   // Have we run out of code space?
1354   if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1355     C->record_failure("CodeCache is full");
1356     return nullptr;
1357   }
1358   // Configure the code buffer.
1359   cb->initialize_consts_size(const_req);
1360   cb->initialize_stubs_size(stub_req);
1361   cb->initialize_oop_recorder(C->env()->oop_recorder());
1362 
1363   // fill in the nop array for bundling computations
1364   MachNode *_nop_list[Bundle::_nop_count];
1365   Bundle::initialize_nops(_nop_list);
1366 
1367   return cb;
1368 }
1369 
1370 //------------------------------fill_buffer------------------------------------
1371 void PhaseOutput::fill_buffer(C2_MacroAssembler* masm, uint* blk_starts) {
1372   // blk_starts[] contains offsets calculated during short branches processing,
1373   // offsets should not be increased during following steps.
1374 
1375   // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1376   // of a loop. It is used to determine the padding for loop alignment.
1377   Compile::TracePhase tp("fill buffer", &timers[_t_fillBuffer]);
1378 
1379   compute_loop_first_inst_sizes();
1380 
1381   // Create oopmap set.
1382   _oop_map_set = new OopMapSet();
1383 
1384   // !!!!! This preserves old handling of oopmaps for now
1385   C->debug_info()->set_oopmaps(_oop_map_set);
1386 
1387   uint nblocks  = C->cfg()->number_of_blocks();
1388   // Count and start of implicit null check instructions
1389   uint inct_cnt = 0;
1390   uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1391 
1392   // Count and start of calls
1393   uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1394 
1395   uint  return_offset = 0;
1396   int nop_size = (new MachNopNode())->size(C->regalloc());
1397 
1398   int previous_offset = 0;
1399   int current_offset  = 0;
1400   int last_call_offset = -1;
1401   int last_avoid_back_to_back_offset = -1;
1402 #ifdef ASSERT
1403   uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1404   uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1405   uint* jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
1406   uint* jmp_rule   = NEW_RESOURCE_ARRAY(uint,nblocks);
1407 #endif
1408 
1409   // Create an array of unused labels, one for each basic block, if printing is enabled
1410 #if defined(SUPPORT_OPTO_ASSEMBLY)
1411   int* node_offsets      = nullptr;
1412   uint node_offset_limit = C->unique();
1413 
1414   if (C->print_assembly()) {
1415     node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1416   }
1417   if (node_offsets != nullptr) {
1418     // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1419     memset(node_offsets, 0, node_offset_limit*sizeof(int));
1420   }
1421 #endif
1422 
1423   NonSafepointEmitter non_safepoints(C);  // emit non-safepoints lazily
1424 
1425   // Emit the constant table.
1426   if (C->has_mach_constant_base_node()) {
1427     if (!constant_table().emit(masm)) {
1428       C->record_failure("consts section overflow");
1429       return;
1430     }
1431   }
1432 
1433   // Create an array of labels, one for each basic block
1434   Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1435   for (uint i = 0; i <= nblocks; i++) {
1436     blk_labels[i].init();
1437   }
1438 
1439   // Now fill in the code buffer
1440   Node* delay_slot = nullptr;
1441   for (uint i = 0; i < nblocks; i++) {
1442     Block* block = C->cfg()->get_block(i);
1443     _block = block;
1444     Node* head = block->head();
1445 
1446     // If this block needs to start aligned (i.e, can be reached other
1447     // than by falling-thru from the previous block), then force the
1448     // start of a new bundle.
1449     if (Pipeline::requires_bundling() && starts_bundle(head)) {
1450       masm->code()->flush_bundle(true);
1451     }
1452 
1453 #ifdef ASSERT
1454     if (!block->is_connector()) {
1455       stringStream st;
1456       block->dump_head(C->cfg(), &st);
1457       masm->block_comment(st.freeze());
1458     }
1459     jmp_target[i] = 0;
1460     jmp_offset[i] = 0;
1461     jmp_size[i]   = 0;
1462     jmp_rule[i]   = 0;
1463 #endif
1464     int blk_offset = current_offset;
1465 
1466     // Define the label at the beginning of the basic block
1467     masm->bind(blk_labels[block->_pre_order]);
1468 
1469     uint last_inst = block->number_of_nodes();
1470 
1471     // Emit block normally, except for last instruction.
1472     // Emit means "dump code bits into code buffer".
1473     for (uint j = 0; j<last_inst; j++) {
1474       _index = j;
1475 
1476       // Get the node
1477       Node* n = block->get_node(j);
1478 
1479       // See if delay slots are supported
1480       if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) {
1481         assert(delay_slot == nullptr, "no use of delay slot node");
1482         assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1483 
1484         delay_slot = n;
1485         continue;
1486       }
1487 
1488       // If this starts a new instruction group, then flush the current one
1489       // (but allow split bundles)
1490       if (Pipeline::requires_bundling() && starts_bundle(n))
1491         masm->code()->flush_bundle(false);
1492 
1493       // Special handling for SafePoint/Call Nodes
1494       bool is_mcall = false;
1495       if (n->is_Mach()) {
1496         MachNode *mach = n->as_Mach();
1497         is_mcall = n->is_MachCall();
1498         bool is_sfn = n->is_MachSafePoint();
1499 
1500         // If this requires all previous instructions be flushed, then do so
1501         if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1502           masm->code()->flush_bundle(true);
1503           current_offset = masm->offset();
1504         }
1505 
1506         // A padding may be needed again since a previous instruction
1507         // could be moved to delay slot.
1508 
1509         // align the instruction if necessary
1510         int padding = mach->compute_padding(current_offset);
1511         // Make sure safepoint node for polling is distinct from a call's
1512         // return by adding a nop if needed.
1513         if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1514           padding = nop_size;
1515         }
1516         if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1517             current_offset == last_avoid_back_to_back_offset) {
1518           // Avoid back to back some instructions.
1519           padding = nop_size;
1520         }
1521 
1522         if (padding > 0) {
1523           assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1524           int nops_cnt = padding / nop_size;
1525           MachNode *nop = new MachNopNode(nops_cnt);
1526           block->insert_node(nop, j++);
1527           last_inst++;
1528           C->cfg()->map_node_to_block(nop, block);
1529           // Ensure enough space.
1530           masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1531           if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1532             C->record_failure("CodeCache is full");
1533             return;
1534           }
1535           nop->emit(masm, C->regalloc());
1536           masm->code()->flush_bundle(true);
1537           current_offset = masm->offset();
1538         }
1539 
1540         bool observe_safepoint = is_sfn;
1541         // Remember the start of the last call in a basic block
1542         if (is_mcall) {
1543           MachCallNode *mcall = mach->as_MachCall();
1544 
1545           // This destination address is NOT PC-relative
1546           mcall->method_set((intptr_t)mcall->entry_point());
1547 
1548           // Save the return address
1549           call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1550 
1551           observe_safepoint = mcall->guaranteed_safepoint();
1552         }
1553 
1554         // sfn will be valid whenever mcall is valid now because of inheritance
1555         if (observe_safepoint) {
1556           // Handle special safepoint nodes for synchronization
1557           if (!is_mcall) {
1558             MachSafePointNode *sfn = mach->as_MachSafePoint();
1559             // !!!!! Stubs only need an oopmap right now, so bail out
1560             if (sfn->jvms()->method() == nullptr) {
1561               // Write the oopmap directly to the code blob??!!
1562               continue;
1563             }
1564           } // End synchronization
1565 
1566           non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1567                                            current_offset);
1568           Process_OopMap_Node(mach, current_offset);
1569         } // End if safepoint
1570 
1571           // If this is a null check, then add the start of the previous instruction to the list
1572         else if( mach->is_MachNullCheck() ) {
1573           inct_starts[inct_cnt++] = previous_offset;
1574         }
1575 
1576           // If this is a branch, then fill in the label with the target BB's label
1577         else if (mach->is_MachBranch()) {
1578           // This requires the TRUE branch target be in succs[0]
1579           uint block_num = block->non_connector_successor(0)->_pre_order;
1580 
1581           // Try to replace long branch if delay slot is not used,
1582           // it is mostly for back branches since forward branch's
1583           // distance is not updated yet.
1584           bool delay_slot_is_used = valid_bundle_info(n) &&
1585                                     C->output()->node_bundling(n)->use_unconditional_delay();
1586           if (!delay_slot_is_used && mach->may_be_short_branch()) {
1587             assert(delay_slot == nullptr, "not expecting delay slot node");
1588             int br_size = n->size(C->regalloc());
1589             int offset = blk_starts[block_num] - current_offset;
1590             if (block_num >= i) {
1591               // Current and following block's offset are not
1592               // finalized yet, adjust distance by the difference
1593               // between calculated and final offsets of current block.
1594               offset -= (blk_starts[i] - blk_offset);
1595             }
1596             // In the following code a nop could be inserted before
1597             // the branch which will increase the backward distance.
1598             bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1599             if (needs_padding && offset <= 0)
1600               offset -= nop_size;
1601 
1602             if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1603               // We've got a winner.  Replace this branch.
1604               MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1605 
1606               // Update the jmp_size.
1607               int new_size = replacement->size(C->regalloc());
1608               assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1609               // Insert padding between avoid_back_to_back branches.
