1 /* 2 * Copyright 2005-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_c1_LinearScan.cpp.incl" 27 28 29 #ifndef PRODUCT 30 31 static LinearScanStatistic _stat_before_alloc; 32 static LinearScanStatistic _stat_after_asign; 33 static LinearScanStatistic _stat_final; 34 35 static LinearScanTimers _total_timer; 36 37 // helper macro for short definition of timer 38 #define TIME_LINEAR_SCAN(timer_name) TraceTime _block_timer("", _total_timer.timer(LinearScanTimers::timer_name), TimeLinearScan || TimeEachLinearScan, Verbose); 39 40 // helper macro for short definition of trace-output inside code 41 #define TRACE_LINEAR_SCAN(level, code) \ 42 if (TraceLinearScanLevel >= level) { \ 43 code; \ 44 } 45 46 #else 47 48 #define TIME_LINEAR_SCAN(timer_name) 49 #define TRACE_LINEAR_SCAN(level, code) 50 51 #endif 52 53 // Map BasicType to spill size in 32-bit words, matching VMReg's notion of words 54 #ifdef _LP64 55 static int type2spill_size[T_CONFLICT+1]={ -1, 0, 0, 0, 1, 1, 1, 2, 1, 1, 1, 2, 2, 2, 0, 1, -1}; 56 #else 57 static int type2spill_size[T_CONFLICT+1]={ -1, 0, 0, 0, 1, 1, 1, 2, 1, 1, 1, 2, 1, 1, 0, 1, -1}; 58 #endif 59 60 61 // Implementation of LinearScan 62 63 LinearScan::LinearScan(IR* ir, LIRGenerator* gen, FrameMap* frame_map) 64 : _compilation(ir->compilation()) 65 , _ir(ir) 66 , _gen(gen) 67 , _frame_map(frame_map) 68 , _num_virtual_regs(gen->max_virtual_register_number()) 69 , _has_fpu_registers(false) 70 , _num_calls(-1) 71 , _max_spills(0) 72 , _unused_spill_slot(-1) 73 , _intervals(0) // initialized later with correct length 74 , _new_intervals_from_allocation(new IntervalList()) 75 , _sorted_intervals(NULL) 76 , _lir_ops(0) // initialized later with correct length 77 , _block_of_op(0) // initialized later with correct length 78 , _has_info(0) 79 , _has_call(0) 80 , _scope_value_cache(0) // initialized later with correct length 81 , _interval_in_loop(0, 0) // initialized later with correct length 82 , _cached_blocks(*ir->linear_scan_order()) 83 #ifdef X86 84 , _fpu_stack_allocator(NULL) 85 #endif 86 { 87 // note: to use more than on instance of LinearScan at a time this function call has to 88 // be moved somewhere outside of this constructor: 89 Interval::initialize(); 90 91 assert(this->ir() != NULL, "check if valid"); 92 assert(this->compilation() != NULL, "check if valid"); 93 assert(this->gen() != NULL, "check if valid"); 94 assert(this->frame_map() != NULL, "check if valid"); 95 } 96 97 98 // ********** functions for converting LIR-Operands to register numbers 99 // 100 // Emulate a flat register file comprising physical integer registers, 101 // physical floating-point registers and virtual registers, in that order. 102 // Virtual registers already have appropriate numbers, since V0 is 103 // the number of physical registers. 104 // Returns -1 for hi word if opr is a single word operand. 105 // 106 // Note: the inverse operation (calculating an operand for register numbers) 107 // is done in calc_operand_for_interval() 108 109 int LinearScan::reg_num(LIR_Opr opr) { 110 assert(opr->is_register(), "should not call this otherwise"); 111 112 if (opr->is_virtual_register()) { 113 assert(opr->vreg_number() >= nof_regs, "found a virtual register with a fixed-register number"); 114 return opr->vreg_number(); 115 } else if (opr->is_single_cpu()) { 116 return opr->cpu_regnr(); 117 } else if (opr->is_double_cpu()) { 118 return opr->cpu_regnrLo(); 119 #ifdef X86 120 } else if (opr->is_single_xmm()) { 121 return opr->fpu_regnr() + pd_first_xmm_reg; 122 } else if (opr->is_double_xmm()) { 123 return opr->fpu_regnrLo() + pd_first_xmm_reg; 124 #endif 125 } else if (opr->is_single_fpu()) { 126 return opr->fpu_regnr() + pd_first_fpu_reg; 127 } else if (opr->is_double_fpu()) { 128 return opr->fpu_regnrLo() + pd_first_fpu_reg; 129 } else { 130 ShouldNotReachHere(); 131 return -1; 132 } 133 } 134 135 int LinearScan::reg_numHi(LIR_Opr opr) { 136 assert(opr->is_register(), "should not call this otherwise"); 137 138 if (opr->is_virtual_register()) { 139 return -1; 140 } else if (opr->is_single_cpu()) { 141 return -1; 142 } else if (opr->is_double_cpu()) { 143 return opr->cpu_regnrHi(); 144 #ifdef X86 145 } else if (opr->is_single_xmm()) { 146 return -1; 147 } else if (opr->is_double_xmm()) { 148 return -1; 149 #endif 150 } else if (opr->is_single_fpu()) { 151 return -1; 152 } else if (opr->is_double_fpu()) { 153 return opr->fpu_regnrHi() + pd_first_fpu_reg; 154 } else { 155 ShouldNotReachHere(); 156 return -1; 157 } 158 } 159 160 161 // ********** functions for classification of intervals 162 163 bool LinearScan::is_precolored_interval(const Interval* i) { 164 return i->reg_num() < LinearScan::nof_regs; 165 } 166 167 bool LinearScan::is_virtual_interval(const Interval* i) { 168 return i->reg_num() >= LIR_OprDesc::vreg_base; 169 } 170 171 bool LinearScan::is_precolored_cpu_interval(const Interval* i) { 172 return i->reg_num() < LinearScan::nof_cpu_regs; 173 } 174 175 bool LinearScan::is_virtual_cpu_interval(const Interval* i) { 176 return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() != T_FLOAT && i->type() != T_DOUBLE); 177 } 178 179 bool LinearScan::is_precolored_fpu_interval(const Interval* i) { 180 return i->reg_num() >= LinearScan::nof_cpu_regs && i->reg_num() < LinearScan::nof_regs; 181 } 182 183 bool LinearScan::is_virtual_fpu_interval(const Interval* i) { 184 return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() == T_FLOAT || i->type() == T_DOUBLE); 185 } 186 187 bool LinearScan::is_in_fpu_register(const Interval* i) { 188 // fixed intervals not needed for FPU stack allocation 189 return i->reg_num() >= nof_regs && pd_first_fpu_reg <= i->assigned_reg() && i->assigned_reg() <= pd_last_fpu_reg; 190 } 191 192 bool LinearScan::is_oop_interval(const Interval* i) { 193 // fixed intervals never contain oops 194 return i->reg_num() >= nof_regs && i->type() == T_OBJECT; 195 } 196 197 198 // ********** General helper functions 199 200 // compute next unused stack index that can be used for spilling 201 int LinearScan::allocate_spill_slot(bool double_word) { 202 int spill_slot; 203 if (double_word) { 204 if ((_max_spills & 1) == 1) { 205 // alignment of double-word values 206 // the hole because of the alignment is filled with the next single-word value 207 assert(_unused_spill_slot == -1, "wasting a spill slot"); 208 _unused_spill_slot = _max_spills; 209 _max_spills++; 210 } 211 spill_slot = _max_spills; 212 _max_spills += 2; 213 214 } else if (_unused_spill_slot != -1) { 215 // re-use hole that was the result of a previous double-word alignment 216 spill_slot = _unused_spill_slot; 217 _unused_spill_slot = -1; 218 219 } else { 220 spill_slot = _max_spills; 221 _max_spills++; 222 } 223 224 int result = spill_slot + LinearScan::nof_regs + frame_map()->argcount(); 225 226 // the class OopMapValue uses only 11 bits for storing the name of the 227 // oop location. So a stack slot bigger than 2^11 leads to an overflow 228 // that is not reported in product builds. Prevent this by checking the 229 // spill slot here (altough this value and the later used location name 230 // are slightly different) 231 if (result > 2000) { 232 bailout("too many stack slots used"); 233 } 234 235 return result; 236 } 237 238 void LinearScan::assign_spill_slot(Interval* it) { 239 // assign the canonical spill slot of the parent (if a part of the interval 240 // is already spilled) or allocate a new spill slot 241 if (it->canonical_spill_slot() >= 0) { 242 it->assign_reg(it->canonical_spill_slot()); 243 } else { 244 int spill = allocate_spill_slot(type2spill_size[it->type()] == 2); 245 it->set_canonical_spill_slot(spill); 246 it->assign_reg(spill); 247 } 248 } 249 250 void LinearScan::propagate_spill_slots() { 251 if (!frame_map()->finalize_frame(max_spills())) { 252 bailout("frame too large"); 253 } 254 } 255 256 // create a new interval with a predefined reg_num 257 // (only used for parent intervals that are created during the building phase) 258 Interval* LinearScan::create_interval(int reg_num) { 259 assert(_intervals.at(reg_num) == NULL, "overwriting exisiting interval"); 260 261 Interval* interval = new Interval(reg_num); 262 _intervals.at_put(reg_num, interval); 263 264 // assign register number for precolored intervals 265 if (reg_num < LIR_OprDesc::vreg_base) { 266 interval->assign_reg(reg_num); 267 } 268 return interval; 269 } 270 271 // assign a new reg_num to the interval and append it to the list of intervals 272 // (only used for child intervals that are created during register allocation) 273 void LinearScan::append_interval(Interval* it) { 274 it->set_reg_num(_intervals.length()); 275 _intervals.append(it); 276 _new_intervals_from_allocation->append(it); 277 } 278 279 // copy the vreg-flags if an interval is split 280 void LinearScan::copy_register_flags(Interval* from, Interval* to) { 281 if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::byte_reg)) { 282 gen()->set_vreg_flag(to->reg_num(), LIRGenerator::byte_reg); 283 } 284 if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::callee_saved)) { 285 gen()->set_vreg_flag(to->reg_num(), LIRGenerator::callee_saved); 286 } 287 288 // Note: do not copy the must_start_in_memory flag because it is not necessary for child 289 // intervals (only the very beginning of the interval must be in memory) 290 } 291 292 293 // ********** spill move optimization 294 // eliminate moves from register to stack if stack slot is known to be correct 295 296 // called during building of intervals 297 void LinearScan::change_spill_definition_pos(Interval* interval, int def_pos) { 298 assert(interval->is_split_parent(), "can only be called for split parents"); 299 300 switch (interval->spill_state()) { 301 case noDefinitionFound: 302 assert(interval->spill_definition_pos() == -1, "must no be set before"); 303 interval->set_spill_definition_pos(def_pos); 304 interval->set_spill_state(oneDefinitionFound); 305 break; 306 307 case oneDefinitionFound: 308 assert(def_pos <= interval->spill_definition_pos(), "positions are processed in reverse order when intervals are created"); 309 if (def_pos < interval->spill_definition_pos() - 2) { 310 // second definition found, so no spill optimization possible for this interval 311 interval->set_spill_state(noOptimization); 312 } else { 313 // two consecutive definitions (because of two-operand LIR form) 314 assert(block_of_op_with_id(def_pos) == block_of_op_with_id(interval->spill_definition_pos()), "block must be equal"); 315 } 316 break; 317 318 case noOptimization: 319 // nothing to do 320 break; 321 322 default: 323 assert(false, "other states not allowed at this time"); 324 } 325 } 326 327 // called during register allocation 328 void LinearScan::change_spill_state(Interval* interval, int spill_pos) { 329 switch (interval->spill_state()) { 330 case oneDefinitionFound: { 331 int def_loop_depth = block_of_op_with_id(interval->spill_definition_pos())->loop_depth(); 332 int spill_loop_depth = block_of_op_with_id(spill_pos)->loop_depth(); 333 334 if (def_loop_depth < spill_loop_depth) { 335 // the loop depth of the spilling position is higher then the loop depth 336 // at the definition of the interval -> move write to memory out of loop 337 // by storing at definitin of the interval 338 interval->set_spill_state(storeAtDefinition); 339 } else { 340 // the interval is currently spilled only once, so for now there is no 341 // reason to store the interval at the definition 342 interval->set_spill_state(oneMoveInserted); 343 } 344 break; 345 } 346 347 case oneMoveInserted: { 348 // the interval is spilled more then once, so it is better to store it to 349 // memory at the definition 350 interval->set_spill_state(storeAtDefinition); 351 break; 352 } 353 354 case storeAtDefinition: 355 case startInMemory: 356 case noOptimization: 357 case noDefinitionFound: 358 // nothing to do 359 break; 360 361 default: 362 assert(false, "other states not allowed at this time"); 363 } 364 } 365 366 367 bool LinearScan::must_store_at_definition(const Interval* i) { 368 return i->is_split_parent() && i->spill_state() == storeAtDefinition; 369 } 370 371 // called once before asignment of register numbers 372 void LinearScan::eliminate_spill_moves() { 373 TIME_LINEAR_SCAN(timer_eliminate_spill_moves); 374 TRACE_LINEAR_SCAN(3, tty->print_cr("***** Eliminating unnecessary spill moves")); 375 376 // collect all intervals that must be stored after their definion. 377 // the list is sorted by Interval::spill_definition_pos 378 Interval* interval; 379 Interval* temp_list; 380 create_unhandled_lists(&interval, &temp_list, must_store_at_definition, NULL); 381 382 #ifdef ASSERT 383 Interval* prev = NULL; 384 Interval* temp = interval; 385 while (temp != Interval::end()) { 386 assert(temp->spill_definition_pos() > 0, "invalid spill definition pos"); 387 if (prev != NULL) { 388 assert(temp->from() >= prev->from(), "intervals not sorted"); 389 assert(temp->spill_definition_pos() >= prev->spill_definition_pos(), "when intervals are sorted by from, then they must also be sorted by spill_definition_pos"); 390 } 391 392 assert(temp->canonical_spill_slot() >= LinearScan::nof_regs, "interval has no spill slot assigned"); 393 assert(temp->spill_definition_pos() >= temp->from(), "invalid order"); 394 assert(temp->spill_definition_pos() <= temp->from() + 2, "only intervals defined once at their start-pos can be optimized"); 395 396 TRACE_LINEAR_SCAN(4, tty->print_cr("interval %d (from %d to %d) must be stored at %d", temp->reg_num(), temp->from(), temp->to(), temp->spill_definition_pos())); 397 398 temp = temp->next(); 399 } 400 #endif 401 402 LIR_InsertionBuffer insertion_buffer; 403 int num_blocks = block_count(); 404 for (int i = 0; i < num_blocks; i++) { 405 BlockBegin* block = block_at(i); 406 LIR_OpList* instructions = block->lir()->instructions_list(); 407 int num_inst = instructions->length(); 408 bool has_new = false; 409 410 // iterate all instructions of the block. skip the first because it is always a label 411 for (int j = 1; j < num_inst; j++) { 412 LIR_Op* op = instructions->at(j); 413 int op_id = op->id(); 414 415 if (op_id == -1) { 416 // remove move from register to stack if the stack slot is guaranteed to be correct. 417 // only moves that have been inserted by LinearScan can be removed. 418 assert(op->code() == lir_move, "only moves can have a op_id of -1"); 419 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); 420 assert(op->as_Op1()->result_opr()->is_virtual(), "LinearScan inserts only moves to virtual registers"); 421 422 LIR_Op1* op1 = (LIR_Op1*)op; 423 Interval* interval = interval_at(op1->result_opr()->vreg_number()); 424 425 if (interval->assigned_reg() >= LinearScan::nof_regs && interval->always_in_memory()) { 426 // move target is a stack slot that is always correct, so eliminate instruction 427 TRACE_LINEAR_SCAN(4, tty->print_cr("eliminating move from interval %d to %d", op1->in_opr()->vreg_number(), op1->result_opr()->vreg_number())); 428 instructions->at_put(j, NULL); // NULL-instructions are deleted by assign_reg_num 429 } 430 431 } else { 432 // insert move from register to stack just after the beginning of the interval 433 assert(interval == Interval::end() || interval->spill_definition_pos() >= op_id, "invalid order"); 434 assert(interval == Interval::end() || (interval->is_split_parent() && interval->spill_state() == storeAtDefinition), "invalid interval"); 435 436 while (interval != Interval::end() && interval->spill_definition_pos() == op_id) { 437 if (!has_new) { 438 // prepare insertion buffer (appended when all instructions of the block are processed) 439 insertion_buffer.init(block->lir()); 440 has_new = true; 441 } 442 443 LIR_Opr from_opr = operand_for_interval(interval); 444 LIR_Opr to_opr = canonical_spill_opr(interval); 445 assert(from_opr->is_fixed_cpu() || from_opr->is_fixed_fpu(), "from operand must be a register"); 446 assert(to_opr->is_stack(), "to operand must be a stack slot"); 447 448 insertion_buffer.move(j, from_opr, to_opr); 449 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting move after definition of interval %d to stack slot %d at op_id %d", interval->reg_num(), interval->canonical_spill_slot() - LinearScan::nof_regs, op_id)); 450 451 interval = interval->next(); 452 } 453 } 454 } // end of instruction iteration 455 456 if (has_new) { 457 block->lir()->append(&insertion_buffer); 458 } 459 } // end of block iteration 460 461 assert(interval == Interval::end(), "missed an interval"); 462 } 463 464 465 // ********** Phase 1: number all instructions in all blocks 466 // Compute depth-first and linear scan block orders, and number LIR_Op nodes for linear scan. 467 468 void LinearScan::number_instructions() { 469 { 470 // dummy-timer to measure the cost of the timer itself 471 // (this time is then subtracted from all other timers to get the real value) 472 TIME_LINEAR_SCAN(timer_do_nothing); 473 } 474 TIME_LINEAR_SCAN(timer_number_instructions); 475 476 // Assign IDs to LIR nodes and build a mapping, lir_ops, from ID to LIR_Op node. 477 int num_blocks = block_count(); 478 int num_instructions = 0; 479 int i; 480 for (i = 0; i < num_blocks; i++) { 481 num_instructions += block_at(i)->lir()->instructions_list()->length(); 482 } 483 484 // initialize with correct length 485 _lir_ops = LIR_OpArray(num_instructions); 486 _block_of_op = BlockBeginArray(num_instructions); 487 488 int op_id = 0; 489 int idx = 0; 490 491 for (i = 0; i < num_blocks; i++) { 492 BlockBegin* block = block_at(i); 493 block->set_first_lir_instruction_id(op_id); 494 LIR_OpList* instructions = block->lir()->instructions_list(); 495 496 int num_inst = instructions->length(); 497 for (int j = 0; j < num_inst; j++) { 498 LIR_Op* op = instructions->at(j); 499 op->set_id(op_id); 500 501 _lir_ops.at_put(idx, op); 502 _block_of_op.at_put(idx, block); 503 assert(lir_op_with_id(op_id) == op, "must match"); 504 505 idx++; 506 op_id += 2; // numbering of lir_ops by two 507 } 508 block->set_last_lir_instruction_id(op_id - 2); 509 } 510 assert(idx == num_instructions, "must match"); 511 assert(idx * 2 == op_id, "must match"); 512 513 _has_call = BitMap(num_instructions); _has_call.clear(); 514 _has_info = BitMap(num_instructions); _has_info.clear(); 515 } 516 517 518 // ********** Phase 2: compute local live sets separately for each block 519 // (sets live_gen and live_kill for each block) 520 521 void LinearScan::set_live_gen_kill(Value value, LIR_Op* op, BitMap& live_gen, BitMap& live_kill) { 522 LIR_Opr opr = value->operand(); 523 Constant* con = value->as_Constant(); 524 525 // check some asumptions about debug information 526 assert(!value->type()->is_illegal(), "if this local is used by the interpreter it shouldn't be of indeterminate type"); 527 assert(con == NULL || opr->is_virtual() || opr->is_constant() || opr->is_illegal(), "asumption: Constant instructions have only constant operands"); 528 assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands"); 529 530 if ((con == NULL || con->is_pinned()) && opr->is_register()) { 531 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 532 int reg = opr->vreg_number(); 533 if (!live_kill.at(reg)) { 534 live_gen.set_bit(reg); 535 TRACE_LINEAR_SCAN(4, tty->print_cr(" Setting live_gen for value %c%d, LIR op_id %d, register number %d", value->type()->tchar(), value->id(), op->id(), reg)); 536 } 537 } 538 } 539 540 541 void LinearScan::compute_local_live_sets() { 542 TIME_LINEAR_SCAN(timer_compute_local_live_sets); 543 544 int num_blocks = block_count(); 545 int live_size = live_set_size(); 546 bool local_has_fpu_registers = false; 547 int local_num_calls = 0; 548 LIR_OpVisitState visitor; 549 550 BitMap2D local_interval_in_loop = BitMap2D(_num_virtual_regs, num_loops()); 551 local_interval_in_loop.clear(); 552 553 // iterate all blocks 554 for (int i = 0; i < num_blocks; i++) { 555 BlockBegin* block = block_at(i); 556 557 BitMap live_gen(live_size); live_gen.clear(); 558 BitMap live_kill(live_size); live_kill.clear(); 559 560 if (block->is_set(BlockBegin::exception_entry_flag)) { 561 // Phi functions at the begin of an exception handler are 562 // implicitly defined (= killed) at the beginning of the block. 563 for_each_phi_fun(block, phi, 564 live_kill.set_bit(phi->operand()->vreg_number()) 565 ); 566 } 567 568 LIR_OpList* instructions = block->lir()->instructions_list(); 569 int num_inst = instructions->length(); 570 571 // iterate all instructions of the block. skip the first because it is always a label 572 assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label"); 573 for (int j = 1; j < num_inst; j++) { 574 LIR_Op* op = instructions->at(j); 575 576 // visit operation to collect all operands 577 visitor.visit(op); 578 579 if (visitor.has_call()) { 580 _has_call.set_bit(op->id() >> 1); 581 local_num_calls++; 582 } 583 if (visitor.info_count() > 0) { 584 _has_info.set_bit(op->id() >> 1); 585 } 586 587 // iterate input operands of instruction 588 int k, n, reg; 589 n = visitor.opr_count(LIR_OpVisitState::inputMode); 590 for (k = 0; k < n; k++) { 591 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k); 592 assert(opr->is_register(), "visitor should only return register operands"); 593 594 if (opr->is_virtual_register()) { 595 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 596 reg = opr->vreg_number(); 597 if (!live_kill.at(reg)) { 598 live_gen.set_bit(reg); 599 TRACE_LINEAR_SCAN(4, tty->print_cr(" Setting live_gen for register %d at instruction %d", reg, op->id())); 600 } 601 if (block->loop_index() >= 0) { 602 local_interval_in_loop.set_bit(reg, block->loop_index()); 603 } 604 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); 605 } 606 607 #ifdef ASSERT 608 // fixed intervals are never live at block boundaries, so 609 // they need not be processed in live sets. 610 // this is checked by these assertions to be sure about it. 611 // the entry block may have incoming values in registers, which is ok. 612 if (!opr->is_virtual_register() && block != ir()->start()) { 613 reg = reg_num(opr); 614 if (is_processed_reg_num(reg)) { 615 assert(live_kill.at(reg), "using fixed register that is not defined in this block"); 616 } 617 reg = reg_numHi(opr); 618 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 619 assert(live_kill.at(reg), "using fixed register that is not defined in this block"); 620 } 621 } 622 #endif 623 } 624 625 // Add uses of live locals from interpreter's point of view for proper debug information generation 626 n = visitor.info_count(); 627 for (k = 0; k < n; k++) { 628 CodeEmitInfo* info = visitor.info_at(k); 629 ValueStack* stack = info->stack(); 630 for_each_state_value(stack, value, 631 set_live_gen_kill(value, op, live_gen, live_kill) 632 ); 633 } 634 635 // iterate temp operands of instruction 636 n = visitor.opr_count(LIR_OpVisitState::tempMode); 637 for (k = 0; k < n; k++) { 638 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k); 639 assert(opr->is_register(), "visitor should only return register operands"); 640 641 if (opr->is_virtual_register()) { 642 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 643 reg = opr->vreg_number(); 644 live_kill.set_bit(reg); 645 if (block->loop_index() >= 0) { 646 local_interval_in_loop.set_bit(reg, block->loop_index()); 647 } 648 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); 649 } 650 651 #ifdef ASSERT 652 // fixed intervals are never live at block boundaries, so 653 // they need not be processed in live sets 654 // process them only in debug mode so that this can be checked 655 if (!opr->is_virtual_register()) { 656 reg = reg_num(opr); 657 if (is_processed_reg_num(reg)) { 658 live_kill.set_bit(reg_num(opr)); 659 } 660 reg = reg_numHi(opr); 661 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 662 live_kill.set_bit(reg); 663 } 664 } 665 #endif 666 } 667 668 // iterate output operands of instruction 669 n = visitor.opr_count(LIR_OpVisitState::outputMode); 670 for (k = 0; k < n; k++) { 671 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k); 672 assert(opr->is_register(), "visitor should only return register operands"); 673 674 if (opr->is_virtual_register()) { 675 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 676 reg = opr->vreg_number(); 677 live_kill.set_bit(reg); 678 if (block->loop_index() >= 0) { 679 local_interval_in_loop.set_bit(reg, block->loop_index()); 680 } 681 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); 682 } 683 684 #ifdef ASSERT 685 // fixed intervals are never live at block boundaries, so 686 // they need not be processed in live sets 687 // process them only in debug mode so that this can be checked 688 if (!opr->is_virtual_register()) { 689 reg = reg_num(opr); 690 if (is_processed_reg_num(reg)) { 691 live_kill.set_bit(reg_num(opr)); 692 } 693 reg = reg_numHi(opr); 694 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 695 live_kill.set_bit(reg); 696 } 697 } 698 #endif 699 } 700 } // end of instruction iteration 701 702 block->set_live_gen (live_gen); 703 block->set_live_kill(live_kill); 704 block->set_live_in (BitMap(live_size)); block->live_in().clear(); 705 block->set_live_out (BitMap(live_size)); block->live_out().clear(); 706 707 TRACE_LINEAR_SCAN(4, tty->print("live_gen B%d ", block->block_id()); print_bitmap(block->live_gen())); 708 TRACE_LINEAR_SCAN(4, tty->print("live_kill B%d ", block->block_id()); print_bitmap(block->live_kill())); 709 } // end of block iteration 710 711 // propagate local calculated information into LinearScan object 712 _has_fpu_registers = local_has_fpu_registers; 713 compilation()->set_has_fpu_code(local_has_fpu_registers); 714 715 _num_calls = local_num_calls; 716 _interval_in_loop = local_interval_in_loop; 717 } 718 719 720 // ********** Phase 3: perform a backward dataflow analysis to compute global live sets 721 // (sets live_in and live_out for each block) 722 723 void LinearScan::compute_global_live_sets() { 724 TIME_LINEAR_SCAN(timer_compute_global_live_sets); 725 726 int num_blocks = block_count(); 727 bool change_occurred; 728 bool change_occurred_in_block; 729 int iteration_count = 0; 730 BitMap live_out(live_set_size()); live_out.clear(); // scratch set for calculations 731 732 // Perform a backward dataflow analysis to compute live_out and live_in for each block. 733 // The loop is executed until a fixpoint is reached (no changes in an iteration) 734 // Exception handlers must be processed because not all live values are 735 // present in the state array, e.g. because of global value numbering 736 do { 737 change_occurred = false; 738 739 // iterate all blocks in reverse order 740 for (int i = num_blocks - 1; i >= 0; i--) { 741 BlockBegin* block = block_at(i); 742 743 change_occurred_in_block = false; 744 745 // live_out(block) is the union of live_in(sux), for successors sux of block 746 int n = block->number_of_sux(); 747 int e = block->number_of_exception_handlers(); 748 if (n + e > 0) { 749 // block has successors 750 if (n > 0) { 751 live_out.set_from(block->sux_at(0)->live_in()); 752 for (int j = 1; j < n; j++) { 753 live_out.set_union(block->sux_at(j)->live_in()); 754 } 755 } else { 756 live_out.clear(); 757 } 758 for (int j = 0; j < e; j++) { 759 live_out.set_union(block->exception_handler_at(j)->live_in()); 760 } 761 762 if (!block->live_out().is_same(live_out)) { 763 // A change occurred. Swap the old and new live out sets to avoid copying. 764 BitMap temp = block->live_out(); 765 block->set_live_out(live_out); 766 live_out = temp; 767 768 change_occurred = true; 769 change_occurred_in_block = true; 770 } 771 } 772 773 if (iteration_count == 0 || change_occurred_in_block) { 774 // live_in(block) is the union of live_gen(block) with (live_out(block) & !live_kill(block)) 775 // note: live_in has to be computed only in first iteration or if live_out has changed! 776 BitMap live_in = block->live_in(); 777 live_in.set_from(block->live_out()); 778 live_in.set_difference(block->live_kill()); 779 live_in.set_union(block->live_gen()); 780 } 781 782 #ifndef PRODUCT 783 if (TraceLinearScanLevel >= 4) { 784 char c = ' '; 785 if (iteration_count == 0 || change_occurred_in_block) { 786 c = '*'; 787 } 788 tty->print("(%d) live_in%c B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_in()); 789 tty->print("(%d) live_out%c B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_out()); 790 } 791 #endif 792 } 793 iteration_count++; 794 795 if (change_occurred && iteration_count > 50) { 796 BAILOUT("too many iterations in compute_global_live_sets"); 797 } 798 } while (change_occurred); 799 800 801 #ifdef ASSERT 802 // check that fixed intervals are not live at block boundaries 803 // (live set must be empty at fixed intervals) 804 for (int i = 0; i < num_blocks; i++) { 805 BlockBegin* block = block_at(i); 806 for (int j = 0; j < LIR_OprDesc::vreg_base; j++) { 807 assert(block->live_in().at(j) == false, "live_in set of fixed register must be empty"); 808 assert(block->live_out().at(j) == false, "live_out set of fixed register must be empty"); 809 assert(block->live_gen().at(j) == false, "live_gen set of fixed register must be empty"); 810 } 811 } 812 #endif 813 814 // check that the live_in set of the first block is empty 815 BitMap live_in_args(ir()->start()->live_in().size()); 816 live_in_args.clear(); 817 if (!ir()->start()->live_in().is_same(live_in_args)) { 818 #ifdef ASSERT 819 tty->print_cr("Error: live_in set of first block must be empty (when this fails, virtual registers are used before they are defined)"); 820 tty->print_cr("affected registers:"); 821 print_bitmap(ir()->start()->live_in()); 822 823 // print some additional information to simplify debugging 824 for (unsigned int i = 0; i < ir()->start()->live_in().size(); i++) { 825 if (ir()->start()->live_in().at(i)) { 826 Instruction* instr = gen()->instruction_for_vreg(i); 827 tty->print_cr("* vreg %d (HIR instruction %c%d)", i, instr == NULL ? ' ' : instr->type()->tchar(), instr == NULL ? 0 : instr->id()); 828 829 for (int j = 0; j < num_blocks; j++) { 830 BlockBegin* block = block_at(j); 831 if (block->live_gen().at(i)) { 832 tty->print_cr(" used in block B%d", block->block_id()); 833 } 834 if (block->live_kill().at(i)) { 835 tty->print_cr(" defined in block B%d", block->block_id()); 836 } 837 } 838 } 839 } 840 841 #endif 842 // when this fails, virtual registers are used before they are defined. 