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/c2/barrierSetC2.hpp" 37 #include "memory/allocation.inline.hpp" 38 #include "memory/allocation.hpp" 39 #include "opto/ad.hpp" 40 #include "opto/block.hpp" 41 #include "opto/c2compiler.hpp" 42 #include "opto/c2_MacroAssembler.hpp" 43 #include "opto/callnode.hpp" 44 #include "opto/cfgnode.hpp" 45 #include "opto/locknode.hpp" 46 #include "opto/machnode.hpp" 47 #include "opto/node.hpp" 48 #include "opto/optoreg.hpp" 49 #include "opto/output.hpp" 50 #include "opto/regalloc.hpp" 51 #include "opto/runtime.hpp" 52 #include "opto/subnode.hpp" 53 #include "opto/type.hpp" 54 #include "runtime/handles.inline.hpp" 55 #include "runtime/sharedRuntime.hpp" 56 #include "utilities/macros.hpp" 57 #include "utilities/powerOfTwo.hpp" 58 #include "utilities/xmlstream.hpp" 59 60 #ifndef PRODUCT 61 #define DEBUG_ARG(x) , x 62 #else 63 #define DEBUG_ARG(x) 64 #endif 65 66 //------------------------------Scheduling---------------------------------- 67 // This class contains all the information necessary to implement instruction 68 // scheduling and bundling. 69 class Scheduling { 70 71 private: 72 // Arena to use 73 Arena *_arena; 74 75 // Control-Flow Graph info 76 PhaseCFG *_cfg; 77 78 // Register Allocation info 79 PhaseRegAlloc *_regalloc; 80 81 // Number of nodes in the method 82 uint _node_bundling_limit; 83 84 // List of scheduled nodes. Generated in reverse order 85 Node_List _scheduled; 86 87 // List of nodes currently available for choosing for scheduling 88 Node_List _available; 89 90 // For each instruction beginning a bundle, the number of following 91 // nodes to be bundled with it. 92 Bundle *_node_bundling_base; 93 94 // Mapping from register to Node 95 Node_List _reg_node; 96 97 // Free list for pinch nodes. 98 Node_List _pinch_free_list; 99 100 // Number of uses of this node within the containing basic block. 101 short *_uses; 102 103 // Schedulable portion of current block. Skips Region/Phi/CreateEx up 104 // front, branch+proj at end. Also skips Catch/CProj (same as 105 // branch-at-end), plus just-prior exception-throwing call. 106 uint _bb_start, _bb_end; 107 108 // Latency from the end of the basic block as scheduled 109 unsigned short *_current_latency; 110 111 // Remember the next node 112 Node *_next_node; 113 114 // Use this for an unconditional branch delay slot 115 Node *_unconditional_delay_slot; 116 117 // Pointer to a Nop 118 MachNopNode *_nop; 119 120 // Length of the current bundle, in instructions 121 uint _bundle_instr_count; 122 123 // Current Cycle number, for computing latencies and bundling 124 uint _bundle_cycle_number; 125 126 // Bundle information 127 Pipeline_Use_Element _bundle_use_elements[resource_count]; 128 Pipeline_Use _bundle_use; 129 130 // Dump the available list 131 void dump_available() const; 132 133 public: 134 Scheduling(Arena *arena, Compile &compile); 135 136 // Destructor 137 NOT_PRODUCT( ~Scheduling(); ) 138 139 // Step ahead "i" cycles 140 void step(uint i); 141 142 // Step ahead 1 cycle, and clear the bundle state (for example, 143 // at a branch target) 144 void step_and_clear(); 145 146 Bundle* node_bundling(const Node *n) { 147 assert(valid_bundle_info(n), "oob"); 148 return (&_node_bundling_base[n->_idx]); 149 } 150 151 bool valid_bundle_info(const Node *n) const { 152 return (_node_bundling_limit > n->_idx); 153 } 154 155 bool starts_bundle(const Node *n) const { 156 return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle()); 157 } 158 159 // Do the scheduling 160 void DoScheduling(); 161 162 // Compute the register antidependencies within a basic block 163 void ComputeRegisterAntidependencies(Block *bb); 164 void verify_do_def( Node *n, OptoReg::Name def, const char *msg ); 165 void verify_good_schedule( Block *b, const char *msg ); 166 void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ); 167 void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ); 168 169 // Add a node to the current bundle 170 void AddNodeToBundle(Node *n, const Block *bb); 171 172 // Return an integer less than, equal to, or greater than zero 173 // if the stack offset of the first argument is respectively 174 // less than, equal to, or greater than the second. 175 int compare_two_spill_nodes(Node* first, Node* second); 176 177 // Add a node to the list of available nodes 178 void AddNodeToAvailableList(Node *n); 179 180 // Compute the local use count for the nodes in a block, and compute 181 // the list of instructions with no uses in the block as available 182 void ComputeUseCount(const Block *bb); 183 184 // Choose an instruction from the available list to add to the bundle 185 Node * ChooseNodeToBundle(); 186 187 // See if this Node fits into the currently accumulating bundle 188 bool NodeFitsInBundle(Node *n); 189 190 // Decrement the use count for a node 191 void DecrementUseCounts(Node *n, const Block *bb); 192 193 // Garbage collect pinch nodes for reuse by other blocks. 194 void garbage_collect_pinch_nodes(); 195 // Clean up a pinch node for reuse (helper for above). 196 void cleanup_pinch( Node *pinch ); 197 198 // Information for statistics gathering 199 #ifndef PRODUCT 200 private: 201 // Gather information on size of nops relative to total 202 uint _branches, _unconditional_delays; 203 204 static uint _total_nop_size, _total_method_size; 205 static uint _total_branches, _total_unconditional_delays; 206 static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 207 208 public: 209 static void print_statistics(); 210 211 static void increment_instructions_per_bundle(uint i) { 212 _total_instructions_per_bundle[i]++; 213 } 214 215 static void increment_nop_size(uint s) { 216 _total_nop_size += s; 217 } 218 219 static void increment_method_size(uint s) { 220 _total_method_size += s; 221 } 222 #endif 223 224 }; 225 226 PhaseOutput::PhaseOutput() 227 : Phase(Phase::Output), 228 _code_buffer("Compile::Fill_buffer"), 229 _first_block_size(0), 230 _handler_table(), 231 _inc_table(), 232 _stub_list(), 233 _oop_map_set(nullptr), 234 _scratch_buffer_blob(nullptr), 235 _scratch_locs_memory(nullptr), 236 _scratch_const_size(-1), 237 _in_scratch_emit_size(false), 238 _frame_slots(0), 239 _code_offsets(), 240 _node_bundling_limit(0), 241 _node_bundling_base(nullptr), 242 _orig_pc_slot(0), 243 _orig_pc_slot_offset_in_bytes(0), 244 _buf_sizes(), 245 _block(nullptr), 246 _index(0) { 247 C->set_output(this); 248 if (C->stub_name() == nullptr) { 249 _orig_pc_slot = C->fixed_slots() - (sizeof(address) / VMRegImpl::stack_slot_size); 250 } 251 } 252 253 PhaseOutput::~PhaseOutput() { 254 C->set_output(nullptr); 255 if (_scratch_buffer_blob != nullptr) { 256 BufferBlob::free(_scratch_buffer_blob); 257 } 258 } 259 260 void PhaseOutput::perform_mach_node_analysis() { 261 // Late barrier analysis must be done after schedule and bundle 262 // Otherwise liveness based spilling will fail 263 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 264 bs->late_barrier_analysis(); 265 266 pd_perform_mach_node_analysis(); 267 268 C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3); 269 } 270 271 // Convert Nodes to instruction bits and pass off to the VM 272 void PhaseOutput::Output() { 273 // RootNode goes 274 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" ); 275 276 // The number of new nodes (mostly MachNop) is proportional to 277 // the number of java calls and inner loops which are aligned. 278 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 279 C->inner_loops()*(OptoLoopAlignment-1)), 280 "out of nodes before code generation" ) ) { 281 return; 282 } 283 // Make sure I can find the Start Node 284 Block *entry = C->cfg()->get_block(1); 285 Block *broot = C->cfg()->get_root_block(); 286 287 const StartNode *start = entry->head()->as_Start(); 288 289 // Replace StartNode with prolog 290 MachPrologNode *prolog = new MachPrologNode(); 291 entry->map_node(prolog, 0); 292 C->cfg()->map_node_to_block(prolog, entry); 293 C->cfg()->unmap_node_from_block(start); // start is no longer in any block 294 295 // Virtual methods need an unverified entry point 296 297 if( C->is_osr_compilation() ) { 298 if( PoisonOSREntry ) { 299 // TODO: Should use a ShouldNotReachHereNode... 300 C->cfg()->insert( broot, 0, new MachBreakpointNode() ); 301 } 302 } else { 303 if( C->method() && !C->method()->flags().is_static() ) { 304 // Insert unvalidated entry point 305 C->cfg()->insert( broot, 0, new MachUEPNode() ); 306 } 307 308 } 309 310 // Break before main entry point 311 if ((C->method() && C->directive()->BreakAtExecuteOption) || 312 (OptoBreakpoint && C->is_method_compilation()) || 313 (OptoBreakpointOSR && C->is_osr_compilation()) || 314 (OptoBreakpointC2R && !C->method()) ) { 315 // checking for C->method() means that OptoBreakpoint does not apply to 316 // runtime stubs or frame converters 317 C->cfg()->insert( entry, 1, new MachBreakpointNode() ); 318 } 319 320 // Insert epilogs before every return 321 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) { 322 Block* block = C->cfg()->get_block(i); 323 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point? 324 Node* m = block->end(); 325 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) { 326 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 327 block->add_inst(epilog); 328 C->cfg()->map_node_to_block(epilog, block); 329 } 330 } 331 } 332 333 // Keeper of sizing aspects 334 _buf_sizes = BufferSizingData(); 335 336 // Initialize code buffer 337 estimate_buffer_size(_buf_sizes._const); 338 if (C->failing()) return; 339 340 // Pre-compute the length of blocks and replace 341 // long branches with short if machine supports it. 342 // Must be done before ScheduleAndBundle due to SPARC delay slots 343 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1); 344 blk_starts[0] = 0; 345 shorten_branches(blk_starts); 346 347 ScheduleAndBundle(); 348 if (C->failing()) { 349 return; 350 } 351 352 perform_mach_node_analysis(); 353 354 // Complete sizing of codebuffer 355 CodeBuffer* cb = init_buffer(); 356 if (cb == nullptr || C->failing()) { 357 return; 358 } 359 360 BuildOopMaps(); 361 362 if (C->failing()) { 363 return; 364 } 365 366 C2_MacroAssembler masm(cb); 367 fill_buffer(&masm, blk_starts); 368 } 369 370 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const { 371 // Determine if we need to generate a stack overflow check. 372 // Do it if the method is not a stub function and 373 // has java calls or has frame size > vm_page_size/8. 374 // The debug VM checks that deoptimization doesn't trigger an 375 // unexpected stack overflow (compiled method stack banging should 376 // guarantee it doesn't happen) so we always need the stack bang in 377 // a debug VM. 378 return (C->stub_function() == nullptr && 379 (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3 380 DEBUG_ONLY(|| true))); 381 } 382 383 bool PhaseOutput::need_register_stack_bang() const { 384 // Determine if we need to generate a register stack overflow check. 385 // This is only used on architectures which have split register 386 // and memory stacks (ie. IA64). 387 // Bang if the method is not a stub function and has java calls 388 return (C->stub_function() == nullptr && C->has_java_calls()); 389 } 390 391 392 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 393 // of a loop. When aligning a loop we need to provide enough instructions 394 // in cpu's fetch buffer to feed decoders. The loop alignment could be 395 // avoided if we have enough instructions in fetch buffer at the head of a loop. 396 // By default, the size is set to 999999 by Block's constructor so that 397 // a loop will be aligned if the size is not reset here. 398 // 399 // Note: Mach instructions could contain several HW instructions 400 // so the size is estimated only. 401 // 402 void PhaseOutput::compute_loop_first_inst_sizes() { 403 // The next condition is used to gate the loop alignment optimization. 404 // Don't aligned a loop if there are enough instructions at the head of a loop 405 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 406 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 407 // equal to 11 bytes which is the largest address NOP instruction. 408 if (MaxLoopPad < OptoLoopAlignment - 1) { 409 uint last_block = C->cfg()->number_of_blocks() - 1; 410 for (uint i = 1; i <= last_block; i++) { 411 Block* block = C->cfg()->get_block(i); 412 // Check the first loop's block which requires an alignment. 413 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) { 414 uint sum_size = 0; 415 uint inst_cnt = NumberOfLoopInstrToAlign; 416 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc()); 417 418 // Check subsequent fallthrough blocks if the loop's first 419 // block(s) does not have enough instructions. 420 Block *nb = block; 421 while(inst_cnt > 0 && 422 i < last_block && 423 !C->cfg()->get_block(i + 1)->has_loop_alignment() && 424 !nb->has_successor(block)) { 425 i++; 426 nb = C->cfg()->get_block(i); 427 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc()); 428 } // while( inst_cnt > 0 && i < last_block ) 429 430 block->set_first_inst_size(sum_size); 431 } // f( b->head()->is_Loop() ) 432 } // for( i <= last_block ) 433 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 434 } 435 436 // The architecture description provides short branch variants for some long 437 // branch instructions. Replace eligible long branches with short branches. 438 void PhaseOutput::shorten_branches(uint* blk_starts) { 439 440 Compile::TracePhase tp("shorten branches", &timers[_t_shortenBranches]); 441 442 // Compute size of each block, method size, and relocation information size 443 uint nblocks = C->cfg()->number_of_blocks(); 444 445 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 446 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 447 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks); 448 449 // Collect worst case block paddings 450 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks); 451 memset(block_worst_case_pad, 0, nblocks * sizeof(int)); 452 453 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 454 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 455 456 bool has_short_branch_candidate = false; 457 458 // Initialize the sizes to 0 459 int code_size = 0; // Size in bytes of generated code 460 int stub_size = 0; // Size in bytes of all stub entries 461 // Size in bytes of all relocation entries, including those in local stubs. 462 // Start with 2-bytes of reloc info for the unvalidated entry point 463 int reloc_size = 1; // Number of relocation entries 464 465 // Make three passes. The first computes pessimistic blk_starts, 466 // relative jmp_offset and reloc_size information. The second performs 467 // short branch substitution using the pessimistic sizing. The 468 // third inserts nops where needed. 469 470 // Step one, perform a pessimistic sizing pass. 471 uint last_call_adr = max_juint; 472 uint last_avoid_back_to_back_adr = max_juint; 473 uint nop_size = (new MachNopNode())->size(C->regalloc()); 474 for (uint i = 0; i < nblocks; i++) { // For all blocks 475 Block* block = C->cfg()->get_block(i); 476 _block = block; 477 478 // During short branch replacement, we store the relative (to blk_starts) 479 // offset of jump in jmp_offset, rather than the absolute offset of jump. 480 // This is so that we do not need to recompute sizes of all nodes when 481 // we compute correct blk_starts in our next sizing pass. 482 jmp_offset[i] = 0; 483 jmp_size[i] = 0; 484 jmp_nidx[i] = -1; 485 DEBUG_ONLY( jmp_target[i] = 0; ) 486 DEBUG_ONLY( jmp_rule[i] = 0; ) 487 488 // Sum all instruction sizes to compute block size 489 uint last_inst = block->number_of_nodes(); 490 uint blk_size = 0; 491 for (uint j = 0; j < last_inst; j++) { 492 _index = j; 493 Node* nj = block->get_node(_index); 494 // Handle machine instruction nodes 495 if (nj->is_Mach()) { 496 MachNode* mach = nj->as_Mach(); 497 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 498 reloc_size += mach->reloc(); 499 if (mach->is_MachCall()) { 500 // add size information for trampoline stub 501 // class CallStubImpl is platform-specific and defined in the *.ad files. 