1 /*
   2  * Copyright (c) 1998, 2025, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "asm/assembler.inline.hpp"
  26 #include "code/aotCodeCache.hpp"
  27 #include "code/compiledIC.hpp"
  28 #include "code/debugInfo.hpp"
  29 #include "code/debugInfoRec.hpp"
  30 #include "compiler/compileBroker.hpp"
  31 #include "compiler/compilerDirectives.hpp"
  32 #include "compiler/disassembler.hpp"
  33 #include "compiler/oopMap.hpp"
  34 #include "gc/shared/barrierSet.hpp"
  35 #include "gc/shared/c2/barrierSetC2.hpp"
  36 #include "memory/allocation.hpp"
  37 #include "opto/ad.hpp"
  38 #include "opto/block.hpp"
  39 #include "opto/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     _orig_pc_slot = C->fixed_slots() - (sizeof(address) / VMRegImpl::stack_slot_size);
 245   }
 246 }
 247 
 248 PhaseOutput::~PhaseOutput() {
 249   C->set_output(nullptr);
 250   if (_scratch_buffer_blob != nullptr) {
 251     BufferBlob::free(_scratch_buffer_blob);
 252   }
 253 }
 254 
 255 void PhaseOutput::perform_mach_node_analysis() {
 256   // Late barrier analysis must be done after schedule and bundle
 257   // Otherwise liveness based spilling will fail
 258   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 259   bs->late_barrier_analysis();
 260 
 261   pd_perform_mach_node_analysis();
 262 
 263   C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
 264 }
 265 
 266 // Convert Nodes to instruction bits and pass off to the VM
 267 void PhaseOutput::Output() {
 268   // RootNode goes
 269   assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
 270 
 271   // The number of new nodes (mostly MachNop) is proportional to
 272   // the number of java calls and inner loops which are aligned.
 273   if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
 274                             C->inner_loops()*(OptoLoopAlignment-1)),
 275                            "out of nodes before code generation" ) ) {
 276     return;
 277   }
 278   // Make sure I can find the Start Node
 279   Block *entry = C->cfg()->get_block(1);
 280   Block *broot = C->cfg()->get_root_block();
 281 
 282   const StartNode *start = entry->head()->as_Start();
 283 
 284   // Replace StartNode with prolog
 285   MachPrologNode *prolog = new MachPrologNode();
 286   entry->map_node(prolog, 0);
 287   C->cfg()->map_node_to_block(prolog, entry);
 288   C->cfg()->unmap_node_from_block(start); // start is no longer in any block
 289 
 290   // Virtual methods need an unverified entry point
 291 
 292   if( C->is_osr_compilation() ) {
 293     if( PoisonOSREntry ) {
 294       // TODO: Should use a ShouldNotReachHereNode...
 295       C->cfg()->insert( broot, 0, new MachBreakpointNode() );
 296     }
 297   } else {
 298     if( C->method() && !C->method()->flags().is_static() ) {
 299       // Insert unvalidated entry point
 300       C->cfg()->insert( broot, 0, new MachUEPNode() );
 301     }
 302 
 303   }
 304 
 305   // Break before main entry point
 306   if ((C->method() && C->directive()->BreakAtExecuteOption) ||
 307       (OptoBreakpoint && C->is_method_compilation())       ||
 308       (OptoBreakpointOSR && C->is_osr_compilation())       ||
 309       (OptoBreakpointC2R && !C->method())                   ) {
 310     // checking for C->method() means that OptoBreakpoint does not apply to
 311     // runtime stubs or frame converters
 312     C->cfg()->insert( entry, 1, new MachBreakpointNode() );
 313   }
 314 
 315   // Insert epilogs before every return
 316   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
 317     Block* block = C->cfg()->get_block(i);
 318     if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
 319       Node* m = block->end();
 320       if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
 321         MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
 322         block->add_inst(epilog);
 323         C->cfg()->map_node_to_block(epilog, block);
 324       }
 325     }
 326   }
 327 
 328   // Keeper of sizing aspects
 329   _buf_sizes = BufferSizingData();
 330 
 331   // Initialize code buffer
 332   estimate_buffer_size(_buf_sizes._const);
 333   if (C->failing()) return;
 334 
 335   // Pre-compute the length of blocks and replace
 336   // long branches with short if machine supports it.
 337   // Must be done before ScheduleAndBundle due to SPARC delay slots
 338   uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
 339   blk_starts[0] = 0;
 340   shorten_branches(blk_starts);
 341 
 342   ScheduleAndBundle();
 343   if (C->failing()) {
 344     return;
 345   }
 346 
 347   perform_mach_node_analysis();
 348 
 349   // Complete sizing of codebuffer
 350   CodeBuffer* cb = init_buffer();
 351   if (cb == nullptr || C->failing()) {
 352     return;
 353   }
 354 
 355   BuildOopMaps();
 356 
 357   if (C->failing())  {
 358     return;
 359   }
 360 
 361   C2_MacroAssembler masm(cb);
 362   fill_buffer(&masm, blk_starts);
 363 }
 364 
 365 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
 366   // Determine if we need to generate a stack overflow check.
 367   // Do it if the method is not a stub function and
 368   // has java calls or has frame size > vm_page_size/8.
 369   // The debug VM checks that deoptimization doesn't trigger an
 370   // unexpected stack overflow (compiled method stack banging should
 371   // guarantee it doesn't happen) so we always need the stack bang in
 372   // a debug VM.
 373   return (C->stub_function() == nullptr &&
 374           (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
 375            DEBUG_ONLY(|| true)));
 376 }
 377 
 378 bool PhaseOutput::need_register_stack_bang() const {
 379   // Determine if we need to generate a register stack overflow check.
 380   // This is only used on architectures which have split register
 381   // and memory stacks.
 382   // Bang if the method is not a stub function and has java calls
 383   return (C->stub_function() == nullptr && C->has_java_calls());
 384 }
 385 
 386 
 387 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
 388 // of a loop. When aligning a loop we need to provide enough instructions
 389 // in cpu's fetch buffer to feed decoders. The loop alignment could be
 390 // avoided if we have enough instructions in fetch buffer at the head of a loop.
 391 // By default, the size is set to 999999 by Block's constructor so that
 392 // a loop will be aligned if the size is not reset here.
 393 //
 394 // Note: Mach instructions could contain several HW instructions
 395 // so the size is estimated only.
 396 //
 397 void PhaseOutput::compute_loop_first_inst_sizes() {
 398   // The next condition is used to gate the loop alignment optimization.
 399   // Don't aligned a loop if there are enough instructions at the head of a loop
 400   // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
 401   // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
 402   // equal to 11 bytes which is the largest address NOP instruction.
 403   if (MaxLoopPad < OptoLoopAlignment - 1) {
 404     uint last_block = C->cfg()->number_of_blocks() - 1;
 405     for (uint i = 1; i <= last_block; i++) {
 406       Block* block = C->cfg()->get_block(i);
 407       // Check the first loop's block which requires an alignment.
 408       if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
 409         uint sum_size = 0;
 410         uint inst_cnt = NumberOfLoopInstrToAlign;
 411         inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
 412 
 413         // Check subsequent fallthrough blocks if the loop's first
 414         // block(s) does not have enough instructions.
 415         Block *nb = block;
 416         while(inst_cnt > 0 &&
 417               i < last_block &&
 418               !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
 419               !nb->has_successor(block)) {
 420           i++;
 421           nb = C->cfg()->get_block(i);
 422           inst_cnt  = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
 423         } // while( inst_cnt > 0 && i < last_block  )
 424 
 425         block->set_first_inst_size(sum_size);
 426       } // f( b->head()->is_Loop() )
 427     } // for( i <= last_block )
 428   } // if( MaxLoopPad < OptoLoopAlignment-1 )
 429 }
 430 
 431 // The architecture description provides short branch variants for some long
 432 // branch instructions. Replace eligible long branches with short branches.
 433 void PhaseOutput::shorten_branches(uint* blk_starts) {
 434 
 435   Compile::TracePhase tp(_t_shortenBranches);
 436 
 437   // Compute size of each block, method size, and relocation information size
 438   uint nblocks  = C->cfg()->number_of_blocks();
 439 
 440   uint*      jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
 441   uint*      jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
 442   int*       jmp_nidx   = NEW_RESOURCE_ARRAY(int ,nblocks);
 443 
 444   // Collect worst case block paddings
 445   int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
 446   memset(block_worst_case_pad, 0, nblocks * sizeof(int));
 447 
 448   DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
 449   DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
 450 
 451   bool has_short_branch_candidate = false;
 452 
 453   // Initialize the sizes to 0
 454   int code_size  = 0;          // Size in bytes of generated code
 455   int stub_size  = 0;          // Size in bytes of all stub entries
 456   // Size in bytes of all relocation entries, including those in local stubs.
 457   // Start with 2-bytes of reloc info for the unvalidated entry point
 458   int reloc_size = 1;          // Number of relocation entries
 459 
 460   // Make three passes.  The first computes pessimistic blk_starts,
 461   // relative jmp_offset and reloc_size information.  The second performs
 462   // short branch substitution using the pessimistic sizing.  The
 463   // third inserts nops where needed.
 464 
 465   // Step one, perform a pessimistic sizing pass.
 466   uint last_call_adr = max_juint;
 467   uint last_avoid_back_to_back_adr = max_juint;
 468   uint nop_size = (new MachNopNode())->size(C->regalloc());
 469   for (uint i = 0; i < nblocks; i++) { // For all blocks
 470     Block* block = C->cfg()->get_block(i);
 471     _block = block;
 472 
 473     // During short branch replacement, we store the relative (to blk_starts)
 474     // offset of jump in jmp_offset, rather than the absolute offset of jump.
 475     // This is so that we do not need to recompute sizes of all nodes when
 476     // we compute correct blk_starts in our next sizing pass.
 477     jmp_offset[i] = 0;
 478     jmp_size[i]   = 0;
 479     jmp_nidx[i]   = -1;
 480     DEBUG_ONLY( jmp_target[i] = 0; )
 481     DEBUG_ONLY( jmp_rule[i]   = 0; )
 482 
 483     // Sum all instruction sizes to compute block size
 484     uint last_inst = block->number_of_nodes();
 485     uint blk_size = 0;
 486     for (uint j = 0; j < last_inst; j++) {
 487       _index = j;
 488       Node* nj = block->get_node(_index);
 489       // Handle machine instruction nodes
 490       if (nj->is_Mach()) {
 491         MachNode* mach = nj->as_Mach();
 492         blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
 493         reloc_size += mach->reloc();
 494         if (mach->is_MachCall()) {
 495           // add size information for trampoline stub
 496           // class CallStubImpl is platform-specific and defined in the *.ad files.
 497           stub_size  += CallStubImpl::size_call_trampoline();
 498           reloc_size += CallStubImpl::reloc_call_trampoline();
 499 
 500           MachCallNode *mcall = mach->as_MachCall();
 501           // This destination address is NOT PC-relative
 502 
 503           mcall->method_set((intptr_t)mcall->entry_point());
 504 
 505           if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
 506             stub_size  += CompiledDirectCall::to_interp_stub_size();
 507             reloc_size += CompiledDirectCall::reloc_to_interp_stub();
 508           }
 509         } else if (mach->is_MachSafePoint()) {
 510           // If call/safepoint are adjacent, account for possible
 511           // nop to disambiguate the two safepoints.
 512           // ScheduleAndBundle() can rearrange nodes in a block,
 513           // check for all offsets inside this block.
 514           if (last_call_adr >= blk_starts[i]) {
 515             blk_size += nop_size;
 516           }
 517         }
 518         if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
 519           // Nop is inserted between "avoid back to back" instructions.
 520           // ScheduleAndBundle() can rearrange nodes in a block,
 521           // check for all offsets inside this block.
 522           if (last_avoid_back_to_back_adr >= blk_starts[i]) {
 523             blk_size += nop_size;
 524           }
 525         }
 526         if (mach->may_be_short_branch()) {
 527           if (!nj->is_MachBranch()) {
 528 #ifndef PRODUCT
 529             nj->dump(3);
 530 #endif
 531             Unimplemented();
 532           }
 533           assert(jmp_nidx[i] == -1, "block should have only one branch");
 534           jmp_offset[i] = blk_size;
 535           jmp_size[i]   = nj->size(C->regalloc());
 536           jmp_nidx[i]   = j;
 537           has_short_branch_candidate = true;
 538         }
 539       }
 540       blk_size += nj->size(C->regalloc());
 541       // Remember end of call offset
 542       if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
 543         last_call_adr = blk_starts[i]+blk_size;
 544       }
 545       // Remember end of avoid_back_to_back offset
 546       if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
 547         last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
 548       }
 549     }
 550 
 551     // When the next block starts a loop, we may insert pad NOP
 552     // instructions.  Since we cannot know our future alignment,
 553     // assume the worst.
 554     if (i < nblocks - 1) {
 555       Block* nb = C->cfg()->get_block(i + 1);
 556       int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
 557       if (max_loop_pad > 0) {
 558         assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
 559         // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
 560         // If either is the last instruction in this block, bump by
 561         // max_loop_pad in lock-step with blk_size, so sizing
 562         // calculations in subsequent blocks still can conservatively
 563         // detect that it may the last instruction in this block.
 564         if (last_call_adr == blk_starts[i]+blk_size) {
 565           last_call_adr += max_loop_pad;
 566         }
 567         if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
 568           last_avoid_back_to_back_adr += max_loop_pad;
 569         }
 570         blk_size += max_loop_pad;
 571         block_worst_case_pad[i + 1] = max_loop_pad;
 572       }
 573     }
 574 
 575     // Save block size; update total method size
 576     blk_starts[i+1] = blk_starts[i]+blk_size;
 577   }
 578 
 579   // Step two, replace eligible long jumps.
 580   bool progress = true;
 581   uint last_may_be_short_branch_adr = max_juint;
 582   while (has_short_branch_candidate && progress) {
 583     progress = false;
 584     has_short_branch_candidate = false;
 585     int adjust_block_start = 0;
 586     for (uint i = 0; i < nblocks; i++) {
 587       Block* block = C->cfg()->get_block(i);
 588       int idx = jmp_nidx[i];
 589       MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach();
 590       if (mach != nullptr && mach->may_be_short_branch()) {
 591 #ifdef ASSERT
 592         assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
 593         int j;
 594         // Find the branch; ignore trailing NOPs.
 595         for (j = block->number_of_nodes()-1; j>=0; j--) {
 596           Node* n = block->get_node(j);
 597           if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
 598             break;
 599         }
 600         assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
 601 #endif
 602         int br_size = jmp_size[i];
 603         int br_offs = blk_starts[i] + jmp_offset[i];
 604 
 605         // This requires the TRUE branch target be in succs[0]
 606         uint bnum = block->non_connector_successor(0)->_pre_order;
 607         int offset = blk_starts[bnum] - br_offs;
 608         if (bnum > i) { // adjust following block's offset
 609           offset -= adjust_block_start;
 610         }
 611 
 612         // This block can be a loop header, account for the padding
 613         // in the previous block.
 614         int block_padding = block_worst_case_pad[i];
 615         assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
 616         // In the following code a nop could be inserted before
 617         // the branch which will increase the backward distance.
 618         bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
 619         assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
 620 
 621         if (needs_padding && offset <= 0)
 622           offset -= nop_size;
 623 
 624         if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
 625           // We've got a winner.  Replace this branch.
 626           MachNode* replacement = mach->as_MachBranch()->short_branch_version();
 627 
 628           // Update the jmp_size.
