1 /* 2 * Copyright (c) 1998, 2025, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "asm/assembler.inline.hpp" 26 #include "code/aotCodeCache.hpp" 27 #include "code/compiledIC.hpp" 28 #include "code/debugInfo.hpp" 29 #include "code/debugInfoRec.hpp" 30 #include "compiler/compileBroker.hpp" 31 #include "compiler/compilerDirectives.hpp" 32 #include "compiler/disassembler.hpp" 33 #include "compiler/oopMap.hpp" 34 #include "gc/shared/barrierSet.hpp" 35 #include "gc/shared/c2/barrierSetC2.hpp" 36 #include "memory/allocation.hpp" 37 #include "opto/ad.hpp" 38 #include "opto/block.hpp" 39 #include "opto/c2_MacroAssembler.hpp" 40 #include "opto/c2compiler.hpp" 41 #include "opto/callnode.hpp" 42 #include "opto/cfgnode.hpp" 43 #include "opto/locknode.hpp" 44 #include "opto/machnode.hpp" 45 #include "opto/node.hpp" 46 #include "opto/optoreg.hpp" 47 #include "opto/output.hpp" 48 #include "opto/regalloc.hpp" 49 #include "opto/type.hpp" 50 #include "runtime/sharedRuntime.hpp" 51 #include "utilities/macros.hpp" 52 #include "utilities/powerOfTwo.hpp" 53 #include "utilities/xmlstream.hpp" 54 55 #ifndef PRODUCT 56 #define DEBUG_ARG(x) , x 57 #else 58 #define DEBUG_ARG(x) 59 #endif 60 61 //------------------------------Scheduling---------------------------------- 62 // This class contains all the information necessary to implement instruction 63 // scheduling and bundling. 64 class Scheduling { 65 66 private: 67 // Arena to use 68 Arena *_arena; 69 70 // Control-Flow Graph info 71 PhaseCFG *_cfg; 72 73 // Register Allocation info 74 PhaseRegAlloc *_regalloc; 75 76 // Number of nodes in the method 77 uint _node_bundling_limit; 78 79 // List of scheduled nodes. Generated in reverse order 80 Node_List _scheduled; 81 82 // List of nodes currently available for choosing for scheduling 83 Node_List _available; 84 85 // For each instruction beginning a bundle, the number of following 86 // nodes to be bundled with it. 87 Bundle *_node_bundling_base; 88 89 // Mapping from register to Node 90 Node_List _reg_node; 91 92 // Free list for pinch nodes. 93 Node_List _pinch_free_list; 94 95 // Number of uses of this node within the containing basic block. 96 short *_uses; 97 98 // Schedulable portion of current block. Skips Region/Phi/CreateEx up 99 // front, branch+proj at end. Also skips Catch/CProj (same as 100 // branch-at-end), plus just-prior exception-throwing call. 101 uint _bb_start, _bb_end; 102 103 // Latency from the end of the basic block as scheduled 104 unsigned short *_current_latency; 105 106 // Remember the next node 107 Node *_next_node; 108 109 // Length of the current bundle, in instructions 110 uint _bundle_instr_count; 111 112 // Current Cycle number, for computing latencies and bundling 113 uint _bundle_cycle_number; 114 115 // Bundle information 116 Pipeline_Use_Element _bundle_use_elements[resource_count]; 117 Pipeline_Use _bundle_use; 118 119 // Dump the available list 120 void dump_available() const; 121 122 public: 123 Scheduling(Arena *arena, Compile &compile); 124 125 // Step ahead "i" cycles 126 void step(uint i); 127 128 // Step ahead 1 cycle, and clear the bundle state (for example, 129 // at a branch target) 130 void step_and_clear(); 131 132 Bundle* node_bundling(const Node *n) { 133 assert(valid_bundle_info(n), "oob"); 134 return (&_node_bundling_base[n->_idx]); 135 } 136 137 bool valid_bundle_info(const Node *n) const { 138 return (_node_bundling_limit > n->_idx); 139 } 140 141 bool starts_bundle(const Node *n) const { 142 return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle()); 143 } 144 145 // Do the scheduling 146 void DoScheduling(); 147 148 // Compute the register antidependencies within a basic block 149 void ComputeRegisterAntidependencies(Block *bb); 150 void verify_do_def( Node *n, OptoReg::Name def, const char *msg ); 151 void verify_good_schedule( Block *b, const char *msg ); 152 void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ); 153 void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ); 154 155 // Add a node to the current bundle 156 void AddNodeToBundle(Node *n, const Block *bb); 157 158 // Return an integer less than, equal to, or greater than zero 159 // if the stack offset of the first argument is respectively 160 // less than, equal to, or greater than the second. 161 int compare_two_spill_nodes(Node* first, Node* second); 162 163 // Add a node to the list of available nodes 164 void AddNodeToAvailableList(Node *n); 165 166 // Compute the local use count for the nodes in a block, and compute 167 // the list of instructions with no uses in the block as available 168 void ComputeUseCount(const Block *bb); 169 170 // Choose an instruction from the available list to add to the bundle 171 Node * ChooseNodeToBundle(); 172 173 // See if this Node fits into the currently accumulating bundle 174 bool NodeFitsInBundle(Node *n); 175 176 // Decrement the use count for a node 177 void DecrementUseCounts(Node *n, const Block *bb); 178 179 // Garbage collect pinch nodes for reuse by other blocks. 180 void garbage_collect_pinch_nodes(); 181 // Clean up a pinch node for reuse (helper for above). 182 void cleanup_pinch( Node *pinch ); 183 184 // Information for statistics gathering 185 #ifndef PRODUCT 186 private: 187 // Gather information on size of nops relative to total 188 static uint _total_nop_size, _total_method_size; 189 static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 190 191 public: 192 static void print_statistics(); 193 194 static void increment_instructions_per_bundle(uint i) { 195 _total_instructions_per_bundle[i]++; 196 } 197 198 static void increment_nop_size(uint s) { 199 _total_nop_size += s; 200 } 201 202 static void increment_method_size(uint s) { 203 _total_method_size += s; 204 } 205 #endif 206 207 }; 208 209 PhaseOutput::PhaseOutput() 210 : Phase(Phase::Output), 211 _code_buffer("Compile::Fill_buffer"), 212 _first_block_size(0), 213 _handler_table(), 214 _inc_table(), 215 _stub_list(), 216 _oop_map_set(nullptr), 217 _scratch_buffer_blob(nullptr), 218 _scratch_locs_memory(nullptr), 219 _scratch_const_size(-1), 220 _in_scratch_emit_size(false), 221 _frame_slots(0), 222 _code_offsets(), 223 _node_bundling_limit(0), 224 _node_bundling_base(nullptr), 225 _orig_pc_slot(0), 226 _orig_pc_slot_offset_in_bytes(0), 227 _buf_sizes(), 228 _block(nullptr), 229 _index(0) { 230 C->set_output(this); 231 if (C->stub_name() == nullptr) { 232 _orig_pc_slot = C->fixed_slots() - (sizeof(address) / VMRegImpl::stack_slot_size); 233 } 234 } 235 236 PhaseOutput::~PhaseOutput() { 237 C->set_output(nullptr); 238 if (_scratch_buffer_blob != nullptr) { 239 BufferBlob::free(_scratch_buffer_blob); 240 } 241 } 242 243 void PhaseOutput::perform_mach_node_analysis() { 244 // Late barrier analysis must be done after schedule and bundle 245 // Otherwise liveness based spilling will fail 246 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 247 bs->late_barrier_analysis(); 248 249 pd_perform_mach_node_analysis(); 250 251 C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3); 252 } 253 254 // Convert Nodes to instruction bits and pass off to the VM 255 void PhaseOutput::Output() { 256 // RootNode goes 257 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" ); 258 259 // The number of new nodes (mostly MachNop) is proportional to 260 // the number of java calls and inner loops which are aligned. 261 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 262 C->inner_loops()*(OptoLoopAlignment-1)), 263 "out of nodes before code generation" ) ) { 264 return; 265 } 266 // Make sure I can find the Start Node 267 Block *entry = C->cfg()->get_block(1); 268 Block *broot = C->cfg()->get_root_block(); 269 270 const StartNode *start = entry->head()->as_Start(); 271 272 // Replace StartNode with prolog 273 MachPrologNode *prolog = new MachPrologNode(); 274 entry->map_node(prolog, 0); 275 C->cfg()->map_node_to_block(prolog, entry); 276 C->cfg()->unmap_node_from_block(start); // start is no longer in any block 277 278 // Virtual methods need an unverified entry point 279 280 if( C->is_osr_compilation() ) { 281 if( PoisonOSREntry ) { 282 // TODO: Should use a ShouldNotReachHereNode... 283 C->cfg()->insert( broot, 0, new MachBreakpointNode() ); 284 } 285 } else { 286 if( C->method() && !C->method()->flags().is_static() ) { 287 // Insert unvalidated entry point 288 C->cfg()->insert( broot, 0, new MachUEPNode() ); 289 } 290 291 } 292 293 // Break before main entry point 294 if ((C->method() && C->directive()->BreakAtExecuteOption) || 295 (OptoBreakpoint && C->is_method_compilation()) || 296 (OptoBreakpointOSR && C->is_osr_compilation()) || 297 (OptoBreakpointC2R && !C->method()) ) { 298 // checking for C->method() means that OptoBreakpoint does not apply to 299 // runtime stubs or frame converters 300 C->cfg()->insert( entry, 1, new MachBreakpointNode() ); 301 } 302 303 // Insert epilogs before every return 304 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) { 305 Block* block = C->cfg()->get_block(i); 306 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point? 307 Node* m = block->end(); 308 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) { 309 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 310 block->add_inst(epilog); 311 C->cfg()->map_node_to_block(epilog, block); 312 } 313 } 314 } 315 316 // Keeper of sizing aspects 317 _buf_sizes = BufferSizingData(); 318 319 // Initialize code buffer 320 estimate_buffer_size(_buf_sizes._const); 321 if (C->failing()) return; 322 323 // Pre-compute the length of blocks and replace 324 // long branches with short if machine supports it. 325 // Must be done before ScheduleAndBundle due to SPARC delay slots 326 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1); 327 blk_starts[0] = 0; 328 shorten_branches(blk_starts); 329 330 ScheduleAndBundle(); 331 if (C->failing()) { 332 return; 333 } 334 335 perform_mach_node_analysis(); 336 337 // Complete sizing of codebuffer 338 CodeBuffer* cb = init_buffer(); 339 if (cb == nullptr || C->failing()) { 340 return; 341 } 342 343 BuildOopMaps(); 344 345 if (C->failing()) { 346 return; 347 } 348 349 C2_MacroAssembler masm(cb); 350 fill_buffer(&masm, blk_starts); 351 } 352 353 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const { 354 // Determine if we need to generate a stack overflow check. 355 // Do it if the method is not a stub function and 356 // has java calls or has frame size > vm_page_size/8. 357 // The debug VM checks that deoptimization doesn't trigger an 358 // unexpected stack overflow (compiled method stack banging should 359 // guarantee it doesn't happen) so we always need the stack bang in 360 // a debug VM. 361 return (C->stub_function() == nullptr && 362 (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3 363 DEBUG_ONLY(|| true))); 364 } 365 366 bool PhaseOutput::need_register_stack_bang() const { 367 // Determine if we need to generate a register stack overflow check. 368 // This is only used on architectures which have split register 369 // and memory stacks. 370 // Bang if the method is not a stub function and has java calls 371 return (C->stub_function() == nullptr && C->has_java_calls()); 372 } 373 374 375 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 376 // of a loop. When aligning a loop we need to provide enough instructions 377 // in cpu's fetch buffer to feed decoders. The loop alignment could be 378 // avoided if we have enough instructions in fetch buffer at the head of a loop. 379 // By default, the size is set to 999999 by Block's constructor so that 380 // a loop will be aligned if the size is not reset here. 381 // 382 // Note: Mach instructions could contain several HW instructions 383 // so the size is estimated only. 384 // 385 void PhaseOutput::compute_loop_first_inst_sizes() { 386 // The next condition is used to gate the loop alignment optimization. 387 // Don't aligned a loop if there are enough instructions at the head of a loop 388 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 389 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 390 // equal to 11 bytes which is the largest address NOP instruction. 391 if (MaxLoopPad < OptoLoopAlignment - 1) { 392 uint last_block = C->cfg()->number_of_blocks() - 1; 393 for (uint i = 1; i <= last_block; i++) { 394 Block* block = C->cfg()->get_block(i); 395 // Check the first loop's block which requires an alignment. 396 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) { 397 uint sum_size = 0; 398 uint inst_cnt = NumberOfLoopInstrToAlign; 399 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc()); 400 401 // Check subsequent fallthrough blocks if the loop's first 402 // block(s) does not have enough instructions. 403 Block *nb = block; 404 while(inst_cnt > 0 && 405 i < last_block && 406 !C->cfg()->get_block(i + 1)->has_loop_alignment() && 407 !nb->has_successor(block)) { 408 i++; 409 nb = C->cfg()->get_block(i); 410 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc()); 411 } // while( inst_cnt > 0 && i < last_block ) 412 413 block->set_first_inst_size(sum_size); 414 } // f( b->head()->is_Loop() ) 415 } // for( i <= last_block ) 416 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 417 } 418 419 // The architecture description provides short branch variants for some long 420 // branch instructions. Replace eligible long branches with short branches. 421 void PhaseOutput::shorten_branches(uint* blk_starts) { 422 423 Compile::TracePhase tp(_t_shortenBranches); 424 425 // Compute size of each block, method size, and relocation information size 426 uint nblocks = C->cfg()->number_of_blocks(); 427 428 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 429 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 430 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks); 431 432 // Collect worst case block paddings 433 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks); 434 memset(block_worst_case_pad, 0, nblocks * sizeof(int)); 435 436 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 437 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 438 439 bool has_short_branch_candidate = false; 440 441 // Initialize the sizes to 0 442 int code_size = 0; // Size in bytes of generated code 443 int stub_size = 0; // Size in bytes of all stub entries 444 // Size in bytes of all relocation entries, including those in local stubs. 445 // Start with 2-bytes of reloc info for the unvalidated entry point 446 int reloc_size = 1; // Number of relocation entries 447 448 // Make three passes. The first computes pessimistic blk_starts, 449 // relative jmp_offset and reloc_size information. The second performs 450 // short branch substitution using the pessimistic sizing. The 451 // third inserts nops where needed. 