1 /* 2 * Copyright (c) 2005, 2015, 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 "code/vmreg.inline.hpp" 35 #include "utilities/bitMap.inline.hpp" 36 37 #ifndef PRODUCT 38 39 static LinearScanStatistic _stat_before_alloc; 40 static LinearScanStatistic _stat_after_asign; 41 static LinearScanStatistic _stat_final; 42 43 static LinearScanTimers _total_timer; 44 45 // helper macro for short definition of timer 46 #define TIME_LINEAR_SCAN(timer_name) TraceTime _block_timer("", _total_timer.timer(LinearScanTimers::timer_name), TimeLinearScan || TimeEachLinearScan, Verbose); 47 48 // helper macro for short definition of trace-output inside code 49 #define TRACE_LINEAR_SCAN(level, code) \ 50 if (TraceLinearScanLevel >= level) { \ 51 code; \ 52 } 53 54 #else 55 56 #define TIME_LINEAR_SCAN(timer_name) 57 #define TRACE_LINEAR_SCAN(level, code) 58 59 #endif 60 61 // Map BasicType to spill size in 32-bit words, matching VMReg's notion of words 62 #ifdef _LP64 63 static int type2spill_size[T_CONFLICT+1]={ -1, 0, 0, 0, 1, 1, 1, 2, 1, 1, 1, 2, 2, 2, 0, 2, 1, 2, 1, -1}; 64 #else 65 static int type2spill_size[T_CONFLICT+1]={ -1, 0, 0, 0, 1, 1, 1, 2, 1, 1, 1, 2, 1, 1, 0, 1, -1, 1, 1, -1}; 66 #endif 67 68 69 // Implementation of LinearScan 70 71 LinearScan::LinearScan(IR* ir, LIRGenerator* gen, FrameMap* frame_map) 72 : _compilation(ir->compilation()) 73 , _ir(ir) 74 , _gen(gen) 75 , _frame_map(frame_map) 76 , _num_virtual_regs(gen->max_virtual_register_number()) 77 , _has_fpu_registers(false) 78 , _num_calls(-1) 79 , _max_spills(0) 80 , _unused_spill_slot(-1) 81 , _intervals(0) // initialized later with correct length 82 , _new_intervals_from_allocation(new IntervalList()) 83 , _sorted_intervals(NULL) 84 , _needs_full_resort(false) 85 , _lir_ops(0) // initialized later with correct length 86 , _block_of_op(0) // initialized later with correct length 87 , _has_info(0) 88 , _has_call(0) 89 , _scope_value_cache(0) // initialized later with correct length 90 , _interval_in_loop(0, 0) // initialized later with correct length 91 , _cached_blocks(*ir->linear_scan_order()) 92 #ifdef X86 93 , _fpu_stack_allocator(NULL) 94 #endif 95 { 96 assert(this->ir() != NULL, "check if valid"); 97 assert(this->compilation() != NULL, "check if valid"); 98 assert(this->gen() != NULL, "check if valid"); 99 assert(this->frame_map() != NULL, "check if valid"); 100 } 101 102 103 // ********** functions for converting LIR-Operands to register numbers 104 // 105 // Emulate a flat register file comprising physical integer registers, 106 // physical floating-point registers and virtual registers, in that order. 107 // Virtual registers already have appropriate numbers, since V0 is 108 // the number of physical registers. 109 // Returns -1 for hi word if opr is a single word operand. 110 // 111 // Note: the inverse operation (calculating an operand for register numbers) 112 // is done in calc_operand_for_interval() 113 114 int LinearScan::reg_num(LIR_Opr opr) { 115 assert(opr->is_register(), "should not call this otherwise"); 116 117 if (opr->is_virtual_register()) { 118 assert(opr->vreg_number() >= nof_regs, "found a virtual register with a fixed-register number"); 119 return opr->vreg_number(); 120 } else if (opr->is_single_cpu()) { 121 return opr->cpu_regnr(); 122 } else if (opr->is_double_cpu()) { 123 return opr->cpu_regnrLo(); 124 #ifdef X86 125 } else if (opr->is_single_xmm()) { 126 return opr->fpu_regnr() + pd_first_xmm_reg; 127 } else if (opr->is_double_xmm()) { 128 return opr->fpu_regnrLo() + pd_first_xmm_reg; 129 #endif 130 } else if (opr->is_single_fpu()) { 131 return opr->fpu_regnr() + pd_first_fpu_reg; 132 } else if (opr->is_double_fpu()) { 133 return opr->fpu_regnrLo() + pd_first_fpu_reg; 134 } else { 135 ShouldNotReachHere(); 136 return -1; 137 } 138 } 139 140 int LinearScan::reg_numHi(LIR_Opr opr) { 141 assert(opr->is_register(), "should not call this otherwise"); 142 143 if (opr->is_virtual_register()) { 144 return -1; 145 } else if (opr->is_single_cpu()) { 146 return -1; 147 } else if (opr->is_double_cpu()) { 148 return opr->cpu_regnrHi(); 149 #ifdef X86 150 } else if (opr->is_single_xmm()) { 151 return -1; 152 } else if (opr->is_double_xmm()) { 153 return -1; 154 #endif 155 } else if (opr->is_single_fpu()) { 156 return -1; 157 } else if (opr->is_double_fpu()) { 158 return opr->fpu_regnrHi() + pd_first_fpu_reg; 159 } else { 160 ShouldNotReachHere(); 161 return -1; 162 } 163 } 164 165 166 // ********** functions for classification of intervals 167 168 bool LinearScan::is_precolored_interval(const Interval* i) { 169 return i->reg_num() < LinearScan::nof_regs; 170 } 171 172 bool LinearScan::is_virtual_interval(const Interval* i) { 173 return i->reg_num() >= LIR_OprDesc::vreg_base; 174 } 175 176 bool LinearScan::is_precolored_cpu_interval(const Interval* i) { 177 return i->reg_num() < LinearScan::nof_cpu_regs; 178 } 179 180 bool LinearScan::is_virtual_cpu_interval(const Interval* i) { 181 #if defined(__SOFTFP__) || defined(E500V2) 182 return i->reg_num() >= LIR_OprDesc::vreg_base; 183 #else 184 return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() != T_FLOAT && i->type() != T_DOUBLE); 185 #endif // __SOFTFP__ or E500V2 186 } 187 188 bool LinearScan::is_precolored_fpu_interval(const Interval* i) { 189 return i->reg_num() >= LinearScan::nof_cpu_regs && i->reg_num() < LinearScan::nof_regs; 190 } 191 192 bool LinearScan::is_virtual_fpu_interval(const Interval* i) { 193 #if defined(__SOFTFP__) || defined(E500V2) 194 return false; 195 #else 196 return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() == T_FLOAT || i->type() == T_DOUBLE); 197 #endif // __SOFTFP__ or E500V2 198 } 199 200 bool LinearScan::is_in_fpu_register(const Interval* i) { 201 // fixed intervals not needed for FPU stack allocation 202 return i->reg_num() >= nof_regs && pd_first_fpu_reg <= i->assigned_reg() && i->assigned_reg() <= pd_last_fpu_reg; 203 } 204 205 bool LinearScan::is_oop_interval(const Interval* i) { 206 // fixed intervals never contain oops 207 return i->reg_num() >= nof_regs && i->type() == T_OBJECT; 208 } 209 210 211 // ********** General helper functions 212 213 // compute next unused stack index that can be used for spilling 214 int LinearScan::allocate_spill_slot(bool double_word) { 215 int spill_slot; 216 if (double_word) { 217 if ((_max_spills & 1) == 1) { 218 // alignment of double-word values 219 // the hole because of the alignment is filled with the next single-word value 220 assert(_unused_spill_slot == -1, "wasting a spill slot"); 221 _unused_spill_slot = _max_spills; 222 _max_spills++; 223 } 224 spill_slot = _max_spills; 225 _max_spills += 2; 226 227 } else if (_unused_spill_slot != -1) { 228 // re-use hole that was the result of a previous double-word alignment 229 spill_slot = _unused_spill_slot; 230 _unused_spill_slot = -1; 231 232 } else { 233 spill_slot = _max_spills; 234 _max_spills++; 235 } 236 237 int result = spill_slot + LinearScan::nof_regs + frame_map()->argcount(); 238 239 // the class OopMapValue uses only 11 bits for storing the name of the 240 // oop location. So a stack slot bigger than 2^11 leads to an overflow 241 // that is not reported in product builds. Prevent this by checking the 242 // spill slot here (altough this value and the later used location name 243 // are slightly different) 244 if (result > 2000) { 245 bailout("too many stack slots used"); 246 } 247 248 return result; 249 } 250 251 void LinearScan::assign_spill_slot(Interval* it) { 252 // assign the canonical spill slot of the parent (if a part of the interval 253 // is already spilled) or allocate a new spill slot 254 if (it->canonical_spill_slot() >= 0) { 255 it->assign_reg(it->canonical_spill_slot()); 256 } else { 257 int spill = allocate_spill_slot(type2spill_size[it->type()] == 2); 258 it->set_canonical_spill_slot(spill); 259 it->assign_reg(spill); 260 } 261 } 262 263 void LinearScan::propagate_spill_slots() { 264 if (!frame_map()->finalize_frame(max_spills())) { 265 bailout("frame too large"); 266 } 267 } 268 269 // create a new interval with a predefined reg_num 270 // (only used for parent intervals that are created during the building phase) 271 Interval* LinearScan::create_interval(int reg_num) { 272 assert(_intervals.at(reg_num) == NULL, "overwriting exisiting interval"); 273 274 Interval* interval = new Interval(reg_num); 275 _intervals.at_put(reg_num, interval); 276 277 // assign register number for precolored intervals 278 if (reg_num < LIR_OprDesc::vreg_base) { 279 interval->assign_reg(reg_num); 280 } 281 return interval; 282 } 283 284 // assign a new reg_num to the interval and append it to the list of intervals 285 // (only used for child intervals that are created during register allocation) 286 void LinearScan::append_interval(Interval* it) { 287 it->set_reg_num(_intervals.length()); 288 _intervals.append(it); 289 _new_intervals_from_allocation->append(it); 290 } 291 292 // copy the vreg-flags if an interval is split 293 void LinearScan::copy_register_flags(Interval* from, Interval* to) { 294 if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::byte_reg)) { 295 gen()->set_vreg_flag(to->reg_num(), LIRGenerator::byte_reg); 296 } 297 if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::callee_saved)) { 298 gen()->set_vreg_flag(to->reg_num(), LIRGenerator::callee_saved); 299 } 300 301 // Note: do not copy the must_start_in_memory flag because it is not necessary for child 302 // intervals (only the very beginning of the interval must be in memory) 303 } 304 305 306 // ********** spill move optimization 307 // eliminate moves from register to stack if stack slot is known to be correct 308 309 // called during building of intervals 310 void LinearScan::change_spill_definition_pos(Interval* interval, int def_pos) { 311 assert(interval->is_split_parent(), "can only be called for split parents"); 312 313 switch (interval->spill_state()) { 314 case noDefinitionFound: 315 assert(interval->spill_definition_pos() == -1, "must no be set before"); 316 interval->set_spill_definition_pos(def_pos); 317 interval->set_spill_state(oneDefinitionFound); 318 break; 319 320 case oneDefinitionFound: 321 assert(def_pos <= interval->spill_definition_pos(), "positions are processed in reverse order when intervals are created"); 322 if (def_pos < interval->spill_definition_pos() - 2) { 323 // second definition found, so no spill optimization possible for this interval 324 interval->set_spill_state(noOptimization); 325 } else { 326 // two consecutive definitions (because of two-operand LIR form) 327 assert(block_of_op_with_id(def_pos) == block_of_op_with_id(interval->spill_definition_pos()), "block must be equal"); 328 } 329 break; 330 331 case noOptimization: 332 // nothing to do 333 break; 334 335 default: 336 assert(false, "other states not allowed at this time"); 337 } 338 } 339 340 // called during register allocation 341 void LinearScan::change_spill_state(Interval* interval, int spill_pos) { 342 switch (interval->spill_state()) { 343 case oneDefinitionFound: { 344 int def_loop_depth = block_of_op_with_id(interval->spill_definition_pos())->loop_depth(); 345 int spill_loop_depth = block_of_op_with_id(spill_pos)->loop_depth(); 346 347 if (def_loop_depth < spill_loop_depth) { 348 // the loop depth of the spilling position is higher then the loop depth 349 // at the definition of the interval -> move write to memory out of loop 350 // by storing at definitin of the interval 351 interval->set_spill_state(storeAtDefinition); 352 } else { 353 // the interval is currently spilled only once, so for now there is no 354 // reason to store the interval at the definition 355 interval->set_spill_state(oneMoveInserted); 356 } 357 break; 358 } 359 360 case oneMoveInserted: { 361 // the interval is spilled more then once, so it is better to store it to 362 // memory at the definition 363 interval->set_spill_state(storeAtDefinition); 364 break; 365 } 366 367 case storeAtDefinition: 368 case startInMemory: 369 case noOptimization: 370 case noDefinitionFound: 371 // nothing to do 372 break; 373 374 default: 375 assert(false, "other states not allowed at this time"); 376 } 377 } 378 379 380 bool LinearScan::must_store_at_definition(const Interval* i) { 381 return i->is_split_parent() && i->spill_state() == storeAtDefinition; 382 } 383 384 // called once before asignment of register numbers 385 void LinearScan::eliminate_spill_moves() { 386 TIME_LINEAR_SCAN(timer_eliminate_spill_moves); 387 TRACE_LINEAR_SCAN(3, tty->print_cr("***** Eliminating unnecessary spill moves")); 388 389 // collect all intervals that must be stored after their definion. 390 // the list is sorted by Interval::spill_definition_pos 391 Interval* interval; 392 Interval* temp_list; 393 create_unhandled_lists(&interval, &temp_list, must_store_at_definition, NULL); 394 395 #ifdef ASSERT 396 Interval* prev = NULL; 397 Interval* temp = interval; 398 while (temp != Interval::end()) { 399 assert(temp->spill_definition_pos() > 0, "invalid spill definition pos"); 400 if (prev != NULL) { 401 assert(temp->from() >= prev->from(), "intervals not sorted"); 402 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"); 403 } 404 405 assert(temp->canonical_spill_slot() >= LinearScan::nof_regs, "interval has no spill slot assigned"); 406 assert(temp->spill_definition_pos() >= temp->from(), "invalid order"); 407 assert(temp->spill_definition_pos() <= temp->from() + 2, "only intervals defined once at their start-pos can be optimized"); 408 409 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())); 410 411 temp = temp->next(); 412 } 413 #endif 414 415 LIR_InsertionBuffer insertion_buffer; 416 int num_blocks = block_count(); 417 for (int i = 0; i < num_blocks; i++) { 418 BlockBegin* block = block_at(i); 419 LIR_OpList* instructions = block->lir()->instructions_list(); 420 int num_inst = instructions->length(); 421 bool has_new = false; 422 423 // iterate all instructions of the block. skip the first because it is always a label 424 for (int j = 1; j < num_inst; j++) { 425 LIR_Op* op = instructions->at(j); 426 int op_id = op->id(); 427 428 if (op_id == -1) { 429 // remove move from register to stack if the stack slot is guaranteed to be correct. 430 // only moves that have been inserted by LinearScan can be removed. 431 assert(op->code() == lir_move, "only moves can have a op_id of -1"); 432 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); 433 assert(op->as_Op1()->result_opr()->is_virtual(), "LinearScan inserts only moves to virtual registers"); 434 435 LIR_Op1* op1 = (LIR_Op1*)op; 436 Interval* interval = interval_at(op1->result_opr()->vreg_number()); 437 438 if (interval->assigned_reg() >= LinearScan::nof_regs && interval->always_in_memory()) { 439 // move target is a stack slot that is always correct, so eliminate instruction 440 TRACE_LINEAR_SCAN(4, tty->print_cr("eliminating move from interval %d to %d", op1->in_opr()->vreg_number(), op1->result_opr()->vreg_number())); 441 instructions->at_put(j, NULL); // NULL-instructions are deleted by assign_reg_num 442 } 443 444 } else { 445 // insert move from register to stack just after the beginning of the interval 446 assert(interval == Interval::end() || interval->spill_definition_pos() >= op_id, "invalid order"); 447 assert(interval == Interval::end() || (interval->is_split_parent() && interval->spill_state() == storeAtDefinition), "invalid interval"); 448 449 while (interval != Interval::end() && interval->spill_definition_pos() == op_id) { 450 if (!has_new) { 451 // prepare insertion buffer (appended when all instructions of the block are processed) 452 insertion_buffer.init(block->lir()); 453 has_new = true; 454 } 455 456 LIR_Opr from_opr = operand_for_interval(interval); 457 LIR_Opr to_opr = canonical_spill_opr(interval); 458 assert(from_opr->is_fixed_cpu() || from_opr->is_fixed_fpu(), "from operand must be a register"); 459 assert(to_opr->is_stack(), "to operand must be a stack slot"); 460 461 insertion_buffer.move(j, from_opr, to_opr); 462 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)); 463 464 interval = interval->next(); 465 } 466 } 467 } // end of instruction iteration 468 469 if (has_new) { 470 block->lir()->append(&insertion_buffer); 471 } 472 } // end of block iteration 473 474 assert(interval == Interval::end(), "missed an interval"); 475 } 476 477 478 // ********** Phase 1: number all instructions in all blocks 479 // Compute depth-first and linear scan block orders, and number LIR_Op nodes for linear scan. 480 481 void LinearScan::number_instructions() { 482 { 483 // dummy-timer to measure the cost of the timer itself 484 // (this time is then subtracted from all other timers to get the real value) 485 TIME_LINEAR_SCAN(timer_do_nothing); 486 } 487 TIME_LINEAR_SCAN(timer_number_instructions); 488 489 // Assign IDs to LIR nodes and build a mapping, lir_ops, from ID to LIR_Op node. 490 int num_blocks = block_count(); 491 int num_instructions = 0; 492 int i; 493 for (i = 0; i < num_blocks; i++) { 494 num_instructions += block_at(i)->lir()->instructions_list()->length(); 495 } 496 497 // initialize with correct length 498 _lir_ops = LIR_OpArray(num_instructions); 499 _block_of_op = BlockBeginArray(num_instructions); 500 501 int op_id = 0; 502 int idx = 0; 503 504 for (i = 0; i < num_blocks; i++) { 505 BlockBegin* block = block_at(i); 506 block->set_first_lir_instruction_id(op_id); 507 LIR_OpList* instructions = block->lir()->instructions_list(); 508 509 int num_inst = instructions->length(); 510 for (int j = 0; j < num_inst; j++) { 511 LIR_Op* op = instructions->at(j); 512 op->set_id(op_id); 513 514 _lir_ops.at_put(idx, op); 515 _block_of_op.at_put(idx, block); 516 assert(lir_op_with_id(op_id) == op, "must match"); 517 518 idx++; 519 op_id += 2; // numbering of lir_ops by two 520 } 521 block->set_last_lir_instruction_id(op_id - 2); 522 } 523 assert(idx == num_instructions, "must match"); 524 assert(idx * 2 == op_id, "must match"); 525 526 _has_call = BitMap(num_instructions); _has_call.clear(); 527 _has_info = BitMap(num_instructions); _has_info.clear(); 528 } 529 530 531 // ********** Phase 2: compute local live sets separately for each block 532 // (sets live_gen and live_kill for each block) 533 534 void LinearScan::set_live_gen_kill(Value value, LIR_Op* op, BitMap& live_gen, BitMap& live_kill) { 535 LIR_Opr opr = value->operand(); 536 Constant* con = value->as_Constant(); 537 538 // check some asumptions about debug information 539 assert(!value->type()->is_illegal(), "if this local is used by the interpreter it shouldn't be of indeterminate type"); 540 assert(con == NULL || opr->is_virtual() || opr->is_constant() || opr->is_illegal(), "asumption: Constant instructions have only constant operands"); 541 assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands"); 542 543 if ((con == NULL || con->is_pinned()) && opr->is_register()) { 544 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 545 int reg = opr->vreg_number(); 546 if (!live_kill.at(reg)) { 547 live_gen.set_bit(reg); 548 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)); 549 } 550 } 551 } 552 553 554 void LinearScan::compute_local_live_sets() { 555 TIME_LINEAR_SCAN(timer_compute_local_live_sets); 556 557 int num_blocks = block_count(); 558 int live_size = live_set_size(); 559 bool local_has_fpu_registers = false; 560 int local_num_calls = 0; 561 LIR_OpVisitState visitor; 562 563 BitMap2D local_interval_in_loop = BitMap2D(_num_virtual_regs, num_loops()); 564 local_interval_in_loop.clear(); 565 566 // iterate all blocks 567 for (int i = 0; i < num_blocks; i++) { 568 BlockBegin* block = block_at(i); 569 570 BitMap live_gen(live_size); live_gen.clear(); 571 BitMap live_kill(live_size); live_kill.clear(); 572 573 if (block->is_set(BlockBegin::exception_entry_flag)) { 574 // Phi functions at the begin of an exception handler are 575 // implicitly defined (= killed) at the beginning of the block. 576 for_each_phi_fun(block, phi, 577 live_kill.set_bit(phi->operand()->vreg_number()) 578 ); 579 } 580 581 LIR_OpList* instructions = block->lir()->instructions_list(); 582 int num_inst = instructions->length(); 583 584 // iterate all instructions of the block. skip the first because it is always a label 585 assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label"); 586 for (int j = 1; j < num_inst; j++) { 587 LIR_Op* op = instructions->at(j); 588 589 // visit operation to collect all operands 590 visitor.visit(op); 591 592 if (visitor.has_call()) { 593 _has_call.set_bit(op->id() >> 1); 594 local_num_calls++; 595 } 596 if (visitor.info_count() > 0) { 597 _has_info.set_bit(op->id() >> 1); 598 } 599 600 // iterate input operands of instruction 601 int k, n, reg; 602 n = visitor.opr_count(LIR_OpVisitState::inputMode); 603 for (k = 0; k < n; k++) { 604 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k); 605 assert(opr->is_register(), "visitor should only return register operands"); 606 607 if (opr->is_virtual_register()) { 608 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 609 reg = opr->vreg_number(); 610 if (!live_kill.at(reg)) { 611 live_gen.set_bit(reg); 612 TRACE_LINEAR_SCAN(4, tty->print_cr(" Setting live_gen for register %d at instruction %d", reg, op->id())); 613 } 614 if (block->loop_index() >= 0) { 615 local_interval_in_loop.set_bit(reg, block->loop_index()); 616 } 617 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); 618 } 619 620 #ifdef ASSERT 621 // fixed intervals are never live at block boundaries, so 622 // they need not be processed in live sets. 623 // this is checked by these assertions to be sure about it. 624 // the entry block may have incoming values in registers, which is ok. 625 if (!opr->is_virtual_register() && block != ir()->start()) { 626 reg = reg_num(opr); 627 if (is_processed_reg_num(reg)) { 628 assert(live_kill.at(reg), "using fixed register that is not defined in this block"); 629 } 630 reg = reg_numHi(opr); 631 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 632 assert(live_kill.at(reg), "using fixed register that is not defined in this block"); 633 } 634 } 635 #endif 636 } 637 638 // Add uses of live locals from interpreter's point of view for proper debug information generation 639 n = visitor.info_count(); 640 for (k = 0; k < n; k++) { 641 CodeEmitInfo* info = visitor.info_at(k); 642 ValueStack* stack = info->stack(); 643 for_each_state_value(stack, value, 644 set_live_gen_kill(value, op, live_gen, live_kill) 645 ); 646 } 647 648 // iterate temp operands of instruction 649 n = visitor.opr_count(LIR_OpVisitState::tempMode); 650 for (k = 0; k < n; k++) { 651 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k); 652 assert(opr->is_register(), "visitor should only return register operands"); 653 654 if (opr->is_virtual_register()) { 655 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 656 reg = opr->vreg_number(); 657 live_kill.set_bit(reg); 658 if (block->loop_index() >= 0) { 659 local_interval_in_loop.set_bit(reg, block->loop_index()); 660 } 661 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); 662 } 663 664 #ifdef ASSERT 665 // fixed intervals are never live at block boundaries, so 666 // they need not be processed in live sets 667 // process them only in debug mode so that this can be checked 668 if (!opr->is_virtual_register()) { 669 reg = reg_num(opr); 670 if (is_processed_reg_num(reg)) { 671 live_kill.set_bit(reg_num(opr)); 672 } 673 reg = reg_numHi(opr); 674 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 675 live_kill.set_bit(reg); 676 } 677 } 678 #endif 679 } 680 681 // iterate output operands of instruction 682 n = visitor.opr_count(LIR_OpVisitState::outputMode); 683 for (k = 0; k < n; k++) { 684 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k); 685 assert(opr->is_register(), "visitor should only return register operands"); 686 687 if (opr->is_virtual_register()) { 688 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 689 reg = opr->vreg_number(); 690 live_kill.set_bit(reg); 691 if (block->loop_index() >= 0) { 692 local_interval_in_loop.set_bit(reg, block->loop_index()); 693 } 694 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); 695 } 696 697 #ifdef ASSERT 698 // fixed intervals are never live at block boundaries, so 699 // they need not be processed in live sets 700 // process them only in debug mode so that this can be checked 701 if (!opr->is_virtual_register()) { 702 reg = reg_num(opr); 703 if (is_processed_reg_num(reg)) { 704 live_kill.set_bit(reg_num(opr)); 705 } 706 reg = reg_numHi(opr); 707 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 708 live_kill.set_bit(reg); 709 } 710 } 711 #endif 712 } 713 } // end of instruction iteration 714 715 block->set_live_gen (live_gen); 716 block->set_live_kill(live_kill); 717 block->set_live_in (BitMap(live_size)); block->live_in().clear(); 718 block->set_live_out (BitMap(live_size)); block->live_out().clear(); 719 720 TRACE_LINEAR_SCAN(4, tty->print("live_gen B%d ", block->block_id()); print_bitmap(block->live_gen())); 721 TRACE_LINEAR_SCAN(4, tty->print("live_kill B%d ", block->block_id()); print_bitmap(block->live_kill())); 722 } // end of block iteration 723 724 // propagate local calculated information into LinearScan object 725 _has_fpu_registers = local_has_fpu_registers; 726 compilation()->set_has_fpu_code(local_has_fpu_registers); 727 728 _num_calls = local_num_calls; 729 _interval_in_loop = local_interval_in_loop; 730 } 731 732 733 // ********** Phase 3: perform a backward dataflow analysis to compute global live sets 734 // (sets live_in and live_out for each block) 735 736 void LinearScan::compute_global_live_sets() { 737 TIME_LINEAR_SCAN(timer_compute_global_live_sets); 738 739 int num_blocks = block_count(); 740 bool change_occurred; 741 bool change_occurred_in_block; 742 int iteration_count = 0; 743 BitMap live_out(live_set_size()); live_out.clear(); // scratch set for calculations 744 745 // Perform a backward dataflow analysis to compute live_out and live_in for each block. 746 // The loop is executed until a fixpoint is reached (no changes in an iteration) 747 // Exception handlers must be processed because not all live values are 748 // present in the state array, e.g. because of global value numbering 749 do { 750 change_occurred = false; 751 752 // iterate all blocks in reverse order 753 for (int i = num_blocks - 1; i >= 0; i--) { 754 BlockBegin* block = block_at(i); 755 756 change_occurred_in_block = false; 757 758 // live_out(block) is the union of live_in(sux), for successors sux of block 759 int n = block->number_of_sux(); 760 int e = block->number_of_exception_handlers(); 761 if (n + e > 0) { 762 // block has successors 763 if (n > 0) { 764 live_out.set_from(block->sux_at(0)->live_in()); 765 for (int j = 1; j < n; j++) { 766 live_out.set_union(block->sux_at(j)->live_in()); 767 } 768 } else { 769 live_out.clear(); 770 } 771 for (int j = 0; j < e; j++) { 772 live_out.set_union(block->exception_handler_at(j)->live_in()); 773 } 774 775 if (!block->live_out().is_same(live_out)) { 776 // A change occurred. Swap the old and new live out sets to avoid copying. 777 BitMap temp = block->live_out(); 778 block->set_live_out(live_out); 779 live_out = temp; 780 781 change_occurred = true; 782 change_occurred_in_block = true; 783 } 784 } 785 786 if (iteration_count == 0 || change_occurred_in_block) { 787 // live_in(block) is the union of live_gen(block) with (live_out(block) & !live_kill(block)) 788 // note: live_in has to be computed only in first iteration or if live_out has changed! 