1 /* 2 * Copyright (c) 2005, 2011, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "c1/c1_CFGPrinter.hpp" 27 #include "c1/c1_CodeStubs.hpp" 28 #include "c1/c1_Compilation.hpp" 29 #include "c1/c1_FrameMap.hpp" 30 #include "c1/c1_IR.hpp" 31 #include "c1/c1_LIRGenerator.hpp" 32 #include "c1/c1_LinearScan.hpp" 33 #include "c1/c1_ValueStack.hpp" 34 #include "utilities/bitMap.inline.hpp" 35 #ifdef TARGET_ARCH_x86 36 # include "vmreg_x86.inline.hpp" 37 #endif 38 #ifdef TARGET_ARCH_sparc 39 # include "vmreg_sparc.inline.hpp" 40 #endif 41 #ifdef TARGET_ARCH_zero 42 # include "vmreg_zero.inline.hpp" 43 #endif 44 #ifdef TARGET_ARCH_arm 45 # include "vmreg_arm.inline.hpp" 46 #endif 47 #ifdef TARGET_ARCH_ppc 48 # include "vmreg_ppc.inline.hpp" 49 #endif 50 51 52 #ifndef PRODUCT 53 54 static LinearScanStatistic _stat_before_alloc; 55 static LinearScanStatistic _stat_after_asign; 56 static LinearScanStatistic _stat_final; 57 58 static LinearScanTimers _total_timer; 59 60 // helper macro for short definition of timer 61 #define TIME_LINEAR_SCAN(timer_name) TraceTime _block_timer("", _total_timer.timer(LinearScanTimers::timer_name), TimeLinearScan || TimeEachLinearScan, Verbose); 62 63 // helper macro for short definition of trace-output inside code 64 #define TRACE_LINEAR_SCAN(level, code) \ 65 if (TraceLinearScanLevel >= level) { \ 66 code; \ 67 } 68 69 #else 70 71 #define TIME_LINEAR_SCAN(timer_name) 72 #define TRACE_LINEAR_SCAN(level, code) 73 74 #endif 75 76 // Map BasicType to spill size in 32-bit words, matching VMReg's notion of words 77 #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, 1, -1}; 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}; 81 #endif 82 83 84 // Implementation of LinearScan 85 86 LinearScan::LinearScan(IR* ir, LIRGenerator* gen, FrameMap* frame_map) 87 : _compilation(ir->compilation()) 88 , _ir(ir) 89 , _gen(gen) 90 , _frame_map(frame_map) 91 , _num_virtual_regs(gen->max_virtual_register_number()) 92 , _has_fpu_registers(false) 93 , _num_calls(-1) 94 , _max_spills(0) 95 , _unused_spill_slot(-1) 96 , _intervals(0) // initialized later with correct length 97 , _new_intervals_from_allocation(new IntervalList()) 98 , _sorted_intervals(NULL) 99 , _needs_full_resort(false) 100 , _lir_ops(0) // initialized later with correct length 101 , _block_of_op(0) // initialized later with correct length 102 , _has_info(0) 103 , _has_call(0) 104 , _scope_value_cache(0) // initialized later with correct length 105 , _interval_in_loop(0, 0) // initialized later with correct length 106 , _cached_blocks(*ir->linear_scan_order()) 107 #ifdef X86 108 , _fpu_stack_allocator(NULL) 109 #endif 110 { 111 assert(this->ir() != NULL, "check if valid"); 112 assert(this->compilation() != NULL, "check if valid"); 113 assert(this->gen() != NULL, "check if valid"); 114 assert(this->frame_map() != NULL, "check if valid"); 115 } 116 117 118 // ********** functions for converting LIR-Operands to register numbers 119 // 120 // Emulate a flat register file comprising physical integer registers, 121 // physical floating-point registers and virtual registers, in that order. 122 // Virtual registers already have appropriate numbers, since V0 is 123 // the number of physical registers. 124 // Returns -1 for hi word if opr is a single word operand. 125 // 126 // Note: the inverse operation (calculating an operand for register numbers) 127 // is done in calc_operand_for_interval() 128 129 int LinearScan::reg_num(LIR_Opr opr) { 130 assert(opr->is_register(), "should not call this otherwise"); 131 132 if (opr->is_virtual_register()) { 133 assert(opr->vreg_number() >= nof_regs, "found a virtual register with a fixed-register number"); 134 return opr->vreg_number(); 135 } else if (opr->is_single_cpu()) { 136 return opr->cpu_regnr(); 137 } else if (opr->is_double_cpu()) { 138 return opr->cpu_regnrLo(); 139 #ifdef X86 140 } else if (opr->is_single_xmm()) { 141 return opr->fpu_regnr() + pd_first_xmm_reg; 142 } else if (opr->is_double_xmm()) { 143 return opr->fpu_regnrLo() + pd_first_xmm_reg; 144 #endif 145 } else if (opr->is_single_fpu()) { 146 return opr->fpu_regnr() + pd_first_fpu_reg; 147 } else if (opr->is_double_fpu()) { 148 return opr->fpu_regnrLo() + pd_first_fpu_reg; 149 } else { 150 ShouldNotReachHere(); 151 return -1; 152 } 153 } 154 155 int LinearScan::reg_numHi(LIR_Opr opr) { 156 assert(opr->is_register(), "should not call this otherwise"); 157 158 if (opr->is_virtual_register()) { 159 return -1; 160 } else if (opr->is_single_cpu()) { 161 return -1; 162 } else if (opr->is_double_cpu()) { 163 return opr->cpu_regnrHi(); 164 #ifdef X86 165 } else if (opr->is_single_xmm()) { 166 return -1; 167 } else if (opr->is_double_xmm()) { 168 return -1; 169 #endif 170 } else if (opr->is_single_fpu()) { 171 return -1; 172 } else if (opr->is_double_fpu()) { 173 return opr->fpu_regnrHi() + pd_first_fpu_reg; 174 } else { 175 ShouldNotReachHere(); 176 return -1; 177 } 178 } 179 180 181 // ********** functions for classification of intervals 182 183 bool LinearScan::is_precolored_interval(const Interval* i) { 184 return i->reg_num() < LinearScan::nof_regs; 185 } 186 187 bool LinearScan::is_virtual_interval(const Interval* i) { 188 return i->reg_num() >= LIR_OprDesc::vreg_base; 189 } 190 191 bool LinearScan::is_precolored_cpu_interval(const Interval* i) { 192 return i->reg_num() < LinearScan::nof_cpu_regs; 193 } 194 195 bool LinearScan::is_virtual_cpu_interval(const Interval* i) { 196 #if defined(__SOFTFP__) || defined(E500V2) 197 return i->reg_num() >= LIR_OprDesc::vreg_base; 198 #else 199 return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() != T_FLOAT && i->type() != T_DOUBLE); 200 #endif // __SOFTFP__ or E500V2 201 } 202 203 bool LinearScan::is_precolored_fpu_interval(const Interval* i) { 204 return i->reg_num() >= LinearScan::nof_cpu_regs && i->reg_num() < LinearScan::nof_regs; 205 } 206 207 bool LinearScan::is_virtual_fpu_interval(const Interval* i) { 208 #if defined(__SOFTFP__) || defined(E500V2) 209 return false; 210 #else 211 return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() == T_FLOAT || i->type() == T_DOUBLE); 212 #endif // __SOFTFP__ or E500V2 213 } 214 215 bool LinearScan::is_in_fpu_register(const Interval* i) { 216 // fixed intervals not needed for FPU stack allocation 217 return i->reg_num() >= nof_regs && pd_first_fpu_reg <= i->assigned_reg() && i->assigned_reg() <= pd_last_fpu_reg; 218 } 219 220 bool LinearScan::is_oop_interval(const Interval* i) { 221 // fixed intervals never contain oops 222 return i->reg_num() >= nof_regs && i->type() == T_OBJECT; 223 } 224 225 226 // ********** General helper functions 227 228 // compute next unused stack index that can be used for spilling 229 int LinearScan::allocate_spill_slot(bool double_word) { 230 int spill_slot; 231 if (double_word) { 232 if ((_max_spills & 1) == 1) { 233 // alignment of double-word values 234 // the hole because of the alignment is filled with the next single-word value 235 assert(_unused_spill_slot == -1, "wasting a spill slot"); 236 _unused_spill_slot = _max_spills; 237 _max_spills++; 238 } 239 spill_slot = _max_spills; 240 _max_spills += 2; 241 242 } else if (_unused_spill_slot != -1) { 243 // re-use hole that was the result of a previous double-word alignment 244 spill_slot = _unused_spill_slot; 245 _unused_spill_slot = -1; 246 247 } else { 248 spill_slot = _max_spills; 249 _max_spills++; 250 } 251 252 int result = spill_slot + LinearScan::nof_regs + frame_map()->argcount(); 253 254 // the class OopMapValue uses only 11 bits for storing the name of the 255 // oop location. So a stack slot bigger than 2^11 leads to an overflow 256 // that is not reported in product builds. Prevent this by checking the 257 // spill slot here (altough this value and the later used location name 258 // are slightly different) 259 if (result > 2000) { 260 bailout("too many stack slots used"); 261 } 262 263 return result; 264 } 265 266 void LinearScan::assign_spill_slot(Interval* it) { 267 // assign the canonical spill slot of the parent (if a part of the interval 268 // is already spilled) or allocate a new spill slot 269 if (it->canonical_spill_slot() >= 0) { 270 it->assign_reg(it->canonical_spill_slot()); 271 } else { 272 int spill = allocate_spill_slot(type2spill_size[it->type()] == 2); 273 it->set_canonical_spill_slot(spill); 274 it->assign_reg(spill); 275 } 276 } 277 278 void LinearScan::propagate_spill_slots() { 279 if (!frame_map()->finalize_frame(max_spills())) { 280 bailout("frame too large"); 281 } 282 } 283 284 // create a new interval with a predefined reg_num 285 // (only used for parent intervals that are created during the building phase) 286 Interval* LinearScan::create_interval(int reg_num) { 287 assert(_intervals.at(reg_num) == NULL, "overwriting exisiting interval"); 288 289 Interval* interval = new Interval(reg_num); 290 _intervals.at_put(reg_num, interval); 291 292 // assign register number for precolored intervals 293 if (reg_num < LIR_OprDesc::vreg_base) { 294 interval->assign_reg(reg_num); 295 } 296 return interval; 297 } 298 299 // assign a new reg_num to the interval and append it to the list of intervals 300 // (only used for child intervals that are created during register allocation) 301 void LinearScan::append_interval(Interval* it) { 302 it->set_reg_num(_intervals.length()); 303 _intervals.append(it); 304 _new_intervals_from_allocation->append(it); 305 } 306 307 // copy the vreg-flags if an interval is split 308 void LinearScan::copy_register_flags(Interval* from, Interval* to) { 309 if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::byte_reg)) { 310 gen()->set_vreg_flag(to->reg_num(), LIRGenerator::byte_reg); 311 } 312 if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::callee_saved)) { 313 gen()->set_vreg_flag(to->reg_num(), LIRGenerator::callee_saved); 314 } 315 316 // Note: do not copy the must_start_in_memory flag because it is not necessary for child 317 // intervals (only the very beginning of the interval must be in memory) 318 } 319 320 321 // ********** spill move optimization 322 // eliminate moves from register to stack if stack slot is known to be correct 323 324 // called during building of intervals 325 void LinearScan::change_spill_definition_pos(Interval* interval, int def_pos) { 326 assert(interval->is_split_parent(), "can only be called for split parents"); 327 328 switch (interval->spill_state()) { 329 case noDefinitionFound: 330 assert(interval->spill_definition_pos() == -1, "must no be set before"); 331 interval->set_spill_definition_pos(def_pos); 332 interval->set_spill_state(oneDefinitionFound); 333 break; 334 335 case oneDefinitionFound: 336 assert(def_pos <= interval->spill_definition_pos(), "positions are processed in reverse order when intervals are created"); 337 if (def_pos < interval->spill_definition_pos() - 2) { 338 // second definition found, so no spill optimization possible for this interval 339 interval->set_spill_state(noOptimization); 340 } else { 341 // two consecutive definitions (because of two-operand LIR form) 342 assert(block_of_op_with_id(def_pos) == block_of_op_with_id(interval->spill_definition_pos()), "block must be equal"); 343 } 344 break; 345 346 case noOptimization: 347 // nothing to do 348 break; 349 350 default: 351 assert(false, "other states not allowed at this time"); 352 } 353 } 354 355 // called during register allocation 356 void LinearScan::change_spill_state(Interval* interval, int spill_pos) { 357 switch (interval->spill_state()) { 358 case oneDefinitionFound: { 359 int def_loop_depth = block_of_op_with_id(interval->spill_definition_pos())->loop_depth(); 360 int spill_loop_depth = block_of_op_with_id(spill_pos)->loop_depth(); 361 362 if (def_loop_depth < spill_loop_depth) { 363 // the loop depth of the spilling position is higher then the loop depth 364 // at the definition of the interval -> move write to memory out of loop 365 // by storing at definitin of the interval 366 interval->set_spill_state(storeAtDefinition); 367 } else { 368 // the interval is currently spilled only once, so for now there is no 369 // reason to store the interval at the definition 370 interval->set_spill_state(oneMoveInserted); 371 } 372 break; 373 } 374 375 case oneMoveInserted: { 376 // the interval is spilled more then once, so it is better to store it to 377 // memory at the definition 378 interval->set_spill_state(storeAtDefinition); 379 break; 380 } 381 382 case storeAtDefinition: 383 case startInMemory: 384 case noOptimization: 385 case noDefinitionFound: 386 // nothing to do 387 break; 388 389 default: 390 assert(false, "other states not allowed at this time"); 391 } 392 } 393 394 395 bool LinearScan::must_store_at_definition(const Interval* i) { 396 return i->is_split_parent() && i->spill_state() == storeAtDefinition; 397 } 398 399 // called once before asignment of register numbers 400 void LinearScan::eliminate_spill_moves() { 401 TIME_LINEAR_SCAN(timer_eliminate_spill_moves); 402 TRACE_LINEAR_SCAN(3, tty->print_cr("***** Eliminating unnecessary spill moves")); 403 404 // collect all intervals that must be stored after their definion. 405 // the list is sorted by Interval::spill_definition_pos 406 Interval* interval; 407 Interval* temp_list; 408 create_unhandled_lists(&interval, &temp_list, must_store_at_definition, NULL); 409 410 #ifdef ASSERT 411 Interval* prev = NULL; 412 Interval* temp = interval; 413 while (temp != Interval::end()) { 414 assert(temp->spill_definition_pos() > 0, "invalid spill definition pos"); 415 if (prev != NULL) { 416 assert(temp->from() >= prev->from(), "intervals not sorted"); 417 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"); 418 } 419 420 assert(temp->canonical_spill_slot() >= LinearScan::nof_regs, "interval has no spill slot assigned"); 421 assert(temp->spill_definition_pos() >= temp->from(), "invalid order"); 422 assert(temp->spill_definition_pos() <= temp->from() + 2, "only intervals defined once at their start-pos can be optimized"); 423 424 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())); 425 426 temp = temp->next(); 427 } 428 #endif 429 430 LIR_InsertionBuffer insertion_buffer; 431 int num_blocks = block_count(); 432 for (int i = 0; i < num_blocks; i++) { 433 BlockBegin* block = block_at(i); 434 LIR_OpList* instructions = block->lir()->instructions_list(); 435 int num_inst = instructions->length(); 436 bool has_new = false; 437 438 // iterate all instructions of the block. skip the first because it is always a label 439 for (int j = 1; j < num_inst; j++) { 440 LIR_Op* op = instructions->at(j); 441 int op_id = op->id(); 442 443 if (op_id == -1) { 444 // remove move from register to stack if the stack slot is guaranteed to be correct. 445 // only moves that have been inserted by LinearScan can be removed. 446 assert(op->code() == lir_move, "only moves can have a op_id of -1"); 447 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); 448 assert(op->as_Op1()->result_opr()->is_virtual(), "LinearScan inserts only moves to virtual registers"); 449 450 LIR_Op1* op1 = (LIR_Op1*)op; 451 Interval* interval = interval_at(op1->result_opr()->vreg_number()); 452 453 if (interval->assigned_reg() >= LinearScan::nof_regs && interval->always_in_memory()) { 454 // move target is a stack slot that is always correct, so eliminate instruction 455 TRACE_LINEAR_SCAN(4, tty->print_cr("eliminating move from interval %d to %d", op1->in_opr()->vreg_number(), op1->result_opr()->vreg_number())); 456 instructions->at_put(j, NULL); // NULL-instructions are deleted by assign_reg_num 457 } 458 459 } else { 460 // insert move from register to stack just after the beginning of the interval 461 assert(interval == Interval::end() || interval->spill_definition_pos() >= op_id, "invalid order"); 462 assert(interval == Interval::end() || (interval->is_split_parent() && interval->spill_state() == storeAtDefinition), "invalid interval"); 463 464 while (interval != Interval::end() && interval->spill_definition_pos() == op_id) { 465 if (!has_new) { 466 // prepare insertion buffer (appended when all instructions of the block are processed) 467 insertion_buffer.init(block->lir()); 468 has_new = true; 469 } 470 471 LIR_Opr from_opr = operand_for_interval(interval); 472 LIR_Opr to_opr = canonical_spill_opr(interval); 473 assert(from_opr->is_fixed_cpu() || from_opr->is_fixed_fpu(), "from operand must be a register"); 474 assert(to_opr->is_stack(), "to operand must be a stack slot"); 475 476 insertion_buffer.move(j, from_opr, to_opr); 477 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)); 478 479 interval = interval->next(); 480 } 481 } 482 } // end of instruction iteration 483 484 if (has_new) { 485 block->lir()->append(&insertion_buffer); 486 } 487 } // end of block iteration 488 489 assert(interval == Interval::end(), "missed an interval"); 490 } 491 492 493 // ********** Phase 1: number all instructions in all blocks 494 // Compute depth-first and linear scan block orders, and number LIR_Op nodes for linear scan. 495 496 void LinearScan::number_instructions() { 497 { 498 // dummy-timer to measure the cost of the timer itself 499 // (this time is then subtracted from all other timers to get the real value) 500 TIME_LINEAR_SCAN(timer_do_nothing); 501 } 502 TIME_LINEAR_SCAN(timer_number_instructions); 503 504 // Assign IDs to LIR nodes and build a mapping, lir_ops, from ID to LIR_Op node. 505 int num_blocks = block_count(); 506 int num_instructions = 0; 507 int i; 508 for (i = 0; i < num_blocks; i++) { 509 num_instructions += block_at(i)->lir()->instructions_list()->length(); 510 } 511 512 // initialize with correct length 513 _lir_ops = LIR_OpArray(num_instructions); 514 _block_of_op = BlockBeginArray(num_instructions); 515 516 int op_id = 0; 517 int idx = 0; 518 519 for (i = 0; i < num_blocks; i++) { 520 BlockBegin* block = block_at(i); 521 block->set_first_lir_instruction_id(op_id); 522 LIR_OpList* instructions = block->lir()->instructions_list(); 523 524 int num_inst = instructions->length(); 525 for (int j = 0; j < num_inst; j++) { 526 LIR_Op* op = instructions->at(j); 527 op->set_id(op_id); 528 529 _lir_ops.at_put(idx, op); 530 _block_of_op.at_put(idx, block); 531 assert(lir_op_with_id(op_id) == op, "must match"); 532 533 idx++; 534 op_id += 2; // numbering of lir_ops by two 535 } 536 block->set_last_lir_instruction_id(op_id - 2); 537 } 538 assert(idx == num_instructions, "must match"); 539 assert(idx * 2 == op_id, "must match"); 540 541 _has_call = BitMap(num_instructions); _has_call.clear(); 542 _has_info = BitMap(num_instructions); _has_info.clear(); 543 } 544 545 546 // ********** Phase 2: compute local live sets separately for each block 547 // (sets live_gen and live_kill for each block) 548 549 void LinearScan::set_live_gen_kill(Value value, LIR_Op* op, BitMap& live_gen, BitMap& live_kill) { 550 LIR_Opr opr = value->operand(); 551 Constant* con = value->as_Constant(); 552 553 // check some asumptions about debug information 554 assert(!value->type()->is_illegal(), "if this local is used by the interpreter it shouldn't be of indeterminate type"); 555 assert(con == NULL || opr->is_virtual() || opr->is_constant() || opr->is_illegal(), "asumption: Constant instructions have only constant operands"); 556 assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands"); 557 558 if ((con == NULL || con->is_pinned()) && opr->is_register()) { 559 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 560 int reg = opr->vreg_number(); 561 if (!live_kill.at(reg)) { 562 live_gen.set_bit(reg); 563 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)); 564 } 565 } 566 } 567 568 569 void LinearScan::compute_local_live_sets() { 570 TIME_LINEAR_SCAN(timer_compute_local_live_sets); 571 572 int num_blocks = block_count(); 573 int live_size = live_set_size(); 574 bool local_has_fpu_registers = false; 575 int local_num_calls = 0; 576 LIR_OpVisitState visitor; 577 578 BitMap2D local_interval_in_loop = BitMap2D(_num_virtual_regs, num_loops()); 579 local_interval_in_loop.clear(); 580 581 // iterate all blocks 582 for (int i = 0; i < num_blocks; i++) { 583 BlockBegin* block = block_at(i); 584 585 BitMap live_gen(live_size); live_gen.clear(); 586 BitMap live_kill(live_size); live_kill.clear(); 587 588 if (block->is_set(BlockBegin::exception_entry_flag)) { 589 // Phi functions at the begin of an exception handler are 590 // implicitly defined (= killed) at the beginning of the block. 591 for_each_phi_fun(block, phi, 592 live_kill.set_bit(phi->operand()->vreg_number()) 593 ); 594 } 595 596 LIR_OpList* instructions = block->lir()->instructions_list(); 597 int num_inst = instructions->length(); 598 599 // iterate all instructions of the block. skip the first because it is always a label 600 assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label"); 601 for (int j = 1; j < num_inst; j++) { 602 LIR_Op* op = instructions->at(j); 603 604 // visit operation to collect all operands 605 visitor.visit(op); 606 607 if (visitor.has_call()) { 608 _has_call.set_bit(op->id() >> 1); 609 local_num_calls++; 610 } 611 if (visitor.info_count() > 0) { 612 _has_info.set_bit(op->id() >> 1); 613 } 614 615 // iterate input operands of instruction 616 int k, n, reg; 617 n = visitor.opr_count(LIR_OpVisitState::inputMode); 618 for (k = 0; k < n; k++) { 619 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k); 620 assert(opr->is_register(), "visitor should only return register operands"); 621 622 if (opr->is_virtual_register()) { 623 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 624 reg = opr->vreg_number(); 625 if (!live_kill.at(reg)) { 626 live_gen.set_bit(reg); 627 TRACE_LINEAR_SCAN(4, tty->print_cr(" Setting live_gen for register %d at instruction %d", reg, op->id())); 628 } 629 if (block->loop_index() >= 0) { 630 local_interval_in_loop.set_bit(reg, block->loop_index()); 631 } 632 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); 633 } 634 635 #ifdef ASSERT 636 // fixed intervals are never live at block boundaries, so 637 // they need not be processed in live sets. 638 // this is checked by these assertions to be sure about it. 639 // the entry block may have incoming values in registers, which is ok. 640 if (!opr->is_virtual_register() && block != ir()->start()) { 641 reg = reg_num(opr); 642 if (is_processed_reg_num(reg)) { 643 assert(live_kill.at(reg), "using fixed register that is not defined in this block"); 644 } 645 reg = reg_numHi(opr); 646 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 647 assert(live_kill.at(reg), "using fixed register that is not defined in this block"); 648 } 649 } 650 #endif 651 } 652 653 // Add uses of live locals from interpreter's point of view for proper debug information generation 654 n = visitor.info_count(); 655 for (k = 0; k < n; k++) { 656 CodeEmitInfo* info = visitor.info_at(k); 657 ValueStack* stack = info->stack(); 658 for_each_state_value(stack, value, 659 set_live_gen_kill(value, op, live_gen, live_kill) 660 ); 661 } 662 663 // iterate temp operands of instruction 664 n = visitor.opr_count(LIR_OpVisitState::tempMode); 665 for (k = 0; k < n; k++) { 666 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k); 667 assert(opr->is_register(), "visitor should only return register operands"); 668 669 if (opr->is_virtual_register()) { 670 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 671 reg = opr->vreg_number(); 672 live_kill.set_bit(reg); 673 if (block->loop_index() >= 0) { 674 local_interval_in_loop.set_bit(reg, block->loop_index()); 675 } 676 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); 677 } 678 679 #ifdef ASSERT 680 // fixed intervals are never live at block boundaries, so 681 // they need not be processed in live sets 682 // process them only in debug mode so that this can be checked 683 if (!opr->is_virtual_register()) { 684 reg = reg_num(opr); 685 if (is_processed_reg_num(reg)) { 686 live_kill.set_bit(reg_num(opr)); 687 } 688 reg = reg_numHi(opr); 689 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 690 live_kill.set_bit(reg); 691 } 692 } 693 #endif 694 } 695 696 // iterate output operands of instruction 697 n = visitor.opr_count(LIR_OpVisitState::outputMode); 698 for (k = 0; k < n; k++) { 699 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k); 700 assert(opr->is_register(), "visitor should only return register operands"); 701 702 if (opr->is_virtual_register()) { 703 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 704 reg = opr->vreg_number(); 705 live_kill.set_bit(reg); 706 if (block->loop_index() >= 0) { 707 local_interval_in_loop.set_bit(reg, block->loop_index()); 708 } 709 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); 710 } 711 712 #ifdef ASSERT 713 // fixed intervals are never live at block boundaries, so 714 // they need not be processed in live sets 715 // process them only in debug mode so that this can be checked 716 if (!opr->is_virtual_register()) { 717 reg = reg_num(opr); 718 if (is_processed_reg_num(reg)) { 719 live_kill.set_bit(reg_num(opr)); 720 } 721 reg = reg_numHi(opr); 722 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 723 live_kill.set_bit(reg); 724 } 725 } 726 #endif 727 } 728 } // end of instruction iteration 729 730 block->set_live_gen (live_gen); 731 block->set_live_kill(live_kill); 732 block->set_live_in (BitMap(live_size)); block->live_in().clear(); 733 block->set_live_out (BitMap(live_size)); block->live_out().clear(); 734 735 TRACE_LINEAR_SCAN(4, tty->print("live_gen B%d ", block->block_id()); print_bitmap(block->live_gen())); 736 TRACE_LINEAR_SCAN(4, tty->print("live_kill B%d ", block->block_id()); print_bitmap(block->live_kill())); 737 } // end of block iteration 738 739 // propagate local calculated information into LinearScan object 740 _has_fpu_registers = local_has_fpu_registers; 741 compilation()->set_has_fpu_code(local_has_fpu_registers); 742 743 _num_calls = local_num_calls; 744 _interval_in_loop = local_interval_in_loop; 745 } 746 747 748 // ********** Phase 3: perform a backward dataflow analysis to compute global live sets 749 // (sets live_in and live_out for each block) 750 751 void LinearScan::compute_global_live_sets() { 752 TIME_LINEAR_SCAN(timer_compute_global_live_sets); 753 754 int num_blocks = block_count(); 755 bool change_occurred; 756 bool change_occurred_in_block; 757 int iteration_count = 0; 758 BitMap live_out(live_set_size()); live_out.clear(); // scratch set for calculations 759 760 // Perform a backward dataflow analysis to compute live_out and live_in for each block. 761 // The loop is executed until a fixpoint is reached (no changes in an iteration) 762 // Exception handlers must be processed because not all live values are 763 // present in the state array, e.g. because of global value numbering 764 do { 765 change_occurred = false; 766 767 // iterate all blocks in reverse order 768 for (int i = num_blocks - 1; i >= 0; i--) { 769 BlockBegin* block = block_at(i); 770 771 change_occurred_in_block = false; 772 773 // live_out(block) is the union of live_in(sux), for successors sux of block 774 int n = block->number_of_sux(); 775 int e = block->number_of_exception_handlers(); 776 if (n + e > 0) { 777 // block has successors 778 if (n > 0) { 779 live_out.set_from(block->sux_at(0)->live_in()); 780 for (int j = 1; j < n; j++) { 781 live_out.set_union(block->sux_at(j)->live_in()); 782 } 783 } else { 784 live_out.clear(); 785 } 786 for (int j = 0; j < e; j++) { 787 live_out.set_union(block->exception_handler_at(j)->live_in()); 788 } 789 790 if (!block->live_out().is_same(live_out)) { 791 // A change occurred. Swap the old and new live out sets to avoid copying. 792 BitMap temp = block->live_out(); 793 block->set_live_out(live_out); 794 live_out = temp; 795 796 change_occurred = true; 797 change_occurred_in_block = true; 798 } 799 } 800 801 if (iteration_count == 0 || change_occurred_in_block) { 802 // live_in(block) is the union of live_gen(block) with (live_out(block) & !live_kill(block)) 803 // note: live_in has to be computed only in first iteration or if live_out has changed! 804 BitMap live_in = block->live_in(); 805 live_in.set_from(block->live_out()); 806 live_in.set_difference(block->live_kill()); 807 live_in.set_union(block->live_gen()); 808 } 809 810 #ifndef PRODUCT 811 if (TraceLinearScanLevel >= 4) { 812 char c = ' '; 813 if (iteration_count == 0 || change_occurred_in_block) { 814 c = '*'; 815 } 816 tty->print("(%d) live_in%c B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_in()); 817 tty->print("(%d) live_out%c B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_out()); 818 } 819 #endif 820 } 821 iteration_count++; 822 823 if (change_occurred && iteration_count > 50) { 824 BAILOUT("too many iterations in compute_global_live_sets"); 825 } 826 } while (change_occurred); 827 828 829 #ifdef ASSERT 830 // check that fixed intervals are not live at block boundaries 831 // (live set must be empty at fixed intervals) 832 for (int i = 0; i < num_blocks; i++) { 833 BlockBegin* block = block_at(i); 834 for (int j = 0; j < LIR_OprDesc::vreg_base; j++) { 835 assert(block->live_in().at(j) == false, "live_in set of fixed register must be empty"); 836 assert(block->live_out().at(j) == false, "live_out set of fixed register must be empty"); 837 assert(block->live_gen().at(j) == false, "live_gen set of fixed register must be empty"); 838 } 839 } 840 #endif 841 842 // check that the live_in set of the first block is empty 843 BitMap live_in_args(ir()->start()->live_in().size()); 844 live_in_args.clear(); 845 if (!ir()->start()->live_in().is_same(live_in_args)) { 846 #ifdef ASSERT 847 tty->print_cr("Error: live_in set of first block must be empty (when this fails, virtual registers are used before they are defined)"); 848 tty->print_cr("affected registers:"); 849 print_bitmap(ir()->start()->live_in()); 850 851 // print some additional information to simplify debugging 852 for (unsigned int i = 0; i < ir()->start()->live_in().size(); i++) { 853 if (ir()->start()->live_in().at(i)) { 854 Instruction* instr = gen()->instruction_for_vreg(i); 855 tty->print_cr("* vreg %d (HIR instruction %c%d)", i, instr == NULL ? ' ' : instr->type()->tchar(), instr == NULL ? 0 : instr->id()); 856 857 for (int j = 0; j < num_blocks; j++) { 858 BlockBegin* block = block_at(j); 859 if (block->live_gen().at(i)) { 860 tty->print_cr(" used in block B%d", block->block_id()); 861 } 862 if (block->live_kill().at(i)) { 863 tty->print_cr(" defined in block B%d", block->block_id()); 864 } 865 } 866 } 867 } 868 869 #endif 870 // when this fails, virtual registers are used before they are defined. 