452 453 // Step one, perform a pessimistic sizing pass. 454 uint last_call_adr = max_juint; 455 uint last_avoid_back_to_back_adr = max_juint; 456 uint nop_size = (new MachNopNode())->size(C->regalloc()); 457 for (uint i = 0; i < nblocks; i++) { // For all blocks 458 Block* block = C->cfg()->get_block(i); 459 _block = block; 460 461 // During short branch replacement, we store the relative (to blk_starts) 462 // offset of jump in jmp_offset, rather than the absolute offset of jump. 463 // This is so that we do not need to recompute sizes of all nodes when 464 // we compute correct blk_starts in our next sizing pass. 465 jmp_offset[i] = 0; 466 jmp_size[i] = 0; 467 jmp_nidx[i] = -1; 468 DEBUG_ONLY( jmp_target[i] = 0; ) 469 DEBUG_ONLY( jmp_rule[i] = 0; ) 470 471 // Sum all instruction sizes to compute block size 472 uint last_inst = block->number_of_nodes(); 473 uint blk_size = 0; 474 for (uint j = 0; j < last_inst; j++) { 475 _index = j; 476 Node* nj = block->get_node(_index); 477 // Handle machine instruction nodes 478 if (nj->is_Mach()) { 479 MachNode* mach = nj->as_Mach(); 480 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 481 reloc_size += mach->reloc(); 482 if (mach->is_MachCall()) { 483 // add size information for trampoline stub 484 // class CallStubImpl is platform-specific and defined in the *.ad files. 485 stub_size += CallStubImpl::size_call_trampoline(); 486 reloc_size += CallStubImpl::reloc_call_trampoline(); 487 488 MachCallNode *mcall = mach->as_MachCall(); 489 // This destination address is NOT PC-relative 490 491 mcall->method_set((intptr_t)mcall->entry_point()); 492 493 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) { 494 stub_size += CompiledDirectCall::to_interp_stub_size(); 495 reloc_size += CompiledDirectCall::reloc_to_interp_stub(); 496 } 497 } else if (mach->is_MachSafePoint()) { 498 // If call/safepoint are adjacent, account for possible 499 // nop to disambiguate the two safepoints. 500 // ScheduleAndBundle() can rearrange nodes in a block, 501 // check for all offsets inside this block. 502 if (last_call_adr >= blk_starts[i]) { 503 blk_size += nop_size; 504 } 505 } 506 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 507 // Nop is inserted between "avoid back to back" instructions. 508 // ScheduleAndBundle() can rearrange nodes in a block, 509 // check for all offsets inside this block. 510 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 511 blk_size += nop_size; 512 } 513 } 514 if (mach->may_be_short_branch()) { 515 if (!nj->is_MachBranch()) { 516 #ifndef PRODUCT 517 nj->dump(3); 518 #endif 519 Unimplemented(); 520 } 521 assert(jmp_nidx[i] == -1, "block should have only one branch"); 522 jmp_offset[i] = blk_size; 523 jmp_size[i] = nj->size(C->regalloc()); 524 jmp_nidx[i] = j; 525 has_short_branch_candidate = true; 526 } 527 } 528 blk_size += nj->size(C->regalloc()); 529 // Remember end of call offset 530 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 531 last_call_adr = blk_starts[i]+blk_size; 532 } 533 // Remember end of avoid_back_to_back offset 534 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 535 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 536 } 537 } 538 539 // When the next block starts a loop, we may insert pad NOP 540 // instructions. Since we cannot know our future alignment, 541 // assume the worst. 542 if (i < nblocks - 1) { 543 Block* nb = C->cfg()->get_block(i + 1); 544 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 545 if (max_loop_pad > 0) { 546 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 547 // Adjust last_call_adr and/or last_avoid_back_to_back_adr. 548 // If either is the last instruction in this block, bump by 549 // max_loop_pad in lock-step with blk_size, so sizing 550 // calculations in subsequent blocks still can conservatively 551 // detect that it may the last instruction in this block. 552 if (last_call_adr == blk_starts[i]+blk_size) { 553 last_call_adr += max_loop_pad; 554 } 555 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) { 556 last_avoid_back_to_back_adr += max_loop_pad; 557 } 558 blk_size += max_loop_pad; 559 block_worst_case_pad[i + 1] = max_loop_pad; 560 } 561 } 562 563 // Save block size; update total method size 564 blk_starts[i+1] = blk_starts[i]+blk_size; 565 } 566 567 // Step two, replace eligible long jumps. 568 bool progress = true; 569 uint last_may_be_short_branch_adr = max_juint; 570 while (has_short_branch_candidate && progress) { 571 progress = false; 572 has_short_branch_candidate = false; 573 int adjust_block_start = 0; 574 for (uint i = 0; i < nblocks; i++) { 575 Block* block = C->cfg()->get_block(i); 576 int idx = jmp_nidx[i]; 577 MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach(); 578 if (mach != nullptr && mach->may_be_short_branch()) { 579 #ifdef ASSERT 580 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 581 int j; 582 // Find the branch; ignore trailing NOPs. 583 for (j = block->number_of_nodes()-1; j>=0; j--) { 584 Node* n = block->get_node(j); 585 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 586 break; 587 } 588 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity"); 589 #endif 590 int br_size = jmp_size[i]; 591 int br_offs = blk_starts[i] + jmp_offset[i]; 592 593 // This requires the TRUE branch target be in succs[0] 594 uint bnum = block->non_connector_successor(0)->_pre_order; 595 int offset = blk_starts[bnum] - br_offs; 596 if (bnum > i) { // adjust following block's offset 597 offset -= adjust_block_start; 598 } 599 600 // This block can be a loop header, account for the padding 601 // in the previous block. 602 int block_padding = block_worst_case_pad[i]; 603 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top"); 604 // In the following code a nop could be inserted before 605 // the branch which will increase the backward distance. 606 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr); 607 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block"); 608 609 if (needs_padding && offset <= 0) 610 offset -= nop_size; 611 612 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) { 613 // We've got a winner. Replace this branch. 614 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 615 616 // Update the jmp_size. 617 int new_size = replacement->size(C->regalloc()); 618 int diff = br_size - new_size; 619 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 620 // Conservatively take into account padding between 621 // avoid_back_to_back branches. Previous branch could be 622 // converted into avoid_back_to_back branch during next 623 // rounds. 624 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 625 jmp_offset[i] += nop_size; 626 diff -= nop_size; 627 } 628 adjust_block_start += diff; 629 block->map_node(replacement, idx); 630 mach->subsume_by(replacement, C); 631 mach = replacement; 632 progress = true; 633 634 jmp_size[i] = new_size; 635 DEBUG_ONLY( jmp_target[i] = bnum; ); 636 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 637 } else { 638 // The jump distance is not short, try again during next iteration. 639 has_short_branch_candidate = true; 640 } 641 } // (mach->may_be_short_branch()) 642 if (mach != nullptr && (mach->may_be_short_branch() || 643 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) { 644 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 645 } 646 blk_starts[i+1] -= adjust_block_start; 647 } 648 } 649 650 #ifdef ASSERT 651 for (uint i = 0; i < nblocks; i++) { // For all blocks 652 if (jmp_target[i] != 0) { 653 int br_size = jmp_size[i]; 654 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 655 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 656 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]); 657 } 658 assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 659 } 660 } 661 #endif 662 663 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 664 // after ScheduleAndBundle(). 665 666 // ------------------ 667 // Compute size for code buffer 668 code_size = blk_starts[nblocks]; 669 670 // Relocation records 671 reloc_size += 1; // Relo entry for exception handler 672 673 // Adjust reloc_size to number of record of relocation info 674 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 675 // a relocation index. 676 // The CodeBuffer will expand the locs array if this estimate is too low. 677 reloc_size *= 10 / sizeof(relocInfo); 678 679 _buf_sizes._reloc = reloc_size; 680 _buf_sizes._code = code_size; 681 _buf_sizes._stub = stub_size; 682 } 683 684 //------------------------------FillLocArray----------------------------------- 685 // Create a bit of debug info and append it to the array. The mapping is from 686 // Java local or expression stack to constant, register or stack-slot. For 687 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 688 // entry has been taken care of and caller should skip it). 689 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 690 // This should never have accepted Bad before 691 assert(OptoReg::is_valid(regnum), "location must be valid"); 692 return (OptoReg::is_reg(regnum)) 693 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 694 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 695 } 696 697 698 ObjectValue* 699 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 700 for (int i = 0; i < objs->length(); i++) { 701 assert(objs->at(i)->is_object(), "corrupt object cache"); 702 ObjectValue* sv = objs->at(i)->as_ObjectValue(); 703 if (sv->id() == id) { 704 return sv; 705 } 706 } 707 // Otherwise.. 708 return nullptr; 709 } 710 711 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 712 ObjectValue* sv ) { 713 assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition"); 714 objs->append(sv); 715 } 716 717 718 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 719 GrowableArray<ScopeValue*> *array, 720 GrowableArray<ScopeValue*> *objs ) { 721 assert( local, "use _top instead of null" ); 722 if (array->length() != idx) { 723 assert(array->length() == idx + 1, "Unexpected array count"); 724 // Old functionality: 725 // return 726 // New functionality: 727 // Assert if the local is not top. In product mode let the new node 728 // override the old entry. 729 assert(local == C->top(), "LocArray collision"); 730 if (local == C->top()) { 731 return; 732 } 733 array->pop(); 734 } 735 const Type *t = local->bottom_type(); 736 737 // Is it a safepoint scalar object node? 738 if (local->is_SafePointScalarObject()) { 739 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 740 741 ObjectValue* sv = sv_for_node_id(objs, spobj->_idx); 742 if (sv == nullptr) { 743 ciKlass* cik = t->is_oopptr()->exact_klass(); 744 assert(cik->is_instance_klass() || 745 cik->is_array_klass(), "Not supported allocation."); 746 sv = new ObjectValue(spobj->_idx, 747 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 748 set_sv_for_object_node(objs, sv); 749 750 uint first_ind = spobj->first_index(sfpt->jvms()); 751 for (uint i = 0; i < spobj->n_fields(); i++) { 752 Node* fld_node = sfpt->in(first_ind+i); 753 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 754 } 755 } 756 array->append(sv); 757 return; 758 } else if (local->is_SafePointScalarMerge()) { 759 SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge(); 760 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx); 761 762 if (mv == nullptr) { 763 GrowableArray<ScopeValue*> deps; 764 765 int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms()); 766 (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs); 767 assert(deps.length() == 1, "missing value"); 768 769 int selector_idx = smerge->selector_idx(sfpt->jvms()); 770 (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr); 771 assert(deps.length() == 2, "missing value"); 772 773 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1)); 774 set_sv_for_object_node(objs, mv); 775 776 for (uint i = 1; i < smerge->req(); i++) { 777 Node* obj_node = smerge->in(i); 778 int idx = mv->possible_objects()->length(); 779 (void)FillLocArray(idx, sfpt, obj_node, mv->possible_objects(), objs); 780 781 // By default ObjectValues that are in 'possible_objects' are not root objects. 782 // They will be marked as root later if they are directly referenced in a JVMS. 783 assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!"); 784 assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue."); 785 mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false); 786 } 787 } 788 array->append(mv); 789 return; 790 } 791 792 // Grab the register number for the local 793 OptoReg::Name regnum = C->regalloc()->get_reg_first(local); 794 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 795 // Record the double as two float registers. 796 // The register mask for such a value always specifies two adjacent 797 // float registers, with the lower register number even. 798 // Normally, the allocation of high and low words to these registers 799 // is irrelevant, because nearly all operations on register pairs 800 // (e.g., StoreD) treat them as a single unit. 801 // Here, we assume in addition that the words in these two registers 802 // stored "naturally" (by operations like StoreD and double stores 803 // within the interpreter) such that the lower-numbered register 804 // is written to the lower memory address. This may seem like 805 // a machine dependency, but it is not--it is a requirement on 806 // the author of the <arch>.ad file to ensure that, for every 807 // even/odd double-register pair to which a double may be allocated, 808 // the word in the even single-register is stored to the first 809 // memory word. (Note that register numbers are completely 810 // arbitrary, and are not tied to any machine-level encodings.) 811 #ifdef _LP64 812 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 813 array->append(new ConstantIntValue((jint)0)); 814 array->append(new_loc_value( C->regalloc(), regnum, Location::dbl )); 815 } else if ( t->base() == Type::Long ) { 816 array->append(new ConstantIntValue((jint)0)); 817 array->append(new_loc_value( C->regalloc(), regnum, Location::lng )); 818 } else if ( t->base() == Type::RawPtr ) { 819 // jsr/ret return address which must be restored into the full 820 // width 64-bit stack slot. 821 array->append(new_loc_value( C->regalloc(), regnum, Location::lng )); 822 } 823 #else //_LP64 824 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 825 // Repack the double/long as two jints. 826 // The convention the interpreter uses is that the second local 827 // holds the first raw word of the native double representation. 828 // This is actually reasonable, since locals and stack arrays 829 // grow downwards in all implementations. 830 // (If, on some machine, the interpreter's Java locals or stack 831 // were to grow upwards, the embedded doubles would be word-swapped.) 