 629           int new_size = replacement->size(C->regalloc());
 630           int diff     = br_size - new_size;
 631           assert(diff >= (int)nop_size, "short_branch size should be smaller");
 632           // Conservatively take into account padding between
 633           // avoid_back_to_back branches. Previous branch could be
 634           // converted into avoid_back_to_back branch during next
 635           // rounds.
 636           if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
 637             jmp_offset[i] += nop_size;
 638             diff -= nop_size;
 639           }
 640           adjust_block_start += diff;
 641           block->map_node(replacement, idx);
 642           mach->subsume_by(replacement, C);
 643           mach = replacement;
 644           progress = true;
 645 
 646           jmp_size[i] = new_size;
 647           DEBUG_ONLY( jmp_target[i] = bnum; );
 648           DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
 649         } else {
 650           // The jump distance is not short, try again during next iteration.
 651           has_short_branch_candidate = true;
 652         }
 653       } // (mach->may_be_short_branch())
 654       if (mach != nullptr && (mach->may_be_short_branch() ||
 655                            mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
 656         last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
 657       }
 658       blk_starts[i+1] -= adjust_block_start;
 659     }
 660   }
 661 
 662 #ifdef ASSERT
 663   for (uint i = 0; i < nblocks; i++) { // For all blocks
 664     if (jmp_target[i] != 0) {
 665       int br_size = jmp_size[i];
 666       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
 667       if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
 668         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]);
 669       }
 670       assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
 671     }
 672   }
 673 #endif
 674 
 675   // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
 676   // after ScheduleAndBundle().
 677 
 678   // ------------------
 679   // Compute size for code buffer
 680   code_size = blk_starts[nblocks];
 681 
 682   // Relocation records
 683   reloc_size += 1;              // Relo entry for exception handler
 684 
 685   // Adjust reloc_size to number of record of relocation info
 686   // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
 687   // a relocation index.
 688   // The CodeBuffer will expand the locs array if this estimate is too low.
 689   reloc_size *= 10 / sizeof(relocInfo);
 690 
 691   _buf_sizes._reloc = reloc_size;
 692   _buf_sizes._code  = code_size;
 693   _buf_sizes._stub  = stub_size;
 694 }
 695 
 696 //------------------------------FillLocArray-----------------------------------
 697 // Create a bit of debug info and append it to the array.  The mapping is from
 698 // Java local or expression stack to constant, register or stack-slot.  For
 699 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
 700 // entry has been taken care of and caller should skip it).
 701 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
 702   // This should never have accepted Bad before
 703   assert(OptoReg::is_valid(regnum), "location must be valid");
 704   return (OptoReg::is_reg(regnum))
 705          ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
 706          : new LocationValue(Location::new_stk_loc(l_type,  ra->reg2offset(regnum)));
 707 }
 708 
 709 
 710 ObjectValue*
 711 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
 712   for (int i = 0; i < objs->length(); i++) {
 713     assert(objs->at(i)->is_object(), "corrupt object cache");
 714     ObjectValue* sv = objs->at(i)->as_ObjectValue();
 715     if (sv->id() == id) {
 716       return sv;
 717     }
 718   }
 719   // Otherwise..
 720   return nullptr;
 721 }
 722 
 723 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
 724                                      ObjectValue* sv ) {
 725   assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition");
 726   objs->append(sv);
 727 }
 728 
 729 
 730 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
 731                             GrowableArray<ScopeValue*> *array,
 732                             GrowableArray<ScopeValue*> *objs ) {
 733   assert( local, "use _top instead of null" );
 734   if (array->length() != idx) {
 735     assert(array->length() == idx + 1, "Unexpected array count");
 736     // Old functionality:
 737     //   return
 738     // New functionality:
 739     //   Assert if the local is not top. In product mode let the new node
 740     //   override the old entry.
 741     assert(local == C->top(), "LocArray collision");
 742     if (local == C->top()) {
 743       return;
 744     }
 745     array->pop();
 746   }
 747   const Type *t = local->bottom_type();
 748 
 749   // Is it a safepoint scalar object node?
 750   if (local->is_SafePointScalarObject()) {
 751     SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
 752 
 753     ObjectValue* sv = sv_for_node_id(objs, spobj->_idx);
 754     if (sv == nullptr) {
 755       ciKlass* cik = t->is_oopptr()->exact_klass();
 756       assert(cik->is_instance_klass() ||
 757              cik->is_array_klass(), "Not supported allocation.");
 758       sv = new ObjectValue(spobj->_idx,
 759                            new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
 760       set_sv_for_object_node(objs, sv);
 761 
 762       uint first_ind = spobj->first_index(sfpt->jvms());
 763       for (uint i = 0; i < spobj->n_fields(); i++) {
 764         Node* fld_node = sfpt->in(first_ind+i);
 765         (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
 766       }
 767     }
 768     array->append(sv);
 769     return;
 770   } else if (local->is_SafePointScalarMerge()) {
 771     SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
 772     ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
 773 
 774     if (mv == nullptr) {
 775       GrowableArray<ScopeValue*> deps;
 776 
 777       int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
 778       (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
 779       assert(deps.length() == 1, "missing value");
 780 
 781       int selector_idx = smerge->selector_idx(sfpt->jvms());
 782       (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
 783       assert(deps.length() == 2, "missing value");
 784 
 785       mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
 786       set_sv_for_object_node(objs, mv);
 787 
 788       for (uint i = 1; i < smerge->req(); i++) {
 789         Node* obj_node = smerge->in(i);
 790         int idx = mv->possible_objects()->length();
 791         (void)FillLocArray(idx, sfpt, obj_node, mv->possible_objects(), objs);
 792 
 793         // By default ObjectValues that are in 'possible_objects' are not root objects.
 794         // They will be marked as root later if they are directly referenced in a JVMS.
 795         assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
 796         assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
 797         mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
 798       }
 799     }
 800     array->append(mv);
 801     return;
 802   }
 803 
 804   // Grab the register number for the local
 805   OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
 806   if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
 807     // Record the double as two float registers.
 808     // The register mask for such a value always specifies two adjacent
 809     // float registers, with the lower register number even.
 810     // Normally, the allocation of high and low words to these registers
 811     // is irrelevant, because nearly all operations on register pairs
 812     // (e.g., StoreD) treat them as a single unit.
 813     // Here, we assume in addition that the words in these two registers
 814     // stored "naturally" (by operations like StoreD and double stores
 815     // within the interpreter) such that the lower-numbered register
 816     // is written to the lower memory address.  This may seem like
 817     // a machine dependency, but it is not--it is a requirement on
 818     // the author of the <arch>.ad file to ensure that, for every
 819     // even/odd double-register pair to which a double may be allocated,
 820     // the word in the even single-register is stored to the first
 821     // memory word.  (Note that register numbers are completely
 822     // arbitrary, and are not tied to any machine-level encodings.)
 823 #ifdef _LP64
 824     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
 825       array->append(new ConstantIntValue((jint)0));
 826       array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
 827     } else if ( t->base() == Type::Long ) {
 828       array->append(new ConstantIntValue((jint)0));
 829       array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
 830     } else if ( t->base() == Type::RawPtr ) {
 831       // jsr/ret return address which must be restored into the full
 832       // width 64-bit stack slot.
 833       array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
 834     }
 835 #else //_LP64
 836     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
 837       // Repack the double/long as two jints.
 838       // The convention the interpreter uses is that the second local
 839       // holds the first raw word of the native double representation.
 840       // This is actually reasonable, since locals and stack arrays
 841       // grow downwards in all implementations.
 842       // (If, on some machine, the interpreter's Java locals or stack
 843       // were to grow upwards, the embedded doubles would be word-swapped.)
 844       array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
 845       array->append(new_loc_value( C->regalloc(),              regnum   , Location::normal ));
 846     }
 847 #endif //_LP64
 848     else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
 849              OptoReg::is_reg(regnum) ) {
 850       array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
 851                                                       ? Location::float_in_dbl : Location::normal ));
 852     } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
 853       array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
 854                                                       ? Location::int_in_long : Location::normal ));
 855     } else if( t->base() == Type::NarrowOop ) {
 856       array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
 857     } else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
 858                t->base() == Type::VectorD || t->base() == Type::VectorX ||
 859                t->base() == Type::VectorY || t->base() == Type::VectorZ) {
 860       array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
 861     } else if (C->regalloc()->is_oop(local)) {
 862       assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
 863              t->base() == Type::AryPtr,
 864              "Unexpected type: %s", t->msg());
 865       array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
 866     } else {
 867       assert(t->base() == Type::Int || t->base() == Type::Half ||
 868              t->base() == Type::FloatCon || t->base() == Type::FloatBot,
 869              "Unexpected type: %s", t->msg());
 870       array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
 871     }
 872     return;
 873   }
 874 
 875   // No register.  It must be constant data.
 876   switch (t->base()) {
 877     case Type::Half:              // Second half of a double
 878       ShouldNotReachHere();       // Caller should skip 2nd halves
 879       break;
 880     case Type::AnyPtr:
 881       array->append(new ConstantOopWriteValue(nullptr));
 882       break;
 883     case Type::AryPtr:
 884     case Type::InstPtr:          // fall through
 885       array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
 886       break;
 887     case Type::NarrowOop:
 888       if (t == TypeNarrowOop::NULL_PTR) {
 889         array->append(new ConstantOopWriteValue(nullptr));
 890       } else {
 891         array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
 892       }
 893       break;
 894     case Type::Int:
 895       array->append(new ConstantIntValue(t->is_int()->get_con()));
 896       break;
 897     case Type::RawPtr:
 898       // A return address (T_ADDRESS).
 899       assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
 900 #ifdef _LP64
 901       // Must be restored to the full-width 64-bit stack slot.
 902       array->append(new ConstantLongValue(t->is_ptr()->get_con()));
 903 #else
 904       array->append(new ConstantIntValue(t->is_ptr()->get_con()));
 905 #endif
 906       break;
 907     case Type::FloatCon: {
 908       float f = t->is_float_constant()->getf();
 909       array->append(new ConstantIntValue(jint_cast(f)));
 910       break;
 911     }
 912     case Type::DoubleCon: {
 913       jdouble d = t->is_double_constant()->getd();
 914 #ifdef _LP64
 915       array->append(new ConstantIntValue((jint)0));
 916       array->append(new ConstantDoubleValue(d));
 917 #else
 918       // Repack the double as two jints.
 919     // The convention the interpreter uses is that the second local
 920     // holds the first raw word of the native double representation.
 921     // This is actually reasonable, since locals and stack arrays
 922     // grow downwards in all implementations.
 923     // (If, on some machine, the interpreter's Java locals or stack
 924     // were to grow upwards, the embedded doubles would be word-swapped.)
 925     jlong_accessor acc;
 926     acc.long_value = jlong_cast(d);
 927     array->append(new ConstantIntValue(acc.words[1]));
 928     array->append(new ConstantIntValue(acc.words[0]));
 929 #endif
 930       break;
 931     }
 932     case Type::Long: {
 933       jlong d = t->is_long()->get_con();
 934 #ifdef _LP64
 935       array->append(new ConstantIntValue((jint)0));
 936       array->append(new ConstantLongValue(d));
 937 #else
 938       // Repack the long as two jints.
 939     // The convention the interpreter uses is that the second local
 940     // holds the first raw word of the native double representation.
 941     // This is actually reasonable, since locals and stack arrays
 942     // grow downwards in all implementations.
 943     // (If, on some machine, the interpreter's Java locals or stack
 944     // were to grow upwards, the embedded doubles would be word-swapped.)
 945     jlong_accessor acc;
 946     acc.long_value = d;
 947     array->append(new ConstantIntValue(acc.words[1]));
 948     array->append(new ConstantIntValue(acc.words[0]));
 949 #endif
 950       break;
 951     }
 952     case Type::Top:               // Add an illegal value here
 953       array->append(new LocationValue(Location()));
 954       break;
 955     default:
 956       ShouldNotReachHere();
 957       break;
 958   }
 959 }
 960 
 961 // Determine if this node starts a bundle
 962 bool PhaseOutput::starts_bundle(const Node *n) const {
 963   return (_node_bundling_limit > n->_idx &&
 964           _node_bundling_base[n->_idx].starts_bundle());
 965 }
 966 
 967 // Determine if there is a monitor that has 'ov' as its owner.
 968 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const {
 969   for (int k = 0; k < monarray->length(); k++) {
 970     MonitorValue* mv = monarray->at(k);
 971     if (mv->owner() == ov) {
 972       return true;
 973     }
 974   }
 975 
 976   return false;
 977 }
 978 
 979 // Determine if there is a scalar replaced object description represented by 'ov'.
 980 bool PhaseOutput::contains_as_scalarized_obj(JVMState* jvms, MachSafePointNode* sfn,
 981                                              GrowableArray<ScopeValue*>* objs,
 982                                              ObjectValue* ov) const {
 983   for (int i = 0; i < jvms->scl_size(); i++) {
 984     Node* n = sfn->scalarized_obj(jvms, i);
 985     // Other kinds of nodes that we may encounter here, for instance constants
 986     // representing values of fields of objects scalarized, aren't relevant for
 987     // us, since they don't map to ObjectValue.
 988     if (!n->is_SafePointScalarObject()) {
 989       continue;
 990     }
 991 
 992     ObjectValue* other = sv_for_node_id(objs, n->_idx);
 993     if (ov == other) {
 994       return true;
 995     }
 996   }
 997   return false;
 998 }
 999 
1000 //--------------------------Process_OopMap_Node--------------------------------
1001 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
1002   // Handle special safepoint nodes for synchronization
1003   MachSafePointNode *sfn   = mach->as_MachSafePoint();
1004   MachCallNode      *mcall;
1005 
1006   int safepoint_pc_offset = current_offset;
1007   bool is_method_handle_invoke = false;
1008   bool return_oop = false;
1009   bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
1010   bool arg_escape = false;
1011 
1012   // Add the safepoint in the DebugInfoRecorder
1013   if( !mach->is_MachCall() ) {
1014     mcall = nullptr;
1015     C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1016   } else {
1017     mcall = mach->as_MachCall();
1018 
1019     // Is the call a MethodHandle call?
1020     if (mcall->is_MachCallJava()) {
1021       if (mcall->as_MachCallJava()->_method_handle_invoke) {
1022         assert(C->has_method_handle_invokes(), "must have been set during call generation");
1023         is_method_handle_invoke = true;
1024       }
1025       arg_escape = mcall->as_MachCallJava()->_arg_escape;
1026     }
1027 
1028     // Check if a call returns an object.
1029     if (mcall->returns_pointer()) {
1030       return_oop = true;
1031     }
1032     safepoint_pc_offset += mcall->ret_addr_offset();
1033     C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1034   }
1035 
1036   // Loop over the JVMState list to add scope information
1037   // Do not skip safepoints with a null method, they need monitor info
1038   JVMState* youngest_jvms = sfn->jvms();
1039   int max_depth = youngest_jvms->depth();
1040 
1041   // Allocate the object pool for scalar-replaced objects -- the map from
1042   // small-integer keys (which can be recorded in the local and ostack
1043   // arrays) to descriptions of the object state.
1044   GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1045 
1046   // Visit scopes from oldest to youngest.
1047   for (int depth = 1; depth <= max_depth; depth++) {
1048     JVMState* jvms = youngest_jvms->of_depth(depth);
1049     int idx;
1050     ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1051     // Safepoints that do not have method() set only provide oop-map and monitor info
1052     // to support GC; these do not support deoptimization.
