1 /*
   2  * Copyright (c) 1998, 2025, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "asm/assembler.inline.hpp"
  26 #include "code/aotCodeCache.hpp"
  27 #include "code/compiledIC.hpp"
  28 #include "code/debugInfo.hpp"
  29 #include "code/debugInfoRec.hpp"
  30 #include "compiler/compileBroker.hpp"
  31 #include "compiler/compilerDirectives.hpp"
  32 #include "compiler/disassembler.hpp"
  33 #include "compiler/oopMap.hpp"
  34 #include "gc/shared/barrierSet.hpp"
  35 #include "gc/shared/c2/barrierSetC2.hpp"
  36 #include "gc/shared/gc_globals.hpp"
  37 #include "memory/allocation.hpp"
  38 #include "opto/ad.hpp"
  39 #include "opto/block.hpp"
  40 #include "opto/c2_MacroAssembler.hpp"
  41 #include "opto/c2compiler.hpp"
  42 #include "opto/callnode.hpp"
  43 #include "opto/cfgnode.hpp"
  44 #include "opto/locknode.hpp"
  45 #include "opto/machnode.hpp"
  46 #include "opto/node.hpp"
  47 #include "opto/optoreg.hpp"
  48 #include "opto/output.hpp"
  49 #include "opto/regalloc.hpp"
  50 #include "opto/type.hpp"
  51 #include "runtime/sharedRuntime.hpp"
  52 #include "utilities/macros.hpp"
  53 #include "utilities/powerOfTwo.hpp"
  54 #include "utilities/xmlstream.hpp"
  55 
  56 #ifndef PRODUCT
  57 #define DEBUG_ARG(x) , x
  58 #else
  59 #define DEBUG_ARG(x)
  60 #endif
  61 
  62 //------------------------------Scheduling----------------------------------
  63 // This class contains all the information necessary to implement instruction
  64 // scheduling and bundling.
  65 class Scheduling {
  66 
  67 private:
  68   // Arena to use
  69   Arena *_arena;
  70 
  71   // Control-Flow Graph info
  72   PhaseCFG *_cfg;
  73 
  74   // Register Allocation info
  75   PhaseRegAlloc *_regalloc;
  76 
  77   // Number of nodes in the method
  78   uint _node_bundling_limit;
  79 
  80   // List of scheduled nodes. Generated in reverse order
  81   Node_List _scheduled;
  82 
  83   // List of nodes currently available for choosing for scheduling
  84   Node_List _available;
  85 
  86   // For each instruction beginning a bundle, the number of following
  87   // nodes to be bundled with it.
  88   Bundle *_node_bundling_base;
  89 
  90   // Mapping from register to Node
  91   Node_List _reg_node;
  92 
  93   // Free list for pinch nodes.
  94   Node_List _pinch_free_list;
  95 
  96   // Number of uses of this node within the containing basic block.
  97   short *_uses;
  98 
  99   // Schedulable portion of current block.  Skips Region/Phi/CreateEx up
 100   // front, branch+proj at end.  Also skips Catch/CProj (same as
 101   // branch-at-end), plus just-prior exception-throwing call.
 102   uint _bb_start, _bb_end;
 103 
 104   // Latency from the end of the basic block as scheduled
 105   unsigned short *_current_latency;
 106 
 107   // Remember the next node
 108   Node *_next_node;
 109 
 110   // Use this for an unconditional branch delay slot
 111   Node *_unconditional_delay_slot;
 112 
 113   // Pointer to a Nop
 114   MachNopNode *_nop;
 115 
 116   // Length of the current bundle, in instructions
 117   uint _bundle_instr_count;
 118 
 119   // Current Cycle number, for computing latencies and bundling
 120   uint _bundle_cycle_number;
 121 
 122   // Bundle information
 123   Pipeline_Use_Element _bundle_use_elements[resource_count];
 124   Pipeline_Use         _bundle_use;
 125 
 126   // Dump the available list
 127   void dump_available() const;
 128 
 129 public:
 130   Scheduling(Arena *arena, Compile &compile);
 131 
 132   // Destructor
 133   NOT_PRODUCT( ~Scheduling(); )
 134 
 135   // Step ahead "i" cycles
 136   void step(uint i);
 137 
 138   // Step ahead 1 cycle, and clear the bundle state (for example,
 139   // at a branch target)
 140   void step_and_clear();
 141 
 142   Bundle* node_bundling(const Node *n) {
 143     assert(valid_bundle_info(n), "oob");
 144     return (&_node_bundling_base[n->_idx]);
 145   }
 146 
 147   bool valid_bundle_info(const Node *n) const {
 148     return (_node_bundling_limit > n->_idx);
 149   }
 150 
 151   bool starts_bundle(const Node *n) const {
 152     return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle());
 153   }
 154 
 155   // Do the scheduling
 156   void DoScheduling();
 157 
 158   // Compute the register antidependencies within a basic block
 159   void ComputeRegisterAntidependencies(Block *bb);
 160   void verify_do_def( Node *n, OptoReg::Name def, const char *msg );
 161   void verify_good_schedule( Block *b, const char *msg );
 162   void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def );
 163   void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg );
 164 
 165   // Add a node to the current bundle
 166   void AddNodeToBundle(Node *n, const Block *bb);
 167 
 168   // Return an integer less than, equal to, or greater than zero
 169   // if the stack offset of the first argument is respectively
 170   // less than, equal to, or greater than the second.
 171   int compare_two_spill_nodes(Node* first, Node* second);
 172 
 173   // Add a node to the list of available nodes
 174   void AddNodeToAvailableList(Node *n);
 175 
 176   // Compute the local use count for the nodes in a block, and compute
 177   // the list of instructions with no uses in the block as available
 178   void ComputeUseCount(const Block *bb);
 179 
 180   // Choose an instruction from the available list to add to the bundle
 181   Node * ChooseNodeToBundle();
 182 
 183   // See if this Node fits into the currently accumulating bundle
 184   bool NodeFitsInBundle(Node *n);
 185 
 186   // Decrement the use count for a node
 187  void DecrementUseCounts(Node *n, const Block *bb);
 188 
 189   // Garbage collect pinch nodes for reuse by other blocks.
 190   void garbage_collect_pinch_nodes();
 191   // Clean up a pinch node for reuse (helper for above).
 192   void cleanup_pinch( Node *pinch );
 193 
 194   // Information for statistics gathering
 195 #ifndef PRODUCT
 196 private:
 197   // Gather information on size of nops relative to total
 198   uint _branches, _unconditional_delays;
 199 
 200   static uint _total_nop_size, _total_method_size;
 201   static uint _total_branches, _total_unconditional_delays;
 202   static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
 203 
 204 public:
 205   static void print_statistics();
 206 
 207   static void increment_instructions_per_bundle(uint i) {
 208     _total_instructions_per_bundle[i]++;
 209   }
 210 
 211   static void increment_nop_size(uint s) {
 212     _total_nop_size += s;
 213   }
 214 
 215   static void increment_method_size(uint s) {
 216     _total_method_size += s;
 217   }
 218 #endif
 219 
 220 };
 221 
 222 PhaseOutput::PhaseOutput()
 223   : Phase(Phase::Output),
 224     _code_buffer("Compile::Fill_buffer"),
 225     _first_block_size(0),
 226     _handler_table(),
 227     _inc_table(),
 228     _stub_list(),
 229     _oop_map_set(nullptr),
 230     _scratch_buffer_blob(nullptr),
 231     _scratch_locs_memory(nullptr),
 232     _scratch_const_size(-1),
 233     _in_scratch_emit_size(false),
 234     _frame_slots(0),
 235     _code_offsets(),
 236     _node_bundling_limit(0),
 237     _node_bundling_base(nullptr),
 238     _orig_pc_slot(0),
 239     _orig_pc_slot_offset_in_bytes(0),
 240     _buf_sizes(),
 241     _block(nullptr),
 242     _index(0) {
 243   C->set_output(this);
 244   if (C->stub_name() == nullptr) {
 245     int fixed_slots = C->fixed_slots();
 246     if (C->needs_stack_repair()) {
 247       fixed_slots -= 2;
 248     }
 249     // TODO 8284443 Only reserve extra slot if needed
 250     if (InlineTypeReturnedAsFields) {
 251       fixed_slots -= 2;
 252     }
 253     _orig_pc_slot = fixed_slots - (sizeof(address) / VMRegImpl::stack_slot_size);
 254   }
 255 }
 256 
 257 PhaseOutput::~PhaseOutput() {
 258   C->set_output(nullptr);
 259   if (_scratch_buffer_blob != nullptr) {
 260     BufferBlob::free(_scratch_buffer_blob);
 261   }
 262 }
 263 
 264 void PhaseOutput::perform_mach_node_analysis() {
 265   // Late barrier analysis must be done after schedule and bundle
 266   // Otherwise liveness based spilling will fail
 267   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 268   bs->late_barrier_analysis();
 269 
 270   pd_perform_mach_node_analysis();
 271 
 272   C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
 273 }
 274 
 275 // Convert Nodes to instruction bits and pass off to the VM
 276 void PhaseOutput::Output() {
 277   // RootNode goes
 278   assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
 279 
 280   // The number of new nodes (mostly MachNop) is proportional to
 281   // the number of java calls and inner loops which are aligned.
 282   if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
 283                             C->inner_loops()*(OptoLoopAlignment-1)),
 284                            "out of nodes before code generation" ) ) {
 285     return;
 286   }
 287   // Make sure I can find the Start Node
 288   Block *entry = C->cfg()->get_block(1);
 289   Block *broot = C->cfg()->get_root_block();
 290 
 291   const StartNode *start = entry->head()->as_Start();
 292 
 293   // Replace StartNode with prolog
 294   Label verified_entry;
 295   MachPrologNode* prolog = new MachPrologNode(&verified_entry);
 296   entry->map_node(prolog, 0);
 297   C->cfg()->map_node_to_block(prolog, entry);
 298   C->cfg()->unmap_node_from_block(start); // start is no longer in any block
 299 
 300   // Virtual methods need an unverified entry point
 301   if (C->is_osr_compilation()) {
 302     if (PoisonOSREntry) {
 303       // TODO: Should use a ShouldNotReachHereNode...
 304       C->cfg()->insert( broot, 0, new MachBreakpointNode() );
 305     }
 306   } else {
 307     if (C->method()) {
 308       if (C->method()->has_scalarized_args()) {
 309         // Add entry point to unpack all inline type arguments
 310         C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ false));
 311         if (!C->method()->is_static()) {
 312           // Add verified/unverified entry points to only unpack inline type receiver at interface calls
 313           C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ false));
 314           C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true,  /* receiver_only */ true));
 315           C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ true));
 316         }
 317       } else if (!C->method()->is_static()) {
 318         // Insert unvalidated entry point
 319         C->cfg()->insert(broot, 0, new MachUEPNode());
 320       }
 321     }
 322   }
 323 
 324   // Break before main entry point
 325   if ((C->method() && C->directive()->BreakAtExecuteOption) ||
 326       (OptoBreakpoint && C->is_method_compilation())       ||
 327       (OptoBreakpointOSR && C->is_osr_compilation())       ||
 328       (OptoBreakpointC2R && !C->method())                   ) {
 329     // checking for C->method() means that OptoBreakpoint does not apply to
 330     // runtime stubs or frame converters
 331     C->cfg()->insert( entry, 1, new MachBreakpointNode() );
 332   }
 333 
 334   // Insert epilogs before every return
 335   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
 336     Block* block = C->cfg()->get_block(i);
 337     if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
 338       Node* m = block->end();
 339       if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
 340         MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
 341         block->add_inst(epilog);
 342         C->cfg()->map_node_to_block(epilog, block);
 343       }
 344     }
 345   }
 346 
 347   // Keeper of sizing aspects
 348   _buf_sizes = BufferSizingData();
 349 
 350   // Initialize code buffer
 351   estimate_buffer_size(_buf_sizes._const);
 352   if (C->failing()) return;
 353 
 354   // Pre-compute the length of blocks and replace
 355   // long branches with short if machine supports it.
 356   // Must be done before ScheduleAndBundle due to SPARC delay slots
 357   uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
 358   blk_starts[0] = 0;
 359   shorten_branches(blk_starts);
 360 
 361   if (!C->is_osr_compilation() && C->has_scalarized_args()) {
 362     // Compute the offsets of the entry points required by the inline type calling convention
 363     if (!C->method()->is_static()) {
 364       // We have entries at the beginning of the method, implemented by the first 4 nodes.
 365       // Entry                     (unverified) @ offset 0
 366       // Verified_Inline_Entry_RO
 367       // Inline_Entry              (unverified)
 368       // Verified_Inline_Entry
 369       uint offset = 0;
 370       _code_offsets.set_value(CodeOffsets::Entry, offset);
 371 
 372       offset += ((MachVEPNode*)broot->get_node(0))->size(C->regalloc());
 373       _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, offset);
 374 
 375       offset += ((MachVEPNode*)broot->get_node(1))->size(C->regalloc());
 376       _code_offsets.set_value(CodeOffsets::Inline_Entry, offset);
 377 
 378       offset += ((MachVEPNode*)broot->get_node(2))->size(C->regalloc());
 379       _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, offset);
 380     } else {
 381       _code_offsets.set_value(CodeOffsets::Entry, -1); // will be patched later
 382       _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, 0);
 383     }
 384   }
 385 
 386   ScheduleAndBundle();
 387   if (C->failing()) {
 388     return;
 389   }
 390 
 391   perform_mach_node_analysis();
 392 
 393   // Complete sizing of codebuffer
 394   CodeBuffer* cb = init_buffer();
 395   if (cb == nullptr || C->failing()) {
 396     return;
 397   }
 398 
 399   BuildOopMaps();
 400 
 401   if (C->failing())  {
 402     return;
 403   }
 404 
 405   C2_MacroAssembler masm(cb);
 406   fill_buffer(&masm, blk_starts);
 407 }
 408 
 409 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
 410   // Determine if we need to generate a stack overflow check.
 411   // Do it if the method is not a stub function and
 412   // has java calls or has frame size > vm_page_size/8.
 413   // The debug VM checks that deoptimization doesn't trigger an
 414   // unexpected stack overflow (compiled method stack banging should
 415   // guarantee it doesn't happen) so we always need the stack bang in
 416   // a debug VM.
 417   return (C->stub_function() == nullptr &&
 418           (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
 419            DEBUG_ONLY(|| true)));
 420 }
 421 
 422 bool PhaseOutput::need_register_stack_bang() const {
 423   // Determine if we need to generate a register stack overflow check.
 424   // This is only used on architectures which have split register
 425   // and memory stacks.
 426   // Bang if the method is not a stub function and has java calls
 427   return (C->stub_function() == nullptr && C->has_java_calls());
 428 }
 429 
 430 
 431 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
 432 // of a loop. When aligning a loop we need to provide enough instructions
 433 // in cpu's fetch buffer to feed decoders. The loop alignment could be
 434 // avoided if we have enough instructions in fetch buffer at the head of a loop.
 435 // By default, the size is set to 999999 by Block's constructor so that
 436 // a loop will be aligned if the size is not reset here.
 437 //
 438 // Note: Mach instructions could contain several HW instructions
 439 // so the size is estimated only.
 440 //
 441 void PhaseOutput::compute_loop_first_inst_sizes() {
 442   // The next condition is used to gate the loop alignment optimization.
 443   // Don't aligned a loop if there are enough instructions at the head of a loop
 444   // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
 445   // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
 446   // equal to 11 bytes which is the largest address NOP instruction.
 447   if (MaxLoopPad < OptoLoopAlignment - 1) {
 448     uint last_block = C->cfg()->number_of_blocks() - 1;
 449     for (uint i = 1; i <= last_block; i++) {
 450       Block* block = C->cfg()->get_block(i);
 451       // Check the first loop's block which requires an alignment.
 452       if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
 453         uint sum_size = 0;
 454         uint inst_cnt = NumberOfLoopInstrToAlign;
 455         inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
 456 
 457         // Check subsequent fallthrough blocks if the loop's first
 458         // block(s) does not have enough instructions.
 459         Block *nb = block;
 460         while(inst_cnt > 0 &&
 461               i < last_block &&
 462               !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
 463               !nb->has_successor(block)) {
 464           i++;
 465           nb = C->cfg()->get_block(i);
 466           inst_cnt  = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
 467         } // while( inst_cnt > 0 && i < last_block  )
 468 
 469         block->set_first_inst_size(sum_size);
 470       } // f( b->head()->is_Loop() )
 471     } // for( i <= last_block )
 472   } // if( MaxLoopPad < OptoLoopAlignment-1 )
 473 }
 474 
 475 // The architecture description provides short branch variants for some long
 476 // branch instructions. Replace eligible long branches with short branches.
