c1_LinearScan.cpp 247.6 KB
Newer Older
D
duke 已提交
1
/*
2
 * Copyright (c) 2005, 2013, Oracle and/or its affiliates. All rights reserved.
D
duke 已提交
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
 *
 * This code is free software; you can redistribute it and/or modify it
 * under the terms of the GNU General Public License version 2 only, as
 * published by the Free Software Foundation.
 *
 * This code is distributed in the hope that it will be useful, but WITHOUT
 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
 * version 2 for more details (a copy is included in the LICENSE file that
 * accompanied this code).
 *
 * You should have received a copy of the GNU General Public License version
 * 2 along with this work; if not, write to the Free Software Foundation,
 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
 *
19 20 21
 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
 * or visit www.oracle.com if you need additional information or have any
 * questions.
D
duke 已提交
22 23 24
 *
 */

25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43
#include "precompiled.hpp"
#include "c1/c1_CFGPrinter.hpp"
#include "c1/c1_CodeStubs.hpp"
#include "c1/c1_Compilation.hpp"
#include "c1/c1_FrameMap.hpp"
#include "c1/c1_IR.hpp"
#include "c1/c1_LIRGenerator.hpp"
#include "c1/c1_LinearScan.hpp"
#include "c1/c1_ValueStack.hpp"
#include "utilities/bitMap.inline.hpp"
#ifdef TARGET_ARCH_x86
# include "vmreg_x86.inline.hpp"
#endif
#ifdef TARGET_ARCH_sparc
# include "vmreg_sparc.inline.hpp"
#endif
#ifdef TARGET_ARCH_zero
# include "vmreg_zero.inline.hpp"
#endif
44 45 46 47 48 49
#ifdef TARGET_ARCH_arm
# include "vmreg_arm.inline.hpp"
#endif
#ifdef TARGET_ARCH_ppc
# include "vmreg_ppc.inline.hpp"
#endif
D
duke 已提交
50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77


#ifndef PRODUCT

  static LinearScanStatistic _stat_before_alloc;
  static LinearScanStatistic _stat_after_asign;
  static LinearScanStatistic _stat_final;

  static LinearScanTimers _total_timer;

  // helper macro for short definition of timer
  #define TIME_LINEAR_SCAN(timer_name)  TraceTime _block_timer("", _total_timer.timer(LinearScanTimers::timer_name), TimeLinearScan || TimeEachLinearScan, Verbose);

  // helper macro for short definition of trace-output inside code
  #define TRACE_LINEAR_SCAN(level, code)       \
    if (TraceLinearScanLevel >= level) {       \
      code;                                    \
    }

#else

  #define TIME_LINEAR_SCAN(timer_name)
  #define TRACE_LINEAR_SCAN(level, code)

#endif

// Map BasicType to spill size in 32-bit words, matching VMReg's notion of words
#ifdef _LP64
78
static int type2spill_size[T_CONFLICT+1]={ -1, 0, 0, 0, 1, 1, 1, 2, 1, 1, 1, 2, 2, 2, 0, 2,  1, 2, 1, -1};
D
duke 已提交
79
#else
80
static int type2spill_size[T_CONFLICT+1]={ -1, 0, 0, 0, 1, 1, 1, 2, 1, 1, 1, 2, 1, 1, 0, 1, -1, 1, 1, -1};
D
duke 已提交
81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98
#endif


// Implementation of LinearScan

LinearScan::LinearScan(IR* ir, LIRGenerator* gen, FrameMap* frame_map)
 : _compilation(ir->compilation())
 , _ir(ir)
 , _gen(gen)
 , _frame_map(frame_map)
 , _num_virtual_regs(gen->max_virtual_register_number())
 , _has_fpu_registers(false)
 , _num_calls(-1)
 , _max_spills(0)
 , _unused_spill_slot(-1)
 , _intervals(0)   // initialized later with correct length
 , _new_intervals_from_allocation(new IntervalList())
 , _sorted_intervals(NULL)
99
 , _needs_full_resort(false)
D
duke 已提交
100 101 102 103 104 105 106
 , _lir_ops(0)     // initialized later with correct length
 , _block_of_op(0) // initialized later with correct length
 , _has_info(0)
 , _has_call(0)
 , _scope_value_cache(0) // initialized later with correct length
 , _interval_in_loop(0, 0) // initialized later with correct length
 , _cached_blocks(*ir->linear_scan_order())
107
#ifdef X86
D
duke 已提交
108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138
 , _fpu_stack_allocator(NULL)
#endif
{
  assert(this->ir() != NULL,          "check if valid");
  assert(this->compilation() != NULL, "check if valid");
  assert(this->gen() != NULL,         "check if valid");
  assert(this->frame_map() != NULL,   "check if valid");
}


// ********** functions for converting LIR-Operands to register numbers
//
// Emulate a flat register file comprising physical integer registers,
// physical floating-point registers and virtual registers, in that order.
// Virtual registers already have appropriate numbers, since V0 is
// the number of physical registers.
// Returns -1 for hi word if opr is a single word operand.
//
// Note: the inverse operation (calculating an operand for register numbers)
//       is done in calc_operand_for_interval()

int LinearScan::reg_num(LIR_Opr opr) {
  assert(opr->is_register(), "should not call this otherwise");

  if (opr->is_virtual_register()) {
    assert(opr->vreg_number() >= nof_regs, "found a virtual register with a fixed-register number");
    return opr->vreg_number();
  } else if (opr->is_single_cpu()) {
    return opr->cpu_regnr();
  } else if (opr->is_double_cpu()) {
    return opr->cpu_regnrLo();
139
#ifdef X86
D
duke 已提交
140 141 142 143 144 145 146 147 148 149 150
  } else if (opr->is_single_xmm()) {
    return opr->fpu_regnr() + pd_first_xmm_reg;
  } else if (opr->is_double_xmm()) {
    return opr->fpu_regnrLo() + pd_first_xmm_reg;
#endif
  } else if (opr->is_single_fpu()) {
    return opr->fpu_regnr() + pd_first_fpu_reg;
  } else if (opr->is_double_fpu()) {
    return opr->fpu_regnrLo() + pd_first_fpu_reg;
  } else {
    ShouldNotReachHere();
151
    return -1;
D
duke 已提交
152 153 154 155 156 157 158 159 160 161 162 163
  }
}

int LinearScan::reg_numHi(LIR_Opr opr) {
  assert(opr->is_register(), "should not call this otherwise");

  if (opr->is_virtual_register()) {
    return -1;
  } else if (opr->is_single_cpu()) {
    return -1;
  } else if (opr->is_double_cpu()) {
    return opr->cpu_regnrHi();
164
#ifdef X86
D
duke 已提交
165 166 167 168 169 170 171 172 173 174 175
  } else if (opr->is_single_xmm()) {
    return -1;
  } else if (opr->is_double_xmm()) {
    return -1;
#endif
  } else if (opr->is_single_fpu()) {
    return -1;
  } else if (opr->is_double_fpu()) {
    return opr->fpu_regnrHi() + pd_first_fpu_reg;
  } else {
    ShouldNotReachHere();
176
    return -1;
D
duke 已提交
177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195
  }
}


// ********** functions for classification of intervals

bool LinearScan::is_precolored_interval(const Interval* i) {
  return i->reg_num() < LinearScan::nof_regs;
}

bool LinearScan::is_virtual_interval(const Interval* i) {
  return i->reg_num() >= LIR_OprDesc::vreg_base;
}

bool LinearScan::is_precolored_cpu_interval(const Interval* i) {
  return i->reg_num() < LinearScan::nof_cpu_regs;
}

bool LinearScan::is_virtual_cpu_interval(const Interval* i) {
196 197 198
#if defined(__SOFTFP__) || defined(E500V2)
  return i->reg_num() >= LIR_OprDesc::vreg_base;
#else
D
duke 已提交
199
  return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() != T_FLOAT && i->type() != T_DOUBLE);
200
#endif // __SOFTFP__ or E500V2
D
duke 已提交
201 202 203 204 205 206 207
}

bool LinearScan::is_precolored_fpu_interval(const Interval* i) {
  return i->reg_num() >= LinearScan::nof_cpu_regs && i->reg_num() < LinearScan::nof_regs;
}

bool LinearScan::is_virtual_fpu_interval(const Interval* i) {
208 209 210
#if defined(__SOFTFP__) || defined(E500V2)
  return false;
#else
D
duke 已提交
211
  return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() == T_FLOAT || i->type() == T_DOUBLE);
212
#endif // __SOFTFP__ or E500V2
D
duke 已提交
213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095
}

bool LinearScan::is_in_fpu_register(const Interval* i) {
  // fixed intervals not needed for FPU stack allocation
  return i->reg_num() >= nof_regs && pd_first_fpu_reg <= i->assigned_reg() && i->assigned_reg() <= pd_last_fpu_reg;
}

bool LinearScan::is_oop_interval(const Interval* i) {
  // fixed intervals never contain oops
  return i->reg_num() >= nof_regs && i->type() == T_OBJECT;
}


// ********** General helper functions

// compute next unused stack index that can be used for spilling
int LinearScan::allocate_spill_slot(bool double_word) {
  int spill_slot;
  if (double_word) {
    if ((_max_spills & 1) == 1) {
      // alignment of double-word values
      // the hole because of the alignment is filled with the next single-word value
      assert(_unused_spill_slot == -1, "wasting a spill slot");
      _unused_spill_slot = _max_spills;
      _max_spills++;
    }
    spill_slot = _max_spills;
    _max_spills += 2;

  } else if (_unused_spill_slot != -1) {
    // re-use hole that was the result of a previous double-word alignment
    spill_slot = _unused_spill_slot;
    _unused_spill_slot = -1;

  } else {
    spill_slot = _max_spills;
    _max_spills++;
  }

  int result = spill_slot + LinearScan::nof_regs + frame_map()->argcount();

  // the class OopMapValue uses only 11 bits for storing the name of the
  // oop location. So a stack slot bigger than 2^11 leads to an overflow
  // that is not reported in product builds. Prevent this by checking the
  // spill slot here (altough this value and the later used location name
  // are slightly different)
  if (result > 2000) {
    bailout("too many stack slots used");
  }

  return result;
}

void LinearScan::assign_spill_slot(Interval* it) {
  // assign the canonical spill slot of the parent (if a part of the interval
  // is already spilled) or allocate a new spill slot
  if (it->canonical_spill_slot() >= 0) {
    it->assign_reg(it->canonical_spill_slot());
  } else {
    int spill = allocate_spill_slot(type2spill_size[it->type()] == 2);
    it->set_canonical_spill_slot(spill);
    it->assign_reg(spill);
  }
}

void LinearScan::propagate_spill_slots() {
  if (!frame_map()->finalize_frame(max_spills())) {
    bailout("frame too large");
  }
}

// create a new interval with a predefined reg_num
// (only used for parent intervals that are created during the building phase)
Interval* LinearScan::create_interval(int reg_num) {
  assert(_intervals.at(reg_num) == NULL, "overwriting exisiting interval");

  Interval* interval = new Interval(reg_num);
  _intervals.at_put(reg_num, interval);

  // assign register number for precolored intervals
  if (reg_num < LIR_OprDesc::vreg_base) {
    interval->assign_reg(reg_num);
  }
  return interval;
}

// assign a new reg_num to the interval and append it to the list of intervals
// (only used for child intervals that are created during register allocation)
void LinearScan::append_interval(Interval* it) {
  it->set_reg_num(_intervals.length());
  _intervals.append(it);
  _new_intervals_from_allocation->append(it);
}

// copy the vreg-flags if an interval is split
void LinearScan::copy_register_flags(Interval* from, Interval* to) {
  if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::byte_reg)) {
    gen()->set_vreg_flag(to->reg_num(), LIRGenerator::byte_reg);
  }
  if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::callee_saved)) {
    gen()->set_vreg_flag(to->reg_num(), LIRGenerator::callee_saved);
  }

  // Note: do not copy the must_start_in_memory flag because it is not necessary for child
  //       intervals (only the very beginning of the interval must be in memory)
}


// ********** spill move optimization
// eliminate moves from register to stack if stack slot is known to be correct

// called during building of intervals
void LinearScan::change_spill_definition_pos(Interval* interval, int def_pos) {
  assert(interval->is_split_parent(), "can only be called for split parents");

  switch (interval->spill_state()) {
    case noDefinitionFound:
      assert(interval->spill_definition_pos() == -1, "must no be set before");
      interval->set_spill_definition_pos(def_pos);
      interval->set_spill_state(oneDefinitionFound);
      break;

    case oneDefinitionFound:
      assert(def_pos <= interval->spill_definition_pos(), "positions are processed in reverse order when intervals are created");
      if (def_pos < interval->spill_definition_pos() - 2) {
        // second definition found, so no spill optimization possible for this interval
        interval->set_spill_state(noOptimization);
      } else {
        // two consecutive definitions (because of two-operand LIR form)
        assert(block_of_op_with_id(def_pos) == block_of_op_with_id(interval->spill_definition_pos()), "block must be equal");
      }
      break;

    case noOptimization:
      // nothing to do
      break;

    default:
      assert(false, "other states not allowed at this time");
  }
}

// called during register allocation
void LinearScan::change_spill_state(Interval* interval, int spill_pos) {
  switch (interval->spill_state()) {
    case oneDefinitionFound: {
      int def_loop_depth = block_of_op_with_id(interval->spill_definition_pos())->loop_depth();
      int spill_loop_depth = block_of_op_with_id(spill_pos)->loop_depth();

      if (def_loop_depth < spill_loop_depth) {
        // the loop depth of the spilling position is higher then the loop depth
        // at the definition of the interval -> move write to memory out of loop
        // by storing at definitin of the interval
        interval->set_spill_state(storeAtDefinition);
      } else {
        // the interval is currently spilled only once, so for now there is no
        // reason to store the interval at the definition
        interval->set_spill_state(oneMoveInserted);
      }
      break;
    }

    case oneMoveInserted: {
      // the interval is spilled more then once, so it is better to store it to
      // memory at the definition
      interval->set_spill_state(storeAtDefinition);
      break;
    }

    case storeAtDefinition:
    case startInMemory:
    case noOptimization:
    case noDefinitionFound:
      // nothing to do
      break;

    default:
      assert(false, "other states not allowed at this time");
  }
}


bool LinearScan::must_store_at_definition(const Interval* i) {
  return i->is_split_parent() && i->spill_state() == storeAtDefinition;
}

// called once before asignment of register numbers
void LinearScan::eliminate_spill_moves() {
  TIME_LINEAR_SCAN(timer_eliminate_spill_moves);
  TRACE_LINEAR_SCAN(3, tty->print_cr("***** Eliminating unnecessary spill moves"));

  // collect all intervals that must be stored after their definion.
  // the list is sorted by Interval::spill_definition_pos
  Interval* interval;
  Interval* temp_list;
  create_unhandled_lists(&interval, &temp_list, must_store_at_definition, NULL);

#ifdef ASSERT
  Interval* prev = NULL;
  Interval* temp = interval;
  while (temp != Interval::end()) {
    assert(temp->spill_definition_pos() > 0, "invalid spill definition pos");
    if (prev != NULL) {
      assert(temp->from() >= prev->from(), "intervals not sorted");
      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");
    }

    assert(temp->canonical_spill_slot() >= LinearScan::nof_regs, "interval has no spill slot assigned");
    assert(temp->spill_definition_pos() >= temp->from(), "invalid order");
    assert(temp->spill_definition_pos() <= temp->from() + 2, "only intervals defined once at their start-pos can be optimized");

    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()));

    temp = temp->next();
  }
#endif

  LIR_InsertionBuffer insertion_buffer;
  int num_blocks = block_count();
  for (int i = 0; i < num_blocks; i++) {
    BlockBegin* block = block_at(i);
    LIR_OpList* instructions = block->lir()->instructions_list();
    int         num_inst = instructions->length();
    bool        has_new = false;

    // iterate all instructions of the block. skip the first because it is always a label
    for (int j = 1; j < num_inst; j++) {
      LIR_Op* op = instructions->at(j);
      int op_id = op->id();

      if (op_id == -1) {
        // remove move from register to stack if the stack slot is guaranteed to be correct.
        // only moves that have been inserted by LinearScan can be removed.
        assert(op->code() == lir_move, "only moves can have a op_id of -1");
        assert(op->as_Op1() != NULL, "move must be LIR_Op1");
        assert(op->as_Op1()->result_opr()->is_virtual(), "LinearScan inserts only moves to virtual registers");

        LIR_Op1* op1 = (LIR_Op1*)op;
        Interval* interval = interval_at(op1->result_opr()->vreg_number());

        if (interval->assigned_reg() >= LinearScan::nof_regs && interval->always_in_memory()) {
          // move target is a stack slot that is always correct, so eliminate instruction
          TRACE_LINEAR_SCAN(4, tty->print_cr("eliminating move from interval %d to %d", op1->in_opr()->vreg_number(), op1->result_opr()->vreg_number()));
          instructions->at_put(j, NULL); // NULL-instructions are deleted by assign_reg_num
        }

      } else {
        // insert move from register to stack just after the beginning of the interval
        assert(interval == Interval::end() || interval->spill_definition_pos() >= op_id, "invalid order");
        assert(interval == Interval::end() || (interval->is_split_parent() && interval->spill_state() == storeAtDefinition), "invalid interval");

        while (interval != Interval::end() && interval->spill_definition_pos() == op_id) {
          if (!has_new) {
            // prepare insertion buffer (appended when all instructions of the block are processed)
            insertion_buffer.init(block->lir());
            has_new = true;
          }

          LIR_Opr from_opr = operand_for_interval(interval);
          LIR_Opr to_opr = canonical_spill_opr(interval);
          assert(from_opr->is_fixed_cpu() || from_opr->is_fixed_fpu(), "from operand must be a register");
          assert(to_opr->is_stack(), "to operand must be a stack slot");

          insertion_buffer.move(j, from_opr, to_opr);
          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));

          interval = interval->next();
        }
      }
    } // end of instruction iteration

    if (has_new) {
      block->lir()->append(&insertion_buffer);
    }
  } // end of block iteration

  assert(interval == Interval::end(), "missed an interval");
}


// ********** Phase 1: number all instructions in all blocks
// Compute depth-first and linear scan block orders, and number LIR_Op nodes for linear scan.

void LinearScan::number_instructions() {
  {
    // dummy-timer to measure the cost of the timer itself
    // (this time is then subtracted from all other timers to get the real value)
    TIME_LINEAR_SCAN(timer_do_nothing);
  }
  TIME_LINEAR_SCAN(timer_number_instructions);

  // Assign IDs to LIR nodes and build a mapping, lir_ops, from ID to LIR_Op node.
  int num_blocks = block_count();
  int num_instructions = 0;
  int i;
  for (i = 0; i < num_blocks; i++) {
    num_instructions += block_at(i)->lir()->instructions_list()->length();
  }

  // initialize with correct length
  _lir_ops = LIR_OpArray(num_instructions);
  _block_of_op = BlockBeginArray(num_instructions);

  int op_id = 0;
  int idx = 0;

  for (i = 0; i < num_blocks; i++) {
    BlockBegin* block = block_at(i);
    block->set_first_lir_instruction_id(op_id);
    LIR_OpList* instructions = block->lir()->instructions_list();

    int num_inst = instructions->length();
    for (int j = 0; j < num_inst; j++) {
      LIR_Op* op = instructions->at(j);
      op->set_id(op_id);

      _lir_ops.at_put(idx, op);
      _block_of_op.at_put(idx, block);
      assert(lir_op_with_id(op_id) == op, "must match");

      idx++;
      op_id += 2; // numbering of lir_ops by two
    }
    block->set_last_lir_instruction_id(op_id - 2);
  }
  assert(idx == num_instructions, "must match");
  assert(idx * 2 == op_id, "must match");

  _has_call = BitMap(num_instructions); _has_call.clear();
  _has_info = BitMap(num_instructions); _has_info.clear();
}


// ********** Phase 2: compute local live sets separately for each block
// (sets live_gen and live_kill for each block)

void LinearScan::set_live_gen_kill(Value value, LIR_Op* op, BitMap& live_gen, BitMap& live_kill) {
  LIR_Opr opr = value->operand();
  Constant* con = value->as_Constant();

  // check some asumptions about debug information
  assert(!value->type()->is_illegal(), "if this local is used by the interpreter it shouldn't be of indeterminate type");
  assert(con == NULL || opr->is_virtual() || opr->is_constant() || opr->is_illegal(), "asumption: Constant instructions have only constant operands");
  assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands");

  if ((con == NULL || con->is_pinned()) && opr->is_register()) {
    assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
    int reg = opr->vreg_number();
    if (!live_kill.at(reg)) {
      live_gen.set_bit(reg);
      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));
    }
  }
}


void LinearScan::compute_local_live_sets() {
  TIME_LINEAR_SCAN(timer_compute_local_live_sets);

  int  num_blocks = block_count();
  int  live_size = live_set_size();
  bool local_has_fpu_registers = false;
  int  local_num_calls = 0;
  LIR_OpVisitState visitor;

  BitMap2D local_interval_in_loop = BitMap2D(_num_virtual_regs, num_loops());
  local_interval_in_loop.clear();

  // iterate all blocks
  for (int i = 0; i < num_blocks; i++) {
    BlockBegin* block = block_at(i);

    BitMap live_gen(live_size);  live_gen.clear();
    BitMap live_kill(live_size); live_kill.clear();

    if (block->is_set(BlockBegin::exception_entry_flag)) {
      // Phi functions at the begin of an exception handler are
      // implicitly defined (= killed) at the beginning of the block.
      for_each_phi_fun(block, phi,
        live_kill.set_bit(phi->operand()->vreg_number())
      );
    }

    LIR_OpList* instructions = block->lir()->instructions_list();
    int num_inst = instructions->length();

    // iterate all instructions of the block. skip the first because it is always a label
    assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label");
    for (int j = 1; j < num_inst; j++) {
      LIR_Op* op = instructions->at(j);

      // visit operation to collect all operands
      visitor.visit(op);

      if (visitor.has_call()) {
        _has_call.set_bit(op->id() >> 1);
        local_num_calls++;
      }
      if (visitor.info_count() > 0) {
        _has_info.set_bit(op->id() >> 1);
      }

      // iterate input operands of instruction
      int k, n, reg;
      n = visitor.opr_count(LIR_OpVisitState::inputMode);
      for (k = 0; k < n; k++) {
        LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k);
        assert(opr->is_register(), "visitor should only return register operands");

        if (opr->is_virtual_register()) {
          assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
          reg = opr->vreg_number();
          if (!live_kill.at(reg)) {
            live_gen.set_bit(reg);
            TRACE_LINEAR_SCAN(4, tty->print_cr("  Setting live_gen for register %d at instruction %d", reg, op->id()));
          }
          if (block->loop_index() >= 0) {
            local_interval_in_loop.set_bit(reg, block->loop_index());
          }
          local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu();
        }

#ifdef ASSERT
        // fixed intervals are never live at block boundaries, so
        // they need not be processed in live sets.
        // this is checked by these assertions to be sure about it.
        // the entry block may have incoming values in registers, which is ok.
        if (!opr->is_virtual_register() && block != ir()->start()) {
          reg = reg_num(opr);
          if (is_processed_reg_num(reg)) {
            assert(live_kill.at(reg), "using fixed register that is not defined in this block");
          }
          reg = reg_numHi(opr);
          if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
            assert(live_kill.at(reg), "using fixed register that is not defined in this block");
          }
        }
#endif
      }

      // Add uses of live locals from interpreter's point of view for proper debug information generation
      n = visitor.info_count();
      for (k = 0; k < n; k++) {
        CodeEmitInfo* info = visitor.info_at(k);
        ValueStack* stack = info->stack();
        for_each_state_value(stack, value,
          set_live_gen_kill(value, op, live_gen, live_kill)
        );
      }

      // iterate temp operands of instruction
      n = visitor.opr_count(LIR_OpVisitState::tempMode);
      for (k = 0; k < n; k++) {
        LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k);
        assert(opr->is_register(), "visitor should only return register operands");

        if (opr->is_virtual_register()) {
          assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
          reg = opr->vreg_number();
          live_kill.set_bit(reg);
          if (block->loop_index() >= 0) {
            local_interval_in_loop.set_bit(reg, block->loop_index());
          }
          local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu();
        }

#ifdef ASSERT
        // fixed intervals are never live at block boundaries, so
        // they need not be processed in live sets
        // process them only in debug mode so that this can be checked
        if (!opr->is_virtual_register()) {
          reg = reg_num(opr);
          if (is_processed_reg_num(reg)) {
            live_kill.set_bit(reg_num(opr));
          }
          reg = reg_numHi(opr);
          if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
            live_kill.set_bit(reg);
          }
        }
#endif
      }

      // iterate output operands of instruction
      n = visitor.opr_count(LIR_OpVisitState::outputMode);
      for (k = 0; k < n; k++) {
        LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k);
        assert(opr->is_register(), "visitor should only return register operands");

        if (opr->is_virtual_register()) {
          assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
          reg = opr->vreg_number();
          live_kill.set_bit(reg);
          if (block->loop_index() >= 0) {
            local_interval_in_loop.set_bit(reg, block->loop_index());
          }
          local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu();
        }

#ifdef ASSERT
        // fixed intervals are never live at block boundaries, so
        // they need not be processed in live sets
        // process them only in debug mode so that this can be checked
        if (!opr->is_virtual_register()) {
          reg = reg_num(opr);
          if (is_processed_reg_num(reg)) {
            live_kill.set_bit(reg_num(opr));
          }
          reg = reg_numHi(opr);
          if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
            live_kill.set_bit(reg);
          }
        }
#endif
      }
    } // end of instruction iteration

    block->set_live_gen (live_gen);
    block->set_live_kill(live_kill);
    block->set_live_in  (BitMap(live_size)); block->live_in().clear();
    block->set_live_out (BitMap(live_size)); block->live_out().clear();

    TRACE_LINEAR_SCAN(4, tty->print("live_gen  B%d ", block->block_id()); print_bitmap(block->live_gen()));
    TRACE_LINEAR_SCAN(4, tty->print("live_kill B%d ", block->block_id()); print_bitmap(block->live_kill()));
  } // end of block iteration

  // propagate local calculated information into LinearScan object
  _has_fpu_registers = local_has_fpu_registers;
  compilation()->set_has_fpu_code(local_has_fpu_registers);

  _num_calls = local_num_calls;
  _interval_in_loop = local_interval_in_loop;
}


// ********** Phase 3: perform a backward dataflow analysis to compute global live sets
// (sets live_in and live_out for each block)

void LinearScan::compute_global_live_sets() {
  TIME_LINEAR_SCAN(timer_compute_global_live_sets);

  int  num_blocks = block_count();
  bool change_occurred;
  bool change_occurred_in_block;
  int  iteration_count = 0;
  BitMap live_out(live_set_size()); live_out.clear(); // scratch set for calculations

  // Perform a backward dataflow analysis to compute live_out and live_in for each block.
  // The loop is executed until a fixpoint is reached (no changes in an iteration)
  // Exception handlers must be processed because not all live values are
  // present in the state array, e.g. because of global value numbering
  do {
    change_occurred = false;

    // iterate all blocks in reverse order
    for (int i = num_blocks - 1; i >= 0; i--) {
      BlockBegin* block = block_at(i);

      change_occurred_in_block = false;

      // live_out(block) is the union of live_in(sux), for successors sux of block
      int n = block->number_of_sux();
      int e = block->number_of_exception_handlers();
      if (n + e > 0) {
        // block has successors
        if (n > 0) {
          live_out.set_from(block->sux_at(0)->live_in());
          for (int j = 1; j < n; j++) {
            live_out.set_union(block->sux_at(j)->live_in());
          }
        } else {
          live_out.clear();
        }
        for (int j = 0; j < e; j++) {
          live_out.set_union(block->exception_handler_at(j)->live_in());
        }

        if (!block->live_out().is_same(live_out)) {
          // A change occurred.  Swap the old and new live out sets to avoid copying.
          BitMap temp = block->live_out();
          block->set_live_out(live_out);
          live_out = temp;

          change_occurred = true;
          change_occurred_in_block = true;
        }
      }

      if (iteration_count == 0 || change_occurred_in_block) {
        // live_in(block) is the union of live_gen(block) with (live_out(block) & !live_kill(block))
        // note: live_in has to be computed only in first iteration or if live_out has changed!
        BitMap live_in = block->live_in();
        live_in.set_from(block->live_out());
        live_in.set_difference(block->live_kill());
        live_in.set_union(block->live_gen());
      }

#ifndef PRODUCT
      if (TraceLinearScanLevel >= 4) {
        char c = ' ';
        if (iteration_count == 0 || change_occurred_in_block) {
          c = '*';
        }
        tty->print("(%d) live_in%c  B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_in());
        tty->print("(%d) live_out%c B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_out());
      }
#endif
    }
    iteration_count++;

    if (change_occurred && iteration_count > 50) {
      BAILOUT("too many iterations in compute_global_live_sets");
    }
  } while (change_occurred);


#ifdef ASSERT
  // check that fixed intervals are not live at block boundaries
  // (live set must be empty at fixed intervals)
  for (int i = 0; i < num_blocks; i++) {
    BlockBegin* block = block_at(i);
    for (int j = 0; j < LIR_OprDesc::vreg_base; j++) {
      assert(block->live_in().at(j)  == false, "live_in  set of fixed register must be empty");
      assert(block->live_out().at(j) == false, "live_out set of fixed register must be empty");
      assert(block->live_gen().at(j) == false, "live_gen set of fixed register must be empty");
    }
  }
#endif

  // check that the live_in set of the first block is empty
  BitMap live_in_args(ir()->start()->live_in().size());
  live_in_args.clear();
  if (!ir()->start()->live_in().is_same(live_in_args)) {
#ifdef ASSERT
    tty->print_cr("Error: live_in set of first block must be empty (when this fails, virtual registers are used before they are defined)");
    tty->print_cr("affected registers:");
    print_bitmap(ir()->start()->live_in());

    // print some additional information to simplify debugging
    for (unsigned int i = 0; i < ir()->start()->live_in().size(); i++) {
      if (ir()->start()->live_in().at(i)) {
        Instruction* instr = gen()->instruction_for_vreg(i);
        tty->print_cr("* vreg %d (HIR instruction %c%d)", i, instr == NULL ? ' ' : instr->type()->tchar(), instr == NULL ? 0 : instr->id());

        for (int j = 0; j < num_blocks; j++) {
          BlockBegin* block = block_at(j);
          if (block->live_gen().at(i)) {
            tty->print_cr("  used in block B%d", block->block_id());
          }
          if (block->live_kill().at(i)) {
            tty->print_cr("  defined in block B%d", block->block_id());
          }
        }
      }
    }

#endif
    // when this fails, virtual registers are used before they are defined.
    assert(false, "live_in set of first block must be empty");
    // bailout of if this occurs in product mode.
    bailout("live_in set of first block not empty");
  }
}


// ********** Phase 4: build intervals
// (fills the list _intervals)

void LinearScan::add_use(Value value, int from, int to, IntervalUseKind use_kind) {
  assert(!value->type()->is_illegal(), "if this value is used by the interpreter it shouldn't be of indeterminate type");
  LIR_Opr opr = value->operand();
  Constant* con = value->as_Constant();

  if ((con == NULL || con->is_pinned()) && opr->is_register()) {
    assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
    add_use(opr, from, to, use_kind);
  }
}


void LinearScan::add_def(LIR_Opr opr, int def_pos, IntervalUseKind use_kind) {
  TRACE_LINEAR_SCAN(2, tty->print(" def "); opr->print(tty); tty->print_cr(" def_pos %d (%d)", def_pos, use_kind));
  assert(opr->is_register(), "should not be called otherwise");

  if (opr->is_virtual_register()) {
    assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
    add_def(opr->vreg_number(), def_pos, use_kind, opr->type_register());

