1 /* 2 * Copyright (c) 1998, 2024, 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 "precompiled.hpp" 26 #include "asm/assembler.inline.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "code/compiledIC.hpp" 29 #include "code/debugInfo.hpp" 30 #include "code/debugInfoRec.hpp" 31 #include "compiler/compileBroker.hpp" 32 #include "compiler/compilerDirectives.hpp" 33 #include "compiler/disassembler.hpp" 34 #include "compiler/oopMap.hpp" 35 #include "gc/shared/barrierSet.hpp" 36 #include "gc/shared/gc_globals.hpp" 37 #include "gc/shared/c2/barrierSetC2.hpp" 38 #include "memory/allocation.inline.hpp" 39 #include "memory/allocation.hpp" 40 #include "opto/ad.hpp" 41 #include "opto/block.hpp" 42 #include "opto/c2compiler.hpp" 43 #include "opto/c2_MacroAssembler.hpp" 44 #include "opto/callnode.hpp" 45 #include "opto/cfgnode.hpp" 46 #include "opto/locknode.hpp" 47 #include "opto/machnode.hpp" 48 #include "opto/node.hpp" 49 #include "opto/optoreg.hpp" 50 #include "opto/output.hpp" 51 #include "opto/regalloc.hpp" 52 #include "opto/runtime.hpp" 53 #include "opto/subnode.hpp" 54 #include "opto/type.hpp" 55 #include "runtime/handles.inline.hpp" 56 #include "runtime/sharedRuntime.hpp" 57 #include "utilities/macros.hpp" 58 #include "utilities/powerOfTwo.hpp" 59 #include "utilities/xmlstream.hpp" 60 61 #ifndef PRODUCT 62 #define DEBUG_ARG(x) , x 63 #else 64 #define DEBUG_ARG(x) 65 #endif 66 67 //------------------------------Scheduling---------------------------------- 68 // This class contains all the information necessary to implement instruction 69 // scheduling and bundling. 70 class Scheduling { 71 72 private: 73 // Arena to use 74 Arena *_arena; 75 76 // Control-Flow Graph info 77 PhaseCFG *_cfg; 78 79 // Register Allocation info 80 PhaseRegAlloc *_regalloc; 81 82 // Number of nodes in the method 83 uint _node_bundling_limit; 84 85 // List of scheduled nodes. Generated in reverse order 86 Node_List _scheduled; 87 88 // List of nodes currently available for choosing for scheduling 89 Node_List _available; 90 91 // For each instruction beginning a bundle, the number of following 92 // nodes to be bundled with it. 93 Bundle *_node_bundling_base; 94 95 // Mapping from register to Node 96 Node_List _reg_node; 97 98 // Free list for pinch nodes. 99 Node_List _pinch_free_list; 100 101 // Number of uses of this node within the containing basic block. 102 short *_uses; 103 104 // Schedulable portion of current block. Skips Region/Phi/CreateEx up 105 // front, branch+proj at end. Also skips Catch/CProj (same as 106 // branch-at-end), plus just-prior exception-throwing call. 107 uint _bb_start, _bb_end; 108 109 // Latency from the end of the basic block as scheduled 110 unsigned short *_current_latency; 111 112 // Remember the next node 113 Node *_next_node; 114 115 // Use this for an unconditional branch delay slot 116 Node *_unconditional_delay_slot; 117 118 // Pointer to a Nop 119 MachNopNode *_nop; 120 121 // Length of the current bundle, in instructions 122 uint _bundle_instr_count; 123 124 // Current Cycle number, for computing latencies and bundling 125 uint _bundle_cycle_number; 126 127 // Bundle information 128 Pipeline_Use_Element _bundle_use_elements[resource_count]; 129 Pipeline_Use _bundle_use; 130 131 // Dump the available list 132 void dump_available() const; 133 134 public: 135 Scheduling(Arena *arena, Compile &compile); 136 137 // Destructor 138 NOT_PRODUCT( ~Scheduling(); ) 139 140 // Step ahead "i" cycles 141 void step(uint i); 142 143 // Step ahead 1 cycle, and clear the bundle state (for example, 144 // at a branch target) 145 void step_and_clear(); 146 147 Bundle* node_bundling(const Node *n) { 148 assert(valid_bundle_info(n), "oob"); 149 return (&_node_bundling_base[n->_idx]); 150 } 151 152 bool valid_bundle_info(const Node *n) const { 153 return (_node_bundling_limit > n->_idx); 154 } 155 156 bool starts_bundle(const Node *n) const { 157 return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle()); 158 } 159 160 // Do the scheduling 161 void DoScheduling(); 162 163 // Compute the register antidependencies within a basic block 164 void ComputeRegisterAntidependencies(Block *bb); 165 void verify_do_def( Node *n, OptoReg::Name def, const char *msg ); 166 void verify_good_schedule( Block *b, const char *msg ); 167 void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ); 168 void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ); 169 170 // Add a node to the current bundle 171 void AddNodeToBundle(Node *n, const Block *bb); 172 173 // Return an integer less than, equal to, or greater than zero 174 // if the stack offset of the first argument is respectively 175 // less than, equal to, or greater than the second. 176 int compare_two_spill_nodes(Node* first, Node* second); 177 178 // Add a node to the list of available nodes 179 void AddNodeToAvailableList(Node *n); 180 181 // Compute the local use count for the nodes in a block, and compute 182 // the list of instructions with no uses in the block as available 183 void ComputeUseCount(const Block *bb); 184 185 // Choose an instruction from the available list to add to the bundle 186 Node * ChooseNodeToBundle(); 187 188 // See if this Node fits into the currently accumulating bundle 189 bool NodeFitsInBundle(Node *n); 190 191 // Decrement the use count for a node 192 void DecrementUseCounts(Node *n, const Block *bb); 193 194 // Garbage collect pinch nodes for reuse by other blocks. 195 void garbage_collect_pinch_nodes(); 196 // Clean up a pinch node for reuse (helper for above). 197 void cleanup_pinch( Node *pinch ); 198 199 // Information for statistics gathering 200 #ifndef PRODUCT 201 private: 202 // Gather information on size of nops relative to total 203 uint _branches, _unconditional_delays; 204 205 static uint _total_nop_size, _total_method_size; 206 static uint _total_branches, _total_unconditional_delays; 207 static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 208 209 public: 210 static void print_statistics(); 211 212 static void increment_instructions_per_bundle(uint i) { 213 _total_instructions_per_bundle[i]++; 214 } 215 216 static void increment_nop_size(uint s) { 217 _total_nop_size += s; 218 } 219 220 static void increment_method_size(uint s) { 221 _total_method_size += s; 222 } 223 #endif 224 225 }; 226 227 PhaseOutput::PhaseOutput() 228 : Phase(Phase::Output), 229 _code_buffer("Compile::Fill_buffer"), 230 _first_block_size(0), 231 _handler_table(), 232 _inc_table(), 233 _stub_list(), 234 _oop_map_set(nullptr), 235 _scratch_buffer_blob(nullptr), 236 _scratch_locs_memory(nullptr), 237 _scratch_const_size(-1), 238 _in_scratch_emit_size(false), 239 _frame_slots(0), 240 _code_offsets(), 241 _node_bundling_limit(0), 242 _node_bundling_base(nullptr), 243 _orig_pc_slot(0), 244 _orig_pc_slot_offset_in_bytes(0), 245 _buf_sizes(), 246 _block(nullptr), 247 _index(0) { 248 C->set_output(this); 249 if (C->stub_name() == nullptr) { 250 int fixed_slots = C->fixed_slots(); 251 if (C->needs_stack_repair()) { 252 fixed_slots -= 2; 253 } 254 // TODO 8284443 Only reserve extra slot if needed 255 if (InlineTypeReturnedAsFields) { 256 fixed_slots -= 2; 257 } 258 _orig_pc_slot = fixed_slots - (sizeof(address) / VMRegImpl::stack_slot_size); 259 } 260 } 261 262 PhaseOutput::~PhaseOutput() { 263 C->set_output(nullptr); 264 if (_scratch_buffer_blob != nullptr) { 265 BufferBlob::free(_scratch_buffer_blob); 266 } 267 } 268 269 void PhaseOutput::perform_mach_node_analysis() { 270 // Late barrier analysis must be done after schedule and bundle 271 // Otherwise liveness based spilling will fail 272 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 273 bs->late_barrier_analysis(); 274 275 pd_perform_mach_node_analysis(); 276 277 C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3); 278 } 279 280 // Convert Nodes to instruction bits and pass off to the VM 281 void PhaseOutput::Output() { 282 // RootNode goes 283 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" ); 284 285 // The number of new nodes (mostly MachNop) is proportional to 286 // the number of java calls and inner loops which are aligned. 287 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 288 C->inner_loops()*(OptoLoopAlignment-1)), 289 "out of nodes before code generation" ) ) { 290 return; 291 } 292 // Make sure I can find the Start Node 293 Block *entry = C->cfg()->get_block(1); 294 Block *broot = C->cfg()->get_root_block(); 295 296 const StartNode *start = entry->head()->as_Start(); 297 298 // Replace StartNode with prolog 299 Label verified_entry; 300 MachPrologNode* prolog = new MachPrologNode(&verified_entry); 301 entry->map_node(prolog, 0); 302 C->cfg()->map_node_to_block(prolog, entry); 303 C->cfg()->unmap_node_from_block(start); // start is no longer in any block 304 305 // Virtual methods need an unverified entry point 306 if (C->is_osr_compilation()) { 307 if (PoisonOSREntry) { 308 // TODO: Should use a ShouldNotReachHereNode... 309 C->cfg()->insert( broot, 0, new MachBreakpointNode() ); 310 } 311 } else { 312 if (C->method()) { 313 if (C->method()->has_scalarized_args()) { 314 // Add entry point to unpack all inline type arguments 315 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ false)); 316 if (!C->method()->is_static()) { 317 // Add verified/unverified entry points to only unpack inline type receiver at interface calls 318 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ false)); 319 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ true)); 320 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ true)); 321 } 322 } else if (!C->method()->is_static()) { 323 // Insert unvalidated entry point 324 C->cfg()->insert(broot, 0, new MachUEPNode()); 325 } 326 } 327 } 328 329 // Break before main entry point 330 if ((C->method() && C->directive()->BreakAtExecuteOption) || 331 (OptoBreakpoint && C->is_method_compilation()) || 332 (OptoBreakpointOSR && C->is_osr_compilation()) || 333 (OptoBreakpointC2R && !C->method()) ) { 334 // checking for C->method() means that OptoBreakpoint does not apply to 335 // runtime stubs or frame converters 336 C->cfg()->insert( entry, 1, new MachBreakpointNode() ); 337 } 338 339 // Insert epilogs before every return 340 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) { 341 Block* block = C->cfg()->get_block(i); 342 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point? 343 Node* m = block->end(); 344 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) { 345 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 346 block->add_inst(epilog); 347 C->cfg()->map_node_to_block(epilog, block); 348 } 349 } 350 } 351 352 // Keeper of sizing aspects 353 _buf_sizes = BufferSizingData(); 354 355 // Initialize code buffer 356 estimate_buffer_size(_buf_sizes._const); 357 if (C->failing()) return; 358 359 // Pre-compute the length of blocks and replace 360 // long branches with short if machine supports it. 361 // Must be done before ScheduleAndBundle due to SPARC delay slots 362 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1); 363 blk_starts[0] = 0; 364 shorten_branches(blk_starts); 365 366 if (!C->is_osr_compilation() && C->has_scalarized_args()) { 367 // Compute the offsets of the entry points required by the inline type calling convention 368 if (!C->method()->is_static()) { 369 // We have entries at the beginning of the method, implemented by the first 4 nodes. 370 // Entry (unverified) @ offset 0 371 // Verified_Inline_Entry_RO 372 // Inline_Entry (unverified) 373 // Verified_Inline_Entry 374 uint offset = 0; 375 _code_offsets.set_value(CodeOffsets::Entry, offset); 376 377 offset += ((MachVEPNode*)broot->get_node(0))->size(C->regalloc()); 378 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, offset); 379 380 offset += ((MachVEPNode*)broot->get_node(1))->size(C->regalloc()); 381 _code_offsets.set_value(CodeOffsets::Inline_Entry, offset); 382 383 offset += ((MachVEPNode*)broot->get_node(2))->size(C->regalloc()); 384 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, offset); 385 } else { 386 _code_offsets.set_value(CodeOffsets::Entry, -1); // will be patched later 387 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, 0); 388 } 389 } 390 391 ScheduleAndBundle(); 392 if (C->failing()) { 393 return; 394 } 395 396 perform_mach_node_analysis(); 397 398 // Complete sizing of codebuffer 399 CodeBuffer* cb = init_buffer(); 400 if (cb == nullptr || C->failing()) { 401 return; 402 } 403 404 BuildOopMaps(); 405 406 if (C->failing()) { 407 return; 408 } 409 410 fill_buffer(cb, blk_starts); 411 } 412 413 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const { 414 // Determine if we need to generate a stack overflow check. 415 // Do it if the method is not a stub function and 416 // has java calls or has frame size > vm_page_size/8. 417 // The debug VM checks that deoptimization doesn't trigger an 418 // unexpected stack overflow (compiled method stack banging should 419 // guarantee it doesn't happen) so we always need the stack bang in 420 // a debug VM. 421 return (C->stub_function() == nullptr && 422 (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3 423 DEBUG_ONLY(|| true))); 424 } 425 426 bool PhaseOutput::need_register_stack_bang() const { 427 // Determine if we need to generate a register stack overflow check. 428 // This is only used on architectures which have split register 429 // and memory stacks (ie. IA64). 430 // Bang if the method is not a stub function and has java calls 431 return (C->stub_function() == nullptr && C->has_java_calls()); 432 } 433 434 435 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 436 // of a loop. When aligning a loop we need to provide enough instructions 437 // in cpu's fetch buffer to feed decoders. The loop alignment could be 438 // avoided if we have enough instructions in fetch buffer at the head of a loop. 439 // By default, the size is set to 999999 by Block's constructor so that 440 // a loop will be aligned if the size is not reset here. 441 // 442 // Note: Mach instructions could contain several HW instructions 443 // so the size is estimated only. 444 // 445 void PhaseOutput::compute_loop_first_inst_sizes() { 446 // The next condition is used to gate the loop alignment optimization. 447 // Don't aligned a loop if there are enough instructions at the head of a loop 448 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 449 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 450 // equal to 11 bytes which is the largest address NOP instruction. 451 if (MaxLoopPad < OptoLoopAlignment - 1) { 452 uint last_block = C->cfg()->number_of_blocks() - 1; 453 for (uint i = 1; i <= last_block; i++) { 454 Block* block = C->cfg()->get_block(i); 455 // Check the first loop's block which requires an alignment. 456 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) { 457 uint sum_size = 0; 458 uint inst_cnt = NumberOfLoopInstrToAlign; 459 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc()); 460 461 // Check subsequent fallthrough blocks if the loop's first 462 // block(s) does not have enough instructions. 463 Block *nb = block; 464 while(inst_cnt > 0 && 465 i < last_block && 466 !C->cfg()->get_block(i + 1)->has_loop_alignment() && 467 !nb->has_successor(block)) { 468 i++; 469 nb = C->cfg()->get_block(i); 470 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc()); 471 } // while( inst_cnt > 0 && i < last_block ) 472 473 block->set_first_inst_size(sum_size); 474 } // f( b->head()->is_Loop() ) 475 } // for( i <= last_block ) 476 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 477 } 478 479 // The architecture description provides short branch variants for some long 480 // branch instructions. Replace eligible long branches with short branches. 481 void PhaseOutput::shorten_branches(uint* blk_starts) { 482 483 Compile::TracePhase tp("shorten branches", &timers[_t_shortenBranches]); 484 485 // Compute size of each block, method size, and relocation information size 486 uint nblocks = C->cfg()->number_of_blocks(); 487 488 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 489 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 490 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks); 491 492 // Collect worst case block paddings 493 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks); 494 memset(block_worst_case_pad, 0, nblocks * sizeof(int)); 495 496 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 497 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 498 499 bool has_short_branch_candidate = false; 500 501 // Initialize the sizes to 0 502 int code_size = 0; // Size in bytes of generated code 503 int stub_size = 0; // Size in bytes of all stub entries 504 // Size in bytes of all relocation entries, including those in local stubs. 