 477 void PhaseOutput::shorten_branches(uint* blk_starts) {
 478 
 479   Compile::TracePhase tp(_t_shortenBranches);
 480 
 481   // Compute size of each block, method size, and relocation information size
 482   uint nblocks  = C->cfg()->number_of_blocks();
 483 
 484   uint*      jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
 485   uint*      jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
 486   int*       jmp_nidx   = NEW_RESOURCE_ARRAY(int ,nblocks);
 487 
 488   // Collect worst case block paddings
 489   int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
 490   memset(block_worst_case_pad, 0, nblocks * sizeof(int));
 491 
 492   DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
 493   DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
 494 
 495   bool has_short_branch_candidate = false;
 496 
 497   // Initialize the sizes to 0
 498   int code_size  = 0;          // Size in bytes of generated code
 499   int stub_size  = 0;          // Size in bytes of all stub entries
 500   // Size in bytes of all relocation entries, including those in local stubs.
 501   // Start with 2-bytes of reloc info for the unvalidated entry point
 502   int reloc_size = 1;          // Number of relocation entries
 503 
 504   // Make three passes.  The first computes pessimistic blk_starts,
 505   // relative jmp_offset and reloc_size information.  The second performs
 506   // short branch substitution using the pessimistic sizing.  The
 507   // third inserts nops where needed.
 508 
 509   // Step one, perform a pessimistic sizing pass.
 510   uint last_call_adr = max_juint;
 511   uint last_avoid_back_to_back_adr = max_juint;
 512   uint nop_size = (new MachNopNode())->size(C->regalloc());
 513   for (uint i = 0; i < nblocks; i++) { // For all blocks
 514     Block* block = C->cfg()->get_block(i);
 515     _block = block;
 516 
 517     // During short branch replacement, we store the relative (to blk_starts)
 518     // offset of jump in jmp_offset, rather than the absolute offset of jump.
 519     // This is so that we do not need to recompute sizes of all nodes when
 520     // we compute correct blk_starts in our next sizing pass.
 521     jmp_offset[i] = 0;
 522     jmp_size[i]   = 0;
 523     jmp_nidx[i]   = -1;
 524     DEBUG_ONLY( jmp_target[i] = 0; )
 525     DEBUG_ONLY( jmp_rule[i]   = 0; )
 526 
 527     // Sum all instruction sizes to compute block size
 528     uint last_inst = block->number_of_nodes();
 529     uint blk_size = 0;
 530     for (uint j = 0; j < last_inst; j++) {
 531       _index = j;
 532       Node* nj = block->get_node(_index);
 533       // Handle machine instruction nodes
 534       if (nj->is_Mach()) {
 535         MachNode* mach = nj->as_Mach();
 536         blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
 537         reloc_size += mach->reloc();
 538         if (mach->is_MachCall()) {
 539           // add size information for trampoline stub
 540           // class CallStubImpl is platform-specific and defined in the *.ad files.
 541           stub_size  += CallStubImpl::size_call_trampoline();
 542           reloc_size += CallStubImpl::reloc_call_trampoline();
 543 
 544           MachCallNode *mcall = mach->as_MachCall();
 545           // This destination address is NOT PC-relative
 546 
 547           if (mcall->entry_point() != nullptr) {
 548             mcall->method_set((intptr_t)mcall->entry_point());
 549           }
 550 
 551           if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
 552             stub_size  += CompiledDirectCall::to_interp_stub_size();
 553             reloc_size += CompiledDirectCall::reloc_to_interp_stub();
 554           }
 555         } else if (mach->is_MachSafePoint()) {
 556           // If call/safepoint are adjacent, account for possible
 557           // nop to disambiguate the two safepoints.
 558           // ScheduleAndBundle() can rearrange nodes in a block,
 559           // check for all offsets inside this block.
 560           if (last_call_adr >= blk_starts[i]) {
 561             blk_size += nop_size;
 562           }
 563         }
 564         if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
 565           // Nop is inserted between "avoid back to back" instructions.
 566           // ScheduleAndBundle() can rearrange nodes in a block,
 567           // check for all offsets inside this block.
 568           if (last_avoid_back_to_back_adr >= blk_starts[i]) {
 569             blk_size += nop_size;
 570           }
 571         }
 572         if (mach->may_be_short_branch()) {
 573           if (!nj->is_MachBranch()) {
 574 #ifndef PRODUCT
 575             nj->dump(3);
 576 #endif
 577             Unimplemented();
 578           }
 579           assert(jmp_nidx[i] == -1, "block should have only one branch");
 580           jmp_offset[i] = blk_size;
 581           jmp_size[i]   = nj->size(C->regalloc());
 582           jmp_nidx[i]   = j;
 583           has_short_branch_candidate = true;
 584         }
 585       }
 586       blk_size += nj->size(C->regalloc());
 587       // Remember end of call offset
 588       if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
 589         last_call_adr = blk_starts[i]+blk_size;
 590       }
 591       // Remember end of avoid_back_to_back offset
 592       if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
 593         last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
 594       }
 595     }
 596 
 597     // When the next block starts a loop, we may insert pad NOP
 598     // instructions.  Since we cannot know our future alignment,
 599     // assume the worst.
 600     if (i < nblocks - 1) {
 601       Block* nb = C->cfg()->get_block(i + 1);
 602       int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
 603       if (max_loop_pad > 0) {
 604         assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
 605         // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
 606         // If either is the last instruction in this block, bump by
 607         // max_loop_pad in lock-step with blk_size, so sizing
 608         // calculations in subsequent blocks still can conservatively
 609         // detect that it may the last instruction in this block.
 610         if (last_call_adr == blk_starts[i]+blk_size) {
 611           last_call_adr += max_loop_pad;
 612         }
 613         if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
 614           last_avoid_back_to_back_adr += max_loop_pad;
 615         }
 616         blk_size += max_loop_pad;
 617         block_worst_case_pad[i + 1] = max_loop_pad;
 618       }
 619     }
 620 
 621     // Save block size; update total method size
 622     blk_starts[i+1] = blk_starts[i]+blk_size;
 623   }
 624 
 625   // Step two, replace eligible long jumps.
 626   bool progress = true;
 627   uint last_may_be_short_branch_adr = max_juint;
 628   while (has_short_branch_candidate && progress) {
 629     progress = false;
 630     has_short_branch_candidate = false;
 631     int adjust_block_start = 0;
 632     for (uint i = 0; i < nblocks; i++) {
 633       Block* block = C->cfg()->get_block(i);
 634       int idx = jmp_nidx[i];
 635       MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach();
 636       if (mach != nullptr && mach->may_be_short_branch()) {
 637 #ifdef ASSERT
 638         assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
 639         int j;
 640         // Find the branch; ignore trailing NOPs.
 641         for (j = block->number_of_nodes()-1; j>=0; j--) {
 642           Node* n = block->get_node(j);
 643           if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
 644             break;
 645         }
 646         assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
 647 #endif
 648         int br_size = jmp_size[i];
 649         int br_offs = blk_starts[i] + jmp_offset[i];
 650 
 651         // This requires the TRUE branch target be in succs[0]
 652         uint bnum = block->non_connector_successor(0)->_pre_order;
 653         int offset = blk_starts[bnum] - br_offs;
 654         if (bnum > i) { // adjust following block's offset
 655           offset -= adjust_block_start;
 656         }
 657 
 658         // This block can be a loop header, account for the padding
 659         // in the previous block.
 660         int block_padding = block_worst_case_pad[i];
 661         assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
 662         // In the following code a nop could be inserted before
 663         // the branch which will increase the backward distance.
 664         bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
 665         assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
 666 
 667         if (needs_padding && offset <= 0)
 668           offset -= nop_size;
 669 
 670         if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
 671           // We've got a winner.  Replace this branch.
 672           MachNode* replacement = mach->as_MachBranch()->short_branch_version();
 673 
 674           // Update the jmp_size.
 675           int new_size = replacement->size(C->regalloc());
 676           int diff     = br_size - new_size;
 677           assert(diff >= (int)nop_size, "short_branch size should be smaller");
 678           // Conservatively take into account padding between
 679           // avoid_back_to_back branches. Previous branch could be
 680           // converted into avoid_back_to_back branch during next
 681           // rounds.
 682           if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
 683             jmp_offset[i] += nop_size;
 684             diff -= nop_size;
 685           }
 686           adjust_block_start += diff;
 687           block->map_node(replacement, idx);
 688           mach->subsume_by(replacement, C);
 689           mach = replacement;
 690           progress = true;
 691 
 692           jmp_size[i] = new_size;
 693           DEBUG_ONLY( jmp_target[i] = bnum; );
 694           DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
 695         } else {
 696           // The jump distance is not short, try again during next iteration.
 697           has_short_branch_candidate = true;
 698         }
 699       } // (mach->may_be_short_branch())
 700       if (mach != nullptr && (mach->may_be_short_branch() ||
 701                            mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
 702         last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
 703       }
 704       blk_starts[i+1] -= adjust_block_start;
 705     }
 706   }
 707 
 708 #ifdef ASSERT
 709   for (uint i = 0; i < nblocks; i++) { // For all blocks
 710     if (jmp_target[i] != 0) {
 711       int br_size = jmp_size[i];
 712       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
 713       if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
 714         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]);
 715       }
 716       assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
 717     }
 718   }
 719 #endif
 720 
 721   // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
 722   // after ScheduleAndBundle().
 723 
 724   // ------------------
 725   // Compute size for code buffer
 726   code_size = blk_starts[nblocks];
 727 
 728   // Relocation records
 729   reloc_size += 1;              // Relo entry for exception handler
 730 
 731   // Adjust reloc_size to number of record of relocation info
 732   // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
 733   // a relocation index.
 734   // The CodeBuffer will expand the locs array if this estimate is too low.
 735   reloc_size *= 10 / sizeof(relocInfo);
 736 
 737   _buf_sizes._reloc = reloc_size;
 738   _buf_sizes._code  = code_size;
 739   _buf_sizes._stub  = stub_size;
 740 }
 741 
 742 //------------------------------FillLocArray-----------------------------------
 743 // Create a bit of debug info and append it to the array.  The mapping is from
 744 // Java local or expression stack to constant, register or stack-slot.  For
 745 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
 746 // entry has been taken care of and caller should skip it).
 747 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
 748   // This should never have accepted Bad before
 749   assert(OptoReg::is_valid(regnum), "location must be valid");
 750   return (OptoReg::is_reg(regnum))
 751          ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
 752          : new LocationValue(Location::new_stk_loc(l_type,  ra->reg2offset(regnum)));
 753 }
 754 
 755 
 756 ObjectValue*
 757 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
 758   for (int i = 0; i < objs->length(); i++) {
 759     assert(objs->at(i)->is_object(), "corrupt object cache");
 760     ObjectValue* sv = objs->at(i)->as_ObjectValue();
 761     if (sv->id() == id) {
 762       return sv;
 763     }
 764   }
 765   // Otherwise..
 766   return nullptr;
 767 }
 768 
 769 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
 770                                      ObjectValue* sv ) {
 771   assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition");
 772   objs->append(sv);
 773 }
 774 
 775 
 776 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
 777                             GrowableArray<ScopeValue*> *array,
 778                             GrowableArray<ScopeValue*> *objs ) {
 779   assert( local, "use _top instead of null" );
 780   if (array->length() != idx) {
 781     assert(array->length() == idx + 1, "Unexpected array count");
 782     // Old functionality:
 783     //   return
 784     // New functionality:
 785     //   Assert if the local is not top. In product mode let the new node
 786     //   override the old entry.
 787     assert(local == C->top(), "LocArray collision");
 788     if (local == C->top()) {
 789       return;
 790     }
 791     array->pop();
 792   }
 793   const Type *t = local->bottom_type();
 794 
 795   // Is it a safepoint scalar object node?
 796   if (local->is_SafePointScalarObject()) {
 797     SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
 798 
 799     ObjectValue* sv = sv_for_node_id(objs, spobj->_idx);
 800     if (sv == nullptr) {
 801       ciKlass* cik = t->is_oopptr()->exact_klass();
 802       assert(cik->is_instance_klass() ||
 803              cik->is_array_klass(), "Not supported allocation.");
 804       uint first_ind = spobj->first_index(sfpt->jvms());
 805       // Nullable, scalarized inline types have a null_marker input
 806       // that needs to be checked before using the field values.
 807       ScopeValue* properties = nullptr;
 808       if (cik->is_inlinetype()) {
 809         Node* null_marker_node = sfpt->in(first_ind++);
 810         assert(null_marker_node != nullptr, "null_marker node not found");
 811         if (!null_marker_node->is_top()) {
 812           const TypeInt* null_marker_type = null_marker_node->bottom_type()->is_int();
 813           if (null_marker_node->is_Con()) {
 814             properties = new ConstantIntValue(null_marker_type->get_con());
 815           } else {
 816             OptoReg::Name null_marker_reg = C->regalloc()->get_reg_first(null_marker_node);
 817             properties = new_loc_value(C->regalloc(), null_marker_reg, Location::normal);
 818           }
 819         }
 820       }
 821       if (cik->is_array_klass() && !cik->is_type_array_klass()) {
 822         jint props = ArrayKlass::ArrayProperties::DEFAULT;
 823         if (cik->as_array_klass()->element_klass()->is_inlinetype()) {
 824           if (cik->as_array_klass()->is_elem_null_free()) {
 825             props |= ArrayKlass::ArrayProperties::NULL_RESTRICTED;
 826           }
 827           if (!cik->as_array_klass()->is_elem_atomic()) {
 828             props |= ArrayKlass::ArrayProperties::NON_ATOMIC;
 829           }
 830         }
 831         properties = new ConstantIntValue(props);
 832       }
 833       sv = new ObjectValue(spobj->_idx,
 834                            new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()), true, properties);
 835       set_sv_for_object_node(objs, sv);
 836 
 837       for (uint i = 0; i < spobj->n_fields(); i++) {
 838         Node* fld_node = sfpt->in(first_ind+i);
 839         (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
 840       }
 841     }
 842     array->append(sv);
 843     return;
 844   } else if (local->is_SafePointScalarMerge()) {
 845     SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
 846     ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
 847 
 848     if (mv == nullptr) {
 849       GrowableArray<ScopeValue*> deps;
 850 
 851       int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
 852       (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
 853       assert(deps.length() == 1, "missing value");
 854 
 855       int selector_idx = smerge->selector_idx(sfpt->jvms());
 856       (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
 857       assert(deps.length() == 2, "missing value");
 858 
 859       mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
 860       set_sv_for_object_node(objs, mv);
 861 
 862       for (uint i = 1; i < smerge->req(); i++) {
 863         Node* obj_node = smerge->in(i);
 864         int idx = mv->possible_objects()->length();
 865         (void)FillLocArray(idx, sfpt, obj_node, mv->possible_objects(), objs);
 866 
 867         // By default ObjectValues that are in 'possible_objects' are not root objects.
 868         // They will be marked as root later if they are directly referenced in a JVMS.
 869         assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
 870         assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
 871         mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
 872       }
 873     }
 874     array->append(mv);
 875     return;
 876   }
 877 
 878   // Grab the register number for the local
 879   OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
 880   if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
 881     // Record the double as two float registers.
 882     // The register mask for such a value always specifies two adjacent
 883     // float registers, with the lower register number even.
 884     // Normally, the allocation of high and low words to these registers
 885     // is irrelevant, because nearly all operations on register pairs
 886     // (e.g., StoreD) treat them as a single unit.
 887     // Here, we assume in addition that the words in these two registers
 888     // stored "naturally" (by operations like StoreD and double stores
 889     // within the interpreter) such that the lower-numbered register
 890     // is written to the lower memory address.  This may seem like
 891     // a machine dependency, but it is not--it is a requirement on
 892     // the author of the <arch>.ad file to ensure that, for every
 893     // even/odd double-register pair to which a double may be allocated,
 894     // the word in the even single-register is stored to the first
 895     // memory word.  (Note that register numbers are completely
 896     // arbitrary, and are not tied to any machine-level encodings.)
 897 #ifdef _LP64
 898     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
 899       array->append(new ConstantIntValue((jint)0));
 900       array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
 901     } else if ( t->base() == Type::Long ) {
 902       array->append(new ConstantIntValue((jint)0));
 903       array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
 904     } else if ( t->base() == Type::RawPtr ) {
 905       // jsr/ret return address which must be restored into the full
 906       // width 64-bit stack slot.
 907       array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
 908     }
 909 #else //_LP64
 910     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
 911       // Repack the double/long as two jints.
 912       // The convention the interpreter uses is that the second local
 913       // holds the first raw word of the native double representation.
 914       // This is actually reasonable, since locals and stack arrays
 915       // grow downwards in all implementations.
 916       // (If, on some machine, the interpreter's Java locals or stack
 917       // were to grow upwards, the embedded doubles would be word-swapped.)