1053     int num_locs = (method == nullptr) ? 0 : jvms->loc_size();
1054     int num_exps = (method == nullptr) ? 0 : jvms->stk_size();
1055     int num_mon  = jvms->nof_monitors();
1056     assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(),
1057            "JVMS local count must match that of the method");
1058 
1059     // Add Local and Expression Stack Information
1060 
1061     // Insert locals into the locarray
1062     GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1063     for( idx = 0; idx < num_locs; idx++ ) {
1064       FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1065     }
1066 
1067     // Insert expression stack entries into the exparray
1068     GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1069     for( idx = 0; idx < num_exps; idx++ ) {
1070       FillLocArray( idx,  sfn, sfn->stack(jvms, idx), exparray, objs );
1071     }
1072 
1073     // Add in mappings of the monitors
1074     assert( !method ||
1075             !method->is_synchronized() ||
1076             method->is_native() ||
1077             num_mon > 0 ||
1078             !GenerateSynchronizationCode,
1079             "monitors must always exist for synchronized methods");
1080 
1081     // Build the growable array of ScopeValues for exp stack
1082     GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1083 
1084     // Loop over monitors and insert into array
1085     for (idx = 0; idx < num_mon; idx++) {
1086       // Grab the node that defines this monitor
1087       Node* box_node = sfn->monitor_box(jvms, idx);
1088       Node* obj_node = sfn->monitor_obj(jvms, idx);
1089 
1090       // Create ScopeValue for object
1091       ScopeValue *scval = nullptr;
1092 
1093       if (obj_node->is_SafePointScalarObject()) {
1094         SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1095         scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1096         if (scval == nullptr) {
1097           const Type *t = spobj->bottom_type();
1098           ciKlass* cik = t->is_oopptr()->exact_klass();
1099           assert(cik->is_instance_klass() ||
1100                  cik->is_array_klass(), "Not supported allocation.");
1101           ObjectValue* sv = new ObjectValue(spobj->_idx,
1102                                             new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1103           PhaseOutput::set_sv_for_object_node(objs, sv);
1104 
1105           uint first_ind = spobj->first_index(youngest_jvms);
1106           for (uint i = 0; i < spobj->n_fields(); i++) {
1107             Node* fld_node = sfn->in(first_ind+i);
1108             (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1109           }
1110           scval = sv;
1111         }
1112       } else if (obj_node->is_SafePointScalarMerge()) {
1113         SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1114         ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1115 
1116         if (mv == nullptr) {
1117           GrowableArray<ScopeValue*> deps;
1118 
1119           int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1120           FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1121           assert(deps.length() == 1, "missing value");
1122 
1123           int selector_idx = smerge->selector_idx(youngest_jvms);
1124           FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr);
1125           assert(deps.length() == 2, "missing value");
1126 
1127           mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
1128           set_sv_for_object_node(objs, mv);
1129 
1130           for (uint i = 1; i < smerge->req(); i++) {
1131             Node* obj_node = smerge->in(i);
1132             int idx = mv->possible_objects()->length();
1133             (void)FillLocArray(idx, sfn, obj_node, mv->possible_objects(), objs);
1134 
1135             // By default ObjectValues that are in 'possible_objects' are not root objects.
1136             // They will be marked as root later if they are directly referenced in a JVMS.
1137             assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
1138             assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
1139             mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
1140           }
1141         }
1142         scval = mv;
1143       } else if (!obj_node->is_Con()) {
1144         OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1145         if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1146           scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1147         } else {
1148           scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1149         }
1150       } else {
1151         const TypePtr *tp = obj_node->get_ptr_type();
1152         scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1153       }
1154 
1155       OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1156       Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1157       bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1158       monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1159     }
1160 
1161     // Mark ObjectValue nodes as root nodes if they are directly
1162     // referenced in the JVMS.
1163     for (int i = 0; i < objs->length(); i++) {
1164       ScopeValue* sv = objs->at(i);
1165       if (sv->is_object_merge()) {
1166         ObjectMergeValue* merge = sv->as_ObjectMergeValue();
1167 
1168         for (int j = 0; j< merge->possible_objects()->length(); j++) {
1169           ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue();
1170           if (ov->is_root()) {
1171             // Already flagged as 'root' by something else. We shouldn't change it
1172             // to non-root in a younger JVMS because it may need to be alive in
1173             // a younger JVMS.
1174           } else {
1175             bool is_root = locarray->contains(ov) ||
1176                            exparray->contains(ov) ||
1177                            contains_as_owner(monarray, ov) ||
1178                            contains_as_scalarized_obj(jvms, sfn, objs, ov);
1179             ov->set_root(is_root);
1180           }
1181         }
1182       }
1183     }
1184 
1185     // We dump the object pool first, since deoptimization reads it in first.
1186     C->debug_info()->dump_object_pool(objs);
1187 
1188     // Build first class objects to pass to scope
1189     DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1190     DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1191     DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1192 
1193     // Make method available for all Safepoints
1194     ciMethod* scope_method = method ? method : C->method();
1195     // Describe the scope here
1196     assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1197     assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1198     // Now we can describe the scope.
1199     methodHandle null_mh;
1200     bool rethrow_exception = false;
1201     C->debug_info()->describe_scope(
1202       safepoint_pc_offset,
1203       null_mh,
1204       scope_method,
1205       jvms->bci(),
1206       jvms->should_reexecute(),
1207       rethrow_exception,
1208       is_method_handle_invoke,
1209       return_oop,
1210       has_ea_local_in_scope,
1211       arg_escape,
1212       locvals,
1213       expvals,
1214       monvals
1215     );
1216   } // End jvms loop
1217 
1218   // Mark the end of the scope set.
1219   C->debug_info()->end_safepoint(safepoint_pc_offset);
1220 }
1221 
1222 
1223 
1224 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1225 class NonSafepointEmitter {
1226     Compile*  C;
1227     JVMState* _pending_jvms;
1228     int       _pending_offset;
1229 
1230     void emit_non_safepoint();
1231 
1232  public:
1233     NonSafepointEmitter(Compile* compile) {
1234       this->C = compile;
1235       _pending_jvms = nullptr;
1236       _pending_offset = 0;
1237     }
1238 
1239     void observe_instruction(Node* n, int pc_offset) {
1240       if (!C->debug_info()->recording_non_safepoints())  return;
1241 
1242       Node_Notes* nn = C->node_notes_at(n->_idx);
1243       if (nn == nullptr || nn->jvms() == nullptr)  return;
1244       if (_pending_jvms != nullptr &&
1245           _pending_jvms->same_calls_as(nn->jvms())) {
1246         // Repeated JVMS?  Stretch it up here.
1247         _pending_offset = pc_offset;
1248       } else {
1249         if (_pending_jvms != nullptr &&
1250             _pending_offset < pc_offset) {
1251           emit_non_safepoint();
1252         }
1253         _pending_jvms = nullptr;
1254         if (pc_offset > C->debug_info()->last_pc_offset()) {
1255           // This is the only way _pending_jvms can become non-null:
1256           _pending_jvms = nn->jvms();
1257           _pending_offset = pc_offset;
1258         }
1259       }
1260     }
1261 
1262     // Stay out of the way of real safepoints:
1263     void observe_safepoint(JVMState* jvms, int pc_offset) {
1264       if (_pending_jvms != nullptr &&
1265           !_pending_jvms->same_calls_as(jvms) &&
1266           _pending_offset < pc_offset) {
1267         emit_non_safepoint();
1268       }
1269       _pending_jvms = nullptr;
1270     }
1271 
1272     void flush_at_end() {
1273       if (_pending_jvms != nullptr) {
1274         emit_non_safepoint();
1275       }
1276       _pending_jvms = nullptr;
1277     }
1278 };
1279 
1280 void NonSafepointEmitter::emit_non_safepoint() {
1281   JVMState* youngest_jvms = _pending_jvms;
1282   int       pc_offset     = _pending_offset;
1283 
1284   // Clear it now:
1285   _pending_jvms = nullptr;
1286 
1287   DebugInformationRecorder* debug_info = C->debug_info();
1288   assert(debug_info->recording_non_safepoints(), "sanity");
1289 
1290   debug_info->add_non_safepoint(pc_offset);
1291   int max_depth = youngest_jvms->depth();
1292 
1293   // Visit scopes from oldest to youngest.
1294   for (int depth = 1; depth <= max_depth; depth++) {
1295     JVMState* jvms = youngest_jvms->of_depth(depth);
1296     ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1297     assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1298     methodHandle null_mh;
1299     debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1300   }
1301 
1302   // Mark the end of the scope set.
1303   debug_info->end_non_safepoint(pc_offset);
1304 }
1305 
1306 //------------------------------init_buffer------------------------------------
1307 void PhaseOutput::estimate_buffer_size(int& const_req) {
1308 
1309   // Set the initially allocated size
1310   const_req = initial_const_capacity;
1311 
1312   // The extra spacing after the code is necessary on some platforms.
1313   // Sometimes we need to patch in a jump after the last instruction,
1314   // if the nmethod has been deoptimized.  (See 4932387, 4894843.)
1315 
1316   // Compute the byte offset where we can store the deopt pc.
1317   if (C->fixed_slots() != 0) {
1318     _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1319   }
1320 
1321   // Compute prolog code size
1322   _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1323   assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1324 
1325   if (C->has_mach_constant_base_node()) {
1326     uint add_size = 0;
1327     // Fill the constant table.
1328     // Note:  This must happen before shorten_branches.
1329     for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1330       Block* b = C->cfg()->get_block(i);
1331 
1332       for (uint j = 0; j < b->number_of_nodes(); j++) {
1333         Node* n = b->get_node(j);
1334 
1335         // If the node is a MachConstantNode evaluate the constant
1336         // value section.
1337         if (n->is_MachConstant()) {
1338           MachConstantNode* machcon = n->as_MachConstant();
1339           machcon->eval_constant(C);
1340         } else if (n->is_Mach()) {
1341           // On Power there are more nodes that issue constants.
1342           add_size += (n->as_Mach()->ins_num_consts() * 8);
1343         }
1344       }
1345     }
1346 
1347     // Calculate the offsets of the constants and the size of the
1348     // constant table (including the padding to the next section).
1349     constant_table().calculate_offsets_and_size();
1350     const_req = constant_table().alignment() + constant_table().size() + add_size;
1351   }
1352 
1353   // Initialize the space for the BufferBlob used to find and verify
1354   // instruction size in MachNode::emit_size()
1355   init_scratch_buffer_blob(const_req);
1356 }
1357 
1358 CodeBuffer* PhaseOutput::init_buffer() {
1359   int stub_req  = _buf_sizes._stub;
1360   int code_req  = _buf_sizes._code;
1361   int const_req = _buf_sizes._const;
1362 
1363   int pad_req   = NativeCall::byte_size();
1364 
1365   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1366   stub_req += bs->estimate_stub_size();
1367 
1368   // nmethod and CodeBuffer count stubs & constants as part of method's code.
1369   // class HandlerImpl is platform-specific and defined in the *.ad files.
1370   int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1371   int deopt_handler_req     = HandlerImpl::size_deopt_handler()     + MAX_stubs_size; // add marginal slop for handler
1372   stub_req += MAX_stubs_size;   // ensure per-stub margin
1373   code_req += max_inst_size();  // ensure per-instruction margin
1374 
1375   if (StressCodeBuffers)
1376     code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10;  // force expansion
1377 
1378   int total_req =
1379           const_req +
1380           code_req +
1381           pad_req +
1382           stub_req +
1383           exception_handler_req +
1384           deopt_handler_req;               // deopt handler
1385 
1386   if (C->has_method_handle_invokes())
1387     total_req += deopt_handler_req;  // deopt MH handler
1388 
1389   CodeBuffer* cb = code_buffer();
1390   cb->set_const_section_alignment(constant_table().alignment());
1391   cb->initialize(total_req, _buf_sizes._reloc);
1392 
1393   // Have we run out of code space?
1394   if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1395     C->record_failure("CodeCache is full");
1396     return nullptr;
1397   }
1398   // Configure the code buffer.
1399   cb->initialize_consts_size(const_req);
1400   cb->initialize_stubs_size(stub_req);
1401   cb->initialize_oop_recorder(C->env()->oop_recorder());
1402 
1403   // fill in the nop array for bundling computations
1404   MachNode *_nop_list[Bundle::_nop_count];
1405   Bundle::initialize_nops(_nop_list);
1406 
1407   return cb;
1408 }
1409 
1410 //------------------------------fill_buffer------------------------------------
1411 void PhaseOutput::fill_buffer(C2_MacroAssembler* masm, uint* blk_starts) {
1412   // blk_starts[] contains offsets calculated during short branches processing,
1413   // offsets should not be increased during following steps.
1414 
1415   // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1416   // of a loop. It is used to determine the padding for loop alignment.
1417   Compile::TracePhase tp(_t_fillBuffer);
1418 
1419   compute_loop_first_inst_sizes();
1420 
1421   // Create oopmap set.
1422   _oop_map_set = new OopMapSet();
1423 
1424   // !!!!! This preserves old handling of oopmaps for now
1425   C->debug_info()->set_oopmaps(_oop_map_set);
1426 
1427   uint nblocks  = C->cfg()->number_of_blocks();
1428   // Count and start of implicit null check instructions
1429   uint inct_cnt = 0;
1430   uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1431 
1432   // Count and start of calls
1433   uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1434 
1435   uint  return_offset = 0;
1436   int nop_size = (new MachNopNode())->size(C->regalloc());
1437 
1438   int previous_offset = 0;
1439   int current_offset  = 0;
1440   int last_call_offset = -1;
1441   int last_avoid_back_to_back_offset = -1;
1442 #ifdef ASSERT
1443   uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1444   uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1445   uint* jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
1446   uint* jmp_rule   = NEW_RESOURCE_ARRAY(uint,nblocks);
1447 #endif
1448 
1449   // Create an array of unused labels, one for each basic block, if printing is enabled
1450 #if defined(SUPPORT_OPTO_ASSEMBLY)
1451   int* node_offsets      = nullptr;
1452   uint node_offset_limit = C->unique();
1453 
1454   if (C->print_assembly()) {
1455     node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1456   }
1457   if (node_offsets != nullptr) {
1458     // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1459     memset(node_offsets, 0, node_offset_limit*sizeof(int));
1460   }
1461 #endif
1462 
1463   NonSafepointEmitter non_safepoints(C);  // emit non-safepoints lazily
1464 
1465   // Emit the constant table.
1466   if (C->has_mach_constant_base_node()) {
1467     if (!constant_table().emit(masm)) {
1468       C->record_failure("consts section overflow");
1469       return;
1470     }
1471   }
1472 
1473   // Create an array of labels, one for each basic block
1474   Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1475   for (uint i = 0; i <= nblocks; i++) {
1476     blk_labels[i].init();
1477   }
1478 
1479   // Now fill in the code buffer
1480   Node* delay_slot = nullptr;
1481   for (uint i = 0; i < nblocks; i++) {
1482     Block* block = C->cfg()->get_block(i);
1483     _block = block;
1484     Node* head = block->head();
1485 
1486     // If this block needs to start aligned (i.e, can be reached other
1487     // than by falling-thru from the previous block), then force the
1488     // start of a new bundle.