505 // Start with 2-bytes of reloc info for the unvalidated entry point 506 int reloc_size = 1; // Number of relocation entries 507 508 // Make three passes. The first computes pessimistic blk_starts, 509 // relative jmp_offset and reloc_size information. The second performs 510 // short branch substitution using the pessimistic sizing. The 511 // third inserts nops where needed. 512 513 // Step one, perform a pessimistic sizing pass. 514 uint last_call_adr = max_juint; 515 uint last_avoid_back_to_back_adr = max_juint; 516 uint nop_size = (new MachNopNode())->size(C->regalloc()); 517 for (uint i = 0; i < nblocks; i++) { // For all blocks 518 Block* block = C->cfg()->get_block(i); 519 _block = block; 520 521 // During short branch replacement, we store the relative (to blk_starts) 522 // offset of jump in jmp_offset, rather than the absolute offset of jump. 523 // This is so that we do not need to recompute sizes of all nodes when 524 // we compute correct blk_starts in our next sizing pass. 525 jmp_offset[i] = 0; 526 jmp_size[i] = 0; 527 jmp_nidx[i] = -1; 528 DEBUG_ONLY( jmp_target[i] = 0; ) 529 DEBUG_ONLY( jmp_rule[i] = 0; ) 530 531 // Sum all instruction sizes to compute block size 532 uint last_inst = block->number_of_nodes(); 533 uint blk_size = 0; 534 for (uint j = 0; j < last_inst; j++) { 535 _index = j; 536 Node* nj = block->get_node(_index); 537 // Handle machine instruction nodes 538 if (nj->is_Mach()) { 539 MachNode* mach = nj->as_Mach(); 540 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 541 reloc_size += mach->reloc(); 542 if (mach->is_MachCall()) { 543 // add size information for trampoline stub 544 // class CallStubImpl is platform-specific and defined in the *.ad files. 545 stub_size += CallStubImpl::size_call_trampoline(); 546 reloc_size += CallStubImpl::reloc_call_trampoline(); 547 548 MachCallNode *mcall = mach->as_MachCall(); 549 // This destination address is NOT PC-relative 550 551 if (mcall->entry_point() != nullptr) { 552 mcall->method_set((intptr_t)mcall->entry_point()); 553 } 554 555 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) { 556 stub_size += CompiledDirectCall::to_interp_stub_size(); 557 reloc_size += CompiledDirectCall::reloc_to_interp_stub(); 558 } 559 } else if (mach->is_MachSafePoint()) { 560 // If call/safepoint are adjacent, account for possible 561 // nop to disambiguate the two safepoints. 562 // ScheduleAndBundle() can rearrange nodes in a block, 563 // check for all offsets inside this block. 564 if (last_call_adr >= blk_starts[i]) { 565 blk_size += nop_size; 566 } 567 } 568 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 569 // Nop is inserted between "avoid back to back" instructions. 570 // ScheduleAndBundle() can rearrange nodes in a block, 571 // check for all offsets inside this block. 572 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 573 blk_size += nop_size; 574 } 575 } 576 if (mach->may_be_short_branch()) { 577 if (!nj->is_MachBranch()) { 578 #ifndef PRODUCT 579 nj->dump(3); 580 #endif 581 Unimplemented(); 582 } 583 assert(jmp_nidx[i] == -1, "block should have only one branch"); 584 jmp_offset[i] = blk_size; 585 jmp_size[i] = nj->size(C->regalloc()); 586 jmp_nidx[i] = j; 587 has_short_branch_candidate = true; 588 } 589 } 590 blk_size += nj->size(C->regalloc()); 591 // Remember end of call offset 592 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 593 last_call_adr = blk_starts[i]+blk_size; 594 } 595 // Remember end of avoid_back_to_back offset 596 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 597 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 598 } 599 } 600 601 // When the next block starts a loop, we may insert pad NOP 602 // instructions. Since we cannot know our future alignment, 603 // assume the worst. 604 if (i < nblocks - 1) { 605 Block* nb = C->cfg()->get_block(i + 1); 606 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 607 if (max_loop_pad > 0) { 608 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 609 // Adjust last_call_adr and/or last_avoid_back_to_back_adr. 610 // If either is the last instruction in this block, bump by 611 // max_loop_pad in lock-step with blk_size, so sizing 612 // calculations in subsequent blocks still can conservatively 613 // detect that it may the last instruction in this block. 614 if (last_call_adr == blk_starts[i]+blk_size) { 615 last_call_adr += max_loop_pad; 616 } 617 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) { 618 last_avoid_back_to_back_adr += max_loop_pad; 619 } 620 blk_size += max_loop_pad; 621 block_worst_case_pad[i + 1] = max_loop_pad; 622 } 623 } 624 625 // Save block size; update total method size 626 blk_starts[i+1] = blk_starts[i]+blk_size; 627 } 628 629 // Step two, replace eligible long jumps. 630 bool progress = true; 631 uint last_may_be_short_branch_adr = max_juint; 632 while (has_short_branch_candidate && progress) { 633 progress = false; 634 has_short_branch_candidate = false; 635 int adjust_block_start = 0; 636 for (uint i = 0; i < nblocks; i++) { 637 Block* block = C->cfg()->get_block(i); 638 int idx = jmp_nidx[i]; 639 MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach(); 640 if (mach != nullptr && mach->may_be_short_branch()) { 641 #ifdef ASSERT 642 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 643 int j; 644 // Find the branch; ignore trailing NOPs. 645 for (j = block->number_of_nodes()-1; j>=0; j--) { 646 Node* n = block->get_node(j); 647 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 648 break; 649 } 650 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity"); 651 #endif 652 int br_size = jmp_size[i]; 653 int br_offs = blk_starts[i] + jmp_offset[i]; 654 655 // This requires the TRUE branch target be in succs[0] 656 uint bnum = block->non_connector_successor(0)->_pre_order; 657 int offset = blk_starts[bnum] - br_offs; 658 if (bnum > i) { // adjust following block's offset 659 offset -= adjust_block_start; 660 } 661 662 // This block can be a loop header, account for the padding 663 // in the previous block. 664 int block_padding = block_worst_case_pad[i]; 665 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top"); 666 // In the following code a nop could be inserted before 667 // the branch which will increase the backward distance. 668 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr); 669 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block"); 670 671 if (needs_padding && offset <= 0) 672 offset -= nop_size; 673 674 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) { 675 // We've got a winner. Replace this branch. 676 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 677 678 // Update the jmp_size. 679 int new_size = replacement->size(C->regalloc()); 680 int diff = br_size - new_size; 681 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 682 // Conservatively take into account padding between 683 // avoid_back_to_back branches. Previous branch could be 684 // converted into avoid_back_to_back branch during next 685 // rounds. 686 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 687 jmp_offset[i] += nop_size; 688 diff -= nop_size; 689 } 690 adjust_block_start += diff; 691 block->map_node(replacement, idx); 692 mach->subsume_by(replacement, C); 693 mach = replacement; 694 progress = true; 695 696 jmp_size[i] = new_size; 697 DEBUG_ONLY( jmp_target[i] = bnum; ); 698 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 699 } else { 700 // The jump distance is not short, try again during next iteration. 701 has_short_branch_candidate = true; 702 } 703 } // (mach->may_be_short_branch()) 704 if (mach != nullptr && (mach->may_be_short_branch() || 705 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) { 706 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 707 } 708 blk_starts[i+1] -= adjust_block_start; 709 } 710 } 711 712 #ifdef ASSERT 713 for (uint i = 0; i < nblocks; i++) { // For all blocks 714 if (jmp_target[i] != 0) { 715 int br_size = jmp_size[i]; 716 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 717 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 718 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]); 719 } 720 assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 721 } 722 } 723 #endif 724 725 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 726 // after ScheduleAndBundle(). 727 728 // ------------------ 729 // Compute size for code buffer 730 code_size = blk_starts[nblocks]; 731 732 // Relocation records 733 reloc_size += 1; // Relo entry for exception handler 734 735 // Adjust reloc_size to number of record of relocation info 736 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 737 // a relocation index. 738 // The CodeBuffer will expand the locs array if this estimate is too low. 739 reloc_size *= 10 / sizeof(relocInfo); 740 741 _buf_sizes._reloc = reloc_size; 742 _buf_sizes._code = code_size; 743 _buf_sizes._stub = stub_size; 744 } 745 746 //------------------------------FillLocArray----------------------------------- 747 // Create a bit of debug info and append it to the array. The mapping is from 748 // Java local or expression stack to constant, register or stack-slot. For 749 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 750 // entry has been taken care of and caller should skip it). 751 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 752 // This should never have accepted Bad before 753 assert(OptoReg::is_valid(regnum), "location must be valid"); 754 return (OptoReg::is_reg(regnum)) 755 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 756 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 757 } 758 759 760 ObjectValue* 761 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 762 for (int i = 0; i < objs->length(); i++) { 763 assert(objs->at(i)->is_object(), "corrupt object cache"); 764 ObjectValue* sv = (ObjectValue*) objs->at(i); 765 if (sv->id() == id) { 766 return sv; 767 } 768 } 769 // Otherwise.. 770 return nullptr; 771 } 772 773 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 774 ObjectValue* sv ) { 775 assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition"); 776 objs->append(sv); 777 } 778 779 780 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 781 GrowableArray<ScopeValue*> *array, 782 GrowableArray<ScopeValue*> *objs ) { 783 assert( local, "use _top instead of null" ); 784 if (array->length() != idx) { 785 assert(array->length() == idx + 1, "Unexpected array count"); 786 // Old functionality: 787 // return 788 // New functionality: 789 // Assert if the local is not top. In product mode let the new node 790 // override the old entry. 791 assert(local == C->top(), "LocArray collision"); 792 if (local == C->top()) { 793 return; 794 } 795 array->pop(); 796 } 797 const Type *t = local->bottom_type(); 798 799 // Is it a safepoint scalar object node? 800 if (local->is_SafePointScalarObject()) { 801 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 802 803 ObjectValue* sv = (ObjectValue*) sv_for_node_id(objs, spobj->_idx); 804 if (sv == nullptr) { 805 ciKlass* cik = t->is_oopptr()->exact_klass(); 806 assert(cik->is_instance_klass() || 807 cik->is_array_klass(), "Not supported allocation."); 808 uint first_ind = spobj->first_index(sfpt->jvms()); 809 // Nullable, scalarized inline types have an is_init input 810 // that needs to be checked before using the field values. 811 ScopeValue* is_init = nullptr; 812 if (cik->is_inlinetype()) { 813 Node* init_node = sfpt->in(first_ind++); 814 assert(init_node != nullptr, "is_init node not found"); 815 if (!init_node->is_top()) { 816 const TypeInt* init_type = init_node->bottom_type()->is_int(); 817 if (init_node->is_Con()) { 818 is_init = new ConstantIntValue(init_type->get_con()); 819 } else { 820 OptoReg::Name init_reg = C->regalloc()->get_reg_first(init_node); 821 is_init = new_loc_value(C->regalloc(), init_reg, Location::normal); 822 } 823 } 824 } 825 sv = new ObjectValue(spobj->_idx, 826 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()), is_init); 827 set_sv_for_object_node(objs, sv); 828 829 for (uint i = 0; i < spobj->n_fields(); i++) { 830 Node* fld_node = sfpt->in(first_ind+i); 831 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 832 } 833 } 834 array->append(sv); 835 return; 836 } else if (local->is_SafePointScalarMerge()) { 837 SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge(); 838 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx); 839 840 if (mv == nullptr) { 841 GrowableArray<ScopeValue*> deps; 842 843 int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms()); 844 (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs); 845 assert(deps.length() == 1, "missing value"); 846 847 int selector_idx = smerge->selector_idx(sfpt->jvms()); 848 (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr); 849 assert(deps.length() == 2, "missing value"); 850 851 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1)); 852 set_sv_for_object_node(objs, mv); 853 854 for (uint i = 1; i < smerge->req(); i++) { 855 Node* obj_node = smerge->in(i); 856 (void)FillLocArray(mv->possible_objects()->length(), sfpt, obj_node, mv->possible_objects(), objs); 857 } 858 } 859 array->append(mv); 860 return; 861 } 862 863 // Grab the register number for the local 864 OptoReg::Name regnum = C->regalloc()->get_reg_first(local); 865 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 866 // Record the double as two float registers. 867 // The register mask for such a value always specifies two adjacent 868 // float registers, with the lower register number even. 869 // Normally, the allocation of high and low words to these registers 870 // is irrelevant, because nearly all operations on register pairs 871 // (e.g., StoreD) treat them as a single unit. 872 // Here, we assume in addition that the words in these two registers 873 // stored "naturally" (by operations like StoreD and double stores 874 // within the interpreter) such that the lower-numbered register 875 // is written to the lower memory address. This may seem like 876 // a machine dependency, but it is not--it is a requirement on 877 // the author of the <arch>.ad file to ensure that, for every 878 // even/odd double-register pair to which a double may be allocated, 879 // the word in the even single-register is stored to the first 880 // memory word. (Note that register numbers are completely 881 // arbitrary, and are not tied to any machine-level encodings.) 882 #ifdef _LP64 883 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 884 array->append(new ConstantIntValue((jint)0)); 885 array->append(new_loc_value( C->regalloc(), regnum, Location::dbl )); 886 } else if ( t->base() == Type::Long ) { 887 array->append(new ConstantIntValue((jint)0)); 888 array->append(new_loc_value( C->regalloc(), regnum, Location::lng )); 889 } else if ( t->base() == Type::RawPtr ) { 890 // jsr/ret return address which must be restored into the full 891 // width 64-bit stack slot. 892 array->append(new_loc_value( C->regalloc(), regnum, Location::lng )); 893 } 894 #else //_LP64 895 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 896 // Repack the double/long as two jints. 897 // The convention the interpreter uses is that the second local 898 // holds the first raw word of the native double representation. 899 // This is actually reasonable, since locals and stack arrays 900 // grow downwards in all implementations. 901 // (If, on some machine, the interpreter's Java locals or stack 902 // were to grow upwards, the embedded doubles would be word-swapped.) 