  } else {
    int reg = reg_num(opr);
    if (is_processed_reg_num(reg)) {
      add_def(reg, def_pos, use_kind, opr->type_register());
    }
    reg = reg_numHi(opr);
    if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
      add_def(reg, def_pos, use_kind, opr->type_register());
    }
  }
}

void LinearScan::add_use(LIR_Opr opr, int from, int to, IntervalUseKind use_kind) {
  TRACE_LINEAR_SCAN(2, tty->print(" use "); opr->print(tty); tty->print_cr(" from %d to %d (%d)", from, to, use_kind));
  assert(opr->is_register(), "should not be called otherwise");

  if (opr->is_virtual_register()) {
    assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
    add_use(opr->vreg_number(), from, to, use_kind, opr->type_register());

  } else {
    int reg = reg_num(opr);
    if (is_processed_reg_num(reg)) {
      add_use(reg, from, to, use_kind, opr->type_register());
    }
    reg = reg_numHi(opr);
    if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
      add_use(reg, from, to, use_kind, opr->type_register());
    }
  }
}

void LinearScan::add_temp(LIR_Opr opr, int temp_pos, IntervalUseKind use_kind) {
  TRACE_LINEAR_SCAN(2, tty->print(" temp "); opr->print(tty); tty->print_cr(" temp_pos %d (%d)", temp_pos, use_kind));
  assert(opr->is_register(), "should not be called otherwise");

  if (opr->is_virtual_register()) {
    assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
    add_temp(opr->vreg_number(), temp_pos, use_kind, opr->type_register());

  } else {
    int reg = reg_num(opr);
    if (is_processed_reg_num(reg)) {
      add_temp(reg, temp_pos, use_kind, opr->type_register());
    }
    reg = reg_numHi(opr);
    if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
      add_temp(reg, temp_pos, use_kind, opr->type_register());
    }
  }
}


void LinearScan::add_def(int reg_num, int def_pos, IntervalUseKind use_kind, BasicType type) {
  Interval* interval = interval_at(reg_num);
  if (interval != NULL) {
    assert(interval->reg_num() == reg_num, "wrong interval");

    if (type != T_ILLEGAL) {
      interval->set_type(type);
    }

    Range* r = interval->first();
    if (r->from() <= def_pos) {
      // Update the starting point (when a range is first created for a use, its
      // start is the beginning of the current block until a def is encountered.)
      r->set_from(def_pos);
      interval->add_use_pos(def_pos, use_kind);

    } else {
      // Dead value - make vacuous interval
      // also add use_kind for dead intervals
      interval->add_range(def_pos, def_pos + 1);
      interval->add_use_pos(def_pos, use_kind);
      TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: def of reg %d at %d occurs without use", reg_num, def_pos));
    }

  } else {
    // Dead value - make vacuous interval
    // also add use_kind for dead intervals
    interval = create_interval(reg_num);
    if (type != T_ILLEGAL) {
      interval->set_type(type);
    }

    interval->add_range(def_pos, def_pos + 1);
    interval->add_use_pos(def_pos, use_kind);
    TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: dead value %d at %d in live intervals", reg_num, def_pos));
  }

  change_spill_definition_pos(interval, def_pos);
  if (use_kind == noUse && interval->spill_state() <= startInMemory) {
        // detection of method-parameters and roundfp-results
        // TODO: move this directly to position where use-kind is computed
    interval->set_spill_state(startInMemory);
  }
}

void LinearScan::add_use(int reg_num, int from, int to, IntervalUseKind use_kind, BasicType type) {
  Interval* interval = interval_at(reg_num);
  if (interval == NULL) {
    interval = create_interval(reg_num);
  }
  assert(interval->reg_num() == reg_num, "wrong interval");

  if (type != T_ILLEGAL) {
    interval->set_type(type);
  }

  interval->add_range(from, to);
  interval->add_use_pos(to, use_kind);
}

void LinearScan::add_temp(int reg_num, int temp_pos, IntervalUseKind use_kind, BasicType type) {
  Interval* interval = interval_at(reg_num);
  if (interval == NULL) {
    interval = create_interval(reg_num);
  }
  assert(interval->reg_num() == reg_num, "wrong interval");

  if (type != T_ILLEGAL) {
    interval->set_type(type);
  }

  interval->add_range(temp_pos, temp_pos + 1);
  interval->add_use_pos(temp_pos, use_kind);
}


// the results of this functions are used for optimizing spilling and reloading
// if the functions return shouldHaveRegister and the interval is spilled,
// it is not reloaded to a register.
IntervalUseKind LinearScan::use_kind_of_output_operand(LIR_Op* op, LIR_Opr opr) {
  if (op->code() == lir_move) {
    assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1");
    LIR_Op1* move = (LIR_Op1*)op;
    LIR_Opr res = move->result_opr();
    bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory);

    if (result_in_memory) {
      // Begin of an interval with must_start_in_memory set.
      // This interval will always get a stack slot first, so return noUse.
      return noUse;

    } else if (move->in_opr()->is_stack()) {
      // method argument (condition must be equal to handle_method_arguments)
      return noUse;

    } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) {
      // Move from register to register
      if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) {
        // special handling of phi-function moves inside osr-entry blocks
        // input operand must have a register instead of output operand (leads to better register allocation)
        return shouldHaveRegister;
      }
    }
  }

  if (opr->is_virtual() &&
      gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::must_start_in_memory)) {
    // result is a stack-slot, so prevent immediate reloading
    return noUse;
  }

  // all other operands require a register
  return mustHaveRegister;
}

IntervalUseKind LinearScan::use_kind_of_input_operand(LIR_Op* op, LIR_Opr opr) {
  if (op->code() == lir_move) {
    assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1");
    LIR_Op1* move = (LIR_Op1*)op;
    LIR_Opr res = move->result_opr();
    bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory);

    if (result_in_memory) {
      // Move to an interval with must_start_in_memory set.
      // To avoid moves from stack to stack (not allowed) force the input operand to a register
      return mustHaveRegister;

    } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) {
      // Move from register to register
      if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) {
        // special handling of phi-function moves inside osr-entry blocks
        // input operand must have a register instead of output operand (leads to better register allocation)
        return mustHaveRegister;
      }

      // The input operand is not forced to a register (moves from stack to register are allowed),
      // but it is faster if the input operand is in a register
      return shouldHaveRegister;
    }
  }


1096
#ifdef X86
D
duke 已提交
1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140
  if (op->code() == lir_cmove) {
    // conditional moves can handle stack operands
    assert(op->result_opr()->is_register(), "result must always be in a register");
    return shouldHaveRegister;
  }

  // optimizations for second input operand of arithmehtic operations on Intel
  // this operand is allowed to be on the stack in some cases
  BasicType opr_type = opr->type_register();
  if (opr_type == T_FLOAT || opr_type == T_DOUBLE) {
    if ((UseSSE == 1 && opr_type == T_FLOAT) || UseSSE >= 2) {
      // SSE float instruction (T_DOUBLE only supported with SSE2)
      switch (op->code()) {
        case lir_cmp:
        case lir_add:
        case lir_sub:
        case lir_mul:
        case lir_div:
        {
          assert(op->as_Op2() != NULL, "must be LIR_Op2");
          LIR_Op2* op2 = (LIR_Op2*)op;
          if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) {
            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");
            return shouldHaveRegister;
          }
        }
      }
    } else {
      // FPU stack float instruction
      switch (op->code()) {
        case lir_add:
        case lir_sub:
        case lir_mul:
        case lir_div:
        {
          assert(op->as_Op2() != NULL, "must be LIR_Op2");
          LIR_Op2* op2 = (LIR_Op2*)op;
          if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) {
            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");
            return shouldHaveRegister;
          }
        }
      }
    }
1141 1142 1143 1144
    // We want to sometimes use logical operations on pointers, in particular in GC barriers.
    // Since 64bit logical operations do not current support operands on stack, we have to make sure
    // T_OBJECT doesn't get spilled along with T_LONG.
  } else if (opr_type != T_LONG LP64_ONLY(&& opr_type != T_OBJECT)) {
D
duke 已提交
1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162
    // integer instruction (note: long operands must always be in register)
    switch (op->code()) {
      case lir_cmp:
      case lir_add:
      case lir_sub:
      case lir_logic_and:
      case lir_logic_or:
      case lir_logic_xor:
      {
        assert(op->as_Op2() != NULL, "must be LIR_Op2");
        LIR_Op2* op2 = (LIR_Op2*)op;
        if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) {
          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");
          return shouldHaveRegister;
        }
      }
    }
  }
1163
#endif // X86
D
duke 已提交
1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284

  // all other operands require a register
  return mustHaveRegister;
}


void LinearScan::handle_method_arguments(LIR_Op* op) {
  // special handling for method arguments (moves from stack to virtual register):
  // the interval gets no register assigned, but the stack slot.
  // it is split before the first use by the register allocator.

  if (op->code() == lir_move) {
    assert(op->as_Op1() != NULL, "must be LIR_Op1");
    LIR_Op1* move = (LIR_Op1*)op;

    if (move->in_opr()->is_stack()) {
#ifdef ASSERT
      int arg_size = compilation()->method()->arg_size();
      LIR_Opr o = move->in_opr();
      if (o->is_single_stack()) {
        assert(o->single_stack_ix() >= 0 && o->single_stack_ix() < arg_size, "out of range");
      } else if (o->is_double_stack()) {
        assert(o->double_stack_ix() >= 0 && o->double_stack_ix() < arg_size, "out of range");
      } else {
        ShouldNotReachHere();
      }

      assert(move->id() > 0, "invalid id");
      assert(block_of_op_with_id(move->id())->number_of_preds() == 0, "move from stack must be in first block");
      assert(move->result_opr()->is_virtual(), "result of move must be a virtual register");

      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())));
#endif

      Interval* interval = interval_at(reg_num(move->result_opr()));

      int stack_slot = LinearScan::nof_regs + (move->in_opr()->is_single_stack() ? move->in_opr()->single_stack_ix() : move->in_opr()->double_stack_ix());
      interval->set_canonical_spill_slot(stack_slot);
      interval->assign_reg(stack_slot);
    }
  }
}

void LinearScan::handle_doubleword_moves(LIR_Op* op) {
  // special handling for doubleword move from memory to register:
  // in this case the registers of the input address and the result
  // registers must not overlap -> add a temp range for the input registers
  if (op->code() == lir_move) {
    assert(op->as_Op1() != NULL, "must be LIR_Op1");
    LIR_Op1* move = (LIR_Op1*)op;

    if (move->result_opr()->is_double_cpu() && move->in_opr()->is_pointer()) {
      LIR_Address* address = move->in_opr()->as_address_ptr();
      if (address != NULL) {
        if (address->base()->is_valid()) {
          add_temp(address->base(), op->id(), noUse);
        }
        if (address->index()->is_valid()) {
          add_temp(address->index(), op->id(), noUse);
        }
      }
    }
  }
}

void LinearScan::add_register_hints(LIR_Op* op) {
  switch (op->code()) {
    case lir_move:      // fall through
    case lir_convert: {
      assert(op->as_Op1() != NULL, "lir_move, lir_convert must be LIR_Op1");
      LIR_Op1* move = (LIR_Op1*)op;

      LIR_Opr move_from = move->in_opr();
      LIR_Opr move_to = move->result_opr();

      if (move_to->is_register() && move_from->is_register()) {
        Interval* from = interval_at(reg_num(move_from));
        Interval* to = interval_at(reg_num(move_to));
        if (from != NULL && to != NULL) {
          to->set_register_hint(from);
          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()));
        }
      }
      break;
    }
    case lir_cmove: {
      assert(op->as_Op2() != NULL, "lir_cmove must be LIR_Op2");
      LIR_Op2* cmove = (LIR_Op2*)op;

      LIR_Opr move_from = cmove->in_opr1();
      LIR_Opr move_to = cmove->result_opr();

      if (move_to->is_register() && move_from->is_register()) {
        Interval* from = interval_at(reg_num(move_from));
        Interval* to = interval_at(reg_num(move_to));
        if (from != NULL && to != NULL) {
          to->set_register_hint(from);
          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()));
        }
      }
      break;
    }
  }
}


void LinearScan::build_intervals() {
  TIME_LINEAR_SCAN(timer_build_intervals);

  // initialize interval list with expected number of intervals
  // (32 is added to have some space for split children without having to resize the list)
  _intervals = IntervalList(num_virtual_regs() + 32);
  // initialize all slots that are used by build_intervals
  _intervals.at_put_grow(num_virtual_regs() - 1, NULL, NULL);

  // create a list with all caller-save registers (cpu, fpu, xmm)
  // when an instruction is a call, a temp range is created for all these registers
  int num_caller_save_registers = 0;
  int caller_save_registers[LinearScan::nof_regs];

  int i;
1285
  for (i = 0; i < FrameMap::nof_caller_save_cpu_regs(); i++) {
D
duke 已提交
1286 1287 1288 1289 1290 1291 1292 1293 1294 1295
    LIR_Opr opr = FrameMap::caller_save_cpu_reg_at(i);
    assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands");
    assert(reg_numHi(opr) == -1, "missing addition of range for hi-register");
    caller_save_registers[num_caller_save_registers++] = reg_num(opr);
  }

  // temp ranges for fpu registers are only created when the method has
  // virtual fpu operands. Otherwise no allocation for fpu registers is
  // perfomed and so the temp ranges would be useless
  if (has_fpu_registers()) {
1296
#ifdef X86
D
duke 已提交
1297 1298 1299 1300 1301 1302 1303 1304
    if (UseSSE < 2) {
#endif
      for (i = 0; i < FrameMap::nof_caller_save_fpu_regs; i++) {
        LIR_Opr opr = FrameMap::caller_save_fpu_reg_at(i);
        assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands");
        assert(reg_numHi(opr) == -1, "missing addition of range for hi-register");
        caller_save_registers[num_caller_save_registers++] = reg_num(opr);
      }
1305
#ifdef X86
D
duke 已提交
1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333
    }
    if (UseSSE > 0) {
      for (i = 0; i < FrameMap::nof_caller_save_xmm_regs; i++) {
        LIR_Opr opr = FrameMap::caller_save_xmm_reg_at(i);
        assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands");
        assert(reg_numHi(opr) == -1, "missing addition of range for hi-register");
        caller_save_registers[num_caller_save_registers++] = reg_num(opr);
      }
    }
#endif
  }
  assert(num_caller_save_registers <= LinearScan::nof_regs, "out of bounds");


  LIR_OpVisitState visitor;

  // iterate all blocks in reverse order
  for (i = block_count() - 1; i >= 0; i--) {
    BlockBegin* block = block_at(i);
    LIR_OpList* instructions = block->lir()->instructions_list();
    int         block_from =   block->first_lir_instruction_id();
    int         block_to =     block->last_lir_instruction_id();

    assert(block_from == instructions->at(0)->id(), "must be");
    assert(block_to   == instructions->at(instructions->length() - 1)->id(), "must be");

    // Update intervals for registers live at the end of this block;
    BitMap live = block->live_out();
1334 1335
    int size = (int)live.size();
    for (int number = (int)live.get_next_one_offset(0, size); number < size; number = (int)live.get_next_one_offset(number + 1, size)) {
D
duke 已提交
1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531
      assert(live.at(number), "should not stop here otherwise");
      assert(number >= LIR_OprDesc::vreg_base, "fixed intervals must not be live on block bounds");
      TRACE_LINEAR_SCAN(2, tty->print_cr("live in %d to %d", number, block_to + 2));

      add_use(number, block_from, block_to + 2, noUse, T_ILLEGAL);

      // add special use positions for loop-end blocks when the
      // interval is used anywhere inside this loop.  It's possible
      // that the block was part of a non-natural loop, so it might
      // have an invalid loop index.
      if (block->is_set(BlockBegin::linear_scan_loop_end_flag) &&
          block->loop_index() != -1 &&
          is_interval_in_loop(number, block->loop_index())) {
        interval_at(number)->add_use_pos(block_to + 1, loopEndMarker);
      }
    }

    // iterate all instructions of the block in reverse order.
    // skip the first instruction because it is always a label
    // definitions of intervals are processed before uses
    assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label");
    for (int j = instructions->length() - 1; j >= 1; j--) {
      LIR_Op* op = instructions->at(j);
      int op_id = op->id();

      // visit operation to collect all operands
      visitor.visit(op);

      // add a temp range for each register if operation destroys caller-save registers
      if (visitor.has_call()) {
        for (int k = 0; k < num_caller_save_registers; k++) {
          add_temp(caller_save_registers[k], op_id, noUse, T_ILLEGAL);
        }
        TRACE_LINEAR_SCAN(4, tty->print_cr("operation destroys all caller-save registers"));
      }

      // Add any platform dependent temps
      pd_add_temps(op);

      // visit definitions (output and temp operands)
      int k, n;
      n = visitor.opr_count(LIR_OpVisitState::outputMode);
      for (k = 0; k < n; k++) {
        LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k);
        assert(opr->is_register(), "visitor should only return register operands");
        add_def(opr, op_id, use_kind_of_output_operand(op, opr));
      }

      n = visitor.opr_count(LIR_OpVisitState::tempMode);
      for (k = 0; k < n; k++) {
        LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k);
        assert(opr->is_register(), "visitor should only return register operands");
        add_temp(opr, op_id, mustHaveRegister);
      }

      // visit uses (input operands)
      n = visitor.opr_count(LIR_OpVisitState::inputMode);
      for (k = 0; k < n; k++) {
        LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k);
        assert(opr->is_register(), "visitor should only return register operands");
        add_use(opr, block_from, op_id, use_kind_of_input_operand(op, opr));
      }

      // Add uses of live locals from interpreter's point of view for proper
      // debug information generation
      // Treat these operands as temp values (if the life range is extended
      // to a call site, the value would be in a register at the call otherwise)
      n = visitor.info_count();
      for (k = 0; k < n; k++) {
        CodeEmitInfo* info = visitor.info_at(k);
        ValueStack* stack = info->stack();
        for_each_state_value(stack, value,
          add_use(value, block_from, op_id + 1, noUse);
        );
      }

      // special steps for some instructions (especially moves)
      handle_method_arguments(op);
      handle_doubleword_moves(op);
      add_register_hints(op);

    } // end of instruction iteration
  } // end of block iteration


  // add the range [0, 1[ to all fixed intervals
  // -> the register allocator need not handle unhandled fixed intervals
  for (int n = 0; n < LinearScan::nof_regs; n++) {
    Interval* interval = interval_at(n);
    if (interval != NULL) {
      interval->add_range(0, 1);
    }
  }
}


// ********** Phase 5: actual register allocation

int LinearScan::interval_cmp(Interval** a, Interval** b) {
  if (*a != NULL) {
    if (*b != NULL) {
      return (*a)->from() - (*b)->from();
    } else {
      return -1;
    }
  } else {
    if (*b != NULL) {
      return 1;
    } else {
      return 0;
    }
  }
}

#ifndef PRODUCT
bool LinearScan::is_sorted(IntervalArray* intervals) {
  int from = -1;
  int i, j;
  for (i = 0; i < intervals->length(); i ++) {
    Interval* it = intervals->at(i);
    if (it != NULL) {
      if (from > it->from()) {
        assert(false, "");
        return false;
      }
      from = it->from();
    }
  }

  // check in both directions if sorted list and unsorted list contain same intervals
  for (i = 0; i < interval_count(); i++) {
    if (interval_at(i) != NULL) {
      int num_found = 0;
      for (j = 0; j < intervals->length(); j++) {
        if (interval_at(i) == intervals->at(j)) {
          num_found++;
        }
      }
      assert(num_found == 1, "lists do not contain same intervals");
    }
  }
  for (j = 0; j < intervals->length(); j++) {
    int num_found = 0;
    for (i = 0; i < interval_count(); i++) {
      if (interval_at(i) == intervals->at(j)) {
        num_found++;
      }
    }
    assert(num_found == 1, "lists do not contain same intervals");
  }

  return true;
}
#endif

void LinearScan::add_to_list(Interval** first, Interval** prev, Interval* interval) {
  if (*prev != NULL) {
    (*prev)->set_next(interval);
  } else {
    *first = interval;
  }
  *prev = interval;
}

void LinearScan::create_unhandled_lists(Interval** list1, Interval** list2, bool (is_list1)(const Interval* i), bool (is_list2)(const Interval* i)) {
  assert(is_sorted(_sorted_intervals), "interval list is not sorted");

  *list1 = *list2 = Interval::end();

  Interval* list1_prev = NULL;
  Interval* list2_prev = NULL;
  Interval* v;

  const int n = _sorted_intervals->length();
  for (int i = 0; i < n; i++) {
    v = _sorted_intervals->at(i);
    if (v == NULL) continue;

    if (is_list1(v)) {
      add_to_list(list1, &list1_prev, v);
    } else if (is_list2 == NULL || is_list2(v)) {
      add_to_list(list2, &list2_prev, v);
    }
  }

  if (list1_prev != NULL) list1_prev->set_next(Interval::end());
  if (list2_prev != NULL) list2_prev->set_next(Interval::end());

  assert(list1_prev == NULL || list1_prev->next() == Interval::end(), "linear list ends not with sentinel");
  assert(list2_prev == NULL || list2_prev->next() == Interval::end(), "linear list ends not with sentinel");
}


void LinearScan::sort_intervals_before_allocation() {
  TIME_LINEAR_SCAN(timer_sort_intervals_before);

1532 1533 1534 1535 1536 1537 1538 1539
  if (_needs_full_resort) {
    // There is no known reason why this should occur but just in case...
    assert(false, "should never occur");
    // Re-sort existing interval list because an Interval::from() has changed
    _sorted_intervals->sort(interval_cmp);
    _needs_full_resort = false;
  }

D
duke 已提交
1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578
  IntervalList* unsorted_list = &_intervals;
  int unsorted_len = unsorted_list->length();
  int sorted_len = 0;
  int unsorted_idx;
  int sorted_idx = 0;
  int sorted_from_max = -1;

  // calc number of items for sorted list (sorted list must not contain NULL values)
  for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) {
    if (unsorted_list->at(unsorted_idx) != NULL) {
      sorted_len++;
    }
  }
  IntervalArray* sorted_list = new IntervalArray(sorted_len);

  // special sorting algorithm: the original interval-list is almost sorted,
  // only some intervals are swapped. So this is much faster than a complete QuickSort
  for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) {
    Interval* cur_interval = unsorted_list->at(unsorted_idx);

    if (cur_interval != NULL) {
      int cur_from = cur_interval->from();

      if (sorted_from_max <= cur_from) {
        sorted_list->at_put(sorted_idx++, cur_interval);
        sorted_from_max = cur_interval->from();
      } else {
        // the asumption that the intervals are already sorted failed,
        // so this interval must be sorted in manually
        int j;
        for (j = sorted_idx - 1; j >= 0 && cur_from < sorted_list->at(j)->from(); j--) {
          sorted_list->at_put(j + 1, sorted_list->at(j));
        }
        sorted_list->at_put(j + 1, cur_interval);
        sorted_idx++;
      }
    }
  }
  _sorted_intervals = sorted_list;
1579
  assert(is_sorted(_sorted_intervals), "intervals unsorted");
D
duke 已提交
1580 1581 1582 1583 1584
}

void LinearScan::sort_intervals_after_allocation() {
  TIME_LINEAR_SCAN(timer_sort_intervals_after);

1585 1586 1587 1588 1589 1590
  if (_needs_full_resort) {
    // Re-sort existing interval list because an Interval::from() has changed
    _sorted_intervals->sort(interval_cmp);
    _needs_full_resort = false;
  }

D
duke 已提交
1591 1592 1593 1594 1595 1596 1597
  IntervalArray* old_list      = _sorted_intervals;
  IntervalList*  new_list      = _new_intervals_from_allocation;
  int old_len = old_list->length();
  int new_len = new_list->length();

  if (new_len == 0) {
    // no intervals have been added during allocation, so sorted list is already up to date
1598
    assert(is_sorted(_sorted_intervals), "intervals unsorted");
D
duke 已提交
1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620
    return;
  }

  // conventional sort-algorithm for new intervals
  new_list->sort(interval_cmp);

  // merge old and new list (both already sorted) into one combined list
  IntervalArray* combined_list = new IntervalArray(old_len + new_len);
  int old_idx = 0;
  int new_idx = 0;

  while (old_idx + new_idx < old_len + new_len) {
    if (new_idx >= new_len || (old_idx < old_len && old_list->at(old_idx)->from() <= new_list->at(new_idx)->from())) {
      combined_list->at_put(old_idx + new_idx, old_list->at(old_idx));
      old_idx++;
    } else {
      combined_list->at_put(old_idx + new_idx, new_list->at(new_idx));
      new_idx++;
    }
  }

  _sorted_intervals = combined_list;
1621
  assert(is_sorted(_sorted_intervals), "intervals unsorted");
D
duke 已提交
1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705
}


void LinearScan::allocate_registers() {
  TIME_LINEAR_SCAN(timer_allocate_registers);

  Interval* precolored_cpu_intervals, *not_precolored_cpu_intervals;
  Interval* precolored_fpu_intervals, *not_precolored_fpu_intervals;

  create_unhandled_lists(&precolored_cpu_intervals, &not_precolored_cpu_intervals, is_precolored_cpu_interval, is_virtual_cpu_interval);
  if (has_fpu_registers()) {
    create_unhandled_lists(&precolored_fpu_intervals, &not_precolored_fpu_intervals, is_precolored_fpu_interval, is_virtual_fpu_interval);
#ifdef ASSERT
  } else {
    // fpu register allocation is omitted because no virtual fpu registers are present
    // just check this again...
    create_unhandled_lists(&precolored_fpu_intervals, &not_precolored_fpu_intervals, is_precolored_fpu_interval, is_virtual_fpu_interval);
    assert(not_precolored_fpu_intervals == Interval::end(), "missed an uncolored fpu interval");
#endif
  }

  // allocate cpu registers
  LinearScanWalker cpu_lsw(this, precolored_cpu_intervals, not_precolored_cpu_intervals);
  cpu_lsw.walk();
  cpu_lsw.finish_allocation();

  if (has_fpu_registers()) {
    // allocate fpu registers
    LinearScanWalker fpu_lsw(this, precolored_fpu_intervals, not_precolored_fpu_intervals);
    fpu_lsw.walk();
    fpu_lsw.finish_allocation();
  }
}


// ********** Phase 6: resolve data flow
// (insert moves at edges between blocks if intervals have been split)

// wrapper for Interval::split_child_at_op_id that performs a bailout in product mode
// instead of returning NULL
Interval* LinearScan::split_child_at_op_id(Interval* interval, int op_id, LIR_OpVisitState::OprMode mode) {
  Interval* result = interval->split_child_at_op_id(op_id, mode);
  if (result != NULL) {
    return result;
  }

  assert(false, "must find an interval, but do a clean bailout in product mode");
  result = new Interval(LIR_OprDesc::vreg_base);
  result->assign_reg(0);
  result->set_type(T_INT);
  BAILOUT_("LinearScan: interval is NULL", result);
}


Interval* LinearScan::interval_at_block_begin(BlockBegin* block, int reg_num) {
  assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds");
  assert(interval_at(reg_num) != NULL, "no interval found");

  return split_child_at_op_id(interval_at(reg_num), block->first_lir_instruction_id(), LIR_OpVisitState::outputMode);
}

Interval* LinearScan::interval_at_block_end(BlockBegin* block, int reg_num) {
  assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds");
  assert(interval_at(reg_num) != NULL, "no interval found");

  return split_child_at_op_id(interval_at(reg_num), block->last_lir_instruction_id() + 1, LIR_OpVisitState::outputMode);
}

Interval* LinearScan::interval_at_op_id(int reg_num, int op_id) {
  assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds");
  assert(interval_at(reg_num) != NULL, "no interval found");

  return split_child_at_op_id(interval_at(reg_num), op_id, LIR_OpVisitState::inputMode);
}


void LinearScan::resolve_collect_mappings(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) {
  DEBUG_ONLY(move_resolver.check_empty());

  const int num_regs = num_virtual_regs();
  const int size = live_set_size();
  const BitMap live_at_edge = to_block->live_in();

  // visit all registers where the live_at_edge bit is set
1706
  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)) {
D
duke 已提交
1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853
    assert(r < num_regs, "live information set for not exisiting interval");
    assert(from_block->live_out().at(r) && to_block->live_in().at(r), "interval not live at this edge");

    Interval* from_interval = interval_at_block_end(from_block, r);
    Interval* to_interval = interval_at_block_begin(to_block, r);

    if (from_interval != to_interval && (from_interval->assigned_reg() != to_interval->assigned_reg() || from_interval->assigned_regHi() != to_interval->assigned_regHi())) {
      // need to insert move instruction
      move_resolver.add_mapping(from_interval, to_interval);
    }
  }
}


void LinearScan::resolve_find_insert_pos(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) {
  if (from_block->number_of_sux() <= 1) {
    TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at end of from_block B%d", from_block->block_id()));

    LIR_OpList* instructions = from_block->lir()->instructions_list();
    LIR_OpBranch* branch = instructions->last()->as_OpBranch();
    if (branch != NULL) {
      // insert moves before branch
      assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump");
      move_resolver.set_insert_position(from_block->lir(), instructions->length() - 2);
    } else {
      move_resolver.set_insert_position(from_block->lir(), instructions->length() - 1);
    }

  } else {
    TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at beginning of to_block B%d", to_block->block_id()));
#ifdef ASSERT
    assert(from_block->lir()->instructions_list()->at(0)->as_OpLabel() != NULL, "block does not start with a label");

    // because the number of predecessor edges matches the number of
    // successor edges, blocks which are reached by switch statements
    // may have be more than one predecessor but it will be guaranteed
    // that all predecessors will be the same.
    for (int i = 0; i < to_block->number_of_preds(); i++) {
      assert(from_block == to_block->pred_at(i), "all critical edges must be broken");
    }
#endif

    move_resolver.set_insert_position(to_block->lir(), 0);
  }
}


// insert necessary moves (spilling or reloading) at edges between blocks if interval has been split
void LinearScan::resolve_data_flow() {
  TIME_LINEAR_SCAN(timer_resolve_data_flow);

  int num_blocks = block_count();
  MoveResolver move_resolver(this);
  BitMap block_completed(num_blocks);  block_completed.clear();
  BitMap already_resolved(num_blocks); already_resolved.clear();

  int i;
  for (i = 0; i < num_blocks; i++) {
    BlockBegin* block = block_at(i);

    // check if block has only one predecessor and only one successor
    if (block->number_of_preds() == 1 && block->number_of_sux() == 1 && block->number_of_exception_handlers() == 0) {
      LIR_OpList* instructions = block->lir()->instructions_list();
      assert(instructions->at(0)->code() == lir_label, "block must start with label");
      assert(instructions->last()->code() == lir_branch, "block with successors must end with branch");
      assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block with successor must end with unconditional branch");

      // check if block is empty (only label and branch)
      if (instructions->length() == 2) {
        BlockBegin* pred = block->pred_at(0);
        BlockBegin* sux = block->sux_at(0);

        // prevent optimization of two consecutive blocks
        if (!block_completed.at(pred->linear_scan_number()) && !block_completed.at(sux->linear_scan_number())) {
          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()));
          block_completed.set_bit(block->linear_scan_number());

          // directly resolve between pred and sux (without looking at the empty block between)
          resolve_collect_mappings(pred, sux, move_resolver);
          if (move_resolver.has_mappings()) {
            move_resolver.set_insert_position(block->lir(), 0);
            move_resolver.resolve_and_append_moves();
          }
        }
      }
    }
  }


  for (i = 0; i < num_blocks; i++) {
    if (!block_completed.at(i)) {
      BlockBegin* from_block = block_at(i);
      already_resolved.set_from(block_completed);

      int num_sux = from_block->number_of_sux();
      for (int s = 0; s < num_sux; s++) {
        BlockBegin* to_block = from_block->sux_at(s);

        // check for duplicate edges between the same blocks (can happen with switch blocks)
        if (!already_resolved.at(to_block->linear_scan_number())) {
          TRACE_LINEAR_SCAN(3, tty->print_cr("**** processing edge between B%d and B%d", from_block->block_id(), to_block->block_id()));
          already_resolved.set_bit(to_block->linear_scan_number());