871 assert(false, "live_in set of first block must be empty"); 872 // bailout of if this occurs in product mode. 873 bailout("live_in set of first block not empty"); 874 } 875 } 876 877 878 // ********** Phase 4: build intervals 879 // (fills the list _intervals) 880 881 void LinearScan::add_use(Value value, int from, int to, IntervalUseKind use_kind) { 882 assert(!value->type()->is_illegal(), "if this value is used by the interpreter it shouldn't be of indeterminate type"); 883 LIR_Opr opr = value->operand(); 884 Constant* con = value->as_Constant(); 885 886 if ((con == NULL || con->is_pinned()) && opr->is_register()) { 887 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 888 add_use(opr, from, to, use_kind); 889 } 890 } 891 892 893 void LinearScan::add_def(LIR_Opr opr, int def_pos, IntervalUseKind use_kind) { 894 TRACE_LINEAR_SCAN(2, tty->print(" def "); opr->print(tty); tty->print_cr(" def_pos %d (%d)", def_pos, use_kind)); 895 assert(opr->is_register(), "should not be called otherwise"); 896 897 if (opr->is_virtual_register()) { 898 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 899 add_def(opr->vreg_number(), def_pos, use_kind, opr->type_register()); 900 901 } else { 902 int reg = reg_num(opr); 903 if (is_processed_reg_num(reg)) { 904 add_def(reg, def_pos, use_kind, opr->type_register()); 905 } 906 reg = reg_numHi(opr); 907 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 908 add_def(reg, def_pos, use_kind, opr->type_register()); 909 } 910 } 911 } 912 913 void LinearScan::add_use(LIR_Opr opr, int from, int to, IntervalUseKind use_kind) { 914 TRACE_LINEAR_SCAN(2, tty->print(" use "); opr->print(tty); tty->print_cr(" from %d to %d (%d)", from, to, use_kind)); 915 assert(opr->is_register(), "should not be called otherwise"); 916 917 if (opr->is_virtual_register()) { 918 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 919 add_use(opr->vreg_number(), from, to, use_kind, opr->type_register()); 920 921 } else { 922 int reg = reg_num(opr); 923 if (is_processed_reg_num(reg)) { 924 add_use(reg, from, to, use_kind, opr->type_register()); 925 } 926 reg = reg_numHi(opr); 927 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 928 add_use(reg, from, to, use_kind, opr->type_register()); 929 } 930 } 931 } 932 933 void LinearScan::add_temp(LIR_Opr opr, int temp_pos, IntervalUseKind use_kind) { 934 TRACE_LINEAR_SCAN(2, tty->print(" temp "); opr->print(tty); tty->print_cr(" temp_pos %d (%d)", temp_pos, use_kind)); 935 assert(opr->is_register(), "should not be called otherwise"); 936 937 if (opr->is_virtual_register()) { 938 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 939 add_temp(opr->vreg_number(), temp_pos, use_kind, opr->type_register()); 940 941 } else { 942 int reg = reg_num(opr); 943 if (is_processed_reg_num(reg)) { 944 add_temp(reg, temp_pos, use_kind, opr->type_register()); 945 } 946 reg = reg_numHi(opr); 947 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 948 add_temp(reg, temp_pos, use_kind, opr->type_register()); 949 } 950 } 951 } 952 953 954 void LinearScan::add_def(int reg_num, int def_pos, IntervalUseKind use_kind, BasicType type) { 955 Interval* interval = interval_at(reg_num); 956 if (interval != NULL) { 957 assert(interval->reg_num() == reg_num, "wrong interval"); 958 959 if (type != T_ILLEGAL) { 960 interval->set_type(type); 961 } 962 963 Range* r = interval->first(); 964 if (r->from() <= def_pos) { 965 // Update the starting point (when a range is first created for a use, its 966 // start is the beginning of the current block until a def is encountered.) 967 r->set_from(def_pos); 968 interval->add_use_pos(def_pos, use_kind); 969 970 } else { 971 // Dead value - make vacuous interval 972 // also add use_kind for dead intervals 973 interval->add_range(def_pos, def_pos + 1); 974 interval->add_use_pos(def_pos, use_kind); 975 TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: def of reg %d at %d occurs without use", reg_num, def_pos)); 976 } 977 978 } else { 979 // Dead value - make vacuous interval 980 // also add use_kind for dead intervals 981 interval = create_interval(reg_num); 982 if (type != T_ILLEGAL) { 983 interval->set_type(type); 984 } 985 986 interval->add_range(def_pos, def_pos + 1); 987 interval->add_use_pos(def_pos, use_kind); 988 TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: dead value %d at %d in live intervals", reg_num, def_pos)); 989 } 990 991 change_spill_definition_pos(interval, def_pos); 992 if (use_kind == noUse && interval->spill_state() <= startInMemory) { 993 // detection of method-parameters and roundfp-results 994 // TODO: move this directly to position where use-kind is computed 995 interval->set_spill_state(startInMemory); 996 } 997 } 998 999 void LinearScan::add_use(int reg_num, int from, int to, IntervalUseKind use_kind, BasicType type) { 1000 Interval* interval = interval_at(reg_num); 1001 if (interval == NULL) { 1002 interval = create_interval(reg_num); 1003 } 1004 assert(interval->reg_num() == reg_num, "wrong interval"); 1005 1006 if (type != T_ILLEGAL) { 1007 interval->set_type(type); 1008 } 1009 1010 interval->add_range(from, to); 1011 interval->add_use_pos(to, use_kind); 1012 } 1013 1014 void LinearScan::add_temp(int reg_num, int temp_pos, IntervalUseKind use_kind, BasicType type) { 1015 Interval* interval = interval_at(reg_num); 1016 if (interval == NULL) { 1017 interval = create_interval(reg_num); 1018 } 1019 assert(interval->reg_num() == reg_num, "wrong interval"); 1020 1021 if (type != T_ILLEGAL) { 1022 interval->set_type(type); 1023 } 1024 1025 interval->add_range(temp_pos, temp_pos + 1); 1026 interval->add_use_pos(temp_pos, use_kind); 1027 } 1028 1029 1030 // the results of this functions are used for optimizing spilling and reloading 1031 // if the functions return shouldHaveRegister and the interval is spilled, 1032 // it is not reloaded to a register. 1033 IntervalUseKind LinearScan::use_kind_of_output_operand(LIR_Op* op, LIR_Opr opr) { 1034 if (op->code() == lir_move) { 1035 assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1"); 1036 LIR_Op1* move = (LIR_Op1*)op; 1037 LIR_Opr res = move->result_opr(); 1038 bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory); 1039 1040 if (result_in_memory) { 1041 // Begin of an interval with must_start_in_memory set. 1042 // This interval will always get a stack slot first, so return noUse. 1043 return noUse; 1044 1045 } else if (move->in_opr()->is_stack()) { 1046 // method argument (condition must be equal to handle_method_arguments) 1047 return noUse; 1048 1049 } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) { 1050 // Move from register to register 1051 if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) { 1052 // special handling of phi-function moves inside osr-entry blocks 1053 // input operand must have a register instead of output operand (leads to better register allocation) 1054 return shouldHaveRegister; 1055 } 1056 } 1057 } 1058 1059 if (opr->is_virtual() && 1060 gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::must_start_in_memory)) { 1061 // result is a stack-slot, so prevent immediate reloading 1062 return noUse; 1063 } 1064 1065 // all other operands require a register 1066 return mustHaveRegister; 1067 } 1068 1069 IntervalUseKind LinearScan::use_kind_of_input_operand(LIR_Op* op, LIR_Opr opr) { 1070 if (op->code() == lir_move) { 1071 assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1"); 1072 LIR_Op1* move = (LIR_Op1*)op; 1073 LIR_Opr res = move->result_opr(); 1074 bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory); 1075 1076 if (result_in_memory) { 1077 // Move to an interval with must_start_in_memory set. 1078 // To avoid moves from stack to stack (not allowed) force the input operand to a register 1079 return mustHaveRegister; 1080 1081 } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) { 1082 // Move from register to register 1083 if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) { 1084 // special handling of phi-function moves inside osr-entry blocks 1085 // input operand must have a register instead of output operand (leads to better register allocation) 1086 return mustHaveRegister; 1087 } 1088 1089 // The input operand is not forced to a register (moves from stack to register are allowed), 1090 // but it is faster if the input operand is in a register 1091 return shouldHaveRegister; 1092 } 1093 } 1094 1095 1096 #ifdef X86 1097 if (op->code() == lir_cmove) { 1098 // conditional moves can handle stack operands 1099 assert(op->result_opr()->is_register(), "result must always be in a register"); 1100 return shouldHaveRegister; 1101 } 1102 1103 // optimizations for second input operand of arithmehtic operations on Intel 1104 // this operand is allowed to be on the stack in some cases 1105 BasicType opr_type = opr->type_register(); 1106 if (opr_type == T_FLOAT || opr_type == T_DOUBLE) { 1107 if ((UseSSE == 1 && opr_type == T_FLOAT) || UseSSE >= 2) { 1108 // SSE float instruction (T_DOUBLE only supported with SSE2) 1109 switch (op->code()) { 1110 case lir_cmp: 1111 case lir_add: 1112 case lir_sub: 1113 case lir_mul: 1114 case lir_div: 1115 { 1116 assert(op->as_Op2() != NULL, "must be LIR_Op2"); 1117 LIR_Op2* op2 = (LIR_Op2*)op; 1118 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { 1119 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"); 1120 return shouldHaveRegister; 1121 } 1122 } 1123 } 1124 } else { 1125 // FPU stack float instruction 1126 switch (op->code()) { 1127 case lir_add: 1128 case lir_sub: 1129 case lir_mul: 1130 case lir_div: 1131 { 1132 assert(op->as_Op2() != NULL, "must be LIR_Op2"); 1133 LIR_Op2* op2 = (LIR_Op2*)op; 1134 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { 1135 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"); 1136 return shouldHaveRegister; 1137 } 1138 } 1139 } 1140 } 1141 1142 } else if (opr_type != T_LONG) { 1143 // integer instruction (note: long operands must always be in register) 1144 switch (op->code()) { 1145 case lir_cmp: 1146 case lir_add: 1147 case lir_sub: 1148 case lir_logic_and: 1149 case lir_logic_or: 1150 case lir_logic_xor: 1151 { 1152 assert(op->as_Op2() != NULL, "must be LIR_Op2"); 1153 LIR_Op2* op2 = (LIR_Op2*)op; 1154 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { 1155 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"); 1156 return shouldHaveRegister; 1157 } 1158 } 1159 } 1160 } 1161 #endif // X86 1162 1163 // all other operands require a register 1164 return mustHaveRegister; 1165 } 1166 1167 1168 void LinearScan::handle_method_arguments(LIR_Op* op) { 1169 // special handling for method arguments (moves from stack to virtual register): 1170 // the interval gets no register assigned, but the stack slot. 1171 // it is split before the first use by the register allocator. 1172 1173 if (op->code() == lir_move) { 1174 assert(op->as_Op1() != NULL, "must be LIR_Op1"); 1175 LIR_Op1* move = (LIR_Op1*)op; 1176 1177 if (move->in_opr()->is_stack()) { 1178 #ifdef ASSERT 1179 int arg_size = compilation()->method()->arg_size(); 1180 LIR_Opr o = move->in_opr(); 1181 if (o->is_single_stack()) { 1182 assert(o->single_stack_ix() >= 0 && o->single_stack_ix() < arg_size, "out of range"); 1183 } else if (o->is_double_stack()) { 1184 assert(o->double_stack_ix() >= 0 && o->double_stack_ix() < arg_size, "out of range"); 1185 } else { 1186 ShouldNotReachHere(); 1187 } 1188 1189 assert(move->id() > 0, "invalid id"); 1190 assert(block_of_op_with_id(move->id())->number_of_preds() == 0, "move from stack must be in first block"); 1191 assert(move->result_opr()->is_virtual(), "result of move must be a virtual register"); 1192 1193 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()))); 1194 #endif 1195 1196 Interval* interval = interval_at(reg_num(move->result_opr())); 1197 1198 int stack_slot = LinearScan::nof_regs + (move->in_opr()->is_single_stack() ? move->in_opr()->single_stack_ix() : move->in_opr()->double_stack_ix()); 1199 interval->set_canonical_spill_slot(stack_slot); 1200 interval->assign_reg(stack_slot); 1201 } 1202 } 1203 } 1204 1205 void LinearScan::handle_doubleword_moves(LIR_Op* op) { 1206 // special handling for doubleword move from memory to register: 1207 // in this case the registers of the input address and the result 1208 // registers must not overlap -> add a temp range for the input registers 1209 if (op->code() == lir_move) { 1210 assert(op->as_Op1() != NULL, "must be LIR_Op1"); 1211 LIR_Op1* move = (LIR_Op1*)op; 1212 1213 if (move->result_opr()->is_double_cpu() && move->in_opr()->is_pointer()) { 1214 LIR_Address* address = move->in_opr()->as_address_ptr(); 1215 if (address != NULL) { 1216 if (address->base()->is_valid()) { 1217 add_temp(address->base(), op->id(), noUse); 1218 } 1219 if (address->index()->is_valid()) { 1220 add_temp(address->index(), op->id(), noUse); 1221 } 1222 } 1223 } 1224 } 1225 } 1226 1227 void LinearScan::add_register_hints(LIR_Op* op) { 1228 switch (op->code()) { 1229 case lir_move: // fall through 1230 case lir_convert: { 1231 assert(op->as_Op1() != NULL, "lir_move, lir_convert must be LIR_Op1"); 1232 LIR_Op1* move = (LIR_Op1*)op; 1233 1234 LIR_Opr move_from = move->in_opr(); 1235 LIR_Opr move_to = move->result_opr(); 1236 1237 if (move_to->is_register() && move_from->is_register()) { 1238 Interval* from = interval_at(reg_num(move_from)); 1239 Interval* to = interval_at(reg_num(move_to)); 1240 if (from != NULL && to != NULL) { 1241 to->set_register_hint(from); 1242 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())); 1243 } 1244 } 1245 break; 1246 } 1247 case lir_cmove: { 1248 assert(op->as_Op2() != NULL, "lir_cmove must be LIR_Op2"); 1249 LIR_Op2* cmove = (LIR_Op2*)op; 1250 1251 LIR_Opr move_from = cmove->in_opr1(); 1252 LIR_Opr move_to = cmove->result_opr(); 1253 1254 if (move_to->is_register() && move_from->is_register()) { 1255 Interval* from = interval_at(reg_num(move_from)); 1256 Interval* to = interval_at(reg_num(move_to)); 1257 if (from != NULL && to != NULL) { 1258 to->set_register_hint(from); 1259 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())); 1260 } 1261 } 1262 break; 1263 } 1264 } 1265 } 1266 1267 1268 void LinearScan::build_intervals() { 1269 TIME_LINEAR_SCAN(timer_build_intervals); 1270 1271 // initialize interval list with expected number of intervals 1272 // (32 is added to have some space for split children without having to resize the list) 1273 _intervals = IntervalList(num_virtual_regs() + 32); 1274 // initialize all slots that are used by build_intervals 1275 _intervals.at_put_grow(num_virtual_regs() - 1, NULL, NULL); 1276 1277 // create a list with all caller-save registers (cpu, fpu, xmm) 1278 // when an instruction is a call, a temp range is created for all these registers 1279 int num_caller_save_registers = 0; 1280 int caller_save_registers[LinearScan::nof_regs]; 1281 1282 int i; 1283 for (i = 0; i < FrameMap::nof_caller_save_cpu_regs(); i++) { 1284 LIR_Opr opr = FrameMap::caller_save_cpu_reg_at(i); 1285 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands"); 1286 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register"); 1287 caller_save_registers[num_caller_save_registers++] = reg_num(opr); 1288 } 1289 1290 // temp ranges for fpu registers are only created when the method has 1291 // virtual fpu operands. Otherwise no allocation for fpu registers is 1292 // perfomed and so the temp ranges would be useless 1293 if (has_fpu_registers()) { 1294 #ifdef X86 1295 if (UseSSE < 2) { 1296 #endif 1297 for (i = 0; i < FrameMap::nof_caller_save_fpu_regs; i++) { 1298 LIR_Opr opr = FrameMap::caller_save_fpu_reg_at(i); 1299 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands"); 1300 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register"); 1301 caller_save_registers[num_caller_save_registers++] = reg_num(opr); 1302 } 1303 #ifdef X86 1304 } 1305 if (UseSSE > 0) { 1306 for (i = 0; i < FrameMap::nof_caller_save_xmm_regs; i++) { 1307 LIR_Opr opr = FrameMap::caller_save_xmm_reg_at(i); 1308 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands"); 1309 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register"); 1310 caller_save_registers[num_caller_save_registers++] = reg_num(opr); 1311 } 1312 } 1313 #endif 1314 } 1315 assert(num_caller_save_registers <= LinearScan::nof_regs, "out of bounds"); 1316 1317 1318 LIR_OpVisitState visitor; 1319 1320 // iterate all blocks in reverse order 1321 for (i = block_count() - 1; i >= 0; i--) { 1322 BlockBegin* block = block_at(i); 1323 LIR_OpList* instructions = block->lir()->instructions_list(); 1324 int block_from = block->first_lir_instruction_id(); 1325 int block_to = block->last_lir_instruction_id(); 1326 1327 assert(block_from == instructions->at(0)->id(), "must be"); 1328 assert(block_to == instructions->at(instructions->length() - 1)->id(), "must be"); 1329 1330 // Update intervals for registers live at the end of this block; 1331 BitMap live = block->live_out(); 1332 int size = (int)live.size(); 1333 for (int number = (int)live.get_next_one_offset(0, size); number < size; number = (int)live.get_next_one_offset(number + 1, size)) { 1334 assert(live.at(number), "should not stop here otherwise"); 1335 assert(number >= LIR_OprDesc::vreg_base, "fixed intervals must not be live on block bounds"); 1336 TRACE_LINEAR_SCAN(2, tty->print_cr("live in %d to %d", number, block_to + 2)); 1337 1338 add_use(number, block_from, block_to + 2, noUse, T_ILLEGAL); 1339 1340 // add special use positions for loop-end blocks when the 1341 // interval is used anywhere inside this loop. It's possible 1342 // that the block was part of a non-natural loop, so it might 1343 // have an invalid loop index. 1344 if (block->is_set(BlockBegin::linear_scan_loop_end_flag) && 1345 block->loop_index() != -1 && 1346 is_interval_in_loop(number, block->loop_index())) { 1347 interval_at(number)->add_use_pos(block_to + 1, loopEndMarker); 1348 } 1349 } 1350 1351 // iterate all instructions of the block in reverse order. 1352 // skip the first instruction because it is always a label 1353 // definitions of intervals are processed before uses 1354 assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label"); 1355 for (int j = instructions->length() - 1; j >= 1; j--) { 1356 LIR_Op* op = instructions->at(j); 1357 int op_id = op->id(); 1358 1359 // visit operation to collect all operands 1360 visitor.visit(op); 1361 1362 // add a temp range for each register if operation destroys caller-save registers 1363 if (visitor.has_call()) { 1364 for (int k = 0; k < num_caller_save_registers; k++) { 1365 add_temp(caller_save_registers[k], op_id, noUse, T_ILLEGAL); 1366 } 1367 TRACE_LINEAR_SCAN(4, tty->print_cr("operation destroys all caller-save registers")); 1368 } 1369 1370 // Add any platform dependent temps 1371 pd_add_temps(op); 1372 1373 // visit definitions (output and temp operands) 1374 int k, n; 1375 n = visitor.opr_count(LIR_OpVisitState::outputMode); 1376 for (k = 0; k < n; k++) { 1377 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k); 1378 assert(opr->is_register(), "visitor should only return register operands"); 1379 add_def(opr, op_id, use_kind_of_output_operand(op, opr)); 1380 } 1381 1382 n = visitor.opr_count(LIR_OpVisitState::tempMode); 1383 for (k = 0; k < n; k++) { 1384 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k); 1385 assert(opr->is_register(), "visitor should only return register operands"); 1386 add_temp(opr, op_id, mustHaveRegister); 1387 } 1388 1389 // visit uses (input operands) 1390 n = visitor.opr_count(LIR_OpVisitState::inputMode); 1391 for (k = 0; k < n; k++) { 1392 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k); 1393 assert(opr->is_register(), "visitor should only return register operands"); 1394 add_use(opr, block_from, op_id, use_kind_of_input_operand(op, opr)); 1395 } 1396 1397 // Add uses of live locals from interpreter's point of view for proper 1398 // debug information generation 1399 // Treat these operands as temp values (if the life range is extended 1400 // to a call site, the value would be in a register at the call otherwise) 1401 n = visitor.info_count(); 1402 for (k = 0; k < n; k++) { 1403 CodeEmitInfo* info = visitor.info_at(k); 1404 ValueStack* stack = info->stack(); 1405 for_each_state_value(stack, value, 1406 add_use(value, block_from, op_id + 1, noUse); 1407 ); 1408 } 1409 1410 // special steps for some instructions (especially moves) 1411 handle_method_arguments(op); 1412 handle_doubleword_moves(op); 1413 add_register_hints(op); 1414 1415 } // end of instruction iteration 1416 } // end of block iteration 1417 1418 1419 // add the range [0, 1[ to all fixed intervals 1420 // -> the register allocator need not handle unhandled fixed intervals 1421 for (int n = 0; n < LinearScan::nof_regs; n++) { 1422 Interval* interval = interval_at(n); 1423 if (interval != NULL) { 1424 interval->add_range(0, 1); 1425 } 1426 } 1427 } 1428 1429 1430 // ********** Phase 5: actual register allocation 1431 1432 int LinearScan::interval_cmp(Interval** a, Interval** b) { 1433 if (*a != NULL) { 1434 if (*b != NULL) { 1435 return (*a)->from() - (*b)->from(); 1436 } else { 1437 return -1; 1438 } 1439 } else { 1440 if (*b != NULL) { 1441 return 1; 1442 } else { 1443 return 0; 1444 } 1445 } 1446 } 1447 1448 #ifndef PRODUCT 1449 bool LinearScan::is_sorted(IntervalArray* intervals) { 1450 int from = -1; 1451 int i, j; 1452 for (i = 0; i < intervals->length(); i ++) { 1453 Interval* it = intervals->at(i); 1454 if (it != NULL) { 1455 if (from > it->from()) { 1456 assert(false, ""); 1457 return false; 1458 } 1459 from = it->from(); 1460 } 1461 } 1462 1463 // check in both directions if sorted list and unsorted list contain same intervals 1464 for (i = 0; i < interval_count(); i++) { 1465 if (interval_at(i) != NULL) { 1466 int num_found = 0; 1467 for (j = 0; j < intervals->length(); j++) { 1468 if (interval_at(i) == intervals->at(j)) { 1469 num_found++; 1470 } 1471 } 1472 assert(num_found == 1, "lists do not contain same intervals"); 1473 } 1474 } 1475 for (j = 0; j < intervals->length(); j++) { 1476 int num_found = 0; 1477 for (i = 0; i < interval_count(); i++) { 1478 if (interval_at(i) == intervals->at(j)) { 1479 num_found++; 1480 } 1481 } 1482 assert(num_found == 1, "lists do not contain same intervals"); 1483 } 1484 1485 return true; 1486 } 1487 #endif 1488 1489 void LinearScan::add_to_list(Interval** first, Interval** prev, Interval* interval) { 1490 if (*prev != NULL) { 1491 (*prev)->set_next(interval); 1492 } else { 1493 *first = interval; 1494 } 1495 *prev = interval; 1496 } 1497 1498 void LinearScan::create_unhandled_lists(Interval** list1, Interval** list2, bool (is_list1)(const Interval* i), bool (is_list2)(const Interval* i)) { 1499 assert(is_sorted(_sorted_intervals), "interval list is not sorted"); 1500 1501 *list1 = *list2 = Interval::end(); 1502 1503 Interval* list1_prev = NULL; 1504 Interval* list2_prev = NULL; 1505 Interval* v; 1506 1507 const int n = _sorted_intervals->length(); 1508 for (int i = 0; i < n; i++) { 1509 v = _sorted_intervals->at(i); 1510 if (v == NULL) continue; 1511 1512 if (is_list1(v)) { 1513 add_to_list(list1, &list1_prev, v); 1514 } else if (is_list2 == NULL || is_list2(v)) { 1515 add_to_list(list2, &list2_prev, v); 1516 } 1517 } 1518 1519 if (list1_prev != NULL) list1_prev->set_next(Interval::end()); 1520 if (list2_prev != NULL) list2_prev->set_next(Interval::end()); 1521 1522 assert(list1_prev == NULL || list1_prev->next() == Interval::end(), "linear list ends not with sentinel"); 1523 assert(list2_prev == NULL || list2_prev->next() == Interval::end(), "linear list ends not with sentinel"); 1524 } 1525 1526 1527 void LinearScan::sort_intervals_before_allocation() { 1528 TIME_LINEAR_SCAN(timer_sort_intervals_before); 1529 1530 if (_needs_full_resort) { 1531 // There is no known reason why this should occur but just in case... 1532 assert(false, "should never occur"); 1533 // Re-sort existing interval list because an Interval::from() has changed 1534 _sorted_intervals->sort(interval_cmp); 1535 _needs_full_resort = false; 1536 } 1537 1538 IntervalList* unsorted_list = &_intervals; 1539 int unsorted_len = unsorted_list->length(); 1540 int sorted_len = 0; 1541 int unsorted_idx; 1542 int sorted_idx = 0; 1543 int sorted_from_max = -1; 1544 1545 // calc number of items for sorted list (sorted list must not contain NULL values) 1546 for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) { 1547 if (unsorted_list->at(unsorted_idx) != NULL) { 1548 sorted_len++; 1549 } 1550 } 1551 IntervalArray* sorted_list = new IntervalArray(sorted_len); 1552 1553 // special sorting algorithm: the original interval-list is almost sorted, 1554 // only some intervals are swapped. So this is much faster than a complete QuickSort 1555 for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) { 1556 Interval* cur_interval = unsorted_list->at(unsorted_idx); 1557 1558 if (cur_interval != NULL) { 1559 int cur_from = cur_interval->from(); 1560 1561 if (sorted_from_max <= cur_from) { 1562 sorted_list->at_put(sorted_idx++, cur_interval); 1563 sorted_from_max = cur_interval->from(); 1564 } else { 1565 // the asumption that the intervals are already sorted failed, 1566 // so this interval must be sorted in manually 1567 int j; 1568 for (j = sorted_idx - 1; j >= 0 && cur_from < sorted_list->at(j)->from(); j--) { 1569 sorted_list->at_put(j + 1, sorted_list->at(j)); 1570 } 1571 sorted_list->at_put(j + 1, cur_interval); 1572 sorted_idx++; 1573 } 1574 } 1575 } 1576 _sorted_intervals = sorted_list; 1577 assert(is_sorted(_sorted_intervals), "intervals unsorted"); 1578 } 1579 1580 void LinearScan::sort_intervals_after_allocation() { 1581 TIME_LINEAR_SCAN(timer_sort_intervals_after); 1582 1583 if (_needs_full_resort) { 1584 // Re-sort existing interval list because an Interval::from() has changed 1585 _sorted_intervals->sort(interval_cmp); 1586 _needs_full_resort = false; 1587 } 1588 1589 IntervalArray* old_list = _sorted_intervals; 1590 IntervalList* new_list = _new_intervals_from_allocation; 1591 int old_len = old_list->length(); 1592 int new_len = new_list->length(); 1593 1594 if (new_len == 0) { 1595 // no intervals have been added during allocation, so sorted list is already up to date 1596 assert(is_sorted(_sorted_intervals), "intervals unsorted"); 1597 return; 1598 } 1599 1600 // conventional sort-algorithm for new intervals 1601 new_list->sort(interval_cmp); 1602 1603 // merge old and new list (both already sorted) into one combined list 1604 IntervalArray* combined_list = new IntervalArray(old_len + new_len); 1605 int old_idx = 0; 1606 int new_idx = 0; 1607 1608 while (old_idx + new_idx < old_len + new_len) { 1609 if (new_idx >= new_len || (old_idx < old_len && old_list->at(old_idx)->from() <= new_list->at(new_idx)->from())) { 1610 combined_list->at_put(old_idx + new_idx, old_list->at(old_idx)); 1611 old_idx++; 1612 } else { 1613 combined_list->at_put(old_idx + new_idx, new_list->at(new_idx)); 1614 new_idx++; 1615 } 1616 } 1617 1618 _sorted_intervals = combined_list; 1619 assert(is_sorted(_sorted_intervals), "intervals unsorted"); 1620 } 1621 1622 1623 void LinearScan::allocate_registers() { 1624 TIME_LINEAR_SCAN(timer_allocate_registers); 1625 1626 Interval* precolored_cpu_intervals, *not_precolored_cpu_intervals; 1627 Interval* precolored_fpu_intervals, *not_precolored_fpu_intervals; 1628 1629 create_unhandled_lists(&precolored_cpu_intervals, ¬_precolored_cpu_intervals, is_precolored_cpu_interval, is_virtual_cpu_interval); 1630 if (has_fpu_registers()) { 1631 create_unhandled_lists(&precolored_fpu_intervals, ¬_precolored_fpu_intervals, is_precolored_fpu_interval, is_virtual_fpu_interval); 1632 #ifdef ASSERT 1633 } else { 1634 // fpu register allocation is omitted because no virtual fpu registers are present 1635 // just check this again... 