789 BitMap live_in = block->live_in(); 790 live_in.set_from(block->live_out()); 791 live_in.set_difference(block->live_kill()); 792 live_in.set_union(block->live_gen()); 793 } 794 795 #ifndef PRODUCT 796 if (TraceLinearScanLevel >= 4) { 797 char c = ' '; 798 if (iteration_count == 0 || change_occurred_in_block) { 799 c = '*'; 800 } 801 tty->print("(%d) live_in%c B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_in()); 802 tty->print("(%d) live_out%c B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_out()); 803 } 804 #endif 805 } 806 iteration_count++; 807 808 if (change_occurred && iteration_count > 50) { 809 BAILOUT("too many iterations in compute_global_live_sets"); 810 } 811 } while (change_occurred); 812 813 814 #ifdef ASSERT 815 // check that fixed intervals are not live at block boundaries 816 // (live set must be empty at fixed intervals) 817 for (int i = 0; i < num_blocks; i++) { 818 BlockBegin* block = block_at(i); 819 for (int j = 0; j < LIR_OprDesc::vreg_base; j++) { 820 assert(block->live_in().at(j) == false, "live_in set of fixed register must be empty"); 821 assert(block->live_out().at(j) == false, "live_out set of fixed register must be empty"); 822 assert(block->live_gen().at(j) == false, "live_gen set of fixed register must be empty"); 823 } 824 } 825 #endif 826 827 // check that the live_in set of the first block is empty 828 BitMap live_in_args(ir()->start()->live_in().size()); 829 live_in_args.clear(); 830 if (!ir()->start()->live_in().is_same(live_in_args)) { 831 #ifdef ASSERT 832 tty->print_cr("Error: live_in set of first block must be empty (when this fails, virtual registers are used before they are defined)"); 833 tty->print_cr("affected registers:"); 834 print_bitmap(ir()->start()->live_in()); 835 836 // print some additional information to simplify debugging 837 for (unsigned int i = 0; i < ir()->start()->live_in().size(); i++) { 838 if (ir()->start()->live_in().at(i)) { 839 Instruction* instr = gen()->instruction_for_vreg(i); 840 tty->print_cr("* vreg %d (HIR instruction %c%d)", i, instr == NULL ? ' ' : instr->type()->tchar(), instr == NULL ? 0 : instr->id()); 841 842 for (int j = 0; j < num_blocks; j++) { 843 BlockBegin* block = block_at(j); 844 if (block->live_gen().at(i)) { 845 tty->print_cr(" used in block B%d", block->block_id()); 846 } 847 if (block->live_kill().at(i)) { 848 tty->print_cr(" defined in block B%d", block->block_id()); 849 } 850 } 851 } 852 } 853 854 #endif 855 // when this fails, virtual registers are used before they are defined. 856 assert(false, "live_in set of first block must be empty"); 857 // bailout of if this occurs in product mode. 858 bailout("live_in set of first block not empty"); 859 } 860 } 861 862 863 // ********** Phase 4: build intervals 864 // (fills the list _intervals) 865 866 void LinearScan::add_use(Value value, int from, int to, IntervalUseKind use_kind) { 867 assert(!value->type()->is_illegal(), "if this value is used by the interpreter it shouldn't be of indeterminate type"); 868 LIR_Opr opr = value->operand(); 869 Constant* con = value->as_Constant(); 870 871 if ((con == NULL || con->is_pinned()) && opr->is_register()) { 872 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 873 add_use(opr, from, to, use_kind); 874 } 875 } 876 877 878 void LinearScan::add_def(LIR_Opr opr, int def_pos, IntervalUseKind use_kind) { 879 TRACE_LINEAR_SCAN(2, tty->print(" def "); opr->print(tty); tty->print_cr(" def_pos %d (%d)", def_pos, use_kind)); 880 assert(opr->is_register(), "should not be called otherwise"); 881 882 if (opr->is_virtual_register()) { 883 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 884 add_def(opr->vreg_number(), def_pos, use_kind, opr->type_register()); 885 886 } else { 887 int reg = reg_num(opr); 888 if (is_processed_reg_num(reg)) { 889 add_def(reg, def_pos, use_kind, opr->type_register()); 890 } 891 reg = reg_numHi(opr); 892 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 893 add_def(reg, def_pos, use_kind, opr->type_register()); 894 } 895 } 896 } 897 898 void LinearScan::add_use(LIR_Opr opr, int from, int to, IntervalUseKind use_kind) { 899 TRACE_LINEAR_SCAN(2, tty->print(" use "); opr->print(tty); tty->print_cr(" from %d to %d (%d)", from, to, use_kind)); 900 assert(opr->is_register(), "should not be called otherwise"); 901 902 if (opr->is_virtual_register()) { 903 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 904 add_use(opr->vreg_number(), from, to, use_kind, opr->type_register()); 905 906 } else { 907 int reg = reg_num(opr); 908 if (is_processed_reg_num(reg)) { 909 add_use(reg, from, to, use_kind, opr->type_register()); 910 } 911 reg = reg_numHi(opr); 912 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 913 add_use(reg, from, to, use_kind, opr->type_register()); 914 } 915 } 916 } 917 918 void LinearScan::add_temp(LIR_Opr opr, int temp_pos, IntervalUseKind use_kind) { 919 TRACE_LINEAR_SCAN(2, tty->print(" temp "); opr->print(tty); tty->print_cr(" temp_pos %d (%d)", temp_pos, use_kind)); 920 assert(opr->is_register(), "should not be called otherwise"); 921 922 if (opr->is_virtual_register()) { 923 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 924 add_temp(opr->vreg_number(), temp_pos, use_kind, opr->type_register()); 925 926 } else { 927 int reg = reg_num(opr); 928 if (is_processed_reg_num(reg)) { 929 add_temp(reg, temp_pos, use_kind, opr->type_register()); 930 } 931 reg = reg_numHi(opr); 932 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 933 add_temp(reg, temp_pos, use_kind, opr->type_register()); 934 } 935 } 936 } 937 938 939 void LinearScan::add_def(int reg_num, int def_pos, IntervalUseKind use_kind, BasicType type) { 940 Interval* interval = interval_at(reg_num); 941 if (interval != NULL) { 942 assert(interval->reg_num() == reg_num, "wrong interval"); 943 944 if (type != T_ILLEGAL) { 945 interval->set_type(type); 946 } 947 948 Range* r = interval->first(); 949 if (r->from() <= def_pos) { 950 // Update the starting point (when a range is first created for a use, its 951 // start is the beginning of the current block until a def is encountered.) 952 r->set_from(def_pos); 953 interval->add_use_pos(def_pos, use_kind); 954 955 } else { 956 // Dead value - make vacuous interval 957 // also add use_kind for dead intervals 958 interval->add_range(def_pos, def_pos + 1); 959 interval->add_use_pos(def_pos, use_kind); 960 TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: def of reg %d at %d occurs without use", reg_num, def_pos)); 961 } 962 963 } else { 964 // Dead value - make vacuous interval 965 // also add use_kind for dead intervals 966 interval = create_interval(reg_num); 967 if (type != T_ILLEGAL) { 968 interval->set_type(type); 969 } 970 971 interval->add_range(def_pos, def_pos + 1); 972 interval->add_use_pos(def_pos, use_kind); 973 TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: dead value %d at %d in live intervals", reg_num, def_pos)); 974 } 975 976 change_spill_definition_pos(interval, def_pos); 977 if (use_kind == noUse && interval->spill_state() <= startInMemory) { 978 // detection of method-parameters and roundfp-results 979 // TODO: move this directly to position where use-kind is computed 980 interval->set_spill_state(startInMemory); 981 } 982 } 983 984 void LinearScan::add_use(int reg_num, int from, int to, IntervalUseKind use_kind, BasicType type) { 985 Interval* interval = interval_at(reg_num); 986 if (interval == NULL) { 987 interval = create_interval(reg_num); 988 } 989 assert(interval->reg_num() == reg_num, "wrong interval"); 990 991 if (type != T_ILLEGAL) { 992 interval->set_type(type); 993 } 994 995 interval->add_range(from, to); 996 interval->add_use_pos(to, use_kind); 997 } 998 999 void LinearScan::add_temp(int reg_num, int temp_pos, 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(temp_pos, temp_pos + 1); 1011 interval->add_use_pos(temp_pos, use_kind); 1012 } 1013 1014 1015 // the results of this functions are used for optimizing spilling and reloading 1016 // if the functions return shouldHaveRegister and the interval is spilled, 1017 // it is not reloaded to a register. 1018 IntervalUseKind LinearScan::use_kind_of_output_operand(LIR_Op* op, LIR_Opr opr) { 1019 if (op->code() == lir_move) { 1020 assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1"); 1021 LIR_Op1* move = (LIR_Op1*)op; 1022 LIR_Opr res = move->result_opr(); 1023 bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory); 1024 1025 if (result_in_memory) { 1026 // Begin of an interval with must_start_in_memory set. 1027 // This interval will always get a stack slot first, so return noUse. 1028 return noUse; 1029 1030 } else if (move->in_opr()->is_stack()) { 1031 // method argument (condition must be equal to handle_method_arguments) 1032 return noUse; 1033 1034 } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) { 1035 // Move from register to register 1036 if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) { 1037 // special handling of phi-function moves inside osr-entry blocks 1038 // input operand must have a register instead of output operand (leads to better register allocation) 1039 return shouldHaveRegister; 1040 } 1041 } 1042 } 1043 1044 if (opr->is_virtual() && 1045 gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::must_start_in_memory)) { 1046 // result is a stack-slot, so prevent immediate reloading 1047 return noUse; 1048 } 1049 1050 // all other operands require a register 1051 return mustHaveRegister; 1052 } 1053 1054 IntervalUseKind LinearScan::use_kind_of_input_operand(LIR_Op* op, LIR_Opr opr) { 1055 if (op->code() == lir_move) { 1056 assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1"); 1057 LIR_Op1* move = (LIR_Op1*)op; 1058 LIR_Opr res = move->result_opr(); 1059 bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory); 1060 1061 if (result_in_memory) { 1062 // Move to an interval with must_start_in_memory set. 1063 // To avoid moves from stack to stack (not allowed) force the input operand to a register 1064 return mustHaveRegister; 1065 1066 } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) { 1067 // Move from register to register 1068 if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) { 1069 // special handling of phi-function moves inside osr-entry blocks 1070 // input operand must have a register instead of output operand (leads to better register allocation) 1071 return mustHaveRegister; 1072 } 1073 1074 // The input operand is not forced to a register (moves from stack to register are allowed), 1075 // but it is faster if the input operand is in a register 1076 return shouldHaveRegister; 1077 } 1078 } 1079 1080 1081 #ifdef X86 1082 if (op->code() == lir_cmove) { 1083 // conditional moves can handle stack operands 1084 assert(op->result_opr()->is_register(), "result must always be in a register"); 1085 return shouldHaveRegister; 1086 } 1087 1088 // optimizations for second input operand of arithmehtic operations on Intel 1089 // this operand is allowed to be on the stack in some cases 1090 BasicType opr_type = opr->type_register(); 1091 if (opr_type == T_FLOAT || opr_type == T_DOUBLE) { 1092 if ((UseSSE == 1 && opr_type == T_FLOAT) || UseSSE >= 2) { 1093 // SSE float instruction (T_DOUBLE only supported with SSE2) 1094 switch (op->code()) { 1095 case lir_cmp: 1096 case lir_add: 1097 case lir_sub: 1098 case lir_mul: 1099 case lir_div: 1100 { 1101 assert(op->as_Op2() != NULL, "must be LIR_Op2"); 1102 LIR_Op2* op2 = (LIR_Op2*)op; 1103 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { 1104 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"); 1105 return shouldHaveRegister; 1106 } 1107 } 1108 } 1109 } else { 1110 // FPU stack float instruction 1111 switch (op->code()) { 1112 case lir_add: 1113 case lir_sub: 1114 case lir_mul: 1115 case lir_div: 1116 { 1117 assert(op->as_Op2() != NULL, "must be LIR_Op2"); 1118 LIR_Op2* op2 = (LIR_Op2*)op; 1119 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { 1120 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"); 1121 return shouldHaveRegister; 1122 } 1123 } 1124 } 1125 } 1126 // We want to sometimes use logical operations on pointers, in particular in GC barriers. 1127 // Since 64bit logical operations do not current support operands on stack, we have to make sure 1128 // T_OBJECT doesn't get spilled along with T_LONG. 1129 } else if (opr_type != T_LONG LP64_ONLY(&& opr_type != T_OBJECT)) { 1130 // integer instruction (note: long operands must always be in register) 1131 switch (op->code()) { 1132 case lir_cmp: 1133 case lir_add: 1134 case lir_sub: 1135 case lir_logic_and: 1136 case lir_logic_or: 1137 case lir_logic_xor: 1138 { 1139 assert(op->as_Op2() != NULL, "must be LIR_Op2"); 1140 LIR_Op2* op2 = (LIR_Op2*)op; 1141 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { 1142 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"); 1143 return shouldHaveRegister; 1144 } 1145 } 1146 } 1147 } 1148 #endif // X86 1149 1150 // all other operands require a register 1151 return mustHaveRegister; 1152 } 1153 1154 1155 void LinearScan::handle_method_arguments(LIR_Op* op) { 1156 // special handling for method arguments (moves from stack to virtual register): 1157 // the interval gets no register assigned, but the stack slot. 1158 // it is split before the first use by the register allocator. 1159 1160 if (op->code() == lir_move) { 1161 assert(op->as_Op1() != NULL, "must be LIR_Op1"); 1162 LIR_Op1* move = (LIR_Op1*)op; 1163 1164 if (move->in_opr()->is_stack()) { 1165 #ifdef ASSERT 1166 int arg_size = compilation()->method()->arg_size(); 1167 LIR_Opr o = move->in_opr(); 1168 if (o->is_single_stack()) { 1169 assert(o->single_stack_ix() >= 0 && o->single_stack_ix() < arg_size, "out of range"); 1170 } else if (o->is_double_stack()) { 1171 assert(o->double_stack_ix() >= 0 && o->double_stack_ix() < arg_size, "out of range"); 1172 } else { 1173 ShouldNotReachHere(); 1174 } 1175 1176 assert(move->id() > 0, "invalid id"); 1177 assert(block_of_op_with_id(move->id())->number_of_preds() == 0, "move from stack must be in first block"); 1178 assert(move->result_opr()->is_virtual(), "result of move must be a virtual register"); 1179 1180 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()))); 1181 #endif 1182 1183 Interval* interval = interval_at(reg_num(move->result_opr())); 1184 1185 int stack_slot = LinearScan::nof_regs + (move->in_opr()->is_single_stack() ? move->in_opr()->single_stack_ix() : move->in_opr()->double_stack_ix()); 1186 interval->set_canonical_spill_slot(stack_slot); 1187 interval->assign_reg(stack_slot); 1188 } 1189 } 1190 } 1191 1192 void LinearScan::handle_doubleword_moves(LIR_Op* op) { 1193 // special handling for doubleword move from memory to register: 1194 // in this case the registers of the input address and the result 1195 // registers must not overlap -> add a temp range for the input registers 1196 if (op->code() == lir_move) { 1197 assert(op->as_Op1() != NULL, "must be LIR_Op1"); 1198 LIR_Op1* move = (LIR_Op1*)op; 1199 1200 if (move->result_opr()->is_double_cpu() && move->in_opr()->is_pointer()) { 1201 LIR_Address* address = move->in_opr()->as_address_ptr(); 1202 if (address != NULL) { 1203 if (address->base()->is_valid()) { 1204 add_temp(address->base(), op->id(), noUse); 1205 } 1206 if (address->index()->is_valid()) { 1207 add_temp(address->index(), op->id(), noUse); 1208 } 1209 } 1210 } 1211 } 1212 } 1213 1214 void LinearScan::add_register_hints(LIR_Op* op) { 1215 switch (op->code()) { 1216 case lir_move: // fall through 1217 case lir_convert: { 1218 assert(op->as_Op1() != NULL, "lir_move, lir_convert must be LIR_Op1"); 1219 LIR_Op1* move = (LIR_Op1*)op; 1220 1221 LIR_Opr move_from = move->in_opr(); 1222 LIR_Opr move_to = move->result_opr(); 1223 1224 if (move_to->is_register() && move_from->is_register()) { 1225 Interval* from = interval_at(reg_num(move_from)); 1226 Interval* to = interval_at(reg_num(move_to)); 1227 if (from != NULL && to != NULL) { 1228 to->set_register_hint(from); 1229 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())); 1230 } 1231 } 1232 break; 1233 } 1234 case lir_cmove: { 1235 assert(op->as_Op2() != NULL, "lir_cmove must be LIR_Op2"); 1236 LIR_Op2* cmove = (LIR_Op2*)op; 1237 1238 LIR_Opr move_from = cmove->in_opr1(); 1239 LIR_Opr move_to = cmove->result_opr(); 1240 1241 if (move_to->is_register() && move_from->is_register()) { 1242 Interval* from = interval_at(reg_num(move_from)); 1243 Interval* to = interval_at(reg_num(move_to)); 1244 if (from != NULL && to != NULL) { 1245 to->set_register_hint(from); 1246 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())); 1247 } 1248 } 1249 break; 1250 } 1251 } 1252 } 1253 1254 1255 void LinearScan::build_intervals() { 1256 TIME_LINEAR_SCAN(timer_build_intervals); 1257 1258 // initialize interval list with expected number of intervals 1259 // (32 is added to have some space for split children without having to resize the list) 1260 _intervals = IntervalList(num_virtual_regs() + 32); 1261 // initialize all slots that are used by build_intervals 1262 _intervals.at_put_grow(num_virtual_regs() - 1, NULL, NULL); 1263 1264 // create a list with all caller-save registers (cpu, fpu, xmm) 1265 // when an instruction is a call, a temp range is created for all these registers 1266 int num_caller_save_registers = 0; 1267 int caller_save_registers[LinearScan::nof_regs]; 1268 1269 int i; 1270 for (i = 0; i < FrameMap::nof_caller_save_cpu_regs(); i++) { 1271 LIR_Opr opr = FrameMap::caller_save_cpu_reg_at(i); 1272 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands"); 1273 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register"); 1274 caller_save_registers[num_caller_save_registers++] = reg_num(opr); 1275 } 1276 1277 // temp ranges for fpu registers are only created when the method has 1278 // virtual fpu operands. Otherwise no allocation for fpu registers is 1279 // perfomed and so the temp ranges would be useless 1280 if (has_fpu_registers()) { 1281 #ifdef X86 1282 if (UseSSE < 2) { 1283 #endif 1284 for (i = 0; i < FrameMap::nof_caller_save_fpu_regs; i++) { 1285 LIR_Opr opr = FrameMap::caller_save_fpu_reg_at(i); 1286 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands"); 1287 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register"); 1288 caller_save_registers[num_caller_save_registers++] = reg_num(opr); 1289 } 1290 #ifdef X86 1291 } 1292 if (UseSSE > 0) { 1293 int num_caller_save_xmm_regs = FrameMap::nof_caller_save_xmm_regs; 1294 #if _LP64 1295 if (UseAVX < 3) { 1296 num_caller_save_xmm_regs = num_caller_save_xmm_regs / 2; 1297 } 1298 #endif 1299 for (i = 0; i < num_caller_save_xmm_regs; i ++) { 1300 LIR_Opr opr = FrameMap::caller_save_xmm_reg_at(i); 1301 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands"); 1302 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register"); 1303 caller_save_registers[num_caller_save_registers++] = reg_num(opr); 1304 } 1305 } 1306 #endif 1307 } 1308 assert(num_caller_save_registers <= LinearScan::nof_regs, "out of bounds"); 1309 1310 1311 LIR_OpVisitState visitor; 1312 1313 // iterate all blocks in reverse order 1314 for (i = block_count() - 1; i >= 0; i--) { 1315 BlockBegin* block = block_at(i); 1316 LIR_OpList* instructions = block->lir()->instructions_list(); 1317 int block_from = block->first_lir_instruction_id(); 1318 int block_to = block->last_lir_instruction_id(); 1319 1320 assert(block_from == instructions->at(0)->id(), "must be"); 1321 assert(block_to == instructions->at(instructions->length() - 1)->id(), "must be"); 1322 1323 // Update intervals for registers live at the end of this block; 1324 BitMap live = block->live_out(); 1325 int size = (int)live.size(); 1326 for (int number = (int)live.get_next_one_offset(0, size); number < size; number = (int)live.get_next_one_offset(number + 1, size)) { 1327 assert(live.at(number), "should not stop here otherwise"); 1328 assert(number >= LIR_OprDesc::vreg_base, "fixed intervals must not be live on block bounds"); 1329 TRACE_LINEAR_SCAN(2, tty->print_cr("live in %d to %d", number, block_to + 2)); 1330 1331 add_use(number, block_from, block_to + 2, noUse, T_ILLEGAL); 1332 1333 // add special use positions for loop-end blocks when the 1334 // interval is used anywhere inside this loop. It's possible 1335 // that the block was part of a non-natural loop, so it might 1336 // have an invalid loop index. 1337 if (block->is_set(BlockBegin::linear_scan_loop_end_flag) && 1338 block->loop_index() != -1 && 1339 is_interval_in_loop(number, block->loop_index())) { 1340 interval_at(number)->add_use_pos(block_to + 1, loopEndMarker); 1341 } 1342 } 1343 1344 // iterate all instructions of the block in reverse order. 1345 // skip the first instruction because it is always a label 1346 // definitions of intervals are processed before uses 1347 assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label"); 1348 for (int j = instructions->length() - 1; j >= 1; j--) { 1349 LIR_Op* op = instructions->at(j); 1350 int op_id = op->id(); 1351 1352 // visit operation to collect all operands 1353 visitor.visit(op); 1354 1355 // add a temp range for each register if operation destroys caller-save registers 1356 if (visitor.has_call()) { 1357 for (int k = 0; k < num_caller_save_registers; k++) { 1358 add_temp(caller_save_registers[k], op_id, noUse, T_ILLEGAL); 1359 } 1360 TRACE_LINEAR_SCAN(4, tty->print_cr("operation destroys all caller-save registers")); 1361 } 1362 1363 // Add any platform dependent temps 1364 pd_add_temps(op); 1365 1366 // visit definitions (output and temp operands) 1367 int k, n; 1368 n = visitor.opr_count(LIR_OpVisitState::outputMode); 1369 for (k = 0; k < n; k++) { 1370 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k); 1371 assert(opr->is_register(), "visitor should only return register operands"); 1372 add_def(opr, op_id, use_kind_of_output_operand(op, opr)); 1373 } 1374 1375 n = visitor.opr_count(LIR_OpVisitState::tempMode); 1376 for (k = 0; k < n; k++) { 1377 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k); 1378 assert(opr->is_register(), "visitor should only return register operands"); 1379 add_temp(opr, op_id, mustHaveRegister); 1380 } 1381 1382 // visit uses (input operands) 1383 n = visitor.opr_count(LIR_OpVisitState::inputMode); 1384 for (k = 0; k < n; k++) { 1385 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k); 1386 assert(opr->is_register(), "visitor should only return register operands"); 1387 add_use(opr, block_from, op_id, use_kind_of_input_operand(op, opr)); 1388 } 1389 1390 // Add uses of live locals from interpreter's point of view for proper 1391 // debug information generation 1392 // Treat these operands as temp values (if the life range is extended 1393 // to a call site, the value would be in a register at the call otherwise) 1394 n = visitor.info_count(); 1395 for (k = 0; k < n; k++) { 1396 CodeEmitInfo* info = visitor.info_at(k); 1397 ValueStack* stack = info->stack(); 1398 for_each_state_value(stack, value, 1399 add_use(value, block_from, op_id + 1, noUse); 1400 ); 1401 } 1402 1403 // special steps for some instructions (especially moves) 1404 handle_method_arguments(op); 1405 handle_doubleword_moves(op); 1406 add_register_hints(op); 1407 1408 } // end of instruction iteration 1409 } // end of block iteration 1410 1411 1412 // add the range [0, 1[ to all fixed intervals 1413 // -> the register allocator need not handle unhandled fixed intervals 1414 for (int n = 0; n < LinearScan::nof_regs; n++) { 1415 Interval* interval = interval_at(n); 1416 if (interval != NULL) { 1417 interval->add_range(0, 1); 1418 } 1419 } 1420 } 1421 1422 1423 // ********** Phase 5: actual register allocation 1424 1425 int LinearScan::interval_cmp(Interval** a, Interval** b) { 1426 if (*a != NULL) { 1427 if (*b != NULL) { 1428 return (*a)->from() - (*b)->from(); 1429 } else { 1430 return -1; 1431 } 1432 } else { 1433 if (*b != NULL) { 1434 return 1; 1435 } else { 1436 return 0; 1437 } 1438 } 1439 } 1440 1441 #ifndef PRODUCT 1442 bool LinearScan::is_sorted(IntervalArray* intervals) { 1443 int from = -1; 1444 int i, j; 1445 for (i = 0; i < intervals->length(); i ++) { 1446 Interval* it = intervals->at(i); 1447 if (it != NULL) { 1448 if (from > it->from()) { 1449 assert(false, ""); 1450 return false; 1451 } 1452 from = it->from(); 1453 } 1454 } 1455 1456 // check in both directions if sorted list and unsorted list contain same intervals 1457 for (i = 0; i < interval_count(); i++) { 1458 if (interval_at(i) != NULL) { 1459 int num_found = 0; 1460 for (j = 0; j < intervals->length(); j++) { 1461 if (interval_at(i) == intervals->at(j)) { 1462 num_found++; 1463 } 1464 } 1465 assert(num_found == 1, "lists do not contain same intervals"); 1466 } 1467 } 1468 for (j = 0; j < intervals->length(); j++) { 1469 int num_found = 0; 1470 for (i = 0; i < interval_count(); i++) { 1471 if (interval_at(i) == intervals->at(j)) { 1472 num_found++; 1473 } 1474 } 1475 assert(num_found == 1, "lists do not contain same intervals"); 1476 } 1477 1478 return true; 1479 } 1480 #endif 1481 1482 void LinearScan::add_to_list(Interval** first, Interval** prev, Interval* interval) { 1483 if (*prev != NULL) { 1484 (*prev)->set_next(interval); 1485 } else { 1486 *first = interval; 1487 } 1488 *prev = interval; 1489 } 1490 1491 void LinearScan::create_unhandled_lists(Interval** list1, Interval** list2, bool (is_list1)(const Interval* i), bool (is_list2)(const Interval* i)) { 1492 assert(is_sorted(_sorted_intervals), "interval list is not sorted"); 1493 1494 *list1 = *list2 = Interval::end(); 1495 1496 Interval* list1_prev = NULL; 1497 Interval* list2_prev = NULL; 1498 Interval* v; 1499 1500 const int n = _sorted_intervals->length(); 1501 for (int i = 0; i < n; i++) { 1502 v = _sorted_intervals->at(i); 1503 if (v == NULL) continue; 1504 1505 if (is_list1(v)) { 1506 add_to_list(list1, &list1_prev, v); 1507 } else if (is_list2 == NULL || is_list2(v)) { 1508 add_to_list(list2, &list2_prev, v); 1509 } 1510 } 1511 1512 if (list1_prev != NULL) list1_prev->set_next(Interval::end()); 1513 if (list2_prev != NULL) list2_prev->set_next(Interval::end()); 1514 1515 assert(list1_prev == NULL || list1_prev->next() == Interval::end(), "linear list ends not with sentinel"); 1516 assert(list2_prev == NULL || list2_prev->next() == Interval::end(), "linear list ends not with sentinel"); 1517 } 1518 1519 1520 void LinearScan::sort_intervals_before_allocation() { 1521 TIME_LINEAR_SCAN(timer_sort_intervals_before); 1522 1523 if (_needs_full_resort) { 1524 // There is no known reason why this should occur but just in case... 1525 assert(false, "should never occur"); 1526 // Re-sort existing interval list because an Interval::from() has changed 1527 _sorted_intervals->sort(interval_cmp); 1528 _needs_full_resort = false; 1529 } 1530 1531 IntervalList* unsorted_list = &_intervals; 1532 int unsorted_len = unsorted_list->length(); 1533 int sorted_len = 0; 1534 int unsorted_idx; 1535 int sorted_idx = 0; 1536 int sorted_from_max = -1; 1537 1538 // calc number of items for sorted list (sorted list must not contain NULL values) 1539 for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) { 1540 if (unsorted_list->at(unsorted_idx) != NULL) { 1541 sorted_len++; 1542 } 1543 } 1544 IntervalArray* sorted_list = new IntervalArray(sorted_len); 1545 1546 // special sorting algorithm: the original interval-list is almost sorted, 1547 // only some intervals are swapped. So this is much faster than a complete QuickSort 1548 for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) { 1549 Interval* cur_interval = unsorted_list->at(unsorted_idx); 1550 1551 if (cur_interval != NULL) { 1552 int cur_from = cur_interval->from(); 1553 1554 if (sorted_from_max <= cur_from) { 1555 sorted_list->at_put(sorted_idx++, cur_interval); 1556 sorted_from_max = cur_interval->from(); 1557 } else { 1558 // the asumption that the intervals are already sorted failed, 1559 // so this interval must be sorted in manually 1560 int j; 1561 for (j = sorted_idx - 1; j >= 0 && cur_from < sorted_list->at(j)->from(); j--) { 1562 sorted_list->at_put(j + 1, sorted_list->at(j)); 1563 } 1564 sorted_list->at_put(j + 1, cur_interval); 1565 sorted_idx++; 1566 } 1567 } 1568 } 1569 _sorted_intervals = sorted_list; 1570 assert(is_sorted(_sorted_intervals), "intervals unsorted"); 1571 } 1572 1573 void LinearScan::sort_intervals_after_allocation() { 1574 TIME_LINEAR_SCAN(timer_sort_intervals_after); 1575 1576 if (_needs_full_resort) { 1577 // Re-sort existing interval list because an Interval::from() has changed 1578 _sorted_intervals->sort(interval_cmp); 1579 _needs_full_resort = false; 1580 } 1581 1582 IntervalArray* old_list = _sorted_intervals; 1583 IntervalList* new_list = _new_intervals_from_allocation; 1584 int old_len = old_list->length(); 1585 int new_len = new_list->length(); 1586 1587 if (new_len == 0) { 1588 // no intervals have been added during allocation, so sorted list is already up to date 1589 assert(is_sorted(_sorted_intervals), "intervals unsorted"); 1590 return; 1591 } 1592 1593 // conventional sort-algorithm for new intervals 1594 new_list->sort(interval_cmp); 1595 1596 // merge old and new list (both already sorted) into one combined list 1597 IntervalArray* combined_list = new IntervalArray(old_len + new_len); 1598 int old_idx = 0; 1599 int new_idx = 0; 1600 1601 while (old_idx + new_idx < old_len + new_len) { 1602 if (new_idx >= new_len || (old_idx < old_len && old_list->at(old_idx)->from() <= new_list->at(new_idx)->from())) { 1603 combined_list->at_put(old_idx + new_idx, old_list->at(old_idx)); 1604 old_idx++; 1605 } else { 1606 combined_list->at_put(old_idx + new_idx, new_list->at(new_idx)); 1607 new_idx++; 1608 } 1609 } 1610 1611 _sorted_intervals = combined_list; 1612 assert(is_sorted(_sorted_intervals), "intervals unsorted"); 1613 } 1614 1615 1616 void LinearScan::allocate_registers() { 1617 TIME_LINEAR_SCAN(timer_allocate_registers); 1618 1619 Interval* precolored_cpu_intervals, *not_precolored_cpu_intervals; 1620 Interval* precolored_fpu_intervals, *not_precolored_fpu_intervals; 1621 1622 // allocate cpu registers 1623 create_unhandled_lists(&precolored_cpu_intervals, ¬_precolored_cpu_intervals, 1624 is_precolored_cpu_interval, is_virtual_cpu_interval); 1625 1626 // allocate fpu registers 1627 create_unhandled_lists(&precolored_fpu_intervals, ¬_precolored_fpu_intervals, 1628 is_precolored_fpu_interval, is_virtual_fpu_interval); 1629 1630 // the fpu interval allocation cannot be moved down below with the fpu section as 1631 // the cpu_lsw.walk() changes interval positions. 