          // collect all intervals that have been split between from_block and to_block
          resolve_collect_mappings(from_block, to_block, move_resolver);
          if (move_resolver.has_mappings()) {
            resolve_find_insert_pos(from_block, to_block, move_resolver);
            move_resolver.resolve_and_append_moves();
          }
        }
      }
    }
  }
}


void LinearScan::resolve_exception_entry(BlockBegin* block, int reg_num, MoveResolver &move_resolver) {
  if (interval_at(reg_num) == NULL) {
    // if a phi function is never used, no interval is created -> ignore this
    return;
  }

  Interval* interval = interval_at_block_begin(block, reg_num);
  int reg = interval->assigned_reg();
  int regHi = interval->assigned_regHi();

  if ((reg < nof_regs && interval->always_in_memory()) ||
      (use_fpu_stack_allocation() && reg >= pd_first_fpu_reg && reg <= pd_last_fpu_reg)) {
    // the interval is split to get a short range that is located on the stack
    // in the following two cases:
    // * the interval started in memory (e.g. method parameter), but is currently in a register
    //   this is an optimization for exception handling that reduces the number of moves that
    //   are necessary for resolving the states when an exception uses this exception handler
    // * the interval would be on the fpu stack at the begin of the exception handler
    //   this is not allowed because of the complicated fpu stack handling on Intel

    // range that will be spilled to memory
    int from_op_id = block->first_lir_instruction_id();
    int to_op_id = from_op_id + 1;  // short live range of length 1
    assert(interval->from() <= from_op_id && interval->to() >= to_op_id,
           "no split allowed between exception entry and first instruction");

    if (interval->from() != from_op_id) {
      // the part before from_op_id is unchanged
      interval = interval->split(from_op_id);
      interval->assign_reg(reg, regHi);
      append_interval(interval);
1854 1855
    } else {
      _needs_full_resort = true;
D
duke 已提交
1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877
    }
    assert(interval->from() == from_op_id, "must be true now");

    Interval* spilled_part = interval;
    if (interval->to() != to_op_id) {
      // the part after to_op_id is unchanged
      spilled_part = interval->split_from_start(to_op_id);
      append_interval(spilled_part);
      move_resolver.add_mapping(spilled_part, interval);
    }
    assign_spill_slot(spilled_part);

    assert(spilled_part->from() == from_op_id && spilled_part->to() == to_op_id, "just checking");
  }
}

void LinearScan::resolve_exception_entry(BlockBegin* block, MoveResolver &move_resolver) {
  assert(block->is_set(BlockBegin::exception_entry_flag), "should not call otherwise");
  DEBUG_ONLY(move_resolver.check_empty());

  // visit all registers where the live_in bit is set
  int size = live_set_size();
1878
  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)) {
D
duke 已提交
1879 1880 1881 1882 1883 1884 1885 1886 1887 1888
    resolve_exception_entry(block, r, move_resolver);
  }

  // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately
  for_each_phi_fun(block, phi,
    resolve_exception_entry(block, phi->operand()->vreg_number(), move_resolver)
  );

  if (move_resolver.has_mappings()) {
    // insert moves after first instruction
1889
    move_resolver.set_insert_position(block->lir(), 0);
D
duke 已提交
1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951
    move_resolver.resolve_and_append_moves();
  }
}


void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, int reg_num, Phi* phi, MoveResolver &move_resolver) {
  if (interval_at(reg_num) == NULL) {
    // if a phi function is never used, no interval is created -> ignore this
    return;
  }

  // the computation of to_interval is equal to resolve_collect_mappings,
  // but from_interval is more complicated because of phi functions
  BlockBegin* to_block = handler->entry_block();
  Interval* to_interval = interval_at_block_begin(to_block, reg_num);

  if (phi != NULL) {
    // phi function of the exception entry block
    // no moves are created for this phi function in the LIR_Generator, so the
    // interval at the throwing instruction must be searched using the operands
    // of the phi function
    Value from_value = phi->operand_at(handler->phi_operand());

    // with phi functions it can happen that the same from_value is used in
    // multiple mappings, so notify move-resolver that this is allowed
    move_resolver.set_multiple_reads_allowed();

    Constant* con = from_value->as_Constant();
    if (con != NULL && !con->is_pinned()) {
      // unpinned constants may have no register, so add mapping from constant to interval
      move_resolver.add_mapping(LIR_OprFact::value_type(con->type()), to_interval);
    } else {
      // search split child at the throwing op_id
      Interval* from_interval = interval_at_op_id(from_value->operand()->vreg_number(), throwing_op_id);
      move_resolver.add_mapping(from_interval, to_interval);
    }

  } else {
    // no phi function, so use reg_num also for from_interval
    // search split child at the throwing op_id
    Interval* from_interval = interval_at_op_id(reg_num, throwing_op_id);
    if (from_interval != to_interval) {
      // optimization to reduce number of moves: when to_interval is on stack and
      // the stack slot is known to be always correct, then no move is necessary
      if (!from_interval->always_in_memory() || from_interval->canonical_spill_slot() != to_interval->assigned_reg()) {
        move_resolver.add_mapping(from_interval, to_interval);
      }
    }
  }
}

void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, MoveResolver &move_resolver) {
  TRACE_LINEAR_SCAN(4, tty->print_cr("resolving exception handler B%d: throwing_op_id=%d", handler->entry_block()->block_id(), throwing_op_id));

  DEBUG_ONLY(move_resolver.check_empty());
  assert(handler->lir_op_id() == -1, "already processed this xhandler");
  DEBUG_ONLY(handler->set_lir_op_id(throwing_op_id));
  assert(handler->entry_code() == NULL, "code already present");

  // visit all registers where the live_in bit is set
  BlockBegin* block = handler->entry_block();
  int size = live_set_size();
1952
  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)) {
D
duke 已提交
1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065
    resolve_exception_edge(handler, throwing_op_id, r, NULL, move_resolver);
  }

  // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately
  for_each_phi_fun(block, phi,
    resolve_exception_edge(handler, throwing_op_id, phi->operand()->vreg_number(), phi, move_resolver)
  );

  if (move_resolver.has_mappings()) {
    LIR_List* entry_code = new LIR_List(compilation());
    move_resolver.set_insert_position(entry_code, 0);
    move_resolver.resolve_and_append_moves();

    entry_code->jump(handler->entry_block());
    handler->set_entry_code(entry_code);
  }
}


void LinearScan::resolve_exception_handlers() {
  MoveResolver move_resolver(this);
  LIR_OpVisitState visitor;
  int num_blocks = block_count();

  int i;
  for (i = 0; i < num_blocks; i++) {
    BlockBegin* block = block_at(i);
    if (block->is_set(BlockBegin::exception_entry_flag)) {
      resolve_exception_entry(block, move_resolver);
    }
  }

  for (i = 0; i < num_blocks; i++) {
    BlockBegin* block = block_at(i);
    LIR_List* ops = block->lir();
    int num_ops = ops->length();

    // iterate all instructions of the block. skip the first because it is always a label
    assert(visitor.no_operands(ops->at(0)), "first operation must always be a label");
    for (int j = 1; j < num_ops; j++) {
      LIR_Op* op = ops->at(j);
      int op_id = op->id();

      if (op_id != -1 && has_info(op_id)) {
        // visit operation to collect all operands
        visitor.visit(op);
        assert(visitor.info_count() > 0, "should not visit otherwise");

        XHandlers* xhandlers = visitor.all_xhandler();
        int n = xhandlers->length();
        for (int k = 0; k < n; k++) {
          resolve_exception_edge(xhandlers->handler_at(k), op_id, move_resolver);
        }

#ifdef ASSERT
      } else {
        visitor.visit(op);
        assert(visitor.all_xhandler()->length() == 0, "missed exception handler");
#endif
      }
    }
  }
}


// ********** Phase 7: assign register numbers back to LIR
// (includes computation of debug information and oop maps)

VMReg LinearScan::vm_reg_for_interval(Interval* interval) {
  VMReg reg = interval->cached_vm_reg();
  if (!reg->is_valid() ) {
    reg = vm_reg_for_operand(operand_for_interval(interval));
    interval->set_cached_vm_reg(reg);
  }
  assert(reg == vm_reg_for_operand(operand_for_interval(interval)), "wrong cached value");
  return reg;
}

VMReg LinearScan::vm_reg_for_operand(LIR_Opr opr) {
  assert(opr->is_oop(), "currently only implemented for oop operands");
  return frame_map()->regname(opr);
}


LIR_Opr LinearScan::operand_for_interval(Interval* interval) {
  LIR_Opr opr = interval->cached_opr();
  if (opr->is_illegal()) {
    opr = calc_operand_for_interval(interval);
    interval->set_cached_opr(opr);
  }

  assert(opr == calc_operand_for_interval(interval), "wrong cached value");
  return opr;
}

LIR_Opr LinearScan::calc_operand_for_interval(const Interval* interval) {
  int assigned_reg = interval->assigned_reg();
  BasicType type = interval->type();

  if (assigned_reg >= nof_regs) {
    // stack slot
    assert(interval->assigned_regHi() == any_reg, "must not have hi register");
    return LIR_OprFact::stack(assigned_reg - nof_regs, type);

  } else {
    // register
    switch (type) {
      case T_OBJECT: {
        assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register");
        assert(interval->assigned_regHi() == any_reg, "must not have hi register");
        return LIR_OprFact::single_cpu_oop(assigned_reg);
      }

2066 2067 2068 2069 2070 2071
      case T_ADDRESS: {
        assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register");
        assert(interval->assigned_regHi() == any_reg, "must not have hi register");
        return LIR_OprFact::single_cpu_address(assigned_reg);
      }

2072 2073 2074 2075 2076 2077
      case T_METADATA: {
        assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register");
        assert(interval->assigned_regHi() == any_reg, "must not have hi register");
        return LIR_OprFact::single_cpu_metadata(assigned_reg);
      }

2078 2079 2080
#ifdef __SOFTFP__
      case T_FLOAT:  // fall through
#endif // __SOFTFP__
D
duke 已提交
2081 2082 2083 2084 2085 2086
      case T_INT: {
        assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register");
        assert(interval->assigned_regHi() == any_reg, "must not have hi register");
        return LIR_OprFact::single_cpu(assigned_reg);
      }

2087 2088 2089
#ifdef __SOFTFP__
      case T_DOUBLE:  // fall through
#endif // __SOFTFP__
D
duke 已提交
2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106
      case T_LONG: {
        int assigned_regHi = interval->assigned_regHi();
        assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register");
        assert(num_physical_regs(T_LONG) == 1 ||
               (assigned_regHi >= pd_first_cpu_reg && assigned_regHi <= pd_last_cpu_reg), "no cpu register");

        assert(assigned_reg != assigned_regHi, "invalid allocation");
        assert(num_physical_regs(T_LONG) == 1 || assigned_reg < assigned_regHi,
               "register numbers must be sorted (ensure that e.g. a move from eax,ebx to ebx,eax can not occur)");
        assert((assigned_regHi != any_reg) ^ (num_physical_regs(T_LONG) == 1), "must be match");
        if (requires_adjacent_regs(T_LONG)) {
          assert(assigned_reg % 2 == 0 && assigned_reg + 1 == assigned_regHi, "must be sequential and even");
        }

#ifdef _LP64
        return LIR_OprFact::double_cpu(assigned_reg, assigned_reg);
#else
2107
#if defined(SPARC) || defined(PPC)
D
duke 已提交
2108 2109 2110
        return LIR_OprFact::double_cpu(assigned_regHi, assigned_reg);
#else
        return LIR_OprFact::double_cpu(assigned_reg, assigned_regHi);
2111 2112
#endif // SPARC
#endif // LP64
D
duke 已提交
2113 2114
      }

2115
#ifndef __SOFTFP__
D
duke 已提交
2116
      case T_FLOAT: {
2117
#ifdef X86
D
duke 已提交
2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130
        if (UseSSE >= 1) {
          assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= pd_last_xmm_reg, "no xmm register");
          assert(interval->assigned_regHi() == any_reg, "must not have hi register");
          return LIR_OprFact::single_xmm(assigned_reg - pd_first_xmm_reg);
        }
#endif

        assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register");
        assert(interval->assigned_regHi() == any_reg, "must not have hi register");
        return LIR_OprFact::single_fpu(assigned_reg - pd_first_fpu_reg);
      }

      case T_DOUBLE: {
2131
#ifdef X86
D
duke 已提交
2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143
        if (UseSSE >= 2) {
          assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= pd_last_xmm_reg, "no xmm register");
          assert(interval->assigned_regHi() == any_reg, "must not have hi register (double xmm values are stored in one register)");
          return LIR_OprFact::double_xmm(assigned_reg - pd_first_xmm_reg);
        }
#endif

#ifdef SPARC
        assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register");
        assert(interval->assigned_regHi() >= pd_first_fpu_reg && interval->assigned_regHi() <= pd_last_fpu_reg, "no fpu register");
        assert(assigned_reg % 2 == 0 && assigned_reg + 1 == interval->assigned_regHi(), "must be sequential and even");
        LIR_Opr result = LIR_OprFact::double_fpu(interval->assigned_regHi() - pd_first_fpu_reg, assigned_reg - pd_first_fpu_reg);
2144 2145 2146 2147 2148
#elif defined(ARM)
        assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register");
        assert(interval->assigned_regHi() >= pd_first_fpu_reg && interval->assigned_regHi() <= pd_last_fpu_reg, "no fpu register");
        assert(assigned_reg % 2 == 0 && assigned_reg + 1 == interval->assigned_regHi(), "must be sequential and even");
        LIR_Opr result = LIR_OprFact::double_fpu(assigned_reg - pd_first_fpu_reg, interval->assigned_regHi() - pd_first_fpu_reg);
D
duke 已提交
2149 2150 2151 2152 2153 2154 2155
#else
        assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register");
        assert(interval->assigned_regHi() == any_reg, "must not have hi register (double fpu values are stored in one register on Intel)");
        LIR_Opr result = LIR_OprFact::double_fpu(assigned_reg - pd_first_fpu_reg);
#endif
        return result;
      }
2156
#endif // __SOFTFP__
D
duke 已提交
2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200

      default: {
        ShouldNotReachHere();
        return LIR_OprFact::illegalOpr;
      }
    }
  }
}

LIR_Opr LinearScan::canonical_spill_opr(Interval* interval) {
  assert(interval->canonical_spill_slot() >= nof_regs, "canonical spill slot not set");
  return LIR_OprFact::stack(interval->canonical_spill_slot() - nof_regs, interval->type());
}

LIR_Opr LinearScan::color_lir_opr(LIR_Opr opr, int op_id, LIR_OpVisitState::OprMode mode) {
  assert(opr->is_virtual(), "should not call this otherwise");

  Interval* interval = interval_at(opr->vreg_number());
  assert(interval != NULL, "interval must exist");

  if (op_id != -1) {
#ifdef ASSERT
    BlockBegin* block = block_of_op_with_id(op_id);
    if (block->number_of_sux() <= 1 && op_id == block->last_lir_instruction_id()) {
      // check if spill moves could have been appended at the end of this block, but
      // before the branch instruction. So the split child information for this branch would
      // be incorrect.
      LIR_OpBranch* branch = block->lir()->instructions_list()->last()->as_OpBranch();
      if (branch != NULL) {
        if (block->live_out().at(opr->vreg_number())) {
          assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump");
          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)");
        }
      }
    }
#endif

    // operands are not changed when an interval is split during allocation,
    // so search the right interval here
    interval = split_child_at_op_id(interval, op_id, mode);
  }

  LIR_Opr res = operand_for_interval(interval);

2201
#ifdef X86
D
duke 已提交
2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327
  // new semantic for is_last_use: not only set on definite end of interval,
  // but also before hole
  // This may still miss some cases (e.g. for dead values), but it is not necessary that the
  // last use information is completely correct
  // information is only needed for fpu stack allocation
  if (res->is_fpu_register()) {
    if (opr->is_last_use() || op_id == interval->to() || (op_id != -1 && interval->has_hole_between(op_id, op_id + 1))) {
      assert(op_id == -1 || !is_block_begin(op_id), "holes at begin of block may also result from control flow");
      res = res->make_last_use();
    }
  }
#endif

  assert(!gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::callee_saved) || !FrameMap::is_caller_save_register(res), "bad allocation");

  return res;
}


#ifdef ASSERT
// some methods used to check correctness of debug information

void assert_no_register_values(GrowableArray<ScopeValue*>* values) {
  if (values == NULL) {
    return;
  }

  for (int i = 0; i < values->length(); i++) {
    ScopeValue* value = values->at(i);

    if (value->is_location()) {
      Location location = ((LocationValue*)value)->location();
      assert(location.where() == Location::on_stack, "value is in register");
    }
  }
}

void assert_no_register_values(GrowableArray<MonitorValue*>* values) {
  if (values == NULL) {
    return;
  }

  for (int i = 0; i < values->length(); i++) {
    MonitorValue* value = values->at(i);

    if (value->owner()->is_location()) {
      Location location = ((LocationValue*)value->owner())->location();
      assert(location.where() == Location::on_stack, "owner is in register");
    }
    assert(value->basic_lock().where() == Location::on_stack, "basic_lock is in register");
  }
}

void assert_equal(Location l1, Location l2) {
  assert(l1.where() == l2.where() && l1.type() == l2.type() && l1.offset() == l2.offset(), "");
}

void assert_equal(ScopeValue* v1, ScopeValue* v2) {
  if (v1->is_location()) {
    assert(v2->is_location(), "");
    assert_equal(((LocationValue*)v1)->location(), ((LocationValue*)v2)->location());
  } else if (v1->is_constant_int()) {
    assert(v2->is_constant_int(), "");
    assert(((ConstantIntValue*)v1)->value() == ((ConstantIntValue*)v2)->value(), "");
  } else if (v1->is_constant_double()) {
    assert(v2->is_constant_double(), "");
    assert(((ConstantDoubleValue*)v1)->value() == ((ConstantDoubleValue*)v2)->value(), "");
  } else if (v1->is_constant_long()) {
    assert(v2->is_constant_long(), "");
    assert(((ConstantLongValue*)v1)->value() == ((ConstantLongValue*)v2)->value(), "");
  } else if (v1->is_constant_oop()) {
    assert(v2->is_constant_oop(), "");
    assert(((ConstantOopWriteValue*)v1)->value() == ((ConstantOopWriteValue*)v2)->value(), "");
  } else {
    ShouldNotReachHere();
  }
}

void assert_equal(MonitorValue* m1, MonitorValue* m2) {
  assert_equal(m1->owner(), m2->owner());
  assert_equal(m1->basic_lock(), m2->basic_lock());
}

void assert_equal(IRScopeDebugInfo* d1, IRScopeDebugInfo* d2) {
  assert(d1->scope() == d2->scope(), "not equal");
  assert(d1->bci() == d2->bci(), "not equal");

  if (d1->locals() != NULL) {
    assert(d1->locals() != NULL && d2->locals() != NULL, "not equal");
    assert(d1->locals()->length() == d2->locals()->length(), "not equal");
    for (int i = 0; i < d1->locals()->length(); i++) {
      assert_equal(d1->locals()->at(i), d2->locals()->at(i));
    }
  } else {
    assert(d1->locals() == NULL && d2->locals() == NULL, "not equal");
  }

  if (d1->expressions() != NULL) {
    assert(d1->expressions() != NULL && d2->expressions() != NULL, "not equal");
    assert(d1->expressions()->length() == d2->expressions()->length(), "not equal");
    for (int i = 0; i < d1->expressions()->length(); i++) {
      assert_equal(d1->expressions()->at(i), d2->expressions()->at(i));
    }
  } else {
    assert(d1->expressions() == NULL && d2->expressions() == NULL, "not equal");
  }

  if (d1->monitors() != NULL) {
    assert(d1->monitors() != NULL && d2->monitors() != NULL, "not equal");
    assert(d1->monitors()->length() == d2->monitors()->length(), "not equal");
    for (int i = 0; i < d1->monitors()->length(); i++) {
      assert_equal(d1->monitors()->at(i), d2->monitors()->at(i));
    }
  } else {
    assert(d1->monitors() == NULL && d2->monitors() == NULL, "not equal");
  }

  if (d1->caller() != NULL) {
    assert(d1->caller() != NULL && d2->caller() != NULL, "not equal");
    assert_equal(d1->caller(), d2->caller());
  } else {
    assert(d1->caller() == NULL && d2->caller() == NULL, "not equal");
  }
}

void check_stack_depth(CodeEmitInfo* info, int stack_end) {
R
roland 已提交
2328 2329
  if (info->stack()->bci() != SynchronizationEntryBCI && !info->scope()->method()->is_native()) {
    Bytecodes::Code code = info->scope()->method()->java_code_at_bci(info->stack()->bci());
D
duke 已提交
2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398
    switch (code) {
      case Bytecodes::_ifnull    : // fall through
      case Bytecodes::_ifnonnull : // fall through
      case Bytecodes::_ifeq      : // fall through
      case Bytecodes::_ifne      : // fall through
      case Bytecodes::_iflt      : // fall through
      case Bytecodes::_ifge      : // fall through
      case Bytecodes::_ifgt      : // fall through
      case Bytecodes::_ifle      : // fall through
      case Bytecodes::_if_icmpeq : // fall through
      case Bytecodes::_if_icmpne : // fall through
      case Bytecodes::_if_icmplt : // fall through
      case Bytecodes::_if_icmpge : // fall through
      case Bytecodes::_if_icmpgt : // fall through
      case Bytecodes::_if_icmple : // fall through
      case Bytecodes::_if_acmpeq : // fall through
      case Bytecodes::_if_acmpne :
        assert(stack_end >= -Bytecodes::depth(code), "must have non-empty expression stack at if bytecode");
        break;
    }
  }
}

#endif // ASSERT


IntervalWalker* LinearScan::init_compute_oop_maps() {
  // setup lists of potential oops for walking
  Interval* oop_intervals;
  Interval* non_oop_intervals;

  create_unhandled_lists(&oop_intervals, &non_oop_intervals, is_oop_interval, NULL);

  // intervals that have no oops inside need not to be processed
  // to ensure a walking until the last instruction id, add a dummy interval
  // with a high operation id
  non_oop_intervals = new Interval(any_reg);
  non_oop_intervals->add_range(max_jint - 2, max_jint - 1);

  return new IntervalWalker(this, oop_intervals, non_oop_intervals);
}


OopMap* LinearScan::compute_oop_map(IntervalWalker* iw, LIR_Op* op, CodeEmitInfo* info, bool is_call_site) {
  TRACE_LINEAR_SCAN(3, tty->print_cr("creating oop map at op_id %d", op->id()));

  // walk before the current operation -> intervals that start at
  // the operation (= output operands of the operation) are not
  // included in the oop map
  iw->walk_before(op->id());

  int frame_size = frame_map()->framesize();
  int arg_count = frame_map()->oop_map_arg_count();
  OopMap* map = new OopMap(frame_size, arg_count);

  // Iterate through active intervals
  for (Interval* interval = iw->active_first(fixedKind); interval != Interval::end(); interval = interval->next()) {
    int assigned_reg = interval->assigned_reg();

    assert(interval->current_from() <= op->id() && op->id() <= interval->current_to(), "interval should not be active otherwise");
    assert(interval->assigned_regHi() == any_reg, "oop must be single word");
    assert(interval->reg_num() >= LIR_OprDesc::vreg_base, "fixed interval found");

    // Check if this range covers the instruction. Intervals that
    // start or end at the current operation are not included in the
    // oop map, except in the case of patching moves.  For patching
    // moves, any intervals which end at this instruction are included
    // in the oop map since we may safepoint while doing the patch
    // before we've consumed the inputs.
2399
    if (op->is_patching() || op->id() < interval->current_to()) {
D
duke 已提交
2400 2401 2402 2403 2404

      // caller-save registers must not be included into oop-maps at calls
      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");

      VMReg name = vm_reg_for_interval(interval);
2405
      set_oop(map, name);
D
duke 已提交
2406 2407 2408 2409 2410 2411 2412 2413 2414 2415

      // Spill optimization: when the stack value is guaranteed to be always correct,
      // then it must be added to the oop map even if the interval is currently in a register
      if (interval->always_in_memory() &&
          op->id() > interval->spill_definition_pos() &&
          interval->assigned_reg() != interval->canonical_spill_slot()) {
        assert(interval->spill_definition_pos() > 0, "position not set correctly");
        assert(interval->canonical_spill_slot() >= LinearScan::nof_regs, "no spill slot assigned");
        assert(interval->assigned_reg() < LinearScan::nof_regs, "interval is on stack, so stack slot is registered twice");

2416
        set_oop(map, frame_map()->slot_regname(interval->canonical_spill_slot() - LinearScan::nof_regs));
D
duke 已提交
2417 2418 2419 2420 2421 2422
      }
    }
  }

  // add oops from lock stack
  assert(info->stack() != NULL, "CodeEmitInfo must always have a stack");
R
roland 已提交
2423
  int locks_count = info->stack()->total_locks_size();
D
duke 已提交
2424
  for (int i = 0; i < locks_count; i++) {
2425
    set_oop(map, frame_map()->monitor_object_regname(i));
D
duke 已提交
2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464
  }

  return map;
}


void LinearScan::compute_oop_map(IntervalWalker* iw, const LIR_OpVisitState &visitor, LIR_Op* op) {
  assert(visitor.info_count() > 0, "no oop map needed");

  // compute oop_map only for first CodeEmitInfo
  // because it is (in most cases) equal for all other infos of the same operation
  CodeEmitInfo* first_info = visitor.info_at(0);
  OopMap* first_oop_map = compute_oop_map(iw, op, first_info, visitor.has_call());

  for (int i = 0; i < visitor.info_count(); i++) {
    CodeEmitInfo* info = visitor.info_at(i);
    OopMap* oop_map = first_oop_map;

    if (info->stack()->locks_size() != first_info->stack()->locks_size()) {
      // this info has a different number of locks then the precomputed oop map
      // (possible for lock and unlock instructions) -> compute oop map with
      // correct lock information
      oop_map = compute_oop_map(iw, op, info, visitor.has_call());
    }

    if (info->_oop_map == NULL) {
      info->_oop_map = oop_map;
    } else {
      // a CodeEmitInfo can not be shared between different LIR-instructions
      // because interval splitting can occur anywhere between two instructions
      // and so the oop maps must be different
      // -> check if the already set oop_map is exactly the one calculated for this operation
      assert(info->_oop_map == oop_map, "same CodeEmitInfo used for multiple LIR instructions");
    }
  }
}


// frequently used constants
2465 2466 2467
// Allocate them with new so they are never destroyed (otherwise, a
// forced exit could destroy these objects while they are still in
// use).
Z
zgu 已提交
2468 2469 2470 2471 2472 2473
ConstantOopWriteValue* LinearScan::_oop_null_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantOopWriteValue(NULL);
ConstantIntValue*      LinearScan::_int_m1_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(-1);
ConstantIntValue*      LinearScan::_int_0_scope_value =  new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(0);
ConstantIntValue*      LinearScan::_int_1_scope_value =  new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(1);
ConstantIntValue*      LinearScan::_int_2_scope_value =  new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(2);
LocationValue*         _illegal_value = new (ResourceObj::C_HEAP, mtCompiler) LocationValue(Location());
D
duke 已提交
2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511

void LinearScan::init_compute_debug_info() {
  // cache for frequently used scope values
  // (cpu registers and stack slots)
  _scope_value_cache = ScopeValueArray((LinearScan::nof_cpu_regs + frame_map()->argcount() + max_spills()) * 2, NULL);
}

MonitorValue* LinearScan::location_for_monitor_index(int monitor_index) {
  Location loc;
  if (!frame_map()->location_for_monitor_object(monitor_index, &loc)) {
    bailout("too large frame");
  }
  ScopeValue* object_scope_value = new LocationValue(loc);

  if (!frame_map()->location_for_monitor_lock(monitor_index, &loc)) {
    bailout("too large frame");
  }
  return new MonitorValue(object_scope_value, loc);
}

LocationValue* LinearScan::location_for_name(int name, Location::Type loc_type) {
  Location loc;
  if (!frame_map()->locations_for_slot(name, loc_type, &loc)) {
    bailout("too large frame");
  }
  return new LocationValue(loc);
}


int LinearScan::append_scope_value_for_constant(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) {
  assert(opr->is_constant(), "should not be called otherwise");

  LIR_Const* c = opr->as_constant_ptr();
  BasicType t = c->type();
  switch (t) {
    case T_OBJECT: {
      jobject value = c->as_jobject();
      if (value == NULL) {
2512
        scope_values->append(_oop_null_scope_value);
D
duke 已提交
2513 2514 2515 2516 2517 2518 2519 2520 2521 2522
      } else {
        scope_values->append(new ConstantOopWriteValue(c->as_jobject()));
      }
      return 1;
    }

    case T_INT: // fall through
    case T_FLOAT: {
      int value = c->as_jint_bits();
      switch (value) {
2523 2524 2525 2526
        case -1: scope_values->append(_int_m1_scope_value); break;
        case 0:  scope_values->append(_int_0_scope_value); break;
        case 1:  scope_values->append(_int_1_scope_value); break;
        case 2:  scope_values->append(_int_2_scope_value); break;
D
duke 已提交
2527 2528 2529 2530 2531 2532 2533
        default: scope_values->append(new ConstantIntValue(c->as_jint_bits())); break;
      }
      return 1;
    }

    case T_LONG: // fall through
    case T_DOUBLE: {
R
roland 已提交
2534
#ifdef _LP64
2535
      scope_values->append(_int_0_scope_value);
R
roland 已提交
2536 2537
      scope_values->append(new ConstantLongValue(c->as_jlong_bits()));
#else
D
duke 已提交
2538 2539 2540 2541 2542 2543 2544
      if (hi_word_offset_in_bytes > lo_word_offset_in_bytes) {
        scope_values->append(new ConstantIntValue(c->as_jint_hi_bits()));
        scope_values->append(new ConstantIntValue(c->as_jint_lo_bits()));
      } else {
        scope_values->append(new ConstantIntValue(c->as_jint_lo_bits()));
        scope_values->append(new ConstantIntValue(c->as_jint_hi_bits()));
      }
R
roland 已提交
2545
#endif
D
duke 已提交
2546 2547 2548
      return 2;
    }

2549 2550 2551 2552 2553 2554 2555 2556 2557
    case T_ADDRESS: {
#ifdef _LP64
      scope_values->append(new ConstantLongValue(c->as_jint()));
#else
      scope_values->append(new ConstantIntValue(c->as_jint()));
#endif
      return 1;
    }

D
duke 已提交
2558 2559
    default:
      ShouldNotReachHere();
2560
      return -1;
D
duke 已提交
2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585
  }
}

int LinearScan::append_scope_value_for_operand(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) {
  if (opr->is_single_stack()) {
    int stack_idx = opr->single_stack_ix();
    bool is_oop = opr->is_oop_register();
    int cache_idx = (stack_idx + LinearScan::nof_cpu_regs) * 2 + (is_oop ? 1 : 0);

    ScopeValue* sv = _scope_value_cache.at(cache_idx);
    if (sv == NULL) {
      Location::Type loc_type = is_oop ? Location::oop : Location::normal;
      sv = location_for_name(stack_idx, loc_type);
      _scope_value_cache.at_put(cache_idx, sv);
    }

    // check if cached value is correct
    DEBUG_ONLY(assert_equal(sv, location_for_name(stack_idx, is_oop ? Location::oop : Location::normal)));

    scope_values->append(sv);
    return 1;

  } else if (opr->is_single_cpu()) {
    bool is_oop = opr->is_oop_register();
    int cache_idx = opr->cpu_regnr() * 2 + (is_oop ? 1 : 0);
R
roland 已提交
2586
    Location::Type int_loc_type = NOT_LP64(Location::normal) LP64_ONLY(Location::int_in_long);
D
duke 已提交
2587 2588 2589

    ScopeValue* sv = _scope_value_cache.at(cache_idx);
    if (sv == NULL) {
R
roland 已提交
2590
      Location::Type loc_type = is_oop ? Location::oop : int_loc_type;
D
duke 已提交
2591 2592 2593 2594 2595 2596
      VMReg rname = frame_map()->regname(opr);
      sv = new LocationValue(Location::new_reg_loc(loc_type, rname));
      _scope_value_cache.at_put(cache_idx, sv);
    }

    // check if cached value is correct
R
roland 已提交
2597
    DEBUG_ONLY(assert_equal(sv, new LocationValue(Location::new_reg_loc(is_oop ? Location::oop : int_loc_type, frame_map()->regname(opr)))));
D
duke 已提交
2598 2599 2600 2601

    scope_values->append(sv);
    return 1;

2602
#ifdef X86
D
duke 已提交
2603 2604 2605 2606 2607 2608 2609 2610 2611
  } else if (opr->is_single_xmm()) {
    VMReg rname = opr->as_xmm_float_reg()->as_VMReg();
    LocationValue* sv = new LocationValue(Location::new_reg_loc(Location::normal, rname));

    scope_values->append(sv);
    return 1;
#endif

  } else if (opr->is_single_fpu()) {
2612
#ifdef X86
D
duke 已提交
2613 2614 2615 2616 2617 2618 2619 2620 2621 2622
    // the exact location of fpu stack values is only known
    // during fpu stack allocation, so the stack allocator object
    // must be present
    assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)");
    assert(_fpu_stack_allocator != NULL, "must be present");
    opr = _fpu_stack_allocator->to_fpu_stack(opr);
#endif

    Location::Type loc_type = float_saved_as_double ? Location::float_in_dbl : Location::normal;
    VMReg rname = frame_map()->fpu_regname(opr->fpu_regnr());
2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640
#ifndef __SOFTFP__
#ifndef VM_LITTLE_ENDIAN
    if (! float_saved_as_double) {
      // On big endian system, we may have an issue if float registers use only
      // the low half of the (same) double registers.
      // Both the float and the double could have the same regnr but would correspond
      // to two different addresses once saved.