502 stub_size += CallStubImpl::size_call_trampoline(); 503 reloc_size += CallStubImpl::reloc_call_trampoline(); 504 505 MachCallNode *mcall = mach->as_MachCall(); 506 // This destination address is NOT PC-relative 507 508 mcall->method_set((intptr_t)mcall->entry_point()); 509 510 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) { 511 stub_size += CompiledDirectCall::to_interp_stub_size(); 512 reloc_size += CompiledDirectCall::reloc_to_interp_stub(); 513 } 514 } else if (mach->is_MachSafePoint()) { 515 // If call/safepoint are adjacent, account for possible 516 // nop to disambiguate the two safepoints. 517 // ScheduleAndBundle() can rearrange nodes in a block, 518 // check for all offsets inside this block. 519 if (last_call_adr >= blk_starts[i]) { 520 blk_size += nop_size; 521 } 522 } 523 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 524 // Nop is inserted between "avoid back to back" instructions. 525 // ScheduleAndBundle() can rearrange nodes in a block, 526 // check for all offsets inside this block. 527 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 528 blk_size += nop_size; 529 } 530 } 531 if (mach->may_be_short_branch()) { 532 if (!nj->is_MachBranch()) { 533 #ifndef PRODUCT 534 nj->dump(3); 535 #endif 536 Unimplemented(); 537 } 538 assert(jmp_nidx[i] == -1, "block should have only one branch"); 539 jmp_offset[i] = blk_size; 540 jmp_size[i] = nj->size(C->regalloc()); 541 jmp_nidx[i] = j; 542 has_short_branch_candidate = true; 543 } 544 } 545 blk_size += nj->size(C->regalloc()); 546 // Remember end of call offset 547 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 548 last_call_adr = blk_starts[i]+blk_size; 549 } 550 // Remember end of avoid_back_to_back offset 551 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 552 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 553 } 554 } 555 556 // When the next block starts a loop, we may insert pad NOP 557 // instructions. Since we cannot know our future alignment, 558 // assume the worst. 559 if (i < nblocks - 1) { 560 Block* nb = C->cfg()->get_block(i + 1); 561 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 562 if (max_loop_pad > 0) { 563 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 564 // Adjust last_call_adr and/or last_avoid_back_to_back_adr. 565 // If either is the last instruction in this block, bump by 566 // max_loop_pad in lock-step with blk_size, so sizing 567 // calculations in subsequent blocks still can conservatively 568 // detect that it may the last instruction in this block. 569 if (last_call_adr == blk_starts[i]+blk_size) { 570 last_call_adr += max_loop_pad; 571 } 572 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) { 573 last_avoid_back_to_back_adr += max_loop_pad; 574 } 575 blk_size += max_loop_pad; 576 block_worst_case_pad[i + 1] = max_loop_pad; 577 } 578 } 579 580 // Save block size; update total method size 581 blk_starts[i+1] = blk_starts[i]+blk_size; 582 } 583 584 // Step two, replace eligible long jumps. 585 bool progress = true; 586 uint last_may_be_short_branch_adr = max_juint; 587 while (has_short_branch_candidate && progress) { 588 progress = false; 589 has_short_branch_candidate = false; 590 int adjust_block_start = 0; 591 for (uint i = 0; i < nblocks; i++) { 592 Block* block = C->cfg()->get_block(i); 593 int idx = jmp_nidx[i]; 594 MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach(); 595 if (mach != nullptr && mach->may_be_short_branch()) { 596 #ifdef ASSERT 597 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 598 int j; 599 // Find the branch; ignore trailing NOPs. 600 for (j = block->number_of_nodes()-1; j>=0; j--) { 601 Node* n = block->get_node(j); 602 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 603 break; 604 } 605 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity"); 606 #endif 607 int br_size = jmp_size[i]; 608 int br_offs = blk_starts[i] + jmp_offset[i]; 609 610 // This requires the TRUE branch target be in succs[0] 611 uint bnum = block->non_connector_successor(0)->_pre_order; 612 int offset = blk_starts[bnum] - br_offs; 613 if (bnum > i) { // adjust following block's offset 614 offset -= adjust_block_start; 615 } 616 617 // This block can be a loop header, account for the padding 618 // in the previous block. 619 int block_padding = block_worst_case_pad[i]; 620 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top"); 621 // In the following code a nop could be inserted before 622 // the branch which will increase the backward distance. 623 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr); 624 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block"); 625 626 if (needs_padding && offset <= 0) 627 offset -= nop_size; 628 629 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) { 630 // We've got a winner. Replace this branch. 631 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 632 633 // Update the jmp_size. 634 int new_size = replacement->size(C->regalloc()); 635 int diff = br_size - new_size; 636 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 637 // Conservatively take into account padding between 638 // avoid_back_to_back branches. Previous branch could be 639 // converted into avoid_back_to_back branch during next 640 // rounds. 641 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 642 jmp_offset[i] += nop_size; 643 diff -= nop_size; 644 } 645 adjust_block_start += diff; 646 block->map_node(replacement, idx); 647 mach->subsume_by(replacement, C); 648 mach = replacement; 649 progress = true; 650 651 jmp_size[i] = new_size; 652 DEBUG_ONLY( jmp_target[i] = bnum; ); 653 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 654 } else { 655 // The jump distance is not short, try again during next iteration. 656 has_short_branch_candidate = true; 657 } 658 } // (mach->may_be_short_branch()) 659 if (mach != nullptr && (mach->may_be_short_branch() || 660 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) { 661 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 662 } 663 blk_starts[i+1] -= adjust_block_start; 664 } 665 } 666 667 #ifdef ASSERT 668 for (uint i = 0; i < nblocks; i++) { // For all blocks 669 if (jmp_target[i] != 0) { 670 int br_size = jmp_size[i]; 671 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 672 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 673 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]); 674 } 675 assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 676 } 677 } 678 #endif 679 680 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 681 // after ScheduleAndBundle(). 682 683 // ------------------ 684 // Compute size for code buffer 685 code_size = blk_starts[nblocks]; 686 687 // Relocation records 688 reloc_size += 1; // Relo entry for exception handler 689 690 // Adjust reloc_size to number of record of relocation info 691 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 692 // a relocation index. 693 // The CodeBuffer will expand the locs array if this estimate is too low. 694 reloc_size *= 10 / sizeof(relocInfo); 695 696 _buf_sizes._reloc = reloc_size; 697 _buf_sizes._code = code_size; 698 _buf_sizes._stub = stub_size; 699 } 700 701 //------------------------------FillLocArray----------------------------------- 702 // Create a bit of debug info and append it to the array. The mapping is from 703 // Java local or expression stack to constant, register or stack-slot. For 704 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 705 // entry has been taken care of and caller should skip it). 706 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 707 // This should never have accepted Bad before 708 assert(OptoReg::is_valid(regnum), "location must be valid"); 709 return (OptoReg::is_reg(regnum)) 710 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 711 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 712 } 713 714 715 ObjectValue* 716 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 717 for (int i = 0; i < objs->length(); i++) { 718 assert(objs->at(i)->is_object(), "corrupt object cache"); 719 ObjectValue* sv = (ObjectValue*) objs->at(i); 720 if (sv->id() == id) { 721 return sv; 722 } 723 } 724 // Otherwise.. 725 return nullptr; 726 } 727 728 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 729 ObjectValue* sv ) { 730 assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition"); 731 objs->append(sv); 732 } 733 734 735 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 736 GrowableArray<ScopeValue*> *array, 737 GrowableArray<ScopeValue*> *objs ) { 738 assert( local, "use _top instead of null" ); 739 if (array->length() != idx) { 740 assert(array->length() == idx + 1, "Unexpected array count"); 741 // Old functionality: 742 // return 743 // New functionality: 744 // Assert if the local is not top. In product mode let the new node 745 // override the old entry. 746 assert(local == C->top(), "LocArray collision"); 747 if (local == C->top()) { 748 return; 749 } 750 array->pop(); 751 } 752 const Type *t = local->bottom_type(); 753 754 // Is it a safepoint scalar object node? 755 if (local->is_SafePointScalarObject()) { 756 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 757 758 ObjectValue* sv = (ObjectValue*) sv_for_node_id(objs, spobj->_idx); 759 if (sv == nullptr) { 760 ciKlass* cik = t->is_oopptr()->exact_klass(); 761 assert(cik->is_instance_klass() || 762 cik->is_array_klass(), "Not supported allocation."); 763 sv = new ObjectValue(spobj->_idx, 764 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 765 set_sv_for_object_node(objs, sv); 766 767 uint first_ind = spobj->first_index(sfpt->jvms()); 768 for (uint i = 0; i < spobj->n_fields(); i++) { 769 Node* fld_node = sfpt->in(first_ind+i); 770 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 771 } 772 } 773 array->append(sv); 774 return; 775 } else if (local->is_SafePointScalarMerge()) { 776 SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge(); 777 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx); 778 779 if (mv == nullptr) { 780 GrowableArray<ScopeValue*> deps; 781 782 int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms()); 783 (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs); 784 assert(deps.length() == 1, "missing value"); 785 786 int selector_idx = smerge->selector_idx(sfpt->jvms()); 787 (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr); 788 assert(deps.length() == 2, "missing value"); 789 790 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1)); 791 set_sv_for_object_node(objs, mv); 792 793 for (uint i = 1; i < smerge->req(); i++) { 794 Node* obj_node = smerge->in(i); 795 (void)FillLocArray(mv->possible_objects()->length(), sfpt, obj_node, mv->possible_objects(), objs); 796 } 797 } 798 array->append(mv); 799 return; 800 } 801 802 // Grab the register number for the local 803 OptoReg::Name regnum = C->regalloc()->get_reg_first(local); 804 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 805 // Record the double as two float registers. 806 // The register mask for such a value always specifies two adjacent 807 // float registers, with the lower register number even. 808 // Normally, the allocation of high and low words to these registers 809 // is irrelevant, because nearly all operations on register pairs 810 // (e.g., StoreD) treat them as a single unit. 811 // Here, we assume in addition that the words in these two registers 812 // stored "naturally" (by operations like StoreD and double stores 813 // within the interpreter) such that the lower-numbered register 814 // is written to the lower memory address. This may seem like 815 // a machine dependency, but it is not--it is a requirement on 816 // the author of the <arch>.ad file to ensure that, for every 817 // even/odd double-register pair to which a double may be allocated, 818 // the word in the even single-register is stored to the first 819 // memory word. (Note that register numbers are completely 820 // arbitrary, and are not tied to any machine-level encodings.) 821 #ifdef _LP64 822 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 823 array->append(new ConstantIntValue((jint)0)); 824 array->append(new_loc_value( C->regalloc(), regnum, Location::dbl )); 825 } else if ( t->base() == Type::Long ) { 826 array->append(new ConstantIntValue((jint)0)); 827 array->append(new_loc_value( C->regalloc(), regnum, Location::lng )); 828 } else if ( t->base() == Type::RawPtr ) { 829 // jsr/ret return address which must be restored into the full 830 // width 64-bit stack slot. 831 array->append(new_loc_value( C->regalloc(), regnum, Location::lng )); 832 } 833 #else //_LP64 834 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 835 // Repack the double/long as two jints. 836 // The convention the interpreter uses is that the second local 837 // holds the first raw word of the native double representation. 838 // This is actually reasonable, since locals and stack arrays 839 // grow downwards in all implementations. 840 // (If, on some machine, the interpreter's Java locals or stack 841 // were to grow upwards, the embedded doubles would be word-swapped.) 842 array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal )); 843 array->append(new_loc_value( C->regalloc(), regnum , Location::normal )); 844 } 845 #endif //_LP64 846 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 847 OptoReg::is_reg(regnum) ) { 848 array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double() 849 ? Location::float_in_dbl : Location::normal )); 850 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 851 array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long 852 ? Location::int_in_long : Location::normal )); 853 } else if( t->base() == Type::NarrowOop ) { 854 array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop )); 855 } else if (t->base() == Type::VectorA || t->base() == Type::VectorS || 856 t->base() == Type::VectorD || t->base() == Type::VectorX || 857 t->base() == Type::VectorY || t->base() == Type::VectorZ) { 858 array->append(new_loc_value( C->regalloc(), regnum, Location::vector )); 859 } else if (C->regalloc()->is_oop(local)) { 860 assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr || 861 t->base() == Type::AryPtr, 862 "Unexpected type: %s", t->msg()); 863 array->append(new_loc_value( C->regalloc(), regnum, Location::oop )); 864 } else { 865 assert(t->base() == Type::Int || t->base() == Type::Half || 866 t->base() == Type::FloatCon || t->base() == Type::FloatBot, 867 "Unexpected type: %s", t->msg()); 868 array->append(new_loc_value( C->regalloc(), regnum, Location::normal )); 869 } 870 return; 871 } 872 873 // No register. It must be constant data. 874 switch (t->base()) { 875 case Type::Half: // Second half of a double 876 ShouldNotReachHere(); // Caller should skip 2nd halves 877 break; 878 case Type::AnyPtr: 879 array->append(new ConstantOopWriteValue(nullptr)); 880 break; 881 case Type::AryPtr: 882 case Type::InstPtr: // fall through 883 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 884 break; 885 case Type::NarrowOop: 886 if (t == TypeNarrowOop::NULL_PTR) { 887 array->append(new ConstantOopWriteValue(nullptr)); 888 } else { 889 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 890 } 891 break; 892 case Type::Int: 893 array->append(new ConstantIntValue(t->is_int()->get_con())); 894 break; 895 case Type::RawPtr: 896 // A return address (T_ADDRESS). 897 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 898 #ifdef _LP64 899 // Must be restored to the full-width 64-bit stack slot. 900 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 901 #else 902 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 903 #endif 904 break; 905 case Type::FloatCon: { 906 float f = t->is_float_constant()->getf(); 907 array->append(new ConstantIntValue(jint_cast(f))); 908 break; 909 } 910 case Type::DoubleCon: { 911 jdouble d = t->is_double_constant()->getd(); 912 #ifdef _LP64 913 array->append(new ConstantIntValue((jint)0)); 914 array->append(new ConstantDoubleValue(d)); 915 #else 916 // Repack the double as two jints. 