832 array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal )); 833 array->append(new_loc_value( C->regalloc(), regnum , Location::normal )); 834 } 835 #endif //_LP64 836 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 837 OptoReg::is_reg(regnum) ) { 838 array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double() 839 ? Location::float_in_dbl : Location::normal )); 840 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 841 array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long 842 ? Location::int_in_long : Location::normal )); 843 } else if( t->base() == Type::NarrowOop ) { 844 array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop )); 845 } else if (t->base() == Type::VectorA || t->base() == Type::VectorS || 846 t->base() == Type::VectorD || t->base() == Type::VectorX || 847 t->base() == Type::VectorY || t->base() == Type::VectorZ) { 848 array->append(new_loc_value( C->regalloc(), regnum, Location::vector )); 849 } else if (C->regalloc()->is_oop(local)) { 850 assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr || 851 t->base() == Type::AryPtr, 852 "Unexpected type: %s", t->msg()); 853 array->append(new_loc_value( C->regalloc(), regnum, Location::oop )); 854 } else { 855 assert(t->base() == Type::Int || t->base() == Type::Half || 856 t->base() == Type::FloatCon || t->base() == Type::FloatBot, 857 "Unexpected type: %s", t->msg()); 858 array->append(new_loc_value( C->regalloc(), regnum, Location::normal )); 859 } 860 return; 861 } 862 863 // No register. It must be constant data. 864 switch (t->base()) { 865 case Type::Half: // Second half of a double 866 ShouldNotReachHere(); // Caller should skip 2nd halves 867 break; 868 case Type::AnyPtr: 869 array->append(new ConstantOopWriteValue(nullptr)); 870 break; 871 case Type::AryPtr: 872 case Type::InstPtr: // fall through 873 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 874 break; 875 case Type::NarrowOop: 876 if (t == TypeNarrowOop::NULL_PTR) { 877 array->append(new ConstantOopWriteValue(nullptr)); 878 } else { 879 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 880 } 881 break; 882 case Type::Int: 883 array->append(new ConstantIntValue(t->is_int()->get_con())); 884 break; 885 case Type::RawPtr: 886 // A return address (T_ADDRESS). 887 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 888 #ifdef _LP64 889 // Must be restored to the full-width 64-bit stack slot. 890 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 891 #else 892 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 893 #endif 894 break; 895 case Type::FloatCon: { 896 float f = t->is_float_constant()->getf(); 897 array->append(new ConstantIntValue(jint_cast(f))); 898 break; 899 } 900 case Type::DoubleCon: { 901 jdouble d = t->is_double_constant()->getd(); 902 #ifdef _LP64 903 array->append(new ConstantIntValue((jint)0)); 904 array->append(new ConstantDoubleValue(d)); 905 #else 906 // Repack the double as two jints. 907 // The convention the interpreter uses is that the second local 908 // holds the first raw word of the native double representation. 909 // This is actually reasonable, since locals and stack arrays 910 // grow downwards in all implementations. 911 // (If, on some machine, the interpreter's Java locals or stack 912 // were to grow upwards, the embedded doubles would be word-swapped.) 913 jlong_accessor acc; 914 acc.long_value = jlong_cast(d); 915 array->append(new ConstantIntValue(acc.words[1])); 916 array->append(new ConstantIntValue(acc.words[0])); 917 #endif 918 break; 919 } 920 case Type::Long: { 921 jlong d = t->is_long()->get_con(); 922 #ifdef _LP64 923 array->append(new ConstantIntValue((jint)0)); 924 array->append(new ConstantLongValue(d)); 925 #else 926 // Repack the long as two jints. 927 // The convention the interpreter uses is that the second local 928 // holds the first raw word of the native double representation. 929 // This is actually reasonable, since locals and stack arrays 930 // grow downwards in all implementations. 931 // (If, on some machine, the interpreter's Java locals or stack 932 // were to grow upwards, the embedded doubles would be word-swapped.) 933 jlong_accessor acc; 934 acc.long_value = d; 935 array->append(new ConstantIntValue(acc.words[1])); 936 array->append(new ConstantIntValue(acc.words[0])); 937 #endif 938 break; 939 } 940 case Type::Top: // Add an illegal value here 941 array->append(new LocationValue(Location())); 942 break; 943 default: 944 ShouldNotReachHere(); 945 break; 946 } 947 } 948 949 // Determine if this node starts a bundle 950 bool PhaseOutput::starts_bundle(const Node *n) const { 951 return (_node_bundling_limit > n->_idx && 952 _node_bundling_base[n->_idx].starts_bundle()); 953 } 954 955 // Determine if there is a monitor that has 'ov' as its owner. 956 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const { 957 for (int k = 0; k < monarray->length(); k++) { 958 MonitorValue* mv = monarray->at(k); 959 if (mv->owner() == ov) { 960 return true; 961 } 962 } 963 964 return false; 965 } 966 967 // Determine if there is a scalar replaced object description represented by 'ov'. 968 bool PhaseOutput::contains_as_scalarized_obj(JVMState* jvms, MachSafePointNode* sfn, 969 GrowableArray<ScopeValue*>* objs, 970 ObjectValue* ov) const { 971 for (int i = 0; i < jvms->scl_size(); i++) { 972 Node* n = sfn->scalarized_obj(jvms, i); 973 // Other kinds of nodes that we may encounter here, for instance constants 974 // representing values of fields of objects scalarized, aren't relevant for 975 // us, since they don't map to ObjectValue. 976 if (!n->is_SafePointScalarObject()) { 977 continue; 978 } 979 980 ObjectValue* other = sv_for_node_id(objs, n->_idx); 981 if (ov == other) { 982 return true; 983 } 984 } 985 return false; 986 } 987 988 //--------------------------Process_OopMap_Node-------------------------------- 989 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) { 990 // Handle special safepoint nodes for synchronization 991 MachSafePointNode *sfn = mach->as_MachSafePoint(); 992 MachCallNode *mcall; 993 994 int safepoint_pc_offset = current_offset; 995 bool is_method_handle_invoke = false; 996 bool return_oop = false; 997 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope; 998 bool arg_escape = false; 999 1000 // Add the safepoint in the DebugInfoRecorder 1001 if( !mach->is_MachCall() ) { 1002 mcall = nullptr; 1003 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 1004 } else { 1005 mcall = mach->as_MachCall(); 1006 1007 // Is the call a MethodHandle call? 1008 if (mcall->is_MachCallJava()) { 1009 if (mcall->as_MachCallJava()->_method_handle_invoke) { 1010 assert(C->has_method_handle_invokes(), "must have been set during call generation"); 1011 is_method_handle_invoke = true; 1012 } 1013 arg_escape = mcall->as_MachCallJava()->_arg_escape; 1014 } 1015 1016 // Check if a call returns an object. 1017 if (mcall->returns_pointer()) { 1018 return_oop = true; 1019 } 1020 safepoint_pc_offset += mcall->ret_addr_offset(); 1021 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 1022 } 1023 1024 // Loop over the JVMState list to add scope information 1025 // Do not skip safepoints with a null method, they need monitor info 1026 JVMState* youngest_jvms = sfn->jvms(); 1027 int max_depth = youngest_jvms->depth(); 1028 1029 // Allocate the object pool for scalar-replaced objects -- the map from 1030 // small-integer keys (which can be recorded in the local and ostack 1031 // arrays) to descriptions of the object state. 1032 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 1033 1034 // Visit scopes from oldest to youngest. 1035 for (int depth = 1; depth <= max_depth; depth++) { 1036 JVMState* jvms = youngest_jvms->of_depth(depth); 1037 int idx; 1038 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr; 1039 // Safepoints that do not have method() set only provide oop-map and monitor info 1040 // to support GC; these do not support deoptimization. 1041 int num_locs = (method == nullptr) ? 0 : jvms->loc_size(); 1042 int num_exps = (method == nullptr) ? 0 : jvms->stk_size(); 1043 int num_mon = jvms->nof_monitors(); 1044 assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(), 1045 "JVMS local count must match that of the method"); 1046 1047 // Add Local and Expression Stack Information 1048 1049 // Insert locals into the locarray 1050 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 1051 for( idx = 0; idx < num_locs; idx++ ) { 1052 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 1053 } 1054 1055 // Insert expression stack entries into the exparray 1056 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 1057 for( idx = 0; idx < num_exps; idx++ ) { 1058 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 1059 } 1060 1061 // Add in mappings of the monitors 1062 assert( !method || 1063 !method->is_synchronized() || 1064 method->is_native() || 1065 num_mon > 0, 1066 "monitors must always exist for synchronized methods"); 1067 1068 // Build the growable array of ScopeValues for exp stack 1069 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 1070 1071 // Loop over monitors and insert into array 1072 for (idx = 0; idx < num_mon; idx++) { 1073 // Grab the node that defines this monitor 1074 Node* box_node = sfn->monitor_box(jvms, idx); 1075 Node* obj_node = sfn->monitor_obj(jvms, idx); 1076 1077 // Create ScopeValue for object 1078 ScopeValue *scval = nullptr; 1079 1080 if (obj_node->is_SafePointScalarObject()) { 1081 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 1082 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx); 1083 if (scval == nullptr) { 1084 const Type *t = spobj->bottom_type(); 1085 ciKlass* cik = t->is_oopptr()->exact_klass(); 1086 assert(cik->is_instance_klass() || 1087 cik->is_array_klass(), "Not supported allocation."); 1088 ObjectValue* sv = new ObjectValue(spobj->_idx, 1089 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 1090 PhaseOutput::set_sv_for_object_node(objs, sv); 1091 1092 uint first_ind = spobj->first_index(youngest_jvms); 1093 for (uint i = 0; i < spobj->n_fields(); i++) { 1094 Node* fld_node = sfn->in(first_ind+i); 1095 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 1096 } 1097 scval = sv; 1098 } 1099 } else if (obj_node->is_SafePointScalarMerge()) { 1100 SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge(); 1101 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx); 1102 1103 if (mv == nullptr) { 1104 GrowableArray<ScopeValue*> deps; 1105 1106 int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms); 1107 FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs); 1108 assert(deps.length() == 1, "missing value"); 1109 1110 int selector_idx = smerge->selector_idx(youngest_jvms); 1111 FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr); 1112 assert(deps.length() == 2, "missing value"); 1113 1114 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1)); 1115 set_sv_for_object_node(objs, mv); 1116 1117 for (uint i = 1; i < smerge->req(); i++) { 1118 Node* obj_node = smerge->in(i); 1119 int idx = mv->possible_objects()->length(); 1120 (void)FillLocArray(idx, sfn, obj_node, mv->possible_objects(), objs); 1121 1122 // By default ObjectValues that are in 'possible_objects' are not root objects. 1123 // They will be marked as root later if they are directly referenced in a JVMS. 1124 assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!"); 1125 assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue."); 1126 mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false); 1127 } 1128 } 1129 scval = mv; 1130 } else if (!obj_node->is_Con()) { 1131 OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node); 1132 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 1133 scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop ); 1134 } else { 1135 scval = new_loc_value( C->regalloc(), obj_reg, Location::oop ); 1136 } 1137 } else { 1138 const TypePtr *tp = obj_node->get_ptr_type(); 1139 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 1140 } 1141 1142 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 1143 Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg)); 1144 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 1145 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 1146 } 1147 1148 // Mark ObjectValue nodes as root nodes if they are directly 1149 // referenced in the JVMS. 1150 for (int i = 0; i < objs->length(); i++) { 1151 ScopeValue* sv = objs->at(i); 1152 if (sv->is_object_merge()) { 1153 ObjectMergeValue* merge = sv->as_ObjectMergeValue(); 1154 1155 for (int j = 0; j< merge->possible_objects()->length(); j++) { 1156 ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue(); 1157 if (ov->is_root()) { 1158 // Already flagged as 'root' by something else. We shouldn't change it 1159 // to non-root in a younger JVMS because it may need to be alive in 1160 // a younger JVMS. 1161 } else { 1162 bool is_root = locarray->contains(ov) || 1163 exparray->contains(ov) || 1164 contains_as_owner(monarray, ov) || 1165 contains_as_scalarized_obj(jvms, sfn, objs, ov); 1166 ov->set_root(is_root); 1167 } 1168 } 1169 } 1170 } 1171 1172 // We dump the object pool first, since deoptimization reads it in first. 1173 C->debug_info()->dump_object_pool(objs); 1174 1175 // Build first class objects to pass to scope 1176 DebugToken *locvals = C->debug_info()->create_scope_values(locarray); 1177 DebugToken *expvals = C->debug_info()->create_scope_values(exparray); 1178 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray); 1179 1180 // Make method available for all Safepoints 1181 ciMethod* scope_method = method ? method : C->method(); 1182 // Describe the scope here 1183 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 1184 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 1185 // Now we can describe the scope. 1186 methodHandle null_mh; 1187 bool rethrow_exception = false; 1188 C->debug_info()->describe_scope( 1189 safepoint_pc_offset, 1190 null_mh, 1191 scope_method, 1192 jvms->bci(), 1193 jvms->should_reexecute(), 1194 rethrow_exception, 1195 is_method_handle_invoke, 1196 return_oop, 1197 has_ea_local_in_scope, 1198 arg_escape, 1199 locvals, 1200 expvals, 1201 monvals 1202 ); 1203 } // End jvms loop 1204 1205 // Mark the end of the scope set. 1206 C->debug_info()->end_safepoint(safepoint_pc_offset); 1207 } 1208 1209 1210 1211 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 1212 class NonSafepointEmitter { 1213 Compile* C; 1214 JVMState* _pending_jvms; 1215 int _pending_offset; 1216 1217 void emit_non_safepoint(); 1218 1219 public: 1220 NonSafepointEmitter(Compile* compile) { 1221 this->C = compile; 1222 _pending_jvms = nullptr; 1223 _pending_offset = 0; 1224 } 1225 1226 void observe_instruction(Node* n, int pc_offset) { 1227 if (!C->debug_info()->recording_non_safepoints()) return; 1228 1229 Node_Notes* nn = C->node_notes_at(n->_idx); 1230 if (nn == nullptr || nn->jvms() == nullptr) return; 1231 if (_pending_jvms != nullptr && 1232 _pending_jvms->same_calls_as(nn->jvms())) { 1233 // Repeated JVMS? Stretch it up here. 1234 _pending_offset = pc_offset; 1235 } else { 1236 if (_pending_jvms != nullptr && 1237 _pending_offset < pc_offset) { 1238 emit_non_safepoint(); 1239 } 1240 _pending_jvms = nullptr; 1241 if (pc_offset > C->debug_info()->last_pc_offset()) { 1242 // This is the only way _pending_jvms can become non-null: 1243 _pending_jvms = nn->jvms(); 1244 _pending_offset = pc_offset; 1245 } 1246 } 1247 } 1248 1249 // Stay out of the way of real safepoints: 1250 void observe_safepoint(JVMState* jvms, int pc_offset) { 1251 if (_pending_jvms != nullptr && 1252 !_pending_jvms->same_calls_as(jvms) && 1253 _pending_offset < pc_offset) { 1254 emit_non_safepoint(); 1255 } 1256 _pending_jvms = nullptr; 1257 } 1258 1259 void flush_at_end() { 1260 if (_pending_jvms != nullptr) { 1261 emit_non_safepoint(); 1262 } 1263 _pending_jvms = nullptr; 1264 } 1265 }; 1266 1267 void NonSafepointEmitter::emit_non_safepoint() { 1268 JVMState* youngest_jvms = _pending_jvms; 1269 int pc_offset = _pending_offset; 1270 1271 // Clear it now: 1272 _pending_jvms = nullptr; 1273 1274 DebugInformationRecorder* debug_info = C->debug_info(); 1275 assert(debug_info->recording_non_safepoints(), "sanity"); 1276 1277 debug_info->add_non_safepoint(pc_offset); 1278 int max_depth = youngest_jvms->depth(); 1279 1280 // Visit scopes from oldest to youngest. 