903 array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal )); 904 array->append(new_loc_value( C->regalloc(), regnum , Location::normal )); 905 } 906 #endif //_LP64 907 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 908 OptoReg::is_reg(regnum) ) { 909 array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double() 910 ? Location::float_in_dbl : Location::normal )); 911 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 912 array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long 913 ? Location::int_in_long : Location::normal )); 914 } else if( t->base() == Type::NarrowOop ) { 915 array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop )); 916 } else if (t->base() == Type::VectorA || t->base() == Type::VectorS || 917 t->base() == Type::VectorD || t->base() == Type::VectorX || 918 t->base() == Type::VectorY || t->base() == Type::VectorZ) { 919 array->append(new_loc_value( C->regalloc(), regnum, Location::vector )); 920 } else if (C->regalloc()->is_oop(local)) { 921 assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr || 922 t->base() == Type::AryPtr, 923 "Unexpected type: %s", t->msg()); 924 array->append(new_loc_value( C->regalloc(), regnum, Location::oop )); 925 } else { 926 assert(t->base() == Type::Int || t->base() == Type::Half || 927 t->base() == Type::FloatCon || t->base() == Type::FloatBot, 928 "Unexpected type: %s", t->msg()); 929 array->append(new_loc_value( C->regalloc(), regnum, Location::normal )); 930 } 931 return; 932 } 933 934 // No register. It must be constant data. 935 switch (t->base()) { 936 case Type::Half: // Second half of a double 937 ShouldNotReachHere(); // Caller should skip 2nd halves 938 break; 939 case Type::AnyPtr: 940 array->append(new ConstantOopWriteValue(nullptr)); 941 break; 942 case Type::AryPtr: 943 case Type::InstPtr: // fall through 944 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 945 break; 946 case Type::NarrowOop: 947 if (t == TypeNarrowOop::NULL_PTR) { 948 array->append(new ConstantOopWriteValue(nullptr)); 949 } else { 950 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 951 } 952 break; 953 case Type::Int: 954 array->append(new ConstantIntValue(t->is_int()->get_con())); 955 break; 956 case Type::RawPtr: 957 // A return address (T_ADDRESS). 958 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 959 #ifdef _LP64 960 // Must be restored to the full-width 64-bit stack slot. 961 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 962 #else 963 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 964 #endif 965 break; 966 case Type::FloatCon: { 967 float f = t->is_float_constant()->getf(); 968 array->append(new ConstantIntValue(jint_cast(f))); 969 break; 970 } 971 case Type::DoubleCon: { 972 jdouble d = t->is_double_constant()->getd(); 973 #ifdef _LP64 974 array->append(new ConstantIntValue((jint)0)); 975 array->append(new ConstantDoubleValue(d)); 976 #else 977 // Repack the double as two jints. 978 // The convention the interpreter uses is that the second local 979 // holds the first raw word of the native double representation. 980 // This is actually reasonable, since locals and stack arrays 981 // grow downwards in all implementations. 982 // (If, on some machine, the interpreter's Java locals or stack 983 // were to grow upwards, the embedded doubles would be word-swapped.) 984 jlong_accessor acc; 985 acc.long_value = jlong_cast(d); 986 array->append(new ConstantIntValue(acc.words[1])); 987 array->append(new ConstantIntValue(acc.words[0])); 988 #endif 989 break; 990 } 991 case Type::Long: { 992 jlong d = t->is_long()->get_con(); 993 #ifdef _LP64 994 array->append(new ConstantIntValue((jint)0)); 995 array->append(new ConstantLongValue(d)); 996 #else 997 // Repack the long as two jints. 998 // The convention the interpreter uses is that the second local 999 // holds the first raw word of the native double representation. 1000 // This is actually reasonable, since locals and stack arrays 1001 // grow downwards in all implementations. 1002 // (If, on some machine, the interpreter's Java locals or stack 1003 // were to grow upwards, the embedded doubles would be word-swapped.) 1004 jlong_accessor acc; 1005 acc.long_value = d; 1006 array->append(new ConstantIntValue(acc.words[1])); 1007 array->append(new ConstantIntValue(acc.words[0])); 1008 #endif 1009 break; 1010 } 1011 case Type::Top: // Add an illegal value here 1012 array->append(new LocationValue(Location())); 1013 break; 1014 default: 1015 ShouldNotReachHere(); 1016 break; 1017 } 1018 } 1019 1020 // Determine if this node starts a bundle 1021 bool PhaseOutput::starts_bundle(const Node *n) const { 1022 return (_node_bundling_limit > n->_idx && 1023 _node_bundling_base[n->_idx].starts_bundle()); 1024 } 1025 1026 // Determine if there is a monitor that has 'ov' as its owner. 1027 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const { 1028 for (int k = 0; k < monarray->length(); k++) { 1029 MonitorValue* mv = monarray->at(k); 1030 if (mv->owner() == ov) { 1031 return true; 1032 } 1033 } 1034 1035 return false; 1036 } 1037 1038 //--------------------------Process_OopMap_Node-------------------------------- 1039 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) { 1040 // Handle special safepoint nodes for synchronization 1041 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1042 MachCallNode *mcall; 1043 1044 int safepoint_pc_offset = current_offset; 1045 bool is_method_handle_invoke = false; 1046 bool return_oop = false; 1047 bool return_scalarized = false; 1048 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope; 1049 bool arg_escape = false; 1050 1051 // Add the safepoint in the DebugInfoRecorder 1052 if( !mach->is_MachCall() ) { 1053 mcall = nullptr; 1054 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 1055 } else { 1056 mcall = mach->as_MachCall(); 1057 1058 // Is the call a MethodHandle call? 1059 if (mcall->is_MachCallJava()) { 1060 if (mcall->as_MachCallJava()->_method_handle_invoke) { 1061 assert(C->has_method_handle_invokes(), "must have been set during call generation"); 1062 is_method_handle_invoke = true; 1063 } 1064 arg_escape = mcall->as_MachCallJava()->_arg_escape; 1065 } 1066 1067 // Check if a call returns an object. 1068 if (mcall->returns_pointer() || mcall->returns_scalarized()) { 1069 return_oop = true; 1070 } 1071 if (mcall->returns_scalarized()) { 1072 return_scalarized = true; 1073 } 1074 safepoint_pc_offset += mcall->ret_addr_offset(); 1075 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 1076 } 1077 1078 // Loop over the JVMState list to add scope information 1079 // Do not skip safepoints with a null method, they need monitor info 1080 JVMState* youngest_jvms = sfn->jvms(); 1081 int max_depth = youngest_jvms->depth(); 1082 1083 // Allocate the object pool for scalar-replaced objects -- the map from 1084 // small-integer keys (which can be recorded in the local and ostack 1085 // arrays) to descriptions of the object state. 1086 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 1087 1088 // Visit scopes from oldest to youngest. 1089 for (int depth = 1; depth <= max_depth; depth++) { 1090 JVMState* jvms = youngest_jvms->of_depth(depth); 1091 int idx; 1092 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr; 1093 // Safepoints that do not have method() set only provide oop-map and monitor info 1094 // to support GC; these do not support deoptimization. 1095 int num_locs = (method == nullptr) ? 0 : jvms->loc_size(); 1096 int num_exps = (method == nullptr) ? 0 : jvms->stk_size(); 1097 int num_mon = jvms->nof_monitors(); 1098 assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(), 1099 "JVMS local count must match that of the method"); 1100 1101 // Add Local and Expression Stack Information 1102 1103 // Insert locals into the locarray 1104 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 1105 for( idx = 0; idx < num_locs; idx++ ) { 1106 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 1107 } 1108 1109 // Insert expression stack entries into the exparray 1110 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 1111 for( idx = 0; idx < num_exps; idx++ ) { 1112 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 1113 } 1114 1115 // Add in mappings of the monitors 1116 assert( !method || 1117 !method->is_synchronized() || 1118 method->is_native() || 1119 num_mon > 0 || 1120 !GenerateSynchronizationCode, 1121 "monitors must always exist for synchronized methods"); 1122 1123 // Build the growable array of ScopeValues for exp stack 1124 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 1125 1126 // Loop over monitors and insert into array 1127 for (idx = 0; idx < num_mon; idx++) { 1128 // Grab the node that defines this monitor 1129 Node* box_node = sfn->monitor_box(jvms, idx); 1130 Node* obj_node = sfn->monitor_obj(jvms, idx); 1131 1132 // Create ScopeValue for object 1133 ScopeValue *scval = nullptr; 1134 1135 if (obj_node->is_SafePointScalarObject()) { 1136 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 1137 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx); 1138 if (scval == nullptr) { 1139 const Type *t = spobj->bottom_type(); 1140 ciKlass* cik = t->is_oopptr()->exact_klass(); 1141 assert(cik->is_instance_klass() || 1142 cik->is_array_klass(), "Not supported allocation."); 1143 ObjectValue* sv = new ObjectValue(spobj->_idx, 1144 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 1145 PhaseOutput::set_sv_for_object_node(objs, sv); 1146 1147 uint first_ind = spobj->first_index(youngest_jvms); 1148 for (uint i = 0; i < spobj->n_fields(); i++) { 1149 Node* fld_node = sfn->in(first_ind+i); 1150 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 1151 } 1152 scval = sv; 1153 } 1154 } else if (obj_node->is_SafePointScalarMerge()) { 1155 SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge(); 1156 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx); 1157 1158 if (mv == nullptr) { 1159 GrowableArray<ScopeValue*> deps; 1160 1161 int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms); 1162 FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs); 1163 assert(deps.length() == 1, "missing value"); 1164 1165 int selector_idx = smerge->selector_idx(youngest_jvms); 1166 FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr); 1167 assert(deps.length() == 2, "missing value"); 1168 1169 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1)); 1170 set_sv_for_object_node(objs, mv); 1171 1172 for (uint i = 1; i < smerge->req(); i++) { 1173 Node* obj_node = smerge->in(i); 1174 FillLocArray(mv->possible_objects()->length(), sfn, obj_node, mv->possible_objects(), objs); 1175 } 1176 } 1177 scval = mv; 1178 } else if (!obj_node->is_Con()) { 1179 OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node); 1180 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 1181 scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop ); 1182 } else { 1183 scval = new_loc_value( C->regalloc(), obj_reg, Location::oop ); 1184 } 1185 } else { 1186 const TypePtr *tp = obj_node->get_ptr_type(); 1187 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 1188 } 1189 1190 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 1191 Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg)); 1192 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 1193 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 1194 } 1195 1196 // Mark ObjectValue nodes as root nodes if they are directly 1197 // referenced in the JVMS. 1198 for (int i = 0; i < objs->length(); i++) { 1199 ScopeValue* sv = objs->at(i); 1200 if (sv->is_object_merge()) { 1201 ObjectMergeValue* merge = sv->as_ObjectMergeValue(); 1202 1203 for (int j = 0; j< merge->possible_objects()->length(); j++) { 1204 ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue(); 1205 bool is_root = locarray->contains(ov) || exparray->contains(ov) || contains_as_owner(monarray, ov); 1206 ov->set_root(is_root); 1207 } 1208 } 1209 } 1210 1211 // We dump the object pool first, since deoptimization reads it in first. 1212 C->debug_info()->dump_object_pool(objs); 1213 1214 // Build first class objects to pass to scope 1215 DebugToken *locvals = C->debug_info()->create_scope_values(locarray); 1216 DebugToken *expvals = C->debug_info()->create_scope_values(exparray); 1217 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray); 1218 1219 // Make method available for all Safepoints 1220 ciMethod* scope_method = method ? method : C->method(); 1221 // Describe the scope here 1222 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 1223 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 1224 // Now we can describe the scope. 1225 methodHandle null_mh; 1226 bool rethrow_exception = false; 1227 C->debug_info()->describe_scope( 1228 safepoint_pc_offset, 1229 null_mh, 1230 scope_method, 1231 jvms->bci(), 1232 jvms->should_reexecute(), 1233 rethrow_exception, 1234 is_method_handle_invoke, 1235 return_oop, 1236 return_scalarized, 1237 has_ea_local_in_scope, 1238 arg_escape, 1239 locvals, 1240 expvals, 1241 monvals 1242 ); 1243 } // End jvms loop 1244 1245 // Mark the end of the scope set. 1246 C->debug_info()->end_safepoint(safepoint_pc_offset); 1247 } 1248 1249 1250 1251 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 1252 class NonSafepointEmitter { 1253 Compile* C; 1254 JVMState* _pending_jvms; 1255 int _pending_offset; 1256 1257 void emit_non_safepoint(); 1258 1259 public: 1260 NonSafepointEmitter(Compile* compile) { 1261 this->C = compile; 1262 _pending_jvms = nullptr; 1263 _pending_offset = 0; 1264 } 1265 1266 void observe_instruction(Node* n, int pc_offset) { 1267 if (!C->debug_info()->recording_non_safepoints()) return; 1268 1269 Node_Notes* nn = C->node_notes_at(n->_idx); 1270 if (nn == nullptr || nn->jvms() == nullptr) return; 1271 if (_pending_jvms != nullptr && 1272 _pending_jvms->same_calls_as(nn->jvms())) { 1273 // Repeated JVMS? Stretch it up here. 1274 _pending_offset = pc_offset; 1275 } else { 1276 if (_pending_jvms != nullptr && 1277 _pending_offset < pc_offset) { 1278 emit_non_safepoint(); 1279 } 1280 _pending_jvms = nullptr; 1281 if (pc_offset > C->debug_info()->last_pc_offset()) { 1282 // This is the only way _pending_jvms can become non-null: 1283 _pending_jvms = nn->jvms(); 1284 _pending_offset = pc_offset; 1285 } 1286 } 1287 } 1288 1289 // Stay out of the way of real safepoints: 1290 void observe_safepoint(JVMState* jvms, int pc_offset) { 1291 if (_pending_jvms != nullptr && 1292 !_pending_jvms->same_calls_as(jvms) && 1293 _pending_offset < pc_offset) { 1294 emit_non_safepoint(); 1295 } 1296 _pending_jvms = nullptr; 1297 } 1298 1299 void flush_at_end() { 1300 if (_pending_jvms != nullptr) { 1301 emit_non_safepoint(); 1302 } 1303 _pending_jvms = nullptr; 1304 } 1305 }; 1306 1307 void NonSafepointEmitter::emit_non_safepoint() { 1308 JVMState* youngest_jvms = _pending_jvms; 1309 int pc_offset = _pending_offset; 1310 1311 // Clear it now: 1312 _pending_jvms = nullptr; 1313 1314 DebugInformationRecorder* debug_info = C->debug_info(); 1315 assert(debug_info->recording_non_safepoints(), "sanity"); 1316 1317 debug_info->add_non_safepoint(pc_offset); 1318 int max_depth = youngest_jvms->depth(); 1319 1320 // Visit scopes from oldest to youngest. 1321 for (int depth = 1; depth <= max_depth; depth++) { 1322 JVMState* jvms = youngest_jvms->of_depth(depth); 1323 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr; 1324 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 1325 methodHandle null_mh; 1326 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute()); 1327 } 1328 1329 // Mark the end of the scope set. 1330 debug_info->end_non_safepoint(pc_offset); 1331 } 1332 1333 //------------------------------init_buffer------------------------------------ 1334 void PhaseOutput::estimate_buffer_size(int& const_req) { 1335 1336 // Set the initially allocated size 1337 const_req = initial_const_capacity; 1338 1339 // The extra spacing after the code is necessary on some platforms. 1340 // Sometimes we need to patch in a jump after the last instruction, 1341 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 1342 1343 // Compute the byte offset where we can store the deopt pc. 1344 if (C->fixed_slots() != 0) { 1345 _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 1346 } 1347 1348 // Compute prolog code size 1349 _method_size = 0; 1350 _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize; 1351 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 1352 1353 if (C->has_mach_constant_base_node()) { 1354 uint add_size = 0; 1355 // Fill the constant table. 1356 // Note: This must happen before shorten_branches. 1357 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) { 1358 Block* b = C->cfg()->get_block(i); 1359 1360 for (uint j = 0; j < b->number_of_nodes(); j++) { 1361 Node* n = b->get_node(j); 1362 1363 // If the node is a MachConstantNode evaluate the constant 1364 // value section. 1365 if (n->is_MachConstant()) { 1366 MachConstantNode* machcon = n->as_MachConstant(); 1367 machcon->eval_constant(C); 1368 } else if (n->is_Mach()) { 1369 // On Power there are more nodes that issue constants. 1370 add_size += (n->as_Mach()->ins_num_consts() * 8); 1371 } 1372 } 1373 } 1374 1375 // Calculate the offsets of the constants and the size of the 1376 // constant table (including the padding to the next section). 