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