1281 for (int depth = 1; depth <= max_depth; depth++) { 1282 JVMState* jvms = youngest_jvms->of_depth(depth); 1283 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr; 1284 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 1285 methodHandle null_mh; 1286 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute()); 1287 } 1288 1289 // Mark the end of the scope set. 1290 debug_info->end_non_safepoint(pc_offset); 1291 } 1292 1293 //------------------------------init_buffer------------------------------------ 1294 void PhaseOutput::estimate_buffer_size(int& const_req) { 1295 1296 // Set the initially allocated size 1297 const_req = initial_const_capacity; 1298 1299 // The extra spacing after the code is necessary on some platforms. 1300 // Sometimes we need to patch in a jump after the last instruction, 1301 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 1302 1303 // Compute the byte offset where we can store the deopt pc. 1304 if (C->fixed_slots() != 0) { 1305 _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 1306 } 1307 1308 // Compute prolog code size 1309 _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize; 1310 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 1311 1312 if (C->has_mach_constant_base_node()) { 1313 uint add_size = 0; 1314 // Fill the constant table. 1315 // Note: This must happen before shorten_branches. 1316 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) { 1317 Block* b = C->cfg()->get_block(i); 1318 1319 for (uint j = 0; j < b->number_of_nodes(); j++) { 1320 Node* n = b->get_node(j); 1321 1322 // If the node is a MachConstantNode evaluate the constant 1323 // value section. 1324 if (n->is_MachConstant()) { 1325 MachConstantNode* machcon = n->as_MachConstant(); 1326 machcon->eval_constant(C); 1327 } else if (n->is_Mach()) { 1328 // On Power there are more nodes that issue constants. 1329 add_size += (n->as_Mach()->ins_num_consts() * 8); 1330 } 1331 } 1332 } 1333 1334 // Calculate the offsets of the constants and the size of the 1335 // constant table (including the padding to the next section). 1336 constant_table().calculate_offsets_and_size(); 1337 const_req = constant_table().alignment() + constant_table().size() + add_size; 1338 } 1339 1340 // Initialize the space for the BufferBlob used to find and verify 1341 // instruction size in MachNode::emit_size() 1342 init_scratch_buffer_blob(const_req); 1343 } 1344 1345 CodeBuffer* PhaseOutput::init_buffer() { 1346 int stub_req = _buf_sizes._stub; 1347 int code_req = _buf_sizes._code; 1348 int const_req = _buf_sizes._const; 1349 1350 int pad_req = NativeCall::byte_size(); 1351 1352 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1353 stub_req += bs->estimate_stub_size(); 1354 1355 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1356 // class HandlerImpl is platform-specific and defined in the *.ad files. 1357 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler 1358 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler 1359 stub_req += MAX_stubs_size; // ensure per-stub margin 1360 code_req += max_inst_size(); // ensure per-instruction margin 1361 1362 if (StressCodeBuffers) 1363 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1364 1365 int total_req = 1366 const_req + 1367 code_req + 1368 pad_req + 1369 stub_req + 1370 exception_handler_req + 1371 deopt_handler_req; // deopt handler 1372 1373 if (C->has_method_handle_invokes()) 1374 total_req += deopt_handler_req; // deopt MH handler 1375 1376 CodeBuffer* cb = code_buffer(); 1377 cb->set_const_section_alignment(constant_table().alignment()); 1378 cb->initialize(total_req, _buf_sizes._reloc); 1379 1380 // Have we run out of code space? 1381 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) { 1382 C->record_failure("CodeCache is full"); 1383 return nullptr; 1384 } 1385 // Configure the code buffer. 1386 cb->initialize_consts_size(const_req); 1387 cb->initialize_stubs_size(stub_req); 1388 cb->initialize_oop_recorder(C->env()->oop_recorder()); 1389 1390 return cb; 1391 } 1392 1393 //------------------------------fill_buffer------------------------------------ 1394 void PhaseOutput::fill_buffer(C2_MacroAssembler* masm, uint* blk_starts) { 1395 // blk_starts[] contains offsets calculated during short branches processing, 1396 // offsets should not be increased during following steps. 1397 1398 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1399 // of a loop. It is used to determine the padding for loop alignment. 1400 Compile::TracePhase tp(_t_fillBuffer); 1401 1402 compute_loop_first_inst_sizes(); 1403 1404 // Create oopmap set. 1405 _oop_map_set = new OopMapSet(); 1406 1407 // !!!!! This preserves old handling of oopmaps for now 1408 C->debug_info()->set_oopmaps(_oop_map_set); 1409 1410 uint nblocks = C->cfg()->number_of_blocks(); 1411 // Count and start of implicit null check instructions 1412 uint inct_cnt = 0; 1413 uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1414 1415 // Count and start of calls 1416 uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1417 1418 uint return_offset = 0; 1419 int nop_size = (new MachNopNode())->size(C->regalloc()); 1420 1421 int previous_offset = 0; 1422 int current_offset = 0; 1423 int last_call_offset = -1; 1424 int last_avoid_back_to_back_offset = -1; 1425 #ifdef ASSERT 1426 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1427 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1428 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1429 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1430 #endif 1431 1432 // Create an array of unused labels, one for each basic block, if printing is enabled 1433 #if defined(SUPPORT_OPTO_ASSEMBLY) 1434 int* node_offsets = nullptr; 1435 uint node_offset_limit = C->unique(); 1436 1437 if (C->print_assembly()) { 1438 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1439 } 1440 if (node_offsets != nullptr) { 1441 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly. 1442 memset(node_offsets, 0, node_offset_limit*sizeof(int)); 1443 } 1444 #endif 1445 1446 NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily 1447 1448 // Emit the constant table. 1449 if (C->has_mach_constant_base_node()) { 1450 if (!constant_table().emit(masm)) { 1451 C->record_failure("consts section overflow"); 1452 return; 1453 } 1454 } 1455 1456 // Create an array of labels, one for each basic block 1457 Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1458 for (uint i = 0; i <= nblocks; i++) { 1459 blk_labels[i].init(); 1460 } 1461 1462 // Now fill in the code buffer 1463 for (uint i = 0; i < nblocks; i++) { 1464 Block* block = C->cfg()->get_block(i); 1465 _block = block; 1466 Node* head = block->head(); 1467 1468 // If this block needs to start aligned (i.e, can be reached other 1469 // than by falling-thru from the previous block), then force the 1470 // start of a new bundle. 1471 if (Pipeline::requires_bundling() && starts_bundle(head)) { 1472 masm->code()->flush_bundle(true); 1473 } 1474 1475 #ifdef ASSERT 1476 if (!block->is_connector()) { 1477 stringStream st; 1478 block->dump_head(C->cfg(), &st); 1479 masm->block_comment(st.freeze()); 1480 } 1481 jmp_target[i] = 0; 1482 jmp_offset[i] = 0; 1483 jmp_size[i] = 0; 1484 jmp_rule[i] = 0; 1485 #endif 1486 int blk_offset = current_offset; 1487 1488 // Define the label at the beginning of the basic block 1489 masm->bind(blk_labels[block->_pre_order]); 1490 1491 uint last_inst = block->number_of_nodes(); 1492 1493 // Emit block normally, except for last instruction. 1494 // Emit means "dump code bits into code buffer". 1495 for (uint j = 0; j<last_inst; j++) { 1496 _index = j; 1497 1498 // Get the node 1499 Node* n = block->get_node(j); 1500 1501 // If this starts a new instruction group, then flush the current one 1502 // (but allow split bundles) 1503 if (Pipeline::requires_bundling() && starts_bundle(n)) 1504 masm->code()->flush_bundle(false); 1505 1506 // Special handling for SafePoint/Call Nodes 1507 bool is_mcall = false; 1508 if (n->is_Mach()) { 1509 MachNode *mach = n->as_Mach(); 1510 is_mcall = n->is_MachCall(); 1511 bool is_sfn = n->is_MachSafePoint(); 1512 1513 // If this requires all previous instructions be flushed, then do so 1514 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1515 masm->code()->flush_bundle(true); 1516 current_offset = masm->offset(); 1517 } 1518 1519 // align the instruction if necessary 1520 int padding = mach->compute_padding(current_offset); 1521 // Make sure safepoint node for polling is distinct from a call's 1522 // return by adding a nop if needed. 1523 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1524 padding = nop_size; 1525 } 1526 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) && 1527 current_offset == last_avoid_back_to_back_offset) { 1528 // Avoid back to back some instructions. 1529 padding = nop_size; 1530 } 1531 1532 if (padding > 0) { 1533 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1534 int nops_cnt = padding / nop_size; 1535 MachNode *nop = new MachNopNode(nops_cnt); 1536 block->insert_node(nop, j++); 1537 last_inst++; 1538 C->cfg()->map_node_to_block(nop, block); 1539 // Ensure enough space. 1540 masm->code()->insts()->maybe_expand_to_ensure_remaining(max_inst_size()); 1541 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) { 1542 C->record_failure("CodeCache is full"); 1543 return; 1544 } 1545 nop->emit(masm, C->regalloc()); 1546 masm->code()->flush_bundle(true); 1547 current_offset = masm->offset(); 1548 } 1549 1550 bool observe_safepoint = is_sfn; 1551 // Remember the start of the last call in a basic block 1552 if (is_mcall) { 1553 MachCallNode *mcall = mach->as_MachCall(); 1554 1555 // This destination address is NOT PC-relative 1556 mcall->method_set((intptr_t)mcall->entry_point()); 1557 1558 // Save the return address 1559 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset(); 1560 1561 observe_safepoint = mcall->guaranteed_safepoint(); 1562 } 1563 1564 // sfn will be valid whenever mcall is valid now because of inheritance 1565 if (observe_safepoint) { 1566 // Handle special safepoint nodes for synchronization 1567 if (!is_mcall) { 1568 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1569 // !!!!! Stubs only need an oopmap right now, so bail out 1570 if (sfn->jvms()->method() == nullptr) { 1571 // Write the oopmap directly to the code blob??!! 1572 continue; 1573 } 1574 } // End synchronization 1575 1576 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1577 current_offset); 1578 Process_OopMap_Node(mach, current_offset); 1579 } // End if safepoint 1580 1581 // If this is a null check, then add the start of the previous instruction to the list 1582 else if( mach->is_MachNullCheck() ) { 1583 inct_starts[inct_cnt++] = previous_offset; 1584 } 1585 1586 // If this is a branch, then fill in the label with the target BB's label 1587 else if (mach->is_MachBranch()) { 1588 // This requires the TRUE branch target be in succs[0] 1589 uint block_num = block->non_connector_successor(0)->_pre_order; 1590 1591 // Try to replace long branch, 1592 // it is mostly for back branches since forward branch's 1593 // distance is not updated yet. 1594 if (mach->may_be_short_branch()) { 1595 int br_size = n->size(C->regalloc()); 1596 int offset = blk_starts[block_num] - current_offset; 1597 if (block_num >= i) { 1598 // Current and following block's offset are not 1599 // finalized yet, adjust distance by the difference 1600 // between calculated and final offsets of current block. 1601 offset -= (blk_starts[i] - blk_offset); 1602 } 1603 // In the following code a nop could be inserted before 1604 // the branch which will increase the backward distance. 1605 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1606 if (needs_padding && offset <= 0) 1607 offset -= nop_size; 1608 1609 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) { 1610 // We've got a winner. Replace this branch. 1611 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 1612 1613 // Update the jmp_size. 1614 int new_size = replacement->size(C->regalloc()); 1615 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1616 // Insert padding between avoid_back_to_back branches. 1617 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 1618 MachNode *nop = new MachNopNode(); 1619 block->insert_node(nop, j++); 1620 C->cfg()->map_node_to_block(nop, block); 1621 last_inst++; 1622 nop->emit(masm, C->regalloc()); 1623 masm->code()->flush_bundle(true); 1624 current_offset = masm->offset(); 1625 } 1626 #ifdef ASSERT 1627 jmp_target[i] = block_num; 1628 jmp_offset[i] = current_offset - blk_offset; 1629 jmp_size[i] = new_size; 1630 jmp_rule[i] = mach->rule(); 1631 #endif 1632 block->map_node(replacement, j); 1633 mach->subsume_by(replacement, C); 1634 n = replacement; 1635 mach = replacement; 1636 } 1637 } 1638 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1639 } else if (mach->ideal_Opcode() == Op_Jump) { 1640 for (uint h = 0; h < block->_num_succs; h++) { 1641 Block* succs_block = block->_succs[h]; 1642 for (uint j = 1; j < succs_block->num_preds(); j++) { 1643 Node* jpn = succs_block->pred(j); 1644 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1645 uint block_num = succs_block->non_connector()->_pre_order; 1646 Label *blkLabel = &blk_labels[block_num]; 1647 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1648 } 1649 } 1650 } 1651 } else if (!n->is_Proj()) { 1652 // Remember the beginning of the previous instruction, in case 1653 // it's followed by a flag-kill and a null-check. Happens on 1654 // Intel all the time, with add-to-memory kind of opcodes. 1655 previous_offset = current_offset; 1656 } 1657 1658 // Not an else-if! 1659 // If this is a trap based cmp then add its offset to the list. 1660 if (mach->is_TrapBasedCheckNode()) { 1661 inct_starts[inct_cnt++] = current_offset; 1662 } 1663 } 1664 1665 // Verify that there is sufficient space remaining 1666 masm->code()->insts()->maybe_expand_to_ensure_remaining(max_inst_size()); 1667 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) { 1668 C->record_failure("CodeCache is full"); 1669 return; 1670 } 1671 1672 // Save the offset for the listing 1673 #if defined(SUPPORT_OPTO_ASSEMBLY) 1674 if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) { 1675 node_offsets[n->_idx] = masm->offset(); 1676 } 1677 #endif 1678 assert(!C->failing_internal() || C->failure_is_artificial(), "Should not reach here if failing."); 1679 1680 // "Normal" instruction case 1681 DEBUG_ONLY(uint instr_offset = masm->offset()); 1682 n->emit(masm, C->regalloc()); 1683 current_offset = masm->offset(); 1684 1685 // Above we only verified that there is enough space in the instruction section. 1686 // However, the instruction may emit stubs that cause code buffer expansion. 1687 // Bail out here if expansion failed due to a lack of code cache space. 