843 assert(false, "live_in set of first block must be empty"); 844 // bailout of if this occurs in product mode. 845 bailout("live_in set of first block not empty"); 846 } 847 } 848 849 850 // ********** Phase 4: build intervals 851 // (fills the list _intervals) 852 853 void LinearScan::add_use(Value value, int from, int to, IntervalUseKind use_kind) { 854 assert(!value->type()->is_illegal(), "if this value is used by the interpreter it shouldn't be of indeterminate type"); 855 LIR_Opr opr = value->operand(); 856 Constant* con = value->as_Constant(); 857 858 if ((con == NULL || con->is_pinned()) && opr->is_register()) { 859 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 860 add_use(opr, from, to, use_kind); 861 } 862 } 863 864 865 void LinearScan::add_def(LIR_Opr opr, int def_pos, IntervalUseKind use_kind) { 866 TRACE_LINEAR_SCAN(2, tty->print(" def "); opr->print(tty); tty->print_cr(" def_pos %d (%d)", def_pos, use_kind)); 867 assert(opr->is_register(), "should not be called otherwise"); 868 869 if (opr->is_virtual_register()) { 870 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 871 add_def(opr->vreg_number(), def_pos, use_kind, opr->type_register()); 872 873 } else { 874 int reg = reg_num(opr); 875 if (is_processed_reg_num(reg)) { 876 add_def(reg, def_pos, use_kind, opr->type_register()); 877 } 878 reg = reg_numHi(opr); 879 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 880 add_def(reg, def_pos, use_kind, opr->type_register()); 881 } 882 } 883 } 884 885 void LinearScan::add_use(LIR_Opr opr, int from, int to, IntervalUseKind use_kind) { 886 TRACE_LINEAR_SCAN(2, tty->print(" use "); opr->print(tty); tty->print_cr(" from %d to %d (%d)", from, to, use_kind)); 887 assert(opr->is_register(), "should not be called otherwise"); 888 889 if (opr->is_virtual_register()) { 890 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 891 add_use(opr->vreg_number(), from, to, use_kind, opr->type_register()); 892 893 } else { 894 int reg = reg_num(opr); 895 if (is_processed_reg_num(reg)) { 896 add_use(reg, from, to, use_kind, opr->type_register()); 897 } 898 reg = reg_numHi(opr); 899 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 900 add_use(reg, from, to, use_kind, opr->type_register()); 901 } 902 } 903 } 904 905 void LinearScan::add_temp(LIR_Opr opr, int temp_pos, IntervalUseKind use_kind) { 906 TRACE_LINEAR_SCAN(2, tty->print(" temp "); opr->print(tty); tty->print_cr(" temp_pos %d (%d)", temp_pos, use_kind)); 907 assert(opr->is_register(), "should not be called otherwise"); 908 909 if (opr->is_virtual_register()) { 910 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 911 add_temp(opr->vreg_number(), temp_pos, use_kind, opr->type_register()); 912 913 } else { 914 int reg = reg_num(opr); 915 if (is_processed_reg_num(reg)) { 916 add_temp(reg, temp_pos, use_kind, opr->type_register()); 917 } 918 reg = reg_numHi(opr); 919 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 920 add_temp(reg, temp_pos, use_kind, opr->type_register()); 921 } 922 } 923 } 924 925 926 void LinearScan::add_def(int reg_num, int def_pos, IntervalUseKind use_kind, BasicType type) { 927 Interval* interval = interval_at(reg_num); 928 if (interval != NULL) { 929 assert(interval->reg_num() == reg_num, "wrong interval"); 930 931 if (type != T_ILLEGAL) { 932 interval->set_type(type); 933 } 934 935 Range* r = interval->first(); 936 if (r->from() <= def_pos) { 937 // Update the starting point (when a range is first created for a use, its 938 // start is the beginning of the current block until a def is encountered.) 939 r->set_from(def_pos); 940 interval->add_use_pos(def_pos, use_kind); 941 942 } else { 943 // Dead value - make vacuous interval 944 // also add use_kind for dead intervals 945 interval->add_range(def_pos, def_pos + 1); 946 interval->add_use_pos(def_pos, use_kind); 947 TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: def of reg %d at %d occurs without use", reg_num, def_pos)); 948 } 949 950 } else { 951 // Dead value - make vacuous interval 952 // also add use_kind for dead intervals 953 interval = create_interval(reg_num); 954 if (type != T_ILLEGAL) { 955 interval->set_type(type); 956 } 957 958 interval->add_range(def_pos, def_pos + 1); 959 interval->add_use_pos(def_pos, use_kind); 960 TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: dead value %d at %d in live intervals", reg_num, def_pos)); 961 } 962 963 change_spill_definition_pos(interval, def_pos); 964 if (use_kind == noUse && interval->spill_state() <= startInMemory) { 965 // detection of method-parameters and roundfp-results 966 // TODO: move this directly to position where use-kind is computed 967 interval->set_spill_state(startInMemory); 968 } 969 } 970 971 void LinearScan::add_use(int reg_num, int from, int to, IntervalUseKind use_kind, BasicType type) { 972 Interval* interval = interval_at(reg_num); 973 if (interval == NULL) { 974 interval = create_interval(reg_num); 975 } 976 assert(interval->reg_num() == reg_num, "wrong interval"); 977 978 if (type != T_ILLEGAL) { 979 interval->set_type(type); 980 } 981 982 interval->add_range(from, to); 983 interval->add_use_pos(to, use_kind); 984 } 985 986 void LinearScan::add_temp(int reg_num, int temp_pos, IntervalUseKind use_kind, BasicType type) { 987 Interval* interval = interval_at(reg_num); 988 if (interval == NULL) { 989 interval = create_interval(reg_num); 990 } 991 assert(interval->reg_num() == reg_num, "wrong interval"); 992 993 if (type != T_ILLEGAL) { 994 interval->set_type(type); 995 } 996 997 interval->add_range(temp_pos, temp_pos + 1); 998 interval->add_use_pos(temp_pos, use_kind); 999 } 1000 1001 1002 // the results of this functions are used for optimizing spilling and reloading 1003 // if the functions return shouldHaveRegister and the interval is spilled, 1004 // it is not reloaded to a register. 1005 IntervalUseKind LinearScan::use_kind_of_output_operand(LIR_Op* op, LIR_Opr opr) { 1006 if (op->code() == lir_move) { 1007 assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1"); 1008 LIR_Op1* move = (LIR_Op1*)op; 1009 LIR_Opr res = move->result_opr(); 1010 bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory); 1011 1012 if (result_in_memory) { 1013 // Begin of an interval with must_start_in_memory set. 1014 // This interval will always get a stack slot first, so return noUse. 1015 return noUse; 1016 1017 } else if (move->in_opr()->is_stack()) { 1018 // method argument (condition must be equal to handle_method_arguments) 1019 return noUse; 1020 1021 } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) { 1022 // Move from register to register 1023 if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) { 1024 // special handling of phi-function moves inside osr-entry blocks 1025 // input operand must have a register instead of output operand (leads to better register allocation) 1026 return shouldHaveRegister; 1027 } 1028 } 1029 } 1030 1031 if (opr->is_virtual() && 1032 gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::must_start_in_memory)) { 1033 // result is a stack-slot, so prevent immediate reloading 1034 return noUse; 1035 } 1036 1037 // all other operands require a register 1038 return mustHaveRegister; 1039 } 1040 1041 IntervalUseKind LinearScan::use_kind_of_input_operand(LIR_Op* op, LIR_Opr opr) { 1042 if (op->code() == lir_move) { 1043 assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1"); 1044 LIR_Op1* move = (LIR_Op1*)op; 1045 LIR_Opr res = move->result_opr(); 1046 bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory); 1047 1048 if (result_in_memory) { 1049 // Move to an interval with must_start_in_memory set. 1050 // To avoid moves from stack to stack (not allowed) force the input operand to a register 1051 return mustHaveRegister; 1052 1053 } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) { 1054 // Move from register to register 1055 if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) { 1056 // special handling of phi-function moves inside osr-entry blocks 1057 // input operand must have a register instead of output operand (leads to better register allocation) 1058 return mustHaveRegister; 1059 } 1060 1061 // The input operand is not forced to a register (moves from stack to register are allowed), 1062 // but it is faster if the input operand is in a register 1063 return shouldHaveRegister; 1064 } 1065 } 1066 1067 1068 #ifdef X86 1069 if (op->code() == lir_cmove) { 1070 // conditional moves can handle stack operands 1071 assert(op->result_opr()->is_register(), "result must always be in a register"); 1072 return shouldHaveRegister; 1073 } 1074 1075 // optimizations for second input operand of arithmehtic operations on Intel 1076 // this operand is allowed to be on the stack in some cases 1077 BasicType opr_type = opr->type_register(); 1078 if (opr_type == T_FLOAT || opr_type == T_DOUBLE) { 1079 if ((UseSSE == 1 && opr_type == T_FLOAT) || UseSSE >= 2) { 1080 // SSE float instruction (T_DOUBLE only supported with SSE2) 1081 switch (op->code()) { 1082 case lir_cmp: 1083 case lir_add: 1084 case lir_sub: 1085 case lir_mul: 1086 case lir_div: 1087 { 1088 assert(op->as_Op2() != NULL, "must be LIR_Op2"); 1089 LIR_Op2* op2 = (LIR_Op2*)op; 1090 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { 1091 assert((op2->result_opr()->is_register() || op->code() == lir_cmp) && op2->in_opr1()->is_register(), "cannot mark second operand as stack if others are not in register"); 1092 return shouldHaveRegister; 1093 } 1094 } 1095 } 1096 } else { 1097 // FPU stack float instruction 1098 switch (op->code()) { 1099 case lir_add: 1100 case lir_sub: 1101 case lir_mul: 1102 case lir_div: 1103 { 1104 assert(op->as_Op2() != NULL, "must be LIR_Op2"); 1105 LIR_Op2* op2 = (LIR_Op2*)op; 1106 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { 1107 assert((op2->result_opr()->is_register() || op->code() == lir_cmp) && op2->in_opr1()->is_register(), "cannot mark second operand as stack if others are not in register"); 1108 return shouldHaveRegister; 1109 } 1110 } 1111 } 1112 } 1113 1114 } else if (opr_type != T_LONG) { 1115 // integer instruction (note: long operands must always be in register) 1116 switch (op->code()) { 1117 case lir_cmp: 1118 case lir_add: 1119 case lir_sub: 1120 case lir_logic_and: 1121 case lir_logic_or: 1122 case lir_logic_xor: 1123 { 1124 assert(op->as_Op2() != NULL, "must be LIR_Op2"); 1125 LIR_Op2* op2 = (LIR_Op2*)op; 1126 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { 1127 assert((op2->result_opr()->is_register() || op->code() == lir_cmp) && op2->in_opr1()->is_register(), "cannot mark second operand as stack if others are not in register"); 1128 return shouldHaveRegister; 1129 } 1130 } 1131 } 1132 } 1133 #endif // X86 1134 1135 // all other operands require a register 1136 return mustHaveRegister; 1137 } 1138 1139 1140 void LinearScan::handle_method_arguments(LIR_Op* op) { 1141 // special handling for method arguments (moves from stack to virtual register): 1142 // the interval gets no register assigned, but the stack slot. 1143 // it is split before the first use by the register allocator. 1144 1145 if (op->code() == lir_move) { 1146 assert(op->as_Op1() != NULL, "must be LIR_Op1"); 1147 LIR_Op1* move = (LIR_Op1*)op; 1148 1149 if (move->in_opr()->is_stack()) { 1150 #ifdef ASSERT 1151 int arg_size = compilation()->method()->arg_size(); 1152 LIR_Opr o = move->in_opr(); 1153 if (o->is_single_stack()) { 1154 assert(o->single_stack_ix() >= 0 && o->single_stack_ix() < arg_size, "out of range"); 1155 } else if (o->is_double_stack()) { 1156 assert(o->double_stack_ix() >= 0 && o->double_stack_ix() < arg_size, "out of range"); 1157 } else { 1158 ShouldNotReachHere(); 1159 } 1160 1161 assert(move->id() > 0, "invalid id"); 1162 assert(block_of_op_with_id(move->id())->number_of_preds() == 0, "move from stack must be in first block"); 1163 assert(move->result_opr()->is_virtual(), "result of move must be a virtual register"); 1164 1165 TRACE_LINEAR_SCAN(4, tty->print_cr("found move from stack slot %d to vreg %d", o->is_single_stack() ? o->single_stack_ix() : o->double_stack_ix(), reg_num(move->result_opr()))); 1166 #endif 1167 1168 Interval* interval = interval_at(reg_num(move->result_opr())); 1169 1170 int stack_slot = LinearScan::nof_regs + (move->in_opr()->is_single_stack() ? move->in_opr()->single_stack_ix() : move->in_opr()->double_stack_ix()); 1171 interval->set_canonical_spill_slot(stack_slot); 1172 interval->assign_reg(stack_slot); 1173 } 1174 } 1175 } 1176 1177 void LinearScan::handle_doubleword_moves(LIR_Op* op) { 1178 // special handling for doubleword move from memory to register: 1179 // in this case the registers of the input address and the result 1180 // registers must not overlap -> add a temp range for the input registers 1181 if (op->code() == lir_move) { 1182 assert(op->as_Op1() != NULL, "must be LIR_Op1"); 1183 LIR_Op1* move = (LIR_Op1*)op; 1184 1185 if (move->result_opr()->is_double_cpu() && move->in_opr()->is_pointer()) { 1186 LIR_Address* address = move->in_opr()->as_address_ptr(); 1187 if (address != NULL) { 1188 if (address->base()->is_valid()) { 1189 add_temp(address->base(), op->id(), noUse); 1190 } 1191 if (address->index()->is_valid()) { 1192 add_temp(address->index(), op->id(), noUse); 1193 } 1194 } 1195 } 1196 } 1197 } 1198 1199 void LinearScan::add_register_hints(LIR_Op* op) { 1200 switch (op->code()) { 1201 case lir_move: // fall through 1202 case lir_convert: { 1203 assert(op->as_Op1() != NULL, "lir_move, lir_convert must be LIR_Op1"); 1204 LIR_Op1* move = (LIR_Op1*)op; 1205 1206 LIR_Opr move_from = move->in_opr(); 1207 LIR_Opr move_to = move->result_opr(); 1208 1209 if (move_to->is_register() && move_from->is_register()) { 1210 Interval* from = interval_at(reg_num(move_from)); 1211 Interval* to = interval_at(reg_num(move_to)); 1212 if (from != NULL && to != NULL) { 1213 to->set_register_hint(from); 1214 TRACE_LINEAR_SCAN(4, tty->print_cr("operation at op_id %d: added hint from interval %d to %d", move->id(), from->reg_num(), to->reg_num())); 1215 } 1216 } 1217 break; 1218 } 1219 case lir_cmove: { 1220 assert(op->as_Op2() != NULL, "lir_cmove must be LIR_Op2"); 1221 LIR_Op2* cmove = (LIR_Op2*)op; 1222 1223 LIR_Opr move_from = cmove->in_opr1(); 1224 LIR_Opr move_to = cmove->result_opr(); 1225 1226 if (move_to->is_register() && move_from->is_register()) { 1227 Interval* from = interval_at(reg_num(move_from)); 1228 Interval* to = interval_at(reg_num(move_to)); 1229 if (from != NULL && to != NULL) { 1230 to->set_register_hint(from); 1231 TRACE_LINEAR_SCAN(4, tty->print_cr("operation at op_id %d: added hint from interval %d to %d", cmove->id(), from->reg_num(), to->reg_num())); 1232 } 1233 } 1234 break; 1235 } 1236 } 1237 } 1238 1239 1240 void LinearScan::build_intervals() { 1241 TIME_LINEAR_SCAN(timer_build_intervals); 1242 1243 // initialize interval list with expected number of intervals 1244 // (32 is added to have some space for split children without having to resize the list) 1245 _intervals = IntervalList(num_virtual_regs() + 32); 1246 // initialize all slots that are used by build_intervals 1247 _intervals.at_put_grow(num_virtual_regs() - 1, NULL, NULL); 1248 1249 // create a list with all caller-save registers (cpu, fpu, xmm) 1250 // when an instruction is a call, a temp range is created for all these registers 1251 int num_caller_save_registers = 0; 1252 int caller_save_registers[LinearScan::nof_regs]; 1253 1254 int i; 1255 for (i = 0; i < FrameMap::nof_caller_save_cpu_regs; i++) { 1256 LIR_Opr opr = FrameMap::caller_save_cpu_reg_at(i); 1257 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands"); 1258 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register"); 1259 caller_save_registers[num_caller_save_registers++] = reg_num(opr); 1260 } 1261 1262 // temp ranges for fpu registers are only created when the method has 1263 // virtual fpu operands. Otherwise no allocation for fpu registers is 1264 // perfomed and so the temp ranges would be useless 1265 if (has_fpu_registers()) { 1266 #ifdef X86 1267 if (UseSSE < 2) { 1268 #endif 1269 for (i = 0; i < FrameMap::nof_caller_save_fpu_regs; i++) { 1270 LIR_Opr opr = FrameMap::caller_save_fpu_reg_at(i); 1271 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands"); 1272 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register"); 1273 caller_save_registers[num_caller_save_registers++] = reg_num(opr); 1274 } 1275 #ifdef X86 1276 } 1277 if (UseSSE > 0) { 1278 for (i = 0; i < FrameMap::nof_caller_save_xmm_regs; i++) { 1279 LIR_Opr opr = FrameMap::caller_save_xmm_reg_at(i); 1280 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands"); 1281 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register"); 1282 caller_save_registers[num_caller_save_registers++] = reg_num(opr); 1283 } 1284 } 1285 #endif 1286 } 1287 assert(num_caller_save_registers <= LinearScan::nof_regs, "out of bounds"); 1288 1289 1290 LIR_OpVisitState visitor; 1291 1292 // iterate all blocks in reverse order 1293 for (i = block_count() - 1; i >= 0; i--) { 1294 BlockBegin* block = block_at(i); 1295 LIR_OpList* instructions = block->lir()->instructions_list(); 1296 int block_from = block->first_lir_instruction_id(); 1297 int block_to = block->last_lir_instruction_id(); 1298 1299 assert(block_from == instructions->at(0)->id(), "must be"); 1300 assert(block_to == instructions->at(instructions->length() - 1)->id(), "must be"); 1301 1302 // Update intervals for registers live at the end of this block; 1303 BitMap live = block->live_out(); 1304 int size = (int)live.size(); 1305 for (int number = (int)live.get_next_one_offset(0, size); number < size; number = (int)live.get_next_one_offset(number + 1, size)) { 1306 assert(live.at(number), "should not stop here otherwise"); 1307 assert(number >= LIR_OprDesc::vreg_base, "fixed intervals must not be live on block bounds"); 1308 TRACE_LINEAR_SCAN(2, tty->print_cr("live in %d to %d", number, block_to + 2)); 1309 1310 add_use(number, block_from, block_to + 2, noUse, T_ILLEGAL); 1311 1312 // add special use positions for loop-end blocks when the 1313 // interval is used anywhere inside this loop. It's possible 1314 // that the block was part of a non-natural loop, so it might 1315 // have an invalid loop index. 1316 if (block->is_set(BlockBegin::linear_scan_loop_end_flag) && 1317 block->loop_index() != -1 && 1318 is_interval_in_loop(number, block->loop_index())) { 1319 interval_at(number)->add_use_pos(block_to + 1, loopEndMarker); 1320 } 1321 } 1322 1323 // iterate all instructions of the block in reverse order. 1324 // skip the first instruction because it is always a label 1325 // definitions of intervals are processed before uses 1326 assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label"); 1327 for (int j = instructions->length() - 1; j >= 1; j--) { 1328 LIR_Op* op = instructions->at(j); 1329 int op_id = op->id(); 1330 1331 // visit operation to collect all operands 1332 visitor.visit(op); 1333 1334 // add a temp range for each register if operation destroys caller-save registers 1335 if (visitor.has_call()) { 1336 for (int k = 0; k < num_caller_save_registers; k++) { 1337 add_temp(caller_save_registers[k], op_id, noUse, T_ILLEGAL); 1338 } 1339 TRACE_LINEAR_SCAN(4, tty->print_cr("operation destroys all caller-save registers")); 1340 } 1341 1342 // Add any platform dependent temps 1343 pd_add_temps(op); 1344 1345 // visit definitions (output and temp operands) 1346 int k, n; 1347 n = visitor.opr_count(LIR_OpVisitState::outputMode); 1348 for (k = 0; k < n; k++) { 1349 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k); 1350 assert(opr->is_register(), "visitor should only return register operands"); 1351 add_def(opr, op_id, use_kind_of_output_operand(op, opr)); 1352 } 1353 1354 n = visitor.opr_count(LIR_OpVisitState::tempMode); 1355 for (k = 0; k < n; k++) { 1356 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k); 1357 assert(opr->is_register(), "visitor should only return register operands"); 1358 add_temp(opr, op_id, mustHaveRegister); 1359 } 1360 1361 // visit uses (input operands) 1362 n = visitor.opr_count(LIR_OpVisitState::inputMode); 1363 for (k = 0; k < n; k++) { 1364 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k); 1365 assert(opr->is_register(), "visitor should only return register operands"); 1366 add_use(opr, block_from, op_id, use_kind_of_input_operand(op, opr)); 1367 } 1368 1369 // Add uses of live locals from interpreter's point of view for proper 1370 // debug information generation 1371 // Treat these operands as temp values (if the life range is extended 1372 // to a call site, the value would be in a register at the call otherwise) 1373 n = visitor.info_count(); 1374 for (k = 0; k < n; k++) { 1375 CodeEmitInfo* info = visitor.info_at(k); 1376 ValueStack* stack = info->stack(); 1377 for_each_state_value(stack, value, 1378 add_use(value, block_from, op_id + 1, noUse); 1379 ); 1380 } 1381 1382 // special steps for some instructions (especially moves) 1383 handle_method_arguments(op); 1384 handle_doubleword_moves(op); 1385 add_register_hints(op); 1386 1387 } // end of instruction iteration 1388 } // end of block iteration 1389 1390 1391 // add the range [0, 1[ to all fixed intervals 1392 // -> the register allocator need not handle unhandled fixed intervals 1393 for (int n = 0; n < LinearScan::nof_regs; n++) { 1394 Interval* interval = interval_at(n); 1395 if (interval != NULL) { 1396 interval->add_range(0, 1); 1397 } 1398 } 1399 } 1400 1401 1402 // ********** Phase 5: actual register allocation 1403 1404 int LinearScan::interval_cmp(Interval** a, Interval** b) { 1405 if (*a != NULL) { 1406 if (*b != NULL) { 1407 return (*a)->from() - (*b)->from(); 1408 } else { 1409 return -1; 1410 } 1411 } else { 1412 if (*b != NULL) { 1413 return 1; 1414 } else { 1415 return 0; 1416 } 1417 } 1418 } 1419 1420 #ifndef PRODUCT 1421 bool LinearScan::is_sorted(IntervalArray* intervals) { 1422 int from = -1; 1423 int i, j; 1424 for (i = 0; i < intervals->length(); i ++) { 1425 Interval* it = intervals->at(i); 1426 if (it != NULL) { 1427 if (from > it->from()) { 1428 assert(false, ""); 1429 return false; 1430 } 1431 from = it->from(); 1432 } 1433 } 1434 1435 // check in both directions if sorted list and unsorted list contain same intervals 1436 for (i = 0; i < interval_count(); i++) { 1437 if (interval_at(i) != NULL) { 1438 int num_found = 0; 1439 for (j = 0; j < intervals->length(); j++) { 1440 if (interval_at(i) == intervals->at(j)) { 1441 num_found++; 1442 } 1443 } 1444 assert(num_found == 1, "lists do not contain same intervals"); 1445 } 1446 } 1447 for (j = 0; j < intervals->length(); j++) { 1448 int num_found = 0; 1449 for (i = 0; i < interval_count(); i++) { 1450 if (interval_at(i) == intervals->at(j)) { 1451 num_found++; 1452 } 1453 } 1454 assert(num_found == 1, "lists do not contain same intervals"); 1455 } 1456 1457 return true; 1458 } 1459 #endif 1460 1461 void LinearScan::add_to_list(Interval** first, Interval** prev, Interval* interval) { 1462 if (*prev != NULL) { 1463 (*prev)->set_next(interval); 1464 } else { 1465 *first = interval; 1466 } 1467 *prev = interval; 1468 } 1469 1470 void LinearScan::create_unhandled_lists(Interval** list1, Interval** list2, bool (is_list1)(const Interval* i), bool (is_list2)(const Interval* i)) { 1471 assert(is_sorted(_sorted_intervals), "interval list is not sorted"); 1472 1473 *list1 = *list2 = Interval::end(); 1474 1475 Interval* list1_prev = NULL; 1476 Interval* list2_prev = NULL; 1477 Interval* v; 1478 1479 const int n = _sorted_intervals->length(); 1480 for (int i = 0; i < n; i++) { 1481 v = _sorted_intervals->at(i); 1482 if (v == NULL) continue; 1483 1484 if (is_list1(v)) { 1485 add_to_list(list1, &list1_prev, v); 1486 } else if (is_list2 == NULL || is_list2(v)) { 1487 add_to_list(list2, &list2_prev, v); 1488 } 1489 } 1490 1491 if (list1_prev != NULL) list1_prev->set_next(Interval::end()); 1492 if (list2_prev != NULL) list2_prev->set_next(Interval::end()); 1493 1494 assert(list1_prev == NULL || list1_prev->next() == Interval::end(), "linear list ends not with sentinel"); 1495 assert(list2_prev == NULL || list2_prev->next() == Interval::end(), "linear list ends not with sentinel"); 1496 } 1497 1498 1499 void LinearScan::sort_intervals_before_allocation() { 1500 TIME_LINEAR_SCAN(timer_sort_intervals_before); 1501 1502 IntervalList* unsorted_list = &_intervals; 1503 int unsorted_len = unsorted_list->length(); 1504 int sorted_len = 0; 1505 int unsorted_idx; 1506 int sorted_idx = 0; 1507 int sorted_from_max = -1; 1508 1509 // calc number of items for sorted list (sorted list must not contain NULL values) 1510 for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) { 1511 if (unsorted_list->at(unsorted_idx) != NULL) { 1512 sorted_len++; 1513 } 1514 } 1515 IntervalArray* sorted_list = new IntervalArray(sorted_len); 1516 1517 // special sorting algorithm: the original interval-list is almost sorted, 1518 // only some intervals are swapped. So this is much faster than a complete QuickSort 1519 for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) { 1520 Interval* cur_interval = unsorted_list->at(unsorted_idx); 1521 1522 if (cur_interval != NULL) { 1523 int cur_from = cur_interval->from(); 1524 1525 if (sorted_from_max <= cur_from) { 1526 sorted_list->at_put(sorted_idx++, cur_interval); 1527 sorted_from_max = cur_interval->from(); 1528 } else { 1529 // the asumption that the intervals are already sorted failed, 1530 // so this interval must be sorted in manually 1531 int j; 1532 for (j = sorted_idx - 1; j >= 0 && cur_from < sorted_list->at(j)->from(); j--) { 1533 sorted_list->at_put(j + 1, sorted_list->at(j)); 1534 } 1535 sorted_list->at_put(j + 1, cur_interval); 1536 sorted_idx++; 1537 } 1538 } 1539 } 1540 _sorted_intervals = sorted_list; 1541 } 1542 1543 void LinearScan::sort_intervals_after_allocation() { 1544 TIME_LINEAR_SCAN(timer_sort_intervals_after); 1545 1546 IntervalArray* old_list = _sorted_intervals; 1547 IntervalList* new_list = _new_intervals_from_allocation; 1548 int old_len = old_list->length(); 1549 int new_len = new_list->length(); 1550 1551 if (new_len == 0) { 1552 // no intervals have been added during allocation, so sorted list is already up to date 1553 return; 1554 } 1555 1556 // conventional sort-algorithm for new intervals 1557 new_list->sort(interval_cmp); 1558 1559 // merge old and new list (both already sorted) into one combined list 1560 IntervalArray* combined_list = new IntervalArray(old_len + new_len); 1561 int old_idx = 0; 1562 int new_idx = 0; 1563 1564 while (old_idx + new_idx < old_len + new_len) { 1565 if (new_idx >= new_len || (old_idx < old_len && old_list->at(old_idx)->from() <= new_list->at(new_idx)->from())) { 1566 combined_list->at_put(old_idx + new_idx, old_list->at(old_idx)); 1567 old_idx++; 1568 } else { 1569 combined_list->at_put(old_idx + new_idx, new_list->at(new_idx)); 1570 new_idx++; 1571 } 1572 } 1573 1574 _sorted_intervals = combined_list; 1575 } 1576 1577 1578 void LinearScan::allocate_registers() { 1579 TIME_LINEAR_SCAN(timer_allocate_registers); 1580 1581 Interval* precolored_cpu_intervals, *not_precolored_cpu_intervals; 1582 Interval* precolored_fpu_intervals, *not_precolored_fpu_intervals; 1583 1584 create_unhandled_lists(&precolored_cpu_intervals, ¬_precolored_cpu_intervals, is_precolored_cpu_interval, is_virtual_cpu_interval); 1585 if (has_fpu_registers()) { 1586 create_unhandled_lists(&precolored_fpu_intervals, ¬_precolored_fpu_intervals, is_precolored_fpu_interval, is_virtual_fpu_interval); 1587 #ifdef ASSERT 1588 } else { 1589 // fpu register allocation is omitted because no virtual fpu registers are present 1590 // just check this again... 