1377 constant_table().calculate_offsets_and_size(); 1378 const_req = constant_table().size() + add_size; 1379 } 1380 1381 // Initialize the space for the BufferBlob used to find and verify 1382 // instruction size in MachNode::emit_size() 1383 init_scratch_buffer_blob(const_req); 1384 } 1385 1386 CodeBuffer* PhaseOutput::init_buffer() { 1387 int stub_req = _buf_sizes._stub; 1388 int code_req = _buf_sizes._code; 1389 int const_req = _buf_sizes._const; 1390 1391 int pad_req = NativeCall::instruction_size; 1392 1393 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1394 stub_req += bs->estimate_stub_size(); 1395 1396 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1397 // class HandlerImpl is platform-specific and defined in the *.ad files. 1398 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler 1399 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler 1400 stub_req += MAX_stubs_size; // ensure per-stub margin 1401 code_req += MAX_inst_size; // ensure per-instruction margin 1402 1403 if (StressCodeBuffers) 1404 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1405 1406 int total_req = 1407 const_req + 1408 code_req + 1409 pad_req + 1410 stub_req + 1411 exception_handler_req + 1412 deopt_handler_req; // deopt handler 1413 1414 if (C->has_method_handle_invokes()) 1415 total_req += deopt_handler_req; // deopt MH handler 1416 1417 CodeBuffer* cb = code_buffer(); 1418 cb->initialize(total_req, _buf_sizes._reloc); 1419 1420 // Have we run out of code space? 1421 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) { 1422 C->record_failure("CodeCache is full"); 1423 return nullptr; 1424 } 1425 // Configure the code buffer. 1426 cb->initialize_consts_size(const_req); 1427 cb->initialize_stubs_size(stub_req); 1428 cb->initialize_oop_recorder(C->env()->oop_recorder()); 1429 1430 // fill in the nop array for bundling computations 1431 MachNode *_nop_list[Bundle::_nop_count]; 1432 Bundle::initialize_nops(_nop_list); 1433 1434 return cb; 1435 } 1436 1437 //------------------------------fill_buffer------------------------------------ 1438 void PhaseOutput::fill_buffer(CodeBuffer* cb, uint* blk_starts) { 1439 // blk_starts[] contains offsets calculated during short branches processing, 1440 // offsets should not be increased during following steps. 1441 1442 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1443 // of a loop. It is used to determine the padding for loop alignment. 1444 Compile::TracePhase tp("fill buffer", &timers[_t_fillBuffer]); 1445 1446 compute_loop_first_inst_sizes(); 1447 1448 // Create oopmap set. 1449 _oop_map_set = new OopMapSet(); 1450 1451 // !!!!! This preserves old handling of oopmaps for now 1452 C->debug_info()->set_oopmaps(_oop_map_set); 1453 1454 uint nblocks = C->cfg()->number_of_blocks(); 1455 // Count and start of implicit null check instructions 1456 uint inct_cnt = 0; 1457 uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1458 1459 // Count and start of calls 1460 uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1461 1462 uint return_offset = 0; 1463 int nop_size = (new MachNopNode())->size(C->regalloc()); 1464 1465 int previous_offset = 0; 1466 int current_offset = 0; 1467 int last_call_offset = -1; 1468 int last_avoid_back_to_back_offset = -1; 1469 #ifdef ASSERT 1470 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1471 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1472 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1473 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1474 #endif 1475 1476 // Create an array of unused labels, one for each basic block, if printing is enabled 1477 #if defined(SUPPORT_OPTO_ASSEMBLY) 1478 int* node_offsets = nullptr; 1479 uint node_offset_limit = C->unique(); 1480 1481 if (C->print_assembly()) { 1482 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1483 } 1484 if (node_offsets != nullptr) { 1485 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly. 1486 memset(node_offsets, 0, node_offset_limit*sizeof(int)); 1487 } 1488 #endif 1489 1490 NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily 1491 1492 // Emit the constant table. 1493 if (C->has_mach_constant_base_node()) { 1494 if (!constant_table().emit(*cb)) { 1495 C->record_failure("consts section overflow"); 1496 return; 1497 } 1498 } 1499 1500 // Create an array of labels, one for each basic block 1501 Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1502 for (uint i = 0; i <= nblocks; i++) { 1503 blk_labels[i].init(); 1504 } 1505 1506 // Now fill in the code buffer 1507 Node* delay_slot = nullptr; 1508 for (uint i = 0; i < nblocks; i++) { 1509 Block* block = C->cfg()->get_block(i); 1510 _block = block; 1511 Node* head = block->head(); 1512 1513 // If this block needs to start aligned (i.e, can be reached other 1514 // than by falling-thru from the previous block), then force the 1515 // start of a new bundle. 1516 if (Pipeline::requires_bundling() && starts_bundle(head)) { 1517 cb->flush_bundle(true); 1518 } 1519 1520 #ifdef ASSERT 1521 if (!block->is_connector()) { 1522 stringStream st; 1523 block->dump_head(C->cfg(), &st); 1524 MacroAssembler(cb).block_comment(st.freeze()); 1525 } 1526 jmp_target[i] = 0; 1527 jmp_offset[i] = 0; 1528 jmp_size[i] = 0; 1529 jmp_rule[i] = 0; 1530 #endif 1531 int blk_offset = current_offset; 1532 1533 // Define the label at the beginning of the basic block 1534 MacroAssembler(cb).bind(blk_labels[block->_pre_order]); 1535 1536 uint last_inst = block->number_of_nodes(); 1537 1538 // Emit block normally, except for last instruction. 1539 // Emit means "dump code bits into code buffer". 1540 for (uint j = 0; j<last_inst; j++) { 1541 _index = j; 1542 1543 // Get the node 1544 Node* n = block->get_node(j); 1545 1546 // See if delay slots are supported 1547 if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) { 1548 assert(delay_slot == nullptr, "no use of delay slot node"); 1549 assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1550 1551 delay_slot = n; 1552 continue; 1553 } 1554 1555 // If this starts a new instruction group, then flush the current one 1556 // (but allow split bundles) 1557 if (Pipeline::requires_bundling() && starts_bundle(n)) 1558 cb->flush_bundle(false); 1559 1560 // Special handling for SafePoint/Call Nodes 1561 bool is_mcall = false; 1562 if (n->is_Mach()) { 1563 MachNode *mach = n->as_Mach(); 1564 is_mcall = n->is_MachCall(); 1565 bool is_sfn = n->is_MachSafePoint(); 1566 1567 // If this requires all previous instructions be flushed, then do so 1568 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1569 cb->flush_bundle(true); 1570 current_offset = cb->insts_size(); 1571 } 1572 1573 // A padding may be needed again since a previous instruction 1574 // could be moved to delay slot. 1575 1576 // align the instruction if necessary 1577 int padding = mach->compute_padding(current_offset); 1578 // Make sure safepoint node for polling is distinct from a call's 1579 // return by adding a nop if needed. 1580 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1581 padding = nop_size; 1582 } 1583 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) && 1584 current_offset == last_avoid_back_to_back_offset) { 1585 // Avoid back to back some instructions. 1586 padding = nop_size; 1587 } 1588 1589 if (padding > 0) { 1590 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1591 int nops_cnt = padding / nop_size; 1592 MachNode *nop = new MachNopNode(nops_cnt); 1593 block->insert_node(nop, j++); 1594 last_inst++; 1595 C->cfg()->map_node_to_block(nop, block); 1596 // Ensure enough space. 1597 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1598 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) { 1599 C->record_failure("CodeCache is full"); 1600 return; 1601 } 1602 nop->emit(*cb, C->regalloc()); 1603 cb->flush_bundle(true); 1604 current_offset = cb->insts_size(); 1605 } 1606 1607 bool observe_safepoint = is_sfn; 1608 // Remember the start of the last call in a basic block 1609 if (is_mcall) { 1610 MachCallNode *mcall = mach->as_MachCall(); 1611 1612 if (mcall->entry_point() != nullptr) { 1613 // This destination address is NOT PC-relative 1614 mcall->method_set((intptr_t)mcall->entry_point()); 1615 } 1616 1617 // Save the return address 1618 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset(); 1619 1620 observe_safepoint = mcall->guaranteed_safepoint(); 1621 } 1622 1623 // sfn will be valid whenever mcall is valid now because of inheritance 1624 if (observe_safepoint) { 1625 // Handle special safepoint nodes for synchronization 1626 if (!is_mcall) { 1627 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1628 // !!!!! Stubs only need an oopmap right now, so bail out 1629 if (sfn->jvms()->method() == nullptr) { 1630 // Write the oopmap directly to the code blob??!! 1631 continue; 1632 } 1633 } // End synchronization 1634 1635 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1636 current_offset); 1637 Process_OopMap_Node(mach, current_offset); 1638 } // End if safepoint 1639 1640 // If this is a null check, then add the start of the previous instruction to the list 1641 else if( mach->is_MachNullCheck() ) { 1642 inct_starts[inct_cnt++] = previous_offset; 1643 } 1644 1645 // If this is a branch, then fill in the label with the target BB's label 1646 else if (mach->is_MachBranch()) { 1647 // This requires the TRUE branch target be in succs[0] 1648 uint block_num = block->non_connector_successor(0)->_pre_order; 1649 1650 // Try to replace long branch if delay slot is not used, 1651 // it is mostly for back branches since forward branch's 1652 // distance is not updated yet. 1653 bool delay_slot_is_used = valid_bundle_info(n) && 1654 C->output()->node_bundling(n)->use_unconditional_delay(); 1655 if (!delay_slot_is_used && mach->may_be_short_branch()) { 1656 assert(delay_slot == nullptr, "not expecting delay slot node"); 1657 int br_size = n->size(C->regalloc()); 1658 int offset = blk_starts[block_num] - current_offset; 1659 if (block_num >= i) { 1660 // Current and following block's offset are not 1661 // finalized yet, adjust distance by the difference 1662 // between calculated and final offsets of current block. 1663 offset -= (blk_starts[i] - blk_offset); 1664 } 1665 // In the following code a nop could be inserted before 1666 // the branch which will increase the backward distance. 1667 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1668 if (needs_padding && offset <= 0) 1669 offset -= nop_size; 1670 1671 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) { 1672 // We've got a winner. Replace this branch. 1673 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 1674 1675 // Update the jmp_size. 1676 int new_size = replacement->size(C->regalloc()); 1677 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1678 // Insert padding between avoid_back_to_back branches. 1679 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 1680 MachNode *nop = new MachNopNode(); 1681 block->insert_node(nop, j++); 1682 C->cfg()->map_node_to_block(nop, block); 1683 last_inst++; 1684 nop->emit(*cb, C->regalloc()); 1685 cb->flush_bundle(true); 1686 current_offset = cb->insts_size(); 1687 } 1688 #ifdef ASSERT 1689 jmp_target[i] = block_num; 1690 jmp_offset[i] = current_offset - blk_offset; 1691 jmp_size[i] = new_size; 1692 jmp_rule[i] = mach->rule(); 1693 #endif 1694 block->map_node(replacement, j); 1695 mach->subsume_by(replacement, C); 1696 n = replacement; 1697 mach = replacement; 1698 } 1699 } 1700 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1701 } else if (mach->ideal_Opcode() == Op_Jump) { 1702 for (uint h = 0; h < block->_num_succs; h++) { 1703 Block* succs_block = block->_succs[h]; 1704 for (uint j = 1; j < succs_block->num_preds(); j++) { 1705 Node* jpn = succs_block->pred(j); 1706 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1707 uint block_num = succs_block->non_connector()->_pre_order; 1708 Label *blkLabel = &blk_labels[block_num]; 1709 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1710 } 1711 } 1712 } 1713 } 1714 #ifdef ASSERT 1715 // Check that oop-store precedes the card-mark 1716 else if (mach->ideal_Opcode() == Op_StoreCM) { 1717 uint storeCM_idx = j; 1718 int count = 0; 1719 for (uint prec = mach->req(); prec < mach->len(); prec++) { 1720 Node *oop_store = mach->in(prec); // Precedence edge 1721 if (oop_store == nullptr) continue; 1722 count++; 1723 uint i4; 1724 for (i4 = 0; i4 < last_inst; ++i4) { 1725 if (block->get_node(i4) == oop_store) { 1726 break; 1727 } 1728 } 1729 // Note: This test can provide a false failure if other precedence 1730 // edges have been added to the storeCMNode. 1731 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1732 } 1733 assert(count > 0, "storeCM expects at least one precedence edge"); 1734 } 1735 #endif 1736 else if (!n->is_Proj()) { 1737 // Remember the beginning of the previous instruction, in case 1738 // it's followed by a flag-kill and a null-check. Happens on 1739 // Intel all the time, with add-to-memory kind of opcodes. 1740 previous_offset = current_offset; 1741 } 1742 1743 // Not an else-if! 1744 // If this is a trap based cmp then add its offset to the list. 1745 if (mach->is_TrapBasedCheckNode()) { 1746 inct_starts[inct_cnt++] = current_offset; 1747 } 1748 } 1749 1750 // Verify that there is sufficient space remaining 1751 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1752 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) { 1753 C->record_failure("CodeCache is full"); 1754 return; 1755 } 1756 1757 // Save the offset for the listing 1758 #if defined(SUPPORT_OPTO_ASSEMBLY) 1759 if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) { 1760 node_offsets[n->_idx] = cb->insts_size(); 1761 } 1762 #endif 1763 assert(!C->failing(), "Should not reach here if failing."); 1764 1765 // "Normal" instruction case 1766 DEBUG_ONLY(uint instr_offset = cb->insts_size()); 1767 n->emit(*cb, C->regalloc()); 1768 current_offset = cb->insts_size(); 1769 1770 // Above we only verified that there is enough space in the instruction section. 1771 // However, the instruction may emit stubs that cause code buffer expansion. 1772 // Bail out here if expansion failed due to a lack of code cache space. 1773 if (C->failing()) { 1774 return; 1775 } 1776 1777 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset), 1778 "ret_addr_offset() not within emitted code"); 1779 #ifdef ASSERT 1780 uint n_size = n->size(C->regalloc()); 1781 if (n_size < (current_offset-instr_offset)) { 1782 MachNode* mach = n->as_Mach(); 1783 n->dump(); 1784 mach->dump_format(C->regalloc(), tty); 1785 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset); 1786 Disassembler::decode(cb->insts_begin() + instr_offset, cb->insts_begin() + current_offset + 1, tty); 1787 tty->print_cr(" ------------------- "); 1788 BufferBlob* blob = this->scratch_buffer_blob(); 1789 address blob_begin = blob->content_begin(); 1790 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty); 1791 assert(false, "wrong size of mach node"); 1792 } 1793 #endif 1794 non_safepoints.observe_instruction(n, current_offset); 1795 1796 // mcall is last "call" that can be a safepoint 1797 // record it so we can see if a poll will directly follow it 1798 // in which case we'll need a pad to make the PcDesc sites unique 1799 // see 5010568. This can be slightly inaccurate but conservative 1800 // in the case that return address is not actually at current_offset. 1801 // This is a small price to pay. 1802 1803 if (is_mcall) { 1804 last_call_offset = current_offset; 1805 } 1806 1807 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 1808 // Avoid back to back some instructions. 1809 last_avoid_back_to_back_offset = current_offset; 1810 } 1811 1812 // See if this instruction has a delay slot 1813 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1814 guarantee(delay_slot != nullptr, "expecting delay slot node"); 1815 1816 // Back up 1 instruction 1817 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1818 1819 // Save the offset for the listing 1820 #if defined(SUPPORT_OPTO_ASSEMBLY) 1821 if ((node_offsets != nullptr) && (delay_slot->_idx < node_offset_limit)) { 1822 node_offsets[delay_slot->_idx] = cb->insts_size(); 1823 } 1824 #endif 1825 1826 // Support a SafePoint in the delay slot 1827 if (delay_slot->is_MachSafePoint()) { 1828 MachNode *mach = delay_slot->as_Mach(); 1829 // !!!!! Stubs only need an oopmap right now, so bail out 1830 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == nullptr) { 1831 // Write the oopmap directly to the code blob??!! 1832 delay_slot = nullptr; 1833 continue; 1834 } 1835 1836 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1837 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1838 adjusted_offset); 1839 // Generate an OopMap entry 1840 Process_OopMap_Node(mach, adjusted_offset); 1841 } 1842 1843 // Insert the delay slot instruction 1844 delay_slot->emit(*cb, C->regalloc()); 1845 1846 // Don't reuse it 1847 delay_slot = nullptr; 1848 } 1849 1850 } // End for all instructions in block 1851 1852 // If the next block is the top of a loop, pad this block out to align 1853 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1854 if (i < nblocks-1) { 1855 Block *nb = C->cfg()->get_block(i + 1); 1856 int padding = nb->alignment_padding(current_offset); 1857 if( padding > 0 ) { 1858 MachNode *nop = new MachNopNode(padding / nop_size); 1859 block->insert_node(nop, block->number_of_nodes()); 1860 C->cfg()->map_node_to_block(nop, block); 1861 nop->emit(*cb, C->regalloc()); 1862 current_offset = cb->insts_size(); 1863 } 1864 } 1865 // Verify that the distance for generated before forward 1866 // short branches is still valid. 