1688 if (C->failing()) { 1689 return; 1690 } 1691 1692 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset), 1693 "ret_addr_offset() not within emitted code"); 1694 1695 #ifdef ASSERT 1696 uint n_size = n->size(C->regalloc()); 1697 if (n_size < (current_offset-instr_offset)) { 1698 MachNode* mach = n->as_Mach(); 1699 n->dump(); 1700 mach->dump_format(C->regalloc(), tty); 1701 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset); 1702 Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty); 1703 tty->print_cr(" ------------------- "); 1704 BufferBlob* blob = this->scratch_buffer_blob(); 1705 address blob_begin = blob->content_begin(); 1706 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty); 1707 assert(false, "wrong size of mach node"); 1708 } 1709 #endif 1710 non_safepoints.observe_instruction(n, current_offset); 1711 1712 // mcall is last "call" that can be a safepoint 1713 // record it so we can see if a poll will directly follow it 1714 // in which case we'll need a pad to make the PcDesc sites unique 1715 // see 5010568. This can be slightly inaccurate but conservative 1716 // in the case that return address is not actually at current_offset. 1717 // This is a small price to pay. 1718 1719 if (is_mcall) { 1720 last_call_offset = current_offset; 1721 } 1722 1723 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 1724 // Avoid back to back some instructions. 1725 last_avoid_back_to_back_offset = current_offset; 1726 } 1727 1728 } // End for all instructions in block 1729 1730 // If the next block is the top of a loop, pad this block out to align 1731 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1732 if (i < nblocks-1) { 1733 Block *nb = C->cfg()->get_block(i + 1); 1734 int padding = nb->alignment_padding(current_offset); 1735 if( padding > 0 ) { 1736 MachNode *nop = new MachNopNode(padding / nop_size); 1737 block->insert_node(nop, block->number_of_nodes()); 1738 C->cfg()->map_node_to_block(nop, block); 1739 nop->emit(masm, C->regalloc()); 1740 current_offset = masm->offset(); 1741 } 1742 } 1743 // Verify that the distance for generated before forward 1744 // short branches is still valid. 1745 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size"); 1746 1747 // Save new block start offset 1748 blk_starts[i] = blk_offset; 1749 } // End of for all blocks 1750 blk_starts[nblocks] = current_offset; 1751 1752 non_safepoints.flush_at_end(); 1753 1754 // Offset too large? 1755 if (C->failing()) return; 1756 1757 // Define a pseudo-label at the end of the code 1758 masm->bind( blk_labels[nblocks] ); 1759 1760 // Compute the size of the first block 1761 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1762 1763 #ifdef ASSERT 1764 for (uint i = 0; i < nblocks; i++) { // For all blocks 1765 if (jmp_target[i] != 0) { 1766 int br_size = jmp_size[i]; 1767 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1768 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1769 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]); 1770 assert(false, "Displacement too large for short jmp"); 1771 } 1772 } 1773 } 1774 #endif 1775 1776 if (!masm->code()->finalize_stubs()) { 1777 C->record_failure("CodeCache is full"); 1778 return; 1779 } 1780 1781 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1782 bs->emit_stubs(*masm->code()); 1783 if (C->failing()) return; 1784 1785 // Fill in stubs. 1786 assert(masm->inst_mark() == nullptr, "should be."); 1787 _stub_list.emit(*masm); 1788 if (C->failing()) return; 1789 1790 #ifndef PRODUCT 1791 // Information on the size of the method, without the extraneous code 1792 Scheduling::increment_method_size(masm->offset()); 1793 #endif 1794 1795 // ------------------ 1796 // Fill in exception table entries. 1797 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1798 1799 // Only java methods have exception handlers and deopt handlers 1800 // class HandlerImpl is platform-specific and defined in the *.ad files. 1801 if (C->method()) { 1802 // Emit the exception handler code. 1803 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(masm)); 1804 if (C->failing()) { 1805 return; // CodeBuffer::expand failed 1806 } 1807 // Emit the deopt handler code. 1808 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(masm)); 1809 1810 // Emit the MethodHandle deopt handler code (if required). 1811 if (C->has_method_handle_invokes() && !C->failing()) { 1812 // We can use the same code as for the normal deopt handler, we 1813 // just need a different entry point address. 1814 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(masm)); 1815 } 1816 } 1817 1818 // One last check for failed CodeBuffer::expand: 1819 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) { 1820 C->record_failure("CodeCache is full"); 1821 return; 1822 } 1823 1824 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY) 1825 if (C->print_assembly()) { 1826 tty->cr(); 1827 tty->print_cr("============================= C2-compiled nmethod =============================="); 1828 } 1829 #endif 1830 1831 #if defined(SUPPORT_OPTO_ASSEMBLY) 1832 // Dump the assembly code, including basic-block numbers 1833 if (C->print_assembly()) { 1834 ttyLocker ttyl; // keep the following output all in one block 1835 if (!VMThread::should_terminate()) { // test this under the tty lock 1836 // print_metadata and dump_asm may safepoint which makes us loose the ttylock. 1837 // We call them first and write to a stringStream, then we retake the lock to 1838 // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe. 1839 ResourceMark rm; 1840 stringStream method_metadata_str; 1841 if (C->method() != nullptr) { 1842 C->method()->print_metadata(&method_metadata_str); 1843 } 1844 stringStream dump_asm_str; 1845 dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit); 1846 1847 NoSafepointVerifier nsv; 1848 ttyLocker ttyl2; 1849 // This output goes directly to the tty, not the compiler log. 1850 // To enable tools to match it up with the compilation activity, 1851 // be sure to tag this tty output with the compile ID. 1852 if (xtty != nullptr) { 1853 xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(), 1854 C->is_osr_compilation() ? " compile_kind='osr'" : 1855 (C->for_preload() ? " compile_kind='AP'" : "")); 1856 } 1857 if (C->method() != nullptr) { 1858 tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id()); 1859 tty->print_raw(method_metadata_str.freeze()); 1860 } else if (C->stub_name() != nullptr) { 1861 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name()); 1862 } 1863 tty->cr(); 1864 tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id()); 1865 tty->print_raw(dump_asm_str.freeze()); 1866 tty->print_cr("--------------------------------------------------------------------------------"); 1867 if (xtty != nullptr) { 1868 xtty->tail("opto_assembly"); 1869 } 1870 } 1871 } 1872 #endif 1873 } 1874 1875 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1876 _inc_table.set_size(cnt); 1877 1878 uint inct_cnt = 0; 1879 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) { 1880 Block* block = C->cfg()->get_block(i); 1881 Node *n = nullptr; 1882 int j; 1883 1884 // Find the branch; ignore trailing NOPs. 1885 for (j = block->number_of_nodes() - 1; j >= 0; j--) { 1886 n = block->get_node(j); 1887 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) { 1888 break; 1889 } 1890 } 1891 1892 // If we didn't find anything, continue 1893 if (j < 0) { 1894 continue; 1895 } 1896 1897 // Compute ExceptionHandlerTable subtable entry and add it 1898 // (skip empty blocks) 1899 if (n->is_Catch()) { 1900 1901 // Get the offset of the return from the call 1902 uint call_return = call_returns[block->_pre_order]; 1903 #ifdef ASSERT 1904 assert( call_return > 0, "no call seen for this basic block" ); 1905 while (block->get_node(--j)->is_MachProj()) ; 1906 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); 1907 #endif 1908 // last instruction is a CatchNode, find it's CatchProjNodes 1909 int nof_succs = block->_num_succs; 1910 // allocate space 1911 GrowableArray<intptr_t> handler_bcis(nof_succs); 1912 GrowableArray<intptr_t> handler_pcos(nof_succs); 1913 // iterate through all successors 1914 for (int j = 0; j < nof_succs; j++) { 1915 Block* s = block->_succs[j]; 1916 bool found_p = false; 1917 for (uint k = 1; k < s->num_preds(); k++) { 1918 Node* pk = s->pred(k); 1919 if (pk->is_CatchProj() && pk->in(0) == n) { 1920 const CatchProjNode* p = pk->as_CatchProj(); 1921 found_p = true; 1922 // add the corresponding handler bci & pco information 1923 if (p->_con != CatchProjNode::fall_through_index) { 1924 // p leads to an exception handler (and is not fall through) 1925 assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering"); 1926 // no duplicates, please 1927 if (!handler_bcis.contains(p->handler_bci())) { 1928 uint block_num = s->non_connector()->_pre_order; 1929 handler_bcis.append(p->handler_bci()); 1930 handler_pcos.append(blk_labels[block_num].loc_pos()); 1931 } 1932 } 1933 } 1934 } 1935 assert(found_p, "no matching predecessor found"); 1936 // Note: Due to empty block removal, one block may have 1937 // several CatchProj inputs, from the same Catch. 1938 } 1939 1940 // Set the offset of the return from the call 1941 assert(handler_bcis.find(-1) != -1, "must have default handler"); 1942 _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos); 1943 continue; 1944 } 1945 1946 // Handle implicit null exception table updates 1947 if (n->is_MachNullCheck()) { 1948 MachNode* access = n->in(1)->as_Mach(); 1949 assert(access->barrier_data() == 0 || 1950 access->is_late_expanded_null_check_candidate(), 1951 "Implicit null checks on memory accesses with barriers are only supported on nodes explicitly marked as null-check candidates"); 1952 uint block_num = block->non_connector_successor(0)->_pre_order; 1953 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1954 continue; 1955 } 1956 // Handle implicit exception table updates: trap instructions. 1957 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) { 1958 uint block_num = block->non_connector_successor(0)->_pre_order; 1959 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1960 continue; 1961 } 1962 } // End of for all blocks fill in exception table entries 1963 } 1964 1965 // Static Variables 1966 #ifndef PRODUCT 1967 uint Scheduling::_total_nop_size = 0; 1968 uint Scheduling::_total_method_size = 0; 1969 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1970 #endif 1971 1972 // Initializer for class Scheduling 1973 1974 Scheduling::Scheduling(Arena *arena, Compile &compile) 1975 : _arena(arena), 1976 _cfg(compile.cfg()), 1977 _regalloc(compile.regalloc()), 1978 _scheduled(arena), 1979 _available(arena), 1980 _reg_node(arena), 1981 _pinch_free_list(arena), 1982 _next_node(nullptr), 1983 _bundle_instr_count(0), 1984 _bundle_cycle_number(0), 1985 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]) 1986 { 1987 // Save the count 1988 _node_bundling_limit = compile.unique(); 1989 uint node_max = _regalloc->node_regs_max_index(); 1990 1991 compile.output()->set_node_bundling_limit(_node_bundling_limit); 1992 1993 // This one is persistent within the Compile class 1994 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1995 1996 // Allocate space for fixed-size arrays 1997 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1998 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1999 2000 // Clear the arrays 2001 for (uint i = 0; i < node_max; i++) { 2002 ::new (&_node_bundling_base[i]) Bundle(); 2003 } 2004 memset(_uses, 0, node_max * sizeof(short)); 2005 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 2006 2007 // Clear the bundling information 2008 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements)); 2009 2010 // Get the last node 2011 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1); 2012 2013 _next_node = block->get_node(block->number_of_nodes() - 1); 2014 } 2015 2016 // Step ahead "i" cycles 2017 void Scheduling::step(uint i) { 2018 2019 Bundle *bundle = node_bundling(_next_node); 2020 bundle->set_starts_bundle(); 2021 2022 // Update the bundle record, but leave the flags information alone 2023 if (_bundle_instr_count > 0) { 2024 bundle->set_instr_count(_bundle_instr_count); 2025 bundle->set_resources_used(_bundle_use.resourcesUsed()); 2026 } 2027 2028 // Update the state information 2029 _bundle_instr_count = 0; 2030 _bundle_cycle_number += i; 2031 _bundle_use.step(i); 2032 } 2033 2034 void Scheduling::step_and_clear() { 2035 Bundle *bundle = node_bundling(_next_node); 2036 bundle->set_starts_bundle(); 2037 2038 // Update the bundle record 2039 if (_bundle_instr_count > 0) { 2040 bundle->set_instr_count(_bundle_instr_count); 2041 bundle->set_resources_used(_bundle_use.resourcesUsed()); 2042 2043 _bundle_cycle_number += 1; 2044 } 2045 2046 // Clear the bundling information 2047 _bundle_instr_count = 0; 2048 _bundle_use.reset(); 2049 2050 memcpy(_bundle_use_elements, 2051 Pipeline_Use::elaborated_elements, 2052 sizeof(Pipeline_Use::elaborated_elements)); 2053 } 2054 2055 // Perform instruction scheduling and bundling over the sequence of 2056 // instructions in backwards order. 2057 void PhaseOutput::ScheduleAndBundle() { 2058 2059 // Don't optimize this if it isn't a method 2060 if (!C->method()) 2061 return; 2062 2063 // Don't optimize this if scheduling is disabled 2064 if (!C->do_scheduling()) 2065 return; 2066 2067 // Scheduling code works only with pairs (8 bytes) maximum. 