1636 create_unhandled_lists(&precolored_fpu_intervals, ¬_precolored_fpu_intervals, is_precolored_fpu_interval, is_virtual_fpu_interval); 1637 assert(not_precolored_fpu_intervals == Interval::end(), "missed an uncolored fpu interval"); 1638 #endif 1639 } 1640 1641 // allocate cpu registers 1642 LinearScanWalker cpu_lsw(this, precolored_cpu_intervals, not_precolored_cpu_intervals); 1643 cpu_lsw.walk(); 1644 cpu_lsw.finish_allocation(); 1645 1646 if (has_fpu_registers()) { 1647 // allocate fpu registers 1648 LinearScanWalker fpu_lsw(this, precolored_fpu_intervals, not_precolored_fpu_intervals); 1649 fpu_lsw.walk(); 1650 fpu_lsw.finish_allocation(); 1651 } 1652 } 1653 1654 1655 // ********** Phase 6: resolve data flow 1656 // (insert moves at edges between blocks if intervals have been split) 1657 1658 // wrapper for Interval::split_child_at_op_id that performs a bailout in product mode 1659 // instead of returning NULL 1660 Interval* LinearScan::split_child_at_op_id(Interval* interval, int op_id, LIR_OpVisitState::OprMode mode) { 1661 Interval* result = interval->split_child_at_op_id(op_id, mode); 1662 if (result != NULL) { 1663 return result; 1664 } 1665 1666 assert(false, "must find an interval, but do a clean bailout in product mode"); 1667 result = new Interval(LIR_OprDesc::vreg_base); 1668 result->assign_reg(0); 1669 result->set_type(T_INT); 1670 BAILOUT_("LinearScan: interval is NULL", result); 1671 } 1672 1673 1674 Interval* LinearScan::interval_at_block_begin(BlockBegin* block, int reg_num) { 1675 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds"); 1676 assert(interval_at(reg_num) != NULL, "no interval found"); 1677 1678 return split_child_at_op_id(interval_at(reg_num), block->first_lir_instruction_id(), LIR_OpVisitState::outputMode); 1679 } 1680 1681 Interval* LinearScan::interval_at_block_end(BlockBegin* block, int reg_num) { 1682 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds"); 1683 assert(interval_at(reg_num) != NULL, "no interval found"); 1684 1685 return split_child_at_op_id(interval_at(reg_num), block->last_lir_instruction_id() + 1, LIR_OpVisitState::outputMode); 1686 } 1687 1688 Interval* LinearScan::interval_at_op_id(int reg_num, int op_id) { 1689 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds"); 1690 assert(interval_at(reg_num) != NULL, "no interval found"); 1691 1692 return split_child_at_op_id(interval_at(reg_num), op_id, LIR_OpVisitState::inputMode); 1693 } 1694 1695 1696 void LinearScan::resolve_collect_mappings(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) { 1697 DEBUG_ONLY(move_resolver.check_empty()); 1698 1699 const int num_regs = num_virtual_regs(); 1700 const int size = live_set_size(); 1701 const BitMap live_at_edge = to_block->live_in(); 1702 1703 // visit all registers where the live_at_edge bit is set 1704 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)) { 1705 assert(r < num_regs, "live information set for not exisiting interval"); 1706 assert(from_block->live_out().at(r) && to_block->live_in().at(r), "interval not live at this edge"); 1707 1708 Interval* from_interval = interval_at_block_end(from_block, r); 1709 Interval* to_interval = interval_at_block_begin(to_block, r); 1710 1711 if (from_interval != to_interval && (from_interval->assigned_reg() != to_interval->assigned_reg() || from_interval->assigned_regHi() != to_interval->assigned_regHi())) { 1712 // need to insert move instruction 1713 move_resolver.add_mapping(from_interval, to_interval); 1714 } 1715 } 1716 } 1717 1718 1719 void LinearScan::resolve_find_insert_pos(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) { 1720 if (from_block->number_of_sux() <= 1) { 1721 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at end of from_block B%d", from_block->block_id())); 1722 1723 LIR_OpList* instructions = from_block->lir()->instructions_list(); 1724 LIR_OpBranch* branch = instructions->last()->as_OpBranch(); 1725 if (branch != NULL) { 1726 // insert moves before branch 1727 assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump"); 1728 move_resolver.set_insert_position(from_block->lir(), instructions->length() - 2); 1729 } else { 1730 move_resolver.set_insert_position(from_block->lir(), instructions->length() - 1); 1731 } 1732 1733 } else { 1734 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at beginning of to_block B%d", to_block->block_id())); 1735 #ifdef ASSERT 1736 assert(from_block->lir()->instructions_list()->at(0)->as_OpLabel() != NULL, "block does not start with a label"); 1737 1738 // because the number of predecessor edges matches the number of 1739 // successor edges, blocks which are reached by switch statements 1740 // may have be more than one predecessor but it will be guaranteed 1741 // that all predecessors will be the same. 1742 for (int i = 0; i < to_block->number_of_preds(); i++) { 1743 assert(from_block == to_block->pred_at(i), "all critical edges must be broken"); 1744 } 1745 #endif 1746 1747 move_resolver.set_insert_position(to_block->lir(), 0); 1748 } 1749 } 1750 1751 1752 // insert necessary moves (spilling or reloading) at edges between blocks if interval has been split 1753 void LinearScan::resolve_data_flow() { 1754 TIME_LINEAR_SCAN(timer_resolve_data_flow); 1755 1756 int num_blocks = block_count(); 1757 MoveResolver move_resolver(this); 1758 BitMap block_completed(num_blocks); block_completed.clear(); 1759 BitMap already_resolved(num_blocks); already_resolved.clear(); 1760 1761 int i; 1762 for (i = 0; i < num_blocks; i++) { 1763 BlockBegin* block = block_at(i); 1764 1765 // check if block has only one predecessor and only one successor 1766 if (block->number_of_preds() == 1 && block->number_of_sux() == 1 && block->number_of_exception_handlers() == 0) { 1767 LIR_OpList* instructions = block->lir()->instructions_list(); 1768 assert(instructions->at(0)->code() == lir_label, "block must start with label"); 1769 assert(instructions->last()->code() == lir_branch, "block with successors must end with branch"); 1770 assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block with successor must end with unconditional branch"); 1771 1772 // check if block is empty (only label and branch) 1773 if (instructions->length() == 2) { 1774 BlockBegin* pred = block->pred_at(0); 1775 BlockBegin* sux = block->sux_at(0); 1776 1777 // prevent optimization of two consecutive blocks 1778 if (!block_completed.at(pred->linear_scan_number()) && !block_completed.at(sux->linear_scan_number())) { 1779 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())); 1780 block_completed.set_bit(block->linear_scan_number()); 1781 1782 // directly resolve between pred and sux (without looking at the empty block between) 1783 resolve_collect_mappings(pred, sux, move_resolver); 1784 if (move_resolver.has_mappings()) { 1785 move_resolver.set_insert_position(block->lir(), 0); 1786 move_resolver.resolve_and_append_moves(); 1787 } 1788 } 1789 } 1790 } 1791 } 1792 1793 1794 for (i = 0; i < num_blocks; i++) { 1795 if (!block_completed.at(i)) { 1796 BlockBegin* from_block = block_at(i); 1797 already_resolved.set_from(block_completed); 1798 1799 int num_sux = from_block->number_of_sux(); 1800 for (int s = 0; s < num_sux; s++) { 1801 BlockBegin* to_block = from_block->sux_at(s); 1802 1803 // check for duplicate edges between the same blocks (can happen with switch blocks) 1804 if (!already_resolved.at(to_block->linear_scan_number())) { 1805 TRACE_LINEAR_SCAN(3, tty->print_cr("**** processing edge between B%d and B%d", from_block->block_id(), to_block->block_id())); 1806 already_resolved.set_bit(to_block->linear_scan_number()); 1807 1808 // collect all intervals that have been split between from_block and to_block 1809 resolve_collect_mappings(from_block, to_block, move_resolver); 1810 if (move_resolver.has_mappings()) { 1811 resolve_find_insert_pos(from_block, to_block, move_resolver); 1812 move_resolver.resolve_and_append_moves(); 1813 } 1814 } 1815 } 1816 } 1817 } 1818 } 1819 1820 1821 void LinearScan::resolve_exception_entry(BlockBegin* block, int reg_num, MoveResolver &move_resolver) { 1822 if (interval_at(reg_num) == NULL) { 1823 // if a phi function is never used, no interval is created -> ignore this 1824 return; 1825 } 1826 1827 Interval* interval = interval_at_block_begin(block, reg_num); 1828 int reg = interval->assigned_reg(); 1829 int regHi = interval->assigned_regHi(); 1830 1831 if ((reg < nof_regs && interval->always_in_memory()) || 1832 (use_fpu_stack_allocation() && reg >= pd_first_fpu_reg && reg <= pd_last_fpu_reg)) { 1833 // the interval is split to get a short range that is located on the stack 1834 // in the following two cases: 1835 // * the interval started in memory (e.g. method parameter), but is currently in a register 1836 // this is an optimization for exception handling that reduces the number of moves that 1837 // are necessary for resolving the states when an exception uses this exception handler 1838 // * the interval would be on the fpu stack at the begin of the exception handler 1839 // this is not allowed because of the complicated fpu stack handling on Intel 1840 1841 // range that will be spilled to memory 1842 int from_op_id = block->first_lir_instruction_id(); 1843 int to_op_id = from_op_id + 1; // short live range of length 1 1844 assert(interval->from() <= from_op_id && interval->to() >= to_op_id, 1845 "no split allowed between exception entry and first instruction"); 1846 1847 if (interval->from() != from_op_id) { 1848 // the part before from_op_id is unchanged 1849 interval = interval->split(from_op_id); 1850 interval->assign_reg(reg, regHi); 1851 append_interval(interval); 1852 } else { 1853 _needs_full_resort = true; 1854 } 1855 assert(interval->from() == from_op_id, "must be true now"); 1856 1857 Interval* spilled_part = interval; 1858 if (interval->to() != to_op_id) { 1859 // the part after to_op_id is unchanged 1860 spilled_part = interval->split_from_start(to_op_id); 1861 append_interval(spilled_part); 1862 move_resolver.add_mapping(spilled_part, interval); 1863 } 1864 assign_spill_slot(spilled_part); 1865 1866 assert(spilled_part->from() == from_op_id && spilled_part->to() == to_op_id, "just checking"); 1867 } 1868 } 1869 1870 void LinearScan::resolve_exception_entry(BlockBegin* block, MoveResolver &move_resolver) { 1871 assert(block->is_set(BlockBegin::exception_entry_flag), "should not call otherwise"); 1872 DEBUG_ONLY(move_resolver.check_empty()); 1873 1874 // visit all registers where the live_in bit is set 1875 int size = live_set_size(); 1876 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)) { 1877 resolve_exception_entry(block, r, move_resolver); 1878 } 1879 1880 // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately 1881 for_each_phi_fun(block, phi, 1882 resolve_exception_entry(block, phi->operand()->vreg_number(), move_resolver) 1883 ); 1884 1885 if (move_resolver.has_mappings()) { 1886 // insert moves after first instruction 1887 move_resolver.set_insert_position(block->lir(), 1); 1888 move_resolver.resolve_and_append_moves(); 1889 } 1890 } 1891 1892 1893 void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, int reg_num, Phi* phi, MoveResolver &move_resolver) { 1894 if (interval_at(reg_num) == NULL) { 1895 // if a phi function is never used, no interval is created -> ignore this 1896 return; 1897 } 1898 1899 // the computation of to_interval is equal to resolve_collect_mappings, 1900 // but from_interval is more complicated because of phi functions 1901 BlockBegin* to_block = handler->entry_block(); 1902 Interval* to_interval = interval_at_block_begin(to_block, reg_num); 1903 1904 if (phi != NULL) { 1905 // phi function of the exception entry block 1906 // no moves are created for this phi function in the LIR_Generator, so the 1907 // interval at the throwing instruction must be searched using the operands 1908 // of the phi function 1909 Value from_value = phi->operand_at(handler->phi_operand()); 1910 1911 // with phi functions it can happen that the same from_value is used in 1912 // multiple mappings, so notify move-resolver that this is allowed 1913 move_resolver.set_multiple_reads_allowed(); 1914 1915 Constant* con = from_value->as_Constant(); 1916 if (con != NULL && !con->is_pinned()) { 1917 // unpinned constants may have no register, so add mapping from constant to interval 1918 move_resolver.add_mapping(LIR_OprFact::value_type(con->type()), to_interval); 1919 } else { 1920 // search split child at the throwing op_id 1921 Interval* from_interval = interval_at_op_id(from_value->operand()->vreg_number(), throwing_op_id); 1922 move_resolver.add_mapping(from_interval, to_interval); 1923 } 1924 1925 } else { 1926 // no phi function, so use reg_num also for from_interval 1927 // search split child at the throwing op_id 1928 Interval* from_interval = interval_at_op_id(reg_num, throwing_op_id); 1929 if (from_interval != to_interval) { 1930 // optimization to reduce number of moves: when to_interval is on stack and 1931 // the stack slot is known to be always correct, then no move is necessary 1932 if (!from_interval->always_in_memory() || from_interval->canonical_spill_slot() != to_interval->assigned_reg()) { 1933 move_resolver.add_mapping(from_interval, to_interval); 1934 } 1935 } 1936 } 1937 } 1938 1939 void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, MoveResolver &move_resolver) { 1940 TRACE_LINEAR_SCAN(4, tty->print_cr("resolving exception handler B%d: throwing_op_id=%d", handler->entry_block()->block_id(), throwing_op_id)); 1941 1942 DEBUG_ONLY(move_resolver.check_empty()); 1943 assert(handler->lir_op_id() == -1, "already processed this xhandler"); 1944 DEBUG_ONLY(handler->set_lir_op_id(throwing_op_id)); 1945 assert(handler->entry_code() == NULL, "code already present"); 1946 1947 // visit all registers where the live_in bit is set 1948 BlockBegin* block = handler->entry_block(); 1949 int size = live_set_size(); 1950 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)) { 1951 resolve_exception_edge(handler, throwing_op_id, r, NULL, move_resolver); 1952 } 1953 1954 // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately 1955 for_each_phi_fun(block, phi, 1956 resolve_exception_edge(handler, throwing_op_id, phi->operand()->vreg_number(), phi, move_resolver) 1957 ); 1958 1959 if (move_resolver.has_mappings()) { 1960 LIR_List* entry_code = new LIR_List(compilation()); 1961 move_resolver.set_insert_position(entry_code, 0); 1962 move_resolver.resolve_and_append_moves(); 1963 1964 entry_code->jump(handler->entry_block()); 1965 handler->set_entry_code(entry_code); 1966 } 1967 } 1968 1969 1970 void LinearScan::resolve_exception_handlers() { 1971 MoveResolver move_resolver(this); 1972 LIR_OpVisitState visitor; 1973 int num_blocks = block_count(); 1974 1975 int i; 1976 for (i = 0; i < num_blocks; i++) { 1977 BlockBegin* block = block_at(i); 1978 if (block->is_set(BlockBegin::exception_entry_flag)) { 1979 resolve_exception_entry(block, move_resolver); 1980 } 1981 } 1982 1983 for (i = 0; i < num_blocks; i++) { 1984 BlockBegin* block = block_at(i); 1985 LIR_List* ops = block->lir(); 1986 int num_ops = ops->length(); 1987 1988 // iterate all instructions of the block. skip the first because it is always a label 1989 assert(visitor.no_operands(ops->at(0)), "first operation must always be a label"); 1990 for (int j = 1; j < num_ops; j++) { 1991 LIR_Op* op = ops->at(j); 1992 int op_id = op->id(); 1993 1994 if (op_id != -1 && has_info(op_id)) { 1995 // visit operation to collect all operands 1996 visitor.visit(op); 1997 assert(visitor.info_count() > 0, "should not visit otherwise"); 1998 1999 XHandlers* xhandlers = visitor.all_xhandler(); 2000 int n = xhandlers->length(); 2001 for (int k = 0; k < n; k++) { 2002 resolve_exception_edge(xhandlers->handler_at(k), op_id, move_resolver); 2003 } 2004 2005 #ifdef ASSERT 2006 } else { 2007 visitor.visit(op); 2008 assert(visitor.all_xhandler()->length() == 0, "missed exception handler"); 2009 #endif 2010 } 2011 } 2012 } 2013 } 2014 2015 2016 // ********** Phase 7: assign register numbers back to LIR 2017 // (includes computation of debug information and oop maps) 2018 2019 VMReg LinearScan::vm_reg_for_interval(Interval* interval) { 2020 VMReg reg = interval->cached_vm_reg(); 2021 if (!reg->is_valid() ) { 2022 reg = vm_reg_for_operand(operand_for_interval(interval)); 2023 interval->set_cached_vm_reg(reg); 2024 } 2025 assert(reg == vm_reg_for_operand(operand_for_interval(interval)), "wrong cached value"); 2026 return reg; 2027 } 2028 2029 VMReg LinearScan::vm_reg_for_operand(LIR_Opr opr) { 2030 assert(opr->is_oop(), "currently only implemented for oop operands"); 2031 return frame_map()->regname(opr); 2032 } 2033 2034 2035 LIR_Opr LinearScan::operand_for_interval(Interval* interval) { 2036 LIR_Opr opr = interval->cached_opr(); 2037 if (opr->is_illegal()) { 2038 opr = calc_operand_for_interval(interval); 2039 interval->set_cached_opr(opr); 2040 } 2041 2042 assert(opr == calc_operand_for_interval(interval), "wrong cached value"); 2043 return opr; 2044 } 2045 2046 LIR_Opr LinearScan::calc_operand_for_interval(const Interval* interval) { 2047 int assigned_reg = interval->assigned_reg(); 2048 BasicType type = interval->type(); 2049 2050 if (assigned_reg >= nof_regs) { 2051 // stack slot 2052 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2053 return LIR_OprFact::stack(assigned_reg - nof_regs, type); 2054 2055 } else { 2056 // register 2057 switch (type) { 2058 case T_OBJECT: { 2059 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2060 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2061 return LIR_OprFact::single_cpu_oop(assigned_reg); 2062 } 2063 2064 case T_ADDRESS: { 2065 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2066 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2067 return LIR_OprFact::single_cpu_address(assigned_reg); 2068 } 2069 2070 #ifdef __SOFTFP__ 2071 case T_FLOAT: // fall through 2072 #endif // __SOFTFP__ 2073 case T_INT: { 2074 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2075 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2076 return LIR_OprFact::single_cpu(assigned_reg); 2077 } 2078 2079 #ifdef __SOFTFP__ 2080 case T_DOUBLE: // fall through 2081 #endif // __SOFTFP__ 2082 case T_LONG: { 2083 int assigned_regHi = interval->assigned_regHi(); 2084 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2085 assert(num_physical_regs(T_LONG) == 1 || 2086 (assigned_regHi >= pd_first_cpu_reg && assigned_regHi <= pd_last_cpu_reg), "no cpu register"); 2087 2088 assert(assigned_reg != assigned_regHi, "invalid allocation"); 2089 assert(num_physical_regs(T_LONG) == 1 || assigned_reg < assigned_regHi, 2090 "register numbers must be sorted (ensure that e.g. a move from eax,ebx to ebx,eax can not occur)"); 2091 assert((assigned_regHi != any_reg) ^ (num_physical_regs(T_LONG) == 1), "must be match"); 2092 if (requires_adjacent_regs(T_LONG)) { 2093 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == assigned_regHi, "must be sequential and even"); 2094 } 2095 2096 #ifdef _LP64 2097 return LIR_OprFact::double_cpu(assigned_reg, assigned_reg); 2098 #else 2099 #if defined(SPARC) || defined(PPC) 2100 return LIR_OprFact::double_cpu(assigned_regHi, assigned_reg); 2101 #else 2102 return LIR_OprFact::double_cpu(assigned_reg, assigned_regHi); 2103 #endif // SPARC 2104 #endif // LP64 2105 } 2106 2107 #ifndef __SOFTFP__ 2108 case T_FLOAT: { 2109 #ifdef X86 2110 if (UseSSE >= 1) { 2111 assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= pd_last_xmm_reg, "no xmm register"); 2112 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2113 return LIR_OprFact::single_xmm(assigned_reg - pd_first_xmm_reg); 2114 } 2115 #endif 2116 2117 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register"); 2118 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2119 return LIR_OprFact::single_fpu(assigned_reg - pd_first_fpu_reg); 2120 } 2121 2122 case T_DOUBLE: { 2123 #ifdef X86 2124 if (UseSSE >= 2) { 2125 assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= pd_last_xmm_reg, "no xmm register"); 2126 assert(interval->assigned_regHi() == any_reg, "must not have hi register (double xmm values are stored in one register)"); 2127 return LIR_OprFact::double_xmm(assigned_reg - pd_first_xmm_reg); 2128 } 2129 #endif 2130 2131 #ifdef SPARC 2132 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register"); 2133 assert(interval->assigned_regHi() >= pd_first_fpu_reg && interval->assigned_regHi() <= pd_last_fpu_reg, "no fpu register"); 2134 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == interval->assigned_regHi(), "must be sequential and even"); 2135 LIR_Opr result = LIR_OprFact::double_fpu(interval->assigned_regHi() - pd_first_fpu_reg, assigned_reg - pd_first_fpu_reg); 2136 #elif defined(ARM) 2137 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register"); 2138 assert(interval->assigned_regHi() >= pd_first_fpu_reg && interval->assigned_regHi() <= pd_last_fpu_reg, "no fpu register"); 2139 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == interval->assigned_regHi(), "must be sequential and even"); 2140 LIR_Opr result = LIR_OprFact::double_fpu(assigned_reg - pd_first_fpu_reg, interval->assigned_regHi() - pd_first_fpu_reg); 2141 #else 2142 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register"); 2143 assert(interval->assigned_regHi() == any_reg, "must not have hi register (double fpu values are stored in one register on Intel)"); 2144 LIR_Opr result = LIR_OprFact::double_fpu(assigned_reg - pd_first_fpu_reg); 2145 #endif 2146 return result; 2147 } 2148 #endif // __SOFTFP__ 2149 2150 default: { 2151 ShouldNotReachHere(); 2152 return LIR_OprFact::illegalOpr; 2153 } 2154 } 2155 } 2156 } 2157 2158 LIR_Opr LinearScan::canonical_spill_opr(Interval* interval) { 2159 assert(interval->canonical_spill_slot() >= nof_regs, "canonical spill slot not set"); 2160 return LIR_OprFact::stack(interval->canonical_spill_slot() - nof_regs, interval->type()); 2161 } 2162 2163 LIR_Opr LinearScan::color_lir_opr(LIR_Opr opr, int op_id, LIR_OpVisitState::OprMode mode) { 2164 assert(opr->is_virtual(), "should not call this otherwise"); 2165 2166 Interval* interval = interval_at(opr->vreg_number()); 2167 assert(interval != NULL, "interval must exist"); 2168 2169 if (op_id != -1) { 2170 #ifdef ASSERT 2171 BlockBegin* block = block_of_op_with_id(op_id); 2172 if (block->number_of_sux() <= 1 && op_id == block->last_lir_instruction_id()) { 2173 // check if spill moves could have been appended at the end of this block, but 2174 // before the branch instruction. So the split child information for this branch would 2175 // be incorrect. 2176 LIR_OpBranch* branch = block->lir()->instructions_list()->last()->as_OpBranch(); 2177 if (branch != NULL) { 2178 if (block->live_out().at(opr->vreg_number())) { 2179 assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump"); 2180 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)"); 2181 } 2182 } 2183 } 2184 #endif 2185 2186 // operands are not changed when an interval is split during allocation, 2187 // so search the right interval here 2188 interval = split_child_at_op_id(interval, op_id, mode); 2189 } 2190 2191 LIR_Opr res = operand_for_interval(interval); 2192 2193 #ifdef X86 2194 // new semantic for is_last_use: not only set on definite end of interval, 2195 // but also before hole 2196 // This may still miss some cases (e.g. for dead values), but it is not necessary that the 2197 // last use information is completely correct 2198 // information is only needed for fpu stack allocation 2199 if (res->is_fpu_register()) { 2200 if (opr->is_last_use() || op_id == interval->to() || (op_id != -1 && interval->has_hole_between(op_id, op_id + 1))) { 2201 assert(op_id == -1 || !is_block_begin(op_id), "holes at begin of block may also result from control flow"); 2202 res = res->make_last_use(); 2203 } 2204 } 2205 #endif 2206 2207 assert(!gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::callee_saved) || !FrameMap::is_caller_save_register(res), "bad allocation"); 2208 2209 return res; 2210 } 2211 2212 2213 #ifdef ASSERT 2214 // some methods used to check correctness of debug information 2215 2216 void assert_no_register_values(GrowableArray<ScopeValue*>* values) { 2217 if (values == NULL) { 2218 return; 2219 } 2220 2221 for (int i = 0; i < values->length(); i++) { 2222 ScopeValue* value = values->at(i); 2223 2224 if (value->is_location()) { 2225 Location location = ((LocationValue*)value)->location(); 2226 assert(location.where() == Location::on_stack, "value is in register"); 2227 } 2228 } 2229 } 2230 2231 void assert_no_register_values(GrowableArray<MonitorValue*>* values) { 2232 if (values == NULL) { 2233 return; 2234 } 2235 2236 for (int i = 0; i < values->length(); i++) { 2237 MonitorValue* value = values->at(i); 2238 2239 if (value->owner()->is_location()) { 2240 Location location = ((LocationValue*)value->owner())->location(); 2241 assert(location.where() == Location::on_stack, "owner is in register"); 2242 } 2243 assert(value->basic_lock().where() == Location::on_stack, "basic_lock is in register"); 2244 } 2245 } 2246 2247 void assert_equal(Location l1, Location l2) { 2248 assert(l1.where() == l2.where() && l1.type() == l2.type() && l1.offset() == l2.offset(), ""); 2249 } 2250 2251 void assert_equal(ScopeValue* v1, ScopeValue* v2) { 2252 if (v1->is_location()) { 2253 assert(v2->is_location(), ""); 2254 assert_equal(((LocationValue*)v1)->location(), ((LocationValue*)v2)->location()); 2255 } else if (v1->is_constant_int()) { 2256 assert(v2->is_constant_int(), ""); 2257 assert(((ConstantIntValue*)v1)->value() == ((ConstantIntValue*)v2)->value(), ""); 2258 } else if (v1->is_constant_double()) { 2259 assert(v2->is_constant_double(), ""); 2260 assert(((ConstantDoubleValue*)v1)->value() == ((ConstantDoubleValue*)v2)->value(), ""); 2261 } else if (v1->is_constant_long()) { 2262 assert(v2->is_constant_long(), ""); 2263 assert(((ConstantLongValue*)v1)->value() == ((ConstantLongValue*)v2)->value(), ""); 2264 } else if (v1->is_constant_oop()) { 2265 assert(v2->is_constant_oop(), ""); 2266 assert(((ConstantOopWriteValue*)v1)->value() == ((ConstantOopWriteValue*)v2)->value(), ""); 2267 } else { 2268 ShouldNotReachHere(); 2269 } 2270 } 2271 2272 void assert_equal(MonitorValue* m1, MonitorValue* m2) { 2273 assert_equal(m1->owner(), m2->owner()); 2274 assert_equal(m1->basic_lock(), m2->basic_lock()); 2275 } 2276 2277 void assert_equal(IRScopeDebugInfo* d1, IRScopeDebugInfo* d2) { 2278 assert(d1->scope() == d2->scope(), "not equal"); 2279 assert(d1->bci() == d2->bci(), "not equal"); 2280 2281 if (d1->locals() != NULL) { 2282 assert(d1->locals() != NULL && d2->locals() != NULL, "not equal"); 2283 assert(d1->locals()->length() == d2->locals()->length(), "not equal"); 2284 for (int i = 0; i < d1->locals()->length(); i++) { 2285 assert_equal(d1->locals()->at(i), d2->locals()->at(i)); 2286 } 2287 } else { 2288 assert(d1->locals() == NULL && d2->locals() == NULL, "not equal"); 2289 } 2290 2291 if (d1->expressions() != NULL) { 2292 assert(d1->expressions() != NULL && d2->expressions() != NULL, "not equal"); 2293 assert(d1->expressions()->length() == d2->expressions()->length(), "not equal"); 2294 for (int i = 0; i < d1->expressions()->length(); i++) { 2295 assert_equal(d1->expressions()->at(i), d2->expressions()->at(i)); 2296 } 2297 } else { 2298 assert(d1->expressions() == NULL && d2->expressions() == NULL, "not equal"); 2299 } 2300 2301 if (d1->monitors() != NULL) { 2302 assert(d1->monitors() != NULL && d2->monitors() != NULL, "not equal"); 2303 assert(d1->monitors()->length() == d2->monitors()->length(), "not equal"); 2304 for (int i = 0; i < d1->monitors()->length(); i++) { 2305 assert_equal(d1->monitors()->at(i), d2->monitors()->at(i)); 2306 } 2307 } else { 2308 assert(d1->monitors() == NULL && d2->monitors() == NULL, "not equal"); 2309 } 2310 2311 if (d1->caller() != NULL) { 2312 assert(d1->caller() != NULL && d2->caller() != NULL, "not equal"); 2313 assert_equal(d1->caller(), d2->caller()); 2314 } else { 2315 assert(d1->caller() == NULL && d2->caller() == NULL, "not equal"); 2316 } 2317 } 2318 2319 void check_stack_depth(CodeEmitInfo* info, int stack_end) { 2320 if (info->stack()->bci() != SynchronizationEntryBCI && !info->scope()->method()->is_native()) { 2321 Bytecodes::Code code = info->scope()->method()->java_code_at_bci(info->stack()->bci()); 2322 switch (code) { 2323 case Bytecodes::_ifnull : // fall through 2324 case Bytecodes::_ifnonnull : // fall through 2325 case Bytecodes::_ifeq : // fall through 2326 case Bytecodes::_ifne : // fall through 2327 case Bytecodes::_iflt : // fall through 2328 case Bytecodes::_ifge : // fall through 2329 case Bytecodes::_ifgt : // fall through 2330 case Bytecodes::_ifle : // fall through 2331 case Bytecodes::_if_icmpeq : // fall through 2332 case Bytecodes::_if_icmpne : // fall through 2333 case Bytecodes::_if_icmplt : // fall through 2334 case Bytecodes::_if_icmpge : // fall through 2335 case Bytecodes::_if_icmpgt : // fall through 2336 case Bytecodes::_if_icmple : // fall through 2337 case Bytecodes::_if_acmpeq : // fall through 2338 case Bytecodes::_if_acmpne : 2339 assert(stack_end >= -Bytecodes::depth(code), "must have non-empty expression stack at if bytecode"); 2340 break; 2341 } 2342 } 2343 } 2344 2345 #endif // ASSERT 2346 2347 2348 IntervalWalker* LinearScan::init_compute_oop_maps() { 2349 // setup lists of potential oops for walking 2350 Interval* oop_intervals; 2351 Interval* non_oop_intervals; 2352 2353 create_unhandled_lists(&oop_intervals, &non_oop_intervals, is_oop_interval, NULL); 2354 2355 // intervals that have no oops inside need not to be processed 2356 // to ensure a walking until the last instruction id, add a dummy interval 2357 // with a high operation id 2358 non_oop_intervals = new Interval(any_reg); 2359 non_oop_intervals->add_range(max_jint - 2, max_jint - 1); 2360 2361 return new IntervalWalker(this, oop_intervals, non_oop_intervals); 2362 } 2363 2364 2365 OopMap* LinearScan::compute_oop_map(IntervalWalker* iw, LIR_Op* op, CodeEmitInfo* info, bool is_call_site) { 2366 TRACE_LINEAR_SCAN(3, tty->print_cr("creating oop map at op_id %d", op->id())); 2367 2368 // walk before the current operation -> intervals that start at 2369 // the operation (= output operands of the operation) are not 2370 // included in the oop map 2371 iw->walk_before(op->id()); 2372 2373 int frame_size = frame_map()->framesize(); 2374 int arg_count = frame_map()->oop_map_arg_count(); 2375 OopMap* map = new OopMap(frame_size, arg_count); 2376 2377 // Check if this is a patch site. 2378 bool is_patch_info = false; 2379 if (op->code() == lir_move) { 2380 assert(!is_call_site, "move must not be a call site"); 2381 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); 2382 LIR_Op1* move = (LIR_Op1*)op; 2383 2384 is_patch_info = move->patch_code() != lir_patch_none; 2385 } 2386 2387 // Iterate through active intervals 2388 for (Interval* interval = iw->active_first(fixedKind); interval != Interval::end(); interval = interval->next()) { 2389 int assigned_reg = interval->assigned_reg(); 2390 2391 assert(interval->current_from() <= op->id() && op->id() <= interval->current_to(), "interval should not be active otherwise"); 2392 assert(interval->assigned_regHi() == any_reg, "oop must be single word"); 2393 assert(interval->reg_num() >= LIR_OprDesc::vreg_base, "fixed interval found"); 2394 2395 // Check if this range covers the instruction. Intervals that 2396 // start or end at the current operation are not included in the 2397 // oop map, except in the case of patching moves. For patching 2398 // moves, any intervals which end at this instruction are included 2399 // in the oop map since we may safepoint while doing the patch 2400 // before we've consumed the inputs. 2401 if (is_patch_info || op->id() < interval->current_to()) { 2402 2403 // caller-save registers must not be included into oop-maps at calls 2404 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"); 2405 2406 VMReg name = vm_reg_for_interval(interval); 2407 set_oop(map, name); 2408 2409 // Spill optimization: when the stack value is guaranteed to be always correct, 2410 // then it must be added to the oop map even if the interval is currently in a register 2411 if (interval->always_in_memory() && 2412 op->id() > interval->spill_definition_pos() && 2413 interval->assigned_reg() != interval->canonical_spill_slot()) { 2414 assert(interval->spill_definition_pos() > 0, "position not set correctly"); 2415 assert(interval->canonical_spill_slot() >= LinearScan::nof_regs, "no spill slot assigned"); 2416 assert(interval->assigned_reg() < LinearScan::nof_regs, "interval is on stack, so stack slot is registered twice"); 2417 2418 set_oop(map, frame_map()->slot_regname(interval->canonical_spill_slot() - LinearScan::nof_regs)); 2419 } 2420 } 2421 } 2422 2423 // add oops from lock stack 2424 assert(info->stack() != NULL, "CodeEmitInfo must always have a stack"); 2425 int locks_count = info->stack()->total_locks_size(); 2426 for (int i = 0; i < locks_count; i++) { 2427 set_oop(map, frame_map()->monitor_object_regname(i)); 2428 } 2429 2430 return map; 2431 } 2432 2433 2434 void LinearScan::compute_oop_map(IntervalWalker* iw, const LIR_OpVisitState &visitor, LIR_Op* op) { 2435 assert(visitor.info_count() > 0, "no oop map needed"); 2436 2437 // compute oop_map only for first CodeEmitInfo 2438 // because it is (in most cases) equal for all other infos of the same operation 2439 CodeEmitInfo* first_info = visitor.info_at(0); 2440 OopMap* first_oop_map = compute_oop_map(iw, op, first_info, visitor.has_call()); 2441 2442 for (int i = 0; i < visitor.info_count(); i++) { 2443 CodeEmitInfo* info = visitor.info_at(i); 2444 OopMap* oop_map = first_oop_map; 2445 2446 if (info->stack()->locks_size() != first_info->stack()->locks_size()) { 2447 // this info has a different number of locks then the precomputed oop map 2448 // (possible for lock and unlock instructions) -> compute oop map with 2449 // correct lock information 2450 oop_map = compute_oop_map(iw, op, info, visitor.has_call()); 2451 } 2452 2453 if (info->_oop_map == NULL) { 2454 info->_oop_map = oop_map; 2455 } else { 2456 // a CodeEmitInfo can not be shared between different LIR-instructions 2457 // because interval splitting can occur anywhere between two instructions 2458 // and so the oop maps must be different 2459 // -> check if the already set oop_map is exactly the one calculated for this operation 2460 assert(info->_oop_map == oop_map, "same CodeEmitInfo used for multiple LIR instructions"); 2461 } 2462 } 2463 } 2464 2465 2466 // frequently used constants 2467 // Allocate them with new so they are never destroyed (otherwise, a 2468 // forced exit could destroy these objects while they are still in 2469 // use). 