1632 1633 LinearScanWalker cpu_lsw(this, precolored_cpu_intervals, not_precolored_cpu_intervals); 1634 cpu_lsw.walk(); 1635 cpu_lsw.finish_allocation(); 1636 1637 if (has_fpu_registers()) { 1638 LinearScanWalker fpu_lsw(this, precolored_fpu_intervals, not_precolored_fpu_intervals); 1639 fpu_lsw.walk(); 1640 fpu_lsw.finish_allocation(); 1641 } 1642 } 1643 1644 1645 // ********** Phase 6: resolve data flow 1646 // (insert moves at edges between blocks if intervals have been split) 1647 1648 // wrapper for Interval::split_child_at_op_id that performs a bailout in product mode 1649 // instead of returning NULL 1650 Interval* LinearScan::split_child_at_op_id(Interval* interval, int op_id, LIR_OpVisitState::OprMode mode) { 1651 Interval* result = interval->split_child_at_op_id(op_id, mode); 1652 if (result != NULL) { 1653 return result; 1654 } 1655 1656 assert(false, "must find an interval, but do a clean bailout in product mode"); 1657 result = new Interval(LIR_OprDesc::vreg_base); 1658 result->assign_reg(0); 1659 result->set_type(T_INT); 1660 BAILOUT_("LinearScan: interval is NULL", result); 1661 } 1662 1663 1664 Interval* LinearScan::interval_at_block_begin(BlockBegin* block, int reg_num) { 1665 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds"); 1666 assert(interval_at(reg_num) != NULL, "no interval found"); 1667 1668 return split_child_at_op_id(interval_at(reg_num), block->first_lir_instruction_id(), LIR_OpVisitState::outputMode); 1669 } 1670 1671 Interval* LinearScan::interval_at_block_end(BlockBegin* block, int reg_num) { 1672 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds"); 1673 assert(interval_at(reg_num) != NULL, "no interval found"); 1674 1675 return split_child_at_op_id(interval_at(reg_num), block->last_lir_instruction_id() + 1, LIR_OpVisitState::outputMode); 1676 } 1677 1678 Interval* LinearScan::interval_at_op_id(int reg_num, int op_id) { 1679 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds"); 1680 assert(interval_at(reg_num) != NULL, "no interval found"); 1681 1682 return split_child_at_op_id(interval_at(reg_num), op_id, LIR_OpVisitState::inputMode); 1683 } 1684 1685 1686 void LinearScan::resolve_collect_mappings(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) { 1687 DEBUG_ONLY(move_resolver.check_empty()); 1688 1689 const int num_regs = num_virtual_regs(); 1690 const int size = live_set_size(); 1691 const BitMap live_at_edge = to_block->live_in(); 1692 1693 // visit all registers where the live_at_edge bit is set 1694 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)) { 1695 assert(r < num_regs, "live information set for not exisiting interval"); 1696 assert(from_block->live_out().at(r) && to_block->live_in().at(r), "interval not live at this edge"); 1697 1698 Interval* from_interval = interval_at_block_end(from_block, r); 1699 Interval* to_interval = interval_at_block_begin(to_block, r); 1700 1701 if (from_interval != to_interval && (from_interval->assigned_reg() != to_interval->assigned_reg() || from_interval->assigned_regHi() != to_interval->assigned_regHi())) { 1702 // need to insert move instruction 1703 move_resolver.add_mapping(from_interval, to_interval); 1704 } 1705 } 1706 } 1707 1708 1709 void LinearScan::resolve_find_insert_pos(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) { 1710 if (from_block->number_of_sux() <= 1) { 1711 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at end of from_block B%d", from_block->block_id())); 1712 1713 LIR_OpList* instructions = from_block->lir()->instructions_list(); 1714 LIR_OpBranch* branch = instructions->last()->as_OpBranch(); 1715 if (branch != NULL) { 1716 // insert moves before branch 1717 assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump"); 1718 move_resolver.set_insert_position(from_block->lir(), instructions->length() - 2); 1719 } else { 1720 move_resolver.set_insert_position(from_block->lir(), instructions->length() - 1); 1721 } 1722 1723 } else { 1724 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at beginning of to_block B%d", to_block->block_id())); 1725 #ifdef ASSERT 1726 assert(from_block->lir()->instructions_list()->at(0)->as_OpLabel() != NULL, "block does not start with a label"); 1727 1728 // because the number of predecessor edges matches the number of 1729 // successor edges, blocks which are reached by switch statements 1730 // may have be more than one predecessor but it will be guaranteed 1731 // that all predecessors will be the same. 1732 for (int i = 0; i < to_block->number_of_preds(); i++) { 1733 assert(from_block == to_block->pred_at(i), "all critical edges must be broken"); 1734 } 1735 #endif 1736 1737 move_resolver.set_insert_position(to_block->lir(), 0); 1738 } 1739 } 1740 1741 1742 // insert necessary moves (spilling or reloading) at edges between blocks if interval has been split 1743 void LinearScan::resolve_data_flow() { 1744 TIME_LINEAR_SCAN(timer_resolve_data_flow); 1745 1746 int num_blocks = block_count(); 1747 MoveResolver move_resolver(this); 1748 BitMap block_completed(num_blocks); block_completed.clear(); 1749 BitMap already_resolved(num_blocks); already_resolved.clear(); 1750 1751 int i; 1752 for (i = 0; i < num_blocks; i++) { 1753 BlockBegin* block = block_at(i); 1754 1755 // check if block has only one predecessor and only one successor 1756 if (block->number_of_preds() == 1 && block->number_of_sux() == 1 && block->number_of_exception_handlers() == 0) { 1757 LIR_OpList* instructions = block->lir()->instructions_list(); 1758 assert(instructions->at(0)->code() == lir_label, "block must start with label"); 1759 assert(instructions->last()->code() == lir_branch, "block with successors must end with branch"); 1760 assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block with successor must end with unconditional branch"); 1761 1762 // check if block is empty (only label and branch) 1763 if (instructions->length() == 2) { 1764 BlockBegin* pred = block->pred_at(0); 1765 BlockBegin* sux = block->sux_at(0); 1766 1767 // prevent optimization of two consecutive blocks 1768 if (!block_completed.at(pred->linear_scan_number()) && !block_completed.at(sux->linear_scan_number())) { 1769 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())); 1770 block_completed.set_bit(block->linear_scan_number()); 1771 1772 // directly resolve between pred and sux (without looking at the empty block between) 1773 resolve_collect_mappings(pred, sux, move_resolver); 1774 if (move_resolver.has_mappings()) { 1775 move_resolver.set_insert_position(block->lir(), 0); 1776 move_resolver.resolve_and_append_moves(); 1777 } 1778 } 1779 } 1780 } 1781 } 1782 1783 1784 for (i = 0; i < num_blocks; i++) { 1785 if (!block_completed.at(i)) { 1786 BlockBegin* from_block = block_at(i); 1787 already_resolved.set_from(block_completed); 1788 1789 int num_sux = from_block->number_of_sux(); 1790 for (int s = 0; s < num_sux; s++) { 1791 BlockBegin* to_block = from_block->sux_at(s); 1792 1793 // check for duplicate edges between the same blocks (can happen with switch blocks) 1794 if (!already_resolved.at(to_block->linear_scan_number())) { 1795 TRACE_LINEAR_SCAN(3, tty->print_cr("**** processing edge between B%d and B%d", from_block->block_id(), to_block->block_id())); 1796 already_resolved.set_bit(to_block->linear_scan_number()); 1797 1798 // collect all intervals that have been split between from_block and to_block 1799 resolve_collect_mappings(from_block, to_block, move_resolver); 1800 if (move_resolver.has_mappings()) { 1801 resolve_find_insert_pos(from_block, to_block, move_resolver); 1802 move_resolver.resolve_and_append_moves(); 1803 } 1804 } 1805 } 1806 } 1807 } 1808 } 1809 1810 1811 void LinearScan::resolve_exception_entry(BlockBegin* block, int reg_num, MoveResolver &move_resolver) { 1812 if (interval_at(reg_num) == NULL) { 1813 // if a phi function is never used, no interval is created -> ignore this 1814 return; 1815 } 1816 1817 Interval* interval = interval_at_block_begin(block, reg_num); 1818 int reg = interval->assigned_reg(); 1819 int regHi = interval->assigned_regHi(); 1820 1821 if ((reg < nof_regs && interval->always_in_memory()) || 1822 (use_fpu_stack_allocation() && reg >= pd_first_fpu_reg && reg <= pd_last_fpu_reg)) { 1823 // the interval is split to get a short range that is located on the stack 1824 // in the following two cases: 1825 // * the interval started in memory (e.g. method parameter), but is currently in a register 1826 // this is an optimization for exception handling that reduces the number of moves that 1827 // are necessary for resolving the states when an exception uses this exception handler 1828 // * the interval would be on the fpu stack at the begin of the exception handler 1829 // this is not allowed because of the complicated fpu stack handling on Intel 1830 1831 // range that will be spilled to memory 1832 int from_op_id = block->first_lir_instruction_id(); 1833 int to_op_id = from_op_id + 1; // short live range of length 1 1834 assert(interval->from() <= from_op_id && interval->to() >= to_op_id, 1835 "no split allowed between exception entry and first instruction"); 1836 1837 if (interval->from() != from_op_id) { 1838 // the part before from_op_id is unchanged 1839 interval = interval->split(from_op_id); 1840 interval->assign_reg(reg, regHi); 1841 append_interval(interval); 1842 } else { 1843 _needs_full_resort = true; 1844 } 1845 assert(interval->from() == from_op_id, "must be true now"); 1846 1847 Interval* spilled_part = interval; 1848 if (interval->to() != to_op_id) { 1849 // the part after to_op_id is unchanged 1850 spilled_part = interval->split_from_start(to_op_id); 1851 append_interval(spilled_part); 1852 move_resolver.add_mapping(spilled_part, interval); 1853 } 1854 assign_spill_slot(spilled_part); 1855 1856 assert(spilled_part->from() == from_op_id && spilled_part->to() == to_op_id, "just checking"); 1857 } 1858 } 1859 1860 void LinearScan::resolve_exception_entry(BlockBegin* block, MoveResolver &move_resolver) { 1861 assert(block->is_set(BlockBegin::exception_entry_flag), "should not call otherwise"); 1862 DEBUG_ONLY(move_resolver.check_empty()); 1863 1864 // visit all registers where the live_in bit is set 1865 int size = live_set_size(); 1866 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)) { 1867 resolve_exception_entry(block, r, move_resolver); 1868 } 1869 1870 // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately 1871 for_each_phi_fun(block, phi, 1872 resolve_exception_entry(block, phi->operand()->vreg_number(), move_resolver) 1873 ); 1874 1875 if (move_resolver.has_mappings()) { 1876 // insert moves after first instruction 1877 move_resolver.set_insert_position(block->lir(), 0); 1878 move_resolver.resolve_and_append_moves(); 1879 } 1880 } 1881 1882 1883 void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, int reg_num, Phi* phi, MoveResolver &move_resolver) { 1884 if (interval_at(reg_num) == NULL) { 1885 // if a phi function is never used, no interval is created -> ignore this 1886 return; 1887 } 1888 1889 // the computation of to_interval is equal to resolve_collect_mappings, 1890 // but from_interval is more complicated because of phi functions 1891 BlockBegin* to_block = handler->entry_block(); 1892 Interval* to_interval = interval_at_block_begin(to_block, reg_num); 1893 1894 if (phi != NULL) { 1895 // phi function of the exception entry block 1896 // no moves are created for this phi function in the LIR_Generator, so the 1897 // interval at the throwing instruction must be searched using the operands 1898 // of the phi function 1899 Value from_value = phi->operand_at(handler->phi_operand()); 1900 1901 // with phi functions it can happen that the same from_value is used in 1902 // multiple mappings, so notify move-resolver that this is allowed 1903 move_resolver.set_multiple_reads_allowed(); 1904 1905 Constant* con = from_value->as_Constant(); 1906 if (con != NULL && !con->is_pinned()) { 1907 // unpinned constants may have no register, so add mapping from constant to interval 1908 move_resolver.add_mapping(LIR_OprFact::value_type(con->type()), to_interval); 1909 } else { 1910 // search split child at the throwing op_id 1911 Interval* from_interval = interval_at_op_id(from_value->operand()->vreg_number(), throwing_op_id); 1912 move_resolver.add_mapping(from_interval, to_interval); 1913 } 1914 1915 } else { 1916 // no phi function, so use reg_num also for from_interval 1917 // search split child at the throwing op_id 1918 Interval* from_interval = interval_at_op_id(reg_num, throwing_op_id); 1919 if (from_interval != to_interval) { 1920 // optimization to reduce number of moves: when to_interval is on stack and 1921 // the stack slot is known to be always correct, then no move is necessary 1922 if (!from_interval->always_in_memory() || from_interval->canonical_spill_slot() != to_interval->assigned_reg()) { 1923 move_resolver.add_mapping(from_interval, to_interval); 1924 } 1925 } 1926 } 1927 } 1928 1929 void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, MoveResolver &move_resolver) { 1930 TRACE_LINEAR_SCAN(4, tty->print_cr("resolving exception handler B%d: throwing_op_id=%d", handler->entry_block()->block_id(), throwing_op_id)); 1931 1932 DEBUG_ONLY(move_resolver.check_empty()); 1933 assert(handler->lir_op_id() == -1, "already processed this xhandler"); 1934 DEBUG_ONLY(handler->set_lir_op_id(throwing_op_id)); 1935 assert(handler->entry_code() == NULL, "code already present"); 1936 1937 // visit all registers where the live_in bit is set 1938 BlockBegin* block = handler->entry_block(); 1939 int size = live_set_size(); 1940 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)) { 1941 resolve_exception_edge(handler, throwing_op_id, r, NULL, move_resolver); 1942 } 1943 1944 // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately 1945 for_each_phi_fun(block, phi, 1946 resolve_exception_edge(handler, throwing_op_id, phi->operand()->vreg_number(), phi, move_resolver) 1947 ); 1948 1949 if (move_resolver.has_mappings()) { 1950 LIR_List* entry_code = new LIR_List(compilation()); 1951 move_resolver.set_insert_position(entry_code, 0); 1952 move_resolver.resolve_and_append_moves(); 1953 1954 entry_code->jump(handler->entry_block()); 1955 handler->set_entry_code(entry_code); 1956 } 1957 } 1958 1959 1960 void LinearScan::resolve_exception_handlers() { 1961 MoveResolver move_resolver(this); 1962 LIR_OpVisitState visitor; 1963 int num_blocks = block_count(); 1964 1965 int i; 1966 for (i = 0; i < num_blocks; i++) { 1967 BlockBegin* block = block_at(i); 1968 if (block->is_set(BlockBegin::exception_entry_flag)) { 1969 resolve_exception_entry(block, move_resolver); 1970 } 1971 } 1972 1973 for (i = 0; i < num_blocks; i++) { 1974 BlockBegin* block = block_at(i); 1975 LIR_List* ops = block->lir(); 1976 int num_ops = ops->length(); 1977 1978 // iterate all instructions of the block. skip the first because it is always a label 1979 assert(visitor.no_operands(ops->at(0)), "first operation must always be a label"); 1980 for (int j = 1; j < num_ops; j++) { 1981 LIR_Op* op = ops->at(j); 1982 int op_id = op->id(); 1983 1984 if (op_id != -1 && has_info(op_id)) { 1985 // visit operation to collect all operands 1986 visitor.visit(op); 1987 assert(visitor.info_count() > 0, "should not visit otherwise"); 1988 1989 XHandlers* xhandlers = visitor.all_xhandler(); 1990 int n = xhandlers->length(); 1991 for (int k = 0; k < n; k++) { 1992 resolve_exception_edge(xhandlers->handler_at(k), op_id, move_resolver); 1993 } 1994 1995 #ifdef ASSERT 1996 } else { 1997 visitor.visit(op); 1998 assert(visitor.all_xhandler()->length() == 0, "missed exception handler"); 1999 #endif 2000 } 2001 } 2002 } 2003 } 2004 2005 2006 // ********** Phase 7: assign register numbers back to LIR 2007 // (includes computation of debug information and oop maps) 2008 2009 VMReg LinearScan::vm_reg_for_interval(Interval* interval) { 2010 VMReg reg = interval->cached_vm_reg(); 2011 if (!reg->is_valid() ) { 2012 reg = vm_reg_for_operand(operand_for_interval(interval)); 2013 interval->set_cached_vm_reg(reg); 2014 } 2015 assert(reg == vm_reg_for_operand(operand_for_interval(interval)), "wrong cached value"); 2016 return reg; 2017 } 2018 2019 VMReg LinearScan::vm_reg_for_operand(LIR_Opr opr) { 2020 assert(opr->is_oop(), "currently only implemented for oop operands"); 2021 return frame_map()->regname(opr); 2022 } 2023 2024 2025 LIR_Opr LinearScan::operand_for_interval(Interval* interval) { 2026 LIR_Opr opr = interval->cached_opr(); 2027 if (opr->is_illegal()) { 2028 opr = calc_operand_for_interval(interval); 2029 interval->set_cached_opr(opr); 2030 } 2031 2032 assert(opr == calc_operand_for_interval(interval), "wrong cached value"); 2033 return opr; 2034 } 2035 2036 LIR_Opr LinearScan::calc_operand_for_interval(const Interval* interval) { 2037 int assigned_reg = interval->assigned_reg(); 2038 BasicType type = interval->type(); 2039 2040 if (assigned_reg >= nof_regs) { 2041 // stack slot 2042 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2043 return LIR_OprFact::stack(assigned_reg - nof_regs, type); 2044 2045 } else { 2046 // register 2047 switch (type) { 2048 case T_OBJECT: { 2049 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2050 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2051 return LIR_OprFact::single_cpu_oop(assigned_reg); 2052 } 2053 2054 case T_ADDRESS: { 2055 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2056 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2057 return LIR_OprFact::single_cpu_address(assigned_reg); 2058 } 2059 2060 case T_METADATA: { 2061 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2062 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2063 return LIR_OprFact::single_cpu_metadata(assigned_reg); 2064 } 2065 2066 #ifdef __SOFTFP__ 2067 case T_FLOAT: // fall through 2068 #endif // __SOFTFP__ 2069 case T_INT: { 2070 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2071 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2072 return LIR_OprFact::single_cpu(assigned_reg); 2073 } 2074 2075 #ifdef __SOFTFP__ 2076 case T_DOUBLE: // fall through 2077 #endif // __SOFTFP__ 2078 case T_LONG: { 2079 int assigned_regHi = interval->assigned_regHi(); 2080 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2081 assert(num_physical_regs(T_LONG) == 1 || 2082 (assigned_regHi >= pd_first_cpu_reg && assigned_regHi <= pd_last_cpu_reg), "no cpu register"); 2083 2084 assert(assigned_reg != assigned_regHi, "invalid allocation"); 2085 assert(num_physical_regs(T_LONG) == 1 || assigned_reg < assigned_regHi, 2086 "register numbers must be sorted (ensure that e.g. a move from eax,ebx to ebx,eax can not occur)"); 2087 assert((assigned_regHi != any_reg) ^ (num_physical_regs(T_LONG) == 1), "must be match"); 2088 if (requires_adjacent_regs(T_LONG)) { 2089 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == assigned_regHi, "must be sequential and even"); 2090 } 2091 2092 #ifdef _LP64 2093 return LIR_OprFact::double_cpu(assigned_reg, assigned_reg); 2094 #else 2095 #if defined(SPARC) || defined(PPC) 2096 return LIR_OprFact::double_cpu(assigned_regHi, assigned_reg); 2097 #else 2098 return LIR_OprFact::double_cpu(assigned_reg, assigned_regHi); 2099 #endif // SPARC 2100 #endif // LP64 2101 } 2102 2103 #ifndef __SOFTFP__ 2104 case T_FLOAT: { 2105 #ifdef X86 2106 if (UseSSE >= 1) { 2107 int last_xmm_reg = pd_last_xmm_reg; 2108 #ifdef _LP64 2109 if (UseAVX < 3) { 2110 last_xmm_reg = pd_first_xmm_reg + (pd_nof_xmm_regs_frame_map / 2) - 1; 2111 } 2112 #endif 2113 assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= last_xmm_reg, "no xmm register"); 2114 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2115 return LIR_OprFact::single_xmm(assigned_reg - pd_first_xmm_reg); 2116 } 2117 #endif 2118 2119 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register"); 2120 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2121 return LIR_OprFact::single_fpu(assigned_reg - pd_first_fpu_reg); 2122 } 2123 2124 case T_DOUBLE: { 2125 #ifdef X86 2126 if (UseSSE >= 2) { 2127 int last_xmm_reg = pd_last_xmm_reg; 2128 #ifdef _LP64 2129 if (UseAVX < 3) { 2130 last_xmm_reg = pd_first_xmm_reg + (pd_nof_xmm_regs_frame_map / 2) - 1; 2131 } 2132 #endif 2133 assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= last_xmm_reg, "no xmm register"); 2134 assert(interval->assigned_regHi() == any_reg, "must not have hi register (double xmm values are stored in one register)"); 2135 return LIR_OprFact::double_xmm(assigned_reg - pd_first_xmm_reg); 2136 } 2137 #endif 2138 2139 #ifdef SPARC 2140 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register"); 2141 assert(interval->assigned_regHi() >= pd_first_fpu_reg && interval->assigned_regHi() <= pd_last_fpu_reg, "no fpu register"); 2142 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == interval->assigned_regHi(), "must be sequential and even"); 2143 LIR_Opr result = LIR_OprFact::double_fpu(interval->assigned_regHi() - pd_first_fpu_reg, assigned_reg - pd_first_fpu_reg); 2144 #elif defined(ARM32) 2145 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register"); 2146 assert(interval->assigned_regHi() >= pd_first_fpu_reg && interval->assigned_regHi() <= pd_last_fpu_reg, "no fpu register"); 2147 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == interval->assigned_regHi(), "must be sequential and even"); 2148 LIR_Opr result = LIR_OprFact::double_fpu(assigned_reg - pd_first_fpu_reg, interval->assigned_regHi() - pd_first_fpu_reg); 2149 #else 2150 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register"); 2151 assert(interval->assigned_regHi() == any_reg, "must not have hi register (double fpu values are stored in one register on Intel)"); 2152 LIR_Opr result = LIR_OprFact::double_fpu(assigned_reg - pd_first_fpu_reg); 2153 #endif 2154 return result; 2155 } 2156 #endif // __SOFTFP__ 2157 2158 default: { 2159 ShouldNotReachHere(); 2160 return LIR_OprFact::illegalOpr; 2161 } 2162 } 2163 } 2164 } 2165 2166 LIR_Opr LinearScan::canonical_spill_opr(Interval* interval) { 2167 assert(interval->canonical_spill_slot() >= nof_regs, "canonical spill slot not set"); 2168 return LIR_OprFact::stack(interval->canonical_spill_slot() - nof_regs, interval->type()); 2169 } 2170 2171 LIR_Opr LinearScan::color_lir_opr(LIR_Opr opr, int op_id, LIR_OpVisitState::OprMode mode) { 2172 assert(opr->is_virtual(), "should not call this otherwise"); 2173 2174 Interval* interval = interval_at(opr->vreg_number()); 2175 assert(interval != NULL, "interval must exist"); 2176 2177 if (op_id != -1) { 2178 #ifdef ASSERT 2179 BlockBegin* block = block_of_op_with_id(op_id); 2180 if (block->number_of_sux() <= 1 && op_id == block->last_lir_instruction_id()) { 2181 // check if spill moves could have been appended at the end of this block, but 2182 // before the branch instruction. So the split child information for this branch would 2183 // be incorrect. 