1610               if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1611                 MachNode *nop = new MachNopNode();
1612                 block->insert_node(nop, j++);
1613                 C->cfg()->map_node_to_block(nop, block);
1614                 last_inst++;
1615                 nop->emit(masm, C->regalloc());
1616                 masm->code()->flush_bundle(true);
1617                 current_offset = masm->offset();
1618               }
1619 #ifdef ASSERT
1620               jmp_target[i] = block_num;
1621               jmp_offset[i] = current_offset - blk_offset;
1622               jmp_size[i]   = new_size;
1623               jmp_rule[i]   = mach->rule();
1624 #endif
1625               block->map_node(replacement, j);
1626               mach->subsume_by(replacement, C);
1627               n    = replacement;
1628               mach = replacement;
1629             }
1630           }
1631           mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1632         } else if (mach->ideal_Opcode() == Op_Jump) {
1633           for (uint h = 0; h < block->_num_succs; h++) {
1634             Block* succs_block = block->_succs[h];
1635             for (uint j = 1; j < succs_block->num_preds(); j++) {
1636               Node* jpn = succs_block->pred(j);
1637               if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1638                 uint block_num = succs_block->non_connector()->_pre_order;
1639                 Label *blkLabel = &blk_labels[block_num];
1640                 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1641               }
1642             }
1643           }
1644         }
1645 #ifdef ASSERT
1646           // Check that oop-store precedes the card-mark
1647         else if (mach->ideal_Opcode() == Op_StoreCM) {
1648           uint storeCM_idx = j;
1649           int count = 0;
1650           for (uint prec = mach->req(); prec < mach->len(); prec++) {
1651             Node *oop_store = mach->in(prec);  // Precedence edge
1652             if (oop_store == nullptr) continue;
1653             count++;
1654             uint i4;
1655             for (i4 = 0; i4 < last_inst; ++i4) {
1656               if (block->get_node(i4) == oop_store) {
1657                 break;
1658               }
1659             }
1660             // Note: This test can provide a false failure if other precedence
1661             // edges have been added to the storeCMNode.
1662             assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1663           }
1664           assert(count > 0, "storeCM expects at least one precedence edge");
1665         }
1666 #endif
1667         else if (!n->is_Proj()) {
1668           // Remember the beginning of the previous instruction, in case
1669           // it's followed by a flag-kill and a null-check.  Happens on
1670           // Intel all the time, with add-to-memory kind of opcodes.
1671           previous_offset = current_offset;
1672         }
1673 
1674         // Not an else-if!
1675         // If this is a trap based cmp then add its offset to the list.
1676         if (mach->is_TrapBasedCheckNode()) {
1677           inct_starts[inct_cnt++] = current_offset;
1678         }
1679       }
1680 
1681       // Verify that there is sufficient space remaining
1682       masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1683       if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1684         C->record_failure("CodeCache is full");
1685         return;
1686       }
1687 
1688       // Save the offset for the listing
1689 #if defined(SUPPORT_OPTO_ASSEMBLY)
1690       if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1691         node_offsets[n->_idx] = masm->offset();
1692       }
1693 #endif
1694       assert(!C->failing(), "Should not reach here if failing.");
1695 
1696       // "Normal" instruction case
1697       DEBUG_ONLY(uint instr_offset = masm->offset());
1698       n->emit(masm, C->regalloc());
1699       current_offset = masm->offset();
1700 
1701       // Above we only verified that there is enough space in the instruction section.
1702       // However, the instruction may emit stubs that cause code buffer expansion.
1703       // Bail out here if expansion failed due to a lack of code cache space.
1704       if (C->failing()) {
1705         return;
1706       }
1707 
1708       assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1709              "ret_addr_offset() not within emitted code");
1710 
1711 #ifdef ASSERT
1712       uint n_size = n->size(C->regalloc());
1713       if (n_size < (current_offset-instr_offset)) {
1714         MachNode* mach = n->as_Mach();
1715         n->dump();
1716         mach->dump_format(C->regalloc(), tty);
1717         tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1718         Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty);
1719         tty->print_cr(" ------------------- ");
1720         BufferBlob* blob = this->scratch_buffer_blob();
1721         address blob_begin = blob->content_begin();
1722         Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1723         assert(false, "wrong size of mach node");
1724       }
1725 #endif
1726       non_safepoints.observe_instruction(n, current_offset);
1727 
1728       // mcall is last "call" that can be a safepoint
1729       // record it so we can see if a poll will directly follow it
1730       // in which case we'll need a pad to make the PcDesc sites unique
1731       // see  5010568. This can be slightly inaccurate but conservative
1732       // in the case that return address is not actually at current_offset.
1733       // This is a small price to pay.
1734 
1735       if (is_mcall) {
1736         last_call_offset = current_offset;
1737       }
1738 
1739       if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1740         // Avoid back to back some instructions.
1741         last_avoid_back_to_back_offset = current_offset;
1742       }
1743 
1744       // See if this instruction has a delay slot
1745       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1746         guarantee(delay_slot != nullptr, "expecting delay slot node");
1747 
1748         // Back up 1 instruction
1749         masm->code()->set_insts_end(masm->code()->insts_end() - Pipeline::instr_unit_size());
1750 
1751         // Save the offset for the listing
1752 #if defined(SUPPORT_OPTO_ASSEMBLY)
1753         if ((node_offsets != nullptr) && (delay_slot->_idx < node_offset_limit)) {
1754           node_offsets[delay_slot->_idx] = masm->offset();
1755         }
1756 #endif
1757 
1758         // Support a SafePoint in the delay slot
1759         if (delay_slot->is_MachSafePoint()) {
1760           MachNode *mach = delay_slot->as_Mach();
1761           // !!!!! Stubs only need an oopmap right now, so bail out
1762           if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == nullptr) {
1763             // Write the oopmap directly to the code blob??!!
1764             delay_slot = nullptr;
1765             continue;
1766           }
1767 
1768           int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1769           non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1770                                            adjusted_offset);
1771           // Generate an OopMap entry
1772           Process_OopMap_Node(mach, adjusted_offset);
1773         }
1774 
1775         // Insert the delay slot instruction
1776         delay_slot->emit(masm, C->regalloc());
1777 
1778         // Don't reuse it
1779         delay_slot = nullptr;
1780       }
1781 
1782     } // End for all instructions in block
1783 
1784     // If the next block is the top of a loop, pad this block out to align
1785     // the loop top a little. Helps prevent pipe stalls at loop back branches.
1786     if (i < nblocks-1) {
1787       Block *nb = C->cfg()->get_block(i + 1);
1788       int padding = nb->alignment_padding(current_offset);
1789       if( padding > 0 ) {
1790         MachNode *nop = new MachNopNode(padding / nop_size);
1791         block->insert_node(nop, block->number_of_nodes());
1792         C->cfg()->map_node_to_block(nop, block);
1793         nop->emit(masm, C->regalloc());
1794         current_offset = masm->offset();
1795       }
1796     }
1797     // Verify that the distance for generated before forward
1798     // short branches is still valid.
1799     guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1800 
1801     // Save new block start offset
1802     blk_starts[i] = blk_offset;
1803   } // End of for all blocks
1804   blk_starts[nblocks] = current_offset;
1805 
1806   non_safepoints.flush_at_end();
1807 
1808   // Offset too large?
1809   if (C->failing())  return;
1810 
1811   // Define a pseudo-label at the end of the code
1812   masm->bind( blk_labels[nblocks] );
1813 
1814   // Compute the size of the first block
1815   _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1816 
1817 #ifdef ASSERT
1818   for (uint i = 0; i < nblocks; i++) { // For all blocks
1819     if (jmp_target[i] != 0) {
1820       int br_size = jmp_size[i];
1821       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1822       if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1823         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]);
1824         assert(false, "Displacement too large for short jmp");
1825       }
1826     }
1827   }
1828 #endif
1829 
1830   if (!masm->code()->finalize_stubs()) {
1831     C->record_failure("CodeCache is full");
1832     return;
1833   }
1834 
1835   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1836   bs->emit_stubs(*masm->code());
1837   if (C->failing())  return;
1838 
1839   // Fill in stubs.
1840   assert(masm->inst_mark() == nullptr, "should be.");
1841   _stub_list.emit(*masm);
1842   if (C->failing())  return;
1843 
1844 #ifndef PRODUCT
1845   // Information on the size of the method, without the extraneous code
1846   Scheduling::increment_method_size(masm->offset());
1847 #endif
1848 
1849   // ------------------
1850   // Fill in exception table entries.
1851   FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1852 
1853   // Only java methods have exception handlers and deopt handlers
1854   // class HandlerImpl is platform-specific and defined in the *.ad files.
1855   if (C->method()) {
1856     // Emit the exception handler code.
1857     _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(masm));
1858     if (C->failing()) {
1859       return; // CodeBuffer::expand failed
1860     }
1861     // Emit the deopt handler code.
1862     _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(masm));
1863 
1864     // Emit the MethodHandle deopt handler code (if required).
1865     if (C->has_method_handle_invokes() && !C->failing()) {
1866       // We can use the same code as for the normal deopt handler, we
1867       // just need a different entry point address.
1868       _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(masm));
1869     }
1870   }
1871 
1872   // One last check for failed CodeBuffer::expand:
1873   if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1874     C->record_failure("CodeCache is full");
1875     return;
1876   }
1877 
1878 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1879   if (C->print_assembly()) {
1880     tty->cr();
1881     tty->print_cr("============================= C2-compiled nmethod ==============================");
1882   }
1883 #endif
1884 
1885 #if defined(SUPPORT_OPTO_ASSEMBLY)
1886   // Dump the assembly code, including basic-block numbers
1887   if (C->print_assembly()) {
1888     ttyLocker ttyl;  // keep the following output all in one block
1889     if (!VMThread::should_terminate()) {  // test this under the tty lock
1890       // print_metadata and dump_asm may safepoint which makes us loose the ttylock.
1891       // We call them first and write to a stringStream, then we retake the lock to
1892       // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
1893       ResourceMark rm;
1894       stringStream method_metadata_str;
1895       if (C->method() != nullptr) {
1896         C->method()->print_metadata(&method_metadata_str);
1897       }
1898       stringStream dump_asm_str;
1899       dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
1900 
1901       NoSafepointVerifier nsv;
1902       ttyLocker ttyl2;
1903       // This output goes directly to the tty, not the compiler log.
1904       // To enable tools to match it up with the compilation activity,
1905       // be sure to tag this tty output with the compile ID.
1906       if (xtty != nullptr) {
1907         xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
1908                    C->is_osr_compilation() ? " compile_kind='osr'" : "");
1909       }
1910       if (C->method() != nullptr) {
1911         tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
1912         tty->print_raw(method_metadata_str.freeze());
1913       } else if (C->stub_name() != nullptr) {
1914         tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
1915       }
1916       tty->cr();
1917       tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
1918       tty->print_raw(dump_asm_str.freeze());
1919       tty->print_cr("--------------------------------------------------------------------------------");
1920       if (xtty != nullptr) {
1921         xtty->tail("opto_assembly");
1922       }
1923     }
1924   }
1925 #endif
1926 }
1927 
1928 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1929   _inc_table.set_size(cnt);
1930 
1931   uint inct_cnt = 0;
1932   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1933     Block* block = C->cfg()->get_block(i);
1934     Node *n = nullptr;
1935     int j;
1936 
1937     // Find the branch; ignore trailing NOPs.