917 // The convention the interpreter uses is that the second local 918 // holds the first raw word of the native double representation. 919 // This is actually reasonable, since locals and stack arrays 920 // grow downwards in all implementations. 921 // (If, on some machine, the interpreter's Java locals or stack 922 // were to grow upwards, the embedded doubles would be word-swapped.) 923 jlong_accessor acc; 924 acc.long_value = jlong_cast(d); 925 array->append(new ConstantIntValue(acc.words[1])); 926 array->append(new ConstantIntValue(acc.words[0])); 927 #endif 928 break; 929 } 930 case Type::Long: { 931 jlong d = t->is_long()->get_con(); 932 #ifdef _LP64 933 array->append(new ConstantIntValue((jint)0)); 934 array->append(new ConstantLongValue(d)); 935 #else 936 // Repack the long as two jints. 937 // The convention the interpreter uses is that the second local 938 // holds the first raw word of the native double representation. 939 // This is actually reasonable, since locals and stack arrays 940 // grow downwards in all implementations. 941 // (If, on some machine, the interpreter's Java locals or stack 942 // were to grow upwards, the embedded doubles would be word-swapped.) 943 jlong_accessor acc; 944 acc.long_value = d; 945 array->append(new ConstantIntValue(acc.words[1])); 946 array->append(new ConstantIntValue(acc.words[0])); 947 #endif 948 break; 949 } 950 case Type::Top: // Add an illegal value here 951 array->append(new LocationValue(Location())); 952 break; 953 default: 954 ShouldNotReachHere(); 955 break; 956 } 957 } 958 959 // Determine if this node starts a bundle 960 bool PhaseOutput::starts_bundle(const Node *n) const { 961 return (_node_bundling_limit > n->_idx && 962 _node_bundling_base[n->_idx].starts_bundle()); 963 } 964 965 // Determine if there is a monitor that has 'ov' as its owner. 966 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const { 967 for (int k = 0; k < monarray->length(); k++) { 968 MonitorValue* mv = monarray->at(k); 969 if (mv->owner() == ov) { 970 return true; 971 } 972 } 973 974 return false; 975 } 976 977 //--------------------------Process_OopMap_Node-------------------------------- 978 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) { 979 // Handle special safepoint nodes for synchronization 980 MachSafePointNode *sfn = mach->as_MachSafePoint(); 981 MachCallNode *mcall; 982 983 int safepoint_pc_offset = current_offset; 984 bool is_method_handle_invoke = false; 985 bool return_oop = false; 986 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope; 987 bool arg_escape = false; 988 989 // Add the safepoint in the DebugInfoRecorder 990 if( !mach->is_MachCall() ) { 991 mcall = nullptr; 992 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 993 } else { 994 mcall = mach->as_MachCall(); 995 996 // Is the call a MethodHandle call? 997 if (mcall->is_MachCallJava()) { 998 if (mcall->as_MachCallJava()->_method_handle_invoke) { 999 assert(C->has_method_handle_invokes(), "must have been set during call generation"); 1000 is_method_handle_invoke = true; 1001 } 1002 arg_escape = mcall->as_MachCallJava()->_arg_escape; 1003 } 1004 1005 // Check if a call returns an object. 1006 if (mcall->returns_pointer()) { 1007 return_oop = true; 1008 } 1009 safepoint_pc_offset += mcall->ret_addr_offset(); 1010 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 1011 } 1012 1013 // Loop over the JVMState list to add scope information 1014 // Do not skip safepoints with a null method, they need monitor info 1015 JVMState* youngest_jvms = sfn->jvms(); 1016 int max_depth = youngest_jvms->depth(); 1017 1018 // Allocate the object pool for scalar-replaced objects -- the map from 1019 // small-integer keys (which can be recorded in the local and ostack 1020 // arrays) to descriptions of the object state. 1021 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 1022 1023 // Visit scopes from oldest to youngest. 1024 for (int depth = 1; depth <= max_depth; depth++) { 1025 JVMState* jvms = youngest_jvms->of_depth(depth); 1026 int idx; 1027 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr; 1028 // Safepoints that do not have method() set only provide oop-map and monitor info 1029 // to support GC; these do not support deoptimization. 1030 int num_locs = (method == nullptr) ? 0 : jvms->loc_size(); 1031 int num_exps = (method == nullptr) ? 0 : jvms->stk_size(); 1032 int num_mon = jvms->nof_monitors(); 1033 assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(), 1034 "JVMS local count must match that of the method"); 1035 1036 // Add Local and Expression Stack Information 1037 1038 // Insert locals into the locarray 1039 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 1040 for( idx = 0; idx < num_locs; idx++ ) { 1041 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 1042 } 1043 1044 // Insert expression stack entries into the exparray 1045 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 1046 for( idx = 0; idx < num_exps; idx++ ) { 1047 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 1048 } 1049 1050 // Add in mappings of the monitors 1051 assert( !method || 1052 !method->is_synchronized() || 1053 method->is_native() || 1054 num_mon > 0 || 1055 !GenerateSynchronizationCode, 1056 "monitors must always exist for synchronized methods"); 1057 1058 // Build the growable array of ScopeValues for exp stack 1059 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 1060 1061 // Loop over monitors and insert into array 1062 for (idx = 0; idx < num_mon; idx++) { 1063 // Grab the node that defines this monitor 1064 Node* box_node = sfn->monitor_box(jvms, idx); 1065 Node* obj_node = sfn->monitor_obj(jvms, idx); 1066 1067 // Create ScopeValue for object 1068 ScopeValue *scval = nullptr; 1069 1070 if (obj_node->is_SafePointScalarObject()) { 1071 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 1072 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx); 1073 if (scval == nullptr) { 1074 const Type *t = spobj->bottom_type(); 1075 ciKlass* cik = t->is_oopptr()->exact_klass(); 1076 assert(cik->is_instance_klass() || 1077 cik->is_array_klass(), "Not supported allocation."); 1078 ObjectValue* sv = new ObjectValue(spobj->_idx, 1079 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 1080 PhaseOutput::set_sv_for_object_node(objs, sv); 1081 1082 uint first_ind = spobj->first_index(youngest_jvms); 1083 for (uint i = 0; i < spobj->n_fields(); i++) { 1084 Node* fld_node = sfn->in(first_ind+i); 1085 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 1086 } 1087 scval = sv; 1088 } 1089 } else if (obj_node->is_SafePointScalarMerge()) { 1090 SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge(); 1091 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx); 1092 1093 if (mv == nullptr) { 1094 GrowableArray<ScopeValue*> deps; 1095 1096 int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms); 1097 FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs); 1098 assert(deps.length() == 1, "missing value"); 1099 1100 int selector_idx = smerge->selector_idx(youngest_jvms); 1101 FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr); 1102 assert(deps.length() == 2, "missing value"); 1103 1104 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1)); 1105 set_sv_for_object_node(objs, mv); 1106 1107 for (uint i = 1; i < smerge->req(); i++) { 1108 Node* obj_node = smerge->in(i); 1109 FillLocArray(mv->possible_objects()->length(), sfn, obj_node, mv->possible_objects(), objs); 1110 } 1111 } 1112 scval = mv; 1113 } else if (!obj_node->is_Con()) { 1114 OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node); 1115 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 1116 scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop ); 1117 } else { 1118 scval = new_loc_value( C->regalloc(), obj_reg, Location::oop ); 1119 } 1120 } else { 1121 const TypePtr *tp = obj_node->get_ptr_type(); 1122 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 1123 } 1124 1125 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 1126 Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg)); 1127 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 1128 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 1129 } 1130 1131 // Mark ObjectValue nodes as root nodes if they are directly 1132 // referenced in the JVMS. 1133 for (int i = 0; i < objs->length(); i++) { 1134 ScopeValue* sv = objs->at(i); 1135 if (sv->is_object_merge()) { 1136 ObjectMergeValue* merge = sv->as_ObjectMergeValue(); 1137 1138 for (int j = 0; j< merge->possible_objects()->length(); j++) { 1139 ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue(); 1140 bool is_root = locarray->contains(ov) || exparray->contains(ov) || contains_as_owner(monarray, ov); 1141 ov->set_root(is_root); 1142 } 1143 } 1144 } 1145 1146 // We dump the object pool first, since deoptimization reads it in first. 1147 C->debug_info()->dump_object_pool(objs); 1148 1149 // Build first class objects to pass to scope 1150 DebugToken *locvals = C->debug_info()->create_scope_values(locarray); 1151 DebugToken *expvals = C->debug_info()->create_scope_values(exparray); 1152 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray); 1153 1154 // Make method available for all Safepoints 1155 ciMethod* scope_method = method ? method : C->method(); 1156 // Describe the scope here 1157 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 1158 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 1159 // Now we can describe the scope. 1160 methodHandle null_mh; 1161 bool rethrow_exception = false; 1162 C->debug_info()->describe_scope( 1163 safepoint_pc_offset, 1164 null_mh, 1165 scope_method, 1166 jvms->bci(), 1167 jvms->should_reexecute(), 1168 rethrow_exception, 1169 is_method_handle_invoke, 1170 return_oop, 1171 has_ea_local_in_scope, 1172 arg_escape, 1173 locvals, 1174 expvals, 1175 monvals 1176 ); 1177 } // End jvms loop 1178 1179 // Mark the end of the scope set. 1180 C->debug_info()->end_safepoint(safepoint_pc_offset); 1181 } 1182 1183 1184 1185 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 1186 class NonSafepointEmitter { 1187 Compile* C; 1188 JVMState* _pending_jvms; 1189 int _pending_offset; 1190 1191 void emit_non_safepoint(); 1192 1193 public: 1194 NonSafepointEmitter(Compile* compile) { 1195 this->C = compile; 1196 _pending_jvms = nullptr; 1197 _pending_offset = 0; 1198 } 1199 1200 void observe_instruction(Node* n, int pc_offset) { 1201 if (!C->debug_info()->recording_non_safepoints()) return; 1202 1203 Node_Notes* nn = C->node_notes_at(n->_idx); 1204 if (nn == nullptr || nn->jvms() == nullptr) return; 1205 if (_pending_jvms != nullptr && 1206 _pending_jvms->same_calls_as(nn->jvms())) { 1207 // Repeated JVMS? Stretch it up here. 1208 _pending_offset = pc_offset; 1209 } else { 1210 if (_pending_jvms != nullptr && 1211 _pending_offset < pc_offset) { 1212 emit_non_safepoint(); 1213 } 1214 _pending_jvms = nullptr; 1215 if (pc_offset > C->debug_info()->last_pc_offset()) { 1216 // This is the only way _pending_jvms can become non-null: 1217 _pending_jvms = nn->jvms(); 1218 _pending_offset = pc_offset; 1219 } 1220 } 1221 } 1222 1223 // Stay out of the way of real safepoints: 1224 void observe_safepoint(JVMState* jvms, int pc_offset) { 1225 if (_pending_jvms != nullptr && 1226 !_pending_jvms->same_calls_as(jvms) && 1227 _pending_offset < pc_offset) { 1228 emit_non_safepoint(); 1229 } 1230 _pending_jvms = nullptr; 1231 } 1232 1233 void flush_at_end() { 1234 if (_pending_jvms != nullptr) { 1235 emit_non_safepoint(); 1236 } 1237 _pending_jvms = nullptr; 1238 } 1239 }; 1240 1241 void NonSafepointEmitter::emit_non_safepoint() { 1242 JVMState* youngest_jvms = _pending_jvms; 1243 int pc_offset = _pending_offset; 1244 1245 // Clear it now: 1246 _pending_jvms = nullptr; 1247 1248 DebugInformationRecorder* debug_info = C->debug_info(); 1249 assert(debug_info->recording_non_safepoints(), "sanity"); 1250 1251 debug_info->add_non_safepoint(pc_offset); 1252 int max_depth = youngest_jvms->depth(); 1253 1254 // Visit scopes from oldest to youngest. 1255 for (int depth = 1; depth <= max_depth; depth++) { 1256 JVMState* jvms = youngest_jvms->of_depth(depth); 1257 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr; 1258 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 1259 methodHandle null_mh; 1260 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute()); 1261 } 1262 1263 // Mark the end of the scope set. 1264 debug_info->end_non_safepoint(pc_offset); 1265 } 1266 1267 //------------------------------init_buffer------------------------------------ 1268 void PhaseOutput::estimate_buffer_size(int& const_req) { 1269 1270 // Set the initially allocated size 1271 const_req = initial_const_capacity; 1272 1273 // The extra spacing after the code is necessary on some platforms. 1274 // Sometimes we need to patch in a jump after the last instruction, 1275 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 1276 1277 // Compute the byte offset where we can store the deopt pc. 1278 if (C->fixed_slots() != 0) { 1279 _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 1280 } 1281 1282 // Compute prolog code size 1283 _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize; 1284 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 1285 1286 if (C->has_mach_constant_base_node()) { 1287 uint add_size = 0; 1288 // Fill the constant table. 1289 // Note: This must happen before shorten_branches. 1290 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) { 1291 Block* b = C->cfg()->get_block(i); 1292 1293 for (uint j = 0; j < b->number_of_nodes(); j++) { 1294 Node* n = b->get_node(j); 1295 1296 // If the node is a MachConstantNode evaluate the constant 1297 // value section. 1298 if (n->is_MachConstant()) { 1299 MachConstantNode* machcon = n->as_MachConstant(); 1300 machcon->eval_constant(C); 1301 } else if (n->is_Mach()) { 1302 // On Power there are more nodes that issue constants. 1303 add_size += (n->as_Mach()->ins_num_consts() * 8); 1304 } 1305 } 1306 } 1307 1308 // Calculate the offsets of the constants and the size of the 1309 // constant table (including the padding to the next section). 1310 constant_table().calculate_offsets_and_size(); 1311 const_req = constant_table().size() + add_size; 1312 } 1313 1314 // Initialize the space for the BufferBlob used to find and verify 1315 // instruction size in MachNode::emit_size() 1316 init_scratch_buffer_blob(const_req); 1317 } 1318 1319 CodeBuffer* PhaseOutput::init_buffer() { 1320 int stub_req = _buf_sizes._stub; 1321 int code_req = _buf_sizes._code; 1322 int const_req = _buf_sizes._const; 1323 1324 int pad_req = NativeCall::instruction_size; 1325 1326 // GC barrier stubs are generated in code section 1327 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1328 code_req += bs->estimate_stub_size(); 1329 1330 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1331 // class HandlerImpl is platform-specific and defined in the *.ad files. 1332 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler 1333 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler 1334 stub_req += MAX_stubs_size; // ensure per-stub margin 1335 code_req += MAX_inst_size; // ensure per-instruction margin 1336 1337 if (StressCodeBuffers) { 1338 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1339 } 1340 int total_req = 1341 const_req + 1342 code_req + 1343 pad_req + 1344 stub_req + 1345 exception_handler_req + 1346 deopt_handler_req; // deopt handler 1347 1348 if (C->has_method_handle_invokes()) { 1349 total_req += deopt_handler_req; // deopt MH handler 1350 stub_req += deopt_handler_req; 1351 } 1352 CodeBuffer* cb = code_buffer(); 1353 cb->initialize(total_req, _buf_sizes._reloc); 1354 1355 // Have we run out of code space? 1356 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) { 1357 C->record_failure("CodeCache is full"); 1358 return nullptr; 1359 } 1360 // Configure the code buffer. 