2470 ConstantOopWriteValue* LinearScan::_oop_null_scope_value = new (ResourceObj::C_HEAP) ConstantOopWriteValue(NULL); 2471 ConstantIntValue* LinearScan::_int_m1_scope_value = new (ResourceObj::C_HEAP) ConstantIntValue(-1); 2472 ConstantIntValue* LinearScan::_int_0_scope_value = new (ResourceObj::C_HEAP) ConstantIntValue(0); 2473 ConstantIntValue* LinearScan::_int_1_scope_value = new (ResourceObj::C_HEAP) ConstantIntValue(1); 2474 ConstantIntValue* LinearScan::_int_2_scope_value = new (ResourceObj::C_HEAP) ConstantIntValue(2); 2475 LocationValue* _illegal_value = new (ResourceObj::C_HEAP) LocationValue(Location()); 2476 2477 void LinearScan::init_compute_debug_info() { 2478 // cache for frequently used scope values 2479 // (cpu registers and stack slots) 2480 _scope_value_cache = ScopeValueArray((LinearScan::nof_cpu_regs + frame_map()->argcount() + max_spills()) * 2, NULL); 2481 } 2482 2483 MonitorValue* LinearScan::location_for_monitor_index(int monitor_index) { 2484 Location loc; 2485 if (!frame_map()->location_for_monitor_object(monitor_index, &loc)) { 2486 bailout("too large frame"); 2487 } 2488 ScopeValue* object_scope_value = new LocationValue(loc); 2489 2490 if (!frame_map()->location_for_monitor_lock(monitor_index, &loc)) { 2491 bailout("too large frame"); 2492 } 2493 return new MonitorValue(object_scope_value, loc); 2494 } 2495 2496 LocationValue* LinearScan::location_for_name(int name, Location::Type loc_type) { 2497 Location loc; 2498 if (!frame_map()->locations_for_slot(name, loc_type, &loc)) { 2499 bailout("too large frame"); 2500 } 2501 return new LocationValue(loc); 2502 } 2503 2504 2505 int LinearScan::append_scope_value_for_constant(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) { 2506 assert(opr->is_constant(), "should not be called otherwise"); 2507 2508 LIR_Const* c = opr->as_constant_ptr(); 2509 BasicType t = c->type(); 2510 switch (t) { 2511 case T_OBJECT: { 2512 jobject value = c->as_jobject(); 2513 if (value == NULL) { 2514 scope_values->append(_oop_null_scope_value); 2515 } else { 2516 scope_values->append(new ConstantOopWriteValue(c->as_jobject())); 2517 } 2518 return 1; 2519 } 2520 2521 case T_INT: // fall through 2522 case T_FLOAT: { 2523 int value = c->as_jint_bits(); 2524 switch (value) { 2525 case -1: scope_values->append(_int_m1_scope_value); break; 2526 case 0: scope_values->append(_int_0_scope_value); break; 2527 case 1: scope_values->append(_int_1_scope_value); break; 2528 case 2: scope_values->append(_int_2_scope_value); break; 2529 default: scope_values->append(new ConstantIntValue(c->as_jint_bits())); break; 2530 } 2531 return 1; 2532 } 2533 2534 case T_LONG: // fall through 2535 case T_DOUBLE: { 2536 #ifdef _LP64 2537 scope_values->append(_int_0_scope_value); 2538 scope_values->append(new ConstantLongValue(c->as_jlong_bits())); 2539 #else 2540 if (hi_word_offset_in_bytes > lo_word_offset_in_bytes) { 2541 scope_values->append(new ConstantIntValue(c->as_jint_hi_bits())); 2542 scope_values->append(new ConstantIntValue(c->as_jint_lo_bits())); 2543 } else { 2544 scope_values->append(new ConstantIntValue(c->as_jint_lo_bits())); 2545 scope_values->append(new ConstantIntValue(c->as_jint_hi_bits())); 2546 } 2547 #endif 2548 return 2; 2549 } 2550 2551 case T_ADDRESS: { 2552 #ifdef _LP64 2553 scope_values->append(new ConstantLongValue(c->as_jint())); 2554 #else 2555 scope_values->append(new ConstantIntValue(c->as_jint())); 2556 #endif 2557 return 1; 2558 } 2559 2560 default: 2561 ShouldNotReachHere(); 2562 return -1; 2563 } 2564 } 2565 2566 int LinearScan::append_scope_value_for_operand(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) { 2567 if (opr->is_single_stack()) { 2568 int stack_idx = opr->single_stack_ix(); 2569 bool is_oop = opr->is_oop_register(); 2570 int cache_idx = (stack_idx + LinearScan::nof_cpu_regs) * 2 + (is_oop ? 1 : 0); 2571 2572 ScopeValue* sv = _scope_value_cache.at(cache_idx); 2573 if (sv == NULL) { 2574 Location::Type loc_type = is_oop ? Location::oop : Location::normal; 2575 sv = location_for_name(stack_idx, loc_type); 2576 _scope_value_cache.at_put(cache_idx, sv); 2577 } 2578 2579 // check if cached value is correct 2580 DEBUG_ONLY(assert_equal(sv, location_for_name(stack_idx, is_oop ? Location::oop : Location::normal))); 2581 2582 scope_values->append(sv); 2583 return 1; 2584 2585 } else if (opr->is_single_cpu()) { 2586 bool is_oop = opr->is_oop_register(); 2587 int cache_idx = opr->cpu_regnr() * 2 + (is_oop ? 1 : 0); 2588 Location::Type int_loc_type = NOT_LP64(Location::normal) LP64_ONLY(Location::int_in_long); 2589 2590 ScopeValue* sv = _scope_value_cache.at(cache_idx); 2591 if (sv == NULL) { 2592 Location::Type loc_type = is_oop ? Location::oop : int_loc_type; 2593 VMReg rname = frame_map()->regname(opr); 2594 sv = new LocationValue(Location::new_reg_loc(loc_type, rname)); 2595 _scope_value_cache.at_put(cache_idx, sv); 2596 } 2597 2598 // check if cached value is correct 2599 DEBUG_ONLY(assert_equal(sv, new LocationValue(Location::new_reg_loc(is_oop ? Location::oop : int_loc_type, frame_map()->regname(opr))))); 2600 2601 scope_values->append(sv); 2602 return 1; 2603 2604 #ifdef X86 2605 } else if (opr->is_single_xmm()) { 2606 VMReg rname = opr->as_xmm_float_reg()->as_VMReg(); 2607 LocationValue* sv = new LocationValue(Location::new_reg_loc(Location::normal, rname)); 2608 2609 scope_values->append(sv); 2610 return 1; 2611 #endif 2612 2613 } else if (opr->is_single_fpu()) { 2614 #ifdef X86 2615 // the exact location of fpu stack values is only known 2616 // during fpu stack allocation, so the stack allocator object 2617 // must be present 2618 assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)"); 2619 assert(_fpu_stack_allocator != NULL, "must be present"); 2620 opr = _fpu_stack_allocator->to_fpu_stack(opr); 2621 #endif 2622 2623 Location::Type loc_type = float_saved_as_double ? Location::float_in_dbl : Location::normal; 2624 VMReg rname = frame_map()->fpu_regname(opr->fpu_regnr()); 2625 #ifndef __SOFTFP__ 2626 #ifndef VM_LITTLE_ENDIAN 2627 if (! float_saved_as_double) { 2628 // On big endian system, we may have an issue if float registers use only 2629 // the low half of the (same) double registers. 2630 // Both the float and the double could have the same regnr but would correspond 2631 // to two different addresses once saved. 2632 2633 // get next safely (no assertion checks) 2634 VMReg next = VMRegImpl::as_VMReg(1+rname->value()); 2635 if (next->is_reg() && 2636 (next->as_FloatRegister() == rname->as_FloatRegister())) { 2637 // the back-end does use the same numbering for the double and the float 2638 rname = next; // VMReg for the low bits, e.g. the real VMReg for the float 2639 } 2640 } 2641 #endif 2642 #endif 2643 LocationValue* sv = new LocationValue(Location::new_reg_loc(loc_type, rname)); 2644 2645 scope_values->append(sv); 2646 return 1; 2647 2648 } else { 2649 // double-size operands 2650 2651 ScopeValue* first; 2652 ScopeValue* second; 2653 2654 if (opr->is_double_stack()) { 2655 #ifdef _LP64 2656 Location loc1; 2657 Location::Type loc_type = opr->type() == T_LONG ? Location::lng : Location::dbl; 2658 if (!frame_map()->locations_for_slot(opr->double_stack_ix(), loc_type, &loc1, NULL)) { 2659 bailout("too large frame"); 2660 } 2661 // Does this reverse on x86 vs. sparc? 2662 first = new LocationValue(loc1); 2663 second = _int_0_scope_value; 2664 #else 2665 Location loc1, loc2; 2666 if (!frame_map()->locations_for_slot(opr->double_stack_ix(), Location::normal, &loc1, &loc2)) { 2667 bailout("too large frame"); 2668 } 2669 first = new LocationValue(loc1); 2670 second = new LocationValue(loc2); 2671 #endif // _LP64 2672 2673 } else if (opr->is_double_cpu()) { 2674 #ifdef _LP64 2675 VMReg rname_first = opr->as_register_lo()->as_VMReg(); 2676 first = new LocationValue(Location::new_reg_loc(Location::lng, rname_first)); 2677 second = _int_0_scope_value; 2678 #else 2679 VMReg rname_first = opr->as_register_lo()->as_VMReg(); 2680 VMReg rname_second = opr->as_register_hi()->as_VMReg(); 2681 2682 if (hi_word_offset_in_bytes < lo_word_offset_in_bytes) { 2683 // lo/hi and swapped relative to first and second, so swap them 2684 VMReg tmp = rname_first; 2685 rname_first = rname_second; 2686 rname_second = tmp; 2687 } 2688 2689 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); 2690 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); 2691 #endif //_LP64 2692 2693 2694 #ifdef X86 2695 } else if (opr->is_double_xmm()) { 2696 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation"); 2697 VMReg rname_first = opr->as_xmm_double_reg()->as_VMReg(); 2698 # ifdef _LP64 2699 first = new LocationValue(Location::new_reg_loc(Location::dbl, rname_first)); 2700 second = _int_0_scope_value; 2701 # else 2702 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); 2703 // %%% This is probably a waste but we'll keep things as they were for now 2704 if (true) { 2705 VMReg rname_second = rname_first->next(); 2706 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); 2707 } 2708 # endif 2709 #endif 2710 2711 } else if (opr->is_double_fpu()) { 2712 // On SPARC, fpu_regnrLo/fpu_regnrHi represents the two halves of 2713 // the double as float registers in the native ordering. On X86, 2714 // fpu_regnrLo is a FPU stack slot whose VMReg represents 2715 // the low-order word of the double and fpu_regnrLo + 1 is the 2716 // name for the other half. *first and *second must represent the 2717 // least and most significant words, respectively. 2718 2719 #ifdef X86 2720 // the exact location of fpu stack values is only known 2721 // during fpu stack allocation, so the stack allocator object 2722 // must be present 2723 assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)"); 2724 assert(_fpu_stack_allocator != NULL, "must be present"); 2725 opr = _fpu_stack_allocator->to_fpu_stack(opr); 2726 2727 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation (only fpu_regnrLo is used)"); 2728 #endif 2729 #ifdef SPARC 2730 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi() + 1, "assumed in calculation (only fpu_regnrHi is used)"); 2731 #endif 2732 #ifdef ARM 2733 assert(opr->fpu_regnrHi() == opr->fpu_regnrLo() + 1, "assumed in calculation (only fpu_regnrLo is used)"); 2734 #endif 2735 #ifdef PPC 2736 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation (only fpu_regnrHi is used)"); 2737 #endif 2738 2739 #ifdef VM_LITTLE_ENDIAN 2740 VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrLo()); 2741 #else 2742 VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrHi()); 2743 #endif 2744 2745 #ifdef _LP64 2746 first = new LocationValue(Location::new_reg_loc(Location::dbl, rname_first)); 2747 second = _int_0_scope_value; 2748 #else 2749 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); 2750 // %%% This is probably a waste but we'll keep things as they were for now 2751 if (true) { 2752 VMReg rname_second = rname_first->next(); 2753 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); 2754 } 2755 #endif 2756 2757 } else { 2758 ShouldNotReachHere(); 2759 first = NULL; 2760 second = NULL; 2761 } 2762 2763 assert(first != NULL && second != NULL, "must be set"); 2764 // The convention the interpreter uses is that the second local 2765 // holds the first raw word of the native double representation. 2766 // This is actually reasonable, since locals and stack arrays 2767 // grow downwards in all implementations. 2768 // (If, on some machine, the interpreter's Java locals or stack 2769 // were to grow upwards, the embedded doubles would be word-swapped.) 2770 scope_values->append(second); 2771 scope_values->append(first); 2772 return 2; 2773 } 2774 } 2775 2776 2777 int LinearScan::append_scope_value(int op_id, Value value, GrowableArray<ScopeValue*>* scope_values) { 2778 if (value != NULL) { 2779 LIR_Opr opr = value->operand(); 2780 Constant* con = value->as_Constant(); 2781 2782 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)"); 2783 assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands"); 2784 2785 if (con != NULL && !con->is_pinned() && !opr->is_constant()) { 2786 // Unpinned constants may have a virtual operand for a part of the lifetime 2787 // or may be illegal when it was optimized away, 2788 // so always use a constant operand 2789 opr = LIR_OprFact::value_type(con->type()); 2790 } 2791 assert(opr->is_virtual() || opr->is_constant(), "other cases not allowed here"); 2792 2793 if (opr->is_virtual()) { 2794 LIR_OpVisitState::OprMode mode = LIR_OpVisitState::inputMode; 2795 2796 BlockBegin* block = block_of_op_with_id(op_id); 2797 if (block->number_of_sux() == 1 && op_id == block->last_lir_instruction_id()) { 2798 // generating debug information for the last instruction of a block. 2799 // if this instruction is a branch, spill moves are inserted before this branch 2800 // and so the wrong operand would be returned (spill moves at block boundaries are not 2801 // considered in the live ranges of intervals) 2802 // Solution: use the first op_id of the branch target block instead. 2803 if (block->lir()->instructions_list()->last()->as_OpBranch() != NULL) { 2804 if (block->live_out().at(opr->vreg_number())) { 2805 op_id = block->sux_at(0)->first_lir_instruction_id(); 2806 mode = LIR_OpVisitState::outputMode; 2807 } 2808 } 2809 } 2810 2811 // Get current location of operand 2812 // The operand must be live because debug information is considered when building the intervals 2813 // if the interval is not live, color_lir_opr will cause an assertion failure 2814 opr = color_lir_opr(opr, op_id, mode); 2815 assert(!has_call(op_id) || opr->is_stack() || !is_caller_save(reg_num(opr)), "can not have caller-save register operands at calls"); 2816 2817 // Append to ScopeValue array 2818 return append_scope_value_for_operand(opr, scope_values); 2819 2820 } else { 2821 assert(value->as_Constant() != NULL, "all other instructions have only virtual operands"); 2822 assert(opr->is_constant(), "operand must be constant"); 2823 2824 return append_scope_value_for_constant(opr, scope_values); 2825 } 2826 } else { 2827 // append a dummy value because real value not needed 2828 scope_values->append(_illegal_value); 2829 return 1; 2830 } 2831 } 2832 2833 2834 IRScopeDebugInfo* LinearScan::compute_debug_info_for_scope(int op_id, IRScope* cur_scope, ValueStack* cur_state, ValueStack* innermost_state) { 2835 IRScopeDebugInfo* caller_debug_info = NULL; 2836 2837 ValueStack* caller_state = cur_state->caller_state(); 2838 if (caller_state != NULL) { 2839 // process recursively to compute outermost scope first 2840 caller_debug_info = compute_debug_info_for_scope(op_id, cur_scope->caller(), caller_state, innermost_state); 2841 } 2842 2843 // initialize these to null. 2844 // If we don't need deopt info or there are no locals, expressions or monitors, 2845 // then these get recorded as no information and avoids the allocation of 0 length arrays. 2846 GrowableArray<ScopeValue*>* locals = NULL; 2847 GrowableArray<ScopeValue*>* expressions = NULL; 2848 GrowableArray<MonitorValue*>* monitors = NULL; 2849 2850 // describe local variable values 2851 int nof_locals = cur_state->locals_size(); 2852 if (nof_locals > 0) { 2853 locals = new GrowableArray<ScopeValue*>(nof_locals); 2854 2855 int pos = 0; 2856 while (pos < nof_locals) { 2857 assert(pos < cur_state->locals_size(), "why not?"); 2858 2859 Value local = cur_state->local_at(pos); 2860 pos += append_scope_value(op_id, local, locals); 2861 2862 assert(locals->length() == pos, "must match"); 2863 } 2864 assert(locals->length() == cur_scope->method()->max_locals(), "wrong number of locals"); 2865 assert(locals->length() == cur_state->locals_size(), "wrong number of locals"); 2866 } else if (cur_scope->method()->max_locals() > 0) { 2867 assert(cur_state->kind() == ValueStack::EmptyExceptionState, "should be"); 2868 nof_locals = cur_scope->method()->max_locals(); 2869 locals = new GrowableArray<ScopeValue*>(nof_locals); 2870 for(int i = 0; i < nof_locals; i++) { 2871 locals->append(_illegal_value); 2872 } 2873 } 2874 2875 // describe expression stack 2876 int nof_stack = cur_state->stack_size(); 2877 if (nof_stack > 0) { 2878 expressions = new GrowableArray<ScopeValue*>(nof_stack); 2879 2880 int pos = 0; 2881 while (pos < nof_stack) { 2882 Value expression = cur_state->stack_at_inc(pos); 2883 append_scope_value(op_id, expression, expressions); 2884 2885 assert(expressions->length() == pos, "must match"); 2886 } 2887 assert(expressions->length() == cur_state->stack_size(), "wrong number of stack entries"); 2888 } 2889 2890 // describe monitors 2891 int nof_locks = cur_state->locks_size(); 2892 if (nof_locks > 0) { 2893 int lock_offset = cur_state->caller_state() != NULL ? cur_state->caller_state()->total_locks_size() : 0; 2894 monitors = new GrowableArray<MonitorValue*>(nof_locks); 2895 for (int i = 0; i < nof_locks; i++) { 2896 monitors->append(location_for_monitor_index(lock_offset + i)); 2897 } 2898 } 2899 2900 return new IRScopeDebugInfo(cur_scope, cur_state->bci(), locals, expressions, monitors, caller_debug_info); 2901 } 2902 2903 2904 void LinearScan::compute_debug_info(CodeEmitInfo* info, int op_id) { 2905 TRACE_LINEAR_SCAN(3, tty->print_cr("creating debug information at op_id %d", op_id)); 2906 2907 IRScope* innermost_scope = info->scope(); 2908 ValueStack* innermost_state = info->stack(); 2909 2910 assert(innermost_scope != NULL && innermost_state != NULL, "why is it missing?"); 2911 2912 DEBUG_ONLY(check_stack_depth(info, innermost_state->stack_size())); 2913 2914 if (info->_scope_debug_info == NULL) { 2915 // compute debug information 2916 info->_scope_debug_info = compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state); 2917 } else { 2918 // debug information already set. Check that it is correct from the current point of view 2919 DEBUG_ONLY(assert_equal(info->_scope_debug_info, compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state))); 2920 } 2921 } 2922 2923 2924 void LinearScan::assign_reg_num(LIR_OpList* instructions, IntervalWalker* iw) { 2925 LIR_OpVisitState visitor; 2926 int num_inst = instructions->length(); 2927 bool has_dead = false; 2928 2929 for (int j = 0; j < num_inst; j++) { 2930 LIR_Op* op = instructions->at(j); 2931 if (op == NULL) { // this can happen when spill-moves are removed in eliminate_spill_moves 2932 has_dead = true; 2933 continue; 2934 } 2935 int op_id = op->id(); 2936 2937 // visit instruction to get list of operands 2938 visitor.visit(op); 2939 2940 // iterate all modes of the visitor and process all virtual operands 2941 for_each_visitor_mode(mode) { 2942 int n = visitor.opr_count(mode); 2943 for (int k = 0; k < n; k++) { 2944 LIR_Opr opr = visitor.opr_at(mode, k); 2945 if (opr->is_virtual_register()) { 2946 visitor.set_opr_at(mode, k, color_lir_opr(opr, op_id, mode)); 2947 } 2948 } 2949 } 2950 2951 if (visitor.info_count() > 0) { 2952 // exception handling 2953 if (compilation()->has_exception_handlers()) { 2954 XHandlers* xhandlers = visitor.all_xhandler(); 2955 int n = xhandlers->length(); 2956 for (int k = 0; k < n; k++) { 2957 XHandler* handler = xhandlers->handler_at(k); 2958 if (handler->entry_code() != NULL) { 2959 assign_reg_num(handler->entry_code()->instructions_list(), NULL); 2960 } 2961 } 2962 } else { 2963 assert(visitor.all_xhandler()->length() == 0, "missed exception handler"); 2964 } 2965 2966 // compute oop map 2967 assert(iw != NULL, "needed for compute_oop_map"); 2968 compute_oop_map(iw, visitor, op); 2969 2970 // compute debug information 2971 if (!use_fpu_stack_allocation()) { 2972 // compute debug information if fpu stack allocation is not needed. 2973 // when fpu stack allocation is needed, the debug information can not 2974 // be computed here because the exact location of fpu operands is not known 2975 // -> debug information is created inside the fpu stack allocator 2976 int n = visitor.info_count(); 2977 for (int k = 0; k < n; k++) { 2978 compute_debug_info(visitor.info_at(k), op_id); 2979 } 2980 } 2981 } 2982 2983 #ifdef ASSERT 2984 // make sure we haven't made the op invalid. 2985 op->verify(); 2986 #endif 2987 2988 // remove useless moves 2989 if (op->code() == lir_move) { 2990 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); 2991 LIR_Op1* move = (LIR_Op1*)op; 2992 LIR_Opr src = move->in_opr(); 2993 LIR_Opr dst = move->result_opr(); 2994 if (dst == src || 2995 !dst->is_pointer() && !src->is_pointer() && 2996 src->is_same_register(dst)) { 2997 instructions->at_put(j, NULL); 2998 has_dead = true; 2999 } 3000 } 3001 } 3002 3003 if (has_dead) { 3004 // iterate all instructions of the block and remove all null-values. 3005 int insert_point = 0; 3006 for (int j = 0; j < num_inst; j++) { 3007 LIR_Op* op = instructions->at(j); 3008 if (op != NULL) { 3009 if (insert_point != j) { 3010 instructions->at_put(insert_point, op); 3011 } 3012 insert_point++; 3013 } 3014 } 3015 instructions->truncate(insert_point); 3016 } 3017 } 3018 3019 void LinearScan::assign_reg_num() { 3020 TIME_LINEAR_SCAN(timer_assign_reg_num); 3021 3022 init_compute_debug_info(); 3023 IntervalWalker* iw = init_compute_oop_maps(); 3024 3025 int num_blocks = block_count(); 3026 for (int i = 0; i < num_blocks; i++) { 3027 BlockBegin* block = block_at(i); 3028 assign_reg_num(block->lir()->instructions_list(), iw); 3029 } 3030 } 3031 3032 3033 void LinearScan::do_linear_scan() { 3034 NOT_PRODUCT(_total_timer.begin_method()); 3035 3036 number_instructions(); 3037 3038 NOT_PRODUCT(print_lir(1, "Before Register Allocation")); 3039 3040 compute_local_live_sets(); 3041 compute_global_live_sets(); 3042 CHECK_BAILOUT(); 3043 3044 build_intervals(); 3045 CHECK_BAILOUT(); 3046 sort_intervals_before_allocation(); 3047 3048 NOT_PRODUCT(print_intervals("Before Register Allocation")); 3049 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_before_alloc)); 3050 3051 allocate_registers(); 3052 CHECK_BAILOUT(); 3053 3054 resolve_data_flow(); 3055 if (compilation()->has_exception_handlers()) { 3056 resolve_exception_handlers(); 3057 } 3058 // fill in number of spill slots into frame_map 3059 propagate_spill_slots(); 3060 CHECK_BAILOUT(); 3061 3062 NOT_PRODUCT(print_intervals("After Register Allocation")); 3063 NOT_PRODUCT(print_lir(2, "LIR after register allocation:")); 3064 3065 sort_intervals_after_allocation(); 3066 3067 DEBUG_ONLY(verify()); 3068 3069 eliminate_spill_moves(); 3070 assign_reg_num(); 3071 CHECK_BAILOUT(); 3072 3073 NOT_PRODUCT(print_lir(2, "LIR after assignment of register numbers:")); 3074 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_after_asign)); 3075 3076 { TIME_LINEAR_SCAN(timer_allocate_fpu_stack); 3077 3078 if (use_fpu_stack_allocation()) { 3079 allocate_fpu_stack(); // Only has effect on Intel 3080 NOT_PRODUCT(print_lir(2, "LIR after FPU stack allocation:")); 3081 } 3082 } 3083 3084 { TIME_LINEAR_SCAN(timer_optimize_lir); 3085 3086 EdgeMoveOptimizer::optimize(ir()->code()); 3087 ControlFlowOptimizer::optimize(ir()->code()); 3088 // check that cfg is still correct after optimizations 3089 ir()->verify(); 3090 } 3091 3092 NOT_PRODUCT(print_lir(1, "Before Code Generation", false)); 3093 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_final)); 3094 NOT_PRODUCT(_total_timer.end_method(this)); 3095 } 3096 3097 3098 // ********** Printing functions 3099 3100 #ifndef PRODUCT 3101 3102 void LinearScan::print_timers(double total) { 3103 _total_timer.print(total); 3104 } 3105 3106 void LinearScan::print_statistics() { 3107 _stat_before_alloc.print("before allocation"); 3108 _stat_after_asign.print("after assignment of register"); 3109 _stat_final.print("after optimization"); 3110 } 3111 3112 void LinearScan::print_bitmap(BitMap& b) { 3113 for (unsigned int i = 0; i < b.size(); i++) { 3114 if (b.at(i)) tty->print("%d ", i); 3115 } 3116 tty->cr(); 3117 } 3118 3119 void LinearScan::print_intervals(const char* label) { 3120 if (TraceLinearScanLevel >= 1) { 3121 int i; 3122 tty->cr(); 3123 tty->print_cr("%s", label); 3124 3125 for (i = 0; i < interval_count(); i++) { 3126 Interval* interval = interval_at(i); 3127 if (interval != NULL) { 3128 interval->print(); 3129 } 3130 } 3131 3132 tty->cr(); 3133 tty->print_cr("--- Basic Blocks ---"); 3134 for (i = 0; i < block_count(); i++) { 3135 BlockBegin* block = block_at(i); 3136 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()); 3137 } 3138 tty->cr(); 3139 tty->cr(); 3140 } 3141 3142 if (PrintCFGToFile) { 3143 CFGPrinter::print_intervals(&_intervals, label); 3144 } 3145 } 3146 3147 void LinearScan::print_lir(int level, const char* label, bool hir_valid) { 3148 if (TraceLinearScanLevel >= level) { 3149 tty->cr(); 3150 tty->print_cr("%s", label); 3151 print_LIR(ir()->linear_scan_order()); 3152 tty->cr(); 3153 } 3154 3155 if (level == 1 && PrintCFGToFile) { 3156 CFGPrinter::print_cfg(ir()->linear_scan_order(), label, hir_valid, true); 3157 } 3158 } 3159 3160 #endif //PRODUCT 3161 3162 3163 // ********** verification functions for allocation 3164 // (check that all intervals have a correct register and that no registers are overwritten) 3165 #ifdef ASSERT 3166 3167 void LinearScan::verify() { 3168 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying intervals ******************************************")); 3169 verify_intervals(); 3170 3171 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that no oops are in fixed intervals ****************")); 3172 verify_no_oops_in_fixed_intervals(); 3173 3174 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that unpinned constants are not alive across block boundaries")); 3175 verify_constants(); 3176 3177 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying register allocation ********************************")); 3178 verify_registers(); 3179 3180 TRACE_LINEAR_SCAN(2, tty->print_cr("********* no errors found **********************************************")); 3181 } 3182 3183 void LinearScan::verify_intervals() { 3184 int len = interval_count(); 3185 bool has_error = false; 3186 3187 for (int i = 0; i < len; i++) { 3188 Interval* i1 = interval_at(i); 3189 if (i1 == NULL) continue; 3190 3191 i1->check_split_children(); 3192 3193 if (i1->reg_num() != i) { 3194 tty->print_cr("Interval %d is on position %d in list", i1->reg_num(), i); i1->print(); tty->cr(); 3195 has_error = true; 3196 } 3197 3198 if (i1->reg_num() >= LIR_OprDesc::vreg_base && i1->type() == T_ILLEGAL) { 3199 tty->print_cr("Interval %d has no type assigned", i1->reg_num()); i1->print(); tty->cr(); 3200 has_error = true; 3201 } 3202 3203 if (i1->assigned_reg() == any_reg) { 3204 tty->print_cr("Interval %d has no register assigned", i1->reg_num()); i1->print(); tty->cr(); 3205 has_error = true; 3206 } 3207 3208 if (i1->assigned_reg() == i1->assigned_regHi()) { 3209 tty->print_cr("Interval %d: low and high register equal", i1->reg_num()); i1->print(); tty->cr(); 3210 has_error = true; 3211 } 3212 3213 if (!is_processed_reg_num(i1->assigned_reg())) { 3214 tty->print_cr("Can not have an Interval for an ignored register"); i1->print(); tty->cr(); 3215 has_error = true; 3216 } 3217 3218 if (i1->first() == Range::end()) { 3219 tty->print_cr("Interval %d has no Range", i1->reg_num()); i1->print(); tty->cr(); 3220 has_error = true; 3221 } 3222 3223 for (Range* r = i1->first(); r != Range::end(); r = r->next()) { 3224 if (r->from() >= r->to()) { 3225 tty->print_cr("Interval %d has zero length range", i1->reg_num()); i1->print(); tty->cr(); 3226 has_error = true; 3227 } 3228 } 3229 3230 for (int j = i + 1; j < len; j++) { 3231 Interval* i2 = interval_at(j); 3232 if (i2 == NULL) continue; 3233 3234 // special intervals that are created in MoveResolver 3235 // -> ignore them because the range information has no meaning there 3236 if (i1->from() == 1 && i1->to() == 2) continue; 3237 if (i2->from() == 1 && i2->to() == 2) continue; 3238 3239 int r1 = i1->assigned_reg(); 3240 int r1Hi = i1->assigned_regHi(); 3241 int r2 = i2->assigned_reg(); 3242 int r2Hi = i2->assigned_regHi(); 3243 if (i1->intersects(i2) && (r1 == r2 || r1 == r2Hi || (r1Hi != any_reg && (r1Hi == r2 || r1Hi == r2Hi)))) { 3244 tty->print_cr("Intervals %d and %d overlap and have the same register assigned", i1->reg_num(), i2->reg_num()); 3245 i1->print(); tty->cr(); 3246 i2->print(); tty->cr(); 3247 has_error = true; 3248 } 3249 } 3250 } 3251 3252 assert(has_error == false, "register allocation invalid"); 3253 } 3254 3255 3256 void LinearScan::verify_no_oops_in_fixed_intervals() { 3257 Interval* fixed_intervals; 3258 Interval* other_intervals; 3259 create_unhandled_lists(&fixed_intervals, &other_intervals, is_precolored_cpu_interval, NULL); 3260 3261 // to ensure a walking until the last instruction id, add a dummy interval 3262 // with a high operation id 3263 other_intervals = new Interval(any_reg); 3264 other_intervals->add_range(max_jint - 2, max_jint - 1); 3265 IntervalWalker* iw = new IntervalWalker(this, fixed_intervals, other_intervals); 3266 3267 LIR_OpVisitState visitor; 3268 for (int i = 0; i < block_count(); i++) { 3269 BlockBegin* block = block_at(i); 3270 3271 LIR_OpList* instructions = block->lir()->instructions_list(); 3272 3273 for (int j = 0; j < instructions->length(); j++) { 3274 LIR_Op* op = instructions->at(j); 3275 int op_id = op->id(); 3276 3277 visitor.visit(op); 3278 3279 if (visitor.info_count() > 0) { 3280 iw->walk_before(op->id()); 3281 bool check_live = true; 3282 if (op->code() == lir_move) { 3283 LIR_Op1* move = (LIR_Op1*)op; 3284 check_live = (move->patch_code() == lir_patch_none); 3285 } 3286 LIR_OpBranch* branch = op->as_OpBranch(); 3287 if (branch != NULL && branch->stub() != NULL && branch->stub()->is_exception_throw_stub()) { 3288 // Don't bother checking the stub in this case since the 3289 // exception stub will never return to normal control flow. 3290 check_live = false; 3291 } 3292 3293 // Make sure none of the fixed registers is live across an 3294 // oopmap since we can't handle that correctly. 3295 if (check_live) { 3296 for (Interval* interval = iw->active_first(fixedKind); 3297 interval != Interval::end(); 3298 interval = interval->next()) { 3299 if (interval->current_to() > op->id() + 1) { 3300 // This interval is live out of this op so make sure 3301 // that this interval represents some value that's 3302 // referenced by this op either as an input or output. 