1867 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size"); 1868 1869 // Save new block start offset 1870 blk_starts[i] = blk_offset; 1871 } // End of for all blocks 1872 blk_starts[nblocks] = current_offset; 1873 1874 non_safepoints.flush_at_end(); 1875 1876 // Offset too large? 1877 if (C->failing()) return; 1878 1879 // Define a pseudo-label at the end of the code 1880 MacroAssembler(cb).bind( blk_labels[nblocks] ); 1881 1882 // Compute the size of the first block 1883 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1884 1885 #ifdef ASSERT 1886 for (uint i = 0; i < nblocks; i++) { // For all blocks 1887 if (jmp_target[i] != 0) { 1888 int br_size = jmp_size[i]; 1889 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1890 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1891 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]); 1892 assert(false, "Displacement too large for short jmp"); 1893 } 1894 } 1895 } 1896 #endif 1897 1898 if (!cb->finalize_stubs()) { 1899 C->record_failure("CodeCache is full"); 1900 return; 1901 } 1902 1903 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1904 bs->emit_stubs(*cb); 1905 if (C->failing()) return; 1906 1907 // Fill in stubs. 1908 _stub_list.emit(*cb); 1909 if (C->failing()) return; 1910 1911 #ifndef PRODUCT 1912 // Information on the size of the method, without the extraneous code 1913 Scheduling::increment_method_size(cb->insts_size()); 1914 #endif 1915 1916 // ------------------ 1917 // Fill in exception table entries. 1918 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1919 1920 // Only java methods have exception handlers and deopt handlers 1921 // class HandlerImpl is platform-specific and defined in the *.ad files. 1922 if (C->method()) { 1923 // Emit the exception handler code. 1924 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb)); 1925 if (C->failing()) { 1926 return; // CodeBuffer::expand failed 1927 } 1928 // Emit the deopt handler code. 1929 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb)); 1930 1931 // Emit the MethodHandle deopt handler code (if required). 1932 if (C->has_method_handle_invokes() && !C->failing()) { 1933 // We can use the same code as for the normal deopt handler, we 1934 // just need a different entry point address. 1935 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb)); 1936 } 1937 } 1938 1939 // One last check for failed CodeBuffer::expand: 1940 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) { 1941 C->record_failure("CodeCache is full"); 1942 return; 1943 } 1944 1945 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY) 1946 if (C->print_assembly()) { 1947 tty->cr(); 1948 tty->print_cr("============================= C2-compiled nmethod =============================="); 1949 } 1950 #endif 1951 1952 #if defined(SUPPORT_OPTO_ASSEMBLY) 1953 // Dump the assembly code, including basic-block numbers 1954 if (C->print_assembly()) { 1955 ttyLocker ttyl; // keep the following output all in one block 1956 if (!VMThread::should_terminate()) { // test this under the tty lock 1957 // print_metadata and dump_asm may safepoint which makes us loose the ttylock. 1958 // We call them first and write to a stringStream, then we retake the lock to 1959 // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe. 1960 ResourceMark rm; 1961 stringStream method_metadata_str; 1962 if (C->method() != nullptr) { 1963 C->method()->print_metadata(&method_metadata_str); 1964 } 1965 stringStream dump_asm_str; 1966 dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit); 1967 1968 NoSafepointVerifier nsv; 1969 ttyLocker ttyl2; 1970 // This output goes directly to the tty, not the compiler log. 1971 // To enable tools to match it up with the compilation activity, 1972 // be sure to tag this tty output with the compile ID. 1973 if (xtty != nullptr) { 1974 xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(), 1975 C->is_osr_compilation() ? " compile_kind='osr'" : ""); 1976 } 1977 if (C->method() != nullptr) { 1978 tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id()); 1979 tty->print_raw(method_metadata_str.freeze()); 1980 } else if (C->stub_name() != nullptr) { 1981 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name()); 1982 } 1983 tty->cr(); 1984 tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id()); 1985 tty->print_raw(dump_asm_str.freeze()); 1986 tty->print_cr("--------------------------------------------------------------------------------"); 1987 if (xtty != nullptr) { 1988 xtty->tail("opto_assembly"); 1989 } 1990 } 1991 } 1992 #endif 1993 } 1994 1995 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1996 _inc_table.set_size(cnt); 1997 1998 uint inct_cnt = 0; 1999 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) { 2000 Block* block = C->cfg()->get_block(i); 2001 Node *n = nullptr; 2002 int j; 2003 2004 // Find the branch; ignore trailing NOPs. 2005 for (j = block->number_of_nodes() - 1; j >= 0; j--) { 2006 n = block->get_node(j); 2007 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) { 2008 break; 2009 } 2010 } 2011 2012 // If we didn't find anything, continue 2013 if (j < 0) { 2014 continue; 2015 } 2016 2017 // Compute ExceptionHandlerTable subtable entry and add it 2018 // (skip empty blocks) 2019 if (n->is_Catch()) { 2020 2021 // Get the offset of the return from the call 2022 uint call_return = call_returns[block->_pre_order]; 2023 #ifdef ASSERT 2024 assert( call_return > 0, "no call seen for this basic block" ); 2025 while (block->get_node(--j)->is_MachProj()) ; 2026 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); 2027 #endif 2028 // last instruction is a CatchNode, find it's CatchProjNodes 2029 int nof_succs = block->_num_succs; 2030 // allocate space 2031 GrowableArray<intptr_t> handler_bcis(nof_succs); 2032 GrowableArray<intptr_t> handler_pcos(nof_succs); 2033 // iterate through all successors 2034 for (int j = 0; j < nof_succs; j++) { 2035 Block* s = block->_succs[j]; 2036 bool found_p = false; 2037 for (uint k = 1; k < s->num_preds(); k++) { 2038 Node* pk = s->pred(k); 2039 if (pk->is_CatchProj() && pk->in(0) == n) { 2040 const CatchProjNode* p = pk->as_CatchProj(); 2041 found_p = true; 2042 // add the corresponding handler bci & pco information 2043 if (p->_con != CatchProjNode::fall_through_index) { 2044 // p leads to an exception handler (and is not fall through) 2045 assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering"); 2046 // no duplicates, please 2047 if (!handler_bcis.contains(p->handler_bci())) { 2048 uint block_num = s->non_connector()->_pre_order; 2049 handler_bcis.append(p->handler_bci()); 2050 handler_pcos.append(blk_labels[block_num].loc_pos()); 2051 } 2052 } 2053 } 2054 } 2055 assert(found_p, "no matching predecessor found"); 2056 // Note: Due to empty block removal, one block may have 2057 // several CatchProj inputs, from the same Catch. 2058 } 2059 2060 // Set the offset of the return from the call 2061 assert(handler_bcis.find(-1) != -1, "must have default handler"); 2062 _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos); 2063 continue; 2064 } 2065 2066 // Handle implicit null exception table updates 2067 if (n->is_MachNullCheck()) { 2068 uint block_num = block->non_connector_successor(0)->_pre_order; 2069 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 2070 continue; 2071 } 2072 // Handle implicit exception table updates: trap instructions. 2073 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) { 2074 uint block_num = block->non_connector_successor(0)->_pre_order; 2075 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 2076 continue; 2077 } 2078 } // End of for all blocks fill in exception table entries 2079 } 2080 2081 // Static Variables 2082 #ifndef PRODUCT 2083 uint Scheduling::_total_nop_size = 0; 2084 uint Scheduling::_total_method_size = 0; 2085 uint Scheduling::_total_branches = 0; 2086 uint Scheduling::_total_unconditional_delays = 0; 2087 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 2088 #endif 2089 2090 // Initializer for class Scheduling 2091 2092 Scheduling::Scheduling(Arena *arena, Compile &compile) 2093 : _arena(arena), 2094 _cfg(compile.cfg()), 2095 _regalloc(compile.regalloc()), 2096 _scheduled(arena), 2097 _available(arena), 2098 _reg_node(arena), 2099 _pinch_free_list(arena), 2100 _next_node(nullptr), 2101 _bundle_instr_count(0), 2102 _bundle_cycle_number(0), 2103 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]) 2104 #ifndef PRODUCT 2105 , _branches(0) 2106 , _unconditional_delays(0) 2107 #endif 2108 { 2109 // Create a MachNopNode 2110 _nop = new MachNopNode(); 2111 2112 // Now that the nops are in the array, save the count 2113 // (but allow entries for the nops) 2114 _node_bundling_limit = compile.unique(); 2115 uint node_max = _regalloc->node_regs_max_index(); 2116 2117 compile.output()->set_node_bundling_limit(_node_bundling_limit); 2118 2119 // This one is persistent within the Compile class 2120 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 2121 2122 // Allocate space for fixed-size arrays 2123 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 2124 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 2125 2126 // Clear the arrays 2127 for (uint i = 0; i < node_max; i++) { 2128 ::new (&_node_bundling_base[i]) Bundle(); 2129 } 2130 memset(_uses, 0, node_max * sizeof(short)); 2131 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 2132 2133 // Clear the bundling information 2134 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements)); 2135 2136 // Get the last node 2137 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1); 2138 2139 _next_node = block->get_node(block->number_of_nodes() - 1); 2140 } 2141 2142 #ifndef PRODUCT 2143 // Scheduling destructor 2144 Scheduling::~Scheduling() { 2145 _total_branches += _branches; 2146 _total_unconditional_delays += _unconditional_delays; 2147 } 2148 #endif 2149 2150 // Step ahead "i" cycles 2151 void Scheduling::step(uint i) { 2152 2153 Bundle *bundle = node_bundling(_next_node); 2154 bundle->set_starts_bundle(); 2155 2156 // Update the bundle record, but leave the flags information alone 2157 if (_bundle_instr_count > 0) { 2158 bundle->set_instr_count(_bundle_instr_count); 2159 bundle->set_resources_used(_bundle_use.resourcesUsed()); 2160 } 2161 2162 // Update the state information 2163 _bundle_instr_count = 0; 2164 _bundle_cycle_number += i; 2165 _bundle_use.step(i); 2166 } 2167 2168 void Scheduling::step_and_clear() { 2169 Bundle *bundle = node_bundling(_next_node); 2170 bundle->set_starts_bundle(); 2171 2172 // Update the bundle record 2173 if (_bundle_instr_count > 0) { 2174 bundle->set_instr_count(_bundle_instr_count); 2175 bundle->set_resources_used(_bundle_use.resourcesUsed()); 2176 2177 _bundle_cycle_number += 1; 2178 } 2179 2180 // Clear the bundling information 2181 _bundle_instr_count = 0; 2182 _bundle_use.reset(); 2183 2184 memcpy(_bundle_use_elements, 2185 Pipeline_Use::elaborated_elements, 2186 sizeof(Pipeline_Use::elaborated_elements)); 2187 } 2188 2189 // Perform instruction scheduling and bundling over the sequence of 2190 // instructions in backwards order. 2191 void PhaseOutput::ScheduleAndBundle() { 2192 2193 // Don't optimize this if it isn't a method 2194 if (!C->method()) 2195 return; 2196 2197 // Don't optimize this if scheduling is disabled 2198 if (!C->do_scheduling()) 2199 return; 2200 2201 // Scheduling code works only with pairs (8 bytes) maximum. 2202 // And when the scalable vector register is used, we may spill/unspill 2203 // the whole reg regardless of the max vector size. 2204 if (C->max_vector_size() > 8 || 2205 (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) { 2206 return; 2207 } 2208 2209 Compile::TracePhase tp("isched", &timers[_t_instrSched]); 2210 2211 // Create a data structure for all the scheduling information 2212 Scheduling scheduling(Thread::current()->resource_area(), *C); 2213 2214 // Walk backwards over each basic block, computing the needed alignment 2215 // Walk over all the basic blocks 2216 scheduling.DoScheduling(); 2217 2218 #ifndef PRODUCT 2219 if (C->trace_opto_output()) { 2220 // Buffer and print all at once 2221 ResourceMark rm; 2222 stringStream ss; 2223 ss.print("\n---- After ScheduleAndBundle ----\n"); 2224 print_scheduling(&ss); 2225 tty->print("%s", ss.as_string()); 2226 } 2227 #endif 2228 } 2229 2230 #ifndef PRODUCT 2231 // Separated out so that it can be called directly from debugger 2232 void PhaseOutput::print_scheduling() { 2233 print_scheduling(tty); 2234 } 2235 2236 void PhaseOutput::print_scheduling(outputStream* output_stream) { 2237 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) { 2238 output_stream->print("\nBB#%03d:\n", i); 2239 Block* block = C->cfg()->get_block(i); 2240 for (uint j = 0; j < block->number_of_nodes(); j++) { 2241 Node* n = block->get_node(j); 2242 OptoReg::Name reg = C->regalloc()->get_reg_first(n); 2243 output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 2244 n->dump("\n", false, output_stream); 2245 } 2246 } 2247 } 2248 #endif 2249 2250 // See if this node fits into the present instruction bundle 2251 bool Scheduling::NodeFitsInBundle(Node *n) { 2252 uint n_idx = n->_idx; 2253 2254 // If this is the unconditional delay instruction, then it fits 2255 if (n == _unconditional_delay_slot) { 2256 #ifndef PRODUCT 2257 if (_cfg->C->trace_opto_output()) 2258 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 2259 #endif 2260 return (true); 2261 } 2262 2263 // If the node cannot be scheduled this cycle, skip it 2264 if (_current_latency[n_idx] > _bundle_cycle_number) { 2265 #ifndef PRODUCT 2266 if (_cfg->C->trace_opto_output()) 2267 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 2268 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 2269 #endif 2270 return (false); 2271 } 2272 2273 const Pipeline *node_pipeline = n->pipeline(); 2274 2275 uint instruction_count = node_pipeline->instructionCount(); 2276 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2277 instruction_count = 0; 2278 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2279 instruction_count++; 2280 2281 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 2282 #ifndef PRODUCT 2283 if (_cfg->C->trace_opto_output()) 2284 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 2285 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 2286 #endif 2287 return (false); 2288 } 2289 2290 // Don't allow non-machine nodes to be handled this way 2291 if (!n->is_Mach() && instruction_count == 0) 2292 return (false); 2293 2294 // See if there is any overlap 2295 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 2296 2297 if (delay > 0) { 2298 #ifndef PRODUCT 2299 if (_cfg->C->trace_opto_output()) 2300 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 2301 #endif 2302 return false; 2303 } 2304 2305 #ifndef PRODUCT 2306 if (_cfg->C->trace_opto_output()) 2307 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 2308 #endif 2309 2310 return true; 2311 } 2312 2313 Node * Scheduling::ChooseNodeToBundle() { 2314 uint siz = _available.size(); 2315 2316 if (siz == 0) { 2317 2318 #ifndef PRODUCT 2319 if (_cfg->C->trace_opto_output()) 2320 tty->print("# ChooseNodeToBundle: null\n"); 2321 #endif 2322 return (nullptr); 2323 } 2324 2325 // Fast path, if only 1 instruction in the bundle 2326 if (siz == 1) { 2327 #ifndef PRODUCT 2328 if (_cfg->C->trace_opto_output()) { 2329 tty->print("# ChooseNodeToBundle (only 1): "); 2330 _available[0]->dump(); 2331 } 2332 #endif 2333 return (_available[0]); 2334 } 2335 2336 // Don't bother, if the bundle is already full 2337 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 2338 for ( uint i = 0; i < siz; i++ ) { 2339 Node *n = _available[i]; 2340 2341 // Skip projections, we'll handle them another way 2342 if (n->is_Proj()) 