      // get next safely (no assertion checks)
      VMReg next = VMRegImpl::as_VMReg(1+rname->value());
      if (next->is_reg() &&
          (next->as_FloatRegister() == rname->as_FloatRegister())) {
        // the back-end does use the same numbering for the double and the float
        rname = next; // VMReg for the low bits, e.g. the real VMReg for the float
      }
    }
#endif
#endif
D
duke 已提交
2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652
    LocationValue* sv = new LocationValue(Location::new_reg_loc(loc_type, rname));

    scope_values->append(sv);
    return 1;

  } else {
    // double-size operands

    ScopeValue* first;
    ScopeValue* second;

    if (opr->is_double_stack()) {
2653 2654 2655 2656 2657 2658 2659 2660
#ifdef _LP64
      Location loc1;
      Location::Type loc_type = opr->type() == T_LONG ? Location::lng : Location::dbl;
      if (!frame_map()->locations_for_slot(opr->double_stack_ix(), loc_type, &loc1, NULL)) {
        bailout("too large frame");
      }
      // Does this reverse on x86 vs. sparc?
      first =  new LocationValue(loc1);
2661
      second = _int_0_scope_value;
2662
#else
D
duke 已提交
2663 2664 2665 2666 2667 2668
      Location loc1, loc2;
      if (!frame_map()->locations_for_slot(opr->double_stack_ix(), Location::normal, &loc1, &loc2)) {
        bailout("too large frame");
      }
      first =  new LocationValue(loc1);
      second = new LocationValue(loc2);
2669
#endif // _LP64
D
duke 已提交
2670 2671 2672 2673 2674

    } else if (opr->is_double_cpu()) {
#ifdef _LP64
      VMReg rname_first = opr->as_register_lo()->as_VMReg();
      first = new LocationValue(Location::new_reg_loc(Location::lng, rname_first));
2675
      second = _int_0_scope_value;
D
duke 已提交
2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688
#else
      VMReg rname_first = opr->as_register_lo()->as_VMReg();
      VMReg rname_second = opr->as_register_hi()->as_VMReg();

      if (hi_word_offset_in_bytes < lo_word_offset_in_bytes) {
        // lo/hi and swapped relative to first and second, so swap them
        VMReg tmp = rname_first;
        rname_first = rname_second;
        rname_second = tmp;
      }

      first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first));
      second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second));
2689
#endif //_LP64
D
duke 已提交
2690

2691 2692

#ifdef X86
D
duke 已提交
2693 2694 2695
    } else if (opr->is_double_xmm()) {
      assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation");
      VMReg rname_first  = opr->as_xmm_double_reg()->as_VMReg();
I
iveresov 已提交
2696 2697
#  ifdef _LP64
      first = new LocationValue(Location::new_reg_loc(Location::dbl, rname_first));
2698
      second = _int_0_scope_value;
I
iveresov 已提交
2699
#  else
D
duke 已提交
2700 2701 2702 2703 2704 2705
      first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first));
      // %%% This is probably a waste but we'll keep things as they were for now
      if (true) {
        VMReg rname_second = rname_first->next();
        second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second));
      }
I
iveresov 已提交
2706
#  endif
D
duke 已提交
2707 2708 2709 2710
#endif

    } else if (opr->is_double_fpu()) {
      // On SPARC, fpu_regnrLo/fpu_regnrHi represents the two halves of
2711
      // the double as float registers in the native ordering. On X86,
D
duke 已提交
2712 2713 2714 2715 2716
      // fpu_regnrLo is a FPU stack slot whose VMReg represents
      // the low-order word of the double and fpu_regnrLo + 1 is the
      // name for the other half.  *first and *second must represent the
      // least and most significant words, respectively.

2717
#ifdef X86
D
duke 已提交
2718 2719 2720 2721 2722 2723 2724
      // the exact location of fpu stack values is only known
      // during fpu stack allocation, so the stack allocator object
      // must be present
      assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)");
      assert(_fpu_stack_allocator != NULL, "must be present");
      opr = _fpu_stack_allocator->to_fpu_stack(opr);

2725
      assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation (only fpu_regnrLo is used)");
D
duke 已提交
2726 2727 2728 2729
#endif
#ifdef SPARC
      assert(opr->fpu_regnrLo() == opr->fpu_regnrHi() + 1, "assumed in calculation (only fpu_regnrHi is used)");
#endif
2730 2731 2732 2733 2734 2735
#ifdef ARM
      assert(opr->fpu_regnrHi() == opr->fpu_regnrLo() + 1, "assumed in calculation (only fpu_regnrLo is used)");
#endif
#ifdef PPC
      assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation (only fpu_regnrHi is used)");
#endif
D
duke 已提交
2736

2737 2738 2739
#ifdef VM_LITTLE_ENDIAN
      VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrLo());
#else
D
duke 已提交
2740
      VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrHi());
2741 2742
#endif

I
iveresov 已提交
2743 2744
#ifdef _LP64
      first = new LocationValue(Location::new_reg_loc(Location::dbl, rname_first));
2745
      second = _int_0_scope_value;
I
iveresov 已提交
2746
#else
D
duke 已提交
2747 2748 2749 2750 2751 2752
      first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first));
      // %%% This is probably a waste but we'll keep things as they were for now
      if (true) {
        VMReg rname_second = rname_first->next();
        second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second));
      }
I
iveresov 已提交
2753
#endif
D
duke 已提交
2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825

    } else {
      ShouldNotReachHere();
      first = NULL;
      second = NULL;
    }

    assert(first != NULL && second != NULL, "must be set");
    // The convention the interpreter uses is that the second local
    // holds the first raw word of the native double representation.
    // This is actually reasonable, since locals and stack arrays
    // grow downwards in all implementations.
    // (If, on some machine, the interpreter's Java locals or stack
    // were to grow upwards, the embedded doubles would be word-swapped.)
    scope_values->append(second);
    scope_values->append(first);
    return 2;
  }
}


int LinearScan::append_scope_value(int op_id, Value value, GrowableArray<ScopeValue*>* scope_values) {
  if (value != NULL) {
    LIR_Opr opr = value->operand();
    Constant* con = value->as_Constant();

    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)");
    assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands");

    if (con != NULL && !con->is_pinned() && !opr->is_constant()) {
      // Unpinned constants may have a virtual operand for a part of the lifetime
      // or may be illegal when it was optimized away,
      // so always use a constant operand
      opr = LIR_OprFact::value_type(con->type());
    }
    assert(opr->is_virtual() || opr->is_constant(), "other cases not allowed here");

    if (opr->is_virtual()) {
      LIR_OpVisitState::OprMode mode = LIR_OpVisitState::inputMode;

      BlockBegin* block = block_of_op_with_id(op_id);
      if (block->number_of_sux() == 1 && op_id == block->last_lir_instruction_id()) {
        // generating debug information for the last instruction of a block.
        // if this instruction is a branch, spill moves are inserted before this branch
        // and so the wrong operand would be returned (spill moves at block boundaries are not
        // considered in the live ranges of intervals)
        // Solution: use the first op_id of the branch target block instead.
        if (block->lir()->instructions_list()->last()->as_OpBranch() != NULL) {
          if (block->live_out().at(opr->vreg_number())) {
            op_id = block->sux_at(0)->first_lir_instruction_id();
            mode = LIR_OpVisitState::outputMode;
          }
        }
      }

      // Get current location of operand
      // The operand must be live because debug information is considered when building the intervals
      // if the interval is not live, color_lir_opr will cause an assertion failure
      opr = color_lir_opr(opr, op_id, mode);
      assert(!has_call(op_id) || opr->is_stack() || !is_caller_save(reg_num(opr)), "can not have caller-save register operands at calls");

      // Append to ScopeValue array
      return append_scope_value_for_operand(opr, scope_values);

    } else {
      assert(value->as_Constant() != NULL, "all other instructions have only virtual operands");
      assert(opr->is_constant(), "operand must be constant");

      return append_scope_value_for_constant(opr, scope_values);
    }
  } else {
    // append a dummy value because real value not needed
2826
    scope_values->append(_illegal_value);
D
duke 已提交
2827 2828 2829 2830 2831
    return 1;
  }
}


R
roland 已提交
2832
IRScopeDebugInfo* LinearScan::compute_debug_info_for_scope(int op_id, IRScope* cur_scope, ValueStack* cur_state, ValueStack* innermost_state) {
D
duke 已提交
2833 2834
  IRScopeDebugInfo* caller_debug_info = NULL;

R
roland 已提交
2835
  ValueStack* caller_state = cur_state->caller_state();
D
duke 已提交
2836 2837
  if (caller_state != NULL) {
    // process recursively to compute outermost scope first
R
roland 已提交
2838
    caller_debug_info = compute_debug_info_for_scope(op_id, cur_scope->caller(), caller_state, innermost_state);
D
duke 已提交
2839 2840 2841 2842 2843 2844 2845 2846 2847 2848
  }

  // initialize these to null.
  // If we don't need deopt info or there are no locals, expressions or monitors,
  // then these get recorded as no information and avoids the allocation of 0 length arrays.
  GrowableArray<ScopeValue*>*   locals      = NULL;
  GrowableArray<ScopeValue*>*   expressions = NULL;
  GrowableArray<MonitorValue*>* monitors    = NULL;

  // describe local variable values
R
roland 已提交
2849
  int nof_locals = cur_state->locals_size();
D
duke 已提交
2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863
  if (nof_locals > 0) {
    locals = new GrowableArray<ScopeValue*>(nof_locals);

    int pos = 0;
    while (pos < nof_locals) {
      assert(pos < cur_state->locals_size(), "why not?");

      Value local = cur_state->local_at(pos);
      pos += append_scope_value(op_id, local, locals);

      assert(locals->length() == pos, "must match");
    }
    assert(locals->length() == cur_scope->method()->max_locals(), "wrong number of locals");
    assert(locals->length() == cur_state->locals_size(), "wrong number of locals");
R
roland 已提交
2864 2865 2866 2867 2868
  } else if (cur_scope->method()->max_locals() > 0) {
    assert(cur_state->kind() == ValueStack::EmptyExceptionState, "should be");
    nof_locals = cur_scope->method()->max_locals();
    locals = new GrowableArray<ScopeValue*>(nof_locals);
    for(int i = 0; i < nof_locals; i++) {
2869
      locals->append(_illegal_value);
R
roland 已提交
2870
    }
D
duke 已提交
2871 2872 2873
  }

  // describe expression stack
R
roland 已提交
2874
  int nof_stack = cur_state->stack_size();
D
duke 已提交
2875 2876 2877
  if (nof_stack > 0) {
    expressions = new GrowableArray<ScopeValue*>(nof_stack);

R
roland 已提交
2878 2879 2880
    int pos = 0;
    while (pos < nof_stack) {
      Value expression = cur_state->stack_at_inc(pos);
D
duke 已提交
2881 2882
      append_scope_value(op_id, expression, expressions);

R
roland 已提交
2883
      assert(expressions->length() == pos, "must match");
D
duke 已提交
2884
    }
R
roland 已提交
2885
    assert(expressions->length() == cur_state->stack_size(), "wrong number of stack entries");
D
duke 已提交
2886 2887 2888
  }

  // describe monitors
R
roland 已提交
2889
  int nof_locks = cur_state->locks_size();
D
duke 已提交
2890
  if (nof_locks > 0) {
R
roland 已提交
2891
    int lock_offset = cur_state->caller_state() != NULL ? cur_state->caller_state()->total_locks_size() : 0;
D
duke 已提交
2892
    monitors = new GrowableArray<MonitorValue*>(nof_locks);
R
roland 已提交
2893 2894
    for (int i = 0; i < nof_locks; i++) {
      monitors->append(location_for_monitor_index(lock_offset + i));
D
duke 已提交
2895 2896 2897
    }
  }

R
roland 已提交
2898
  return new IRScopeDebugInfo(cur_scope, cur_state->bci(), locals, expressions, monitors, caller_debug_info);
D
duke 已提交
2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909
}


void LinearScan::compute_debug_info(CodeEmitInfo* info, int op_id) {
  TRACE_LINEAR_SCAN(3, tty->print_cr("creating debug information at op_id %d", op_id));

  IRScope* innermost_scope = info->scope();
  ValueStack* innermost_state = info->stack();

  assert(innermost_scope != NULL && innermost_state != NULL, "why is it missing?");

R
roland 已提交
2910
  DEBUG_ONLY(check_stack_depth(info, innermost_state->stack_size()));
D
duke 已提交
2911 2912 2913

  if (info->_scope_debug_info == NULL) {
    // compute debug information
R
roland 已提交
2914
    info->_scope_debug_info = compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state);
D
duke 已提交
2915 2916
  } else {
    // debug information already set. Check that it is correct from the current point of view
R
roland 已提交
2917
    DEBUG_ONLY(assert_equal(info->_scope_debug_info, compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state)));
D
duke 已提交
2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063
  }
}


void LinearScan::assign_reg_num(LIR_OpList* instructions, IntervalWalker* iw) {
  LIR_OpVisitState visitor;
  int num_inst = instructions->length();
  bool has_dead = false;

  for (int j = 0; j < num_inst; j++) {
    LIR_Op* op = instructions->at(j);
    if (op == NULL) {  // this can happen when spill-moves are removed in eliminate_spill_moves
      has_dead = true;
      continue;
    }
    int op_id = op->id();

    // visit instruction to get list of operands
    visitor.visit(op);

    // iterate all modes of the visitor and process all virtual operands
    for_each_visitor_mode(mode) {
      int n = visitor.opr_count(mode);
      for (int k = 0; k < n; k++) {
        LIR_Opr opr = visitor.opr_at(mode, k);
        if (opr->is_virtual_register()) {
          visitor.set_opr_at(mode, k, color_lir_opr(opr, op_id, mode));
        }
      }
    }

    if (visitor.info_count() > 0) {
      // exception handling
      if (compilation()->has_exception_handlers()) {
        XHandlers* xhandlers = visitor.all_xhandler();
        int n = xhandlers->length();
        for (int k = 0; k < n; k++) {
          XHandler* handler = xhandlers->handler_at(k);
          if (handler->entry_code() != NULL) {
            assign_reg_num(handler->entry_code()->instructions_list(), NULL);
          }
        }
      } else {
        assert(visitor.all_xhandler()->length() == 0, "missed exception handler");
      }

      // compute oop map
      assert(iw != NULL, "needed for compute_oop_map");
      compute_oop_map(iw, visitor, op);

      // compute debug information
      if (!use_fpu_stack_allocation()) {
        // compute debug information if fpu stack allocation is not needed.
        // when fpu stack allocation is needed, the debug information can not
        // be computed here because the exact location of fpu operands is not known
        // -> debug information is created inside the fpu stack allocator
        int n = visitor.info_count();
        for (int k = 0; k < n; k++) {
          compute_debug_info(visitor.info_at(k), op_id);
        }
      }
    }

#ifdef ASSERT
    // make sure we haven't made the op invalid.
    op->verify();
#endif

    // remove useless moves
    if (op->code() == lir_move) {
      assert(op->as_Op1() != NULL, "move must be LIR_Op1");
      LIR_Op1* move = (LIR_Op1*)op;
      LIR_Opr src = move->in_opr();
      LIR_Opr dst = move->result_opr();
      if (dst == src ||
          !dst->is_pointer() && !src->is_pointer() &&
          src->is_same_register(dst)) {
        instructions->at_put(j, NULL);
        has_dead = true;
      }
    }
  }

  if (has_dead) {
    // iterate all instructions of the block and remove all null-values.
    int insert_point = 0;
    for (int j = 0; j < num_inst; j++) {
      LIR_Op* op = instructions->at(j);
      if (op != NULL) {
        if (insert_point != j) {
          instructions->at_put(insert_point, op);
        }
        insert_point++;
      }
    }
    instructions->truncate(insert_point);
  }
}

void LinearScan::assign_reg_num() {
  TIME_LINEAR_SCAN(timer_assign_reg_num);

  init_compute_debug_info();
  IntervalWalker* iw = init_compute_oop_maps();

  int num_blocks = block_count();
  for (int i = 0; i < num_blocks; i++) {
    BlockBegin* block = block_at(i);
    assign_reg_num(block->lir()->instructions_list(), iw);
  }
}


void LinearScan::do_linear_scan() {
  NOT_PRODUCT(_total_timer.begin_method());

  number_instructions();

  NOT_PRODUCT(print_lir(1, "Before Register Allocation"));

  compute_local_live_sets();
  compute_global_live_sets();
  CHECK_BAILOUT();

  build_intervals();
  CHECK_BAILOUT();
  sort_intervals_before_allocation();

  NOT_PRODUCT(print_intervals("Before Register Allocation"));
  NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_before_alloc));

  allocate_registers();
  CHECK_BAILOUT();

  resolve_data_flow();
  if (compilation()->has_exception_handlers()) {
    resolve_exception_handlers();
  }
  // fill in number of spill slots into frame_map
  propagate_spill_slots();
  CHECK_BAILOUT();

  NOT_PRODUCT(print_intervals("After Register Allocation"));
  NOT_PRODUCT(print_lir(2, "LIR after register allocation:"));

  sort_intervals_after_allocation();
3064 3065 3066

  DEBUG_ONLY(verify());

D
duke 已提交
3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251 3252 3253 3254
  eliminate_spill_moves();
  assign_reg_num();
  CHECK_BAILOUT();

  NOT_PRODUCT(print_lir(2, "LIR after assignment of register numbers:"));
  NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_after_asign));

  { TIME_LINEAR_SCAN(timer_allocate_fpu_stack);

    if (use_fpu_stack_allocation()) {
      allocate_fpu_stack(); // Only has effect on Intel
      NOT_PRODUCT(print_lir(2, "LIR after FPU stack allocation:"));
    }
  }

  { TIME_LINEAR_SCAN(timer_optimize_lir);

    EdgeMoveOptimizer::optimize(ir()->code());
    ControlFlowOptimizer::optimize(ir()->code());
    // check that cfg is still correct after optimizations
    ir()->verify();
  }

  NOT_PRODUCT(print_lir(1, "Before Code Generation", false));
  NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_final));
  NOT_PRODUCT(_total_timer.end_method(this));
}


// ********** Printing functions

#ifndef PRODUCT

void LinearScan::print_timers(double total) {
  _total_timer.print(total);
}

void LinearScan::print_statistics() {
  _stat_before_alloc.print("before allocation");
  _stat_after_asign.print("after assignment of register");
  _stat_final.print("after optimization");
}

void LinearScan::print_bitmap(BitMap& b) {
  for (unsigned int i = 0; i < b.size(); i++) {
    if (b.at(i)) tty->print("%d ", i);
  }
  tty->cr();
}

void LinearScan::print_intervals(const char* label) {
  if (TraceLinearScanLevel >= 1) {
    int i;
    tty->cr();
    tty->print_cr("%s", label);

    for (i = 0; i < interval_count(); i++) {
      Interval* interval = interval_at(i);
      if (interval != NULL) {
        interval->print();
      }
    }

    tty->cr();
    tty->print_cr("--- Basic Blocks ---");
    for (i = 0; i < block_count(); i++) {
      BlockBegin* block = block_at(i);
      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());
    }
    tty->cr();
    tty->cr();
  }

  if (PrintCFGToFile) {
    CFGPrinter::print_intervals(&_intervals, label);
  }
}

void LinearScan::print_lir(int level, const char* label, bool hir_valid) {
  if (TraceLinearScanLevel >= level) {
    tty->cr();
    tty->print_cr("%s", label);
    print_LIR(ir()->linear_scan_order());
    tty->cr();
  }

  if (level == 1 && PrintCFGToFile) {
    CFGPrinter::print_cfg(ir()->linear_scan_order(), label, hir_valid, true);
  }
}

#endif //PRODUCT


// ********** verification functions for allocation
// (check that all intervals have a correct register and that no registers are overwritten)
#ifdef ASSERT

void LinearScan::verify() {
  TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying intervals ******************************************"));
  verify_intervals();

  TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that no oops are in fixed intervals ****************"));
  verify_no_oops_in_fixed_intervals();

  TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that unpinned constants are not alive across block boundaries"));
  verify_constants();

  TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying register allocation ********************************"));
  verify_registers();

  TRACE_LINEAR_SCAN(2, tty->print_cr("********* no errors found **********************************************"));
}

void LinearScan::verify_intervals() {
  int len = interval_count();
  bool has_error = false;

  for (int i = 0; i < len; i++) {
    Interval* i1 = interval_at(i);
    if (i1 == NULL) continue;

    i1->check_split_children();

    if (i1->reg_num() != i) {
      tty->print_cr("Interval %d is on position %d in list", i1->reg_num(), i); i1->print(); tty->cr();
      has_error = true;
    }

    if (i1->reg_num() >= LIR_OprDesc::vreg_base && i1->type() == T_ILLEGAL) {
      tty->print_cr("Interval %d has no type assigned", i1->reg_num()); i1->print(); tty->cr();
      has_error = true;
    }

    if (i1->assigned_reg() == any_reg) {
      tty->print_cr("Interval %d has no register assigned", i1->reg_num()); i1->print(); tty->cr();
      has_error = true;
    }

    if (i1->assigned_reg() == i1->assigned_regHi()) {
      tty->print_cr("Interval %d: low and high register equal", i1->reg_num()); i1->print(); tty->cr();
      has_error = true;
    }

    if (!is_processed_reg_num(i1->assigned_reg())) {
      tty->print_cr("Can not have an Interval for an ignored register"); i1->print(); tty->cr();
      has_error = true;
    }

    if (i1->first() == Range::end()) {
      tty->print_cr("Interval %d has no Range", i1->reg_num()); i1->print(); tty->cr();
      has_error = true;
    }

    for (Range* r = i1->first(); r != Range::end(); r = r->next()) {
      if (r->from() >= r->to()) {
        tty->print_cr("Interval %d has zero length range", i1->reg_num()); i1->print(); tty->cr();
        has_error = true;
      }
    }

    for (int j = i + 1; j < len; j++) {
      Interval* i2 = interval_at(j);
      if (i2 == NULL) continue;

      // special intervals that are created in MoveResolver
      // -> ignore them because the range information has no meaning there
      if (i1->from() == 1 && i1->to() == 2) continue;
      if (i2->from() == 1 && i2->to() == 2) continue;

      int r1 = i1->assigned_reg();
      int r1Hi = i1->assigned_regHi();
      int r2 = i2->assigned_reg();
      int r2Hi = i2->assigned_regHi();
      if (i1->intersects(i2) && (r1 == r2 || r1 == r2Hi || (r1Hi != any_reg && (r1Hi == r2 || r1Hi == r2Hi)))) {
        tty->print_cr("Intervals %d and %d overlap and have the same register assigned", i1->reg_num(), i2->reg_num());
        i1->print(); tty->cr();
        i2->print(); tty->cr();
        has_error = true;
      }
    }
  }

  assert(has_error == false, "register allocation invalid");
}


void LinearScan::verify_no_oops_in_fixed_intervals() {
3255 3256 3257 3258 3259 3260 3261 3262 3263 3264
  Interval* fixed_intervals;
  Interval* other_intervals;
  create_unhandled_lists(&fixed_intervals, &other_intervals, is_precolored_cpu_interval, NULL);

  // to ensure a walking until the last instruction id, add a dummy interval
  // with a high operation id
  other_intervals = new Interval(any_reg);
  other_intervals->add_range(max_jint - 2, max_jint - 1);
  IntervalWalker* iw = new IntervalWalker(this, fixed_intervals, other_intervals);

D
duke 已提交
3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276
  LIR_OpVisitState visitor;
  for (int i = 0; i < block_count(); i++) {
    BlockBegin* block = block_at(i);

    LIR_OpList* instructions = block->lir()->instructions_list();

    for (int j = 0; j < instructions->length(); j++) {
      LIR_Op* op = instructions->at(j);
      int op_id = op->id();

      visitor.visit(op);

3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324
      if (visitor.info_count() > 0) {
        iw->walk_before(op->id());
        bool check_live = true;
        if (op->code() == lir_move) {
          LIR_Op1* move = (LIR_Op1*)op;
          check_live = (move->patch_code() == lir_patch_none);
        }
        LIR_OpBranch* branch = op->as_OpBranch();
        if (branch != NULL && branch->stub() != NULL && branch->stub()->is_exception_throw_stub()) {
          // Don't bother checking the stub in this case since the
          // exception stub will never return to normal control flow.
          check_live = false;
        }

        // Make sure none of the fixed registers is live across an
        // oopmap since we can't handle that correctly.
        if (check_live) {
          for (Interval* interval = iw->active_first(fixedKind);
               interval != Interval::end();
               interval = interval->next()) {
            if (interval->current_to() > op->id() + 1) {
              // This interval is live out of this op so make sure
              // that this interval represents some value that's
              // referenced by this op either as an input or output.
              bool ok = false;
              for_each_visitor_mode(mode) {
                int n = visitor.opr_count(mode);
                for (int k = 0; k < n; k++) {
                  LIR_Opr opr = visitor.opr_at(mode, k);
                  if (opr->is_fixed_cpu()) {
                    if (interval_at(reg_num(opr)) == interval) {
                      ok = true;
                      break;
                    }
                    int hi = reg_numHi(opr);
                    if (hi != -1 && interval_at(hi) == interval) {
                      ok = true;
                      break;
                    }
                  }
                }
              }
              assert(ok, "fixed intervals should never be live across an oopmap point");
            }
          }
        }
      }

D
duke 已提交
3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370
      // oop-maps at calls do not contain registers, so check is not needed
      if (!visitor.has_call()) {

        for_each_visitor_mode(mode) {
          int n = visitor.opr_count(mode);
          for (int k = 0; k < n; k++) {
            LIR_Opr opr = visitor.opr_at(mode, k);

            if (opr->is_fixed_cpu() && opr->is_oop()) {
              // operand is a non-virtual cpu register and contains an oop
              TRACE_LINEAR_SCAN(4, op->print_on(tty); tty->print("checking operand "); opr->print(); tty->cr());

              Interval* interval = interval_at(reg_num(opr));
              assert(interval != NULL, "no interval");

              if (mode == LIR_OpVisitState::inputMode) {
                if (interval->to() >= op_id + 1) {
                  assert(interval->to() < op_id + 2 ||
                         interval->has_hole_between(op_id, op_id + 2),
                         "oop input operand live after instruction");
                }
              } else if (mode == LIR_OpVisitState::outputMode) {
                if (interval->from() <= op_id - 1) {
                  assert(interval->has_hole_between(op_id - 1, op_id),
                         "oop input operand live after instruction");
                }
              }
            }
          }
        }
      }
    }
  }
}


void LinearScan::verify_constants() {
  int num_regs = num_virtual_regs();
  int size = live_set_size();
  int num_blocks = block_count();

  for (int i = 0; i < num_blocks; i++) {
    BlockBegin* block = block_at(i);
    BitMap live_at_edge = block->live_in();

    // visit all registers where the live_at_edge bit is set
3371
    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)) {
D
duke 已提交
3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609
      TRACE_LINEAR_SCAN(4, tty->print("checking interval %d of block B%d", r, block->block_id()));

      Value value = gen()->instruction_for_vreg(r);

      assert(value != NULL, "all intervals live across block boundaries must have Value");
      assert(value->operand()->is_register() && value->operand()->is_virtual(), "value must have virtual operand");
      assert(value->operand()->vreg_number() == r, "register number must match");
      // TKR assert(value->as_Constant() == NULL || value->is_pinned(), "only pinned constants can be alive accross block boundaries");
    }
  }
}


class RegisterVerifier: public StackObj {
 private:
  LinearScan*   _allocator;
  BlockList     _work_list;      // all blocks that must be processed
  IntervalsList _saved_states;   // saved information of previous check

  // simplified access to methods of LinearScan
  Compilation*  compilation() const              { return _allocator->compilation(); }
  Interval*     interval_at(int reg_num) const   { return _allocator->interval_at(reg_num); }
  int           reg_num(LIR_Opr opr) const       { return _allocator->reg_num(opr); }

  // currently, only registers are processed
  int           state_size()                     { return LinearScan::nof_regs; }

  // accessors
  IntervalList* state_for_block(BlockBegin* block) { return _saved_states.at(block->block_id()); }
  void          set_state_for_block(BlockBegin* block, IntervalList* saved_state) { _saved_states.at_put(block->block_id(), saved_state); }
  void          add_to_work_list(BlockBegin* block) { if (!_work_list.contains(block)) _work_list.append(block); }

  // helper functions
  IntervalList* copy(IntervalList* input_state);
  void          state_put(IntervalList* input_state, int reg, Interval* interval);
  bool          check_state(IntervalList* input_state, int reg, Interval* interval);

  void process_block(BlockBegin* block);
  void process_xhandler(XHandler* xhandler, IntervalList* input_state);
  void process_successor(BlockBegin* block, IntervalList* input_state);
  void process_operations(LIR_List* ops, IntervalList* input_state);

 public:
  RegisterVerifier(LinearScan* allocator)
    : _allocator(allocator)
    , _work_list(16)
    , _saved_states(BlockBegin::number_of_blocks(), NULL)
  { }

  void verify(BlockBegin* start);
};


// entry function from LinearScan that starts the verification
void LinearScan::verify_registers() {
  RegisterVerifier verifier(this);
  verifier.verify(block_at(0));
}


void RegisterVerifier::verify(BlockBegin* start) {
  // setup input registers (method arguments) for first block
  IntervalList* input_state = new IntervalList(state_size(), NULL);
  CallingConvention* args = compilation()->frame_map()->incoming_arguments();
  for (int n = 0; n < args->length(); n++) {
    LIR_Opr opr = args->at(n);
    if (opr->is_register()) {
      Interval* interval = interval_at(reg_num(opr));

      if (interval->assigned_reg() < state_size()) {
        input_state->at_put(interval->assigned_reg(), interval);
      }
      if (interval->assigned_regHi() != LinearScan::any_reg && interval->assigned_regHi() < state_size()) {
        input_state->at_put(interval->assigned_regHi(), interval);
      }
    }
  }

  set_state_for_block(start, input_state);
  add_to_work_list(start);

  // main loop for verification
  do {
    BlockBegin* block = _work_list.at(0);
    _work_list.remove_at(0);

    process_block(block);
  } while (!_work_list.is_empty());
}

void RegisterVerifier::process_block(BlockBegin* block) {
  TRACE_LINEAR_SCAN(2, tty->cr(); tty->print_cr("process_block B%d", block->block_id()));