1361 cb->initialize_consts_size(const_req); 1362 cb->initialize_stubs_size(stub_req); 1363 cb->initialize_oop_recorder(C->env()->oop_recorder()); 1364 1365 // fill in the nop array for bundling computations 1366 MachNode *_nop_list[Bundle::_nop_count]; 1367 Bundle::initialize_nops(_nop_list); 1368 1369 return cb; 1370 } 1371 1372 //------------------------------fill_buffer------------------------------------ 1373 void PhaseOutput::fill_buffer(C2_MacroAssembler* masm, uint* blk_starts) { 1374 // blk_starts[] contains offsets calculated during short branches processing, 1375 // offsets should not be increased during following steps. 1376 1377 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1378 // of a loop. It is used to determine the padding for loop alignment. 1379 Compile::TracePhase tp("fill buffer", &timers[_t_fillBuffer]); 1380 1381 compute_loop_first_inst_sizes(); 1382 1383 // Create oopmap set. 1384 _oop_map_set = new OopMapSet(); 1385 1386 // !!!!! This preserves old handling of oopmaps for now 1387 C->debug_info()->set_oopmaps(_oop_map_set); 1388 1389 uint nblocks = C->cfg()->number_of_blocks(); 1390 // Count and start of implicit null check instructions 1391 uint inct_cnt = 0; 1392 uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1393 1394 // Count and start of calls 1395 uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1396 1397 uint return_offset = 0; 1398 int nop_size = (new MachNopNode())->size(C->regalloc()); 1399 1400 int previous_offset = 0; 1401 int current_offset = 0; 1402 int last_call_offset = -1; 1403 int last_avoid_back_to_back_offset = -1; 1404 #ifdef ASSERT 1405 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1406 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1407 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1408 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1409 #endif 1410 1411 // Create an array of unused labels, one for each basic block, if printing is enabled 1412 #if defined(SUPPORT_OPTO_ASSEMBLY) 1413 int* node_offsets = nullptr; 1414 uint node_offset_limit = C->unique(); 1415 1416 if (C->print_assembly()) { 1417 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1418 } 1419 if (node_offsets != nullptr) { 1420 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly. 1421 memset(node_offsets, 0, node_offset_limit*sizeof(int)); 1422 } 1423 #endif 1424 1425 NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily 1426 1427 // Emit the constant table. 1428 if (C->has_mach_constant_base_node()) { 1429 if (!constant_table().emit(masm)) { 1430 C->record_failure("consts section overflow"); 1431 return; 1432 } 1433 } 1434 1435 // Create an array of labels, one for each basic block 1436 Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1437 for (uint i = 0; i <= nblocks; i++) { 1438 blk_labels[i].init(); 1439 } 1440 1441 // Now fill in the code buffer 1442 Node* delay_slot = nullptr; 1443 for (uint i = 0; i < nblocks; i++) { 1444 Block* block = C->cfg()->get_block(i); 1445 _block = block; 1446 Node* head = block->head(); 1447 1448 // If this block needs to start aligned (i.e, can be reached other 1449 // than by falling-thru from the previous block), then force the 1450 // start of a new bundle. 1451 if (Pipeline::requires_bundling() && starts_bundle(head)) { 1452 masm->code()->flush_bundle(true); 1453 } 1454 1455 #ifdef ASSERT 1456 if (!block->is_connector()) { 1457 stringStream st; 1458 block->dump_head(C->cfg(), &st); 1459 masm->block_comment(st.freeze()); 1460 } 1461 jmp_target[i] = 0; 1462 jmp_offset[i] = 0; 1463 jmp_size[i] = 0; 1464 jmp_rule[i] = 0; 1465 #endif 1466 int blk_offset = current_offset; 1467 1468 // Define the label at the beginning of the basic block 1469 masm->bind(blk_labels[block->_pre_order]); 1470 1471 uint last_inst = block->number_of_nodes(); 1472 1473 // Emit block normally, except for last instruction. 1474 // Emit means "dump code bits into code buffer". 1475 for (uint j = 0; j<last_inst; j++) { 1476 _index = j; 1477 1478 // Get the node 1479 Node* n = block->get_node(j); 1480 1481 // See if delay slots are supported 1482 if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) { 1483 assert(delay_slot == nullptr, "no use of delay slot node"); 1484 assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1485 1486 delay_slot = n; 1487 continue; 1488 } 1489 1490 // If this starts a new instruction group, then flush the current one 1491 // (but allow split bundles) 1492 if (Pipeline::requires_bundling() && starts_bundle(n)) 1493 masm->code()->flush_bundle(false); 1494 1495 // Special handling for SafePoint/Call Nodes 1496 bool is_mcall = false; 1497 if (n->is_Mach()) { 1498 MachNode *mach = n->as_Mach(); 1499 is_mcall = n->is_MachCall(); 1500 bool is_sfn = n->is_MachSafePoint(); 1501 1502 // If this requires all previous instructions be flushed, then do so 1503 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1504 masm->code()->flush_bundle(true); 1505 current_offset = masm->offset(); 1506 } 1507 1508 // A padding may be needed again since a previous instruction 1509 // could be moved to delay slot. 1510 1511 // align the instruction if necessary 1512 int padding = mach->compute_padding(current_offset); 1513 // Make sure safepoint node for polling is distinct from a call's 1514 // return by adding a nop if needed. 1515 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1516 padding = nop_size; 1517 } 1518 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) && 1519 current_offset == last_avoid_back_to_back_offset) { 1520 // Avoid back to back some instructions. 1521 padding = nop_size; 1522 } 1523 1524 if (padding > 0) { 1525 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1526 int nops_cnt = padding / nop_size; 1527 MachNode *nop = new MachNopNode(nops_cnt); 1528 block->insert_node(nop, j++); 1529 last_inst++; 1530 C->cfg()->map_node_to_block(nop, block); 1531 // Ensure enough space. 1532 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1533 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) { 1534 C->record_failure("CodeCache is full"); 1535 return; 1536 } 1537 nop->emit(masm, C->regalloc()); 1538 masm->code()->flush_bundle(true); 1539 current_offset = masm->offset(); 1540 } 1541 1542 bool observe_safepoint = is_sfn; 1543 // Remember the start of the last call in a basic block 1544 if (is_mcall) { 1545 MachCallNode *mcall = mach->as_MachCall(); 1546 1547 // This destination address is NOT PC-relative 1548 mcall->method_set((intptr_t)mcall->entry_point()); 1549 1550 // Save the return address 1551 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset(); 1552 1553 observe_safepoint = mcall->guaranteed_safepoint(); 1554 } 1555 1556 // sfn will be valid whenever mcall is valid now because of inheritance 1557 if (observe_safepoint) { 1558 // Handle special safepoint nodes for synchronization 1559 if (!is_mcall) { 1560 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1561 // !!!!! Stubs only need an oopmap right now, so bail out 1562 if (sfn->jvms()->method() == nullptr) { 1563 // Write the oopmap directly to the code blob??!! 1564 continue; 1565 } 1566 } // End synchronization 1567 1568 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1569 current_offset); 1570 Process_OopMap_Node(mach, current_offset); 1571 } // End if safepoint 1572 1573 // If this is a null check, then add the start of the previous instruction to the list 1574 else if( mach->is_MachNullCheck() ) { 1575 inct_starts[inct_cnt++] = previous_offset; 1576 } 1577 1578 // If this is a branch, then fill in the label with the target BB's label 1579 else if (mach->is_MachBranch()) { 1580 // This requires the TRUE branch target be in succs[0] 1581 uint block_num = block->non_connector_successor(0)->_pre_order; 1582 1583 // Try to replace long branch if delay slot is not used, 1584 // it is mostly for back branches since forward branch's 1585 // distance is not updated yet. 1586 bool delay_slot_is_used = valid_bundle_info(n) && 1587 C->output()->node_bundling(n)->use_unconditional_delay(); 1588 if (!delay_slot_is_used && mach->may_be_short_branch()) { 1589 assert(delay_slot == nullptr, "not expecting delay slot node"); 1590 int br_size = n->size(C->regalloc()); 1591 int offset = blk_starts[block_num] - current_offset; 1592 if (block_num >= i) { 1593 // Current and following block's offset are not 1594 // finalized yet, adjust distance by the difference 1595 // between calculated and final offsets of current block. 1596 offset -= (blk_starts[i] - blk_offset); 1597 } 1598 // In the following code a nop could be inserted before 1599 // the branch which will increase the backward distance. 1600 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1601 if (needs_padding && offset <= 0) 1602 offset -= nop_size; 1603 1604 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) { 1605 // We've got a winner. Replace this branch. 1606 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 1607 1608 // Update the jmp_size. 1609 int new_size = replacement->size(C->regalloc()); 1610 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1611 // Insert padding between avoid_back_to_back branches. 1612 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 1613 MachNode *nop = new MachNopNode(); 1614 block->insert_node(nop, j++); 1615 C->cfg()->map_node_to_block(nop, block); 1616 last_inst++; 1617 nop->emit(masm, C->regalloc()); 1618 masm->code()->flush_bundle(true); 1619 current_offset = masm->offset(); 1620 } 1621 #ifdef ASSERT 1622 jmp_target[i] = block_num; 1623 jmp_offset[i] = current_offset - blk_offset; 1624 jmp_size[i] = new_size; 1625 jmp_rule[i] = mach->rule(); 1626 #endif 1627 block->map_node(replacement, j); 1628 mach->subsume_by(replacement, C); 1629 n = replacement; 1630 mach = replacement; 1631 } 1632 } 1633 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1634 } else if (mach->ideal_Opcode() == Op_Jump) { 1635 for (uint h = 0; h < block->_num_succs; h++) { 1636 Block* succs_block = block->_succs[h]; 1637 for (uint j = 1; j < succs_block->num_preds(); j++) { 1638 Node* jpn = succs_block->pred(j); 1639 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1640 uint block_num = succs_block->non_connector()->_pre_order; 1641 Label *blkLabel = &blk_labels[block_num]; 1642 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1643 } 1644 } 1645 } 1646 } 1647 #ifdef ASSERT 1648 // Check that oop-store precedes the card-mark 1649 else if (mach->ideal_Opcode() == Op_StoreCM) { 1650 uint storeCM_idx = j; 1651 int count = 0; 1652 for (uint prec = mach->req(); prec < mach->len(); prec++) { 1653 Node *oop_store = mach->in(prec); // Precedence edge 1654 if (oop_store == nullptr) continue; 1655 count++; 1656 uint i4; 1657 for (i4 = 0; i4 < last_inst; ++i4) { 1658 if (block->get_node(i4) == oop_store) { 1659 break; 1660 } 1661 } 1662 // Note: This test can provide a false failure if other precedence 1663 // edges have been added to the storeCMNode. 1664 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1665 } 1666 assert(count > 0, "storeCM expects at least one precedence edge"); 1667 } 1668 #endif 1669 else if (!n->is_Proj()) { 1670 // Remember the beginning of the previous instruction, in case 1671 // it's followed by a flag-kill and a null-check. Happens on 1672 // Intel all the time, with add-to-memory kind of opcodes. 1673 previous_offset = current_offset; 1674 } 1675 1676 // Not an else-if! 1677 // If this is a trap based cmp then add its offset to the list. 1678 if (mach->is_TrapBasedCheckNode()) { 1679 inct_starts[inct_cnt++] = current_offset; 1680 } 1681 } 1682 1683 // Verify that there is sufficient space remaining 1684 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1685 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) { 1686 C->record_failure("CodeCache is full"); 1687 return; 1688 } 1689 1690 // Save the offset for the listing 1691 #if defined(SUPPORT_OPTO_ASSEMBLY) 1692 if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) { 1693 node_offsets[n->_idx] = masm->offset(); 1694 } 1695 #endif 1696 assert(!C->failing(), "Should not reach here if failing."); 1697 1698 // "Normal" instruction case 1699 DEBUG_ONLY(uint instr_offset = masm->offset()); 1700 n->emit(masm, C->regalloc()); 1701 current_offset = masm->offset(); 1702 1703 // Above we only verified that there is enough space in the instruction section. 1704 // However, the instruction may emit stubs that cause code buffer expansion. 1705 // Bail out here if expansion failed due to a lack of code cache space. 1706 if (C->failing()) { 1707 return; 1708 } 1709 1710 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset), 1711 "ret_addr_offset() not within emitted code"); 1712 1713 #ifdef ASSERT 1714 uint n_size = n->size(C->regalloc()); 1715 if (n_size < (current_offset-instr_offset)) { 1716 MachNode* mach = n->as_Mach(); 1717 n->dump(); 1718 mach->dump_format(C->regalloc(), tty); 1719 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset); 1720 Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty); 1721 tty->print_cr(" ------------------- "); 1722 BufferBlob* blob = this->scratch_buffer_blob(); 1723 address blob_begin = blob->content_begin(); 1724 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty); 1725 assert(false, "wrong size of mach node"); 1726 } 1727 #endif 1728 non_safepoints.observe_instruction(n, current_offset); 1729 1730 // mcall is last "call" that can be a safepoint 1731 // record it so we can see if a poll will directly follow it 1732 // in which case we'll need a pad to make the PcDesc sites unique 1733 // see 5010568. This can be slightly inaccurate but conservative 1734 // in the case that return address is not actually at current_offset. 1735 // This is a small price to pay. 1736 1737 if (is_mcall) { 1738 last_call_offset = current_offset; 1739 } 1740 1741 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 1742 // Avoid back to back some instructions. 1743 last_avoid_back_to_back_offset = current_offset; 1744 } 1745 1746 // See if this instruction has a delay slot 1747 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1748 guarantee(delay_slot != nullptr, "expecting delay slot node"); 1749 1750 // Back up 1 instruction 1751 masm->code()->set_insts_end(masm->code()->insts_end() - Pipeline::instr_unit_size()); 1752 1753 // Save the offset for the listing 1754 #if defined(SUPPORT_OPTO_ASSEMBLY) 1755 if ((node_offsets != nullptr) && (delay_slot->_idx < node_offset_limit)) { 1756 node_offsets[delay_slot->_idx] = masm->offset(); 1757 } 1758 #endif 1759 1760 // Support a SafePoint in the delay slot 1761 if (delay_slot->is_MachSafePoint()) { 1762 MachNode *mach = delay_slot->as_Mach(); 1763 // !!!!! Stubs only need an oopmap right now, so bail out 1764 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == nullptr) { 1765 // Write the oopmap directly to the code blob??!! 1766 delay_slot = nullptr; 1767 continue; 1768 } 1769 1770 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1771 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1772 adjusted_offset); 1773 // Generate an OopMap entry 1774 Process_OopMap_Node(mach, adjusted_offset); 1775 } 1776 1777 // Insert the delay slot instruction 1778 delay_slot->emit(masm, C->regalloc()); 1779 1780 // Don't reuse it 1781 delay_slot = nullptr; 1782 } 1783 1784 } // End for all instructions in block 1785 1786 // If the next block is the top of a loop, pad this block out to align 1787 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1788 if (i < nblocks-1) { 1789 Block *nb = C->cfg()->get_block(i + 1); 1790 int padding = nb->alignment_padding(current_offset); 1791 if( padding > 0 ) { 1792 MachNode *nop = new MachNopNode(padding / nop_size); 1793 block->insert_node(nop, block->number_of_nodes()); 1794 C->cfg()->map_node_to_block(nop, block); 1795 nop->emit(masm, C->regalloc()); 1796 current_offset = masm->offset(); 1797 } 1798 } 1799 // Verify that the distance for generated before forward 1800 // short branches is still valid. 