1938     for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1939       n = block->get_node(j);
1940       if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1941         break;
1942       }
1943     }
1944 
1945     // If we didn't find anything, continue
1946     if (j < 0) {
1947       continue;
1948     }
1949 
1950     // Compute ExceptionHandlerTable subtable entry and add it
1951     // (skip empty blocks)
1952     if (n->is_Catch()) {
1953 
1954       // Get the offset of the return from the call
1955       uint call_return = call_returns[block->_pre_order];
1956 #ifdef ASSERT
1957       assert( call_return > 0, "no call seen for this basic block" );
1958       while (block->get_node(--j)->is_MachProj()) ;
1959       assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1960 #endif
1961       // last instruction is a CatchNode, find it's CatchProjNodes
1962       int nof_succs = block->_num_succs;
1963       // allocate space
1964       GrowableArray<intptr_t> handler_bcis(nof_succs);
1965       GrowableArray<intptr_t> handler_pcos(nof_succs);
1966       // iterate through all successors
1967       for (int j = 0; j < nof_succs; j++) {
1968         Block* s = block->_succs[j];
1969         bool found_p = false;
1970         for (uint k = 1; k < s->num_preds(); k++) {
1971           Node* pk = s->pred(k);
1972           if (pk->is_CatchProj() && pk->in(0) == n) {
1973             const CatchProjNode* p = pk->as_CatchProj();
1974             found_p = true;
1975             // add the corresponding handler bci & pco information
1976             if (p->_con != CatchProjNode::fall_through_index) {
1977               // p leads to an exception handler (and is not fall through)
1978               assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
1979               // no duplicates, please
1980               if (!handler_bcis.contains(p->handler_bci())) {
1981                 uint block_num = s->non_connector()->_pre_order;
1982                 handler_bcis.append(p->handler_bci());
1983                 handler_pcos.append(blk_labels[block_num].loc_pos());
1984               }
1985             }
1986           }
1987         }
1988         assert(found_p, "no matching predecessor found");
1989         // Note:  Due to empty block removal, one block may have
1990         // several CatchProj inputs, from the same Catch.
1991       }
1992 
1993       // Set the offset of the return from the call
1994       assert(handler_bcis.find(-1) != -1, "must have default handler");
1995       _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos);
1996       continue;
1997     }
1998 
1999     // Handle implicit null exception table updates
2000     if (n->is_MachNullCheck()) {
2001       uint block_num = block->non_connector_successor(0)->_pre_order;
2002       _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2003       continue;
2004     }
2005     // Handle implicit exception table updates: trap instructions.
2006     if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
2007       uint block_num = block->non_connector_successor(0)->_pre_order;
2008       _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2009       continue;
2010     }
2011   } // End of for all blocks fill in exception table entries
2012 }
2013 
2014 // Static Variables
2015 #ifndef PRODUCT
2016 uint Scheduling::_total_nop_size = 0;
2017 uint Scheduling::_total_method_size = 0;
2018 uint Scheduling::_total_branches = 0;
2019 uint Scheduling::_total_unconditional_delays = 0;
2020 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
2021 #endif
2022 
2023 // Initializer for class Scheduling
2024 
2025 Scheduling::Scheduling(Arena *arena, Compile &compile)
2026         : _arena(arena),
2027           _cfg(compile.cfg()),
2028           _regalloc(compile.regalloc()),
2029           _scheduled(arena),
2030           _available(arena),
2031           _reg_node(arena),
2032           _pinch_free_list(arena),
2033           _next_node(nullptr),
2034           _bundle_instr_count(0),
2035           _bundle_cycle_number(0),
2036           _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
2037 #ifndef PRODUCT
2038         , _branches(0)
2039         , _unconditional_delays(0)
2040 #endif
2041 {
2042   // Create a MachNopNode
2043   _nop = new MachNopNode();
2044 
2045   // Now that the nops are in the array, save the count
2046   // (but allow entries for the nops)
2047   _node_bundling_limit = compile.unique();
2048   uint node_max = _regalloc->node_regs_max_index();
2049 
2050   compile.output()->set_node_bundling_limit(_node_bundling_limit);
2051 
2052   // This one is persistent within the Compile class
2053   _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
2054 
2055   // Allocate space for fixed-size arrays
2056   _uses            = NEW_ARENA_ARRAY(arena, short,          node_max);
2057   _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
2058 
2059   // Clear the arrays
2060   for (uint i = 0; i < node_max; i++) {
2061     ::new (&_node_bundling_base[i]) Bundle();
2062   }
2063   memset(_uses,               0, node_max * sizeof(short));
2064   memset(_current_latency,    0, node_max * sizeof(unsigned short));
2065 
2066   // Clear the bundling information
2067   memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2068 
2069   // Get the last node
2070   Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2071 
2072   _next_node = block->get_node(block->number_of_nodes() - 1);
2073 }
2074 
2075 #ifndef PRODUCT
2076 // Scheduling destructor
2077 Scheduling::~Scheduling() {
2078   _total_branches             += _branches;
2079   _total_unconditional_delays += _unconditional_delays;
2080 }
2081 #endif
2082 
2083 // Step ahead "i" cycles
2084 void Scheduling::step(uint i) {
2085 
2086   Bundle *bundle = node_bundling(_next_node);
2087   bundle->set_starts_bundle();
2088 
2089   // Update the bundle record, but leave the flags information alone
2090   if (_bundle_instr_count > 0) {
2091     bundle->set_instr_count(_bundle_instr_count);
2092     bundle->set_resources_used(_bundle_use.resourcesUsed());
2093   }
2094 
2095   // Update the state information
2096   _bundle_instr_count = 0;
2097   _bundle_cycle_number += i;
2098   _bundle_use.step(i);
2099 }
2100 
2101 void Scheduling::step_and_clear() {
2102   Bundle *bundle = node_bundling(_next_node);
2103   bundle->set_starts_bundle();
2104 
2105   // Update the bundle record
2106   if (_bundle_instr_count > 0) {
2107     bundle->set_instr_count(_bundle_instr_count);
2108     bundle->set_resources_used(_bundle_use.resourcesUsed());
2109 
2110     _bundle_cycle_number += 1;
2111   }
2112 
2113   // Clear the bundling information
2114   _bundle_instr_count = 0;
2115   _bundle_use.reset();
2116 
2117   memcpy(_bundle_use_elements,
2118          Pipeline_Use::elaborated_elements,
2119          sizeof(Pipeline_Use::elaborated_elements));
2120 }
2121 
2122 // Perform instruction scheduling and bundling over the sequence of
2123 // instructions in backwards order.
2124 void PhaseOutput::ScheduleAndBundle() {
2125 
2126   // Don't optimize this if it isn't a method
2127   if (!C->method())
2128     return;
2129 
2130   // Don't optimize this if scheduling is disabled
2131   if (!C->do_scheduling())
2132     return;
2133 
2134   // Scheduling code works only with pairs (8 bytes) maximum.
2135   // And when the scalable vector register is used, we may spill/unspill
2136   // the whole reg regardless of the max vector size.