2068 // And when the scalable vector register is used, we may spill/unspill 2069 // the whole reg regardless of the max vector size. 2070 if (C->max_vector_size() > 8 || 2071 (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) { 2072 return; 2073 } 2074 2075 Compile::TracePhase tp(_t_instrSched); 2076 2077 // Create a data structure for all the scheduling information 2078 Scheduling scheduling(Thread::current()->resource_area(), *C); 2079 2080 // Walk backwards over each basic block, computing the needed alignment 2081 // Walk over all the basic blocks 2082 scheduling.DoScheduling(); 2083 2084 #ifndef PRODUCT 2085 if (C->trace_opto_output()) { 2086 // Buffer and print all at once 2087 ResourceMark rm; 2088 stringStream ss; 2089 ss.print("\n---- After ScheduleAndBundle ----\n"); 2090 print_scheduling(&ss); 2091 tty->print("%s", ss.as_string()); 2092 } 2093 #endif 2094 } 2095 2096 #ifndef PRODUCT 2097 // Separated out so that it can be called directly from debugger 2098 void PhaseOutput::print_scheduling() { 2099 print_scheduling(tty); 2100 } 2101 2102 void PhaseOutput::print_scheduling(outputStream* output_stream) { 2103 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) { 2104 output_stream->print("\nBB#%03d:\n", i); 2105 Block* block = C->cfg()->get_block(i); 2106 for (uint j = 0; j < block->number_of_nodes(); j++) { 2107 Node* n = block->get_node(j); 2108 OptoReg::Name reg = C->regalloc()->get_reg_first(n); 2109 output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 2110 n->dump("\n", false, output_stream); 2111 } 2112 } 2113 } 2114 #endif 2115 2116 // See if this node fits into the present instruction bundle 2117 bool Scheduling::NodeFitsInBundle(Node *n) { 2118 uint n_idx = n->_idx; 2119 2120 // If the node cannot be scheduled this cycle, skip it 2121 if (_current_latency[n_idx] > _bundle_cycle_number) { 2122 #ifndef PRODUCT 2123 if (_cfg->C->trace_opto_output()) 2124 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 2125 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 2126 #endif 2127 return (false); 2128 } 2129 2130 const Pipeline *node_pipeline = n->pipeline(); 2131 2132 uint instruction_count = node_pipeline->instructionCount(); 2133 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2134 instruction_count = 0; 2135 2136 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 2137 #ifndef PRODUCT 2138 if (_cfg->C->trace_opto_output()) 2139 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 2140 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 2141 #endif 2142 return (false); 2143 } 2144 2145 // Don't allow non-machine nodes to be handled this way 2146 if (!n->is_Mach() && instruction_count == 0) 2147 return (false); 2148 2149 // See if there is any overlap 2150 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 2151 2152 if (delay > 0) { 2153 #ifndef PRODUCT 2154 if (_cfg->C->trace_opto_output()) 2155 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 2156 #endif 2157 return false; 2158 } 2159 2160 #ifndef PRODUCT 2161 if (_cfg->C->trace_opto_output()) 2162 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 2163 #endif 2164 2165 return true; 2166 } 2167 2168 Node * Scheduling::ChooseNodeToBundle() { 2169 uint siz = _available.size(); 2170 2171 if (siz == 0) { 2172 2173 #ifndef PRODUCT 2174 if (_cfg->C->trace_opto_output()) 2175 tty->print("# ChooseNodeToBundle: null\n"); 2176 #endif 2177 return (nullptr); 2178 } 2179 2180 // Fast path, if only 1 instruction in the bundle 2181 if (siz == 1) { 2182 #ifndef PRODUCT 2183 if (_cfg->C->trace_opto_output()) { 2184 tty->print("# ChooseNodeToBundle (only 1): "); 2185 _available[0]->dump(); 2186 } 2187 #endif 2188 return (_available[0]); 2189 } 2190 2191 // Don't bother, if the bundle is already full 2192 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 2193 for ( uint i = 0; i < siz; i++ ) { 2194 Node *n = _available[i]; 2195 2196 // Skip projections, we'll handle them another way 2197 if (n->is_Proj()) 2198 continue; 2199 2200 // This presupposed that instructions are inserted into the 2201 // available list in a legality order; i.e. instructions that 2202 // must be inserted first are at the head of the list 2203 if (NodeFitsInBundle(n)) { 2204 #ifndef PRODUCT 2205 if (_cfg->C->trace_opto_output()) { 2206 tty->print("# ChooseNodeToBundle: "); 2207 n->dump(); 2208 } 2209 #endif 2210 return (n); 2211 } 2212 } 2213 } 2214 2215 // Nothing fits in this bundle, choose the highest priority 2216 #ifndef PRODUCT 2217 if (_cfg->C->trace_opto_output()) { 2218 tty->print("# ChooseNodeToBundle: "); 2219 _available[0]->dump(); 2220 } 2221 #endif 2222 2223 return _available[0]; 2224 } 2225 2226 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) { 2227 assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), ""); 2228 2229 OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1)); 2230 OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first); 2231 OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1)); 2232 OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second); 2233 2234 // Comparison between stack -> reg and stack -> reg 2235 if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) && 2236 OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) { 2237 return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo); 2238 } 2239 2240 // Comparison between reg -> stack and reg -> stack 2241 if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) && 2242 OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) { 2243 return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo); 2244 } 2245 2246 return 0; // Not comparable 2247 } 2248 2249 void Scheduling::AddNodeToAvailableList(Node *n) { 2250 assert( !n->is_Proj(), "projections never directly made available" ); 2251 #ifndef PRODUCT 2252 if (_cfg->C->trace_opto_output()) { 2253 tty->print("# AddNodeToAvailableList: "); 2254 n->dump(); 2255 } 2256 #endif 2257 2258 int latency = _current_latency[n->_idx]; 2259 2260 // Insert in latency order (insertion sort). If two MachSpillCopyNodes 2261 // for stack spilling or unspilling have the same latency, we sort 2262 // them in the order of stack offset. Some ports (e.g. aarch64) may also 2263 // have more opportunities to do ld/st merging 2264 uint i; 2265 for (i = 0; i < _available.size(); i++) { 2266 if (_current_latency[_available[i]->_idx] > latency) { 2267 break; 2268 } else if (_current_latency[_available[i]->_idx] == latency && 2269 n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() && 2270 compare_two_spill_nodes(n, _available[i]) > 0) { 2271 break; 2272 } 2273 } 2274 2275 // Special Check for compares following branches 2276 if( n->is_Mach() && _scheduled.size() > 0 ) { 2277 int op = n->as_Mach()->ideal_Opcode(); 2278 Node *last = _scheduled[0]; 2279 if( last->is_MachIf() && last->in(1) == n && 2280 ( op == Op_CmpI || 2281 op == Op_CmpU || 2282 op == Op_CmpUL || 2283 op == Op_CmpP || 2284 op == Op_CmpF || 2285 op == Op_CmpD || 2286 op == Op_CmpL ) ) { 2287 2288 // Recalculate position, moving to front of same latency 2289 for ( i=0 ; i < _available.size(); i++ ) 2290 if (_current_latency[_available[i]->_idx] >= latency) 2291 break; 2292 } 2293 } 2294 2295 // Insert the node in the available list 2296 _available.insert(i, n); 2297 2298 #ifndef PRODUCT 2299 if (_cfg->C->trace_opto_output()) 2300 dump_available(); 2301 #endif 2302 } 2303 2304 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 2305 for ( uint i=0; i < n->len(); i++ ) { 2306 Node *def = n->in(i); 2307 if (!def) continue; 2308 if( def->is_Proj() ) // If this is a machine projection, then 2309 def = def->in(0); // propagate usage thru to the base instruction 2310 2311 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local 2312 continue; 2313 } 2314 2315 // Compute the latency 2316 uint l = _bundle_cycle_number + n->latency(i); 2317 if (_current_latency[def->_idx] < l) 2318 _current_latency[def->_idx] = l; 2319 2320 // If this does not have uses then schedule it 2321 if ((--_uses[def->_idx]) == 0) 2322 AddNodeToAvailableList(def); 2323 } 2324 } 2325 2326 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2327 #ifndef PRODUCT 2328 if (_cfg->C->trace_opto_output()) { 2329 tty->print("# AddNodeToBundle: "); 2330 n->dump(); 2331 } 2332 #endif 2333 2334 // Remove this from the available list 2335 uint i; 2336 for (i = 0; i < _available.size(); i++) 2337 if (_available[i] == n) 2338 break; 2339 assert(i < _available.size(), "entry in _available list not found"); 2340 _available.remove(i); 2341 2342 // See if this fits in the current bundle 2343 const Pipeline *node_pipeline = n->pipeline(); 2344 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2345 2346 2347 // Get the number of instructions 2348 uint instruction_count = node_pipeline->instructionCount(); 2349 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2350 instruction_count = 0; 2351 2352 // Compute the latency information 2353 uint delay = 0; 2354 2355 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2356 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2357 if (relative_latency < 0) 2358 relative_latency = 0; 2359 2360 delay = _bundle_use.full_latency(relative_latency, node_usage); 2361 2362 // Does not fit in this bundle, start a new one 2363 if (delay > 0) { 2364 step(delay); 2365 2366 #ifndef PRODUCT 2367 if (_cfg->C->trace_opto_output()) 2368 tty->print("# *** STEP(%d) ***\n", delay); 2369 #endif 2370 } 2371 } 2372 2373 if (delay == 0) { 2374 if (node_pipeline->hasMultipleBundles()) { 2375 #ifndef PRODUCT 2376 if (_cfg->C->trace_opto_output()) 2377 tty->print("# *** STEP(multiple instructions) ***\n"); 2378 #endif 2379 step(1); 2380 } 2381 2382 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2383 #ifndef PRODUCT 2384 if (_cfg->C->trace_opto_output()) 2385 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2386 instruction_count + _bundle_instr_count, 2387 Pipeline::_max_instrs_per_cycle); 2388 #endif 2389 step(1); 2390 } 2391 } 2392 2393 // Set the node's latency 2394 _current_latency[n->_idx] = _bundle_cycle_number; 2395 2396 // Now merge the functional unit information 2397 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2398 _bundle_use.add_usage(node_usage); 2399 2400 // Increment the number of instructions in this bundle 2401 _bundle_instr_count += instruction_count; 2402 2403 // Remember this node for later 2404 if (n->is_Mach()) 2405 _next_node = n; 2406 2407 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2408 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2409 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2410 // into the block. All other scheduled nodes get put in the schedule here. 2411 int op = n->Opcode(); 2412 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2413 (op != Op_Node && // Not an unused antidepedence node and 2414 // not an unallocated boxlock 2415 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2416 2417 // Push any trailing projections 2418 if( bb->get_node(bb->number_of_nodes()-1) != n ) { 2419 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2420 Node *foi = n->fast_out(i); 2421 if( foi->is_Proj() ) 2422 _scheduled.push(foi); 2423 } 2424 } 2425 2426 // Put the instruction in the schedule list 2427 _scheduled.push(n); 2428 } 2429 2430 #ifndef PRODUCT 2431 if (_cfg->C->trace_opto_output()) 2432 dump_available(); 2433 #endif 2434 2435 // Walk all the definitions, decrementing use counts, and 2436 // if a definition has a 0 use count, place it in the available list. 2437 DecrementUseCounts(n,bb); 2438 } 2439 2440 // This method sets the use count within a basic block. We will ignore all 2441 // uses outside the current basic block. As we are doing a backwards walk, 2442 // any node we reach that has a use count of 0 may be scheduled. This also 2443 // avoids the problem of cyclic references from phi nodes, as long as phi 2444 // nodes are at the front of the basic block. This method also initializes 2445 // the available list to the set of instructions that have no uses within this 2446 // basic block. 2447 void Scheduling::ComputeUseCount(const Block *bb) { 2448 #ifndef PRODUCT 2449 if (_cfg->C->trace_opto_output()) 2450 tty->print("# -> ComputeUseCount\n"); 2451 #endif 2452 2453 // Clear the list of available and scheduled instructions, just in case 2454 _available.clear(); 2455 _scheduled.clear(); 2456 2457 #ifdef ASSERT 2458 for( uint i=0; i < bb->number_of_nodes(); i++ ) 2459 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" ); 2460 #endif 2461 2462 // Force the _uses count to never go to zero for unscheduable pieces 2463 // of the block 2464 for( uint k = 0; k < _bb_start; k++ ) 2465 _uses[bb->get_node(k)->_idx] = 1; 2466 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ ) 2467 _uses[bb->get_node(l)->_idx] = 1; 2468 2469 // Iterate backwards over the instructions in the block. Don't count the 2470 // branch projections at end or the block header instructions. 2471 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2472 Node *n = bb->get_node(j); 2473 if( n->is_Proj() ) continue; // Projections handled another way 2474 2475 // Account for all uses 2476 for ( uint k = 0; k < n->len(); k++ ) { 2477 Node *inp = n->in(k); 2478 if (!inp) continue; 2479 assert(inp != n, "no cycles allowed" ); 2480 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use? 2481 if (inp->is_Proj()) { // Skip through Proj's 2482 inp = inp->in(0); 2483 } 2484 ++_uses[inp->_idx]; // Count 1 block-local use 2485 } 2486 } 2487 2488 // If this instruction has a 0 use count, then it is available 2489 if (!_uses[n->_idx]) { 2490 _current_latency[n->_idx] = _bundle_cycle_number; 2491 AddNodeToAvailableList(n); 2492 } 2493 2494 #ifndef PRODUCT 2495 if (_cfg->C->trace_opto_output()) { 2496 tty->print("# uses: %3d: ", _uses[n->_idx]); 2497 n->dump(); 2498 } 2499 #endif 2500 } 2501 2502 #ifndef PRODUCT 2503 if (_cfg->C->trace_opto_output()) 2504 tty->print("# <- ComputeUseCount\n"); 2505 #endif 2506 } 2507 2508 // This routine performs scheduling on each basic block in reverse order, 2509 // using instruction latencies and taking into account function unit 2510 // availability. 2511 void Scheduling::DoScheduling() { 2512 #ifndef PRODUCT 2513 if (_cfg->C->trace_opto_output()) 2514 tty->print("# -> DoScheduling\n"); 2515 #endif 2516 2517 Block *succ_bb = nullptr; 2518 Block *bb; 2519 Compile* C = Compile::current(); 2520 2521 // Walk over all the basic blocks in reverse order 2522 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) { 2523 bb = _cfg->get_block(i); 2524 2525 #ifndef PRODUCT 2526 if (_cfg->C->trace_opto_output()) { 2527 tty->print("# Schedule BB#%03d (initial)\n", i); 2528 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2529 bb->get_node(j)->dump(); 2530 } 2531 } 2532 #endif 2533 2534 // On the head node, skip processing 2535 if (bb == _cfg->get_root_block()) { 2536 continue; 2537 } 2538 2539 // Skip empty, connector blocks 2540 if (bb->is_connector()) 2541 continue; 2542 2543 // If the following block is not the sole successor of 2544 // this one, then reset the pipeline information 2545 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2546 #ifndef PRODUCT 2547 if (_cfg->C->trace_opto_output()) { 2548 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2549 _next_node->_idx, _bundle_instr_count); 2550 } 2551 #endif 2552 step_and_clear(); 2553 } 2554 2555 // Leave untouched the starting instruction, any Phis, a CreateEx node 2556 // or Top. bb->get_node(_bb_start) is the first schedulable instruction. 2557 _bb_end = bb->number_of_nodes()-1; 2558 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2559 Node *n = bb->get_node(_bb_start); 2560 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2561 // Also, MachIdealNodes do not get scheduled 2562 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2563 MachNode *mach = n->as_Mach(); 2564 int iop = mach->ideal_Opcode(); 2565 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2566 if( iop == Op_Con ) continue; // Do not schedule Top 2567 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2568 mach->pipeline() == MachNode::pipeline_class() && 2569 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc 2570 continue; 2571 break; // Funny loop structure to be sure... 2572 } 2573 // Compute last "interesting" instruction in block - last instruction we 2574 // might schedule. _bb_end points just after last schedulable inst. 2575 Node *last = bb->get_node(_bb_end); 2576 // Ignore trailing NOPs. 2577 while (_bb_end > 0 && last->is_Mach() && 2578 last->as_Mach()->ideal_Opcode() == Op_Con) { 2579 last = bb->get_node(--_bb_end); 2580 } 2581 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2582 if( last->is_Catch() || 2583 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2584 // There might be a prior call. Skip it. 2585 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj()); 2586 } else if( last->is_MachNullCheck() ) { 2587 // Backup so the last null-checked memory instruction is 2588 // outside the schedulable range. Skip over the nullcheck, 2589 // projection, and the memory nodes. 