1801 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size"); 1802 1803 // Save new block start offset 1804 blk_starts[i] = blk_offset; 1805 } // End of for all blocks 1806 blk_starts[nblocks] = current_offset; 1807 1808 non_safepoints.flush_at_end(); 1809 1810 // Offset too large? 1811 if (C->failing()) return; 1812 1813 // Define a pseudo-label at the end of the code 1814 masm->bind( blk_labels[nblocks] ); 1815 1816 // Compute the size of the first block 1817 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1818 1819 #ifdef ASSERT 1820 for (uint i = 0; i < nblocks; i++) { // For all blocks 1821 if (jmp_target[i] != 0) { 1822 int br_size = jmp_size[i]; 1823 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1824 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1825 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]); 1826 assert(false, "Displacement too large for short jmp"); 1827 } 1828 } 1829 } 1830 #endif 1831 1832 if (!masm->code()->finalize_stubs()) { 1833 C->record_failure("CodeCache is full"); 1834 return; 1835 } 1836 1837 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1838 bs->emit_stubs(*masm->code()); 1839 if (C->failing()) return; 1840 1841 // Fill in stubs. 1842 assert(masm->inst_mark() == nullptr, "should be."); 1843 _stub_list.emit(*masm); 1844 if (C->failing()) return; 1845 1846 #ifndef PRODUCT 1847 // Information on the size of the method, without the extraneous code 1848 Scheduling::increment_method_size(masm->offset()); 1849 #endif 1850 1851 // ------------------ 1852 // Fill in exception table entries. 1853 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1854 1855 // Only java methods have exception handlers and deopt handlers 1856 // class HandlerImpl is platform-specific and defined in the *.ad files. 1857 if (C->method()) { 1858 // Emit the exception handler code. 1859 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(masm)); 1860 if (C->failing()) { 1861 return; // CodeBuffer::expand failed 1862 } 1863 // Emit the deopt handler code. 1864 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(masm)); 1865 1866 // Emit the MethodHandle deopt handler code (if required). 1867 if (C->has_method_handle_invokes() && !C->failing()) { 1868 // We can use the same code as for the normal deopt handler, we 1869 // just need a different entry point address. 1870 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(masm)); 1871 } 1872 } 1873 1874 // One last check for failed CodeBuffer::expand: 1875 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) { 1876 C->record_failure("CodeCache is full"); 1877 return; 1878 } 1879 1880 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY) 1881 if (C->print_assembly()) { 1882 tty->cr(); 1883 tty->print_cr("============================= C2-compiled nmethod =============================="); 1884 } 1885 #endif 1886 1887 #if defined(SUPPORT_OPTO_ASSEMBLY) 1888 // Dump the assembly code, including basic-block numbers 1889 if (C->print_assembly()) { 1890 ttyLocker ttyl; // keep the following output all in one block 1891 if (!VMThread::should_terminate()) { // test this under the tty lock 1892 // print_metadata and dump_asm may safepoint which makes us loose the ttylock. 1893 // We call them first and write to a stringStream, then we retake the lock to 1894 // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe. 1895 ResourceMark rm; 1896 stringStream method_metadata_str; 1897 if (C->method() != nullptr) { 1898 C->method()->print_metadata(&method_metadata_str); 1899 } 1900 stringStream dump_asm_str; 1901 dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit); 1902 1903 NoSafepointVerifier nsv; 1904 ttyLocker ttyl2; 1905 // This output goes directly to the tty, not the compiler log. 1906 // To enable tools to match it up with the compilation activity, 1907 // be sure to tag this tty output with the compile ID. 1908 if (xtty != nullptr) { 1909 xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(), 1910 C->is_osr_compilation() ? " compile_kind='osr'" : ""); 1911 } 1912 if (C->method() != nullptr) { 1913 tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id()); 1914 tty->print_raw(method_metadata_str.freeze()); 1915 } else if (C->stub_name() != nullptr) { 1916 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name()); 1917 } 1918 tty->cr(); 1919 tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id()); 1920 tty->print_raw(dump_asm_str.freeze()); 1921 tty->print_cr("--------------------------------------------------------------------------------"); 1922 if (xtty != nullptr) { 1923 xtty->tail("opto_assembly"); 1924 } 1925 } 1926 } 1927 #endif 1928 } 1929 1930 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1931 _inc_table.set_size(cnt); 1932 1933 uint inct_cnt = 0; 1934 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) { 1935 Block* block = C->cfg()->get_block(i); 1936 Node *n = nullptr; 1937 int j; 1938 1939 // Find the branch; ignore trailing NOPs. 1940 for (j = block->number_of_nodes() - 1; j >= 0; j--) { 1941 n = block->get_node(j); 1942 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) { 1943 break; 1944 } 1945 } 1946 1947 // If we didn't find anything, continue 1948 if (j < 0) { 1949 continue; 1950 } 1951 1952 // Compute ExceptionHandlerTable subtable entry and add it 1953 // (skip empty blocks) 1954 if (n->is_Catch()) { 1955 1956 // Get the offset of the return from the call 1957 uint call_return = call_returns[block->_pre_order]; 1958 #ifdef ASSERT 1959 assert( call_return > 0, "no call seen for this basic block" ); 1960 while (block->get_node(--j)->is_MachProj()) ; 1961 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); 1962 #endif 1963 // last instruction is a CatchNode, find it's CatchProjNodes 1964 int nof_succs = block->_num_succs; 1965 // allocate space 1966 GrowableArray<intptr_t> handler_bcis(nof_succs); 1967 GrowableArray<intptr_t> handler_pcos(nof_succs); 1968 // iterate through all successors 1969 for (int j = 0; j < nof_succs; j++) { 1970 Block* s = block->_succs[j]; 1971 bool found_p = false; 1972 for (uint k = 1; k < s->num_preds(); k++) { 1973 Node* pk = s->pred(k); 1974 if (pk->is_CatchProj() && pk->in(0) == n) { 1975 const CatchProjNode* p = pk->as_CatchProj(); 1976 found_p = true; 1977 // add the corresponding handler bci & pco information 1978 if (p->_con != CatchProjNode::fall_through_index) { 1979 // p leads to an exception handler (and is not fall through) 1980 assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering"); 1981 // no duplicates, please 1982 if (!handler_bcis.contains(p->handler_bci())) { 1983 uint block_num = s->non_connector()->_pre_order; 1984 handler_bcis.append(p->handler_bci()); 1985 handler_pcos.append(blk_labels[block_num].loc_pos()); 1986 } 1987 } 1988 } 1989 } 1990 assert(found_p, "no matching predecessor found"); 1991 // Note: Due to empty block removal, one block may have 1992 // several CatchProj inputs, from the same Catch. 1993 } 1994 1995 // Set the offset of the return from the call 1996 assert(handler_bcis.find(-1) != -1, "must have default handler"); 1997 _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos); 1998 continue; 1999 } 2000 2001 // Handle implicit null exception table updates 2002 if (n->is_MachNullCheck()) { 2003 uint block_num = block->non_connector_successor(0)->_pre_order; 2004 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 2005 continue; 2006 } 2007 // Handle implicit exception table updates: trap instructions. 2008 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) { 2009 uint block_num = block->non_connector_successor(0)->_pre_order; 2010 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 2011 continue; 2012 } 2013 } // End of for all blocks fill in exception table entries 2014 } 2015 2016 // Static Variables 2017 #ifndef PRODUCT 2018 uint Scheduling::_total_nop_size = 0; 2019 uint Scheduling::_total_method_size = 0; 2020 uint Scheduling::_total_branches = 0; 2021 uint Scheduling::_total_unconditional_delays = 0; 2022 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 2023 #endif 2024 2025 // Initializer for class Scheduling 2026 2027 Scheduling::Scheduling(Arena *arena, Compile &compile) 2028 : _arena(arena), 2029 _cfg(compile.cfg()), 2030 _regalloc(compile.regalloc()), 2031 _scheduled(arena), 2032 _available(arena), 2033 _reg_node(arena), 2034 _pinch_free_list(arena), 2035 _next_node(nullptr), 2036 _bundle_instr_count(0), 2037 _bundle_cycle_number(0), 2038 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]) 2039 #ifndef PRODUCT 2040 , _branches(0) 2041 , _unconditional_delays(0) 2042 #endif 2043 { 2044 // Create a MachNopNode 2045 _nop = new MachNopNode(); 2046 2047 // Now that the nops are in the array, save the count 2048 // (but allow entries for the nops) 2049 _node_bundling_limit = compile.unique(); 2050 uint node_max = _regalloc->node_regs_max_index(); 2051 2052 compile.output()->set_node_bundling_limit(_node_bundling_limit); 2053 2054 // This one is persistent within the Compile class 2055 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 2056 2057 // Allocate space for fixed-size arrays 2058 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 2059 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 2060 2061 // Clear the arrays 2062 for (uint i = 0; i < node_max; i++) { 2063 ::new (&_node_bundling_base[i]) Bundle(); 2064 } 2065 memset(_uses, 0, node_max * sizeof(short)); 2066 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 2067 2068 // Clear the bundling information 2069 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements)); 2070 2071 // Get the last node 2072 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1); 2073 2074 _next_node = block->get_node(block->number_of_nodes() - 1); 2075 } 2076 2077 #ifndef PRODUCT 2078 // Scheduling destructor 2079 Scheduling::~Scheduling() { 2080 _total_branches += _branches; 2081 _total_unconditional_delays += _unconditional_delays; 2082 } 2083 #endif 2084 2085 // Step ahead "i" cycles 2086 void Scheduling::step(uint i) { 2087 2088 Bundle *bundle = node_bundling(_next_node); 2089 bundle->set_starts_bundle(); 2090 2091 // Update the bundle record, but leave the flags information alone 2092 if (_bundle_instr_count > 0) { 2093 bundle->set_instr_count(_bundle_instr_count); 2094 bundle->set_resources_used(_bundle_use.resourcesUsed()); 2095 } 2096 2097 // Update the state information 2098 _bundle_instr_count = 0; 2099 _bundle_cycle_number += i; 2100 _bundle_use.step(i); 2101 } 2102 2103 void Scheduling::step_and_clear() { 2104 Bundle *bundle = node_bundling(_next_node); 2105 bundle->set_starts_bundle(); 2106 2107 // Update the bundle record 2108 if (_bundle_instr_count > 0) { 2109 bundle->set_instr_count(_bundle_instr_count); 2110 bundle->set_resources_used(_bundle_use.resourcesUsed()); 2111 2112 _bundle_cycle_number += 1; 2113 } 2114 2115 // Clear the bundling information 2116 _bundle_instr_count = 0; 2117 _bundle_use.reset(); 2118 2119 memcpy(_bundle_use_elements, 2120 Pipeline_Use::elaborated_elements, 2121 sizeof(Pipeline_Use::elaborated_elements)); 2122 } 2123 2124 // Perform instruction scheduling and bundling over the sequence of 2125 // instructions in backwards order. 2126 void PhaseOutput::ScheduleAndBundle() { 2127 2128 // Don't optimize this if it isn't a method 2129 if (!C->method()) 2130 return; 2131 2132 // Don't optimize this if scheduling is disabled 2133 if (!C->do_scheduling()) 2134 return; 2135 2136 // Scheduling code works only with pairs (8 bytes) maximum. 2137 // And when the scalable vector register is used, we may spill/unspill 2138 // the whole reg regardless of the max vector size. 2139 if (C->max_vector_size() > 8 || 2140 (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) { 2141 return; 2142 } 2143 2144 Compile::TracePhase tp("isched", &timers[_t_instrSched]); 2145 2146 // Create a data structure for all the scheduling information 2147 Scheduling scheduling(Thread::current()->resource_area(), *C); 2148 2149 // Walk backwards over each basic block, computing the needed alignment 2150 // Walk over all the basic blocks 2151 scheduling.DoScheduling(); 2152 2153 #ifndef PRODUCT 2154 if (C->trace_opto_output()) { 2155 // Buffer and print all at once 2156 ResourceMark rm; 2157 stringStream ss; 2158 ss.print("\n---- After ScheduleAndBundle ----\n"); 2159 print_scheduling(&ss); 2160 tty->print("%s", ss.as_string()); 2161 } 2162 #endif 2163 } 2164 2165 #ifndef PRODUCT 2166 // Separated out so that it can be called directly from debugger 2167 void PhaseOutput::print_scheduling() { 2168 print_scheduling(tty); 2169 } 2170 2171 void PhaseOutput::print_scheduling(outputStream* output_stream) { 2172 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) { 2173 output_stream->print("\nBB#%03d:\n", i); 2174 Block* block = C->cfg()->get_block(i); 2175 for (uint j = 0; j < block->number_of_nodes(); j++) { 2176 Node* n = block->get_node(j); 2177 OptoReg::Name reg = C->regalloc()->get_reg_first(n); 2178 output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 2179 n->dump("\n", false, output_stream); 2180 } 2181 } 2182 } 2183 #endif 2184 2185 // See if this node fits into the present instruction bundle 2186 bool Scheduling::NodeFitsInBundle(Node *n) { 2187 uint n_idx = n->_idx; 2188 2189 // If this is the unconditional delay instruction, then it fits 2190 if (n == _unconditional_delay_slot) { 2191 #ifndef PRODUCT 2192 if (_cfg->C->trace_opto_output()) 2193 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 2194 #endif 2195 return (true); 2196 } 2197 2198 // If the node cannot be scheduled this cycle, skip it 2199 if (_current_latency[n_idx] > _bundle_cycle_number) { 2200 #ifndef PRODUCT 2201 if (_cfg->C->trace_opto_output()) 2202 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 2203 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 2204 #endif 2205 return (false); 2206 } 2207 2208 const Pipeline *node_pipeline = n->pipeline(); 2209 2210 uint instruction_count = node_pipeline->instructionCount(); 2211 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2212 instruction_count = 0; 2213 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2214 instruction_count++; 2215 2216 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 2217 #ifndef PRODUCT 2218 if (_cfg->C->trace_opto_output()) 2219 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 2220 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 2221 #endif 2222 return (false); 2223 } 2224 2225 // Don't allow non-machine nodes to be handled this way 2226 if (!n->is_Mach() && instruction_count == 0) 2227 return (false); 2228 2229 // See if there is any overlap 2230 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 2231 2232 if (delay > 0) { 2233 #ifndef PRODUCT 2234 if (_cfg->C->trace_opto_output()) 2235 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 2236 #endif 2237 return false; 2238 } 2239 2240 #ifndef PRODUCT 2241 if (_cfg->C->trace_opto_output()) 2242 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 2243 #endif 2244 2245 return true; 2246 } 2247 2248 Node * Scheduling::ChooseNodeToBundle() { 2249 uint siz = _available.size(); 2250 2251 if (siz == 0) { 2252 2253 #ifndef PRODUCT 2254 if (_cfg->C->trace_opto_output()) 2255 tty->print("# ChooseNodeToBundle: null\n"); 2256 #endif 2257 return (nullptr); 2258 } 2259 2260 // Fast path, if only 1 instruction in the bundle 2261 if (siz == 1) { 2262 #ifndef PRODUCT 2263 if (_cfg->C->trace_opto_output()) { 2264 tty->print("# ChooseNodeToBundle (only 1): "); 2265 _available[0]->dump(); 2266 } 2267 #endif 2268 return (_available[0]); 2269 } 2270 2271 // Don't bother, if the bundle is already full 2272 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 2273 for ( uint i = 0; i < siz; i++ ) { 2274 Node *n = _available[i]; 2275 2276 // Skip projections, we'll handle them another way 2277 if (n->is_Proj()) 2278 continue; 2279 2280 // This presupposed that instructions are inserted into the 2281 // available list in a legality order; i.e. instructions that 2282 // must be inserted first are at the head of the list 2283 if (NodeFitsInBundle(n)) { 2284 #ifndef PRODUCT 2285 if (_cfg->C->trace_opto_output()) { 2286 tty->print("# ChooseNodeToBundle: "); 2287 n->dump(); 2288 } 2289 #endif 2290 return (n); 2291 } 2292 } 2293 } 2294 2295 // Nothing fits in this bundle, choose the highest priority 2296 #ifndef PRODUCT 2297 if (_cfg->C->trace_opto_output()) { 2298 tty->print("# ChooseNodeToBundle: "); 2299 _available[0]->dump(); 2300 } 2301 #endif 2302 2303 return _available[0]; 2304 } 2305 2306 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) { 2307 assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), ""); 2308 2309 OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1)); 2310 OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first); 2311 OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1)); 2312 OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second); 2313 2314 // Comparison between stack -> reg and stack -> reg 2315 if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) && 2316 OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) { 2317 return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo); 2318 } 2319 2320 // Comparison between reg -> stack and reg -> stack 2321 if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) && 2322 OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) { 2323 return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo); 2324 } 2325 2326 return 0; // Not comparable 2327 } 2328 2329 void Scheduling::AddNodeToAvailableList(Node *n) { 2330 assert( !n->is_Proj(), "projections never directly made available" ); 2331 #ifndef PRODUCT 2332 if (_cfg->C->trace_opto_output()) { 2333 tty->print("# AddNodeToAvailableList: "); 2334 n->dump(); 2335 } 2336 #endif 2337 2338 int latency = _current_latency[n->_idx]; 2339 2340 // Insert in latency order (insertion sort). If two MachSpillCopyNodes 2341 // for stack spilling or unspilling have the same latency, we sort 2342 // them in the order of stack offset. Some ports (e.g. aarch64) may also 2343 // have more opportunities to do ld/st merging 2344 uint i; 2345 for (i = 0; i < _available.size(); i++) { 2346 if (_current_latency[_available[i]->_idx] > latency) { 2347 break; 2348 } else if (_current_latency[_available[i]->_idx] == latency && 2349 n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() && 2350 compare_two_spill_nodes(n, _available[i]) > 0) { 2351 break; 2352 } 2353 } 2354 2355 // Special Check for compares following branches 2356 if( n->is_Mach() && _scheduled.size() > 0 ) { 2357 int op = n->as_Mach()->ideal_Opcode(); 2358 Node *last = _scheduled[0]; 2359 if( last->is_MachIf() && last->in(1) == n && 2360 ( op == Op_CmpI || 2361 op == Op_CmpU || 2362 op == Op_CmpUL || 2363 op == Op_CmpP || 2364 op == Op_CmpF || 2365 op == Op_CmpD || 2366 op == Op_CmpL ) ) { 2367 2368 // Recalculate position, moving to front of same latency 2369 for ( i=0 ; i < _available.size(); i++ ) 2370 if (_current_latency[_available[i]->_idx] >= latency) 2371 break; 2372 } 2373 } 2374 2375 // Insert the node in the available list 2376 _available.insert(i, n); 2377 2378 #ifndef PRODUCT 2379 if (_cfg->C->trace_opto_output()) 2380 dump_available(); 2381 #endif 2382 } 2383 2384 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 2385 for ( uint i=0; i < n->len(); i++ ) { 2386 Node *def = n->in(i); 2387 if (!def) continue; 2388 if( def->is_Proj() ) // If this is a machine projection, then 2389 def = def->in(0); // propagate usage thru to the base instruction 2390 2391 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local 2392 continue; 2393 } 2394 2395 // Compute the latency 2396 uint l = _bundle_cycle_number + n->latency(i); 2397 if (_current_latency[def->_idx] < l) 2398 _current_latency[def->_idx] = l; 2399 2400 // If this does not have uses then schedule it 2401 if ((--_uses[def->_idx]) == 0) 2402 AddNodeToAvailableList(def); 2403 } 2404 } 2405 2406 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2407 #ifndef PRODUCT 2408 if (_cfg->C->trace_opto_output()) { 2409 tty->print("# AddNodeToBundle: "); 2410 n->dump(); 2411 } 2412 #endif 2413 2414 // Remove this from the available list 2415 uint i; 2416 for (i = 0; i < _available.size(); i++) 2417 if (_available[i] == n) 2418 break; 2419 assert(i < _available.size(), "entry in _available list not found"); 2420 _available.remove(i); 2421 2422 // See if this fits in the current bundle 2423 const Pipeline *node_pipeline = n->pipeline(); 2424 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2425 2426 // Check for instructions to be placed in the delay slot. We 2427 // do this before we actually schedule the current instruction, 2428 // because the delay slot follows the current instruction. 2429 if (Pipeline::_branch_has_delay_slot && 2430 node_pipeline->hasBranchDelay() && 2431 !_unconditional_delay_slot) { 2432 2433 uint siz = _available.size(); 2434 2435 // Conditional branches can support an instruction that 2436 // is unconditionally executed and not dependent by the 2437 // branch, OR a conditionally executed instruction if 2438 // the branch is taken. In practice, this means that 2439 // the first instruction at the branch target is 2440 // copied to the delay slot, and the branch goes to 2441 // the instruction after that at the branch target 2442 if ( n->is_MachBranch() ) { 2443 2444 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2445 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2446 2447 #ifndef PRODUCT 2448 _branches++; 2449 #endif 2450 2451 // At least 1 instruction is on the available list 2452 // that is not dependent on the branch 2453 for (uint i = 0; i < siz; i++) { 2454 Node *d = _available[i]; 2455 const Pipeline *avail_pipeline = d->pipeline(); 2456 2457 // Don't allow safepoints in the branch shadow, that will 2458 // cause a number of difficulties 2459 if ( avail_pipeline->instructionCount() == 1 && 2460 !avail_pipeline->hasMultipleBundles() && 2461 !avail_pipeline->hasBranchDelay() && 2462 Pipeline::instr_has_unit_size() && 2463 d->size(_regalloc) == Pipeline::instr_unit_size() && 2464 NodeFitsInBundle(d) && 2465 !node_bundling(d)->used_in_delay()) { 2466 2467 if (d->is_Mach() && !d->is_MachSafePoint()) { 2468 // A node that fits in the delay slot was found, so we need to 2469 // set the appropriate bits in the bundle pipeline information so 2470 // that it correctly indicates resource usage. Later, when we 2471 // attempt to add this instruction to the bundle, we will skip 2472 // setting the resource usage. 2473 _unconditional_delay_slot = d; 2474 node_bundling(n)->set_use_unconditional_delay(); 2475 node_bundling(d)->set_used_in_unconditional_delay(); 2476 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2477 _current_latency[d->_idx] = _bundle_cycle_number; 2478 _next_node = d; 2479 ++_bundle_instr_count; 2480 #ifndef PRODUCT 2481 _unconditional_delays++; 2482 #endif 2483 break; 2484 } 2485 } 2486 } 2487 } 2488 2489 // No delay slot, add a nop to the usage 2490 if (!_unconditional_delay_slot) { 2491 // See if adding an instruction in the delay slot will overflow 2492 // the bundle. 2493 if (!NodeFitsInBundle(_nop)) { 2494 #ifndef PRODUCT 2495 if (_cfg->C->trace_opto_output()) 2496 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2497 #endif 2498 step(1); 2499 } 2500 2501 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2502 _next_node = _nop; 2503 ++_bundle_instr_count; 2504 } 2505 2506 // See if the instruction in the delay slot requires a 2507 // step of the bundles 2508 if (!NodeFitsInBundle(n)) { 2509 #ifndef PRODUCT 2510 if (_cfg->C->trace_opto_output()) 2511 tty->print("# *** STEP(branch won't fit) ***\n"); 2512 #endif 2513 // Update the state information 2514 _bundle_instr_count = 0; 2515 _bundle_cycle_number += 1; 2516 _bundle_use.step(1); 2517 } 2518 } 2519 2520 // Get the number of instructions 2521 uint instruction_count = node_pipeline->instructionCount(); 2522 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2523 instruction_count = 0; 2524 2525 // Compute the latency information 2526 uint delay = 0; 2527 2528 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2529 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2530 if (relative_latency < 0) 2531 relative_latency = 0; 2532 2533 delay = _bundle_use.full_latency(relative_latency, node_usage); 2534 2535 // Does not fit in this bundle, start a new one 2536 if (delay > 0) { 2537 step(delay); 2538 2539 #ifndef PRODUCT 2540 if (_cfg->C->trace_opto_output()) 2541 tty->print("# *** STEP(%d) ***\n", delay); 2542 #endif 2543 } 2544 } 2545 2546 // If this was placed in the delay slot, ignore it 2547 if (n != _unconditional_delay_slot) { 2548 2549 if (delay == 0) { 2550 if (node_pipeline->hasMultipleBundles()) { 2551 #ifndef PRODUCT 2552 if (_cfg->C->trace_opto_output()) 2553 tty->print("# *** STEP(multiple instructions) ***\n"); 2554 #endif 2555 step(1); 2556 } 2557 2558 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2559 #ifndef PRODUCT 2560 if (_cfg->C->trace_opto_output()) 2561 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2562 instruction_count + _bundle_instr_count, 2563 Pipeline::_max_instrs_per_cycle); 2564 #endif 2565 step(1); 2566 } 2567 } 2568 2569 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2570 _bundle_instr_count++; 2571 2572 // Set the node's latency 2573 _current_latency[n->_idx] = _bundle_cycle_number; 2574 2575 // Now merge the functional unit information 2576 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2577 _bundle_use.add_usage(node_usage); 2578 2579 // Increment the number of instructions in this bundle 2580 _bundle_instr_count += instruction_count; 2581 2582 // Remember this node for later 2583 if (n->is_Mach()) 2584 _next_node = n; 2585 } 2586 2587 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2588 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2589 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2590 // into the block. All other scheduled nodes get put in the schedule here. 2591 int op = n->Opcode(); 2592 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2593 (op != Op_Node && // Not an unused antidepedence node and 2594 // not an unallocated boxlock 2595 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2596 2597 // Push any trailing projections 2598 if( bb->get_node(bb->number_of_nodes()-1) != n ) { 2599 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2600 Node *foi = n->fast_out(i); 2601 if( foi->is_Proj() ) 2602 _scheduled.push(foi); 2603 } 2604 } 2605 2606 // Put the instruction in the schedule list 2607 _scheduled.push(n); 2608 } 2609 2610 #ifndef PRODUCT 2611 if (_cfg->C->trace_opto_output()) 2612 dump_available(); 2613 #endif 2614 2615 // Walk all the definitions, decrementing use counts, and 2616 // if a definition has a 0 use count, place it in the available list. 2617 DecrementUseCounts(n,bb); 2618 } 2619 2620 // This method sets the use count within a basic block. We will ignore all 2621 // uses outside the current basic block. As we are doing a backwards walk, 2622 // any node we reach that has a use count of 0 may be scheduled. This also 2623 // avoids the problem of cyclic references from phi nodes, as long as phi 2624 // nodes are at the front of the basic block. This method also initializes 2625 // the available list to the set of instructions that have no uses within this 2626 // basic block. 2627 void Scheduling::ComputeUseCount(const Block *bb) { 2628 #ifndef PRODUCT 2629 if (_cfg->C->trace_opto_output()) 2630 tty->print("# -> ComputeUseCount\n"); 2631 #endif 2632 2633 // Clear the list of available and scheduled instructions, just in case 2634 _available.clear(); 2635 _scheduled.clear(); 2636 2637 // No delay slot specified 2638 _unconditional_delay_slot = nullptr; 2639 2640 #ifdef ASSERT 2641 for( uint i=0; i < bb->number_of_nodes(); i++ ) 2642 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" ); 2643 #endif 2644 2645 // Force the _uses count to never go to zero for unscheduable pieces 2646 // of the block 2647 for( uint k = 0; k < _bb_start; k++ ) 2648 _uses[bb->get_node(k)->_idx] = 1; 2649 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ ) 2650 _uses[bb->get_node(l)->_idx] = 1; 2651 2652 // Iterate backwards over the instructions in the block. Don't count the 2653 // branch projections at end or the block header instructions. 2654 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2655 Node *n = bb->get_node(j); 2656 if( n->is_Proj() ) continue; // Projections handled another way 2657 2658 // Account for all uses 2659 for ( uint k = 0; k < n->len(); k++ ) { 2660 Node *inp = n->in(k); 2661 if (!inp) continue; 2662 assert(inp != n, "no cycles allowed" ); 2663 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use? 2664 if (inp->is_Proj()) { // Skip through Proj's 2665 inp = inp->in(0); 2666 } 2667 ++_uses[inp->_idx]; // Count 1 block-local use 2668 } 2669 } 2670 2671 // If this instruction has a 0 use count, then it is available 2672 if (!_uses[n->_idx]) { 2673 _current_latency[n->_idx] = _bundle_cycle_number; 2674 AddNodeToAvailableList(n); 2675 } 2676 2677 #ifndef PRODUCT 2678 if (_cfg->C->trace_opto_output()) { 2679 tty->print("# uses: %3d: ", _uses[n->_idx]); 2680 n->dump(); 2681 } 2682 #endif 2683 } 2684 2685 #ifndef PRODUCT 2686 if (_cfg->C->trace_opto_output()) 2687 tty->print("# <- ComputeUseCount\n"); 2688 #endif 2689 } 2690 2691 // This routine performs scheduling on each basic block in reverse order, 2692 // using instruction latencies and taking into account function unit 2693 // availability. 2694 void Scheduling::DoScheduling() { 2695 #ifndef PRODUCT 2696 if (_cfg->C->trace_opto_output()) 2697 tty->print("# -> DoScheduling\n"); 2698 #endif 2699 2700 Block *succ_bb = nullptr; 2701 Block *bb; 2702 Compile* C = Compile::current(); 2703 2704 // Walk over all the basic blocks in reverse order 2705 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) { 2706 bb = _cfg->get_block(i); 2707 2708 #ifndef PRODUCT 2709 if (_cfg->C->trace_opto_output()) { 2710 tty->print("# Schedule BB#%03d (initial)\n", i); 2711 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2712 bb->get_node(j)->dump(); 2713 } 2714 } 2715 #endif 2716 2717 // On the head node, skip processing 2718 if (bb == _cfg->get_root_block()) { 2719 continue; 2720 } 2721 2722 // Skip empty, connector blocks 2723 if (bb->is_connector()) 2724 continue; 2725 2726 // If the following block is not the sole successor of 2727 // this one, then reset the pipeline information 2728 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2729 #ifndef PRODUCT 2730 if (_cfg->C->trace_opto_output()) { 2731 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2732 _next_node->_idx, _bundle_instr_count); 2733 } 2734 #endif 2735 step_and_clear(); 2736 } 2737 2738 // Leave untouched the starting instruction, any Phis, a CreateEx node 2739 // or Top. bb->get_node(_bb_start) is the first schedulable instruction. 2740 _bb_end = bb->number_of_nodes()-1; 2741 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2742 Node *n = bb->get_node(_bb_start); 2743 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2744 // Also, MachIdealNodes do not get scheduled 2745 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2746 MachNode *mach = n->as_Mach(); 2747 int iop = mach->ideal_Opcode(); 2748 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2749 if( iop == Op_Con ) continue; // Do not schedule Top 2750 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2751 mach->pipeline() == MachNode::pipeline_class() && 2752 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc 2753 continue; 2754 break; // Funny loop structure to be sure... 2755 } 2756 // Compute last "interesting" instruction in block - last instruction we 2757 // might schedule. _bb_end points just after last schedulable inst. We 2758 // normally schedule conditional branches (despite them being forced last 2759 // in the block), because they have delay slots we can fill. Calls all 2760 // have their delay slots filled in the template expansions, so we don't 2761 // bother scheduling them. 2762 Node *last = bb->get_node(_bb_end); 2763 // Ignore trailing NOPs. 2764 while (_bb_end > 0 && last->is_Mach() && 2765 last->as_Mach()->ideal_Opcode() == Op_Con) { 2766 last = bb->get_node(--_bb_end); 2767 } 2768 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2769 if( last->is_Catch() || 2770 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2771 // There might be a prior call. Skip it. 2772 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj()); 2773 } else if( last->is_MachNullCheck() ) { 2774 // Backup so the last null-checked memory instruction is 2775 // outside the schedulable range. Skip over the nullcheck, 2776 // projection, and the memory nodes. 