1591 create_unhandled_lists(&precolored_fpu_intervals, ¬_precolored_fpu_intervals, is_precolored_fpu_interval, is_virtual_fpu_interval); 1592 assert(not_precolored_fpu_intervals == Interval::end(), "missed an uncolored fpu interval"); 1593 #endif 1594 } 1595 1596 // allocate cpu registers 1597 LinearScanWalker cpu_lsw(this, precolored_cpu_intervals, not_precolored_cpu_intervals); 1598 cpu_lsw.walk(); 1599 cpu_lsw.finish_allocation(); 1600 1601 if (has_fpu_registers()) { 1602 // allocate fpu registers 1603 LinearScanWalker fpu_lsw(this, precolored_fpu_intervals, not_precolored_fpu_intervals); 1604 fpu_lsw.walk(); 1605 fpu_lsw.finish_allocation(); 1606 } 1607 } 1608 1609 1610 // ********** Phase 6: resolve data flow 1611 // (insert moves at edges between blocks if intervals have been split) 1612 1613 // wrapper for Interval::split_child_at_op_id that performs a bailout in product mode 1614 // instead of returning NULL 1615 Interval* LinearScan::split_child_at_op_id(Interval* interval, int op_id, LIR_OpVisitState::OprMode mode) { 1616 Interval* result = interval->split_child_at_op_id(op_id, mode); 1617 if (result != NULL) { 1618 return result; 1619 } 1620 1621 assert(false, "must find an interval, but do a clean bailout in product mode"); 1622 result = new Interval(LIR_OprDesc::vreg_base); 1623 result->assign_reg(0); 1624 result->set_type(T_INT); 1625 BAILOUT_("LinearScan: interval is NULL", result); 1626 } 1627 1628 1629 Interval* LinearScan::interval_at_block_begin(BlockBegin* block, int reg_num) { 1630 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds"); 1631 assert(interval_at(reg_num) != NULL, "no interval found"); 1632 1633 return split_child_at_op_id(interval_at(reg_num), block->first_lir_instruction_id(), LIR_OpVisitState::outputMode); 1634 } 1635 1636 Interval* LinearScan::interval_at_block_end(BlockBegin* block, int reg_num) { 1637 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds"); 1638 assert(interval_at(reg_num) != NULL, "no interval found"); 1639 1640 return split_child_at_op_id(interval_at(reg_num), block->last_lir_instruction_id() + 1, LIR_OpVisitState::outputMode); 1641 } 1642 1643 Interval* LinearScan::interval_at_op_id(int reg_num, int op_id) { 1644 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds"); 1645 assert(interval_at(reg_num) != NULL, "no interval found"); 1646 1647 return split_child_at_op_id(interval_at(reg_num), op_id, LIR_OpVisitState::inputMode); 1648 } 1649 1650 1651 void LinearScan::resolve_collect_mappings(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) { 1652 DEBUG_ONLY(move_resolver.check_empty()); 1653 1654 const int num_regs = num_virtual_regs(); 1655 const int size = live_set_size(); 1656 const BitMap live_at_edge = to_block->live_in(); 1657 1658 // visit all registers where the live_at_edge bit is set 1659 for (int r = (int)live_at_edge.get_next_one_offset(0, size); r < size; r = (int)live_at_edge.get_next_one_offset(r + 1, size)) { 1660 assert(r < num_regs, "live information set for not exisiting interval"); 1661 assert(from_block->live_out().at(r) && to_block->live_in().at(r), "interval not live at this edge"); 1662 1663 Interval* from_interval = interval_at_block_end(from_block, r); 1664 Interval* to_interval = interval_at_block_begin(to_block, r); 1665 1666 if (from_interval != to_interval && (from_interval->assigned_reg() != to_interval->assigned_reg() || from_interval->assigned_regHi() != to_interval->assigned_regHi())) { 1667 // need to insert move instruction 1668 move_resolver.add_mapping(from_interval, to_interval); 1669 } 1670 } 1671 } 1672 1673 1674 void LinearScan::resolve_find_insert_pos(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) { 1675 if (from_block->number_of_sux() <= 1) { 1676 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at end of from_block B%d", from_block->block_id())); 1677 1678 LIR_OpList* instructions = from_block->lir()->instructions_list(); 1679 LIR_OpBranch* branch = instructions->last()->as_OpBranch(); 1680 if (branch != NULL) { 1681 // insert moves before branch 1682 assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump"); 1683 move_resolver.set_insert_position(from_block->lir(), instructions->length() - 2); 1684 } else { 1685 move_resolver.set_insert_position(from_block->lir(), instructions->length() - 1); 1686 } 1687 1688 } else { 1689 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at beginning of to_block B%d", to_block->block_id())); 1690 #ifdef ASSERT 1691 assert(from_block->lir()->instructions_list()->at(0)->as_OpLabel() != NULL, "block does not start with a label"); 1692 1693 // because the number of predecessor edges matches the number of 1694 // successor edges, blocks which are reached by switch statements 1695 // may have be more than one predecessor but it will be guaranteed 1696 // that all predecessors will be the same. 1697 for (int i = 0; i < to_block->number_of_preds(); i++) { 1698 assert(from_block == to_block->pred_at(i), "all critical edges must be broken"); 1699 } 1700 #endif 1701 1702 move_resolver.set_insert_position(to_block->lir(), 0); 1703 } 1704 } 1705 1706 1707 // insert necessary moves (spilling or reloading) at edges between blocks if interval has been split 1708 void LinearScan::resolve_data_flow() { 1709 TIME_LINEAR_SCAN(timer_resolve_data_flow); 1710 1711 int num_blocks = block_count(); 1712 MoveResolver move_resolver(this); 1713 BitMap block_completed(num_blocks); block_completed.clear(); 1714 BitMap already_resolved(num_blocks); already_resolved.clear(); 1715 1716 int i; 1717 for (i = 0; i < num_blocks; i++) { 1718 BlockBegin* block = block_at(i); 1719 1720 // check if block has only one predecessor and only one successor 1721 if (block->number_of_preds() == 1 && block->number_of_sux() == 1 && block->number_of_exception_handlers() == 0) { 1722 LIR_OpList* instructions = block->lir()->instructions_list(); 1723 assert(instructions->at(0)->code() == lir_label, "block must start with label"); 1724 assert(instructions->last()->code() == lir_branch, "block with successors must end with branch"); 1725 assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block with successor must end with unconditional branch"); 1726 1727 // check if block is empty (only label and branch) 1728 if (instructions->length() == 2) { 1729 BlockBegin* pred = block->pred_at(0); 1730 BlockBegin* sux = block->sux_at(0); 1731 1732 // prevent optimization of two consecutive blocks 1733 if (!block_completed.at(pred->linear_scan_number()) && !block_completed.at(sux->linear_scan_number())) { 1734 TRACE_LINEAR_SCAN(3, tty->print_cr("**** optimizing empty block B%d (pred: B%d, sux: B%d)", block->block_id(), pred->block_id(), sux->block_id())); 1735 block_completed.set_bit(block->linear_scan_number()); 1736 1737 // directly resolve between pred and sux (without looking at the empty block between) 1738 resolve_collect_mappings(pred, sux, move_resolver); 1739 if (move_resolver.has_mappings()) { 1740 move_resolver.set_insert_position(block->lir(), 0); 1741 move_resolver.resolve_and_append_moves(); 1742 } 1743 } 1744 } 1745 } 1746 } 1747 1748 1749 for (i = 0; i < num_blocks; i++) { 1750 if (!block_completed.at(i)) { 1751 BlockBegin* from_block = block_at(i); 1752 already_resolved.set_from(block_completed); 1753 1754 int num_sux = from_block->number_of_sux(); 1755 for (int s = 0; s < num_sux; s++) { 1756 BlockBegin* to_block = from_block->sux_at(s); 1757 1758 // check for duplicate edges between the same blocks (can happen with switch blocks) 1759 if (!already_resolved.at(to_block->linear_scan_number())) { 1760 TRACE_LINEAR_SCAN(3, tty->print_cr("**** processing edge between B%d and B%d", from_block->block_id(), to_block->block_id())); 1761 already_resolved.set_bit(to_block->linear_scan_number()); 1762 1763 // collect all intervals that have been split between from_block and to_block 1764 resolve_collect_mappings(from_block, to_block, move_resolver); 1765 if (move_resolver.has_mappings()) { 1766 resolve_find_insert_pos(from_block, to_block, move_resolver); 1767 move_resolver.resolve_and_append_moves(); 1768 } 1769 } 1770 } 1771 } 1772 } 1773 } 1774 1775 1776 void LinearScan::resolve_exception_entry(BlockBegin* block, int reg_num, MoveResolver &move_resolver) { 1777 if (interval_at(reg_num) == NULL) { 1778 // if a phi function is never used, no interval is created -> ignore this 1779 return; 1780 } 1781 1782 Interval* interval = interval_at_block_begin(block, reg_num); 1783 int reg = interval->assigned_reg(); 1784 int regHi = interval->assigned_regHi(); 1785 1786 if ((reg < nof_regs && interval->always_in_memory()) || 1787 (use_fpu_stack_allocation() && reg >= pd_first_fpu_reg && reg <= pd_last_fpu_reg)) { 1788 // the interval is split to get a short range that is located on the stack 1789 // in the following two cases: 1790 // * the interval started in memory (e.g. method parameter), but is currently in a register 1791 // this is an optimization for exception handling that reduces the number of moves that 1792 // are necessary for resolving the states when an exception uses this exception handler 1793 // * the interval would be on the fpu stack at the begin of the exception handler 1794 // this is not allowed because of the complicated fpu stack handling on Intel 1795 1796 // range that will be spilled to memory 1797 int from_op_id = block->first_lir_instruction_id(); 1798 int to_op_id = from_op_id + 1; // short live range of length 1 1799 assert(interval->from() <= from_op_id && interval->to() >= to_op_id, 1800 "no split allowed between exception entry and first instruction"); 1801 1802 if (interval->from() != from_op_id) { 1803 // the part before from_op_id is unchanged 1804 interval = interval->split(from_op_id); 1805 interval->assign_reg(reg, regHi); 1806 append_interval(interval); 1807 } 1808 assert(interval->from() == from_op_id, "must be true now"); 1809 1810 Interval* spilled_part = interval; 1811 if (interval->to() != to_op_id) { 1812 // the part after to_op_id is unchanged 1813 spilled_part = interval->split_from_start(to_op_id); 1814 append_interval(spilled_part); 1815 move_resolver.add_mapping(spilled_part, interval); 1816 } 1817 assign_spill_slot(spilled_part); 1818 1819 assert(spilled_part->from() == from_op_id && spilled_part->to() == to_op_id, "just checking"); 1820 } 1821 } 1822 1823 void LinearScan::resolve_exception_entry(BlockBegin* block, MoveResolver &move_resolver) { 1824 assert(block->is_set(BlockBegin::exception_entry_flag), "should not call otherwise"); 1825 DEBUG_ONLY(move_resolver.check_empty()); 1826 1827 // visit all registers where the live_in bit is set 1828 int size = live_set_size(); 1829 for (int r = (int)block->live_in().get_next_one_offset(0, size); r < size; r = (int)block->live_in().get_next_one_offset(r + 1, size)) { 1830 resolve_exception_entry(block, r, move_resolver); 1831 } 1832 1833 // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately 1834 for_each_phi_fun(block, phi, 1835 resolve_exception_entry(block, phi->operand()->vreg_number(), move_resolver) 1836 ); 1837 1838 if (move_resolver.has_mappings()) { 1839 // insert moves after first instruction 1840 move_resolver.set_insert_position(block->lir(), 1); 1841 move_resolver.resolve_and_append_moves(); 1842 } 1843 } 1844 1845 1846 void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, int reg_num, Phi* phi, MoveResolver &move_resolver) { 1847 if (interval_at(reg_num) == NULL) { 1848 // if a phi function is never used, no interval is created -> ignore this 1849 return; 1850 } 1851 1852 // the computation of to_interval is equal to resolve_collect_mappings, 1853 // but from_interval is more complicated because of phi functions 1854 BlockBegin* to_block = handler->entry_block(); 1855 Interval* to_interval = interval_at_block_begin(to_block, reg_num); 1856 1857 if (phi != NULL) { 1858 // phi function of the exception entry block 1859 // no moves are created for this phi function in the LIR_Generator, so the 1860 // interval at the throwing instruction must be searched using the operands 1861 // of the phi function 1862 Value from_value = phi->operand_at(handler->phi_operand()); 1863 1864 // with phi functions it can happen that the same from_value is used in 1865 // multiple mappings, so notify move-resolver that this is allowed 1866 move_resolver.set_multiple_reads_allowed(); 1867 1868 Constant* con = from_value->as_Constant(); 1869 if (con != NULL && !con->is_pinned()) { 1870 // unpinned constants may have no register, so add mapping from constant to interval 1871 move_resolver.add_mapping(LIR_OprFact::value_type(con->type()), to_interval); 1872 } else { 1873 // search split child at the throwing op_id 1874 Interval* from_interval = interval_at_op_id(from_value->operand()->vreg_number(), throwing_op_id); 1875 move_resolver.add_mapping(from_interval, to_interval); 1876 } 1877 1878 } else { 1879 // no phi function, so use reg_num also for from_interval 1880 // search split child at the throwing op_id 1881 Interval* from_interval = interval_at_op_id(reg_num, throwing_op_id); 1882 if (from_interval != to_interval) { 1883 // optimization to reduce number of moves: when to_interval is on stack and 1884 // the stack slot is known to be always correct, then no move is necessary 1885 if (!from_interval->always_in_memory() || from_interval->canonical_spill_slot() != to_interval->assigned_reg()) { 1886 move_resolver.add_mapping(from_interval, to_interval); 1887 } 1888 } 1889 } 1890 } 1891 1892 void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, MoveResolver &move_resolver) { 1893 TRACE_LINEAR_SCAN(4, tty->print_cr("resolving exception handler B%d: throwing_op_id=%d", handler->entry_block()->block_id(), throwing_op_id)); 1894 1895 DEBUG_ONLY(move_resolver.check_empty()); 1896 assert(handler->lir_op_id() == -1, "already processed this xhandler"); 1897 DEBUG_ONLY(handler->set_lir_op_id(throwing_op_id)); 1898 assert(handler->entry_code() == NULL, "code already present"); 1899 1900 // visit all registers where the live_in bit is set 1901 BlockBegin* block = handler->entry_block(); 1902 int size = live_set_size(); 1903 for (int r = (int)block->live_in().get_next_one_offset(0, size); r < size; r = (int)block->live_in().get_next_one_offset(r + 1, size)) { 1904 resolve_exception_edge(handler, throwing_op_id, r, NULL, move_resolver); 1905 } 1906 1907 // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately 1908 for_each_phi_fun(block, phi, 1909 resolve_exception_edge(handler, throwing_op_id, phi->operand()->vreg_number(), phi, move_resolver) 1910 ); 1911 1912 if (move_resolver.has_mappings()) { 1913 LIR_List* entry_code = new LIR_List(compilation()); 1914 move_resolver.set_insert_position(entry_code, 0); 1915 move_resolver.resolve_and_append_moves(); 1916 1917 entry_code->jump(handler->entry_block()); 1918 handler->set_entry_code(entry_code); 1919 } 1920 } 1921 1922 1923 void LinearScan::resolve_exception_handlers() { 1924 MoveResolver move_resolver(this); 1925 LIR_OpVisitState visitor; 1926 int num_blocks = block_count(); 1927 1928 int i; 1929 for (i = 0; i < num_blocks; i++) { 1930 BlockBegin* block = block_at(i); 1931 if (block->is_set(BlockBegin::exception_entry_flag)) { 1932 resolve_exception_entry(block, move_resolver); 1933 } 1934 } 1935 1936 for (i = 0; i < num_blocks; i++) { 1937 BlockBegin* block = block_at(i); 1938 LIR_List* ops = block->lir(); 1939 int num_ops = ops->length(); 1940 1941 // iterate all instructions of the block. skip the first because it is always a label 1942 assert(visitor.no_operands(ops->at(0)), "first operation must always be a label"); 1943 for (int j = 1; j < num_ops; j++) { 1944 LIR_Op* op = ops->at(j); 1945 int op_id = op->id(); 1946 1947 if (op_id != -1 && has_info(op_id)) { 1948 // visit operation to collect all operands 1949 visitor.visit(op); 1950 assert(visitor.info_count() > 0, "should not visit otherwise"); 1951 1952 XHandlers* xhandlers = visitor.all_xhandler(); 1953 int n = xhandlers->length(); 1954 for (int k = 0; k < n; k++) { 1955 resolve_exception_edge(xhandlers->handler_at(k), op_id, move_resolver); 1956 } 1957 1958 #ifdef ASSERT 1959 } else { 1960 visitor.visit(op); 1961 assert(visitor.all_xhandler()->length() == 0, "missed exception handler"); 1962 #endif 1963 } 1964 } 1965 } 1966 } 1967 1968 1969 // ********** Phase 7: assign register numbers back to LIR 1970 // (includes computation of debug information and oop maps) 1971 1972 VMReg LinearScan::vm_reg_for_interval(Interval* interval) { 1973 VMReg reg = interval->cached_vm_reg(); 1974 if (!reg->is_valid() ) { 1975 reg = vm_reg_for_operand(operand_for_interval(interval)); 1976 interval->set_cached_vm_reg(reg); 1977 } 1978 assert(reg == vm_reg_for_operand(operand_for_interval(interval)), "wrong cached value"); 1979 return reg; 1980 } 1981 1982 VMReg LinearScan::vm_reg_for_operand(LIR_Opr opr) { 1983 assert(opr->is_oop(), "currently only implemented for oop operands"); 1984 return frame_map()->regname(opr); 1985 } 1986 1987 1988 LIR_Opr LinearScan::operand_for_interval(Interval* interval) { 1989 LIR_Opr opr = interval->cached_opr(); 1990 if (opr->is_illegal()) { 1991 opr = calc_operand_for_interval(interval); 1992 interval->set_cached_opr(opr); 1993 } 1994 1995 assert(opr == calc_operand_for_interval(interval), "wrong cached value"); 1996 return opr; 1997 } 1998 1999 LIR_Opr LinearScan::calc_operand_for_interval(const Interval* interval) { 2000 int assigned_reg = interval->assigned_reg(); 2001 BasicType type = interval->type(); 2002 2003 if (assigned_reg >= nof_regs) { 2004 // stack slot 2005 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2006 return LIR_OprFact::stack(assigned_reg - nof_regs, type); 2007 2008 } else { 2009 // register 2010 switch (type) { 2011 case T_OBJECT: { 2012 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2013 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2014 return LIR_OprFact::single_cpu_oop(assigned_reg); 2015 } 2016 2017 case T_INT: { 2018 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2019 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2020 return LIR_OprFact::single_cpu(assigned_reg); 2021 } 2022 2023 case T_LONG: { 2024 int assigned_regHi = interval->assigned_regHi(); 2025 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2026 assert(num_physical_regs(T_LONG) == 1 || 2027 (assigned_regHi >= pd_first_cpu_reg && assigned_regHi <= pd_last_cpu_reg), "no cpu register"); 2028 2029 assert(assigned_reg != assigned_regHi, "invalid allocation"); 2030 assert(num_physical_regs(T_LONG) == 1 || assigned_reg < assigned_regHi, 2031 "register numbers must be sorted (ensure that e.g. a move from eax,ebx to ebx,eax can not occur)"); 2032 assert((assigned_regHi != any_reg) ^ (num_physical_regs(T_LONG) == 1), "must be match"); 2033 if (requires_adjacent_regs(T_LONG)) { 2034 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == assigned_regHi, "must be sequential and even"); 2035 } 2036 2037 #ifdef _LP64 2038 return LIR_OprFact::double_cpu(assigned_reg, assigned_reg); 2039 #else 2040 #ifdef SPARC 2041 return LIR_OprFact::double_cpu(assigned_regHi, assigned_reg); 2042 #else 2043 return LIR_OprFact::double_cpu(assigned_reg, assigned_regHi); 2044 #endif // SPARC 2045 #endif // LP64 2046 } 2047 2048 case T_FLOAT: { 2049 #ifdef X86 2050 if (UseSSE >= 1) { 2051 assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= pd_last_xmm_reg, "no xmm register"); 2052 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2053 return LIR_OprFact::single_xmm(assigned_reg - pd_first_xmm_reg); 2054 } 2055 #endif 2056 2057 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register"); 2058 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2059 return LIR_OprFact::single_fpu(assigned_reg - pd_first_fpu_reg); 2060 } 2061 2062 case T_DOUBLE: { 2063 #ifdef X86 2064 if (UseSSE >= 2) { 2065 assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= pd_last_xmm_reg, "no xmm register"); 2066 assert(interval->assigned_regHi() == any_reg, "must not have hi register (double xmm values are stored in one register)"); 2067 return LIR_OprFact::double_xmm(assigned_reg - pd_first_xmm_reg); 2068 } 2069 #endif 2070 2071 #ifdef SPARC 2072 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register"); 2073 assert(interval->assigned_regHi() >= pd_first_fpu_reg && interval->assigned_regHi() <= pd_last_fpu_reg, "no fpu register"); 2074 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == interval->assigned_regHi(), "must be sequential and even"); 2075 LIR_Opr result = LIR_OprFact::double_fpu(interval->assigned_regHi() - pd_first_fpu_reg, assigned_reg - pd_first_fpu_reg); 2076 #else 2077 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register"); 2078 assert(interval->assigned_regHi() == any_reg, "must not have hi register (double fpu values are stored in one register on Intel)"); 2079 LIR_Opr result = LIR_OprFact::double_fpu(assigned_reg - pd_first_fpu_reg); 2080 #endif 2081 return result; 2082 } 2083 2084 default: { 2085 ShouldNotReachHere(); 2086 return LIR_OprFact::illegalOpr; 2087 } 2088 } 2089 } 2090 } 2091 2092 LIR_Opr LinearScan::canonical_spill_opr(Interval* interval) { 2093 assert(interval->canonical_spill_slot() >= nof_regs, "canonical spill slot not set"); 2094 return LIR_OprFact::stack(interval->canonical_spill_slot() - nof_regs, interval->type()); 2095 } 2096 2097 LIR_Opr LinearScan::color_lir_opr(LIR_Opr opr, int op_id, LIR_OpVisitState::OprMode mode) { 2098 assert(opr->is_virtual(), "should not call this otherwise"); 2099 2100 Interval* interval = interval_at(opr->vreg_number()); 2101 assert(interval != NULL, "interval must exist"); 2102 2103 if (op_id != -1) { 2104 #ifdef ASSERT 2105 BlockBegin* block = block_of_op_with_id(op_id); 2106 if (block->number_of_sux() <= 1 && op_id == block->last_lir_instruction_id()) { 2107 // check if spill moves could have been appended at the end of this block, but 2108 // before the branch instruction. So the split child information for this branch would 2109 // be incorrect. 2110 LIR_OpBranch* branch = block->lir()->instructions_list()->last()->as_OpBranch(); 2111 if (branch != NULL) { 2112 if (block->live_out().at(opr->vreg_number())) { 2113 assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump"); 2114 assert(false, "can't get split child for the last branch of a block because the information would be incorrect (moves are inserted before the branch in resolve_data_flow)"); 2115 } 2116 } 2117 } 2118 #endif 2119 2120 // operands are not changed when an interval is split during allocation, 2121 // so search the right interval here 2122 interval = split_child_at_op_id(interval, op_id, mode); 2123 } 2124 2125 LIR_Opr res = operand_for_interval(interval); 2126 2127 #ifdef X86 2128 // new semantic for is_last_use: not only set on definite end of interval, 2129 // but also before hole 2130 // This may still miss some cases (e.g. for dead values), but it is not necessary that the 2131 // last use information is completely correct 2132 // information is only needed for fpu stack allocation 2133 if (res->is_fpu_register()) { 2134 if (opr->is_last_use() || op_id == interval->to() || (op_id != -1 && interval->has_hole_between(op_id, op_id + 1))) { 2135 assert(op_id == -1 || !is_block_begin(op_id), "holes at begin of block may also result from control flow"); 2136 res = res->make_last_use(); 2137 } 2138 } 2139 #endif 2140 2141 assert(!gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::callee_saved) || !FrameMap::is_caller_save_register(res), "bad allocation"); 2142 2143 return res; 2144 } 2145 2146 2147 #ifdef ASSERT 2148 // some methods used to check correctness of debug information 2149 2150 void assert_no_register_values(GrowableArray<ScopeValue*>* values) { 2151 if (values == NULL) { 2152 return; 2153 } 2154 2155 for (int i = 0; i < values->length(); i++) { 2156 ScopeValue* value = values->at(i); 2157 2158 if (value->is_location()) { 2159 Location location = ((LocationValue*)value)->location(); 2160 assert(location.where() == Location::on_stack, "value is in register"); 2161 } 2162 } 2163 } 2164 2165 void assert_no_register_values(GrowableArray<MonitorValue*>* values) { 2166 if (values == NULL) { 2167 return; 2168 } 2169 2170 for (int i = 0; i < values->length(); i++) { 2171 MonitorValue* value = values->at(i); 2172 2173 if (value->owner()->is_location()) { 2174 Location location = ((LocationValue*)value->owner())->location(); 2175 assert(location.where() == Location::on_stack, "owner is in register"); 2176 } 2177 assert(value->basic_lock().where() == Location::on_stack, "basic_lock is in register"); 2178 } 2179 } 2180 2181 void assert_equal(Location l1, Location l2) { 2182 assert(l1.where() == l2.where() && l1.type() == l2.type() && l1.offset() == l2.offset(), ""); 2183 } 2184 2185 void assert_equal(ScopeValue* v1, ScopeValue* v2) { 2186 if (v1->is_location()) { 2187 assert(v2->is_location(), ""); 2188 assert_equal(((LocationValue*)v1)->location(), ((LocationValue*)v2)->location()); 2189 } else if (v1->is_constant_int()) { 2190 assert(v2->is_constant_int(), ""); 2191 assert(((ConstantIntValue*)v1)->value() == ((ConstantIntValue*)v2)->value(), ""); 2192 } else if (v1->is_constant_double()) { 2193 assert(v2->is_constant_double(), ""); 2194 assert(((ConstantDoubleValue*)v1)->value() == ((ConstantDoubleValue*)v2)->value(), ""); 2195 } else if (v1->is_constant_long()) { 2196 assert(v2->is_constant_long(), ""); 2197 assert(((ConstantLongValue*)v1)->value() == ((ConstantLongValue*)v2)->value(), ""); 2198 } else if (v1->is_constant_oop()) { 2199 assert(v2->is_constant_oop(), ""); 2200 assert(((ConstantOopWriteValue*)v1)->value() == ((ConstantOopWriteValue*)v2)->value(), ""); 2201 } else { 2202 ShouldNotReachHere(); 2203 } 2204 } 2205 2206 void assert_equal(MonitorValue* m1, MonitorValue* m2) { 2207 assert_equal(m1->owner(), m2->owner()); 2208 assert_equal(m1->basic_lock(), m2->basic_lock()); 2209 } 2210 2211 void assert_equal(IRScopeDebugInfo* d1, IRScopeDebugInfo* d2) { 2212 assert(d1->scope() == d2->scope(), "not equal"); 2213 assert(d1->bci() == d2->bci(), "not equal"); 2214 2215 if (d1->locals() != NULL) { 2216 assert(d1->locals() != NULL && d2->locals() != NULL, "not equal"); 2217 assert(d1->locals()->length() == d2->locals()->length(), "not equal"); 2218 for (int i = 0; i < d1->locals()->length(); i++) { 2219 assert_equal(d1->locals()->at(i), d2->locals()->at(i)); 2220 } 2221 } else { 2222 assert(d1->locals() == NULL && d2->locals() == NULL, "not equal"); 2223 } 2224 2225 if (d1->expressions() != NULL) { 2226 assert(d1->expressions() != NULL && d2->expressions() != NULL, "not equal"); 2227 assert(d1->expressions()->length() == d2->expressions()->length(), "not equal"); 2228 for (int i = 0; i < d1->expressions()->length(); i++) { 2229 assert_equal(d1->expressions()->at(i), d2->expressions()->at(i)); 2230 } 2231 } else { 2232 assert(d1->expressions() == NULL && d2->expressions() == NULL, "not equal"); 2233 } 2234 2235 if (d1->monitors() != NULL) { 2236 assert(d1->monitors() != NULL && d2->monitors() != NULL, "not equal"); 2237 assert(d1->monitors()->length() == d2->monitors()->length(), "not equal"); 2238 for (int i = 0; i < d1->monitors()->length(); i++) { 2239 assert_equal(d1->monitors()->at(i), d2->monitors()->at(i)); 2240 } 2241 } else { 2242 assert(d1->monitors() == NULL && d2->monitors() == NULL, "not equal"); 2243 } 2244 2245 if (d1->caller() != NULL) { 2246 assert(d1->caller() != NULL && d2->caller() != NULL, "not equal"); 2247 assert_equal(d1->caller(), d2->caller()); 2248 } else { 2249 assert(d1->caller() == NULL && d2->caller() == NULL, "not equal"); 2250 } 2251 } 2252 2253 void check_stack_depth(CodeEmitInfo* info, int stack_end) { 2254 if (info->bci() != SynchronizationEntryBCI && !info->scope()->method()->is_native()) { 2255 Bytecodes::Code code = info->scope()->method()->java_code_at_bci(info->bci()); 2256 switch (code) { 2257 case Bytecodes::_ifnull : // fall through 2258 case Bytecodes::_ifnonnull : // fall through 2259 case Bytecodes::_ifeq : // fall through 2260 case Bytecodes::_ifne : // fall through 2261 case Bytecodes::_iflt : // fall through 2262 case Bytecodes::_ifge : // fall through 2263 case Bytecodes::_ifgt : // fall through 2264 case Bytecodes::_ifle : // fall through 2265 case Bytecodes::_if_icmpeq : // fall through 2266 case Bytecodes::_if_icmpne : // fall through 2267 case Bytecodes::_if_icmplt : // fall through 2268 case Bytecodes::_if_icmpge : // fall through 2269 case Bytecodes::_if_icmpgt : // fall through 2270 case Bytecodes::_if_icmple : // fall through 2271 case Bytecodes::_if_acmpeq : // fall through 2272 case Bytecodes::_if_acmpne : 2273 assert(stack_end >= -Bytecodes::depth(code), "must have non-empty expression stack at if bytecode"); 2274 break; 2275 } 2276 } 2277 } 2278 2279 #endif // ASSERT 2280 2281 2282 IntervalWalker* LinearScan::init_compute_oop_maps() { 2283 // setup lists of potential oops for walking 2284 Interval* oop_intervals; 2285 Interval* non_oop_intervals; 2286 2287 create_unhandled_lists(&oop_intervals, &non_oop_intervals, is_oop_interval, NULL); 2288 2289 // intervals that have no oops inside need not to be processed 2290 // to ensure a walking until the last instruction id, add a dummy interval 2291 // with a high operation id 2292 non_oop_intervals = new Interval(any_reg); 2293 non_oop_intervals->add_range(max_jint - 2, max_jint - 1); 2294 2295 return new IntervalWalker(this, oop_intervals, non_oop_intervals); 2296 } 2297 2298 2299 OopMap* LinearScan::compute_oop_map(IntervalWalker* iw, LIR_Op* op, CodeEmitInfo* info, bool is_call_site) { 2300 TRACE_LINEAR_SCAN(3, tty->print_cr("creating oop map at op_id %d", op->id())); 2301 2302 // walk before the current operation -> intervals that start at 2303 // the operation (= output operands of the operation) are not 2304 // included in the oop map 2305 iw->walk_before(op->id()); 2306 2307 int frame_size = frame_map()->framesize(); 2308 int arg_count = frame_map()->oop_map_arg_count(); 2309 OopMap* map = new OopMap(frame_size, arg_count); 2310 2311 // Check if this is a patch site. 2312 bool is_patch_info = false; 2313 if (op->code() == lir_move) { 2314 assert(!is_call_site, "move must not be a call site"); 2315 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); 2316 LIR_Op1* move = (LIR_Op1*)op; 2317 2318 is_patch_info = move->patch_code() != lir_patch_none; 2319 } 2320 2321 // Iterate through active intervals 2322 for (Interval* interval = iw->active_first(fixedKind); interval != Interval::end(); interval = interval->next()) { 2323 int assigned_reg = interval->assigned_reg(); 2324 2325 assert(interval->current_from() <= op->id() && op->id() <= interval->current_to(), "interval should not be active otherwise"); 2326 assert(interval->assigned_regHi() == any_reg, "oop must be single word"); 2327 assert(interval->reg_num() >= LIR_OprDesc::vreg_base, "fixed interval found"); 2328 2329 // Check if this range covers the instruction. Intervals that 2330 // start or end at the current operation are not included in the 2331 // oop map, except in the case of patching moves. For patching 2332 // moves, any intervals which end at this instruction are included 2333 // in the oop map since we may safepoint while doing the patch 2334 // before we've consumed the inputs. 2335 if (is_patch_info || op->id() < interval->current_to()) { 2336 2337 // caller-save registers must not be included into oop-maps at calls 2338 assert(!is_call_site || assigned_reg >= nof_regs || !is_caller_save(assigned_reg), "interval is in a caller-save register at a call -> register will be overwritten"); 2339 2340 VMReg name = vm_reg_for_interval(interval); 2341 map->set_oop(name); 2342 2343 // Spill optimization: when the stack value is guaranteed to be always correct, 2344 // then it must be added to the oop map even if the interval is currently in a register 2345 if (interval->always_in_memory() && 2346 op->id() > interval->spill_definition_pos() && 2347 interval->assigned_reg() != interval->canonical_spill_slot()) { 2348 assert(interval->spill_definition_pos() > 0, "position not set correctly"); 2349 assert(interval->canonical_spill_slot() >= LinearScan::nof_regs, "no spill slot assigned"); 2350 assert(interval->assigned_reg() < LinearScan::nof_regs, "interval is on stack, so stack slot is registered twice"); 2351 2352 map->set_oop(frame_map()->slot_regname(interval->canonical_spill_slot() - LinearScan::nof_regs)); 2353 } 2354 } 2355 } 2356 2357 // add oops from lock stack 2358 assert(info->stack() != NULL, "CodeEmitInfo must always have a stack"); 2359 int locks_count = info->stack()->locks_size(); 2360 for (int i = 0; i < locks_count; i++) { 2361 map->set_oop(frame_map()->monitor_object_regname(i)); 2362 } 2363 2364 return map; 2365 } 2366 2367 2368 void LinearScan::compute_oop_map(IntervalWalker* iw, const LIR_OpVisitState &visitor, LIR_Op* op) { 2369 assert(visitor.info_count() > 0, "no oop map needed"); 2370 2371 // compute oop_map only for first CodeEmitInfo 2372 // because it is (in most cases) equal for all other infos of the same operation 2373 CodeEmitInfo* first_info = visitor.info_at(0); 2374 OopMap* first_oop_map = compute_oop_map(iw, op, first_info, visitor.has_call()); 2375 2376 for (int i = 0; i < visitor.info_count(); i++) { 2377 CodeEmitInfo* info = visitor.info_at(i); 2378 OopMap* oop_map = first_oop_map; 2379 2380 if (info->stack()->locks_size() != first_info->stack()->locks_size()) { 2381 // this info has a different number of locks then the precomputed oop map 2382 // (possible for lock and unlock instructions) -> compute oop map with 2383 // correct lock information 2384 oop_map = compute_oop_map(iw, op, info, visitor.has_call()); 2385 } 2386 2387 if (info->_oop_map == NULL) { 2388 info->_oop_map = oop_map; 2389 } else { 2390 // a CodeEmitInfo can not be shared between different LIR-instructions 2391 // because interval splitting can occur anywhere between two instructions 2392 // and so the oop maps must be different 2393 // -> check if the already set oop_map is exactly the one calculated for this operation 2394 assert(info->_oop_map == oop_map, "same CodeEmitInfo used for multiple LIR instructions"); 2395 } 2396 } 2397 } 2398 2399 2400 // frequently used constants 2401 ConstantOopWriteValue LinearScan::_oop_null_scope_value = ConstantOopWriteValue(NULL); 2402 ConstantIntValue LinearScan::_int_m1_scope_value = ConstantIntValue(-1); 2403 ConstantIntValue LinearScan::_int_0_scope_value = ConstantIntValue(0); 2404 ConstantIntValue LinearScan::_int_1_scope_value = ConstantIntValue(1); 2405 ConstantIntValue LinearScan::_int_2_scope_value = ConstantIntValue(2); 2406 LocationValue _illegal_value = LocationValue(Location()); 2407 2408 void LinearScan::init_compute_debug_info() { 2409 // cache for frequently used scope values 2410 // (cpu registers and stack slots) 2411 _scope_value_cache = ScopeValueArray((LinearScan::nof_cpu_regs + frame_map()->argcount() + max_spills()) * 2, NULL); 2412 } 2413 2414 MonitorValue* LinearScan::location_for_monitor_index(int monitor_index) { 2415 Location loc; 2416 if (!frame_map()->location_for_monitor_object(monitor_index, &loc)) { 2417 bailout("too large frame"); 2418 } 2419 ScopeValue* object_scope_value = new LocationValue(loc); 2420 2421 if (!frame_map()->location_for_monitor_lock(monitor_index, &loc)) { 2422 bailout("too large frame"); 2423 } 2424 return new MonitorValue(object_scope_value, loc); 2425 } 2426 2427 LocationValue* LinearScan::location_for_name(int name, Location::Type loc_type) { 2428 Location loc; 2429 if (!frame_map()->locations_for_slot(name, loc_type, &loc)) { 2430 bailout("too large frame"); 2431 } 2432 return new LocationValue(loc); 2433 } 2434 2435 2436 int LinearScan::append_scope_value_for_constant(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) { 2437 assert(opr->is_constant(), "should not be called otherwise"); 2438 2439 LIR_Const* c = opr->as_constant_ptr(); 2440 BasicType t = c->type(); 2441 switch (t) { 2442 case T_OBJECT: { 2443 jobject value = c->as_jobject(); 2444 if (value == NULL) { 2445 scope_values->append(&_oop_null_scope_value); 2446 } else { 2447 scope_values->append(new ConstantOopWriteValue(c->as_jobject())); 2448 } 2449 return 1; 2450 } 2451 2452 case T_INT: // fall through 2453 case T_FLOAT: { 2454 int value = c->as_jint_bits(); 2455 switch (value) { 2456 case -1: scope_values->append(&_int_m1_scope_value); break; 2457 case 0: scope_values->append(&_int_0_scope_value); break; 2458 case 1: scope_values->append(&_int_1_scope_value); break; 2459 case 2: scope_values->append(&_int_2_scope_value); break; 2460 default: scope_values->append(new ConstantIntValue(c->as_jint_bits())); break; 2461 } 2462 return 1; 2463 } 2464 2465 case T_LONG: // fall through 2466 case T_DOUBLE: { 2467 #ifdef _LP64 2468 scope_values->append(&_int_0_scope_value); 2469 scope_values->append(new ConstantLongValue(c->as_jlong_bits())); 2470 #else 2471 if (hi_word_offset_in_bytes > lo_word_offset_in_bytes) { 2472 scope_values->append(new ConstantIntValue(c->as_jint_hi_bits())); 2473 scope_values->append(new ConstantIntValue(c->as_jint_lo_bits())); 2474 } else { 2475 scope_values->append(new ConstantIntValue(c->as_jint_lo_bits())); 2476 scope_values->append(new ConstantIntValue(c->as_jint_hi_bits())); 2477 } 2478 #endif 2479 return 2; 2480 } 2481 2482 case T_ADDRESS: { 2483 #ifdef _LP64 2484 scope_values->append(new ConstantLongValue(c->as_jint())); 2485 #else 2486 scope_values->append(new ConstantIntValue(c->as_jint())); 2487 #endif 2488 return 1; 2489 } 2490 2491 default: 2492 ShouldNotReachHere(); 2493 return -1; 2494 } 2495 } 2496 2497 int LinearScan::append_scope_value_for_operand(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) { 2498 if (opr->is_single_stack()) { 2499 int stack_idx = opr->single_stack_ix(); 2500 bool is_oop = opr->is_oop_register(); 2501 int cache_idx = (stack_idx + LinearScan::nof_cpu_regs) * 2 + (is_oop ? 