2137   if (C->max_vector_size() > 8 ||
2138       (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2139     return;
2140   }
2141 
2142   Compile::TracePhase tp("isched", &timers[_t_instrSched]);
2143 
2144   // Create a data structure for all the scheduling information
2145   Scheduling scheduling(Thread::current()->resource_area(), *C);
2146 
2147   // Walk backwards over each basic block, computing the needed alignment
2148   // Walk over all the basic blocks
2149   scheduling.DoScheduling();
2150 
2151 #ifndef PRODUCT
2152   if (C->trace_opto_output()) {
2153     // Buffer and print all at once
2154     ResourceMark rm;
2155     stringStream ss;
2156     ss.print("\n---- After ScheduleAndBundle ----\n");
2157     print_scheduling(&ss);
2158     tty->print("%s", ss.as_string());
2159   }
2160 #endif
2161 }
2162 
2163 #ifndef PRODUCT
2164 // Separated out so that it can be called directly from debugger
2165 void PhaseOutput::print_scheduling() {
2166   print_scheduling(tty);
2167 }
2168 
2169 void PhaseOutput::print_scheduling(outputStream* output_stream) {
2170   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2171     output_stream->print("\nBB#%03d:\n", i);
2172     Block* block = C->cfg()->get_block(i);
2173     for (uint j = 0; j < block->number_of_nodes(); j++) {
2174       Node* n = block->get_node(j);
2175       OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2176       output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2177       n->dump("\n", false, output_stream);
2178     }
2179   }
2180 }
2181 #endif
2182 
2183 // See if this node fits into the present instruction bundle
2184 bool Scheduling::NodeFitsInBundle(Node *n) {
2185   uint n_idx = n->_idx;
2186 
2187   // If this is the unconditional delay instruction, then it fits
2188   if (n == _unconditional_delay_slot) {
2189 #ifndef PRODUCT
2190     if (_cfg->C->trace_opto_output())
2191       tty->print("#     NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2192 #endif
2193     return (true);
2194   }
2195 
2196   // If the node cannot be scheduled this cycle, skip it
2197   if (_current_latency[n_idx] > _bundle_cycle_number) {
2198 #ifndef PRODUCT
2199     if (_cfg->C->trace_opto_output())
2200       tty->print("#     NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2201                  n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2202 #endif
2203     return (false);
2204   }
2205 
2206   const Pipeline *node_pipeline = n->pipeline();
2207 
2208   uint instruction_count = node_pipeline->instructionCount();
2209   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2210     instruction_count = 0;
2211   else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2212     instruction_count++;
2213 
2214   if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2215 #ifndef PRODUCT
2216     if (_cfg->C->trace_opto_output())
2217       tty->print("#     NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2218                  n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2219 #endif
2220     return (false);
2221   }
2222 
2223   // Don't allow non-machine nodes to be handled this way
2224   if (!n->is_Mach() && instruction_count == 0)
2225     return (false);
2226 
2227   // See if there is any overlap
2228   uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2229 
2230   if (delay > 0) {
2231 #ifndef PRODUCT
2232     if (_cfg->C->trace_opto_output())
2233       tty->print("#     NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2234 #endif
2235     return false;
2236   }
2237 
2238 #ifndef PRODUCT
2239   if (_cfg->C->trace_opto_output())
2240     tty->print("#     NodeFitsInBundle [%4d]:  TRUE\n", n_idx);
2241 #endif
2242 
2243   return true;
2244 }
2245 
2246 Node * Scheduling::ChooseNodeToBundle() {
2247   uint siz = _available.size();
2248 
2249   if (siz == 0) {
2250 
2251 #ifndef PRODUCT
2252     if (_cfg->C->trace_opto_output())
2253       tty->print("#   ChooseNodeToBundle: null\n");
2254 #endif
2255     return (nullptr);
2256   }
2257 
2258   // Fast path, if only 1 instruction in the bundle
2259   if (siz == 1) {
2260 #ifndef PRODUCT
2261     if (_cfg->C->trace_opto_output()) {
2262       tty->print("#   ChooseNodeToBundle (only 1): ");
2263       _available[0]->dump();
2264     }
2265 #endif
2266     return (_available[0]);
2267   }
2268 
2269   // Don't bother, if the bundle is already full
2270   if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2271     for ( uint i = 0; i < siz; i++ ) {
2272       Node *n = _available[i];
2273 
2274       // Skip projections, we'll handle them another way
2275       if (n->is_Proj())
2276         continue;
2277 
2278       // This presupposed that instructions are inserted into the
2279       // available list in a legality order; i.e. instructions that
2280       // must be inserted first are at the head of the list
2281       if (NodeFitsInBundle(n)) {
2282 #ifndef PRODUCT
2283         if (_cfg->C->trace_opto_output()) {
2284           tty->print("#   ChooseNodeToBundle: ");
2285           n->dump();
2286         }
2287 #endif
2288         return (n);
2289       }
2290     }
2291   }
2292 
2293   // Nothing fits in this bundle, choose the highest priority
2294 #ifndef PRODUCT
2295   if (_cfg->C->trace_opto_output()) {
2296     tty->print("#   ChooseNodeToBundle: ");
2297     _available[0]->dump();
2298   }
2299 #endif
2300 
2301   return _available[0];
2302 }
2303 
2304 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) {
2305   assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), "");
2306 
2307   OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1));
2308   OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first);
2309   OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1));
2310   OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second);
2311 
2312   // Comparison between stack -> reg and stack -> reg
2313   if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) &&
2314       OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) {
2315     return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo);
2316   }
2317 
2318   // Comparison between reg -> stack and reg -> stack
2319   if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) &&
2320       OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) {
2321     return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo);
2322   }
2323 
2324   return 0; // Not comparable
2325 }
2326 
2327 void Scheduling::AddNodeToAvailableList(Node *n) {
2328   assert( !n->is_Proj(), "projections never directly made available" );
2329 #ifndef PRODUCT
2330   if (_cfg->C->trace_opto_output()) {
2331     tty->print("#   AddNodeToAvailableList: ");
2332     n->dump();
2333   }
2334 #endif
2335 
2336   int latency = _current_latency[n->_idx];
2337 
2338   // Insert in latency order (insertion sort). If two MachSpillCopyNodes
2339   // for stack spilling or unspilling have the same latency, we sort
2340   // them in the order of stack offset. Some ports (e.g. aarch64) may also
2341   // have more opportunities to do ld/st merging
2342   uint i;
2343   for (i = 0; i < _available.size(); i++) {
2344     if (_current_latency[_available[i]->_idx] > latency) {
2345       break;
2346     } else if (_current_latency[_available[i]->_idx] == latency &&
2347                n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() &&
2348                compare_two_spill_nodes(n, _available[i]) > 0) {
2349       break;
2350     }
2351   }
2352 
2353   // Special Check for compares following branches
2354   if( n->is_Mach() && _scheduled.size() > 0 ) {
2355     int op = n->as_Mach()->ideal_Opcode();
2356     Node *last = _scheduled[0];
2357     if( last->is_MachIf() && last->in(1) == n &&
2358         ( op == Op_CmpI ||
2359           op == Op_CmpU ||
2360           op == Op_CmpUL ||
2361           op == Op_CmpP ||
2362           op == Op_CmpF ||
2363           op == Op_CmpD ||
2364           op == Op_CmpL ) ) {
2365 
2366       // Recalculate position, moving to front of same latency
2367       for ( i=0 ; i < _available.size(); i++ )
2368         if (_current_latency[_available[i]->_idx] >= latency)
2369           break;
2370     }
2371   }
2372 
2373   // Insert the node in the available list
2374   _available.insert(i, n);
2375 
2376 #ifndef PRODUCT
2377   if (_cfg->C->trace_opto_output())
2378     dump_available();
2379 #endif
2380 }
2381 
2382 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2383   for ( uint i=0; i < n->len(); i++ ) {
2384     Node *def = n->in(i);
2385     if (!def) continue;
2386     if( def->is_Proj() )        // If this is a machine projection, then
2387       def = def->in(0);         // propagate usage thru to the base instruction
2388 
2389     if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2390       continue;
2391     }
2392 
2393     // Compute the latency
2394     uint l = _bundle_cycle_number + n->latency(i);
2395     if (_current_latency[def->_idx] < l)
2396       _current_latency[def->_idx] = l;
2397 
2398     // If this does not have uses then schedule it
2399     if ((--_uses[def->_idx]) == 0)
2400       AddNodeToAvailableList(def);
2401   }
2402 }
2403 
2404 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2405 #ifndef PRODUCT
2406   if (_cfg->C->trace_opto_output()) {
2407     tty->print("#   AddNodeToBundle: ");
2408     n->dump();
2409   }
2410 #endif
2411 
2412   // Remove this from the available list
2413   uint i;
2414   for (i = 0; i < _available.size(); i++)
2415     if (_available[i] == n)
2416       break;
2417   assert(i < _available.size(), "entry in _available list not found");
2418   _available.remove(i);
2419 
2420   // See if this fits in the current bundle
2421   const Pipeline *node_pipeline = n->pipeline();
2422   const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2423 
2424   // Check for instructions to be placed in the delay slot. We
2425   // do this before we actually schedule the current instruction,
2426   // because the delay slot follows the current instruction.
2427   if (Pipeline::_branch_has_delay_slot &&
2428       node_pipeline->hasBranchDelay() &&
2429       !_unconditional_delay_slot) {
2430 
2431     uint siz = _available.size();
2432 
2433     // Conditional branches can support an instruction that
2434     // is unconditionally executed and not dependent by the
2435     // branch, OR a conditionally executed instruction if
2436     // the branch is taken.  In practice, this means that
2437     // the first instruction at the branch target is
2438     // copied to the delay slot, and the branch goes to
2439     // the instruction after that at the branch target
2440     if ( n->is_MachBranch() ) {
2441 
2442       assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2443       assert( !n->is_Catch(),         "should not look for delay slot for Catch" );
2444 
2445 #ifndef PRODUCT
2446       _branches++;
2447 #endif
2448 
2449       // At least 1 instruction is on the available list
2450       // that is not dependent on the branch
2451       for (uint i = 0; i < siz; i++) {
2452         Node *d = _available[i];
2453         const Pipeline *avail_pipeline = d->pipeline();
2454 
2455         // Don't allow safepoints in the branch shadow, that will
2456         // cause a number of difficulties
2457         if ( avail_pipeline->instructionCount() == 1 &&
2458              !avail_pipeline->hasMultipleBundles() &&
2459              !avail_pipeline->hasBranchDelay() &&
2460              Pipeline::instr_has_unit_size() &&
2461              d->size(_regalloc) == Pipeline::instr_unit_size() &&
2462              NodeFitsInBundle(d) &&
2463              !node_bundling(d)->used_in_delay()) {
2464 
2465           if (d->is_Mach() && !d->is_MachSafePoint()) {
2466             // A node that fits in the delay slot was found, so we need to
2467             // set the appropriate bits in the bundle pipeline information so
2468             // that it correctly indicates resource usage.  Later, when we
2469             // attempt to add this instruction to the bundle, we will skip
2470             // setting the resource usage.
2471             _unconditional_delay_slot = d;
2472             node_bundling(n)->set_use_unconditional_delay();
2473             node_bundling(d)->set_used_in_unconditional_delay();
2474             _bundle_use.add_usage(avail_pipeline->resourceUse());
2475             _current_latency[d->_idx] = _bundle_cycle_number;
2476             _next_node = d;
2477             ++_bundle_instr_count;
2478 #ifndef PRODUCT
2479             _unconditional_delays++;
2480 #endif
2481             break;
2482           }
2483         }
2484       }
2485     }
2486 
2487     // No delay slot, add a nop to the usage
2488     if (!_unconditional_delay_slot) {
2489       // See if adding an instruction in the delay slot will overflow
2490       // the bundle.