2777 Node *mem = last->in(1); 2778 do { 2779 _bb_end--; 2780 } while (mem != bb->get_node(_bb_end)); 2781 } else { 2782 // Set _bb_end to point after last schedulable inst. 2783 _bb_end++; 2784 } 2785 2786 assert( _bb_start <= _bb_end, "inverted block ends" ); 2787 2788 // Compute the register antidependencies for the basic block 2789 ComputeRegisterAntidependencies(bb); 2790 if (C->failing()) return; // too many D-U pinch points 2791 2792 // Compute the usage within the block, and set the list of all nodes 2793 // in the block that have no uses within the block. 2794 ComputeUseCount(bb); 2795 2796 // Schedule the remaining instructions in the block 2797 while ( _available.size() > 0 ) { 2798 Node *n = ChooseNodeToBundle(); 2799 guarantee(n != nullptr, "no nodes available"); 2800 AddNodeToBundle(n,bb); 2801 } 2802 2803 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2804 #ifdef ASSERT 2805 for( uint l = _bb_start; l < _bb_end; l++ ) { 2806 Node *n = bb->get_node(l); 2807 uint m; 2808 for( m = 0; m < _bb_end-_bb_start; m++ ) 2809 if( _scheduled[m] == n ) 2810 break; 2811 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2812 } 2813 #endif 2814 2815 // Now copy the instructions (in reverse order) back to the block 2816 for ( uint k = _bb_start; k < _bb_end; k++ ) 2817 bb->map_node(_scheduled[_bb_end-k-1], k); 2818 2819 #ifndef PRODUCT 2820 if (_cfg->C->trace_opto_output()) { 2821 tty->print("# Schedule BB#%03d (final)\n", i); 2822 uint current = 0; 2823 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2824 Node *n = bb->get_node(j); 2825 if( valid_bundle_info(n) ) { 2826 Bundle *bundle = node_bundling(n); 2827 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2828 tty->print("*** Bundle: "); 2829 bundle->dump(); 2830 } 2831 n->dump(); 2832 } 2833 } 2834 } 2835 #endif 2836 #ifdef ASSERT 2837 verify_good_schedule(bb,"after block local scheduling"); 2838 #endif 2839 } 2840 2841 #ifndef PRODUCT 2842 if (_cfg->C->trace_opto_output()) 2843 tty->print("# <- DoScheduling\n"); 2844 #endif 2845 2846 // Record final node-bundling array location 2847 _regalloc->C->output()->set_node_bundling_base(_node_bundling_base); 2848 2849 } // end DoScheduling 2850 2851 // Verify that no live-range used in the block is killed in the block by a 2852 // wrong DEF. This doesn't verify live-ranges that span blocks. 2853 2854 // Check for edge existence. Used to avoid adding redundant precedence edges. 2855 static bool edge_from_to( Node *from, Node *to ) { 2856 for( uint i=0; i<from->len(); i++ ) 2857 if( from->in(i) == to ) 2858 return true; 2859 return false; 2860 } 2861 2862 #ifdef ASSERT 2863 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2864 // Check for bad kills 2865 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2866 Node *prior_use = _reg_node[def]; 2867 if( prior_use && !edge_from_to(prior_use,n) ) { 2868 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2869 n->dump(); 2870 tty->print_cr("..."); 2871 prior_use->dump(); 2872 assert(edge_from_to(prior_use,n), "%s", msg); 2873 } 2874 _reg_node.map(def,nullptr); // Kill live USEs 2875 } 2876 } 2877 2878 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2879 2880 // Zap to something reasonable for the verify code 2881 _reg_node.clear(); 2882 2883 // Walk over the block backwards. Check to make sure each DEF doesn't 2884 // kill a live value (other than the one it's supposed to). Add each 2885 // USE to the live set. 2886 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) { 2887 Node *n = b->get_node(i); 2888 int n_op = n->Opcode(); 2889 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2890 // Fat-proj kills a slew of registers 2891 RegMaskIterator rmi(n->out_RegMask()); 2892 while (rmi.has_next()) { 2893 OptoReg::Name kill = rmi.next(); 2894 verify_do_def(n, kill, msg); 2895 } 2896 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2897 // Get DEF'd registers the normal way 2898 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2899 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2900 } 2901 2902 // Now make all USEs live 2903 for( uint i=1; i<n->req(); i++ ) { 2904 Node *def = n->in(i); 2905 assert(def != 0, "input edge required"); 2906 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2907 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2908 if( OptoReg::is_valid(reg_lo) ) { 2909 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg); 2910 _reg_node.map(reg_lo,n); 2911 } 2912 if( OptoReg::is_valid(reg_hi) ) { 2913 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg); 2914 _reg_node.map(reg_hi,n); 2915 } 2916 } 2917 2918 } 2919 2920 // Zap to something reasonable for the Antidependence code 2921 _reg_node.clear(); 2922 } 2923 #endif 2924 2925 // Conditionally add precedence edges. Avoid putting edges on Projs. 2926 static void add_prec_edge_from_to( Node *from, Node *to ) { 2927 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2928 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2929 from = from->in(0); 2930 } 2931 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2932 !edge_from_to( from, to ) ) // Avoid duplicate edge 2933 from->add_prec(to); 2934 } 2935 2936 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2937 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2938 return; 2939 2940 if (OptoReg::is_reg(def_reg)) { 2941 VMReg vmreg = OptoReg::as_VMReg(def_reg); 2942 if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) { 2943 // This is one of the high slots of a vector register. 2944 // ScheduleAndBundle already checked there are no live wide 2945 // vectors in this method so it can be safely ignored. 2946 return; 2947 } 2948 } 2949 2950 Node *pinch = _reg_node[def_reg]; // Get pinch point 2951 if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet? 2952 is_def ) { // Check for a true def (not a kill) 2953 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2954 return; 2955 } 2956 2957 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2958 debug_only( def = (Node*)((intptr_t)0xdeadbeef); ) 2959 2960 // After some number of kills there _may_ be a later def 2961 Node *later_def = nullptr; 2962 2963 Compile* C = Compile::current(); 2964 2965 // Finding a kill requires a real pinch-point. 2966 // Check for not already having a pinch-point. 2967 // Pinch points are Op_Node's. 2968 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2969 later_def = pinch; // Must be def/kill as optimistic pinch-point 2970 if ( _pinch_free_list.size() > 0) { 2971 pinch = _pinch_free_list.pop(); 2972 } else { 2973 pinch = new Node(1); // Pinch point to-be 2974 } 2975 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2976 DEBUG_ONLY( pinch->dump(); ); 2977 assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index()); 2978 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2979 return; 2980 } 2981 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init) 2982 _reg_node.map(def_reg,pinch); // Record pinch-point 2983 //regalloc()->set_bad(pinch->_idx); // Already initialized this way. 2984 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2985 pinch->init_req(0, C->top()); // set not null for the next call 2986 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2987 later_def = nullptr; // and no later def 2988 } 2989 pinch->set_req(0,later_def); // Hook later def so we can find it 2990 } else { // Else have valid pinch point 2991 if( pinch->in(0) ) // If there is a later-def 2992 later_def = pinch->in(0); // Get it 2993 } 2994 2995 // Add output-dependence edge from later def to kill 2996 if( later_def ) // If there is some original def 2997 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2998 2999 // See if current kill is also a use, and so is forced to be the pinch-point. 3000 if( pinch->Opcode() == Op_Node ) { 3001 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 3002 for( uint i=1; i<uses->req(); i++ ) { 3003 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 3004 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 3005 // Yes, found a use/kill pinch-point 3006 pinch->set_req(0,nullptr); // 3007 pinch->replace_by(kill); // Move anti-dep edges up 3008 pinch = kill; 3009 _reg_node.map(def_reg,pinch); 3010 return; 3011 } 3012 } 3013 } 3014 3015 // Add edge from kill to pinch-point 3016 add_prec_edge_from_to(kill,pinch); 3017 } 3018 3019 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 3020 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 3021 return; 3022 Node *pinch = _reg_node[use_reg]; // Get pinch point 3023 // Check for no later def_reg/kill in block 3024 if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b && 3025 // Use has to be block-local as well 3026 _cfg->get_block_for_node(use) == b) { 3027 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 3028 pinch->req() == 1 ) { // pinch not yet in block? 3029 pinch->del_req(0); // yank pointer to later-def, also set flag 3030 // Insert the pinch-point in the block just after the last use 3031 b->insert_node(pinch, b->find_node(use) + 1); 3032 _bb_end++; // Increase size scheduled region in block 3033 } 3034 3035 add_prec_edge_from_to(pinch,use); 3036 } 3037 } 3038 3039 // We insert antidependences between the reads and following write of 3040 // allocated registers to prevent illegal code motion. Hopefully, the 3041 // number of added references should be fairly small, especially as we 3042 // are only adding references within the current basic block. 3043 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 3044 3045 #ifdef ASSERT 3046 verify_good_schedule(b,"before block local scheduling"); 3047 #endif 3048 3049 // A valid schedule, for each register independently, is an endless cycle 3050 // of: a def, then some uses (connected to the def by true dependencies), 3051 // then some kills (defs with no uses), finally the cycle repeats with a new 3052 // def. The uses are allowed to float relative to each other, as are the 3053 // kills. No use is allowed to slide past a kill (or def). This requires 3054 // antidependencies between all uses of a single def and all kills that 3055 // follow, up to the next def. More edges are redundant, because later defs 3056 // & kills are already serialized with true or antidependencies. To keep 3057 // the edge count down, we add a 'pinch point' node if there's more than 3058 // one use or more than one kill/def. 3059 3060 // We add dependencies in one bottom-up pass. 3061 3062 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 3063 3064 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 3065 // register. If not, we record the DEF/KILL in _reg_node, the 3066 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 3067 // "pinch point", a new Node that's in the graph but not in the block. 3068 // We put edges from the prior and current DEF/KILLs to the pinch point. 3069 // We put the pinch point in _reg_node. If there's already a pinch point 3070 // we merely add an edge from the current DEF/KILL to the pinch point. 3071 3072 // After doing the DEF/KILLs, we handle USEs. For each used register, we 3073 // put an edge from the pinch point to the USE. 3074 3075 // To be expedient, the _reg_node array is pre-allocated for the whole 3076 // compilation. _reg_node is lazily initialized; it either contains a null, 3077 // or a valid def/kill/pinch-point, or a leftover node from some prior 3078 // block. Leftover node from some prior block is treated like a null (no 3079 // prior def, so no anti-dependence needed). Valid def is distinguished by 3080 // it being in the current block. 3081 bool fat_proj_seen = false; 3082 uint last_safept = _bb_end-1; 3083 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr; 3084 Node* last_safept_node = end_node; 3085 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 3086 Node *n = b->get_node(i); 3087 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 3088 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 3089 // Fat-proj kills a slew of registers 3090 // This can add edges to 'n' and obscure whether or not it was a def, 3091 // hence the is_def flag. 3092 fat_proj_seen = true; 3093 RegMaskIterator rmi(n->out_RegMask()); 3094 while (rmi.has_next()) { 3095 OptoReg::Name kill = rmi.next(); 3096 anti_do_def(b, n, kill, is_def); 3097 } 3098 } else { 3099 // Get DEF'd registers the normal way 3100 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 3101 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 3102 } 3103 3104 // Kill projections on a branch should appear to occur on the 3105 // branch, not afterwards, so grab the masks from the projections 3106 // and process them. 3107 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) { 3108 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3109 Node* use = n->fast_out(i); 3110 if (use->is_Proj()) { 3111 RegMaskIterator rmi(use->out_RegMask()); 3112 while (rmi.has_next()) { 3113 OptoReg::Name kill = rmi.next(); 3114 anti_do_def(b, n, kill, false); 3115 } 3116 } 3117 } 3118 } 3119 3120 // Check each register used by this instruction for a following DEF/KILL 3121 // that must occur afterward and requires an anti-dependence edge. 3122 for( uint j=0; j<n->req(); j++ ) { 3123 Node *def = n->in(j); 3124 if( def ) { 3125 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 3126 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 3127 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 3128 } 3129 } 3130 // Do not allow defs of new derived values to float above GC 3131 // points unless the base is definitely available at the GC point. 3132 3133 Node *m = b->get_node(i); 3134 3135 // Add precedence edge from following safepoint to use of derived pointer 3136 if( last_safept_node != end_node && 3137 m != last_safept_node) { 3138 for (uint k = 1; k < m->req(); k++) { 3139 const Type *t = m->in(k)->bottom_type(); 3140 if( t->isa_oop_ptr() && 3141 t->is_ptr()->offset() != 0 ) { 3142 last_safept_node->add_prec( m ); 3143 break; 3144 } 3145 } 3146 } 3147 3148 if( n->jvms() ) { // Precedence edge from derived to safept 3149 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 3150 if( b->get_node(last_safept) != last_safept_node ) { 3151 last_safept = b->find_node(last_safept_node); 3152 } 3153 for( uint j=last_safept; j > i; j-- ) { 3154 Node *mach = b->get_node(j); 3155 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 3156 mach->add_prec( n ); 3157 } 3158 last_safept = i; 3159 last_safept_node = m; 3160 } 3161 } 3162 3163 if (fat_proj_seen) { 3164 // Garbage collect pinch nodes that were not consumed. 3165 // They are usually created by a fat kill MachProj for a call. 3166 garbage_collect_pinch_nodes(); 3167 } 3168 } 3169 3170 // Garbage collect pinch nodes for reuse by other blocks. 3171 // 3172 // The block scheduler's insertion of anti-dependence 3173 // edges creates many pinch nodes when the block contains 3174 // 2 or more Calls. A pinch node is used to prevent a 3175 // combinatorial explosion of edges. If a set of kills for a 3176 // register is anti-dependent on a set of uses (or defs), rather 3177 // than adding an edge in the graph between each pair of kill 3178 // and use (or def), a pinch is inserted between them: 3179 // 3180 // use1 use2 use3 3181 // \ | / 3182 // \ | / 3183 // pinch 3184 // / | \ 3185 // / | \ 3186 // kill1 kill2 kill3 3187 // 3188 // One pinch node is created per register killed when 3189 // the second call is encountered during a backwards pass 3190 // over the block. Most of these pinch nodes are never 3191 // wired into the graph because the register is never 3192 // used or def'ed in the block. 