 918       array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
 919       array->append(new_loc_value( C->regalloc(),              regnum   , Location::normal ));
 920     }
 921 #endif //_LP64
 922     else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
 923              OptoReg::is_reg(regnum) ) {
 924       array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
 925                                                       ? Location::float_in_dbl : Location::normal ));
 926     } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
 927       array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
 928                                                       ? Location::int_in_long : Location::normal ));
 929     } else if( t->base() == Type::NarrowOop ) {
 930       array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
 931     } else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
 932                t->base() == Type::VectorD || t->base() == Type::VectorX ||
 933                t->base() == Type::VectorY || t->base() == Type::VectorZ) {
 934       array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
 935     } else if (C->regalloc()->is_oop(local)) {
 936       assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
 937              t->base() == Type::AryPtr,
 938              "Unexpected type: %s", t->msg());
 939       array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
 940     } else {
 941       assert(t->base() == Type::Int || t->base() == Type::Half ||
 942              t->base() == Type::FloatCon || t->base() == Type::FloatBot,
 943              "Unexpected type: %s", t->msg());
 944       array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
 945     }
 946     return;
 947   }
 948 
 949   // No register.  It must be constant data.
 950   switch (t->base()) {
 951     case Type::Half:              // Second half of a double
 952       ShouldNotReachHere();       // Caller should skip 2nd halves
 953       break;
 954     case Type::AnyPtr:
 955       array->append(new ConstantOopWriteValue(nullptr));
 956       break;
 957     case Type::AryPtr:
 958     case Type::InstPtr:          // fall through
 959       array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
 960       break;
 961     case Type::NarrowOop:
 962       if (t == TypeNarrowOop::NULL_PTR) {
 963         array->append(new ConstantOopWriteValue(nullptr));
 964       } else {
 965         array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
 966       }
 967       break;
 968     case Type::Int:
 969       array->append(new ConstantIntValue(t->is_int()->get_con()));
 970       break;
 971     case Type::RawPtr:
 972       // A return address (T_ADDRESS).
 973       assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
 974 #ifdef _LP64
 975       // Must be restored to the full-width 64-bit stack slot.
 976       array->append(new ConstantLongValue(t->is_ptr()->get_con()));
 977 #else
 978       array->append(new ConstantIntValue(t->is_ptr()->get_con()));
 979 #endif
 980       break;
 981     case Type::FloatCon: {
 982       float f = t->is_float_constant()->getf();
 983       array->append(new ConstantIntValue(jint_cast(f)));
 984       break;
 985     }
 986     case Type::DoubleCon: {
 987       jdouble d = t->is_double_constant()->getd();
 988 #ifdef _LP64
 989       array->append(new ConstantIntValue((jint)0));
 990       array->append(new ConstantDoubleValue(d));
 991 #else
 992       // Repack the double as two jints.
 993     // The convention the interpreter uses is that the second local
 994     // holds the first raw word of the native double representation.
 995     // This is actually reasonable, since locals and stack arrays
 996     // grow downwards in all implementations.
 997     // (If, on some machine, the interpreter's Java locals or stack
 998     // were to grow upwards, the embedded doubles would be word-swapped.)
 999     jlong_accessor acc;
1000     acc.long_value = jlong_cast(d);
1001     array->append(new ConstantIntValue(acc.words[1]));
1002     array->append(new ConstantIntValue(acc.words[0]));
1003 #endif
1004       break;
1005     }
1006     case Type::Long: {
1007       jlong d = t->is_long()->get_con();
1008 #ifdef _LP64
1009       array->append(new ConstantIntValue((jint)0));
1010       array->append(new ConstantLongValue(d));
1011 #else
1012       // Repack the long as two jints.
1013     // The convention the interpreter uses is that the second local
1014     // holds the first raw word of the native double representation.
1015     // This is actually reasonable, since locals and stack arrays
1016     // grow downwards in all implementations.
1017     // (If, on some machine, the interpreter's Java locals or stack
1018     // were to grow upwards, the embedded doubles would be word-swapped.)
1019     jlong_accessor acc;
1020     acc.long_value = d;
1021     array->append(new ConstantIntValue(acc.words[1]));
1022     array->append(new ConstantIntValue(acc.words[0]));
1023 #endif
1024       break;
1025     }
1026     case Type::Top:               // Add an illegal value here
1027       array->append(new LocationValue(Location()));
1028       break;
1029     default:
1030       ShouldNotReachHere();
1031       break;
1032   }
1033 }
1034 
1035 // Determine if this node starts a bundle
1036 bool PhaseOutput::starts_bundle(const Node *n) const {
1037   return (_node_bundling_limit > n->_idx &&
1038           _node_bundling_base[n->_idx].starts_bundle());
1039 }
1040 
1041 // Determine if there is a monitor that has 'ov' as its owner.
1042 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const {
1043   for (int k = 0; k < monarray->length(); k++) {
1044     MonitorValue* mv = monarray->at(k);
1045     if (mv->owner() == ov) {
1046       return true;
1047     }
1048   }
1049 
1050   return false;
1051 }
1052 
1053 // Determine if there is a scalar replaced object description represented by 'ov'.
1054 bool PhaseOutput::contains_as_scalarized_obj(JVMState* jvms, MachSafePointNode* sfn,
1055                                              GrowableArray<ScopeValue*>* objs,
1056                                              ObjectValue* ov) const {
1057   for (int i = 0; i < jvms->scl_size(); i++) {
1058     Node* n = sfn->scalarized_obj(jvms, i);
1059     // Other kinds of nodes that we may encounter here, for instance constants
1060     // representing values of fields of objects scalarized, aren't relevant for
1061     // us, since they don't map to ObjectValue.
1062     if (!n->is_SafePointScalarObject()) {
1063       continue;
1064     }
1065 
1066     ObjectValue* other = sv_for_node_id(objs, n->_idx);
1067     if (ov == other) {
1068       return true;
1069     }
1070   }
1071   return false;
1072 }
1073 
1074 //--------------------------Process_OopMap_Node--------------------------------
1075 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
1076   // Handle special safepoint nodes for synchronization
1077   MachSafePointNode *sfn   = mach->as_MachSafePoint();
1078   MachCallNode      *mcall;
1079 
1080   int safepoint_pc_offset = current_offset;
1081   bool is_method_handle_invoke = false;
1082   bool return_oop = false;
1083   bool return_scalarized = false;
1084   bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
1085   bool arg_escape = false;
1086 
1087   // Add the safepoint in the DebugInfoRecorder
1088   if( !mach->is_MachCall() ) {
1089     mcall = nullptr;
1090     C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1091   } else {
1092     mcall = mach->as_MachCall();
1093 
1094     // Is the call a MethodHandle call?
1095     if (mcall->is_MachCallJava()) {
1096       if (mcall->as_MachCallJava()->_method_handle_invoke) {
1097         assert(C->has_method_handle_invokes(), "must have been set during call generation");
1098         is_method_handle_invoke = true;
1099       }
1100       arg_escape = mcall->as_MachCallJava()->_arg_escape;
1101     }
1102 
1103     // Check if a call returns an object.
1104     if (mcall->returns_pointer() || mcall->returns_scalarized()) {
1105       return_oop = true;
1106     }
1107     if (mcall->returns_scalarized()) {
1108       return_scalarized = true;
1109     }
1110     safepoint_pc_offset += mcall->ret_addr_offset();
1111     C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1112   }
1113 
1114   // Loop over the JVMState list to add scope information
1115   // Do not skip safepoints with a null method, they need monitor info
1116   JVMState* youngest_jvms = sfn->jvms();
1117   int max_depth = youngest_jvms->depth();
1118 
1119   // Allocate the object pool for scalar-replaced objects -- the map from
1120   // small-integer keys (which can be recorded in the local and ostack
1121   // arrays) to descriptions of the object state.
1122   GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1123 
1124   // Visit scopes from oldest to youngest.
1125   for (int depth = 1; depth <= max_depth; depth++) {
1126     JVMState* jvms = youngest_jvms->of_depth(depth);
1127     int idx;
1128     ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1129     // Safepoints that do not have method() set only provide oop-map and monitor info
1130     // to support GC; these do not support deoptimization.
1131     int num_locs = (method == nullptr) ? 0 : jvms->loc_size();
1132     int num_exps = (method == nullptr) ? 0 : jvms->stk_size();
1133     int num_mon  = jvms->nof_monitors();
1134     assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(),
1135            "JVMS local count must match that of the method");
1136 
1137     // Add Local and Expression Stack Information
1138 
1139     // Insert locals into the locarray
1140     GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1141     for( idx = 0; idx < num_locs; idx++ ) {
1142       FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1143     }
1144 
1145     // Insert expression stack entries into the exparray
1146     GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1147     for( idx = 0; idx < num_exps; idx++ ) {
1148       FillLocArray( idx,  sfn, sfn->stack(jvms, idx), exparray, objs );
1149     }
1150 
1151     // Add in mappings of the monitors
1152     assert( !method ||
1153             !method->is_synchronized() ||
1154             method->is_native() ||
1155             num_mon > 0 ||
1156             !GenerateSynchronizationCode,
1157             "monitors must always exist for synchronized methods");
1158 
1159     // Build the growable array of ScopeValues for exp stack
1160     GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1161 
1162     // Loop over monitors and insert into array
1163     for (idx = 0; idx < num_mon; idx++) {
1164       // Grab the node that defines this monitor
1165       Node* box_node = sfn->monitor_box(jvms, idx);
1166       Node* obj_node = sfn->monitor_obj(jvms, idx);
1167 
1168       // Create ScopeValue for object
1169       ScopeValue *scval = nullptr;
1170 
1171       if (obj_node->is_SafePointScalarObject()) {
1172         SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1173         scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1174         if (scval == nullptr) {
1175           const Type *t = spobj->bottom_type();
1176           ciKlass* cik = t->is_oopptr()->exact_klass();
1177           assert(cik->is_instance_klass() ||
1178                  cik->is_array_klass(), "Not supported allocation.");
1179           assert(!cik->is_inlinetype(), "Synchronization on value object?");
1180           ScopeValue* properties = nullptr;
1181           if (cik->is_array_klass() && !cik->is_type_array_klass()) {
1182             jint props = ArrayKlass::ArrayProperties::DEFAULT;
1183             if (cik->as_array_klass()->element_klass()->is_inlinetype()) {
1184               if (cik->as_array_klass()->is_elem_null_free()) {
1185                 props |= ArrayKlass::ArrayProperties::NULL_RESTRICTED;
1186               }
1187               if (!cik->as_array_klass()->is_elem_atomic()) {
1188                 props |= ArrayKlass::ArrayProperties::NON_ATOMIC;
1189               }
1190             }
1191             properties = new ConstantIntValue(props);
1192           }
1193           ObjectValue* sv = new ObjectValue(spobj->_idx,
1194                                             new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()), true, properties);
1195           PhaseOutput::set_sv_for_object_node(objs, sv);
1196 
1197           uint first_ind = spobj->first_index(youngest_jvms);
1198           for (uint i = 0; i < spobj->n_fields(); i++) {
1199             Node* fld_node = sfn->in(first_ind+i);
1200             (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1201           }
1202           scval = sv;
1203         }
1204       } else if (obj_node->is_SafePointScalarMerge()) {
1205         SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1206         ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1207 
1208         if (mv == nullptr) {
1209           GrowableArray<ScopeValue*> deps;
1210 
1211           int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1212           FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1213           assert(deps.length() == 1, "missing value");
1214 
1215           int selector_idx = smerge->selector_idx(youngest_jvms);
1216           FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr);
1217           assert(deps.length() == 2, "missing value");
1218 
1219           mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
1220           set_sv_for_object_node(objs, mv);
1221 
1222           for (uint i = 1; i < smerge->req(); i++) {
1223             Node* obj_node = smerge->in(i);
1224             int idx = mv->possible_objects()->length();
1225             (void)FillLocArray(idx, sfn, obj_node, mv->possible_objects(), objs);
1226 
1227             // By default ObjectValues that are in 'possible_objects' are not root objects.
1228             // They will be marked as root later if they are directly referenced in a JVMS.
1229             assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
1230             assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
1231             mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
1232           }
1233         }
1234         scval = mv;
1235       } else if (!obj_node->is_Con()) {
1236         OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1237         if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1238           scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1239         } else {
1240           scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1241         }
1242       } else {
1243         const TypePtr *tp = obj_node->get_ptr_type();
1244         scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1245       }
1246 
1247       OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1248       Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1249       bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1250       monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1251     }
1252 
1253     // Mark ObjectValue nodes as root nodes if they are directly
1254     // referenced in the JVMS.
1255     for (int i = 0; i < objs->length(); i++) {
1256       ScopeValue* sv = objs->at(i);
1257       if (sv->is_object_merge()) {
1258         ObjectMergeValue* merge = sv->as_ObjectMergeValue();
1259 
1260         for (int j = 0; j< merge->possible_objects()->length(); j++) {
1261           ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue();
1262           if (ov->is_root()) {
1263             // Already flagged as 'root' by something else. We shouldn't change it
1264             // to non-root in a younger JVMS because it may need to be alive in
1265             // a younger JVMS.
1266           } else {
1267             bool is_root = locarray->contains(ov) ||
1268                            exparray->contains(ov) ||
1269                            contains_as_owner(monarray, ov) ||
1270                            contains_as_scalarized_obj(jvms, sfn, objs, ov);
1271             ov->set_root(is_root);
1272           }
1273         }
1274       }
1275     }
1276 
1277     // We dump the object pool first, since deoptimization reads it in first.
1278     C->debug_info()->dump_object_pool(objs);
1279 
1280     // Build first class objects to pass to scope
1281     DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1282     DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1283     DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1284 
1285     // Make method available for all Safepoints
1286     ciMethod* scope_method = method ? method : C->method();
1287     // Describe the scope here
1288     assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1289     assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1290     // Now we can describe the scope.
1291     methodHandle null_mh;
1292     bool rethrow_exception = false;
1293     C->debug_info()->describe_scope(
1294       safepoint_pc_offset,
1295       null_mh,
1296       scope_method,
1297       jvms->bci(),
1298       jvms->should_reexecute(),
1299       rethrow_exception,
1300       is_method_handle_invoke,
1301       return_oop,
1302       return_scalarized,
1303       has_ea_local_in_scope,
1304       arg_escape,
1305       locvals,
1306       expvals,
1307       monvals
1308     );
1309   } // End jvms loop
1310 
1311   // Mark the end of the scope set.
1312   C->debug_info()->end_safepoint(safepoint_pc_offset);
1313 }
1314 
1315 
1316 
1317 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1318 class NonSafepointEmitter {
1319     Compile*  C;
1320     JVMState* _pending_jvms;
1321     int       _pending_offset;
1322 
1323     void emit_non_safepoint();
1324 
1325  public:
1326     NonSafepointEmitter(Compile* compile) {
1327       this->C = compile;
1328       _pending_jvms = nullptr;
1329       _pending_offset = 0;
1330     }
1331 
1332     void observe_instruction(Node* n, int pc_offset) {
1333       if (!C->debug_info()->recording_non_safepoints())  return;
1334 
1335       Node_Notes* nn = C->node_notes_at(n->_idx);
1336       if (nn == nullptr || nn->jvms() == nullptr)  return;
1337       if (_pending_jvms != nullptr &&
1338           _pending_jvms->same_calls_as(nn->jvms())) {
1339         // Repeated JVMS?  Stretch it up here.
1340         _pending_offset = pc_offset;
1341       } else {
1342         if (_pending_jvms != nullptr &&
1343             _pending_offset < pc_offset) {
1344           emit_non_safepoint();
1345         }
1346         _pending_jvms = nullptr;
1347         if (pc_offset > C->debug_info()->last_pc_offset()) {
1348           // This is the only way _pending_jvms can become non-null:
1349           _pending_jvms = nn->jvms();
1350           _pending_offset = pc_offset;
1351         }
1352       }
1353     }
1354 
1355     // Stay out of the way of real safepoints:
1356     void observe_safepoint(JVMState* jvms, int pc_offset) {
1357       if (_pending_jvms != nullptr &&
1358           !_pending_jvms->same_calls_as(jvms) &&
1359           _pending_offset < pc_offset) {
1360         emit_non_safepoint();
1361       }
1362       _pending_jvms = nullptr;
1363     }
1364 
1365     void flush_at_end() {
1366       if (_pending_jvms != nullptr) {
1367         emit_non_safepoint();
1368       }
1369       _pending_jvms = nullptr;
1370     }
1371 };
1372 
1373 void NonSafepointEmitter::emit_non_safepoint() {
1374   JVMState* youngest_jvms = _pending_jvms;
1375   int       pc_offset     = _pending_offset;
1376 
1377   // Clear it now:
1378   _pending_jvms = nullptr;
1379 
1380   DebugInformationRecorder* debug_info = C->debug_info();
1381   assert(debug_info->recording_non_safepoints(), "sanity");
1382 
1383   debug_info->add_non_safepoint(pc_offset);
1384   int max_depth = youngest_jvms->depth();
1385 
1386   // Visit scopes from oldest to youngest.
1387   for (int depth = 1; depth <= max_depth; depth++) {
1388     JVMState* jvms = youngest_jvms->of_depth(depth);
1389     ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1390     assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1391     methodHandle null_mh;
1392     debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1393   }
1394 
1395   // Mark the end of the scope set.
1396   debug_info->end_non_safepoint(pc_offset);
1397 }
1398 
1399 //------------------------------init_buffer------------------------------------
1400 void PhaseOutput::estimate_buffer_size(int& const_req) {
1401 
1402   // Set the initially allocated size
1403   const_req = initial_const_capacity;
1404 
1405   // The extra spacing after the code is necessary on some platforms.