1 : 0); 2502 2503 ScopeValue* sv = _scope_value_cache.at(cache_idx); 2504 if (sv == NULL) { 2505 Location::Type loc_type = is_oop ? Location::oop : Location::normal; 2506 sv = location_for_name(stack_idx, loc_type); 2507 _scope_value_cache.at_put(cache_idx, sv); 2508 } 2509 2510 // check if cached value is correct 2511 DEBUG_ONLY(assert_equal(sv, location_for_name(stack_idx, is_oop ? Location::oop : Location::normal))); 2512 2513 scope_values->append(sv); 2514 return 1; 2515 2516 } else if (opr->is_single_cpu()) { 2517 bool is_oop = opr->is_oop_register(); 2518 int cache_idx = opr->cpu_regnr() * 2 + (is_oop ? 1 : 0); 2519 Location::Type int_loc_type = NOT_LP64(Location::normal) LP64_ONLY(Location::int_in_long); 2520 2521 ScopeValue* sv = _scope_value_cache.at(cache_idx); 2522 if (sv == NULL) { 2523 Location::Type loc_type = is_oop ? Location::oop : int_loc_type; 2524 VMReg rname = frame_map()->regname(opr); 2525 sv = new LocationValue(Location::new_reg_loc(loc_type, rname)); 2526 _scope_value_cache.at_put(cache_idx, sv); 2527 } 2528 2529 // check if cached value is correct 2530 DEBUG_ONLY(assert_equal(sv, new LocationValue(Location::new_reg_loc(is_oop ? Location::oop : int_loc_type, frame_map()->regname(opr))))); 2531 2532 scope_values->append(sv); 2533 return 1; 2534 2535 #ifdef X86 2536 } else if (opr->is_single_xmm()) { 2537 VMReg rname = opr->as_xmm_float_reg()->as_VMReg(); 2538 LocationValue* sv = new LocationValue(Location::new_reg_loc(Location::normal, rname)); 2539 2540 scope_values->append(sv); 2541 return 1; 2542 #endif 2543 2544 } else if (opr->is_single_fpu()) { 2545 #ifdef X86 2546 // the exact location of fpu stack values is only known 2547 // during fpu stack allocation, so the stack allocator object 2548 // must be present 2549 assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)"); 2550 assert(_fpu_stack_allocator != NULL, "must be present"); 2551 opr = _fpu_stack_allocator->to_fpu_stack(opr); 2552 #endif 2553 2554 Location::Type loc_type = float_saved_as_double ? Location::float_in_dbl : Location::normal; 2555 VMReg rname = frame_map()->fpu_regname(opr->fpu_regnr()); 2556 LocationValue* sv = new LocationValue(Location::new_reg_loc(loc_type, rname)); 2557 2558 scope_values->append(sv); 2559 return 1; 2560 2561 } else { 2562 // double-size operands 2563 2564 ScopeValue* first; 2565 ScopeValue* second; 2566 2567 if (opr->is_double_stack()) { 2568 #ifdef _LP64 2569 Location loc1; 2570 Location::Type loc_type = opr->type() == T_LONG ? Location::lng : Location::dbl; 2571 if (!frame_map()->locations_for_slot(opr->double_stack_ix(), loc_type, &loc1, NULL)) { 2572 bailout("too large frame"); 2573 } 2574 // Does this reverse on x86 vs. sparc? 2575 first = new LocationValue(loc1); 2576 second = &_int_0_scope_value; 2577 #else 2578 Location loc1, loc2; 2579 if (!frame_map()->locations_for_slot(opr->double_stack_ix(), Location::normal, &loc1, &loc2)) { 2580 bailout("too large frame"); 2581 } 2582 first = new LocationValue(loc1); 2583 second = new LocationValue(loc2); 2584 #endif // _LP64 2585 2586 } else if (opr->is_double_cpu()) { 2587 #ifdef _LP64 2588 VMReg rname_first = opr->as_register_lo()->as_VMReg(); 2589 first = new LocationValue(Location::new_reg_loc(Location::lng, rname_first)); 2590 second = &_int_0_scope_value; 2591 #else 2592 VMReg rname_first = opr->as_register_lo()->as_VMReg(); 2593 VMReg rname_second = opr->as_register_hi()->as_VMReg(); 2594 2595 if (hi_word_offset_in_bytes < lo_word_offset_in_bytes) { 2596 // lo/hi and swapped relative to first and second, so swap them 2597 VMReg tmp = rname_first; 2598 rname_first = rname_second; 2599 rname_second = tmp; 2600 } 2601 2602 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); 2603 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); 2604 #endif //_LP64 2605 2606 2607 #ifdef X86 2608 } else if (opr->is_double_xmm()) { 2609 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation"); 2610 VMReg rname_first = opr->as_xmm_double_reg()->as_VMReg(); 2611 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); 2612 // %%% This is probably a waste but we'll keep things as they were for now 2613 if (true) { 2614 VMReg rname_second = rname_first->next(); 2615 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); 2616 } 2617 #endif 2618 2619 } else if (opr->is_double_fpu()) { 2620 // On SPARC, fpu_regnrLo/fpu_regnrHi represents the two halves of 2621 // the double as float registers in the native ordering. On X86, 2622 // fpu_regnrLo is a FPU stack slot whose VMReg represents 2623 // the low-order word of the double and fpu_regnrLo + 1 is the 2624 // name for the other half. *first and *second must represent the 2625 // least and most significant words, respectively. 2626 2627 #ifdef X86 2628 // the exact location of fpu stack values is only known 2629 // during fpu stack allocation, so the stack allocator object 2630 // must be present 2631 assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)"); 2632 assert(_fpu_stack_allocator != NULL, "must be present"); 2633 opr = _fpu_stack_allocator->to_fpu_stack(opr); 2634 2635 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation (only fpu_regnrHi is used)"); 2636 #endif 2637 #ifdef SPARC 2638 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi() + 1, "assumed in calculation (only fpu_regnrHi is used)"); 2639 #endif 2640 2641 VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrHi()); 2642 2643 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); 2644 // %%% This is probably a waste but we'll keep things as they were for now 2645 if (true) { 2646 VMReg rname_second = rname_first->next(); 2647 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); 2648 } 2649 2650 } else { 2651 ShouldNotReachHere(); 2652 first = NULL; 2653 second = NULL; 2654 } 2655 2656 assert(first != NULL && second != NULL, "must be set"); 2657 // The convention the interpreter uses is that the second local 2658 // holds the first raw word of the native double representation. 2659 // This is actually reasonable, since locals and stack arrays 2660 // grow downwards in all implementations. 2661 // (If, on some machine, the interpreter's Java locals or stack 2662 // were to grow upwards, the embedded doubles would be word-swapped.) 2663 scope_values->append(second); 2664 scope_values->append(first); 2665 return 2; 2666 } 2667 } 2668 2669 2670 int LinearScan::append_scope_value(int op_id, Value value, GrowableArray<ScopeValue*>* scope_values) { 2671 if (value != NULL) { 2672 LIR_Opr opr = value->operand(); 2673 Constant* con = value->as_Constant(); 2674 2675 assert(con == NULL || opr->is_virtual() || opr->is_constant() || opr->is_illegal(), "asumption: Constant instructions have only constant operands (or illegal if constant is optimized away)"); 2676 assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands"); 2677 2678 if (con != NULL && !con->is_pinned() && !opr->is_constant()) { 2679 // Unpinned constants may have a virtual operand for a part of the lifetime 2680 // or may be illegal when it was optimized away, 2681 // so always use a constant operand 2682 opr = LIR_OprFact::value_type(con->type()); 2683 } 2684 assert(opr->is_virtual() || opr->is_constant(), "other cases not allowed here"); 2685 2686 if (opr->is_virtual()) { 2687 LIR_OpVisitState::OprMode mode = LIR_OpVisitState::inputMode; 2688 2689 BlockBegin* block = block_of_op_with_id(op_id); 2690 if (block->number_of_sux() == 1 && op_id == block->last_lir_instruction_id()) { 2691 // generating debug information for the last instruction of a block. 2692 // if this instruction is a branch, spill moves are inserted before this branch 2693 // and so the wrong operand would be returned (spill moves at block boundaries are not 2694 // considered in the live ranges of intervals) 2695 // Solution: use the first op_id of the branch target block instead. 2696 if (block->lir()->instructions_list()->last()->as_OpBranch() != NULL) { 2697 if (block->live_out().at(opr->vreg_number())) { 2698 op_id = block->sux_at(0)->first_lir_instruction_id(); 2699 mode = LIR_OpVisitState::outputMode; 2700 } 2701 } 2702 } 2703 2704 // Get current location of operand 2705 // The operand must be live because debug information is considered when building the intervals 2706 // if the interval is not live, color_lir_opr will cause an assertion failure 2707 opr = color_lir_opr(opr, op_id, mode); 2708 assert(!has_call(op_id) || opr->is_stack() || !is_caller_save(reg_num(opr)), "can not have caller-save register operands at calls"); 2709 2710 // Append to ScopeValue array 2711 return append_scope_value_for_operand(opr, scope_values); 2712 2713 } else { 2714 assert(value->as_Constant() != NULL, "all other instructions have only virtual operands"); 2715 assert(opr->is_constant(), "operand must be constant"); 2716 2717 return append_scope_value_for_constant(opr, scope_values); 2718 } 2719 } else { 2720 // append a dummy value because real value not needed 2721 scope_values->append(&_illegal_value); 2722 return 1; 2723 } 2724 } 2725 2726 2727 IRScopeDebugInfo* LinearScan::compute_debug_info_for_scope(int op_id, IRScope* cur_scope, ValueStack* cur_state, ValueStack* innermost_state, int cur_bci, int stack_end, int locks_end) { 2728 IRScopeDebugInfo* caller_debug_info = NULL; 2729 int stack_begin, locks_begin; 2730 2731 ValueStack* caller_state = cur_scope->caller_state(); 2732 if (caller_state != NULL) { 2733 // process recursively to compute outermost scope first 2734 stack_begin = caller_state->stack_size(); 2735 locks_begin = caller_state->locks_size(); 2736 caller_debug_info = compute_debug_info_for_scope(op_id, cur_scope->caller(), caller_state, innermost_state, cur_scope->caller_bci(), stack_begin, locks_begin); 2737 } else { 2738 stack_begin = 0; 2739 locks_begin = 0; 2740 } 2741 2742 // initialize these to null. 2743 // If we don't need deopt info or there are no locals, expressions or monitors, 2744 // then these get recorded as no information and avoids the allocation of 0 length arrays. 2745 GrowableArray<ScopeValue*>* locals = NULL; 2746 GrowableArray<ScopeValue*>* expressions = NULL; 2747 GrowableArray<MonitorValue*>* monitors = NULL; 2748 2749 // describe local variable values 2750 int nof_locals = cur_scope->method()->max_locals(); 2751 if (nof_locals > 0) { 2752 locals = new GrowableArray<ScopeValue*>(nof_locals); 2753 2754 int pos = 0; 2755 while (pos < nof_locals) { 2756 assert(pos < cur_state->locals_size(), "why not?"); 2757 2758 Value local = cur_state->local_at(pos); 2759 pos += append_scope_value(op_id, local, locals); 2760 2761 assert(locals->length() == pos, "must match"); 2762 } 2763 assert(locals->length() == cur_scope->method()->max_locals(), "wrong number of locals"); 2764 assert(locals->length() == cur_state->locals_size(), "wrong number of locals"); 2765 } 2766 2767 2768 // describe expression stack 2769 // 2770 // When we inline methods containing exception handlers, the 2771 // "lock_stacks" are changed to preserve expression stack values 2772 // in caller scopes when exception handlers are present. This 2773 // can cause callee stacks to be smaller than caller stacks. 2774 if (stack_end > innermost_state->stack_size()) { 2775 stack_end = innermost_state->stack_size(); 2776 } 2777 2778 2779 2780 int nof_stack = stack_end - stack_begin; 2781 if (nof_stack > 0) { 2782 expressions = new GrowableArray<ScopeValue*>(nof_stack); 2783 2784 int pos = stack_begin; 2785 while (pos < stack_end) { 2786 Value expression = innermost_state->stack_at_inc(pos); 2787 append_scope_value(op_id, expression, expressions); 2788 2789 assert(expressions->length() + stack_begin == pos, "must match"); 2790 } 2791 } 2792 2793 // describe monitors 2794 assert(locks_begin <= locks_end, "error in scope iteration"); 2795 int nof_locks = locks_end - locks_begin; 2796 if (nof_locks > 0) { 2797 monitors = new GrowableArray<MonitorValue*>(nof_locks); 2798 for (int i = locks_begin; i < locks_end; i++) { 2799 monitors->append(location_for_monitor_index(i)); 2800 } 2801 } 2802 2803 return new IRScopeDebugInfo(cur_scope, cur_bci, locals, expressions, monitors, caller_debug_info); 2804 } 2805 2806 2807 void LinearScan::compute_debug_info(CodeEmitInfo* info, int op_id) { 2808 if (!compilation()->needs_debug_information()) { 2809 return; 2810 } 2811 TRACE_LINEAR_SCAN(3, tty->print_cr("creating debug information at op_id %d", op_id)); 2812 2813 IRScope* innermost_scope = info->scope(); 2814 ValueStack* innermost_state = info->stack(); 2815 2816 assert(innermost_scope != NULL && innermost_state != NULL, "why is it missing?"); 2817 2818 int stack_end = innermost_state->stack_size(); 2819 int locks_end = innermost_state->locks_size(); 2820 2821 DEBUG_ONLY(check_stack_depth(info, stack_end)); 2822 2823 if (info->_scope_debug_info == NULL) { 2824 // compute debug information 2825 info->_scope_debug_info = compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state, info->bci(), stack_end, locks_end); 2826 } else { 2827 // debug information already set. Check that it is correct from the current point of view 2828 DEBUG_ONLY(assert_equal(info->_scope_debug_info, compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state, info->bci(), stack_end, locks_end))); 2829 } 2830 } 2831 2832 2833 void LinearScan::assign_reg_num(LIR_OpList* instructions, IntervalWalker* iw) { 2834 LIR_OpVisitState visitor; 2835 int num_inst = instructions->length(); 2836 bool has_dead = false; 2837 2838 for (int j = 0; j < num_inst; j++) { 2839 LIR_Op* op = instructions->at(j); 2840 if (op == NULL) { // this can happen when spill-moves are removed in eliminate_spill_moves 2841 has_dead = true; 2842 continue; 2843 } 2844 int op_id = op->id(); 2845 2846 // visit instruction to get list of operands 2847 visitor.visit(op); 2848 2849 // iterate all modes of the visitor and process all virtual operands 2850 for_each_visitor_mode(mode) { 2851 int n = visitor.opr_count(mode); 2852 for (int k = 0; k < n; k++) { 2853 LIR_Opr opr = visitor.opr_at(mode, k); 2854 if (opr->is_virtual_register()) { 2855 visitor.set_opr_at(mode, k, color_lir_opr(opr, op_id, mode)); 2856 } 2857 } 2858 } 2859 2860 if (visitor.info_count() > 0) { 2861 // exception handling 2862 if (compilation()->has_exception_handlers()) { 2863 XHandlers* xhandlers = visitor.all_xhandler(); 2864 int n = xhandlers->length(); 2865 for (int k = 0; k < n; k++) { 2866 XHandler* handler = xhandlers->handler_at(k); 2867 if (handler->entry_code() != NULL) { 2868 assign_reg_num(handler->entry_code()->instructions_list(), NULL); 2869 } 2870 } 2871 } else { 2872 assert(visitor.all_xhandler()->length() == 0, "missed exception handler"); 2873 } 2874 2875 // compute oop map 2876 assert(iw != NULL, "needed for compute_oop_map"); 2877 compute_oop_map(iw, visitor, op); 2878 2879 // compute debug information 2880 if (!use_fpu_stack_allocation()) { 2881 // compute debug information if fpu stack allocation is not needed. 2882 // when fpu stack allocation is needed, the debug information can not 2883 // be computed here because the exact location of fpu operands is not known 2884 // -> debug information is created inside the fpu stack allocator 2885 int n = visitor.info_count(); 2886 for (int k = 0; k < n; k++) { 2887 compute_debug_info(visitor.info_at(k), op_id); 2888 } 2889 } 2890 } 2891 2892 #ifdef ASSERT 2893 // make sure we haven't made the op invalid. 2894 op->verify(); 2895 #endif 2896 2897 // remove useless moves 2898 if (op->code() == lir_move) { 2899 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); 2900 LIR_Op1* move = (LIR_Op1*)op; 2901 LIR_Opr src = move->in_opr(); 2902 LIR_Opr dst = move->result_opr(); 2903 if (dst == src || 2904 !dst->is_pointer() && !src->is_pointer() && 2905 src->is_same_register(dst)) { 2906 instructions->at_put(j, NULL); 2907 has_dead = true; 2908 } 2909 } 2910 } 2911 2912 if (has_dead) { 2913 // iterate all instructions of the block and remove all null-values. 2914 int insert_point = 0; 2915 for (int j = 0; j < num_inst; j++) { 2916 LIR_Op* op = instructions->at(j); 2917 if (op != NULL) { 2918 if (insert_point != j) { 2919 instructions->at_put(insert_point, op); 2920 } 2921 insert_point++; 2922 } 2923 } 2924 instructions->truncate(insert_point); 2925 } 2926 } 2927 2928 void LinearScan::assign_reg_num() { 2929 TIME_LINEAR_SCAN(timer_assign_reg_num); 2930 2931 init_compute_debug_info(); 2932 IntervalWalker* iw = init_compute_oop_maps(); 2933 2934 int num_blocks = block_count(); 2935 for (int i = 0; i < num_blocks; i++) { 2936 BlockBegin* block = block_at(i); 2937 assign_reg_num(block->lir()->instructions_list(), iw); 2938 } 2939 } 2940 2941 2942 void LinearScan::do_linear_scan() { 2943 NOT_PRODUCT(_total_timer.begin_method()); 2944 2945 number_instructions(); 2946 2947 NOT_PRODUCT(print_lir(1, "Before Register Allocation")); 2948 2949 compute_local_live_sets(); 2950 compute_global_live_sets(); 2951 CHECK_BAILOUT(); 2952 2953 build_intervals(); 2954 CHECK_BAILOUT(); 2955 sort_intervals_before_allocation(); 2956 2957 NOT_PRODUCT(print_intervals("Before Register Allocation")); 2958 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_before_alloc)); 2959 2960 allocate_registers(); 2961 CHECK_BAILOUT(); 2962 2963 resolve_data_flow(); 2964 if (compilation()->has_exception_handlers()) { 2965 resolve_exception_handlers(); 2966 } 2967 // fill in number of spill slots into frame_map 2968 propagate_spill_slots(); 2969 CHECK_BAILOUT(); 2970 2971 NOT_PRODUCT(print_intervals("After Register Allocation")); 2972 NOT_PRODUCT(print_lir(2, "LIR after register allocation:")); 2973 2974 sort_intervals_after_allocation(); 2975 2976 DEBUG_ONLY(verify()); 2977 2978 eliminate_spill_moves(); 2979 assign_reg_num(); 2980 CHECK_BAILOUT(); 2981 2982 NOT_PRODUCT(print_lir(2, "LIR after assignment of register numbers:")); 2983 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_after_asign)); 2984 2985 { TIME_LINEAR_SCAN(timer_allocate_fpu_stack); 2986 2987 if (use_fpu_stack_allocation()) { 2988 allocate_fpu_stack(); // Only has effect on Intel 2989 NOT_PRODUCT(print_lir(2, "LIR after FPU stack allocation:")); 2990 } 2991 } 2992 2993 { TIME_LINEAR_SCAN(timer_optimize_lir); 2994 2995 EdgeMoveOptimizer::optimize(ir()->code()); 2996 ControlFlowOptimizer::optimize(ir()->code()); 2997 // check that cfg is still correct after optimizations 2998 ir()->verify(); 2999 } 3000 3001 NOT_PRODUCT(print_lir(1, "Before Code Generation", false)); 3002 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_final)); 3003 NOT_PRODUCT(_total_timer.end_method(this)); 3004 } 3005 3006 3007 // ********** Printing functions 3008 3009 #ifndef PRODUCT 3010 3011 void LinearScan::print_timers(double total) { 3012 _total_timer.print(total); 3013 } 3014 3015 void LinearScan::print_statistics() { 3016 _stat_before_alloc.print("before allocation"); 3017 _stat_after_asign.print("after assignment of register"); 3018 _stat_final.print("after optimization"); 3019 } 3020 3021 void LinearScan::print_bitmap(BitMap& b) { 3022 for (unsigned int i = 0; i < b.size(); i++) { 3023 if (b.at(i)) tty->print("%d ", i); 3024 } 3025 tty->cr(); 3026 } 3027 3028 void LinearScan::print_intervals(const char* label) { 3029 if (TraceLinearScanLevel >= 1) { 3030 int i; 3031 tty->cr(); 3032 tty->print_cr("%s", label); 3033 3034 for (i = 0; i < interval_count(); i++) { 3035 Interval* interval = interval_at(i); 3036 if (interval != NULL) { 3037 interval->print(); 3038 } 3039 } 3040 3041 tty->cr(); 3042 tty->print_cr("--- Basic Blocks ---"); 3043 for (i = 0; i < block_count(); i++) { 3044 BlockBegin* block = block_at(i); 3045 tty->print("B%d [%d, %d, %d, %d] ", block->block_id(), block->first_lir_instruction_id(), block->last_lir_instruction_id(), block->loop_index(), block->loop_depth()); 3046 } 3047 tty->cr(); 3048 tty->cr(); 3049 } 3050 3051 if (PrintCFGToFile) { 3052 CFGPrinter::print_intervals(&_intervals, label); 3053 } 3054 } 3055 3056 void LinearScan::print_lir(int level, const char* label, bool hir_valid) { 3057 if (TraceLinearScanLevel >= level) { 3058 tty->cr(); 3059 tty->print_cr("%s", label); 3060 print_LIR(ir()->linear_scan_order()); 3061 tty->cr(); 3062 } 3063 3064 if (level == 1 && PrintCFGToFile) { 3065 CFGPrinter::print_cfg(ir()->linear_scan_order(), label, hir_valid, true); 3066 } 3067 } 3068 3069 #endif //PRODUCT 3070 3071 3072 // ********** verification functions for allocation 3073 // (check that all intervals have a correct register and that no registers are overwritten) 3074 #ifdef ASSERT 3075 3076 void LinearScan::verify() { 3077 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying intervals ******************************************")); 3078 verify_intervals(); 3079 3080 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that no oops are in fixed intervals ****************")); 3081 verify_no_oops_in_fixed_intervals(); 3082 3083 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that unpinned constants are not alive across block boundaries")); 3084 verify_constants(); 3085 3086 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying register allocation ********************************")); 3087 verify_registers(); 3088 3089 TRACE_LINEAR_SCAN(2, tty->print_cr("********* no errors found **********************************************")); 3090 } 3091 3092 void LinearScan::verify_intervals() { 3093 int len = interval_count(); 3094 bool has_error = false; 3095 3096 for (int i = 0; i < len; i++) { 3097 Interval* i1 = interval_at(i); 3098 if (i1 == NULL) continue; 3099 3100 i1->check_split_children(); 3101 3102 if (i1->reg_num() != i) { 3103 tty->print_cr("Interval %d is on position %d in list", i1->reg_num(), i); i1->print(); tty->cr(); 3104 has_error = true; 3105 } 3106 3107 if (i1->reg_num() >= LIR_OprDesc::vreg_base && i1->type() == T_ILLEGAL) { 3108 tty->print_cr("Interval %d has no type assigned", i1->reg_num()); i1->print(); tty->cr(); 3109 has_error = true; 3110 } 3111 3112 if (i1->assigned_reg() == any_reg) { 3113 tty->print_cr("Interval %d has no register assigned", i1->reg_num()); i1->print(); tty->cr(); 3114 has_error = true; 3115 } 3116 3117 if (i1->assigned_reg() == i1->assigned_regHi()) { 3118 tty->print_cr("Interval %d: low and high register equal", i1->reg_num()); i1->print(); tty->cr(); 3119 has_error = true; 3120 } 3121 3122 if (!is_processed_reg_num(i1->assigned_reg())) { 3123 tty->print_cr("Can not have an Interval for an ignored register"); i1->print(); tty->cr(); 3124 has_error = true; 3125 } 3126 3127 if (i1->first() == Range::end()) { 3128 tty->print_cr("Interval %d has no Range", i1->reg_num()); i1->print(); tty->cr(); 3129 has_error = true; 3130 } 3131 3132 for (Range* r = i1->first(); r != Range::end(); r = r->next()) { 3133 if (r->from() >= r->to()) { 3134 tty->print_cr("Interval %d has zero length range", i1->reg_num()); i1->print(); tty->cr(); 3135 has_error = true; 3136 } 3137 } 3138 3139 for (int j = i + 1; j < len; j++) { 3140 Interval* i2 = interval_at(j); 3141 if (i2 == NULL) continue; 3142 3143 // special intervals that are created in MoveResolver 3144 // -> ignore them because the range information has no meaning there 3145 if (i1->from() == 1 && i1->to() == 2) continue; 3146 if (i2->from() == 1 && i2->to() == 2) continue; 3147 3148 int r1 = i1->assigned_reg(); 3149 int r1Hi = i1->assigned_regHi(); 3150 int r2 = i2->assigned_reg(); 3151 int r2Hi = i2->assigned_regHi(); 3152 if (i1->intersects(i2) && (r1 == r2 || r1 == r2Hi || (r1Hi != any_reg && (r1Hi == r2 || r1Hi == r2Hi)))) { 3153 tty->print_cr("Intervals %d and %d overlap and have the same register assigned", i1->reg_num(), i2->reg_num()); 3154 i1->print(); tty->cr(); 3155 i2->print(); tty->cr(); 3156 has_error = true; 3157 } 3158 } 3159 } 3160 3161 assert(has_error == false, "register allocation invalid"); 3162 } 3163 3164 3165 void LinearScan::verify_no_oops_in_fixed_intervals() { 3166 Interval* fixed_intervals; 3167 Interval* other_intervals; 3168 create_unhandled_lists(&fixed_intervals, &other_intervals, is_precolored_cpu_interval, NULL); 3169 3170 // to ensure a walking until the last instruction id, add a dummy interval 3171 // with a high operation id 3172 other_intervals = new Interval(any_reg); 3173 other_intervals->add_range(max_jint - 2, max_jint - 1); 3174 IntervalWalker* iw = new IntervalWalker(this, fixed_intervals, other_intervals); 3175 3176 LIR_OpVisitState visitor; 3177 for (int i = 0; i < block_count(); i++) { 3178 BlockBegin* block = block_at(i); 3179 3180 LIR_OpList* instructions = block->lir()->instructions_list(); 3181 3182 for (int j = 0; j < instructions->length(); j++) { 3183 LIR_Op* op = instructions->at(j); 3184 int op_id = op->id(); 3185 3186 visitor.visit(op); 3187 3188 if (visitor.info_count() > 0) { 3189 iw->walk_before(op->id()); 3190 bool check_live = true; 3191 if (op->code() == lir_move) { 3192 LIR_Op1* move = (LIR_Op1*)op; 3193 check_live = (move->patch_code() == lir_patch_none); 3194 } 3195 LIR_OpBranch* branch = op->as_OpBranch(); 3196 if (branch != NULL && branch->stub() != NULL && branch->stub()->is_exception_throw_stub()) { 3197 // Don't bother checking the stub in this case since the 3198 // exception stub will never return to normal control flow. 3199 check_live = false; 3200 } 3201 3202 // Make sure none of the fixed registers is live across an 3203 // oopmap since we can't handle that correctly. 3204 if (check_live) { 3205 for (Interval* interval = iw->active_first(fixedKind); 3206 interval != Interval::end(); 3207 interval = interval->next()) { 3208 if (interval->current_to() > op->id() + 1) { 3209 // This interval is live out of this op so make sure 3210 // that this interval represents some value that's 3211 // referenced by this op either as an input or output. 3212 bool ok = false; 3213 for_each_visitor_mode(mode) { 3214 int n = visitor.opr_count(mode); 3215 for (int k = 0; k < n; k++) { 3216 LIR_Opr opr = visitor.opr_at(mode, k); 3217 if (opr->is_fixed_cpu()) { 3218 if (interval_at(reg_num(opr)) == interval) { 3219 ok = true; 3220 break; 3221 } 3222 int hi = reg_numHi(opr); 3223 if (hi != -1 && interval_at(hi) == interval) { 3224 ok = true; 3225 break; 3226 } 3227 } 3228 } 3229 } 3230 assert(ok, "fixed intervals should never be live across an oopmap point"); 3231 } 3232 } 3233 } 3234 } 3235 3236 // oop-maps at calls do not contain registers, so check is not needed 3237 if (!visitor.has_call()) { 3238 3239 for_each_visitor_mode(mode) { 3240 int n = visitor.opr_count(mode); 3241 for (int k = 0; k < n; k++) { 3242 LIR_Opr opr = visitor.opr_at(mode, k); 3243 3244 if (opr->is_fixed_cpu() && opr->is_oop()) { 3245 // operand is a non-virtual cpu register and contains an oop 3246 TRACE_LINEAR_SCAN(4, op->print_on(tty); tty->print("checking operand "); opr->print(); tty->cr()); 3247 3248 Interval* interval = interval_at(reg_num(opr)); 3249 assert(interval != NULL, "no interval"); 3250 3251 if (mode == LIR_OpVisitState::inputMode) { 3252 if (interval->to() >= op_id + 1) { 3253 assert(interval->to() < op_id + 2 || 3254 interval->has_hole_between(op_id, op_id + 2), 3255 "oop input operand live after instruction"); 3256 } 3257 } else if (mode == LIR_OpVisitState::outputMode) { 3258 if (interval->from() <= op_id - 1) { 3259 assert(interval->has_hole_between(op_id - 1, op_id), 3260 "oop input operand live after instruction"); 3261 } 3262 } 3263 } 3264 } 3265 } 3266 } 3267 } 3268 } 3269 } 3270 3271 3272 void LinearScan::verify_constants() { 3273 int num_regs = num_virtual_regs(); 3274 int size = live_set_size(); 3275 int num_blocks = block_count(); 3276 3277 for (int i = 0; i < num_blocks; i++) { 3278 BlockBegin* block = block_at(i); 3279 BitMap live_at_edge = block->live_in(); 3280 3281 // visit all registers where the live_at_edge bit is set 3282 for (int r = (int)live_at_edge.get_next_one_offset(0, size); r < size; r = (int)live_at_edge.get_next_one_offset(r + 1, size)) { 3283 TRACE_LINEAR_SCAN(4, tty->print("checking interval %d of block B%d", r, block->block_id())); 3284 3285 Value value = gen()->instruction_for_vreg(r); 3286 3287 assert(value != NULL, "all intervals live across block boundaries must have Value"); 3288 assert(value->operand()->is_register() && value->operand()->is_virtual(), "value must have virtual operand"); 3289 assert(value->operand()->vreg_number() == r, "register number must match"); 3290 // TKR assert(value->as_Constant() == NULL || value->is_pinned(), "only pinned constants can be alive accross block boundaries"); 3291 } 3292 } 3293 } 3294 3295 3296 class RegisterVerifier: public StackObj { 3297 private: 3298 LinearScan* _allocator; 3299 BlockList _work_list; // all blocks that must be processed 3300 IntervalsList _saved_states; // saved information of previous check 3301 3302 // simplified access to methods of LinearScan 3303 Compilation* compilation() const { return _allocator->compilation(); } 3304 Interval* interval_at(int reg_num) const { return _allocator->interval_at(reg_num); } 3305 int reg_num(LIR_Opr opr) const { return _allocator->reg_num(opr); } 3306 3307 // currently, only registers are processed 3308 int state_size() { return LinearScan::nof_regs; } 3309 3310 // accessors 3311 IntervalList* state_for_block(BlockBegin* block) { return _saved_states.at(block->block_id()); } 3312 void set_state_for_block(BlockBegin* block, IntervalList* saved_state) { _saved_states.at_put(block->block_id(), saved_state); } 3313 void add_to_work_list(BlockBegin* block) { if (!