2590 Node *mem = last->in(1); 2591 do { 2592 _bb_end--; 2593 } while (mem != bb->get_node(_bb_end)); 2594 } else { 2595 // Set _bb_end to point after last schedulable inst. 2596 _bb_end++; 2597 } 2598 2599 assert( _bb_start <= _bb_end, "inverted block ends" ); 2600 2601 // Compute the register antidependencies for the basic block 2602 ComputeRegisterAntidependencies(bb); 2603 if (C->failing()) return; // too many D-U pinch points 2604 2605 // Compute the usage within the block, and set the list of all nodes 2606 // in the block that have no uses within the block. 2607 ComputeUseCount(bb); 2608 2609 // Schedule the remaining instructions in the block 2610 while ( _available.size() > 0 ) { 2611 Node *n = ChooseNodeToBundle(); 2612 guarantee(n != nullptr, "no nodes available"); 2613 AddNodeToBundle(n,bb); 2614 } 2615 2616 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2617 #ifdef ASSERT 2618 for( uint l = _bb_start; l < _bb_end; l++ ) { 2619 Node *n = bb->get_node(l); 2620 uint m; 2621 for( m = 0; m < _bb_end-_bb_start; m++ ) 2622 if( _scheduled[m] == n ) 2623 break; 2624 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2625 } 2626 #endif 2627 2628 // Now copy the instructions (in reverse order) back to the block 2629 for ( uint k = _bb_start; k < _bb_end; k++ ) 2630 bb->map_node(_scheduled[_bb_end-k-1], k); 2631 2632 #ifndef PRODUCT 2633 if (_cfg->C->trace_opto_output()) { 2634 tty->print("# Schedule BB#%03d (final)\n", i); 2635 uint current = 0; 2636 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2637 Node *n = bb->get_node(j); 2638 if( valid_bundle_info(n) ) { 2639 Bundle *bundle = node_bundling(n); 2640 if (bundle->instr_count() > 0) { 2641 tty->print("*** Bundle: "); 2642 bundle->dump(); 2643 } 2644 n->dump(); 2645 } 2646 } 2647 } 2648 #endif 2649 #ifdef ASSERT 2650 verify_good_schedule(bb,"after block local scheduling"); 2651 #endif 2652 } 2653 2654 #ifndef PRODUCT 2655 if (_cfg->C->trace_opto_output()) 2656 tty->print("# <- DoScheduling\n"); 2657 #endif 2658 2659 // Record final node-bundling array location 2660 _regalloc->C->output()->set_node_bundling_base(_node_bundling_base); 2661 2662 } // end DoScheduling 2663 2664 // Verify that no live-range used in the block is killed in the block by a 2665 // wrong DEF. This doesn't verify live-ranges that span blocks. 2666 2667 // Check for edge existence. Used to avoid adding redundant precedence edges. 2668 static bool edge_from_to( Node *from, Node *to ) { 2669 for( uint i=0; i<from->len(); i++ ) 2670 if( from->in(i) == to ) 2671 return true; 2672 return false; 2673 } 2674 2675 #ifdef ASSERT 2676 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2677 // Check for bad kills 2678 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2679 Node *prior_use = _reg_node[def]; 2680 if( prior_use && !edge_from_to(prior_use,n) ) { 2681 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2682 n->dump(); 2683 tty->print_cr("..."); 2684 prior_use->dump(); 2685 assert(edge_from_to(prior_use,n), "%s", msg); 2686 } 2687 _reg_node.map(def,nullptr); // Kill live USEs 2688 } 2689 } 2690 2691 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2692 2693 // Zap to something reasonable for the verify code 2694 _reg_node.clear(); 2695 2696 // Walk over the block backwards. Check to make sure each DEF doesn't 2697 // kill a live value (other than the one it's supposed to). Add each 2698 // USE to the live set. 2699 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) { 2700 Node *n = b->get_node(i); 2701 int n_op = n->Opcode(); 2702 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2703 // Fat-proj kills a slew of registers 2704 RegMaskIterator rmi(n->out_RegMask()); 2705 while (rmi.has_next()) { 2706 OptoReg::Name kill = rmi.next(); 2707 verify_do_def(n, kill, msg); 2708 } 2709 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2710 // Get DEF'd registers the normal way 2711 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2712 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2713 } 2714 2715 // Now make all USEs live 2716 for( uint i=1; i<n->req(); i++ ) { 2717 Node *def = n->in(i); 2718 assert(def != nullptr, "input edge required"); 2719 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2720 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2721 if( OptoReg::is_valid(reg_lo) ) { 2722 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg); 2723 _reg_node.map(reg_lo,n); 2724 } 2725 if( OptoReg::is_valid(reg_hi) ) { 2726 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg); 2727 _reg_node.map(reg_hi,n); 2728 } 2729 } 2730 2731 } 2732 2733 // Zap to something reasonable for the Antidependence code 2734 _reg_node.clear(); 2735 } 2736 #endif 2737 2738 // Conditionally add precedence edges. Avoid putting edges on Projs. 2739 static void add_prec_edge_from_to( Node *from, Node *to ) { 2740 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2741 assert( from->req() == 1 && (from->len() == 1 || from->in(1) == nullptr), "no precedence edges on projections" ); 2742 from = from->in(0); 2743 } 2744 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2745 !edge_from_to( from, to ) ) // Avoid duplicate edge 2746 from->add_prec(to); 2747 } 2748 2749 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2750 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2751 return; 2752 2753 if (OptoReg::is_reg(def_reg)) { 2754 VMReg vmreg = OptoReg::as_VMReg(def_reg); 2755 if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) { 2756 // This is one of the high slots of a vector register. 2757 // ScheduleAndBundle already checked there are no live wide 2758 // vectors in this method so it can be safely ignored. 2759 return; 2760 } 2761 } 2762 2763 Node *pinch = _reg_node[def_reg]; // Get pinch point 2764 if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet? 2765 is_def ) { // Check for a true def (not a kill) 2766 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2767 return; 2768 } 2769 2770 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2771 DEBUG_ONLY( def = (Node*)((intptr_t)0xdeadbeef); ) 2772 2773 // After some number of kills there _may_ be a later def 2774 Node *later_def = nullptr; 2775 2776 Compile* C = Compile::current(); 2777 2778 // Finding a kill requires a real pinch-point. 2779 // Check for not already having a pinch-point. 2780 // Pinch points are Op_Node's. 2781 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2782 later_def = pinch; // Must be def/kill as optimistic pinch-point 2783 if ( _pinch_free_list.size() > 0) { 2784 pinch = _pinch_free_list.pop(); 2785 } else { 2786 pinch = new Node(1); // Pinch point to-be 2787 } 2788 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2789 DEBUG_ONLY( pinch->dump(); ); 2790 assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index()); 2791 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2792 return; 2793 } 2794 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init) 2795 _reg_node.map(def_reg,pinch); // Record pinch-point 2796 //regalloc()->set_bad(pinch->_idx); // Already initialized this way. 2797 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2798 pinch->init_req(0, C->top()); // set not null for the next call 2799 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2800 later_def = nullptr; // and no later def 2801 } 2802 pinch->set_req(0,later_def); // Hook later def so we can find it 2803 } else { // Else have valid pinch point 2804 if( pinch->in(0) ) // If there is a later-def 2805 later_def = pinch->in(0); // Get it 2806 } 2807 2808 // Add output-dependence edge from later def to kill 2809 if( later_def ) // If there is some original def 2810 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2811 2812 // See if current kill is also a use, and so is forced to be the pinch-point. 2813 if( pinch->Opcode() == Op_Node ) { 2814 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2815 for( uint i=1; i<uses->req(); i++ ) { 2816 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2817 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2818 // Yes, found a use/kill pinch-point 2819 pinch->set_req(0,nullptr); // 2820 pinch->replace_by(kill); // Move anti-dep edges up 2821 pinch = kill; 2822 _reg_node.map(def_reg,pinch); 2823 return; 2824 } 2825 } 2826 } 2827 2828 // Add edge from kill to pinch-point 2829 add_prec_edge_from_to(kill,pinch); 2830 } 2831 2832 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2833 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2834 return; 2835 Node *pinch = _reg_node[use_reg]; // Get pinch point 2836 // Check for no later def_reg/kill in block 2837 if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b && 2838 // Use has to be block-local as well 2839 _cfg->get_block_for_node(use) == b) { 2840 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2841 pinch->req() == 1 ) { // pinch not yet in block? 2842 pinch->del_req(0); // yank pointer to later-def, also set flag 2843 // Insert the pinch-point in the block just after the last use 2844 b->insert_node(pinch, b->find_node(use) + 1); 2845 _bb_end++; // Increase size scheduled region in block 2846 } 2847 2848 add_prec_edge_from_to(pinch,use); 2849 } 2850 } 2851 2852 // We insert antidependences between the reads and following write of 2853 // allocated registers to prevent illegal code motion. Hopefully, the 2854 // number of added references should be fairly small, especially as we 2855 // are only adding references within the current basic block. 2856 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2857 2858 #ifdef ASSERT 2859 verify_good_schedule(b,"before block local scheduling"); 2860 #endif 2861 2862 // A valid schedule, for each register independently, is an endless cycle 2863 // of: a def, then some uses (connected to the def by true dependencies), 2864 // then some kills (defs with no uses), finally the cycle repeats with a new 2865 // def. The uses are allowed to float relative to each other, as are the 2866 // kills. No use is allowed to slide past a kill (or def). This requires 2867 // antidependencies between all uses of a single def and all kills that 2868 // follow, up to the next def. More edges are redundant, because later defs 2869 // & kills are already serialized with true or antidependencies. To keep 2870 // the edge count down, we add a 'pinch point' node if there's more than 2871 // one use or more than one kill/def. 2872 2873 // We add dependencies in one bottom-up pass. 2874 2875 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2876 2877 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2878 // register. If not, we record the DEF/KILL in _reg_node, the 2879 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2880 // "pinch point", a new Node that's in the graph but not in the block. 2881 // We put edges from the prior and current DEF/KILLs to the pinch point. 2882 // We put the pinch point in _reg_node. If there's already a pinch point 2883 // we merely add an edge from the current DEF/KILL to the pinch point. 2884 2885 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2886 // put an edge from the pinch point to the USE. 2887 2888 // To be expedient, the _reg_node array is pre-allocated for the whole 2889 // compilation. _reg_node is lazily initialized; it either contains a null, 2890 // or a valid def/kill/pinch-point, or a leftover node from some prior 2891 // block. Leftover node from some prior block is treated like a null (no 2892 // prior def, so no anti-dependence needed). Valid def is distinguished by 2893 // it being in the current block. 2894 bool fat_proj_seen = false; 2895 uint last_safept = _bb_end-1; 2896 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr; 2897 Node* last_safept_node = end_node; 2898 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2899 Node *n = b->get_node(i); 2900 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2901 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 2902 // Fat-proj kills a slew of registers 2903 // This can add edges to 'n' and obscure whether or not it was a def, 2904 // hence the is_def flag. 2905 fat_proj_seen = true; 2906 RegMaskIterator rmi(n->out_RegMask()); 2907 while (rmi.has_next()) { 2908 OptoReg::Name kill = rmi.next(); 2909 anti_do_def(b, n, kill, is_def); 2910 } 2911 } else { 2912 // Get DEF'd registers the normal way 2913 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2914 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2915 } 2916 2917 // Kill projections on a branch should appear to occur on the 2918 // branch, not afterwards, so grab the masks from the projections 2919 // and process them. 2920 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) { 2921 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2922 Node* use = n->fast_out(i); 2923 if (use->is_Proj()) { 2924 RegMaskIterator rmi(use->out_RegMask()); 2925 while (rmi.has_next()) { 2926 OptoReg::Name kill = rmi.next(); 2927 anti_do_def(b, n, kill, false); 2928 } 2929 } 2930 } 2931 } 2932 2933 // Check each register used by this instruction for a following DEF/KILL 2934 // that must occur afterward and requires an anti-dependence edge. 2935 for( uint j=0; j<n->req(); j++ ) { 2936 Node *def = n->in(j); 2937 if( def ) { 2938 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2939 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2940 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2941 } 2942 } 2943 // Do not allow defs of new derived values to float above GC 2944 // points unless the base is definitely available at the GC point. 2945 2946 Node *m = b->get_node(i); 2947 2948 // Add precedence edge from following safepoint to use of derived pointer 2949 if( last_safept_node != end_node && 2950 m != last_safept_node) { 2951 for (uint k = 1; k < m->req(); k++) { 2952 const Type *t = m->in(k)->bottom_type(); 2953 if( t->isa_oop_ptr() && 2954 t->is_ptr()->offset() != 0 ) { 2955 last_safept_node->add_prec( m ); 2956 break; 2957 } 2958 } 2959 } 2960 2961 if( n->jvms() ) { // Precedence edge from derived to safept 2962 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2963 if( b->get_node(last_safept) != last_safept_node ) { 2964 last_safept = b->find_node(last_safept_node); 2965 } 2966 for( uint j=last_safept; j > i; j-- ) { 2967 Node *mach = b->get_node(j); 2968 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2969 mach->add_prec( n ); 2970 } 2971 last_safept = i; 2972 last_safept_node = m; 2973 } 2974 } 2975 2976 if (fat_proj_seen) { 2977 // Garbage collect pinch nodes that were not consumed. 2978 // They are usually created by a fat kill MachProj for a call. 2979 garbage_collect_pinch_nodes(); 2980 } 2981 } 2982 2983 // Garbage collect pinch nodes for reuse by other blocks. 