1406   // Sometimes we need to patch in a jump after the last instruction,
1407   // if the nmethod has been deoptimized.  (See 4932387, 4894843.)
1408 
1409   // Compute the byte offset where we can store the deopt pc.
1410   if (C->fixed_slots() != 0) {
1411     _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1412   }
1413 
1414   // Compute prolog code size
1415   _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1416   assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1417 
1418   if (C->has_mach_constant_base_node()) {
1419     uint add_size = 0;
1420     // Fill the constant table.
1421     // Note:  This must happen before shorten_branches.
1422     for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1423       Block* b = C->cfg()->get_block(i);
1424 
1425       for (uint j = 0; j < b->number_of_nodes(); j++) {
1426         Node* n = b->get_node(j);
1427 
1428         // If the node is a MachConstantNode evaluate the constant
1429         // value section.
1430         if (n->is_MachConstant()) {
1431           MachConstantNode* machcon = n->as_MachConstant();
1432           machcon->eval_constant(C);
1433         } else if (n->is_Mach()) {
1434           // On Power there are more nodes that issue constants.
1435           add_size += (n->as_Mach()->ins_num_consts() * 8);
1436         }
1437       }
1438     }
1439 
1440     // Calculate the offsets of the constants and the size of the
1441     // constant table (including the padding to the next section).
1442     constant_table().calculate_offsets_and_size();
1443     const_req = constant_table().alignment() + constant_table().size() + add_size;
1444   }
1445 
1446   // Initialize the space for the BufferBlob used to find and verify
1447   // instruction size in MachNode::emit_size()
1448   init_scratch_buffer_blob(const_req);
1449 }
1450 
1451 CodeBuffer* PhaseOutput::init_buffer() {
1452   int stub_req  = _buf_sizes._stub;
1453   int code_req  = _buf_sizes._code;
1454   int const_req = _buf_sizes._const;
1455 
1456   int pad_req   = NativeCall::byte_size();
1457 
1458   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1459   stub_req += bs->estimate_stub_size();
1460 
1461   // nmethod and CodeBuffer count stubs & constants as part of method's code.
1462   // class HandlerImpl is platform-specific and defined in the *.ad files.
1463   int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1464   int deopt_handler_req     = HandlerImpl::size_deopt_handler()     + MAX_stubs_size; // add marginal slop for handler
1465   stub_req += MAX_stubs_size;   // ensure per-stub margin
1466   code_req += MAX_inst_size;    // ensure per-instruction margin
1467 
1468   if (StressCodeBuffers)
1469     code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10;  // force expansion
1470 
1471   int total_req =
1472           const_req +
1473           code_req +
1474           pad_req +
1475           stub_req +
1476           exception_handler_req +
1477           deopt_handler_req;               // deopt handler
1478 
1479   if (C->has_method_handle_invokes())
1480     total_req += deopt_handler_req;  // deopt MH handler
1481 
1482   CodeBuffer* cb = code_buffer();
1483   cb->set_const_section_alignment(constant_table().alignment());
1484   cb->initialize(total_req, _buf_sizes._reloc);
1485 
1486   // Have we run out of code space?
1487   if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1488     C->record_failure("CodeCache is full");
1489     return nullptr;
1490   }
1491   // Configure the code buffer.
1492   cb->initialize_consts_size(const_req);
1493   cb->initialize_stubs_size(stub_req);
1494   cb->initialize_oop_recorder(C->env()->oop_recorder());
1495 
1496   // fill in the nop array for bundling computations
1497   MachNode *_nop_list[Bundle::_nop_count];
1498   Bundle::initialize_nops(_nop_list);
1499 
1500   return cb;
1501 }
1502 
1503 //------------------------------fill_buffer------------------------------------
1504 void PhaseOutput::fill_buffer(C2_MacroAssembler* masm, uint* blk_starts) {
1505   // blk_starts[] contains offsets calculated during short branches processing,
1506   // offsets should not be increased during following steps.
1507 
1508   // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1509   // of a loop. It is used to determine the padding for loop alignment.
1510   Compile::TracePhase tp(_t_fillBuffer);
1511 
1512   compute_loop_first_inst_sizes();
1513 
1514   // Create oopmap set.
1515   _oop_map_set = new OopMapSet();
1516 
1517   // !!!!! This preserves old handling of oopmaps for now
1518   C->debug_info()->set_oopmaps(_oop_map_set);
1519 
1520   uint nblocks  = C->cfg()->number_of_blocks();
1521   // Count and start of implicit null check instructions
1522   uint inct_cnt = 0;
1523   uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1524 
1525   // Count and start of calls
1526   uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1527 
1528   uint  return_offset = 0;
1529   int nop_size = (new MachNopNode())->size(C->regalloc());
1530 
1531   int previous_offset = 0;
1532   int current_offset  = 0;
1533   int last_call_offset = -1;
1534   int last_avoid_back_to_back_offset = -1;
1535 #ifdef ASSERT
1536   uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1537   uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1538   uint* jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
1539   uint* jmp_rule   = NEW_RESOURCE_ARRAY(uint,nblocks);
1540 #endif
1541 
1542   // Create an array of unused labels, one for each basic block, if printing is enabled
1543 #if defined(SUPPORT_OPTO_ASSEMBLY)
1544   int* node_offsets      = nullptr;
1545   uint node_offset_limit = C->unique();
1546 
1547   if (C->print_assembly()) {
1548     node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1549   }
1550   if (node_offsets != nullptr) {
1551     // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1552     memset(node_offsets, 0, node_offset_limit*sizeof(int));
1553   }
1554 #endif
1555 
1556   NonSafepointEmitter non_safepoints(C);  // emit non-safepoints lazily
1557 
1558   // Emit the constant table.
1559   if (C->has_mach_constant_base_node()) {
1560     if (!constant_table().emit(masm)) {
1561       C->record_failure("consts section overflow");
1562       return;
1563     }
1564   }
1565 
1566   // Create an array of labels, one for each basic block
1567   Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1568   for (uint i = 0; i <= nblocks; i++) {
1569     blk_labels[i].init();
1570   }
1571 
1572   // Now fill in the code buffer
1573   Node* delay_slot = nullptr;
1574   for (uint i = 0; i < nblocks; i++) {
1575     Block* block = C->cfg()->get_block(i);
1576     _block = block;
1577     Node* head = block->head();
1578 
1579     // If this block needs to start aligned (i.e, can be reached other
1580     // than by falling-thru from the previous block), then force the
1581     // start of a new bundle.
1582     if (Pipeline::requires_bundling() && starts_bundle(head)) {
1583       masm->code()->flush_bundle(true);
1584     }
1585 
1586 #ifdef ASSERT
1587     if (!block->is_connector()) {
1588       stringStream st;
1589       block->dump_head(C->cfg(), &st);
1590       masm->block_comment(st.freeze());
1591     }
1592     jmp_target[i] = 0;
1593     jmp_offset[i] = 0;
1594     jmp_size[i]   = 0;
1595     jmp_rule[i]   = 0;
1596 #endif
1597     int blk_offset = current_offset;
1598 
1599     // Define the label at the beginning of the basic block
1600     masm->bind(blk_labels[block->_pre_order]);
1601 
1602     uint last_inst = block->number_of_nodes();
1603 
1604     // Emit block normally, except for last instruction.
1605     // Emit means "dump code bits into code buffer".
1606     for (uint j = 0; j<last_inst; j++) {
1607       _index = j;
1608 
1609       // Get the node
1610       Node* n = block->get_node(j);
1611 
1612       // See if delay slots are supported
1613       if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) {
1614         assert(delay_slot == nullptr, "no use of delay slot node");
1615         assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1616 
1617         delay_slot = n;
1618         continue;
1619       }
1620 
1621       // If this starts a new instruction group, then flush the current one
1622       // (but allow split bundles)
1623       if (Pipeline::requires_bundling() && starts_bundle(n))
1624         masm->code()->flush_bundle(false);
1625 
1626       // Special handling for SafePoint/Call Nodes
1627       bool is_mcall = false;
1628       if (n->is_Mach()) {
1629         MachNode *mach = n->as_Mach();
1630         is_mcall = n->is_MachCall();
1631         bool is_sfn = n->is_MachSafePoint();
1632 
1633         // If this requires all previous instructions be flushed, then do so
1634         if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1635           masm->code()->flush_bundle(true);
1636           current_offset = masm->offset();
1637         }
1638 
1639         // A padding may be needed again since a previous instruction
1640         // could be moved to delay slot.
1641 
1642         // align the instruction if necessary
1643         int padding = mach->compute_padding(current_offset);
1644         // Make sure safepoint node for polling is distinct from a call's
1645         // return by adding a nop if needed.
1646         if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1647           padding = nop_size;
1648         }
1649         if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1650             current_offset == last_avoid_back_to_back_offset) {
1651           // Avoid back to back some instructions.
1652           padding = nop_size;
1653         }
1654 
1655         if (padding > 0) {
1656           assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1657           int nops_cnt = padding / nop_size;
1658           MachNode *nop = new MachNopNode(nops_cnt);
1659           block->insert_node(nop, j++);
1660           last_inst++;
1661           C->cfg()->map_node_to_block(nop, block);
1662           // Ensure enough space.
1663           masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1664           if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1665             C->record_failure("CodeCache is full");
1666             return;
1667           }
1668           nop->emit(masm, C->regalloc());
1669           masm->code()->flush_bundle(true);
1670           current_offset = masm->offset();
1671         }
1672 
1673         bool observe_safepoint = is_sfn;
1674         // Remember the start of the last call in a basic block
1675         if (is_mcall) {
1676           MachCallNode *mcall = mach->as_MachCall();
1677 
1678           if (mcall->entry_point() != nullptr) {
1679             // This destination address is NOT PC-relative
1680             mcall->method_set((intptr_t)mcall->entry_point());
1681           }
1682 
1683           // Save the return address
1684           call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1685 
1686           observe_safepoint = mcall->guaranteed_safepoint();
1687         }
1688 
1689         // sfn will be valid whenever mcall is valid now because of inheritance
1690         if (observe_safepoint) {
1691           // Handle special safepoint nodes for synchronization
1692           if (!is_mcall) {
1693             MachSafePointNode *sfn = mach->as_MachSafePoint();
1694             // !!!!! Stubs only need an oopmap right now, so bail out
1695             if (sfn->jvms()->method() == nullptr) {
1696               // Write the oopmap directly to the code blob??!!
1697               continue;
1698             }
1699           } // End synchronization
1700 
1701           non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1702                                            current_offset);
1703           Process_OopMap_Node(mach, current_offset);
1704         } // End if safepoint
1705 
1706           // If this is a null check, then add the start of the previous instruction to the list
1707         else if( mach->is_MachNullCheck() ) {
1708           inct_starts[inct_cnt++] = previous_offset;
1709         }
1710 
1711           // If this is a branch, then fill in the label with the target BB's label
1712         else if (mach->is_MachBranch()) {
1713           // This requires the TRUE branch target be in succs[0]
1714           uint block_num = block->non_connector_successor(0)->_pre_order;
1715 
1716           // Try to replace long branch if delay slot is not used,
1717           // it is mostly for back branches since forward branch's
1718           // distance is not updated yet.
1719           bool delay_slot_is_used = valid_bundle_info(n) &&
1720                                     C->output()->node_bundling(n)->use_unconditional_delay();
1721           if (!delay_slot_is_used && mach->may_be_short_branch()) {
1722             assert(delay_slot == nullptr, "not expecting delay slot node");
1723             int br_size = n->size(C->regalloc());
1724             int offset = blk_starts[block_num] - current_offset;
1725             if (block_num >= i) {
1726               // Current and following block's offset are not
1727               // finalized yet, adjust distance by the difference
1728               // between calculated and final offsets of current block.
1729               offset -= (blk_starts[i] - blk_offset);
1730             }
1731             // In the following code a nop could be inserted before
1732             // the branch which will increase the backward distance.
1733             bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1734             if (needs_padding && offset <= 0)
1735               offset -= nop_size;
1736 
1737             if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1738               // We've got a winner.  Replace this branch.
1739               MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1740 
1741               // Update the jmp_size.
1742               int new_size = replacement->size(C->regalloc());
1743               assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1744               // Insert padding between avoid_back_to_back branches.
1745               if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1746                 MachNode *nop = new MachNopNode();
1747                 block->insert_node(nop, j++);
1748                 C->cfg()->map_node_to_block(nop, block);
1749                 last_inst++;
1750                 nop->emit(masm, C->regalloc());
1751                 masm->code()->flush_bundle(true);
1752                 current_offset = masm->offset();
1753               }
1754 #ifdef ASSERT
1755               jmp_target[i] = block_num;
1756               jmp_offset[i] = current_offset - blk_offset;
1757               jmp_size[i]   = new_size;
1758               jmp_rule[i]   = mach->rule();
1759 #endif
1760               block->map_node(replacement, j);
1761               mach->subsume_by(replacement, C);
1762               n    = replacement;
1763               mach = replacement;
1764             }
1765           }
1766           mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1767         } else if (mach->ideal_Opcode() == Op_Jump) {
1768           for (uint h = 0; h < block->_num_succs; h++) {
1769             Block* succs_block = block->_succs[h];
1770             for (uint j = 1; j < succs_block->num_preds(); j++) {
1771               Node* jpn = succs_block->pred(j);
1772               if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1773                 uint block_num = succs_block->non_connector()->_pre_order;
1774                 Label *blkLabel = &blk_labels[block_num];
1775                 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1776               }
1777             }
1778           }
1779         } else if (!n->is_Proj()) {
1780           // Remember the beginning of the previous instruction, in case
1781           // it's followed by a flag-kill and a null-check.  Happens on
1782           // Intel all the time, with add-to-memory kind of opcodes.
1783           previous_offset = current_offset;
1784         }
1785 
1786         // Not an else-if!
1787         // If this is a trap based cmp then add its offset to the list.
1788         if (mach->is_TrapBasedCheckNode()) {
1789           inct_starts[inct_cnt++] = current_offset;
1790         }
1791       }
1792 
1793       // Verify that there is sufficient space remaining
1794       masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1795       if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1796         C->record_failure("CodeCache is full");
1797         return;
1798       }
1799 
1800       // Save the offset for the listing
1801 #if defined(SUPPORT_OPTO_ASSEMBLY)
1802       if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1803         node_offsets[n->_idx] = masm->offset();
1804       }
1805 #endif
1806       assert(!C->failing_internal() || C->failure_is_artificial(), "Should not reach here if failing.");
1807 
1808       // "Normal" instruction case
1809       DEBUG_ONLY(uint instr_offset = masm->offset());
1810       n->emit(masm, C->regalloc());
1811       current_offset = masm->offset();
1812 
1813       // Above we only verified that there is enough space in the instruction section.
1814       // However, the instruction may emit stubs that cause code buffer expansion.
1815       // Bail out here if expansion failed due to a lack of code cache space.
1816       if (C->failing()) {
1817         return;
1818       }
1819 
1820       assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1821              "ret_addr_offset() not within emitted code");
1822 #ifdef ASSERT
1823       uint n_size = n->size(C->regalloc());
1824       if (n_size < (current_offset-instr_offset)) {
1825         MachNode* mach = n->as_Mach();
1826         n->dump();
1827         mach->dump_format(C->regalloc(), tty);
1828         tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1829         Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty);
1830         tty->print_cr(" ------------------- ");
1831         BufferBlob* blob = this->scratch_buffer_blob();
1832         address blob_begin = blob->content_begin();
1833         Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1834         assert(false, "wrong size of mach node");
1835       }
1836 #endif
1837       non_safepoints.observe_instruction(n, current_offset);
1838 
1839       // mcall is last "call" that can be a safepoint
1840       // record it so we can see if a poll will directly follow it
1841       // in which case we'll need a pad to make the PcDesc sites unique
1842       // see  5010568. This can be slightly inaccurate but conservative
1843       // in the case that return address is not actually at current_offset.
1844       // This is a small price to pay.
1845 
1846       if (is_mcall) {
1847         last_call_offset = current_offset;
1848       }
1849 
1850       if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1851         // Avoid back to back some instructions.
1852         last_avoid_back_to_back_offset = current_offset;
1853       }
1854 
1855       // See if this instruction has a delay slot
1856       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1857         guarantee(delay_slot != nullptr, "expecting delay slot node");
1858 
1859         // Back up 1 instruction
1860         masm->code()->set_insts_end(masm->code()->insts_end() - Pipeline::instr_unit_size());
1861 
1862         // Save the offset for the listing
1863 #if defined(SUPPORT_OPTO_ASSEMBLY)
1864         if ((node_offsets != nullptr) && (delay_slot->_idx < node_offset_limit)) {
1865           node_offsets[delay_slot->_idx] = masm->offset();
1866         }
1867 #endif
1868 
1869         // Support a SafePoint in the delay slot
1870         if (delay_slot->is_MachSafePoint()) {
1871           MachNode *mach = delay_slot->as_Mach();
1872           // !!!!! Stubs only need an oopmap right now, so bail out
1873           if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == nullptr) {
1874             // Write the oopmap directly to the code blob??!!