3303 bool ok = false; 3304 for_each_visitor_mode(mode) { 3305 int n = visitor.opr_count(mode); 3306 for (int k = 0; k < n; k++) { 3307 LIR_Opr opr = visitor.opr_at(mode, k); 3308 if (opr->is_fixed_cpu()) { 3309 if (interval_at(reg_num(opr)) == interval) { 3310 ok = true; 3311 break; 3312 } 3313 int hi = reg_numHi(opr); 3314 if (hi != -1 && interval_at(hi) == interval) { 3315 ok = true; 3316 break; 3317 } 3318 } 3319 } 3320 } 3321 assert(ok, "fixed intervals should never be live across an oopmap point"); 3322 } 3323 } 3324 } 3325 } 3326 3327 // oop-maps at calls do not contain registers, so check is not needed 3328 if (!visitor.has_call()) { 3329 3330 for_each_visitor_mode(mode) { 3331 int n = visitor.opr_count(mode); 3332 for (int k = 0; k < n; k++) { 3333 LIR_Opr opr = visitor.opr_at(mode, k); 3334 3335 if (opr->is_fixed_cpu() && opr->is_oop()) { 3336 // operand is a non-virtual cpu register and contains an oop 3337 TRACE_LINEAR_SCAN(4, op->print_on(tty); tty->print("checking operand "); opr->print(); tty->cr()); 3338 3339 Interval* interval = interval_at(reg_num(opr)); 3340 assert(interval != NULL, "no interval"); 3341 3342 if (mode == LIR_OpVisitState::inputMode) { 3343 if (interval->to() >= op_id + 1) { 3344 assert(interval->to() < op_id + 2 || 3345 interval->has_hole_between(op_id, op_id + 2), 3346 "oop input operand live after instruction"); 3347 } 3348 } else if (mode == LIR_OpVisitState::outputMode) { 3349 if (interval->from() <= op_id - 1) { 3350 assert(interval->has_hole_between(op_id - 1, op_id), 3351 "oop input operand live after instruction"); 3352 } 3353 } 3354 } 3355 } 3356 } 3357 } 3358 } 3359 } 3360 } 3361 3362 3363 void LinearScan::verify_constants() { 3364 int num_regs = num_virtual_regs(); 3365 int size = live_set_size(); 3366 int num_blocks = block_count(); 3367 3368 for (int i = 0; i < num_blocks; i++) { 3369 BlockBegin* block = block_at(i); 3370 BitMap live_at_edge = block->live_in(); 3371 3372 // visit all registers where the live_at_edge bit is set 3373 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)) { 3374 TRACE_LINEAR_SCAN(4, tty->print("checking interval %d of block B%d", r, block->block_id())); 3375 3376 Value value = gen()->instruction_for_vreg(r); 3377 3378 assert(value != NULL, "all intervals live across block boundaries must have Value"); 3379 assert(value->operand()->is_register() && value->operand()->is_virtual(), "value must have virtual operand"); 3380 assert(value->operand()->vreg_number() == r, "register number must match"); 3381 // TKR assert(value->as_Constant() == NULL || value->is_pinned(), "only pinned constants can be alive accross block boundaries"); 3382 } 3383 } 3384 } 3385 3386 3387 class RegisterVerifier: public StackObj { 3388 private: 3389 LinearScan* _allocator; 3390 BlockList _work_list; // all blocks that must be processed 3391 IntervalsList _saved_states; // saved information of previous check 3392 3393 // simplified access to methods of LinearScan 3394 Compilation* compilation() const { return _allocator->compilation(); } 3395 Interval* interval_at(int reg_num) const { return _allocator->interval_at(reg_num); } 3396 int reg_num(LIR_Opr opr) const { return _allocator->reg_num(opr); } 3397 3398 // currently, only registers are processed 3399 int state_size() { return LinearScan::nof_regs; } 3400 3401 // accessors 3402 IntervalList* state_for_block(BlockBegin* block) { return _saved_states.at(block->block_id()); } 3403 void set_state_for_block(BlockBegin* block, IntervalList* saved_state) { _saved_states.at_put(block->block_id(), saved_state); } 3404 void add_to_work_list(BlockBegin* block) { if (!_work_list.contains(block)) _work_list.append(block); } 3405 3406 // helper functions 3407 IntervalList* copy(IntervalList* input_state); 3408 void state_put(IntervalList* input_state, int reg, Interval* interval); 3409 bool check_state(IntervalList* input_state, int reg, Interval* interval); 3410 3411 void process_block(BlockBegin* block); 3412 void process_xhandler(XHandler* xhandler, IntervalList* input_state); 3413 void process_successor(BlockBegin* block, IntervalList* input_state); 3414 void process_operations(LIR_List* ops, IntervalList* input_state); 3415 3416 public: 3417 RegisterVerifier(LinearScan* allocator) 3418 : _allocator(allocator) 3419 , _work_list(16) 3420 , _saved_states(BlockBegin::number_of_blocks(), NULL) 3421 { } 3422 3423 void verify(BlockBegin* start); 3424 }; 3425 3426 3427 // entry function from LinearScan that starts the verification 3428 void LinearScan::verify_registers() { 3429 RegisterVerifier verifier(this); 3430 verifier.verify(block_at(0)); 3431 } 3432 3433 3434 void RegisterVerifier::verify(BlockBegin* start) { 3435 // setup input registers (method arguments) for first block 3436 IntervalList* input_state = new IntervalList(state_size(), NULL); 3437 CallingConvention* args = compilation()->frame_map()->incoming_arguments(); 3438 for (int n = 0; n < args->length(); n++) { 3439 LIR_Opr opr = args->at(n); 3440 if (opr->is_register()) { 3441 Interval* interval = interval_at(reg_num(opr)); 3442 3443 if (interval->assigned_reg() < state_size()) { 3444 input_state->at_put(interval->assigned_reg(), interval); 3445 } 3446 if (interval->assigned_regHi() != LinearScan::any_reg && interval->assigned_regHi() < state_size()) { 3447 input_state->at_put(interval->assigned_regHi(), interval); 3448 } 3449 } 3450 } 3451 3452 set_state_for_block(start, input_state); 3453 add_to_work_list(start); 3454 3455 // main loop for verification 3456 do { 3457 BlockBegin* block = _work_list.at(0); 3458 _work_list.remove_at(0); 3459 3460 process_block(block); 3461 } while (!_work_list.is_empty()); 3462 } 3463 3464 void RegisterVerifier::process_block(BlockBegin* block) { 3465 TRACE_LINEAR_SCAN(2, tty->cr(); tty->print_cr("process_block B%d", block->block_id())); 3466 3467 // must copy state because it is modified 3468 IntervalList* input_state = copy(state_for_block(block)); 3469 3470 if (TraceLinearScanLevel >= 4) { 3471 tty->print_cr("Input-State of intervals:"); 3472 tty->print(" "); 3473 for (int i = 0; i < state_size(); i++) { 3474 if (input_state->at(i) != NULL) { 3475 tty->print(" %4d", input_state->at(i)->reg_num()); 3476 } else { 3477 tty->print(" __"); 3478 } 3479 } 3480 tty->cr(); 3481 tty->cr(); 3482 } 3483 3484 // process all operations of the block 3485 process_operations(block->lir(), input_state); 3486 3487 // iterate all successors 3488 for (int i = 0; i < block->number_of_sux(); i++) { 3489 process_successor(block->sux_at(i), input_state); 3490 } 3491 } 3492 3493 void RegisterVerifier::process_xhandler(XHandler* xhandler, IntervalList* input_state) { 3494 TRACE_LINEAR_SCAN(2, tty->print_cr("process_xhandler B%d", xhandler->entry_block()->block_id())); 3495 3496 // must copy state because it is modified 3497 input_state = copy(input_state); 3498 3499 if (xhandler->entry_code() != NULL) { 3500 process_operations(xhandler->entry_code(), input_state); 3501 } 3502 process_successor(xhandler->entry_block(), input_state); 3503 } 3504 3505 void RegisterVerifier::process_successor(BlockBegin* block, IntervalList* input_state) { 3506 IntervalList* saved_state = state_for_block(block); 3507 3508 if (saved_state != NULL) { 3509 // this block was already processed before. 3510 // check if new input_state is consistent with saved_state 3511 3512 bool saved_state_correct = true; 3513 for (int i = 0; i < state_size(); i++) { 3514 if (input_state->at(i) != saved_state->at(i)) { 3515 // current input_state and previous saved_state assume a different 3516 // interval in this register -> assume that this register is invalid 3517 if (saved_state->at(i) != NULL) { 3518 // invalidate old calculation only if it assumed that 3519 // register was valid. when the register was already invalid, 3520 // then the old calculation was correct. 3521 saved_state_correct = false; 3522 saved_state->at_put(i, NULL); 3523 3524 TRACE_LINEAR_SCAN(4, tty->print_cr("process_successor B%d: invalidating slot %d", block->block_id(), i)); 3525 } 3526 } 3527 } 3528 3529 if (saved_state_correct) { 3530 // already processed block with correct input_state 3531 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: previous visit already correct", block->block_id())); 3532 } else { 3533 // must re-visit this block 3534 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: must re-visit because input state changed", block->block_id())); 3535 add_to_work_list(block); 3536 } 3537 3538 } else { 3539 // block was not processed before, so set initial input_state 3540 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: initial visit", block->block_id())); 3541 3542 set_state_for_block(block, copy(input_state)); 3543 add_to_work_list(block); 3544 } 3545 } 3546 3547 3548 IntervalList* RegisterVerifier::copy(IntervalList* input_state) { 3549 IntervalList* copy_state = new IntervalList(input_state->length()); 3550 copy_state->push_all(input_state); 3551 return copy_state; 3552 } 3553 3554 void RegisterVerifier::state_put(IntervalList* input_state, int reg, Interval* interval) { 3555 if (reg != LinearScan::any_reg && reg < state_size()) { 3556 if (interval != NULL) { 3557 TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = %d", reg, interval->reg_num())); 3558 } else if (input_state->at(reg) != NULL) { 3559 TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = NULL", reg)); 3560 } 3561 3562 input_state->at_put(reg, interval); 3563 } 3564 } 3565 3566 bool RegisterVerifier::check_state(IntervalList* input_state, int reg, Interval* interval) { 3567 if (reg != LinearScan::any_reg && reg < state_size()) { 3568 if (input_state->at(reg) != interval) { 3569 tty->print_cr("!! Error in register allocation: register %d does not contain interval %d", reg, interval->reg_num()); 3570 return true; 3571 } 3572 } 3573 return false; 3574 } 3575 3576 void RegisterVerifier::process_operations(LIR_List* ops, IntervalList* input_state) { 3577 // visit all instructions of the block 3578 LIR_OpVisitState visitor; 3579 bool has_error = false; 3580 3581 for (int i = 0; i < ops->length(); i++) { 3582 LIR_Op* op = ops->at(i); 3583 visitor.visit(op); 3584 3585 TRACE_LINEAR_SCAN(4, op->print_on(tty)); 3586 3587 // check if input operands are correct 3588 int j; 3589 int n = visitor.opr_count(LIR_OpVisitState::inputMode); 3590 for (j = 0; j < n; j++) { 3591 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, j); 3592 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { 3593 Interval* interval = interval_at(reg_num(opr)); 3594 if (op->id() != -1) { 3595 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::inputMode); 3596 } 3597 3598 has_error |= check_state(input_state, interval->assigned_reg(), interval->split_parent()); 3599 has_error |= check_state(input_state, interval->assigned_regHi(), interval->split_parent()); 3600 3601 // When an operand is marked with is_last_use, then the fpu stack allocator 3602 // removes the register from the fpu stack -> the register contains no value 3603 if (opr->is_last_use()) { 3604 state_put(input_state, interval->assigned_reg(), NULL); 3605 state_put(input_state, interval->assigned_regHi(), NULL); 3606 } 3607 } 3608 } 3609 3610 // invalidate all caller save registers at calls 3611 if (visitor.has_call()) { 3612 for (j = 0; j < FrameMap::nof_caller_save_cpu_regs(); j++) { 3613 state_put(input_state, reg_num(FrameMap::caller_save_cpu_reg_at(j)), NULL); 3614 } 3615 for (j = 0; j < FrameMap::nof_caller_save_fpu_regs; j++) { 3616 state_put(input_state, reg_num(FrameMap::caller_save_fpu_reg_at(j)), NULL); 3617 } 3618 3619 #ifdef X86 3620 for (j = 0; j < FrameMap::nof_caller_save_xmm_regs; j++) { 3621 state_put(input_state, reg_num(FrameMap::caller_save_xmm_reg_at(j)), NULL); 3622 } 3623 #endif 3624 } 3625 3626 // process xhandler before output and temp operands 3627 XHandlers* xhandlers = visitor.all_xhandler(); 3628 n = xhandlers->length(); 3629 for (int k = 0; k < n; k++) { 3630 process_xhandler(xhandlers->handler_at(k), input_state); 3631 } 3632 3633 // set temp operands (some operations use temp operands also as output operands, so can't set them NULL) 3634 n = visitor.opr_count(LIR_OpVisitState::tempMode); 3635 for (j = 0; j < n; j++) { 3636 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, j); 3637 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { 3638 Interval* interval = interval_at(reg_num(opr)); 3639 if (op->id() != -1) { 3640 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::tempMode); 3641 } 3642 3643 state_put(input_state, interval->assigned_reg(), interval->split_parent()); 3644 state_put(input_state, interval->assigned_regHi(), interval->split_parent()); 3645 } 3646 } 3647 3648 // set output operands 3649 n = visitor.opr_count(LIR_OpVisitState::outputMode); 3650 for (j = 0; j < n; j++) { 3651 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, j); 3652 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { 3653 Interval* interval = interval_at(reg_num(opr)); 3654 if (op->id() != -1) { 3655 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::outputMode); 3656 } 3657 3658 state_put(input_state, interval->assigned_reg(), interval->split_parent()); 3659 state_put(input_state, interval->assigned_regHi(), interval->split_parent()); 3660 } 3661 } 3662 } 3663 assert(has_error == false, "Error in register allocation"); 3664 } 3665 3666 #endif // ASSERT 3667 3668 3669 3670 // **** Implementation of MoveResolver ****************************** 3671 3672 MoveResolver::MoveResolver(LinearScan* allocator) : 3673 _allocator(allocator), 3674 _multiple_reads_allowed(false), 3675 _mapping_from(8), 3676 _mapping_from_opr(8), 3677 _mapping_to(8), 3678 _insert_list(NULL), 3679 _insert_idx(-1), 3680 _insertion_buffer() 3681 { 3682 for (int i = 0; i < LinearScan::nof_regs; i++) { 3683 _register_blocked[i] = 0; 3684 } 3685 DEBUG_ONLY(check_empty()); 3686 } 3687 3688 3689 #ifdef ASSERT 3690 3691 void MoveResolver::check_empty() { 3692 assert(_mapping_from.length() == 0 && _mapping_from_opr.length() == 0 && _mapping_to.length() == 0, "list must be empty before and after processing"); 3693 for (int i = 0; i < LinearScan::nof_regs; i++) { 3694 assert(register_blocked(i) == 0, "register map must be empty before and after processing"); 3695 } 3696 assert(_multiple_reads_allowed == false, "must have default value"); 3697 } 3698 3699 void MoveResolver::verify_before_resolve() { 3700 assert(_mapping_from.length() == _mapping_from_opr.length(), "length must be equal"); 3701 assert(_mapping_from.length() == _mapping_to.length(), "length must be equal"); 3702 assert(_insert_list != NULL && _insert_idx != -1, "insert position not set"); 3703 3704 int i, j; 3705 if (!_multiple_reads_allowed) { 3706 for (i = 0; i < _mapping_from.length(); i++) { 3707 for (j = i + 1; j < _mapping_from.length(); j++) { 3708 assert(_mapping_from.at(i) == NULL || _mapping_from.at(i) != _mapping_from.at(j), "cannot read from same interval twice"); 3709 } 3710 } 3711 } 3712 3713 for (i = 0; i < _mapping_to.length(); i++) { 3714 for (j = i + 1; j < _mapping_to.length(); j++) { 3715 assert(_mapping_to.at(i) != _mapping_to.at(j), "cannot write to same interval twice"); 3716 } 3717 } 3718 3719 3720 BitMap used_regs(LinearScan::nof_regs + allocator()->frame_map()->argcount() + allocator()->max_spills()); 3721 used_regs.clear(); 3722 if (!_multiple_reads_allowed) { 3723 for (i = 0; i < _mapping_from.length(); i++) { 3724 Interval* it = _mapping_from.at(i); 3725 if (it != NULL) { 3726 assert(!used_regs.at(it->assigned_reg()), "cannot read from same register twice"); 3727 used_regs.set_bit(it->assigned_reg()); 3728 3729 if (it->assigned_regHi() != LinearScan::any_reg) { 3730 assert(!used_regs.at(it->assigned_regHi()), "cannot read from same register twice"); 3731 used_regs.set_bit(it->assigned_regHi()); 3732 } 3733 } 3734 } 3735 } 3736 3737 used_regs.clear(); 3738 for (i = 0; i < _mapping_to.length(); i++) { 3739 Interval* it = _mapping_to.at(i); 3740 assert(!used_regs.at(it->assigned_reg()), "cannot write to same register twice"); 3741 used_regs.set_bit(it->assigned_reg()); 3742 3743 if (it->assigned_regHi() != LinearScan::any_reg) { 3744 assert(!used_regs.at(it->assigned_regHi()), "cannot write to same register twice"); 3745 used_regs.set_bit(it->assigned_regHi()); 3746 } 3747 } 3748 3749 used_regs.clear(); 3750 for (i = 0; i < _mapping_from.length(); i++) { 3751 Interval* it = _mapping_from.at(i); 3752 if (it != NULL && it->assigned_reg() >= LinearScan::nof_regs) { 3753 used_regs.set_bit(it->assigned_reg()); 3754 } 3755 } 3756 for (i = 0; i < _mapping_to.length(); i++) { 3757 Interval* it = _mapping_to.at(i); 3758 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"); 3759 } 3760 } 3761 3762 #endif // ASSERT 3763 3764 3765 // mark assigned_reg and assigned_regHi of the interval as blocked 3766 void MoveResolver::block_registers(Interval* it) { 3767 int reg = it->assigned_reg(); 3768 if (reg < LinearScan::nof_regs) { 3769 assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used"); 3770 set_register_blocked(reg, 1); 3771 } 3772 reg = it->assigned_regHi(); 3773 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { 3774 assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used"); 3775 set_register_blocked(reg, 1); 3776 } 3777 } 3778 3779 // mark assigned_reg and assigned_regHi of the interval as unblocked 3780 void MoveResolver::unblock_registers(Interval* it) { 3781 int reg = it->assigned_reg(); 3782 if (reg < LinearScan::nof_regs) { 3783 assert(register_blocked(reg) > 0, "register already marked as unused"); 3784 set_register_blocked(reg, -1); 3785 } 3786 reg = it->assigned_regHi(); 3787 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { 3788 assert(register_blocked(reg) > 0, "register already marked as unused"); 3789 set_register_blocked(reg, -1); 3790 } 3791 } 3792 3793 // check if assigned_reg and assigned_regHi of the to-interval are not blocked (or only blocked by from) 3794 bool MoveResolver::save_to_process_move(Interval* from, Interval* to) { 3795 int from_reg = -1; 3796 int from_regHi = -1; 3797 if (from != NULL) { 3798 from_reg = from->assigned_reg(); 3799 from_regHi = from->assigned_regHi(); 3800 } 3801 3802 int reg = to->assigned_reg(); 3803 if (reg < LinearScan::nof_regs) { 3804 if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) { 3805 return false; 3806 } 3807 } 3808 reg = to->assigned_regHi(); 3809 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { 3810 if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) { 3811 return false; 3812 } 3813 } 3814 3815 return true; 3816 } 3817 3818 3819 void MoveResolver::create_insertion_buffer(LIR_List* list) { 3820 assert(!_insertion_buffer.initialized(), "overwriting existing buffer"); 3821 _insertion_buffer.init(list); 3822 } 3823 3824 void MoveResolver::append_insertion_buffer() { 3825 if (_insertion_buffer.initialized()) { 3826 _insertion_buffer.lir_list()->append(&_insertion_buffer); 3827 } 3828 assert(!_insertion_buffer.initialized(), "must be uninitialized now"); 3829 3830 _insert_list = NULL; 3831 _insert_idx = -1; 3832 } 3833 3834 void MoveResolver::insert_move(Interval* from_interval, Interval* to_interval) { 3835 assert(from_interval->reg_num() != to_interval->reg_num(), "from and to interval equal"); 3836 assert(from_interval->type() == to_interval->type(), "move between different types"); 3837 assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first"); 3838 assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer"); 3839 3840 LIR_Opr from_opr = LIR_OprFact::virtual_register(from_interval->reg_num(), from_interval->type()); 3841 LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type()); 3842 3843 if (!_multiple_reads_allowed) { 3844 // the last_use flag is an optimization for FPU stack allocation. When the same 3845 // input interval is used in more than one move, then it is too difficult to determine 3846 // if this move is really the last use. 3847 from_opr = from_opr->make_last_use(); 3848 } 3849 _insertion_buffer.move(_insert_idx, from_opr, to_opr); 3850 3851 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())); 3852 } 3853 3854 void MoveResolver::insert_move(LIR_Opr from_opr, Interval* to_interval) { 3855 assert(from_opr->type() == to_interval->type(), "move between different types"); 3856 assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first"); 3857 assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer"); 3858 3859 LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type()); 3860 _insertion_buffer.move(_insert_idx, from_opr, to_opr); 3861 3862 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())); 3863 } 3864 3865 3866 void MoveResolver::resolve_mappings() { 3867 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)); 3868 DEBUG_ONLY(verify_before_resolve()); 3869 3870 // Block all registers that are used as input operands of a move. 3871 // When a register is blocked, no move to this register is emitted. 3872 // This is necessary for detecting cycles in moves. 3873 int i; 3874 for (i = _mapping_from.length() - 1; i >= 0; i--) { 3875 Interval* from_interval = _mapping_from.at(i); 3876 if (from_interval != NULL) { 3877 block_registers(from_interval); 3878 } 3879 } 3880 3881 int spill_candidate = -1; 3882 while (_mapping_from.length() > 0) { 3883 bool processed_interval = false; 3884 3885 for (i = _mapping_from.length() - 1; i >= 0; i--) { 3886 Interval* from_interval = _mapping_from.at(i); 3887 Interval* to_interval = _mapping_to.at(i); 3888 3889 if (save_to_process_move(from_interval, to_interval)) { 3890 // this inverval can be processed because target is free 3891 if (from_interval != NULL) { 3892 insert_move(from_interval, to_interval); 3893 unblock_registers(from_interval); 3894 } else { 3895 insert_move(_mapping_from_opr.at(i), to_interval); 3896 } 3897 _mapping_from.remove_at(i); 3898 _mapping_from_opr.remove_at(i); 3899 _mapping_to.remove_at(i); 3900 3901 processed_interval = true; 3902 } else if (from_interval != NULL && from_interval->assigned_reg() < LinearScan::nof_regs) { 3903 // this interval cannot be processed now because target is not free 3904 // it starts in a register, so it is a possible candidate for spilling 3905 spill_candidate = i; 3906 } 3907 } 3908 3909 if (!processed_interval) { 3910 // no move could be processed because there is a cycle in the move list 3911 // (e.g. r1 -> r2, r2 -> r1), so one interval must be spilled to memory 3912 assert(spill_candidate != -1, "no interval in register for spilling found"); 3913 3914 // create a new spill interval and assign a stack slot to it 3915 Interval* from_interval = _mapping_from.at(spill_candidate); 3916 Interval* spill_interval = new Interval(-1); 3917 spill_interval->set_type(from_interval->type()); 3918 3919 // add a dummy range because real position is difficult to calculate 3920 // Note: this range is a special case when the integrity of the allocation is checked 3921 spill_interval->add_range(1, 2); 3922 3923 // do not allocate a new spill slot for temporary interval, but 3924 // use spill slot assigned to from_interval. Otherwise moves from 3925 // one stack slot to another can happen (not allowed by LIR_Assembler 3926 int spill_slot = from_interval->canonical_spill_slot(); 3927 if (spill_slot < 0) { 3928 spill_slot = allocator()->allocate_spill_slot(type2spill_size[spill_interval->type()] == 2); 3929 from_interval->set_canonical_spill_slot(spill_slot); 3930 } 3931 spill_interval->assign_reg(spill_slot); 3932 allocator()->append_interval(spill_interval); 3933 3934 TRACE_LINEAR_SCAN(4, tty->print_cr("created new Interval %d for spilling", spill_interval->reg_num())); 3935 3936 // insert a move from register to stack and update the mapping 3937 insert_move(from_interval, spill_interval); 3938 _mapping_from.at_put(spill_candidate, spill_interval); 3939 unblock_registers(from_interval); 3940 } 3941 } 3942 3943 // reset to default value 3944 _multiple_reads_allowed = false; 3945 3946 // check that all intervals have been processed 3947 DEBUG_ONLY(check_empty()); 3948 } 3949 3950 3951 void MoveResolver::set_insert_position(LIR_List* insert_list, int insert_idx) { 3952 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)); 3953 assert(_insert_list == NULL && _insert_idx == -1, "use move_insert_position instead of set_insert_position when data already set"); 3954 3955 create_insertion_buffer(insert_list); 3956 _insert_list = insert_list; 3957 _insert_idx = insert_idx; 3958 } 3959 3960 void MoveResolver::move_insert_position(LIR_List* insert_list, int insert_idx) { 3961 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)); 3962 3963 if (_insert_list != NULL && (insert_list != _insert_list || insert_idx != _insert_idx)) { 3964 // insert position changed -> resolve current mappings 3965 resolve_mappings(); 3966 } 3967 3968 if (insert_list != _insert_list) { 3969 // block changed -> append insertion_buffer because it is 3970 // bound to a specific block and create a new insertion_buffer 3971 append_insertion_buffer(); 3972 create_insertion_buffer(insert_list); 3973 } 3974 3975 _insert_list = insert_list; 3976 _insert_idx = insert_idx; 3977 } 3978 3979 void MoveResolver::add_mapping(Interval* from_interval, Interval* to_interval) { 3980 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())); 3981 3982 _mapping_from.append(from_interval); 3983 _mapping_from_opr.append(LIR_OprFact::illegalOpr); 3984 _mapping_to.append(to_interval); 3985 } 3986 3987 3988 void MoveResolver::add_mapping(LIR_Opr from_opr, Interval* to_interval) { 3989 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())); 3990 assert(from_opr->is_constant(), "only for constants"); 3991 3992 _mapping_from.append(NULL); 3993 _mapping_from_opr.append(from_opr); 3994 _mapping_to.append(to_interval); 3995 } 3996 3997 void MoveResolver::resolve_and_append_moves() { 3998 if (has_mappings()) { 3999 resolve_mappings(); 4000 } 4001 append_insertion_buffer(); 4002 } 4003 4004 4005 4006 // **** Implementation of Range ************************************* 4007 4008 Range::Range(int from, int to, Range* next) : 4009 _from(from), 4010 _to(to), 4011 _next(next) 4012 { 4013 } 4014 4015 // initialize sentinel 4016 Range* Range::_end = NULL; 4017 void Range::initialize(Arena* arena) { 4018 _end = new (arena) Range(max_jint, max_jint, NULL); 4019 } 4020 4021 int Range::intersects_at(Range* r2) const { 4022 const Range* r1 = this; 4023 4024 assert(r1 != NULL && r2 != NULL, "null ranges not allowed"); 4025 assert(r1 != _end && r2 != _end, "empty ranges not allowed"); 4026 4027 do { 4028 if (r1->from() < r2->from()) { 4029 if (r1->to() <= r2->from()) { 4030 r1 = r1->next(); if (r1 == _end) return -1; 4031 } else { 4032 return r2->from(); 4033 } 4034 } else if (r2->from() < r1->from()) { 4035 if (r2->to() <= r1->from()) { 4036 r2 = r2->next(); if (r2 == _end) return -1; 4037 } else { 4038 return r1->from(); 4039 } 4040 } else { // r1->from() == r2->from() 4041 if (r1->from() == r1->to()) { 4042 r1 = r1->next(); if (r1 == _end) return -1; 4043 } else if (r2->from() == r2->to()) { 4044 r2 = r2->next(); if (r2 == _end) return -1; 4045 } else { 4046 return r1->from(); 4047 } 4048 } 4049 } while (true); 4050 } 4051 4052 #ifndef PRODUCT 4053 void Range::print(outputStream* out) const { 4054 out->print("[%d, %d[ ", _from, _to); 4055 } 4056 #endif 4057 4058 4059 4060 // **** Implementation of Interval ********************************** 4061 4062 // initialize sentinel 4063 Interval* Interval::_end = NULL; 4064 void Interval::initialize(Arena* arena) { 4065 Range::initialize(arena); 4066 _end = new (arena) Interval(-1); 4067 } 4068 4069 Interval::Interval(int reg_num) : 4070 _reg_num(reg_num), 4071 _type(T_ILLEGAL), 4072 _first(Range::end()), 4073 _use_pos_and_kinds(12), 4074 _current(Range::end()), 4075 _next(_end), 4076 _state(invalidState), 4077 _assigned_reg(LinearScan::any_reg), 4078 _assigned_regHi(LinearScan::any_reg), 4079 _cached_to(-1), 4080 _cached_opr(LIR_OprFact::illegalOpr), 4081 _cached_vm_reg(VMRegImpl::Bad()), 4082 _split_children(0), 4083 _canonical_spill_slot(-1), 4084 _insert_move_when_activated(false), 4085 _register_hint(NULL), 4086 _spill_state(noDefinitionFound), 4087 _spill_definition_pos(-1) 4088 { 4089 _split_parent = this; 4090 _current_split_child = this; 4091 } 4092 4093 int Interval::calc_to() { 4094 assert(_first != Range::end(), "interval has no range"); 4095 4096 Range* r = _first; 4097 while (r->next() != Range::end()) { 4098 r = r->next(); 4099 } 4100 return r->to(); 4101 } 4102 4103 4104 #ifdef ASSERT 4105 // consistency check of split-children 4106 void Interval::check_split_children() { 4107 if (_split_children.length() > 0) { 4108 assert(is_split_parent(), "only split parents can have children"); 4109 4110 for (int i = 0; i < _split_children.length(); i++) { 4111 Interval* i1 = _split_children.at(i); 4112 4113 assert(i1->split_parent() == this, "not a split child of this interval"); 4114 assert(i1->type() == type(), "must be equal for all split children"); 4115 assert(i1->canonical_spill_slot() == canonical_spill_slot(), "must be equal for all split children"); 4116 4117 for (int j = i + 1; j < _split_children.length(); j++) { 4118 Interval* i2 = _split_children.at(j); 4119 4120 assert(i1->reg_num() != i2->reg_num(), "same register number"); 4121 4122 if (i1->from() < i2->from()) { 4123 assert(i1->to() <= i2->from() && i1->to() < i2->to(), "intervals overlapping"); 4124 } else { 4125 assert(i2->from() < i1->from(), "intervals start at same op_id"); 4126 assert(i2->to() <= i1->from() && i2->to() < i1->to(), "intervals overlapping"); 4127 } 4128 } 4129 } 4130 } 4131 } 4132 #endif // ASSERT 4133 4134 Interval* Interval::register_hint(bool search_split_child) const { 4135 if (!search_split_child) { 4136 return _register_hint; 4137 } 4138 4139 if (_register_hint != NULL) { 4140 assert(_register_hint->is_split_parent(), "ony split parents are valid hint registers"); 4141 4142 if (_register_hint->assigned_reg() >= 0 && _register_hint->assigned_reg() < LinearScan::nof_regs) { 4143 return _register_hint; 4144 4145 } else if (_register_hint->_split_children.length() > 0) { 4146 // search the first split child that has a register assigned 4147 int len = _register_hint->_split_children.length(); 4148 for (int i = 0; i < len; i++) { 4149 Interval* cur = _register_hint->_split_children.at(i); 4150 4151 if (cur->assigned_reg() >= 0 && cur->assigned_reg() < LinearScan::nof_regs) { 4152 return cur; 4153 } 4154 } 4155 } 4156 } 4157 4158 // no hint interval found that has a register assigned 4159 return NULL; 4160 } 4161 4162 4163 Interval* Interval::split_child_at_op_id(int op_id, LIR_OpVisitState::OprMode mode) { 4164 assert(is_split_parent(), "can only be called for split parents"); 4165 assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)"); 4166 4167 Interval* result; 4168 if (_split_children.length() == 0) { 4169 result = this; 4170 } else { 4171 result = NULL; 4172 int len = _split_children.length(); 4173 4174 // in outputMode, the end of the interval (op_id == cur->to()) is not valid 4175 int to_offset = (mode == LIR_OpVisitState::outputMode ? 