2184 LIR_OpBranch* branch = block->lir()->instructions_list()->last()->as_OpBranch(); 2185 if (branch != NULL) { 2186 if (block->live_out().at(opr->vreg_number())) { 2187 assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump"); 2188 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)"); 2189 } 2190 } 2191 } 2192 #endif 2193 2194 // operands are not changed when an interval is split during allocation, 2195 // so search the right interval here 2196 interval = split_child_at_op_id(interval, op_id, mode); 2197 } 2198 2199 LIR_Opr res = operand_for_interval(interval); 2200 2201 #ifdef X86 2202 // new semantic for is_last_use: not only set on definite end of interval, 2203 // but also before hole 2204 // This may still miss some cases (e.g. for dead values), but it is not necessary that the 2205 // last use information is completely correct 2206 // information is only needed for fpu stack allocation 2207 if (res->is_fpu_register()) { 2208 if (opr->is_last_use() || op_id == interval->to() || (op_id != -1 && interval->has_hole_between(op_id, op_id + 1))) { 2209 assert(op_id == -1 || !is_block_begin(op_id), "holes at begin of block may also result from control flow"); 2210 res = res->make_last_use(); 2211 } 2212 } 2213 #endif 2214 2215 assert(!gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::callee_saved) || !FrameMap::is_caller_save_register(res), "bad allocation"); 2216 2217 return res; 2218 } 2219 2220 2221 #ifdef ASSERT 2222 // some methods used to check correctness of debug information 2223 2224 void assert_no_register_values(GrowableArray<ScopeValue*>* values) { 2225 if (values == NULL) { 2226 return; 2227 } 2228 2229 for (int i = 0; i < values->length(); i++) { 2230 ScopeValue* value = values->at(i); 2231 2232 if (value->is_location()) { 2233 Location location = ((LocationValue*)value)->location(); 2234 assert(location.where() == Location::on_stack, "value is in register"); 2235 } 2236 } 2237 } 2238 2239 void assert_no_register_values(GrowableArray<MonitorValue*>* values) { 2240 if (values == NULL) { 2241 return; 2242 } 2243 2244 for (int i = 0; i < values->length(); i++) { 2245 MonitorValue* value = values->at(i); 2246 2247 if (value->owner()->is_location()) { 2248 Location location = ((LocationValue*)value->owner())->location(); 2249 assert(location.where() == Location::on_stack, "owner is in register"); 2250 } 2251 assert(value->basic_lock().where() == Location::on_stack, "basic_lock is in register"); 2252 } 2253 } 2254 2255 void assert_equal(Location l1, Location l2) { 2256 assert(l1.where() == l2.where() && l1.type() == l2.type() && l1.offset() == l2.offset(), ""); 2257 } 2258 2259 void assert_equal(ScopeValue* v1, ScopeValue* v2) { 2260 if (v1->is_location()) { 2261 assert(v2->is_location(), ""); 2262 assert_equal(((LocationValue*)v1)->location(), ((LocationValue*)v2)->location()); 2263 } else if (v1->is_constant_int()) { 2264 assert(v2->is_constant_int(), ""); 2265 assert(((ConstantIntValue*)v1)->value() == ((ConstantIntValue*)v2)->value(), ""); 2266 } else if (v1->is_constant_double()) { 2267 assert(v2->is_constant_double(), ""); 2268 assert(((ConstantDoubleValue*)v1)->value() == ((ConstantDoubleValue*)v2)->value(), ""); 2269 } else if (v1->is_constant_long()) { 2270 assert(v2->is_constant_long(), ""); 2271 assert(((ConstantLongValue*)v1)->value() == ((ConstantLongValue*)v2)->value(), ""); 2272 } else if (v1->is_constant_oop()) { 2273 assert(v2->is_constant_oop(), ""); 2274 assert(((ConstantOopWriteValue*)v1)->value() == ((ConstantOopWriteValue*)v2)->value(), ""); 2275 } else { 2276 ShouldNotReachHere(); 2277 } 2278 } 2279 2280 void assert_equal(MonitorValue* m1, MonitorValue* m2) { 2281 assert_equal(m1->owner(), m2->owner()); 2282 assert_equal(m1->basic_lock(), m2->basic_lock()); 2283 } 2284 2285 void assert_equal(IRScopeDebugInfo* d1, IRScopeDebugInfo* d2) { 2286 assert(d1->scope() == d2->scope(), "not equal"); 2287 assert(d1->bci() == d2->bci(), "not equal"); 2288 2289 if (d1->locals() != NULL) { 2290 assert(d1->locals() != NULL && d2->locals() != NULL, "not equal"); 2291 assert(d1->locals()->length() == d2->locals()->length(), "not equal"); 2292 for (int i = 0; i < d1->locals()->length(); i++) { 2293 assert_equal(d1->locals()->at(i), d2->locals()->at(i)); 2294 } 2295 } else { 2296 assert(d1->locals() == NULL && d2->locals() == NULL, "not equal"); 2297 } 2298 2299 if (d1->expressions() != NULL) { 2300 assert(d1->expressions() != NULL && d2->expressions() != NULL, "not equal"); 2301 assert(d1->expressions()->length() == d2->expressions()->length(), "not equal"); 2302 for (int i = 0; i < d1->expressions()->length(); i++) { 2303 assert_equal(d1->expressions()->at(i), d2->expressions()->at(i)); 2304 } 2305 } else { 2306 assert(d1->expressions() == NULL && d2->expressions() == NULL, "not equal"); 2307 } 2308 2309 if (d1->monitors() != NULL) { 2310 assert(d1->monitors() != NULL && d2->monitors() != NULL, "not equal"); 2311 assert(d1->monitors()->length() == d2->monitors()->length(), "not equal"); 2312 for (int i = 0; i < d1->monitors()->length(); i++) { 2313 assert_equal(d1->monitors()->at(i), d2->monitors()->at(i)); 2314 } 2315 } else { 2316 assert(d1->monitors() == NULL && d2->monitors() == NULL, "not equal"); 2317 } 2318 2319 if (d1->caller() != NULL) { 2320 assert(d1->caller() != NULL && d2->caller() != NULL, "not equal"); 2321 assert_equal(d1->caller(), d2->caller()); 2322 } else { 2323 assert(d1->caller() == NULL && d2->caller() == NULL, "not equal"); 2324 } 2325 } 2326 2327 void check_stack_depth(CodeEmitInfo* info, int stack_end) { 2328 if (info->stack()->bci() != SynchronizationEntryBCI && !info->scope()->method()->is_native()) { 2329 Bytecodes::Code code = info->scope()->method()->java_code_at_bci(info->stack()->bci()); 2330 switch (code) { 2331 case Bytecodes::_ifnull : // fall through 2332 case Bytecodes::_ifnonnull : // fall through 2333 case Bytecodes::_ifeq : // fall through 2334 case Bytecodes::_ifne : // fall through 2335 case Bytecodes::_iflt : // fall through 2336 case Bytecodes::_ifge : // fall through 2337 case Bytecodes::_ifgt : // fall through 2338 case Bytecodes::_ifle : // fall through 2339 case Bytecodes::_if_icmpeq : // fall through 2340 case Bytecodes::_if_icmpne : // fall through 2341 case Bytecodes::_if_icmplt : // fall through 2342 case Bytecodes::_if_icmpge : // fall through 2343 case Bytecodes::_if_icmpgt : // fall through 2344 case Bytecodes::_if_icmple : // fall through 2345 case Bytecodes::_if_acmpeq : // fall through 2346 case Bytecodes::_if_acmpne : 2347 assert(stack_end >= -Bytecodes::depth(code), "must have non-empty expression stack at if bytecode"); 2348 break; 2349 } 2350 } 2351 } 2352 2353 #endif // ASSERT 2354 2355 2356 IntervalWalker* LinearScan::init_compute_oop_maps() { 2357 // setup lists of potential oops for walking 2358 Interval* oop_intervals; 2359 Interval* non_oop_intervals; 2360 2361 create_unhandled_lists(&oop_intervals, &non_oop_intervals, is_oop_interval, NULL); 2362 2363 // intervals that have no oops inside need not to be processed 2364 // to ensure a walking until the last instruction id, add a dummy interval 2365 // with a high operation id 2366 non_oop_intervals = new Interval(any_reg); 2367 non_oop_intervals->add_range(max_jint - 2, max_jint - 1); 2368 2369 return new IntervalWalker(this, oop_intervals, non_oop_intervals); 2370 } 2371 2372 2373 OopMap* LinearScan::compute_oop_map(IntervalWalker* iw, LIR_Op* op, CodeEmitInfo* info, bool is_call_site) { 2374 TRACE_LINEAR_SCAN(3, tty->print_cr("creating oop map at op_id %d", op->id())); 2375 2376 // walk before the current operation -> intervals that start at 2377 // the operation (= output operands of the operation) are not 2378 // included in the oop map 2379 iw->walk_before(op->id()); 2380 2381 int frame_size = frame_map()->framesize(); 2382 int arg_count = frame_map()->oop_map_arg_count(); 2383 OopMap* map = new OopMap(frame_size, arg_count); 2384 2385 // Iterate through active intervals 2386 for (Interval* interval = iw->active_first(fixedKind); interval != Interval::end(); interval = interval->next()) { 2387 int assigned_reg = interval->assigned_reg(); 2388 2389 assert(interval->current_from() <= op->id() && op->id() <= interval->current_to(), "interval should not be active otherwise"); 2390 assert(interval->assigned_regHi() == any_reg, "oop must be single word"); 2391 assert(interval->reg_num() >= LIR_OprDesc::vreg_base, "fixed interval found"); 2392 2393 // Check if this range covers the instruction. Intervals that 2394 // start or end at the current operation are not included in the 2395 // oop map, except in the case of patching moves. For patching 2396 // moves, any intervals which end at this instruction are included 2397 // in the oop map since we may safepoint while doing the patch 2398 // before we've consumed the inputs. 2399 if (op->is_patching() || op->id() < interval->current_to()) { 2400 2401 // caller-save registers must not be included into oop-maps at calls 2402 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"); 2403 2404 VMReg name = vm_reg_for_interval(interval); 2405 set_oop(map, name); 2406 2407 // Spill optimization: when the stack value is guaranteed to be always correct, 2408 // then it must be added to the oop map even if the interval is currently in a register 2409 if (interval->always_in_memory() && 2410 op->id() > interval->spill_definition_pos() && 2411 interval->assigned_reg() != interval->canonical_spill_slot()) { 2412 assert(interval->spill_definition_pos() > 0, "position not set correctly"); 2413 assert(interval->canonical_spill_slot() >= LinearScan::nof_regs, "no spill slot assigned"); 2414 assert(interval->assigned_reg() < LinearScan::nof_regs, "interval is on stack, so stack slot is registered twice"); 2415 2416 set_oop(map, frame_map()->slot_regname(interval->canonical_spill_slot() - LinearScan::nof_regs)); 2417 } 2418 } 2419 } 2420 2421 // add oops from lock stack 2422 assert(info->stack() != NULL, "CodeEmitInfo must always have a stack"); 2423 int locks_count = info->stack()->total_locks_size(); 2424 for (int i = 0; i < locks_count; i++) { 2425 set_oop(map, frame_map()->monitor_object_regname(i)); 2426 } 2427 2428 return map; 2429 } 2430 2431 2432 void LinearScan::compute_oop_map(IntervalWalker* iw, const LIR_OpVisitState &visitor, LIR_Op* op) { 2433 assert(visitor.info_count() > 0, "no oop map needed"); 2434 2435 // compute oop_map only for first CodeEmitInfo 2436 // because it is (in most cases) equal for all other infos of the same operation 2437 CodeEmitInfo* first_info = visitor.info_at(0); 2438 OopMap* first_oop_map = compute_oop_map(iw, op, first_info, visitor.has_call()); 2439 2440 for (int i = 0; i < visitor.info_count(); i++) { 2441 CodeEmitInfo* info = visitor.info_at(i); 2442 OopMap* oop_map = first_oop_map; 2443 2444 // compute worst case interpreter size in case of a deoptimization 2445 _compilation->update_interpreter_frame_size(info->interpreter_frame_size()); 2446 2447 if (info->stack()->locks_size() != first_info->stack()->locks_size()) { 2448 // this info has a different number of locks then the precomputed oop map 2449 // (possible for lock and unlock instructions) -> compute oop map with 2450 // correct lock information 2451 oop_map = compute_oop_map(iw, op, info, visitor.has_call()); 2452 } 2453 2454 if (info->_oop_map == NULL) { 2455 info->_oop_map = oop_map; 2456 } else { 2457 // a CodeEmitInfo can not be shared between different LIR-instructions 2458 // because interval splitting can occur anywhere between two instructions 2459 // and so the oop maps must be different 2460 // -> check if the already set oop_map is exactly the one calculated for this operation 2461 assert(info->_oop_map == oop_map, "same CodeEmitInfo used for multiple LIR instructions"); 2462 } 2463 } 2464 } 2465 2466 2467 // frequently used constants 2468 // Allocate them with new so they are never destroyed (otherwise, a 2469 // forced exit could destroy these objects while they are still in 2470 // use). 2471 ConstantOopWriteValue* LinearScan::_oop_null_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantOopWriteValue(NULL); 2472 ConstantIntValue* LinearScan::_int_m1_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(-1); 2473 ConstantIntValue* LinearScan::_int_0_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(0); 2474 ConstantIntValue* LinearScan::_int_1_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(1); 2475 ConstantIntValue* LinearScan::_int_2_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(2); 2476 LocationValue* _illegal_value = new (ResourceObj::C_HEAP, mtCompiler) LocationValue(Location()); 2477 2478 void LinearScan::init_compute_debug_info() { 2479 // cache for frequently used scope values 2480 // (cpu registers and stack slots) 2481 _scope_value_cache = ScopeValueArray((LinearScan::nof_cpu_regs + frame_map()->argcount() + max_spills()) * 2, NULL); 2482 } 2483 2484 MonitorValue* LinearScan::location_for_monitor_index(int monitor_index) { 2485 Location loc; 2486 if (!frame_map()->location_for_monitor_object(monitor_index, &loc)) { 2487 bailout("too large frame"); 2488 } 2489 ScopeValue* object_scope_value = new LocationValue(loc); 2490 2491 if (!frame_map()->location_for_monitor_lock(monitor_index, &loc)) { 2492 bailout("too large frame"); 2493 } 2494 return new MonitorValue(object_scope_value, loc); 2495 } 2496 2497 LocationValue* LinearScan::location_for_name(int name, Location::Type loc_type) { 2498 Location loc; 2499 if (!frame_map()->locations_for_slot(name, loc_type, &loc)) { 2500 bailout("too large frame"); 2501 } 2502 return new LocationValue(loc); 2503 } 2504 2505 2506 int LinearScan::append_scope_value_for_constant(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) { 2507 assert(opr->is_constant(), "should not be called otherwise"); 2508 2509 LIR_Const* c = opr->as_constant_ptr(); 2510 BasicType t = c->type(); 2511 switch (t) { 2512 case T_OBJECT: { 2513 jobject value = c->as_jobject(); 2514 if (value == NULL) { 2515 scope_values->append(_oop_null_scope_value); 2516 } else { 2517 scope_values->append(new ConstantOopWriteValue(c->as_jobject())); 2518 } 2519 return 1; 2520 } 2521 2522 case T_INT: // fall through 2523 case T_FLOAT: { 2524 int value = c->as_jint_bits(); 2525 switch (value) { 2526 case -1: scope_values->append(_int_m1_scope_value); break; 2527 case 0: scope_values->append(_int_0_scope_value); break; 2528 case 1: scope_values->append(_int_1_scope_value); break; 2529 case 2: scope_values->append(_int_2_scope_value); break; 2530 default: scope_values->append(new ConstantIntValue(c->as_jint_bits())); break; 2531 } 2532 return 1; 2533 } 2534 2535 case T_LONG: // fall through 2536 case T_DOUBLE: { 2537 #ifdef _LP64 2538 scope_values->append(_int_0_scope_value); 2539 scope_values->append(new ConstantLongValue(c->as_jlong_bits())); 2540 #else 2541 if (hi_word_offset_in_bytes > lo_word_offset_in_bytes) { 2542 scope_values->append(new ConstantIntValue(c->as_jint_hi_bits())); 2543 scope_values->append(new ConstantIntValue(c->as_jint_lo_bits())); 2544 } else { 2545 scope_values->append(new ConstantIntValue(c->as_jint_lo_bits())); 2546 scope_values->append(new ConstantIntValue(c->as_jint_hi_bits())); 2547 } 2548 #endif 2549 return 2; 2550 } 2551 2552 case T_ADDRESS: { 2553 #ifdef _LP64 2554 scope_values->append(new ConstantLongValue(c->as_jint())); 2555 #else 2556 scope_values->append(new ConstantIntValue(c->as_jint())); 2557 #endif 2558 return 1; 2559 } 2560 2561 default: 2562 ShouldNotReachHere(); 2563 return -1; 2564 } 2565 } 2566 2567 int LinearScan::append_scope_value_for_operand(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) { 2568 if (opr->is_single_stack()) { 2569 int stack_idx = opr->single_stack_ix(); 2570 bool is_oop = opr->is_oop_register(); 2571 int cache_idx = (stack_idx + LinearScan::nof_cpu_regs) * 2 + (is_oop ? 1 : 0); 2572 2573 ScopeValue* sv = _scope_value_cache.at(cache_idx); 2574 if (sv == NULL) { 2575 Location::Type loc_type = is_oop ? Location::oop : Location::normal; 2576 sv = location_for_name(stack_idx, loc_type); 2577 _scope_value_cache.at_put(cache_idx, sv); 2578 } 2579 2580 // check if cached value is correct 2581 DEBUG_ONLY(assert_equal(sv, location_for_name(stack_idx, is_oop ? Location::oop : Location::normal))); 2582 2583 scope_values->append(sv); 2584 return 1; 2585 2586 } else if (opr->is_single_cpu()) { 2587 bool is_oop = opr->is_oop_register(); 2588 int cache_idx = opr->cpu_regnr() * 2 + (is_oop ? 1 : 0); 2589 Location::Type int_loc_type = NOT_LP64(Location::normal) LP64_ONLY(Location::int_in_long); 2590 2591 ScopeValue* sv = _scope_value_cache.at(cache_idx); 2592 if (sv == NULL) { 2593 Location::Type loc_type = is_oop ? Location::oop : int_loc_type; 2594 VMReg rname = frame_map()->regname(opr); 2595 sv = new LocationValue(Location::new_reg_loc(loc_type, rname)); 2596 _scope_value_cache.at_put(cache_idx, sv); 2597 } 2598 2599 // check if cached value is correct 2600 DEBUG_ONLY(assert_equal(sv, new LocationValue(Location::new_reg_loc(is_oop ? Location::oop : int_loc_type, frame_map()->regname(opr))))); 2601 2602 scope_values->append(sv); 2603 return 1; 2604 2605 #ifdef X86 2606 } else if (opr->is_single_xmm()) { 2607 VMReg rname = opr->as_xmm_float_reg()->as_VMReg(); 2608 LocationValue* sv = new LocationValue(Location::new_reg_loc(Location::normal, rname)); 2609 2610 scope_values->append(sv); 2611 return 1; 2612 #endif 2613 2614 } else if (opr->is_single_fpu()) { 2615 #ifdef X86 2616 // the exact location of fpu stack values is only known 2617 // during fpu stack allocation, so the stack allocator object 2618 // must be present 2619 assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)"); 2620 assert(_fpu_stack_allocator != NULL, "must be present"); 2621 opr = _fpu_stack_allocator->to_fpu_stack(opr); 2622 #endif 2623 2624 Location::Type loc_type = float_saved_as_double ? Location::float_in_dbl : Location::normal; 2625 VMReg rname = frame_map()->fpu_regname(opr->fpu_regnr()); 2626 #ifndef __SOFTFP__ 2627 #ifndef VM_LITTLE_ENDIAN 2628 if (! float_saved_as_double) { 2629 // On big endian system, we may have an issue if float registers use only 2630 // the low half of the (same) double registers. 2631 // Both the float and the double could have the same regnr but would correspond 2632 // to two different addresses once saved. 2633 2634 // get next safely (no assertion checks) 2635 VMReg next = VMRegImpl::as_VMReg(1+rname->value()); 2636 if (next->is_reg() && 2637 (next->as_FloatRegister() == rname->as_FloatRegister())) { 2638 // the back-end does use the same numbering for the double and the float 2639 rname = next; // VMReg for the low bits, e.g. the real VMReg for the float 2640 } 2641 } 2642 #endif 2643 #endif 2644 LocationValue* sv = new LocationValue(Location::new_reg_loc(loc_type, rname)); 2645 2646 scope_values->append(sv); 2647 return 1; 2648 2649 } else { 2650 // double-size operands 2651 2652 ScopeValue* first; 2653 ScopeValue* second; 2654 2655 if (opr->is_double_stack()) { 2656 #ifdef _LP64 2657 Location loc1; 2658 Location::Type loc_type = opr->type() == T_LONG ? Location::lng : Location::dbl; 2659 if (!frame_map()->locations_for_slot(opr->double_stack_ix(), loc_type, &loc1, NULL)) { 2660 bailout("too large frame"); 2661 } 2662 // Does this reverse on x86 vs. sparc? 2663 first = new LocationValue(loc1); 2664 second = _int_0_scope_value; 2665 #else 2666 Location loc1, loc2; 2667 if (!frame_map()->locations_for_slot(opr->double_stack_ix(), Location::normal, &loc1, &loc2)) { 2668 bailout("too large frame"); 2669 } 2670 first = new LocationValue(loc1); 2671 second = new LocationValue(loc2); 2672 #endif // _LP64 2673 2674 } else if (opr->is_double_cpu()) { 2675 #ifdef _LP64 2676 VMReg rname_first = opr->as_register_lo()->as_VMReg(); 2677 first = new LocationValue(Location::new_reg_loc(Location::lng, rname_first)); 2678 second = _int_0_scope_value; 2679 #else 2680 VMReg rname_first = opr->as_register_lo()->as_VMReg(); 2681 VMReg rname_second = opr->as_register_hi()->as_VMReg(); 2682 2683 if (hi_word_offset_in_bytes < lo_word_offset_in_bytes) { 2684 // lo/hi and swapped relative to first and second, so swap them 2685 VMReg tmp = rname_first; 2686 rname_first = rname_second; 2687 rname_second = tmp; 2688 } 2689 2690 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); 2691 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); 2692 #endif //_LP64 2693 2694 2695 #ifdef X86 2696 } else if (opr->is_double_xmm()) { 2697 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation"); 2698 VMReg rname_first = opr->as_xmm_double_reg()->as_VMReg(); 2699 # ifdef _LP64 2700 first = new LocationValue(Location::new_reg_loc(Location::dbl, rname_first)); 2701 second = _int_0_scope_value; 2702 # else 2703 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); 2704 // %%% This is probably a waste but we'll keep things as they were for now 2705 if (true) { 2706 VMReg rname_second = rname_first->next(); 2707 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); 2708 } 2709 # endif 2710 #endif 2711 2712 } else if (opr->is_double_fpu()) { 2713 // On SPARC, fpu_regnrLo/fpu_regnrHi represents the two halves of 2714 // the double as float registers in the native ordering. On X86, 2715 // fpu_regnrLo is a FPU stack slot whose VMReg represents 2716 // the low-order word of the double and fpu_regnrLo + 1 is the 2717 // name for the other half. *first and *second must represent the 2718 // least and most significant words, respectively. 2719 2720 #ifdef X86 2721 // the exact location of fpu stack values is only known 2722 // during fpu stack allocation, so the stack allocator object 2723 // must be present 2724 assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)"); 2725 assert(_fpu_stack_allocator != NULL, "must be present"); 2726 opr = _fpu_stack_allocator->to_fpu_stack(opr); 2727 2728 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation (only fpu_regnrLo is used)"); 2729 #endif 2730 #ifdef SPARC 2731 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi() + 1, "assumed in calculation (only fpu_regnrHi is used)"); 2732 #endif 2733 #ifdef ARM32 2734 assert(opr->fpu_regnrHi() == opr->fpu_regnrLo() + 1, "assumed in calculation (only fpu_regnrLo is used)"); 2735 #endif 2736 #ifdef PPC 2737 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation (only fpu_regnrHi is used)"); 2738 #endif 2739 2740 #ifdef VM_LITTLE_ENDIAN 2741 VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrLo()); 2742 #else 2743 VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrHi()); 2744 #endif 2745 2746 #ifdef _LP64 2747 first = new LocationValue(Location::new_reg_loc(Location::dbl, rname_first)); 2748 second = _int_0_scope_value; 2749 #else 2750 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); 2751 // %%% This is probably a waste but we'll keep things as they were for now 2752 if (true) { 2753 VMReg rname_second = rname_first->next(); 2754 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); 2755 } 2756 #endif 2757 2758 } else { 2759 ShouldNotReachHere(); 2760 first = NULL; 2761 second = NULL; 2762 } 2763 2764 assert(first != NULL && second != NULL, "must be set"); 2765 // The convention the interpreter uses is that the second local 2766 // holds the first raw word of the native double representation. 2767 // This is actually reasonable, since locals and stack arrays 2768 // grow downwards in all implementations. 2769 // (If, on some machine, the interpreter's Java locals or stack 2770 // were to grow upwards, the embedded doubles would be word-swapped.) 2771 scope_values->append(second); 2772 scope_values->append(first); 2773 return 2; 2774 } 2775 } 2776 2777 2778 int LinearScan::append_scope_value(int op_id, Value value, GrowableArray<ScopeValue*>* scope_values) { 2779 if (value != NULL) { 2780 LIR_Opr opr = value->operand(); 2781 Constant* con = value->as_Constant(); 2782 2783 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)"); 2784 assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands"); 2785 2786 if (con != NULL && !con->is_pinned() && !opr->is_constant()) { 2787 // Unpinned constants may have a virtual operand for a part of the lifetime 2788 // or may be illegal when it was optimized away, 2789 // so always use a constant operand 2790 opr = LIR_OprFact::value_type(con->type()); 2791 } 2792 assert(opr->is_virtual() || opr->is_constant(), "other cases not allowed here"); 2793 2794 if (opr->is_virtual()) { 2795 LIR_OpVisitState::OprMode mode = LIR_OpVisitState::inputMode; 2796 2797 BlockBegin* block = block_of_op_with_id(op_id); 2798 if (block->number_of_sux() == 1 && op_id == block->last_lir_instruction_id()) { 2799 // generating debug information for the last instruction of a block. 2800 // if this instruction is a branch, spill moves are inserted before this branch 2801 // and so the wrong operand would be returned (spill moves at block boundaries are not 2802 // considered in the live ranges of intervals) 2803 // Solution: use the first op_id of the branch target block instead. 2804 if (block->lir()->instructions_list()->last()->as_OpBranch() != NULL) { 2805 if (block->live_out().at(opr->vreg_number())) { 2806 op_id = block->sux_at(0)->first_lir_instruction_id(); 2807 mode = LIR_OpVisitState::outputMode; 2808 } 2809 } 2810 } 2811 2812 // Get current location of operand 2813 // The operand must be live because debug information is considered when building the intervals 2814 // if the interval is not live, color_lir_opr will cause an assertion failure 2815 opr = color_lir_opr(opr, op_id, mode); 2816 assert(!has_call(op_id) || opr->is_stack() || !is_caller_save(reg_num(opr)), "can not have caller-save register operands at calls"); 2817 2818 // Append to ScopeValue array 2819 return append_scope_value_for_operand(opr, scope_values); 2820 2821 } else { 2822 assert(value->as_Constant() != NULL, "all other instructions have only virtual operands"); 2823 assert(opr->is_constant(), "operand must be constant"); 2824 2825 return append_scope_value_for_constant(opr, scope_values); 2826 } 2827 } else { 2828 // append a dummy value because real value not needed 2829 scope_values->append(_illegal_value); 2830 return 1; 2831 } 2832 } 2833 2834 2835 IRScopeDebugInfo* LinearScan::compute_debug_info_for_scope(int op_id, IRScope* cur_scope, ValueStack* cur_state, ValueStack* innermost_state) { 2836 IRScopeDebugInfo* caller_debug_info = NULL; 2837 2838 ValueStack* caller_state = cur_state->caller_state(); 2839 if (caller_state != NULL) { 2840 // process recursively to compute outermost scope first 2841 caller_debug_info = compute_debug_info_for_scope(op_id, cur_scope->caller(), caller_state, innermost_state); 2842 } 2843 2844 // initialize these to null. 2845 // If we don't need deopt info or there are no locals, expressions or monitors, 2846 // then these get recorded as no information and avoids the allocation of 0 length arrays. 2847 GrowableArray<ScopeValue*>* locals = NULL; 2848 GrowableArray<ScopeValue*>* expressions = NULL; 2849 GrowableArray<MonitorValue*>* monitors = NULL; 2850 2851 // describe local variable values 2852 int nof_locals = cur_state->locals_size(); 2853 if (nof_locals > 0) { 2854 locals = new GrowableArray<ScopeValue*>(nof_locals); 2855 2856 int pos = 0; 2857 while (pos < nof_locals) { 2858 assert(pos < cur_state->locals_size(), "why not?"); 2859 2860 Value local = cur_state->local_at(pos); 2861 pos += append_scope_value(op_id, local, locals); 2862 2863 assert(locals->length() == pos, "must match"); 2864 } 2865 assert(locals->length() == cur_scope->method()->max_locals(), "wrong number of locals"); 2866 assert(locals->length() == cur_state->locals_size(), "wrong number of locals"); 2867 } else if (cur_scope->method()->max_locals() > 0) { 2868 assert(cur_state->kind() == ValueStack::EmptyExceptionState, "should be"); 2869 nof_locals = cur_scope->method()->max_locals(); 2870 locals = new GrowableArray<ScopeValue*>(nof_locals); 2871 for(int i = 0; i < nof_locals; i++) { 2872 locals->append(_illegal_value); 2873 } 2874 } 2875 2876 // describe expression stack 2877 int nof_stack = cur_state->stack_size(); 2878 if (nof_stack > 0) { 2879 expressions = new GrowableArray<ScopeValue*>(nof_stack); 2880 2881 int pos = 0; 2882 while (pos < nof_stack) { 2883 Value expression = cur_state->stack_at_inc(pos); 2884 append_scope_value(op_id, expression, expressions); 2885 2886 assert(expressions->length() == pos, "must match"); 2887 } 2888 assert(expressions->length() == cur_state->stack_size(), "wrong number of stack entries"); 2889 } 2890 2891 // describe monitors 2892 int nof_locks = cur_state->locks_size(); 2893 if (nof_locks > 0) { 2894 int lock_offset = cur_state->caller_state() != NULL ? cur_state->caller_state()->total_locks_size() : 0; 2895 monitors = new GrowableArray<MonitorValue*>(nof_locks); 2896 for (int i = 0; i < nof_locks; i++) { 2897 monitors->append(location_for_monitor_index(lock_offset + i)); 2898 } 2899 } 2900 2901 return new IRScopeDebugInfo(cur_scope, cur_state->bci(), locals, expressions, monitors, caller_debug_info); 2902 } 2903 2904 2905 void LinearScan::compute_debug_info(CodeEmitInfo* info, int op_id) { 2906 TRACE_LINEAR_SCAN(3, tty->print_cr("creating debug information at op_id %d", op_id)); 2907 2908 IRScope* innermost_scope = info->scope(); 2909 ValueStack* innermost_state = info->stack(); 2910 2911 assert(innermost_scope != NULL && innermost_state != NULL, "why is it missing?"); 2912 2913 DEBUG_ONLY(check_stack_depth(info, innermost_state->stack_size())); 2914 2915 if (info->_scope_debug_info == NULL) { 2916 // compute debug information 2917 info->_scope_debug_info = compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state); 2918 } else { 2919 // debug information already set. Check that it is correct from the current point of view 2920 DEBUG_ONLY(assert_equal(info->_scope_debug_info, compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state))); 2921 } 2922 } 2923 2924 2925 void LinearScan::assign_reg_num(LIR_OpList* instructions, IntervalWalker* iw) { 2926 LIR_OpVisitState visitor; 2927 int num_inst = instructions->length(); 2928 bool has_dead = false; 2929 2930 for (int j = 0; j < num_inst; j++) { 2931 LIR_Op* op = instructions->at(j); 2932 if (op == NULL) { // this can happen when spill-moves are removed in eliminate_spill_moves 2933 has_dead = true; 2934 continue; 2935 } 2936 int op_id = op->id(); 2937 2938 // visit instruction to get list of operands 2939 visitor.visit(op); 2940 2941 // iterate all modes of the visitor and process all virtual operands 2942 for_each_visitor_mode(mode) { 2943 int n = visitor.opr_count(mode); 2944 for (int k = 0; k < n; k++) { 2945 LIR_Opr opr = visitor.opr_at(mode, k); 2946 if (opr->is_virtual_register()) { 2947 visitor.set_opr_at(mode, k, color_lir_opr(opr, op_id, mode)); 2948 } 2949 } 2950 } 2951 2952 if (visitor.info_count() > 0) { 2953 // exception handling 2954 if (compilation()->has_exception_handlers()) { 2955 XHandlers* xhandlers = visitor.all_xhandler(); 2956 int n = xhandlers->length(); 2957 for (int k = 0; k < n; k++) { 2958 XHandler* handler = xhandlers->handler_at(k); 2959 if (handler->entry_code() != NULL) { 2960 assign_reg_num(handler->entry_code()->instructions_list(), NULL); 2961 } 2962 } 2963 } else { 2964 assert(visitor.all_xhandler()->length() == 0, "missed exception handler"); 2965 } 2966 2967 // compute oop map 2968 assert(iw != NULL, "needed for compute_oop_map"); 2969 compute_oop_map(iw, visitor, op); 2970 2971 // compute debug information 2972 if (!use_fpu_stack_allocation()) { 2973 // compute debug information if fpu stack allocation is not needed. 