_work_list.contains(block)) _work_list.append(block); } 3314 3315 // helper functions 3316 IntervalList* copy(IntervalList* input_state); 3317 void state_put(IntervalList* input_state, int reg, Interval* interval); 3318 bool check_state(IntervalList* input_state, int reg, Interval* interval); 3319 3320 void process_block(BlockBegin* block); 3321 void process_xhandler(XHandler* xhandler, IntervalList* input_state); 3322 void process_successor(BlockBegin* block, IntervalList* input_state); 3323 void process_operations(LIR_List* ops, IntervalList* input_state); 3324 3325 public: 3326 RegisterVerifier(LinearScan* allocator) 3327 : _allocator(allocator) 3328 , _work_list(16) 3329 , _saved_states(BlockBegin::number_of_blocks(), NULL) 3330 { } 3331 3332 void verify(BlockBegin* start); 3333 }; 3334 3335 3336 // entry function from LinearScan that starts the verification 3337 void LinearScan::verify_registers() { 3338 RegisterVerifier verifier(this); 3339 verifier.verify(block_at(0)); 3340 } 3341 3342 3343 void RegisterVerifier::verify(BlockBegin* start) { 3344 // setup input registers (method arguments) for first block 3345 IntervalList* input_state = new IntervalList(state_size(), NULL); 3346 CallingConvention* args = compilation()->frame_map()->incoming_arguments(); 3347 for (int n = 0; n < args->length(); n++) { 3348 LIR_Opr opr = args->at(n); 3349 if (opr->is_register()) { 3350 Interval* interval = interval_at(reg_num(opr)); 3351 3352 if (interval->assigned_reg() < state_size()) { 3353 input_state->at_put(interval->assigned_reg(), interval); 3354 } 3355 if (interval->assigned_regHi() != LinearScan::any_reg && interval->assigned_regHi() < state_size()) { 3356 input_state->at_put(interval->assigned_regHi(), interval); 3357 } 3358 } 3359 } 3360 3361 set_state_for_block(start, input_state); 3362 add_to_work_list(start); 3363 3364 // main loop for verification 3365 do { 3366 BlockBegin* block = _work_list.at(0); 3367 _work_list.remove_at(0); 3368 3369 process_block(block); 3370 } while (!_work_list.is_empty()); 3371 } 3372 3373 void RegisterVerifier::process_block(BlockBegin* block) { 3374 TRACE_LINEAR_SCAN(2, tty->cr(); tty->print_cr("process_block B%d", block->block_id())); 3375 3376 // must copy state because it is modified 3377 IntervalList* input_state = copy(state_for_block(block)); 3378 3379 if (TraceLinearScanLevel >= 4) { 3380 tty->print_cr("Input-State of intervals:"); 3381 tty->print(" "); 3382 for (int i = 0; i < state_size(); i++) { 3383 if (input_state->at(i) != NULL) { 3384 tty->print(" %4d", input_state->at(i)->reg_num()); 3385 } else { 3386 tty->print(" __"); 3387 } 3388 } 3389 tty->cr(); 3390 tty->cr(); 3391 } 3392 3393 // process all operations of the block 3394 process_operations(block->lir(), input_state); 3395 3396 // iterate all successors 3397 for (int i = 0; i < block->number_of_sux(); i++) { 3398 process_successor(block->sux_at(i), input_state); 3399 } 3400 } 3401 3402 void RegisterVerifier::process_xhandler(XHandler* xhandler, IntervalList* input_state) { 3403 TRACE_LINEAR_SCAN(2, tty->print_cr("process_xhandler B%d", xhandler->entry_block()->block_id())); 3404 3405 // must copy state because it is modified 3406 input_state = copy(input_state); 3407 3408 if (xhandler->entry_code() != NULL) { 3409 process_operations(xhandler->entry_code(), input_state); 3410 } 3411 process_successor(xhandler->entry_block(), input_state); 3412 } 3413 3414 void RegisterVerifier::process_successor(BlockBegin* block, IntervalList* input_state) { 3415 IntervalList* saved_state = state_for_block(block); 3416 3417 if (saved_state != NULL) { 3418 // this block was already processed before. 3419 // check if new input_state is consistent with saved_state 3420 3421 bool saved_state_correct = true; 3422 for (int i = 0; i < state_size(); i++) { 3423 if (input_state->at(i) != saved_state->at(i)) { 3424 // current input_state and previous saved_state assume a different 3425 // interval in this register -> assume that this register is invalid 3426 if (saved_state->at(i) != NULL) { 3427 // invalidate old calculation only if it assumed that 3428 // register was valid. when the register was already invalid, 3429 // then the old calculation was correct. 3430 saved_state_correct = false; 3431 saved_state->at_put(i, NULL); 3432 3433 TRACE_LINEAR_SCAN(4, tty->print_cr("process_successor B%d: invalidating slot %d", block->block_id(), i)); 3434 } 3435 } 3436 } 3437 3438 if (saved_state_correct) { 3439 // already processed block with correct input_state 3440 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: previous visit already correct", block->block_id())); 3441 } else { 3442 // must re-visit this block 3443 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: must re-visit because input state changed", block->block_id())); 3444 add_to_work_list(block); 3445 } 3446 3447 } else { 3448 // block was not processed before, so set initial input_state 3449 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: initial visit", block->block_id())); 3450 3451 set_state_for_block(block, copy(input_state)); 3452 add_to_work_list(block); 3453 } 3454 } 3455 3456 3457 IntervalList* RegisterVerifier::copy(IntervalList* input_state) { 3458 IntervalList* copy_state = new IntervalList(input_state->length()); 3459 copy_state->push_all(input_state); 3460 return copy_state; 3461 } 3462 3463 void RegisterVerifier::state_put(IntervalList* input_state, int reg, Interval* interval) { 3464 if (reg != LinearScan::any_reg && reg < state_size()) { 3465 if (interval != NULL) { 3466 TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = %d", reg, interval->reg_num())); 3467 } else if (input_state->at(reg) != NULL) { 3468 TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = NULL", reg)); 3469 } 3470 3471 input_state->at_put(reg, interval); 3472 } 3473 } 3474 3475 bool RegisterVerifier::check_state(IntervalList* input_state, int reg, Interval* interval) { 3476 if (reg != LinearScan::any_reg && reg < state_size()) { 3477 if (input_state->at(reg) != interval) { 3478 tty->print_cr("!! Error in register allocation: register %d does not contain interval %d", reg, interval->reg_num()); 3479 return true; 3480 } 3481 } 3482 return false; 3483 } 3484 3485 void RegisterVerifier::process_operations(LIR_List* ops, IntervalList* input_state) { 3486 // visit all instructions of the block 3487 LIR_OpVisitState visitor; 3488 bool has_error = false; 3489 3490 for (int i = 0; i < ops->length(); i++) { 3491 LIR_Op* op = ops->at(i); 3492 visitor.visit(op); 3493 3494 TRACE_LINEAR_SCAN(4, op->print_on(tty)); 3495 3496 // check if input operands are correct 3497 int j; 3498 int n = visitor.opr_count(LIR_OpVisitState::inputMode); 3499 for (j = 0; j < n; j++) { 3500 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, j); 3501 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { 3502 Interval* interval = interval_at(reg_num(opr)); 3503 if (op->id() != -1) { 3504 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::inputMode); 3505 } 3506 3507 has_error |= check_state(input_state, interval->assigned_reg(), interval->split_parent()); 3508 has_error |= check_state(input_state, interval->assigned_regHi(), interval->split_parent()); 3509 3510 // When an operand is marked with is_last_use, then the fpu stack allocator 3511 // removes the register from the fpu stack -> the register contains no value 3512 if (opr->is_last_use()) { 3513 state_put(input_state, interval->assigned_reg(), NULL); 3514 state_put(input_state, interval->assigned_regHi(), NULL); 3515 } 3516 } 3517 } 3518 3519 // invalidate all caller save registers at calls 3520 if (visitor.has_call()) { 3521 for (j = 0; j < FrameMap::nof_caller_save_cpu_regs; j++) { 3522 state_put(input_state, reg_num(FrameMap::caller_save_cpu_reg_at(j)), NULL); 3523 } 3524 for (j = 0; j < FrameMap::nof_caller_save_fpu_regs; j++) { 3525 state_put(input_state, reg_num(FrameMap::caller_save_fpu_reg_at(j)), NULL); 3526 } 3527 3528 #ifdef X86 3529 for (j = 0; j < FrameMap::nof_caller_save_xmm_regs; j++) { 3530 state_put(input_state, reg_num(FrameMap::caller_save_xmm_reg_at(j)), NULL); 3531 } 3532 #endif 3533 } 3534 3535 // process xhandler before output and temp operands 3536 XHandlers* xhandlers = visitor.all_xhandler(); 3537 n = xhandlers->length(); 3538 for (int k = 0; k < n; k++) { 3539 process_xhandler(xhandlers->handler_at(k), input_state); 3540 } 3541 3542 // set temp operands (some operations use temp operands also as output operands, so can't set them NULL) 3543 n = visitor.opr_count(LIR_OpVisitState::tempMode); 3544 for (j = 0; j < n; j++) { 3545 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, j); 3546 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { 3547 Interval* interval = interval_at(reg_num(opr)); 3548 if (op->id() != -1) { 3549 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::tempMode); 3550 } 3551 3552 state_put(input_state, interval->assigned_reg(), interval->split_parent()); 3553 state_put(input_state, interval->assigned_regHi(), interval->split_parent()); 3554 } 3555 } 3556 3557 // set output operands 3558 n = visitor.opr_count(LIR_OpVisitState::outputMode); 3559 for (j = 0; j < n; j++) { 3560 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, j); 3561 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { 3562 Interval* interval = interval_at(reg_num(opr)); 3563 if (op->id() != -1) { 3564 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::outputMode); 3565 } 3566 3567 state_put(input_state, interval->assigned_reg(), interval->split_parent()); 3568 state_put(input_state, interval->assigned_regHi(), interval->split_parent()); 3569 } 3570 } 3571 } 3572 assert(has_error == false, "Error in register allocation"); 3573 } 3574 3575 #endif // ASSERT 3576 3577 3578 3579 // **** Implementation of MoveResolver ****************************** 3580 3581 MoveResolver::MoveResolver(LinearScan* allocator) : 3582 _allocator(allocator), 3583 _multiple_reads_allowed(false), 3584 _mapping_from(8), 3585 _mapping_from_opr(8), 3586 _mapping_to(8), 3587 _insert_list(NULL), 3588 _insert_idx(-1), 3589 _insertion_buffer() 3590 { 3591 for (int i = 0; i < LinearScan::nof_regs; i++) { 3592 _register_blocked[i] = 0; 3593 } 3594 DEBUG_ONLY(check_empty()); 3595 } 3596 3597 3598 #ifdef ASSERT 3599 3600 void MoveResolver::check_empty() { 3601 assert(_mapping_from.length() == 0 && _mapping_from_opr.length() == 0 && _mapping_to.length() == 0, "list must be empty before and after processing"); 3602 for (int i = 0; i < LinearScan::nof_regs; i++) { 3603 assert(register_blocked(i) == 0, "register map must be empty before and after processing"); 3604 } 3605 assert(_multiple_reads_allowed == false, "must have default value"); 3606 } 3607 3608 void MoveResolver::verify_before_resolve() { 3609 assert(_mapping_from.length() == _mapping_from_opr.length(), "length must be equal"); 3610 assert(_mapping_from.length() == _mapping_to.length(), "length must be equal"); 3611 assert(_insert_list != NULL && _insert_idx != -1, "insert position not set"); 3612 3613 int i, j; 3614 if (!_multiple_reads_allowed) { 3615 for (i = 0; i < _mapping_from.length(); i++) { 3616 for (j = i + 1; j < _mapping_from.length(); j++) { 3617 assert(_mapping_from.at(i) == NULL || _mapping_from.at(i) != _mapping_from.at(j), "cannot read from same interval twice"); 3618 } 3619 } 3620 } 3621 3622 for (i = 0; i < _mapping_to.length(); i++) { 3623 for (j = i + 1; j < _mapping_to.length(); j++) { 3624 assert(_mapping_to.at(i) != _mapping_to.at(j), "cannot write to same interval twice"); 3625 } 3626 } 3627 3628 3629 BitMap used_regs(LinearScan::nof_regs + allocator()->frame_map()->argcount() + allocator()->max_spills()); 3630 used_regs.clear(); 3631 if (!_multiple_reads_allowed) { 3632 for (i = 0; i < _mapping_from.length(); i++) { 3633 Interval* it = _mapping_from.at(i); 3634 if (it != NULL) { 3635 assert(!used_regs.at(it->assigned_reg()), "cannot read from same register twice"); 3636 used_regs.set_bit(it->assigned_reg()); 3637 3638 if (it->assigned_regHi() != LinearScan::any_reg) { 3639 assert(!used_regs.at(it->assigned_regHi()), "cannot read from same register twice"); 3640 used_regs.set_bit(it->assigned_regHi()); 3641 } 3642 } 3643 } 3644 } 3645 3646 used_regs.clear(); 3647 for (i = 0; i < _mapping_to.length(); i++) { 3648 Interval* it = _mapping_to.at(i); 3649 assert(!used_regs.at(it->assigned_reg()), "cannot write to same register twice"); 3650 used_regs.set_bit(it->assigned_reg()); 3651 3652 if (it->assigned_regHi() != LinearScan::any_reg) { 3653 assert(!used_regs.at(it->assigned_regHi()), "cannot write to same register twice"); 3654 used_regs.set_bit(it->assigned_regHi()); 3655 } 3656 } 3657 3658 used_regs.clear(); 3659 for (i = 0; i < _mapping_from.length(); i++) { 3660 Interval* it = _mapping_from.at(i); 3661 if (it != NULL && it->assigned_reg() >= LinearScan::nof_regs) { 3662 used_regs.set_bit(it->assigned_reg()); 3663 } 3664 } 3665 for (i = 0; i < _mapping_to.length(); i++) { 3666 Interval* it = _mapping_to.at(i); 3667 assert(!used_regs.at(it->assigned_reg()) || it->assigned_reg() == _mapping_from.at(i)->assigned_reg(), "stack slots used in _mapping_from must be disjoint to _mapping_to"); 3668 } 3669 } 3670 3671 #endif // ASSERT 3672 3673 3674 // mark assigned_reg and assigned_regHi of the interval as blocked 3675 void MoveResolver::block_registers(Interval* it) { 3676 int reg = it->assigned_reg(); 3677 if (reg < LinearScan::nof_regs) { 3678 assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used"); 3679 set_register_blocked(reg, 1); 3680 } 3681 reg = it->assigned_regHi(); 3682 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { 3683 assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used"); 3684 set_register_blocked(reg, 1); 3685 } 3686 } 3687 3688 // mark assigned_reg and assigned_regHi of the interval as unblocked 3689 void MoveResolver::unblock_registers(Interval* it) { 3690 int reg = it->assigned_reg(); 3691 if (reg < LinearScan::nof_regs) { 3692 assert(register_blocked(reg) > 0, "register already marked as unused"); 3693 set_register_blocked(reg, -1); 3694 } 3695 reg = it->assigned_regHi(); 3696 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { 3697 assert(register_blocked(reg) > 0, "register already marked as unused"); 3698 set_register_blocked(reg, -1); 3699 } 3700 } 3701 3702 // check if assigned_reg and assigned_regHi of the to-interval are not blocked (or only blocked by from) 3703 bool MoveResolver::save_to_process_move(Interval* from, Interval* to) { 3704 int from_reg = -1; 3705 int from_regHi = -1; 3706 if (from != NULL) { 3707 from_reg = from->assigned_reg(); 3708 from_regHi = from->assigned_regHi(); 3709 } 3710 3711 int reg = to->assigned_reg(); 3712 if (reg < LinearScan::nof_regs) { 3713 if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) { 3714 return false; 3715 } 3716 } 3717 reg = to->assigned_regHi(); 3718 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { 3719 if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) { 3720 return false; 3721 } 3722 } 3723 3724 return true; 3725 } 3726 3727 3728 void MoveResolver::create_insertion_buffer(LIR_List* list) { 3729 assert(!_insertion_buffer.initialized(), "overwriting existing buffer"); 3730 _insertion_buffer.init(list); 3731 } 3732 3733 void MoveResolver::append_insertion_buffer() { 3734 if (_insertion_buffer.initialized()) { 3735 _insertion_buffer.lir_list()->append(&_insertion_buffer); 3736 } 3737 assert(!_insertion_buffer.initialized(), "must be uninitialized now"); 3738 3739 _insert_list = NULL; 3740 _insert_idx = -1; 3741 } 3742 3743 void MoveResolver::insert_move(Interval* from_interval, Interval* to_interval) { 3744 assert(from_interval->reg_num() != to_interval->reg_num(), "from and to interval equal"); 3745 assert(from_interval->type() == to_interval->type(), "move between different types"); 3746 assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first"); 3747 assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer"); 3748 3749 LIR_Opr from_opr = LIR_OprFact::virtual_register(from_interval->reg_num(), from_interval->type()); 3750 LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type()); 3751 3752 if (!_multiple_reads_allowed) { 3753 // the last_use flag is an optimization for FPU stack allocation. When the same 3754 // input interval is used in more than one move, then it is too difficult to determine 3755 // if this move is really the last use. 3756 from_opr = from_opr->make_last_use(); 3757 } 3758 _insertion_buffer.move(_insert_idx, from_opr, to_opr); 3759 3760 TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: inserted move from register %d (%d, %d) to %d (%d, %d)", from_interval->reg_num(), from_interval->assigned_reg(), from_interval->assigned_regHi(), to_interval->reg_num(), to_interval->assigned_reg(), to_interval->assigned_regHi())); 3761 } 3762 3763 void MoveResolver::insert_move(LIR_Opr from_opr, Interval* to_interval) { 3764 assert(from_opr->type() == to_interval->type(), "move between different types"); 3765 assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first"); 3766 assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer"); 3767 3768 LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type()); 3769 _insertion_buffer.move(_insert_idx, from_opr, to_opr); 3770 3771 TRACE_LINEAR_SCAN(4, tty->print("MoveResolver: inserted move from constant "); from_opr->print(); tty->print_cr(" to %d (%d, %d)", to_interval->reg_num(), to_interval->assigned_reg(), to_interval->assigned_regHi())); 3772 } 3773 3774 3775 void MoveResolver::resolve_mappings() { 3776 TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: resolving mappings for Block B%d, index %d", _insert_list->block() != NULL ? _insert_list->block()->block_id() : -1, _insert_idx)); 3777 DEBUG_ONLY(verify_before_resolve()); 3778 3779 // Block all registers that are used as input operands of a move. 3780 // When a register is blocked, no move to this register is emitted. 3781 // This is necessary for detecting cycles in moves. 3782 int i; 3783 for (i = _mapping_from.length() - 1; i >= 0; i--) { 3784 Interval* from_interval = _mapping_from.at(i); 3785 if (from_interval != NULL) { 3786 block_registers(from_interval); 3787 } 3788 } 3789 3790 int spill_candidate = -1; 3791 while (_mapping_from.length() > 0) { 3792 bool processed_interval = false; 3793 3794 for (i = _mapping_from.length() - 1; i >= 0; i--) { 3795 Interval* from_interval = _mapping_from.at(i); 3796 Interval* to_interval = _mapping_to.at(i); 3797 3798 if (save_to_process_move(from_interval, to_interval)) { 3799 // this inverval can be processed because target is free 3800 if (from_interval != NULL) { 3801 insert_move(from_interval, to_interval); 3802 unblock_registers(from_interval); 3803 } else { 3804 insert_move(_mapping_from_opr.at(i), to_interval); 3805 } 3806 _mapping_from.remove_at(i); 3807 _mapping_from_opr.remove_at(i); 3808 _mapping_to.remove_at(i); 3809 3810 processed_interval = true; 3811 } else if (from_interval != NULL && from_interval->assigned_reg() < LinearScan::nof_regs) { 3812 // this interval cannot be processed now because target is not free 3813 // it starts in a register, so it is a possible candidate for spilling 3814 spill_candidate = i; 3815 } 3816 } 3817 3818 if (!processed_interval) { 3819 // no move could be processed because there is a cycle in the move list 3820 // (e.g. r1 -> r2, r2 -> r1), so one interval must be spilled to memory 3821 assert(spill_candidate != -1, "no interval in register for spilling found"); 3822 3823 // create a new spill interval and assign a stack slot to it 3824 Interval* from_interval = _mapping_from.at(spill_candidate); 3825 Interval* spill_interval = new Interval(-1); 3826 spill_interval->set_type(from_interval->type()); 3827 3828 // add a dummy range because real position is difficult to calculate 3829 // Note: this range is a special case when the integrity of the allocation is checked 3830 spill_interval->add_range(1, 2); 3831 3832 // do not allocate a new spill slot for temporary interval, but 3833 // use spill slot assigned to from_interval. Otherwise moves from 3834 // one stack slot to another can happen (not allowed by LIR_Assembler 3835 int spill_slot = from_interval->canonical_spill_slot(); 3836 if (spill_slot < 0) { 3837 spill_slot = allocator()->allocate_spill_slot(type2spill_size[spill_interval->type()] == 2); 3838 from_interval->set_canonical_spill_slot(spill_slot); 3839 } 3840 spill_interval->assign_reg(spill_slot); 3841 allocator()->append_interval(spill_interval); 3842 3843 TRACE_LINEAR_SCAN(4, tty->print_cr("created new Interval %d for spilling", spill_interval->reg_num())); 3844 3845 // insert a move from register to stack and update the mapping 3846 insert_move(from_interval, spill_interval); 3847 _mapping_from.at_put(spill_candidate, spill_interval); 3848 unblock_registers(from_interval); 3849 } 3850 } 3851 3852 // reset to default value 3853 _multiple_reads_allowed = false; 3854 3855 // check that all intervals have been processed 3856 DEBUG_ONLY(check_empty()); 3857 } 3858 3859 3860 void MoveResolver::set_insert_position(LIR_List* insert_list, int insert_idx) { 3861 TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: setting insert position to Block B%d, index %d", insert_list->block() != NULL ? insert_list->block()->block_id() : -1, insert_idx)); 3862 assert(_insert_list == NULL && _insert_idx == -1, "use move_insert_position instead of set_insert_position when data already set"); 3863 3864 create_insertion_buffer(insert_list); 3865 _insert_list = insert_list; 3866 _insert_idx = insert_idx; 3867 } 3868 3869 void MoveResolver::move_insert_position(LIR_List* insert_list, int insert_idx) { 3870 TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: moving insert position to Block B%d, index %d", insert_list->block() != NULL ? insert_list->block()->block_id() : -1, insert_idx)); 3871 3872 if (_insert_list != NULL && (insert_list != _insert_list || insert_idx != _insert_idx)) { 3873 // insert position changed -> resolve current mappings 3874 resolve_mappings(); 3875 } 3876 3877 if (insert_list != _insert_list) { 3878 // block changed -> append insertion_buffer because it is 3879 // bound to a specific block and create a new insertion_buffer 3880 append_insertion_buffer(); 3881 create_insertion_buffer(insert_list); 3882 } 3883 3884 _insert_list = insert_list; 3885 _insert_idx = insert_idx; 3886 } 3887 3888 void MoveResolver::add_mapping(Interval* from_interval, Interval* to_interval) { 3889 TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: adding mapping from %d (%d, %d) to %d (%d, %d)", from_interval->reg_num(), from_interval->assigned_reg(), from_interval->assigned_regHi(), to_interval->reg_num(), to_interval->assigned_reg(), to_interval->assigned_regHi())); 3890 3891 _mapping_from.append(from_interval); 3892 _mapping_from_opr.append(LIR_OprFact::illegalOpr); 3893 _mapping_to.append(to_interval); 3894 } 3895 3896 3897 void MoveResolver::add_mapping(LIR_Opr from_opr, Interval* to_interval) { 3898 TRACE_LINEAR_SCAN(4, tty->print("MoveResolver: adding mapping from "); from_opr->print(); tty->print_cr(" to %d (%d, %d)", to_interval->reg_num(), to_interval->assigned_reg(), to_interval->assigned_regHi())); 3899 assert(from_opr->is_constant(), "only for constants"); 3900 3901 _mapping_from.append(NULL); 3902 _mapping_from_opr.append(from_opr); 3903 _mapping_to.append(to_interval); 3904 } 3905 3906 void MoveResolver::resolve_and_append_moves() { 3907 if (has_mappings()) { 3908 resolve_mappings(); 3909 } 3910 append_insertion_buffer(); 3911 } 3912 3913 3914 3915 // **** Implementation of Range ************************************* 3916 3917 Range::Range(int from, int to, Range* next) : 3918 _from(from), 3919 _to(to), 3920 _next(next) 3921 { 3922 } 3923 3924 // initialize sentinel 3925 Range* Range::_end = NULL; 3926 void Range::initialize() { 3927 _end = new Range(max_jint, max_jint, NULL); 3928 } 3929 3930 int Range::intersects_at(Range* r2) const { 3931 const Range* r1 = this; 3932 3933 assert(r1 != NULL && r2 != NULL, "null ranges not allowed"); 3934 assert(r1 != _end && r2 != _end, "empty ranges not allowed"); 3935 3936 do { 3937 if (r1->from() < r2->from()) { 3938 if (r1->to() <= r2->from()) { 3939 r1 = r1->next(); if (r1 == _end) return -1; 3940 } else { 3941 return r2->from(); 3942 } 3943 } else if (r2->from() < r1->from()) { 3944 if (r2->to() <= r1->from()) { 3945 r2 = r2->next(); if (r2 == _end) return -1; 3946 } else { 3947 return r1->from(); 3948 } 3949 } else { // r1->from() == r2->from() 3950 if (r1->from() == r1->to()) { 3951 r1 = r1->next(); if (r1 == _end) return -1; 3952 } else if (r2->from() == r2->to()) { 3953 r2 = r2->next(); if (r2 == _end) return -1; 3954 } else { 3955 return r1->from(); 3956 } 3957 } 3958 } while (true); 3959 } 3960 3961 #ifndef PRODUCT 3962 void Range::print(outputStream* out) const { 3963 out->print("[%d, %d[ ", _from, _to); 3964 } 3965 #endif 3966 3967 3968 3969 // **** Implementation of Interval ********************************** 3970 3971 // initialize sentinel 3972 Interval* Interval::_end = NULL; 3973 void Interval::initialize() { 3974 Range::initialize(); 3975 _end = new Interval(-1); 3976 } 3977 3978 Interval::Interval(int reg_num) : 3979 _reg_num(reg_num), 3980 _type(T_ILLEGAL), 3981 _first(Range::end()), 3982 _use_pos_and_kinds(12), 3983 _current(Range::end()), 3984 _next(_end), 3985 _state(invalidState), 3986 _assigned_reg(LinearScan::any_reg), 3987 _assigned_regHi(LinearScan::any_reg), 3988 _cached_to(-1), 3989 _cached_opr(LIR_OprFact::illegalOpr), 3990 _cached_vm_reg(VMRegImpl::Bad()), 3991 _split_children(0), 3992 _canonical_spill_slot(-1), 3993 _insert_move_when_activated(false), 3994 _register_hint(NULL), 3995 _spill_state(noDefinitionFound), 3996 _spill_definition_pos(-1) 3997 { 3998 _split_parent = this; 3999 _current_split_child = this; 4000 } 4001 4002 int Interval::calc_to() { 4003 assert(_first != Range::end(), "interval has no range"); 4004 4005 Range* r = _first; 4006 while (r->next() != Range::end()) { 4007 r = r->next(); 4008 } 4009 return r->to(); 4010 } 4011 4012 4013 #ifdef ASSERT 4014 // consistency check of split-children 4015 void Interval::check_split_children() { 4016 if (_split_children.length() > 0) { 4017 assert(is_split_parent(), "only split parents can have children"); 4018 4019 for (int i = 0; i < _split_children.length(); i++) { 4020 Interval* i1 = _split_children.at(i); 4021 4022 assert(i1->split_parent() == this, "not a split child of this interval"); 4023 assert(i1->type() == type(), "must be equal for all split children"); 4024 assert(i1->canonical_spill_slot() == canonical_spill_slot(), "must be equal for all split children"); 4025 4026 for (int j = i + 1; j < _split_children.length(); j++) { 4027 Interval* i2 = _split_children.at(j); 4028 4029 assert(i1->reg_num() != i2->reg_num(), "same register number"); 4030 4031 if (i1->from() < i2->from()) { 4032 assert(i1->to() <= i2->from() && i1->to() < i2->to(), "intervals overlapping"); 4033 } else { 4034 assert(i2->from() < i1->from(), "intervals start at same op_id"); 4035 assert(i2->to() <= i1->from() && i2->to() < i1->to(), "intervals overlapping"); 4036 } 4037 } 4038 } 4039 } 4040 } 4041 #endif // ASSERT 4042 4043 Interval* Interval::register_hint(bool search_split_child) const { 4044 if (!search_split_child) { 4045 return _register_hint; 4046 } 4047 4048 if (_register_hint != NULL) { 4049 assert(_register_hint->is_split_parent(), "ony split parents are valid hint registers"); 4050 4051 if (_register_hint->assigned_reg() >= 0 && _register_hint->assigned_reg() < LinearScan::nof_regs) { 4052 return _register_hint; 4053 4054 } else if (_register_hint->_split_children.length() > 0) { 4055 // search the first split child that has a register assigned 4056 int len = _register_hint->_split_children.length(); 4057 for (int i = 0; i < len; i++) { 4058 Interval* cur = _register_hint->_split_children.at(i); 4059 4060 if (cur->assigned_reg() >= 0 && cur->assigned_reg() < LinearScan::nof_regs) { 4061 return cur; 4062 } 4063 } 4064 } 4065 } 4066 4067 // no hint interval found that has a register assigned 4068 return NULL; 4069 } 4070 4071 4072 Interval* Interval::split_child_at_op_id(int op_id, LIR_OpVisitState::OprMode mode) { 4073 assert(is_split_parent(), "can only be called for split parents"); 4074 assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)"); 4075 4076 Interval* result; 4077 if (_split_children.length() == 0) { 4078 result = this; 4079 } else { 4080 result = NULL; 4081 int len = _split_children.length(); 4082 4083 // in outputMode, the end of the interval (op_id == cur->to()) is not valid 4084 int to_offset = (mode == LIR_OpVisitState::outputMode ? 