2491       if (!NodeFitsInBundle(_nop)) {
2492 #ifndef PRODUCT
2493         if (_cfg->C->trace_opto_output())
2494           tty->print("#  *** STEP(1 instruction for delay slot) ***\n");
2495 #endif
2496         step(1);
2497       }
2498 
2499       _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2500       _next_node = _nop;
2501       ++_bundle_instr_count;
2502     }
2503 
2504     // See if the instruction in the delay slot requires a
2505     // step of the bundles
2506     if (!NodeFitsInBundle(n)) {
2507 #ifndef PRODUCT
2508       if (_cfg->C->trace_opto_output())
2509         tty->print("#  *** STEP(branch won't fit) ***\n");
2510 #endif
2511       // Update the state information
2512       _bundle_instr_count = 0;
2513       _bundle_cycle_number += 1;
2514       _bundle_use.step(1);
2515     }
2516   }
2517 
2518   // Get the number of instructions
2519   uint instruction_count = node_pipeline->instructionCount();
2520   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2521     instruction_count = 0;
2522 
2523   // Compute the latency information
2524   uint delay = 0;
2525 
2526   if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2527     int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2528     if (relative_latency < 0)
2529       relative_latency = 0;
2530 
2531     delay = _bundle_use.full_latency(relative_latency, node_usage);
2532 
2533     // Does not fit in this bundle, start a new one
2534     if (delay > 0) {
2535       step(delay);
2536 
2537 #ifndef PRODUCT
2538       if (_cfg->C->trace_opto_output())
2539         tty->print("#  *** STEP(%d) ***\n", delay);
2540 #endif
2541     }
2542   }
2543 
2544   // If this was placed in the delay slot, ignore it
2545   if (n != _unconditional_delay_slot) {
2546 
2547     if (delay == 0) {
2548       if (node_pipeline->hasMultipleBundles()) {
2549 #ifndef PRODUCT
2550         if (_cfg->C->trace_opto_output())
2551           tty->print("#  *** STEP(multiple instructions) ***\n");
2552 #endif
2553         step(1);
2554       }
2555 
2556       else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2557 #ifndef PRODUCT
2558         if (_cfg->C->trace_opto_output())
2559           tty->print("#  *** STEP(%d >= %d instructions) ***\n",
2560                      instruction_count + _bundle_instr_count,
2561                      Pipeline::_max_instrs_per_cycle);
2562 #endif
2563         step(1);
2564       }
2565     }
2566 
2567     if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2568       _bundle_instr_count++;
2569 
2570     // Set the node's latency
2571     _current_latency[n->_idx] = _bundle_cycle_number;
2572 
2573     // Now merge the functional unit information
2574     if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2575       _bundle_use.add_usage(node_usage);
2576 
2577     // Increment the number of instructions in this bundle
2578     _bundle_instr_count += instruction_count;
2579 
2580     // Remember this node for later
2581     if (n->is_Mach())
2582       _next_node = n;
2583   }
2584 
2585   // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2586   // not in the bb->_nodes array.  This happens for debug-info-only BoxLocks.
2587   // 'Schedule' them (basically ignore in the schedule) but do not insert them
2588   // into the block.  All other scheduled nodes get put in the schedule here.
2589   int op = n->Opcode();
2590   if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2591       (op != Op_Node &&         // Not an unused antidepedence node and
2592        // not an unallocated boxlock
2593        (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2594 
2595     // Push any trailing projections
2596     if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2597       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2598         Node *foi = n->fast_out(i);
2599         if( foi->is_Proj() )
2600           _scheduled.push(foi);
2601       }
2602     }
2603 
2604     // Put the instruction in the schedule list
2605     _scheduled.push(n);
2606   }
2607 
2608 #ifndef PRODUCT
2609   if (_cfg->C->trace_opto_output())
2610     dump_available();
2611 #endif
2612 
2613   // Walk all the definitions, decrementing use counts, and
2614   // if a definition has a 0 use count, place it in the available list.
2615   DecrementUseCounts(n,bb);
2616 }
2617 
2618 // This method sets the use count within a basic block.  We will ignore all
2619 // uses outside the current basic block.  As we are doing a backwards walk,
2620 // any node we reach that has a use count of 0 may be scheduled.  This also
2621 // avoids the problem of cyclic references from phi nodes, as long as phi
2622 // nodes are at the front of the basic block.  This method also initializes
2623 // the available list to the set of instructions that have no uses within this
2624 // basic block.
2625 void Scheduling::ComputeUseCount(const Block *bb) {
2626 #ifndef PRODUCT
2627   if (_cfg->C->trace_opto_output())
2628     tty->print("# -> ComputeUseCount\n");
2629 #endif
2630 
2631   // Clear the list of available and scheduled instructions, just in case
2632   _available.clear();
2633   _scheduled.clear();
2634 
2635   // No delay slot specified
2636   _unconditional_delay_slot = nullptr;
2637 
2638 #ifdef ASSERT
2639   for( uint i=0; i < bb->number_of_nodes(); i++ )
2640     assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2641 #endif
2642 
2643   // Force the _uses count to never go to zero for unscheduable pieces
2644   // of the block
2645   for( uint k = 0; k < _bb_start; k++ )
2646     _uses[bb->get_node(k)->_idx] = 1;
2647   for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2648     _uses[bb->get_node(l)->_idx] = 1;
2649 
2650   // Iterate backwards over the instructions in the block.  Don't count the
2651   // branch projections at end or the block header instructions.
2652   for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2653     Node *n = bb->get_node(j);
2654     if( n->is_Proj() ) continue; // Projections handled another way
2655 
2656     // Account for all uses
2657     for ( uint k = 0; k < n->len(); k++ ) {
2658       Node *inp = n->in(k);
2659       if (!inp) continue;
2660       assert(inp != n, "no cycles allowed" );
2661       if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2662         if (inp->is_Proj()) { // Skip through Proj's
2663           inp = inp->in(0);
2664         }
2665         ++_uses[inp->_idx];     // Count 1 block-local use
2666       }
2667     }
2668 
2669     // If this instruction has a 0 use count, then it is available
2670     if (!_uses[n->_idx]) {
2671       _current_latency[n->_idx] = _bundle_cycle_number;
2672       AddNodeToAvailableList(n);
2673     }
2674 
2675 #ifndef PRODUCT
2676     if (_cfg->C->trace_opto_output()) {
2677       tty->print("#   uses: %3d: ", _uses[n->_idx]);
2678       n->dump();
2679     }
2680 #endif
2681   }
2682 
2683 #ifndef PRODUCT
2684   if (_cfg->C->trace_opto_output())
2685     tty->print("# <- ComputeUseCount\n");
2686 #endif
2687 }
2688 
2689 // This routine performs scheduling on each basic block in reverse order,
2690 // using instruction latencies and taking into account function unit
2691 // availability.
2692 void Scheduling::DoScheduling() {
2693 #ifndef PRODUCT
2694   if (_cfg->C->trace_opto_output())
2695     tty->print("# -> DoScheduling\n");
2696 #endif
2697 
2698   Block *succ_bb = nullptr;
2699   Block *bb;
2700   Compile* C = Compile::current();
2701 
2702   // Walk over all the basic blocks in reverse order
2703   for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2704     bb = _cfg->get_block(i);
2705 
2706 #ifndef PRODUCT
2707     if (_cfg->C->trace_opto_output()) {
2708       tty->print("#  Schedule BB#%03d (initial)\n", i);
2709       for (uint j = 0; j < bb->number_of_nodes(); j++) {
2710         bb->get_node(j)->dump();
2711       }
2712     }
2713 #endif
2714 
2715     // On the head node, skip processing
2716     if (bb == _cfg->get_root_block()) {
2717       continue;
2718     }
2719 
2720     // Skip empty, connector blocks
2721     if (bb->is_connector())
2722       continue;
2723 
2724     // If the following block is not the sole successor of
2725     // this one, then reset the pipeline information
2726     if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2727 #ifndef PRODUCT
2728       if (_cfg->C->trace_opto_output()) {
2729         tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2730                    _next_node->_idx, _bundle_instr_count);
2731       }
2732 #endif
2733       step_and_clear();
2734     }
2735 
2736     // Leave untouched the starting instruction, any Phis, a CreateEx node
2737     // or Top.  bb->get_node(_bb_start) is the first schedulable instruction.
2738     _bb_end = bb->number_of_nodes()-1;
2739     for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2740       Node *n = bb->get_node(_bb_start);
2741       // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2742       // Also, MachIdealNodes do not get scheduled
2743       if( !n->is_Mach() ) continue;     // Skip non-machine nodes
2744       MachNode *mach = n->as_Mach();
2745       int iop = mach->ideal_Opcode();
2746       if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2747       if( iop == Op_Con ) continue;      // Do not schedule Top
2748       if( iop == Op_Node &&     // Do not schedule PhiNodes, ProjNodes
2749           mach->pipeline() == MachNode::pipeline_class() &&
2750           !n->is_SpillCopy() && !n->is_MachMerge() )  // Breakpoints, Prolog, etc
2751         continue;
2752       break;                    // Funny loop structure to be sure...
2753     }
2754     // Compute last "interesting" instruction in block - last instruction we
2755     // might schedule.  _bb_end points just after last schedulable inst.  We
2756     // normally schedule conditional branches (despite them being forced last
2757     // in the block), because they have delay slots we can fill.  Calls all
2758     // have their delay slots filled in the template expansions, so we don't
2759     // bother scheduling them.
2760     Node *last = bb->get_node(_bb_end);
2761     // Ignore trailing NOPs.
2762     while (_bb_end > 0 && last->is_Mach() &&
2763            last->as_Mach()->ideal_Opcode() == Op_Con) {
2764       last = bb->get_node(--_bb_end);
2765     }
2766     assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2767     if( last->is_Catch() ||
2768         (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2769       // There might be a prior call.  Skip it.
2770       while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2771     } else if( last->is_MachNullCheck() ) {
2772       // Backup so the last null-checked memory instruction is
2773       // outside the schedulable range. Skip over the nullcheck,
2774       // projection, and the memory nodes.
2775       Node *mem = last->in(1);
2776       do {
2777         _bb_end--;
2778       } while (mem != bb->get_node(_bb_end));
2779     } else {
2780       // Set _bb_end to point after last schedulable inst.
2781       _bb_end++;
2782     }
2783 
2784     assert( _bb_start <= _bb_end, "inverted block ends" );
2785 
2786     // Compute the register antidependencies for the basic block
2787     ComputeRegisterAntidependencies(bb);
2788     if (C->failing())  return;  // too many D-U pinch points
2789 
2790     // Compute the usage within the block, and set the list of all nodes
2791     // in the block that have no uses within the block.