1875             delay_slot = nullptr;
1876             continue;
1877           }
1878 
1879           int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1880           non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1881                                            adjusted_offset);
1882           // Generate an OopMap entry
1883           Process_OopMap_Node(mach, adjusted_offset);
1884         }
1885 
1886         // Insert the delay slot instruction
1887         delay_slot->emit(masm, C->regalloc());
1888 
1889         // Don't reuse it
1890         delay_slot = nullptr;
1891       }
1892 
1893     } // End for all instructions in block
1894 
1895     // If the next block is the top of a loop, pad this block out to align
1896     // the loop top a little. Helps prevent pipe stalls at loop back branches.
1897     if (i < nblocks-1) {
1898       Block *nb = C->cfg()->get_block(i + 1);
1899       int padding = nb->alignment_padding(current_offset);
1900       if( padding > 0 ) {
1901         MachNode *nop = new MachNopNode(padding / nop_size);
1902         block->insert_node(nop, block->number_of_nodes());
1903         C->cfg()->map_node_to_block(nop, block);
1904         nop->emit(masm, C->regalloc());
1905         current_offset = masm->offset();
1906       }
1907     }
1908     // Verify that the distance for generated before forward
1909     // short branches is still valid.
1910     guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1911 
1912     // Save new block start offset
1913     blk_starts[i] = blk_offset;
1914   } // End of for all blocks
1915   blk_starts[nblocks] = current_offset;
1916 
1917   non_safepoints.flush_at_end();
1918 
1919   // Offset too large?
1920   if (C->failing())  return;
1921 
1922   // Define a pseudo-label at the end of the code
1923   masm->bind( blk_labels[nblocks] );
1924 
1925   // Compute the size of the first block
1926   _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1927 
1928 #ifdef ASSERT
1929   for (uint i = 0; i < nblocks; i++) { // For all blocks
1930     if (jmp_target[i] != 0) {
1931       int br_size = jmp_size[i];
1932       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1933       if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1934         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]);
1935         assert(false, "Displacement too large for short jmp");
1936       }
1937     }
1938   }
1939 #endif
1940 
1941   if (!masm->code()->finalize_stubs()) {
1942     C->record_failure("CodeCache is full");
1943     return;
1944   }
1945 
1946   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1947   bs->emit_stubs(*masm->code());
1948   if (C->failing())  return;
1949 
1950   // Fill in stubs.
1951   assert(masm->inst_mark() == nullptr, "should be.");
1952   _stub_list.emit(*masm);
1953   if (C->failing())  return;
1954 
1955 #ifndef PRODUCT
1956   // Information on the size of the method, without the extraneous code
1957   Scheduling::increment_method_size(masm->offset());
1958 #endif
1959 
1960   // ------------------
1961   // Fill in exception table entries.
1962   FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1963 
1964   // Only java methods have exception handlers and deopt handlers
1965   // class HandlerImpl is platform-specific and defined in the *.ad files.
1966   if (C->method()) {
1967     // Emit the exception handler code.
1968     _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(masm));
1969     if (C->failing()) {
1970       return; // CodeBuffer::expand failed
1971     }
1972     // Emit the deopt handler code.
1973     _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(masm));
1974 
1975     // Emit the MethodHandle deopt handler code (if required).
1976     if (C->has_method_handle_invokes() && !C->failing()) {
1977       // We can use the same code as for the normal deopt handler, we
1978       // just need a different entry point address.
1979       _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(masm));
1980     }
1981   }
1982 
1983   // One last check for failed CodeBuffer::expand:
1984   if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1985     C->record_failure("CodeCache is full");
1986     return;
1987   }
1988 
1989 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1990   if (C->print_assembly()) {
1991     tty->cr();
1992     tty->print_cr("============================= C2-compiled nmethod ==============================");
1993   }
1994 #endif
1995 
1996 #if defined(SUPPORT_OPTO_ASSEMBLY)
1997   // Dump the assembly code, including basic-block numbers
1998   if (C->print_assembly()) {
1999     ttyLocker ttyl;  // keep the following output all in one block
2000     if (!VMThread::should_terminate()) {  // test this under the tty lock
2001       // print_metadata and dump_asm may safepoint which makes us loose the ttylock.
2002       // We call them first and write to a stringStream, then we retake the lock to
2003       // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
2004       ResourceMark rm;
2005       stringStream method_metadata_str;
2006       if (C->method() != nullptr) {
2007         C->method()->print_metadata(&method_metadata_str);
2008       }
2009       stringStream dump_asm_str;
2010       dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
2011 
2012       NoSafepointVerifier nsv;
2013       ttyLocker ttyl2;
2014       // This output goes directly to the tty, not the compiler log.
2015       // To enable tools to match it up with the compilation activity,
2016       // be sure to tag this tty output with the compile ID.
2017       if (xtty != nullptr) {
2018         xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
2019                    C->is_osr_compilation() ? " compile_kind='osr'" : "");
2020       }
2021       if (C->method() != nullptr) {
2022         tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
2023         tty->print_raw(method_metadata_str.freeze());
2024       } else if (C->stub_name() != nullptr) {
2025         tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
2026       }
2027       tty->cr();
2028       tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
2029       tty->print_raw(dump_asm_str.freeze());
2030       tty->print_cr("--------------------------------------------------------------------------------");
2031       if (xtty != nullptr) {
2032         xtty->tail("opto_assembly");
2033       }
2034     }
2035   }
2036 #endif
2037 }
2038 
2039 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
2040   _inc_table.set_size(cnt);
2041 
2042   uint inct_cnt = 0;
2043   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2044     Block* block = C->cfg()->get_block(i);
2045     Node *n = nullptr;
2046     int j;
2047 
2048     // Find the branch; ignore trailing NOPs.
2049     for (j = block->number_of_nodes() - 1; j >= 0; j--) {
2050       n = block->get_node(j);
2051       if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
2052         break;
2053       }
2054     }
2055 
2056     // If we didn't find anything, continue
2057     if (j < 0) {
2058       continue;
2059     }
2060 
2061     // Compute ExceptionHandlerTable subtable entry and add it
2062     // (skip empty blocks)
2063     if (n->is_Catch()) {
2064 
2065       // Get the offset of the return from the call
2066       uint call_return = call_returns[block->_pre_order];
2067 #ifdef ASSERT
2068       assert( call_return > 0, "no call seen for this basic block" );
2069       while (block->get_node(--j)->is_MachProj()) ;
2070       assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
2071 #endif
2072       // last instruction is a CatchNode, find it's CatchProjNodes
2073       int nof_succs = block->_num_succs;
2074       // allocate space
2075       GrowableArray<intptr_t> handler_bcis(nof_succs);
2076       GrowableArray<intptr_t> handler_pcos(nof_succs);
2077       // iterate through all successors
2078       for (int j = 0; j < nof_succs; j++) {
2079         Block* s = block->_succs[j];
2080         bool found_p = false;
2081         for (uint k = 1; k < s->num_preds(); k++) {
2082           Node* pk = s->pred(k);
2083           if (pk->is_CatchProj() && pk->in(0) == n) {
2084             const CatchProjNode* p = pk->as_CatchProj();
2085             found_p = true;
2086             // add the corresponding handler bci & pco information
2087             if (p->_con != CatchProjNode::fall_through_index) {
2088               // p leads to an exception handler (and is not fall through)
2089               assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
2090               // no duplicates, please
2091               if (!handler_bcis.contains(p->handler_bci())) {
2092                 uint block_num = s->non_connector()->_pre_order;
2093                 handler_bcis.append(p->handler_bci());
2094                 handler_pcos.append(blk_labels[block_num].loc_pos());
2095               }
2096             }
2097           }
2098         }
2099         assert(found_p, "no matching predecessor found");
2100         // Note:  Due to empty block removal, one block may have
2101         // several CatchProj inputs, from the same Catch.
2102       }
2103 
2104       // Set the offset of the return from the call
2105       assert(handler_bcis.find(-1) != -1, "must have default handler");
2106       _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos);
2107       continue;
2108     }
2109 
2110     // Handle implicit null exception table updates
2111     if (n->is_MachNullCheck()) {
2112       MachNode* access = n->in(1)->as_Mach();
2113       assert(access->barrier_data() == 0 ||
2114              access->is_late_expanded_null_check_candidate(),
2115              "Implicit null checks on memory accesses with barriers are only supported on nodes explicitly marked as null-check candidates");
2116       uint block_num = block->non_connector_successor(0)->_pre_order;
2117       _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2118       continue;
2119     }
2120     // Handle implicit exception table updates: trap instructions.
2121     if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
2122       uint block_num = block->non_connector_successor(0)->_pre_order;
2123       _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2124       continue;
2125     }
2126   } // End of for all blocks fill in exception table entries
2127 }
2128 
2129 // Static Variables
2130 #ifndef PRODUCT
2131 uint Scheduling::_total_nop_size = 0;
2132 uint Scheduling::_total_method_size = 0;
2133 uint Scheduling::_total_branches = 0;
2134 uint Scheduling::_total_unconditional_delays = 0;
2135 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
2136 #endif
2137 
2138 // Initializer for class Scheduling
2139 
2140 Scheduling::Scheduling(Arena *arena, Compile &compile)
2141         : _arena(arena),
2142           _cfg(compile.cfg()),
2143           _regalloc(compile.regalloc()),
2144           _scheduled(arena),
2145           _available(arena),
2146           _reg_node(arena),
2147           _pinch_free_list(arena),
2148           _next_node(nullptr),
2149           _bundle_instr_count(0),
2150           _bundle_cycle_number(0),
2151           _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
2152 #ifndef PRODUCT
2153         , _branches(0)
2154         , _unconditional_delays(0)
2155 #endif
2156 {
2157   // Create a MachNopNode
2158   _nop = new MachNopNode();
2159 
2160   // Now that the nops are in the array, save the count
2161   // (but allow entries for the nops)
2162   _node_bundling_limit = compile.unique();
2163   uint node_max = _regalloc->node_regs_max_index();
2164 
2165   compile.output()->set_node_bundling_limit(_node_bundling_limit);
2166 
2167   // This one is persistent within the Compile class
2168   _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
2169 
2170   // Allocate space for fixed-size arrays
2171   _uses            = NEW_ARENA_ARRAY(arena, short,          node_max);
2172   _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
2173 
2174   // Clear the arrays
2175   for (uint i = 0; i < node_max; i++) {
2176     ::new (&_node_bundling_base[i]) Bundle();
2177   }
2178   memset(_uses,               0, node_max * sizeof(short));
2179   memset(_current_latency,    0, node_max * sizeof(unsigned short));
2180 
2181   // Clear the bundling information
2182   memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2183 
2184   // Get the last node
2185   Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2186 
2187   _next_node = block->get_node(block->number_of_nodes() - 1);
2188 }
2189 
2190 #ifndef PRODUCT
2191 // Scheduling destructor
2192 Scheduling::~Scheduling() {
2193   _total_branches             += _branches;
2194   _total_unconditional_delays += _unconditional_delays;
2195 }
2196 #endif
2197 
2198 // Step ahead "i" cycles
2199 void Scheduling::step(uint i) {
2200 
2201   Bundle *bundle = node_bundling(_next_node);
2202   bundle->set_starts_bundle();
2203 
2204   // Update the bundle record, but leave the flags information alone
2205   if (_bundle_instr_count > 0) {
2206     bundle->set_instr_count(_bundle_instr_count);
2207     bundle->set_resources_used(_bundle_use.resourcesUsed());
2208   }
2209 
2210   // Update the state information
2211   _bundle_instr_count = 0;
2212   _bundle_cycle_number += i;
2213   _bundle_use.step(i);
2214 }
2215 
2216 void Scheduling::step_and_clear() {
2217   Bundle *bundle = node_bundling(_next_node);
2218   bundle->set_starts_bundle();
2219 
2220   // Update the bundle record
2221   if (_bundle_instr_count > 0) {
2222     bundle->set_instr_count(_bundle_instr_count);
2223     bundle->set_resources_used(_bundle_use.resourcesUsed());
2224 
2225     _bundle_cycle_number += 1;
2226   }
2227 
2228   // Clear the bundling information
2229   _bundle_instr_count = 0;
2230   _bundle_use.reset();
2231 
2232   memcpy(_bundle_use_elements,
2233          Pipeline_Use::elaborated_elements,
2234          sizeof(Pipeline_Use::elaborated_elements));
2235 }
2236 
2237 // Perform instruction scheduling and bundling over the sequence of
2238 // instructions in backwards order.
2239 void PhaseOutput::ScheduleAndBundle() {
2240 
2241   // Don't optimize this if it isn't a method
2242   if (!C->method())
2243     return;
2244 
2245   // Don't optimize this if scheduling is disabled
2246   if (!C->do_scheduling())
2247     return;
2248 
2249   // Scheduling code works only with pairs (8 bytes) maximum.
2250   // And when the scalable vector register is used, we may spill/unspill
2251   // the whole reg regardless of the max vector size.