0 : 1); 4176 4177 int i; 4178 for (i = 0; i < len; i++) { 4179 Interval* cur = _split_children.at(i); 4180 if (cur->from() <= op_id && op_id < cur->to() + to_offset) { 4181 if (i > 0) { 4182 // exchange current split child to start of list (faster access for next call) 4183 _split_children.at_put(i, _split_children.at(0)); 4184 _split_children.at_put(0, cur); 4185 } 4186 4187 // interval found 4188 result = cur; 4189 break; 4190 } 4191 } 4192 4193 #ifdef ASSERT 4194 for (i = 0; i < len; i++) { 4195 Interval* tmp = _split_children.at(i); 4196 if (tmp != result && tmp->from() <= op_id && op_id < tmp->to() + to_offset) { 4197 tty->print_cr("two valid result intervals found for op_id %d: %d and %d", op_id, result->reg_num(), tmp->reg_num()); 4198 result->print(); 4199 tmp->print(); 4200 assert(false, "two valid result intervals found"); 4201 } 4202 } 4203 #endif 4204 } 4205 4206 assert(result != NULL, "no matching interval found"); 4207 assert(result->covers(op_id, mode), "op_id not covered by interval"); 4208 4209 return result; 4210 } 4211 4212 4213 // returns the last split child that ends before the given op_id 4214 Interval* Interval::split_child_before_op_id(int op_id) { 4215 assert(op_id >= 0, "invalid op_id"); 4216 4217 Interval* parent = split_parent(); 4218 Interval* result = NULL; 4219 4220 int len = parent->_split_children.length(); 4221 assert(len > 0, "no split children available"); 4222 4223 for (int i = len - 1; i >= 0; i--) { 4224 Interval* cur = parent->_split_children.at(i); 4225 if (cur->to() <= op_id && (result == NULL || result->to() < cur->to())) { 4226 result = cur; 4227 } 4228 } 4229 4230 assert(result != NULL, "no split child found"); 4231 return result; 4232 } 4233 4234 4235 // checks if op_id is covered by any split child 4236 bool Interval::split_child_covers(int op_id, LIR_OpVisitState::OprMode mode) { 4237 assert(is_split_parent(), "can only be called for split parents"); 4238 assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)"); 4239 4240 if (_split_children.length() == 0) { 4241 // simple case if interval was not split 4242 return covers(op_id, mode); 4243 4244 } else { 4245 // extended case: check all split children 4246 int len = _split_children.length(); 4247 for (int i = 0; i < len; i++) { 4248 Interval* cur = _split_children.at(i); 4249 if (cur->covers(op_id, mode)) { 4250 return true; 4251 } 4252 } 4253 return false; 4254 } 4255 } 4256 4257 4258 // Note: use positions are sorted descending -> first use has highest index 4259 int Interval::first_usage(IntervalUseKind min_use_kind) const { 4260 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4261 4262 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4263 if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) { 4264 return _use_pos_and_kinds.at(i); 4265 } 4266 } 4267 return max_jint; 4268 } 4269 4270 int Interval::next_usage(IntervalUseKind min_use_kind, int from) const { 4271 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4272 4273 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4274 if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) >= min_use_kind) { 4275 return _use_pos_and_kinds.at(i); 4276 } 4277 } 4278 return max_jint; 4279 } 4280 4281 int Interval::next_usage_exact(IntervalUseKind exact_use_kind, int from) const { 4282 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4283 4284 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4285 if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) == exact_use_kind) { 4286 return _use_pos_and_kinds.at(i); 4287 } 4288 } 4289 return max_jint; 4290 } 4291 4292 int Interval::previous_usage(IntervalUseKind min_use_kind, int from) const { 4293 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4294 4295 int prev = 0; 4296 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4297 if (_use_pos_and_kinds.at(i) > from) { 4298 return prev; 4299 } 4300 if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) { 4301 prev = _use_pos_and_kinds.at(i); 4302 } 4303 } 4304 return prev; 4305 } 4306 4307 void Interval::add_use_pos(int pos, IntervalUseKind use_kind) { 4308 assert(covers(pos, LIR_OpVisitState::inputMode), "use position not covered by live range"); 4309 4310 // do not add use positions for precolored intervals because 4311 // they are never used 4312 if (use_kind != noUse && reg_num() >= LIR_OprDesc::vreg_base) { 4313 #ifdef ASSERT 4314 assert(_use_pos_and_kinds.length() % 2 == 0, "must be"); 4315 for (int i = 0; i < _use_pos_and_kinds.length(); i += 2) { 4316 assert(pos <= _use_pos_and_kinds.at(i), "already added a use-position with lower position"); 4317 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4318 if (i > 0) { 4319 assert(_use_pos_and_kinds.at(i) < _use_pos_and_kinds.at(i - 2), "not sorted descending"); 4320 } 4321 } 4322 #endif 4323 4324 // Note: add_use is called in descending order, so list gets sorted 4325 // automatically by just appending new use positions 4326 int len = _use_pos_and_kinds.length(); 4327 if (len == 0 || _use_pos_and_kinds.at(len - 2) > pos) { 4328 _use_pos_and_kinds.append(pos); 4329 _use_pos_and_kinds.append(use_kind); 4330 } else if (_use_pos_and_kinds.at(len - 1) < use_kind) { 4331 assert(_use_pos_and_kinds.at(len - 2) == pos, "list not sorted correctly"); 4332 _use_pos_and_kinds.at_put(len - 1, use_kind); 4333 } 4334 } 4335 } 4336 4337 void Interval::add_range(int from, int to) { 4338 assert(from < to, "invalid range"); 4339 assert(first() == Range::end() || to < first()->next()->from(), "not inserting at begin of interval"); 4340 assert(from <= first()->to(), "not inserting at begin of interval"); 4341 4342 if (first()->from() <= to) { 4343 // join intersecting ranges 4344 first()->set_from(MIN2(from, first()->from())); 4345 first()->set_to (MAX2(to, first()->to())); 4346 } else { 4347 // insert new range 4348 _first = new Range(from, to, first()); 4349 } 4350 } 4351 4352 Interval* Interval::new_split_child() { 4353 // allocate new interval 4354 Interval* result = new Interval(-1); 4355 result->set_type(type()); 4356 4357 Interval* parent = split_parent(); 4358 result->_split_parent = parent; 4359 result->set_register_hint(parent); 4360 4361 // insert new interval in children-list of parent 4362 if (parent->_split_children.length() == 0) { 4363 assert(is_split_parent(), "list must be initialized at first split"); 4364 4365 parent->_split_children = IntervalList(4); 4366 parent->_split_children.append(this); 4367 } 4368 parent->_split_children.append(result); 4369 4370 return result; 4371 } 4372 4373 // split this interval at the specified position and return 4374 // the remainder as a new interval. 4375 // 4376 // when an interval is split, a bi-directional link is established between the original interval 4377 // (the split parent) and the intervals that are split off this interval (the split children) 4378 // When a split child is split again, the new created interval is also a direct child 4379 // of the original parent (there is no tree of split children stored, but a flat list) 4380 // All split children are spilled to the same stack slot (stored in _canonical_spill_slot) 4381 // 4382 // Note: The new interval has no valid reg_num 4383 Interval* Interval::split(int split_pos) { 4384 assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals"); 4385 4386 // allocate new interval 4387 Interval* result = new_split_child(); 4388 4389 // split the ranges 4390 Range* prev = NULL; 4391 Range* cur = _first; 4392 while (cur != Range::end() && cur->to() <= split_pos) { 4393 prev = cur; 4394 cur = cur->next(); 4395 } 4396 assert(cur != Range::end(), "split interval after end of last range"); 4397 4398 if (cur->from() < split_pos) { 4399 result->_first = new Range(split_pos, cur->to(), cur->next()); 4400 cur->set_to(split_pos); 4401 cur->set_next(Range::end()); 4402 4403 } else { 4404 assert(prev != NULL, "split before start of first range"); 4405 result->_first = cur; 4406 prev->set_next(Range::end()); 4407 } 4408 result->_current = result->_first; 4409 _cached_to = -1; // clear cached value 4410 4411 // split list of use positions 4412 int total_len = _use_pos_and_kinds.length(); 4413 int start_idx = total_len - 2; 4414 while (start_idx >= 0 && _use_pos_and_kinds.at(start_idx) < split_pos) { 4415 start_idx -= 2; 4416 } 4417 4418 intStack new_use_pos_and_kinds(total_len - start_idx); 4419 int i; 4420 for (i = start_idx + 2; i < total_len; i++) { 4421 new_use_pos_and_kinds.append(_use_pos_and_kinds.at(i)); 4422 } 4423 4424 _use_pos_and_kinds.truncate(start_idx + 2); 4425 result->_use_pos_and_kinds = _use_pos_and_kinds; 4426 _use_pos_and_kinds = new_use_pos_and_kinds; 4427 4428 #ifdef ASSERT 4429 assert(_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos"); 4430 assert(result->_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos"); 4431 assert(_use_pos_and_kinds.length() + result->_use_pos_and_kinds.length() == total_len, "missed some entries"); 4432 4433 for (i = 0; i < _use_pos_and_kinds.length(); i += 2) { 4434 assert(_use_pos_and_kinds.at(i) < split_pos, "must be"); 4435 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4436 } 4437 for (i = 0; i < result->_use_pos_and_kinds.length(); i += 2) { 4438 assert(result->_use_pos_and_kinds.at(i) >= split_pos, "must be"); 4439 assert(result->_use_pos_and_kinds.at(i + 1) >= firstValidKind && result->_use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4440 } 4441 #endif 4442 4443 return result; 4444 } 4445 4446 // split this interval at the specified position and return 4447 // the head as a new interval (the original interval is the tail) 4448 // 4449 // Currently, only the first range can be split, and the new interval 4450 // must not have split positions 4451 Interval* Interval::split_from_start(int split_pos) { 4452 assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals"); 4453 assert(split_pos > from() && split_pos < to(), "can only split inside interval"); 4454 assert(split_pos > _first->from() && split_pos <= _first->to(), "can only split inside first range"); 4455 assert(first_usage(noUse) > split_pos, "can not split when use positions are present"); 4456 4457 // allocate new interval 4458 Interval* result = new_split_child(); 4459 4460 // the new created interval has only one range (checked by assertion above), 4461 // so the splitting of the ranges is very simple 4462 result->add_range(_first->from(), split_pos); 4463 4464 if (split_pos == _first->to()) { 4465 assert(_first->next() != Range::end(), "must not be at end"); 4466 _first = _first->next(); 4467 } else { 4468 _first->set_from(split_pos); 4469 } 4470 4471 return result; 4472 } 4473 4474 4475 // returns true if the op_id is inside the interval 4476 bool Interval::covers(int op_id, LIR_OpVisitState::OprMode mode) const { 4477 Range* cur = _first; 4478 4479 while (cur != Range::end() && cur->to() < op_id) { 4480 cur = cur->next(); 4481 } 4482 if (cur != Range::end()) { 4483 assert(cur->to() != cur->next()->from(), "ranges not separated"); 4484 4485 if (mode == LIR_OpVisitState::outputMode) { 4486 return cur->from() <= op_id && op_id < cur->to(); 4487 } else { 4488 return cur->from() <= op_id && op_id <= cur->to(); 4489 } 4490 } 4491 return false; 4492 } 4493 4494 // returns true if the interval has any hole between hole_from and hole_to 4495 // (even if the hole has only the length 1) 4496 bool Interval::has_hole_between(int hole_from, int hole_to) { 4497 assert(hole_from < hole_to, "check"); 4498 assert(from() <= hole_from && hole_to <= to(), "index out of interval"); 4499 4500 Range* cur = _first; 4501 while (cur != Range::end()) { 4502 assert(cur->to() < cur->next()->from(), "no space between ranges"); 4503 4504 // hole-range starts before this range -> hole 4505 if (hole_from < cur->from()) { 4506 return true; 4507 4508 // hole-range completely inside this range -> no hole 4509 } else if (hole_to <= cur->to()) { 4510 return false; 4511 4512 // overlapping of hole-range with this range -> hole 4513 } else if (hole_from <= cur->to()) { 4514 return true; 4515 } 4516 4517 cur = cur->next(); 4518 } 4519 4520 return false; 4521 } 4522 4523 4524 #ifndef PRODUCT 4525 void Interval::print(outputStream* out) const { 4526 const char* SpillState2Name[] = { "no definition", "no spill store", "one spill store", "store at definition", "start in memory", "no optimization" }; 4527 const char* UseKind2Name[] = { "N", "L", "S", "M" }; 4528 4529 const char* type_name; 4530 LIR_Opr opr = LIR_OprFact::illegal(); 4531 if (reg_num() < LIR_OprDesc::vreg_base) { 4532 type_name = "fixed"; 4533 // need a temporary operand for fixed intervals because type() cannot be called 4534 if (assigned_reg() >= pd_first_cpu_reg && assigned_reg() <= pd_last_cpu_reg) { 4535 opr = LIR_OprFact::single_cpu(assigned_reg()); 4536 } else if (assigned_reg() >= pd_first_fpu_reg && assigned_reg() <= pd_last_fpu_reg) { 4537 opr = LIR_OprFact::single_fpu(assigned_reg() - pd_first_fpu_reg); 4538 #ifdef X86 4539 } else if (assigned_reg() >= pd_first_xmm_reg && assigned_reg() <= pd_last_xmm_reg) { 4540 opr = LIR_OprFact::single_xmm(assigned_reg() - pd_first_xmm_reg); 4541 #endif 4542 } else { 4543 ShouldNotReachHere(); 4544 } 4545 } else { 4546 type_name = type2name(type()); 4547 if (assigned_reg() != -1 && 4548 (LinearScan::num_physical_regs(type()) == 1 || assigned_regHi() != -1)) { 4549 opr = LinearScan::calc_operand_for_interval(this); 4550 } 4551 } 4552 4553 out->print("%d %s ", reg_num(), type_name); 4554 if (opr->is_valid()) { 4555 out->print("\""); 4556 opr->print(out); 4557 out->print("\" "); 4558 } 4559 out->print("%d %d ", split_parent()->reg_num(), (register_hint(false) != NULL ? register_hint(false)->reg_num() : -1)); 4560 4561 // print ranges 4562 Range* cur = _first; 4563 while (cur != Range::end()) { 4564 cur->print(out); 4565 cur = cur->next(); 4566 assert(cur != NULL, "range list not closed with range sentinel"); 4567 } 4568 4569 // print use positions 4570 int prev = 0; 4571 assert(_use_pos_and_kinds.length() % 2 == 0, "must be"); 4572 for (int i =_use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4573 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4574 assert(prev < _use_pos_and_kinds.at(i), "use positions not sorted"); 4575 4576 out->print("%d %s ", _use_pos_and_kinds.at(i), UseKind2Name[_use_pos_and_kinds.at(i + 1)]); 4577 prev = _use_pos_and_kinds.at(i); 4578 } 4579 4580 out->print(" \"%s\"", SpillState2Name[spill_state()]); 4581 out->cr(); 4582 } 4583 #endif 4584 4585 4586 4587 // **** Implementation of IntervalWalker **************************** 4588 4589 IntervalWalker::IntervalWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first) 4590 : _compilation(allocator->compilation()) 4591 , _allocator(allocator) 4592 { 4593 _unhandled_first[fixedKind] = unhandled_fixed_first; 4594 _unhandled_first[anyKind] = unhandled_any_first; 4595 _active_first[fixedKind] = Interval::end(); 4596 _inactive_first[fixedKind] = Interval::end(); 4597 _active_first[anyKind] = Interval::end(); 4598 _inactive_first[anyKind] = Interval::end(); 4599 _current_position = -1; 4600 _current = NULL; 4601 next_interval(); 4602 } 4603 4604 4605 // append interval at top of list 4606 void IntervalWalker::append_unsorted(Interval** list, Interval* interval) { 4607 interval->set_next(*list); *list = interval; 4608 } 4609 4610 4611 // append interval in order of current range from() 4612 void IntervalWalker::append_sorted(Interval** list, Interval* interval) { 4613 Interval* prev = NULL; 4614 Interval* cur = *list; 4615 while (cur->current_from() < interval->current_from()) { 4616 prev = cur; cur = cur->next(); 4617 } 4618 if (prev == NULL) { 4619 *list = interval; 4620 } else { 4621 prev->set_next(interval); 4622 } 4623 interval->set_next(cur); 4624 } 4625 4626 void IntervalWalker::append_to_unhandled(Interval** list, Interval* interval) { 4627 assert(interval->from() >= current()->current_from(), "cannot append new interval before current walk position"); 4628 4629 Interval* prev = NULL; 4630 Interval* cur = *list; 4631 while (cur->from() < interval->from() || (cur->from() == interval->from() && cur->first_usage(noUse) < interval->first_usage(noUse))) { 4632 prev = cur; cur = cur->next(); 4633 } 4634 if (prev == NULL) { 4635 *list = interval; 4636 } else { 4637 prev->set_next(interval); 4638 } 4639 interval->set_next(cur); 4640 } 4641 4642 4643 inline bool IntervalWalker::remove_from_list(Interval** list, Interval* i) { 4644 while (*list != Interval::end() && *list != i) { 4645 list = (*list)->next_addr(); 4646 } 4647 if (*list != Interval::end()) { 4648 assert(*list == i, "check"); 4649 *list = (*list)->next(); 4650 return true; 4651 } else { 4652 return false; 4653 } 4654 } 4655 4656 void IntervalWalker::remove_from_list(Interval* i) { 4657 bool deleted; 4658 4659 if (i->state() == activeState) { 4660 deleted = remove_from_list(active_first_addr(anyKind), i); 4661 } else { 4662 assert(i->state() == inactiveState, "invalid state"); 4663 deleted = remove_from_list(inactive_first_addr(anyKind), i); 4664 } 4665 4666 assert(deleted, "interval has not been found in list"); 4667 } 4668 4669 4670 void IntervalWalker::walk_to(IntervalState state, int from) { 4671 assert (state == activeState || state == inactiveState, "wrong state"); 4672 for_each_interval_kind(kind) { 4673 Interval** prev = state == activeState ? active_first_addr(kind) : inactive_first_addr(kind); 4674 Interval* next = *prev; 4675 while (next->current_from() <= from) { 4676 Interval* cur = next; 4677 next = cur->next(); 4678 4679 bool range_has_changed = false; 4680 while (cur->current_to() <= from) { 4681 cur->next_range(); 4682 range_has_changed = true; 4683 } 4684 4685 // also handle move from inactive list to active list 4686 range_has_changed = range_has_changed || (state == inactiveState && cur->current_from() <= from); 4687 4688 if (range_has_changed) { 4689 // remove cur from list 4690 *prev = next; 4691 if (cur->current_at_end()) { 4692 // move to handled state (not maintained as a list) 4693 cur->set_state(handledState); 4694 interval_moved(cur, kind, state, handledState); 4695 } else if (cur->current_from() <= from){ 4696 // sort into active list 4697 append_sorted(active_first_addr(kind), cur); 4698 cur->set_state(activeState); 4699 if (*prev == cur) { 4700 assert(state == activeState, "check"); 4701 prev = cur->next_addr(); 4702 } 4703 interval_moved(cur, kind, state, activeState); 4704 } else { 4705 // sort into inactive list 4706 append_sorted(inactive_first_addr(kind), cur); 4707 cur->set_state(inactiveState); 4708 if (*prev == cur) { 4709 assert(state == inactiveState, "check"); 4710 prev = cur->next_addr(); 4711 } 4712 interval_moved(cur, kind, state, inactiveState); 4713 } 4714 } else { 4715 prev = cur->next_addr(); 4716 continue; 4717 } 4718 } 4719 } 4720 } 4721 4722 4723 void IntervalWalker::next_interval() { 4724 IntervalKind kind; 4725 Interval* any = _unhandled_first[anyKind]; 4726 Interval* fixed = _unhandled_first[fixedKind]; 4727 4728 if (any != Interval::end()) { 4729 // intervals may start at same position -> prefer fixed interval 4730 kind = fixed != Interval::end() && fixed->from() <= any->from() ? fixedKind : anyKind; 4731 4732 assert (kind == fixedKind && fixed->from() <= any->from() || 4733 kind == anyKind && any->from() <= fixed->from(), "wrong interval!!!"); 4734 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"); 4735 4736 } else if (fixed != Interval::end()) { 4737 kind = fixedKind; 4738 } else { 4739 _current = NULL; return; 4740 } 4741 _current_kind = kind; 4742 _current = _unhandled_first[kind]; 4743 _unhandled_first[kind] = _current->next(); 4744 _current->set_next(Interval::end()); 4745 _current->rewind_range(); 4746 } 4747 4748 4749 void IntervalWalker::walk_to(int lir_op_id) { 4750 assert(_current_position <= lir_op_id, "can not walk backwards"); 4751 while (current() != NULL) { 4752 bool is_active = current()->from() <= lir_op_id; 4753 int id = is_active ? current()->from() : lir_op_id; 4754 4755 TRACE_LINEAR_SCAN(2, if (_current_position < id) { tty->cr(); tty->print_cr("walk_to(%d) **************************************************************", id); }) 4756 4757 // set _current_position prior to call of walk_to 4758 _current_position = id; 4759 4760 // call walk_to even if _current_position == id 4761 walk_to(activeState, id); 4762 walk_to(inactiveState, id); 4763 4764 if (is_active) { 4765 current()->set_state(activeState); 4766 if (activate_current()) { 4767 append_sorted(active_first_addr(current_kind()), current()); 4768 interval_moved(current(), current_kind(), unhandledState, activeState); 4769 } 4770 4771 next_interval(); 4772 } else { 4773 return; 4774 } 4775 } 4776 } 4777 4778 void IntervalWalker::interval_moved(Interval* interval, IntervalKind kind, IntervalState from, IntervalState to) { 4779 #ifndef PRODUCT 4780 if (TraceLinearScanLevel >= 4) { 4781 #define print_state(state) \ 4782 switch(state) {\ 4783 case unhandledState: tty->print("unhandled"); break;\ 4784 case activeState: tty->print("active"); break;\ 4785 case inactiveState: tty->print("inactive"); break;\ 4786 case handledState: tty->print("handled"); break;\ 4787 default: ShouldNotReachHere(); \ 4788 } 4789 4790 print_state(from); tty->print(" to "); print_state(to); 4791 tty->fill_to(23); 4792 interval->print(); 4793 4794 #undef print_state 4795 } 4796 #endif 4797 } 4798 4799 4800 4801 // **** Implementation of LinearScanWalker ************************** 4802 4803 LinearScanWalker::LinearScanWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first) 4804 : IntervalWalker(allocator, unhandled_fixed_first, unhandled_any_first) 4805 , _move_resolver(allocator) 4806 { 4807 for (int i = 0; i < LinearScan::nof_regs; i++) { 4808 _spill_intervals[i] = new IntervalList(2); 4809 } 4810 } 4811 4812 4813 inline void LinearScanWalker::init_use_lists(bool only_process_use_pos) { 4814 for (int i = _first_reg; i <= _last_reg; i++) { 4815 _use_pos[i] = max_jint; 4816 4817 if (!only_process_use_pos) { 4818 _block_pos[i] = max_jint; 4819 _spill_intervals[i]->clear(); 4820 } 4821 } 4822 } 4823 4824 inline void LinearScanWalker::exclude_from_use(int reg) { 4825 assert(reg < LinearScan::nof_regs, "interval must have a register assigned (stack slots not allowed)"); 4826 if (reg >= _first_reg && reg <= _last_reg) { 4827 _use_pos[reg] = 0; 4828 } 4829 } 4830 inline void LinearScanWalker::exclude_from_use(Interval* i) { 4831 assert(i->assigned_reg() != any_reg, "interval has no register assigned"); 4832 4833 exclude_from_use(i->assigned_reg()); 4834 exclude_from_use(i->assigned_regHi()); 4835 } 4836 4837 inline void LinearScanWalker::set_use_pos(int reg, Interval* i, int use_pos, bool only_process_use_pos) { 4838 assert(use_pos != 0, "must use exclude_from_use to set use_pos to 0"); 4839 4840 if (reg >= _first_reg && reg <= _last_reg) { 4841 if (_use_pos[reg] > use_pos) { 4842 _use_pos[reg] = use_pos; 4843 } 4844 if (!only_process_use_pos) { 4845 _spill_intervals[reg]->append(i); 4846 } 4847 } 4848 } 4849 inline void LinearScanWalker::set_use_pos(Interval* i, int use_pos, bool only_process_use_pos) { 4850 assert(i->assigned_reg() != any_reg, "interval has no register assigned"); 4851 if (use_pos != -1) { 4852 set_use_pos(i->assigned_reg(), i, use_pos, only_process_use_pos); 4853 set_use_pos(i->assigned_regHi(), i, use_pos, only_process_use_pos); 4854 } 4855 } 4856 4857 inline void LinearScanWalker::set_block_pos(int reg, Interval* i, int block_pos) { 4858 if (reg >= _first_reg && reg <= _last_reg) { 4859 if (_block_pos[reg] > block_pos) { 4860 _block_pos[reg] = block_pos; 4861 } 4862 if (_use_pos[reg] > block_pos) { 4863 _use_pos[reg] = block_pos; 4864 } 4865 } 4866 } 4867 inline void LinearScanWalker::set_block_pos(Interval* i, int block_pos) { 4868 assert(i->assigned_reg() != any_reg, "interval has no register assigned"); 4869 if (block_pos != -1) { 4870 set_block_pos(i->assigned_reg(), i, block_pos); 4871 set_block_pos(i->assigned_regHi(), i, block_pos); 4872 } 4873 } 4874 4875 4876 void LinearScanWalker::free_exclude_active_fixed() { 4877 Interval* list = active_first(fixedKind); 4878 while (list != Interval::end()) { 4879 assert(list->assigned_reg() < LinearScan::nof_regs, "active interval must have a register assigned"); 4880 exclude_from_use(list); 4881 list = list->next(); 4882 } 4883 } 4884 4885 void LinearScanWalker::free_exclude_active_any() { 4886 Interval* list = active_first(anyKind); 4887 while (list != Interval::end()) { 4888 exclude_from_use(list); 4889 list = list->next(); 4890 } 4891 } 4892 4893 void LinearScanWalker::free_collect_inactive_fixed(Interval* cur) { 4894 Interval* list = inactive_first(fixedKind); 4895 while (list != Interval::end()) { 4896 if (cur->to() <= list->current_from()) { 4897 assert(list->current_intersects_at(cur) == -1, "must not intersect"); 4898 set_use_pos(list, list->current_from(), true); 4899 } else { 4900 set_use_pos(list, list->current_intersects_at(cur), true); 4901 } 4902 list = list->next(); 4903 } 4904 } 4905 4906 void LinearScanWalker::free_collect_inactive_any(Interval* cur) { 4907 Interval* list = inactive_first(anyKind); 4908 while (list != Interval::end()) { 4909 set_use_pos(list, list->current_intersects_at(cur), true); 4910 list = list->next(); 4911 } 4912 } 4913 4914 void LinearScanWalker::free_collect_unhandled(IntervalKind kind, Interval* cur) { 4915 Interval* list = unhandled_first(kind); 4916 while (list != Interval::end()) { 4917 set_use_pos(list, list->intersects_at(cur), true); 4918 if (kind == fixedKind && cur->to() <= list->from()) { 4919 set_use_pos(list, list->from(), true); 4920 } 4921 list = list->next(); 4922 } 4923 } 4924 4925 void LinearScanWalker::spill_exclude_active_fixed() { 4926 Interval* list = active_first(fixedKind); 4927 while (list != Interval::end()) { 4928 exclude_from_use(list); 4929 list = list->next(); 4930 } 4931 } 4932 4933 void LinearScanWalker::spill_block_unhandled_fixed(Interval* cur) { 4934 Interval* list = unhandled_first(fixedKind); 4935 while (list != Interval::end()) { 4936 set_block_pos(list, list->intersects_at(cur)); 4937 list = list->next(); 4938 } 4939 } 4940 4941 void LinearScanWalker::spill_block_inactive_fixed(Interval* cur) { 4942 Interval* list = inactive_first(fixedKind); 4943 while (list != Interval::end()) { 4944 if (cur->to() > list->current_from()) { 4945 set_block_pos(list, list->current_intersects_at(cur)); 4946 } else { 4947 assert(list->current_intersects_at(cur) == -1, "invalid optimization: intervals intersect"); 4948 } 4949 4950 list = list->next(); 4951 } 4952 } 4953 4954 void LinearScanWalker::spill_collect_active_any() { 4955 Interval* list = active_first(anyKind); 4956 while (list != Interval::end()) { 4957 set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false); 4958 list = list->next(); 4959 } 4960 } 4961 4962 void LinearScanWalker::spill_collect_inactive_any(Interval* cur) { 4963 Interval* list = inactive_first(anyKind); 4964 while (list != Interval::end()) { 4965 if (list->current_intersects(cur)) { 4966 set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false); 4967 } 4968 list = list->next(); 4969 } 4970 } 4971 4972 4973 void LinearScanWalker::insert_move(int op_id, Interval* src_it, Interval* dst_it) { 4974 // output all moves here. When source and target are equal, the move is 4975 // optimized away later in assign_reg_nums 4976 4977 op_id = (op_id + 1) & ~1; 4978 BlockBegin* op_block = allocator()->block_of_op_with_id(op_id); 4979 assert(op_id > 0 && allocator()->block_of_op_with_id(op_id - 2) == op_block, "cannot insert move at block boundary"); 4980 4981 // calculate index of instruction inside instruction list of current block 4982 // the minimal index (for a block with no spill moves) can be calculated because the 4983 // numbering of instructions is known. 4984 // When the block already contains spill moves, the index must be increased until the 4985 // correct index is reached. 4986 LIR_OpList* list = op_block->lir()->instructions_list(); 4987 int index = (op_id - list->at(0)->id()) / 2; 4988 assert(list->at(index)->id() <= op_id, "error in calculation"); 4989 4990 while (list->at(index)->id() != op_id) { 4991 index++; 4992 assert(0 <= index && index < list->length(), "index out of bounds"); 4993 } 4994 assert(1 <= index && index < list->length(), "index out of bounds"); 4995 assert(list->at(index)->id() == op_id, "error in calculation"); 4996 4997 // insert new instruction before instruction at position index 4998 _move_resolver.move_insert_position(op_block->lir(), index - 1); 4999 _move_resolver.add_mapping(src_it, dst_it); 5000 } 5001 5002 5003 int LinearScanWalker::find_optimal_split_pos(BlockBegin* min_block, BlockBegin* max_block, int max_split_pos) { 5004 int from_block_nr = min_block->linear_scan_number(); 5005 int to_block_nr = max_block->linear_scan_number(); 5006 5007 assert(0 <= from_block_nr && from_block_nr < block_count(), "out of range"); 5008 assert(0 <= to_block_nr && to_block_nr < block_count(), "out of range"); 5009 assert(from_block_nr < to_block_nr, "must cross block boundary"); 5010 5011 // Try to split at end of max_block. If this would be after 5012 // max_split_pos, then use the begin of max_block 5013 int optimal_split_pos = max_block->last_lir_instruction_id() + 2; 5014 if (optimal_split_pos > max_split_pos) { 5015 optimal_split_pos = max_block->first_lir_instruction_id(); 5016 } 5017 5018 int min_loop_depth = max_block->loop_depth(); 5019 for (int i = to_block_nr - 1; i >= from_block_nr; i--) { 5020 BlockBegin* cur = block_at(i); 5021 5022 if (cur->loop_depth() < min_loop_depth) { 5023 // block with lower loop-depth found -> split at the end of this block 5024 min_loop_depth = cur->loop_depth(); 5025 optimal_split_pos = cur->last_lir_instruction_id() + 2; 5026 } 5027 } 5028 assert(optimal_split_pos > allocator()->max_lir_op_id() || allocator()->is_block_begin(optimal_split_pos), "algorithm must move split pos to block boundary"); 5029 5030 return optimal_split_pos; 5031 } 5032 5033 5034 int LinearScanWalker::find_optimal_split_pos(Interval* it, int min_split_pos, int max_split_pos, bool do_loop_optimization) { 5035 int optimal_split_pos = -1; 5036 if (min_split_pos == max_split_pos) { 5037 // trivial case, no optimization of split position possible 5038 TRACE_LINEAR_SCAN(4, tty->print_cr(" min-pos and max-pos are equal, no optimization possible")); 5039 optimal_split_pos = min_split_pos; 5040 5041 } else { 5042 assert(min_split_pos < max_split_pos, "must be true then"); 5043 assert(min_split_pos > 0, "cannot access min_split_pos - 1 otherwise"); 5044 5045 // reason for using min_split_pos - 1: when the minimal split pos is exactly at the 5046 // beginning of a block, then min_split_pos is also a possible split position. 5047 // Use the block before as min_block, because then min_block->last_lir_instruction_id() + 2 == min_split_pos 5048 BlockBegin* min_block = allocator()->block_of_op_with_id(min_split_pos - 1); 5049 5050 // reason for using max_split_pos - 1: otherwise there would be an assertion failure 5051 // when an interval ends at the end of the last block of the method 5052 // (in this case, max_split_pos == allocator()->max_lir_op_id() + 2, and there is no 5053 // block at this op_id) 5054 BlockBegin* max_block = allocator()->block_of_op_with_id(max_split_pos - 1); 5055 5056 assert(min_block->linear_scan_number() <= max_block->linear_scan_number(), "invalid order"); 5057 if (min_block == max_block) { 5058 // split position cannot be moved to block boundary, so split as late as possible 5059 TRACE_LINEAR_SCAN(4, tty->print_cr(" cannot move split pos to block boundary because min_pos and max_pos are in same block")); 5060 optimal_split_pos = max_split_pos; 5061 5062 } else if (it->has_hole_between(max_split_pos - 1, max_split_pos) && !allocator()->is_block_begin(max_split_pos)) { 5063 // Do not move split position if the interval has a hole before max_split_pos. 5064 // Intervals resulting from Phi-Functions have more than one definition (marked 5065 // as mustHaveRegister) with a hole before each definition. When the register is needed 5066 // for the second definition, an earlier reloading is unnecessary. 5067 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has hole just before max_split_pos, so splitting at max_split_pos")); 5068 optimal_split_pos = max_split_pos; 5069 5070 } else { 5071 // seach optimal block boundary between min_split_pos and max_split_pos 5072 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())); 5073 5074 if (do_loop_optimization) { 5075 // Loop optimization: if a loop-end marker is found between min- and max-position, 5076 // then split before this loop 5077 int loop_end_pos = it->next_usage_exact(loopEndMarker, min_block->last_lir_instruction_id() + 2); 5078 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization: loop end found at pos %d", loop_end_pos)); 5079 5080 assert(loop_end_pos > min_split_pos, "invalid order"); 5081 if (loop_end_pos < max_split_pos) { 5082 // loop-end marker found between min- and max-position 5083 // if it is not the end marker for the same loop as the min-position, then move 5084 // the max-position to this loop block. 