3193 // 3194 void Scheduling::garbage_collect_pinch_nodes() { 3195 #ifndef PRODUCT 3196 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 3197 #endif 3198 int trace_cnt = 0; 3199 for (uint k = 0; k < _reg_node.max(); k++) { 3200 Node* pinch = _reg_node[k]; 3201 if ((pinch != nullptr) && pinch->Opcode() == Op_Node && 3202 // no predecence input edges 3203 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) { 3204 cleanup_pinch(pinch); 3205 _pinch_free_list.push(pinch); 3206 _reg_node.map(k, nullptr); 3207 #ifndef PRODUCT 3208 if (_cfg->C->trace_opto_output()) { 3209 trace_cnt++; 3210 if (trace_cnt > 40) { 3211 tty->print("\n"); 3212 trace_cnt = 0; 3213 } 3214 tty->print(" %d", pinch->_idx); 3215 } 3216 #endif 3217 } 3218 } 3219 #ifndef PRODUCT 3220 if (_cfg->C->trace_opto_output()) tty->print("\n"); 3221 #endif 3222 } 3223 3224 // Clean up a pinch node for reuse. 3225 void Scheduling::cleanup_pinch( Node *pinch ) { 3226 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 3227 3228 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 3229 Node* use = pinch->last_out(i); 3230 uint uses_found = 0; 3231 for (uint j = use->req(); j < use->len(); j++) { 3232 if (use->in(j) == pinch) { 3233 use->rm_prec(j); 3234 uses_found++; 3235 } 3236 } 3237 assert(uses_found > 0, "must be a precedence edge"); 3238 i -= uses_found; // we deleted 1 or more copies of this edge 3239 } 3240 // May have a later_def entry 3241 pinch->set_req(0, nullptr); 3242 } 3243 3244 #ifndef PRODUCT 3245 3246 void Scheduling::dump_available() const { 3247 tty->print("#Availist "); 3248 for (uint i = 0; i < _available.size(); i++) 3249 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 3250 tty->cr(); 3251 } 3252 3253 // Print Scheduling Statistics 3254 void Scheduling::print_statistics() { 3255 // Print the size added by nops for bundling 3256 tty->print("Nops added %d bytes to total of %d bytes", 3257 _total_nop_size, _total_method_size); 3258 if (_total_method_size > 0) 3259 tty->print(", for %.2f%%", 3260 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 3261 tty->print("\n"); 3262 3263 // Print the number of branch shadows filled 3264 if (Pipeline::_branch_has_delay_slot) { 3265 tty->print("Of %d branches, %d had unconditional delay slots filled", 3266 _total_branches, _total_unconditional_delays); 3267 if (_total_branches > 0) 3268 tty->print(", for %.2f%%", 3269 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 3270 tty->print("\n"); 3271 } 3272 3273 uint total_instructions = 0, total_bundles = 0; 3274 3275 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 3276 uint bundle_count = _total_instructions_per_bundle[i]; 3277 total_instructions += bundle_count * i; 3278 total_bundles += bundle_count; 3279 } 3280 3281 if (total_bundles > 0) 3282 tty->print("Average ILP (excluding nops) is %.2f\n", 3283 ((double)total_instructions) / ((double)total_bundles)); 3284 } 3285 #endif 3286 3287 //-----------------------init_scratch_buffer_blob------------------------------ 3288 // Construct a temporary BufferBlob and cache it for this compile. 3289 void PhaseOutput::init_scratch_buffer_blob(int const_size) { 3290 // If there is already a scratch buffer blob allocated and the 3291 // constant section is big enough, use it. Otherwise free the 3292 // current and allocate a new one. 3293 BufferBlob* blob = scratch_buffer_blob(); 3294 if ((blob != nullptr) && (const_size <= _scratch_const_size)) { 3295 // Use the current blob. 3296 } else { 3297 if (blob != nullptr) { 3298 BufferBlob::free(blob); 3299 } 3300 3301 ResourceMark rm; 3302 _scratch_const_size = const_size; 3303 int size = C2Compiler::initial_code_buffer_size(const_size); 3304 blob = BufferBlob::create("Compile::scratch_buffer", size); 3305 // Record the buffer blob for next time. 3306 set_scratch_buffer_blob(blob); 3307 // Have we run out of code space? 3308 if (scratch_buffer_blob() == nullptr) { 3309 // Let CompilerBroker disable further compilations. 3310 C->record_failure("Not enough space for scratch buffer in CodeCache"); 3311 return; 3312 } 3313 } 3314 3315 // Initialize the relocation buffers 3316 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 3317 set_scratch_locs_memory(locs_buf); 3318 } 3319 3320 3321 //-----------------------scratch_emit_size------------------------------------- 3322 // Helper function that computes size by emitting code 3323 uint PhaseOutput::scratch_emit_size(const Node* n) { 3324 // Start scratch_emit_size section. 3325 set_in_scratch_emit_size(true); 3326 3327 // Emit into a trash buffer and count bytes emitted. 3328 // This is a pretty expensive way to compute a size, 3329 // but it works well enough if seldom used. 3330 // All common fixed-size instructions are given a size 3331 // method by the AD file. 3332 // Note that the scratch buffer blob and locs memory are 3333 // allocated at the beginning of the compile task, and 3334 // may be shared by several calls to scratch_emit_size. 3335 // The allocation of the scratch buffer blob is particularly 3336 // expensive, since it has to grab the code cache lock. 3337 BufferBlob* blob = this->scratch_buffer_blob(); 3338 assert(blob != nullptr, "Initialize BufferBlob at start"); 3339 assert(blob->size() > MAX_inst_size, "sanity"); 3340 relocInfo* locs_buf = scratch_locs_memory(); 3341 address blob_begin = blob->content_begin(); 3342 address blob_end = (address)locs_buf; 3343 assert(blob->contains(blob_end), "sanity"); 3344 CodeBuffer buf(blob_begin, blob_end - blob_begin); 3345 buf.initialize_consts_size(_scratch_const_size); 3346 buf.initialize_stubs_size(MAX_stubs_size); 3347 assert(locs_buf != nullptr, "sanity"); 3348 int lsize = MAX_locs_size / 3; 3349 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 3350 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 3351 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 3352 // Mark as scratch buffer. 3353 buf.consts()->set_scratch_emit(); 3354 buf.insts()->set_scratch_emit(); 3355 buf.stubs()->set_scratch_emit(); 3356 3357 // Do the emission. 3358 3359 Label fakeL; // Fake label for branch instructions. 3360 Label* saveL = nullptr; 3361 uint save_bnum = 0; 3362 bool is_branch = n->is_MachBranch(); 3363 C2_MacroAssembler masm(&buf); 3364 masm.bind(fakeL); 3365 if (is_branch) { 3366 n->as_MachBranch()->save_label(&saveL, &save_bnum); 3367 n->as_MachBranch()->label_set(&fakeL, 0); 3368 } 3369 n->emit(&masm, C->regalloc()); 3370 3371 // Emitting into the scratch buffer should not fail 3372 assert (!C->failing(), "Must not have pending failure. Reason is: %s", C->failure_reason()); 3373 3374 if (is_branch) // Restore label. 3375 n->as_MachBranch()->label_set(saveL, save_bnum); 3376 3377 // End scratch_emit_size section. 3378 set_in_scratch_emit_size(false); 3379 3380 return buf.insts_size(); 3381 } 3382 3383 void PhaseOutput::install() { 3384 if (C->should_install_code() && C->stub_function() != nullptr) { 3385 install_stub(C->stub_name()); 3386 } else { 3387 install_code(C->method(), 3388 C->entry_bci(), 3389 CompilerThread::current()->compiler(), 3390 C->has_unsafe_access(), 3391 SharedRuntime::is_wide_vector(C->max_vector_size()), 3392 C->rtm_state()); 3393 } 3394 } 3395 3396 void PhaseOutput::install_code(ciMethod* target, 3397 int entry_bci, 3398 AbstractCompiler* compiler, 3399 bool has_unsafe_access, 3400 bool has_wide_vectors, 3401 RTMState rtm_state) { 3402 // Check if we want to skip execution of all compiled code. 3403 { 3404 #ifndef PRODUCT 3405 if (OptoNoExecute) { 3406 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed 3407 return; 3408 } 3409 #endif 3410 Compile::TracePhase tp("install_code", &timers[_t_registerMethod]); 3411 3412 if (C->is_osr_compilation()) { 3413 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 3414 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 3415 } else { 3416 if (!target->is_static()) { 3417 // The UEP of an nmethod ensures that the VEP is padded. However, the padding of the UEP is placed 3418 // before the inline cache check, so we don't have to execute any nop instructions when dispatching 3419 // through the UEP, yet we can ensure that the VEP is aligned appropriately. 3420 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size - MacroAssembler::ic_check_size()); 3421 } 3422 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 3423 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 3424 } 3425 3426 C->env()->register_method(target, 3427 entry_bci, 3428 &_code_offsets, 3429 _orig_pc_slot_offset_in_bytes, 3430 code_buffer(), 3431 frame_size_in_words(), 3432 oop_map_set(), 3433 &_handler_table, 3434 inc_table(), 3435 compiler, 3436 C->has_clinit_barriers(), 3437 C->for_preload(), 3438 has_unsafe_access, 3439 SharedRuntime::is_wide_vector(C->max_vector_size()), 3440 C->has_monitors(), 3441 0, 3442 C->should_install_code(), 3443 C->rtm_state()); 3444 3445 if (C->log() != nullptr) { // Print code cache state into compiler log 3446 C->log()->code_cache_state(); 3447 } 3448 if (C->has_clinit_barriers()) { 3449 assert(C->for_preload(), "sanity"); 3450 // Build second version of code without class initialization barriers 3451 if (C->env()->task()->compile_reason() == CompileTask::Reason_PrecompileForPreload) { 3452 // don't automatically precompile a barrier-free version unless explicitly asked 3453 } else { 3454 C->record_failure(C2Compiler::retry_no_clinit_barriers()); 3455 } 3456 } 3457 } 3458 } 3459 void PhaseOutput::install_stub(const char* stub_name) { 3460 // Entry point will be accessed using stub_entry_point(); 3461 if (code_buffer() == nullptr) { 3462 Matcher::soft_match_failure(); 3463 } else { 3464 if (PrintAssembly && (WizardMode || Verbose)) 3465 tty->print_cr("### Stub::%s", stub_name); 3466 3467 if (!C->failing()) { 3468 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs"); 3469 3470 // Make the NMethod 3471 // For now we mark the frame as never safe for profile stackwalking 3472 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 3473 code_buffer(), 3474 CodeOffsets::frame_never_safe, 3475 // _code_offsets.value(CodeOffsets::Frame_Complete), 3476 frame_size_in_words(), 3477 oop_map_set(), 3478 false); 3479 assert(rs != nullptr && rs->is_runtime_stub(), "sanity check"); 3480 3481 C->set_stub_entry_point(rs->entry_point()); 3482 } 3483 } 3484 } 3485 3486 // Support for bundling info 3487 Bundle* PhaseOutput::node_bundling(const Node *n) { 3488 assert(valid_bundle_info(n), "oob"); 3489 return &_node_bundling_base[n->_idx]; 3490 } 3491 3492 bool PhaseOutput::valid_bundle_info(const Node *n) { 3493 return (_node_bundling_limit > n->_idx); 3494 } 3495 3496 //------------------------------frame_size_in_words----------------------------- 3497 // frame_slots in units of words 3498 int PhaseOutput::frame_size_in_words() const { 3499 // shift is 0 in LP32 and 1 in LP64 3500 const int shift = (LogBytesPerWord - LogBytesPerInt); 3501 int words = _frame_slots >> shift; 3502 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 3503 return words; 3504 } 3505 3506 // To bang the stack of this compiled method we use the stack size 3507 // that the interpreter would need in case of a deoptimization. This 3508 // removes the need to bang the stack in the deoptimization blob which 3509 // in turn simplifies stack overflow handling. 3510 int PhaseOutput::bang_size_in_bytes() const { 3511 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size()); 3512 } 3513 3514 //------------------------------dump_asm--------------------------------------- 3515 // Dump formatted assembly 3516 #if defined(SUPPORT_OPTO_ASSEMBLY) 3517 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) { 3518 3519 int pc_digits = 3; // #chars required for pc 3520 int sb_chars = 3; // #chars for "start bundle" indicator 3521 int tab_size = 8; 3522 if (pcs != nullptr) { 3523 int max_pc = 0; 3524 for (uint i = 0; i < pc_limit; i++) { 3525 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc; 3526 } 3527 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc 3528 } 3529 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size; 3530 3531 bool cut_short = false; 3532 st->print_cr("#"); 3533 st->print("# "); C->tf()->dump_on(st); st->cr(); 3534 st->print_cr("#"); 3535 3536 // For all blocks 3537 int pc = 0x0; // Program counter 3538 char starts_bundle = ' '; 3539 C->regalloc()->dump_frame(); 3540 3541 Node *n = nullptr; 3542 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) { 3543 if (VMThread::should_terminate()) { 3544 cut_short = true; 3545 break; 3546 } 3547 Block* block = C->cfg()->get_block(i); 3548 if (block->is_connector() && !Verbose) { 3549 continue; 3550 } 3551 n = block->head(); 3552 if ((pcs != nullptr) && (n->_idx < pc_limit)) { 3553 pc = pcs[n->_idx]; 3554 st->print("%*.*x", pc_digits, pc_digits, pc); 3555 } 3556 st->fill_to(prefix_len); 3557 block->dump_head(C->cfg(), st); 3558 if (block->is_connector()) { 3559 st->fill_to(prefix_len); 3560 st->print_cr("# Empty connector block"); 3561 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 3562 st->fill_to(prefix_len); 3563 st->print_cr("# Block is sole successor of call"); 3564 } 3565 3566 // For all instructions 3567 Node *delay = nullptr; 3568 for (uint j = 0; j < block->number_of_nodes(); j++) { 3569 if (VMThread::should_terminate()) { 3570 cut_short = true; 3571 break; 3572 } 3573 n = block->get_node(j); 3574 if (valid_bundle_info(n)) { 3575 Bundle* bundle = node_bundling(n); 3576 if (bundle->used_in_unconditional_delay()) { 3577 delay = n; 3578 continue; 3579 } 3580 if (bundle->starts_bundle()) { 3581 starts_bundle = '+'; 3582 } 3583 } 3584 3585 if (WizardMode) { 3586 n->dump(); 3587 } 3588 3589 if( !n->is_Region() && // Dont print in the Assembly 3590 !n->is_Phi() && // a few noisely useless nodes 3591 !n->is_Proj() && 3592 !n->is_MachTemp() && 3593 !n->is_SafePointScalarObject() && 3594 !n->is_Catch() && // Would be nice to print exception table targets 3595 !n->is_MergeMem() && // Not very interesting 3596 !n->is_top() && // Debug info table constants 3597 !(n->is_Con() && !n->is_Mach())// Debug info table constants 3598 ) { 3599 if ((pcs != nullptr) && (n->_idx < pc_limit)) { 3600 pc = pcs[n->_idx]; 3601 st->print("%*.*x", pc_digits, pc_digits, pc); 3602 } else { 3603 st->fill_to(pc_digits); 3604 } 3605 st->print(" %c ", starts_bundle); 3606 starts_bundle = ' '; 3607 st->fill_to(prefix_len); 3608 n->format(C->regalloc(), st); 3609 st->cr(); 3610 } 3611 3612 // If we have an instruction with a delay slot, and have seen a delay, 3613 // then back up and print it 3614 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 3615 // Coverity finding - Explicit null dereferenced. 3616 guarantee(delay != nullptr, "no unconditional delay instruction"); 3617 if (WizardMode) delay->dump(); 3618 3619 if (node_bundling(delay)->starts_bundle()) 3620 starts_bundle = '+'; 3621 if ((pcs != nullptr) && (n->_idx < pc_limit)) { 3622 pc = pcs[n->_idx]; 3623 st->print("%*.*x", pc_digits, pc_digits, pc); 3624 } else { 3625 st->fill_to(pc_digits); 3626 } 3627 st->print(" %c ", starts_bundle); 3628 starts_bundle = ' '; 3629 st->fill_to(prefix_len); 3630 delay->format(C->regalloc(), st); 3631 st->cr(); 3632 delay = nullptr; 3633 } 3634 3635 // Dump the exception table as well 3636 if( n->is_Catch() && (Verbose || WizardMode) ) { 3637 // Print the exception table for this offset 3638 _handler_table.print_subtable_for(pc); 3639 } 3640 st->bol(); // Make sure we start on a new line 3641 } 3642 st->cr(); // one empty line between blocks 3643 assert(cut_short || delay == nullptr, "no unconditional delay branch"); 3644 } // End of per-block dump 3645 3646 if (cut_short) st->print_cr("*** disassembly is cut short ***"); 3647 } 3648 #endif 3649 3650 #ifndef PRODUCT 3651 void PhaseOutput::print_statistics() { 3652 Scheduling::print_statistics(); 3653 } 3654 #endif