2252   if (C->max_vector_size() > 8 ||
2253       (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2254     return;
2255   }
2256 
2257   Compile::TracePhase tp(_t_instrSched);
2258 
2259   // Create a data structure for all the scheduling information
2260   Scheduling scheduling(Thread::current()->resource_area(), *C);
2261 
2262   // Walk backwards over each basic block, computing the needed alignment
2263   // Walk over all the basic blocks
2264   scheduling.DoScheduling();
2265 
2266 #ifndef PRODUCT
2267   if (C->trace_opto_output()) {
2268     // Buffer and print all at once
2269     ResourceMark rm;
2270     stringStream ss;
2271     ss.print("\n---- After ScheduleAndBundle ----\n");
2272     print_scheduling(&ss);
2273     tty->print("%s", ss.as_string());
2274   }
2275 #endif
2276 }
2277 
2278 #ifndef PRODUCT
2279 // Separated out so that it can be called directly from debugger
2280 void PhaseOutput::print_scheduling() {
2281   print_scheduling(tty);
2282 }
2283 
2284 void PhaseOutput::print_scheduling(outputStream* output_stream) {
2285   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2286     output_stream->print("\nBB#%03d:\n", i);
2287     Block* block = C->cfg()->get_block(i);
2288     for (uint j = 0; j < block->number_of_nodes(); j++) {
2289       Node* n = block->get_node(j);
2290       OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2291       output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2292       n->dump("\n", false, output_stream);
2293     }
2294   }
2295 }
2296 #endif
2297 
2298 // See if this node fits into the present instruction bundle
2299 bool Scheduling::NodeFitsInBundle(Node *n) {
2300   uint n_idx = n->_idx;
2301 
2302   // If this is the unconditional delay instruction, then it fits
2303   if (n == _unconditional_delay_slot) {
2304 #ifndef PRODUCT
2305     if (_cfg->C->trace_opto_output())
2306       tty->print("#     NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2307 #endif
2308     return (true);
2309   }
2310 
2311   // If the node cannot be scheduled this cycle, skip it
2312   if (_current_latency[n_idx] > _bundle_cycle_number) {
2313 #ifndef PRODUCT
2314     if (_cfg->C->trace_opto_output())
2315       tty->print("#     NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2316                  n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2317 #endif
2318     return (false);
2319   }
2320 
2321   const Pipeline *node_pipeline = n->pipeline();
2322 
2323   uint instruction_count = node_pipeline->instructionCount();
2324   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2325     instruction_count = 0;
2326   else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2327     instruction_count++;
2328 
2329   if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2330 #ifndef PRODUCT
2331     if (_cfg->C->trace_opto_output())
2332       tty->print("#     NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2333                  n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2334 #endif
2335     return (false);
2336   }
2337 
2338   // Don't allow non-machine nodes to be handled this way
2339   if (!n->is_Mach() && instruction_count == 0)
2340     return (false);
2341 
2342   // See if there is any overlap
2343   uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2344 
2345   if (delay > 0) {
2346 #ifndef PRODUCT
2347     if (_cfg->C->trace_opto_output())
2348       tty->print("#     NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2349 #endif
2350     return false;
2351   }
2352 
2353 #ifndef PRODUCT
2354   if (_cfg->C->trace_opto_output())
2355     tty->print("#     NodeFitsInBundle [%4d]:  TRUE\n", n_idx);
2356 #endif
2357 
2358   return true;
2359 }
2360 
2361 Node * Scheduling::ChooseNodeToBundle() {
2362   uint siz = _available.size();
2363 
2364   if (siz == 0) {
2365 
2366 #ifndef PRODUCT
2367     if (_cfg->C->trace_opto_output())
2368       tty->print("#   ChooseNodeToBundle: null\n");
2369 #endif
2370     return (nullptr);
2371   }
2372 
2373   // Fast path, if only 1 instruction in the bundle
2374   if (siz == 1) {
2375 #ifndef PRODUCT
2376     if (_cfg->C->trace_opto_output()) {
2377       tty->print("#   ChooseNodeToBundle (only 1): ");
2378       _available[0]->dump();
2379     }
2380 #endif
2381     return (_available[0]);
2382   }
2383 
2384   // Don't bother, if the bundle is already full
2385   if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2386     for ( uint i = 0; i < siz; i++ ) {
2387       Node *n = _available[i];
2388 
2389       // Skip projections, we'll handle them another way
2390       if (n->is_Proj())
2391         continue;
2392 
2393       // This presupposed that instructions are inserted into the
2394       // available list in a legality order; i.e. instructions that
2395       // must be inserted first are at the head of the list
2396       if (NodeFitsInBundle(n)) {
2397 #ifndef PRODUCT
2398         if (_cfg->C->trace_opto_output()) {
2399           tty->print("#   ChooseNodeToBundle: ");
2400           n->dump();
2401         }
2402 #endif
2403         return (n);
2404       }
2405     }
2406   }
2407 
2408   // Nothing fits in this bundle, choose the highest priority
2409 #ifndef PRODUCT
2410   if (_cfg->C->trace_opto_output()) {
2411     tty->print("#   ChooseNodeToBundle: ");
2412     _available[0]->dump();
2413   }
2414 #endif
2415 
2416   return _available[0];
2417 }
2418 
2419 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) {
2420   assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), "");
2421 
2422   OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1));
2423   OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first);
2424   OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1));
2425   OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second);
2426 
2427   // Comparison between stack -> reg and stack -> reg
2428   if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) &&
2429       OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) {
2430     return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo);
2431   }
2432 
2433   // Comparison between reg -> stack and reg -> stack
2434   if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) &&
2435       OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) {
2436     return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo);
2437   }
2438 
2439   return 0; // Not comparable
2440 }
2441 
2442 void Scheduling::AddNodeToAvailableList(Node *n) {
2443   assert( !n->is_Proj(), "projections never directly made available" );
2444 #ifndef PRODUCT
2445   if (_cfg->C->trace_opto_output()) {
2446     tty->print("#   AddNodeToAvailableList: ");
2447     n->dump();
2448   }
2449 #endif
2450 
2451   int latency = _current_latency[n->_idx];
2452 
2453   // Insert in latency order (insertion sort). If two MachSpillCopyNodes
2454   // for stack spilling or unspilling have the same latency, we sort
2455   // them in the order of stack offset. Some ports (e.g. aarch64) may also
2456   // have more opportunities to do ld/st merging
2457   uint i;
2458   for (i = 0; i < _available.size(); i++) {
2459     if (_current_latency[_available[i]->_idx] > latency) {
2460       break;
2461     } else if (_current_latency[_available[i]->_idx] == latency &&
2462                n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() &&
2463                compare_two_spill_nodes(n, _available[i]) > 0) {
2464       break;
2465     }
2466   }
2467 
2468   // Special Check for compares following branches
2469   if( n->is_Mach() && _scheduled.size() > 0 ) {
2470     int op = n->as_Mach()->ideal_Opcode();
2471     Node *last = _scheduled[0];
2472     if( last->is_MachIf() && last->in(1) == n &&
2473         ( op == Op_CmpI ||
2474           op == Op_CmpU ||
2475           op == Op_CmpUL ||
2476           op == Op_CmpP ||
2477           op == Op_CmpF ||
2478           op == Op_CmpD ||
2479           op == Op_CmpL ) ) {
2480 
2481       // Recalculate position, moving to front of same latency
2482       for ( i=0 ; i < _available.size(); i++ )
2483         if (_current_latency[_available[i]->_idx] >= latency)
2484           break;
2485     }
2486   }
2487 
2488   // Insert the node in the available list
2489   _available.insert(i, n);
2490 
2491 #ifndef PRODUCT
2492   if (_cfg->C->trace_opto_output())
2493     dump_available();
2494 #endif
2495 }
2496 
2497 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2498   for ( uint i=0; i < n->len(); i++ ) {
2499     Node *def = n->in(i);
2500     if (!def) continue;
2501     if( def->is_Proj() )        // If this is a machine projection, then
2502       def = def->in(0);         // propagate usage thru to the base instruction
2503 
2504     if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2505       continue;
2506     }
2507 
2508     // Compute the latency
2509     uint l = _bundle_cycle_number + n->latency(i);
2510     if (_current_latency[def->_idx] < l)
2511       _current_latency[def->_idx] = l;
2512 
2513     // If this does not have uses then schedule it
2514     if ((--_uses[def->_idx]) == 0)
2515       AddNodeToAvailableList(def);
2516   }
2517 }
2518 
2519 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2520 #ifndef PRODUCT
2521   if (_cfg->C->trace_opto_output()) {
2522     tty->print("#   AddNodeToBundle: ");
2523     n->dump();
2524   }
2525 #endif
2526 
2527   // Remove this from the available list
2528   uint i;
2529   for (i = 0; i < _available.size(); i++)
2530     if (_available[i] == n)
2531       break;
2532   assert(i < _available.size(), "entry in _available list not found");
2533   _available.remove(i);
2534 
2535   // See if this fits in the current bundle
2536   const Pipeline *node_pipeline = n->pipeline();
2537   const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2538 
2539   // Check for instructions to be placed in the delay slot. We
2540   // do this before we actually schedule the current instruction,
2541   // because the delay slot follows the current instruction.
2542   if (Pipeline::_branch_has_delay_slot &&
2543       node_pipeline->hasBranchDelay() &&
2544       !_unconditional_delay_slot) {
2545 
2546     uint siz = _available.size();
2547 
2548     // Conditional branches can support an instruction that
2549     // is unconditionally executed and not dependent by the
2550     // branch, OR a conditionally executed instruction if
2551     // the branch is taken.  In practice, this means that
2552     // the first instruction at the branch target is
2553     // copied to the delay slot, and the branch goes to
2554     // the instruction after that at the branch target
2555     if ( n->is_MachBranch() ) {
2556 
2557       assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2558       assert( !n->is_Catch(),         "should not look for delay slot for Catch" );
2559 
2560 #ifndef PRODUCT
2561       _branches++;
2562 #endif
2563 
2564       // At least 1 instruction is on the available list
2565       // that is not dependent on the branch
2566       for (uint i = 0; i < siz; i++) {
2567         Node *d = _available[i];
2568         const Pipeline *avail_pipeline = d->pipeline();
2569 
2570         // Don't allow safepoints in the branch shadow, that will
2571         // cause a number of difficulties
2572         if ( avail_pipeline->instructionCount() == 1 &&
2573              !avail_pipeline->hasMultipleBundles() &&
2574              !avail_pipeline->hasBranchDelay() &&
2575              Pipeline::instr_has_unit_size() &&
2576              d->size(_regalloc) == Pipeline::instr_unit_size() &&
2577              NodeFitsInBundle(d) &&
2578              !node_bundling(d)->used_in_delay()) {
2579 
2580           if (d->is_Mach() && !d->is_MachSafePoint()) {
2581             // A node that fits in the delay slot was found, so we need to
2582             // set the appropriate bits in the bundle pipeline information so
2583             // that it correctly indicates resource usage.  Later, when we
2584             // attempt to add this instruction to the bundle, we will skip
2585             // setting the resource usage.
2586             _unconditional_delay_slot = d;
2587             node_bundling(n)->set_use_unconditional_delay();
2588             node_bundling(d)->set_used_in_unconditional_delay();
2589             _bundle_use.add_usage(avail_pipeline->resourceUse());
2590             _current_latency[d->_idx] = _bundle_cycle_number;
2591             _next_node = d;
2592             ++_bundle_instr_count;
2593 #ifndef PRODUCT
2594             _unconditional_delays++;
2595 #endif
2596             break;
2597           }
2598         }
2599       }
2600     }
2601 
2602     // No delay slot, add a nop to the usage
2603     if (!_unconditional_delay_slot) {
2604       // See if adding an instruction in the delay slot will overflow
2605       // the bundle.
2606       if (!NodeFitsInBundle(_nop)) {
2607 #ifndef PRODUCT
2608         if (_cfg->C->trace_opto_output())
2609           tty->print("#  *** STEP(1 instruction for delay slot) ***\n");
2610 #endif
2611         step(1);
2612       }
2613 
2614       _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2615       _next_node = _nop;
2616       ++_bundle_instr_count;
2617     }
2618 
2619     // See if the instruction in the delay slot requires a
2620     // step of the bundles
2621     if (!NodeFitsInBundle(n)) {
2622 #ifndef PRODUCT
2623       if (_cfg->C->trace_opto_output())
2624         tty->print("#  *** STEP(branch won't fit) ***\n");
2625 #endif
2626       // Update the state information
2627       _bundle_instr_count = 0;
2628       _bundle_cycle_number += 1;
2629       _bundle_use.step(1);
2630     }
2631   }
2632 
2633   // Get the number of instructions
2634   uint instruction_count = node_pipeline->instructionCount();
2635   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2636     instruction_count = 0;
2637 
2638   // Compute the latency information
2639   uint delay = 0;
2640 
2641   if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2642     int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2643     if (relative_latency < 0)
2644       relative_latency = 0;
2645 
2646     delay = _bundle_use.full_latency(relative_latency, node_usage);
2647 
2648     // Does not fit in this bundle, start a new one
2649     if (delay > 0) {
2650       step(delay);
2651 
2652 #ifndef PRODUCT
2653       if (_cfg->C->trace_opto_output())
2654         tty->print("#  *** STEP(%d) ***\n", delay);
2655 #endif
2656     }
2657   }
2658 
2659   // If this was placed in the delay slot, ignore it
2660   if (n != _unconditional_delay_slot) {
2661 
2662     if (delay == 0) {
2663       if (node_pipeline->hasMultipleBundles()) {
2664 #ifndef PRODUCT
2665         if (_cfg->C->trace_opto_output())
2666           tty->print("#  *** STEP(multiple instructions) ***\n");
2667 #endif
2668         step(1);
2669       }
2670 
2671       else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2672 #ifndef PRODUCT
2673         if (_cfg->C->trace_opto_output())
2674           tty->print("#  *** STEP(%d >= %d instructions) ***\n",
2675                      instruction_count + _bundle_instr_count,
2676                      Pipeline::_max_instrs_per_cycle);
2677 #endif
2678         step(1);
2679       }
2680     }
2681 
2682     if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2683       _bundle_instr_count++;
2684 
2685     // Set the node's latency
2686     _current_latency[n->_idx] = _bundle_cycle_number;
2687 
2688     // Now merge the functional unit information
2689     if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2690       _bundle_use.add_usage(node_usage);
2691 
2692     // Increment the number of instructions in this bundle
2693     _bundle_instr_count += instruction_count;
2694 
2695     // Remember this node for later
2696     if (n->is_Mach())
2697       _next_node = n;
2698   }
2699 
2700   // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2701   // not in the bb->_nodes array.  This happens for debug-info-only BoxLocks.
2702   // 'Schedule' them (basically ignore in the schedule) but do not insert them
2703   // into the block.  All other scheduled nodes get put in the schedule here.
2704   int op = n->Opcode();
2705   if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2706       (op != Op_Node &&         // Not an unused antidepedence node and
2707        // not an unallocated boxlock
2708        (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2709 
2710     // Push any trailing projections
2711     if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2712       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2713         Node *foi = n->fast_out(i);
2714         if( foi->is_Proj() )
2715           _scheduled.push(foi);
2716       }
2717     }
2718 
2719     // Put the instruction in the schedule list
2720     _scheduled.push(n);
2721   }
2722 
2723 #ifndef PRODUCT
2724   if (_cfg->C->trace_opto_output())
2725     dump_available();
2726 #endif
2727 
2728   // Walk all the definitions, decrementing use counts, and
2729   // if a definition has a 0 use count, place it in the available list.
2730   DecrementUseCounts(n,bb);
2731 }
2732 
2733 // This method sets the use count within a basic block.  We will ignore all
2734 // uses outside the current basic block.  As we are doing a backwards walk,
2735 // any node we reach that has a use count of 0 may be scheduled.  This also
2736 // avoids the problem of cyclic references from phi nodes, as long as phi
2737 // nodes are at the front of the basic block.  This method also initializes
2738 // the available list to the set of instructions that have no uses within this
2739 // basic block.
2740 void Scheduling::ComputeUseCount(const Block *bb) {
2741 #ifndef PRODUCT
2742   if (_cfg->C->trace_opto_output())
2743     tty->print("# -> ComputeUseCount\n");
2744 #endif
2745 
2746   // Clear the list of available and scheduled instructions, just in case
2747   _available.clear();
2748   _scheduled.clear();
2749 
2750   // No delay slot specified
2751   _unconditional_delay_slot = nullptr;
2752 
2753 #ifdef ASSERT
2754   for( uint i=0; i < bb->number_of_nodes(); i++ )
2755     assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2756 #endif
2757 
2758   // Force the _uses count to never go to zero for unscheduable pieces
2759   // of the block
2760   for( uint k = 0; k < _bb_start; k++ )
2761     _uses[bb->get_node(k)->_idx] = 1;
2762   for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2763     _uses[bb->get_node(l)->_idx] = 1;
2764 
2765   // Iterate backwards over the instructions in the block.  Don't count the
2766   // branch projections at end or the block header instructions.
2767   for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2768     Node *n = bb->get_node(j);
2769     if( n->is_Proj() ) continue; // Projections handled another way
2770 
2771     // Account for all uses
2772     for ( uint k = 0; k < n->len(); k++ ) {
2773       Node *inp = n->in(k);
2774       if (!inp) continue;
2775       assert(inp != n, "no cycles allowed" );
2776       if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2777         if (inp->is_Proj()) { // Skip through Proj's
2778           inp = inp->in(0);
2779         }
2780         ++_uses[inp->_idx];     // Count 1 block-local use
2781       }
2782     }
2783 
2784     // If this instruction has a 0 use count, then it is available
2785     if (!_uses[n->_idx]) {
2786       _current_latency[n->_idx] = _bundle_cycle_number;
2787       AddNodeToAvailableList(n);
2788     }
2789 
2790 #ifndef PRODUCT
2791     if (_cfg->C->trace_opto_output()) {
2792       tty->print("#   uses: %3d: ", _uses[n->_idx]);
2793       n->dump();
2794     }
2795 #endif
2796   }
2797 
2798 #ifndef PRODUCT
2799   if (_cfg->C->trace_opto_output())
2800     tty->print("# <- ComputeUseCount\n");
2801 #endif
2802 }
2803 
2804 // This routine performs scheduling on each basic block in reverse order,
2805 // using instruction latencies and taking into account function unit
2806 // availability.
2807 void Scheduling::DoScheduling() {
2808 #ifndef PRODUCT
2809   if (_cfg->C->trace_opto_output())
2810     tty->print("# -> DoScheduling\n");
2811 #endif
2812 
2813   Block *succ_bb = nullptr;
2814   Block *bb;
2815   Compile* C = Compile::current();
2816 
2817   // Walk over all the basic blocks in reverse order
2818   for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2819     bb = _cfg->get_block(i);
2820 
2821 #ifndef PRODUCT
2822     if (_cfg->C->trace_opto_output()) {
2823       tty->print("#  Schedule BB#%03d (initial)\n", i);
2824       for (uint j = 0; j < bb->number_of_nodes(); j++) {
2825         bb->get_node(j)->dump();
2826       }
2827     }
2828 #endif
2829 
2830     // On the head node, skip processing
2831     if (bb == _cfg->get_root_block()) {
2832       continue;
2833     }
2834 
2835     // Skip empty, connector blocks
2836     if (bb->is_connector())
2837       continue;
2838 
2839     // If the following block is not the sole successor of
2840     // this one, then reset the pipeline information
2841     if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2842 #ifndef PRODUCT
2843       if (_cfg->C->trace_opto_output()) {
2844         tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2845                    _next_node->_idx, _bundle_instr_count);
2846       }
2847 #endif
2848       step_and_clear();
2849     }
2850 
2851     // Leave untouched the starting instruction, any Phis, a CreateEx node
2852     // or Top.  bb->get_node(_bb_start) is the first schedulable instruction.
2853     _bb_end = bb->number_of_nodes()-1;
2854     for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2855       Node *n = bb->get_node(_bb_start);
2856       // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2857       // Also, MachIdealNodes do not get scheduled
2858       if( !n->is_Mach() ) continue;     // Skip non-machine nodes
2859       MachNode *mach = n->as_Mach();
2860       int iop = mach->ideal_Opcode();
2861       if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2862       if( iop == Op_Con ) continue;      // Do not schedule Top
2863       if( iop == Op_Node &&     // Do not schedule PhiNodes, ProjNodes
2864           mach->pipeline() == MachNode::pipeline_class() &&
2865           !n->is_SpillCopy() && !n->is_MachMerge() )  // Breakpoints, Prolog, etc
2866         continue;
2867       break;                    // Funny loop structure to be sure...
2868     }
2869     // Compute last "interesting" instruction in block - last instruction we
2870     // might schedule.  _bb_end points just after last schedulable inst.  We
2871     // normally schedule conditional branches (despite them being forced last
2872     // in the block), because they have delay slots we can fill.  Calls all
2873     // have their delay slots filled in the template expansions, so we don't
2874     // bother scheduling them.