5085 // Desired result: uses tagged as shouldHaveRegister inside a loop cause a reloading 5086 // of the interval (normally, only mustHaveRegister causes a reloading) 5087 BlockBegin* loop_block = allocator()->block_of_op_with_id(loop_end_pos); 5088 5089 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())); 5090 assert(loop_block != min_block, "loop_block and min_block must be different because block boundary is needed between"); 5091 5092 optimal_split_pos = find_optimal_split_pos(min_block, loop_block, loop_block->last_lir_instruction_id() + 2); 5093 if (optimal_split_pos == loop_block->last_lir_instruction_id() + 2) { 5094 optimal_split_pos = -1; 5095 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization not necessary")); 5096 } else { 5097 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization successful")); 5098 } 5099 } 5100 } 5101 5102 if (optimal_split_pos == -1) { 5103 // not calculated by loop optimization 5104 optimal_split_pos = find_optimal_split_pos(min_block, max_block, max_split_pos); 5105 } 5106 } 5107 } 5108 TRACE_LINEAR_SCAN(4, tty->print_cr(" optimal split position: %d", optimal_split_pos)); 5109 5110 return optimal_split_pos; 5111 } 5112 5113 5114 /* 5115 split an interval at the optimal position between min_split_pos and 5116 max_split_pos in two parts: 5117 1) the left part has already a location assigned 5118 2) the right part is sorted into to the unhandled-list 5119 */ 5120 void LinearScanWalker::split_before_usage(Interval* it, int min_split_pos, int max_split_pos) { 5121 TRACE_LINEAR_SCAN(2, tty->print ("----- splitting interval: "); it->print()); 5122 TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d", min_split_pos, max_split_pos)); 5123 5124 assert(it->from() < min_split_pos, "cannot split at start of interval"); 5125 assert(current_position() < min_split_pos, "cannot split before current position"); 5126 assert(min_split_pos <= max_split_pos, "invalid order"); 5127 assert(max_split_pos <= it->to(), "cannot split after end of interval"); 5128 5129 int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, true); 5130 5131 assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range"); 5132 assert(optimal_split_pos <= it->to(), "cannot split after end of interval"); 5133 assert(optimal_split_pos > it->from(), "cannot split at start of interval"); 5134 5135 if (optimal_split_pos == it->to() && it->next_usage(mustHaveRegister, min_split_pos) == max_jint) { 5136 // the split position would be just before the end of the interval 5137 // -> no split at all necessary 5138 TRACE_LINEAR_SCAN(4, tty->print_cr(" no split necessary because optimal split position is at end of interval")); 5139 return; 5140 } 5141 5142 // must calculate this before the actual split is performed and before split position is moved to odd op_id 5143 bool move_necessary = !allocator()->is_block_begin(optimal_split_pos) && !it->has_hole_between(optimal_split_pos - 1, optimal_split_pos); 5144 5145 if (!allocator()->is_block_begin(optimal_split_pos)) { 5146 // move position before actual instruction (odd op_id) 5147 optimal_split_pos = (optimal_split_pos - 1) | 1; 5148 } 5149 5150 TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d", optimal_split_pos)); 5151 assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary"); 5152 assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary"); 5153 5154 Interval* split_part = it->split(optimal_split_pos); 5155 5156 allocator()->append_interval(split_part); 5157 allocator()->copy_register_flags(it, split_part); 5158 split_part->set_insert_move_when_activated(move_necessary); 5159 append_to_unhandled(unhandled_first_addr(anyKind), split_part); 5160 5161 TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts (insert_move_when_activated: %d)", move_necessary)); 5162 TRACE_LINEAR_SCAN(2, tty->print (" "); it->print()); 5163 TRACE_LINEAR_SCAN(2, tty->print (" "); split_part->print()); 5164 } 5165 5166 /* 5167 split an interval at the optimal position between min_split_pos and 5168 max_split_pos in two parts: 5169 1) the left part has already a location assigned 5170 2) the right part is always on the stack and therefore ignored in further processing 5171 */ 5172 void LinearScanWalker::split_for_spilling(Interval* it) { 5173 // calculate allowed range of splitting position 5174 int max_split_pos = current_position(); 5175 int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, max_split_pos) + 1, it->from()); 5176 5177 TRACE_LINEAR_SCAN(2, tty->print ("----- splitting and spilling interval: "); it->print()); 5178 TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d", min_split_pos, max_split_pos)); 5179 5180 assert(it->state() == activeState, "why spill interval that is not active?"); 5181 assert(it->from() <= min_split_pos, "cannot split before start of interval"); 5182 assert(min_split_pos <= max_split_pos, "invalid order"); 5183 assert(max_split_pos < it->to(), "cannot split at end end of interval"); 5184 assert(current_position() < it->to(), "interval must not end before current position"); 5185 5186 if (min_split_pos == it->from()) { 5187 // the whole interval is never used, so spill it entirely to memory 5188 TRACE_LINEAR_SCAN(2, tty->print_cr(" spilling entire interval because split pos is at beginning of interval")); 5189 assert(it->first_usage(shouldHaveRegister) > current_position(), "interval must not have use position before current_position"); 5190 5191 allocator()->assign_spill_slot(it); 5192 allocator()->change_spill_state(it, min_split_pos); 5193 5194 // Also kick parent intervals out of register to memory when they have no use 5195 // position. This avoids short interval in register surrounded by intervals in 5196 // memory -> avoid useless moves from memory to register and back 5197 Interval* parent = it; 5198 while (parent != NULL && parent->is_split_child()) { 5199 parent = parent->split_child_before_op_id(parent->from()); 5200 5201 if (parent->assigned_reg() < LinearScan::nof_regs) { 5202 if (parent->first_usage(shouldHaveRegister) == max_jint) { 5203 // parent is never used, so kick it out of its assigned register 5204 TRACE_LINEAR_SCAN(4, tty->print_cr(" kicking out interval %d out of its register because it is never used", parent->reg_num())); 5205 allocator()->assign_spill_slot(parent); 5206 } else { 5207 // do not go further back because the register is actually used by the interval 5208 parent = NULL; 5209 } 5210 } 5211 } 5212 5213 } else { 5214 // search optimal split pos, split interval and spill only the right hand part 5215 int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, false); 5216 5217 assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range"); 5218 assert(optimal_split_pos < it->to(), "cannot split at end of interval"); 5219 assert(optimal_split_pos >= it->from(), "cannot split before start of interval"); 5220 5221 if (!allocator()->is_block_begin(optimal_split_pos)) { 5222 // move position before actual instruction (odd op_id) 5223 optimal_split_pos = (optimal_split_pos - 1) | 1; 5224 } 5225 5226 TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d", optimal_split_pos)); 5227 assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary"); 5228 assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary"); 5229 5230 Interval* spilled_part = it->split(optimal_split_pos); 5231 allocator()->append_interval(spilled_part); 5232 allocator()->assign_spill_slot(spilled_part); 5233 allocator()->change_spill_state(spilled_part, optimal_split_pos); 5234 5235 if (!allocator()->is_block_begin(optimal_split_pos)) { 5236 TRACE_LINEAR_SCAN(4, tty->print_cr(" inserting move from interval %d to %d", it->reg_num(), spilled_part->reg_num())); 5237 insert_move(optimal_split_pos, it, spilled_part); 5238 } 5239 5240 // the current_split_child is needed later when moves are inserted for reloading 5241 assert(spilled_part->current_split_child() == it, "overwriting wrong current_split_child"); 5242 spilled_part->make_current_split_child(); 5243 5244 TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts")); 5245 TRACE_LINEAR_SCAN(2, tty->print (" "); it->print()); 5246 TRACE_LINEAR_SCAN(2, tty->print (" "); spilled_part->print()); 5247 } 5248 } 5249 5250 5251 void LinearScanWalker::split_stack_interval(Interval* it) { 5252 int min_split_pos = current_position() + 1; 5253 int max_split_pos = MIN2(it->first_usage(shouldHaveRegister), it->to()); 5254 5255 split_before_usage(it, min_split_pos, max_split_pos); 5256 } 5257 5258 void LinearScanWalker::split_when_partial_register_available(Interval* it, int register_available_until) { 5259 int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, register_available_until), it->from() + 1); 5260 int max_split_pos = register_available_until; 5261 5262 split_before_usage(it, min_split_pos, max_split_pos); 5263 } 5264 5265 void LinearScanWalker::split_and_spill_interval(Interval* it) { 5266 assert(it->state() == activeState || it->state() == inactiveState, "other states not allowed"); 5267 5268 int current_pos = current_position(); 5269 if (it->state() == inactiveState) { 5270 // the interval is currently inactive, so no spill slot is needed for now. 5271 // when the split part is activated, the interval has a new chance to get a register, 5272 // so in the best case no stack slot is necessary 5273 assert(it->has_hole_between(current_pos - 1, current_pos + 1), "interval can not be inactive otherwise"); 5274 split_before_usage(it, current_pos + 1, current_pos + 1); 5275 5276 } else { 5277 // search the position where the interval must have a register and split 5278 // at the optimal position before. 5279 // The new created part is added to the unhandled list and will get a register 5280 // when it is activated 5281 int min_split_pos = current_pos + 1; 5282 int max_split_pos = MIN2(it->next_usage(mustHaveRegister, min_split_pos), it->to()); 5283 5284 split_before_usage(it, min_split_pos, max_split_pos); 5285 5286 assert(it->next_usage(mustHaveRegister, current_pos) == max_jint, "the remaining part is spilled to stack and therefore has no register"); 5287 split_for_spilling(it); 5288 } 5289 } 5290 5291 5292 int LinearScanWalker::find_free_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) { 5293 int min_full_reg = any_reg; 5294 int max_partial_reg = any_reg; 5295 5296 for (int i = _first_reg; i <= _last_reg; i++) { 5297 if (i == ignore_reg) { 5298 // this register must be ignored 5299 5300 } else if (_use_pos[i] >= interval_to) { 5301 // this register is free for the full interval 5302 if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) { 5303 min_full_reg = i; 5304 } 5305 } else if (_use_pos[i] > reg_needed_until) { 5306 // this register is at least free until reg_needed_until 5307 if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) { 5308 max_partial_reg = i; 5309 } 5310 } 5311 } 5312 5313 if (min_full_reg != any_reg) { 5314 return min_full_reg; 5315 } else if (max_partial_reg != any_reg) { 5316 *need_split = true; 5317 return max_partial_reg; 5318 } else { 5319 return any_reg; 5320 } 5321 } 5322 5323 int LinearScanWalker::find_free_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) { 5324 assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm"); 5325 5326 int min_full_reg = any_reg; 5327 int max_partial_reg = any_reg; 5328 5329 for (int i = _first_reg; i < _last_reg; i+=2) { 5330 if (_use_pos[i] >= interval_to && _use_pos[i + 1] >= interval_to) { 5331 // this register is free for the full interval 5332 if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) { 5333 min_full_reg = i; 5334 } 5335 } else if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) { 5336 // this register is at least free until reg_needed_until 5337 if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) { 5338 max_partial_reg = i; 5339 } 5340 } 5341 } 5342 5343 if (min_full_reg != any_reg) { 5344 return min_full_reg; 5345 } else if (max_partial_reg != any_reg) { 5346 *need_split = true; 5347 return max_partial_reg; 5348 } else { 5349 return any_reg; 5350 } 5351 } 5352 5353 5354 bool LinearScanWalker::alloc_free_reg(Interval* cur) { 5355 TRACE_LINEAR_SCAN(2, tty->print("trying to find free register for "); cur->print()); 5356 5357 init_use_lists(true); 5358 free_exclude_active_fixed(); 5359 free_exclude_active_any(); 5360 free_collect_inactive_fixed(cur); 5361 free_collect_inactive_any(cur); 5362 // free_collect_unhandled(fixedKind, cur); 5363 assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0"); 5364 5365 // _use_pos contains the start of the next interval that has this register assigned 5366 // (either as a fixed register or a normal allocated register in the past) 5367 // only intervals overlapping with cur are processed, non-overlapping invervals can be ignored safely 5368 TRACE_LINEAR_SCAN(4, tty->print_cr(" state of registers:")); 5369 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])); 5370 5371 int hint_reg, hint_regHi; 5372 Interval* register_hint = cur->register_hint(); 5373 if (register_hint != NULL) { 5374 hint_reg = register_hint->assigned_reg(); 5375 hint_regHi = register_hint->assigned_regHi(); 5376 5377 if (allocator()->is_precolored_cpu_interval(register_hint)) { 5378 assert(hint_reg != any_reg && hint_regHi == any_reg, "must be for fixed intervals"); 5379 hint_regHi = hint_reg + 1; // connect e.g. eax-edx 5380 } 5381 TRACE_LINEAR_SCAN(4, tty->print(" hint registers %d, %d from interval ", hint_reg, hint_regHi); register_hint->print()); 5382 5383 } else { 5384 hint_reg = any_reg; 5385 hint_regHi = any_reg; 5386 } 5387 assert(hint_reg == any_reg || hint_reg != hint_regHi, "hint reg and regHi equal"); 5388 assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned to interval"); 5389 5390 // the register must be free at least until this position 5391 int reg_needed_until = cur->from() + 1; 5392 int interval_to = cur->to(); 5393 5394 bool need_split = false; 5395 int split_pos = -1; 5396 int reg = any_reg; 5397 int regHi = any_reg; 5398 5399 if (_adjacent_regs) { 5400 reg = find_free_double_reg(reg_needed_until, interval_to, hint_reg, &need_split); 5401 regHi = reg + 1; 5402 if (reg == any_reg) { 5403 return false; 5404 } 5405 split_pos = MIN2(_use_pos[reg], _use_pos[regHi]); 5406 5407 } else { 5408 reg = find_free_reg(reg_needed_until, interval_to, hint_reg, any_reg, &need_split); 5409 if (reg == any_reg) { 5410 return false; 5411 } 5412 split_pos = _use_pos[reg]; 5413 5414 if (_num_phys_regs == 2) { 5415 regHi = find_free_reg(reg_needed_until, interval_to, hint_regHi, reg, &need_split); 5416 5417 if (_use_pos[reg] < interval_to && regHi == any_reg) { 5418 // do not split interval if only one register can be assigned until the split pos 5419 // (when one register is found for the whole interval, split&spill is only 5420 // performed for the hi register) 5421 return false; 5422 5423 } else if (regHi != any_reg) { 5424 split_pos = MIN2(split_pos, _use_pos[regHi]); 5425 5426 // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax 5427 if (reg > regHi) { 5428 int temp = reg; 5429 reg = regHi; 5430 regHi = temp; 5431 } 5432 } 5433 } 5434 } 5435 5436 cur->assign_reg(reg, regHi); 5437 TRACE_LINEAR_SCAN(2, tty->print_cr("selected register %d, %d", reg, regHi)); 5438 5439 assert(split_pos > 0, "invalid split_pos"); 5440 if (need_split) { 5441 // register not available for full interval, so split it 5442 split_when_partial_register_available(cur, split_pos); 5443 } 5444 5445 // only return true if interval is completely assigned 5446 return _num_phys_regs == 1 || regHi != any_reg; 5447 } 5448 5449 5450 int LinearScanWalker::find_locked_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) { 5451 int max_reg = any_reg; 5452 5453 for (int i = _first_reg; i <= _last_reg; i++) { 5454 if (i == ignore_reg) { 5455 // this register must be ignored 5456 5457 } else if (_use_pos[i] > reg_needed_until) { 5458 if (max_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_reg] && max_reg != hint_reg)) { 5459 max_reg = i; 5460 } 5461 } 5462 } 5463 5464 if (max_reg != any_reg && _block_pos[max_reg] <= interval_to) { 5465 *need_split = true; 5466 } 5467 5468 return max_reg; 5469 } 5470 5471 int LinearScanWalker::find_locked_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) { 5472 assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm"); 5473 5474 int max_reg = any_reg; 5475 5476 for (int i = _first_reg; i < _last_reg; i+=2) { 5477 if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) { 5478 if (max_reg == any_reg || _use_pos[i] > _use_pos[max_reg]) { 5479 max_reg = i; 5480 } 5481 } 5482 } 5483 5484 if (_block_pos[max_reg] <= interval_to || _block_pos[max_reg + 1] <= interval_to) { 5485 *need_split = true; 5486 } 5487 5488 return max_reg; 5489 } 5490 5491 void LinearScanWalker::split_and_spill_intersecting_intervals(int reg, int regHi) { 5492 assert(reg != any_reg, "no register assigned"); 5493 5494 for (int i = 0; i < _spill_intervals[reg]->length(); i++) { 5495 Interval* it = _spill_intervals[reg]->at(i); 5496 remove_from_list(it); 5497 split_and_spill_interval(it); 5498 } 5499 5500 if (regHi != any_reg) { 5501 IntervalList* processed = _spill_intervals[reg]; 5502 for (int i = 0; i < _spill_intervals[regHi]->length(); i++) { 5503 Interval* it = _spill_intervals[regHi]->at(i); 5504 if (processed->index_of(it) == -1) { 5505 remove_from_list(it); 5506 split_and_spill_interval(it); 5507 } 5508 } 5509 } 5510 } 5511 5512 5513 // Split an Interval and spill it to memory so that cur can be placed in a register 5514 void LinearScanWalker::alloc_locked_reg(Interval* cur) { 5515 TRACE_LINEAR_SCAN(2, tty->print("need to split and spill to get register for "); cur->print()); 5516 5517 // collect current usage of registers 5518 init_use_lists(false); 5519 spill_exclude_active_fixed(); 5520 // spill_block_unhandled_fixed(cur); 5521 assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0"); 5522 spill_block_inactive_fixed(cur); 5523 spill_collect_active_any(); 5524 spill_collect_inactive_any(cur); 5525 5526 #ifndef PRODUCT 5527 if (TraceLinearScanLevel >= 4) { 5528 tty->print_cr(" state of registers:"); 5529 for (int i = _first_reg; i <= _last_reg; i++) { 5530 tty->print(" reg %d: use_pos: %d, block_pos: %d, intervals: ", i, _use_pos[i], _block_pos[i]); 5531 for (int j = 0; j < _spill_intervals[i]->length(); j++) { 5532 tty->print("%d ", _spill_intervals[i]->at(j)->reg_num()); 5533 } 5534 tty->cr(); 5535 } 5536 } 5537 #endif 5538 5539 // the register must be free at least until this position 5540 int reg_needed_until = MIN2(cur->first_usage(mustHaveRegister), cur->from() + 1); 5541 int interval_to = cur->to(); 5542 assert (reg_needed_until > 0 && reg_needed_until < max_jint, "interval has no use"); 5543 5544 int split_pos = 0; 5545 int use_pos = 0; 5546 bool need_split = false; 5547 int reg, regHi; 5548 5549 if (_adjacent_regs) { 5550 reg = find_locked_double_reg(reg_needed_until, interval_to, any_reg, &need_split); 5551 regHi = reg + 1; 5552 5553 if (reg != any_reg) { 5554 use_pos = MIN2(_use_pos[reg], _use_pos[regHi]); 5555 split_pos = MIN2(_block_pos[reg], _block_pos[regHi]); 5556 } 5557 } else { 5558 reg = find_locked_reg(reg_needed_until, interval_to, any_reg, cur->assigned_reg(), &need_split); 5559 regHi = any_reg; 5560 5561 if (reg != any_reg) { 5562 use_pos = _use_pos[reg]; 5563 split_pos = _block_pos[reg]; 5564 5565 if (_num_phys_regs == 2) { 5566 if (cur->assigned_reg() != any_reg) { 5567 regHi = reg; 5568 reg = cur->assigned_reg(); 5569 } else { 5570 regHi = find_locked_reg(reg_needed_until, interval_to, any_reg, reg, &need_split); 5571 if (regHi != any_reg) { 5572 use_pos = MIN2(use_pos, _use_pos[regHi]); 5573 split_pos = MIN2(split_pos, _block_pos[regHi]); 5574 } 5575 } 5576 5577 if (regHi != any_reg && reg > regHi) { 5578 // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax 5579 int temp = reg; 5580 reg = regHi; 5581 regHi = temp; 5582 } 5583 } 5584 } 5585 } 5586 5587 if (reg == any_reg || (_num_phys_regs == 2 && regHi == any_reg) || use_pos <= cur->first_usage(mustHaveRegister)) { 5588 // the first use of cur is later than the spilling position -> spill cur 5589 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)); 5590 5591 if (cur->first_usage(mustHaveRegister) <= cur->from() + 1) { 5592 assert(false, "cannot spill interval that is used in first instruction (possible reason: no register found)"); 5593 // assign a reasonable register and do a bailout in product mode to avoid errors 5594 allocator()->assign_spill_slot(cur); 5595 BAILOUT("LinearScan: no register found"); 5596 } 5597 5598 split_and_spill_interval(cur); 5599 } else { 5600 TRACE_LINEAR_SCAN(4, tty->print_cr("decided to use register %d, %d", reg, regHi)); 5601 assert(reg != any_reg && (_num_phys_regs == 1 || regHi != any_reg), "no register found"); 5602 assert(split_pos > 0, "invalid split_pos"); 5603 assert(need_split == false || split_pos > cur->from(), "splitting interval at from"); 5604 5605 cur->assign_reg(reg, regHi); 5606 if (need_split) { 5607 // register not available for full interval, so split it 5608 split_when_partial_register_available(cur, split_pos); 5609 } 5610 5611 // perform splitting and spilling for all affected intervalls 5612 split_and_spill_intersecting_intervals(reg, regHi); 5613 } 5614 } 5615 5616 bool LinearScanWalker::no_allocation_possible(Interval* cur) { 5617 #ifdef X86 5618 // fast calculation of intervals that can never get a register because the 5619 // the next instruction is a call that blocks all registers 5620 // Note: this does not work if callee-saved registers are available (e.g. on Sparc) 5621 5622 // check if this interval is the result of a split operation 5623 // (an interval got a register until this position) 5624 int pos = cur->from(); 5625 if ((pos & 1) == 1) { 5626 // the current instruction is a call that blocks all registers 5627 if (pos < allocator()->max_lir_op_id() && allocator()->has_call(pos + 1)) { 5628 TRACE_LINEAR_SCAN(4, tty->print_cr(" free register cannot be available because all registers blocked by following call")); 5629 5630 // safety check that there is really no register available 5631 assert(alloc_free_reg(cur) == false, "found a register for this interval"); 5632 return true; 5633 } 5634 5635 } 5636 #endif 5637 return false; 5638 } 5639 5640 void LinearScanWalker::init_vars_for_alloc(Interval* cur) { 5641 BasicType type = cur->type(); 5642 _num_phys_regs = LinearScan::num_physical_regs(type); 5643 _adjacent_regs = LinearScan::requires_adjacent_regs(type); 5644 5645 if (pd_init_regs_for_alloc(cur)) { 5646 // the appropriate register range was selected. 5647 } else if (type == T_FLOAT || type == T_DOUBLE) { 5648 _first_reg = pd_first_fpu_reg; 5649 _last_reg = pd_last_fpu_reg; 5650 } else { 5651 _first_reg = pd_first_cpu_reg; 5652 _last_reg = FrameMap::last_cpu_reg(); 5653 } 5654 5655 assert(0 <= _first_reg && _first_reg < LinearScan::nof_regs, "out of range"); 5656 assert(0 <= _last_reg && _last_reg < LinearScan::nof_regs, "out of range"); 5657 } 5658 5659 5660 bool LinearScanWalker::is_move(LIR_Op* op, Interval* from, Interval* to) { 5661 if (op->code() != lir_move) { 5662 return false; 5663 } 5664 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); 5665 5666 LIR_Opr in = ((LIR_Op1*)op)->in_opr(); 5667 LIR_Opr res = ((LIR_Op1*)op)->result_opr(); 5668 return in->is_virtual() && res->is_virtual() && in->vreg_number() == from->reg_num() && res->vreg_number() == to->reg_num(); 5669 } 5670 5671 // optimization (especially for phi functions of nested loops): 5672 // assign same spill slot to non-intersecting intervals 5673 void LinearScanWalker::combine_spilled_intervals(Interval* cur) { 5674 if (cur->is_split_child()) { 5675 // optimization is only suitable for split parents 5676 return; 5677 } 5678 5679 Interval* register_hint = cur->register_hint(false); 5680 if (register_hint == NULL) { 5681 // cur is not the target of a move, otherwise register_hint would be set 5682 return; 5683 } 5684 assert(register_hint->is_split_parent(), "register hint must be split parent"); 5685 5686 if (cur->spill_state() != noOptimization || register_hint->spill_state() != noOptimization) { 5687 // combining the stack slots for intervals where spill move optimization is applied 5688 // is not benefitial and would cause problems 5689 return; 5690 } 5691 5692 int begin_pos = cur->from(); 5693 int end_pos = cur->to(); 5694 if (end_pos > allocator()->max_lir_op_id() || (begin_pos & 1) != 0 || (end_pos & 1) != 0) { 5695 // safety check that lir_op_with_id is allowed 5696 return; 5697 } 5698 5699 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)) { 5700 // cur and register_hint are not connected with two moves 5701 return; 5702 } 5703 5704 Interval* begin_hint = register_hint->split_child_at_op_id(begin_pos, LIR_OpVisitState::inputMode); 5705 Interval* end_hint = register_hint->split_child_at_op_id(end_pos, LIR_OpVisitState::outputMode); 5706 if (begin_hint == end_hint || begin_hint->to() != begin_pos || end_hint->from() != end_pos) { 5707 // register_hint must be split, otherwise the re-writing of use positions does not work 5708 return; 5709 } 5710 5711 assert(begin_hint->assigned_reg() != any_reg, "must have register assigned"); 5712 assert(end_hint->assigned_reg() == any_reg, "must not have register assigned"); 5713 assert(cur->first_usage(mustHaveRegister) == begin_pos, "must have use position at begin of interval because of move"); 5714 assert(end_hint->first_usage(mustHaveRegister) == end_pos, "must have use position at begin of interval because of move"); 5715 5716 if (begin_hint->assigned_reg() < LinearScan::nof_regs) { 5717 // register_hint is not spilled at begin_pos, so it would not be benefitial to immediately spill cur 5718 return; 5719 } 5720 assert(register_hint->canonical_spill_slot() != -1, "must be set when part of interval was spilled"); 5721 5722 // modify intervals such that cur gets the same stack slot as register_hint 5723 // delete use positions to prevent the intervals to get a register at beginning 5724 cur->set_canonical_spill_slot(register_hint->canonical_spill_slot()); 5725 cur->remove_first_use_pos(); 5726 end_hint->remove_first_use_pos(); 5727 } 5728 5729 5730 // allocate a physical register or memory location to an interval 5731 bool LinearScanWalker::activate_current() { 5732 Interval* cur = current(); 5733 bool result = true; 5734 5735 TRACE_LINEAR_SCAN(2, tty->print ("+++++ activating interval "); cur->print()); 5736 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())); 5737 5738 if (cur->assigned_reg() >= LinearScan::nof_regs) { 5739 // activating an interval that has a stack slot assigned -> split it at first use position 5740 // used for method parameters 5741 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has spill slot assigned (method parameter) -> split it before first use")); 5742 5743 split_stack_interval(cur); 5744 result = false; 5745 5746 } else if (allocator()->gen()->is_vreg_flag_set(cur->reg_num(), LIRGenerator::must_start_in_memory)) { 5747 // activating an interval that must start in a stack slot, but may get a register later 5748 // used for lir_roundfp: rounding is done by store to stack and reload later 5749 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval must start in stack slot -> split it before first use")); 5750 assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned"); 5751 5752 allocator()->assign_spill_slot(cur); 5753 split_stack_interval(cur); 5754 result = false; 5755 5756 } else if (cur->assigned_reg() == any_reg) { 5757 // interval has not assigned register -> normal allocation 5758 // (this is the normal case for most intervals) 5759 TRACE_LINEAR_SCAN(4, tty->print_cr(" normal allocation of register")); 5760 5761 // assign same spill slot to non-intersecting intervals 5762 combine_spilled_intervals(cur); 5763 5764 init_vars_for_alloc(cur); 5765 if (no_allocation_possible(cur) || !alloc_free_reg(cur)) { 5766 // no empty register available. 5767 // split and spill another interval so that this interval gets a register 5768 alloc_locked_reg(cur); 5769 } 5770 5771 // spilled intervals need not be move to active-list 5772 if (cur->assigned_reg() >= LinearScan::nof_regs) { 5773 result = false; 5774 } 5775 } 5776 5777 // load spilled values that become active from stack slot to register 5778 if (cur->insert_move_when_activated()) { 5779 assert(cur->is_split_child(), "must be"); 5780 assert(cur->current_split_child() != NULL, "must be"); 5781 assert(cur->current_split_child()->reg_num() != cur->reg_num(), "cannot insert move between same interval"); 5782 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())); 5783 5784 insert_move(cur->from(), cur->current_split_child(), cur); 5785 } 5786 cur->make_current_split_child(); 5787 5788 return result; // true = interval is moved to active list 5789 } 5790 5791 5792 // Implementation of EdgeMoveOptimizer 5793 5794 EdgeMoveOptimizer::EdgeMoveOptimizer() : 5795 _edge_instructions(4), 5796 _edge_instructions_idx(4) 5797 { 5798 } 5799 5800 void EdgeMoveOptimizer::optimize(BlockList* code) { 5801 EdgeMoveOptimizer optimizer = EdgeMoveOptimizer(); 5802 5803 // ignore the first block in the list (index 0 is not processed) 5804 for (int i = code->length() - 1; i >= 1; i--) { 5805 BlockBegin* block = code->at(i); 5806 5807 if (block->number_of_preds() > 1 && !block->is_set(BlockBegin::exception_entry_flag)) { 5808 optimizer.optimize_moves_at_block_end(block); 5809 } 5810 if (block->number_of_sux() == 2) { 5811 optimizer.optimize_moves_at_block_begin(block); 5812 } 5813 } 5814 } 5815 5816 5817 // clear all internal data structures 5818 void EdgeMoveOptimizer::init_instructions() { 5819 _edge_instructions.clear(); 5820 _edge_instructions_idx.clear(); 5821 } 5822 5823 // append a lir-instruction-list and the index of the current operation in to the list 5824 void EdgeMoveOptimizer::append_instructions(LIR_OpList* instructions, int instructions_idx) { 5825 _edge_instructions.append(instructions); 5826 _edge_instructions_idx.append(instructions_idx); 5827 } 5828 5829 // return the current operation of the given edge (predecessor or successor) 5830 LIR_Op* EdgeMoveOptimizer::instruction_at(int edge) { 5831 LIR_OpList* instructions = _edge_instructions.at(edge); 5832 int idx = _edge_instructions_idx.at(edge); 5833 5834 if (idx < instructions->length()) { 5835 return instructions->at(idx); 5836 } else { 5837 return NULL; 5838 } 5839 } 5840 5841 // removes the current operation of the given edge (predecessor or successor) 5842 void EdgeMoveOptimizer::remove_cur_instruction(int edge, bool decrement_index) { 5843 LIR_OpList* instructions = _edge_instructions.at(edge); 5844 int idx = _edge_instructions_idx.at(edge); 5845 instructions->remove_at(idx); 5846 5847 if (decrement_index) { 5848 _edge_instructions_idx.at_put(edge, idx - 1); 5849 } 5850 } 5851 5852 5853 bool EdgeMoveOptimizer::operations_different(LIR_Op* op1, LIR_Op* op2) { 5854 if (op1 == NULL || op2 == NULL) { 5855 // at least one block is already empty -> no optimization possible 5856 return true; 5857 } 5858 5859 if (op1->code() == lir_move && op2->code() == lir_move) { 5860 assert(op1->as_Op1() != NULL, "move must be LIR_Op1"); 5861 assert(op2->as_Op1() != NULL, "move must be LIR_Op1"); 5862 LIR_Op1* move1 = (LIR_Op1*)op1; 5863 LIR_Op1* move2 = (LIR_Op1*)op2; 5864 if (move1->info() == move2->info() && move1->in_opr() == move2->in_opr() && move1->result_opr() == move2->result_opr()) { 5865 // these moves are exactly equal and can be optimized 5866 return false; 5867 } 5868 5869 } else if (op1->code() == lir_fxch && op2->code() == lir_fxch) { 5870 assert(op1->as_Op1() != NULL, "fxch must be LIR_Op1"); 5871 assert(op2->as_Op1() != NULL, "fxch must be LIR_Op1"); 5872 LIR_Op1* fxch1 = (LIR_Op1*)op1; 5873 LIR_Op1* fxch2 = (LIR_Op1*)op2; 5874 if (fxch1->in_opr()->as_jint() == fxch2->in_opr()->as_jint()) { 5875 // equal FPU stack operations can be optimized 5876 return false; 5877 } 5878 5879 } else if (op1->code() == lir_fpop_raw && op2->code() == lir_fpop_raw) { 5880 // equal FPU stack operations can be optimized 5881 return false; 5882 } 5883 5884 // no optimization possible 5885 return true; 5886 } 5887 5888 void EdgeMoveOptimizer::optimize_moves_at_block_end(BlockBegin* block) { 5889 TRACE_LINEAR_SCAN(4, tty->print_cr("optimizing moves at end of block B%d", block->block_id())); 5890 5891 if (block->is_predecessor(block)) { 5892 // currently we can't handle this correctly. 