0 : 1); 4085 4086 int i; 4087 for (i = 0; i < len; i++) { 4088 Interval* cur = _split_children.at(i); 4089 if (cur->from() <= op_id && op_id < cur->to() + to_offset) { 4090 if (i > 0) { 4091 // exchange current split child to start of list (faster access for next call) 4092 _split_children.at_put(i, _split_children.at(0)); 4093 _split_children.at_put(0, cur); 4094 } 4095 4096 // interval found 4097 result = cur; 4098 break; 4099 } 4100 } 4101 4102 #ifdef ASSERT 4103 for (i = 0; i < len; i++) { 4104 Interval* tmp = _split_children.at(i); 4105 if (tmp != result && tmp->from() <= op_id && op_id < tmp->to() + to_offset) { 4106 tty->print_cr("two valid result intervals found for op_id %d: %d and %d", op_id, result->reg_num(), tmp->reg_num()); 4107 result->print(); 4108 tmp->print(); 4109 assert(false, "two valid result intervals found"); 4110 } 4111 } 4112 #endif 4113 } 4114 4115 assert(result != NULL, "no matching interval found"); 4116 assert(result->covers(op_id, mode), "op_id not covered by interval"); 4117 4118 return result; 4119 } 4120 4121 4122 // returns the last split child that ends before the given op_id 4123 Interval* Interval::split_child_before_op_id(int op_id) { 4124 assert(op_id >= 0, "invalid op_id"); 4125 4126 Interval* parent = split_parent(); 4127 Interval* result = NULL; 4128 4129 int len = parent->_split_children.length(); 4130 assert(len > 0, "no split children available"); 4131 4132 for (int i = len - 1; i >= 0; i--) { 4133 Interval* cur = parent->_split_children.at(i); 4134 if (cur->to() <= op_id && (result == NULL || result->to() < cur->to())) { 4135 result = cur; 4136 } 4137 } 4138 4139 assert(result != NULL, "no split child found"); 4140 return result; 4141 } 4142 4143 4144 // checks if op_id is covered by any split child 4145 bool Interval::split_child_covers(int op_id, LIR_OpVisitState::OprMode mode) { 4146 assert(is_split_parent(), "can only be called for split parents"); 4147 assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)"); 4148 4149 if (_split_children.length() == 0) { 4150 // simple case if interval was not split 4151 return covers(op_id, mode); 4152 4153 } else { 4154 // extended case: check all split children 4155 int len = _split_children.length(); 4156 for (int i = 0; i < len; i++) { 4157 Interval* cur = _split_children.at(i); 4158 if (cur->covers(op_id, mode)) { 4159 return true; 4160 } 4161 } 4162 return false; 4163 } 4164 } 4165 4166 4167 // Note: use positions are sorted descending -> first use has highest index 4168 int Interval::first_usage(IntervalUseKind min_use_kind) const { 4169 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4170 4171 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4172 if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) { 4173 return _use_pos_and_kinds.at(i); 4174 } 4175 } 4176 return max_jint; 4177 } 4178 4179 int Interval::next_usage(IntervalUseKind min_use_kind, int from) const { 4180 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4181 4182 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4183 if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) >= min_use_kind) { 4184 return _use_pos_and_kinds.at(i); 4185 } 4186 } 4187 return max_jint; 4188 } 4189 4190 int Interval::next_usage_exact(IntervalUseKind exact_use_kind, int from) const { 4191 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4192 4193 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4194 if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) == exact_use_kind) { 4195 return _use_pos_and_kinds.at(i); 4196 } 4197 } 4198 return max_jint; 4199 } 4200 4201 int Interval::previous_usage(IntervalUseKind min_use_kind, int from) const { 4202 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4203 4204 int prev = 0; 4205 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4206 if (_use_pos_and_kinds.at(i) > from) { 4207 return prev; 4208 } 4209 if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) { 4210 prev = _use_pos_and_kinds.at(i); 4211 } 4212 } 4213 return prev; 4214 } 4215 4216 void Interval::add_use_pos(int pos, IntervalUseKind use_kind) { 4217 assert(covers(pos, LIR_OpVisitState::inputMode), "use position not covered by live range"); 4218 4219 // do not add use positions for precolored intervals because 4220 // they are never used 4221 if (use_kind != noUse && reg_num() >= LIR_OprDesc::vreg_base) { 4222 #ifdef ASSERT 4223 assert(_use_pos_and_kinds.length() % 2 == 0, "must be"); 4224 for (int i = 0; i < _use_pos_and_kinds.length(); i += 2) { 4225 assert(pos <= _use_pos_and_kinds.at(i), "already added a use-position with lower position"); 4226 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4227 if (i > 0) { 4228 assert(_use_pos_and_kinds.at(i) < _use_pos_and_kinds.at(i - 2), "not sorted descending"); 4229 } 4230 } 4231 #endif 4232 4233 // Note: add_use is called in descending order, so list gets sorted 4234 // automatically by just appending new use positions 4235 int len = _use_pos_and_kinds.length(); 4236 if (len == 0 || _use_pos_and_kinds.at(len - 2) > pos) { 4237 _use_pos_and_kinds.append(pos); 4238 _use_pos_and_kinds.append(use_kind); 4239 } else if (_use_pos_and_kinds.at(len - 1) < use_kind) { 4240 assert(_use_pos_and_kinds.at(len - 2) == pos, "list not sorted correctly"); 4241 _use_pos_and_kinds.at_put(len - 1, use_kind); 4242 } 4243 } 4244 } 4245 4246 void Interval::add_range(int from, int to) { 4247 assert(from < to, "invalid range"); 4248 assert(first() == Range::end() || to < first()->next()->from(), "not inserting at begin of interval"); 4249 assert(from <= first()->to(), "not inserting at begin of interval"); 4250 4251 if (first()->from() <= to) { 4252 // join intersecting ranges 4253 first()->set_from(MIN2(from, first()->from())); 4254 first()->set_to (MAX2(to, first()->to())); 4255 } else { 4256 // insert new range 4257 _first = new Range(from, to, first()); 4258 } 4259 } 4260 4261 Interval* Interval::new_split_child() { 4262 // allocate new interval 4263 Interval* result = new Interval(-1); 4264 result->set_type(type()); 4265 4266 Interval* parent = split_parent(); 4267 result->_split_parent = parent; 4268 result->set_register_hint(parent); 4269 4270 // insert new interval in children-list of parent 4271 if (parent->_split_children.length() == 0) { 4272 assert(is_split_parent(), "list must be initialized at first split"); 4273 4274 parent->_split_children = IntervalList(4); 4275 parent->_split_children.append(this); 4276 } 4277 parent->_split_children.append(result); 4278 4279 return result; 4280 } 4281 4282 // split this interval at the specified position and return 4283 // the remainder as a new interval. 4284 // 4285 // when an interval is split, a bi-directional link is established between the original interval 4286 // (the split parent) and the intervals that are split off this interval (the split children) 4287 // When a split child is split again, the new created interval is also a direct child 4288 // of the original parent (there is no tree of split children stored, but a flat list) 4289 // All split children are spilled to the same stack slot (stored in _canonical_spill_slot) 4290 // 4291 // Note: The new interval has no valid reg_num 4292 Interval* Interval::split(int split_pos) { 4293 assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals"); 4294 4295 // allocate new interval 4296 Interval* result = new_split_child(); 4297 4298 // split the ranges 4299 Range* prev = NULL; 4300 Range* cur = _first; 4301 while (cur != Range::end() && cur->to() <= split_pos) { 4302 prev = cur; 4303 cur = cur->next(); 4304 } 4305 assert(cur != Range::end(), "split interval after end of last range"); 4306 4307 if (cur->from() < split_pos) { 4308 result->_first = new Range(split_pos, cur->to(), cur->next()); 4309 cur->set_to(split_pos); 4310 cur->set_next(Range::end()); 4311 4312 } else { 4313 assert(prev != NULL, "split before start of first range"); 4314 result->_first = cur; 4315 prev->set_next(Range::end()); 4316 } 4317 result->_current = result->_first; 4318 _cached_to = -1; // clear cached value 4319 4320 // split list of use positions 4321 int total_len = _use_pos_and_kinds.length(); 4322 int start_idx = total_len - 2; 4323 while (start_idx >= 0 && _use_pos_and_kinds.at(start_idx) < split_pos) { 4324 start_idx -= 2; 4325 } 4326 4327 intStack new_use_pos_and_kinds(total_len - start_idx); 4328 int i; 4329 for (i = start_idx + 2; i < total_len; i++) { 4330 new_use_pos_and_kinds.append(_use_pos_and_kinds.at(i)); 4331 } 4332 4333 _use_pos_and_kinds.truncate(start_idx + 2); 4334 result->_use_pos_and_kinds = _use_pos_and_kinds; 4335 _use_pos_and_kinds = new_use_pos_and_kinds; 4336 4337 #ifdef ASSERT 4338 assert(_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos"); 4339 assert(result->_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos"); 4340 assert(_use_pos_and_kinds.length() + result->_use_pos_and_kinds.length() == total_len, "missed some entries"); 4341 4342 for (i = 0; i < _use_pos_and_kinds.length(); i += 2) { 4343 assert(_use_pos_and_kinds.at(i) < split_pos, "must be"); 4344 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4345 } 4346 for (i = 0; i < result->_use_pos_and_kinds.length(); i += 2) { 4347 assert(result->_use_pos_and_kinds.at(i) >= split_pos, "must be"); 4348 assert(result->_use_pos_and_kinds.at(i + 1) >= firstValidKind && result->_use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4349 } 4350 #endif 4351 4352 return result; 4353 } 4354 4355 // split this interval at the specified position and return 4356 // the head as a new interval (the original interval is the tail) 4357 // 4358 // Currently, only the first range can be split, and the new interval 4359 // must not have split positions 4360 Interval* Interval::split_from_start(int split_pos) { 4361 assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals"); 4362 assert(split_pos > from() && split_pos < to(), "can only split inside interval"); 4363 assert(split_pos > _first->from() && split_pos <= _first->to(), "can only split inside first range"); 4364 assert(first_usage(noUse) > split_pos, "can not split when use positions are present"); 4365 4366 // allocate new interval 4367 Interval* result = new_split_child(); 4368 4369 // the new created interval has only one range (checked by assertion above), 4370 // so the splitting of the ranges is very simple 4371 result->add_range(_first->from(), split_pos); 4372 4373 if (split_pos == _first->to()) { 4374 assert(_first->next() != Range::end(), "must not be at end"); 4375 _first = _first->next(); 4376 } else { 4377 _first->set_from(split_pos); 4378 } 4379 4380 return result; 4381 } 4382 4383 4384 // returns true if the op_id is inside the interval 4385 bool Interval::covers(int op_id, LIR_OpVisitState::OprMode mode) const { 4386 Range* cur = _first; 4387 4388 while (cur != Range::end() && cur->to() < op_id) { 4389 cur = cur->next(); 4390 } 4391 if (cur != Range::end()) { 4392 assert(cur->to() != cur->next()->from(), "ranges not separated"); 4393 4394 if (mode == LIR_OpVisitState::outputMode) { 4395 return cur->from() <= op_id && op_id < cur->to(); 4396 } else { 4397 return cur->from() <= op_id && op_id <= cur->to(); 4398 } 4399 } 4400 return false; 4401 } 4402 4403 // returns true if the interval has any hole between hole_from and hole_to 4404 // (even if the hole has only the length 1) 4405 bool Interval::has_hole_between(int hole_from, int hole_to) { 4406 assert(hole_from < hole_to, "check"); 4407 assert(from() <= hole_from && hole_to <= to(), "index out of interval"); 4408 4409 Range* cur = _first; 4410 while (cur != Range::end()) { 4411 assert(cur->to() < cur->next()->from(), "no space between ranges"); 4412 4413 // hole-range starts before this range -> hole 4414 if (hole_from < cur->from()) { 4415 return true; 4416 4417 // hole-range completely inside this range -> no hole 4418 } else if (hole_to <= cur->to()) { 4419 return false; 4420 4421 // overlapping of hole-range with this range -> hole 4422 } else if (hole_from <= cur->to()) { 4423 return true; 4424 } 4425 4426 cur = cur->next(); 4427 } 4428 4429 return false; 4430 } 4431 4432 4433 #ifndef PRODUCT 4434 void Interval::print(outputStream* out) const { 4435 const char* SpillState2Name[] = { "no definition", "no spill store", "one spill store", "store at definition", "start in memory", "no optimization" }; 4436 const char* UseKind2Name[] = { "N", "L", "S", "M" }; 4437 4438 const char* type_name; 4439 LIR_Opr opr = LIR_OprFact::illegal(); 4440 if (reg_num() < LIR_OprDesc::vreg_base) { 4441 type_name = "fixed"; 4442 // need a temporary operand for fixed intervals because type() cannot be called 4443 if (assigned_reg() >= pd_first_cpu_reg && assigned_reg() <= pd_last_cpu_reg) { 4444 opr = LIR_OprFact::single_cpu(assigned_reg()); 4445 } else if (assigned_reg() >= pd_first_fpu_reg && assigned_reg() <= pd_last_fpu_reg) { 4446 opr = LIR_OprFact::single_fpu(assigned_reg() - pd_first_fpu_reg); 4447 #ifdef X86 4448 } else if (assigned_reg() >= pd_first_xmm_reg && assigned_reg() <= pd_last_xmm_reg) { 4449 opr = LIR_OprFact::single_xmm(assigned_reg() - pd_first_xmm_reg); 4450 #endif 4451 } else { 4452 ShouldNotReachHere(); 4453 } 4454 } else { 4455 type_name = type2name(type()); 4456 if (assigned_reg() != -1) { 4457 opr = LinearScan::calc_operand_for_interval(this); 4458 } 4459 } 4460 4461 out->print("%d %s ", reg_num(), type_name); 4462 if (opr->is_valid()) { 4463 out->print("\""); 4464 opr->print(out); 4465 out->print("\" "); 4466 } 4467 out->print("%d %d ", split_parent()->reg_num(), (register_hint(false) != NULL ? register_hint(false)->reg_num() : -1)); 4468 4469 // print ranges 4470 Range* cur = _first; 4471 while (cur != Range::end()) { 4472 cur->print(out); 4473 cur = cur->next(); 4474 assert(cur != NULL, "range list not closed with range sentinel"); 4475 } 4476 4477 // print use positions 4478 int prev = 0; 4479 assert(_use_pos_and_kinds.length() % 2 == 0, "must be"); 4480 for (int i =_use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4481 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4482 assert(prev < _use_pos_and_kinds.at(i), "use positions not sorted"); 4483 4484 out->print("%d %s ", _use_pos_and_kinds.at(i), UseKind2Name[_use_pos_and_kinds.at(i + 1)]); 4485 prev = _use_pos_and_kinds.at(i); 4486 } 4487 4488 out->print(" \"%s\"", SpillState2Name[spill_state()]); 4489 out->cr(); 4490 } 4491 #endif 4492 4493 4494 4495 // **** Implementation of IntervalWalker **************************** 4496 4497 IntervalWalker::IntervalWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first) 4498 : _compilation(allocator->compilation()) 4499 , _allocator(allocator) 4500 { 4501 _unhandled_first[fixedKind] = unhandled_fixed_first; 4502 _unhandled_first[anyKind] = unhandled_any_first; 4503 _active_first[fixedKind] = Interval::end(); 4504 _inactive_first[fixedKind] = Interval::end(); 4505 _active_first[anyKind] = Interval::end(); 4506 _inactive_first[anyKind] = Interval::end(); 4507 _current_position = -1; 4508 _current = NULL; 4509 next_interval(); 4510 } 4511 4512 4513 // append interval at top of list 4514 void IntervalWalker::append_unsorted(Interval** list, Interval* interval) { 4515 interval->set_next(*list); *list = interval; 4516 } 4517 4518 4519 // append interval in order of current range from() 4520 void IntervalWalker::append_sorted(Interval** list, Interval* interval) { 4521 Interval* prev = NULL; 4522 Interval* cur = *list; 4523 while (cur->current_from() < interval->current_from()) { 4524 prev = cur; cur = cur->next(); 4525 } 4526 if (prev == NULL) { 4527 *list = interval; 4528 } else { 4529 prev->set_next(interval); 4530 } 4531 interval->set_next(cur); 4532 } 4533 4534 void IntervalWalker::append_to_unhandled(Interval** list, Interval* interval) { 4535 assert(interval->from() >= current()->current_from(), "cannot append new interval before current walk position"); 4536 4537 Interval* prev = NULL; 4538 Interval* cur = *list; 4539 while (cur->from() < interval->from() || (cur->from() == interval->from() && cur->first_usage(noUse) < interval->first_usage(noUse))) { 4540 prev = cur; cur = cur->next(); 4541 } 4542 if (prev == NULL) { 4543 *list = interval; 4544 } else { 4545 prev->set_next(interval); 4546 } 4547 interval->set_next(cur); 4548 } 4549 4550 4551 inline bool IntervalWalker::remove_from_list(Interval** list, Interval* i) { 4552 while (*list != Interval::end() && *list != i) { 4553 list = (*list)->next_addr(); 4554 } 4555 if (*list != Interval::end()) { 4556 assert(*list == i, "check"); 4557 *list = (*list)->next(); 4558 return true; 4559 } else { 4560 return false; 4561 } 4562 } 4563 4564 void IntervalWalker::remove_from_list(Interval* i) { 4565 bool deleted; 4566 4567 if (i->state() == activeState) { 4568 deleted = remove_from_list(active_first_addr(anyKind), i); 4569 } else { 4570 assert(i->state() == inactiveState, "invalid state"); 4571 deleted = remove_from_list(inactive_first_addr(anyKind), i); 4572 } 4573 4574 assert(deleted, "interval has not been found in list"); 4575 } 4576 4577 4578 void IntervalWalker::walk_to(IntervalState state, int from) { 4579 assert (state == activeState || state == inactiveState, "wrong state"); 4580 for_each_interval_kind(kind) { 4581 Interval** prev = state == activeState ? active_first_addr(kind) : inactive_first_addr(kind); 4582 Interval* next = *prev; 4583 while (next->current_from() <= from) { 4584 Interval* cur = next; 4585 next = cur->next(); 4586 4587 bool range_has_changed = false; 4588 while (cur->current_to() <= from) { 4589 cur->next_range(); 4590 range_has_changed = true; 4591 } 4592 4593 // also handle move from inactive list to active list 4594 range_has_changed = range_has_changed || (state == inactiveState && cur->current_from() <= from); 4595 4596 if (range_has_changed) { 4597 // remove cur from list 4598 *prev = next; 4599 if (cur->current_at_end()) { 4600 // move to handled state (not maintained as a list) 4601 cur->set_state(handledState); 4602 interval_moved(cur, kind, state, handledState); 4603 } else if (cur->current_from() <= from){ 4604 // sort into active list 4605 append_sorted(active_first_addr(kind), cur); 4606 cur->set_state(activeState); 4607 if (*prev == cur) { 4608 assert(state == activeState, "check"); 4609 prev = cur->next_addr(); 4610 } 4611 interval_moved(cur, kind, state, activeState); 4612 } else { 4613 // sort into inactive list 4614 append_sorted(inactive_first_addr(kind), cur); 4615 cur->set_state(inactiveState); 4616 if (*prev == cur) { 4617 assert(state == inactiveState, "check"); 4618 prev = cur->next_addr(); 4619 } 4620 interval_moved(cur, kind, state, inactiveState); 4621 } 4622 } else { 4623 prev = cur->next_addr(); 4624 continue; 4625 } 4626 } 4627 } 4628 } 4629 4630 4631 void IntervalWalker::next_interval() { 4632 IntervalKind kind; 4633 Interval* any = _unhandled_first[anyKind]; 4634 Interval* fixed = _unhandled_first[fixedKind]; 4635 4636 if (any != Interval::end()) { 4637 // intervals may start at same position -> prefer fixed interval 4638 kind = fixed != Interval::end() && fixed->from() <= any->from() ? fixedKind : anyKind; 4639 4640 assert (kind == fixedKind && fixed->from() <= any->from() || 4641 kind == anyKind && any->from() <= fixed->from(), "wrong interval!!!"); 4642 assert(any == Interval::end() || fixed == Interval::end() || any->from() != fixed->from() || kind == fixedKind, "if fixed and any-Interval start at same position, fixed must be processed first"); 4643 4644 } else if (fixed != Interval::end()) { 4645 kind = fixedKind; 4646 } else { 4647 _current = NULL; return; 4648 } 4649 _current_kind = kind; 4650 _current = _unhandled_first[kind]; 4651 _unhandled_first[kind] = _current->next(); 4652 _current->set_next(Interval::end()); 4653 _current->rewind_range(); 4654 } 4655 4656 4657 void IntervalWalker::walk_to(int lir_op_id) { 4658 assert(_current_position <= lir_op_id, "can not walk backwards"); 4659 while (current() != NULL) { 4660 bool is_active = current()->from() <= lir_op_id; 4661 int id = is_active ? current()->from() : lir_op_id; 4662 4663 TRACE_LINEAR_SCAN(2, if (_current_position < id) { tty->cr(); tty->print_cr("walk_to(%d) **************************************************************", id); }) 4664 4665 // set _current_position prior to call of walk_to 4666 _current_position = id; 4667 4668 // call walk_to even if _current_position == id 4669 walk_to(activeState, id); 4670 walk_to(inactiveState, id); 4671 4672 if (is_active) { 4673 current()->set_state(activeState); 4674 if (activate_current()) { 4675 append_sorted(active_first_addr(current_kind()), current()); 4676 interval_moved(current(), current_kind(), unhandledState, activeState); 4677 } 4678 4679 next_interval(); 4680 } else { 4681 return; 4682 } 4683 } 4684 } 4685 4686 void IntervalWalker::interval_moved(Interval* interval, IntervalKind kind, IntervalState from, IntervalState to) { 4687 #ifndef PRODUCT 4688 if (TraceLinearScanLevel >= 4) { 4689 #define print_state(state) \ 4690 switch(state) {\ 4691 case unhandledState: tty->print("unhandled"); break;\ 4692 case activeState: tty->print("active"); break;\ 4693 case inactiveState: tty->print("inactive"); break;\ 4694 case handledState: tty->print("handled"); break;\ 4695 default: ShouldNotReachHere(); \ 4696 } 4697 4698 print_state(from); tty->print(" to "); print_state(to); 4699 tty->fill_to(23); 4700 interval->print(); 4701 4702 #undef print_state 4703 } 4704 #endif 4705 } 4706 4707 4708 4709 // **** Implementation of LinearScanWalker ************************** 4710 4711 LinearScanWalker::LinearScanWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first) 4712 : IntervalWalker(allocator, unhandled_fixed_first, unhandled_any_first) 4713 , _move_resolver(allocator) 4714 { 4715 for (int i = 0; i < LinearScan::nof_regs; i++) { 4716 _spill_intervals[i] = new IntervalList(2); 4717 } 4718 } 4719 4720 4721 inline void LinearScanWalker::init_use_lists(bool only_process_use_pos) { 4722 for (int i = _first_reg; i <= _last_reg; i++) { 4723 _use_pos[i] = max_jint; 4724 4725 if (!only_process_use_pos) { 4726 _block_pos[i] = max_jint; 4727 _spill_intervals[i]->clear(); 4728 } 4729 } 4730 } 4731 4732 inline void LinearScanWalker::exclude_from_use(int reg) { 4733 assert(reg < LinearScan::nof_regs, "interval must have a register assigned (stack slots not allowed)"); 4734 if (reg >= _first_reg && reg <= _last_reg) { 4735 _use_pos[reg] = 0; 4736 } 4737 } 4738 inline void LinearScanWalker::exclude_from_use(Interval* i) { 4739 assert(i->assigned_reg() != any_reg, "interval has no register assigned"); 4740 4741 exclude_from_use(i->assigned_reg()); 4742 exclude_from_use(i->assigned_regHi()); 4743 } 4744 4745 inline void LinearScanWalker::set_use_pos(int reg, Interval* i, int use_pos, bool only_process_use_pos) { 4746 assert(use_pos != 0, "must use exclude_from_use to set use_pos to 0"); 4747 4748 if (reg >= _first_reg && reg <= _last_reg) { 4749 if (_use_pos[reg] > use_pos) { 4750 _use_pos[reg] = use_pos; 4751 } 4752 if (!only_process_use_pos) { 4753 _spill_intervals[reg]->append(i); 4754 } 4755 } 4756 } 4757 inline void LinearScanWalker::set_use_pos(Interval* i, int use_pos, bool only_process_use_pos) { 4758 assert(i->assigned_reg() != any_reg, "interval has no register assigned"); 4759 if (use_pos != -1) { 4760 set_use_pos(i->assigned_reg(), i, use_pos, only_process_use_pos); 4761 set_use_pos(i->assigned_regHi(), i, use_pos, only_process_use_pos); 4762 } 4763 } 4764 4765 inline void LinearScanWalker::set_block_pos(int reg, Interval* i, int block_pos) { 4766 if (reg >= _first_reg && reg <= _last_reg) { 4767 if (_block_pos[reg] > block_pos) { 4768 _block_pos[reg] = block_pos; 4769 } 4770 if (_use_pos[reg] > block_pos) { 4771 _use_pos[reg] = block_pos; 4772 } 4773 } 4774 } 4775 inline void LinearScanWalker::set_block_pos(Interval* i, int block_pos) { 4776 assert(i->assigned_reg() != any_reg, "interval has no register assigned"); 4777 if (block_pos != -1) { 4778 set_block_pos(i->assigned_reg(), i, block_pos); 4779 set_block_pos(i->assigned_regHi(), i, block_pos); 4780 } 4781 } 4782 4783 4784 void LinearScanWalker::free_exclude_active_fixed() { 4785 Interval* list = active_first(fixedKind); 4786 while (list != Interval::end()) { 4787 assert(list->assigned_reg() < LinearScan::nof_regs, "active interval must have a register assigned"); 4788 exclude_from_use(list); 4789 list = list->next(); 4790 } 4791 } 4792 4793 void LinearScanWalker::free_exclude_active_any() { 4794 Interval* list = active_first(anyKind); 4795 while (list != Interval::end()) { 4796 exclude_from_use(list); 4797 list = list->next(); 4798 } 4799 } 4800 4801 void LinearScanWalker::free_collect_inactive_fixed(Interval* cur) { 4802 Interval* list = inactive_first(fixedKind); 4803 while (list != Interval::end()) { 4804 if (cur->to() <= list->current_from()) { 4805 assert(list->current_intersects_at(cur) == -1, "must not intersect"); 4806 set_use_pos(list, list->current_from(), true); 4807 } else { 4808 set_use_pos(list, list->current_intersects_at(cur), true); 4809 } 4810 list = list->next(); 4811 } 4812 } 4813 4814 void LinearScanWalker::free_collect_inactive_any(Interval* cur) { 4815 Interval* list = inactive_first(anyKind); 4816 while (list != Interval::end()) { 4817 set_use_pos(list, list->current_intersects_at(cur), true); 4818 list = list->next(); 4819 } 4820 } 4821 4822 void LinearScanWalker::free_collect_unhandled(IntervalKind kind, Interval* cur) { 4823 Interval* list = unhandled_first(kind); 4824 while (list != Interval::end()) { 4825 set_use_pos(list, list->intersects_at(cur), true); 4826 if (kind == fixedKind && cur->to() <= list->from()) { 4827 set_use_pos(list, list->from(), true); 4828 } 4829 list = list->next(); 4830 } 4831 } 4832 4833 void LinearScanWalker::spill_exclude_active_fixed() { 4834 Interval* list = active_first(fixedKind); 4835 while (list != Interval::end()) { 4836 exclude_from_use(list); 4837 list = list->next(); 4838 } 4839 } 4840 4841 void LinearScanWalker::spill_block_unhandled_fixed(Interval* cur) { 4842 Interval* list = unhandled_first(fixedKind); 4843 while (list != Interval::end()) { 4844 set_block_pos(list, list->intersects_at(cur)); 4845 list = list->next(); 4846 } 4847 } 4848 4849 void LinearScanWalker::spill_block_inactive_fixed(Interval* cur) { 4850 Interval* list = inactive_first(fixedKind); 4851 while (list != Interval::end()) { 4852 if (cur->to() > list->current_from()) { 4853 set_block_pos(list, list->current_intersects_at(cur)); 4854 } else { 4855 assert(list->current_intersects_at(cur) == -1, "invalid optimization: intervals intersect"); 4856 } 4857 4858 list = list->next(); 4859 } 4860 } 4861 4862 void LinearScanWalker::spill_collect_active_any() { 4863 Interval* list = active_first(anyKind); 4864 while (list != Interval::end()) { 4865 set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false); 4866 list = list->next(); 4867 } 4868 } 4869 4870 void LinearScanWalker::spill_collect_inactive_any(Interval* cur) { 4871 Interval* list = inactive_first(anyKind); 4872 while (list != Interval::end()) { 4873 if (list->current_intersects(cur)) { 4874 set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false); 4875 } 4876 list = list->next(); 4877 } 4878 } 4879 4880 4881 void LinearScanWalker::insert_move(int op_id, Interval* src_it, Interval* dst_it) { 4882 // output all moves here. When source and target are equal, the move is 4883 // optimized away later in assign_reg_nums 4884 4885 op_id = (op_id + 1) & ~1; 4886 BlockBegin* op_block = allocator()->block_of_op_with_id(op_id); 4887 assert(op_id > 0 && allocator()->block_of_op_with_id(op_id - 2) == op_block, "cannot insert move at block boundary"); 4888 4889 // calculate index of instruction inside instruction list of current block 4890 // the minimal index (for a block with no spill moves) can be calculated because the 4891 // numbering of instructions is known. 4892 // When the block already contains spill moves, the index must be increased until the 4893 // correct index is reached. 4894 LIR_OpList* list = op_block->lir()->instructions_list(); 4895 int index = (op_id - list->at(0)->id()) / 2; 4896 assert(list->at(index)->id() <= op_id, "error in calculation"); 4897 4898 while (list->at(index)->id() != op_id) { 4899 index++; 4900 assert(0 <= index && index < list->length(), "index out of bounds"); 4901 } 4902 assert(1 <= index && index < list->length(), "index out of bounds"); 4903 assert(list->at(index)->id() == op_id, "error in calculation"); 4904 4905 // insert new instruction before instruction at position index 4906 _move_resolver.move_insert_position(op_block->lir(), index - 1); 4907 _move_resolver.add_mapping(src_it, dst_it); 4908 } 4909 4910 4911 int LinearScanWalker::find_optimal_split_pos(BlockBegin* min_block, BlockBegin* max_block, int max_split_pos) { 4912 int from_block_nr = min_block->linear_scan_number(); 4913 int to_block_nr = max_block->linear_scan_number(); 4914 4915 assert(0 <= from_block_nr && from_block_nr < block_count(), "out of range"); 4916 assert(0 <= to_block_nr && to_block_nr < block_count(), "out of range"); 4917 assert(from_block_nr < to_block_nr, "must cross block boundary"); 4918 4919 // Try to split at end of max_block. If this would be after 4920 // max_split_pos, then use the begin of max_block 4921 int optimal_split_pos = max_block->last_lir_instruction_id() + 2; 4922 if (optimal_split_pos > max_split_pos) { 4923 optimal_split_pos = max_block->first_lir_instruction_id(); 4924 } 4925 4926 int min_loop_depth = max_block->loop_depth(); 4927 for (int i = to_block_nr - 1; i >= from_block_nr; i--) { 4928 BlockBegin* cur = block_at(i); 4929 4930 if (cur->loop_depth() < min_loop_depth) { 4931 // block with lower loop-depth found -> split at the end of this block 4932 min_loop_depth = cur->loop_depth(); 4933 optimal_split_pos = cur->last_lir_instruction_id() + 2; 4934 } 4935 } 4936 assert(optimal_split_pos > allocator()->max_lir_op_id() || allocator()->is_block_begin(optimal_split_pos), "algorithm must move split pos to block boundary"); 4937 4938 return optimal_split_pos; 4939 } 4940 4941 4942 int LinearScanWalker::find_optimal_split_pos(Interval* it, int min_split_pos, int max_split_pos, bool do_loop_optimization) { 4943 int optimal_split_pos = -1; 4944 if (min_split_pos == max_split_pos) { 4945 // trivial case, no optimization of split position possible 4946 TRACE_LINEAR_SCAN(4, tty->print_cr(" min-pos and max-pos are equal, no optimization possible")); 4947 optimal_split_pos = min_split_pos; 4948 4949 } else { 4950 assert(min_split_pos < max_split_pos, "must be true then"); 4951 assert(min_split_pos > 0, "cannot access min_split_pos - 1 otherwise"); 4952 4953 // reason for using min_split_pos - 1: when the minimal split pos is exactly at the 4954 // beginning of a block, then min_split_pos is also a possible split position. 4955 // Use the block before as min_block, because then min_block->last_lir_instruction_id() + 2 == min_split_pos 4956 BlockBegin* min_block = allocator()->block_of_op_with_id(min_split_pos - 1); 4957 4958 // reason for using max_split_pos - 1: otherwise there would be an assertion failure 4959 // when an interval ends at the end of the last block of the method 4960 // (in this case, max_split_pos == allocator()->max_lir_op_id() + 2, and there is no 4961 // block at this op_id) 4962 BlockBegin* max_block = allocator()->block_of_op_with_id(max_split_pos - 1); 4963 4964 assert(min_block->linear_scan_number() <= max_block->linear_scan_number(), "invalid order"); 4965 if (min_block == max_block) { 4966 // split position cannot be moved to block boundary, so split as late as possible 4967 TRACE_LINEAR_SCAN(4, tty->print_cr(" cannot move split pos to block boundary because min_pos and max_pos are in same block")); 4968 optimal_split_pos = max_split_pos; 4969 4970 } else if (it->has_hole_between(max_split_pos - 1, max_split_pos) && !allocator()->is_block_begin(max_split_pos)) { 4971 // Do not move split position if the interval has a hole before max_split_pos. 4972 // Intervals resulting from Phi-Functions have more than one definition (marked 4973 // as mustHaveRegister) with a hole before each definition. When the register is needed 4974 // for the second definition, an earlier reloading is unnecessary. 4975 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has hole just before max_split_pos, so splitting at max_split_pos")); 4976 optimal_split_pos = max_split_pos; 4977 4978 } else { 4979 // seach optimal block boundary between min_split_pos and max_split_pos 4980 TRACE_LINEAR_SCAN(4, tty->print_cr(" moving split pos to optimal block boundary between block B%d and B%d", min_block->block_id(), max_block->block_id())); 4981 4982 if (do_loop_optimization) { 4983 // Loop optimization: if a loop-end marker is found between min- and max-position, 4984 // then split before this loop 4985 int loop_end_pos = it->next_usage_exact(loopEndMarker, min_block->last_lir_instruction_id() + 2); 4986 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization: loop end found at pos %d", loop_end_pos)); 4987 4988 assert(loop_end_pos > min_split_pos, "invalid order"); 4989 if (loop_end_pos < max_split_pos) { 4990 // loop-end marker found between min- and max-position 4991 // if it is not the end marker for the same loop as the min-position, then move 4992 // the max-position to this loop block. 