2875     Node *last = bb->get_node(_bb_end);
2876     // Ignore trailing NOPs.
2877     while (_bb_end > 0 && last->is_Mach() &&
2878            last->as_Mach()->ideal_Opcode() == Op_Con) {
2879       last = bb->get_node(--_bb_end);
2880     }
2881     assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2882     if( last->is_Catch() ||
2883         (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2884       // There might be a prior call.  Skip it.
2885       while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2886     } else if( last->is_MachNullCheck() ) {
2887       // Backup so the last null-checked memory instruction is
2888       // outside the schedulable range. Skip over the nullcheck,
2889       // projection, and the memory nodes.
2890       Node *mem = last->in(1);
2891       do {
2892         _bb_end--;
2893       } while (mem != bb->get_node(_bb_end));
2894     } else {
2895       // Set _bb_end to point after last schedulable inst.
2896       _bb_end++;
2897     }
2898 
2899     assert( _bb_start <= _bb_end, "inverted block ends" );
2900 
2901     // Compute the register antidependencies for the basic block
2902     ComputeRegisterAntidependencies(bb);
2903     if (C->failing())  return;  // too many D-U pinch points
2904 
2905     // Compute the usage within the block, and set the list of all nodes
2906     // in the block that have no uses within the block.
2907     ComputeUseCount(bb);
2908 
2909     // Schedule the remaining instructions in the block
2910     while ( _available.size() > 0 ) {
2911       Node *n = ChooseNodeToBundle();
2912       guarantee(n != nullptr, "no nodes available");
2913       AddNodeToBundle(n,bb);
2914     }
2915 
2916     assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2917 #ifdef ASSERT
2918     for( uint l = _bb_start; l < _bb_end; l++ ) {
2919       Node *n = bb->get_node(l);
2920       uint m;
2921       for( m = 0; m < _bb_end-_bb_start; m++ )
2922         if( _scheduled[m] == n )
2923           break;
2924       assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2925     }
2926 #endif
2927 
2928     // Now copy the instructions (in reverse order) back to the block
2929     for ( uint k = _bb_start; k < _bb_end; k++ )
2930       bb->map_node(_scheduled[_bb_end-k-1], k);
2931 
2932 #ifndef PRODUCT
2933     if (_cfg->C->trace_opto_output()) {
2934       tty->print("#  Schedule BB#%03d (final)\n", i);
2935       uint current = 0;
2936       for (uint j = 0; j < bb->number_of_nodes(); j++) {
2937         Node *n = bb->get_node(j);
2938         if( valid_bundle_info(n) ) {
2939           Bundle *bundle = node_bundling(n);
2940           if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2941             tty->print("*** Bundle: ");
2942             bundle->dump();
2943           }
2944           n->dump();
2945         }
2946       }
2947     }
2948 #endif
2949 #ifdef ASSERT
2950     verify_good_schedule(bb,"after block local scheduling");
2951 #endif
2952   }
2953 
2954 #ifndef PRODUCT
2955   if (_cfg->C->trace_opto_output())
2956     tty->print("# <- DoScheduling\n");
2957 #endif
2958 
2959   // Record final node-bundling array location
2960   _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2961 
2962 } // end DoScheduling
2963 
2964 // Verify that no live-range used in the block is killed in the block by a
2965 // wrong DEF.  This doesn't verify live-ranges that span blocks.
2966 
2967 // Check for edge existence.  Used to avoid adding redundant precedence edges.
2968 static bool edge_from_to( Node *from, Node *to ) {
2969   for( uint i=0; i<from->len(); i++ )
2970     if( from->in(i) == to )
2971       return true;
2972   return false;
2973 }
2974 
2975 #ifdef ASSERT
2976 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2977   // Check for bad kills
2978   if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2979     Node *prior_use = _reg_node[def];
2980     if( prior_use && !edge_from_to(prior_use,n) ) {
2981       tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2982       n->dump();
2983       tty->print_cr("...");
2984       prior_use->dump();
2985       assert(edge_from_to(prior_use,n), "%s", msg);
2986     }
2987     _reg_node.map(def,nullptr); // Kill live USEs
2988   }
2989 }
2990 
2991 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2992 
2993   // Zap to something reasonable for the verify code
2994   _reg_node.clear();
2995 
2996   // Walk over the block backwards.  Check to make sure each DEF doesn't
2997   // kill a live value (other than the one it's supposed to).  Add each
2998   // USE to the live set.
2999   for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
3000     Node *n = b->get_node(i);
3001     int n_op = n->Opcode();
3002     if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
3003       // Fat-proj kills a slew of registers
3004       RegMaskIterator rmi(n->out_RegMask());
3005       while (rmi.has_next()) {
3006         OptoReg::Name kill = rmi.next();
3007         verify_do_def(n, kill, msg);
3008       }
3009     } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
3010       // Get DEF'd registers the normal way
3011       verify_do_def( n, _regalloc->get_reg_first(n), msg );
3012       verify_do_def( n, _regalloc->get_reg_second(n), msg );
3013     }
3014 
3015     // Now make all USEs live
3016     for( uint i=1; i<n->req(); i++ ) {
3017       Node *def = n->in(i);
3018       assert(def != nullptr, "input edge required");
3019       OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
3020       OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
3021       if( OptoReg::is_valid(reg_lo) ) {
3022         assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
3023         _reg_node.map(reg_lo,n);
3024       }
3025       if( OptoReg::is_valid(reg_hi) ) {
3026         assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
3027         _reg_node.map(reg_hi,n);
3028       }
3029     }
3030 
3031   }
3032 
3033   // Zap to something reasonable for the Antidependence code
3034   _reg_node.clear();
3035 }
3036 #endif
3037 
3038 // Conditionally add precedence edges.  Avoid putting edges on Projs.
3039 static void add_prec_edge_from_to( Node *from, Node *to ) {
3040   if( from->is_Proj() ) {       // Put precedence edge on Proj's input
3041     assert( from->req() == 1 && (from->len() == 1 || from->in(1) == nullptr), "no precedence edges on projections" );
3042     from = from->in(0);
3043   }
3044   if( from != to &&             // No cycles (for things like LD L0,[L0+4] )
3045       !edge_from_to( from, to ) ) // Avoid duplicate edge
3046     from->add_prec(to);
3047 }
3048 
3049 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
3050   if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
3051     return;
3052 
3053   if (OptoReg::is_reg(def_reg)) {
3054     VMReg vmreg = OptoReg::as_VMReg(def_reg);
3055     if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
3056       // This is one of the high slots of a vector register.
3057       // ScheduleAndBundle already checked there are no live wide
3058       // vectors in this method so it can be safely ignored.
3059       return;
3060     }
3061   }
3062 
3063   Node *pinch = _reg_node[def_reg]; // Get pinch point
3064   if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
3065       is_def ) {    // Check for a true def (not a kill)
3066     _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
3067     return;
3068   }
3069 
3070   Node *kill = def;             // Rename 'def' to more descriptive 'kill'
3071   DEBUG_ONLY( def = (Node*)((intptr_t)0xdeadbeef); )
3072 
3073   // After some number of kills there _may_ be a later def
3074   Node *later_def = nullptr;
3075 
3076   Compile* C = Compile::current();
3077 
3078   // Finding a kill requires a real pinch-point.
3079   // Check for not already having a pinch-point.
3080   // Pinch points are Op_Node's.
3081   if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
3082     later_def = pinch;            // Must be def/kill as optimistic pinch-point
3083     if ( _pinch_free_list.size() > 0) {
3084       pinch = _pinch_free_list.pop();
3085     } else {
3086       pinch = new Node(1); // Pinch point to-be
3087     }
3088     if (pinch->_idx >= _regalloc->node_regs_max_index()) {
3089       DEBUG_ONLY( pinch->dump(); );
3090       assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index());
3091       _cfg->C->record_method_not_compilable("too many D-U pinch points");
3092       return;
3093     }
3094     _cfg->map_node_to_block(pinch, b);      // Pretend it's valid in this block (lazy init)
3095     _reg_node.map(def_reg,pinch); // Record pinch-point
3096     //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
3097     if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
3098       pinch->init_req(0, C->top());     // set not null for the next call
3099       add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
3100       later_def = nullptr;           // and no later def
3101     }
3102     pinch->set_req(0,later_def);  // Hook later def so we can find it
3103   } else {                        // Else have valid pinch point
3104     if( pinch->in(0) )            // If there is a later-def
3105       later_def = pinch->in(0);   // Get it
3106   }
3107 
3108   // Add output-dependence edge from later def to kill
3109   if( later_def )               // If there is some original def
3110     add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
3111 
3112   // See if current kill is also a use, and so is forced to be the pinch-point.
3113   if( pinch->Opcode() == Op_Node ) {
3114     Node *uses = kill->is_Proj() ? kill->in(0) : kill;
3115     for( uint i=1; i<uses->req(); i++ ) {
3116       if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
3117           _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
3118         // Yes, found a use/kill pinch-point
3119         pinch->set_req(0,nullptr);  //
3120         pinch->replace_by(kill); // Move anti-dep edges up
3121         pinch = kill;
3122         _reg_node.map(def_reg,pinch);
3123         return;
3124       }
3125     }
3126   }
3127 
3128   // Add edge from kill to pinch-point
3129   add_prec_edge_from_to(kill,pinch);
3130 }
3131 
3132 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
3133   if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
3134     return;
3135   Node *pinch = _reg_node[use_reg]; // Get pinch point
3136   // Check for no later def_reg/kill in block
3137   if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b &&
3138       // Use has to be block-local as well
3139       _cfg->get_block_for_node(use) == b) {
3140     if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
3141         pinch->req() == 1 ) {   // pinch not yet in block?
3142       pinch->del_req(0);        // yank pointer to later-def, also set flag
3143       // Insert the pinch-point in the block just after the last use
3144       b->insert_node(pinch, b->find_node(use) + 1);
3145       _bb_end++;                // Increase size scheduled region in block
3146     }
3147 
3148     add_prec_edge_from_to(pinch,use);
3149   }
3150 }
3151 
3152 // We insert antidependences between the reads and following write of
3153 // allocated registers to prevent illegal code motion. Hopefully, the
3154 // number of added references should be fairly small, especially as we
3155 // are only adding references within the current basic block.
3156 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
3157 
3158 #ifdef ASSERT
3159   verify_good_schedule(b,"before block local scheduling");
3160 #endif
3161 
3162   // A valid schedule, for each register independently, is an endless cycle
3163   // of: a def, then some uses (connected to the def by true dependencies),
3164   // then some kills (defs with no uses), finally the cycle repeats with a new
3165   // def.  The uses are allowed to float relative to each other, as are the
3166   // kills.  No use is allowed to slide past a kill (or def).  This requires
3167   // antidependencies between all uses of a single def and all kills that
3168   // follow, up to the next def.  More edges are redundant, because later defs
3169   // & kills are already serialized with true or antidependencies.  To keep
3170   // the edge count down, we add a 'pinch point' node if there's more than
3171   // one use or more than one kill/def.
3172 
3173   // We add dependencies in one bottom-up pass.
3174 
3175   // For each instruction we handle it's DEFs/KILLs, then it's USEs.
3176 
3177   // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
3178   // register.  If not, we record the DEF/KILL in _reg_node, the
3179   // register-to-def mapping.  If there is a prior DEF/KILL, we insert a
3180   // "pinch point", a new Node that's in the graph but not in the block.
3181   // We put edges from the prior and current DEF/KILLs to the pinch point.
3182   // We put the pinch point in _reg_node.  If there's already a pinch point
3183   // we merely add an edge from the current DEF/KILL to the pinch point.
3184 
3185   // After doing the DEF/KILLs, we handle USEs.  For each used register, we
3186   // put an edge from the pinch point to the USE.
3187 
3188   // To be expedient, the _reg_node array is pre-allocated for the whole
3189   // compilation.  _reg_node is lazily initialized; it either contains a null,
3190   // or a valid def/kill/pinch-point, or a leftover node from some prior
3191   // block.  Leftover node from some prior block is treated like a null (no
3192   // prior def, so no anti-dependence needed).  Valid def is distinguished by
3193   // it being in the current block.
3194   bool fat_proj_seen = false;
3195   uint last_safept = _bb_end-1;
3196   Node* end_node         = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr;
3197   Node* last_safept_node = end_node;
3198   for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
3199     Node *n = b->get_node(i);
3200     int is_def = n->outcnt();   // def if some uses prior to adding precedence edges
3201     if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
3202       // Fat-proj kills a slew of registers
3203       // This can add edges to 'n' and obscure whether or not it was a def,
3204       // hence the is_def flag.
3205       fat_proj_seen = true;
3206       RegMaskIterator rmi(n->out_RegMask());
3207       while (rmi.has_next()) {
3208         OptoReg::Name kill = rmi.next();
3209         anti_do_def(b, n, kill, is_def);
3210       }
3211     } else {
3212       // Get DEF'd registers the normal way
3213       anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
3214       anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
3215     }
3216 
3217     // Kill projections on a branch should appear to occur on the
3218     // branch, not afterwards, so grab the masks from the projections
3219     // and process them.
3220     if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
3221       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3222         Node* use = n->fast_out(i);
3223         if (use->is_Proj()) {
3224           RegMaskIterator rmi(use->out_RegMask());
3225           while (rmi.has_next()) {
3226             OptoReg::Name kill = rmi.next();
3227             anti_do_def(b, n, kill, false);
3228           }
3229         }
3230       }
3231     }
3232 
3233     // Check each register used by this instruction for a following DEF/KILL
3234     // that must occur afterward and requires an anti-dependence edge.
3235     for( uint j=0; j<n->req(); j++ ) {
3236       Node *def = n->in(j);
3237       if( def ) {
3238         assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
3239         anti_do_use( b, n, _regalloc->get_reg_first(def) );
3240         anti_do_use( b, n, _regalloc->get_reg_second(def) );
3241       }
3242     }
3243     // Do not allow defs of new derived values to float above GC
3244     // points unless the base is definitely available at the GC point.
3245 
3246     Node *m = b->get_node(i);
3247 
3248     // Add precedence edge from following safepoint to use of derived pointer
3249     if( last_safept_node != end_node &&
3250         m != last_safept_node) {
3251       for (uint k = 1; k < m->req(); k++) {
3252         const Type *t = m->in(k)->bottom_type();
3253         if( t->isa_oop_ptr() &&
3254             t->is_ptr()->offset() != 0 ) {
3255           last_safept_node->add_prec( m );
3256           break;
3257         }
3258       }
3259 
3260       // Do not allow a CheckCastPP node whose input is a raw pointer to
3261       // float past a safepoint.  This can occur when a buffered inline
3262       // type is allocated in a loop and the CheckCastPP from that
3263       // allocation is reused outside the loop.  If the use inside the
3264       // loop is scalarized the CheckCastPP will no longer be connected
3265       // to the loop safepoint.  See JDK-8264340.
3266       if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CheckCastPP) {
3267         Node *def = m->in(1);
3268         if (def != nullptr && def->bottom_type()->base() == Type::RawPtr) {
3269           last_safept_node->add_prec(m);
3270         }
3271       }
3272     }
3273 
3274     if( n->jvms() ) {           // Precedence edge from derived to safept
3275       // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3276       if( b->get_node(last_safept) != last_safept_node ) {
3277         last_safept = b->find_node(last_safept_node);
3278       }
3279       for( uint j=last_safept; j > i; j-- ) {
3280         Node *mach = b->get_node(j);
3281         if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3282           mach->add_prec( n );
3283       }
3284       last_safept = i;
3285       last_safept_node = m;
3286     }
3287   }
3288 
3289   if (fat_proj_seen) {
3290     // Garbage collect pinch nodes that were not consumed.
3291     // They are usually created by a fat kill MachProj for a call.
3292     garbage_collect_pinch_nodes();
3293   }
3294 }
3295 
3296 // Garbage collect pinch nodes for reuse by other blocks.
3297 //
3298 // The block scheduler's insertion of anti-dependence
3299 // edges creates many pinch nodes when the block contains
3300 // 2 or more Calls.  A pinch node is used to prevent a
3301 // combinatorial explosion of edges.  If a set of kills for a
3302 // register is anti-dependent on a set of uses (or defs), rather
3303 // than adding an edge in the graph between each pair of kill
3304 // and use (or def), a pinch is inserted between them:
3305 //
3306 //            use1   use2  use3
3307 //                \   |   /
3308 //                 \  |  /
3309 //                  pinch
3310 //                 /  |  \
3311 //                /   |   \
3312 //            kill1 kill2 kill3
3313 //
3314 // One pinch node is created per register killed when
3315 // the second call is encountered during a backwards pass
3316 // over the block.  Most of these pinch nodes are never
3317 // wired into the graph because the register is never
3318 // used or def'ed in the block.