5893 return; 5894 } 5895 5896 init_instructions(); 5897 int num_preds = block->number_of_preds(); 5898 assert(num_preds > 1, "do not call otherwise"); 5899 assert(!block->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed"); 5900 5901 // setup a list with the lir-instructions of all predecessors 5902 int i; 5903 for (i = 0; i < num_preds; i++) { 5904 BlockBegin* pred = block->pred_at(i); 5905 LIR_OpList* pred_instructions = pred->lir()->instructions_list(); 5906 5907 if (pred->number_of_sux() != 1) { 5908 // this can happen with switch-statements where multiple edges are between 5909 // the same blocks. 5910 return; 5911 } 5912 5913 assert(pred->number_of_sux() == 1, "can handle only one successor"); 5914 assert(pred->sux_at(0) == block, "invalid control flow"); 5915 assert(pred_instructions->last()->code() == lir_branch, "block with successor must end with branch"); 5916 assert(pred_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch"); 5917 assert(pred_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch"); 5918 5919 if (pred_instructions->last()->info() != NULL) { 5920 // can not optimize instructions when debug info is needed 5921 return; 5922 } 5923 5924 // ignore the unconditional branch at the end of the block 5925 append_instructions(pred_instructions, pred_instructions->length() - 2); 5926 } 5927 5928 5929 // process lir-instructions while all predecessors end with the same instruction 5930 while (true) { 5931 LIR_Op* op = instruction_at(0); 5932 for (i = 1; i < num_preds; i++) { 5933 if (operations_different(op, instruction_at(i))) { 5934 // these instructions are different and cannot be optimized -> 5935 // no further optimization possible 5936 return; 5937 } 5938 } 5939 5940 TRACE_LINEAR_SCAN(4, tty->print("found instruction that is equal in all %d predecessors: ", num_preds); op->print()); 5941 5942 // insert the instruction at the beginning of the current block 5943 block->lir()->insert_before(1, op); 5944 5945 // delete the instruction at the end of all predecessors 5946 for (i = 0; i < num_preds; i++) { 5947 remove_cur_instruction(i, true); 5948 } 5949 } 5950 } 5951 5952 5953 void EdgeMoveOptimizer::optimize_moves_at_block_begin(BlockBegin* block) { 5954 TRACE_LINEAR_SCAN(4, tty->print_cr("optimization moves at begin of block B%d", block->block_id())); 5955 5956 init_instructions(); 5957 int num_sux = block->number_of_sux(); 5958 5959 LIR_OpList* cur_instructions = block->lir()->instructions_list(); 5960 5961 assert(num_sux == 2, "method should not be called otherwise"); 5962 assert(cur_instructions->last()->code() == lir_branch, "block with successor must end with branch"); 5963 assert(cur_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch"); 5964 assert(cur_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch"); 5965 5966 if (cur_instructions->last()->info() != NULL) { 5967 // can no optimize instructions when debug info is needed 5968 return; 5969 } 5970 5971 LIR_Op* branch = cur_instructions->at(cur_instructions->length() - 2); 5972 if (branch->info() != NULL || (branch->code() != lir_branch && branch->code() != lir_cond_float_branch)) { 5973 // not a valid case for optimization 5974 // currently, only blocks that end with two branches (conditional branch followed 5975 // by unconditional branch) are optimized 5976 return; 5977 } 5978 5979 // now it is guaranteed that the block ends with two branch instructions. 5980 // the instructions are inserted at the end of the block before these two branches 5981 int insert_idx = cur_instructions->length() - 2; 5982 5983 int i; 5984 #ifdef ASSERT 5985 for (i = insert_idx - 1; i >= 0; i--) { 5986 LIR_Op* op = cur_instructions->at(i); 5987 if ((op->code() == lir_branch || op->code() == lir_cond_float_branch) && ((LIR_OpBranch*)op)->block() != NULL) { 5988 assert(false, "block with two successors can have only two branch instructions"); 5989 } 5990 } 5991 #endif 5992 5993 // setup a list with the lir-instructions of all successors 5994 for (i = 0; i < num_sux; i++) { 5995 BlockBegin* sux = block->sux_at(i); 5996 LIR_OpList* sux_instructions = sux->lir()->instructions_list(); 5997 5998 assert(sux_instructions->at(0)->code() == lir_label, "block must start with label"); 5999 6000 if (sux->number_of_preds() != 1) { 6001 // this can happen with switch-statements where multiple edges are between 6002 // the same blocks. 6003 return; 6004 } 6005 assert(sux->pred_at(0) == block, "invalid control flow"); 6006 assert(!sux->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed"); 6007 6008 // ignore the label at the beginning of the block 6009 append_instructions(sux_instructions, 1); 6010 } 6011 6012 // process lir-instructions while all successors begin with the same instruction 6013 while (true) { 6014 LIR_Op* op = instruction_at(0); 6015 for (i = 1; i < num_sux; i++) { 6016 if (operations_different(op, instruction_at(i))) { 6017 // these instructions are different and cannot be optimized -> 6018 // no further optimization possible 6019 return; 6020 } 6021 } 6022 6023 TRACE_LINEAR_SCAN(4, tty->print("----- found instruction that is equal in all %d successors: ", num_sux); op->print()); 6024 6025 // insert instruction at end of current block 6026 block->lir()->insert_before(insert_idx, op); 6027 insert_idx++; 6028 6029 // delete the instructions at the beginning of all successors 6030 for (i = 0; i < num_sux; i++) { 6031 remove_cur_instruction(i, false); 6032 } 6033 } 6034 } 6035 6036 6037 // Implementation of ControlFlowOptimizer 6038 6039 ControlFlowOptimizer::ControlFlowOptimizer() : 6040 _original_preds(4) 6041 { 6042 } 6043 6044 void ControlFlowOptimizer::optimize(BlockList* code) { 6045 ControlFlowOptimizer optimizer = ControlFlowOptimizer(); 6046 6047 // push the OSR entry block to the end so that we're not jumping over it. 6048 BlockBegin* osr_entry = code->at(0)->end()->as_Base()->osr_entry(); 6049 if (osr_entry) { 6050 int index = osr_entry->linear_scan_number(); 6051 assert(code->at(index) == osr_entry, "wrong index"); 6052 code->remove_at(index); 6053 code->append(osr_entry); 6054 } 6055 6056 optimizer.reorder_short_loops(code); 6057 optimizer.delete_empty_blocks(code); 6058 optimizer.delete_unnecessary_jumps(code); 6059 optimizer.delete_jumps_to_return(code); 6060 } 6061 6062 void ControlFlowOptimizer::reorder_short_loop(BlockList* code, BlockBegin* header_block, int header_idx) { 6063 int i = header_idx + 1; 6064 int max_end = MIN2(header_idx + ShortLoopSize, code->length()); 6065 while (i < max_end && code->at(i)->loop_depth() >= header_block->loop_depth()) { 6066 i++; 6067 } 6068 6069 if (i == code->length() || code->at(i)->loop_depth() < header_block->loop_depth()) { 6070 int end_idx = i - 1; 6071 BlockBegin* end_block = code->at(end_idx); 6072 6073 if (end_block->number_of_sux() == 1 && end_block->sux_at(0) == header_block) { 6074 // short loop from header_idx to end_idx found -> reorder blocks such that 6075 // the header_block is the last block instead of the first block of the loop 6076 TRACE_LINEAR_SCAN(1, tty->print_cr("Reordering short loop: length %d, header B%d, end B%d", 6077 end_idx - header_idx + 1, 6078 header_block->block_id(), end_block->block_id())); 6079 6080 for (int j = header_idx; j < end_idx; j++) { 6081 code->at_put(j, code->at(j + 1)); 6082 } 6083 code->at_put(end_idx, header_block); 6084 6085 // correct the flags so that any loop alignment occurs in the right place. 6086 assert(code->at(end_idx)->is_set(BlockBegin::backward_branch_target_flag), "must be backward branch target"); 6087 code->at(end_idx)->clear(BlockBegin::backward_branch_target_flag); 6088 code->at(header_idx)->set(BlockBegin::backward_branch_target_flag); 6089 } 6090 } 6091 } 6092 6093 void ControlFlowOptimizer::reorder_short_loops(BlockList* code) { 6094 for (int i = code->length() - 1; i >= 0; i--) { 6095 BlockBegin* block = code->at(i); 6096 6097 if (block->is_set(BlockBegin::linear_scan_loop_header_flag)) { 6098 reorder_short_loop(code, block, i); 6099 } 6100 } 6101 6102 DEBUG_ONLY(verify(code)); 6103 } 6104 6105 // only blocks with exactly one successor can be deleted. Such blocks 6106 // must always end with an unconditional branch to this successor 6107 bool ControlFlowOptimizer::can_delete_block(BlockBegin* block) { 6108 if (block->number_of_sux() != 1 || block->number_of_exception_handlers() != 0 || block->is_entry_block()) { 6109 return false; 6110 } 6111 6112 LIR_OpList* instructions = block->lir()->instructions_list(); 6113 6114 assert(instructions->length() >= 2, "block must have label and branch"); 6115 assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label"); 6116 assert(instructions->last()->as_OpBranch() != NULL, "last instrcution must always be a branch"); 6117 assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "branch must be unconditional"); 6118 assert(instructions->last()->as_OpBranch()->block() == block->sux_at(0), "branch target must be the successor"); 6119 6120 // block must have exactly one successor 6121 6122 if (instructions->length() == 2 && instructions->last()->info() == NULL) { 6123 return true; 6124 } 6125 return false; 6126 } 6127 6128 // substitute branch targets in all branch-instructions of this blocks 6129 void ControlFlowOptimizer::substitute_branch_target(BlockBegin* block, BlockBegin* target_from, BlockBegin* target_to) { 6130 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())); 6131 6132 LIR_OpList* instructions = block->lir()->instructions_list(); 6133 6134 assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label"); 6135 for (int i = instructions->length() - 1; i >= 1; i--) { 6136 LIR_Op* op = instructions->at(i); 6137 6138 if (op->code() == lir_branch || op->code() == lir_cond_float_branch) { 6139 assert(op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch"); 6140 LIR_OpBranch* branch = (LIR_OpBranch*)op; 6141 6142 if (branch->block() == target_from) { 6143 branch->change_block(target_to); 6144 } 6145 if (branch->ublock() == target_from) { 6146 branch->change_ublock(target_to); 6147 } 6148 } 6149 } 6150 } 6151 6152 void ControlFlowOptimizer::delete_empty_blocks(BlockList* code) { 6153 int old_pos = 0; 6154 int new_pos = 0; 6155 int num_blocks = code->length(); 6156 6157 while (old_pos < num_blocks) { 6158 BlockBegin* block = code->at(old_pos); 6159 6160 if (can_delete_block(block)) { 6161 BlockBegin* new_target = block->sux_at(0); 6162 6163 // propagate backward branch target flag for correct code alignment 6164 if (block->is_set(BlockBegin::backward_branch_target_flag)) { 6165 new_target->set(BlockBegin::backward_branch_target_flag); 6166 } 6167 6168 // collect a list with all predecessors that contains each predecessor only once 6169 // the predecessors of cur are changed during the substitution, so a copy of the 6170 // predecessor list is necessary 6171 int j; 6172 _original_preds.clear(); 6173 for (j = block->number_of_preds() - 1; j >= 0; j--) { 6174 BlockBegin* pred = block->pred_at(j); 6175 if (_original_preds.index_of(pred) == -1) { 6176 _original_preds.append(pred); 6177 } 6178 } 6179 6180 for (j = _original_preds.length() - 1; j >= 0; j--) { 6181 BlockBegin* pred = _original_preds.at(j); 6182 substitute_branch_target(pred, block, new_target); 6183 pred->substitute_sux(block, new_target); 6184 } 6185 } else { 6186 // adjust position of this block in the block list if blocks before 6187 // have been deleted 6188 if (new_pos != old_pos) { 6189 code->at_put(new_pos, code->at(old_pos)); 6190 } 6191 new_pos++; 6192 } 6193 old_pos++; 6194 } 6195 code->truncate(new_pos); 6196 6197 DEBUG_ONLY(verify(code)); 6198 } 6199 6200 void ControlFlowOptimizer::delete_unnecessary_jumps(BlockList* code) { 6201 // skip the last block because there a branch is always necessary 6202 for (int i = code->length() - 2; i >= 0; i--) { 6203 BlockBegin* block = code->at(i); 6204 LIR_OpList* instructions = block->lir()->instructions_list(); 6205 6206 LIR_Op* last_op = instructions->last(); 6207 if (last_op->code() == lir_branch) { 6208 assert(last_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch"); 6209 LIR_OpBranch* last_branch = (LIR_OpBranch*)last_op; 6210 6211 assert(last_branch->block() != NULL, "last branch must always have a block as target"); 6212 assert(last_branch->label() == last_branch->block()->label(), "must be equal"); 6213 6214 if (last_branch->info() == NULL) { 6215 if (last_branch->block() == code->at(i + 1)) { 6216 6217 TRACE_LINEAR_SCAN(3, tty->print_cr("Deleting unconditional branch at end of block B%d", block->block_id())); 6218 6219 // delete last branch instruction 6220 instructions->truncate(instructions->length() - 1); 6221 6222 } else { 6223 LIR_Op* prev_op = instructions->at(instructions->length() - 2); 6224 if (prev_op->code() == lir_branch || prev_op->code() == lir_cond_float_branch) { 6225 assert(prev_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch"); 6226 LIR_OpBranch* prev_branch = (LIR_OpBranch*)prev_op; 6227 6228 LIR_Op2* prev_cmp = NULL; 6229 6230 for(int j = instructions->length() - 3; j >= 0 && prev_cmp == NULL; j--) { 6231 prev_op = instructions->at(j); 6232 if(prev_op->code() == lir_cmp) { 6233 assert(prev_op->as_Op2() != NULL, "branch must be of type LIR_Op2"); 6234 prev_cmp = (LIR_Op2*)prev_op; 6235 assert(prev_branch->cond() == prev_cmp->condition(), "should be the same"); 6236 } 6237 } 6238 assert(prev_cmp != NULL, "should have found comp instruction for branch"); 6239 if (prev_branch->block() == code->at(i + 1) && prev_branch->info() == NULL) { 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())); 6242 6243 // eliminate a conditional branch to the immediate successor 6244 prev_branch->change_block(last_branch->block()); 6245 prev_branch->negate_cond(); 6246 prev_cmp->set_condition(prev_branch->cond()); 6247 instructions->truncate(instructions->length() - 1); 6248 } 6249 } 6250 } 6251 } 6252 } 6253 } 6254 6255 DEBUG_ONLY(verify(code)); 6256 } 6257 6258 void ControlFlowOptimizer::delete_jumps_to_return(BlockList* code) { 6259 #ifdef ASSERT 6260 BitMap return_converted(BlockBegin::number_of_blocks()); 6261 return_converted.clear(); 6262 #endif 6263 6264 for (int i = code->length() - 1; i >= 0; i--) { 6265 BlockBegin* block = code->at(i); 6266 LIR_OpList* cur_instructions = block->lir()->instructions_list(); 6267 LIR_Op* cur_last_op = cur_instructions->last(); 6268 6269 assert(cur_instructions->at(0)->code() == lir_label, "first instruction must always be a label"); 6270 if (cur_instructions->length() == 2 && cur_last_op->code() == lir_return) { 6271 // the block contains only a label and a return 6272 // if a predecessor ends with an unconditional jump to this block, then the jump 6273 // can be replaced with a return instruction 6274 // 6275 // Note: the original block with only a return statement cannot be deleted completely 6276 // because the predecessors might have other (conditional) jumps to this block 6277 // -> this may lead to unnecesary return instructions in the final code 6278 6279 assert(cur_last_op->info() == NULL, "return instructions do not have debug information"); 6280 assert(block->number_of_sux() == 0 || 6281 (return_converted.at(block->block_id()) && block->number_of_sux() == 1), 6282 "blocks that end with return must not have successors"); 6283 6284 assert(cur_last_op->as_Op1() != NULL, "return must be LIR_Op1"); 6285 LIR_Opr return_opr = ((LIR_Op1*)cur_last_op)->in_opr(); 6286 6287 for (int j = block->number_of_preds() - 1; j >= 0; j--) { 6288 BlockBegin* pred = block->pred_at(j); 6289 LIR_OpList* pred_instructions = pred->lir()->instructions_list(); 6290 LIR_Op* pred_last_op = pred_instructions->last(); 6291 6292 if (pred_last_op->code() == lir_branch) { 6293 assert(pred_last_op->as_OpBranch() != NULL, "branch must be LIR_OpBranch"); 6294 LIR_OpBranch* pred_last_branch = (LIR_OpBranch*)pred_last_op; 6295 6296 if (pred_last_branch->block() == block && pred_last_branch->cond() == lir_cond_always && pred_last_branch->info() == NULL) { 6297 // replace the jump to a return with a direct return 6298 // Note: currently the edge between the blocks is not deleted 6299 pred_instructions->at_put(pred_instructions->length() - 1, new LIR_Op1(lir_return, return_opr)); 6300 #ifdef ASSERT 6301 return_converted.set_bit(pred->block_id()); 6302 #endif 6303 } 6304 } 6305 } 6306 } 6307 } 6308 } 6309 6310 6311 #ifdef ASSERT 6312 void ControlFlowOptimizer::verify(BlockList* code) { 6313 for (int i = 0; i < code->length(); i++) { 6314 BlockBegin* block = code->at(i); 6315 LIR_OpList* instructions = block->lir()->instructions_list(); 6316 6317 int j; 6318 for (j = 0; j < instructions->length(); j++) { 6319 LIR_OpBranch* op_branch = instructions->at(j)->as_OpBranch(); 6320 6321 if (op_branch != NULL) { 6322 assert(op_branch->block() == NULL || code->index_of(op_branch->block()) != -1, "branch target not valid"); 6323 assert(op_branch->ublock() == NULL || code->index_of(op_branch->ublock()) != -1, "branch target not valid"); 6324 } 6325 } 6326 6327 for (j = 0; j < block->number_of_sux() - 1; j++) { 6328 BlockBegin* sux = block->sux_at(j); 6329 assert(code->index_of(sux) != -1, "successor not valid"); 6330 } 6331 6332 for (j = 0; j < block->number_of_preds() - 1; j++) { 6333 BlockBegin* pred = block->pred_at(j); 6334 assert(code->index_of(pred) != -1, "successor not valid"); 6335 } 6336 } 6337 } 6338 #endif 6339 6340 6341 #ifndef PRODUCT 6342 6343 // Implementation of LinearStatistic 6344 6345 const char* LinearScanStatistic::counter_name(int counter_idx) { 6346 switch (counter_idx) { 6347 case counter_method: return "compiled methods"; 6348 case counter_fpu_method: return "methods using fpu"; 6349 case counter_loop_method: return "methods with loops"; 6350 case counter_exception_method:return "methods with xhandler"; 6351 6352 case counter_loop: return "loops"; 6353 case counter_block: return "blocks"; 6354 case counter_loop_block: return "blocks inside loop"; 6355 case counter_exception_block: return "exception handler entries"; 6356 case counter_interval: return "intervals"; 6357 case counter_fixed_interval: return "fixed intervals"; 6358 case counter_range: return "ranges"; 6359 case counter_fixed_range: return "fixed ranges"; 6360 case counter_use_pos: return "use positions"; 6361 case counter_fixed_use_pos: return "fixed use positions"; 6362 case counter_spill_slots: return "spill slots"; 6363 6364 // counter for classes of lir instructions 6365 case counter_instruction: return "total instructions"; 6366 case counter_label: return "labels"; 6367 case counter_entry: return "method entries"; 6368 case counter_return: return "method returns"; 6369 case counter_call: return "method calls"; 6370 case counter_move: return "moves"; 6371 case counter_cmp: return "compare"; 6372 case counter_cond_branch: return "conditional branches"; 6373 case counter_uncond_branch: return "unconditional branches"; 6374 case counter_stub_branch: return "branches to stub"; 6375 case counter_alu: return "artithmetic + logic"; 6376 case counter_alloc: return "allocations"; 6377 case counter_sync: return "synchronisation"; 6378 case counter_throw: return "throw"; 6379 case counter_unwind: return "unwind"; 6380 case counter_typecheck: return "type+null-checks"; 6381 case counter_fpu_stack: return "fpu-stack"; 6382 case counter_misc_inst: return "other instructions"; 6383 case counter_other_inst: return "misc. instructions"; 6384 6385 // counter for different types of moves 6386 case counter_move_total: return "total moves"; 6387 case counter_move_reg_reg: return "register->register"; 6388 case counter_move_reg_stack: return "register->stack"; 6389 case counter_move_stack_reg: return "stack->register"; 6390 case counter_move_stack_stack:return "stack->stack"; 6391 case counter_move_reg_mem: return "register->memory"; 6392 case counter_move_mem_reg: return "memory->register"; 6393 case counter_move_const_any: return "constant->any"; 6394 6395 case blank_line_1: return ""; 6396 case blank_line_2: return ""; 6397 6398 default: ShouldNotReachHere(); return ""; 6399 } 6400 } 6401 6402 LinearScanStatistic::Counter LinearScanStatistic::base_counter(int counter_idx) { 6403 if (counter_idx == counter_fpu_method || counter_idx == counter_loop_method || counter_idx == counter_exception_method) { 6404 return counter_method; 6405 } else if (counter_idx == counter_loop_block || counter_idx == counter_exception_block) { 6406 return counter_block; 6407 } else if (counter_idx >= counter_instruction && counter_idx <= counter_other_inst) { 6408 return counter_instruction; 6409 } else if (counter_idx >= counter_move_total && counter_idx <= counter_move_const_any) { 6410 return counter_move_total; 6411 } 6412 return invalid_counter; 6413 } 6414 6415 LinearScanStatistic::LinearScanStatistic() { 6416 for (int i = 0; i < number_of_counters; i++) { 6417 _counters_sum[i] = 0; 6418 _counters_max[i] = -1; 6419 } 6420 6421 } 6422 6423 // add the method-local numbers to the total sum 6424 void LinearScanStatistic::sum_up(LinearScanStatistic &method_statistic) { 6425 for (int i = 0; i < number_of_counters; i++) { 6426 _counters_sum[i] += method_statistic._counters_sum[i]; 6427 _counters_max[i] = MAX2(_counters_max[i], method_statistic._counters_sum[i]); 6428 } 6429 } 6430 6431 void LinearScanStatistic::print(const char* title) { 6432 if (CountLinearScan || TraceLinearScanLevel > 0) { 6433 tty->cr(); 6434 tty->print_cr("***** LinearScan statistic - %s *****", title); 6435 6436 for (int i = 0; i < number_of_counters; i++) { 6437 if (_counters_sum[i] > 0 || _counters_max[i] >= 0) { 6438 tty->print("%25s: %8d", counter_name(i), _counters_sum[i]); 6439 6440 if (base_counter(i) != invalid_counter) { 6441 tty->print(" (%5.1f%%) ", _counters_sum[i] * 100.0 / _counters_sum[base_counter(i)]); 6442 } else { 6443 tty->print(" "); 6444 } 6445 6446 if (_counters_max[i] >= 0) { 6447 tty->print("%8d", _counters_max[i]); 6448 } 6449 } 6450 tty->cr(); 6451 } 6452 } 6453 } 6454 6455 void LinearScanStatistic::collect(LinearScan* allocator) { 6456 inc_counter(counter_method); 6457 if (allocator->has_fpu_registers()) { 6458 inc_counter(counter_fpu_method); 6459 } 6460 if (allocator->num_loops() > 0) { 6461 inc_counter(counter_loop_method); 6462 } 6463 inc_counter(counter_loop, allocator->num_loops()); 6464 inc_counter(counter_spill_slots, allocator->max_spills()); 6465 6466 int i; 6467 for (i = 0; i < allocator->interval_count(); i++) { 6468 Interval* cur = allocator->interval_at(i); 6469 6470 if (cur != NULL) { 6471 inc_counter(counter_interval); 6472 inc_counter(counter_use_pos, cur->num_use_positions()); 6473 if (LinearScan::is_precolored_interval(cur)) { 6474 inc_counter(counter_fixed_interval); 6475 inc_counter(counter_fixed_use_pos, cur->num_use_positions()); 6476 } 6477 6478 Range* range = cur->first(); 6479 while (range != Range::end()) { 6480 inc_counter(counter_range); 6481 if (LinearScan::is_precolored_interval(cur)) { 6482 inc_counter(counter_fixed_range); 6483 } 6484 range = range->next(); 6485 } 6486 } 6487 } 6488 6489 bool has_xhandlers = false; 6490 // Note: only count blocks that are in code-emit order 6491 for (i = 0; i < allocator->ir()->code()->length(); i++) { 6492 BlockBegin* cur = allocator->ir()->code()->at(i); 6493 6494 inc_counter(counter_block); 6495 if (cur->loop_depth() > 0) { 6496 inc_counter(counter_loop_block); 6497 } 6498 if (cur->is_set(BlockBegin::exception_entry_flag)) { 6499 inc_counter(counter_exception_block); 6500 has_xhandlers = true; 6501 } 6502 6503 LIR_OpList* instructions = cur->lir()->instructions_list(); 6504 for (int j = 0; j < instructions->length(); j++) { 6505 LIR_Op* op = instructions->at(j); 6506 6507 inc_counter(counter_instruction); 6508 6509 switch (op->code()) { 6510 case lir_label: inc_counter(counter_label); break; 6511 case lir_std_entry: 6512 case lir_osr_entry: inc_counter(counter_entry); break; 6513 case lir_return: inc_counter(counter_return); break; 6514 6515 case lir_rtcall: 6516 case lir_static_call: 6517 case lir_optvirtual_call: 6518 case lir_virtual_call: inc_counter(counter_call); break; 6519 6520 case lir_move: { 6521 inc_counter(counter_move); 6522 inc_counter(counter_move_total); 6523 6524 LIR_Opr in = op->as_Op1()->in_opr(); 6525 LIR_Opr res = op->as_Op1()->result_opr(); 6526 if (in->is_register()) { 6527 if (res->is_register()) { 6528 inc_counter(counter_move_reg_reg); 6529 } else if (res->is_stack()) { 6530 inc_counter(counter_move_reg_stack); 6531 } else if (res->is_address()) { 6532 inc_counter(counter_move_reg_mem); 6533 } else { 6534 ShouldNotReachHere(); 6535 } 6536 } else if (in->is_stack()) { 6537 if (res->is_register()) { 6538 inc_counter(counter_move_stack_reg); 6539 } else { 6540 inc_counter(counter_move_stack_stack); 6541 } 6542 } else if (in->is_address()) { 6543 assert(res->is_register(), "must be"); 6544 inc_counter(counter_move_mem_reg); 6545 } else if (in->is_constant()) { 6546 inc_counter(counter_move_const_any); 6547 } else { 6548 ShouldNotReachHere(); 6549 } 6550 break; 6551 } 6552 6553 case lir_cmp: inc_counter(counter_cmp); break; 6554 6555 case lir_branch: 6556 case lir_cond_float_branch: { 6557 LIR_OpBranch* branch = op->as_OpBranch(); 6558 if (branch->block() == NULL) { 6559 inc_counter(counter_stub_branch); 6560 } else if (branch->cond() == lir_cond_always) { 6561 inc_counter(counter_uncond_branch); 6562 } else { 6563 inc_counter(counter_cond_branch); 6564 } 6565 break; 6566 } 6567 6568 case lir_neg: 6569 case lir_add: 6570 case lir_sub: 6571 case lir_mul: 6572 case lir_mul_strictfp: 6573 case lir_div: 6574 case lir_div_strictfp: 6575 case lir_rem: 6576 case lir_sqrt: 6577 case lir_sin: 6578 case lir_cos: 6579 case lir_abs: 6580 case lir_log10: 6581 case lir_log: 6582 case lir_logic_and: 6583 case lir_logic_or: 6584 case lir_logic_xor: 6585 case lir_shl: 6586 case lir_shr: 6587 case lir_ushr: inc_counter(counter_alu); break; 6588 6589 case lir_alloc_object: 6590 case lir_alloc_array: inc_counter(counter_alloc); break; 6591 6592 case lir_monaddr: 6593 case lir_lock: 6594 case lir_unlock: inc_counter(counter_sync); break; 6595 6596 case lir_throw: inc_counter(counter_throw); break; 6597 6598 case lir_unwind: inc_counter(counter_unwind); break; 6599 6600 case lir_null_check: 6601 case lir_leal: 6602 case lir_instanceof: 6603 case lir_checkcast: 6604 case lir_store_check: inc_counter(counter_typecheck); break; 6605 6606 case lir_fpop_raw: 6607 case lir_fxch: 6608 case lir_fld: inc_counter(counter_fpu_stack); break; 6609 6610 case lir_nop: 6611 case lir_push: 6612 case lir_pop: 6613 case lir_convert: 6614 case lir_roundfp: 6615 case lir_cmove: inc_counter(counter_misc_inst); break; 6616 6617 default: inc_counter(counter_other_inst); break; 6618 } 6619 } 6620 } 6621 6622 if (has_xhandlers) { 6623 inc_counter(counter_exception_method); 6624 } 6625 } 6626 6627 void LinearScanStatistic::compute(LinearScan* allocator, LinearScanStatistic &global_statistic) { 6628 if (CountLinearScan || TraceLinearScanLevel > 0) { 6629 6630 LinearScanStatistic local_statistic = LinearScanStatistic(); 6631 6632 local_statistic.collect(allocator); 6633 global_statistic.sum_up(local_statistic); 6634 6635 if (TraceLinearScanLevel > 2) { 6636 local_statistic.print("current local statistic"); 6637 } 6638 } 6639 } 6640 6641 6642 // Implementation of LinearTimers 6643 6644 LinearScanTimers::LinearScanTimers() { 6645 for (int i = 0; i < number_of_timers; i++) { 6646 timer(i)->reset(); 6647 } 6648 } 6649 6650 const char* LinearScanTimers::timer_name(int idx) { 6651 switch (idx) { 6652 case timer_do_nothing: return "Nothing (Time Check)"; 6653 case timer_number_instructions: return "Number Instructions"; 6654 case timer_compute_local_live_sets: return "Local Live Sets"; 6655 case timer_compute_global_live_sets: return "Global Live Sets"; 6656 case timer_build_intervals: return "Build Intervals"; 6657 case timer_sort_intervals_before: return "Sort Intervals Before"; 6658 case timer_allocate_registers: return "Allocate Registers"; 6659 case timer_resolve_data_flow: return "Resolve Data Flow"; 6660 case timer_sort_intervals_after: return "Sort Intervals After"; 6661 case timer_eliminate_spill_moves: return "Spill optimization"; 6662 case timer_assign_reg_num: return "Assign Reg Num"; 6663 case timer_allocate_fpu_stack: return "Allocate FPU Stack"; 6664 case timer_optimize_lir: return "Optimize LIR"; 6665 default: ShouldNotReachHere(); return ""; 6666 } 6667 } 6668 6669 void LinearScanTimers::begin_method() { 6670 if (TimeEachLinearScan) { 6671 // reset all timers to measure only current method 6672 for (int i = 0; i < number_of_timers; i++) { 6673 timer(i)->reset(); 6674 } 6675 } 6676 } 6677 6678 void LinearScanTimers::end_method(LinearScan* allocator) { 6679 if (TimeEachLinearScan) { 6680 6681 double c = timer(timer_do_nothing)->seconds(); 6682 double total = 0; 6683 for (int i = 1; i < number_of_timers; i++) { 6684 total += timer(i)->seconds() - c; 6685 } 6686 6687 if (total >= 0.0005) { 6688 // print all information in one line for automatic processing 6689 tty->print("@"); allocator->compilation()->method()->print_name(); 6690 6691 tty->print("@ %d ", allocator->compilation()->method()->code_size()); 6692 tty->print("@ %d ", allocator->block_at(allocator->block_count() - 1)->last_lir_instruction_id() / 2); 6693 tty->print("@ %d ", allocator->block_count()); 6694 tty->print("@ %d ", allocator->num_virtual_regs()); 6695 tty->print("@ %d ", allocator->interval_count()); 6696 tty->print("@ %d ", allocator->_num_calls); 6697 tty->print("@ %d ", allocator->num_loops()); 6698 6699 tty->print("@ %6.6f ", total); 6700 for (int i = 1; i < number_of_timers; i++) { 6701 tty->print("@ %4.1f ", ((timer(i)->seconds() - c) / total) * 100); 6702 } 6703 tty->cr(); 6704 } 6705 } 6706 } 6707 6708 void LinearScanTimers::print(double total_time) { 6709 if (TimeLinearScan) { 6710 // correction value: sum of dummy-timer that only measures the time that 6711 // is necesary to start and stop itself 6712 double c = timer(timer_do_nothing)->seconds(); 6713 6714 for (int i = 0; i < number_of_timers; i++) { 6715 double t = timer(i)->seconds(); 6716 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); 6717 } 6718 } 6719 } 6720 6721 #endif // #ifndef PRODUCT