1489     if (Pipeline::requires_bundling() && starts_bundle(head)) {
1490       masm->code()->flush_bundle(true);
1491     }
1492 
1493 #ifdef ASSERT
1494     if (!block->is_connector()) {
1495       stringStream st;
1496       block->dump_head(C->cfg(), &st);
1497       masm->block_comment(st.freeze());
1498     }
1499     jmp_target[i] = 0;
1500     jmp_offset[i] = 0;
1501     jmp_size[i]   = 0;
1502     jmp_rule[i]   = 0;
1503 #endif
1504     int blk_offset = current_offset;
1505 
1506     // Define the label at the beginning of the basic block
1507     masm->bind(blk_labels[block->_pre_order]);
1508 
1509     uint last_inst = block->number_of_nodes();
1510 
1511     // Emit block normally, except for last instruction.
1512     // Emit means "dump code bits into code buffer".
1513     for (uint j = 0; j<last_inst; j++) {
1514       _index = j;
1515 
1516       // Get the node
1517       Node* n = block->get_node(j);
1518 
1519       // See if delay slots are supported
1520       if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) {
1521         assert(delay_slot == nullptr, "no use of delay slot node");
1522         assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1523 
1524         delay_slot = n;
1525         continue;
1526       }
1527 
1528       // If this starts a new instruction group, then flush the current one
1529       // (but allow split bundles)
1530       if (Pipeline::requires_bundling() && starts_bundle(n))
1531         masm->code()->flush_bundle(false);
1532 
1533       // Special handling for SafePoint/Call Nodes
1534       bool is_mcall = false;
1535       if (n->is_Mach()) {
1536         MachNode *mach = n->as_Mach();
1537         is_mcall = n->is_MachCall();
1538         bool is_sfn = n->is_MachSafePoint();
1539 
1540         // If this requires all previous instructions be flushed, then do so
1541         if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1542           masm->code()->flush_bundle(true);
1543           current_offset = masm->offset();
1544         }
1545 
1546         // A padding may be needed again since a previous instruction
1547         // could be moved to delay slot.
1548 
1549         // align the instruction if necessary
1550         int padding = mach->compute_padding(current_offset);
1551         // Make sure safepoint node for polling is distinct from a call's
1552         // return by adding a nop if needed.
1553         if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1554           padding = nop_size;
1555         }
1556         if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1557             current_offset == last_avoid_back_to_back_offset) {
1558           // Avoid back to back some instructions.
1559           padding = nop_size;
1560         }
1561 
1562         if (padding > 0) {
1563           assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1564           int nops_cnt = padding / nop_size;
1565           MachNode *nop = new MachNopNode(nops_cnt);
1566           block->insert_node(nop, j++);
1567           last_inst++;
1568           C->cfg()->map_node_to_block(nop, block);
1569           // Ensure enough space.
1570           masm->code()->insts()->maybe_expand_to_ensure_remaining(max_inst_size());
1571           if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1572             C->record_failure("CodeCache is full");
1573             return;
1574           }
1575           nop->emit(masm, C->regalloc());
1576           masm->code()->flush_bundle(true);
1577           current_offset = masm->offset();
1578         }
1579 
1580         bool observe_safepoint = is_sfn;
1581         // Remember the start of the last call in a basic block
1582         if (is_mcall) {
1583           MachCallNode *mcall = mach->as_MachCall();
1584 
1585           // This destination address is NOT PC-relative
1586           mcall->method_set((intptr_t)mcall->entry_point());
1587 
1588           // Save the return address
1589           call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1590 
1591           observe_safepoint = mcall->guaranteed_safepoint();
1592         }
1593 
1594         // sfn will be valid whenever mcall is valid now because of inheritance
1595         if (observe_safepoint) {
1596           // Handle special safepoint nodes for synchronization
1597           if (!is_mcall) {
1598             MachSafePointNode *sfn = mach->as_MachSafePoint();
1599             // !!!!! Stubs only need an oopmap right now, so bail out
1600             if (sfn->jvms()->method() == nullptr) {
1601               // Write the oopmap directly to the code blob??!!
1602               continue;
1603             }
1604           } // End synchronization
1605 
1606           non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1607                                            current_offset);
1608           Process_OopMap_Node(mach, current_offset);
1609         } // End if safepoint
1610 
1611           // If this is a null check, then add the start of the previous instruction to the list
1612         else if( mach->is_MachNullCheck() ) {
1613           inct_starts[inct_cnt++] = previous_offset;
1614         }
1615 
1616           // If this is a branch, then fill in the label with the target BB's label
1617         else if (mach->is_MachBranch()) {
1618           // This requires the TRUE branch target be in succs[0]
1619           uint block_num = block->non_connector_successor(0)->_pre_order;
1620 
1621           // Try to replace long branch if delay slot is not used,
1622           // it is mostly for back branches since forward branch's
1623           // distance is not updated yet.
1624           bool delay_slot_is_used = valid_bundle_info(n) &&
1625                                     C->output()->node_bundling(n)->use_unconditional_delay();
1626           if (!delay_slot_is_used && mach->may_be_short_branch()) {
1627             assert(delay_slot == nullptr, "not expecting delay slot node");
1628             int br_size = n->size(C->regalloc());
1629             int offset = blk_starts[block_num] - current_offset;
1630             if (block_num >= i) {
1631               // Current and following block's offset are not
1632               // finalized yet, adjust distance by the difference
1633               // between calculated and final offsets of current block.
1634               offset -= (blk_starts[i] - blk_offset);
1635             }
1636             // In the following code a nop could be inserted before
1637             // the branch which will increase the backward distance.
1638             bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1639             if (needs_padding && offset <= 0)
1640               offset -= nop_size;
1641 
1642             if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1643               // We've got a winner.  Replace this branch.
1644               MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1645 
1646               // Update the jmp_size.
1647               int new_size = replacement->size(C->regalloc());
1648               assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1649               // Insert padding between avoid_back_to_back branches.
1650               if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1651                 MachNode *nop = new MachNopNode();
1652                 block->insert_node(nop, j++);
1653                 C->cfg()->map_node_to_block(nop, block);
1654                 last_inst++;
1655                 nop->emit(masm, C->regalloc());
1656                 masm->code()->flush_bundle(true);
1657                 current_offset = masm->offset();
1658               }
1659 #ifdef ASSERT
1660               jmp_target[i] = block_num;
1661               jmp_offset[i] = current_offset - blk_offset;
1662               jmp_size[i]   = new_size;
1663               jmp_rule[i]   = mach->rule();
1664 #endif
1665               block->map_node(replacement, j);
1666               mach->subsume_by(replacement, C);
1667               n    = replacement;
1668               mach = replacement;
1669             }
1670           }
1671           mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1672         } else if (mach->ideal_Opcode() == Op_Jump) {
1673           for (uint h = 0; h < block->_num_succs; h++) {
1674             Block* succs_block = block->_succs[h];
1675             for (uint j = 1; j < succs_block->num_preds(); j++) {
1676               Node* jpn = succs_block->pred(j);
1677               if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1678                 uint block_num = succs_block->non_connector()->_pre_order;
1679                 Label *blkLabel = &blk_labels[block_num];
1680                 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1681               }
1682             }
1683           }
1684         } else if (!n->is_Proj()) {
1685           // Remember the beginning of the previous instruction, in case
1686           // it's followed by a flag-kill and a null-check.  Happens on
1687           // Intel all the time, with add-to-memory kind of opcodes.
1688           previous_offset = current_offset;
1689         }
1690 
1691         // Not an else-if!
1692         // If this is a trap based cmp then add its offset to the list.
1693         if (mach->is_TrapBasedCheckNode()) {
1694           inct_starts[inct_cnt++] = current_offset;
1695         }
1696       }
1697 
1698       // Verify that there is sufficient space remaining
1699       masm->code()->insts()->maybe_expand_to_ensure_remaining(max_inst_size());
1700       if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1701         C->record_failure("CodeCache is full");
1702         return;
1703       }
1704 
1705       // Save the offset for the listing
1706 #if defined(SUPPORT_OPTO_ASSEMBLY)
1707       if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1708         node_offsets[n->_idx] = masm->offset();
1709       }
1710 #endif
1711       assert(!C->failing_internal() || C->failure_is_artificial(), "Should not reach here if failing.");
1712 
1713       // "Normal" instruction case
1714       DEBUG_ONLY(uint instr_offset = masm->offset());
1715       n->emit(masm, C->regalloc());
1716       current_offset = masm->offset();
1717 
1718       // Above we only verified that there is enough space in the instruction section.
1719       // However, the instruction may emit stubs that cause code buffer expansion.
1720       // Bail out here if expansion failed due to a lack of code cache space.
1721       if (C->failing()) {
1722         return;
1723       }
1724 
1725       assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1726              "ret_addr_offset() not within emitted code");
1727 
1728 #ifdef ASSERT
1729       uint n_size = n->size(C->regalloc());
1730       if (n_size < (current_offset-instr_offset)) {
1731         MachNode* mach = n->as_Mach();
1732         n->dump();
1733         mach->dump_format(C->regalloc(), tty);
1734         tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1735         Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty);
1736         tty->print_cr(" ------------------- ");
1737         BufferBlob* blob = this->scratch_buffer_blob();
1738         address blob_begin = blob->content_begin();
1739         Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1740         assert(false, "wrong size of mach node");
1741       }
1742 #endif
1743       non_safepoints.observe_instruction(n, current_offset);
1744 
1745       // mcall is last "call" that can be a safepoint
1746       // record it so we can see if a poll will directly follow it
1747       // in which case we'll need a pad to make the PcDesc sites unique
1748       // see  5010568. This can be slightly inaccurate but conservative
1749       // in the case that return address is not actually at current_offset.
1750       // This is a small price to pay.
1751 
1752       if (is_mcall) {
1753         last_call_offset = current_offset;
1754       }
1755 
1756       if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1757         // Avoid back to back some instructions.
1758         last_avoid_back_to_back_offset = current_offset;
1759       }
1760 
1761       // See if this instruction has a delay slot
1762       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1763         guarantee(delay_slot != nullptr, "expecting delay slot node");
1764 
1765         // Back up 1 instruction
1766         masm->code()->set_insts_end(masm->code()->insts_end() - Pipeline::instr_unit_size());
1767 
1768         // Save the offset for the listing
1769 #if defined(SUPPORT_OPTO_ASSEMBLY)
1770         if ((node_offsets != nullptr) && (delay_slot->_idx < node_offset_limit)) {
1771           node_offsets[delay_slot->_idx] = masm->offset();
1772         }
1773 #endif
1774 
1775         // Support a SafePoint in the delay slot
1776         if (delay_slot->is_MachSafePoint()) {
1777           MachNode *mach = delay_slot->as_Mach();
1778           // !!!!! Stubs only need an oopmap right now, so bail out
1779           if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == nullptr) {
1780             // Write the oopmap directly to the code blob??!!
1781             delay_slot = nullptr;
1782             continue;
1783           }
1784 
1785           int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1786           non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1787                                            adjusted_offset);
1788           // Generate an OopMap entry
1789           Process_OopMap_Node(mach, adjusted_offset);
1790         }
1791 
1792         // Insert the delay slot instruction
1793         delay_slot->emit(masm, C->regalloc());
1794 
1795         // Don't reuse it
1796         delay_slot = nullptr;
1797       }
1798 
1799     } // End for all instructions in block
1800 
1801     // If the next block is the top of a loop, pad this block out to align
1802     // the loop top a little. Helps prevent pipe stalls at loop back branches.
1803     if (i < nblocks-1) {
1804       Block *nb = C->cfg()->get_block(i + 1);
1805       int padding = nb->alignment_padding(current_offset);
1806       if( padding > 0 ) {
1807         MachNode *nop = new MachNopNode(padding / nop_size);
1808         block->insert_node(nop, block->number_of_nodes());
1809         C->cfg()->map_node_to_block(nop, block);
1810         nop->emit(masm, C->regalloc());
1811         current_offset = masm->offset();
1812       }
1813     }
1814     // Verify that the distance for generated before forward
1815     // short branches is still valid.
1816     guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1817 
1818     // Save new block start offset
1819     blk_starts[i] = blk_offset;
1820   } // End of for all blocks
1821   blk_starts[nblocks] = current_offset;
1822 
1823   non_safepoints.flush_at_end();
1824 
1825   // Offset too large?
1826   if (C->failing())  return;
1827 
1828   // Define a pseudo-label at the end of the code
1829   masm->bind( blk_labels[nblocks] );
1830 
1831   // Compute the size of the first block
1832   _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1833 
1834 #ifdef ASSERT
1835   for (uint i = 0; i < nblocks; i++) { // For all blocks
1836     if (jmp_target[i] != 0) {
1837       int br_size = jmp_size[i];
1838       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1839       if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1840         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]);
1841         assert(false, "Displacement too large for short jmp");
1842       }
1843     }
1844   }
1845 #endif
1846 
1847   if (!masm->code()->finalize_stubs()) {
1848     C->record_failure("CodeCache is full");
1849     return;
1850   }
1851 
1852   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1853   bs->emit_stubs(*masm->code());
1854   if (C->failing())  return;
1855 
1856   // Fill in stubs.
1857   assert(masm->inst_mark() == nullptr, "should be.");
1858   _stub_list.emit(*masm);
1859   if (C->failing())  return;
1860 
1861 #ifndef PRODUCT
1862   // Information on the size of the method, without the extraneous code
1863   Scheduling::increment_method_size(masm->offset());
1864 #endif
1865 
1866   // ------------------
1867   // Fill in exception table entries.
1868   FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1869 
1870   // Only java methods have exception handlers and deopt handlers
1871   // class HandlerImpl is platform-specific and defined in the *.ad files.
1872   if (C->method()) {
1873     // Emit the exception handler code.
1874     _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(masm));
1875     if (C->failing()) {
1876       return; // CodeBuffer::expand failed
1877     }
1878     // Emit the deopt handler code.
1879     _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(masm));
1880 
1881     // Emit the MethodHandle deopt handler code (if required).
1882     if (C->has_method_handle_invokes() && !C->failing()) {
1883       // We can use the same code as for the normal deopt handler, we
1884       // just need a different entry point address.
1885       _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(masm));
1886     }
1887   }
1888 
1889   // One last check for failed CodeBuffer::expand:
1890   if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1891     C->record_failure("CodeCache is full");
1892     return;
1893   }
1894 
1895 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1896   if (C->print_assembly()) {
1897     tty->cr();
1898     tty->print_cr("============================= C2-compiled nmethod ==============================");
1899   }
1900 #endif
1901 
1902 #if defined(SUPPORT_OPTO_ASSEMBLY)
1903   // Dump the assembly code, including basic-block numbers
1904   if (C->print_assembly()) {
1905     ttyLocker ttyl;  // keep the following output all in one block
1906     if (!VMThread::should_terminate()) {  // test this under the tty lock
1907       // print_metadata and dump_asm may safepoint which makes us loose the ttylock.
1908       // We call them first and write to a stringStream, then we retake the lock to
1909       // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
1910       ResourceMark rm;
1911       stringStream method_metadata_str;
1912       if (C->method() != nullptr) {
1913         C->method()->print_metadata(&method_metadata_str);
1914       }
1915       stringStream dump_asm_str;
1916       dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
1917 
1918       NoSafepointVerifier nsv;
1919       ttyLocker ttyl2;
1920       // This output goes directly to the tty, not the compiler log.
1921       // To enable tools to match it up with the compilation activity,
1922       // be sure to tag this tty output with the compile ID.