2792     ComputeUseCount(bb);
2793 
2794     // Schedule the remaining instructions in the block
2795     while ( _available.size() > 0 ) {
2796       Node *n = ChooseNodeToBundle();
2797       guarantee(n != nullptr, "no nodes available");
2798       AddNodeToBundle(n,bb);
2799     }
2800 
2801     assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2802 #ifdef ASSERT
2803     for( uint l = _bb_start; l < _bb_end; l++ ) {
2804       Node *n = bb->get_node(l);
2805       uint m;
2806       for( m = 0; m < _bb_end-_bb_start; m++ )
2807         if( _scheduled[m] == n )
2808           break;
2809       assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2810     }
2811 #endif
2812 
2813     // Now copy the instructions (in reverse order) back to the block
2814     for ( uint k = _bb_start; k < _bb_end; k++ )
2815       bb->map_node(_scheduled[_bb_end-k-1], k);
2816 
2817 #ifndef PRODUCT
2818     if (_cfg->C->trace_opto_output()) {
2819       tty->print("#  Schedule BB#%03d (final)\n", i);
2820       uint current = 0;
2821       for (uint j = 0; j < bb->number_of_nodes(); j++) {
2822         Node *n = bb->get_node(j);
2823         if( valid_bundle_info(n) ) {
2824           Bundle *bundle = node_bundling(n);
2825           if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2826             tty->print("*** Bundle: ");
2827             bundle->dump();
2828           }
2829           n->dump();
2830         }
2831       }
2832     }
2833 #endif
2834 #ifdef ASSERT
2835     verify_good_schedule(bb,"after block local scheduling");
2836 #endif
2837   }
2838 
2839 #ifndef PRODUCT
2840   if (_cfg->C->trace_opto_output())
2841     tty->print("# <- DoScheduling\n");
2842 #endif
2843 
2844   // Record final node-bundling array location
2845   _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2846 
2847 } // end DoScheduling
2848 
2849 // Verify that no live-range used in the block is killed in the block by a
2850 // wrong DEF.  This doesn't verify live-ranges that span blocks.
2851 
2852 // Check for edge existence.  Used to avoid adding redundant precedence edges.
2853 static bool edge_from_to( Node *from, Node *to ) {
2854   for( uint i=0; i<from->len(); i++ )
2855     if( from->in(i) == to )
2856       return true;
2857   return false;
2858 }
2859 
2860 #ifdef ASSERT
2861 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2862   // Check for bad kills
2863   if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2864     Node *prior_use = _reg_node[def];
2865     if( prior_use && !edge_from_to(prior_use,n) ) {
2866       tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2867       n->dump();
2868       tty->print_cr("...");
2869       prior_use->dump();
2870       assert(edge_from_to(prior_use,n), "%s", msg);
2871     }
2872     _reg_node.map(def,nullptr); // Kill live USEs
2873   }
2874 }
2875 
2876 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2877 
2878   // Zap to something reasonable for the verify code
2879   _reg_node.clear();
2880 
2881   // Walk over the block backwards.  Check to make sure each DEF doesn't
2882   // kill a live value (other than the one it's supposed to).  Add each
2883   // USE to the live set.
2884   for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2885     Node *n = b->get_node(i);
2886     int n_op = n->Opcode();
2887     if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2888       // Fat-proj kills a slew of registers
2889       RegMaskIterator rmi(n->out_RegMask());
2890       while (rmi.has_next()) {
2891         OptoReg::Name kill = rmi.next();
2892         verify_do_def(n, kill, msg);
2893       }
2894     } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2895       // Get DEF'd registers the normal way
2896       verify_do_def( n, _regalloc->get_reg_first(n), msg );
2897       verify_do_def( n, _regalloc->get_reg_second(n), msg );
2898     }
2899 
2900     // Now make all USEs live
2901     for( uint i=1; i<n->req(); i++ ) {
2902       Node *def = n->in(i);
2903       assert(def != 0, "input edge required");
2904       OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2905       OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2906       if( OptoReg::is_valid(reg_lo) ) {
2907         assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2908         _reg_node.map(reg_lo,n);
2909       }
2910       if( OptoReg::is_valid(reg_hi) ) {
2911         assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2912         _reg_node.map(reg_hi,n);
2913       }
2914     }
2915 
2916   }
2917 
2918   // Zap to something reasonable for the Antidependence code
2919   _reg_node.clear();
2920 }
2921 #endif
2922 
2923 // Conditionally add precedence edges.  Avoid putting edges on Projs.
2924 static void add_prec_edge_from_to( Node *from, Node *to ) {
2925   if( from->is_Proj() ) {       // Put precedence edge on Proj's input
2926     assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2927     from = from->in(0);
2928   }
2929   if( from != to &&             // No cycles (for things like LD L0,[L0+4] )
2930       !edge_from_to( from, to ) ) // Avoid duplicate edge
2931     from->add_prec(to);
2932 }
2933 
2934 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2935   if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2936     return;
2937 
2938   if (OptoReg::is_reg(def_reg)) {
2939     VMReg vmreg = OptoReg::as_VMReg(def_reg);
2940     if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
2941       // This is one of the high slots of a vector register.
2942       // ScheduleAndBundle already checked there are no live wide
2943       // vectors in this method so it can be safely ignored.
2944       return;
2945     }
2946   }
2947 
2948   Node *pinch = _reg_node[def_reg]; // Get pinch point
2949   if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2950       is_def ) {    // Check for a true def (not a kill)
2951     _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2952     return;
2953   }
2954 
2955   Node *kill = def;             // Rename 'def' to more descriptive 'kill'
2956   debug_only( def = (Node*)((intptr_t)0xdeadbeef); )
2957 
2958   // After some number of kills there _may_ be a later def
2959   Node *later_def = nullptr;
2960 
2961   Compile* C = Compile::current();
2962 
2963   // Finding a kill requires a real pinch-point.
2964   // Check for not already having a pinch-point.
2965   // Pinch points are Op_Node's.
2966   if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2967     later_def = pinch;            // Must be def/kill as optimistic pinch-point
2968     if ( _pinch_free_list.size() > 0) {
2969       pinch = _pinch_free_list.pop();
2970     } else {
2971       pinch = new Node(1); // Pinch point to-be
2972     }
2973     if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2974       DEBUG_ONLY( pinch->dump(); );
2975       assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index());
2976       _cfg->C->record_method_not_compilable("too many D-U pinch points");
2977       return;
2978     }
2979     _cfg->map_node_to_block(pinch, b);      // Pretend it's valid in this block (lazy init)
2980     _reg_node.map(def_reg,pinch); // Record pinch-point
2981     //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
2982     if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2983       pinch->init_req(0, C->top());     // set not null for the next call
2984       add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2985       later_def = nullptr;           // and no later def
2986     }
2987     pinch->set_req(0,later_def);  // Hook later def so we can find it
2988   } else {                        // Else have valid pinch point
2989     if( pinch->in(0) )            // If there is a later-def
2990       later_def = pinch->in(0);   // Get it
2991   }
2992 
2993   // Add output-dependence edge from later def to kill
2994   if( later_def )               // If there is some original def
2995     add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2996 
2997   // See if current kill is also a use, and so is forced to be the pinch-point.
2998   if( pinch->Opcode() == Op_Node ) {
2999     Node *uses = kill->is_Proj() ? kill->in(0) : kill;
3000     for( uint i=1; i<uses->req(); i++ ) {
3001       if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
3002           _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
3003         // Yes, found a use/kill pinch-point
3004         pinch->set_req(0,nullptr);  //
3005         pinch->replace_by(kill); // Move anti-dep edges up
3006         pinch = kill;
3007         _reg_node.map(def_reg,pinch);
3008         return;
3009       }
3010     }
3011   }
3012 
3013   // Add edge from kill to pinch-point
3014   add_prec_edge_from_to(kill,pinch);
3015 }
3016 
3017 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
3018   if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
3019     return;
3020   Node *pinch = _reg_node[use_reg]; // Get pinch point
3021   // Check for no later def_reg/kill in block
3022   if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b &&
3023       // Use has to be block-local as well
3024       _cfg->get_block_for_node(use) == b) {
3025     if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
3026         pinch->req() == 1 ) {   // pinch not yet in block?
3027       pinch->del_req(0);        // yank pointer to later-def, also set flag
3028       // Insert the pinch-point in the block just after the last use
3029       b->insert_node(pinch, b->find_node(use) + 1);
3030       _bb_end++;                // Increase size scheduled region in block
3031     }
3032 
3033     add_prec_edge_from_to(pinch,use);
3034   }
3035 }
3036 
3037 // We insert antidependences between the reads and following write of
3038 // allocated registers to prevent illegal code motion. Hopefully, the
3039 // number of added references should be fairly small, especially as we
3040 // are only adding references within the current basic block.
3041 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
3042 
3043 #ifdef ASSERT
3044   verify_good_schedule(b,"before block local scheduling");
3045 #endif
3046 
3047   // A valid schedule, for each register independently, is an endless cycle
3048   // of: a def, then some uses (connected to the def by true dependencies),
3049   // then some kills (defs with no uses), finally the cycle repeats with a new
3050   // def.  The uses are allowed to float relative to each other, as are the
3051   // kills.  No use is allowed to slide past a kill (or def).  This requires
3052   // antidependencies between all uses of a single def and all kills that
3053   // follow, up to the next def.  More edges are redundant, because later defs
3054   // & kills are already serialized with true or antidependencies.  To keep
3055   // the edge count down, we add a 'pinch point' node if there's more than
3056   // one use or more than one kill/def.
3057 
3058   // We add dependencies in one bottom-up pass.
3059 
3060   // For each instruction we handle it's DEFs/KILLs, then it's USEs.
3061 
3062   // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
3063   // register.  If not, we record the DEF/KILL in _reg_node, the
3064   // register-to-def mapping.  If there is a prior DEF/KILL, we insert a
3065   // "pinch point", a new Node that's in the graph but not in the block.
3066   // We put edges from the prior and current DEF/KILLs to the pinch point.
3067   // We put the pinch point in _reg_node.  If there's already a pinch point
3068   // we merely add an edge from the current DEF/KILL to the pinch point.
3069 
3070   // After doing the DEF/KILLs, we handle USEs.  For each used register, we
3071   // put an edge from the pinch point to the USE.
3072 
3073   // To be expedient, the _reg_node array is pre-allocated for the whole
3074   // compilation.  _reg_node is lazily initialized; it either contains a null,
3075   // or a valid def/kill/pinch-point, or a leftover node from some prior
3076   // block.  Leftover node from some prior block is treated like a null (no
3077   // prior def, so no anti-dependence needed).  Valid def is distinguished by
3078   // it being in the current block.