  // must copy state because it is modified
  IntervalList* input_state = copy(state_for_block(block));

  if (TraceLinearScanLevel >= 4) {
    tty->print_cr("Input-State of intervals:");
    tty->print("    ");
    for (int i = 0; i < state_size(); i++) {
      if (input_state->at(i) != NULL) {
        tty->print(" %4d", input_state->at(i)->reg_num());
      } else {
        tty->print("   __");
      }
    }
    tty->cr();
    tty->cr();
  }

  // process all operations of the block
  process_operations(block->lir(), input_state);

  // iterate all successors
  for (int i = 0; i < block->number_of_sux(); i++) {
    process_successor(block->sux_at(i), input_state);
  }
}

void RegisterVerifier::process_xhandler(XHandler* xhandler, IntervalList* input_state) {
  TRACE_LINEAR_SCAN(2, tty->print_cr("process_xhandler B%d", xhandler->entry_block()->block_id()));

  // must copy state because it is modified
  input_state = copy(input_state);

  if (xhandler->entry_code() != NULL) {
    process_operations(xhandler->entry_code(), input_state);
  }
  process_successor(xhandler->entry_block(), input_state);
}

void RegisterVerifier::process_successor(BlockBegin* block, IntervalList* input_state) {
  IntervalList* saved_state = state_for_block(block);

  if (saved_state != NULL) {
    // this block was already processed before.
    // check if new input_state is consistent with saved_state

    bool saved_state_correct = true;
    for (int i = 0; i < state_size(); i++) {
      if (input_state->at(i) != saved_state->at(i)) {
        // current input_state and previous saved_state assume a different
        // interval in this register -> assume that this register is invalid
        if (saved_state->at(i) != NULL) {
          // invalidate old calculation only if it assumed that
          // register was valid. when the register was already invalid,
          // then the old calculation was correct.
          saved_state_correct = false;
          saved_state->at_put(i, NULL);

          TRACE_LINEAR_SCAN(4, tty->print_cr("process_successor B%d: invalidating slot %d", block->block_id(), i));
        }
      }
    }

    if (saved_state_correct) {
      // already processed block with correct input_state
      TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: previous visit already correct", block->block_id()));
    } else {
      // must re-visit this block
      TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: must re-visit because input state changed", block->block_id()));
      add_to_work_list(block);
    }

  } else {
    // block was not processed before, so set initial input_state
    TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: initial visit", block->block_id()));

    set_state_for_block(block, copy(input_state));
    add_to_work_list(block);
  }
}


IntervalList* RegisterVerifier::copy(IntervalList* input_state) {
  IntervalList* copy_state = new IntervalList(input_state->length());
  copy_state->push_all(input_state);
  return copy_state;
}

void RegisterVerifier::state_put(IntervalList* input_state, int reg, Interval* interval) {
  if (reg != LinearScan::any_reg && reg < state_size()) {
    if (interval != NULL) {
      TRACE_LINEAR_SCAN(4, tty->print_cr("        reg[%d] = %d", reg, interval->reg_num()));
    } else if (input_state->at(reg) != NULL) {
      TRACE_LINEAR_SCAN(4, tty->print_cr("        reg[%d] = NULL", reg));
    }

    input_state->at_put(reg, interval);
  }
}

bool RegisterVerifier::check_state(IntervalList* input_state, int reg, Interval* interval) {
  if (reg != LinearScan::any_reg && reg < state_size()) {
    if (input_state->at(reg) != interval) {
      tty->print_cr("!! Error in register allocation: register %d does not contain interval %d", reg, interval->reg_num());
      return true;
    }
  }
  return false;
}

void RegisterVerifier::process_operations(LIR_List* ops, IntervalList* input_state) {
  // visit all instructions of the block
  LIR_OpVisitState visitor;
  bool has_error = false;

  for (int i = 0; i < ops->length(); i++) {
    LIR_Op* op = ops->at(i);
    visitor.visit(op);

    TRACE_LINEAR_SCAN(4, op->print_on(tty));

    // check if input operands are correct
    int j;
    int n = visitor.opr_count(LIR_OpVisitState::inputMode);
    for (j = 0; j < n; j++) {
      LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, j);
      if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) {
        Interval* interval = interval_at(reg_num(opr));
        if (op->id() != -1) {
          interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::inputMode);
        }

        has_error |= check_state(input_state, interval->assigned_reg(),   interval->split_parent());
        has_error |= check_state(input_state, interval->assigned_regHi(), interval->split_parent());

        // When an operand is marked with is_last_use, then the fpu stack allocator
        // removes the register from the fpu stack -> the register contains no value
        if (opr->is_last_use()) {
          state_put(input_state, interval->assigned_reg(),   NULL);
          state_put(input_state, interval->assigned_regHi(), NULL);
        }
      }
    }

    // invalidate all caller save registers at calls
    if (visitor.has_call()) {
3610
      for (j = 0; j < FrameMap::nof_caller_save_cpu_regs(); j++) {
D
duke 已提交
3611 3612 3613 3614 3615 3616
        state_put(input_state, reg_num(FrameMap::caller_save_cpu_reg_at(j)), NULL);
      }
      for (j = 0; j < FrameMap::nof_caller_save_fpu_regs; j++) {
        state_put(input_state, reg_num(FrameMap::caller_save_fpu_reg_at(j)), NULL);
      }

3617
#ifdef X86
D
duke 已提交
3618 3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636 3637 3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648 3649 3650 3651 3652 3653 3654 3655 3656 3657 3658 3659 3660 3661 3662 3663 3664 3665 3666 3667 3668 3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713 3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748 3749 3750 3751 3752 3753 3754 3755 3756 3757 3758 3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826 3827 3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848 3849 3850 3851 3852 3853 3854 3855 3856 3857 3858 3859 3860 3861 3862 3863 3864 3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892 3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951 3952 3953 3954 3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982 3983 3984 3985 3986 3987 3988 3989 3990 3991 3992 3993 3994 3995 3996 3997 3998 3999 4000 4001 4002 4003 4004 4005 4006 4007 4008 4009 4010 4011 4012 4013 4014
      for (j = 0; j < FrameMap::nof_caller_save_xmm_regs; j++) {
        state_put(input_state, reg_num(FrameMap::caller_save_xmm_reg_at(j)), NULL);
      }
#endif
    }

    // process xhandler before output and temp operands
    XHandlers* xhandlers = visitor.all_xhandler();
    n = xhandlers->length();
    for (int k = 0; k < n; k++) {
      process_xhandler(xhandlers->handler_at(k), input_state);
    }

    // set temp operands (some operations use temp operands also as output operands, so can't set them NULL)
    n = visitor.opr_count(LIR_OpVisitState::tempMode);
    for (j = 0; j < n; j++) {
      LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, j);
      if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) {
        Interval* interval = interval_at(reg_num(opr));
        if (op->id() != -1) {
          interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::tempMode);
        }

        state_put(input_state, interval->assigned_reg(),   interval->split_parent());
        state_put(input_state, interval->assigned_regHi(), interval->split_parent());
      }
    }

    // set output operands
    n = visitor.opr_count(LIR_OpVisitState::outputMode);
    for (j = 0; j < n; j++) {
      LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, j);
      if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) {
        Interval* interval = interval_at(reg_num(opr));
        if (op->id() != -1) {
          interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::outputMode);
        }

        state_put(input_state, interval->assigned_reg(),   interval->split_parent());
        state_put(input_state, interval->assigned_regHi(), interval->split_parent());
      }
    }
  }
  assert(has_error == false, "Error in register allocation");
}

#endif // ASSERT



// **** Implementation of MoveResolver ******************************

MoveResolver::MoveResolver(LinearScan* allocator) :
  _allocator(allocator),
  _multiple_reads_allowed(false),
  _mapping_from(8),
  _mapping_from_opr(8),
  _mapping_to(8),
  _insert_list(NULL),
  _insert_idx(-1),
  _insertion_buffer()
{
  for (int i = 0; i < LinearScan::nof_regs; i++) {
    _register_blocked[i] = 0;
  }
  DEBUG_ONLY(check_empty());
}


#ifdef ASSERT

void MoveResolver::check_empty() {
  assert(_mapping_from.length() == 0 && _mapping_from_opr.length() == 0 && _mapping_to.length() == 0, "list must be empty before and after processing");
  for (int i = 0; i < LinearScan::nof_regs; i++) {
    assert(register_blocked(i) == 0, "register map must be empty before and after processing");
  }
  assert(_multiple_reads_allowed == false, "must have default value");
}

void MoveResolver::verify_before_resolve() {
  assert(_mapping_from.length() == _mapping_from_opr.length(), "length must be equal");
  assert(_mapping_from.length() == _mapping_to.length(), "length must be equal");
  assert(_insert_list != NULL && _insert_idx != -1, "insert position not set");

  int i, j;
  if (!_multiple_reads_allowed) {
    for (i = 0; i < _mapping_from.length(); i++) {
      for (j = i + 1; j < _mapping_from.length(); j++) {
        assert(_mapping_from.at(i) == NULL || _mapping_from.at(i) != _mapping_from.at(j), "cannot read from same interval twice");
      }
    }
  }

  for (i = 0; i < _mapping_to.length(); i++) {
    for (j = i + 1; j < _mapping_to.length(); j++) {
      assert(_mapping_to.at(i) != _mapping_to.at(j), "cannot write to same interval twice");
    }
  }


  BitMap used_regs(LinearScan::nof_regs + allocator()->frame_map()->argcount() + allocator()->max_spills());
  used_regs.clear();
  if (!_multiple_reads_allowed) {
    for (i = 0; i < _mapping_from.length(); i++) {
      Interval* it = _mapping_from.at(i);
      if (it != NULL) {
        assert(!used_regs.at(it->assigned_reg()), "cannot read from same register twice");
        used_regs.set_bit(it->assigned_reg());

        if (it->assigned_regHi() != LinearScan::any_reg) {
          assert(!used_regs.at(it->assigned_regHi()), "cannot read from same register twice");
          used_regs.set_bit(it->assigned_regHi());
        }
      }
    }
  }

  used_regs.clear();
  for (i = 0; i < _mapping_to.length(); i++) {
    Interval* it = _mapping_to.at(i);
    assert(!used_regs.at(it->assigned_reg()), "cannot write to same register twice");
    used_regs.set_bit(it->assigned_reg());

    if (it->assigned_regHi() != LinearScan::any_reg) {
      assert(!used_regs.at(it->assigned_regHi()), "cannot write to same register twice");
      used_regs.set_bit(it->assigned_regHi());
    }
  }

  used_regs.clear();
  for (i = 0; i < _mapping_from.length(); i++) {
    Interval* it = _mapping_from.at(i);
    if (it != NULL && it->assigned_reg() >= LinearScan::nof_regs) {
      used_regs.set_bit(it->assigned_reg());
    }
  }
  for (i = 0; i < _mapping_to.length(); i++) {
    Interval* it = _mapping_to.at(i);
    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");
  }
}

#endif // ASSERT


// mark assigned_reg and assigned_regHi of the interval as blocked
void MoveResolver::block_registers(Interval* it) {
  int reg = it->assigned_reg();
  if (reg < LinearScan::nof_regs) {
    assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used");
    set_register_blocked(reg, 1);
  }
  reg = it->assigned_regHi();
  if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) {
    assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used");
    set_register_blocked(reg, 1);
  }
}

// mark assigned_reg and assigned_regHi of the interval as unblocked
void MoveResolver::unblock_registers(Interval* it) {
  int reg = it->assigned_reg();
  if (reg < LinearScan::nof_regs) {
    assert(register_blocked(reg) > 0, "register already marked as unused");
    set_register_blocked(reg, -1);
  }
  reg = it->assigned_regHi();
  if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) {
    assert(register_blocked(reg) > 0, "register already marked as unused");
    set_register_blocked(reg, -1);
  }
}

// check if assigned_reg and assigned_regHi of the to-interval are not blocked (or only blocked by from)
bool MoveResolver::save_to_process_move(Interval* from, Interval* to) {
  int from_reg = -1;
  int from_regHi = -1;
  if (from != NULL) {
    from_reg = from->assigned_reg();
    from_regHi = from->assigned_regHi();
  }

  int reg = to->assigned_reg();
  if (reg < LinearScan::nof_regs) {
    if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) {
      return false;
    }
  }
  reg = to->assigned_regHi();
  if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) {
    if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) {
      return false;
    }
  }

  return true;
}


void MoveResolver::create_insertion_buffer(LIR_List* list) {
  assert(!_insertion_buffer.initialized(), "overwriting existing buffer");
  _insertion_buffer.init(list);
}

void MoveResolver::append_insertion_buffer() {
  if (_insertion_buffer.initialized()) {
    _insertion_buffer.lir_list()->append(&_insertion_buffer);
  }
  assert(!_insertion_buffer.initialized(), "must be uninitialized now");

  _insert_list = NULL;
  _insert_idx = -1;
}

void MoveResolver::insert_move(Interval* from_interval, Interval* to_interval) {
  assert(from_interval->reg_num() != to_interval->reg_num(), "from and to interval equal");
  assert(from_interval->type() == to_interval->type(), "move between different types");
  assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first");
  assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer");

  LIR_Opr from_opr = LIR_OprFact::virtual_register(from_interval->reg_num(), from_interval->type());
  LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type());

  if (!_multiple_reads_allowed) {
    // the last_use flag is an optimization for FPU stack allocation. When the same
    // input interval is used in more than one move, then it is too difficult to determine
    // if this move is really the last use.
    from_opr = from_opr->make_last_use();
  }
  _insertion_buffer.move(_insert_idx, from_opr, to_opr);

  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()));
}

void MoveResolver::insert_move(LIR_Opr from_opr, Interval* to_interval) {
  assert(from_opr->type() == to_interval->type(), "move between different types");
  assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first");
  assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer");

  LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type());
  _insertion_buffer.move(_insert_idx, from_opr, to_opr);

  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()));
}


void MoveResolver::resolve_mappings() {
  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));
  DEBUG_ONLY(verify_before_resolve());

  // Block all registers that are used as input operands of a move.
  // When a register is blocked, no move to this register is emitted.
  // This is necessary for detecting cycles in moves.
  int i;
  for (i = _mapping_from.length() - 1; i >= 0; i--) {
    Interval* from_interval = _mapping_from.at(i);
    if (from_interval != NULL) {
      block_registers(from_interval);
    }
  }

  int spill_candidate = -1;
  while (_mapping_from.length() > 0) {
    bool processed_interval = false;

    for (i = _mapping_from.length() - 1; i >= 0; i--) {
      Interval* from_interval = _mapping_from.at(i);
      Interval* to_interval = _mapping_to.at(i);

      if (save_to_process_move(from_interval, to_interval)) {
        // this inverval can be processed because target is free
        if (from_interval != NULL) {
          insert_move(from_interval, to_interval);
          unblock_registers(from_interval);
        } else {
          insert_move(_mapping_from_opr.at(i), to_interval);
        }
        _mapping_from.remove_at(i);
        _mapping_from_opr.remove_at(i);
        _mapping_to.remove_at(i);

        processed_interval = true;
      } else if (from_interval != NULL && from_interval->assigned_reg() < LinearScan::nof_regs) {
        // this interval cannot be processed now because target is not free
        // it starts in a register, so it is a possible candidate for spilling
        spill_candidate = i;
      }
    }

    if (!processed_interval) {
      // no move could be processed because there is a cycle in the move list
      // (e.g. r1 -> r2, r2 -> r1), so one interval must be spilled to memory
      assert(spill_candidate != -1, "no interval in register for spilling found");

      // create a new spill interval and assign a stack slot to it
      Interval* from_interval = _mapping_from.at(spill_candidate);
      Interval* spill_interval = new Interval(-1);
      spill_interval->set_type(from_interval->type());

      // add a dummy range because real position is difficult to calculate
      // Note: this range is a special case when the integrity of the allocation is checked
      spill_interval->add_range(1, 2);

      //       do not allocate a new spill slot for temporary interval, but
      //       use spill slot assigned to from_interval. Otherwise moves from
      //       one stack slot to another can happen (not allowed by LIR_Assembler
      int spill_slot = from_interval->canonical_spill_slot();
      if (spill_slot < 0) {
        spill_slot = allocator()->allocate_spill_slot(type2spill_size[spill_interval->type()] == 2);
        from_interval->set_canonical_spill_slot(spill_slot);
      }
      spill_interval->assign_reg(spill_slot);
      allocator()->append_interval(spill_interval);

      TRACE_LINEAR_SCAN(4, tty->print_cr("created new Interval %d for spilling", spill_interval->reg_num()));

      // insert a move from register to stack and update the mapping
      insert_move(from_interval, spill_interval);
      _mapping_from.at_put(spill_candidate, spill_interval);
      unblock_registers(from_interval);
    }
  }

  // reset to default value
  _multiple_reads_allowed = false;

  // check that all intervals have been processed
  DEBUG_ONLY(check_empty());
}


void MoveResolver::set_insert_position(LIR_List* insert_list, int insert_idx) {
  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));
  assert(_insert_list == NULL && _insert_idx == -1, "use move_insert_position instead of set_insert_position when data already set");

  create_insertion_buffer(insert_list);
  _insert_list = insert_list;
  _insert_idx = insert_idx;
}

void MoveResolver::move_insert_position(LIR_List* insert_list, int insert_idx) {
  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));

  if (_insert_list != NULL && (insert_list != _insert_list || insert_idx != _insert_idx)) {
    // insert position changed -> resolve current mappings
    resolve_mappings();
  }

  if (insert_list != _insert_list) {
    // block changed -> append insertion_buffer because it is
    // bound to a specific block and create a new insertion_buffer
    append_insertion_buffer();
    create_insertion_buffer(insert_list);
  }

  _insert_list = insert_list;
  _insert_idx = insert_idx;
}

void MoveResolver::add_mapping(Interval* from_interval, Interval* to_interval) {
  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()));

  _mapping_from.append(from_interval);
  _mapping_from_opr.append(LIR_OprFact::illegalOpr);
  _mapping_to.append(to_interval);
}


void MoveResolver::add_mapping(LIR_Opr from_opr, Interval* to_interval) {
  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()));
  assert(from_opr->is_constant(), "only for constants");

  _mapping_from.append(NULL);
  _mapping_from_opr.append(from_opr);
  _mapping_to.append(to_interval);
}

void MoveResolver::resolve_and_append_moves() {
  if (has_mappings()) {
    resolve_mappings();
  }
  append_insertion_buffer();
}



// **** Implementation of Range *************************************

Range::Range(int from, int to, Range* next) :
  _from(from),
  _to(to),
  _next(next)
{
}

// initialize sentinel
Range* Range::_end = NULL;
4015 4016
void Range::initialize(Arena* arena) {
  _end = new (arena) Range(max_jint, max_jint, NULL);
D
duke 已提交
4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031 4032 4033 4034 4035 4036 4037 4038 4039 4040 4041 4042 4043 4044 4045 4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057 4058 4059 4060 4061
}

int Range::intersects_at(Range* r2) const {
  const Range* r1 = this;

  assert(r1 != NULL && r2 != NULL, "null ranges not allowed");
  assert(r1 != _end && r2 != _end, "empty ranges not allowed");

  do {
    if (r1->from() < r2->from()) {
      if (r1->to() <= r2->from()) {
        r1 = r1->next(); if (r1 == _end) return -1;
      } else {
        return r2->from();
      }
    } else if (r2->from() < r1->from()) {
      if (r2->to() <= r1->from()) {
        r2 = r2->next(); if (r2 == _end) return -1;
      } else {
        return r1->from();
      }
    } else { // r1->from() == r2->from()
      if (r1->from() == r1->to()) {
        r1 = r1->next(); if (r1 == _end) return -1;
      } else if (r2->from() == r2->to()) {
        r2 = r2->next(); if (r2 == _end) return -1;
      } else {
        return r1->from();
      }
    }
  } while (true);
}

#ifndef PRODUCT
void Range::print(outputStream* out) const {
  out->print("[%d, %d[ ", _from, _to);
}
#endif



// **** Implementation of Interval **********************************

// initialize sentinel
Interval* Interval::_end = NULL;
4062 4063 4064
void Interval::initialize(Arena* arena) {
  Range::initialize(arena);
  _end = new (arena) Interval(-1);
D
duke 已提交
4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077 4078 4079 4080 4081 4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095 4096 4097 4098 4099 4100 4101 4102 4103 4104 4105 4106 4107 4108 4109 4110 4111 4112 4113 4114 4115 4116 4117 4118 4119 4120 4121 4122 4123 4124 4125 4126 4127 4128 4129 4130 4131 4132 4133 4134 4135 4136 4137 4138 4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167 4168 4169 4170 4171 4172 4173 4174 4175 4176 4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190 4191 4192 4193 4194 4195 4196 4197 4198 4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238 4239 4240 4241 4242 4243 4244 4245 4246 4247 4248 4249 4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261 4262 4263 4264 4265 4266 4267 4268 4269 4270 4271 4272 4273 4274 4275 4276 4277 4278 4279 4280 4281 4282 4283 4284 4285 4286 4287 4288 4289 4290 4291 4292 4293 4294 4295 4296 4297 4298 4299 4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312 4313 4314 4315 4316 4317 4318 4319 4320 4321 4322 4323 4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429 4430 4431 4432 4433 4434 4435 4436 4437 4438 4439 4440 4441 4442 4443 4444 4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458 4459 4460 4461 4462 4463 4464 4465 4466 4467 4468 4469 4470 4471 4472 4473 4474 4475 4476 4477 4478 4479 4480 4481 4482 4483 4484 4485 4486 4487 4488 4489 4490 4491 4492 4493 4494 4495 4496 4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525 4526 4527 4528 4529 4530 4531 4532 4533 4534 4535
}

Interval::Interval(int reg_num) :
  _reg_num(reg_num),
  _type(T_ILLEGAL),
  _first(Range::end()),
  _use_pos_and_kinds(12),
  _current(Range::end()),
  _next(_end),
  _state(invalidState),
  _assigned_reg(LinearScan::any_reg),
  _assigned_regHi(LinearScan::any_reg),
  _cached_to(-1),
  _cached_opr(LIR_OprFact::illegalOpr),
  _cached_vm_reg(VMRegImpl::Bad()),
  _split_children(0),
  _canonical_spill_slot(-1),
  _insert_move_when_activated(false),
  _register_hint(NULL),
  _spill_state(noDefinitionFound),
  _spill_definition_pos(-1)
{
  _split_parent = this;
  _current_split_child = this;
}

int Interval::calc_to() {
  assert(_first != Range::end(), "interval has no range");

  Range* r = _first;
  while (r->next() != Range::end()) {
    r = r->next();
  }
  return r->to();
}


#ifdef ASSERT
// consistency check of split-children
void Interval::check_split_children() {
  if (_split_children.length() > 0) {
    assert(is_split_parent(), "only split parents can have children");

    for (int i = 0; i < _split_children.length(); i++) {
      Interval* i1 = _split_children.at(i);

      assert(i1->split_parent() == this, "not a split child of this interval");
      assert(i1->type() == type(), "must be equal for all split children");
      assert(i1->canonical_spill_slot() == canonical_spill_slot(), "must be equal for all split children");

      for (int j = i + 1; j < _split_children.length(); j++) {
        Interval* i2 = _split_children.at(j);

        assert(i1->reg_num() != i2->reg_num(), "same register number");

        if (i1->from() < i2->from()) {
          assert(i1->to() <= i2->from() && i1->to() < i2->to(), "intervals overlapping");
        } else {
          assert(i2->from() < i1->from(), "intervals start at same op_id");
          assert(i2->to() <= i1->from() && i2->to() < i1->to(), "intervals overlapping");
        }
      }
    }
  }
}
#endif // ASSERT

Interval* Interval::register_hint(bool search_split_child) const {
  if (!search_split_child) {
    return _register_hint;
  }

  if (_register_hint != NULL) {
    assert(_register_hint->is_split_parent(), "ony split parents are valid hint registers");

    if (_register_hint->assigned_reg() >= 0 && _register_hint->assigned_reg() < LinearScan::nof_regs) {
      return _register_hint;

    } else if (_register_hint->_split_children.length() > 0) {
      // search the first split child that has a register assigned
      int len = _register_hint->_split_children.length();
      for (int i = 0; i < len; i++) {
        Interval* cur = _register_hint->_split_children.at(i);

        if (cur->assigned_reg() >= 0 && cur->assigned_reg() < LinearScan::nof_regs) {
          return cur;
        }
      }
    }
  }

  // no hint interval found that has a register assigned
  return NULL;
}


Interval* Interval::split_child_at_op_id(int op_id, LIR_OpVisitState::OprMode mode) {
  assert(is_split_parent(), "can only be called for split parents");
  assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)");

  Interval* result;
  if (_split_children.length() == 0) {
    result = this;
  } else {
    result = NULL;
    int len = _split_children.length();

    // in outputMode, the end of the interval (op_id == cur->to()) is not valid
    int to_offset = (mode == LIR_OpVisitState::outputMode ? 0 : 1);

    int i;
    for (i = 0; i < len; i++) {
      Interval* cur = _split_children.at(i);
      if (cur->from() <= op_id && op_id < cur->to() + to_offset) {
        if (i > 0) {
          // exchange current split child to start of list (faster access for next call)
          _split_children.at_put(i, _split_children.at(0));
          _split_children.at_put(0, cur);
        }

        // interval found
        result = cur;
        break;
      }
    }

#ifdef ASSERT
    for (i = 0; i < len; i++) {
      Interval* tmp = _split_children.at(i);
      if (tmp != result && tmp->from() <= op_id && op_id < tmp->to() + to_offset) {
        tty->print_cr("two valid result intervals found for op_id %d: %d and %d", op_id, result->reg_num(), tmp->reg_num());
        result->print();
        tmp->print();
        assert(false, "two valid result intervals found");
      }
    }
#endif
  }

  assert(result != NULL, "no matching interval found");
  assert(result->covers(op_id, mode), "op_id not covered by interval");

  return result;
}


// returns the last split child that ends before the given op_id
Interval* Interval::split_child_before_op_id(int op_id) {
  assert(op_id >= 0, "invalid op_id");

  Interval* parent = split_parent();
  Interval* result = NULL;

  int len = parent->_split_children.length();
  assert(len > 0, "no split children available");

  for (int i = len - 1; i >= 0; i--) {
    Interval* cur = parent->_split_children.at(i);
    if (cur->to() <= op_id && (result == NULL || result->to() < cur->to())) {
      result = cur;
    }
  }

  assert(result != NULL, "no split child found");
  return result;
}


// checks if op_id is covered by any split child
bool Interval::split_child_covers(int op_id, LIR_OpVisitState::OprMode mode) {
  assert(is_split_parent(), "can only be called for split parents");
  assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)");

  if (_split_children.length() == 0) {
    // simple case if interval was not split
    return covers(op_id, mode);

  } else {
    // extended case: check all split children
    int len = _split_children.length();
    for (int i = 0; i < len; i++) {
      Interval* cur = _split_children.at(i);
      if (cur->covers(op_id, mode)) {
        return true;
      }
    }
    return false;
  }
}


// Note: use positions are sorted descending -> first use has highest index
int Interval::first_usage(IntervalUseKind min_use_kind) const {
  assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals");

  for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) {
    if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) {
      return _use_pos_and_kinds.at(i);
    }
  }
  return max_jint;
}

int Interval::next_usage(IntervalUseKind min_use_kind, int from) const {
  assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals");

  for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) {
    if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) >= min_use_kind) {
      return _use_pos_and_kinds.at(i);
    }
  }
  return max_jint;
}

int Interval::next_usage_exact(IntervalUseKind exact_use_kind, int from) const {
  assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals");

  for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) {
    if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) == exact_use_kind) {
      return _use_pos_and_kinds.at(i);
    }
  }
  return max_jint;
}

int Interval::previous_usage(IntervalUseKind min_use_kind, int from) const {
  assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals");

  int prev = 0;
  for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) {
    if (_use_pos_and_kinds.at(i) > from) {
      return prev;
    }
    if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) {
      prev = _use_pos_and_kinds.at(i);
    }
  }
  return prev;
}

void Interval::add_use_pos(int pos, IntervalUseKind use_kind) {
  assert(covers(pos, LIR_OpVisitState::inputMode), "use position not covered by live range");

  // do not add use positions for precolored intervals because
  // they are never used
  if (use_kind != noUse && reg_num() >= LIR_OprDesc::vreg_base) {
#ifdef ASSERT
    assert(_use_pos_and_kinds.length() % 2 == 0, "must be");
    for (int i = 0; i < _use_pos_and_kinds.length(); i += 2) {
      assert(pos <= _use_pos_and_kinds.at(i), "already added a use-position with lower position");
      assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind");
      if (i > 0) {
        assert(_use_pos_and_kinds.at(i) < _use_pos_and_kinds.at(i - 2), "not sorted descending");
      }
    }
#endif

    // Note: add_use is called in descending order, so list gets sorted
    //       automatically by just appending new use positions
    int len = _use_pos_and_kinds.length();
    if (len == 0 || _use_pos_and_kinds.at(len - 2) > pos) {
      _use_pos_and_kinds.append(pos);
      _use_pos_and_kinds.append(use_kind);
    } else if (_use_pos_and_kinds.at(len - 1) < use_kind) {
      assert(_use_pos_and_kinds.at(len - 2) == pos, "list not sorted correctly");
      _use_pos_and_kinds.at_put(len - 1, use_kind);
    }
  }
}

void Interval::add_range(int from, int to) {
  assert(from < to, "invalid range");
  assert(first() == Range::end() || to < first()->next()->from(), "not inserting at begin of interval");
  assert(from <= first()->to(), "not inserting at begin of interval");

  if (first()->from() <= to) {
    // join intersecting ranges
    first()->set_from(MIN2(from, first()->from()));
    first()->set_to  (MAX2(to,   first()->to()));
  } else {
    // insert new range
    _first = new Range(from, to, first());
  }
}

Interval* Interval::new_split_child() {
  // allocate new interval
  Interval* result = new Interval(-1);
  result->set_type(type());

  Interval* parent = split_parent();
  result->_split_parent = parent;
  result->set_register_hint(parent);

  // insert new interval in children-list of parent
  if (parent->_split_children.length() == 0) {
    assert(is_split_parent(), "list must be initialized at first split");

    parent->_split_children = IntervalList(4);
    parent->_split_children.append(this);
  }
  parent->_split_children.append(result);

  return result;
}

// split this interval at the specified position and return
// the remainder as a new interval.
//
// when an interval is split, a bi-directional link is established between the original interval
// (the split parent) and the intervals that are split off this interval (the split children)
// When a split child is split again, the new created interval is also a direct child
// of the original parent (there is no tree of split children stored, but a flat list)
// All split children are spilled to the same stack slot (stored in _canonical_spill_slot)
//
// Note: The new interval has no valid reg_num
Interval* Interval::split(int split_pos) {
  assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals");

  // allocate new interval
  Interval* result = new_split_child();

  // split the ranges
  Range* prev = NULL;
  Range* cur = _first;
  while (cur != Range::end() && cur->to() <= split_pos) {
    prev = cur;
    cur = cur->next();
  }
  assert(cur != Range::end(), "split interval after end of last range");

  if (cur->from() < split_pos) {
    result->_first = new Range(split_pos, cur->to(), cur->next());
    cur->set_to(split_pos);
    cur->set_next(Range::end());

  } else {
    assert(prev != NULL, "split before start of first range");
    result->_first = cur;
    prev->set_next(Range::end());
  }
  result->_current = result->_first;
  _cached_to = -1; // clear cached value

  // split list of use positions
  int total_len = _use_pos_and_kinds.length();
  int start_idx = total_len - 2;
  while (start_idx >= 0 && _use_pos_and_kinds.at(start_idx) < split_pos) {
    start_idx -= 2;
  }

  intStack new_use_pos_and_kinds(total_len - start_idx);
  int i;
  for (i = start_idx + 2; i < total_len; i++) {
    new_use_pos_and_kinds.append(_use_pos_and_kinds.at(i));
  }