1923       if (xtty != nullptr) {
1924         xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
1925                    C->is_osr_compilation() ? " compile_kind='osr'" :
1926                    (C->for_preload() ? " compile_kind='AP'" : ""));
1927       }
1928       if (C->method() != nullptr) {
1929         tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
1930         tty->print_raw(method_metadata_str.freeze());
1931       } else if (C->stub_name() != nullptr) {
1932         tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
1933       }
1934       tty->cr();
1935       tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
1936       tty->print_raw(dump_asm_str.freeze());
1937       tty->print_cr("--------------------------------------------------------------------------------");
1938       if (xtty != nullptr) {
1939         xtty->tail("opto_assembly");
1940       }
1941     }
1942   }
1943 #endif
1944 }
1945 
1946 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1947   _inc_table.set_size(cnt);
1948 
1949   uint inct_cnt = 0;
1950   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1951     Block* block = C->cfg()->get_block(i);
1952     Node *n = nullptr;
1953     int j;
1954 
1955     // Find the branch; ignore trailing NOPs.
1956     for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1957       n = block->get_node(j);
1958       if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1959         break;
1960       }
1961     }
1962 
1963     // If we didn't find anything, continue
1964     if (j < 0) {
1965       continue;
1966     }
1967 
1968     // Compute ExceptionHandlerTable subtable entry and add it
1969     // (skip empty blocks)
1970     if (n->is_Catch()) {
1971 
1972       // Get the offset of the return from the call
1973       uint call_return = call_returns[block->_pre_order];
1974 #ifdef ASSERT
1975       assert( call_return > 0, "no call seen for this basic block" );
1976       while (block->get_node(--j)->is_MachProj()) ;
1977       assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1978 #endif
1979       // last instruction is a CatchNode, find it's CatchProjNodes
1980       int nof_succs = block->_num_succs;
1981       // allocate space
1982       GrowableArray<intptr_t> handler_bcis(nof_succs);
1983       GrowableArray<intptr_t> handler_pcos(nof_succs);
1984       // iterate through all successors
1985       for (int j = 0; j < nof_succs; j++) {
1986         Block* s = block->_succs[j];
1987         bool found_p = false;
1988         for (uint k = 1; k < s->num_preds(); k++) {
1989           Node* pk = s->pred(k);
1990           if (pk->is_CatchProj() && pk->in(0) == n) {
1991             const CatchProjNode* p = pk->as_CatchProj();
1992             found_p = true;
1993             // add the corresponding handler bci & pco information
1994             if (p->_con != CatchProjNode::fall_through_index) {
1995               // p leads to an exception handler (and is not fall through)
1996               assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
1997               // no duplicates, please
1998               if (!handler_bcis.contains(p->handler_bci())) {
1999                 uint block_num = s->non_connector()->_pre_order;
2000                 handler_bcis.append(p->handler_bci());
2001                 handler_pcos.append(blk_labels[block_num].loc_pos());
2002               }
2003             }
2004           }
2005         }
2006         assert(found_p, "no matching predecessor found");
2007         // Note:  Due to empty block removal, one block may have
2008         // several CatchProj inputs, from the same Catch.
2009       }
2010 
2011       // Set the offset of the return from the call
2012       assert(handler_bcis.find(-1) != -1, "must have default handler");
2013       _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos);
2014       continue;
2015     }
2016 
2017     // Handle implicit null exception table updates
2018     if (n->is_MachNullCheck()) {
2019       assert(n->in(1)->as_Mach()->barrier_data() == 0,
2020              "Implicit null checks on memory accesses with barriers are not yet supported");
2021       uint block_num = block->non_connector_successor(0)->_pre_order;
2022       _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2023       continue;
2024     }
2025     // Handle implicit exception table updates: trap instructions.
2026     if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
2027       uint block_num = block->non_connector_successor(0)->_pre_order;
2028       _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2029       continue;
2030     }
2031   } // End of for all blocks fill in exception table entries
2032 }
2033 
2034 // Static Variables
2035 #ifndef PRODUCT
2036 uint Scheduling::_total_nop_size = 0;
2037 uint Scheduling::_total_method_size = 0;
2038 uint Scheduling::_total_branches = 0;
2039 uint Scheduling::_total_unconditional_delays = 0;
2040 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
2041 #endif
2042 
2043 // Initializer for class Scheduling
2044 
2045 Scheduling::Scheduling(Arena *arena, Compile &compile)
2046         : _arena(arena),
2047           _cfg(compile.cfg()),
2048           _regalloc(compile.regalloc()),
2049           _scheduled(arena),
2050           _available(arena),
2051           _reg_node(arena),
2052           _pinch_free_list(arena),
2053           _next_node(nullptr),
2054           _bundle_instr_count(0),
2055           _bundle_cycle_number(0),
2056           _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
2057 #ifndef PRODUCT
2058         , _branches(0)
2059         , _unconditional_delays(0)
2060 #endif
2061 {
2062   // Create a MachNopNode
2063   _nop = new MachNopNode();
2064 
2065   // Now that the nops are in the array, save the count
2066   // (but allow entries for the nops)
2067   _node_bundling_limit = compile.unique();
2068   uint node_max = _regalloc->node_regs_max_index();
2069 
2070   compile.output()->set_node_bundling_limit(_node_bundling_limit);
2071 
2072   // This one is persistent within the Compile class
2073   _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
2074 
2075   // Allocate space for fixed-size arrays
2076   _uses            = NEW_ARENA_ARRAY(arena, short,          node_max);
2077   _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
2078 
2079   // Clear the arrays
2080   for (uint i = 0; i < node_max; i++) {
2081     ::new (&_node_bundling_base[i]) Bundle();
2082   }
2083   memset(_uses,               0, node_max * sizeof(short));
2084   memset(_current_latency,    0, node_max * sizeof(unsigned short));
2085 
2086   // Clear the bundling information
2087   memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2088 
2089   // Get the last node
2090   Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2091 
2092   _next_node = block->get_node(block->number_of_nodes() - 1);
2093 }
2094 
2095 #ifndef PRODUCT
2096 // Scheduling destructor
2097 Scheduling::~Scheduling() {
2098   _total_branches             += _branches;
2099   _total_unconditional_delays += _unconditional_delays;
2100 }
2101 #endif
2102 
2103 // Step ahead "i" cycles
2104 void Scheduling::step(uint i) {
2105 
2106   Bundle *bundle = node_bundling(_next_node);
2107   bundle->set_starts_bundle();
2108 
2109   // Update the bundle record, but leave the flags information alone
2110   if (_bundle_instr_count > 0) {
2111     bundle->set_instr_count(_bundle_instr_count);
2112     bundle->set_resources_used(_bundle_use.resourcesUsed());
2113   }
2114 
2115   // Update the state information
2116   _bundle_instr_count = 0;
2117   _bundle_cycle_number += i;
2118   _bundle_use.step(i);
2119 }
2120 
2121 void Scheduling::step_and_clear() {
2122   Bundle *bundle = node_bundling(_next_node);
2123   bundle->set_starts_bundle();
2124 
2125   // Update the bundle record
2126   if (_bundle_instr_count > 0) {
2127     bundle->set_instr_count(_bundle_instr_count);
2128     bundle->set_resources_used(_bundle_use.resourcesUsed());
2129 
2130     _bundle_cycle_number += 1;
2131   }
2132 
2133   // Clear the bundling information
2134   _bundle_instr_count = 0;
2135   _bundle_use.reset();
2136 
2137   memcpy(_bundle_use_elements,
2138          Pipeline_Use::elaborated_elements,
2139          sizeof(Pipeline_Use::elaborated_elements));
2140 }
2141 
2142 // Perform instruction scheduling and bundling over the sequence of
2143 // instructions in backwards order.
2144 void PhaseOutput::ScheduleAndBundle() {
2145 
2146   // Don't optimize this if it isn't a method
2147   if (!C->method())
2148     return;
2149 
2150   // Don't optimize this if scheduling is disabled
2151   if (!C->do_scheduling())
2152     return;
2153 
2154   // Scheduling code works only with pairs (8 bytes) maximum.
2155   // And when the scalable vector register is used, we may spill/unspill
2156   // the whole reg regardless of the max vector size.
2157   if (C->max_vector_size() > 8 ||
2158       (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2159     return;
2160   }
2161 
2162   Compile::TracePhase tp(_t_instrSched);
2163 
2164   // Create a data structure for all the scheduling information
2165   Scheduling scheduling(Thread::current()->resource_area(), *C);
2166 
2167   // Walk backwards over each basic block, computing the needed alignment
2168   // Walk over all the basic blocks
2169   scheduling.DoScheduling();
2170 
2171 #ifndef PRODUCT
2172   if (C->trace_opto_output()) {
2173     // Buffer and print all at once
2174     ResourceMark rm;
2175     stringStream ss;
2176     ss.print("\n---- After ScheduleAndBundle ----\n");
2177     print_scheduling(&ss);
2178     tty->print("%s", ss.as_string());
2179   }
2180 #endif
2181 }
2182 
2183 #ifndef PRODUCT
2184 // Separated out so that it can be called directly from debugger
2185 void PhaseOutput::print_scheduling() {
2186   print_scheduling(tty);
2187 }
2188 
2189 void PhaseOutput::print_scheduling(outputStream* output_stream) {
2190   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2191     output_stream->print("\nBB#%03d:\n", i);
2192     Block* block = C->cfg()->get_block(i);
2193     for (uint j = 0; j < block->number_of_nodes(); j++) {
2194       Node* n = block->get_node(j);
2195       OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2196       output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2197       n->dump("\n", false, output_stream);
2198     }
2199   }
2200 }
2201 #endif
2202 
2203 // See if this node fits into the present instruction bundle
2204 bool Scheduling::NodeFitsInBundle(Node *n) {
2205   uint n_idx = n->_idx;
2206 
2207   // If this is the unconditional delay instruction, then it fits
2208   if (n == _unconditional_delay_slot) {
2209 #ifndef PRODUCT
2210     if (_cfg->C->trace_opto_output())
2211       tty->print("#     NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2212 #endif
2213     return (true);
2214   }
2215 
2216   // If the node cannot be scheduled this cycle, skip it
2217   if (_current_latency[n_idx] > _bundle_cycle_number) {
2218 #ifndef PRODUCT
2219     if (_cfg->C->trace_opto_output())
2220       tty->print("#     NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2221                  n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2222 #endif
2223     return (false);
2224   }
2225 
2226   const Pipeline *node_pipeline = n->pipeline();
2227 
2228   uint instruction_count = node_pipeline->instructionCount();
2229   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2230     instruction_count = 0;
2231   else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2232     instruction_count++;
2233 
2234   if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2235 #ifndef PRODUCT
2236     if (_cfg->C->trace_opto_output())
2237       tty->print("#     NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2238                  n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2239 #endif
2240     return (false);
2241   }
2242 
2243   // Don't allow non-machine nodes to be handled this way
2244   if (!n->is_Mach() && instruction_count == 0)
2245     return (false);
2246 
2247   // See if there is any overlap
2248   uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2249 
2250   if (delay > 0) {
2251 #ifndef PRODUCT
2252     if (_cfg->C->trace_opto_output())
2253       tty->print("#     NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2254 #endif
2255     return false;
2256   }
2257 
2258 #ifndef PRODUCT
2259   if (_cfg->C->trace_opto_output())
2260     tty->print("#     NodeFitsInBundle [%4d]:  TRUE\n", n_idx);
2261 #endif
2262 
2263   return true;
2264 }
2265 
2266 Node * Scheduling::ChooseNodeToBundle() {
2267   uint siz = _available.size();
2268 
2269   if (siz == 0) {
2270 
2271 #ifndef PRODUCT
2272     if (_cfg->C->trace_opto_output())
2273       tty->print("#   ChooseNodeToBundle: null\n");
2274 #endif
2275     return (nullptr);
2276   }
2277 
2278   // Fast path, if only 1 instruction in the bundle
2279   if (siz == 1) {
2280 #ifndef PRODUCT
2281     if (_cfg->C->trace_opto_output()) {
2282       tty->print("#   ChooseNodeToBundle (only 1): ");
2283       _available[0]->dump();
2284     }
2285 #endif
2286     return (_available[0]);
2287   }
2288 
2289   // Don't bother, if the bundle is already full
2290   if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2291     for ( uint i = 0; i < siz; i++ ) {
2292       Node *n = _available[i];
2293 
2294       // Skip projections, we'll handle them another way
2295       if (n->is_Proj())
2296         continue;
2297 
2298       // This presupposed that instructions are inserted into the
2299       // available list in a legality order; i.e. instructions that
2300       // must be inserted first are at the head of the list
2301       if (NodeFitsInBundle(n)) {
2302 #ifndef PRODUCT
2303         if (_cfg->C->trace_opto_output()) {
2304           tty->print("#   ChooseNodeToBundle: ");
2305           n->dump();
2306         }
2307 #endif
2308         return (n);
2309       }
2310     }
2311   }
2312 
2313   // Nothing fits in this bundle, choose the highest priority
2314 #ifndef PRODUCT
2315   if (_cfg->C->trace_opto_output()) {
2316     tty->print("#   ChooseNodeToBundle: ");
2317     _available[0]->dump();
2318   }
2319 #endif
2320 
2321   return _available[0];
2322 }
2323 
2324 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) {
2325   assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), "");
2326 
2327   OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1));
2328   OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first);
2329   OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1));
2330   OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second);
2331 
2332   // Comparison between stack -> reg and stack -> reg
2333   if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) &&
2334       OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) {
2335     return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo);
2336   }
2337 
2338   // Comparison between reg -> stack and reg -> stack
2339   if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) &&
2340       OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) {
2341     return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo);
2342   }
2343 
2344   return 0; // Not comparable
2345 }
2346 
2347 void Scheduling::AddNodeToAvailableList(Node *n) {
2348   assert( !n->is_Proj(), "projections never directly made available" );
2349 #ifndef PRODUCT
2350   if (_cfg->C->trace_opto_output()) {
2351     tty->print("#   AddNodeToAvailableList: ");
2352     n->dump();
2353   }
2354 #endif
2355 
2356   int latency = _current_latency[n->_idx];
2357 
2358   // Insert in latency order (insertion sort). If two MachSpillCopyNodes
2359   // for stack spilling or unspilling have the same latency, we sort
2360   // them in the order of stack offset. Some ports (e.g. aarch64) may also
2361   // have more opportunities to do ld/st merging
2362   uint i;
2363   for (i = 0; i < _available.size(); i++) {
2364     if (_current_latency[_available[i]->_idx] > latency) {
2365       break;
2366     } else if (_current_latency[_available[i]->_idx] == latency &&
2367                n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() &&
2368                compare_two_spill_nodes(n, _available[i]) > 0) {
2369       break;
2370     }
2371   }
2372 
2373   // Special Check for compares following branches
2374   if( n->is_Mach() && _scheduled.size() > 0 ) {
2375     int op = n->as_Mach()->ideal_Opcode();
2376     Node *last = _scheduled[0];
2377     if( last->is_MachIf() && last->in(1) == n &&
2378         ( op == Op_CmpI ||
2379           op == Op_CmpU ||
2380           op == Op_CmpUL ||
2381           op == Op_CmpP ||
2382           op == Op_CmpF ||
2383           op == Op_CmpD ||
2384           op == Op_CmpL ) ) {
2385 
2386       // Recalculate position, moving to front of same latency
2387       for ( i=0 ; i < _available.size(); i++ )
2388         if (_current_latency[_available[i]->_idx] >= latency)
2389           break;
2390     }
2391   }
2392 
2393   // Insert the node in the available list
2394   _available.insert(i, n);
2395 
2396 #ifndef PRODUCT
2397   if (_cfg->C->trace_opto_output())
2398     dump_available();
2399 #endif
2400 }
2401 
2402 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2403   for ( uint i=0; i < n->len(); i++ ) {
2404     Node *def = n->in(i);
2405     if (!def) continue;
2406     if( def->is_Proj() )        // If this is a machine projection, then
2407       def = def->in(0);         // propagate usage thru to the base instruction
2408 
2409     if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2410       continue;
2411     }
2412 
2413     // Compute the latency
2414     uint l = _bundle_cycle_number + n->latency(i);
2415     if (_current_latency[def->_idx] < l)
2416       _current_latency[def->_idx] = l;
2417 
2418     // If this does not have uses then schedule it
2419     if ((--_uses[def->_idx]) == 0)
2420       AddNodeToAvailableList(def);
2421   }
2422 }
2423 
2424 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2425 #ifndef PRODUCT
2426   if (_cfg->C->trace_opto_output()) {
2427     tty->print("#   AddNodeToBundle: ");
2428     n->dump();
2429   }
2430 #endif
2431 
2432   // Remove this from the available list
2433   uint i;
2434   for (i = 0; i < _available.size(); i++)
2435     if (_available[i] == n)
2436       break;
2437   assert(i < _available.size(), "entry in _available list not found");
2438   _available.remove(i);
2439 
2440   // See if this fits in the current bundle
2441   const Pipeline *node_pipeline = n->pipeline();
2442   const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2443 
2444   // Check for instructions to be placed in the delay slot. We
2445   // do this before we actually schedule the current instruction,
2446   // because the delay slot follows the current instruction.