3079   bool fat_proj_seen = false;
3080   uint last_safept = _bb_end-1;
3081   Node* end_node         = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr;
3082   Node* last_safept_node = end_node;
3083   for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
3084     Node *n = b->get_node(i);
3085     int is_def = n->outcnt();   // def if some uses prior to adding precedence edges
3086     if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
3087       // Fat-proj kills a slew of registers
3088       // This can add edges to 'n' and obscure whether or not it was a def,
3089       // hence the is_def flag.
3090       fat_proj_seen = true;
3091       RegMaskIterator rmi(n->out_RegMask());
3092       while (rmi.has_next()) {
3093         OptoReg::Name kill = rmi.next();
3094         anti_do_def(b, n, kill, is_def);
3095       }
3096     } else {
3097       // Get DEF'd registers the normal way
3098       anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
3099       anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
3100     }
3101 
3102     // Kill projections on a branch should appear to occur on the
3103     // branch, not afterwards, so grab the masks from the projections
3104     // and process them.
3105     if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
3106       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3107         Node* use = n->fast_out(i);
3108         if (use->is_Proj()) {
3109           RegMaskIterator rmi(use->out_RegMask());
3110           while (rmi.has_next()) {
3111             OptoReg::Name kill = rmi.next();
3112             anti_do_def(b, n, kill, false);
3113           }
3114         }
3115       }
3116     }
3117 
3118     // Check each register used by this instruction for a following DEF/KILL
3119     // that must occur afterward and requires an anti-dependence edge.
3120     for( uint j=0; j<n->req(); j++ ) {
3121       Node *def = n->in(j);
3122       if( def ) {
3123         assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
3124         anti_do_use( b, n, _regalloc->get_reg_first(def) );
3125         anti_do_use( b, n, _regalloc->get_reg_second(def) );
3126       }
3127     }
3128     // Do not allow defs of new derived values to float above GC
3129     // points unless the base is definitely available at the GC point.
3130 
3131     Node *m = b->get_node(i);
3132 
3133     // Add precedence edge from following safepoint to use of derived pointer
3134     if( last_safept_node != end_node &&
3135         m != last_safept_node) {
3136       for (uint k = 1; k < m->req(); k++) {
3137         const Type *t = m->in(k)->bottom_type();
3138         if( t->isa_oop_ptr() &&
3139             t->is_ptr()->offset() != 0 ) {
3140           last_safept_node->add_prec( m );
3141           break;
3142         }
3143       }
3144     }
3145 
3146     if( n->jvms() ) {           // Precedence edge from derived to safept
3147       // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3148       if( b->get_node(last_safept) != last_safept_node ) {
3149         last_safept = b->find_node(last_safept_node);
3150       }
3151       for( uint j=last_safept; j > i; j-- ) {
3152         Node *mach = b->get_node(j);
3153         if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3154           mach->add_prec( n );
3155       }
3156       last_safept = i;
3157       last_safept_node = m;
3158     }
3159   }
3160 
3161   if (fat_proj_seen) {
3162     // Garbage collect pinch nodes that were not consumed.
3163     // They are usually created by a fat kill MachProj for a call.
3164     garbage_collect_pinch_nodes();
3165   }
3166 }
3167 
3168 // Garbage collect pinch nodes for reuse by other blocks.
3169 //
3170 // The block scheduler's insertion of anti-dependence
3171 // edges creates many pinch nodes when the block contains
3172 // 2 or more Calls.  A pinch node is used to prevent a
3173 // combinatorial explosion of edges.  If a set of kills for a
3174 // register is anti-dependent on a set of uses (or defs), rather
3175 // than adding an edge in the graph between each pair of kill
3176 // and use (or def), a pinch is inserted between them:
3177 //
3178 //            use1   use2  use3
3179 //                \   |   /
3180 //                 \  |  /
3181 //                  pinch
3182 //                 /  |  \
3183 //                /   |   \
3184 //            kill1 kill2 kill3
3185 //
3186 // One pinch node is created per register killed when
3187 // the second call is encountered during a backwards pass
3188 // over the block.  Most of these pinch nodes are never
3189 // wired into the graph because the register is never
3190 // used or def'ed in the block.
3191 //
3192 void Scheduling::garbage_collect_pinch_nodes() {
3193 #ifndef PRODUCT
3194   if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3195 #endif
3196   int trace_cnt = 0;
3197   for (uint k = 0; k < _reg_node.max(); k++) {
3198     Node* pinch = _reg_node[k];
3199     if ((pinch != nullptr) && pinch->Opcode() == Op_Node &&
3200         // no predecence input edges
3201         (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) {
3202       cleanup_pinch(pinch);
3203       _pinch_free_list.push(pinch);
3204       _reg_node.map(k, nullptr);
3205 #ifndef PRODUCT
3206       if (_cfg->C->trace_opto_output()) {
3207         trace_cnt++;
3208         if (trace_cnt > 40) {
3209           tty->print("\n");
3210           trace_cnt = 0;
3211         }
3212         tty->print(" %d", pinch->_idx);
3213       }
3214 #endif
3215     }
3216   }
3217 #ifndef PRODUCT
3218   if (_cfg->C->trace_opto_output()) tty->print("\n");
3219 #endif
3220 }
3221 
3222 // Clean up a pinch node for reuse.
3223 void Scheduling::cleanup_pinch( Node *pinch ) {
3224   assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3225 
3226   for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3227     Node* use = pinch->last_out(i);
3228     uint uses_found = 0;
3229     for (uint j = use->req(); j < use->len(); j++) {
3230       if (use->in(j) == pinch) {
3231         use->rm_prec(j);
3232         uses_found++;
3233       }
3234     }
3235     assert(uses_found > 0, "must be a precedence edge");
3236     i -= uses_found;    // we deleted 1 or more copies of this edge
3237   }
3238   // May have a later_def entry
3239   pinch->set_req(0, nullptr);
3240 }
3241 
3242 #ifndef PRODUCT
3243 
3244 void Scheduling::dump_available() const {
3245   tty->print("#Availist  ");
3246   for (uint i = 0; i < _available.size(); i++)
3247     tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3248   tty->cr();
3249 }
3250 
3251 // Print Scheduling Statistics
3252 void Scheduling::print_statistics() {
3253   // Print the size added by nops for bundling
3254   tty->print("Nops added %d bytes to total of %d bytes",
3255              _total_nop_size, _total_method_size);
3256   if (_total_method_size > 0)
3257     tty->print(", for %.2f%%",
3258                ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3259   tty->print("\n");
3260 
3261   // Print the number of branch shadows filled
3262   if (Pipeline::_branch_has_delay_slot) {
3263     tty->print("Of %d branches, %d had unconditional delay slots filled",
3264                _total_branches, _total_unconditional_delays);
3265     if (_total_branches > 0)
3266       tty->print(", for %.2f%%",
3267                  ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3268     tty->print("\n");
3269   }
3270 
3271   uint total_instructions = 0, total_bundles = 0;
3272 
3273   for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3274     uint bundle_count   = _total_instructions_per_bundle[i];
3275     total_instructions += bundle_count * i;
3276     total_bundles      += bundle_count;
3277   }
3278 
3279   if (total_bundles > 0)
3280     tty->print("Average ILP (excluding nops) is %.2f\n",
3281                ((double)total_instructions) / ((double)total_bundles));
3282 }
3283 #endif
3284 
3285 //-----------------------init_scratch_buffer_blob------------------------------
3286 // Construct a temporary BufferBlob and cache it for this compile.
3287 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3288   // If there is already a scratch buffer blob allocated and the
3289   // constant section is big enough, use it.  Otherwise free the
3290   // current and allocate a new one.
3291   BufferBlob* blob = scratch_buffer_blob();
3292   if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3293     // Use the current blob.
3294   } else {
3295     if (blob != nullptr) {
3296       BufferBlob::free(blob);
3297     }
3298 
3299     ResourceMark rm;
3300     _scratch_const_size = const_size;
3301     int size = C2Compiler::initial_code_buffer_size(const_size);
3302     blob = BufferBlob::create("Compile::scratch_buffer", size);
3303     // Record the buffer blob for next time.
3304     set_scratch_buffer_blob(blob);
3305     // Have we run out of code space?
3306     if (scratch_buffer_blob() == nullptr) {
3307       // Let CompilerBroker disable further compilations.
3308       C->record_failure("Not enough space for scratch buffer in CodeCache");
3309       return;
3310     }
3311   }
3312 
3313   // Initialize the relocation buffers
3314   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3315   set_scratch_locs_memory(locs_buf);
3316 }
3317 
3318 
3319 //-----------------------scratch_emit_size-------------------------------------
3320 // Helper function that computes size by emitting code
3321 uint PhaseOutput::scratch_emit_size(const Node* n) {
3322   // Start scratch_emit_size section.
3323   set_in_scratch_emit_size(true);
3324 
3325   // Emit into a trash buffer and count bytes emitted.
3326   // This is a pretty expensive way to compute a size,
3327   // but it works well enough if seldom used.
3328   // All common fixed-size instructions are given a size
3329   // method by the AD file.
3330   // Note that the scratch buffer blob and locs memory are
3331   // allocated at the beginning of the compile task, and
3332   // may be shared by several calls to scratch_emit_size.
3333   // The allocation of the scratch buffer blob is particularly
3334   // expensive, since it has to grab the code cache lock.
3335   BufferBlob* blob = this->scratch_buffer_blob();
3336   assert(blob != nullptr, "Initialize BufferBlob at start");
3337   assert(blob->size() > MAX_inst_size, "sanity");
3338   relocInfo* locs_buf = scratch_locs_memory();
3339   address blob_begin = blob->content_begin();
3340   address blob_end   = (address)locs_buf;
3341   assert(blob->contains(blob_end), "sanity");
3342   CodeBuffer buf(blob_begin, blob_end - blob_begin);
3343   buf.initialize_consts_size(_scratch_const_size);
3344   buf.initialize_stubs_size(MAX_stubs_size);
3345   assert(locs_buf != nullptr, "sanity");
3346   int lsize = MAX_locs_size / 3;
3347   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3348   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3349   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3350   // Mark as scratch buffer.
3351   buf.consts()->set_scratch_emit();
3352   buf.insts()->set_scratch_emit();
3353   buf.stubs()->set_scratch_emit();
3354 
3355   // Do the emission.