4993 // Desired result: uses tagged as shouldHaveRegister inside a loop cause a reloading 4994 // of the interval (normally, only mustHaveRegister causes a reloading) 4995 BlockBegin* loop_block = allocator()->block_of_op_with_id(loop_end_pos); 4996 4997 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval is used in loop that ends in block B%d, so trying to move max_block back from B%d to B%d", loop_block->block_id(), max_block->block_id(), loop_block->block_id())); 4998 assert(loop_block != min_block, "loop_block and min_block must be different because block boundary is needed between"); 4999 5000 optimal_split_pos = find_optimal_split_pos(min_block, loop_block, loop_block->last_lir_instruction_id() + 2); 5001 if (optimal_split_pos == loop_block->last_lir_instruction_id() + 2) { 5002 optimal_split_pos = -1; 5003 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization not necessary")); 5004 } else { 5005 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization successful")); 5006 } 5007 } 5008 } 5009 5010 if (optimal_split_pos == -1) { 5011 // not calculated by loop optimization 5012 optimal_split_pos = find_optimal_split_pos(min_block, max_block, max_split_pos); 5013 } 5014 } 5015 } 5016 TRACE_LINEAR_SCAN(4, tty->print_cr(" optimal split position: %d", optimal_split_pos)); 5017 5018 return optimal_split_pos; 5019 } 5020 5021 5022 /* 5023 split an interval at the optimal position between min_split_pos and 5024 max_split_pos in two parts: 5025 1) the left part has already a location assigned 5026 2) the right part is sorted into to the unhandled-list 5027 */ 5028 void LinearScanWalker::split_before_usage(Interval* it, int min_split_pos, int max_split_pos) { 5029 TRACE_LINEAR_SCAN(2, tty->print ("----- splitting interval: "); it->print()); 5030 TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d", min_split_pos, max_split_pos)); 5031 5032 assert(it->from() < min_split_pos, "cannot split at start of interval"); 5033 assert(current_position() < min_split_pos, "cannot split before current position"); 5034 assert(min_split_pos <= max_split_pos, "invalid order"); 5035 assert(max_split_pos <= it->to(), "cannot split after end of interval"); 5036 5037 int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, true); 5038 5039 assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range"); 5040 assert(optimal_split_pos <= it->to(), "cannot split after end of interval"); 5041 assert(optimal_split_pos > it->from(), "cannot split at start of interval"); 5042 5043 if (optimal_split_pos == it->to() && it->next_usage(mustHaveRegister, min_split_pos) == max_jint) { 5044 // the split position would be just before the end of the interval 5045 // -> no split at all necessary 5046 TRACE_LINEAR_SCAN(4, tty->print_cr(" no split necessary because optimal split position is at end of interval")); 5047 return; 5048 } 5049 5050 // must calculate this before the actual split is performed and before split position is moved to odd op_id 5051 bool move_necessary = !allocator()->is_block_begin(optimal_split_pos) && !it->has_hole_between(optimal_split_pos - 1, optimal_split_pos); 5052 5053 if (!allocator()->is_block_begin(optimal_split_pos)) { 5054 // move position before actual instruction (odd op_id) 5055 optimal_split_pos = (optimal_split_pos - 1) | 1; 5056 } 5057 5058 TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d", optimal_split_pos)); 5059 assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary"); 5060 assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary"); 5061 5062 Interval* split_part = it->split(optimal_split_pos); 5063 5064 allocator()->append_interval(split_part); 5065 allocator()->copy_register_flags(it, split_part); 5066 split_part->set_insert_move_when_activated(move_necessary); 5067 append_to_unhandled(unhandled_first_addr(anyKind), split_part); 5068 5069 TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts (insert_move_when_activated: %d)", move_necessary)); 5070 TRACE_LINEAR_SCAN(2, tty->print (" "); it->print()); 5071 TRACE_LINEAR_SCAN(2, tty->print (" "); split_part->print()); 5072 } 5073 5074 /* 5075 split an interval at the optimal position between min_split_pos and 5076 max_split_pos in two parts: 5077 1) the left part has already a location assigned 5078 2) the right part is always on the stack and therefore ignored in further processing 5079 */ 5080 void LinearScanWalker::split_for_spilling(Interval* it) { 5081 // calculate allowed range of splitting position 5082 int max_split_pos = current_position(); 5083 int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, max_split_pos) + 1, it->from()); 5084 5085 TRACE_LINEAR_SCAN(2, tty->print ("----- splitting and spilling interval: "); it->print()); 5086 TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d", min_split_pos, max_split_pos)); 5087 5088 assert(it->state() == activeState, "why spill interval that is not active?"); 5089 assert(it->from() <= min_split_pos, "cannot split before start of interval"); 5090 assert(min_split_pos <= max_split_pos, "invalid order"); 5091 assert(max_split_pos < it->to(), "cannot split at end end of interval"); 5092 assert(current_position() < it->to(), "interval must not end before current position"); 5093 5094 if (min_split_pos == it->from()) { 5095 // the whole interval is never used, so spill it entirely to memory 5096 TRACE_LINEAR_SCAN(2, tty->print_cr(" spilling entire interval because split pos is at beginning of interval")); 5097 assert(it->first_usage(shouldHaveRegister) > current_position(), "interval must not have use position before current_position"); 5098 5099 allocator()->assign_spill_slot(it); 5100 allocator()->change_spill_state(it, min_split_pos); 5101 5102 // Also kick parent intervals out of register to memory when they have no use 5103 // position. This avoids short interval in register surrounded by intervals in 5104 // memory -> avoid useless moves from memory to register and back 5105 Interval* parent = it; 5106 while (parent != NULL && parent->is_split_child()) { 5107 parent = parent->split_child_before_op_id(parent->from()); 5108 5109 if (parent->assigned_reg() < LinearScan::nof_regs) { 5110 if (parent->first_usage(shouldHaveRegister) == max_jint) { 5111 // parent is never used, so kick it out of its assigned register 5112 TRACE_LINEAR_SCAN(4, tty->print_cr(" kicking out interval %d out of its register because it is never used", parent->reg_num())); 5113 allocator()->assign_spill_slot(parent); 5114 } else { 5115 // do not go further back because the register is actually used by the interval 5116 parent = NULL; 5117 } 5118 } 5119 } 5120 5121 } else { 5122 // search optimal split pos, split interval and spill only the right hand part 5123 int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, false); 5124 5125 assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range"); 5126 assert(optimal_split_pos < it->to(), "cannot split at end of interval"); 5127 assert(optimal_split_pos >= it->from(), "cannot split before start of interval"); 5128 5129 if (!allocator()->is_block_begin(optimal_split_pos)) { 5130 // move position before actual instruction (odd op_id) 5131 optimal_split_pos = (optimal_split_pos - 1) | 1; 5132 } 5133 5134 TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d", optimal_split_pos)); 5135 assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary"); 5136 assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary"); 5137 5138 Interval* spilled_part = it->split(optimal_split_pos); 5139 allocator()->append_interval(spilled_part); 5140 allocator()->assign_spill_slot(spilled_part); 5141 allocator()->change_spill_state(spilled_part, optimal_split_pos); 5142 5143 if (!allocator()->is_block_begin(optimal_split_pos)) { 5144 TRACE_LINEAR_SCAN(4, tty->print_cr(" inserting move from interval %d to %d", it->reg_num(), spilled_part->reg_num())); 5145 insert_move(optimal_split_pos, it, spilled_part); 5146 } 5147 5148 // the current_split_child is needed later when moves are inserted for reloading 5149 assert(spilled_part->current_split_child() == it, "overwriting wrong current_split_child"); 5150 spilled_part->make_current_split_child(); 5151 5152 TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts")); 5153 TRACE_LINEAR_SCAN(2, tty->print (" "); it->print()); 5154 TRACE_LINEAR_SCAN(2, tty->print (" "); spilled_part->print()); 5155 } 5156 } 5157 5158 5159 void LinearScanWalker::split_stack_interval(Interval* it) { 5160 int min_split_pos = current_position() + 1; 5161 int max_split_pos = MIN2(it->first_usage(shouldHaveRegister), it->to()); 5162 5163 split_before_usage(it, min_split_pos, max_split_pos); 5164 } 5165 5166 void LinearScanWalker::split_when_partial_register_available(Interval* it, int register_available_until) { 5167 int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, register_available_until), it->from() + 1); 5168 int max_split_pos = register_available_until; 5169 5170 split_before_usage(it, min_split_pos, max_split_pos); 5171 } 5172 5173 void LinearScanWalker::split_and_spill_interval(Interval* it) { 5174 assert(it->state() == activeState || it->state() == inactiveState, "other states not allowed"); 5175 5176 int current_pos = current_position(); 5177 if (it->state() == inactiveState) { 5178 // the interval is currently inactive, so no spill slot is needed for now. 5179 // when the split part is activated, the interval has a new chance to get a register, 5180 // so in the best case no stack slot is necessary 5181 assert(it->has_hole_between(current_pos - 1, current_pos + 1), "interval can not be inactive otherwise"); 5182 split_before_usage(it, current_pos + 1, current_pos + 1); 5183 5184 } else { 5185 // search the position where the interval must have a register and split 5186 // at the optimal position before. 5187 // The new created part is added to the unhandled list and will get a register 5188 // when it is activated 5189 int min_split_pos = current_pos + 1; 5190 int max_split_pos = MIN2(it->next_usage(mustHaveRegister, min_split_pos), it->to()); 5191 5192 split_before_usage(it, min_split_pos, max_split_pos); 5193 5194 assert(it->next_usage(mustHaveRegister, current_pos) == max_jint, "the remaining part is spilled to stack and therefore has no register"); 5195 split_for_spilling(it); 5196 } 5197 } 5198 5199 5200 int LinearScanWalker::find_free_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) { 5201 int min_full_reg = any_reg; 5202 int max_partial_reg = any_reg; 5203 5204 for (int i = _first_reg; i <= _last_reg; i++) { 5205 if (i == ignore_reg) { 5206 // this register must be ignored 5207 5208 } else if (_use_pos[i] >= interval_to) { 5209 // this register is free for the full interval 5210 if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) { 5211 min_full_reg = i; 5212 } 5213 } else if (_use_pos[i] > reg_needed_until) { 5214 // this register is at least free until reg_needed_until 5215 if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) { 5216 max_partial_reg = i; 5217 } 5218 } 5219 } 5220 5221 if (min_full_reg != any_reg) { 5222 return min_full_reg; 5223 } else if (max_partial_reg != any_reg) { 5224 *need_split = true; 5225 return max_partial_reg; 5226 } else { 5227 return any_reg; 5228 } 5229 } 5230 5231 int LinearScanWalker::find_free_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) { 5232 assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm"); 5233 5234 int min_full_reg = any_reg; 5235 int max_partial_reg = any_reg; 5236 5237 for (int i = _first_reg; i < _last_reg; i+=2) { 5238 if (_use_pos[i] >= interval_to && _use_pos[i + 1] >= interval_to) { 5239 // this register is free for the full interval 5240 if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) { 5241 min_full_reg = i; 5242 } 5243 } else if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) { 5244 // this register is at least free until reg_needed_until 5245 if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) { 5246 max_partial_reg = i; 5247 } 5248 } 5249 } 5250 5251 if (min_full_reg != any_reg) { 5252 return min_full_reg; 5253 } else if (max_partial_reg != any_reg) { 5254 *need_split = true; 5255 return max_partial_reg; 5256 } else { 5257 return any_reg; 5258 } 5259 } 5260 5261 5262 bool LinearScanWalker::alloc_free_reg(Interval* cur) { 5263 TRACE_LINEAR_SCAN(2, tty->print("trying to find free register for "); cur->print()); 5264 5265 init_use_lists(true); 5266 free_exclude_active_fixed(); 5267 free_exclude_active_any(); 5268 free_collect_inactive_fixed(cur); 5269 free_collect_inactive_any(cur); 5270 // free_collect_unhandled(fixedKind, cur); 5271 assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0"); 5272 5273 // _use_pos contains the start of the next interval that has this register assigned 5274 // (either as a fixed register or a normal allocated register in the past) 5275 // only intervals overlapping with cur are processed, non-overlapping invervals can be ignored safely 5276 TRACE_LINEAR_SCAN(4, tty->print_cr(" state of registers:")); 5277 TRACE_LINEAR_SCAN(4, for (int i = _first_reg; i <= _last_reg; i++) tty->print_cr(" reg %d: use_pos: %d", i, _use_pos[i])); 5278 5279 int hint_reg, hint_regHi; 5280 Interval* register_hint = cur->register_hint(); 5281 if (register_hint != NULL) { 5282 hint_reg = register_hint->assigned_reg(); 5283 hint_regHi = register_hint->assigned_regHi(); 5284 5285 if (allocator()->is_precolored_cpu_interval(register_hint)) { 5286 assert(hint_reg != any_reg && hint_regHi == any_reg, "must be for fixed intervals"); 5287 hint_regHi = hint_reg + 1; // connect e.g. eax-edx 5288 } 5289 TRACE_LINEAR_SCAN(4, tty->print(" hint registers %d, %d from interval ", hint_reg, hint_regHi); register_hint->print()); 5290 5291 } else { 5292 hint_reg = any_reg; 5293 hint_regHi = any_reg; 5294 } 5295 assert(hint_reg == any_reg || hint_reg != hint_regHi, "hint reg and regHi equal"); 5296 assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned to interval"); 5297 5298 // the register must be free at least until this position 5299 int reg_needed_until = cur->from() + 1; 5300 int interval_to = cur->to(); 5301 5302 bool need_split = false; 5303 int split_pos = -1; 5304 int reg = any_reg; 5305 int regHi = any_reg; 5306 5307 if (_adjacent_regs) { 5308 reg = find_free_double_reg(reg_needed_until, interval_to, hint_reg, &need_split); 5309 regHi = reg + 1; 5310 if (reg == any_reg) { 5311 return false; 5312 } 5313 split_pos = MIN2(_use_pos[reg], _use_pos[regHi]); 5314 5315 } else { 5316 reg = find_free_reg(reg_needed_until, interval_to, hint_reg, any_reg, &need_split); 5317 if (reg == any_reg) { 5318 return false; 5319 } 5320 split_pos = _use_pos[reg]; 5321 5322 if (_num_phys_regs == 2) { 5323 regHi = find_free_reg(reg_needed_until, interval_to, hint_regHi, reg, &need_split); 5324 5325 if (_use_pos[reg] < interval_to && regHi == any_reg) { 5326 // do not split interval if only one register can be assigned until the split pos 5327 // (when one register is found for the whole interval, split&spill is only 5328 // performed for the hi register) 5329 return false; 5330 5331 } else if (regHi != any_reg) { 5332 split_pos = MIN2(split_pos, _use_pos[regHi]); 5333 5334 // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax 5335 if (reg > regHi) { 5336 int temp = reg; 5337 reg = regHi; 5338 regHi = temp; 5339 } 5340 } 5341 } 5342 } 5343 5344 cur->assign_reg(reg, regHi); 5345 TRACE_LINEAR_SCAN(2, tty->print_cr("selected register %d, %d", reg, regHi)); 5346 5347 assert(split_pos > 0, "invalid split_pos"); 5348 if (need_split) { 5349 // register not available for full interval, so split it 5350 split_when_partial_register_available(cur, split_pos); 5351 } 5352 5353 // only return true if interval is completely assigned 5354 return _num_phys_regs == 1 || regHi != any_reg; 5355 } 5356 5357 5358 int LinearScanWalker::find_locked_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) { 5359 int max_reg = any_reg; 5360 5361 for (int i = _first_reg; i <= _last_reg; i++) { 5362 if (i == ignore_reg) { 5363 // this register must be ignored 5364 5365 } else if (_use_pos[i] > reg_needed_until) { 5366 if (max_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_reg] && max_reg != hint_reg)) { 5367 max_reg = i; 5368 } 5369 } 5370 } 5371 5372 if (max_reg != any_reg && _block_pos[max_reg] <= interval_to) { 5373 *need_split = true; 5374 } 5375 5376 return max_reg; 5377 } 5378 5379 int LinearScanWalker::find_locked_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) { 5380 assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm"); 5381 5382 int max_reg = any_reg; 5383 5384 for (int i = _first_reg; i < _last_reg; i+=2) { 5385 if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) { 5386 if (max_reg == any_reg || _use_pos[i] > _use_pos[max_reg]) { 5387 max_reg = i; 5388 } 5389 } 5390 } 5391 5392 if (_block_pos[max_reg] <= interval_to || _block_pos[max_reg + 1] <= interval_to) { 5393 *need_split = true; 5394 } 5395 5396 return max_reg; 5397 } 5398 5399 void LinearScanWalker::split_and_spill_intersecting_intervals(int reg, int regHi) { 5400 assert(reg != any_reg, "no register assigned"); 5401 5402 for (int i = 0; i < _spill_intervals[reg]->length(); i++) { 5403 Interval* it = _spill_intervals[reg]->at(i); 5404 remove_from_list(it); 5405 split_and_spill_interval(it); 5406 } 5407 5408 if (regHi != any_reg) { 5409 IntervalList* processed = _spill_intervals[reg]; 5410 for (int i = 0; i < _spill_intervals[regHi]->length(); i++) { 5411 Interval* it = _spill_intervals[regHi]->at(i); 5412 if (processed->index_of(it) == -1) { 5413 remove_from_list(it); 5414 split_and_spill_interval(it); 5415 } 5416 } 5417 } 5418 } 5419 5420 5421 // Split an Interval and spill it to memory so that cur can be placed in a register 5422 void LinearScanWalker::alloc_locked_reg(Interval* cur) { 5423 TRACE_LINEAR_SCAN(2, tty->print("need to split and spill to get register for "); cur->print()); 5424 5425 // collect current usage of registers 5426 init_use_lists(false); 5427 spill_exclude_active_fixed(); 5428 // spill_block_unhandled_fixed(cur); 5429 assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0"); 5430 spill_block_inactive_fixed(cur); 5431 spill_collect_active_any(); 5432 spill_collect_inactive_any(cur); 5433 5434 #ifndef PRODUCT 5435 if (TraceLinearScanLevel >= 4) { 5436 tty->print_cr(" state of registers:"); 5437 for (int i = _first_reg; i <= _last_reg; i++) { 5438 tty->print(" reg %d: use_pos: %d, block_pos: %d, intervals: ", i, _use_pos[i], _block_pos[i]); 5439 for (int j = 0; j < _spill_intervals[i]->length(); j++) { 5440 tty->print("%d ", _spill_intervals[i]->at(j)->reg_num()); 5441 } 5442 tty->cr(); 5443 } 5444 } 5445 #endif 5446 5447 // the register must be free at least until this position 5448 int reg_needed_until = MIN2(cur->first_usage(mustHaveRegister), cur->from() + 1); 5449 int interval_to = cur->to(); 5450 assert (reg_needed_until > 0 && reg_needed_until < max_jint, "interval has no use"); 5451 5452 int split_pos = 0; 5453 int use_pos = 0; 5454 bool need_split = false; 5455 int reg, regHi; 5456 5457 if (_adjacent_regs) { 5458 reg = find_locked_double_reg(reg_needed_until, interval_to, any_reg, &need_split); 5459 regHi = reg + 1; 5460 5461 if (reg != any_reg) { 5462 use_pos = MIN2(_use_pos[reg], _use_pos[regHi]); 5463 split_pos = MIN2(_block_pos[reg], _block_pos[regHi]); 5464 } 5465 } else { 5466 reg = find_locked_reg(reg_needed_until, interval_to, any_reg, cur->assigned_reg(), &need_split); 5467 regHi = any_reg; 5468 5469 if (reg != any_reg) { 5470 use_pos = _use_pos[reg]; 5471 split_pos = _block_pos[reg]; 5472 5473 if (_num_phys_regs == 2) { 5474 if (cur->assigned_reg() != any_reg) { 5475 regHi = reg; 5476 reg = cur->assigned_reg(); 5477 } else { 5478 regHi = find_locked_reg(reg_needed_until, interval_to, any_reg, reg, &need_split); 5479 if (regHi != any_reg) { 5480 use_pos = MIN2(use_pos, _use_pos[regHi]); 5481 split_pos = MIN2(split_pos, _block_pos[regHi]); 5482 } 5483 } 5484 5485 if (regHi != any_reg && reg > regHi) { 5486 // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax 5487 int temp = reg; 5488 reg = regHi; 5489 regHi = temp; 5490 } 5491 } 5492 } 5493 } 5494 5495 if (reg == any_reg || (_num_phys_regs == 2 && regHi == any_reg) || use_pos <= cur->first_usage(mustHaveRegister)) { 5496 // the first use of cur is later than the spilling position -> spill cur 5497 TRACE_LINEAR_SCAN(4, tty->print_cr("able to spill current interval. first_usage(register): %d, use_pos: %d", cur->first_usage(mustHaveRegister), use_pos)); 5498 5499 if (cur->first_usage(mustHaveRegister) <= cur->from() + 1) { 5500 assert(false, "cannot spill interval that is used in first instruction (possible reason: no register found)"); 5501 // assign a reasonable register and do a bailout in product mode to avoid errors 5502 allocator()->assign_spill_slot(cur); 5503 BAILOUT("LinearScan: no register found"); 5504 } 5505 5506 split_and_spill_interval(cur); 5507 } else { 5508 TRACE_LINEAR_SCAN(4, tty->print_cr("decided to use register %d, %d", reg, regHi)); 5509 assert(reg != any_reg && (_num_phys_regs == 1 || regHi != any_reg), "no register found"); 5510 assert(split_pos > 0, "invalid split_pos"); 5511 assert(need_split == false || split_pos > cur->from(), "splitting interval at from"); 5512 5513 cur->assign_reg(reg, regHi); 5514 if (need_split) { 5515 // register not available for full interval, so split it 5516 split_when_partial_register_available(cur, split_pos); 5517 } 5518 5519 // perform splitting and spilling for all affected intervalls 5520 split_and_spill_intersecting_intervals(reg, regHi); 5521 } 5522 } 5523 5524 bool LinearScanWalker::no_allocation_possible(Interval* cur) { 5525 #ifdef X86 5526 // fast calculation of intervals that can never get a register because the 5527 // the next instruction is a call that blocks all registers 5528 // Note: this does not work if callee-saved registers are available (e.g. on Sparc) 5529 5530 // check if this interval is the result of a split operation 5531 // (an interval got a register until this position) 5532 int pos = cur->from(); 5533 if ((pos & 1) == 1) { 5534 // the current instruction is a call that blocks all registers 5535 if (pos < allocator()->max_lir_op_id() && allocator()->has_call(pos + 1)) { 5536 TRACE_LINEAR_SCAN(4, tty->print_cr(" free register cannot be available because all registers blocked by following call")); 5537 5538 // safety check that there is really no register available 5539 assert(alloc_free_reg(cur) == false, "found a register for this interval"); 5540 return true; 5541 } 5542 5543 } 5544 #endif 5545 return false; 5546 } 5547 5548 void LinearScanWalker::init_vars_for_alloc(Interval* cur) { 5549 BasicType type = cur->type(); 5550 _num_phys_regs = LinearScan::num_physical_regs(type); 5551 _adjacent_regs = LinearScan::requires_adjacent_regs(type); 5552 5553 if (pd_init_regs_for_alloc(cur)) { 5554 // the appropriate register range was selected. 5555 } else if (type == T_FLOAT || type == T_DOUBLE) { 5556 _first_reg = pd_first_fpu_reg; 5557 _last_reg = pd_last_fpu_reg; 5558 } else { 5559 _first_reg = pd_first_cpu_reg; 5560 _last_reg = pd_last_cpu_reg; 5561 } 5562 5563 assert(0 <= _first_reg && _first_reg < LinearScan::nof_regs, "out of range"); 5564 assert(0 <= _last_reg && _last_reg < LinearScan::nof_regs, "out of range"); 5565 } 5566 5567 5568 bool LinearScanWalker::is_move(LIR_Op* op, Interval* from, Interval* to) { 5569 if (op->code() != lir_move) { 5570 return false; 5571 } 5572 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); 5573 5574 LIR_Opr in = ((LIR_Op1*)op)->in_opr(); 5575 LIR_Opr res = ((LIR_Op1*)op)->result_opr(); 5576 return in->is_virtual() && res->is_virtual() && in->vreg_number() == from->reg_num() && res->vreg_number() == to->reg_num(); 5577 } 5578 5579 // optimization (especially for phi functions of nested loops): 5580 // assign same spill slot to non-intersecting intervals 5581 void LinearScanWalker::combine_spilled_intervals(Interval* cur) { 5582 if (cur->is_split_child()) { 5583 // optimization is only suitable for split parents 5584 return; 5585 } 5586 5587 Interval* register_hint = cur->register_hint(false); 5588 if (register_hint == NULL) { 5589 // cur is not the target of a move, otherwise register_hint would be set 5590 return; 5591 } 5592 assert(register_hint->is_split_parent(), "register hint must be split parent"); 5593 5594 if (cur->spill_state() != noOptimization || register_hint->spill_state() != noOptimization) { 5595 // combining the stack slots for intervals where spill move optimization is applied 5596 // is not benefitial and would cause problems 5597 return; 5598 } 5599 5600 int begin_pos = cur->from(); 5601 int end_pos = cur->to(); 5602 if (end_pos > allocator()->max_lir_op_id() || (begin_pos & 1) != 0 || (end_pos & 1) != 0) { 5603 // safety check that lir_op_with_id is allowed 5604 return; 5605 } 5606 5607 if (!is_move(allocator()->lir_op_with_id(begin_pos), register_hint, cur) || !is_move(allocator()->lir_op_with_id(end_pos), cur, register_hint)) { 5608 // cur and register_hint are not connected with two moves 5609 return; 5610 } 5611 5612 Interval* begin_hint = register_hint->split_child_at_op_id(begin_pos, LIR_OpVisitState::inputMode); 5613 Interval* end_hint = register_hint->split_child_at_op_id(end_pos, LIR_OpVisitState::outputMode); 5614 if (begin_hint == end_hint || begin_hint->to() != begin_pos || end_hint->from() != end_pos) { 5615 // register_hint must be split, otherwise the re-writing of use positions does not work 5616 return; 5617 } 5618 5619 assert(begin_hint->assigned_reg() != any_reg, "must have register assigned"); 5620 assert(end_hint->assigned_reg() == any_reg, "must not have register assigned"); 5621 assert(cur->first_usage(mustHaveRegister) == begin_pos, "must have use position at begin of interval because of move"); 5622 assert(end_hint->first_usage(mustHaveRegister) == end_pos, "must have use position at begin of interval because of move"); 5623 5624 if (begin_hint->assigned_reg() < LinearScan::nof_regs) { 5625 // register_hint is not spilled at begin_pos, so it would not be benefitial to immediately spill cur 5626 return; 5627 } 5628 assert(register_hint->canonical_spill_slot() != -1, "must be set when part of interval was spilled"); 5629 5630 // modify intervals such that cur gets the same stack slot as register_hint 5631 // delete use positions to prevent the intervals to get a register at beginning 5632 cur->set_canonical_spill_slot(register_hint->canonical_spill_slot()); 5633 cur->remove_first_use_pos(); 5634 end_hint->remove_first_use_pos(); 5635 } 5636 5637 5638 // allocate a physical register or memory location to an interval 5639 bool LinearScanWalker::activate_current() { 5640 Interval* cur = current(); 5641 bool result = true; 5642 5643 TRACE_LINEAR_SCAN(2, tty->print ("+++++ activating interval "); cur->print()); 5644 TRACE_LINEAR_SCAN(4, tty->print_cr(" split_parent: %d, insert_move_when_activated: %d", cur->split_parent()->reg_num(), cur->insert_move_when_activated())); 5645 5646 if (cur->assigned_reg() >= LinearScan::nof_regs) { 5647 // activating an interval that has a stack slot assigned -> split it at first use position 5648 // used for method parameters 5649 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has spill slot assigned (method parameter) -> split it before first use")); 5650 5651 split_stack_interval(cur); 5652 result = false; 5653 5654 } else if (allocator()->gen()->is_vreg_flag_set(cur->reg_num(), LIRGenerator::must_start_in_memory)) { 5655 // activating an interval that must start in a stack slot, but may get a register later 5656 // used for lir_roundfp: rounding is done by store to stack and reload later 5657 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval must start in stack slot -> split it before first use")); 5658 assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned"); 5659 5660 allocator()->assign_spill_slot(cur); 5661 split_stack_interval(cur); 5662 result = false; 5663 5664 } else if (cur->assigned_reg() == any_reg) { 5665 // interval has not assigned register -> normal allocation 5666 // (this is the normal case for most intervals) 5667 TRACE_LINEAR_SCAN(4, tty->print_cr(" normal allocation of register")); 5668 5669 // assign same spill slot to non-intersecting intervals 5670 combine_spilled_intervals(cur); 5671 5672 init_vars_for_alloc(cur); 5673 if (no_allocation_possible(cur) || !alloc_free_reg(cur)) { 5674 // no empty register available. 5675 // split and spill another interval so that this interval gets a register 5676 alloc_locked_reg(cur); 5677 } 5678 5679 // spilled intervals need not be move to active-list 5680 if (cur->assigned_reg() >= LinearScan::nof_regs) { 5681 result = false; 5682 } 5683 } 5684 5685 // load spilled values that become active from stack slot to register 5686 if (cur->insert_move_when_activated()) { 5687 assert(cur->is_split_child(), "must be"); 5688 assert(cur->current_split_child() != NULL, "must be"); 5689 assert(cur->current_split_child()->reg_num() != cur->reg_num(), "cannot insert move between same interval"); 5690 TRACE_LINEAR_SCAN(4, tty->print_cr("Inserting move from interval %d to %d because insert_move_when_activated is set", cur->current_split_child()->reg_num(), cur->reg_num())); 5691 5692 insert_move(cur->from(), cur->current_split_child(), cur); 5693 } 5694 cur->make_current_split_child(); 5695 5696 return result; // true = interval is moved to active list 5697 } 5698 5699 5700 // Implementation of EdgeMoveOptimizer 5701 5702 EdgeMoveOptimizer::EdgeMoveOptimizer() : 5703 _edge_instructions(4), 5704 _edge_instructions_idx(4) 5705 { 5706 } 5707 5708 void EdgeMoveOptimizer::optimize(BlockList* code) { 5709 EdgeMoveOptimizer optimizer = EdgeMoveOptimizer(); 5710 5711 // ignore the first block in the list (index 0 is not processed) 5712 for (int i = code->length() - 1; i >= 1; i--) { 5713 BlockBegin* block = code->at(i); 5714 5715 if (block->number_of_preds() > 1 && !block->is_set(BlockBegin::exception_entry_flag)) { 5716 optimizer.optimize_moves_at_block_end(block); 5717 } 5718 if (block->number_of_sux() == 2) { 5719 optimizer.optimize_moves_at_block_begin(block); 5720 } 5721 } 5722 } 5723 5724 5725 // clear all internal data structures 5726 void EdgeMoveOptimizer::init_instructions() { 5727 _edge_instructions.clear(); 5728 _edge_instructions_idx.clear(); 5729 } 5730 5731 // append a lir-instruction-list and the index of the current operation in to the list 5732 void EdgeMoveOptimizer::append_instructions(LIR_OpList* instructions, int instructions_idx) { 5733 _edge_instructions.append(instructions); 5734 _edge_instructions_idx.append(instructions_idx); 5735 } 5736 5737 // return the current operation of the given edge (predecessor or successor) 5738 LIR_Op* EdgeMoveOptimizer::instruction_at(int edge) { 5739 LIR_OpList* instructions = _edge_instructions.at(edge); 5740 int idx = _edge_instructions_idx.at(edge); 5741 5742 if (idx < instructions->length()) { 5743 return instructions->at(idx); 5744 } else { 5745 return NULL; 5746 } 5747 } 5748 5749 // removes the current operation of the given edge (predecessor or successor) 5750 void EdgeMoveOptimizer::remove_cur_instruction(int edge, bool decrement_index) { 5751 LIR_OpList* instructions = _edge_instructions.at(edge); 5752 int idx = _edge_instructions_idx.at(edge); 5753 instructions->remove_at(idx); 5754 5755 if (decrement_index) { 5756 _edge_instructions_idx.at_put(edge, idx - 1); 5757 } 5758 } 5759 5760 5761 bool EdgeMoveOptimizer::operations_different(LIR_Op* op1, LIR_Op* op2) { 5762 if (op1 == NULL || op2 == NULL) { 5763 // at least one block is already empty -> no optimization possible 5764 return true; 5765 } 5766 5767 if (op1->code() == lir_move && op2->code() == lir_move) { 5768 assert(op1->as_Op1() != NULL, "move must be LIR_Op1"); 5769 assert(op2->as_Op1() != NULL, "move must be LIR_Op1"); 5770 LIR_Op1* move1 = (LIR_Op1*)op1; 5771 LIR_Op1* move2 = (LIR_Op1*)op2; 5772 if (move1->info() == move2->info() && move1->in_opr() == move2->in_opr() && move1->result_opr() == move2->result_opr()) { 5773 // these moves are exactly equal and can be optimized 5774 return false; 5775 } 5776 5777 } else if (op1->code() == lir_fxch && op2->code() == lir_fxch) { 5778 assert(op1->as_Op1() != NULL, "fxch must be LIR_Op1"); 5779 assert(op2->as_Op1() != NULL, "fxch must be LIR_Op1"); 5780 LIR_Op1* fxch1 = (LIR_Op1*)op1; 5781 LIR_Op1* fxch2 = (LIR_Op1*)op2; 5782 if (fxch1->in_opr()->as_jint() == fxch2->in_opr()->as_jint()) { 5783 // equal FPU stack operations can be optimized 5784 return false; 5785 } 5786 5787 } else if (op1->code() == lir_fpop_raw && op2->code() == lir_fpop_raw) { 5788 // equal FPU stack operations can be optimized 5789 return false; 5790 } 5791 5792 // no optimization possible 5793 return true; 5794 } 5795 5796 void EdgeMoveOptimizer::optimize_moves_at_block_end(BlockBegin* block) { 5797 TRACE_LINEAR_SCAN(4, tty->print_cr("optimizing moves at end of block B%d", block->block_id())); 5798 5799 if (block->is_predecessor(block)) { 5800 // currently we can't handle this correctly. 