2447   if (Pipeline::_branch_has_delay_slot &&
2448       node_pipeline->hasBranchDelay() &&
2449       !_unconditional_delay_slot) {
2450 
2451     uint siz = _available.size();
2452 
2453     // Conditional branches can support an instruction that
2454     // is unconditionally executed and not dependent by the
2455     // branch, OR a conditionally executed instruction if
2456     // the branch is taken.  In practice, this means that
2457     // the first instruction at the branch target is
2458     // copied to the delay slot, and the branch goes to
2459     // the instruction after that at the branch target
2460     if ( n->is_MachBranch() ) {
2461 
2462       assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2463       assert( !n->is_Catch(),         "should not look for delay slot for Catch" );
2464 
2465 #ifndef PRODUCT
2466       _branches++;
2467 #endif
2468 
2469       // At least 1 instruction is on the available list
2470       // that is not dependent on the branch
2471       for (uint i = 0; i < siz; i++) {
2472         Node *d = _available[i];
2473         const Pipeline *avail_pipeline = d->pipeline();
2474 
2475         // Don't allow safepoints in the branch shadow, that will
2476         // cause a number of difficulties
2477         if ( avail_pipeline->instructionCount() == 1 &&
2478              !avail_pipeline->hasMultipleBundles() &&
2479              !avail_pipeline->hasBranchDelay() &&
2480              Pipeline::instr_has_unit_size() &&
2481              d->size(_regalloc) == Pipeline::instr_unit_size() &&
2482              NodeFitsInBundle(d) &&
2483              !node_bundling(d)->used_in_delay()) {
2484 
2485           if (d->is_Mach() && !d->is_MachSafePoint()) {
2486             // A node that fits in the delay slot was found, so we need to
2487             // set the appropriate bits in the bundle pipeline information so
2488             // that it correctly indicates resource usage.  Later, when we
2489             // attempt to add this instruction to the bundle, we will skip
2490             // setting the resource usage.
2491             _unconditional_delay_slot = d;
2492             node_bundling(n)->set_use_unconditional_delay();
2493             node_bundling(d)->set_used_in_unconditional_delay();
2494             _bundle_use.add_usage(avail_pipeline->resourceUse());
2495             _current_latency[d->_idx] = _bundle_cycle_number;
2496             _next_node = d;
2497             ++_bundle_instr_count;
2498 #ifndef PRODUCT
2499             _unconditional_delays++;
2500 #endif
2501             break;
2502           }
2503         }
2504       }
2505     }
2506 
2507     // No delay slot, add a nop to the usage
2508     if (!_unconditional_delay_slot) {
2509       // See if adding an instruction in the delay slot will overflow
2510       // the bundle.
2511       if (!NodeFitsInBundle(_nop)) {
2512 #ifndef PRODUCT
2513         if (_cfg->C->trace_opto_output())
2514           tty->print("#  *** STEP(1 instruction for delay slot) ***\n");
2515 #endif
2516         step(1);
2517       }
2518 
2519       _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2520       _next_node = _nop;
2521       ++_bundle_instr_count;
2522     }
2523 
2524     // See if the instruction in the delay slot requires a
2525     // step of the bundles
2526     if (!NodeFitsInBundle(n)) {
2527 #ifndef PRODUCT
2528       if (_cfg->C->trace_opto_output())
2529         tty->print("#  *** STEP(branch won't fit) ***\n");
2530 #endif
2531       // Update the state information
2532       _bundle_instr_count = 0;
2533       _bundle_cycle_number += 1;
2534       _bundle_use.step(1);
2535     }
2536   }
2537 
2538   // Get the number of instructions
2539   uint instruction_count = node_pipeline->instructionCount();
2540   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2541     instruction_count = 0;
2542 
2543   // Compute the latency information
2544   uint delay = 0;
2545 
2546   if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2547     int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2548     if (relative_latency < 0)
2549       relative_latency = 0;
2550 
2551     delay = _bundle_use.full_latency(relative_latency, node_usage);
2552 
2553     // Does not fit in this bundle, start a new one
2554     if (delay > 0) {
2555       step(delay);
2556 
2557 #ifndef PRODUCT
2558       if (_cfg->C->trace_opto_output())
2559         tty->print("#  *** STEP(%d) ***\n", delay);
2560 #endif
2561     }
2562   }
2563 
2564   // If this was placed in the delay slot, ignore it
2565   if (n != _unconditional_delay_slot) {
2566 
2567     if (delay == 0) {
2568       if (node_pipeline->hasMultipleBundles()) {
2569 #ifndef PRODUCT
2570         if (_cfg->C->trace_opto_output())
2571           tty->print("#  *** STEP(multiple instructions) ***\n");
2572 #endif
2573         step(1);
2574       }
2575 
2576       else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2577 #ifndef PRODUCT
2578         if (_cfg->C->trace_opto_output())
2579           tty->print("#  *** STEP(%d >= %d instructions) ***\n",
2580                      instruction_count + _bundle_instr_count,
2581                      Pipeline::_max_instrs_per_cycle);
2582 #endif
2583         step(1);
2584       }
2585     }
2586 
2587     if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2588       _bundle_instr_count++;
2589 
2590     // Set the node's latency
2591     _current_latency[n->_idx] = _bundle_cycle_number;
2592 
2593     // Now merge the functional unit information
2594     if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2595       _bundle_use.add_usage(node_usage);
2596 
2597     // Increment the number of instructions in this bundle
2598     _bundle_instr_count += instruction_count;
2599 
2600     // Remember this node for later
2601     if (n->is_Mach())
2602       _next_node = n;
2603   }
2604 
2605   // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2606   // not in the bb->_nodes array.  This happens for debug-info-only BoxLocks.
2607   // 'Schedule' them (basically ignore in the schedule) but do not insert them
2608   // into the block.  All other scheduled nodes get put in the schedule here.
2609   int op = n->Opcode();
2610   if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2611       (op != Op_Node &&         // Not an unused antidepedence node and
2612        // not an unallocated boxlock
2613        (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2614 
2615     // Push any trailing projections
2616     if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2617       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2618         Node *foi = n->fast_out(i);
2619         if( foi->is_Proj() )
2620           _scheduled.push(foi);
2621       }
2622     }
2623 
2624     // Put the instruction in the schedule list
2625     _scheduled.push(n);
2626   }
2627 
2628 #ifndef PRODUCT
2629   if (_cfg->C->trace_opto_output())
2630     dump_available();
2631 #endif
2632 
2633   // Walk all the definitions, decrementing use counts, and
2634   // if a definition has a 0 use count, place it in the available list.
2635   DecrementUseCounts(n,bb);
2636 }
2637 
2638 // This method sets the use count within a basic block.  We will ignore all
2639 // uses outside the current basic block.  As we are doing a backwards walk,
2640 // any node we reach that has a use count of 0 may be scheduled.  This also
2641 // avoids the problem of cyclic references from phi nodes, as long as phi
2642 // nodes are at the front of the basic block.  This method also initializes
2643 // the available list to the set of instructions that have no uses within this
2644 // basic block.
2645 void Scheduling::ComputeUseCount(const Block *bb) {
2646 #ifndef PRODUCT
2647   if (_cfg->C->trace_opto_output())
2648     tty->print("# -> ComputeUseCount\n");
2649 #endif
2650 
2651   // Clear the list of available and scheduled instructions, just in case
2652   _available.clear();
2653   _scheduled.clear();
2654 
2655   // No delay slot specified
2656   _unconditional_delay_slot = nullptr;
2657 
2658 #ifdef ASSERT
2659   for( uint i=0; i < bb->number_of_nodes(); i++ )
2660     assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2661 #endif
2662 
2663   // Force the _uses count to never go to zero for unscheduable pieces
2664   // of the block
2665   for( uint k = 0; k < _bb_start; k++ )
2666     _uses[bb->get_node(k)->_idx] = 1;
2667   for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2668     _uses[bb->get_node(l)->_idx] = 1;
2669 
2670   // Iterate backwards over the instructions in the block.  Don't count the
2671   // branch projections at end or the block header instructions.
2672   for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2673     Node *n = bb->get_node(j);
2674     if( n->is_Proj() ) continue; // Projections handled another way
2675 
2676     // Account for all uses
2677     for ( uint k = 0; k < n->len(); k++ ) {
2678       Node *inp = n->in(k);
2679       if (!inp) continue;
2680       assert(inp != n, "no cycles allowed" );
2681       if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2682         if (inp->is_Proj()) { // Skip through Proj's
2683           inp = inp->in(0);
2684         }
2685         ++_uses[inp->_idx];     // Count 1 block-local use
2686       }
2687     }
2688 
2689     // If this instruction has a 0 use count, then it is available
2690     if (!_uses[n->_idx]) {
2691       _current_latency[n->_idx] = _bundle_cycle_number;
2692       AddNodeToAvailableList(n);
2693     }
2694 
2695 #ifndef PRODUCT
2696     if (_cfg->C->trace_opto_output()) {
2697       tty->print("#   uses: %3d: ", _uses[n->_idx]);
2698       n->dump();
2699     }
2700 #endif
2701   }
2702 
2703 #ifndef PRODUCT
2704   if (_cfg->C->trace_opto_output())
2705     tty->print("# <- ComputeUseCount\n");
2706 #endif
2707 }
2708 
2709 // This routine performs scheduling on each basic block in reverse order,
2710 // using instruction latencies and taking into account function unit
2711 // availability.
2712 void Scheduling::DoScheduling() {
2713 #ifndef PRODUCT
2714   if (_cfg->C->trace_opto_output())
2715     tty->print("# -> DoScheduling\n");
2716 #endif
2717 
2718   Block *succ_bb = nullptr;
2719   Block *bb;
2720   Compile* C = Compile::current();
2721 
2722   // Walk over all the basic blocks in reverse order
2723   for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2724     bb = _cfg->get_block(i);
2725 
2726 #ifndef PRODUCT
2727     if (_cfg->C->trace_opto_output()) {
2728       tty->print("#  Schedule BB#%03d (initial)\n", i);
2729       for (uint j = 0; j < bb->number_of_nodes(); j++) {
2730         bb->get_node(j)->dump();
2731       }
2732     }
2733 #endif
2734 
2735     // On the head node, skip processing
2736     if (bb == _cfg->get_root_block()) {
2737       continue;
2738     }
2739 
2740     // Skip empty, connector blocks
2741     if (bb->is_connector())
2742       continue;
2743 
2744     // If the following block is not the sole successor of
2745     // this one, then reset the pipeline information
2746     if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2747 #ifndef PRODUCT
2748       if (_cfg->C->trace_opto_output()) {
2749         tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2750                    _next_node->_idx, _bundle_instr_count);
2751       }
2752 #endif
2753       step_and_clear();
2754     }
2755 
2756     // Leave untouched the starting instruction, any Phis, a CreateEx node
2757     // or Top.  bb->get_node(_bb_start) is the first schedulable instruction.
2758     _bb_end = bb->number_of_nodes()-1;
2759     for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2760       Node *n = bb->get_node(_bb_start);
2761       // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2762       // Also, MachIdealNodes do not get scheduled
2763       if( !n->is_Mach() ) continue;     // Skip non-machine nodes
2764       MachNode *mach = n->as_Mach();
2765       int iop = mach->ideal_Opcode();
2766       if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2767       if( iop == Op_Con ) continue;      // Do not schedule Top
2768       if( iop == Op_Node &&     // Do not schedule PhiNodes, ProjNodes
2769           mach->pipeline() == MachNode::pipeline_class() &&
2770           !n->is_SpillCopy() && !n->is_MachMerge() )  // Breakpoints, Prolog, etc
2771         continue;
2772       break;                    // Funny loop structure to be sure...
2773     }
2774     // Compute last "interesting" instruction in block - last instruction we
2775     // might schedule.  _bb_end points just after last schedulable inst.  We
2776     // normally schedule conditional branches (despite them being forced last
2777     // in the block), because they have delay slots we can fill.  Calls all
2778     // have their delay slots filled in the template expansions, so we don't
2779     // bother scheduling them.
2780     Node *last = bb->get_node(_bb_end);
2781     // Ignore trailing NOPs.
2782     while (_bb_end > 0 && last->is_Mach() &&
2783            last->as_Mach()->ideal_Opcode() == Op_Con) {
2784       last = bb->get_node(--_bb_end);
2785     }
2786     assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2787     if( last->is_Catch() ||
2788         (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2789       // There might be a prior call.  Skip it.
2790       while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2791     } else if( last->is_MachNullCheck() ) {
2792       // Backup so the last null-checked memory instruction is
2793       // outside the schedulable range. Skip over the nullcheck,
2794       // projection, and the memory nodes.
2795       Node *mem = last->in(1);
2796       do {
2797         _bb_end--;
2798       } while (mem != bb->get_node(_bb_end));
2799     } else {
2800       // Set _bb_end to point after last schedulable inst.
2801       _bb_end++;
2802     }
2803 
2804     assert( _bb_start <= _bb_end, "inverted block ends" );
2805 
2806     // Compute the register antidependencies for the basic block
2807     ComputeRegisterAntidependencies(bb);
2808     if (C->failing())  return;  // too many D-U pinch points
2809 
2810     // Compute the usage within the block, and set the list of all nodes
2811     // in the block that have no uses within the block.