3356 
3357   Label fakeL; // Fake label for branch instructions.
3358   Label*   saveL = nullptr;
3359   uint save_bnum = 0;
3360   bool is_branch = n->is_MachBranch();
3361   C2_MacroAssembler masm(&buf);
3362   masm.bind(fakeL);
3363   if (is_branch) {
3364     n->as_MachBranch()->save_label(&saveL, &save_bnum);
3365     n->as_MachBranch()->label_set(&fakeL, 0);
3366   }
3367   n->emit(&masm, C->regalloc());
3368 
3369   // Emitting into the scratch buffer should not fail
3370   assert (!C->failing(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3371 
3372   if (is_branch) // Restore label.
3373     n->as_MachBranch()->label_set(saveL, save_bnum);
3374 
3375   // End scratch_emit_size section.
3376   set_in_scratch_emit_size(false);
3377 
3378   return buf.insts_size();
3379 }
3380 
3381 void PhaseOutput::install() {
3382   if (!C->should_install_code()) {
3383     return;
3384   } else if (C->stub_function() != nullptr) {
3385     install_stub(C->stub_name());
3386   } else {
3387     install_code(C->method(),
3388                  C->entry_bci(),
3389                  CompileBroker::compiler2(),
3390                  C->has_unsafe_access(),
3391                  SharedRuntime::is_wide_vector(C->max_vector_size()),
3392                  C->rtm_state());
3393   }
3394 }
3395 
3396 void PhaseOutput::install_code(ciMethod*         target,
3397                                int               entry_bci,
3398                                AbstractCompiler* compiler,
3399                                bool              has_unsafe_access,
3400                                bool              has_wide_vectors,
3401                                RTMState          rtm_state) {
3402   // Check if we want to skip execution of all compiled code.
3403   {
3404 #ifndef PRODUCT
3405     if (OptoNoExecute) {
3406       C->record_method_not_compilable("+OptoNoExecute");  // Flag as failed
3407       return;
3408     }
3409 #endif
3410     Compile::TracePhase tp("install_code", &timers[_t_registerMethod]);
3411 
3412     if (C->is_osr_compilation()) {
3413       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3414       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3415     } else {
3416       if (!target->is_static()) {
3417         // The UEP of an nmethod ensures that the VEP is padded. However, the padding of the UEP is placed
3418         // before the inline cache check, so we don't have to execute any nop instructions when dispatching
3419         // through the UEP, yet we can ensure that the VEP is aligned appropriately.
3420         _code_offsets.set_value(CodeOffsets::Entry, _first_block_size - MacroAssembler::ic_check_size());
3421       }
3422       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3423       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3424     }
3425 
3426     C->env()->register_method(target,
3427                                      entry_bci,
3428                                      &_code_offsets,
3429                                      _orig_pc_slot_offset_in_bytes,
3430                                      code_buffer(),
3431                                      frame_size_in_words(),
3432                                      oop_map_set(),
3433                                      &_handler_table,
3434                                      inc_table(),
3435                                      compiler,
3436                                      has_unsafe_access,
3437                                      SharedRuntime::is_wide_vector(C->max_vector_size()),
3438                                      C->has_monitors(),
3439                                      0,
3440                                      C->rtm_state());
3441 
3442     if (C->log() != nullptr) { // Print code cache state into compiler log
3443       C->log()->code_cache_state();
3444     }
3445   }
3446 }
3447 void PhaseOutput::install_stub(const char* stub_name) {
3448   // Entry point will be accessed using stub_entry_point();
3449   if (code_buffer() == nullptr) {
3450     Matcher::soft_match_failure();
3451   } else {
3452     if (PrintAssembly && (WizardMode || Verbose))
3453       tty->print_cr("### Stub::%s", stub_name);
3454 
3455     if (!C->failing()) {
3456       assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3457 
3458       // Make the NMethod
3459       // For now we mark the frame as never safe for profile stackwalking
3460       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3461                                                       code_buffer(),
3462                                                       CodeOffsets::frame_never_safe,
3463                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
3464                                                       frame_size_in_words(),
3465                                                       oop_map_set(),
3466                                                       false);
3467       assert(rs != nullptr && rs->is_runtime_stub(), "sanity check");
3468 
3469       C->set_stub_entry_point(rs->entry_point());
3470     }
3471   }
3472 }
3473 
3474 // Support for bundling info
3475 Bundle* PhaseOutput::node_bundling(const Node *n) {
3476   assert(valid_bundle_info(n), "oob");
3477   return &_node_bundling_base[n->_idx];
3478 }
3479 
3480 bool PhaseOutput::valid_bundle_info(const Node *n) {
3481   return (_node_bundling_limit > n->_idx);
3482 }
3483 
3484 //------------------------------frame_size_in_words-----------------------------
3485 // frame_slots in units of words
3486 int PhaseOutput::frame_size_in_words() const {
3487   // shift is 0 in LP32 and 1 in LP64
3488   const int shift = (LogBytesPerWord - LogBytesPerInt);
3489   int words = _frame_slots >> shift;
3490   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3491   return words;
3492 }
3493 
3494 // To bang the stack of this compiled method we use the stack size
3495 // that the interpreter would need in case of a deoptimization. This
3496 // removes the need to bang the stack in the deoptimization blob which
3497 // in turn simplifies stack overflow handling.
3498 int PhaseOutput::bang_size_in_bytes() const {
3499   return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3500 }
3501 
3502 //------------------------------dump_asm---------------------------------------
3503 // Dump formatted assembly
3504 #if defined(SUPPORT_OPTO_ASSEMBLY)
3505 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3506 
3507   int pc_digits = 3; // #chars required for pc
3508   int sb_chars  = 3; // #chars for "start bundle" indicator
3509   int tab_size  = 8;
3510   if (pcs != nullptr) {
3511     int max_pc = 0;
3512     for (uint i = 0; i < pc_limit; i++) {
3513       max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3514     }
3515     pc_digits  = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3516   }
3517   int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3518 
3519   bool cut_short = false;
3520   st->print_cr("#");
3521   st->print("#  ");  C->tf()->dump_on(st);  st->cr();
3522   st->print_cr("#");
3523 
3524   // For all blocks
3525   int pc = 0x0;                 // Program counter
3526   char starts_bundle = ' ';
3527   C->regalloc()->dump_frame();
3528 
3529   Node *n = nullptr;
3530   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3531     if (VMThread::should_terminate()) {
3532       cut_short = true;
3533       break;
3534     }
3535     Block* block = C->cfg()->get_block(i);
3536     if (block->is_connector() && !Verbose) {
3537       continue;
3538     }
3539     n = block->head();
3540     if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3541       pc = pcs[n->_idx];
3542       st->print("%*.*x", pc_digits, pc_digits, pc);
3543     }
3544     st->fill_to(prefix_len);
3545     block->dump_head(C->cfg(), st);
3546     if (block->is_connector()) {
3547       st->fill_to(prefix_len);
3548       st->print_cr("# Empty connector block");
3549     } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3550       st->fill_to(prefix_len);
3551       st->print_cr("# Block is sole successor of call");
3552     }
3553 
3554     // For all instructions
3555     Node *delay = nullptr;
3556     for (uint j = 0; j < block->number_of_nodes(); j++) {
3557       if (VMThread::should_terminate()) {
3558         cut_short = true;
3559         break;
3560       }
3561       n = block->get_node(j);
3562       if (valid_bundle_info(n)) {
3563         Bundle* bundle = node_bundling(n);
3564         if (bundle->used_in_unconditional_delay()) {
3565           delay = n;
3566           continue;
3567         }
3568         if (bundle->starts_bundle()) {
3569           starts_bundle = '+';
3570         }
3571       }
3572 
3573       if (WizardMode) {
3574         n->dump();
3575       }
3576 
3577       if( !n->is_Region() &&    // Dont print in the Assembly
3578           !n->is_Phi() &&       // a few noisely useless nodes
3579           !n->is_Proj() &&
3580           !n->is_MachTemp() &&
3581           !n->is_SafePointScalarObject() &&
3582           !n->is_Catch() &&     // Would be nice to print exception table targets
3583           !n->is_MergeMem() &&  // Not very interesting
3584           !n->is_top() &&       // Debug info table constants
3585           !(n->is_Con() && !n->is_Mach())// Debug info table constants
3586           ) {
3587         if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3588           pc = pcs[n->_idx];
3589           st->print("%*.*x", pc_digits, pc_digits, pc);
3590         } else {
3591           st->fill_to(pc_digits);
3592         }
3593         st->print(" %c ", starts_bundle);
3594         starts_bundle = ' ';
3595         st->fill_to(prefix_len);
3596         n->format(C->regalloc(), st);
3597         st->cr();
3598       }
3599 
3600       // If we have an instruction with a delay slot, and have seen a delay,
3601       // then back up and print it
3602       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3603         // Coverity finding - Explicit null dereferenced.
3604         guarantee(delay != nullptr, "no unconditional delay instruction");
3605         if (WizardMode) delay->dump();
3606 
3607         if (node_bundling(delay)->starts_bundle())
3608           starts_bundle = '+';
3609         if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3610           pc = pcs[n->_idx];
3611           st->print("%*.*x", pc_digits, pc_digits, pc);
3612         } else {
3613           st->fill_to(pc_digits);
3614         }
3615         st->print(" %c ", starts_bundle);
3616         starts_bundle = ' ';
3617         st->fill_to(prefix_len);
3618         delay->format(C->regalloc(), st);
3619         st->cr();
3620         delay = nullptr;
3621       }
3622 
3623       // Dump the exception table as well
3624       if( n->is_Catch() && (Verbose || WizardMode) ) {
3625         // Print the exception table for this offset
3626         _handler_table.print_subtable_for(pc);
3627       }
3628       st->bol(); // Make sure we start on a new line
3629     }
3630     st->cr(); // one empty line between blocks
3631     assert(cut_short || delay == nullptr, "no unconditional delay branch");
3632   } // End of per-block dump
3633 
3634   if (cut_short)  st->print_cr("*** disassembly is cut short ***");
3635 }
3636 #endif
3637 
3638 #ifndef PRODUCT
3639 void PhaseOutput::print_statistics() {
3640   Scheduling::print_statistics();
3641 }
3642 #endif