2974 // when fpu stack allocation is needed, the debug information can not 2975 // be computed here because the exact location of fpu operands is not known 2976 // -> debug information is created inside the fpu stack allocator 2977 int n = visitor.info_count(); 2978 for (int k = 0; k < n; k++) { 2979 compute_debug_info(visitor.info_at(k), op_id); 2980 } 2981 } 2982 } 2983 2984 #ifdef ASSERT 2985 // make sure we haven't made the op invalid. 2986 op->verify(); 2987 #endif 2988 2989 // remove useless moves 2990 if (op->code() == lir_move) { 2991 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); 2992 LIR_Op1* move = (LIR_Op1*)op; 2993 LIR_Opr src = move->in_opr(); 2994 LIR_Opr dst = move->result_opr(); 2995 if (dst == src || 2996 !dst->is_pointer() && !src->is_pointer() && 2997 src->is_same_register(dst)) { 2998 instructions->at_put(j, NULL); 2999 has_dead = true; 3000 } 3001 } 3002 } 3003 3004 if (has_dead) { 3005 // iterate all instructions of the block and remove all null-values. 3006 int insert_point = 0; 3007 for (int j = 0; j < num_inst; j++) { 3008 LIR_Op* op = instructions->at(j); 3009 if (op != NULL) { 3010 if (insert_point != j) { 3011 instructions->at_put(insert_point, op); 3012 } 3013 insert_point++; 3014 } 3015 } 3016 instructions->truncate(insert_point); 3017 } 3018 } 3019 3020 void LinearScan::assign_reg_num() { 3021 TIME_LINEAR_SCAN(timer_assign_reg_num); 3022 3023 init_compute_debug_info(); 3024 IntervalWalker* iw = init_compute_oop_maps(); 3025 3026 int num_blocks = block_count(); 3027 for (int i = 0; i < num_blocks; i++) { 3028 BlockBegin* block = block_at(i); 3029 assign_reg_num(block->lir()->instructions_list(), iw); 3030 } 3031 } 3032 3033 3034 void LinearScan::do_linear_scan() { 3035 NOT_PRODUCT(_total_timer.begin_method()); 3036 3037 number_instructions(); 3038 3039 NOT_PRODUCT(print_lir(1, "Before Register Allocation")); 3040 3041 compute_local_live_sets(); 3042 compute_global_live_sets(); 3043 CHECK_BAILOUT(); 3044 3045 build_intervals(); 3046 CHECK_BAILOUT(); 3047 sort_intervals_before_allocation(); 3048 3049 NOT_PRODUCT(print_intervals("Before Register Allocation")); 3050 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_before_alloc)); 3051 3052 allocate_registers(); 3053 CHECK_BAILOUT(); 3054 3055 resolve_data_flow(); 3056 if (compilation()->has_exception_handlers()) { 3057 resolve_exception_handlers(); 3058 } 3059 // fill in number of spill slots into frame_map 3060 propagate_spill_slots(); 3061 CHECK_BAILOUT(); 3062 3063 NOT_PRODUCT(print_intervals("After Register Allocation")); 3064 NOT_PRODUCT(print_lir(2, "LIR after register allocation:")); 3065 3066 sort_intervals_after_allocation(); 3067 3068 DEBUG_ONLY(verify()); 3069 3070 eliminate_spill_moves(); 3071 assign_reg_num(); 3072 CHECK_BAILOUT(); 3073 3074 NOT_PRODUCT(print_lir(2, "LIR after assignment of register numbers:")); 3075 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_after_asign)); 3076 3077 { TIME_LINEAR_SCAN(timer_allocate_fpu_stack); 3078 3079 if (use_fpu_stack_allocation()) { 3080 allocate_fpu_stack(); // Only has effect on Intel 3081 NOT_PRODUCT(print_lir(2, "LIR after FPU stack allocation:")); 3082 } 3083 } 3084 3085 { TIME_LINEAR_SCAN(timer_optimize_lir); 3086 3087 EdgeMoveOptimizer::optimize(ir()->code()); 3088 ControlFlowOptimizer::optimize(ir()->code()); 3089 // check that cfg is still correct after optimizations 3090 ir()->verify(); 3091 } 3092 3093 NOT_PRODUCT(print_lir(1, "Before Code Generation", false)); 3094 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_final)); 3095 NOT_PRODUCT(_total_timer.end_method(this)); 3096 } 3097 3098 3099 // ********** Printing functions 3100 3101 #ifndef PRODUCT 3102 3103 void LinearScan::print_timers(double total) { 3104 _total_timer.print(total); 3105 } 3106 3107 void LinearScan::print_statistics() { 3108 _stat_before_alloc.print("before allocation"); 3109 _stat_after_asign.print("after assignment of register"); 3110 _stat_final.print("after optimization"); 3111 } 3112 3113 void LinearScan::print_bitmap(BitMap& b) { 3114 for (unsigned int i = 0; i < b.size(); i++) { 3115 if (b.at(i)) tty->print("%d ", i); 3116 } 3117 tty->cr(); 3118 } 3119 3120 void LinearScan::print_intervals(const char* label) { 3121 if (TraceLinearScanLevel >= 1) { 3122 int i; 3123 tty->cr(); 3124 tty->print_cr("%s", label); 3125 3126 for (i = 0; i < interval_count(); i++) { 3127 Interval* interval = interval_at(i); 3128 if (interval != NULL) { 3129 interval->print(); 3130 } 3131 } 3132 3133 tty->cr(); 3134 tty->print_cr("--- Basic Blocks ---"); 3135 for (i = 0; i < block_count(); i++) { 3136 BlockBegin* block = block_at(i); 3137 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()); 3138 } 3139 tty->cr(); 3140 tty->cr(); 3141 } 3142 3143 if (PrintCFGToFile) { 3144 CFGPrinter::print_intervals(&_intervals, label); 3145 } 3146 } 3147 3148 void LinearScan::print_lir(int level, const char* label, bool hir_valid) { 3149 if (TraceLinearScanLevel >= level) { 3150 tty->cr(); 3151 tty->print_cr("%s", label); 3152 print_LIR(ir()->linear_scan_order()); 3153 tty->cr(); 3154 } 3155 3156 if (level == 1 && PrintCFGToFile) { 3157 CFGPrinter::print_cfg(ir()->linear_scan_order(), label, hir_valid, true); 3158 } 3159 } 3160 3161 #endif //PRODUCT 3162 3163 3164 // ********** verification functions for allocation 3165 // (check that all intervals have a correct register and that no registers are overwritten) 3166 #ifdef ASSERT 3167 3168 void LinearScan::verify() { 3169 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying intervals ******************************************")); 3170 verify_intervals(); 3171 3172 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that no oops are in fixed intervals ****************")); 3173 verify_no_oops_in_fixed_intervals(); 3174 3175 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that unpinned constants are not alive across block boundaries")); 3176 verify_constants(); 3177 3178 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying register allocation ********************************")); 3179 verify_registers(); 3180 3181 TRACE_LINEAR_SCAN(2, tty->print_cr("********* no errors found **********************************************")); 3182 } 3183 3184 void LinearScan::verify_intervals() { 3185 int len = interval_count(); 3186 bool has_error = false; 3187 3188 for (int i = 0; i < len; i++) { 3189 Interval* i1 = interval_at(i); 3190 if (i1 == NULL) continue; 3191 3192 i1->check_split_children(); 3193 3194 if (i1->reg_num() != i) { 3195 tty->print_cr("Interval %d is on position %d in list", i1->reg_num(), i); i1->print(); tty->cr(); 3196 has_error = true; 3197 } 3198 3199 if (i1->reg_num() >= LIR_OprDesc::vreg_base && i1->type() == T_ILLEGAL) { 3200 tty->print_cr("Interval %d has no type assigned", i1->reg_num()); i1->print(); tty->cr(); 3201 has_error = true; 3202 } 3203 3204 if (i1->assigned_reg() == any_reg) { 3205 tty->print_cr("Interval %d has no register assigned", i1->reg_num()); i1->print(); tty->cr(); 3206 has_error = true; 3207 } 3208 3209 if (i1->assigned_reg() == i1->assigned_regHi()) { 3210 tty->print_cr("Interval %d: low and high register equal", i1->reg_num()); i1->print(); tty->cr(); 3211 has_error = true; 3212 } 3213 3214 if (!is_processed_reg_num(i1->assigned_reg())) { 3215 tty->print_cr("Can not have an Interval for an ignored register"); i1->print(); tty->cr(); 3216 has_error = true; 3217 } 3218 3219 if (i1->first() == Range::end()) { 3220 tty->print_cr("Interval %d has no Range", i1->reg_num()); i1->print(); tty->cr(); 3221 has_error = true; 3222 } 3223 3224 for (Range* r = i1->first(); r != Range::end(); r = r->next()) { 3225 if (r->from() >= r->to()) { 3226 tty->print_cr("Interval %d has zero length range", i1->reg_num()); i1->print(); tty->cr(); 3227 has_error = true; 3228 } 3229 } 3230 3231 for (int j = i + 1; j < len; j++) { 3232 Interval* i2 = interval_at(j); 3233 if (i2 == NULL) continue; 3234 3235 // special intervals that are created in MoveResolver 3236 // -> ignore them because the range information has no meaning there 3237 if (i1->from() == 1 && i1->to() == 2) continue; 3238 if (i2->from() == 1 && i2->to() == 2) continue; 3239 3240 int r1 = i1->assigned_reg(); 3241 int r1Hi = i1->assigned_regHi(); 3242 int r2 = i2->assigned_reg(); 3243 int r2Hi = i2->assigned_regHi(); 3244 if (i1->intersects(i2) && (r1 == r2 || r1 == r2Hi || (r1Hi != any_reg && (r1Hi == r2 || r1Hi == r2Hi)))) { 3245 tty->print_cr("Intervals %d and %d overlap and have the same register assigned", i1->reg_num(), i2->reg_num()); 3246 i1->print(); tty->cr(); 3247 i2->print(); tty->cr(); 3248 has_error = true; 3249 } 3250 } 3251 } 3252 3253 assert(has_error == false, "register allocation invalid"); 3254 } 3255 3256 3257 void LinearScan::verify_no_oops_in_fixed_intervals() { 3258 Interval* fixed_intervals; 3259 Interval* other_intervals; 3260 create_unhandled_lists(&fixed_intervals, &other_intervals, is_precolored_cpu_interval, NULL); 3261 3262 // to ensure a walking until the last instruction id, add a dummy interval 3263 // with a high operation id 3264 other_intervals = new Interval(any_reg); 3265 other_intervals->add_range(max_jint - 2, max_jint - 1); 3266 IntervalWalker* iw = new IntervalWalker(this, fixed_intervals, other_intervals); 3267 3268 LIR_OpVisitState visitor; 3269 for (int i = 0; i < block_count(); i++) { 3270 BlockBegin* block = block_at(i); 3271 3272 LIR_OpList* instructions = block->lir()->instructions_list(); 3273 3274 for (int j = 0; j < instructions->length(); j++) { 3275 LIR_Op* op = instructions->at(j); 3276 int op_id = op->id(); 3277 3278 visitor.visit(op); 3279 3280 if (visitor.info_count() > 0) { 3281 iw->walk_before(op->id()); 3282 bool check_live = true; 3283 if (op->code() == lir_move) { 3284 LIR_Op1* move = (LIR_Op1*)op; 3285 check_live = (move->patch_code() == lir_patch_none); 3286 } 3287 LIR_OpBranch* branch = op->as_OpBranch(); 3288 if (branch != NULL && branch->stub() != NULL && branch->stub()->is_exception_throw_stub()) { 3289 // Don't bother checking the stub in this case since the 3290 // exception stub will never return to normal control flow. 3291 check_live = false; 3292 } 3293 3294 // Make sure none of the fixed registers is live across an 3295 // oopmap since we can't handle that correctly. 3296 if (check_live) { 3297 for (Interval* interval = iw->active_first(fixedKind); 3298 interval != Interval::end(); 3299 interval = interval->next()) { 3300 if (interval->current_to() > op->id() + 1) { 3301 // This interval is live out of this op so make sure 3302 // that this interval represents some value that's 3303 // referenced by this op either as an input or output. 3304 bool ok = false; 3305 for_each_visitor_mode(mode) { 3306 int n = visitor.opr_count(mode); 3307 for (int k = 0; k < n; k++) { 3308 LIR_Opr opr = visitor.opr_at(mode, k); 3309 if (opr->is_fixed_cpu()) { 3310 if (interval_at(reg_num(opr)) == interval) { 3311 ok = true; 3312 break; 3313 } 3314 int hi = reg_numHi(opr); 3315 if (hi != -1 && interval_at(hi) == interval) { 3316 ok = true; 3317 break; 3318 } 3319 } 3320 } 3321 } 3322 assert(ok, "fixed intervals should never be live across an oopmap point"); 3323 } 3324 } 3325 } 3326 } 3327 3328 // oop-maps at calls do not contain registers, so check is not needed 3329 if (!visitor.has_call()) { 3330 3331 for_each_visitor_mode(mode) { 3332 int n = visitor.opr_count(mode); 3333 for (int k = 0; k < n; k++) { 3334 LIR_Opr opr = visitor.opr_at(mode, k); 3335 3336 if (opr->is_fixed_cpu() && opr->is_oop()) { 3337 // operand is a non-virtual cpu register and contains an oop 3338 TRACE_LINEAR_SCAN(4, op->print_on(tty); tty->print("checking operand "); opr->print(); tty->cr()); 3339 3340 Interval* interval = interval_at(reg_num(opr)); 3341 assert(interval != NULL, "no interval"); 3342 3343 if (mode == LIR_OpVisitState::inputMode) { 3344 if (interval->to() >= op_id + 1) { 3345 assert(interval->to() < op_id + 2 || 3346 interval->has_hole_between(op_id, op_id + 2), 3347 "oop input operand live after instruction"); 3348 } 3349 } else if (mode == LIR_OpVisitState::outputMode) { 3350 if (interval->from() <= op_id - 1) { 3351 assert(interval->has_hole_between(op_id - 1, op_id), 3352 "oop input operand live after instruction"); 3353 } 3354 } 3355 } 3356 } 3357 } 3358 } 3359 } 3360 } 3361 } 3362 3363 3364 void LinearScan::verify_constants() { 3365 int num_regs = num_virtual_regs(); 3366 int size = live_set_size(); 3367 int num_blocks = block_count(); 3368 3369 for (int i = 0; i < num_blocks; i++) { 3370 BlockBegin* block = block_at(i); 3371 BitMap live_at_edge = block->live_in(); 3372 3373 // visit all registers where the live_at_edge bit is set 3374 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)) { 3375 TRACE_LINEAR_SCAN(4, tty->print("checking interval %d of block B%d", r, block->block_id())); 3376 3377 Value value = gen()->instruction_for_vreg(r); 3378 3379 assert(value != NULL, "all intervals live across block boundaries must have Value"); 3380 assert(value->operand()->is_register() && value->operand()->is_virtual(), "value must have virtual operand"); 3381 assert(value->operand()->vreg_number() == r, "register number must match"); 3382 // TKR assert(value->as_Constant() == NULL || value->is_pinned(), "only pinned constants can be alive accross block boundaries"); 3383 } 3384 } 3385 } 3386 3387 3388 class RegisterVerifier: public StackObj { 3389 private: 3390 LinearScan* _allocator; 3391 BlockList _work_list; // all blocks that must be processed 3392 IntervalsList _saved_states; // saved information of previous check 3393 3394 // simplified access to methods of LinearScan 3395 Compilation* compilation() const { return _allocator->compilation(); } 3396 Interval* interval_at(int reg_num) const { return _allocator->interval_at(reg_num); } 3397 int reg_num(LIR_Opr opr) const { return _allocator->reg_num(opr); } 3398 3399 // currently, only registers are processed 3400 int state_size() { return LinearScan::nof_regs; } 3401 3402 // accessors 3403 IntervalList* state_for_block(BlockBegin* block) { return _saved_states.at(block->block_id()); } 3404 void set_state_for_block(BlockBegin* block, IntervalList* saved_state) { _saved_states.at_put(block->block_id(), saved_state); } 3405 void add_to_work_list(BlockBegin* block) { if (!_work_list.contains(block)) _work_list.append(block); } 3406 3407 // helper functions 3408 IntervalList* copy(IntervalList* input_state); 3409 void state_put(IntervalList* input_state, int reg, Interval* interval); 3410 bool check_state(IntervalList* input_state, int reg, Interval* interval); 3411 3412 void process_block(BlockBegin* block); 3413 void process_xhandler(XHandler* xhandler, IntervalList* input_state); 3414 void process_successor(BlockBegin* block, IntervalList* input_state); 3415 void process_operations(LIR_List* ops, IntervalList* input_state); 3416 3417 public: 3418 RegisterVerifier(LinearScan* allocator) 3419 : _allocator(allocator) 3420 , _work_list(16) 3421 , _saved_states(BlockBegin::number_of_blocks(), NULL) 3422 { } 3423 3424 void verify(BlockBegin* start); 3425 }; 3426 3427 3428 // entry function from LinearScan that starts the verification 3429 void LinearScan::verify_registers() { 3430 RegisterVerifier verifier(this); 3431 verifier.verify(block_at(0)); 3432 } 3433 3434 3435 void RegisterVerifier::verify(BlockBegin* start) { 3436 // setup input registers (method arguments) for first block 3437 IntervalList* input_state = new IntervalList(state_size(), NULL); 3438 CallingConvention* args = compilation()->frame_map()->incoming_arguments(); 3439 for (int n = 0; n < args->length(); n++) { 3440 LIR_Opr opr = args->at(n); 3441 if (opr->is_register()) { 3442 Interval* interval = interval_at(reg_num(opr)); 3443 3444 if (interval->assigned_reg() < state_size()) { 3445 input_state->at_put(interval->assigned_reg(), interval); 3446 } 3447 if (interval->assigned_regHi() != LinearScan::any_reg && interval->assigned_regHi() < state_size()) { 3448 input_state->at_put(interval->assigned_regHi(), interval); 3449 } 3450 } 3451 } 3452 3453 set_state_for_block(start, input_state); 3454 add_to_work_list(start); 3455 3456 // main loop for verification 3457 do { 3458 BlockBegin* block = _work_list.at(0); 3459 _work_list.remove_at(0); 3460 3461 process_block(block); 3462 } while (!_work_list.is_empty()); 3463 } 3464 3465 void RegisterVerifier::process_block(BlockBegin* block) { 3466 TRACE_LINEAR_SCAN(2, tty->cr(); tty->print_cr("process_block B%d", block->block_id())); 3467 3468 // must copy state because it is modified 3469 IntervalList* input_state = copy(state_for_block(block)); 3470 3471 if (TraceLinearScanLevel >= 4) { 3472 tty->print_cr("Input-State of intervals:"); 3473 tty->print(" "); 3474 for (int i = 0; i < state_size(); i++) { 3475 if (input_state->at(i) != NULL) { 3476 tty->print(" %4d", input_state->at(i)->reg_num()); 3477 } else { 3478 tty->print(" __"); 3479 } 3480 } 3481 tty->cr(); 3482 tty->cr(); 3483 } 3484 3485 // process all operations of the block 3486 process_operations(block->lir(), input_state); 3487 3488 // iterate all successors 3489 for (int i = 0; i < block->number_of_sux(); i++) { 3490 process_successor(block->sux_at(i), input_state); 3491 } 3492 } 3493 3494 void RegisterVerifier::process_xhandler(XHandler* xhandler, IntervalList* input_state) { 3495 TRACE_LINEAR_SCAN(2, tty->print_cr("process_xhandler B%d", xhandler->entry_block()->block_id())); 3496 3497 // must copy state because it is modified 3498 input_state = copy(input_state); 3499 3500 if (xhandler->entry_code() != NULL) { 3501 process_operations(xhandler->entry_code(), input_state); 3502 } 3503 process_successor(xhandler->entry_block(), input_state); 3504 } 3505 3506 void RegisterVerifier::process_successor(BlockBegin* block, IntervalList* input_state) { 3507 IntervalList* saved_state = state_for_block(block); 3508 3509 if (saved_state != NULL) { 3510 // this block was already processed before. 3511 // check if new input_state is consistent with saved_state 3512 3513 bool saved_state_correct = true; 3514 for (int i = 0; i < state_size(); i++) { 3515 if (input_state->at(i) != saved_state->at(i)) { 3516 // current input_state and previous saved_state assume a different 3517 // interval in this register -> assume that this register is invalid 3518 if (saved_state->at(i) != NULL) { 3519 // invalidate old calculation only if it assumed that 3520 // register was valid. when the register was already invalid, 3521 // then the old calculation was correct. 3522 saved_state_correct = false; 3523 saved_state->at_put(i, NULL); 3524 3525 TRACE_LINEAR_SCAN(4, tty->print_cr("process_successor B%d: invalidating slot %d", block->block_id(), i)); 3526 } 3527 } 3528 } 3529 3530 if (saved_state_correct) { 3531 // already processed block with correct input_state 3532 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: previous visit already correct", block->block_id())); 3533 } else { 3534 // must re-visit this block 3535 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: must re-visit because input state changed", block->block_id())); 3536 add_to_work_list(block); 3537 } 3538 3539 } else { 3540 // block was not processed before, so set initial input_state 3541 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: initial visit", block->block_id())); 3542 3543 set_state_for_block(block, copy(input_state)); 3544 add_to_work_list(block); 3545 } 3546 } 3547 3548 3549 IntervalList* RegisterVerifier::copy(IntervalList* input_state) { 3550 IntervalList* copy_state = new IntervalList(input_state->length()); 3551 copy_state->push_all(input_state); 3552 return copy_state; 3553 } 3554 3555 void RegisterVerifier::state_put(IntervalList* input_state, int reg, Interval* interval) { 3556 if (reg != LinearScan::any_reg && reg < state_size()) { 3557 if (interval != NULL) { 3558 TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = %d", reg, interval->reg_num())); 3559 } else if (input_state->at(reg) != NULL) { 3560 TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = NULL", reg)); 3561 } 3562 3563 input_state->at_put(reg, interval); 3564 } 3565 } 3566 3567 bool RegisterVerifier::check_state(IntervalList* input_state, int reg, Interval* interval) { 3568 if (reg != LinearScan::any_reg && reg < state_size()) { 3569 if (input_state->at(reg) != interval) { 3570 tty->print_cr("!! Error in register allocation: register %d does not contain interval %d", reg, interval->reg_num()); 3571 return true; 3572 } 3573 } 3574 return false; 3575 } 3576 3577 void RegisterVerifier::process_operations(LIR_List* ops, IntervalList* input_state) { 3578 // visit all instructions of the block 3579 LIR_OpVisitState visitor; 3580 bool has_error = false; 3581 3582 for (int i = 0; i < ops->length(); i++) { 3583 LIR_Op* op = ops->at(i); 3584 visitor.visit(op); 3585 3586 TRACE_LINEAR_SCAN(4, op->print_on(tty)); 3587 3588 // check if input operands are correct 3589 int j; 3590 int n = visitor.opr_count(LIR_OpVisitState::inputMode); 3591 for (j = 0; j < n; j++) { 3592 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, j); 3593 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { 3594 Interval* interval = interval_at(reg_num(opr)); 3595 if (op->id() != -1) { 3596 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::inputMode); 3597 } 3598 3599 has_error |= check_state(input_state, interval->assigned_reg(), interval->split_parent()); 3600 has_error |= check_state(input_state, interval->assigned_regHi(), interval->split_parent()); 3601 3602 // When an operand is marked with is_last_use, then the fpu stack allocator 3603 // removes the register from the fpu stack -> the register contains no value 3604 if (opr->is_last_use()) { 3605 state_put(input_state, interval->assigned_reg(), NULL); 3606 state_put(input_state, interval->assigned_regHi(), NULL); 3607 } 3608 } 3609 } 3610 3611 // invalidate all caller save registers at calls 3612 if (visitor.has_call()) { 3613 for (j = 0; j < FrameMap::nof_caller_save_cpu_regs(); j++) { 3614 state_put(input_state, reg_num(FrameMap::caller_save_cpu_reg_at(j)), NULL); 3615 } 3616 for (j = 0; j < FrameMap::nof_caller_save_fpu_regs; j++) { 3617 state_put(input_state, reg_num(FrameMap::caller_save_fpu_reg_at(j)), NULL); 3618 } 3619 3620 #ifdef X86 3621 int num_caller_save_xmm_regs = FrameMap::nof_caller_save_xmm_regs; 3622 #if _LP64 3623 if (UseAVX < 3) { 3624 num_caller_save_xmm_regs = num_caller_save_xmm_regs / 2; 3625 } 3626 #endif 3627 for (j = 0; j < num_caller_save_xmm_regs; j++) { 3628 state_put(input_state, reg_num(FrameMap::caller_save_xmm_reg_at(j)), NULL); 3629 } 3630 #endif 3631 } 3632 3633 // process xhandler before output and temp operands 3634 XHandlers* xhandlers = visitor.all_xhandler(); 3635 n = xhandlers->length(); 3636 for (int k = 0; k < n; k++) { 3637 process_xhandler(xhandlers->handler_at(k), input_state); 3638 } 3639 3640 // set temp operands (some operations use temp operands also as output operands, so can't set them NULL) 3641 n = visitor.opr_count(LIR_OpVisitState::tempMode); 3642 for (j = 0; j < n; j++) { 3643 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, j); 3644 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { 3645 Interval* interval = interval_at(reg_num(opr)); 3646 if (op->id() != -1) { 3647 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::tempMode); 3648 } 3649 3650 state_put(input_state, interval->assigned_reg(), interval->split_parent()); 3651 state_put(input_state, interval->assigned_regHi(), interval->split_parent()); 3652 } 3653 } 3654 3655 // set output operands 3656 n = visitor.opr_count(LIR_OpVisitState::outputMode); 3657 for (j = 0; j < n; j++) { 3658 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, j); 3659 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { 3660 Interval* interval = interval_at(reg_num(opr)); 3661 if (op->id() != -1) { 3662 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::outputMode); 3663 } 3664 3665 state_put(input_state, interval->assigned_reg(), interval->split_parent()); 3666 state_put(input_state, interval->assigned_regHi(), interval->split_parent()); 3667 } 3668 } 3669 } 3670 assert(has_error == false, "Error in register allocation"); 3671 } 3672 3673 #endif // ASSERT 3674 3675 3676 3677 // **** Implementation of MoveResolver ****************************** 3678 3679 MoveResolver::MoveResolver(LinearScan* allocator) : 3680 _allocator(allocator), 3681 _multiple_reads_allowed(false), 3682 _mapping_from(8), 3683 _mapping_from_opr(8), 3684 _mapping_to(8), 3685 _insert_list(NULL), 3686 _insert_idx(-1), 3687 _insertion_buffer() 3688 { 3689 for (int i = 0; i < LinearScan::nof_regs; i++) { 3690 _register_blocked[i] = 0; 3691 } 3692 DEBUG_ONLY(check_empty()); 3693 } 3694 3695 3696 #ifdef ASSERT 3697 3698 void MoveResolver::check_empty() { 3699 assert(_mapping_from.length() == 0 && _mapping_from_opr.length() == 0 && _mapping_to.length() == 0, "list must be empty before and after processing"); 3700 for (int i = 0; i < LinearScan::nof_regs; i++) { 3701 assert(register_blocked(i) == 0, "register map must be empty before and after processing"); 3702 } 3703 assert(_multiple_reads_allowed == false, "must have default value"); 3704 } 3705 3706 void MoveResolver::verify_before_resolve() { 3707 assert(_mapping_from.length() == _mapping_from_opr.length(), "length must be equal"); 3708 assert(_mapping_from.length() == _mapping_to.length(), "length must be equal"); 3709 assert(_insert_list != NULL && _insert_idx != -1, "insert position not set"); 3710 3711 int i, j; 3712 if (!_multiple_reads_allowed) { 3713 for (i = 0; i < _mapping_from.length(); i++) { 3714 for (j = i + 1; j < _mapping_from.length(); j++) { 3715 assert(_mapping_from.at(i) == NULL || _mapping_from.at(i) != _mapping_from.at(j), "cannot read from same interval twice"); 3716 } 3717 } 3718 } 3719 3720 for (i = 0; i < _mapping_to.length(); i++) { 3721 for (j = i + 1; j < _mapping_to.length(); j++) { 3722 assert(_mapping_to.at(i) != _mapping_to.at(j), "cannot write to same interval twice"); 3723 } 3724 } 3725 3726 3727 BitMap used_regs(LinearScan::nof_regs + allocator()->frame_map()->argcount() + allocator()->max_spills()); 3728 used_regs.clear(); 3729 if (!