5801 return; 5802 } 5803 5804 init_instructions(); 5805 int num_preds = block->number_of_preds(); 5806 assert(num_preds > 1, "do not call otherwise"); 5807 assert(!block->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed"); 5808 5809 // setup a list with the lir-instructions of all predecessors 5810 int i; 5811 for (i = 0; i < num_preds; i++) { 5812 BlockBegin* pred = block->pred_at(i); 5813 LIR_OpList* pred_instructions = pred->lir()->instructions_list(); 5814 5815 if (pred->number_of_sux() != 1) { 5816 // this can happen with switch-statements where multiple edges are between 5817 // the same blocks. 5818 return; 5819 } 5820 5821 assert(pred->number_of_sux() == 1, "can handle only one successor"); 5822 assert(pred->sux_at(0) == block, "invalid control flow"); 5823 assert(pred_instructions->last()->code() == lir_branch, "block with successor must end with branch"); 5824 assert(pred_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch"); 5825 assert(pred_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch"); 5826 5827 if (pred_instructions->last()->info() != NULL) { 5828 // can not optimize instructions when debug info is needed 5829 return; 5830 } 5831 5832 // ignore the unconditional branch at the end of the block 5833 append_instructions(pred_instructions, pred_instructions->length() - 2); 5834 } 5835 5836 5837 // process lir-instructions while all predecessors end with the same instruction 5838 while (true) { 5839 LIR_Op* op = instruction_at(0); 5840 for (i = 1; i < num_preds; i++) { 5841 if (operations_different(op, instruction_at(i))) { 5842 // these instructions are different and cannot be optimized -> 5843 // no further optimization possible 5844 return; 5845 } 5846 } 5847 5848 TRACE_LINEAR_SCAN(4, tty->print("found instruction that is equal in all %d predecessors: ", num_preds); op->print()); 5849 5850 // insert the instruction at the beginning of the current block 5851 block->lir()->insert_before(1, op); 5852 5853 // delete the instruction at the end of all predecessors 5854 for (i = 0; i < num_preds; i++) { 5855 remove_cur_instruction(i, true); 5856 } 5857 } 5858 } 5859 5860 5861 void EdgeMoveOptimizer::optimize_moves_at_block_begin(BlockBegin* block) { 5862 TRACE_LINEAR_SCAN(4, tty->print_cr("optimization moves at begin of block B%d", block->block_id())); 5863 5864 init_instructions(); 5865 int num_sux = block->number_of_sux(); 5866 5867 LIR_OpList* cur_instructions = block->lir()->instructions_list(); 5868 5869 assert(num_sux == 2, "method should not be called otherwise"); 5870 assert(cur_instructions->last()->code() == lir_branch, "block with successor must end with branch"); 5871 assert(cur_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch"); 5872 assert(cur_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch"); 5873 5874 if (cur_instructions->last()->info() != NULL) { 5875 // can no optimize instructions when debug info is needed 5876 return; 5877 } 5878 5879 LIR_Op* branch = cur_instructions->at(cur_instructions->length() - 2); 5880 if (branch->info() != NULL || (branch->code() != lir_branch && branch->code() != lir_cond_float_branch)) { 5881 // not a valid case for optimization 5882 // currently, only blocks that end with two branches (conditional branch followed 5883 // by unconditional branch) are optimized 5884 return; 5885 } 5886 5887 // now it is guaranteed that the block ends with two branch instructions. 5888 // the instructions are inserted at the end of the block before these two branches 5889 int insert_idx = cur_instructions->length() - 2; 5890 5891 int i; 5892 #ifdef ASSERT 5893 for (i = insert_idx - 1; i >= 0; i--) { 5894 LIR_Op* op = cur_instructions->at(i); 5895 if ((op->code() == lir_branch || op->code() == lir_cond_float_branch) && ((LIR_OpBranch*)op)->block() != NULL) { 5896 assert(false, "block with two successors can have only two branch instructions"); 5897 } 5898 } 5899 #endif 5900 5901 // setup a list with the lir-instructions of all successors 5902 for (i = 0; i < num_sux; i++) { 5903 BlockBegin* sux = block->sux_at(i); 5904 LIR_OpList* sux_instructions = sux->lir()->instructions_list(); 5905 5906 assert(sux_instructions->at(0)->code() == lir_label, "block must start with label"); 5907 5908 if (sux->number_of_preds() != 1) { 5909 // this can happen with switch-statements where multiple edges are between 5910 // the same blocks. 5911 return; 5912 } 5913 assert(sux->pred_at(0) == block, "invalid control flow"); 5914 assert(!sux->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed"); 5915 5916 // ignore the label at the beginning of the block 5917 append_instructions(sux_instructions, 1); 5918 } 5919 5920 // process lir-instructions while all successors begin with the same instruction 5921 while (true) { 5922 LIR_Op* op = instruction_at(0); 5923 for (i = 1; i < num_sux; i++) { 5924 if (operations_different(op, instruction_at(i))) { 5925 // these instructions are different and cannot be optimized -> 5926 // no further optimization possible 5927 return; 5928 } 5929 } 5930 5931 TRACE_LINEAR_SCAN(4, tty->print("----- found instruction that is equal in all %d successors: ", num_sux); op->print()); 5932 5933 // insert instruction at end of current block 5934 block->lir()->insert_before(insert_idx, op); 5935 insert_idx++; 5936 5937 // delete the instructions at the beginning of all successors 5938 for (i = 0; i < num_sux; i++) { 5939 remove_cur_instruction(i, false); 5940 } 5941 } 5942 } 5943 5944 5945 // Implementation of ControlFlowOptimizer 5946 5947 ControlFlowOptimizer::ControlFlowOptimizer() : 5948 _original_preds(4) 5949 { 5950 } 5951 5952 void ControlFlowOptimizer::optimize(BlockList* code) { 5953 ControlFlowOptimizer optimizer = ControlFlowOptimizer(); 5954 5955 // push the OSR entry block to the end so that we're not jumping over it. 5956 BlockBegin* osr_entry = code->at(0)->end()->as_Base()->osr_entry(); 5957 if (osr_entry) { 5958 int index = osr_entry->linear_scan_number(); 5959 assert(code->at(index) == osr_entry, "wrong index"); 5960 code->remove_at(index); 5961 code->append(osr_entry); 5962 } 5963 5964 optimizer.reorder_short_loops(code); 5965 optimizer.delete_empty_blocks(code); 5966 optimizer.delete_unnecessary_jumps(code); 5967 optimizer.delete_jumps_to_return(code); 5968 } 5969 5970 void ControlFlowOptimizer::reorder_short_loop(BlockList* code, BlockBegin* header_block, int header_idx) { 5971 int i = header_idx + 1; 5972 int max_end = MIN2(header_idx + ShortLoopSize, code->length()); 5973 while (i < max_end && code->at(i)->loop_depth() >= header_block->loop_depth()) { 5974 i++; 5975 } 5976 5977 if (i == code->length() || code->at(i)->loop_depth() < header_block->loop_depth()) { 5978 int end_idx = i - 1; 5979 BlockBegin* end_block = code->at(end_idx); 5980 5981 if (end_block->number_of_sux() == 1 && end_block->sux_at(0) == header_block) { 5982 // short loop from header_idx to end_idx found -> reorder blocks such that 5983 // the header_block is the last block instead of the first block of the loop 5984 TRACE_LINEAR_SCAN(1, tty->print_cr("Reordering short loop: length %d, header B%d, end B%d", 5985 end_idx - header_idx + 1, 5986 header_block->block_id(), end_block->block_id())); 5987 5988 for (int j = header_idx; j < end_idx; j++) { 5989 code->at_put(j, code->at(j + 1)); 5990 } 5991 code->at_put(end_idx, header_block); 5992 5993 // correct the flags so that any loop alignment occurs in the right place. 5994 assert(code->at(end_idx)->is_set(BlockBegin::backward_branch_target_flag), "must be backward branch target"); 5995 code->at(end_idx)->clear(BlockBegin::backward_branch_target_flag); 5996 code->at(header_idx)->set(BlockBegin::backward_branch_target_flag); 5997 } 5998 } 5999 } 6000 6001 void ControlFlowOptimizer::reorder_short_loops(BlockList* code) { 6002 for (int i = code->length() - 1; i >= 0; i--) { 6003 BlockBegin* block = code->at(i); 6004 6005 if (block->is_set(BlockBegin::linear_scan_loop_header_flag)) { 6006 reorder_short_loop(code, block, i); 6007 } 6008 } 6009 6010 DEBUG_ONLY(verify(code)); 6011 } 6012 6013 // only blocks with exactly one successor can be deleted. Such blocks 6014 // must always end with an unconditional branch to this successor 6015 bool ControlFlowOptimizer::can_delete_block(BlockBegin* block) { 6016 if (block->number_of_sux() != 1 || block->number_of_exception_handlers() != 0 || block->is_entry_block()) { 6017 return false; 6018 } 6019 6020 LIR_OpList* instructions = block->lir()->instructions_list(); 6021 6022 assert(instructions->length() >= 2, "block must have label and branch"); 6023 assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label"); 6024 assert(instructions->last()->as_OpBranch() != NULL, "last instrcution must always be a branch"); 6025 assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "branch must be unconditional"); 6026 assert(instructions->last()->as_OpBranch()->block() == block->sux_at(0), "branch target must be the successor"); 6027 6028 // block must have exactly one successor 6029 6030 if (instructions->length() == 2 && instructions->last()->info() == NULL) { 6031 return true; 6032 } 6033 return false; 6034 } 6035 6036 // substitute branch targets in all branch-instructions of this blocks 6037 void ControlFlowOptimizer::substitute_branch_target(BlockBegin* block, BlockBegin* target_from, BlockBegin* target_to) { 6038 TRACE_LINEAR_SCAN(3, tty->print_cr("Deleting empty block: substituting from B%d to B%d inside B%d", target_from->block_id(), target_to->block_id(), block->block_id())); 6039 6040 LIR_OpList* instructions = block->lir()->instructions_list(); 6041 6042 assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label"); 6043 for (int i = instructions->length() - 1; i >= 1; i--) { 6044 LIR_Op* op = instructions->at(i); 6045 6046 if (op->code() == lir_branch || op->code() == lir_cond_float_branch) { 6047 assert(op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch"); 6048 LIR_OpBranch* branch = (LIR_OpBranch*)op; 6049 6050 if (branch->block() == target_from) { 6051 branch->change_block(target_to); 6052 } 6053 if (branch->ublock() == target_from) { 6054 branch->change_ublock(target_to); 6055 } 6056 } 6057 } 6058 } 6059 6060 void ControlFlowOptimizer::delete_empty_blocks(BlockList* code) { 6061 int old_pos = 0; 6062 int new_pos = 0; 6063 int num_blocks = code->length(); 6064 6065 while (old_pos < num_blocks) { 6066 BlockBegin* block = code->at(old_pos); 6067 6068 if (can_delete_block(block)) { 6069 BlockBegin* new_target = block->sux_at(0); 6070 6071 // propagate backward branch target flag for correct code alignment 6072 if (block->is_set(BlockBegin::backward_branch_target_flag)) { 6073 new_target->set(BlockBegin::backward_branch_target_flag); 6074 } 6075 6076 // collect a list with all predecessors that contains each predecessor only once 6077 // the predecessors of cur are changed during the substitution, so a copy of the 6078 // predecessor list is necessary 6079 int j; 6080 _original_preds.clear(); 6081 for (j = block->number_of_preds() - 1; j >= 0; j--) { 6082 BlockBegin* pred = block->pred_at(j); 6083 if (_original_preds.index_of(pred) == -1) { 6084 _original_preds.append(pred); 6085 } 6086 } 6087 6088 for (j = _original_preds.length() - 1; j >= 0; j--) { 6089 BlockBegin* pred = _original_preds.at(j); 6090 substitute_branch_target(pred, block, new_target); 6091 pred->substitute_sux(block, new_target); 6092 } 6093 } else { 6094 // adjust position of this block in the block list if blocks before 6095 // have been deleted 6096 if (new_pos != old_pos) { 6097 code->at_put(new_pos, code->at(old_pos)); 6098 } 6099 new_pos++; 6100 } 6101 old_pos++; 6102 } 6103 code->truncate(new_pos); 6104 6105 DEBUG_ONLY(verify(code)); 6106 } 6107 6108 void ControlFlowOptimizer::delete_unnecessary_jumps(BlockList* code) { 6109 // skip the last block because there a branch is always necessary 6110 for (int i = code->length() - 2; i >= 0; i--) { 6111 BlockBegin* block = code->at(i); 6112 LIR_OpList* instructions = block->lir()->instructions_list(); 6113 6114 LIR_Op* last_op = instructions->last(); 6115 if (last_op->code() == lir_branch) { 6116 assert(last_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch"); 6117 LIR_OpBranch* last_branch = (LIR_OpBranch*)last_op; 6118 6119 assert(last_branch->block() != NULL, "last branch must always have a block as target"); 6120 assert(last_branch->label() == last_branch->block()->label(), "must be equal"); 6121 6122 if (last_branch->info() == NULL) { 6123 if (last_branch->block() == code->at(i + 1)) { 6124 6125 TRACE_LINEAR_SCAN(3, tty->print_cr("Deleting unconditional branch at end of block B%d", block->block_id())); 6126 6127 // delete last branch instruction 6128 instructions->truncate(instructions->length() - 1); 6129 6130 } else { 6131 LIR_Op* prev_op = instructions->at(instructions->length() - 2); 6132 if (prev_op->code() == lir_branch || prev_op->code() == lir_cond_float_branch) { 6133 assert(prev_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch"); 6134 LIR_OpBranch* prev_branch = (LIR_OpBranch*)prev_op; 6135 6136 if (prev_branch->block() == code->at(i + 1) && prev_branch->info() == NULL) { 6137 6138 TRACE_LINEAR_SCAN(3, tty->print_cr("Negating conditional branch and deleting unconditional branch at end of block B%d", block->block_id())); 6139 6140 // eliminate a conditional branch to the immediate successor 6141 prev_branch->change_block(last_branch->block()); 6142 prev_branch->negate_cond(); 6143 instructions->truncate(instructions->length() - 1); 6144 } 6145 } 6146 } 6147 } 6148 } 6149 } 6150 6151 DEBUG_ONLY(verify(code)); 6152 } 6153 6154 void ControlFlowOptimizer::delete_jumps_to_return(BlockList* code) { 6155 #ifdef ASSERT 6156 BitMap return_converted(BlockBegin::number_of_blocks()); 6157 return_converted.clear(); 6158 #endif 6159 6160 for (int i = code->length() - 1; i >= 0; i--) { 6161 BlockBegin* block = code->at(i); 6162 LIR_OpList* cur_instructions = block->lir()->instructions_list(); 6163 LIR_Op* cur_last_op = cur_instructions->last(); 6164 6165 assert(cur_instructions->at(0)->code() == lir_label, "first instruction must always be a label"); 6166 if (cur_instructions->length() == 2 && cur_last_op->code() == lir_return) { 6167 // the block contains only a label and a return 6168 // if a predecessor ends with an unconditional jump to this block, then the jump 6169 // can be replaced with a return instruction 6170 // 6171 // Note: the original block with only a return statement cannot be deleted completely 6172 // because the predecessors might have other (conditional) jumps to this block 6173 // -> this may lead to unnecesary return instructions in the final code 6174 6175 assert(cur_last_op->info() == NULL, "return instructions do not have debug information"); 6176 assert(block->number_of_sux() == 0 || 6177 (return_converted.at(block->block_id()) && block->number_of_sux() == 1), 6178 "blocks that end with return must not have successors"); 6179 6180 assert(cur_last_op->as_Op1() != NULL, "return must be LIR_Op1"); 6181 LIR_Opr return_opr = ((LIR_Op1*)cur_last_op)->in_opr(); 6182 6183 for (int j = block->number_of_preds() - 1; j >= 0; j--) { 6184 BlockBegin* pred = block->pred_at(j); 6185 LIR_OpList* pred_instructions = pred->lir()->instructions_list(); 6186 LIR_Op* pred_last_op = pred_instructions->last(); 6187 6188 if (pred_last_op->code() == lir_branch) { 6189 assert(pred_last_op->as_OpBranch() != NULL, "branch must be LIR_OpBranch"); 6190 LIR_OpBranch* pred_last_branch = (LIR_OpBranch*)pred_last_op; 6191 6192 if (pred_last_branch->block() == block && pred_last_branch->cond() == lir_cond_always && pred_last_branch->info() == NULL) { 6193 // replace the jump to a return with a direct return 6194 // Note: currently the edge between the blocks is not deleted 6195 pred_instructions->at_put(pred_instructions->length() - 1, new LIR_Op1(lir_return, return_opr)); 6196 #ifdef ASSERT 6197 return_converted.set_bit(pred->block_id()); 6198 #endif 6199 } 6200 } 6201 } 6202 } 6203 } 6204 } 6205 6206 6207 #ifdef ASSERT 6208 void ControlFlowOptimizer::verify(BlockList* code) { 6209 for (int i = 0; i < code->length(); i++) { 6210 BlockBegin* block = code->at(i); 6211 LIR_OpList* instructions = block->lir()->instructions_list(); 6212 6213 int j; 6214 for (j = 0; j < instructions->length(); j++) { 6215 LIR_OpBranch* op_branch = instructions->at(j)->as_OpBranch(); 6216 6217 if (op_branch != NULL) { 6218 assert(op_branch->block() == NULL || code->index_of(op_branch->block()) != -1, "branch target not valid"); 6219 assert(op_branch->ublock() == NULL || code->index_of(op_branch->ublock()) != -1, "branch target not valid"); 6220 } 6221 } 6222 6223 for (j = 0; j < block->number_of_sux() - 1; j++) { 6224 BlockBegin* sux = block->sux_at(j); 6225 assert(code->index_of(sux) != -1, "successor not valid"); 6226 } 6227 6228 for (j = 0; j < block->number_of_preds() - 1; j++) { 6229 BlockBegin* pred = block->pred_at(j); 6230 assert(code->index_of(pred) != -1, "successor not valid"); 6231 } 6232 } 6233 } 6234 #endif 6235 6236 6237 #ifndef PRODUCT 6238 6239 // Implementation of LinearStatistic 6240 6241 const char* LinearScanStatistic::counter_name(int counter_idx) { 6242 switch (counter_idx) { 6243 case counter_method: return "compiled methods"; 6244 case counter_fpu_method: return "methods using fpu"; 6245 case counter_loop_method: return "methods with loops"; 6246 case counter_exception_method:return "methods with xhandler"; 6247 6248 case counter_loop: return "loops"; 6249 case counter_block: return "blocks"; 6250 case counter_loop_block: return "blocks inside loop"; 6251 case counter_exception_block: return "exception handler entries"; 6252 case counter_interval: return "intervals"; 6253 case counter_fixed_interval: return "fixed intervals"; 6254 case counter_range: return "ranges"; 6255 case counter_fixed_range: return "fixed ranges"; 6256 case counter_use_pos: return "use positions"; 6257 case counter_fixed_use_pos: return "fixed use positions"; 6258 case counter_spill_slots: return "spill slots"; 6259 6260 // counter for classes of lir instructions 6261 case counter_instruction: return "total instructions"; 6262 case counter_label: return "labels"; 6263 case counter_entry: return "method entries"; 6264 case counter_return: return "method returns"; 6265 case counter_call: return "method calls"; 6266 case counter_move: return "moves"; 6267 case counter_cmp: return "compare"; 6268 case counter_cond_branch: return "conditional branches"; 6269 case counter_uncond_branch: return "unconditional branches"; 6270 case counter_stub_branch: return "branches to stub"; 6271 case counter_alu: return "artithmetic + logic"; 6272 case counter_alloc: return "allocations"; 6273 case counter_sync: return "synchronisation"; 6274 case counter_throw: return "throw"; 6275 case counter_unwind: return "unwind"; 6276 case counter_typecheck: return "type+null-checks"; 6277 case counter_fpu_stack: return "fpu-stack"; 6278 case counter_misc_inst: return "other instructions"; 6279 case counter_other_inst: return "misc. instructions"; 6280 6281 // counter for different types of moves 6282 case counter_move_total: return "total moves"; 6283 case counter_move_reg_reg: return "register->register"; 6284 case counter_move_reg_stack: return "register->stack"; 6285 case counter_move_stack_reg: return "stack->register"; 6286 case counter_move_stack_stack:return "stack->stack"; 6287 case counter_move_reg_mem: return "register->memory"; 6288 case counter_move_mem_reg: return "memory->register"; 6289 case counter_move_const_any: return "constant->any"; 6290 6291 case blank_line_1: return ""; 6292 case blank_line_2: return ""; 6293 6294 default: ShouldNotReachHere(); return ""; 6295 } 6296 } 6297 6298 LinearScanStatistic::Counter LinearScanStatistic::base_counter(int counter_idx) { 6299 if (counter_idx == counter_fpu_method || counter_idx == counter_loop_method || counter_idx == counter_exception_method) { 6300 return counter_method; 6301 } else if (counter_idx == counter_loop_block || counter_idx == counter_exception_block) { 6302 return counter_block; 6303 } else if (counter_idx >= counter_instruction && counter_idx <= counter_other_inst) { 6304 return counter_instruction; 6305 } else if (counter_idx >= counter_move_total && counter_idx <= counter_move_const_any) { 6306 return counter_move_total; 6307 } 6308 return invalid_counter; 6309 } 6310 6311 LinearScanStatistic::LinearScanStatistic() { 6312 for (int i = 0; i < number_of_counters; i++) { 6313 _counters_sum[i] = 0; 6314 _counters_max[i] = -1; 6315 } 6316 6317 } 6318 6319 // add the method-local numbers to the total sum 6320 void LinearScanStatistic::sum_up(LinearScanStatistic &method_statistic) { 6321 for (int i = 0; i < number_of_counters; i++) { 6322 _counters_sum[i] += method_statistic._counters_sum[i]; 6323 _counters_max[i] = MAX2(_counters_max[i], method_statistic._counters_sum[i]); 6324 } 6325 } 6326 6327 void LinearScanStatistic::print(const char* title) { 6328 if (CountLinearScan || TraceLinearScanLevel > 0) { 6329 tty->cr(); 6330 tty->print_cr("***** LinearScan statistic - %s *****", title); 6331 6332 for (int i = 0; i < number_of_counters; i++) { 6333 if (_counters_sum[i] > 0 || _counters_max[i] >= 0) { 6334 tty->print("%25s: %8d", counter_name(i), _counters_sum[i]); 6335 6336 if (base_counter(i) != invalid_counter) { 6337 tty->print(" (%5.1f%%) ", _counters_sum[i] * 100.0 / _counters_sum[base_counter(i)]); 6338 } else { 6339 tty->print(" "); 6340 } 6341 6342 if (_counters_max[i] >= 0) { 6343 tty->print("%8d", _counters_max[i]); 6344 } 6345 } 6346 tty->cr(); 6347 } 6348 } 6349 } 6350 6351 void LinearScanStatistic::collect(LinearScan* allocator) { 6352 inc_counter(counter_method); 6353 if (allocator->has_fpu_registers()) { 6354 inc_counter(counter_fpu_method); 6355 } 6356 if (allocator->num_loops() > 0) { 6357 inc_counter(counter_loop_method); 6358 } 6359 inc_counter(counter_loop, allocator->num_loops()); 6360 inc_counter(counter_spill_slots, allocator->max_spills()); 6361 6362 int i; 6363 for (i = 0; i < allocator->interval_count(); i++) { 6364 Interval* cur = allocator->interval_at(i); 6365 6366 if (cur != NULL) { 6367 inc_counter(counter_interval); 6368 inc_counter(counter_use_pos, cur->num_use_positions()); 6369 if (LinearScan::is_precolored_interval(cur)) { 6370 inc_counter(counter_fixed_interval); 6371 inc_counter(counter_fixed_use_pos, cur->num_use_positions()); 6372 } 6373 6374 Range* range = cur->first(); 6375 while (range != Range::end()) { 6376 inc_counter(counter_range); 6377 if (LinearScan::is_precolored_interval(cur)) { 6378 inc_counter(counter_fixed_range); 6379 } 6380 range = range->next(); 6381 } 6382 } 6383 } 6384 6385 bool has_xhandlers = false; 6386 // Note: only count blocks that are in code-emit order 6387 for (i = 0; i < allocator->ir()->code()->length(); i++) { 6388 BlockBegin* cur = allocator->ir()->code()->at(i); 6389 6390 inc_counter(counter_block); 6391 if (cur->loop_depth() > 0) { 6392 inc_counter(counter_loop_block); 6393 } 6394 if (cur->is_set(BlockBegin::exception_entry_flag)) { 6395 inc_counter(counter_exception_block); 6396 has_xhandlers = true; 6397 } 6398 6399 LIR_OpList* instructions = cur->lir()->instructions_list(); 6400 for (int j = 0; j < instructions->length(); j++) { 6401 LIR_Op* op = instructions->at(j); 6402 6403 inc_counter(counter_instruction); 6404 6405 switch (op->code()) { 6406 case lir_label: inc_counter(counter_label); break; 6407 case lir_std_entry: 6408 case lir_osr_entry: inc_counter(counter_entry); break; 6409 case lir_return: inc_counter(counter_return); break; 6410 6411 case lir_rtcall: 6412 case lir_static_call: 6413 case lir_optvirtual_call: 6414 case lir_virtual_call: inc_counter(counter_call); break; 6415 6416 case lir_move: { 6417 inc_counter(counter_move); 6418 inc_counter(counter_move_total); 6419 6420 LIR_Opr in = op->as_Op1()->in_opr(); 6421 LIR_Opr res = op->as_Op1()->result_opr(); 6422 if (in->is_register()) { 6423 if (res->is_register()) { 6424 inc_counter(counter_move_reg_reg); 6425 } else if (res->is_stack()) { 6426 inc_counter(counter_move_reg_stack); 6427 } else if (res->is_address()) { 6428 inc_counter(counter_move_reg_mem); 6429 } else { 6430 ShouldNotReachHere(); 6431 } 6432 } else if (in->is_stack()) { 6433 if (res->is_register()) { 6434 inc_counter(counter_move_stack_reg); 6435 } else { 6436 inc_counter(counter_move_stack_stack); 6437 } 6438 } else if (in->is_address()) { 6439 assert(res->is_register(), "must be"); 6440 inc_counter(counter_move_mem_reg); 6441 } else if (in->is_constant()) { 6442 inc_counter(counter_move_const_any); 6443 } else { 6444 ShouldNotReachHere(); 6445 } 6446 break; 6447 } 6448 6449 case lir_cmp: inc_counter(counter_cmp); break; 6450 6451 case lir_branch: 6452 case lir_cond_float_branch: { 6453 LIR_OpBranch* branch = op->as_OpBranch(); 6454 if (branch->block() == NULL) { 6455 inc_counter(counter_stub_branch); 6456 } else if (branch->cond() == lir_cond_always) { 6457 inc_counter(counter_uncond_branch); 6458 } else { 6459 inc_counter(counter_cond_branch); 6460 } 6461 break; 6462 } 6463 6464 case lir_neg: 6465 case lir_add: 6466 case lir_sub: 6467 case lir_mul: 6468 case lir_mul_strictfp: 6469 case lir_div: 6470 case lir_div_strictfp: 6471 case lir_rem: 6472 case lir_sqrt: 6473 case lir_sin: 6474 case lir_cos: 6475 case lir_abs: 6476 case lir_log10: 6477 case lir_log: 6478 case lir_logic_and: 6479 case lir_logic_or: 6480 case lir_logic_xor: 6481 case lir_shl: 6482 case lir_shr: 6483 case lir_ushr: inc_counter(counter_alu); break; 6484 6485 case lir_alloc_object: 6486 case lir_alloc_array: inc_counter(counter_alloc); break; 6487 6488 case lir_monaddr: 6489 case lir_lock: 6490 case lir_unlock: inc_counter(counter_sync); break; 6491 6492 case lir_throw: inc_counter(counter_throw); break; 6493 6494 case lir_unwind: inc_counter(counter_unwind); break; 6495 6496 case lir_null_check: 6497 case lir_leal: 6498 case lir_instanceof: 6499 case lir_checkcast: 6500 case lir_store_check: inc_counter(counter_typecheck); break; 6501 6502 case lir_fpop_raw: 6503 case lir_fxch: 6504 case lir_fld: inc_counter(counter_fpu_stack); break; 6505 6506 case lir_nop: 6507 case lir_push: 6508 case lir_pop: 6509 case lir_convert: 6510 case lir_roundfp: 6511 case lir_cmove: inc_counter(counter_misc_inst); break; 6512 6513 default: inc_counter(counter_other_inst); break; 6514 } 6515 } 6516 } 6517 6518 if (has_xhandlers) { 6519 inc_counter(counter_exception_method); 6520 } 6521 } 6522 6523 void LinearScanStatistic::compute(LinearScan* allocator, LinearScanStatistic &global_statistic) { 6524 if (CountLinearScan || TraceLinearScanLevel > 0) { 6525 6526 LinearScanStatistic local_statistic = LinearScanStatistic(); 6527 6528 local_statistic.collect(allocator); 6529 global_statistic.sum_up(local_statistic); 6530 6531 if (TraceLinearScanLevel > 2) { 6532 local_statistic.print("current local statistic"); 6533 } 6534 } 6535 } 6536 6537 6538 // Implementation of LinearTimers 6539 6540 LinearScanTimers::LinearScanTimers() { 6541 for (int i = 0; i < number_of_timers; i++) { 6542 timer(i)->reset(); 6543 } 6544 } 6545 6546 const char* LinearScanTimers::timer_name(int idx) { 6547 switch (idx) { 6548 case timer_do_nothing: return "Nothing (Time Check)"; 6549 case timer_number_instructions: return "Number Instructions"; 6550 case timer_compute_local_live_sets: return "Local Live Sets"; 6551 case timer_compute_global_live_sets: return "Global Live Sets"; 6552 case timer_build_intervals: return "Build Intervals"; 6553 case timer_sort_intervals_before: return "Sort Intervals Before"; 6554 case timer_allocate_registers: return "Allocate Registers"; 6555 case timer_resolve_data_flow: return "Resolve Data Flow"; 6556 case timer_sort_intervals_after: return "Sort Intervals After"; 6557 case timer_eliminate_spill_moves: return "Spill optimization"; 6558 case timer_assign_reg_num: return "Assign Reg Num"; 6559 case timer_allocate_fpu_stack: return "Allocate FPU Stack"; 6560 case timer_optimize_lir: return "Optimize LIR"; 6561 default: ShouldNotReachHere(); return ""; 6562 } 6563 } 6564 6565 void LinearScanTimers::begin_method() { 6566 if (TimeEachLinearScan) { 6567 // reset all timers to measure only current method 6568 for (int i = 0; i < number_of_timers; i++) { 6569 timer(i)->reset(); 6570 } 6571 } 6572 } 6573 6574 void LinearScanTimers::end_method(LinearScan* allocator) { 6575 if (TimeEachLinearScan) { 6576 6577 double c = timer(timer_do_nothing)->seconds(); 6578 double total = 0; 6579 for (int i = 1; i < number_of_timers; i++) { 6580 total += timer(i)->seconds() - c; 6581 } 6582 6583 if (total >= 0.0005) { 6584 // print all information in one line for automatic processing 6585 tty->print("@"); allocator->compilation()->method()->print_name(); 6586 6587 tty->print("@ %d ", allocator->compilation()->method()->code_size()); 6588 tty->print("@ %d ", allocator->block_at(allocator->block_count() - 1)->last_lir_instruction_id() / 2); 6589 tty->print("@ %d ", allocator->block_count()); 6590 tty->print("@ %d ", allocator->num_virtual_regs()); 6591 tty->print("@ %d ", allocator->interval_count()); 6592 tty->print("@ %d ", allocator->_num_calls); 6593 tty->print("@ %d ", allocator->num_loops()); 6594 6595 tty->print("@ %6.6f ", total); 6596 for (int i = 1; i < number_of_timers; i++) { 6597 tty->print("@ %4.1f ", ((timer(i)->seconds() - c) / total) * 100); 6598 } 6599 tty->cr(); 6600 } 6601 } 6602 } 6603 6604 void LinearScanTimers::print(double total_time) { 6605 if (TimeLinearScan) { 6606 // correction value: sum of dummy-timer that only measures the time that 6607 // is necesary to start and stop itself 6608 double c = timer(timer_do_nothing)->seconds(); 6609 6610 for (int i = 0; i < number_of_timers; i++) { 6611 double t = timer(i)->seconds(); 6612 tty->print_cr(" %25s: %6.3f s (%4.1f%%) corrected: %6.3f s (%4.1f%%)", timer_name(i), t, (t / total_time) * 100.0, t - c, (t - c) / (total_time - 2 * number_of_timers * c) * 100); 6613 } 6614 } 6615 } 6616 6617 #endif // #ifndef PRODUCT