  _use_pos_and_kinds.truncate(start_idx + 2);
  result->_use_pos_and_kinds = _use_pos_and_kinds;
  _use_pos_and_kinds = new_use_pos_and_kinds;

#ifdef ASSERT
  assert(_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos");
  assert(result->_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos");
  assert(_use_pos_and_kinds.length() + result->_use_pos_and_kinds.length() == total_len, "missed some entries");

  for (i = 0; i < _use_pos_and_kinds.length(); i += 2) {
    assert(_use_pos_and_kinds.at(i) < split_pos, "must be");
    assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind");
  }
  for (i = 0; i < result->_use_pos_and_kinds.length(); i += 2) {
    assert(result->_use_pos_and_kinds.at(i) >= split_pos, "must be");
    assert(result->_use_pos_and_kinds.at(i + 1) >= firstValidKind && result->_use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind");
  }
#endif

  return result;
}

// split this interval at the specified position and return
// the head as a new interval (the original interval is the tail)
//
// Currently, only the first range can be split, and the new interval
// must not have split positions
Interval* Interval::split_from_start(int split_pos) {
  assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals");
  assert(split_pos > from() && split_pos < to(), "can only split inside interval");
  assert(split_pos > _first->from() && split_pos <= _first->to(), "can only split inside first range");
  assert(first_usage(noUse) > split_pos, "can not split when use positions are present");

  // allocate new interval
  Interval* result = new_split_child();

  // the new created interval has only one range (checked by assertion above),
  // so the splitting of the ranges is very simple
  result->add_range(_first->from(), split_pos);

  if (split_pos == _first->to()) {
    assert(_first->next() != Range::end(), "must not be at end");
    _first = _first->next();
  } else {
    _first->set_from(split_pos);
  }

  return result;
}


// returns true if the op_id is inside the interval
bool Interval::covers(int op_id, LIR_OpVisitState::OprMode mode) const {
  Range* cur  = _first;

  while (cur != Range::end() && cur->to() < op_id) {
    cur = cur->next();
  }
  if (cur != Range::end()) {
    assert(cur->to() != cur->next()->from(), "ranges not separated");

    if (mode == LIR_OpVisitState::outputMode) {
      return cur->from() <= op_id && op_id < cur->to();
    } else {
      return cur->from() <= op_id && op_id <= cur->to();
    }
  }
  return false;
}

// returns true if the interval has any hole between hole_from and hole_to
// (even if the hole has only the length 1)
bool Interval::has_hole_between(int hole_from, int hole_to) {
  assert(hole_from < hole_to, "check");
  assert(from() <= hole_from && hole_to <= to(), "index out of interval");

  Range* cur  = _first;
  while (cur != Range::end()) {
    assert(cur->to() < cur->next()->from(), "no space between ranges");

    // hole-range starts before this range -> hole
    if (hole_from < cur->from()) {
      return true;

    // hole-range completely inside this range -> no hole
    } else if (hole_to <= cur->to()) {
      return false;

    // overlapping of hole-range with this range -> hole
    } else if (hole_from <= cur->to()) {
      return true;
    }

    cur = cur->next();
  }

  return false;
}


#ifndef PRODUCT
void Interval::print(outputStream* out) const {
  const char* SpillState2Name[] = { "no definition", "no spill store", "one spill store", "store at definition", "start in memory", "no optimization" };
  const char* UseKind2Name[] = { "N", "L", "S", "M" };

  const char* type_name;
  LIR_Opr opr = LIR_OprFact::illegal();
  if (reg_num() < LIR_OprDesc::vreg_base) {
    type_name = "fixed";
    // need a temporary operand for fixed intervals because type() cannot be called
    if (assigned_reg() >= pd_first_cpu_reg && assigned_reg() <= pd_last_cpu_reg) {
      opr = LIR_OprFact::single_cpu(assigned_reg());
    } else if (assigned_reg() >= pd_first_fpu_reg && assigned_reg() <= pd_last_fpu_reg) {
      opr = LIR_OprFact::single_fpu(assigned_reg() - pd_first_fpu_reg);
4536
#ifdef X86
D
duke 已提交
4537 4538 4539 4540 4541 4542 4543 4544
    } else if (assigned_reg() >= pd_first_xmm_reg && assigned_reg() <= pd_last_xmm_reg) {
      opr = LIR_OprFact::single_xmm(assigned_reg() - pd_first_xmm_reg);
#endif
    } else {
      ShouldNotReachHere();
    }
  } else {
    type_name = type2name(type());
4545 4546
    if (assigned_reg() != -1 &&
        (LinearScan::num_physical_regs(type()) == 1 || assigned_regHi() != -1)) {
D
duke 已提交
4547 4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576 4577 4578 4579 4580 4581 4582 4583 4584 4585 4586 4587 4588 4589 4590 4591 4592 4593 4594 4595 4596 4597 4598 4599 4600 4601 4602 4603 4604 4605 4606 4607 4608 4609 4610 4611 4612 4613 4614 4615 4616 4617 4618 4619 4620 4621 4622 4623 4624 4625 4626 4627 4628 4629 4630 4631 4632 4633 4634 4635 4636 4637 4638 4639 4640 4641 4642 4643 4644 4645 4646 4647 4648 4649 4650 4651 4652 4653 4654 4655 4656 4657 4658 4659 4660 4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675 4676 4677 4678 4679 4680 4681 4682 4683 4684 4685 4686 4687 4688 4689 4690 4691 4692 4693 4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706 4707 4708 4709 4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722 4723 4724 4725 4726 4727 4728 4729 4730 4731 4732 4733 4734 4735 4736 4737 4738 4739 4740 4741 4742 4743 4744 4745 4746 4747 4748 4749 4750 4751 4752 4753 4754 4755 4756 4757 4758 4759 4760 4761 4762 4763 4764 4765 4766 4767 4768 4769 4770 4771 4772 4773 4774 4775 4776 4777 4778 4779 4780 4781 4782 4783 4784 4785 4786 4787 4788 4789 4790 4791 4792 4793 4794 4795 4796 4797 4798 4799 4800 4801 4802 4803 4804 4805 4806 4807 4808 4809 4810 4811 4812 4813 4814 4815 4816 4817 4818 4819 4820 4821 4822 4823 4824 4825 4826 4827 4828 4829 4830 4831 4832 4833 4834 4835 4836 4837 4838 4839 4840 4841 4842 4843 4844 4845 4846 4847 4848 4849 4850 4851 4852 4853 4854 4855 4856 4857 4858 4859 4860 4861 4862 4863 4864 4865 4866 4867 4868 4869 4870 4871 4872 4873 4874 4875 4876 4877 4878 4879 4880 4881 4882 4883 4884 4885 4886 4887 4888 4889 4890 4891 4892 4893 4894 4895 4896 4897 4898 4899 4900 4901 4902 4903 4904 4905 4906 4907 4908 4909 4910 4911 4912 4913 4914 4915 4916 4917 4918 4919 4920 4921 4922 4923 4924 4925 4926 4927 4928 4929 4930 4931 4932 4933 4934 4935 4936 4937 4938 4939 4940 4941 4942 4943 4944 4945 4946 4947 4948 4949 4950 4951 4952 4953 4954 4955 4956 4957 4958 4959 4960 4961 4962 4963 4964 4965 4966 4967 4968 4969 4970 4971 4972 4973 4974 4975 4976 4977 4978 4979 4980 4981 4982 4983 4984 4985 4986 4987 4988 4989 4990 4991 4992 4993 4994 4995 4996 4997 4998 4999 5000 5001 5002 5003 5004 5005 5006 5007 5008 5009 5010 5011 5012 5013 5014 5015 5016 5017 5018 5019 5020 5021 5022 5023 5024 5025 5026 5027 5028 5029 5030 5031 5032 5033 5034 5035 5036 5037 5038 5039 5040 5041 5042 5043 5044 5045 5046 5047 5048 5049 5050 5051 5052 5053 5054 5055 5056 5057 5058 5059 5060 5061 5062 5063 5064 5065 5066 5067 5068 5069 5070 5071 5072 5073 5074 5075 5076 5077 5078 5079 5080 5081 5082 5083 5084 5085 5086 5087 5088 5089 5090 5091 5092 5093 5094 5095 5096 5097 5098 5099 5100 5101 5102 5103 5104 5105 5106 5107 5108 5109 5110 5111 5112 5113 5114 5115 5116 5117 5118 5119 5120 5121 5122 5123 5124 5125 5126 5127 5128 5129 5130 5131 5132 5133 5134 5135 5136 5137 5138 5139 5140 5141 5142 5143 5144 5145 5146 5147 5148 5149 5150 5151 5152 5153 5154 5155 5156 5157 5158 5159 5160 5161 5162 5163 5164 5165 5166 5167 5168 5169 5170 5171 5172 5173 5174 5175 5176 5177 5178 5179 5180 5181 5182 5183 5184 5185 5186 5187 5188 5189 5190 5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201 5202 5203 5204 5205 5206 5207 5208 5209 5210 5211 5212 5213 5214 5215 5216 5217 5218 5219 5220 5221 5222 5223 5224 5225 5226 5227 5228 5229 5230 5231 5232 5233 5234 5235 5236 5237 5238 5239 5240 5241 5242 5243 5244 5245 5246 5247 5248 5249 5250 5251 5252 5253 5254 5255 5256 5257 5258 5259 5260 5261 5262 5263 5264 5265 5266 5267 5268 5269 5270 5271 5272 5273 5274 5275 5276 5277 5278 5279 5280 5281 5282 5283 5284 5285 5286 5287 5288 5289 5290 5291 5292 5293 5294 5295 5296 5297 5298 5299 5300 5301 5302 5303 5304 5305 5306 5307 5308 5309 5310 5311 5312 5313 5314 5315 5316 5317 5318 5319 5320 5321 5322 5323 5324 5325 5326 5327 5328 5329 5330 5331 5332 5333 5334 5335 5336 5337 5338 5339 5340 5341 5342 5343 5344 5345 5346 5347 5348 5349 5350 5351 5352 5353 5354 5355 5356 5357 5358 5359 5360 5361 5362 5363 5364 5365 5366 5367 5368 5369 5370 5371 5372 5373 5374 5375 5376 5377 5378 5379 5380 5381 5382 5383 5384 5385 5386 5387 5388 5389 5390 5391 5392 5393 5394 5395 5396 5397 5398 5399 5400 5401 5402 5403 5404 5405 5406 5407 5408 5409 5410 5411 5412 5413 5414 5415 5416 5417 5418 5419 5420 5421 5422 5423 5424 5425 5426 5427 5428 5429 5430 5431 5432 5433 5434 5435 5436 5437 5438 5439 5440 5441 5442 5443 5444 5445 5446 5447 5448 5449 5450 5451 5452 5453 5454 5455 5456 5457 5458 5459 5460 5461 5462 5463 5464 5465 5466 5467 5468 5469 5470 5471 5472 5473 5474 5475 5476 5477 5478 5479 5480 5481 5482 5483 5484 5485 5486 5487 5488 5489 5490 5491 5492 5493 5494 5495 5496 5497 5498 5499 5500 5501 5502 5503 5504 5505 5506 5507 5508 5509 5510 5511 5512 5513 5514 5515 5516 5517 5518 5519 5520 5521 5522 5523 5524 5525 5526 5527 5528 5529 5530 5531 5532 5533 5534 5535 5536 5537 5538 5539 5540 5541 5542 5543 5544 5545 5546 5547 5548 5549 5550 5551 5552 5553 5554 5555 5556 5557 5558 5559 5560 5561 5562 5563 5564 5565 5566 5567 5568 5569 5570 5571 5572 5573 5574 5575 5576 5577 5578 5579 5580 5581 5582 5583 5584 5585 5586 5587 5588 5589 5590 5591 5592 5593 5594 5595 5596 5597 5598 5599 5600 5601 5602 5603 5604 5605 5606 5607 5608 5609 5610 5611 5612 5613 5614
      opr = LinearScan::calc_operand_for_interval(this);
    }
  }

  out->print("%d %s ", reg_num(), type_name);
  if (opr->is_valid()) {
    out->print("\"");
    opr->print(out);
    out->print("\" ");
  }
  out->print("%d %d ", split_parent()->reg_num(), (register_hint(false) != NULL ? register_hint(false)->reg_num() : -1));

  // print ranges
  Range* cur = _first;
  while (cur != Range::end()) {
    cur->print(out);
    cur = cur->next();
    assert(cur != NULL, "range list not closed with range sentinel");
  }

  // print use positions
  int prev = 0;
  assert(_use_pos_and_kinds.length() % 2 == 0, "must be");
  for (int i =_use_pos_and_kinds.length() - 2; i >= 0; i -= 2) {
    assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind");
    assert(prev < _use_pos_and_kinds.at(i), "use positions not sorted");

    out->print("%d %s ", _use_pos_and_kinds.at(i), UseKind2Name[_use_pos_and_kinds.at(i + 1)]);
    prev = _use_pos_and_kinds.at(i);
  }

  out->print(" \"%s\"", SpillState2Name[spill_state()]);
  out->cr();
}
#endif



// **** Implementation of IntervalWalker ****************************

IntervalWalker::IntervalWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first)
 : _compilation(allocator->compilation())
 , _allocator(allocator)
{
  _unhandled_first[fixedKind] = unhandled_fixed_first;
  _unhandled_first[anyKind]   = unhandled_any_first;
  _active_first[fixedKind]    = Interval::end();
  _inactive_first[fixedKind]  = Interval::end();
  _active_first[anyKind]      = Interval::end();
  _inactive_first[anyKind]    = Interval::end();
  _current_position = -1;
  _current = NULL;
  next_interval();
}


// append interval at top of list
void IntervalWalker::append_unsorted(Interval** list, Interval* interval) {
  interval->set_next(*list); *list = interval;
}


// append interval in order of current range from()
void IntervalWalker::append_sorted(Interval** list, Interval* interval) {
  Interval* prev = NULL;
  Interval* cur  = *list;
  while (cur->current_from() < interval->current_from()) {
    prev = cur; cur = cur->next();
  }
  if (prev == NULL) {
    *list = interval;
  } else {
    prev->set_next(interval);
  }
  interval->set_next(cur);
}

void IntervalWalker::append_to_unhandled(Interval** list, Interval* interval) {
  assert(interval->from() >= current()->current_from(), "cannot append new interval before current walk position");

  Interval* prev = NULL;
  Interval* cur  = *list;
  while (cur->from() < interval->from() || (cur->from() == interval->from() && cur->first_usage(noUse) < interval->first_usage(noUse))) {
    prev = cur; cur = cur->next();
  }
  if (prev == NULL) {
    *list = interval;
  } else {
    prev->set_next(interval);
  }
  interval->set_next(cur);
}


inline bool IntervalWalker::remove_from_list(Interval** list, Interval* i) {
  while (*list != Interval::end() && *list != i) {
    list = (*list)->next_addr();
  }
  if (*list != Interval::end()) {
    assert(*list == i, "check");
    *list = (*list)->next();
    return true;
  } else {
    return false;
  }
}

void IntervalWalker::remove_from_list(Interval* i) {
  bool deleted;

  if (i->state() == activeState) {
    deleted = remove_from_list(active_first_addr(anyKind), i);
  } else {
    assert(i->state() == inactiveState, "invalid state");
    deleted = remove_from_list(inactive_first_addr(anyKind), i);
  }

  assert(deleted, "interval has not been found in list");
}


void IntervalWalker::walk_to(IntervalState state, int from) {
  assert (state == activeState || state == inactiveState, "wrong state");
  for_each_interval_kind(kind) {
    Interval** prev = state == activeState ? active_first_addr(kind) : inactive_first_addr(kind);
    Interval* next   = *prev;
    while (next->current_from() <= from) {
      Interval* cur = next;
      next = cur->next();

      bool range_has_changed = false;
      while (cur->current_to() <= from) {
        cur->next_range();
        range_has_changed = true;
      }

      // also handle move from inactive list to active list
      range_has_changed = range_has_changed || (state == inactiveState && cur->current_from() <= from);

      if (range_has_changed) {
        // remove cur from list
        *prev = next;
        if (cur->current_at_end()) {
          // move to handled state (not maintained as a list)
          cur->set_state(handledState);
          interval_moved(cur, kind, state, handledState);
        } else if (cur->current_from() <= from){
          // sort into active list
          append_sorted(active_first_addr(kind), cur);
          cur->set_state(activeState);
          if (*prev == cur) {
            assert(state == activeState, "check");
            prev = cur->next_addr();
          }
          interval_moved(cur, kind, state, activeState);
        } else {
          // sort into inactive list
          append_sorted(inactive_first_addr(kind), cur);
          cur->set_state(inactiveState);
          if (*prev == cur) {
            assert(state == inactiveState, "check");
            prev = cur->next_addr();
          }
          interval_moved(cur, kind, state, inactiveState);
        }
      } else {
        prev = cur->next_addr();
        continue;
      }
    }
  }
}


void IntervalWalker::next_interval() {
  IntervalKind kind;
  Interval* any   = _unhandled_first[anyKind];
  Interval* fixed = _unhandled_first[fixedKind];

  if (any != Interval::end()) {
    // intervals may start at same position -> prefer fixed interval
    kind = fixed != Interval::end() && fixed->from() <= any->from() ? fixedKind : anyKind;

    assert (kind == fixedKind && fixed->from() <= any->from() ||
            kind == anyKind   && any->from() <= fixed->from(), "wrong interval!!!");
    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");

  } else if (fixed != Interval::end()) {
    kind = fixedKind;
  } else {
    _current = NULL; return;
  }
  _current_kind = kind;
  _current = _unhandled_first[kind];
  _unhandled_first[kind] = _current->next();
  _current->set_next(Interval::end());
  _current->rewind_range();
}


void IntervalWalker::walk_to(int lir_op_id) {
  assert(_current_position <= lir_op_id, "can not walk backwards");
  while (current() != NULL) {
    bool is_active = current()->from() <= lir_op_id;
    int id = is_active ? current()->from() : lir_op_id;

    TRACE_LINEAR_SCAN(2, if (_current_position < id) { tty->cr(); tty->print_cr("walk_to(%d) **************************************************************", id); })

    // set _current_position prior to call of walk_to
    _current_position = id;

    // call walk_to even if _current_position == id
    walk_to(activeState, id);
    walk_to(inactiveState, id);

    if (is_active) {
      current()->set_state(activeState);
      if (activate_current()) {
        append_sorted(active_first_addr(current_kind()), current());
        interval_moved(current(), current_kind(), unhandledState, activeState);
      }

      next_interval();
    } else {
      return;
    }
  }
}

void IntervalWalker::interval_moved(Interval* interval, IntervalKind kind, IntervalState from, IntervalState to) {
#ifndef PRODUCT
  if (TraceLinearScanLevel >= 4) {
    #define print_state(state) \
    switch(state) {\
      case unhandledState: tty->print("unhandled"); break;\
      case activeState: tty->print("active"); break;\
      case inactiveState: tty->print("inactive"); break;\
      case handledState: tty->print("handled"); break;\
      default: ShouldNotReachHere(); \
    }

    print_state(from); tty->print(" to "); print_state(to);
    tty->fill_to(23);
    interval->print();

    #undef print_state
  }
#endif
}



// **** Implementation of LinearScanWalker **************************

LinearScanWalker::LinearScanWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first)
  : IntervalWalker(allocator, unhandled_fixed_first, unhandled_any_first)
  , _move_resolver(allocator)
{
  for (int i = 0; i < LinearScan::nof_regs; i++) {
    _spill_intervals[i] = new IntervalList(2);
  }
}


inline void LinearScanWalker::init_use_lists(bool only_process_use_pos) {
  for (int i = _first_reg; i <= _last_reg; i++) {
    _use_pos[i] = max_jint;

    if (!only_process_use_pos) {
      _block_pos[i] = max_jint;
      _spill_intervals[i]->clear();
    }
  }
}

inline void LinearScanWalker::exclude_from_use(int reg) {
  assert(reg < LinearScan::nof_regs, "interval must have a register assigned (stack slots not allowed)");
  if (reg >= _first_reg && reg <= _last_reg) {
    _use_pos[reg] = 0;
  }
}
inline void LinearScanWalker::exclude_from_use(Interval* i) {
  assert(i->assigned_reg() != any_reg, "interval has no register assigned");

  exclude_from_use(i->assigned_reg());
  exclude_from_use(i->assigned_regHi());
}

inline void LinearScanWalker::set_use_pos(int reg, Interval* i, int use_pos, bool only_process_use_pos) {
  assert(use_pos != 0, "must use exclude_from_use to set use_pos to 0");

  if (reg >= _first_reg && reg <= _last_reg) {
    if (_use_pos[reg] > use_pos) {
      _use_pos[reg] = use_pos;
    }
    if (!only_process_use_pos) {
      _spill_intervals[reg]->append(i);
    }
  }
}
inline void LinearScanWalker::set_use_pos(Interval* i, int use_pos, bool only_process_use_pos) {
  assert(i->assigned_reg() != any_reg, "interval has no register assigned");
  if (use_pos != -1) {
    set_use_pos(i->assigned_reg(), i, use_pos, only_process_use_pos);
    set_use_pos(i->assigned_regHi(), i, use_pos, only_process_use_pos);
  }
}

inline void LinearScanWalker::set_block_pos(int reg, Interval* i, int block_pos) {
  if (reg >= _first_reg && reg <= _last_reg) {
    if (_block_pos[reg] > block_pos) {
      _block_pos[reg] = block_pos;
    }
    if (_use_pos[reg] > block_pos) {
      _use_pos[reg] = block_pos;
    }
  }
}
inline void LinearScanWalker::set_block_pos(Interval* i, int block_pos) {
  assert(i->assigned_reg() != any_reg, "interval has no register assigned");
  if (block_pos != -1) {
    set_block_pos(i->assigned_reg(), i, block_pos);
    set_block_pos(i->assigned_regHi(), i, block_pos);
  }
}


void LinearScanWalker::free_exclude_active_fixed() {
  Interval* list = active_first(fixedKind);
  while (list != Interval::end()) {
    assert(list->assigned_reg() < LinearScan::nof_regs, "active interval must have a register assigned");
    exclude_from_use(list);
    list = list->next();
  }
}

void LinearScanWalker::free_exclude_active_any() {
  Interval* list = active_first(anyKind);
  while (list != Interval::end()) {
    exclude_from_use(list);
    list = list->next();
  }
}

void LinearScanWalker::free_collect_inactive_fixed(Interval* cur) {
  Interval* list = inactive_first(fixedKind);
  while (list != Interval::end()) {
    if (cur->to() <= list->current_from()) {
      assert(list->current_intersects_at(cur) == -1, "must not intersect");
      set_use_pos(list, list->current_from(), true);
    } else {
      set_use_pos(list, list->current_intersects_at(cur), true);
    }
    list = list->next();
  }
}

void LinearScanWalker::free_collect_inactive_any(Interval* cur) {
  Interval* list = inactive_first(anyKind);
  while (list != Interval::end()) {
    set_use_pos(list, list->current_intersects_at(cur), true);
    list = list->next();
  }
}

void LinearScanWalker::free_collect_unhandled(IntervalKind kind, Interval* cur) {
  Interval* list = unhandled_first(kind);
  while (list != Interval::end()) {
    set_use_pos(list, list->intersects_at(cur), true);
    if (kind == fixedKind && cur->to() <= list->from()) {
      set_use_pos(list, list->from(), true);
    }
    list = list->next();
  }
}

void LinearScanWalker::spill_exclude_active_fixed() {
  Interval* list = active_first(fixedKind);
  while (list != Interval::end()) {
    exclude_from_use(list);
    list = list->next();
  }
}

void LinearScanWalker::spill_block_unhandled_fixed(Interval* cur) {
  Interval* list = unhandled_first(fixedKind);
  while (list != Interval::end()) {
    set_block_pos(list, list->intersects_at(cur));
    list = list->next();
  }
}

void LinearScanWalker::spill_block_inactive_fixed(Interval* cur) {
  Interval* list = inactive_first(fixedKind);
  while (list != Interval::end()) {
    if (cur->to() > list->current_from()) {
      set_block_pos(list, list->current_intersects_at(cur));
    } else {
      assert(list->current_intersects_at(cur) == -1, "invalid optimization: intervals intersect");
    }

    list = list->next();
  }
}

void LinearScanWalker::spill_collect_active_any() {
  Interval* list = active_first(anyKind);
  while (list != Interval::end()) {
    set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false);
    list = list->next();
  }
}

void LinearScanWalker::spill_collect_inactive_any(Interval* cur) {
  Interval* list = inactive_first(anyKind);
  while (list != Interval::end()) {
    if (list->current_intersects(cur)) {
      set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false);
    }
    list = list->next();
  }
}


void LinearScanWalker::insert_move(int op_id, Interval* src_it, Interval* dst_it) {
  // output all moves here. When source and target are equal, the move is
  // optimized away later in assign_reg_nums

  op_id = (op_id + 1) & ~1;
  BlockBegin* op_block = allocator()->block_of_op_with_id(op_id);
  assert(op_id > 0 && allocator()->block_of_op_with_id(op_id - 2) == op_block, "cannot insert move at block boundary");

  // calculate index of instruction inside instruction list of current block
  // the minimal index (for a block with no spill moves) can be calculated because the
  // numbering of instructions is known.
  // When the block already contains spill moves, the index must be increased until the
  // correct index is reached.
  LIR_OpList* list = op_block->lir()->instructions_list();
  int index = (op_id - list->at(0)->id()) / 2;
  assert(list->at(index)->id() <= op_id, "error in calculation");

  while (list->at(index)->id() != op_id) {
    index++;
    assert(0 <= index && index < list->length(), "index out of bounds");
  }
  assert(1 <= index && index < list->length(), "index out of bounds");
  assert(list->at(index)->id() == op_id, "error in calculation");

  // insert new instruction before instruction at position index
  _move_resolver.move_insert_position(op_block->lir(), index - 1);
  _move_resolver.add_mapping(src_it, dst_it);
}


int LinearScanWalker::find_optimal_split_pos(BlockBegin* min_block, BlockBegin* max_block, int max_split_pos) {
  int from_block_nr = min_block->linear_scan_number();
  int to_block_nr = max_block->linear_scan_number();

  assert(0 <= from_block_nr && from_block_nr < block_count(), "out of range");
  assert(0 <= to_block_nr && to_block_nr < block_count(), "out of range");
  assert(from_block_nr < to_block_nr, "must cross block boundary");

  // Try to split at end of max_block. If this would be after
  // max_split_pos, then use the begin of max_block
  int optimal_split_pos = max_block->last_lir_instruction_id() + 2;
  if (optimal_split_pos > max_split_pos) {
    optimal_split_pos = max_block->first_lir_instruction_id();
  }

  int min_loop_depth = max_block->loop_depth();
  for (int i = to_block_nr - 1; i >= from_block_nr; i--) {
    BlockBegin* cur = block_at(i);

    if (cur->loop_depth() < min_loop_depth) {
      // block with lower loop-depth found -> split at the end of this block
      min_loop_depth = cur->loop_depth();
      optimal_split_pos = cur->last_lir_instruction_id() + 2;
    }
  }
  assert(optimal_split_pos > allocator()->max_lir_op_id() || allocator()->is_block_begin(optimal_split_pos), "algorithm must move split pos to block boundary");

  return optimal_split_pos;
}


int LinearScanWalker::find_optimal_split_pos(Interval* it, int min_split_pos, int max_split_pos, bool do_loop_optimization) {
  int optimal_split_pos = -1;
  if (min_split_pos == max_split_pos) {
    // trivial case, no optimization of split position possible
    TRACE_LINEAR_SCAN(4, tty->print_cr("      min-pos and max-pos are equal, no optimization possible"));
    optimal_split_pos = min_split_pos;

  } else {
    assert(min_split_pos < max_split_pos, "must be true then");
    assert(min_split_pos > 0, "cannot access min_split_pos - 1 otherwise");

    // reason for using min_split_pos - 1: when the minimal split pos is exactly at the
    // beginning of a block, then min_split_pos is also a possible split position.
    // Use the block before as min_block, because then min_block->last_lir_instruction_id() + 2 == min_split_pos
    BlockBegin* min_block = allocator()->block_of_op_with_id(min_split_pos - 1);

    // reason for using max_split_pos - 1: otherwise there would be an assertion failure
    // when an interval ends at the end of the last block of the method
    // (in this case, max_split_pos == allocator()->max_lir_op_id() + 2, and there is no
    // block at this op_id)
    BlockBegin* max_block = allocator()->block_of_op_with_id(max_split_pos - 1);

    assert(min_block->linear_scan_number() <= max_block->linear_scan_number(), "invalid order");
    if (min_block == max_block) {
      // split position cannot be moved to block boundary, so split as late as possible
      TRACE_LINEAR_SCAN(4, tty->print_cr("      cannot move split pos to block boundary because min_pos and max_pos are in same block"));
      optimal_split_pos = max_split_pos;

    } else if (it->has_hole_between(max_split_pos - 1, max_split_pos) && !allocator()->is_block_begin(max_split_pos)) {
      // Do not move split position if the interval has a hole before max_split_pos.
      // Intervals resulting from Phi-Functions have more than one definition (marked
      // as mustHaveRegister) with a hole before each definition. When the register is needed
      // for the second definition, an earlier reloading is unnecessary.
      TRACE_LINEAR_SCAN(4, tty->print_cr("      interval has hole just before max_split_pos, so splitting at max_split_pos"));
      optimal_split_pos = max_split_pos;

    } else {
      // seach optimal block boundary between min_split_pos and max_split_pos
      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()));

      if (do_loop_optimization) {
        // Loop optimization: if a loop-end marker is found between min- and max-position,
        // then split before this loop
        int loop_end_pos = it->next_usage_exact(loopEndMarker, min_block->last_lir_instruction_id() + 2);
        TRACE_LINEAR_SCAN(4, tty->print_cr("      loop optimization: loop end found at pos %d", loop_end_pos));

        assert(loop_end_pos > min_split_pos, "invalid order");
        if (loop_end_pos < max_split_pos) {
          // loop-end marker found between min- and max-position
          // if it is not the end marker for the same loop as the min-position, then move
          // the max-position to this loop block.
          // Desired result: uses tagged as shouldHaveRegister inside a loop cause a reloading
          // of the interval (normally, only mustHaveRegister causes a reloading)
          BlockBegin* loop_block = allocator()->block_of_op_with_id(loop_end_pos);

          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()));
          assert(loop_block != min_block, "loop_block and min_block must be different because block boundary is needed between");

          optimal_split_pos = find_optimal_split_pos(min_block, loop_block, loop_block->last_lir_instruction_id() + 2);
          if (optimal_split_pos == loop_block->last_lir_instruction_id() + 2) {
            optimal_split_pos = -1;
            TRACE_LINEAR_SCAN(4, tty->print_cr("      loop optimization not necessary"));
          } else {
            TRACE_LINEAR_SCAN(4, tty->print_cr("      loop optimization successful"));
          }
        }
      }

      if (optimal_split_pos == -1) {
        // not calculated by loop optimization
        optimal_split_pos = find_optimal_split_pos(min_block, max_block, max_split_pos);
      }
    }
  }
  TRACE_LINEAR_SCAN(4, tty->print_cr("      optimal split position: %d", optimal_split_pos));

  return optimal_split_pos;
}


/*
  split an interval at the optimal position between min_split_pos and
  max_split_pos in two parts:
  1) the left part has already a location assigned
  2) the right part is sorted into to the unhandled-list
*/
void LinearScanWalker::split_before_usage(Interval* it, int min_split_pos, int max_split_pos) {
  TRACE_LINEAR_SCAN(2, tty->print   ("----- splitting interval: "); it->print());
  TRACE_LINEAR_SCAN(2, tty->print_cr("      between %d and %d", min_split_pos, max_split_pos));

  assert(it->from() < min_split_pos,         "cannot split at start of interval");
  assert(current_position() < min_split_pos, "cannot split before current position");
  assert(min_split_pos <= max_split_pos,     "invalid order");
  assert(max_split_pos <= it->to(),          "cannot split after end of interval");

  int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, true);

  assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range");
  assert(optimal_split_pos <= it->to(),  "cannot split after end of interval");
  assert(optimal_split_pos > it->from(), "cannot split at start of interval");

  if (optimal_split_pos == it->to() && it->next_usage(mustHaveRegister, min_split_pos) == max_jint) {
    // the split position would be just before the end of the interval
    // -> no split at all necessary
    TRACE_LINEAR_SCAN(4, tty->print_cr("      no split necessary because optimal split position is at end of interval"));
    return;
  }