2343 continue; 2344 2345 // This presupposed that instructions are inserted into the 2346 // available list in a legality order; i.e. instructions that 2347 // must be inserted first are at the head of the list 2348 if (NodeFitsInBundle(n)) { 2349 #ifndef PRODUCT 2350 if (_cfg->C->trace_opto_output()) { 2351 tty->print("# ChooseNodeToBundle: "); 2352 n->dump(); 2353 } 2354 #endif 2355 return (n); 2356 } 2357 } 2358 } 2359 2360 // Nothing fits in this bundle, choose the highest priority 2361 #ifndef PRODUCT 2362 if (_cfg->C->trace_opto_output()) { 2363 tty->print("# ChooseNodeToBundle: "); 2364 _available[0]->dump(); 2365 } 2366 #endif 2367 2368 return _available[0]; 2369 } 2370 2371 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) { 2372 assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), ""); 2373 2374 OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1)); 2375 OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first); 2376 OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1)); 2377 OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second); 2378 2379 // Comparison between stack -> reg and stack -> reg 2380 if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) && 2381 OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) { 2382 return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo); 2383 } 2384 2385 // Comparison between reg -> stack and reg -> stack 2386 if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) && 2387 OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) { 2388 return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo); 2389 } 2390 2391 return 0; // Not comparable 2392 } 2393 2394 void Scheduling::AddNodeToAvailableList(Node *n) { 2395 assert( !n->is_Proj(), "projections never directly made available" ); 2396 #ifndef PRODUCT 2397 if (_cfg->C->trace_opto_output()) { 2398 tty->print("# AddNodeToAvailableList: "); 2399 n->dump(); 2400 } 2401 #endif 2402 2403 int latency = _current_latency[n->_idx]; 2404 2405 // Insert in latency order (insertion sort). If two MachSpillCopyNodes 2406 // for stack spilling or unspilling have the same latency, we sort 2407 // them in the order of stack offset. Some ports (e.g. aarch64) may also 2408 // have more opportunities to do ld/st merging 2409 uint i; 2410 for (i = 0; i < _available.size(); i++) { 2411 if (_current_latency[_available[i]->_idx] > latency) { 2412 break; 2413 } else if (_current_latency[_available[i]->_idx] == latency && 2414 n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() && 2415 compare_two_spill_nodes(n, _available[i]) > 0) { 2416 break; 2417 } 2418 } 2419 2420 // Special Check for compares following branches 2421 if( n->is_Mach() && _scheduled.size() > 0 ) { 2422 int op = n->as_Mach()->ideal_Opcode(); 2423 Node *last = _scheduled[0]; 2424 if( last->is_MachIf() && last->in(1) == n && 2425 ( op == Op_CmpI || 2426 op == Op_CmpU || 2427 op == Op_CmpUL || 2428 op == Op_CmpP || 2429 op == Op_CmpF || 2430 op == Op_CmpD || 2431 op == Op_CmpL ) ) { 2432 2433 // Recalculate position, moving to front of same latency 2434 for ( i=0 ; i < _available.size(); i++ ) 2435 if (_current_latency[_available[i]->_idx] >= latency) 2436 break; 2437 } 2438 } 2439 2440 // Insert the node in the available list 2441 _available.insert(i, n); 2442 2443 #ifndef PRODUCT 2444 if (_cfg->C->trace_opto_output()) 2445 dump_available(); 2446 #endif 2447 } 2448 2449 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 2450 for ( uint i=0; i < n->len(); i++ ) { 2451 Node *def = n->in(i); 2452 if (!def) continue; 2453 if( def->is_Proj() ) // If this is a machine projection, then 2454 def = def->in(0); // propagate usage thru to the base instruction 2455 2456 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local 2457 continue; 2458 } 2459 2460 // Compute the latency 2461 uint l = _bundle_cycle_number + n->latency(i); 2462 if (_current_latency[def->_idx] < l) 2463 _current_latency[def->_idx] = l; 2464 2465 // If this does not have uses then schedule it 2466 if ((--_uses[def->_idx]) == 0) 2467 AddNodeToAvailableList(def); 2468 } 2469 } 2470 2471 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2472 #ifndef PRODUCT 2473 if (_cfg->C->trace_opto_output()) { 2474 tty->print("# AddNodeToBundle: "); 2475 n->dump(); 2476 } 2477 #endif 2478 2479 // Remove this from the available list 2480 uint i; 2481 for (i = 0; i < _available.size(); i++) 2482 if (_available[i] == n) 2483 break; 2484 assert(i < _available.size(), "entry in _available list not found"); 2485 _available.remove(i); 2486 2487 // See if this fits in the current bundle 2488 const Pipeline *node_pipeline = n->pipeline(); 2489 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2490 2491 // Check for instructions to be placed in the delay slot. We 2492 // do this before we actually schedule the current instruction, 2493 // because the delay slot follows the current instruction. 2494 if (Pipeline::_branch_has_delay_slot && 2495 node_pipeline->hasBranchDelay() && 2496 !_unconditional_delay_slot) { 2497 2498 uint siz = _available.size(); 2499 2500 // Conditional branches can support an instruction that 2501 // is unconditionally executed and not dependent by the 2502 // branch, OR a conditionally executed instruction if 2503 // the branch is taken. In practice, this means that 2504 // the first instruction at the branch target is 2505 // copied to the delay slot, and the branch goes to 2506 // the instruction after that at the branch target 2507 if ( n->is_MachBranch() ) { 2508 2509 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2510 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2511 2512 #ifndef PRODUCT 2513 _branches++; 2514 #endif 2515 2516 // At least 1 instruction is on the available list 2517 // that is not dependent on the branch 2518 for (uint i = 0; i < siz; i++) { 2519 Node *d = _available[i]; 2520 const Pipeline *avail_pipeline = d->pipeline(); 2521 2522 // Don't allow safepoints in the branch shadow, that will 2523 // cause a number of difficulties 2524 if ( avail_pipeline->instructionCount() == 1 && 2525 !avail_pipeline->hasMultipleBundles() && 2526 !avail_pipeline->hasBranchDelay() && 2527 Pipeline::instr_has_unit_size() && 2528 d->size(_regalloc) == Pipeline::instr_unit_size() && 2529 NodeFitsInBundle(d) && 2530 !node_bundling(d)->used_in_delay()) { 2531 2532 if (d->is_Mach() && !d->is_MachSafePoint()) { 2533 // A node that fits in the delay slot was found, so we need to 2534 // set the appropriate bits in the bundle pipeline information so 2535 // that it correctly indicates resource usage. Later, when we 2536 // attempt to add this instruction to the bundle, we will skip 2537 // setting the resource usage. 2538 _unconditional_delay_slot = d; 2539 node_bundling(n)->set_use_unconditional_delay(); 2540 node_bundling(d)->set_used_in_unconditional_delay(); 2541 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2542 _current_latency[d->_idx] = _bundle_cycle_number; 2543 _next_node = d; 2544 ++_bundle_instr_count; 2545 #ifndef PRODUCT 2546 _unconditional_delays++; 2547 #endif 2548 break; 2549 } 2550 } 2551 } 2552 } 2553 2554 // No delay slot, add a nop to the usage 2555 if (!_unconditional_delay_slot) { 2556 // See if adding an instruction in the delay slot will overflow 2557 // the bundle. 2558 if (!NodeFitsInBundle(_nop)) { 2559 #ifndef PRODUCT 2560 if (_cfg->C->trace_opto_output()) 2561 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2562 #endif 2563 step(1); 2564 } 2565 2566 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2567 _next_node = _nop; 2568 ++_bundle_instr_count; 2569 } 2570 2571 // See if the instruction in the delay slot requires a 2572 // step of the bundles 2573 if (!NodeFitsInBundle(n)) { 2574 #ifndef PRODUCT 2575 if (_cfg->C->trace_opto_output()) 2576 tty->print("# *** STEP(branch won't fit) ***\n"); 2577 #endif 2578 // Update the state information 2579 _bundle_instr_count = 0; 2580 _bundle_cycle_number += 1; 2581 _bundle_use.step(1); 2582 } 2583 } 2584 2585 // Get the number of instructions 2586 uint instruction_count = node_pipeline->instructionCount(); 2587 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2588 instruction_count = 0; 2589 2590 // Compute the latency information 2591 uint delay = 0; 2592 2593 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2594 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2595 if (relative_latency < 0) 2596 relative_latency = 0; 2597 2598 delay = _bundle_use.full_latency(relative_latency, node_usage); 2599 2600 // Does not fit in this bundle, start a new one 2601 if (delay > 0) { 2602 step(delay); 2603 2604 #ifndef PRODUCT 2605 if (_cfg->C->trace_opto_output()) 2606 tty->print("# *** STEP(%d) ***\n", delay); 2607 #endif 2608 } 2609 } 2610 2611 // If this was placed in the delay slot, ignore it 2612 if (n != _unconditional_delay_slot) { 2613 2614 if (delay == 0) { 2615 if (node_pipeline->hasMultipleBundles()) { 2616 #ifndef PRODUCT 2617 if (_cfg->C->trace_opto_output()) 2618 tty->print("# *** STEP(multiple instructions) ***\n"); 2619 #endif 2620 step(1); 2621 } 2622 2623 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2624 #ifndef PRODUCT 2625 if (_cfg->C->trace_opto_output()) 2626 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2627 instruction_count + _bundle_instr_count, 2628 Pipeline::_max_instrs_per_cycle); 2629 #endif 2630 step(1); 2631 } 2632 } 2633 2634 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2635 _bundle_instr_count++; 2636 2637 // Set the node's latency 2638 _current_latency[n->_idx] = _bundle_cycle_number; 2639 2640 // Now merge the functional unit information 2641 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2642 _bundle_use.add_usage(node_usage); 2643 2644 // Increment the number of instructions in this bundle 2645 _bundle_instr_count += instruction_count; 2646 2647 // Remember this node for later 2648 if (n->is_Mach()) 2649 _next_node = n; 2650 } 2651 2652 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2653 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2654 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2655 // into the block. All other scheduled nodes get put in the schedule here. 2656 int op = n->Opcode(); 2657 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2658 (op != Op_Node && // Not an unused antidepedence node and 2659 // not an unallocated boxlock 2660 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2661 2662 // Push any trailing projections 2663 if( bb->get_node(bb->number_of_nodes()-1) != n ) { 2664 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2665 Node *foi = n->fast_out(i); 2666 if( foi->is_Proj() ) 2667 _scheduled.push(foi); 2668 } 2669 } 2670 2671 // Put the instruction in the schedule list 2672 _scheduled.push(n); 2673 } 2674 2675 #ifndef PRODUCT 2676 if (_cfg->C->trace_opto_output()) 2677 dump_available(); 2678 #endif 2679 2680 // Walk all the definitions, decrementing use counts, and 2681 // if a definition has a 0 use count, place it in the available list. 2682 DecrementUseCounts(n,bb); 2683 } 2684 2685 // This method sets the use count within a basic block. We will ignore all 2686 // uses outside the current basic block. As we are doing a backwards walk, 2687 // any node we reach that has a use count of 0 may be scheduled. This also 2688 // avoids the problem of cyclic references from phi nodes, as long as phi 2689 // nodes are at the front of the basic block. This method also initializes 2690 // the available list to the set of instructions that have no uses within this 2691 // basic block. 2692 void Scheduling::ComputeUseCount(const Block *bb) { 2693 #ifndef PRODUCT 2694 if (_cfg->C->trace_opto_output()) 2695 tty->print("# -> ComputeUseCount\n"); 2696 #endif 2697 2698 // Clear the list of available and scheduled instructions, just in case 2699 _available.clear(); 2700 _scheduled.clear(); 2701 2702 // No delay slot specified 2703 _unconditional_delay_slot = nullptr; 2704 2705 #ifdef ASSERT 2706 for( uint i=0; i < bb->number_of_nodes(); i++ ) 2707 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" ); 2708 #endif 2709 2710 // Force the _uses count to never go to zero for unscheduable pieces 2711 // of the block 2712 for( uint k = 0; k < _bb_start; k++ ) 2713 _uses[bb->get_node(k)->_idx] = 1; 2714 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ ) 2715 _uses[bb->get_node(l)->_idx] = 1; 2716 2717 // Iterate backwards over the instructions in the block. Don't count the 2718 // branch projections at end or the block header instructions. 2719 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2720 Node *n = bb->get_node(j); 2721 if( n->is_Proj() ) continue; // Projections handled another way 2722 2723 // Account for all uses 2724 for ( uint k = 0; k < n->len(); k++ ) { 2725 Node *inp = n->in(k); 2726 if (!inp) continue; 2727 assert(inp != n, "no cycles allowed" ); 2728 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use? 2729 if (inp->is_Proj()) { // Skip through Proj's 2730 inp = inp->in(0); 2731 } 2732 ++_uses[inp->_idx]; // Count 1 block-local use 2733 } 2734 } 2735 2736 // If this instruction has a 0 use count, then it is available 2737 if (!_uses[n->_idx]) { 2738 _current_latency[n->_idx] = _bundle_cycle_number; 2739 AddNodeToAvailableList(n); 2740 } 2741 2742 #ifndef PRODUCT 2743 if (_cfg->C->trace_opto_output()) { 2744 tty->print("# uses: %3d: ", _uses[n->_idx]); 2745 n->dump(); 2746 } 2747 #endif 2748 } 2749 2750 #ifndef PRODUCT 2751 if (_cfg->C->trace_opto_output()) 2752 tty->print("# <- ComputeUseCount\n"); 2753 #endif 2754 } 2755 2756 // This routine performs scheduling on each basic block in reverse order, 2757 // using instruction latencies and taking into account function unit 2758 // availability. 2759 void Scheduling::DoScheduling() { 2760 #ifndef PRODUCT 2761 if (_cfg->C->trace_opto_output()) 2762 tty->print("# -> DoScheduling\n"); 2763 #endif 2764 2765 Block *succ_bb = nullptr; 2766 Block *bb; 2767 Compile* C = Compile::current(); 2768 2769 // Walk over all the basic blocks in reverse order 2770 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) { 2771 bb = _cfg->get_block(i); 2772 2773 #ifndef PRODUCT 2774 if (_cfg->C->trace_opto_output()) { 2775 tty->print("# Schedule BB#%03d (initial)\n", i); 2776 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2777 bb->get_node(j)->dump(); 2778 } 2779 } 2780 #endif 2781 2782 // On the head node, skip processing 2783 if (bb == _cfg->get_root_block()) { 2784 continue; 2785 } 2786 2787 // Skip empty, connector blocks 2788 if (bb->is_connector()) 2789 continue; 2790 2791 // If the following block is not the sole successor of 2792 // this one, then reset the pipeline information 2793 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2794 #ifndef PRODUCT 2795 if (_cfg->C->trace_opto_output()) { 2796 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2797 _next_node->_idx, _bundle_instr_count); 2798 } 2799 #endif 2800 step_and_clear(); 2801 } 2802 2803 // Leave untouched the starting instruction, any Phis, a CreateEx node 2804 // or Top. bb->get_node(_bb_start) is the first schedulable instruction. 2805 _bb_end = bb->number_of_nodes()-1; 2806 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2807 Node *n = bb->get_node(_bb_start); 2808 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2809 // Also, MachIdealNodes do not get scheduled 2810 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2811 MachNode *mach = n->as_Mach(); 2812 int iop = mach->ideal_Opcode(); 2813 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2814 if( iop == Op_Con ) continue; // Do not schedule Top 2815 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2816 mach->pipeline() == MachNode::pipeline_class() && 2817 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc 2818 continue; 2819 break; // Funny loop structure to be sure... 2820 } 2821 // Compute last "interesting" instruction in block - last instruction we 2822 // might schedule. _bb_end points just after last schedulable inst. We 2823 // normally schedule conditional branches (despite them being forced last 2824 // in the block), because they have delay slots we can fill. Calls all 2825 // have their delay slots filled in the template expansions, so we don't 2826 // bother scheduling them. 2827 Node *last = bb->get_node(_bb_end); 2828 // Ignore trailing NOPs. 2829 while (_bb_end > 0 && last->is_Mach() && 2830 last->as_Mach()->ideal_Opcode() == Op_Con) { 2831 last = bb->get_node(--_bb_end); 2832 } 2833 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2834 if( last->is_Catch() || 2835 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2836 // There might be a prior call. Skip it. 2837 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj()); 2838 } else if( last->is_MachNullCheck() ) { 2839 // Backup so the last null-checked memory instruction is 2840 // outside the schedulable range. Skip over the nullcheck, 2841 // projection, and the memory nodes. 