3319 //
3320 void Scheduling::garbage_collect_pinch_nodes() {
3321 #ifndef PRODUCT
3322   if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3323 #endif
3324   int trace_cnt = 0;
3325   for (uint k = 0; k < _reg_node.max(); k++) {
3326     Node* pinch = _reg_node[k];
3327     if ((pinch != nullptr) && pinch->Opcode() == Op_Node &&
3328         // no predecence input edges
3329         (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) {
3330       cleanup_pinch(pinch);
3331       _pinch_free_list.push(pinch);
3332       _reg_node.map(k, nullptr);
3333 #ifndef PRODUCT
3334       if (_cfg->C->trace_opto_output()) {
3335         trace_cnt++;
3336         if (trace_cnt > 40) {
3337           tty->print("\n");
3338           trace_cnt = 0;
3339         }
3340         tty->print(" %d", pinch->_idx);
3341       }
3342 #endif
3343     }
3344   }
3345 #ifndef PRODUCT
3346   if (_cfg->C->trace_opto_output()) tty->print("\n");
3347 #endif
3348 }
3349 
3350 // Clean up a pinch node for reuse.
3351 void Scheduling::cleanup_pinch( Node *pinch ) {
3352   assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3353 
3354   for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3355     Node* use = pinch->last_out(i);
3356     uint uses_found = 0;
3357     for (uint j = use->req(); j < use->len(); j++) {
3358       if (use->in(j) == pinch) {
3359         use->rm_prec(j);
3360         uses_found++;
3361       }
3362     }
3363     assert(uses_found > 0, "must be a precedence edge");
3364     i -= uses_found;    // we deleted 1 or more copies of this edge
3365   }
3366   // May have a later_def entry
3367   pinch->set_req(0, nullptr);
3368 }
3369 
3370 #ifndef PRODUCT
3371 
3372 void Scheduling::dump_available() const {
3373   tty->print("#Availist  ");
3374   for (uint i = 0; i < _available.size(); i++)
3375     tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3376   tty->cr();
3377 }
3378 
3379 // Print Scheduling Statistics
3380 void Scheduling::print_statistics() {
3381   // Print the size added by nops for bundling
3382   tty->print("Nops added %d bytes to total of %d bytes",
3383              _total_nop_size, _total_method_size);
3384   if (_total_method_size > 0)
3385     tty->print(", for %.2f%%",
3386                ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3387   tty->print("\n");
3388 
3389   // Print the number of branch shadows filled
3390   if (Pipeline::_branch_has_delay_slot) {
3391     tty->print("Of %d branches, %d had unconditional delay slots filled",
3392                _total_branches, _total_unconditional_delays);
3393     if (_total_branches > 0)
3394       tty->print(", for %.2f%%",
3395                  ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3396     tty->print("\n");
3397   }
3398 
3399   uint total_instructions = 0, total_bundles = 0;
3400 
3401   for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3402     uint bundle_count   = _total_instructions_per_bundle[i];
3403     total_instructions += bundle_count * i;
3404     total_bundles      += bundle_count;
3405   }
3406 
3407   if (total_bundles > 0)
3408     tty->print("Average ILP (excluding nops) is %.2f\n",
3409                ((double)total_instructions) / ((double)total_bundles));
3410 }
3411 #endif
3412 
3413 //-----------------------init_scratch_buffer_blob------------------------------
3414 // Construct a temporary BufferBlob and cache it for this compile.
3415 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3416   // If there is already a scratch buffer blob allocated and the
3417   // constant section is big enough, use it.  Otherwise free the
3418   // current and allocate a new one.
3419   BufferBlob* blob = scratch_buffer_blob();
3420   if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3421     // Use the current blob.
3422   } else {
3423     if (blob != nullptr) {
3424       BufferBlob::free(blob);
3425     }
3426 
3427     ResourceMark rm;
3428     _scratch_const_size = const_size;
3429     int size = C2Compiler::initial_code_buffer_size(const_size);
3430     if (C->has_scalarized_args()) {
3431       // Inline type entry points (MachVEPNodes) require lots of space for GC barriers and oop verification
3432       // when loading object fields from the buffered argument. Increase scratch buffer size accordingly.
3433       ciMethod* method = C->method();
3434       int barrier_size = UseZGC ? 200 : (7 DEBUG_ONLY(+ 37));
3435       int arg_num = 0;
3436       if (!method->is_static()) {
3437         if (method->is_scalarized_arg(arg_num)) {
3438           size += method->holder()->as_inline_klass()->oop_count() * barrier_size;
3439         }
3440         arg_num++;
3441       }
3442       for (ciSignatureStream str(method->signature()); !str.at_return_type(); str.next()) {
3443         if (method->is_scalarized_arg(arg_num)) {
3444           size += str.type()->as_inline_klass()->oop_count() * barrier_size;
3445         }
3446         arg_num++;
3447       }
3448     }
3449     blob = BufferBlob::create("Compile::scratch_buffer", size);
3450     // Record the buffer blob for next time.
3451     set_scratch_buffer_blob(blob);
3452     // Have we run out of code space?
3453     if (scratch_buffer_blob() == nullptr) {
3454       // Let CompilerBroker disable further compilations.
3455       C->record_failure("Not enough space for scratch buffer in CodeCache");
3456       return;
3457     }
3458   }
3459 
3460   // Initialize the relocation buffers
3461   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3462   set_scratch_locs_memory(locs_buf);
3463 }
3464 
3465 
3466 //-----------------------scratch_emit_size-------------------------------------
3467 // Helper function that computes size by emitting code
3468 uint PhaseOutput::scratch_emit_size(const Node* n) {
3469   // Start scratch_emit_size section.
3470   set_in_scratch_emit_size(true);
3471 
3472   // Emit into a trash buffer and count bytes emitted.
3473   // This is a pretty expensive way to compute a size,
3474   // but it works well enough if seldom used.
3475   // All common fixed-size instructions are given a size
3476   // method by the AD file.
3477   // Note that the scratch buffer blob and locs memory are
3478   // allocated at the beginning of the compile task, and
3479   // may be shared by several calls to scratch_emit_size.
3480   // The allocation of the scratch buffer blob is particularly
3481   // expensive, since it has to grab the code cache lock.
3482   BufferBlob* blob = this->scratch_buffer_blob();
3483   assert(blob != nullptr, "Initialize BufferBlob at start");
3484   assert(blob->size() > MAX_inst_size, "sanity");
3485   relocInfo* locs_buf = scratch_locs_memory();
3486   address blob_begin = blob->content_begin();
3487   address blob_end   = (address)locs_buf;
3488   assert(blob->contains(blob_end), "sanity");
3489   CodeBuffer buf(blob_begin, blob_end - blob_begin);
3490   buf.initialize_consts_size(_scratch_const_size);
3491   buf.initialize_stubs_size(MAX_stubs_size);
3492   assert(locs_buf != nullptr, "sanity");
3493   int lsize = MAX_locs_size / 3;
3494   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3495   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3496   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3497   // Mark as scratch buffer.
3498   buf.consts()->set_scratch_emit();
3499   buf.insts()->set_scratch_emit();
3500   buf.stubs()->set_scratch_emit();
3501 
3502   // Do the emission.
3503 
3504   Label fakeL; // Fake label for branch instructions.
3505   Label*   saveL = nullptr;
3506   uint save_bnum = 0;
3507   bool is_branch = n->is_MachBranch();
3508   C2_MacroAssembler masm(&buf);
3509   masm.bind(fakeL);
3510   if (is_branch) {
3511     n->as_MachBranch()->save_label(&saveL, &save_bnum);
3512     n->as_MachBranch()->label_set(&fakeL, 0);
3513   }
3514   n->emit(&masm, C->regalloc());
3515 
3516   // Emitting into the scratch buffer should not fail
3517   assert(!C->failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3518 
3519   // Restore label.
3520   if (is_branch) {
3521     n->as_MachBranch()->label_set(saveL, save_bnum);
3522   }
3523 
3524   // End scratch_emit_size section.
3525   set_in_scratch_emit_size(false);
3526 
3527   return buf.insts_size();
3528 }
3529 
3530 void PhaseOutput::install() {
3531   if (!C->should_install_code()) {
3532     return;
3533   } else if (C->stub_function() != nullptr) {
3534     install_stub(C->stub_name());
3535   } else {
3536     install_code(C->method(),
3537                  C->entry_bci(),
3538                  CompileBroker::compiler2(),
3539                  C->has_unsafe_access(),
3540                  SharedRuntime::is_wide_vector(C->max_vector_size()));
3541   }
3542 }
3543 
3544 void PhaseOutput::install_code(ciMethod*         target,
3545                                int               entry_bci,
3546                                AbstractCompiler* compiler,
3547                                bool              has_unsafe_access,
3548                                bool              has_wide_vectors) {
3549   // Check if we want to skip execution of all compiled code.
3550   {
3551 #ifndef PRODUCT
3552     if (OptoNoExecute) {
3553       C->record_method_not_compilable("+OptoNoExecute");  // Flag as failed
3554       return;
3555     }
3556 #endif
3557     Compile::TracePhase tp(_t_registerMethod);
3558 
3559     if (C->is_osr_compilation()) {
3560       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3561       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3562     } else {
3563       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3564       if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry) == -1) {
3565         _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, _first_block_size);
3566       }
3567       if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry_RO) == -1) {
3568         _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, _first_block_size);
3569       }
3570       if (_code_offsets.value(CodeOffsets::Entry) == -1) {
3571         _code_offsets.set_value(CodeOffsets::Entry, _first_block_size);
3572       }
3573       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3574     }
3575 
3576     C->env()->register_method(target,
3577                               entry_bci,
3578                               &_code_offsets,
3579                               _orig_pc_slot_offset_in_bytes,
3580                               code_buffer(),
3581                               frame_size_in_words(),
3582                               _oop_map_set,
3583                               &_handler_table,
3584                               inc_table(),
3585                               compiler,
3586                               has_unsafe_access,
3587                               SharedRuntime::is_wide_vector(C->max_vector_size()),
3588                               C->has_monitors(),
3589                               C->has_scoped_access(),
3590                               0);
3591 
3592     if (C->log() != nullptr) { // Print code cache state into compiler log
3593       C->log()->code_cache_state();
3594     }
3595   }
3596 }
3597 void PhaseOutput::install_stub(const char* stub_name) {
3598   // Entry point will be accessed using stub_entry_point();
3599   if (code_buffer() == nullptr) {
3600     Matcher::soft_match_failure();
3601   } else {
3602     if (PrintAssembly && (WizardMode || Verbose))
3603       tty->print_cr("### Stub::%s", stub_name);
3604 
3605     if (!C->failing()) {
3606       assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3607 
3608       // Make the NMethod
3609       // For now we mark the frame as never safe for profile stackwalking
3610       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3611                                                       code_buffer(),
3612                                                       CodeOffsets::frame_never_safe,
3613                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
3614                                                       frame_size_in_words(),
3615                                                       oop_map_set(),
3616                                                       false,
3617                                                       false);
3618 
3619       if (rs == nullptr) {
3620         C->record_failure("CodeCache is full");
3621       } else {
3622         assert(rs->is_runtime_stub(), "sanity check");
3623         C->set_stub_entry_point(rs->entry_point());
3624         AOTCodeCache::store_code_blob(*rs, AOTCodeEntry::C2Blob, C->stub_id(), stub_name);
3625       }
3626     }
3627   }
3628 }
3629 
3630 // Support for bundling info
3631 Bundle* PhaseOutput::node_bundling(const Node *n) {
3632   assert(valid_bundle_info(n), "oob");
3633   return &_node_bundling_base[n->_idx];
3634 }
3635 
3636 bool PhaseOutput::valid_bundle_info(const Node *n) {
3637   return (_node_bundling_limit > n->_idx);
3638 }
3639 
3640 //------------------------------frame_size_in_words-----------------------------
3641 // frame_slots in units of words
3642 int PhaseOutput::frame_size_in_words() const {
3643   // shift is 0 in LP32 and 1 in LP64
3644   const int shift = (LogBytesPerWord - LogBytesPerInt);
3645   int words = _frame_slots >> shift;
3646   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3647   return words;
3648 }
3649 
3650 // To bang the stack of this compiled method we use the stack size
3651 // that the interpreter would need in case of a deoptimization. This
3652 // removes the need to bang the stack in the deoptimization blob which
3653 // in turn simplifies stack overflow handling.
3654 int PhaseOutput::bang_size_in_bytes() const {
3655   return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3656 }
3657 
3658 //------------------------------dump_asm---------------------------------------
3659 // Dump formatted assembly
3660 #if defined(SUPPORT_OPTO_ASSEMBLY)
3661 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3662 
3663   int pc_digits = 3; // #chars required for pc
3664   int sb_chars  = 3; // #chars for "start bundle" indicator
3665   int tab_size  = 8;
3666   if (pcs != nullptr) {
3667     int max_pc = 0;
3668     for (uint i = 0; i < pc_limit; i++) {
3669       max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3670     }
3671     pc_digits  = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3672   }
3673   int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3674 
3675   bool cut_short = false;
3676   st->print_cr("#");
3677   st->print("#  ");  C->tf()->dump_on(st);  st->cr();
3678   st->print_cr("#");
3679 
3680   // For all blocks
3681   int pc = 0x0;                 // Program counter
3682   char starts_bundle = ' ';
3683   C->regalloc()->dump_frame();
3684 
3685   Node *n = nullptr;
3686   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3687     if (VMThread::should_terminate()) {
3688       cut_short = true;
3689       break;
3690     }
3691     Block* block = C->cfg()->get_block(i);
3692     if (block->is_connector() && !Verbose) {
3693       continue;
3694     }
3695     n = block->head();
3696     if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3697       pc = pcs[n->_idx];
3698       st->print("%*.*x", pc_digits, pc_digits, pc);
3699     }
3700     st->fill_to(prefix_len);
3701     block->dump_head(C->cfg(), st);
3702     if (block->is_connector()) {
3703       st->fill_to(prefix_len);
3704       st->print_cr("# Empty connector block");
3705     } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3706       st->fill_to(prefix_len);
3707       st->print_cr("# Block is sole successor of call");
3708     }
3709 
3710     // For all instructions
3711     Node *delay = nullptr;
3712     for (uint j = 0; j < block->number_of_nodes(); j++) {
3713       if (VMThread::should_terminate()) {
3714         cut_short = true;
3715         break;
3716       }
3717       n = block->get_node(j);
3718       if (valid_bundle_info(n)) {
3719         Bundle* bundle = node_bundling(n);
3720         if (bundle->used_in_unconditional_delay()) {
3721           delay = n;
3722           continue;
3723         }
3724         if (bundle->starts_bundle()) {
3725           starts_bundle = '+';
3726         }
3727       }
3728 
3729       if (WizardMode) {
3730         n->dump();
3731       }
3732 
3733       if( !n->is_Region() &&    // Dont print in the Assembly
3734           !n->is_Phi() &&       // a few noisely useless nodes
3735           !n->is_Proj() &&
3736           !n->is_MachTemp() &&
3737           !n->is_SafePointScalarObject() &&
3738           !n->is_Catch() &&     // Would be nice to print exception table targets
3739           !n->is_MergeMem() &&  // Not very interesting
3740           !n->is_top() &&       // Debug info table constants
3741           !(n->is_Con() && !n->is_Mach())// Debug info table constants
3742           ) {
3743         if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3744           pc = pcs[n->_idx];
3745           st->print("%*.*x", pc_digits, pc_digits, pc);
3746         } else {
3747           st->fill_to(pc_digits);
3748         }
3749         st->print(" %c ", starts_bundle);
3750         starts_bundle = ' ';
3751         st->fill_to(prefix_len);
3752         n->format(C->regalloc(), st);
3753         st->cr();
3754       }
3755 
3756       // If we have an instruction with a delay slot, and have seen a delay,
3757       // then back up and print it
3758       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3759         // Coverity finding - Explicit null dereferenced.
3760         guarantee(delay != nullptr, "no unconditional delay instruction");
3761         if (WizardMode) delay->dump();
3762 
3763         if (node_bundling(delay)->starts_bundle())
3764           starts_bundle = '+';
3765         if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3766           pc = pcs[n->_idx];
3767           st->print("%*.*x", pc_digits, pc_digits, pc);
3768         } else {
3769           st->fill_to(pc_digits);
3770         }
3771         st->print(" %c ", starts_bundle);
3772         starts_bundle = ' ';
3773         st->fill_to(prefix_len);
3774         delay->format(C->regalloc(), st);
3775         st->cr();
3776         delay = nullptr;
3777       }
3778 
3779       // Dump the exception table as well
3780       if( n->is_Catch() && (Verbose || WizardMode) ) {
3781         // Print the exception table for this offset
3782         _handler_table.print_subtable_for(pc);
3783       }
3784       st->bol(); // Make sure we start on a new line
3785     }
3786     st->cr(); // one empty line between blocks
3787     assert(cut_short || delay == nullptr, "no unconditional delay branch");
3788   } // End of per-block dump
3789 
3790   if (cut_short)  st->print_cr("*** disassembly is cut short ***");
3791 }
3792 #endif
3793 
3794 #ifndef PRODUCT
3795 void PhaseOutput::print_statistics() {
3796   Scheduling::print_statistics();
3797 }
3798 #endif