2812     ComputeUseCount(bb);
2813 
2814     // Schedule the remaining instructions in the block
2815     while ( _available.size() > 0 ) {
2816       Node *n = ChooseNodeToBundle();
2817       guarantee(n != nullptr, "no nodes available");
2818       AddNodeToBundle(n,bb);
2819     }
2820 
2821     assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2822 #ifdef ASSERT
2823     for( uint l = _bb_start; l < _bb_end; l++ ) {
2824       Node *n = bb->get_node(l);
2825       uint m;
2826       for( m = 0; m < _bb_end-_bb_start; m++ )
2827         if( _scheduled[m] == n )
2828           break;
2829       assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2830     }
2831 #endif
2832 
2833     // Now copy the instructions (in reverse order) back to the block
2834     for ( uint k = _bb_start; k < _bb_end; k++ )
2835       bb->map_node(_scheduled[_bb_end-k-1], k);
2836 
2837 #ifndef PRODUCT
2838     if (_cfg->C->trace_opto_output()) {
2839       tty->print("#  Schedule BB#%03d (final)\n", i);
2840       uint current = 0;
2841       for (uint j = 0; j < bb->number_of_nodes(); j++) {
2842         Node *n = bb->get_node(j);
2843         if( valid_bundle_info(n) ) {
2844           Bundle *bundle = node_bundling(n);
2845           if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2846             tty->print("*** Bundle: ");
2847             bundle->dump();
2848           }
2849           n->dump();
2850         }
2851       }
2852     }
2853 #endif
2854 #ifdef ASSERT
2855     verify_good_schedule(bb,"after block local scheduling");
2856 #endif
2857   }
2858 
2859 #ifndef PRODUCT
2860   if (_cfg->C->trace_opto_output())
2861     tty->print("# <- DoScheduling\n");
2862 #endif
2863 
2864   // Record final node-bundling array location
2865   _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2866 
2867 } // end DoScheduling
2868 
2869 // Verify that no live-range used in the block is killed in the block by a
2870 // wrong DEF.  This doesn't verify live-ranges that span blocks.
2871 
2872 // Check for edge existence.  Used to avoid adding redundant precedence edges.
2873 static bool edge_from_to( Node *from, Node *to ) {
2874   for( uint i=0; i<from->len(); i++ )
2875     if( from->in(i) == to )
2876       return true;
2877   return false;
2878 }
2879 
2880 #ifdef ASSERT
2881 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2882   // Check for bad kills
2883   if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2884     Node *prior_use = _reg_node[def];
2885     if( prior_use && !edge_from_to(prior_use,n) ) {
2886       tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2887       n->dump();
2888       tty->print_cr("...");
2889       prior_use->dump();
2890       assert(edge_from_to(prior_use,n), "%s", msg);
2891     }
2892     _reg_node.map(def,nullptr); // Kill live USEs
2893   }
2894 }
2895 
2896 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2897 
2898   // Zap to something reasonable for the verify code
2899   _reg_node.clear();
2900 
2901   // Walk over the block backwards.  Check to make sure each DEF doesn't
2902   // kill a live value (other than the one it's supposed to).  Add each
2903   // USE to the live set.
2904   for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2905     Node *n = b->get_node(i);
2906     int n_op = n->Opcode();
2907     if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2908       // Fat-proj kills a slew of registers
2909       RegMaskIterator rmi(n->out_RegMask());
2910       while (rmi.has_next()) {
2911         OptoReg::Name kill = rmi.next();
2912         verify_do_def(n, kill, msg);
2913       }
2914     } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2915       // Get DEF'd registers the normal way
2916       verify_do_def( n, _regalloc->get_reg_first(n), msg );
2917       verify_do_def( n, _regalloc->get_reg_second(n), msg );
2918     }
2919 
2920     // Now make all USEs live
2921     for( uint i=1; i<n->req(); i++ ) {
2922       Node *def = n->in(i);
2923       assert(def != nullptr, "input edge required");
2924       OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2925       OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2926       if( OptoReg::is_valid(reg_lo) ) {
2927         assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2928         _reg_node.map(reg_lo,n);
2929       }
2930       if( OptoReg::is_valid(reg_hi) ) {
2931         assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2932         _reg_node.map(reg_hi,n);
2933       }
2934     }
2935 
2936   }
2937 
2938   // Zap to something reasonable for the Antidependence code
2939   _reg_node.clear();
2940 }
2941 #endif
2942 
2943 // Conditionally add precedence edges.  Avoid putting edges on Projs.
2944 static void add_prec_edge_from_to( Node *from, Node *to ) {
2945   if( from->is_Proj() ) {       // Put precedence edge on Proj's input
2946     assert( from->req() == 1 && (from->len() == 1 || from->in(1) == nullptr), "no precedence edges on projections" );
2947     from = from->in(0);
2948   }
2949   if( from != to &&             // No cycles (for things like LD L0,[L0+4] )
2950       !edge_from_to( from, to ) ) // Avoid duplicate edge
2951     from->add_prec(to);
2952 }
2953 
2954 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2955   if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2956     return;
2957 
2958   if (OptoReg::is_reg(def_reg)) {
2959     VMReg vmreg = OptoReg::as_VMReg(def_reg);
2960     if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
2961       // This is one of the high slots of a vector register.
2962       // ScheduleAndBundle already checked there are no live wide
2963       // vectors in this method so it can be safely ignored.
2964       return;
2965     }
2966   }
2967 
2968   Node *pinch = _reg_node[def_reg]; // Get pinch point
2969   if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2970       is_def ) {    // Check for a true def (not a kill)
2971     _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2972     return;
2973   }
2974 
2975   Node *kill = def;             // Rename 'def' to more descriptive 'kill'
2976   DEBUG_ONLY( def = (Node*)((intptr_t)0xdeadbeef); )
2977 
2978   // After some number of kills there _may_ be a later def
2979   Node *later_def = nullptr;
2980 
2981   Compile* C = Compile::current();
2982 
2983   // Finding a kill requires a real pinch-point.
2984   // Check for not already having a pinch-point.
2985   // Pinch points are Op_Node's.
2986   if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2987     later_def = pinch;            // Must be def/kill as optimistic pinch-point
2988     if ( _pinch_free_list.size() > 0) {
2989       pinch = _pinch_free_list.pop();
2990     } else {
2991       pinch = new Node(1); // Pinch point to-be
2992     }
2993     if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2994       DEBUG_ONLY( pinch->dump(); );
2995       assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index());
2996       _cfg->C->record_method_not_compilable("too many D-U pinch points");
2997       return;
2998     }
2999     _cfg->map_node_to_block(pinch, b);      // Pretend it's valid in this block (lazy init)
3000     _reg_node.map(def_reg,pinch); // Record pinch-point
3001     //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
3002     if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
3003       pinch->init_req(0, C->top());     // set not null for the next call
3004       add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
3005       later_def = nullptr;           // and no later def
3006     }
3007     pinch->set_req(0,later_def);  // Hook later def so we can find it
3008   } else {                        // Else have valid pinch point
3009     if( pinch->in(0) )            // If there is a later-def
3010       later_def = pinch->in(0);   // Get it
3011   }
3012 
3013   // Add output-dependence edge from later def to kill
3014   if( later_def )               // If there is some original def
3015     add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
3016 
3017   // See if current kill is also a use, and so is forced to be the pinch-point.
3018   if( pinch->Opcode() == Op_Node ) {
3019     Node *uses = kill->is_Proj() ? kill->in(0) : kill;
3020     for( uint i=1; i<uses->req(); i++ ) {
3021       if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
3022           _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
3023         // Yes, found a use/kill pinch-point
3024         pinch->set_req(0,nullptr);  //
3025         pinch->replace_by(kill); // Move anti-dep edges up
3026         pinch = kill;
3027         _reg_node.map(def_reg,pinch);
3028         return;
3029       }
3030     }
3031   }
3032 
3033   // Add edge from kill to pinch-point
3034   add_prec_edge_from_to(kill,pinch);
3035 }
3036 
3037 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
3038   if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
3039     return;
3040   Node *pinch = _reg_node[use_reg]; // Get pinch point
3041   // Check for no later def_reg/kill in block
3042   if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b &&
3043       // Use has to be block-local as well
3044       _cfg->get_block_for_node(use) == b) {
3045     if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
3046         pinch->req() == 1 ) {   // pinch not yet in block?
3047       pinch->del_req(0);        // yank pointer to later-def, also set flag
3048       // Insert the pinch-point in the block just after the last use
3049       b->insert_node(pinch, b->find_node(use) + 1);
3050       _bb_end++;                // Increase size scheduled region in block
3051     }
3052 
3053     add_prec_edge_from_to(pinch,use);
3054   }
3055 }
3056 
3057 // We insert antidependences between the reads and following write of
3058 // allocated registers to prevent illegal code motion. Hopefully, the
3059 // number of added references should be fairly small, especially as we
3060 // are only adding references within the current basic block.
3061 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
3062 
3063 #ifdef ASSERT
3064   verify_good_schedule(b,"before block local scheduling");
3065 #endif
3066 
3067   // A valid schedule, for each register independently, is an endless cycle
3068   // of: a def, then some uses (connected to the def by true dependencies),
3069   // then some kills (defs with no uses), finally the cycle repeats with a new
3070   // def.  The uses are allowed to float relative to each other, as are the
3071   // kills.  No use is allowed to slide past a kill (or def).  This requires
3072   // antidependencies between all uses of a single def and all kills that
3073   // follow, up to the next def.  More edges are redundant, because later defs
3074   // & kills are already serialized with true or antidependencies.  To keep
3075   // the edge count down, we add a 'pinch point' node if there's more than
3076   // one use or more than one kill/def.
3077 
3078   // We add dependencies in one bottom-up pass.
3079 
3080   // For each instruction we handle it's DEFs/KILLs, then it's USEs.
3081 
3082   // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
3083   // register.  If not, we record the DEF/KILL in _reg_node, the
3084   // register-to-def mapping.  If there is a prior DEF/KILL, we insert a
3085   // "pinch point", a new Node that's in the graph but not in the block.
3086   // We put edges from the prior and current DEF/KILLs to the pinch point.
3087   // We put the pinch point in _reg_node.  If there's already a pinch point
3088   // we merely add an edge from the current DEF/KILL to the pinch point.
3089 
3090   // After doing the DEF/KILLs, we handle USEs.  For each used register, we
3091   // put an edge from the pinch point to the USE.
3092 
3093   // To be expedient, the _reg_node array is pre-allocated for the whole
3094   // compilation.  _reg_node is lazily initialized; it either contains a null,
3095   // or a valid def/kill/pinch-point, or a leftover node from some prior
3096   // block.  Leftover node from some prior block is treated like a null (no
3097   // prior def, so no anti-dependence needed).  Valid def is distinguished by
3098   // it being in the current block.
3099   bool fat_proj_seen = false;
3100   uint last_safept = _bb_end-1;
3101   Node* end_node         = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr;
3102   Node* last_safept_node = end_node;
3103   for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
3104     Node *n = b->get_node(i);
3105     int is_def = n->outcnt();   // def if some uses prior to adding precedence edges
3106     if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
3107       // Fat-proj kills a slew of registers
3108       // This can add edges to 'n' and obscure whether or not it was a def,
3109       // hence the is_def flag.
3110       fat_proj_seen = true;
3111       RegMaskIterator rmi(n->out_RegMask());
3112       while (rmi.has_next()) {
3113         OptoReg::Name kill = rmi.next();
3114         anti_do_def(b, n, kill, is_def);
3115       }
3116     } else {
3117       // Get DEF'd registers the normal way
3118       anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
3119       anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
3120     }
3121 
3122     // Kill projections on a branch should appear to occur on the
3123     // branch, not afterwards, so grab the masks from the projections
3124     // and process them.
3125     if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
3126       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3127         Node* use = n->fast_out(i);
3128         if (use->is_Proj()) {
3129           RegMaskIterator rmi(use->out_RegMask());
3130           while (rmi.has_next()) {
3131             OptoReg::Name kill = rmi.next();
3132             anti_do_def(b, n, kill, false);
3133           }
3134         }
3135       }
3136     }
3137 
3138     // Check each register used by this instruction for a following DEF/KILL
3139     // that must occur afterward and requires an anti-dependence edge.
3140     for( uint j=0; j<n->req(); j++ ) {
3141       Node *def = n->in(j);
3142       if( def ) {
3143         assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
3144         anti_do_use( b, n, _regalloc->get_reg_first(def) );
3145         anti_do_use( b, n, _regalloc->get_reg_second(def) );
3146       }
3147     }
3148     // Do not allow defs of new derived values to float above GC
3149     // points unless the base is definitely available at the GC point.
3150 
3151     Node *m = b->get_node(i);
3152 
3153     // Add precedence edge from following safepoint to use of derived pointer
3154     if( last_safept_node != end_node &&
3155         m != last_safept_node) {
3156       for (uint k = 1; k < m->req(); k++) {
3157         const Type *t = m->in(k)->bottom_type();
3158         if( t->isa_oop_ptr() &&
3159             t->is_ptr()->offset() != 0 ) {
3160           last_safept_node->add_prec( m );
3161           break;
3162         }
3163       }
3164     }
3165 
3166     if( n->jvms() ) {           // Precedence edge from derived to safept
3167       // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3168       if( b->get_node(last_safept) != last_safept_node ) {
3169         last_safept = b->find_node(last_safept_node);
3170       }
3171       for( uint j=last_safept; j > i; j-- ) {
3172         Node *mach = b->get_node(j);
3173         if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3174           mach->add_prec( n );
3175       }
3176       last_safept = i;
3177       last_safept_node = m;
3178     }
3179   }
3180 
3181   if (fat_proj_seen) {
3182     // Garbage collect pinch nodes that were not consumed.
3183     // They are usually created by a fat kill MachProj for a call.
3184     garbage_collect_pinch_nodes();
3185   }
3186 }
3187 
3188 // Garbage collect pinch nodes for reuse by other blocks.
3189 //
3190 // The block scheduler's insertion of anti-dependence
3191 // edges creates many pinch nodes when the block contains
3192 // 2 or more Calls.  A pinch node is used to prevent a
3193 // combinatorial explosion of edges.  If a set of kills for a
3194 // register is anti-dependent on a set of uses (or defs), rather
3195 // than adding an edge in the graph between each pair of kill
3196 // and use (or def), a pinch is inserted between them:
3197 //
3198 //            use1   use2  use3
3199 //                \   |   /
3200 //                 \  |  /
3201 //                  pinch
3202 //                 /  |  \
3203 //                /   |   \
3204 //            kill1 kill2 kill3
3205 //
3206 // One pinch node is created per register killed when
3207 // the second call is encountered during a backwards pass
3208 // over the block.  Most of these pinch nodes are never
3209 // wired into the graph because the register is never
3210 // used or def'ed in the block.
3211 //
3212 void Scheduling::garbage_collect_pinch_nodes() {
3213 #ifndef PRODUCT
3214   if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3215 #endif
3216   int trace_cnt = 0;
3217   for (uint k = 0; k < _reg_node.max(); k++) {
3218     Node* pinch = _reg_node[k];
3219     if ((pinch != nullptr) && pinch->Opcode() == Op_Node &&
3220         // no predecence input edges
3221         (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) {
3222       cleanup_pinch(pinch);
3223       _pinch_free_list.push(pinch);
3224       _reg_node.map(k, nullptr);
3225 #ifndef PRODUCT
3226       if (_cfg->C->trace_opto_output()) {
3227         trace_cnt++;
3228         if (trace_cnt > 40) {
3229           tty->print("\n");
3230           trace_cnt = 0;
3231         }
3232         tty->print(" %d", pinch->_idx);
3233       }
3234 #endif
3235     }
3236   }
3237 #ifndef PRODUCT
3238   if (_cfg->C->trace_opto_output()) tty->print("\n");
3239 #endif
3240 }
3241 
3242 // Clean up a pinch node for reuse.