2842 Node *mem = last->in(1); 2843 do { 2844 _bb_end--; 2845 } while (mem != bb->get_node(_bb_end)); 2846 } else { 2847 // Set _bb_end to point after last schedulable inst. 2848 _bb_end++; 2849 } 2850 2851 assert( _bb_start <= _bb_end, "inverted block ends" ); 2852 2853 // Compute the register antidependencies for the basic block 2854 ComputeRegisterAntidependencies(bb); 2855 if (C->failing()) return; // too many D-U pinch points 2856 2857 // Compute the usage within the block, and set the list of all nodes 2858 // in the block that have no uses within the block. 2859 ComputeUseCount(bb); 2860 2861 // Schedule the remaining instructions in the block 2862 while ( _available.size() > 0 ) { 2863 Node *n = ChooseNodeToBundle(); 2864 guarantee(n != nullptr, "no nodes available"); 2865 AddNodeToBundle(n,bb); 2866 } 2867 2868 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2869 #ifdef ASSERT 2870 for( uint l = _bb_start; l < _bb_end; l++ ) { 2871 Node *n = bb->get_node(l); 2872 uint m; 2873 for( m = 0; m < _bb_end-_bb_start; m++ ) 2874 if( _scheduled[m] == n ) 2875 break; 2876 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2877 } 2878 #endif 2879 2880 // Now copy the instructions (in reverse order) back to the block 2881 for ( uint k = _bb_start; k < _bb_end; k++ ) 2882 bb->map_node(_scheduled[_bb_end-k-1], k); 2883 2884 #ifndef PRODUCT 2885 if (_cfg->C->trace_opto_output()) { 2886 tty->print("# Schedule BB#%03d (final)\n", i); 2887 uint current = 0; 2888 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2889 Node *n = bb->get_node(j); 2890 if( valid_bundle_info(n) ) { 2891 Bundle *bundle = node_bundling(n); 2892 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2893 tty->print("*** Bundle: "); 2894 bundle->dump(); 2895 } 2896 n->dump(); 2897 } 2898 } 2899 } 2900 #endif 2901 #ifdef ASSERT 2902 verify_good_schedule(bb,"after block local scheduling"); 2903 #endif 2904 } 2905 2906 #ifndef PRODUCT 2907 if (_cfg->C->trace_opto_output()) 2908 tty->print("# <- DoScheduling\n"); 2909 #endif 2910 2911 // Record final node-bundling array location 2912 _regalloc->C->output()->set_node_bundling_base(_node_bundling_base); 2913 2914 } // end DoScheduling 2915 2916 // Verify that no live-range used in the block is killed in the block by a 2917 // wrong DEF. This doesn't verify live-ranges that span blocks. 2918 2919 // Check for edge existence. Used to avoid adding redundant precedence edges. 2920 static bool edge_from_to( Node *from, Node *to ) { 2921 for( uint i=0; i<from->len(); i++ ) 2922 if( from->in(i) == to ) 2923 return true; 2924 return false; 2925 } 2926 2927 #ifdef ASSERT 2928 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2929 // Check for bad kills 2930 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2931 Node *prior_use = _reg_node[def]; 2932 if( prior_use && !edge_from_to(prior_use,n) ) { 2933 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2934 n->dump(); 2935 tty->print_cr("..."); 2936 prior_use->dump(); 2937 assert(edge_from_to(prior_use,n), "%s", msg); 2938 } 2939 _reg_node.map(def,nullptr); // Kill live USEs 2940 } 2941 } 2942 2943 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2944 2945 // Zap to something reasonable for the verify code 2946 _reg_node.clear(); 2947 2948 // Walk over the block backwards. Check to make sure each DEF doesn't 2949 // kill a live value (other than the one it's supposed to). Add each 2950 // USE to the live set. 2951 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) { 2952 Node *n = b->get_node(i); 2953 int n_op = n->Opcode(); 2954 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2955 // Fat-proj kills a slew of registers 2956 RegMaskIterator rmi(n->out_RegMask()); 2957 while (rmi.has_next()) { 2958 OptoReg::Name kill = rmi.next(); 2959 verify_do_def(n, kill, msg); 2960 } 2961 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2962 // Get DEF'd registers the normal way 2963 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2964 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2965 } 2966 2967 // Now make all USEs live 2968 for( uint i=1; i<n->req(); i++ ) { 2969 Node *def = n->in(i); 2970 assert(def != 0, "input edge required"); 2971 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2972 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2973 if( OptoReg::is_valid(reg_lo) ) { 2974 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg); 2975 _reg_node.map(reg_lo,n); 2976 } 2977 if( OptoReg::is_valid(reg_hi) ) { 2978 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg); 2979 _reg_node.map(reg_hi,n); 2980 } 2981 } 2982 2983 } 2984 2985 // Zap to something reasonable for the Antidependence code 2986 _reg_node.clear(); 2987 } 2988 #endif 2989 2990 // Conditionally add precedence edges. Avoid putting edges on Projs. 2991 static void add_prec_edge_from_to( Node *from, Node *to ) { 2992 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2993 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2994 from = from->in(0); 2995 } 2996 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2997 !edge_from_to( from, to ) ) // Avoid duplicate edge 2998 from->add_prec(to); 2999 } 3000 3001 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 3002 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 3003 return; 3004 3005 if (OptoReg::is_reg(def_reg)) { 3006 VMReg vmreg = OptoReg::as_VMReg(def_reg); 3007 if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) { 3008 // This is one of the high slots of a vector register. 3009 // ScheduleAndBundle already checked there are no live wide 3010 // vectors in this method so it can be safely ignored. 3011 return; 3012 } 3013 } 3014 3015 Node *pinch = _reg_node[def_reg]; // Get pinch point 3016 if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet? 3017 is_def ) { // Check for a true def (not a kill) 3018 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 3019 return; 3020 } 3021 3022 Node *kill = def; // Rename 'def' to more descriptive 'kill' 3023 debug_only( def = (Node*)((intptr_t)0xdeadbeef); ) 3024 3025 // After some number of kills there _may_ be a later def 3026 Node *later_def = nullptr; 3027 3028 Compile* C = Compile::current(); 3029 3030 // Finding a kill requires a real pinch-point. 3031 // Check for not already having a pinch-point. 3032 // Pinch points are Op_Node's. 3033 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 3034 later_def = pinch; // Must be def/kill as optimistic pinch-point 3035 if ( _pinch_free_list.size() > 0) { 3036 pinch = _pinch_free_list.pop(); 3037 } else { 3038 pinch = new Node(1); // Pinch point to-be 3039 } 3040 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 3041 DEBUG_ONLY( pinch->dump(); ); 3042 assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index()); 3043 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 3044 return; 3045 } 3046 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init) 3047 _reg_node.map(def_reg,pinch); // Record pinch-point 3048 //regalloc()->set_bad(pinch->_idx); // Already initialized this way. 3049 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 3050 pinch->init_req(0, C->top()); // set not null for the next call 3051 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 3052 later_def = nullptr; // and no later def 3053 } 3054 pinch->set_req(0,later_def); // Hook later def so we can find it 3055 } else { // Else have valid pinch point 3056 if( pinch->in(0) ) // If there is a later-def 3057 later_def = pinch->in(0); // Get it 3058 } 3059 3060 // Add output-dependence edge from later def to kill 3061 if( later_def ) // If there is some original def 3062 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 3063 3064 // See if current kill is also a use, and so is forced to be the pinch-point. 3065 if( pinch->Opcode() == Op_Node ) { 3066 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 3067 for( uint i=1; i<uses->req(); i++ ) { 3068 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 3069 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 3070 // Yes, found a use/kill pinch-point 3071 pinch->set_req(0,nullptr); // 3072 pinch->replace_by(kill); // Move anti-dep edges up 3073 pinch = kill; 3074 _reg_node.map(def_reg,pinch); 3075 return; 3076 } 3077 } 3078 } 3079 3080 // Add edge from kill to pinch-point 3081 add_prec_edge_from_to(kill,pinch); 3082 } 3083 3084 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 3085 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 3086 return; 3087 Node *pinch = _reg_node[use_reg]; // Get pinch point 3088 // Check for no later def_reg/kill in block 3089 if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b && 3090 // Use has to be block-local as well 3091 _cfg->get_block_for_node(use) == b) { 3092 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 3093 pinch->req() == 1 ) { // pinch not yet in block? 3094 pinch->del_req(0); // yank pointer to later-def, also set flag 3095 // Insert the pinch-point in the block just after the last use 3096 b->insert_node(pinch, b->find_node(use) + 1); 3097 _bb_end++; // Increase size scheduled region in block 3098 } 3099 3100 add_prec_edge_from_to(pinch,use); 3101 } 3102 } 3103 3104 // We insert antidependences between the reads and following write of 3105 // allocated registers to prevent illegal code motion. Hopefully, the 3106 // number of added references should be fairly small, especially as we 3107 // are only adding references within the current basic block. 3108 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 3109 3110 #ifdef ASSERT 3111 verify_good_schedule(b,"before block local scheduling"); 3112 #endif 3113 3114 // A valid schedule, for each register independently, is an endless cycle 3115 // of: a def, then some uses (connected to the def by true dependencies), 3116 // then some kills (defs with no uses), finally the cycle repeats with a new 3117 // def. The uses are allowed to float relative to each other, as are the 3118 // kills. No use is allowed to slide past a kill (or def). This requires 3119 // antidependencies between all uses of a single def and all kills that 3120 // follow, up to the next def. More edges are redundant, because later defs 3121 // & kills are already serialized with true or antidependencies. To keep 3122 // the edge count down, we add a 'pinch point' node if there's more than 3123 // one use or more than one kill/def. 3124 3125 // We add dependencies in one bottom-up pass. 3126 3127 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 3128 3129 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 3130 // register. If not, we record the DEF/KILL in _reg_node, the 3131 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 3132 // "pinch point", a new Node that's in the graph but not in the block. 3133 // We put edges from the prior and current DEF/KILLs to the pinch point. 3134 // We put the pinch point in _reg_node. If there's already a pinch point 3135 // we merely add an edge from the current DEF/KILL to the pinch point. 3136 3137 // After doing the DEF/KILLs, we handle USEs. For each used register, we 3138 // put an edge from the pinch point to the USE. 3139 3140 // To be expedient, the _reg_node array is pre-allocated for the whole 3141 // compilation. _reg_node is lazily initialized; it either contains a null, 3142 // or a valid def/kill/pinch-point, or a leftover node from some prior 3143 // block. Leftover node from some prior block is treated like a null (no 3144 // prior def, so no anti-dependence needed). Valid def is distinguished by 3145 // it being in the current block. 3146 bool fat_proj_seen = false; 3147 uint last_safept = _bb_end-1; 3148 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr; 3149 Node* last_safept_node = end_node; 3150 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 3151 Node *n = b->get_node(i); 3152 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 3153 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 3154 // Fat-proj kills a slew of registers 3155 // This can add edges to 'n' and obscure whether or not it was a def, 3156 // hence the is_def flag. 3157 fat_proj_seen = true; 3158 RegMaskIterator rmi(n->out_RegMask()); 3159 while (rmi.has_next()) { 3160 OptoReg::Name kill = rmi.next(); 3161 anti_do_def(b, n, kill, is_def); 3162 } 3163 } else { 3164 // Get DEF'd registers the normal way 3165 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 3166 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 3167 } 3168 3169 // Kill projections on a branch should appear to occur on the 3170 // branch, not afterwards, so grab the masks from the projections 3171 // and process them. 3172 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) { 3173 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3174 Node* use = n->fast_out(i); 3175 if (use->is_Proj()) { 3176 RegMaskIterator rmi(use->out_RegMask()); 3177 while (rmi.has_next()) { 3178 OptoReg::Name kill = rmi.next(); 3179 anti_do_def(b, n, kill, false); 3180 } 3181 } 3182 } 3183 } 3184 3185 // Check each register used by this instruction for a following DEF/KILL 3186 // that must occur afterward and requires an anti-dependence edge. 3187 for( uint j=0; j<n->req(); j++ ) { 3188 Node *def = n->in(j); 3189 if( def ) { 3190 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 3191 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 3192 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 3193 } 3194 } 3195 // Do not allow defs of new derived values to float above GC 3196 // points unless the base is definitely available at the GC point. 3197 3198 Node *m = b->get_node(i); 3199 3200 // Add precedence edge from following safepoint to use of derived pointer 3201 if( last_safept_node != end_node && 3202 m != last_safept_node) { 3203 for (uint k = 1; k < m->req(); k++) { 3204 const Type *t = m->in(k)->bottom_type(); 3205 if( t->isa_oop_ptr() && 3206 t->is_ptr()->offset() != 0 ) { 3207 last_safept_node->add_prec( m ); 3208 break; 3209 } 3210 } 3211 3212 // Do not allow a CheckCastPP node whose input is a raw pointer to 3213 // float past a safepoint. This can occur when a buffered inline 3214 // type is allocated in a loop and the CheckCastPP from that 3215 // allocation is reused outside the loop. If the use inside the 3216 // loop is scalarized the CheckCastPP will no longer be connected 3217 // to the loop safepoint. See JDK-8264340. 3218 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CheckCastPP) { 3219 Node *def = m->in(1); 3220 if (def != nullptr && def->bottom_type()->base() == Type::RawPtr) { 3221 last_safept_node->add_prec(m); 3222 } 3223 } 3224 } 3225 3226 if( n->jvms() ) { // Precedence edge from derived to safept 3227 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 3228 if( b->get_node(last_safept) != last_safept_node ) { 3229 last_safept = b->find_node(last_safept_node); 3230 } 3231 for( uint j=last_safept; j > i; j-- ) { 3232 Node *mach = b->get_node(j); 3233 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 3234 mach->add_prec( n ); 3235 } 3236 last_safept = i; 3237 last_safept_node = m; 3238 } 3239 } 3240 3241 if (fat_proj_seen) { 3242 // Garbage collect pinch nodes that were not consumed. 3243 // They are usually created by a fat kill MachProj for a call. 3244 garbage_collect_pinch_nodes(); 3245 } 3246 } 3247 3248 // Garbage collect pinch nodes for reuse by other blocks. 3249 // 3250 // The block scheduler's insertion of anti-dependence 3251 // edges creates many pinch nodes when the block contains 3252 // 2 or more Calls. A pinch node is used to prevent a 3253 // combinatorial explosion of edges. If a set of kills for a 3254 // register is anti-dependent on a set of uses (or defs), rather 3255 // than adding an edge in the graph between each pair of kill 3256 // and use (or def), a pinch is inserted between them: 3257 // 3258 // use1 use2 use3 3259 // \ | / 3260 // \ | / 3261 // pinch 3262 // / | \ 3263 // / | \ 3264 // kill1 kill2 kill3 3265 // 3266 // One pinch node is created per register killed when 3267 // the second call is encountered during a backwards pass 3268 // over the block. Most of these pinch nodes are never 3269 // wired into the graph because the register is never 3270 // used or def'ed in the block. 