  // must calculate this before the actual split is performed and before split position is moved to odd op_id
  bool move_necessary = !allocator()->is_block_begin(optimal_split_pos) && !it->has_hole_between(optimal_split_pos - 1, optimal_split_pos);

  if (!allocator()->is_block_begin(optimal_split_pos)) {
    // move position before actual instruction (odd op_id)
    optimal_split_pos = (optimal_split_pos - 1) | 1;
  }

  TRACE_LINEAR_SCAN(4, tty->print_cr("      splitting at position %d", optimal_split_pos));
  assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary");
  assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary");

  Interval* split_part = it->split(optimal_split_pos);

  allocator()->append_interval(split_part);
  allocator()->copy_register_flags(it, split_part);
  split_part->set_insert_move_when_activated(move_necessary);
  append_to_unhandled(unhandled_first_addr(anyKind), split_part);

  TRACE_LINEAR_SCAN(2, tty->print_cr("      split interval in two parts (insert_move_when_activated: %d)", move_necessary));
  TRACE_LINEAR_SCAN(2, tty->print   ("      "); it->print());
  TRACE_LINEAR_SCAN(2, tty->print   ("      "); split_part->print());
}

/*
  split an interval at the optimal position between min_split_pos and
  max_split_pos in two parts:
  1) the left part has already a location assigned
  2) the right part is always on the stack and therefore ignored in further processing
*/
void LinearScanWalker::split_for_spilling(Interval* it) {
  // calculate allowed range of splitting position
  int max_split_pos = current_position();
  int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, max_split_pos) + 1, it->from());

  TRACE_LINEAR_SCAN(2, tty->print   ("----- splitting and spilling interval: "); it->print());
  TRACE_LINEAR_SCAN(2, tty->print_cr("      between %d and %d", min_split_pos, max_split_pos));

  assert(it->state() == activeState,     "why spill interval that is not active?");
  assert(it->from() <= min_split_pos,    "cannot split before start of interval");
  assert(min_split_pos <= max_split_pos, "invalid order");
  assert(max_split_pos < it->to(),       "cannot split at end end of interval");
  assert(current_position() < it->to(),  "interval must not end before current position");

  if (min_split_pos == it->from()) {
    // the whole interval is never used, so spill it entirely to memory
    TRACE_LINEAR_SCAN(2, tty->print_cr("      spilling entire interval because split pos is at beginning of interval"));
    assert(it->first_usage(shouldHaveRegister) > current_position(), "interval must not have use position before current_position");

    allocator()->assign_spill_slot(it);
    allocator()->change_spill_state(it, min_split_pos);

    // Also kick parent intervals out of register to memory when they have no use
    // position. This avoids short interval in register surrounded by intervals in
    // memory -> avoid useless moves from memory to register and back
    Interval* parent = it;
    while (parent != NULL && parent->is_split_child()) {
      parent = parent->split_child_before_op_id(parent->from());

      if (parent->assigned_reg() < LinearScan::nof_regs) {
        if (parent->first_usage(shouldHaveRegister) == max_jint) {
          // parent is never used, so kick it out of its assigned register
          TRACE_LINEAR_SCAN(4, tty->print_cr("      kicking out interval %d out of its register because it is never used", parent->reg_num()));
          allocator()->assign_spill_slot(parent);
        } else {
          // do not go further back because the register is actually used by the interval
          parent = NULL;
        }
      }
    }

  } else {
    // search optimal split pos, split interval and spill only the right hand part
    int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, false);

    assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range");
    assert(optimal_split_pos < it->to(), "cannot split at end of interval");
    assert(optimal_split_pos >= it->from(), "cannot split before start of interval");

    if (!allocator()->is_block_begin(optimal_split_pos)) {
      // move position before actual instruction (odd op_id)
      optimal_split_pos = (optimal_split_pos - 1) | 1;
    }

    TRACE_LINEAR_SCAN(4, tty->print_cr("      splitting at position %d", optimal_split_pos));
    assert(allocator()->is_block_begin(optimal_split_pos)  || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary");
    assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary");

    Interval* spilled_part = it->split(optimal_split_pos);
    allocator()->append_interval(spilled_part);
    allocator()->assign_spill_slot(spilled_part);
    allocator()->change_spill_state(spilled_part, optimal_split_pos);

    if (!allocator()->is_block_begin(optimal_split_pos)) {
      TRACE_LINEAR_SCAN(4, tty->print_cr("      inserting move from interval %d to %d", it->reg_num(), spilled_part->reg_num()));
      insert_move(optimal_split_pos, it, spilled_part);
    }

    // the current_split_child is needed later when moves are inserted for reloading
    assert(spilled_part->current_split_child() == it, "overwriting wrong current_split_child");
    spilled_part->make_current_split_child();

    TRACE_LINEAR_SCAN(2, tty->print_cr("      split interval in two parts"));
    TRACE_LINEAR_SCAN(2, tty->print   ("      "); it->print());
    TRACE_LINEAR_SCAN(2, tty->print   ("      "); spilled_part->print());
  }
}


void LinearScanWalker::split_stack_interval(Interval* it) {
  int min_split_pos = current_position() + 1;
  int max_split_pos = MIN2(it->first_usage(shouldHaveRegister), it->to());

  split_before_usage(it, min_split_pos, max_split_pos);
}

void LinearScanWalker::split_when_partial_register_available(Interval* it, int register_available_until) {
  int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, register_available_until), it->from() + 1);
  int max_split_pos = register_available_until;

  split_before_usage(it, min_split_pos, max_split_pos);
}

void LinearScanWalker::split_and_spill_interval(Interval* it) {
  assert(it->state() == activeState || it->state() == inactiveState, "other states not allowed");

  int current_pos = current_position();
  if (it->state() == inactiveState) {
    // the interval is currently inactive, so no spill slot is needed for now.
    // when the split part is activated, the interval has a new chance to get a register,
    // so in the best case no stack slot is necessary
    assert(it->has_hole_between(current_pos - 1, current_pos + 1), "interval can not be inactive otherwise");
    split_before_usage(it, current_pos + 1, current_pos + 1);

  } else {
    // search the position where the interval must have a register and split
    // at the optimal position before.
    // The new created part is added to the unhandled list and will get a register
    // when it is activated
    int min_split_pos = current_pos + 1;
    int max_split_pos = MIN2(it->next_usage(mustHaveRegister, min_split_pos), it->to());

    split_before_usage(it, min_split_pos, max_split_pos);

    assert(it->next_usage(mustHaveRegister, current_pos) == max_jint, "the remaining part is spilled to stack and therefore has no register");
    split_for_spilling(it);
  }
}


int LinearScanWalker::find_free_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) {
  int min_full_reg = any_reg;
  int max_partial_reg = any_reg;

  for (int i = _first_reg; i <= _last_reg; i++) {
    if (i == ignore_reg) {
      // this register must be ignored

    } else if (_use_pos[i] >= interval_to) {
      // this register is free for the full interval
      if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) {
        min_full_reg = i;
      }
    } else if (_use_pos[i] > reg_needed_until) {
      // this register is at least free until reg_needed_until
      if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) {
        max_partial_reg = i;
      }
    }
  }

  if (min_full_reg != any_reg) {
    return min_full_reg;
  } else if (max_partial_reg != any_reg) {
    *need_split = true;
    return max_partial_reg;
  } else {
    return any_reg;
  }
}

int LinearScanWalker::find_free_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) {
  assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm");

  int min_full_reg = any_reg;
  int max_partial_reg = any_reg;

  for (int i = _first_reg; i < _last_reg; i+=2) {
    if (_use_pos[i] >= interval_to && _use_pos[i + 1] >= interval_to) {
      // this register is free for the full interval
      if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) {
        min_full_reg = i;
      }
    } else if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) {
      // this register is at least free until reg_needed_until
      if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) {
        max_partial_reg = i;
      }
    }
  }

  if (min_full_reg != any_reg) {
    return min_full_reg;
  } else if (max_partial_reg != any_reg) {
    *need_split = true;
    return max_partial_reg;
  } else {
    return any_reg;
  }
}


bool LinearScanWalker::alloc_free_reg(Interval* cur) {
  TRACE_LINEAR_SCAN(2, tty->print("trying to find free register for "); cur->print());

  init_use_lists(true);
  free_exclude_active_fixed();
  free_exclude_active_any();
  free_collect_inactive_fixed(cur);
  free_collect_inactive_any(cur);
//  free_collect_unhandled(fixedKind, cur);
  assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0");

  // _use_pos contains the start of the next interval that has this register assigned
  // (either as a fixed register or a normal allocated register in the past)
  // only intervals overlapping with cur are processed, non-overlapping invervals can be ignored safely
  TRACE_LINEAR_SCAN(4, tty->print_cr("      state of registers:"));
  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]));

  int hint_reg, hint_regHi;
  Interval* register_hint = cur->register_hint();
  if (register_hint != NULL) {
    hint_reg = register_hint->assigned_reg();
    hint_regHi = register_hint->assigned_regHi();

    if (allocator()->is_precolored_cpu_interval(register_hint)) {
      assert(hint_reg != any_reg && hint_regHi == any_reg, "must be for fixed intervals");
      hint_regHi = hint_reg + 1;  // connect e.g. eax-edx
    }
    TRACE_LINEAR_SCAN(4, tty->print("      hint registers %d, %d from interval ", hint_reg, hint_regHi); register_hint->print());

  } else {
    hint_reg = any_reg;
    hint_regHi = any_reg;
  }
  assert(hint_reg == any_reg || hint_reg != hint_regHi, "hint reg and regHi equal");
  assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned to interval");

  // the register must be free at least until this position
  int reg_needed_until = cur->from() + 1;
  int interval_to = cur->to();

  bool need_split = false;
  int split_pos = -1;
  int reg = any_reg;
  int regHi = any_reg;

  if (_adjacent_regs) {
    reg = find_free_double_reg(reg_needed_until, interval_to, hint_reg, &need_split);
    regHi = reg + 1;
    if (reg == any_reg) {
      return false;
    }
    split_pos = MIN2(_use_pos[reg], _use_pos[regHi]);

  } else {
    reg = find_free_reg(reg_needed_until, interval_to, hint_reg, any_reg, &need_split);
    if (reg == any_reg) {
      return false;
    }
    split_pos = _use_pos[reg];

    if (_num_phys_regs == 2) {
      regHi = find_free_reg(reg_needed_until, interval_to, hint_regHi, reg, &need_split);

      if (_use_pos[reg] < interval_to && regHi == any_reg) {
        // do not split interval if only one register can be assigned until the split pos
        // (when one register is found for the whole interval, split&spill is only
        // performed for the hi register)
        return false;

      } else if (regHi != any_reg) {
        split_pos = MIN2(split_pos, _use_pos[regHi]);

        // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax
        if (reg > regHi) {
          int temp = reg;
          reg = regHi;
          regHi = temp;
        }
      }
    }
  }

  cur->assign_reg(reg, regHi);
  TRACE_LINEAR_SCAN(2, tty->print_cr("selected register %d, %d", reg, regHi));

  assert(split_pos > 0, "invalid split_pos");
  if (need_split) {
    // register not available for full interval, so split it
    split_when_partial_register_available(cur, split_pos);
  }

  // only return true if interval is completely assigned
  return _num_phys_regs == 1 || regHi != any_reg;
}


int LinearScanWalker::find_locked_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) {
  int max_reg = any_reg;

  for (int i = _first_reg; i <= _last_reg; i++) {
    if (i == ignore_reg) {
      // this register must be ignored

    } else if (_use_pos[i] > reg_needed_until) {
      if (max_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_reg] && max_reg != hint_reg)) {
        max_reg = i;
      }
    }
  }

  if (max_reg != any_reg && _block_pos[max_reg] <= interval_to) {
    *need_split = true;
  }

  return max_reg;
}

int LinearScanWalker::find_locked_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) {
  assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm");

  int max_reg = any_reg;

  for (int i = _first_reg; i < _last_reg; i+=2) {
    if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) {
      if (max_reg == any_reg || _use_pos[i] > _use_pos[max_reg]) {
        max_reg = i;
      }
    }
  }

  if (_block_pos[max_reg] <= interval_to || _block_pos[max_reg + 1] <= interval_to) {
    *need_split = true;
  }

  return max_reg;
}

void LinearScanWalker::split_and_spill_intersecting_intervals(int reg, int regHi) {
  assert(reg != any_reg, "no register assigned");

  for (int i = 0; i < _spill_intervals[reg]->length(); i++) {
    Interval* it = _spill_intervals[reg]->at(i);
    remove_from_list(it);
    split_and_spill_interval(it);
  }

  if (regHi != any_reg) {
    IntervalList* processed = _spill_intervals[reg];
    for (int i = 0; i < _spill_intervals[regHi]->length(); i++) {
      Interval* it = _spill_intervals[regHi]->at(i);
      if (processed->index_of(it) == -1) {
        remove_from_list(it);
        split_and_spill_interval(it);
      }
    }
  }
}


// Split an Interval and spill it to memory so that cur can be placed in a register
void LinearScanWalker::alloc_locked_reg(Interval* cur) {
  TRACE_LINEAR_SCAN(2, tty->print("need to split and spill to get register for "); cur->print());

  // collect current usage of registers
  init_use_lists(false);
  spill_exclude_active_fixed();
//  spill_block_unhandled_fixed(cur);
  assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0");
  spill_block_inactive_fixed(cur);
  spill_collect_active_any();
  spill_collect_inactive_any(cur);

#ifndef PRODUCT
  if (TraceLinearScanLevel >= 4) {
    tty->print_cr("      state of registers:");
    for (int i = _first_reg; i <= _last_reg; i++) {
      tty->print("      reg %d: use_pos: %d, block_pos: %d, intervals: ", i, _use_pos[i], _block_pos[i]);
      for (int j = 0; j < _spill_intervals[i]->length(); j++) {
        tty->print("%d ", _spill_intervals[i]->at(j)->reg_num());
      }
      tty->cr();
    }
  }
#endif

  // the register must be free at least until this position
  int reg_needed_until = MIN2(cur->first_usage(mustHaveRegister), cur->from() + 1);
  int interval_to = cur->to();
  assert (reg_needed_until > 0 && reg_needed_until < max_jint, "interval has no use");

  int split_pos = 0;
  int use_pos = 0;
  bool need_split = false;
  int reg, regHi;

  if (_adjacent_regs) {
    reg = find_locked_double_reg(reg_needed_until, interval_to, any_reg, &need_split);
    regHi = reg + 1;

    if (reg != any_reg) {
      use_pos = MIN2(_use_pos[reg], _use_pos[regHi]);
      split_pos = MIN2(_block_pos[reg], _block_pos[regHi]);
    }
  } else {
    reg = find_locked_reg(reg_needed_until, interval_to, any_reg, cur->assigned_reg(), &need_split);
    regHi = any_reg;

    if (reg != any_reg) {
      use_pos = _use_pos[reg];
      split_pos = _block_pos[reg];

      if (_num_phys_regs == 2) {
        if (cur->assigned_reg() != any_reg) {
          regHi = reg;
          reg = cur->assigned_reg();
        } else {
          regHi = find_locked_reg(reg_needed_until, interval_to, any_reg, reg, &need_split);
          if (regHi != any_reg) {
            use_pos = MIN2(use_pos, _use_pos[regHi]);
            split_pos = MIN2(split_pos, _block_pos[regHi]);
          }
        }

        if (regHi != any_reg && reg > regHi) {
          // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax
          int temp = reg;
          reg = regHi;
          regHi = temp;
        }
      }
    }
  }

  if (reg == any_reg || (_num_phys_regs == 2 && regHi == any_reg) || use_pos <= cur->first_usage(mustHaveRegister)) {
    // the first use of cur is later than the spilling position -> spill cur
    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));

    if (cur->first_usage(mustHaveRegister) <= cur->from() + 1) {
      assert(false, "cannot spill interval that is used in first instruction (possible reason: no register found)");
      // assign a reasonable register and do a bailout in product mode to avoid errors
      allocator()->assign_spill_slot(cur);
      BAILOUT("LinearScan: no register found");
    }

    split_and_spill_interval(cur);
  } else {
    TRACE_LINEAR_SCAN(4, tty->print_cr("decided to use register %d, %d", reg, regHi));
    assert(reg != any_reg && (_num_phys_regs == 1 || regHi != any_reg), "no register found");
    assert(split_pos > 0, "invalid split_pos");
    assert(need_split == false || split_pos > cur->from(), "splitting interval at from");

    cur->assign_reg(reg, regHi);
    if (need_split) {
      // register not available for full interval, so split it
      split_when_partial_register_available(cur, split_pos);
    }

    // perform splitting and spilling for all affected intervalls
    split_and_spill_intersecting_intervals(reg, regHi);
  }
}

bool LinearScanWalker::no_allocation_possible(Interval* cur) {
5615
#ifdef X86
D
duke 已提交
5616 5617 5618 5619 5620 5621 5622 5623 5624 5625 5626 5627 5628 5629 5630 5631 5632 5633 5634 5635 5636 5637 5638 5639 5640 5641 5642 5643 5644 5645 5646 5647 5648 5649
  // fast calculation of intervals that can never get a register because the
  // the next instruction is a call that blocks all registers
  // Note: this does not work if callee-saved registers are available (e.g. on Sparc)

  // check if this interval is the result of a split operation
  // (an interval got a register until this position)
  int pos = cur->from();
  if ((pos & 1) == 1) {
    // the current instruction is a call that blocks all registers
    if (pos < allocator()->max_lir_op_id() && allocator()->has_call(pos + 1)) {
      TRACE_LINEAR_SCAN(4, tty->print_cr("      free register cannot be available because all registers blocked by following call"));

      // safety check that there is really no register available
      assert(alloc_free_reg(cur) == false, "found a register for this interval");
      return true;
    }

  }
#endif
  return false;
}

void LinearScanWalker::init_vars_for_alloc(Interval* cur) {
  BasicType type = cur->type();
  _num_phys_regs = LinearScan::num_physical_regs(type);
  _adjacent_regs = LinearScan::requires_adjacent_regs(type);

  if (pd_init_regs_for_alloc(cur)) {
    // the appropriate register range was selected.
  } else if (type == T_FLOAT || type == T_DOUBLE) {
    _first_reg = pd_first_fpu_reg;
    _last_reg = pd_last_fpu_reg;
  } else {
    _first_reg = pd_first_cpu_reg;
5650
    _last_reg = FrameMap::last_cpu_reg();
D
duke 已提交
5651 5652 5653 5654 5655 5656 5657 5658 5659 5660 5661 5662 5663 5664 5665 5666 5667 5668 5669 5670 5671 5672 5673 5674 5675 5676 5677 5678 5679 5680 5681 5682 5683 5684 5685 5686 5687 5688 5689 5690 5691 5692 5693 5694 5695 5696 5697 5698 5699 5700 5701 5702 5703 5704 5705 5706 5707 5708 5709 5710 5711 5712 5713 5714 5715 5716 5717 5718 5719 5720 5721 5722 5723 5724 5725 5726 5727 5728 5729 5730 5731 5732 5733 5734 5735 5736 5737 5738 5739 5740 5741 5742 5743 5744 5745 5746 5747 5748 5749 5750 5751 5752 5753 5754 5755 5756 5757 5758 5759 5760 5761 5762 5763 5764 5765 5766 5767 5768 5769 5770 5771 5772 5773 5774 5775 5776 5777 5778 5779 5780 5781 5782 5783 5784 5785 5786 5787 5788 5789 5790 5791 5792 5793 5794 5795 5796 5797 5798 5799 5800 5801 5802 5803 5804 5805 5806 5807 5808 5809 5810 5811 5812 5813 5814 5815 5816 5817 5818 5819 5820 5821 5822 5823 5824 5825 5826 5827 5828 5829 5830 5831 5832 5833 5834 5835 5836 5837 5838 5839 5840 5841 5842 5843 5844 5845 5846 5847 5848 5849 5850 5851 5852 5853 5854 5855 5856 5857 5858 5859 5860 5861 5862 5863 5864 5865 5866 5867 5868 5869 5870 5871 5872 5873 5874 5875 5876 5877 5878 5879 5880 5881 5882 5883 5884 5885 5886 5887 5888 5889 5890 5891 5892 5893 5894 5895 5896 5897 5898 5899 5900 5901 5902 5903 5904 5905 5906 5907 5908 5909 5910 5911 5912 5913 5914 5915 5916 5917 5918 5919 5920 5921 5922 5923 5924 5925 5926 5927 5928 5929 5930 5931 5932 5933 5934 5935 5936 5937 5938 5939 5940 5941 5942 5943 5944 5945 5946 5947 5948 5949 5950 5951 5952 5953 5954 5955 5956 5957 5958 5959 5960 5961 5962 5963 5964 5965 5966 5967 5968 5969 5970 5971 5972 5973 5974 5975 5976 5977 5978 5979 5980 5981 5982 5983 5984 5985 5986 5987 5988 5989 5990 5991 5992 5993 5994 5995 5996 5997 5998 5999 6000 6001 6002 6003 6004 6005 6006 6007 6008 6009 6010 6011 6012 6013 6014 6015 6016 6017 6018 6019 6020 6021 6022 6023 6024 6025 6026 6027 6028 6029 6030 6031 6032 6033 6034 6035 6036 6037 6038 6039 6040 6041 6042 6043 6044 6045 6046 6047 6048 6049 6050 6051 6052 6053 6054 6055 6056 6057 6058 6059 6060 6061 6062 6063 6064 6065 6066 6067 6068 6069 6070 6071 6072 6073 6074 6075 6076 6077 6078 6079 6080 6081 6082 6083 6084 6085 6086 6087 6088 6089 6090 6091 6092 6093 6094 6095 6096 6097 6098 6099 6100 6101 6102 6103 6104 6105 6106 6107 6108 6109 6110 6111 6112 6113 6114 6115 6116 6117 6118 6119 6120 6121 6122 6123 6124 6125 6126 6127 6128 6129 6130 6131 6132 6133 6134 6135 6136 6137 6138 6139 6140 6141 6142 6143 6144 6145 6146 6147 6148 6149 6150 6151 6152 6153 6154 6155 6156 6157 6158 6159 6160 6161 6162 6163 6164 6165 6166 6167 6168 6169 6170 6171 6172 6173 6174 6175 6176 6177 6178 6179 6180 6181 6182 6183 6184 6185 6186 6187 6188 6189 6190 6191 6192 6193 6194 6195 6196 6197 6198 6199 6200 6201 6202 6203 6204 6205 6206 6207 6208 6209 6210 6211 6212 6213 6214 6215 6216 6217 6218 6219 6220 6221 6222 6223 6224 6225
  }

  assert(0 <= _first_reg && _first_reg < LinearScan::nof_regs, "out of range");
  assert(0 <= _last_reg && _last_reg < LinearScan::nof_regs, "out of range");
}


bool LinearScanWalker::is_move(LIR_Op* op, Interval* from, Interval* to) {
  if (op->code() != lir_move) {
    return false;
  }
  assert(op->as_Op1() != NULL, "move must be LIR_Op1");

  LIR_Opr in = ((LIR_Op1*)op)->in_opr();
  LIR_Opr res = ((LIR_Op1*)op)->result_opr();
  return in->is_virtual() && res->is_virtual() && in->vreg_number() == from->reg_num() && res->vreg_number() == to->reg_num();
}

// optimization (especially for phi functions of nested loops):
// assign same spill slot to non-intersecting intervals
void LinearScanWalker::combine_spilled_intervals(Interval* cur) {
  if (cur->is_split_child()) {
    // optimization is only suitable for split parents
    return;
  }

  Interval* register_hint = cur->register_hint(false);
  if (register_hint == NULL) {
    // cur is not the target of a move, otherwise register_hint would be set
    return;
  }
  assert(register_hint->is_split_parent(), "register hint must be split parent");

  if (cur->spill_state() != noOptimization || register_hint->spill_state() != noOptimization) {
    // combining the stack slots for intervals where spill move optimization is applied
    // is not benefitial and would cause problems
    return;
  }

  int begin_pos = cur->from();
  int end_pos = cur->to();
  if (end_pos > allocator()->max_lir_op_id() || (begin_pos & 1) != 0 || (end_pos & 1) != 0) {
    // safety check that lir_op_with_id is allowed
    return;
  }

  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)) {
    // cur and register_hint are not connected with two moves
    return;
  }

  Interval* begin_hint = register_hint->split_child_at_op_id(begin_pos, LIR_OpVisitState::inputMode);
  Interval* end_hint = register_hint->split_child_at_op_id(end_pos, LIR_OpVisitState::outputMode);
  if (begin_hint == end_hint || begin_hint->to() != begin_pos || end_hint->from() != end_pos) {
    // register_hint must be split, otherwise the re-writing of use positions does not work
    return;
  }

  assert(begin_hint->assigned_reg() != any_reg, "must have register assigned");
  assert(end_hint->assigned_reg() == any_reg, "must not have register assigned");
  assert(cur->first_usage(mustHaveRegister) == begin_pos, "must have use position at begin of interval because of move");
  assert(end_hint->first_usage(mustHaveRegister) == end_pos, "must have use position at begin of interval because of move");

  if (begin_hint->assigned_reg() < LinearScan::nof_regs) {
    // register_hint is not spilled at begin_pos, so it would not be benefitial to immediately spill cur
    return;
  }
  assert(register_hint->canonical_spill_slot() != -1, "must be set when part of interval was spilled");

  // modify intervals such that cur gets the same stack slot as register_hint
  // delete use positions to prevent the intervals to get a register at beginning
  cur->set_canonical_spill_slot(register_hint->canonical_spill_slot());
  cur->remove_first_use_pos();
  end_hint->remove_first_use_pos();
}


// allocate a physical register or memory location to an interval
bool LinearScanWalker::activate_current() {
  Interval* cur = current();
  bool result = true;

  TRACE_LINEAR_SCAN(2, tty->print   ("+++++ activating interval "); cur->print());
  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()));

  if (cur->assigned_reg() >= LinearScan::nof_regs) {
    // activating an interval that has a stack slot assigned -> split it at first use position
    // used for method parameters
    TRACE_LINEAR_SCAN(4, tty->print_cr("      interval has spill slot assigned (method parameter) -> split it before first use"));

    split_stack_interval(cur);
    result = false;

  } else if (allocator()->gen()->is_vreg_flag_set(cur->reg_num(), LIRGenerator::must_start_in_memory)) {
    // activating an interval that must start in a stack slot, but may get a register later
    // used for lir_roundfp: rounding is done by store to stack and reload later
    TRACE_LINEAR_SCAN(4, tty->print_cr("      interval must start in stack slot -> split it before first use"));
    assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned");

    allocator()->assign_spill_slot(cur);
    split_stack_interval(cur);
    result = false;

  } else if (cur->assigned_reg() == any_reg) {
    // interval has not assigned register -> normal allocation
    // (this is the normal case for most intervals)
    TRACE_LINEAR_SCAN(4, tty->print_cr("      normal allocation of register"));

    // assign same spill slot to non-intersecting intervals
    combine_spilled_intervals(cur);

    init_vars_for_alloc(cur);
    if (no_allocation_possible(cur) || !alloc_free_reg(cur)) {
      // no empty register available.
      // split and spill another interval so that this interval gets a register
      alloc_locked_reg(cur);
    }

    // spilled intervals need not be move to active-list
    if (cur->assigned_reg() >= LinearScan::nof_regs) {
      result = false;
    }
  }

  // load spilled values that become active from stack slot to register
  if (cur->insert_move_when_activated()) {
    assert(cur->is_split_child(), "must be");
    assert(cur->current_split_child() != NULL, "must be");
    assert(cur->current_split_child()->reg_num() != cur->reg_num(), "cannot insert move between same interval");
    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()));

    insert_move(cur->from(), cur->current_split_child(), cur);
  }
  cur->make_current_split_child();

  return result; // true = interval is moved to active list
}


// Implementation of EdgeMoveOptimizer

EdgeMoveOptimizer::EdgeMoveOptimizer() :
  _edge_instructions(4),
  _edge_instructions_idx(4)
{
}

void EdgeMoveOptimizer::optimize(BlockList* code) {
  EdgeMoveOptimizer optimizer = EdgeMoveOptimizer();

  // ignore the first block in the list (index 0 is not processed)
  for (int i = code->length() - 1; i >= 1; i--) {
    BlockBegin* block = code->at(i);

    if (block->number_of_preds() > 1 && !block->is_set(BlockBegin::exception_entry_flag)) {
      optimizer.optimize_moves_at_block_end(block);
    }
    if (block->number_of_sux() == 2) {
      optimizer.optimize_moves_at_block_begin(block);
    }
  }
}


// clear all internal data structures
void EdgeMoveOptimizer::init_instructions() {
  _edge_instructions.clear();
  _edge_instructions_idx.clear();
}

// append a lir-instruction-list and the index of the current operation in to the list
void EdgeMoveOptimizer::append_instructions(LIR_OpList* instructions, int instructions_idx) {
  _edge_instructions.append(instructions);
  _edge_instructions_idx.append(instructions_idx);
}

// return the current operation of the given edge (predecessor or successor)
LIR_Op* EdgeMoveOptimizer::instruction_at(int edge) {
  LIR_OpList* instructions = _edge_instructions.at(edge);
  int idx = _edge_instructions_idx.at(edge);

  if (idx < instructions->length()) {
    return instructions->at(idx);
  } else {
    return NULL;
  }
}

// removes the current operation of the given edge (predecessor or successor)
void EdgeMoveOptimizer::remove_cur_instruction(int edge, bool decrement_index) {
  LIR_OpList* instructions = _edge_instructions.at(edge);
  int idx = _edge_instructions_idx.at(edge);
  instructions->remove_at(idx);

  if (decrement_index) {
    _edge_instructions_idx.at_put(edge, idx - 1);
  }
}


bool EdgeMoveOptimizer::operations_different(LIR_Op* op1, LIR_Op* op2) {
  if (op1 == NULL || op2 == NULL) {
    // at least one block is already empty -> no optimization possible
    return true;
  }

  if (op1->code() == lir_move && op2->code() == lir_move) {
    assert(op1->as_Op1() != NULL, "move must be LIR_Op1");
    assert(op2->as_Op1() != NULL, "move must be LIR_Op1");
    LIR_Op1* move1 = (LIR_Op1*)op1;
    LIR_Op1* move2 = (LIR_Op1*)op2;
    if (move1->info() == move2->info() && move1->in_opr() == move2->in_opr() && move1->result_opr() == move2->result_opr()) {
      // these moves are exactly equal and can be optimized
      return false;
    }

  } else if (op1->code() == lir_fxch && op2->code() == lir_fxch) {
    assert(op1->as_Op1() != NULL, "fxch must be LIR_Op1");
    assert(op2->as_Op1() != NULL, "fxch must be LIR_Op1");
    LIR_Op1* fxch1 = (LIR_Op1*)op1;
    LIR_Op1* fxch2 = (LIR_Op1*)op2;
    if (fxch1->in_opr()->as_jint() == fxch2->in_opr()->as_jint()) {
      // equal FPU stack operations can be optimized
      return false;
    }

  } else if (op1->code() == lir_fpop_raw && op2->code() == lir_fpop_raw) {
    // equal FPU stack operations can be optimized
    return false;
  }

  // no optimization possible
  return true;
}

void EdgeMoveOptimizer::optimize_moves_at_block_end(BlockBegin* block) {
  TRACE_LINEAR_SCAN(4, tty->print_cr("optimizing moves at end of block B%d", block->block_id()));

  if (block->is_predecessor(block)) {
    // currently we can't handle this correctly.
    return;
  }

  init_instructions();
  int num_preds = block->number_of_preds();
  assert(num_preds > 1, "do not call otherwise");
  assert(!block->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed");

