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