3271 // 3272 void Scheduling::garbage_collect_pinch_nodes() { 3273 #ifndef PRODUCT 3274 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 3275 #endif 3276 int trace_cnt = 0; 3277 for (uint k = 0; k < _reg_node.max(); k++) { 3278 Node* pinch = _reg_node[k]; 3279 if ((pinch != nullptr) && pinch->Opcode() == Op_Node && 3280 // no predecence input edges 3281 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) { 3282 cleanup_pinch(pinch); 3283 _pinch_free_list.push(pinch); 3284 _reg_node.map(k, nullptr); 3285 #ifndef PRODUCT 3286 if (_cfg->C->trace_opto_output()) { 3287 trace_cnt++; 3288 if (trace_cnt > 40) { 3289 tty->print("\n"); 3290 trace_cnt = 0; 3291 } 3292 tty->print(" %d", pinch->_idx); 3293 } 3294 #endif 3295 } 3296 } 3297 #ifndef PRODUCT 3298 if (_cfg->C->trace_opto_output()) tty->print("\n"); 3299 #endif 3300 } 3301 3302 // Clean up a pinch node for reuse. 3303 void Scheduling::cleanup_pinch( Node *pinch ) { 3304 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 3305 3306 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 3307 Node* use = pinch->last_out(i); 3308 uint uses_found = 0; 3309 for (uint j = use->req(); j < use->len(); j++) { 3310 if (use->in(j) == pinch) { 3311 use->rm_prec(j); 3312 uses_found++; 3313 } 3314 } 3315 assert(uses_found > 0, "must be a precedence edge"); 3316 i -= uses_found; // we deleted 1 or more copies of this edge 3317 } 3318 // May have a later_def entry 3319 pinch->set_req(0, nullptr); 3320 } 3321 3322 #ifndef PRODUCT 3323 3324 void Scheduling::dump_available() const { 3325 tty->print("#Availist "); 3326 for (uint i = 0; i < _available.size(); i++) 3327 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 3328 tty->cr(); 3329 } 3330 3331 // Print Scheduling Statistics 3332 void Scheduling::print_statistics() { 3333 // Print the size added by nops for bundling 3334 tty->print("Nops added %d bytes to total of %d bytes", 3335 _total_nop_size, _total_method_size); 3336 if (_total_method_size > 0) 3337 tty->print(", for %.2f%%", 3338 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 3339 tty->print("\n"); 3340 3341 // Print the number of branch shadows filled 3342 if (Pipeline::_branch_has_delay_slot) { 3343 tty->print("Of %d branches, %d had unconditional delay slots filled", 3344 _total_branches, _total_unconditional_delays); 3345 if (_total_branches > 0) 3346 tty->print(", for %.2f%%", 3347 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 3348 tty->print("\n"); 3349 } 3350 3351 uint total_instructions = 0, total_bundles = 0; 3352 3353 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 3354 uint bundle_count = _total_instructions_per_bundle[i]; 3355 total_instructions += bundle_count * i; 3356 total_bundles += bundle_count; 3357 } 3358 3359 if (total_bundles > 0) 3360 tty->print("Average ILP (excluding nops) is %.2f\n", 3361 ((double)total_instructions) / ((double)total_bundles)); 3362 } 3363 #endif 3364 3365 //-----------------------init_scratch_buffer_blob------------------------------ 3366 // Construct a temporary BufferBlob and cache it for this compile. 3367 void PhaseOutput::init_scratch_buffer_blob(int const_size) { 3368 // If there is already a scratch buffer blob allocated and the 3369 // constant section is big enough, use it. Otherwise free the 3370 // current and allocate a new one. 3371 BufferBlob* blob = scratch_buffer_blob(); 3372 if ((blob != nullptr) && (const_size <= _scratch_const_size)) { 3373 // Use the current blob. 3374 } else { 3375 if (blob != nullptr) { 3376 BufferBlob::free(blob); 3377 } 3378 3379 ResourceMark rm; 3380 _scratch_const_size = const_size; 3381 int size = C2Compiler::initial_code_buffer_size(const_size); 3382 if (C->has_scalarized_args()) { 3383 // Inline type entry points (MachVEPNodes) require lots of space for GC barriers and oop verification 3384 // when loading object fields from the buffered argument. Increase scratch buffer size accordingly. 3385 ciMethod* method = C->method(); 3386 int barrier_size = UseZGC ? 200 : (7 DEBUG_ONLY(+ 37)); 3387 int arg_num = 0; 3388 if (!method->is_static()) { 3389 if (method->is_scalarized_arg(arg_num)) { 3390 size += method->holder()->as_inline_klass()->oop_count() * barrier_size; 3391 } 3392 arg_num++; 3393 } 3394 for (ciSignatureStream str(method->signature()); !str.at_return_type(); str.next()) { 3395 if (method->is_scalarized_arg(arg_num)) { 3396 size += str.type()->as_inline_klass()->oop_count() * barrier_size; 3397 } 3398 arg_num++; 3399 } 3400 } 3401 blob = BufferBlob::create("Compile::scratch_buffer", size); 3402 // Record the buffer blob for next time. 3403 set_scratch_buffer_blob(blob); 3404 // Have we run out of code space? 3405 if (scratch_buffer_blob() == nullptr) { 3406 // Let CompilerBroker disable further compilations. 3407 C->record_failure("Not enough space for scratch buffer in CodeCache"); 3408 return; 3409 } 3410 } 3411 3412 // Initialize the relocation buffers 3413 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 3414 set_scratch_locs_memory(locs_buf); 3415 } 3416 3417 3418 //-----------------------scratch_emit_size------------------------------------- 3419 // Helper function that computes size by emitting code 3420 uint PhaseOutput::scratch_emit_size(const Node* n) { 3421 // Start scratch_emit_size section. 3422 set_in_scratch_emit_size(true); 3423 3424 // Emit into a trash buffer and count bytes emitted. 3425 // This is a pretty expensive way to compute a size, 3426 // but it works well enough if seldom used. 3427 // All common fixed-size instructions are given a size 3428 // method by the AD file. 3429 // Note that the scratch buffer blob and locs memory are 3430 // allocated at the beginning of the compile task, and 3431 // may be shared by several calls to scratch_emit_size. 3432 // The allocation of the scratch buffer blob is particularly 3433 // expensive, since it has to grab the code cache lock. 3434 BufferBlob* blob = this->scratch_buffer_blob(); 3435 assert(blob != nullptr, "Initialize BufferBlob at start"); 3436 assert(blob->size() > MAX_inst_size, "sanity"); 3437 relocInfo* locs_buf = scratch_locs_memory(); 3438 address blob_begin = blob->content_begin(); 3439 address blob_end = (address)locs_buf; 3440 assert(blob->contains(blob_end), "sanity"); 3441 CodeBuffer buf(blob_begin, blob_end - blob_begin); 3442 buf.initialize_consts_size(_scratch_const_size); 3443 buf.initialize_stubs_size(MAX_stubs_size); 3444 assert(locs_buf != nullptr, "sanity"); 3445 int lsize = MAX_locs_size / 3; 3446 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 3447 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 3448 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 3449 // Mark as scratch buffer. 3450 buf.consts()->set_scratch_emit(); 3451 buf.insts()->set_scratch_emit(); 3452 buf.stubs()->set_scratch_emit(); 3453 3454 // Do the emission. 3455 3456 Label fakeL; // Fake label for branch instructions. 3457 Label* saveL = nullptr; 3458 uint save_bnum = 0; 3459 bool is_branch = n->is_MachBranch(); 3460 if (is_branch) { 3461 MacroAssembler masm(&buf); 3462 masm.bind(fakeL); 3463 n->as_MachBranch()->save_label(&saveL, &save_bnum); 3464 n->as_MachBranch()->label_set(&fakeL, 0); 3465 } 3466 n->emit(buf, C->regalloc()); 3467 3468 // Emitting into the scratch buffer should not fail 3469 assert (!C->failing(), "Must not have pending failure. Reason is: %s", C->failure_reason()); 3470 3471 // Restore label. 3472 if (is_branch) { 3473 n->as_MachBranch()->label_set(saveL, save_bnum); 3474 } 3475 3476 // End scratch_emit_size section. 3477 set_in_scratch_emit_size(false); 3478 3479 return buf.insts_size(); 3480 } 3481 3482 void PhaseOutput::install() { 3483 if (!C->should_install_code()) { 3484 return; 3485 } else if (C->stub_function() != nullptr) { 3486 install_stub(C->stub_name()); 3487 } else { 3488 install_code(C->method(), 3489 C->entry_bci(), 3490 CompileBroker::compiler2(), 3491 C->has_unsafe_access(), 3492 SharedRuntime::is_wide_vector(C->max_vector_size()), 3493 C->rtm_state()); 3494 } 3495 } 3496 3497 void PhaseOutput::install_code(ciMethod* target, 3498 int entry_bci, 3499 AbstractCompiler* compiler, 3500 bool has_unsafe_access, 3501 bool has_wide_vectors, 3502 RTMState rtm_state) { 3503 // Check if we want to skip execution of all compiled code. 3504 { 3505 #ifndef PRODUCT 3506 if (OptoNoExecute) { 3507 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed 3508 return; 3509 } 3510 #endif 3511 Compile::TracePhase tp("install_code", &timers[_t_registerMethod]); 3512 3513 if (C->is_osr_compilation()) { 3514 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 3515 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 3516 } else { 3517 if (!target->is_static()) { 3518 // The UEP of an nmethod ensures that the VEP is padded. However, the padding of the UEP is placed 3519 // before the inline cache check, so we don't have to execute any nop instructions when dispatching 3520 // through the UEP, yet we can ensure that the VEP is aligned appropriately. 3521 // TODO 8325106 Check this 3522 // _code_offsets.set_value(CodeOffsets::Entry, _first_block_size - MacroAssembler::ic_check_size()); 3523 } 3524 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 3525 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry) == -1) { 3526 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, _first_block_size); 3527 } 3528 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry_RO) == -1) { 3529 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, _first_block_size); 3530 } 3531 if (_code_offsets.value(CodeOffsets::Entry) == -1) { 3532 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size); 3533 } 3534 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 3535 } 3536 3537 C->env()->register_method(target, 3538 entry_bci, 3539 &_code_offsets, 3540 _orig_pc_slot_offset_in_bytes, 3541 code_buffer(), 3542 frame_size_in_words(), 3543 _oop_map_set, 3544 &_handler_table, 3545 inc_table(), 3546 compiler, 3547 has_unsafe_access, 3548 SharedRuntime::is_wide_vector(C->max_vector_size()), 3549 C->has_monitors(), 3550 0, 3551 C->rtm_state()); 3552 3553 if (C->log() != nullptr) { // Print code cache state into compiler log 3554 C->log()->code_cache_state(); 3555 } 3556 } 3557 } 3558 void PhaseOutput::install_stub(const char* stub_name) { 3559 // Entry point will be accessed using stub_entry_point(); 3560 if (code_buffer() == nullptr) { 3561 Matcher::soft_match_failure(); 3562 } else { 3563 if (PrintAssembly && (WizardMode || Verbose)) 3564 tty->print_cr("### Stub::%s", stub_name); 3565 3566 if (!C->failing()) { 3567 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs"); 3568 3569 // Make the NMethod 3570 // For now we mark the frame as never safe for profile stackwalking 3571 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 3572 code_buffer(), 3573 CodeOffsets::frame_never_safe, 3574 // _code_offsets.value(CodeOffsets::Frame_Complete), 3575 frame_size_in_words(), 3576 oop_map_set(), 3577 false); 3578 assert(rs != nullptr && rs->is_runtime_stub(), "sanity check"); 3579 3580 C->set_stub_entry_point(rs->entry_point()); 3581 } 3582 } 3583 } 3584 3585 // Support for bundling info 3586 Bundle* PhaseOutput::node_bundling(const Node *n) { 3587 assert(valid_bundle_info(n), "oob"); 3588 return &_node_bundling_base[n->_idx]; 3589 } 3590 3591 bool PhaseOutput::valid_bundle_info(const Node *n) { 3592 return (_node_bundling_limit > n->_idx); 3593 } 3594 3595 //------------------------------frame_size_in_words----------------------------- 3596 // frame_slots in units of words 3597 int PhaseOutput::frame_size_in_words() const { 3598 // shift is 0 in LP32 and 1 in LP64 3599 const int shift = (LogBytesPerWord - LogBytesPerInt); 3600 int words = _frame_slots >> shift; 3601 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 3602 return words; 3603 } 3604 3605 // To bang the stack of this compiled method we use the stack size 3606 // that the interpreter would need in case of a deoptimization. This 3607 // removes the need to bang the stack in the deoptimization blob which 3608 // in turn simplifies stack overflow handling. 3609 int PhaseOutput::bang_size_in_bytes() const { 3610 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size()); 3611 } 3612 3613 //------------------------------dump_asm--------------------------------------- 3614 // Dump formatted assembly 3615 #if defined(SUPPORT_OPTO_ASSEMBLY) 3616 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) { 3617 3618 int pc_digits = 3; // #chars required for pc 3619 int sb_chars = 3; // #chars for "start bundle" indicator 3620 int tab_size = 8; 3621 if (pcs != nullptr) { 3622 int max_pc = 0; 3623 for (uint i = 0; i < pc_limit; i++) { 3624 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc; 3625 } 3626 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc 3627 } 3628 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size; 3629 3630 bool cut_short = false; 3631 st->print_cr("#"); 3632 st->print("# "); C->tf()->dump_on(st); st->cr(); 3633 st->print_cr("#"); 3634 3635 // For all blocks 3636 int pc = 0x0; // Program counter 3637 char starts_bundle = ' '; 3638 C->regalloc()->dump_frame(); 3639 3640 Node *n = nullptr; 3641 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) { 3642 if (VMThread::should_terminate()) { 3643 cut_short = true; 3644 break; 3645 } 3646 Block* block = C->cfg()->get_block(i); 3647 if (block->is_connector() && !Verbose) { 3648 continue; 3649 } 3650 n = block->head(); 3651 if ((pcs != nullptr) && (n->_idx < pc_limit)) { 3652 pc = pcs[n->_idx]; 3653 st->print("%*.*x", pc_digits, pc_digits, pc); 3654 } 3655 st->fill_to(prefix_len); 3656 block->dump_head(C->cfg(), st); 3657 if (block->is_connector()) { 3658 st->fill_to(prefix_len); 3659 st->print_cr("# Empty connector block"); 3660 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 3661 st->fill_to(prefix_len); 3662 st->print_cr("# Block is sole successor of call"); 3663 } 3664 3665 // For all instructions 3666 Node *delay = nullptr; 3667 for (uint j = 0; j < block->number_of_nodes(); j++) { 3668 if (VMThread::should_terminate()) { 3669 cut_short = true; 3670 break; 3671 } 3672 n = block->get_node(j); 3673 if (valid_bundle_info(n)) { 3674 Bundle* bundle = node_bundling(n); 3675 if (bundle->used_in_unconditional_delay()) { 3676 delay = n; 3677 continue; 3678 } 3679 if (bundle->starts_bundle()) { 3680 starts_bundle = '+'; 3681 } 3682 } 3683 3684 if (WizardMode) { 3685 n->dump(); 3686 } 3687 3688 if( !n->is_Region() && // Dont print in the Assembly 3689 !n->is_Phi() && // a few noisely useless nodes 3690 !n->is_Proj() && 3691 !n->is_MachTemp() && 3692 !n->is_SafePointScalarObject() && 3693 !n->is_Catch() && // Would be nice to print exception table targets 3694 !n->is_MergeMem() && // Not very interesting 3695 !n->is_top() && // Debug info table constants 3696 !(n->is_Con() && !n->is_Mach())// Debug info table constants 3697 ) { 3698 if ((pcs != nullptr) && (n->_idx < pc_limit)) { 3699 pc = pcs[n->_idx]; 3700 st->print("%*.*x", pc_digits, pc_digits, pc); 3701 } else { 3702 st->fill_to(pc_digits); 3703 } 3704 st->print(" %c ", starts_bundle); 3705 starts_bundle = ' '; 3706 st->fill_to(prefix_len); 3707 n->format(C->regalloc(), st); 3708 st->cr(); 3709 } 3710 3711 // If we have an instruction with a delay slot, and have seen a delay, 3712 // then back up and print it 3713 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 3714 // Coverity finding - Explicit null dereferenced. 3715 guarantee(delay != nullptr, "no unconditional delay instruction"); 3716 if (WizardMode) delay->dump(); 3717 3718 if (node_bundling(delay)->starts_bundle()) 3719 starts_bundle = '+'; 3720 if ((pcs != nullptr) && (n->_idx < pc_limit)) { 3721 pc = pcs[n->_idx]; 3722 st->print("%*.*x", pc_digits, pc_digits, pc); 3723 } else { 3724 st->fill_to(pc_digits); 3725 } 3726 st->print(" %c ", starts_bundle); 3727 starts_bundle = ' '; 3728 st->fill_to(prefix_len); 3729 delay->format(C->regalloc(), st); 3730 st->cr(); 3731 delay = nullptr; 3732 } 3733 3734 // Dump the exception table as well 3735 if( n->is_Catch() && (Verbose || WizardMode) ) { 3736 // Print the exception table for this offset 3737 _handler_table.print_subtable_for(pc); 3738 } 3739 st->bol(); // Make sure we start on a new line 3740 } 3741 st->cr(); // one empty line between blocks 3742 assert(cut_short || delay == nullptr, "no unconditional delay branch"); 3743 } // End of per-block dump 3744 3745 if (cut_short) st->print_cr("*** disassembly is cut short ***"); 3746 } 3747 #endif 3748 3749 #ifndef PRODUCT 3750 void PhaseOutput::print_statistics() { 3751 Scheduling::print_statistics(); 3752 } 3753 #endif