3243 void Scheduling::cleanup_pinch( Node *pinch ) {
3244   assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3245 
3246   for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3247     Node* use = pinch->last_out(i);
3248     uint uses_found = 0;
3249     for (uint j = use->req(); j < use->len(); j++) {
3250       if (use->in(j) == pinch) {
3251         use->rm_prec(j);
3252         uses_found++;
3253       }
3254     }
3255     assert(uses_found > 0, "must be a precedence edge");
3256     i -= uses_found;    // we deleted 1 or more copies of this edge
3257   }
3258   // May have a later_def entry
3259   pinch->set_req(0, nullptr);
3260 }
3261 
3262 #ifndef PRODUCT
3263 
3264 void Scheduling::dump_available() const {
3265   tty->print("#Availist  ");
3266   for (uint i = 0; i < _available.size(); i++)
3267     tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3268   tty->cr();
3269 }
3270 
3271 // Print Scheduling Statistics
3272 void Scheduling::print_statistics() {
3273   // Print the size added by nops for bundling
3274   tty->print("Nops added %d bytes to total of %d bytes",
3275              _total_nop_size, _total_method_size);
3276   if (_total_method_size > 0)
3277     tty->print(", for %.2f%%",
3278                ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3279   tty->print("\n");
3280 
3281   // Print the number of branch shadows filled
3282   if (Pipeline::_branch_has_delay_slot) {
3283     tty->print("Of %d branches, %d had unconditional delay slots filled",
3284                _total_branches, _total_unconditional_delays);
3285     if (_total_branches > 0)
3286       tty->print(", for %.2f%%",
3287                  ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3288     tty->print("\n");
3289   }
3290 
3291   uint total_instructions = 0, total_bundles = 0;
3292 
3293   for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3294     uint bundle_count   = _total_instructions_per_bundle[i];
3295     total_instructions += bundle_count * i;
3296     total_bundles      += bundle_count;
3297   }
3298 
3299   if (total_bundles > 0)
3300     tty->print("Average ILP (excluding nops) is %.2f\n",
3301                ((double)total_instructions) / ((double)total_bundles));
3302 }
3303 #endif
3304 
3305 //-----------------------init_scratch_buffer_blob------------------------------
3306 // Construct a temporary BufferBlob and cache it for this compile.
3307 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3308   // If there is already a scratch buffer blob allocated and the
3309   // constant section is big enough, use it.  Otherwise free the
3310   // current and allocate a new one.
3311   BufferBlob* blob = scratch_buffer_blob();
3312   if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3313     // Use the current blob.
3314   } else {
3315     if (blob != nullptr) {
3316       BufferBlob::free(blob);
3317     }
3318 
3319     ResourceMark rm;
3320     _scratch_const_size = const_size;
3321     int size = C2Compiler::initial_code_buffer_size(const_size);
3322     blob = BufferBlob::create("Compile::scratch_buffer", size);
3323     // Record the buffer blob for next time.
3324     set_scratch_buffer_blob(blob);
3325     // Have we run out of code space?
3326     if (scratch_buffer_blob() == nullptr) {
3327       // Let CompilerBroker disable further compilations.
3328       C->record_failure("Not enough space for scratch buffer in CodeCache");
3329       return;
3330     }
3331   }
3332 
3333   // Initialize the relocation buffers
3334   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3335   set_scratch_locs_memory(locs_buf);
3336 }
3337 
3338 
3339 //-----------------------scratch_emit_size-------------------------------------
3340 // Helper function that computes size by emitting code
3341 uint PhaseOutput::scratch_emit_size(const Node* n) {
3342   // Start scratch_emit_size section.
3343   set_in_scratch_emit_size(true);
3344 
3345   // Emit into a trash buffer and count bytes emitted.
3346   // This is a pretty expensive way to compute a size,
3347   // but it works well enough if seldom used.
3348   // All common fixed-size instructions are given a size
3349   // method by the AD file.
3350   // Note that the scratch buffer blob and locs memory are
3351   // allocated at the beginning of the compile task, and
3352   // may be shared by several calls to scratch_emit_size.
3353   // The allocation of the scratch buffer blob is particularly
3354   // expensive, since it has to grab the code cache lock.
3355   BufferBlob* blob = this->scratch_buffer_blob();
3356   assert(blob != nullptr, "Initialize BufferBlob at start");
3357   assert(blob->size() > max_inst_size(), "sanity");
3358   relocInfo* locs_buf = scratch_locs_memory();
3359   address blob_begin = blob->content_begin();
3360   address blob_end   = (address)locs_buf;
3361   assert(blob->contains(blob_end), "sanity");
3362   CodeBuffer buf(blob_begin, blob_end - blob_begin);
3363   buf.initialize_consts_size(_scratch_const_size);
3364   buf.initialize_stubs_size(MAX_stubs_size);
3365   assert(locs_buf != nullptr, "sanity");
3366   int lsize = MAX_locs_size / 3;
3367   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3368   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3369   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3370   // Mark as scratch buffer.
3371   buf.consts()->set_scratch_emit();
3372   buf.insts()->set_scratch_emit();
3373   buf.stubs()->set_scratch_emit();
3374 
3375   // Do the emission.
3376 
3377   Label fakeL; // Fake label for branch instructions.
3378   Label*   saveL = nullptr;
3379   uint save_bnum = 0;
3380   bool is_branch = n->is_MachBranch();
3381   C2_MacroAssembler masm(&buf);
3382   masm.bind(fakeL);
3383   if (is_branch) {
3384     n->as_MachBranch()->save_label(&saveL, &save_bnum);
3385     n->as_MachBranch()->label_set(&fakeL, 0);
3386   }
3387   n->emit(&masm, C->regalloc());
3388 
3389   // Emitting into the scratch buffer should not fail
3390   assert(!C->failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3391 
3392   if (is_branch) // Restore label.
3393     n->as_MachBranch()->label_set(saveL, save_bnum);
3394 
3395   // End scratch_emit_size section.
3396   set_in_scratch_emit_size(false);
3397 
3398   return buf.insts_size();
3399 }
3400 
3401 void PhaseOutput::install() {
3402   if (C->should_install_code() && C->stub_function() != nullptr) {
3403     install_stub(C->stub_name());
3404   } else {
3405     install_code(C->method(),
3406                  C->entry_bci(),
3407                  CompilerThread::current()->compiler(),
3408                  C->has_unsafe_access(),
3409                  SharedRuntime::is_wide_vector(C->max_vector_size()));
3410   }
3411 }
3412 
3413 void PhaseOutput::install_code(ciMethod*         target,
3414                                int               entry_bci,
3415                                AbstractCompiler* compiler,
3416                                bool              has_unsafe_access,
3417                                bool              has_wide_vectors) {
3418   // Check if we want to skip execution of all compiled code.
3419   {
3420 #ifndef PRODUCT
3421     if (OptoNoExecute) {
3422       C->record_method_not_compilable("+OptoNoExecute");  // Flag as failed
3423       return;
3424     }
3425 #endif
3426     Compile::TracePhase tp(_t_registerMethod);
3427 
3428     if (C->is_osr_compilation()) {
3429       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3430       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3431     } else {
3432       if (!target->is_static()) {
3433         // The UEP of an nmethod ensures that the VEP is padded. However, the padding of the UEP is placed
3434         // before the inline cache check, so we don't have to execute any nop instructions when dispatching
3435         // through the UEP, yet we can ensure that the VEP is aligned appropriately.
3436         _code_offsets.set_value(CodeOffsets::Entry, _first_block_size - MacroAssembler::ic_check_size());
3437       }
3438       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3439       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3440     }
3441 
3442     C->env()->register_method(target,
3443                                      entry_bci,
3444                                      &_code_offsets,
3445                                      _orig_pc_slot_offset_in_bytes,
3446                                      code_buffer(),
3447                                      frame_size_in_words(),
3448                                      oop_map_set(),
3449                                      &_handler_table,
3450                                      inc_table(),
3451                                      compiler,
3452                                      C->has_clinit_barriers(),
3453                                      C->for_preload(),
3454                                      has_unsafe_access,
3455                                      SharedRuntime::is_wide_vector(C->max_vector_size()),
3456                                      C->has_monitors(),
3457                                      C->has_scoped_access(),
3458                                      0,
3459                                      C->should_install_code());
3460 
3461     if (C->log() != nullptr) { // Print code cache state into compiler log
3462       C->log()->code_cache_state();
3463     }
3464     assert(!C->has_clinit_barriers() || C->for_preload(), "class init barriers should be only in preload code");
3465   }
3466 }
3467 void PhaseOutput::install_stub(const char* stub_name) {
3468   // Entry point will be accessed using stub_entry_point();
3469   if (code_buffer() == nullptr) {
3470     Matcher::soft_match_failure();
3471   } else {
3472     if (PrintAssembly && (WizardMode || Verbose))
3473       tty->print_cr("### Stub::%s", stub_name);
3474 
3475     if (!C->failing()) {
3476       assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3477 
3478       // Make the NMethod
3479       // For now we mark the frame as never safe for profile stackwalking
3480       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3481                                                       code_buffer(),
3482                                                       CodeOffsets::frame_never_safe,
3483                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
3484                                                       frame_size_in_words(),
3485                                                       oop_map_set(),
3486                                                       false,
3487                                                       false);
3488 
3489       if (rs == nullptr) {
3490         C->record_failure("CodeCache is full");
3491       } else {
3492         assert(rs->is_runtime_stub(), "sanity check");
3493         C->set_stub_entry_point(rs->entry_point());
3494         AOTCodeCache::store_code_blob(*rs, AOTCodeEntry::C2Blob, C->stub_id(), stub_name);
3495       }
3496     }
3497   }
3498 }
3499 
3500 // Support for bundling info
3501 Bundle* PhaseOutput::node_bundling(const Node *n) {
3502   assert(valid_bundle_info(n), "oob");
3503   return &_node_bundling_base[n->_idx];
3504 }
3505 
3506 bool PhaseOutput::valid_bundle_info(const Node *n) {
3507   return (_node_bundling_limit > n->_idx);
3508 }
3509 
3510 //------------------------------frame_size_in_words-----------------------------
3511 // frame_slots in units of words
3512 int PhaseOutput::frame_size_in_words() const {
3513   // shift is 0 in LP32 and 1 in LP64
3514   const int shift = (LogBytesPerWord - LogBytesPerInt);
3515   int words = _frame_slots >> shift;
3516   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3517   return words;
3518 }
3519 
3520 // To bang the stack of this compiled method we use the stack size
3521 // that the interpreter would need in case of a deoptimization. This
3522 // removes the need to bang the stack in the deoptimization blob which
3523 // in turn simplifies stack overflow handling.
3524 int PhaseOutput::bang_size_in_bytes() const {
3525   return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3526 }
3527 
3528 //------------------------------dump_asm---------------------------------------
3529 // Dump formatted assembly
3530 #if defined(SUPPORT_OPTO_ASSEMBLY)
3531 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3532 
3533   int pc_digits = 3; // #chars required for pc
3534   int sb_chars  = 3; // #chars for "start bundle" indicator
3535   int tab_size  = 8;
3536   if (pcs != nullptr) {
3537     int max_pc = 0;
3538     for (uint i = 0; i < pc_limit; i++) {
3539       max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3540     }
3541     pc_digits  = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3542   }
3543   int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3544 
3545   bool cut_short = false;
3546   st->print_cr("#");
3547   st->print("#  ");  C->tf()->dump_on(st);  st->cr();
3548   st->print_cr("#");
3549 
3550   // For all blocks
3551   int pc = 0x0;                 // Program counter
3552   char starts_bundle = ' ';
3553   C->regalloc()->dump_frame();
3554 
3555   Node *n = nullptr;
3556   for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3557     if (VMThread::should_terminate()) {
3558       cut_short = true;
3559       break;
3560     }
3561     Block* block = C->cfg()->get_block(i);
3562     if (block->is_connector() && !Verbose) {
3563       continue;
3564     }
3565     n = block->head();
3566     if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3567       pc = pcs[n->_idx];
3568       st->print("%*.*x", pc_digits, pc_digits, pc);
3569     }
3570     st->fill_to(prefix_len);
3571     block->dump_head(C->cfg(), st);
3572     if (block->is_connector()) {
3573       st->fill_to(prefix_len);
3574       st->print_cr("# Empty connector block");
3575     } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3576       st->fill_to(prefix_len);
3577       st->print_cr("# Block is sole successor of call");
3578     }
3579 
3580     // For all instructions
3581     Node *delay = nullptr;
3582     for (uint j = 0; j < block->number_of_nodes(); j++) {
3583       if (VMThread::should_terminate()) {
3584         cut_short = true;
3585         break;
3586       }
3587       n = block->get_node(j);
3588       if (valid_bundle_info(n)) {
3589         Bundle* bundle = node_bundling(n);
3590         if (bundle->used_in_unconditional_delay()) {
3591           delay = n;
3592           continue;
3593         }
3594         if (bundle->starts_bundle()) {
3595           starts_bundle = '+';
3596         }
3597       }
3598 
3599       if (WizardMode) {
3600         n->dump();
3601       }
3602 
3603       if( !n->is_Region() &&    // Dont print in the Assembly
3604           !n->is_Phi() &&       // a few noisely useless nodes
3605           !n->is_Proj() &&
3606           !n->is_MachTemp() &&
3607           !n->is_SafePointScalarObject() &&
3608           !n->is_Catch() &&     // Would be nice to print exception table targets
3609           !n->is_MergeMem() &&  // Not very interesting
3610           !n->is_top() &&       // Debug info table constants
3611           !(n->is_Con() && !n->is_Mach())// Debug info table constants
3612           ) {
3613         if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3614           pc = pcs[n->_idx];
3615           st->print("%*.*x", pc_digits, pc_digits, pc);
3616         } else {
3617           st->fill_to(pc_digits);
3618         }
3619         st->print(" %c ", starts_bundle);
3620         starts_bundle = ' ';
3621         st->fill_to(prefix_len);
3622         n->format(C->regalloc(), st);
3623         st->cr();
3624       }
3625 
3626       // If we have an instruction with a delay slot, and have seen a delay,
3627       // then back up and print it
3628       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3629         // Coverity finding - Explicit null dereferenced.
3630         guarantee(delay != nullptr, "no unconditional delay instruction");
3631         if (WizardMode) delay->dump();
3632 
3633         if (node_bundling(delay)->starts_bundle())
3634           starts_bundle = '+';
3635         if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3636           pc = pcs[n->_idx];
3637           st->print("%*.*x", pc_digits, pc_digits, pc);
3638         } else {
3639           st->fill_to(pc_digits);
3640         }
3641         st->print(" %c ", starts_bundle);
3642         starts_bundle = ' ';
3643         st->fill_to(prefix_len);
3644         delay->format(C->regalloc(), st);
3645         st->cr();
3646         delay = nullptr;
3647       }
3648 
3649       // Dump the exception table as well
3650       if( n->is_Catch() && (Verbose || WizardMode) ) {
3651         // Print the exception table for this offset
3652         _handler_table.print_subtable_for(pc);
3653       }
3654       st->bol(); // Make sure we start on a new line
3655     }
3656     st->cr(); // one empty line between blocks
3657     assert(cut_short || delay == nullptr, "no unconditional delay branch");
3658   } // End of per-block dump
3659 
3660   if (cut_short)  st->print_cr("*** disassembly is cut short ***");
3661 }
3662 #endif
3663 
3664 #ifndef PRODUCT
3665 void PhaseOutput::print_statistics() {
3666   Scheduling::print_statistics();
3667 }
3668 #endif
3669 
3670 int PhaseOutput::max_inst_size() {
3671   if (AOTCodeCache::maybe_dumping_code()) {
3672     // See the comment in output.hpp.
3673     return 16384;
3674   } else {
3675     return mainline_MAX_inst_size;
3676   }
3677 }
3678 
3679 int PhaseOutput::max_inst_gcstub_size() {
3680   assert(mainline_MAX_inst_size <= max_inst_size(), "Sanity");
3681   return mainline_MAX_inst_size;
3682 }