_multiple_reads_allowed) { 3730 for (i = 0; i < _mapping_from.length(); i++) { 3731 Interval* it = _mapping_from.at(i); 3732 if (it != NULL) { 3733 assert(!used_regs.at(it->assigned_reg()), "cannot read from same register twice"); 3734 used_regs.set_bit(it->assigned_reg()); 3735 3736 if (it->assigned_regHi() != LinearScan::any_reg) { 3737 assert(!used_regs.at(it->assigned_regHi()), "cannot read from same register twice"); 3738 used_regs.set_bit(it->assigned_regHi()); 3739 } 3740 } 3741 } 3742 } 3743 3744 used_regs.clear(); 3745 for (i = 0; i < _mapping_to.length(); i++) { 3746 Interval* it = _mapping_to.at(i); 3747 assert(!used_regs.at(it->assigned_reg()), "cannot write to same register twice"); 3748 used_regs.set_bit(it->assigned_reg()); 3749 3750 if (it->assigned_regHi() != LinearScan::any_reg) { 3751 assert(!used_regs.at(it->assigned_regHi()), "cannot write to same register twice"); 3752 used_regs.set_bit(it->assigned_regHi()); 3753 } 3754 } 3755 3756 used_regs.clear(); 3757 for (i = 0; i < _mapping_from.length(); i++) { 3758 Interval* it = _mapping_from.at(i); 3759 if (it != NULL && it->assigned_reg() >= LinearScan::nof_regs) { 3760 used_regs.set_bit(it->assigned_reg()); 3761 } 3762 } 3763 for (i = 0; i < _mapping_to.length(); i++) { 3764 Interval* it = _mapping_to.at(i); 3765 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"); 3766 } 3767 } 3768 3769 #endif // ASSERT 3770 3771 3772 // mark assigned_reg and assigned_regHi of the interval as blocked 3773 void MoveResolver::block_registers(Interval* it) { 3774 int reg = it->assigned_reg(); 3775 if (reg < LinearScan::nof_regs) { 3776 assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used"); 3777 set_register_blocked(reg, 1); 3778 } 3779 reg = it->assigned_regHi(); 3780 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { 3781 assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used"); 3782 set_register_blocked(reg, 1); 3783 } 3784 } 3785 3786 // mark assigned_reg and assigned_regHi of the interval as unblocked 3787 void MoveResolver::unblock_registers(Interval* it) { 3788 int reg = it->assigned_reg(); 3789 if (reg < LinearScan::nof_regs) { 3790 assert(register_blocked(reg) > 0, "register already marked as unused"); 3791 set_register_blocked(reg, -1); 3792 } 3793 reg = it->assigned_regHi(); 3794 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { 3795 assert(register_blocked(reg) > 0, "register already marked as unused"); 3796 set_register_blocked(reg, -1); 3797 } 3798 } 3799 3800 // check if assigned_reg and assigned_regHi of the to-interval are not blocked (or only blocked by from) 3801 bool MoveResolver::save_to_process_move(Interval* from, Interval* to) { 3802 int from_reg = -1; 3803 int from_regHi = -1; 3804 if (from != NULL) { 3805 from_reg = from->assigned_reg(); 3806 from_regHi = from->assigned_regHi(); 3807 } 3808 3809 int reg = to->assigned_reg(); 3810 if (reg < LinearScan::nof_regs) { 3811 if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) { 3812 return false; 3813 } 3814 } 3815 reg = to->assigned_regHi(); 3816 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { 3817 if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) { 3818 return false; 3819 } 3820 } 3821 3822 return true; 3823 } 3824 3825 3826 void MoveResolver::create_insertion_buffer(LIR_List* list) { 3827 assert(!_insertion_buffer.initialized(), "overwriting existing buffer"); 3828 _insertion_buffer.init(list); 3829 } 3830 3831 void MoveResolver::append_insertion_buffer() { 3832 if (_insertion_buffer.initialized()) { 3833 _insertion_buffer.lir_list()->append(&_insertion_buffer); 3834 } 3835 assert(!_insertion_buffer.initialized(), "must be uninitialized now"); 3836 3837 _insert_list = NULL; 3838 _insert_idx = -1; 3839 } 3840 3841 void MoveResolver::insert_move(Interval* from_interval, Interval* to_interval) { 3842 assert(from_interval->reg_num() != to_interval->reg_num(), "from and to interval equal"); 3843 assert(from_interval->type() == to_interval->type(), "move between different types"); 3844 assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first"); 3845 assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer"); 3846 3847 LIR_Opr from_opr = LIR_OprFact::virtual_register(from_interval->reg_num(), from_interval->type()); 3848 LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type()); 3849 3850 if (!_multiple_reads_allowed) { 3851 // the last_use flag is an optimization for FPU stack allocation. When the same 3852 // input interval is used in more than one move, then it is too difficult to determine 3853 // if this move is really the last use. 3854 from_opr = from_opr->make_last_use(); 3855 } 3856 _insertion_buffer.move(_insert_idx, from_opr, to_opr); 3857 3858 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())); 3859 } 3860 3861 void MoveResolver::insert_move(LIR_Opr from_opr, Interval* to_interval) { 3862 assert(from_opr->type() == to_interval->type(), "move between different types"); 3863 assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first"); 3864 assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer"); 3865 3866 LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type()); 3867 _insertion_buffer.move(_insert_idx, from_opr, to_opr); 3868 3869 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())); 3870 } 3871 3872 3873 void MoveResolver::resolve_mappings() { 3874 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)); 3875 DEBUG_ONLY(verify_before_resolve()); 3876 3877 // Block all registers that are used as input operands of a move. 3878 // When a register is blocked, no move to this register is emitted. 3879 // This is necessary for detecting cycles in moves. 3880 int i; 3881 for (i = _mapping_from.length() - 1; i >= 0; i--) { 3882 Interval* from_interval = _mapping_from.at(i); 3883 if (from_interval != NULL) { 3884 block_registers(from_interval); 3885 } 3886 } 3887 3888 int spill_candidate = -1; 3889 while (_mapping_from.length() > 0) { 3890 bool processed_interval = false; 3891 3892 for (i = _mapping_from.length() - 1; i >= 0; i--) { 3893 Interval* from_interval = _mapping_from.at(i); 3894 Interval* to_interval = _mapping_to.at(i); 3895 3896 if (save_to_process_move(from_interval, to_interval)) { 3897 // this inverval can be processed because target is free 3898 if (from_interval != NULL) { 3899 insert_move(from_interval, to_interval); 3900 unblock_registers(from_interval); 3901 } else { 3902 insert_move(_mapping_from_opr.at(i), to_interval); 3903 } 3904 _mapping_from.remove_at(i); 3905 _mapping_from_opr.remove_at(i); 3906 _mapping_to.remove_at(i); 3907 3908 processed_interval = true; 3909 } else if (from_interval != NULL && from_interval->assigned_reg() < LinearScan::nof_regs) { 3910 // this interval cannot be processed now because target is not free 3911 // it starts in a register, so it is a possible candidate for spilling 3912 spill_candidate = i; 3913 } 3914 } 3915 3916 if (!processed_interval) { 3917 // no move could be processed because there is a cycle in the move list 3918 // (e.g. r1 -> r2, r2 -> r1), so one interval must be spilled to memory 3919 assert(spill_candidate != -1, "no interval in register for spilling found"); 3920 3921 // create a new spill interval and assign a stack slot to it 3922 Interval* from_interval = _mapping_from.at(spill_candidate); 3923 Interval* spill_interval = new Interval(-1); 3924 spill_interval->set_type(from_interval->type()); 3925 3926 // add a dummy range because real position is difficult to calculate 3927 // Note: this range is a special case when the integrity of the allocation is checked 3928 spill_interval->add_range(1, 2); 3929 3930 // do not allocate a new spill slot for temporary interval, but 3931 // use spill slot assigned to from_interval. Otherwise moves from 3932 // one stack slot to another can happen (not allowed by LIR_Assembler 3933 int spill_slot = from_interval->canonical_spill_slot(); 3934 if (spill_slot < 0) { 3935 spill_slot = allocator()->allocate_spill_slot(type2spill_size[spill_interval->type()] == 2); 3936 from_interval->set_canonical_spill_slot(spill_slot); 3937 } 3938 spill_interval->assign_reg(spill_slot); 3939 allocator()->append_interval(spill_interval); 3940 3941 TRACE_LINEAR_SCAN(4, tty->print_cr("created new Interval %d for spilling", spill_interval->reg_num())); 3942 3943 // insert a move from register to stack and update the mapping 3944 insert_move(from_interval, spill_interval); 3945 _mapping_from.at_put(spill_candidate, spill_interval); 3946 unblock_registers(from_interval); 3947 } 3948 } 3949 3950 // reset to default value 3951 _multiple_reads_allowed = false; 3952 3953 // check that all intervals have been processed 3954 DEBUG_ONLY(check_empty()); 3955 } 3956 3957 3958 void MoveResolver::set_insert_position(LIR_List* insert_list, int insert_idx) { 3959 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)); 3960 assert(_insert_list == NULL && _insert_idx == -1, "use move_insert_position instead of set_insert_position when data already set"); 3961 3962 create_insertion_buffer(insert_list); 3963 _insert_list = insert_list; 3964 _insert_idx = insert_idx; 3965 } 3966 3967 void MoveResolver::move_insert_position(LIR_List* insert_list, int insert_idx) { 3968 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)); 3969 3970 if (_insert_list != NULL && (insert_list != _insert_list || insert_idx != _insert_idx)) { 3971 // insert position changed -> resolve current mappings 3972 resolve_mappings(); 3973 } 3974 3975 if (insert_list != _insert_list) { 3976 // block changed -> append insertion_buffer because it is 3977 // bound to a specific block and create a new insertion_buffer 3978 append_insertion_buffer(); 3979 create_insertion_buffer(insert_list); 3980 } 3981 3982 _insert_list = insert_list; 3983 _insert_idx = insert_idx; 3984 } 3985 3986 void MoveResolver::add_mapping(Interval* from_interval, Interval* to_interval) { 3987 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())); 3988 3989 _mapping_from.append(from_interval); 3990 _mapping_from_opr.append(LIR_OprFact::illegalOpr); 3991 _mapping_to.append(to_interval); 3992 } 3993 3994 3995 void MoveResolver::add_mapping(LIR_Opr from_opr, Interval* to_interval) { 3996 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())); 3997 assert(from_opr->is_constant(), "only for constants"); 3998 3999 _mapping_from.append(NULL); 4000 _mapping_from_opr.append(from_opr); 4001 _mapping_to.append(to_interval); 4002 } 4003 4004 void MoveResolver::resolve_and_append_moves() { 4005 if (has_mappings()) { 4006 resolve_mappings(); 4007 } 4008 append_insertion_buffer(); 4009 } 4010 4011 4012 4013 // **** Implementation of Range ************************************* 4014 4015 Range::Range(int from, int to, Range* next) : 4016 _from(from), 4017 _to(to), 4018 _next(next) 4019 { 4020 } 4021 4022 // initialize sentinel 4023 Range* Range::_end = NULL; 4024 void Range::initialize(Arena* arena) { 4025 _end = new (arena) Range(max_jint, max_jint, NULL); 4026 } 4027 4028 int Range::intersects_at(Range* r2) const { 4029 const Range* r1 = this; 4030 4031 assert(r1 != NULL && r2 != NULL, "null ranges not allowed"); 4032 assert(r1 != _end && r2 != _end, "empty ranges not allowed"); 4033 4034 do { 4035 if (r1->from() < r2->from()) { 4036 if (r1->to() <= r2->from()) { 4037 r1 = r1->next(); if (r1 == _end) return -1; 4038 } else { 4039 return r2->from(); 4040 } 4041 } else if (r2->from() < r1->from()) { 4042 if (r2->to() <= r1->from()) { 4043 r2 = r2->next(); if (r2 == _end) return -1; 4044 } else { 4045 return r1->from(); 4046 } 4047 } else { // r1->from() == r2->from() 4048 if (r1->from() == r1->to()) { 4049 r1 = r1->next(); if (r1 == _end) return -1; 4050 } else if (r2->from() == r2->to()) { 4051 r2 = r2->next(); if (r2 == _end) return -1; 4052 } else { 4053 return r1->from(); 4054 } 4055 } 4056 } while (true); 4057 } 4058 4059 #ifndef PRODUCT 4060 void Range::print(outputStream* out) const { 4061 out->print("[%d, %d[ ", _from, _to); 4062 } 4063 #endif 4064 4065 4066 4067 // **** Implementation of Interval ********************************** 4068 4069 // initialize sentinel 4070 Interval* Interval::_end = NULL; 4071 void Interval::initialize(Arena* arena) { 4072 Range::initialize(arena); 4073 _end = new (arena) Interval(-1); 4074 } 4075 4076 Interval::Interval(int reg_num) : 4077 _reg_num(reg_num), 4078 _type(T_ILLEGAL), 4079 _first(Range::end()), 4080 _use_pos_and_kinds(12), 4081 _current(Range::end()), 4082 _next(_end), 4083 _state(invalidState), 4084 _assigned_reg(LinearScan::any_reg), 4085 _assigned_regHi(LinearScan::any_reg), 4086 _cached_to(-1), 4087 _cached_opr(LIR_OprFact::illegalOpr), 4088 _cached_vm_reg(VMRegImpl::Bad()), 4089 _split_children(0), 4090 _canonical_spill_slot(-1), 4091 _insert_move_when_activated(false), 4092 _register_hint(NULL), 4093 _spill_state(noDefinitionFound), 4094 _spill_definition_pos(-1) 4095 { 4096 _split_parent = this; 4097 _current_split_child = this; 4098 } 4099 4100 int Interval::calc_to() { 4101 assert(_first != Range::end(), "interval has no range"); 4102 4103 Range* r = _first; 4104 while (r->next() != Range::end()) { 4105 r = r->next(); 4106 } 4107 return r->to(); 4108 } 4109 4110 4111 #ifdef ASSERT 4112 // consistency check of split-children 4113 void Interval::check_split_children() { 4114 if (_split_children.length() > 0) { 4115 assert(is_split_parent(), "only split parents can have children"); 4116 4117 for (int i = 0; i < _split_children.length(); i++) { 4118 Interval* i1 = _split_children.at(i); 4119 4120 assert(i1->split_parent() == this, "not a split child of this interval"); 4121 assert(i1->type() == type(), "must be equal for all split children"); 4122 assert(i1->canonical_spill_slot() == canonical_spill_slot(), "must be equal for all split children"); 4123 4124 for (int j = i + 1; j < _split_children.length(); j++) { 4125 Interval* i2 = _split_children.at(j); 4126 4127 assert(i1->reg_num() != i2->reg_num(), "same register number"); 4128 4129 if (i1->from() < i2->from()) { 4130 assert(i1->to() <= i2->from() && i1->to() < i2->to(), "intervals overlapping"); 4131 } else { 4132 assert(i2->from() < i1->from(), "intervals start at same op_id"); 4133 assert(i2->to() <= i1->from() && i2->to() < i1->to(), "intervals overlapping"); 4134 } 4135 } 4136 } 4137 } 4138 } 4139 #endif // ASSERT 4140 4141 Interval* Interval::register_hint(bool search_split_child) const { 4142 if (!search_split_child) { 4143 return _register_hint; 4144 } 4145 4146 if (_register_hint != NULL) { 4147 assert(_register_hint->is_split_parent(), "ony split parents are valid hint registers"); 4148 4149 if (_register_hint->assigned_reg() >= 0 && _register_hint->assigned_reg() < LinearScan::nof_regs) { 4150 return _register_hint; 4151 4152 } else if (_register_hint->_split_children.length() > 0) { 4153 // search the first split child that has a register assigned 4154 int len = _register_hint->_split_children.length(); 4155 for (int i = 0; i < len; i++) { 4156 Interval* cur = _register_hint->_split_children.at(i); 4157 4158 if (cur->assigned_reg() >= 0 && cur->assigned_reg() < LinearScan::nof_regs) { 4159 return cur; 4160 } 4161 } 4162 } 4163 } 4164 4165 // no hint interval found that has a register assigned 4166 return NULL; 4167 } 4168 4169 4170 Interval* Interval::split_child_at_op_id(int op_id, LIR_OpVisitState::OprMode mode) { 4171 assert(is_split_parent(), "can only be called for split parents"); 4172 assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)"); 4173 4174 Interval* result; 4175 if (_split_children.length() == 0) { 4176 result = this; 4177 } else { 4178 result = NULL; 4179 int len = _split_children.length(); 4180 4181 // in outputMode, the end of the interval (op_id == cur->to()) is not valid 4182 int to_offset = (mode == LIR_OpVisitState::outputMode ? 0 : 1); 4183 4184 int i; 4185 for (i = 0; i < len; i++) { 4186 Interval* cur = _split_children.at(i); 4187 if (cur->from() <= op_id && op_id < cur->to() + to_offset) { 4188 if (i > 0) { 4189 // exchange current split child to start of list (faster access for next call) 4190 _split_children.at_put(i, _split_children.at(0)); 4191 _split_children.at_put(0, cur); 4192 } 4193 4194 // interval found 4195 result = cur; 4196 break; 4197 } 4198 } 4199 4200 #ifdef ASSERT 4201 for (i = 0; i < len; i++) { 4202 Interval* tmp = _split_children.at(i); 4203 if (tmp != result && tmp->from() <= op_id && op_id < tmp->to() + to_offset) { 4204 tty->print_cr("two valid result intervals found for op_id %d: %d and %d", op_id, result->reg_num(), tmp->reg_num()); 4205 result->print(); 4206 tmp->print(); 4207 assert(false, "two valid result intervals found"); 4208 } 4209 } 4210 #endif 4211 } 4212 4213 assert(result != NULL, "no matching interval found"); 4214 assert(result->covers(op_id, mode), "op_id not covered by interval"); 4215 4216 return result; 4217 } 4218 4219 4220 // returns the last split child that ends before the given op_id 4221 Interval* Interval::split_child_before_op_id(int op_id) { 4222 assert(op_id >= 0, "invalid op_id"); 4223 4224 Interval* parent = split_parent(); 4225 Interval* result = NULL; 4226 4227 int len = parent->_split_children.length(); 4228 assert(len > 0, "no split children available"); 4229 4230 for (int i = len - 1; i >= 0; i--) { 4231 Interval* cur = parent->_split_children.at(i); 4232 if (cur->to() <= op_id && (result == NULL || result->to() < cur->to())) { 4233 result = cur; 4234 } 4235 } 4236 4237 assert(result != NULL, "no split child found"); 4238 return result; 4239 } 4240 4241 4242 // checks if op_id is covered by any split child 4243 bool Interval::split_child_covers(int op_id, LIR_OpVisitState::OprMode mode) { 4244 assert(is_split_parent(), "can only be called for split parents"); 4245 assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)"); 4246 4247 if (_split_children.length() == 0) { 4248 // simple case if interval was not split 4249 return covers(op_id, mode); 4250 4251 } else { 4252 // extended case: check all split children 4253 int len = _split_children.length(); 4254 for (int i = 0; i < len; i++) { 4255 Interval* cur = _split_children.at(i); 4256 if (cur->covers(op_id, mode)) { 4257 return true; 4258 } 4259 } 4260 return false; 4261 } 4262 } 4263 4264 4265 // Note: use positions are sorted descending -> first use has highest index 4266 int Interval::first_usage(IntervalUseKind min_use_kind) const { 4267 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4268 4269 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4270 if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) { 4271 return _use_pos_and_kinds.at(i); 4272 } 4273 } 4274 return max_jint; 4275 } 4276 4277 int Interval::next_usage(IntervalUseKind min_use_kind, int from) const { 4278 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4279 4280 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4281 if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) >= min_use_kind) { 4282 return _use_pos_and_kinds.at(i); 4283 } 4284 } 4285 return max_jint; 4286 } 4287 4288 int Interval::next_usage_exact(IntervalUseKind exact_use_kind, int from) const { 4289 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4290 4291 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4292 if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) == exact_use_kind) { 4293 return _use_pos_and_kinds.at(i); 4294 } 4295 } 4296 return max_jint; 4297 } 4298 4299 int Interval::previous_usage(IntervalUseKind min_use_kind, int from) const { 4300 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4301 4302 int prev = 0; 4303 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4304 if (_use_pos_and_kinds.at(i) > from) { 4305 return prev; 4306 } 4307 if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) { 4308 prev = _use_pos_and_kinds.at(i); 4309 } 4310 } 4311 return prev; 4312 } 4313 4314 void Interval::add_use_pos(int pos, IntervalUseKind use_kind) { 4315 assert(covers(pos, LIR_OpVisitState::inputMode), "use position not covered by live range"); 4316 4317 // do not add use positions for precolored intervals because 4318 // they are never used 4319 if (use_kind != noUse && reg_num() >= LIR_OprDesc::vreg_base) { 4320 #ifdef ASSERT 4321 assert(_use_pos_and_kinds.length() % 2 == 0, "must be"); 4322 for (int i = 0; i < _use_pos_and_kinds.length(); i += 2) { 4323 assert(pos <= _use_pos_and_kinds.at(i), "already added a use-position with lower position"); 4324 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4325 if (i > 0) { 4326 assert(_use_pos_and_kinds.at(i) < _use_pos_and_kinds.at(i - 2), "not sorted descending"); 4327 } 4328 } 4329 #endif 4330 4331 // Note: add_use is called in descending order, so list gets sorted 4332 // automatically by just appending new use positions 4333 int len = _use_pos_and_kinds.length(); 4334 if (len == 0 || _use_pos_and_kinds.at(len - 2) > pos) { 4335 _use_pos_and_kinds.append(pos); 4336 _use_pos_and_kinds.append(use_kind); 4337 } else if (_use_pos_and_kinds.at(len - 1) < use_kind) { 4338 assert(_use_pos_and_kinds.at(len - 2) == pos, "list not sorted correctly"); 4339 _use_pos_and_kinds.at_put(len - 1, use_kind); 4340 } 4341 } 4342 } 4343 4344 void Interval::add_range(int from, int to) { 4345 assert(from < to, "invalid range"); 4346 assert(first() == Range::end() || to < first()->next()->from(), "not inserting at begin of interval"); 4347 assert(from <= first()->to(), "not inserting at begin of interval"); 4348 4349 if (first()->from() <= to) { 4350 // join intersecting ranges 4351 first()->set_from(MIN2(from, first()->from())); 4352 first()->set_to (MAX2(to, first()->to())); 4353 } else { 4354 // insert new range 4355 _first = new Range(from, to, first()); 4356 } 4357 } 4358 4359 Interval* Interval::new_split_child() { 4360 // allocate new interval 4361 Interval* result = new Interval(-1); 4362 result->set_type(type()); 4363 4364 Interval* parent = split_parent(); 4365 result->_split_parent = parent; 4366 result->set_register_hint(parent); 4367 4368 // insert new interval in children-list of parent 4369 if (parent->_split_children.length() == 0) { 4370 assert(is_split_parent(), "list must be initialized at first split"); 4371 4372 parent->_split_children = IntervalList(4); 4373 parent->_split_children.append(this); 4374 } 4375 parent->_split_children.append(result); 4376 4377 return result; 4378 } 4379 4380 // split this interval at the specified position and return 4381 // the remainder as a new interval. 4382 // 4383 // when an interval is split, a bi-directional link is established between the original interval 4384 // (the split parent) and the intervals that are split off this interval (the split children) 4385 // When a split child is split again, the new created interval is also a direct child 4386 // of the original parent (there is no tree of split children stored, but a flat list) 4387 // All split children are spilled to the same stack slot (stored in _canonical_spill_slot) 4388 // 4389 // Note: The new interval has no valid reg_num 4390 Interval* Interval::split(int split_pos) { 4391 assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals"); 4392 4393 // allocate new interval 4394 Interval* result = new_split_child(); 4395 4396 // split the ranges 4397 Range* prev = NULL; 4398 Range* cur = _first; 4399 while (cur != Range::end() && cur->to() <= split_pos) { 4400 prev = cur; 4401 cur = cur->next(); 4402 } 4403 assert(cur != Range::end(), "split interval after end of last range"); 4404 4405 if (cur->from() < split_pos) { 4406 result->_first = new Range(split_pos, cur->to(), cur->next()); 4407 cur->set_to(split_pos); 4408 cur->set_next(Range::end()); 4409 4410 } else { 4411 assert(prev != NULL, "split before start of first range"); 4412 result->_first = cur; 4413 prev->set_next(Range::end()); 4414 } 4415 result->_current = result->_first; 4416 _cached_to = -1; // clear cached value 4417 4418 // split list of use positions 4419 int total_len = _use_pos_and_kinds.length(); 4420 int start_idx = total_len - 2; 4421 while (start_idx >= 0 && _use_pos_and_kinds.at(start_idx) < split_pos) { 4422 start_idx -= 2; 4423 } 4424 4425 intStack new_use_pos_and_kinds(total_len - start_idx); 4426 int i; 4427 for (i = start_idx + 2; i < total_len; i++) { 4428 new_use_pos_and_kinds.append(_use_pos_and_kinds.at(i)); 4429 } 4430 4431 _use_pos_and_kinds.truncate(start_idx + 2); 4432 result->_use_pos_and_kinds = _use_pos_and_kinds; 4433 _use_pos_and_kinds = new_use_pos_and_kinds; 4434 4435 #ifdef ASSERT 4436 assert(_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos"); 4437 assert(result->_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos"); 4438 assert(_use_pos_and_kinds.length() + result->_use_pos_and_kinds.length() == total_len, "missed some entries"); 4439 4440 for (i = 0; i < _use_pos_and_kinds.length(); i += 2) { 4441 assert(_use_pos_and_kinds.at(i) < split_pos, "must be"); 4442 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4443 } 4444 for (i = 0; i < result->_use_pos_and_kinds.length(); i += 2) { 4445 assert(result->_use_pos_and_kinds.at(i) >= split_pos, "must be"); 4446 assert(result->_use_pos_and_kinds.at(i + 1) >= firstValidKind && result->_use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4447 } 4448 #endif 4449 4450 return result; 4451 } 4452 4453 // split this interval at the specified position and return 4454 // the head as a new interval (the original interval is the tail) 4455 // 4456 // Currently, only the first range can be split, and the new interval 4457 // must not have split positions 4458 Interval* Interval::split_from_start(int split_pos) { 4459 assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals"); 4460 assert(split_pos > from() && split_pos < to(), "can only split inside interval"); 4461 assert(split_pos > _first->from() && split_pos <= _first->to(), "can only split inside first range"); 4462 assert(first_usage(noUse) > split_pos, "can not split when use positions are present"); 4463 4464 // allocate new interval 4465 Interval* result = new_split_child(); 4466 4467 // the new created interval has only one range (checked by assertion above), 4468 // so the splitting of the ranges is very simple 4469 result->add_range(_first->from(), split_pos); 4470 4471 if (split_pos == _first->to()) { 4472 assert(_first->next() != Range::end(), "must not be at end"); 4473 _first = _first->next(); 4474 } else { 4475 _first->set_from(split_pos); 4476 } 4477 4478 return result; 4479 } 4480 4481 4482 // returns true if the op_id is inside the interval 4483 bool Interval::covers(int op_id, LIR_OpVisitState::OprMode mode) const { 4484 Range* cur = _first; 4485 4486 while (cur != Range::end() && cur->to() < op_id) { 4487 cur = cur->next(); 4488 } 4489 if (cur != Range::end()) { 4490 assert(cur->to() != cur->next()->from(), "ranges not separated"); 4491 4492 if (mode == LIR_OpVisitState::outputMode) { 4493 return cur->from() <= op_id && op_id < cur->to(); 4494 } else { 4495 return cur->from() <= op_id && op_id <= cur->to(); 4496 } 4497 } 4498 return false; 4499 } 4500 4501 // returns true if the interval has any hole between hole_from and hole_to 4502 // (even if the hole has only the length 1) 4503 bool Interval::has_hole_between(int hole_from, int hole_to) { 4504 assert(hole_from < hole_to, "check"); 4505 assert(from() <= hole_from && hole_to <= to(), "index out of interval"); 4506 4507 Range* cur = _first; 4508 while (cur != Range::end()) { 4509 assert(cur->to() < cur->next()->from(), "no space between ranges"); 4510 4511 // hole-range starts before this range -> hole 4512 if (hole_from < cur->from()) { 4513 return true; 4514 4515 // hole-range completely inside this range -> no hole 4516 } else if (hole_to <= cur->to()) { 4517 return false; 4518 4519 // overlapping of hole-range with this range -> hole 4520 } else if (hole_from <= cur->to()) { 4521 return true; 4522 } 4523 4524 cur = cur->next(); 4525 } 4526 4527 return false; 4528 } 4529 4530 4531 #ifndef PRODUCT 4532 void Interval::print(outputStream* out) const { 4533 const char* SpillState2Name[] = { "no definition", "no spill store", "one spill store", "store at definition", "start in memory", "no optimization" }; 4534 const char* UseKind2Name[] = { "N", "L", "S", "M" }; 4535 4536 const char* type_name; 4537 LIR_Opr opr = LIR_OprFact::illegal(); 4538 if (reg_num() < LIR_OprDesc::vreg_base) { 4539 type_name = "fixed"; 4540 // need a temporary operand for fixed intervals because type() cannot be called 4541 #ifdef X86 4542 int last_xmm_reg = pd_last_xmm_reg; 4543 #ifdef _LP64 4544 if (UseAVX < 3) { 4545 last_xmm_reg = pd_first_xmm_reg + (pd_nof_xmm_regs_frame_map / 2) - 1; 4546 } 4547 #endif 4548 #endif 4549 if (assigned_reg() >= pd_first_cpu_reg && assigned_reg() <= pd_last_cpu_reg) { 4550 opr = LIR_OprFact::single_cpu(assigned_reg()); 4551 } else if (assigned_reg() >= pd_first_fpu_reg && assigned_reg() <= pd_last_fpu_reg) { 4552 opr = LIR_OprFact::single_fpu(assigned_reg() - pd_first_fpu_reg); 4553 #ifdef X86 4554 } else if (assigned_reg() >= pd_first_xmm_reg && assigned_reg() <= last_xmm_reg) { 4555 opr = LIR_OprFact::single_xmm(assigned_reg() - pd_first_xmm_reg); 4556 #endif 4557 } else { 4558 ShouldNotReachHere(); 4559 } 4560 } else { 4561 type_name = type2name(type()); 4562 if (assigned_reg() != -1 && 4563 (LinearScan::num_physical_regs(type()) == 1 || assigned_regHi() != -1)) { 4564 opr = LinearScan::calc_operand_for_interval(this); 4565 } 4566 } 4567 4568 out->print("%d %s ", reg_num(), type_name); 4569 if (opr->is_valid()) { 4570 out->print("\""); 4571 opr->print(out); 4572 out->print("\" "); 4573 } 4574 out->print("%d %d ", split_parent()->reg_num(), (register_hint(false) != NULL ? register_hint(false)->reg_num() : -1)); 4575 4576 // print ranges 4577 Range* cur = _first; 4578 while (cur != Range::end()) { 4579 cur->print(out); 4580 cur = cur->next(); 4581 assert(cur != NULL, "range list not closed with range sentinel"); 4582 } 4583 4584 // print use positions 4585 int prev = 0; 4586 assert(_use_pos_and_kinds.length() % 2 == 0, "must be"); 4587 for (int i =_use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4588 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4589 assert(prev < _use_pos_and_kinds.at(i), "use positions not sorted"); 4590 4591 out->print("%d %s ", _use_pos_and_kinds.at(i), UseKind2Name[_use_pos_and_kinds.at(i + 1)]); 4592 prev = _use_pos_and_kinds.at(i); 4593 } 4594 4595 out->print(" \"%s\"", SpillState2Name[spill_state()]); 4596 out->cr(); 4597 } 4598 #endif 4599 4600 4601 4602 // **** Implementation of IntervalWalker **************************** 4603 4604 IntervalWalker::IntervalWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first) 4605 : _compilation(allocator->compilation()) 4606 , _allocator(allocator) 4607 { 4608 _unhandled_first[fixedKind] = unhandled_fixed_first; 4609 _unhandled_first[anyKind] = unhandled_any_first; 4610 _active_first[fixedKind] = Interval::end(); 4611 _inactive_first[fixedKind] = Interval::end(); 4612 _active_first[anyKind] = Interval::end(); 4613 _inactive_first[anyKind] = Interval::end(); 4614 _current_position = -1; 4615 _current = NULL; 4616 next_interval(); 4617 } 4618 4619 4620 // append interval at top of list 4621 void IntervalWalker::append_unsorted(Interval** list, Interval* interval) { 4622 interval->set_next(*list); *list = interval; 4623 } 4624 4625 4626 // append interval in order of current range from() 4627 void IntervalWalker::append_sorted(Interval** list, Interval* interval) { 4628 Interval* prev = NULL; 4629 Interval* cur = *list; 4630 while (cur->current_from() < interval->current_from()) { 4631 prev = cur; cur = cur->next(); 4632 } 4633 if (prev == NULL) { 4634 *list = interval; 4635 } else { 4636 prev->set_next(interval); 4637 } 4638 interval->set_next(cur); 4639 } 4640 4641 void IntervalWalker::append_to_unhandled(Interval** list, Interval* interval) { 4642 assert(interval->from() >= current()->current_from(), "cannot append new interval before current walk position"); 4643 4644 Interval* prev = NULL; 4645 Interval* cur = *list; 4646 while (cur->from() < interval->from() || (cur->from() == interval->from() && cur->first_usage(noUse) < interval->first_usage(noUse))) { 4647 prev = cur; cur = cur->next(); 4648 } 4649 if (prev == NULL) { 4650 *list = interval; 4651 } else { 4652 prev->set_next(interval); 4653 } 4654 interval->set_next(cur); 4655 } 4656 4657 4658 inline bool IntervalWalker::remove_from_list(Interval** list, Interval* i) { 4659 while (*list != Interval::end() && *list != i) { 4660 list = (*list)->next_addr(); 4661 } 4662 if (*list != Interval::end()) { 4663 assert(*list == i, "check"); 4664 *list = (*list)->next(); 4665 return true; 4666 } else { 4667 return false; 4668 } 4669 } 4670 4671 void IntervalWalker::remove_from_list(Interval* i) { 4672 bool deleted; 4673 4674 if (i->state() == activeState) { 4675 deleted = remove_from_list(active_first_addr(anyKind), i); 4676 } else { 4677 assert(i->state() == inactiveState, "invalid state"); 4678 deleted = remove_from_list(inactive_first_addr(anyKind), i); 4679 } 4680 4681 assert(deleted, "interval has not been found in list"); 4682 } 4683 4684 4685 void IntervalWalker::walk_to(IntervalState state, int from) { 4686 assert (state == activeState || state == inactiveState, "wrong state"); 4687 for_each_interval_kind(kind) { 4688 Interval** prev = state == activeState ? active_first_addr(kind) : inactive_first_addr(kind); 4689 Interval* next = *prev; 4690 while (next->current_from() <= from) { 4691 Interval* cur = next; 4692 next = cur->next(); 4693 4694 bool range_has_changed = false; 4695 while (cur->current_to() <= from) { 4696 cur->next_range(); 4697 range_has_changed = true; 4698 } 4699 4700 // also handle move from inactive list to active list 4701 range_has_changed = range_has_changed || (state == inactiveState && cur->current_from() <= from); 4702 4703 if (range_has_changed) { 4704 // remove cur from list 4705 *prev = next; 4706 if (cur->current_at_end()) { 4707 // move to handled state (not maintained as a list) 4708 cur->set_state(handledState); 4709 interval_moved(cur, kind, state, handledState); 4710 } else if (cur->current_from() <= from){ 4711 // sort into active list 4712 append_sorted(active_first_addr(kind), cur); 4713 cur->set_state(activeState); 4714 if (*prev == cur) { 4715 assert(state == activeState, "check"); 4716 prev = cur->next_addr(); 4717 } 4718 interval_moved(cur, kind, state, activeState); 4719 } else { 4720 // sort into inactive list 4721 append_sorted(inactive_first_addr(kind), cur); 4722 cur->set_state(inactiveState); 4723 if (*prev == cur) { 4724 assert(state == inactiveState, "check"); 4725 prev = cur->next_addr(); 4726 } 4727 interval_moved(cur, kind, state, inactiveState); 4728 } 4729 } else { 4730 prev = cur->next_addr(); 4731 continue; 4732 } 4733 } 4734 } 4735 } 4736 4737 4738 void IntervalWalker::next_interval() { 4739 IntervalKind kind; 4740 Interval* any = _unhandled_first[anyKind]; 4741 Interval* fixed = _unhandled_first[fixedKind]; 4742 4743 if (any != Interval::end()) { 4744 // intervals may start at same position -> prefer fixed interval 4745 kind = fixed != Interval::end() && fixed->from() <= any->from() ? fixedKind : anyKind; 4746 4747 assert (kind == fixedKind && fixed->from() <= any->from() || 4748 kind == anyKind && any->from() <= fixed->from(), "wrong interval!!!"); 4749 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"); 4750 4751 } else if (fixed != Interval::end()) { 4752 kind = fixedKind; 4753 } else { 4754 _current = NULL; return; 4755 } 4756 _current_kind = kind; 4757 _current = _unhandled_first[kind]; 4758 _unhandled_first[kind] = _current->next(); 4759 _current->set_next(Interval::end()); 4760 _current->rewind_range(); 4761 } 4762 4763 4764 void IntervalWalker::walk_to(int lir_op_id) { 4765 assert(_current_position <= lir_op_id, "can not walk backwards"); 4766 while (current() != NULL) { 4767 bool is_active = current()->from() <= lir_op_id; 4768 int id = is_active ? current()->from() : lir_op_id; 4769 4770 TRACE_LINEAR_SCAN(2, if (_current_position < id) { tty->cr(); tty->print_cr("walk_to(%d) **************************************************************", id); }) 4771 4772 // set _current_position prior to call of walk_to 4773 _current_position = id; 4774 4775 // call walk_to even if _current_position == id 4776 walk_to(activeState, id); 4777 walk_to(inactiveState, id); 4778 4779 if (is_active) { 4780 current()->set_state(activeState); 4781 if (activate_current()) { 4782 append_sorted(active_first_addr(current_kind()), current()); 4783 interval_moved(current(), current_kind(), unhandledState, activeState); 4784 } 4785 4786 next_interval(); 4787 } else { 4788 return; 4789 } 4790 } 4791 } 4792 4793 void IntervalWalker::interval_moved(Interval* interval, IntervalKind kind, IntervalState from, IntervalState to) { 4794 #ifndef PRODUCT 4795 if (TraceLinearScanLevel >= 4) { 4796 #define print_state(state) \ 4797 switch(state) {\ 4798 case unhandledState: tty->print("unhandled"); break;\ 4799 case activeState: tty->print("active"); break;\ 4800 case inactiveState: tty->print("inactive"); break;\ 4801 case handledState: tty->print("handled"); break;\ 4802 default: ShouldNotReachHere(); \ 4803 } 4804 4805 print_state(from); tty->print(" to "); print_state(to); 4806 tty->fill_to(23); 4807 interval->print(); 4808 4809 #undef print_state 4810 } 4811 #endif 4812 } 4813 4814 4815 4816 // **** Implementation of LinearScanWalker ************************** 4817 4818 LinearScanWalker::LinearScanWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first) 4819 : IntervalWalker(allocator, unhandled_fixed_first, unhandled_any_first) 4820 , _move_resolver(allocator) 4821 { 4822 for (int i = 0; i < LinearScan::nof_regs; i++) { 4823 _spill_intervals[i] = new IntervalList(2); 4824 } 4825 } 4826 4827 4828 inline void LinearScanWalker::init_use_lists(bool only_process_use_pos) { 4829 for (int i = _first_reg; i <= _last_reg; i++) { 4830 _use_pos[i] = max_jint; 4831 4832 if (!only_process_use_pos) { 4833 _block_pos[i] = max_jint; 4834 _spill_intervals[i]->clear(); 4835 } 4836 } 4837 } 4838 4839 inline void LinearScanWalker::exclude_from_use(int reg) { 4840 assert(reg < LinearScan::nof_regs, "interval must have a register assigned (stack slots not allowed)"); 4841 if (reg >= _first_reg && reg <= _last_reg) { 4842 _use_pos[reg] = 0; 4843 } 4844 } 4845 inline void LinearScanWalker::exclude_from_use(Interval* i) { 4846 assert(i->assigned_reg() != any_reg, "interval has no register assigned"); 4847 4848 exclude_from_use(i->assigned_reg()); 4849 exclude_from_use(i->assigned_regHi()); 4850 } 4851 4852 inline void LinearScanWalker::set_use_pos(int reg, Interval* i, int use_pos, bool only_process_use_pos) { 4853 assert(use_pos != 0, "must use exclude_from_use to set use_pos to 0"); 4854 4855 if (reg >= _first_reg && reg <= _last_reg) { 4856 if (_use_pos[reg] > use_pos) { 4857 _use_pos[reg] = use_pos; 4858 } 4859 if (!only_process_use_pos) { 4860 _spill_intervals[reg]->append(i); 4861 } 4862 } 4863 } 4864 inline void LinearScanWalker::set_use_pos(Interval* i, int use_pos, bool only_process_use_pos) { 4865 assert(i->assigned_reg() != any_reg, "interval has no register assigned"); 4866 if (use_pos != -1) { 4867 set_use_pos(i->assigned_reg(), i, use_pos, only_process_use_pos); 4868 set_use_pos(i->assigned_regHi(), i, use_pos, only_process_use_pos); 4869 } 4870 } 4871 4872 inline void LinearScanWalker::set_block_pos(int reg, Interval* i, int block_pos) { 4873 if (reg >= _first_reg && reg <= _last_reg) { 4874 if (_block_pos[reg] > block_pos) { 4875 _block_pos[reg] = block_pos; 4876 } 4877 if (_use_pos[reg] > block_pos) { 4878 _use_pos[reg] = block_pos; 4879 } 4880 } 4881 } 4882 inline void LinearScanWalker::set_block_pos(Interval* i, int block_pos) { 4883 assert(i->assigned_reg() != any_reg, "interval has no register assigned"); 4884 if (block_pos != -1) { 4885 set_block_pos(i->assigned_reg(), i, block_pos); 4886 set_block_pos(i->assigned_regHi(), i, block_pos); 4887 } 4888 } 4889 4890 4891 void LinearScanWalker::free_exclude_active_fixed() { 4892 Interval* list = active_first(fixedKind); 4893 while (list != Interval::end()) { 4894 assert(list->assigned_reg() < LinearScan::nof_regs, "active interval must have a register assigned"); 4895 exclude_from_use(list); 4896 list = list->next(); 4897 } 4898 } 4899 4900 void LinearScanWalker::free_exclude_active_any() { 4901 Interval* list = active_first(anyKind); 4902 while (list != Interval::end()) { 4903 exclude_from_use(list); 4904 list = list->next(); 4905 } 4906 } 4907 4908 void LinearScanWalker::free_collect_inactive_fixed(Interval* cur) { 4909 Interval* list = inactive_first(fixedKind); 4910 while (list != Interval::end()) { 4911 if (cur->to() <= list->current_from()) { 4912 assert(list->current_intersects_at(cur) == -1, "must not intersect"); 4913 set_use_pos(list, list->current_from(), true); 4914 } else { 4915 set_use_pos(list, list->current_intersects_at(cur), true); 4916 } 4917 list = list->next(); 4918 } 4919 } 4920 4921 void LinearScanWalker::free_collect_inactive_any(Interval* cur) { 4922 Interval* list = inactive_first(anyKind); 4923 while (list != Interval::end()) { 4924 set_use_pos(list, list->current_intersects_at(cur), true); 4925 list = list->next(); 4926 } 4927 } 4928 4929 void LinearScanWalker::free_collect_unhandled(IntervalKind kind, Interval* cur) { 4930 Interval* list = unhandled_first(kind); 4931 while (list != Interval::end()) { 4932 set_use_pos(list, list->intersects_at(cur), true); 4933 if (kind == fixedKind && cur->to() <= list->from()) { 4934 set_use_pos(list, list->from(), true); 4935 } 4936 list = list->next(); 4937 } 4938 } 4939 4940 void LinearScanWalker::spill_exclude_active_fixed() { 4941 Interval* list = active_first(fixedKind); 4942 while (list != Interval::end()) { 4943 exclude_from_use(list); 4944 list = list->next(); 4945 } 4946 } 4947 4948 void LinearScanWalker::spill_block_unhandled_fixed(Interval* cur) { 4949 Interval* list = unhandled_first(fixedKind); 4950 while (list != Interval::end()) { 4951 set_block_pos(list, list->intersects_at(cur)); 4952 list = list->next(); 4953 } 4954 } 4955 4956 void LinearScanWalker::spill_block_inactive_fixed(Interval* cur) { 4957 Interval* list = inactive_first(fixedKind); 4958 while (list != Interval::end()) { 4959 if (cur->to() > list->current_from()) { 4960 set_block_pos(list, list->current_intersects_at(cur)); 4961 } else { 4962 assert(list->current_intersects_at(cur) == -1, "invalid optimization: intervals intersect"); 4963 } 4964 4965 list = list->next(); 4966 } 4967 } 4968 4969 void LinearScanWalker::spill_collect_active_any() { 4970 Interval* list = active_first(anyKind); 4971 while (list != Interval::end()) { 4972 set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false); 4973 list = list->next(); 4974 } 4975 } 4976 4977 void LinearScanWalker::spill_collect_inactive_any(Interval* cur) { 4978 Interval* list = inactive_first(anyKind); 4979 while (list != Interval::end()) { 4980 if (list->current_intersects(cur)) { 4981 set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false); 4982 } 4983 list = list->next(); 4984 } 4985 } 4986 4987 4988 void LinearScanWalker::insert_move(int op_id, Interval* src_it, Interval* dst_it) { 4989 // output all moves here. When source and target are equal, the move is 4990 // optimized away later in assign_reg_nums 4991 4992 op_id = (op_id + 1) & ~1; 4993 BlockBegin* op_block = allocator()->block_of_op_with_id(op_id); 4994 assert(op_id > 0 && allocator()->block_of_op_with_id(op_id - 2) == op_block, "cannot insert move at block boundary"); 4995 4996 // calculate index of instruction inside instruction list of current block 4997 // the minimal index (for a block with no spill moves) can be calculated because the 4998 // numbering of instructions is known. 4999 // When the block already contains spill moves, the index must be increased until the 5000 // correct index is reached. 5001 LIR_OpList* list = op_block->lir()->instructions_list(); 5002 int index = (op_id - list->at(0)->id()) / 2; 5003 assert(list->at(index)->id() <= op_id, "error in calculation"); 5004 5005 while (list->at(index)->id() != op_id) { 5006 index++; 5007 assert(0 <= index && index < list->length(), "index out of bounds"); 5008 } 5009 assert(1 <= index && index < list->length(), "index out of bounds"); 5010 assert(list->at(index)->id() == op_id, "error in calculation"); 5011 5012 // insert new instruction before instruction at position index 5013 _move_resolver.move_insert_position(op_block->lir(), index - 1); 5014 _move_resolver.add_mapping(src_it, dst_it); 5015 } 5016 5017 5018 int LinearScanWalker::find_optimal_split_pos(BlockBegin* min_block, BlockBegin* max_block, int max_split_pos) { 5019 int from_block_nr = min_block->linear_scan_number(); 5020 int to_block_nr = max_block->linear_scan_number(); 5021 5022 assert(0 <= from_block_nr && from_block_nr < block_count(), "out of range"); 5023 assert(0 <= to_block_nr && to_block_nr < block_count(), "out of range"); 5024 assert(from_block_nr < to_block_nr, "must cross block boundary"); 5025 5026 // Try to split at end of max_block. If this would be after 5027 // max_split_pos, then use the begin of max_block 5028 int optimal_split_pos = max_block->last_lir_instruction_id() + 2; 5029 if (optimal_split_pos > max_split_pos) { 5030 optimal_split_pos = max_block->first_lir_instruction_id(); 5031 } 5032 5033 int min_loop_depth = max_block->loop_depth(); 5034 for (int i = to_block_nr - 1; i >= from_block_nr; i--) { 5035 BlockBegin* cur = block_at(i); 5036 5037 if (cur->loop_depth() < min_loop_depth) { 5038 // block with lower loop-depth found -> split at the end of this block 5039 min_loop_depth = cur->loop_depth(); 5040 optimal_split_pos = cur->last_lir_instruction_id() + 2; 5041 } 5042 } 5043 assert(optimal_split_pos > allocator()->max_lir_op_id() || allocator()->is_block_begin(optimal_split_pos), "algorithm must move split pos to block boundary"); 5044 5045 return optimal_split_pos; 5046 } 5047 5048 5049 int LinearScanWalker::find_optimal_split_pos(Interval* it, int min_split_pos, int max_split_pos, bool do_loop_optimization) { 5050 int optimal_split_pos = -1; 5051 if (min_split_pos == max_split_pos) { 5052 // trivial case, no optimization of split position possible 5053 TRACE_LINEAR_SCAN(4, tty->print_cr(" min-pos and max-pos are equal, no optimization possible")); 5054 optimal_split_pos = min_split_pos; 5055 5056 } else { 5057 assert(min_split_pos < max_split_pos, "must be true then"); 5058 assert(min_split_pos > 0, "cannot access min_split_pos - 1 otherwise"); 5059 5060 // reason for using min_split_pos - 1: when the minimal split pos is exactly at the 5061 // beginning of a block, then min_split_pos is also a possible split position. 5062 // Use the block before as min_block, because then min_block->last_lir_instruction_id() + 2 == min_split_pos 5063 BlockBegin* min_block = allocator()->block_of_op_with_id(min_split_pos - 1); 5064 5065 // reason for using max_split_pos - 1: otherwise there would be an assertion failure 5066 // when an interval ends at the end of the last block of the method 5067 // (in this case, max_split_pos == allocator()->max_lir_op_id() + 2, and there is no 5068 // block at this op_id) 5069 BlockBegin* max_block = allocator()->block_of_op_with_id(max_split_pos - 1); 5070 5071 assert(min_block->linear_scan_number() <= max_block->linear_scan_number(), "invalid order"); 5072 if (min_block == max_block) { 5073 // split position cannot be moved to block boundary, so split as late as possible 5074 TRACE_LINEAR_SCAN(4, tty->print_cr(" cannot move split pos to block boundary because min_pos and max_pos are in same block")); 5075 optimal_split_pos = max_split_pos; 5076 5077 } else if (it->has_hole_between(max_split_pos - 1, max_split_pos) && !allocator()->is_block_begin(max_split_pos)) { 5078 // Do not move split position if the interval has a hole before max_split_pos. 5079 // Intervals resulting from Phi-Functions have more than one definition (marked 5080 // as mustHaveRegister) with a hole before each definition. When the register is needed 5081 // for the second definition, an earlier reloading is unnecessary. 5082 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has hole just before max_split_pos, so splitting at max_split_pos")); 5083 optimal_split_pos = max_split_pos; 5084 5085 } else { 5086 // seach optimal block boundary between min_split_pos and max_split_pos 5087 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())); 5088 5089 if (do_loop_optimization) { 5090 // Loop optimization: if a loop-end marker is found between min- and max-position, 5091 // then split before this loop 5092 int loop_end_pos = it->next_usage_exact(loopEndMarker, min_block->last_lir_instruction_id() + 2); 5093 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization: loop end found at pos %d", loop_end_pos)); 5094 5095 assert(loop_end_pos > min_split_pos, "invalid order"); 5096 if (loop_end_pos < max_split_pos) { 5097 // loop-end marker found between min- and max-position 5098 // if it is not the end marker for the same loop as the min-position, then move 5099 // the max-position to this loop block. 5100 // Desired result: uses tagged as shouldHaveRegister inside a loop cause a reloading 5101 // of the interval (normally, only mustHaveRegister causes a reloading) 5102 BlockBegin* loop_block = allocator()->block_of_op_with_id(loop_end_pos); 5103 5104 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())); 5105 assert(loop_block != min_block, "loop_block and min_block must be different because block boundary is needed between"); 5106 5107 optimal_split_pos = find_optimal_split_pos(min_block, loop_block, loop_block->last_lir_instruction_id() + 2); 5108 if (optimal_split_pos == loop_block->last_lir_instruction_id() + 2) { 5109 optimal_split_pos = -1; 5110 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization not necessary")); 5111 } else { 5112 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization successful")); 5113 } 5114 } 5115 } 5116 5117 if (optimal_split_pos == -1) { 5118 // not calculated by loop optimization 5119 optimal_split_pos = find_optimal_split_pos(min_block, max_block, max_split_pos); 5120 } 5121 } 5122 } 5123 TRACE_LINEAR_SCAN(4, tty->print_cr(" optimal split position: %d", optimal_split_pos)); 5124 5125 return optimal_split_pos; 5126 } 5127 5128 5129 /* 5130 split an interval at the optimal position between min_split_pos and 5131 max_split_pos in two parts: 5132 1) the left part has already a location assigned 5133 2) the right part is sorted into to the unhandled-list 5134 */ 5135 void LinearScanWalker::split_before_usage(Interval* it, int min_split_pos, int max_split_pos) { 5136 TRACE_LINEAR_SCAN(2, tty->print ("----- splitting interval: "); it->print()); 5137 TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d", min_split_pos, max_split_pos)); 5138 5139 assert(it->from() < min_split_pos, "cannot split at start of interval"); 5140 assert(current_position() < min_split_pos, "cannot split before current position"); 5141 assert(min_split_pos <= max_split_pos, "invalid order"); 5142 assert(max_split_pos <= it->to(), "cannot split after end of interval"); 5143 5144 int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, true); 5145 5146 assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range"); 5147 assert(optimal_split_pos <= it->to(), "cannot split after end of interval"); 5148 assert(optimal_split_pos > it->from(), "cannot split at start of interval"); 5149 5150 if (optimal_split_pos == it->to() && it->next_usage(mustHaveRegister, min_split_pos) == max_jint) { 5151 // the split position would be just before the end of the interval 5152 // -> no split at all necessary 5153 TRACE_LINEAR_SCAN(4, tty->print_cr(" no split necessary because optimal split position is at end of interval")); 5154 return; 5155 } 5156 5157 // must calculate this before the actual split is performed and before split position is moved to odd op_id 5158 bool move_necessary = !allocator()->is_block_begin(optimal_split_pos) && !it->has_hole_between(optimal_split_pos - 1, optimal_split_pos); 5159 5160 if (!allocator()->is_block_begin(optimal_split_pos)) { 5161 // move position before actual instruction (odd op_id) 5162 optimal_split_pos = (optimal_split_pos - 1) | 1; 5163 } 5164 5165 TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d", optimal_split_pos)); 5166 assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary"); 5167 assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary"); 5168 5169 Interval* split_part = it->split(optimal_split_pos); 5170 5171 allocator()->append_interval(split_part); 5172 allocator()->copy_register_flags(it, split_part); 5173 split_part->set_insert_move_when_activated(move_necessary); 5174 append_to_unhandled(unhandled_first_addr(anyKind), split_part); 5175 5176 TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts (insert_move_when_activated: %d)", move_necessary)); 5177 TRACE_LINEAR_SCAN(2, tty->print (" "); it->print()); 5178 TRACE_LINEAR_SCAN(2, tty->print (" "); split_part->print()); 5179 } 5180 5181 /* 5182 split an interval at the optimal position between min_split_pos and 5183 max_split_pos in two parts: 5184 1) the left part has already a location assigned 5185 2) the right part is always on the stack and therefore ignored in further processing 5186 */ 5187 void LinearScanWalker::split_for_spilling(Interval* it) { 5188 // calculate allowed range of splitting position 5189 int max_split_pos = current_position(); 5190 int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, max_split_pos) + 1, it->from()); 5191 5192 TRACE_LINEAR_SCAN(2, tty->print ("----- splitting and spilling interval: "); it->print()); 5193 TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d", min_split_pos, max_split_pos)); 5194 5195 assert(it->state() == activeState, "why spill interval that is not active?"); 5196 assert(it->from() <= min_split_pos, "cannot split before start of interval"); 5197 assert(min_split_pos <= max_split_pos, "invalid order"); 5198 assert(max_split_pos < it->to(), "cannot split at end end of interval"); 5199 assert(current_position() < it->to(), "interval must not end before current position"); 5200 5201 if (min_split_pos == it->from()) { 5202 // the whole interval is never used, so spill it entirely to memory 5203 TRACE_LINEAR_SCAN(2, tty->print_cr(" spilling entire interval because split pos is at beginning of interval")); 5204 assert(it->first_usage(shouldHaveRegister) > current_position(), "interval must not have use position before current_position"); 5205 5206 allocator()->assign_spill_slot(it); 5207 allocator()->change_spill_state(it, min_split_pos); 5208 5209 // Also kick parent intervals out of register to memory when they have no use 5210 // position. This avoids short interval in register surrounded by intervals in 5211 // memory -> avoid useless moves from memory to register and back 5212 Interval* parent = it; 5213 while (parent != NULL && parent->is_split_child()) { 5214 parent = parent->split_child_before_op_id(parent->from()); 5215 5216 if (parent->assigned_reg() < LinearScan::nof_regs) { 5217 if (parent->first_usage(shouldHaveRegister) == max_jint) { 5218 // parent is never used, so kick it out of its assigned register 5219 TRACE_LINEAR_SCAN(4, tty->print_cr(" kicking out interval %d out of its register because it is never used", parent->reg_num())); 5220 allocator()->assign_spill_slot(parent); 5221 } else { 5222 // do not go further back because the register is actually used by the interval 5223 parent = NULL; 5224 } 5225 } 5226 } 5227 5228 } else { 5229 // search optimal split pos, split interval and spill only the right hand part 5230 int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, false); 5231 5232 assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range"); 5233 assert(optimal_split_pos < it->to(), "cannot split at end of interval"); 5234 assert(optimal_split_pos >= it->from(), "cannot split before start of interval"); 5235 5236 if (!allocator()->is_block_begin(optimal_split_pos)) { 5237 // move position before actual instruction (odd op_id) 5238 optimal_split_pos = (optimal_split_pos - 1) | 1; 5239 } 5240 5241 TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d", optimal_split_pos)); 5242 assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary"); 5243 assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary"); 5244 5245 Interval* spilled_part = it->split(optimal_split_pos); 5246 allocator()->append_interval(spilled_part); 5247 allocator()->assign_spill_slot(spilled_part); 5248 allocator()->change_spill_state(spilled_part, optimal_split_pos); 5249 5250 if (!allocator()->is_block_begin(optimal_split_pos)) { 5251 TRACE_LINEAR_SCAN(4, tty->print_cr(" inserting move from interval %d to %d", it->reg_num(), spilled_part->reg_num())); 5252 insert_move(optimal_split_pos, it, spilled_part); 5253 } 5254 5255 // the current_split_child is needed later when moves are inserted for reloading 5256 assert(spilled_part->current_split_child() == it, "overwriting wrong current_split_child"); 5257 spilled_part->make_current_split_child(); 5258 5259 TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts")); 5260 TRACE_LINEAR_SCAN(2, tty->print (" "); it->print()); 5261 TRACE_LINEAR_SCAN(2, tty->print (" "); spilled_part->print()); 5262 } 5263 } 5264 5265 5266 void LinearScanWalker::split_stack_interval(Interval* it) { 5267 int min_split_pos = current_position() + 1; 5268 int max_split_pos = MIN2(it->first_usage(shouldHaveRegister), it->to()); 5269 5270 split_before_usage(it, min_split_pos, max_split_pos); 5271 } 5272 5273 void LinearScanWalker::split_when_partial_register_available(Interval* it, int register_available_until) { 5274 int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, register_available_until), it->from() + 1); 5275 int max_split_pos = register_available_until; 5276 5277 split_before_usage(it, min_split_pos, max_split_pos); 5278 } 5279 5280 void LinearScanWalker::split_and_spill_interval(Interval* it) { 5281 assert(it->state() == activeState || it->state() == inactiveState, "other states not allowed"); 5282 5283 int current_pos = current_position(); 5284 if (it->state() == inactiveState) { 5285 // the interval is currently inactive, so no spill slot is needed for now. 5286 // when the split part is activated, the interval has a new chance to get a register, 5287 // so in the best case no stack slot is necessary 5288 assert(it->has_hole_between(current_pos - 1, current_pos + 1), "interval can not be inactive otherwise"); 5289 split_before_usage(it, current_pos + 1, current_pos + 1); 5290 5291 } else { 5292 // search the position where the interval must have a register and split 5293 // at the optimal position before. 