  // setup a list with the lir-instructions of all predecessors
  int i;
  for (i = 0; i < num_preds; i++) {
    BlockBegin* pred = block->pred_at(i);
    LIR_OpList* pred_instructions = pred->lir()->instructions_list();

    if (pred->number_of_sux() != 1) {
      // this can happen with switch-statements where multiple edges are between
      // the same blocks.
      return;
    }

    assert(pred->number_of_sux() == 1, "can handle only one successor");
    assert(pred->sux_at(0) == block, "invalid control flow");
    assert(pred_instructions->last()->code() == lir_branch, "block with successor must end with branch");
    assert(pred_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch");
    assert(pred_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch");

    if (pred_instructions->last()->info() != NULL) {
      // can not optimize instructions when debug info is needed
      return;
    }

    // ignore the unconditional branch at the end of the block
    append_instructions(pred_instructions, pred_instructions->length() - 2);
  }


  // process lir-instructions while all predecessors end with the same instruction
  while (true) {
    LIR_Op* op = instruction_at(0);
    for (i = 1; i < num_preds; i++) {
      if (operations_different(op, instruction_at(i))) {
        // these instructions are different and cannot be optimized ->
        // no further optimization possible
        return;
      }
    }

    TRACE_LINEAR_SCAN(4, tty->print("found instruction that is equal in all %d predecessors: ", num_preds); op->print());

    // insert the instruction at the beginning of the current block
    block->lir()->insert_before(1, op);

    // delete the instruction at the end of all predecessors
    for (i = 0; i < num_preds; i++) {
      remove_cur_instruction(i, true);
    }
  }
}


void EdgeMoveOptimizer::optimize_moves_at_block_begin(BlockBegin* block) {
  TRACE_LINEAR_SCAN(4, tty->print_cr("optimization moves at begin of block B%d", block->block_id()));

  init_instructions();
  int num_sux = block->number_of_sux();

  LIR_OpList* cur_instructions = block->lir()->instructions_list();

  assert(num_sux == 2, "method should not be called otherwise");
  assert(cur_instructions->last()->code() == lir_branch, "block with successor must end with branch");
  assert(cur_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch");
  assert(cur_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch");

  if (cur_instructions->last()->info() != NULL) {
    // can no optimize instructions when debug info is needed
    return;
  }

  LIR_Op* branch = cur_instructions->at(cur_instructions->length() - 2);
  if (branch->info() != NULL || (branch->code() != lir_branch && branch->code() != lir_cond_float_branch)) {
    // not a valid case for optimization
    // currently, only blocks that end with two branches (conditional branch followed
    // by unconditional branch) are optimized
    return;
  }

  // now it is guaranteed that the block ends with two branch instructions.
  // the instructions are inserted at the end of the block before these two branches
  int insert_idx = cur_instructions->length() - 2;

  int i;
#ifdef ASSERT
  for (i = insert_idx - 1; i >= 0; i--) {
    LIR_Op* op = cur_instructions->at(i);
    if ((op->code() == lir_branch || op->code() == lir_cond_float_branch) && ((LIR_OpBranch*)op)->block() != NULL) {
      assert(false, "block with two successors can have only two branch instructions");
    }
  }
#endif

  // setup a list with the lir-instructions of all successors
  for (i = 0; i < num_sux; i++) {
    BlockBegin* sux = block->sux_at(i);
    LIR_OpList* sux_instructions = sux->lir()->instructions_list();

    assert(sux_instructions->at(0)->code() == lir_label, "block must start with label");

    if (sux->number_of_preds() != 1) {
      // this can happen with switch-statements where multiple edges are between
      // the same blocks.
      return;
    }
    assert(sux->pred_at(0) == block, "invalid control flow");
    assert(!sux->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed");

    // ignore the label at the beginning of the block
    append_instructions(sux_instructions, 1);
  }

  // process lir-instructions while all successors begin with the same instruction
  while (true) {
    LIR_Op* op = instruction_at(0);
    for (i = 1; i < num_sux; i++) {
      if (operations_different(op, instruction_at(i))) {
        // these instructions are different and cannot be optimized ->
        // no further optimization possible
        return;
      }
    }

    TRACE_LINEAR_SCAN(4, tty->print("----- found instruction that is equal in all %d successors: ", num_sux); op->print());

    // insert instruction at end of current block
    block->lir()->insert_before(insert_idx, op);
    insert_idx++;

    // delete the instructions at the beginning of all successors
    for (i = 0; i < num_sux; i++) {
      remove_cur_instruction(i, false);
    }
  }
}


// Implementation of ControlFlowOptimizer

ControlFlowOptimizer::ControlFlowOptimizer() :
  _original_preds(4)
{
}

void ControlFlowOptimizer::optimize(BlockList* code) {
  ControlFlowOptimizer optimizer = ControlFlowOptimizer();

  // push the OSR entry block to the end so that we're not jumping over it.
  BlockBegin* osr_entry = code->at(0)->end()->as_Base()->osr_entry();
  if (osr_entry) {
    int index = osr_entry->linear_scan_number();
    assert(code->at(index) == osr_entry, "wrong index");
    code->remove_at(index);
    code->append(osr_entry);
  }

  optimizer.reorder_short_loops(code);
  optimizer.delete_empty_blocks(code);
  optimizer.delete_unnecessary_jumps(code);
  optimizer.delete_jumps_to_return(code);
}

void ControlFlowOptimizer::reorder_short_loop(BlockList* code, BlockBegin* header_block, int header_idx) {
  int i = header_idx + 1;
  int max_end = MIN2(header_idx + ShortLoopSize, code->length());
  while (i < max_end && code->at(i)->loop_depth() >= header_block->loop_depth()) {
    i++;
  }

  if (i == code->length() || code->at(i)->loop_depth() < header_block->loop_depth()) {
    int end_idx = i - 1;
    BlockBegin* end_block = code->at(end_idx);

    if (end_block->number_of_sux() == 1 && end_block->sux_at(0) == header_block) {
      // short loop from header_idx to end_idx found -> reorder blocks such that
      // the header_block is the last block instead of the first block of the loop
      TRACE_LINEAR_SCAN(1, tty->print_cr("Reordering short loop: length %d, header B%d, end B%d",
                                         end_idx - header_idx + 1,
                                         header_block->block_id(), end_block->block_id()));

      for (int j = header_idx; j < end_idx; j++) {
        code->at_put(j, code->at(j + 1));
      }
      code->at_put(end_idx, header_block);

      // correct the flags so that any loop alignment occurs in the right place.
      assert(code->at(end_idx)->is_set(BlockBegin::backward_branch_target_flag), "must be backward branch target");
      code->at(end_idx)->clear(BlockBegin::backward_branch_target_flag);
      code->at(header_idx)->set(BlockBegin::backward_branch_target_flag);
    }
  }
}

void ControlFlowOptimizer::reorder_short_loops(BlockList* code) {
  for (int i = code->length() - 1; i >= 0; i--) {
    BlockBegin* block = code->at(i);

    if (block->is_set(BlockBegin::linear_scan_loop_header_flag)) {
      reorder_short_loop(code, block, i);
    }
  }

  DEBUG_ONLY(verify(code));
}

// only blocks with exactly one successor can be deleted. Such blocks
// must always end with an unconditional branch to this successor
bool ControlFlowOptimizer::can_delete_block(BlockBegin* block) {
  if (block->number_of_sux() != 1 || block->number_of_exception_handlers() != 0 || block->is_entry_block()) {
    return false;
  }

  LIR_OpList* instructions = block->lir()->instructions_list();

  assert(instructions->length() >= 2, "block must have label and branch");
  assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label");
  assert(instructions->last()->as_OpBranch() != NULL, "last instrcution must always be a branch");
  assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "branch must be unconditional");
  assert(instructions->last()->as_OpBranch()->block() == block->sux_at(0), "branch target must be the successor");

  // block must have exactly one successor

  if (instructions->length() == 2 && instructions->last()->info() == NULL) {
    return true;
  }
  return false;
}

// substitute branch targets in all branch-instructions of this blocks
void ControlFlowOptimizer::substitute_branch_target(BlockBegin* block, BlockBegin* target_from, BlockBegin* target_to) {
  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()));

  LIR_OpList* instructions = block->lir()->instructions_list();

  assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label");
  for (int i = instructions->length() - 1; i >= 1; i--) {
    LIR_Op* op = instructions->at(i);

    if (op->code() == lir_branch || op->code() == lir_cond_float_branch) {
      assert(op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch");
      LIR_OpBranch* branch = (LIR_OpBranch*)op;

      if (branch->block() == target_from) {
        branch->change_block(target_to);
      }
      if (branch->ublock() == target_from) {
        branch->change_ublock(target_to);
      }
    }
  }
}

void ControlFlowOptimizer::delete_empty_blocks(BlockList* code) {
  int old_pos = 0;
  int new_pos = 0;
  int num_blocks = code->length();

  while (old_pos < num_blocks) {
    BlockBegin* block = code->at(old_pos);

    if (can_delete_block(block)) {
      BlockBegin* new_target = block->sux_at(0);

      // propagate backward branch target flag for correct code alignment
      if (block->is_set(BlockBegin::backward_branch_target_flag)) {
        new_target->set(BlockBegin::backward_branch_target_flag);
      }

      // collect a list with all predecessors that contains each predecessor only once
      // the predecessors of cur are changed during the substitution, so a copy of the
      // predecessor list is necessary
      int j;
      _original_preds.clear();
      for (j = block->number_of_preds() - 1; j >= 0; j--) {
        BlockBegin* pred = block->pred_at(j);
        if (_original_preds.index_of(pred) == -1) {
          _original_preds.append(pred);
        }
      }

      for (j = _original_preds.length() - 1; j >= 0; j--) {
        BlockBegin* pred = _original_preds.at(j);
        substitute_branch_target(pred, block, new_target);
        pred->substitute_sux(block, new_target);
      }
    } else {
      // adjust position of this block in the block list if blocks before
      // have been deleted
      if (new_pos != old_pos) {
        code->at_put(new_pos, code->at(old_pos));
      }
      new_pos++;
    }
    old_pos++;
  }
  code->truncate(new_pos);

  DEBUG_ONLY(verify(code));
}

void ControlFlowOptimizer::delete_unnecessary_jumps(BlockList* code) {
  // skip the last block because there a branch is always necessary
  for (int i = code->length() - 2; i >= 0; i--) {
    BlockBegin* block = code->at(i);
    LIR_OpList* instructions = block->lir()->instructions_list();

    LIR_Op* last_op = instructions->last();
    if (last_op->code() == lir_branch) {
      assert(last_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch");
      LIR_OpBranch* last_branch = (LIR_OpBranch*)last_op;

      assert(last_branch->block() != NULL, "last branch must always have a block as target");
      assert(last_branch->label() == last_branch->block()->label(), "must be equal");

      if (last_branch->info() == NULL) {
        if (last_branch->block() == code->at(i + 1)) {

          TRACE_LINEAR_SCAN(3, tty->print_cr("Deleting unconditional branch at end of block B%d", block->block_id()));

          // delete last branch instruction
          instructions->truncate(instructions->length() - 1);

        } else {
          LIR_Op* prev_op = instructions->at(instructions->length() - 2);
          if (prev_op->code() == lir_branch || prev_op->code() == lir_cond_float_branch) {
            assert(prev_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch");
            LIR_OpBranch* prev_branch = (LIR_OpBranch*)prev_op;

6226 6227 6228 6229 6230 6231 6232 6233 6234 6235 6236
            if (prev_branch->stub() == NULL) {

              LIR_Op2* prev_cmp = NULL;

              for(int j = instructions->length() - 3; j >= 0 && prev_cmp == NULL; j--) {
                prev_op = instructions->at(j);
                if (prev_op->code() == lir_cmp) {
                  assert(prev_op->as_Op2() != NULL, "branch must be of type LIR_Op2");
                  prev_cmp = (LIR_Op2*)prev_op;
                  assert(prev_branch->cond() == prev_cmp->condition(), "should be the same");
                }
6237
              }
6238 6239
              assert(prev_cmp != NULL, "should have found comp instruction for branch");
              if (prev_branch->block() == code->at(i + 1) && prev_branch->info() == NULL) {
D
duke 已提交
6240

6241
                TRACE_LINEAR_SCAN(3, tty->print_cr("Negating conditional branch and deleting unconditional branch at end of block B%d", block->block_id()));
D
duke 已提交
6242

6243 6244 6245 6246 6247 6248
                // eliminate a conditional branch to the immediate successor
                prev_branch->change_block(last_branch->block());
                prev_branch->negate_cond();
                prev_cmp->set_condition(prev_branch->cond());
                instructions->truncate(instructions->length() - 1);
              }
D
duke 已提交
6249 6250 6251 6252 6253 6254 6255 6256 6257 6258 6259 6260 6261 6262 6263 6264 6265 6266 6267 6268 6269 6270 6271 6272 6273 6274 6275 6276 6277 6278 6279 6280 6281 6282 6283 6284 6285 6286 6287 6288 6289 6290 6291 6292 6293 6294 6295 6296 6297 6298 6299 6300 6301 6302 6303 6304 6305 6306 6307 6308 6309 6310 6311 6312 6313 6314 6315 6316 6317 6318 6319 6320 6321 6322 6323 6324 6325 6326 6327 6328 6329 6330 6331 6332 6333 6334 6335 6336 6337 6338 6339 6340 6341 6342 6343 6344 6345 6346 6347 6348 6349 6350 6351 6352 6353 6354 6355 6356 6357 6358 6359 6360 6361 6362 6363 6364 6365 6366 6367 6368 6369 6370 6371 6372 6373 6374 6375 6376 6377 6378 6379 6380 6381 6382 6383 6384 6385 6386 6387 6388 6389 6390 6391 6392 6393 6394 6395 6396 6397 6398 6399 6400 6401 6402 6403 6404 6405 6406 6407 6408 6409 6410 6411 6412 6413 6414 6415 6416 6417 6418 6419 6420 6421 6422 6423 6424 6425 6426 6427 6428 6429 6430 6431 6432 6433 6434 6435 6436 6437 6438 6439 6440 6441 6442 6443 6444 6445 6446 6447 6448 6449 6450 6451 6452 6453 6454 6455 6456 6457 6458 6459 6460 6461 6462 6463 6464 6465 6466 6467 6468 6469 6470 6471 6472 6473 6474 6475 6476 6477 6478 6479 6480 6481 6482 6483 6484 6485 6486 6487 6488 6489 6490 6491 6492 6493 6494 6495 6496 6497 6498 6499 6500 6501 6502 6503 6504 6505 6506 6507 6508 6509 6510 6511 6512 6513 6514 6515 6516 6517 6518 6519 6520 6521 6522 6523 6524 6525 6526 6527 6528 6529 6530 6531 6532 6533 6534 6535 6536 6537 6538 6539 6540 6541 6542 6543 6544 6545 6546 6547 6548 6549 6550 6551 6552 6553 6554 6555 6556 6557 6558 6559 6560 6561 6562 6563 6564 6565 6566 6567 6568 6569 6570 6571 6572 6573 6574 6575 6576 6577 6578 6579 6580 6581 6582
            }
          }
        }
      }
    }
  }

  DEBUG_ONLY(verify(code));
}

void ControlFlowOptimizer::delete_jumps_to_return(BlockList* code) {
#ifdef ASSERT
  BitMap return_converted(BlockBegin::number_of_blocks());
  return_converted.clear();
#endif

  for (int i = code->length() - 1; i >= 0; i--) {
    BlockBegin* block = code->at(i);
    LIR_OpList* cur_instructions = block->lir()->instructions_list();
    LIR_Op*     cur_last_op = cur_instructions->last();

    assert(cur_instructions->at(0)->code() == lir_label, "first instruction must always be a label");
    if (cur_instructions->length() == 2 && cur_last_op->code() == lir_return) {
      // the block contains only a label and a return
      // if a predecessor ends with an unconditional jump to this block, then the jump
      // can be replaced with a return instruction
      //
      // Note: the original block with only a return statement cannot be deleted completely
      //       because the predecessors might have other (conditional) jumps to this block
      //       -> this may lead to unnecesary return instructions in the final code

      assert(cur_last_op->info() == NULL, "return instructions do not have debug information");
      assert(block->number_of_sux() == 0 ||
             (return_converted.at(block->block_id()) && block->number_of_sux() == 1),
             "blocks that end with return must not have successors");

      assert(cur_last_op->as_Op1() != NULL, "return must be LIR_Op1");
      LIR_Opr return_opr = ((LIR_Op1*)cur_last_op)->in_opr();

      for (int j = block->number_of_preds() - 1; j >= 0; j--) {
        BlockBegin* pred = block->pred_at(j);
        LIR_OpList* pred_instructions = pred->lir()->instructions_list();
        LIR_Op*     pred_last_op = pred_instructions->last();

        if (pred_last_op->code() == lir_branch) {
          assert(pred_last_op->as_OpBranch() != NULL, "branch must be LIR_OpBranch");
          LIR_OpBranch* pred_last_branch = (LIR_OpBranch*)pred_last_op;

          if (pred_last_branch->block() == block && pred_last_branch->cond() == lir_cond_always && pred_last_branch->info() == NULL) {
            // replace the jump to a return with a direct return
            // Note: currently the edge between the blocks is not deleted
            pred_instructions->at_put(pred_instructions->length() - 1, new LIR_Op1(lir_return, return_opr));
#ifdef ASSERT
            return_converted.set_bit(pred->block_id());
#endif
          }
        }
      }
    }
  }
}


#ifdef ASSERT
void ControlFlowOptimizer::verify(BlockList* code) {
  for (int i = 0; i < code->length(); i++) {
    BlockBegin* block = code->at(i);
    LIR_OpList* instructions = block->lir()->instructions_list();

    int j;
    for (j = 0; j < instructions->length(); j++) {
      LIR_OpBranch* op_branch = instructions->at(j)->as_OpBranch();

      if (op_branch != NULL) {
        assert(op_branch->block() == NULL || code->index_of(op_branch->block()) != -1, "branch target not valid");
        assert(op_branch->ublock() == NULL || code->index_of(op_branch->ublock()) != -1, "branch target not valid");
      }
    }

    for (j = 0; j < block->number_of_sux() - 1; j++) {
      BlockBegin* sux = block->sux_at(j);
      assert(code->index_of(sux) != -1, "successor not valid");
    }

    for (j = 0; j < block->number_of_preds() - 1; j++) {
      BlockBegin* pred = block->pred_at(j);
      assert(code->index_of(pred) != -1, "successor not valid");
    }
  }
}
#endif


#ifndef PRODUCT

// Implementation of LinearStatistic

const char* LinearScanStatistic::counter_name(int counter_idx) {
  switch (counter_idx) {
    case counter_method:          return "compiled methods";
    case counter_fpu_method:      return "methods using fpu";
    case counter_loop_method:     return "methods with loops";
    case counter_exception_method:return "methods with xhandler";

    case counter_loop:            return "loops";
    case counter_block:           return "blocks";
    case counter_loop_block:      return "blocks inside loop";
    case counter_exception_block: return "exception handler entries";
    case counter_interval:        return "intervals";
    case counter_fixed_interval:  return "fixed intervals";
    case counter_range:           return "ranges";
    case counter_fixed_range:     return "fixed ranges";
    case counter_use_pos:         return "use positions";
    case counter_fixed_use_pos:   return "fixed use positions";
    case counter_spill_slots:     return "spill slots";

    // counter for classes of lir instructions
    case counter_instruction:     return "total instructions";
    case counter_label:           return "labels";
    case counter_entry:           return "method entries";
    case counter_return:          return "method returns";
    case counter_call:            return "method calls";
    case counter_move:            return "moves";
    case counter_cmp:             return "compare";
    case counter_cond_branch:     return "conditional branches";
    case counter_uncond_branch:   return "unconditional branches";
    case counter_stub_branch:     return "branches to stub";
    case counter_alu:             return "artithmetic + logic";
    case counter_alloc:           return "allocations";
    case counter_sync:            return "synchronisation";
    case counter_throw:           return "throw";
    case counter_unwind:          return "unwind";
    case counter_typecheck:       return "type+null-checks";
    case counter_fpu_stack:       return "fpu-stack";
    case counter_misc_inst:       return "other instructions";
    case counter_other_inst:      return "misc. instructions";

    // counter for different types of moves
    case counter_move_total:      return "total moves";
    case counter_move_reg_reg:    return "register->register";
    case counter_move_reg_stack:  return "register->stack";
    case counter_move_stack_reg:  return "stack->register";
    case counter_move_stack_stack:return "stack->stack";
    case counter_move_reg_mem:    return "register->memory";
    case counter_move_mem_reg:    return "memory->register";
    case counter_move_const_any:  return "constant->any";

    case blank_line_1:            return "";
    case blank_line_2:            return "";

    default: ShouldNotReachHere(); return "";
  }
}

LinearScanStatistic::Counter LinearScanStatistic::base_counter(int counter_idx) {
  if (counter_idx == counter_fpu_method || counter_idx == counter_loop_method || counter_idx == counter_exception_method) {
    return counter_method;
  } else if (counter_idx == counter_loop_block || counter_idx == counter_exception_block) {
    return counter_block;
  } else if (counter_idx >= counter_instruction && counter_idx <= counter_other_inst) {
    return counter_instruction;
  } else if (counter_idx >= counter_move_total && counter_idx <= counter_move_const_any) {
    return counter_move_total;
  }
  return invalid_counter;
}

LinearScanStatistic::LinearScanStatistic() {
  for (int i = 0; i < number_of_counters; i++) {
    _counters_sum[i] = 0;
    _counters_max[i] = -1;
  }

}

// add the method-local numbers to the total sum
void LinearScanStatistic::sum_up(LinearScanStatistic &method_statistic) {
  for (int i = 0; i < number_of_counters; i++) {
    _counters_sum[i] += method_statistic._counters_sum[i];
    _counters_max[i] = MAX2(_counters_max[i], method_statistic._counters_sum[i]);
  }
}

void LinearScanStatistic::print(const char* title) {
  if (CountLinearScan || TraceLinearScanLevel > 0) {
    tty->cr();
    tty->print_cr("***** LinearScan statistic - %s *****", title);

    for (int i = 0; i < number_of_counters; i++) {
      if (_counters_sum[i] > 0 || _counters_max[i] >= 0) {
        tty->print("%25s: %8d", counter_name(i), _counters_sum[i]);

        if (base_counter(i) != invalid_counter) {
          tty->print("  (%5.1f%%) ", _counters_sum[i] * 100.0 / _counters_sum[base_counter(i)]);
        } else {
          tty->print("           ");
        }

        if (_counters_max[i] >= 0) {
          tty->print("%8d", _counters_max[i]);
        }
      }
      tty->cr();
    }
  }
}

void LinearScanStatistic::collect(LinearScan* allocator) {
  inc_counter(counter_method);
  if (allocator->has_fpu_registers()) {
    inc_counter(counter_fpu_method);
  }
  if (allocator->num_loops() > 0) {
    inc_counter(counter_loop_method);
  }
  inc_counter(counter_loop, allocator->num_loops());
  inc_counter(counter_spill_slots, allocator->max_spills());

  int i;
  for (i = 0; i < allocator->interval_count(); i++) {
    Interval* cur = allocator->interval_at(i);

    if (cur != NULL) {
      inc_counter(counter_interval);
      inc_counter(counter_use_pos, cur->num_use_positions());
      if (LinearScan::is_precolored_interval(cur)) {
        inc_counter(counter_fixed_interval);
        inc_counter(counter_fixed_use_pos, cur->num_use_positions());
      }

      Range* range = cur->first();
      while (range != Range::end()) {
        inc_counter(counter_range);
        if (LinearScan::is_precolored_interval(cur)) {
          inc_counter(counter_fixed_range);
        }
        range = range->next();
      }
    }
  }

  bool has_xhandlers = false;
  // Note: only count blocks that are in code-emit order
  for (i = 0; i < allocator->ir()->code()->length(); i++) {
    BlockBegin* cur = allocator->ir()->code()->at(i);

    inc_counter(counter_block);
    if (cur->loop_depth() > 0) {
      inc_counter(counter_loop_block);
    }
    if (cur->is_set(BlockBegin::exception_entry_flag)) {
      inc_counter(counter_exception_block);
      has_xhandlers = true;
    }

    LIR_OpList* instructions = cur->lir()->instructions_list();
    for (int j = 0; j < instructions->length(); j++) {
      LIR_Op* op = instructions->at(j);

      inc_counter(counter_instruction);

      switch (op->code()) {
        case lir_label:           inc_counter(counter_label); break;
        case lir_std_entry:
        case lir_osr_entry:       inc_counter(counter_entry); break;
        case lir_return:          inc_counter(counter_return); break;

        case lir_rtcall:
        case lir_static_call:
        case lir_optvirtual_call:
        case lir_virtual_call:    inc_counter(counter_call); break;

        case lir_move: {
          inc_counter(counter_move);
          inc_counter(counter_move_total);

          LIR_Opr in = op->as_Op1()->in_opr();
          LIR_Opr res = op->as_Op1()->result_opr();
          if (in->is_register()) {
            if (res->is_register()) {
              inc_counter(counter_move_reg_reg);
            } else if (res->is_stack()) {
              inc_counter(counter_move_reg_stack);
            } else if (res->is_address()) {
              inc_counter(counter_move_reg_mem);
            } else {
              ShouldNotReachHere();
            }
          } else if (in->is_stack()) {
            if (res->is_register()) {
              inc_counter(counter_move_stack_reg);
            } else {
              inc_counter(counter_move_stack_stack);
            }
          } else if (in->is_address()) {
            assert(res->is_register(), "must be");
            inc_counter(counter_move_mem_reg);
          } else if (in->is_constant()) {
            inc_counter(counter_move_const_any);
          } else {
            ShouldNotReachHere();
          }
          break;
        }

        case lir_cmp:             inc_counter(counter_cmp); break;

        case lir_branch:
        case lir_cond_float_branch: {
          LIR_OpBranch* branch = op->as_OpBranch();
          if (branch->block() == NULL) {
            inc_counter(counter_stub_branch);
          } else if (branch->cond() == lir_cond_always) {
            inc_counter(counter_uncond_branch);
          } else {
            inc_counter(counter_cond_branch);
          }
          break;
        }

        case lir_neg:
        case lir_add:
        case lir_sub:
        case lir_mul:
        case lir_mul_strictfp:
        case lir_div:
        case lir_div_strictfp:
        case lir_rem:
        case lir_sqrt:
        case lir_sin:
        case lir_cos:
        case lir_abs:
        case lir_log10:
        case lir_log:
6583 6584
        case lir_pow:
        case lir_exp:
D
duke 已提交
6585 6586 6587 6588 6589 6590 6591 6592 6593 6594 6595 6596 6597 6598 6599 6600 6601 6602 6603 6604 6605 6606 6607 6608 6609 6610 6611 6612 6613 6614 6615 6616 6617 6618 6619 6620 6621 6622 6623 6624 6625 6626 6627 6628 6629 6630 6631 6632 6633 6634 6635 6636 6637 6638 6639 6640 6641 6642 6643 6644 6645 6646 6647 6648 6649 6650 6651 6652 6653 6654 6655 6656 6657 6658 6659 6660 6661 6662 6663 6664 6665 6666 6667 6668 6669 6670 6671 6672 6673 6674 6675 6676 6677 6678 6679 6680 6681 6682 6683 6684 6685 6686 6687 6688 6689 6690 6691 6692 6693 6694 6695 6696 6697 6698 6699 6700 6701 6702 6703 6704 6705 6706 6707 6708 6709 6710 6711 6712 6713 6714 6715 6716 6717 6718 6719 6720 6721 6722 6723 6724
        case lir_logic_and:
        case lir_logic_or:
        case lir_logic_xor:
        case lir_shl:
        case lir_shr:
        case lir_ushr:            inc_counter(counter_alu); break;

        case lir_alloc_object:
        case lir_alloc_array:     inc_counter(counter_alloc); break;

        case lir_monaddr:
        case lir_lock:
        case lir_unlock:          inc_counter(counter_sync); break;

        case lir_throw:           inc_counter(counter_throw); break;

        case lir_unwind:          inc_counter(counter_unwind); break;

        case lir_null_check:
        case lir_leal:
        case lir_instanceof:
        case lir_checkcast:
        case lir_store_check:     inc_counter(counter_typecheck); break;

        case lir_fpop_raw:
        case lir_fxch:
        case lir_fld:             inc_counter(counter_fpu_stack); break;

        case lir_nop:
        case lir_push:
        case lir_pop:
        case lir_convert:
        case lir_roundfp:
        case lir_cmove:           inc_counter(counter_misc_inst); break;

        default:                  inc_counter(counter_other_inst); break;
      }
    }
  }

  if (has_xhandlers) {
    inc_counter(counter_exception_method);
  }
}

void LinearScanStatistic::compute(LinearScan* allocator, LinearScanStatistic &global_statistic) {
  if (CountLinearScan || TraceLinearScanLevel > 0) {

    LinearScanStatistic local_statistic = LinearScanStatistic();

    local_statistic.collect(allocator);
    global_statistic.sum_up(local_statistic);

    if (TraceLinearScanLevel > 2) {
      local_statistic.print("current local statistic");
    }
  }
}


// Implementation of LinearTimers

LinearScanTimers::LinearScanTimers() {
  for (int i = 0; i < number_of_timers; i++) {
    timer(i)->reset();
  }
}

const char* LinearScanTimers::timer_name(int idx) {
  switch (idx) {
    case timer_do_nothing:               return "Nothing (Time Check)";
    case timer_number_instructions:      return "Number Instructions";
    case timer_compute_local_live_sets:  return "Local Live Sets";
    case timer_compute_global_live_sets: return "Global Live Sets";
    case timer_build_intervals:          return "Build Intervals";
    case timer_sort_intervals_before:    return "Sort Intervals Before";
    case timer_allocate_registers:       return "Allocate Registers";
    case timer_resolve_data_flow:        return "Resolve Data Flow";
    case timer_sort_intervals_after:     return "Sort Intervals After";
    case timer_eliminate_spill_moves:    return "Spill optimization";
    case timer_assign_reg_num:           return "Assign Reg Num";
    case timer_allocate_fpu_stack:       return "Allocate FPU Stack";
    case timer_optimize_lir:             return "Optimize LIR";
    default: ShouldNotReachHere();       return "";
  }
}

void LinearScanTimers::begin_method() {
  if (TimeEachLinearScan) {
    // reset all timers to measure only current method
    for (int i = 0; i < number_of_timers; i++) {
      timer(i)->reset();
    }
  }
}

void LinearScanTimers::end_method(LinearScan* allocator) {
  if (TimeEachLinearScan) {

    double c = timer(timer_do_nothing)->seconds();
    double total = 0;
    for (int i = 1; i < number_of_timers; i++) {
      total += timer(i)->seconds() - c;
    }

    if (total >= 0.0005) {
      // print all information in one line for automatic processing
      tty->print("@"); allocator->compilation()->method()->print_name();

      tty->print("@ %d ", allocator->compilation()->method()->code_size());
      tty->print("@ %d ", allocator->block_at(allocator->block_count() - 1)->last_lir_instruction_id() / 2);
      tty->print("@ %d ", allocator->block_count());
      tty->print("@ %d ", allocator->num_virtual_regs());
      tty->print("@ %d ", allocator->interval_count());
      tty->print("@ %d ", allocator->_num_calls);
      tty->print("@ %d ", allocator->num_loops());

      tty->print("@ %6.6f ", total);
      for (int i = 1; i < number_of_timers; i++) {
        tty->print("@ %4.1f ", ((timer(i)->seconds() - c) / total) * 100);
      }
      tty->cr();
    }
  }
}

void LinearScanTimers::print(double total_time) {
  if (TimeLinearScan) {
    // correction value: sum of dummy-timer that only measures the time that
    // is necesary to start and stop itself
    double c = timer(timer_do_nothing)->seconds();

    for (int i = 0; i < number_of_timers; i++) {
      double t = timer(i)->seconds();
      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);
    }
  }
}

#endif // #ifndef PRODUCT