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