2984 // 2985 // The block scheduler's insertion of anti-dependence 2986 // edges creates many pinch nodes when the block contains 2987 // 2 or more Calls. A pinch node is used to prevent a 2988 // combinatorial explosion of edges. If a set of kills for a 2989 // register is anti-dependent on a set of uses (or defs), rather 2990 // than adding an edge in the graph between each pair of kill 2991 // and use (or def), a pinch is inserted between them: 2992 // 2993 // use1 use2 use3 2994 // \ | / 2995 // \ | / 2996 // pinch 2997 // / | \ 2998 // / | \ 2999 // kill1 kill2 kill3 3000 // 3001 // One pinch node is created per register killed when 3002 // the second call is encountered during a backwards pass 3003 // over the block. Most of these pinch nodes are never 3004 // wired into the graph because the register is never 3005 // used or def'ed in the block. 3006 // 3007 void Scheduling::garbage_collect_pinch_nodes() { 3008 #ifndef PRODUCT 3009 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 3010 #endif 3011 int trace_cnt = 0; 3012 for (uint k = 0; k < _reg_node.max(); k++) { 3013 Node* pinch = _reg_node[k]; 3014 if ((pinch != nullptr) && pinch->Opcode() == Op_Node && 3015 // no predecence input edges 3016 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) { 3017 cleanup_pinch(pinch); 3018 _pinch_free_list.push(pinch); 3019 _reg_node.map(k, nullptr); 3020 #ifndef PRODUCT 3021 if (_cfg->C->trace_opto_output()) { 3022 trace_cnt++; 3023 if (trace_cnt > 40) { 3024 tty->print("\n"); 3025 trace_cnt = 0; 3026 } 3027 tty->print(" %d", pinch->_idx); 3028 } 3029 #endif 3030 } 3031 } 3032 #ifndef PRODUCT 3033 if (_cfg->C->trace_opto_output()) tty->print("\n"); 3034 #endif 3035 } 3036 3037 // Clean up a pinch node for reuse. 3038 void Scheduling::cleanup_pinch( Node *pinch ) { 3039 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 3040 3041 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 3042 Node* use = pinch->last_out(i); 3043 uint uses_found = 0; 3044 for (uint j = use->req(); j < use->len(); j++) { 3045 if (use->in(j) == pinch) { 3046 use->rm_prec(j); 3047 uses_found++; 3048 } 3049 } 3050 assert(uses_found > 0, "must be a precedence edge"); 3051 i -= uses_found; // we deleted 1 or more copies of this edge 3052 } 3053 // May have a later_def entry 3054 pinch->set_req(0, nullptr); 3055 } 3056 3057 #ifndef PRODUCT 3058 3059 void Scheduling::dump_available() const { 3060 tty->print("#Availist "); 3061 for (uint i = 0; i < _available.size(); i++) 3062 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 3063 tty->cr(); 3064 } 3065 3066 // Print Scheduling Statistics 3067 void Scheduling::print_statistics() { 3068 // Print the size added by nops for bundling 3069 tty->print("Nops added %d bytes to total of %d bytes", 3070 _total_nop_size, _total_method_size); 3071 if (_total_method_size > 0) 3072 tty->print(", for %.2f%%", 3073 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 3074 tty->print("\n"); 3075 3076 uint total_instructions = 0, total_bundles = 0; 3077 3078 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 3079 uint bundle_count = _total_instructions_per_bundle[i]; 3080 total_instructions += bundle_count * i; 3081 total_bundles += bundle_count; 3082 } 3083 3084 if (total_bundles > 0) 3085 tty->print("Average ILP (excluding nops) is %.2f\n", 3086 ((double)total_instructions) / ((double)total_bundles)); 3087 } 3088 #endif 3089 3090 //-----------------------init_scratch_buffer_blob------------------------------ 3091 // Construct a temporary BufferBlob and cache it for this compile. 3092 void PhaseOutput::init_scratch_buffer_blob(int const_size) { 3093 // If there is already a scratch buffer blob allocated and the 3094 // constant section is big enough, use it. Otherwise free the 3095 // current and allocate a new one. 3096 BufferBlob* blob = scratch_buffer_blob(); 3097 if ((blob != nullptr) && (const_size <= _scratch_const_size)) { 3098 // Use the current blob. 3099 } else { 3100 if (blob != nullptr) { 3101 BufferBlob::free(blob); 3102 } 3103 3104 ResourceMark rm; 3105 _scratch_const_size = const_size; 3106 int size = C2Compiler::initial_code_buffer_size(const_size); 3107 blob = BufferBlob::create("Compile::scratch_buffer", size); 3108 // Record the buffer blob for next time. 3109 set_scratch_buffer_blob(blob); 3110 // Have we run out of code space? 3111 if (scratch_buffer_blob() == nullptr) { 3112 // Let CompilerBroker disable further compilations. 3113 C->record_failure("Not enough space for scratch buffer in CodeCache"); 3114 return; 3115 } 3116 } 3117 3118 // Initialize the relocation buffers 3119 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 3120 set_scratch_locs_memory(locs_buf); 3121 } 3122 3123 3124 //-----------------------scratch_emit_size------------------------------------- 3125 // Helper function that computes size by emitting code 3126 uint PhaseOutput::scratch_emit_size(const Node* n) { 3127 // Start scratch_emit_size section. 3128 set_in_scratch_emit_size(true); 3129 3130 // Emit into a trash buffer and count bytes emitted. 3131 // This is a pretty expensive way to compute a size, 3132 // but it works well enough if seldom used. 3133 // All common fixed-size instructions are given a size 3134 // method by the AD file. 3135 // Note that the scratch buffer blob and locs memory are 3136 // allocated at the beginning of the compile task, and 3137 // may be shared by several calls to scratch_emit_size. 3138 // The allocation of the scratch buffer blob is particularly 3139 // expensive, since it has to grab the code cache lock. 3140 BufferBlob* blob = this->scratch_buffer_blob(); 3141 assert(blob != nullptr, "Initialize BufferBlob at start"); 3142 assert(blob->size() > max_inst_size(), "sanity"); 3143 relocInfo* locs_buf = scratch_locs_memory(); 3144 address blob_begin = blob->content_begin(); 3145 address blob_end = (address)locs_buf; 3146 assert(blob->contains(blob_end), "sanity"); 3147 CodeBuffer buf(blob_begin, blob_end - blob_begin); 3148 buf.initialize_consts_size(_scratch_const_size); 3149 buf.initialize_stubs_size(MAX_stubs_size); 3150 assert(locs_buf != nullptr, "sanity"); 3151 int lsize = MAX_locs_size / 3; 3152 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 3153 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 3154 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 3155 // Mark as scratch buffer. 3156 buf.consts()->set_scratch_emit(); 3157 buf.insts()->set_scratch_emit(); 3158 buf.stubs()->set_scratch_emit(); 3159 3160 // Do the emission. 3161 3162 Label fakeL; // Fake label for branch instructions. 3163 Label* saveL = nullptr; 3164 uint save_bnum = 0; 3165 bool is_branch = n->is_MachBranch(); 3166 C2_MacroAssembler masm(&buf); 3167 masm.bind(fakeL); 3168 if (is_branch) { 3169 n->as_MachBranch()->save_label(&saveL, &save_bnum); 3170 n->as_MachBranch()->label_set(&fakeL, 0); 3171 } 3172 n->emit(&masm, C->regalloc()); 3173 3174 // Emitting into the scratch buffer should not fail 3175 assert(!C->failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", C->failure_reason()); 3176 3177 if (is_branch) // Restore label. 3178 n->as_MachBranch()->label_set(saveL, save_bnum); 3179 3180 // End scratch_emit_size section. 3181 set_in_scratch_emit_size(false); 3182 3183 return buf.insts_size(); 3184 } 3185 3186 void PhaseOutput::install() { 3187 if (C->should_install_code() && C->stub_function() != nullptr) { 3188 install_stub(C->stub_name()); 3189 } else { 3190 install_code(C->method(), 3191 C->entry_bci(), 3192 CompilerThread::current()->compiler(), 3193 C->has_unsafe_access(), 3194 SharedRuntime::is_wide_vector(C->max_vector_size())); 3195 } 3196 } 3197 3198 void PhaseOutput::install_code(ciMethod* target, 3199 int entry_bci, 3200 AbstractCompiler* compiler, 3201 bool has_unsafe_access, 3202 bool has_wide_vectors) { 3203 // Check if we want to skip execution of all compiled code. 3204 { 3205 #ifndef PRODUCT 3206 if (OptoNoExecute) { 3207 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed 3208 return; 3209 } 3210 #endif 3211 Compile::TracePhase tp(_t_registerMethod); 3212 3213 if (C->is_osr_compilation()) { 3214 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 3215 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 3216 } else { 3217 if (!target->is_static()) { 3218 // The UEP of an nmethod ensures that the VEP is padded. However, the padding of the UEP is placed 3219 // before the inline cache check, so we don't have to execute any nop instructions when dispatching 3220 // through the UEP, yet we can ensure that the VEP is aligned appropriately. 3221 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size - MacroAssembler::ic_check_size()); 3222 } 3223 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 3224 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 3225 } 3226 3227 C->env()->register_method(target, 3228 entry_bci, 3229 &_code_offsets, 3230 _orig_pc_slot_offset_in_bytes, 3231 code_buffer(), 3232 frame_size_in_words(), 3233 oop_map_set(), 3234 &_handler_table, 3235 inc_table(), 3236 compiler, 3237 C->has_clinit_barriers(), 3238 C->for_preload(), 3239 has_unsafe_access, 3240 SharedRuntime::is_wide_vector(C->max_vector_size()), 3241 C->has_monitors(), 3242 C->has_scoped_access(), 3243 0, 3244 C->should_install_code()); 3245 3246 if (C->log() != nullptr) { // Print code cache state into compiler log 3247 C->log()->code_cache_state(); 3248 } 3249 assert(!C->has_clinit_barriers() || C->for_preload(), "class init barriers should be only in preload code"); 3250 } 3251 } 3252 void PhaseOutput::install_stub(const char* stub_name) { 3253 // Entry point will be accessed using stub_entry_point(); 3254 if (code_buffer() == nullptr) { 3255 Matcher::soft_match_failure(); 3256 } else { 3257 if (PrintAssembly && (WizardMode || Verbose)) 3258 tty->print_cr("### Stub::%s", stub_name); 3259 3260 if (!C->failing()) { 3261 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs"); 3262 3263 // Make the NMethod 3264 // For now we mark the frame as never safe for profile stackwalking 3265 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 3266 code_buffer(), 3267 CodeOffsets::frame_never_safe, 3268 // _code_offsets.value(CodeOffsets::Frame_Complete), 3269 frame_size_in_words(), 3270 oop_map_set(), 3271 false, 3272 false); 3273 3274 if (rs == nullptr) { 3275 C->record_failure("CodeCache is full"); 3276 } else { 3277 assert(rs->is_runtime_stub(), "sanity check"); 3278 C->set_stub_entry_point(rs->entry_point()); 3279 BlobId blob_id = StubInfo::blob(C->stub_id()); 3280 AOTCodeCache::store_code_blob(*rs, AOTCodeEntry::C2Blob, blob_id); 3281 } 3282 } 3283 } 3284 } 3285 3286 // Support for bundling info 3287 Bundle* PhaseOutput::node_bundling(const Node *n) { 3288 assert(valid_bundle_info(n), "oob"); 3289 return &_node_bundling_base[n->_idx]; 3290 } 3291 3292 bool PhaseOutput::valid_bundle_info(const Node *n) { 3293 return (_node_bundling_limit > n->_idx); 3294 } 3295 3296 //------------------------------frame_size_in_words----------------------------- 3297 // frame_slots in units of words 3298 int PhaseOutput::frame_size_in_words() const { 3299 // shift is 0 in LP32 and 1 in LP64 3300 const int shift = (LogBytesPerWord - LogBytesPerInt); 3301 int words = _frame_slots >> shift; 3302 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 3303 return words; 3304 } 3305 3306 // To bang the stack of this compiled method we use the stack size 3307 // that the interpreter would need in case of a deoptimization. This 3308 // removes the need to bang the stack in the deoptimization blob which 3309 // in turn simplifies stack overflow handling. 3310 int PhaseOutput::bang_size_in_bytes() const { 3311 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size()); 3312 } 3313 3314 //------------------------------dump_asm--------------------------------------- 3315 // Dump formatted assembly 3316 #if defined(SUPPORT_OPTO_ASSEMBLY) 3317 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) { 3318 3319 int pc_digits = 3; // #chars required for pc 3320 int sb_chars = 3; // #chars for "start bundle" indicator 3321 int tab_size = 8; 3322 if (pcs != nullptr) { 3323 int max_pc = 0; 3324 for (uint i = 0; i < pc_limit; i++) { 3325 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc; 3326 } 3327 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc 3328 } 3329 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size; 3330 3331 bool cut_short = false; 3332 st->print_cr("#"); 3333 st->print("# "); C->tf()->dump_on(st); st->cr(); 3334 st->print_cr("#"); 3335 3336 // For all blocks 3337 int pc = 0x0; // Program counter 3338 char starts_bundle = ' '; 3339 C->regalloc()->dump_frame(); 3340 3341 Node *n = nullptr; 3342 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) { 3343 if (VMThread::should_terminate()) { 3344 cut_short = true; 3345 break; 3346 } 3347 Block* block = C->cfg()->get_block(i); 3348 if (block->is_connector() && !Verbose) { 3349 continue; 3350 } 3351 n = block->head(); 3352 if ((pcs != nullptr) && (n->_idx < pc_limit)) { 3353 pc = pcs[n->_idx]; 3354 st->print("%*.*x", pc_digits, pc_digits, pc); 3355 } 3356 st->fill_to(prefix_len); 3357 block->dump_head(C->cfg(), st); 3358 if (block->is_connector()) { 3359 st->fill_to(prefix_len); 3360 st->print_cr("# Empty connector block"); 3361 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 3362 st->fill_to(prefix_len); 3363 st->print_cr("# Block is sole successor of call"); 3364 } 3365 3366 // For all instructions 3367 for (uint j = 0; j < block->number_of_nodes(); j++) { 3368 if (VMThread::should_terminate()) { 3369 cut_short = true; 3370 break; 3371 } 3372 n = block->get_node(j); 3373 if (valid_bundle_info(n)) { 3374 Bundle* bundle = node_bundling(n); 3375 if (bundle->starts_bundle()) { 3376 starts_bundle = '+'; 3377 } 3378 } 3379 3380 if (WizardMode) { 3381 n->dump(); 3382 } 3383 3384 if( !n->is_Region() && // Dont print in the Assembly 3385 !n->is_Phi() && // a few noisely useless nodes 3386 !n->is_Proj() && 3387 !n->is_MachTemp() && 3388 !n->is_SafePointScalarObject() && 3389 !n->is_Catch() && // Would be nice to print exception table targets 3390 !n->is_MergeMem() && // Not very interesting 3391 !n->is_top() && // Debug info table constants 3392 !(n->is_Con() && !n->is_Mach())// Debug info table constants 3393 ) { 3394 if ((pcs != nullptr) && (n->_idx < pc_limit)) { 3395 pc = pcs[n->_idx]; 3396 st->print("%*.*x", pc_digits, pc_digits, pc); 3397 } else { 3398 st->fill_to(pc_digits); 3399 } 3400 st->print(" %c ", starts_bundle); 3401 starts_bundle = ' '; 3402 st->fill_to(prefix_len); 3403 n->format(C->regalloc(), st); 3404 st->cr(); 3405 } 3406 3407 // Dump the exception table as well 3408 if( n->is_Catch() && (Verbose || WizardMode) ) { 3409 // Print the exception table for this offset 3410 _handler_table.print_subtable_for(pc); 3411 } 3412 st->bol(); // Make sure we start on a new line 3413 } 3414 st->cr(); // one empty line between blocks 3415 } // End of per-block dump 3416 3417 if (cut_short) st->print_cr("*** disassembly is cut short ***"); 3418 } 3419 #endif 3420 3421 #ifndef PRODUCT 3422 void PhaseOutput::print_statistics() { 3423 Scheduling::print_statistics(); 3424 } 3425 #endif 3426 3427 int PhaseOutput::max_inst_size() { 3428 if (AOTCodeCache::maybe_dumping_code()) { 3429 // See the comment in output.hpp. 3430 return 16384; 3431 } else { 3432 return mainline_MAX_inst_size; 3433 } 3434 } 3435 3436 int PhaseOutput::max_inst_gcstub_size() { 3437 assert(mainline_MAX_inst_size <= max_inst_size(), "Sanity"); 3438 return mainline_MAX_inst_size; 3439 }