5294 // The new created part is added to the unhandled list and will get a register 5295 // when it is activated 5296 int min_split_pos = current_pos + 1; 5297 int max_split_pos = MIN2(it->next_usage(mustHaveRegister, min_split_pos), it->to()); 5298 5299 split_before_usage(it, min_split_pos, max_split_pos); 5300 5301 assert(it->next_usage(mustHaveRegister, current_pos) == max_jint, "the remaining part is spilled to stack and therefore has no register"); 5302 split_for_spilling(it); 5303 } 5304 } 5305 5306 5307 int LinearScanWalker::find_free_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) { 5308 int min_full_reg = any_reg; 5309 int max_partial_reg = any_reg; 5310 5311 for (int i = _first_reg; i <= _last_reg; i++) { 5312 if (i == ignore_reg) { 5313 // this register must be ignored 5314 5315 } else if (_use_pos[i] >= interval_to) { 5316 // this register is free for the full interval 5317 if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) { 5318 min_full_reg = i; 5319 } 5320 } else if (_use_pos[i] > reg_needed_until) { 5321 // this register is at least free until reg_needed_until 5322 if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) { 5323 max_partial_reg = i; 5324 } 5325 } 5326 } 5327 5328 if (min_full_reg != any_reg) { 5329 return min_full_reg; 5330 } else if (max_partial_reg != any_reg) { 5331 *need_split = true; 5332 return max_partial_reg; 5333 } else { 5334 return any_reg; 5335 } 5336 } 5337 5338 int LinearScanWalker::find_free_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) { 5339 assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm"); 5340 5341 int min_full_reg = any_reg; 5342 int max_partial_reg = any_reg; 5343 5344 for (int i = _first_reg; i < _last_reg; i+=2) { 5345 if (_use_pos[i] >= interval_to && _use_pos[i + 1] >= interval_to) { 5346 // this register is free for the full interval 5347 if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) { 5348 min_full_reg = i; 5349 } 5350 } else if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) { 5351 // this register is at least free until reg_needed_until 5352 if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) { 5353 max_partial_reg = i; 5354 } 5355 } 5356 } 5357 5358 if (min_full_reg != any_reg) { 5359 return min_full_reg; 5360 } else if (max_partial_reg != any_reg) { 5361 *need_split = true; 5362 return max_partial_reg; 5363 } else { 5364 return any_reg; 5365 } 5366 } 5367 5368 5369 bool LinearScanWalker::alloc_free_reg(Interval* cur) { 5370 TRACE_LINEAR_SCAN(2, tty->print("trying to find free register for "); cur->print()); 5371 5372 init_use_lists(true); 5373 free_exclude_active_fixed(); 5374 free_exclude_active_any(); 5375 free_collect_inactive_fixed(cur); 5376 free_collect_inactive_any(cur); 5377 // free_collect_unhandled(fixedKind, cur); 5378 assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0"); 5379 5380 // _use_pos contains the start of the next interval that has this register assigned 5381 // (either as a fixed register or a normal allocated register in the past) 5382 // only intervals overlapping with cur are processed, non-overlapping invervals can be ignored safely 5383 TRACE_LINEAR_SCAN(4, tty->print_cr(" state of registers:")); 5384 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])); 5385 5386 int hint_reg, hint_regHi; 5387 Interval* register_hint = cur->register_hint(); 5388 if (register_hint != NULL) { 5389 hint_reg = register_hint->assigned_reg(); 5390 hint_regHi = register_hint->assigned_regHi(); 5391 5392 if (allocator()->is_precolored_cpu_interval(register_hint)) { 5393 assert(hint_reg != any_reg && hint_regHi == any_reg, "must be for fixed intervals"); 5394 hint_regHi = hint_reg + 1; // connect e.g. eax-edx 5395 } 5396 TRACE_LINEAR_SCAN(4, tty->print(" hint registers %d, %d from interval ", hint_reg, hint_regHi); register_hint->print()); 5397 5398 } else { 5399 hint_reg = any_reg; 5400 hint_regHi = any_reg; 5401 } 5402 assert(hint_reg == any_reg || hint_reg != hint_regHi, "hint reg and regHi equal"); 5403 assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned to interval"); 5404 5405 // the register must be free at least until this position 5406 int reg_needed_until = cur->from() + 1; 5407 int interval_to = cur->to(); 5408 5409 bool need_split = false; 5410 int split_pos = -1; 5411 int reg = any_reg; 5412 int regHi = any_reg; 5413 5414 if (_adjacent_regs) { 5415 reg = find_free_double_reg(reg_needed_until, interval_to, hint_reg, &need_split); 5416 regHi = reg + 1; 5417 if (reg == any_reg) { 5418 return false; 5419 } 5420 split_pos = MIN2(_use_pos[reg], _use_pos[regHi]); 5421 5422 } else { 5423 reg = find_free_reg(reg_needed_until, interval_to, hint_reg, any_reg, &need_split); 5424 if (reg == any_reg) { 5425 return false; 5426 } 5427 split_pos = _use_pos[reg]; 5428 5429 if (_num_phys_regs == 2) { 5430 regHi = find_free_reg(reg_needed_until, interval_to, hint_regHi, reg, &need_split); 5431 5432 if (_use_pos[reg] < interval_to && regHi == any_reg) { 5433 // do not split interval if only one register can be assigned until the split pos 5434 // (when one register is found for the whole interval, split&spill is only 5435 // performed for the hi register) 5436 return false; 5437 5438 } else if (regHi != any_reg) { 5439 split_pos = MIN2(split_pos, _use_pos[regHi]); 5440 5441 // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax 5442 if (reg > regHi) { 5443 int temp = reg; 5444 reg = regHi; 5445 regHi = temp; 5446 } 5447 } 5448 } 5449 } 5450 5451 cur->assign_reg(reg, regHi); 5452 TRACE_LINEAR_SCAN(2, tty->print_cr("selected register %d, %d", reg, regHi)); 5453 5454 assert(split_pos > 0, "invalid split_pos"); 5455 if (need_split) { 5456 // register not available for full interval, so split it 5457 split_when_partial_register_available(cur, split_pos); 5458 } 5459 5460 // only return true if interval is completely assigned 5461 return _num_phys_regs == 1 || regHi != any_reg; 5462 } 5463 5464 5465 int LinearScanWalker::find_locked_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) { 5466 int max_reg = any_reg; 5467 5468 for (int i = _first_reg; i <= _last_reg; i++) { 5469 if (i == ignore_reg) { 5470 // this register must be ignored 5471 5472 } else if (_use_pos[i] > reg_needed_until) { 5473 if (max_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_reg] && max_reg != hint_reg)) { 5474 max_reg = i; 5475 } 5476 } 5477 } 5478 5479 if (max_reg != any_reg && _block_pos[max_reg] <= interval_to) { 5480 *need_split = true; 5481 } 5482 5483 return max_reg; 5484 } 5485 5486 int LinearScanWalker::find_locked_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) { 5487 assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm"); 5488 5489 int max_reg = any_reg; 5490 5491 for (int i = _first_reg; i < _last_reg; i+=2) { 5492 if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) { 5493 if (max_reg == any_reg || _use_pos[i] > _use_pos[max_reg]) { 5494 max_reg = i; 5495 } 5496 } 5497 } 5498 5499 if (_block_pos[max_reg] <= interval_to || _block_pos[max_reg + 1] <= interval_to) { 5500 *need_split = true; 5501 } 5502 5503 return max_reg; 5504 } 5505 5506 void LinearScanWalker::split_and_spill_intersecting_intervals(int reg, int regHi) { 5507 assert(reg != any_reg, "no register assigned"); 5508 5509 for (int i = 0; i < _spill_intervals[reg]->length(); i++) { 5510 Interval* it = _spill_intervals[reg]->at(i); 5511 remove_from_list(it); 5512 split_and_spill_interval(it); 5513 } 5514 5515 if (regHi != any_reg) { 5516 IntervalList* processed = _spill_intervals[reg]; 5517 for (int i = 0; i < _spill_intervals[regHi]->length(); i++) { 5518 Interval* it = _spill_intervals[regHi]->at(i); 5519 if (processed->index_of(it) == -1) { 5520 remove_from_list(it); 5521 split_and_spill_interval(it); 5522 } 5523 } 5524 } 5525 } 5526 5527 5528 // Split an Interval and spill it to memory so that cur can be placed in a register 5529 void LinearScanWalker::alloc_locked_reg(Interval* cur) { 5530 TRACE_LINEAR_SCAN(2, tty->print("need to split and spill to get register for "); cur->print()); 5531 5532 // collect current usage of registers 5533 init_use_lists(false); 5534 spill_exclude_active_fixed(); 5535 // spill_block_unhandled_fixed(cur); 5536 assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0"); 5537 spill_block_inactive_fixed(cur); 5538 spill_collect_active_any(); 5539 spill_collect_inactive_any(cur); 5540 5541 #ifndef PRODUCT 5542 if (TraceLinearScanLevel >= 4) { 5543 tty->print_cr(" state of registers:"); 5544 for (int i = _first_reg; i <= _last_reg; i++) { 5545 tty->print(" reg %d: use_pos: %d, block_pos: %d, intervals: ", i, _use_pos[i], _block_pos[i]); 5546 for (int j = 0; j < _spill_intervals[i]->length(); j++) { 5547 tty->print("%d ", _spill_intervals[i]->at(j)->reg_num()); 5548 } 5549 tty->cr(); 5550 } 5551 } 5552 #endif 5553 5554 // the register must be free at least until this position 5555 int reg_needed_until = MIN2(cur->first_usage(mustHaveRegister), cur->from() + 1); 5556 int interval_to = cur->to(); 5557 assert (reg_needed_until > 0 && reg_needed_until < max_jint, "interval has no use"); 5558 5559 int split_pos = 0; 5560 int use_pos = 0; 5561 bool need_split = false; 5562 int reg, regHi; 5563 5564 if (_adjacent_regs) { 5565 reg = find_locked_double_reg(reg_needed_until, interval_to, any_reg, &need_split); 5566 regHi = reg + 1; 5567 5568 if (reg != any_reg) { 5569 use_pos = MIN2(_use_pos[reg], _use_pos[regHi]); 5570 split_pos = MIN2(_block_pos[reg], _block_pos[regHi]); 5571 } 5572 } else { 5573 reg = find_locked_reg(reg_needed_until, interval_to, any_reg, cur->assigned_reg(), &need_split); 5574 regHi = any_reg; 5575 5576 if (reg != any_reg) { 5577 use_pos = _use_pos[reg]; 5578 split_pos = _block_pos[reg]; 5579 5580 if (_num_phys_regs == 2) { 5581 if (cur->assigned_reg() != any_reg) { 5582 regHi = reg; 5583 reg = cur->assigned_reg(); 5584 } else { 5585 regHi = find_locked_reg(reg_needed_until, interval_to, any_reg, reg, &need_split); 5586 if (regHi != any_reg) { 5587 use_pos = MIN2(use_pos, _use_pos[regHi]); 5588 split_pos = MIN2(split_pos, _block_pos[regHi]); 5589 } 5590 } 5591 5592 if (regHi != any_reg && reg > regHi) { 5593 // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax 5594 int temp = reg; 5595 reg = regHi; 5596 regHi = temp; 5597 } 5598 } 5599 } 5600 } 5601 5602 if (reg == any_reg || (_num_phys_regs == 2 && regHi == any_reg) || use_pos <= cur->first_usage(mustHaveRegister)) { 5603 // the first use of cur is later than the spilling position -> spill cur 5604 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)); 5605 5606 if (cur->first_usage(mustHaveRegister) <= cur->from() + 1) { 5607 assert(false, "cannot spill interval that is used in first instruction (possible reason: no register found)"); 5608 // assign a reasonable register and do a bailout in product mode to avoid errors 5609 allocator()->assign_spill_slot(cur); 5610 BAILOUT("LinearScan: no register found"); 5611 } 5612 5613 split_and_spill_interval(cur); 5614 } else { 5615 TRACE_LINEAR_SCAN(4, tty->print_cr("decided to use register %d, %d", reg, regHi)); 5616 assert(reg != any_reg && (_num_phys_regs == 1 || regHi != any_reg), "no register found"); 5617 assert(split_pos > 0, "invalid split_pos"); 5618 assert(need_split == false || split_pos > cur->from(), "splitting interval at from"); 5619 5620 cur->assign_reg(reg, regHi); 5621 if (need_split) { 5622 // register not available for full interval, so split it 5623 split_when_partial_register_available(cur, split_pos); 5624 } 5625 5626 // perform splitting and spilling for all affected intervalls 5627 split_and_spill_intersecting_intervals(reg, regHi); 5628 } 5629 } 5630 5631 bool LinearScanWalker::no_allocation_possible(Interval* cur) { 5632 #ifdef X86 5633 // fast calculation of intervals that can never get a register because the 5634 // the next instruction is a call that blocks all registers 5635 // Note: this does not work if callee-saved registers are available (e.g. on Sparc) 5636 5637 // check if this interval is the result of a split operation 5638 // (an interval got a register until this position) 5639 int pos = cur->from(); 5640 if ((pos & 1) == 1) { 5641 // the current instruction is a call that blocks all registers 5642 if (pos < allocator()->max_lir_op_id() && allocator()->has_call(pos + 1)) { 5643 TRACE_LINEAR_SCAN(4, tty->print_cr(" free register cannot be available because all registers blocked by following call")); 5644 5645 // safety check that there is really no register available 5646 assert(alloc_free_reg(cur) == false, "found a register for this interval"); 5647 return true; 5648 } 5649 5650 } 5651 #endif 5652 return false; 5653 } 5654 5655 void LinearScanWalker::init_vars_for_alloc(Interval* cur) { 5656 BasicType type = cur->type(); 5657 _num_phys_regs = LinearScan::num_physical_regs(type); 5658 _adjacent_regs = LinearScan::requires_adjacent_regs(type); 5659 5660 if (pd_init_regs_for_alloc(cur)) { 5661 // the appropriate register range was selected. 5662 } else if (type == T_FLOAT || type == T_DOUBLE) { 5663 _first_reg = pd_first_fpu_reg; 5664 _last_reg = pd_last_fpu_reg; 5665 } else { 5666 _first_reg = pd_first_cpu_reg; 5667 _last_reg = FrameMap::last_cpu_reg(); 5668 } 5669 5670 assert(0 <= _first_reg && _first_reg < LinearScan::nof_regs, "out of range"); 5671 assert(0 <= _last_reg && _last_reg < LinearScan::nof_regs, "out of range"); 5672 } 5673 5674 5675 bool LinearScanWalker::is_move(LIR_Op* op, Interval* from, Interval* to) { 5676 if (op->code() != lir_move) { 5677 return false; 5678 } 5679 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); 5680 5681 LIR_Opr in = ((LIR_Op1*)op)->in_opr(); 5682 LIR_Opr res = ((LIR_Op1*)op)->result_opr(); 5683 return in->is_virtual() && res->is_virtual() && in->vreg_number() == from->reg_num() && res->vreg_number() == to->reg_num(); 5684 } 5685 5686 // optimization (especially for phi functions of nested loops): 5687 // assign same spill slot to non-intersecting intervals 5688 void LinearScanWalker::combine_spilled_intervals(Interval* cur) { 5689 if (cur->is_split_child()) { 5690 // optimization is only suitable for split parents 5691 return; 5692 } 5693 5694 Interval* register_hint = cur->register_hint(false); 5695 if (register_hint == NULL) { 5696 // cur is not the target of a move, otherwise register_hint would be set 5697 return; 5698 } 5699 assert(register_hint->is_split_parent(), "register hint must be split parent"); 5700 5701 if (cur->spill_state() != noOptimization || register_hint->spill_state() != noOptimization) { 5702 // combining the stack slots for intervals where spill move optimization is applied 5703 // is not benefitial and would cause problems 5704 return; 5705 } 5706 5707 int begin_pos = cur->from(); 5708 int end_pos = cur->to(); 5709 if (end_pos > allocator()->max_lir_op_id() || (begin_pos & 1) != 0 || (end_pos & 1) != 0) { 5710 // safety check that lir_op_with_id is allowed 5711 return; 5712 } 5713 5714 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)) { 5715 // cur and register_hint are not connected with two moves 5716 return; 5717 } 5718 5719 Interval* begin_hint = register_hint->split_child_at_op_id(begin_pos, LIR_OpVisitState::inputMode); 5720 Interval* end_hint = register_hint->split_child_at_op_id(end_pos, LIR_OpVisitState::outputMode); 5721 if (begin_hint == end_hint || begin_hint->to() != begin_pos || end_hint->from() != end_pos) { 5722 // register_hint must be split, otherwise the re-writing of use positions does not work 5723 return; 5724 } 5725 5726 assert(begin_hint->assigned_reg() != any_reg, "must have register assigned"); 5727 assert(end_hint->assigned_reg() == any_reg, "must not have register assigned"); 5728 assert(cur->first_usage(mustHaveRegister) == begin_pos, "must have use position at begin of interval because of move"); 5729 assert(end_hint->first_usage(mustHaveRegister) == end_pos, "must have use position at begin of interval because of move"); 5730 5731 if (begin_hint->assigned_reg() < LinearScan::nof_regs) { 5732 // register_hint is not spilled at begin_pos, so it would not be benefitial to immediately spill cur 5733 return; 5734 } 5735 assert(register_hint->canonical_spill_slot() != -1, "must be set when part of interval was spilled"); 5736 5737 // modify intervals such that cur gets the same stack slot as register_hint 5738 // delete use positions to prevent the intervals to get a register at beginning 5739 cur->set_canonical_spill_slot(register_hint->canonical_spill_slot()); 5740 cur->remove_first_use_pos(); 5741 end_hint->remove_first_use_pos(); 5742 } 5743 5744 5745 // allocate a physical register or memory location to an interval 5746 bool LinearScanWalker::activate_current() { 5747 Interval* cur = current(); 5748 bool result = true; 5749 5750 TRACE_LINEAR_SCAN(2, tty->print ("+++++ activating interval "); cur->print()); 5751 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())); 5752 5753 if (cur->assigned_reg() >= LinearScan::nof_regs) { 5754 // activating an interval that has a stack slot assigned -> split it at first use position 5755 // used for method parameters 5756 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has spill slot assigned (method parameter) -> split it before first use")); 5757 5758 split_stack_interval(cur); 5759 result = false; 5760 5761 } else if (allocator()->gen()->is_vreg_flag_set(cur->reg_num(), LIRGenerator::must_start_in_memory)) { 5762 // activating an interval that must start in a stack slot, but may get a register later 5763 // used for lir_roundfp: rounding is done by store to stack and reload later 5764 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval must start in stack slot -> split it before first use")); 5765 assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned"); 5766 5767 allocator()->assign_spill_slot(cur); 5768 split_stack_interval(cur); 5769 result = false; 5770 5771 } else if (cur->assigned_reg() == any_reg) { 5772 // interval has not assigned register -> normal allocation 5773 // (this is the normal case for most intervals) 5774 TRACE_LINEAR_SCAN(4, tty->print_cr(" normal allocation of register")); 5775 5776 // assign same spill slot to non-intersecting intervals 5777 combine_spilled_intervals(cur); 5778 5779 init_vars_for_alloc(cur); 5780 if (no_allocation_possible(cur) || !alloc_free_reg(cur)) { 5781 // no empty register available. 5782 // split and spill another interval so that this interval gets a register 5783 alloc_locked_reg(cur); 5784 } 5785 5786 // spilled intervals need not be move to active-list 5787 if (cur->assigned_reg() >= LinearScan::nof_regs) { 5788 result = false; 5789 } 5790 } 5791 5792 // load spilled values that become active from stack slot to register 5793 if (cur->insert_move_when_activated()) { 5794 assert(cur->is_split_child(), "must be"); 5795 assert(cur->current_split_child() != NULL, "must be"); 5796 assert(cur->current_split_child()->reg_num() != cur->reg_num(), "cannot insert move between same interval"); 5797 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())); 5798 5799 insert_move(cur->from(), cur->current_split_child(), cur); 5800 } 5801 cur->make_current_split_child(); 5802 5803 return result; // true = interval is moved to active list 5804 } 5805 5806 5807 // Implementation of EdgeMoveOptimizer 5808 5809 EdgeMoveOptimizer::EdgeMoveOptimizer() : 5810 _edge_instructions(4), 5811 _edge_instructions_idx(4) 5812 { 5813 } 5814 5815 void EdgeMoveOptimizer::optimize(BlockList* code) { 5816 EdgeMoveOptimizer optimizer = EdgeMoveOptimizer(); 5817 5818 // ignore the first block in the list (index 0 is not processed) 5819 for (int i = code->length() - 1; i >= 1; i--) { 5820 BlockBegin* block = code->at(i); 5821 5822 if (block->number_of_preds() > 1 && !block->is_set(BlockBegin::exception_entry_flag)) { 5823 optimizer.optimize_moves_at_block_end(block); 5824 } 5825 if (block->number_of_sux() == 2) { 5826 optimizer.optimize_moves_at_block_begin(block); 5827 } 5828 } 5829 } 5830 5831 5832 // clear all internal data structures 5833 void EdgeMoveOptimizer::init_instructions() { 5834 _edge_instructions.clear(); 5835 _edge_instructions_idx.clear(); 5836 } 5837 5838 // append a lir-instruction-list and the index of the current operation in to the list 5839 void EdgeMoveOptimizer::append_instructions(LIR_OpList* instructions, int instructions_idx) { 5840 _edge_instructions.append(instructions); 5841 _edge_instructions_idx.append(instructions_idx); 5842 } 5843 5844 // return the current operation of the given edge (predecessor or successor) 5845 LIR_Op* EdgeMoveOptimizer::instruction_at(int edge) { 5846 LIR_OpList* instructions = _edge_instructions.at(edge); 5847 int idx = _edge_instructions_idx.at(edge); 5848 5849 if (idx < instructions->length()) { 5850 return instructions->at(idx); 5851 } else { 5852 return NULL; 5853 } 5854 } 5855 5856 // removes the current operation of the given edge (predecessor or successor) 5857 void EdgeMoveOptimizer::remove_cur_instruction(int edge, bool decrement_index) { 5858 LIR_OpList* instructions = _edge_instructions.at(edge); 5859 int idx = _edge_instructions_idx.at(edge); 5860 instructions->remove_at(idx); 5861 5862 if (decrement_index) { 5863 _edge_instructions_idx.at_put(edge, idx - 1); 5864 } 5865 } 5866 5867 5868 bool EdgeMoveOptimizer::operations_different(LIR_Op* op1, LIR_Op* op2) { 5869 if (op1 == NULL || op2 == NULL) { 5870 // at least one block is already empty -> no optimization possible 5871 return true; 5872 } 5873 5874 if (op1->code() == lir_move && op2->code() == lir_move) { 5875 assert(op1->as_Op1() != NULL, "move must be LIR_Op1"); 5876 assert(op2->as_Op1() != NULL, "move must be LIR_Op1"); 5877 LIR_Op1* move1 = (LIR_Op1*)op1; 5878 LIR_Op1* move2 = (LIR_Op1*)op2; 5879 if (move1->info() == move2->info() && move1->in_opr() == move2->in_opr() && move1->result_opr() == move2->result_opr()) { 5880 // these moves are exactly equal and can be optimized 5881 return false; 5882 } 5883 5884 } else if (op1->code() == lir_fxch && op2->code() == lir_fxch) { 5885 assert(op1->as_Op1() != NULL, "fxch must be LIR_Op1"); 5886 assert(op2->as_Op1() != NULL, "fxch must be LIR_Op1"); 5887 LIR_Op1* fxch1 = (LIR_Op1*)op1; 5888 LIR_Op1* fxch2 = (LIR_Op1*)op2; 5889 if (fxch1->in_opr()->as_jint() == fxch2->in_opr()->as_jint()) { 5890 // equal FPU stack operations can be optimized 5891 return false; 5892 } 5893 5894 } else if (op1->code() == lir_fpop_raw && op2->code() == lir_fpop_raw) { 5895 // equal FPU stack operations can be optimized 5896 return false; 5897 } 5898 5899 // no optimization possible 5900 return true; 5901 } 5902 5903 void EdgeMoveOptimizer::optimize_moves_at_block_end(BlockBegin* block) { 5904 TRACE_LINEAR_SCAN(4, tty->print_cr("optimizing moves at end of block B%d", block->block_id())); 5905 5906 if (block->is_predecessor(block)) { 5907 // currently we can't handle this correctly. 5908 return; 5909 } 5910 5911 init_instructions(); 5912 int num_preds = block->number_of_preds(); 5913 assert(num_preds > 1, "do not call otherwise"); 5914 assert(!block->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed"); 5915 5916 // setup a list with the lir-instructions of all predecessors 5917 int i; 5918 for (i = 0; i < num_preds; i++) { 5919 BlockBegin* pred = block->pred_at(i); 5920 LIR_OpList* pred_instructions = pred->lir()->instructions_list(); 5921 5922 if (pred->number_of_sux() != 1) { 5923 // this can happen with switch-statements where multiple edges are between 5924 // the same blocks. 5925 return; 5926 } 5927 5928 assert(pred->number_of_sux() == 1, "can handle only one successor"); 5929 assert(pred->sux_at(0) == block, "invalid control flow"); 5930 assert(pred_instructions->last()->code() == lir_branch, "block with successor must end with branch"); 5931 assert(pred_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch"); 5932 assert(pred_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch"); 5933 5934 if (pred_instructions->last()->info() != NULL) { 5935 // can not optimize instructions when debug info is needed 5936 return; 5937 } 5938 5939 // ignore the unconditional branch at the end of the block 5940 append_instructions(pred_instructions, pred_instructions->length() - 2); 5941 } 5942 5943 5944 // process lir-instructions while all predecessors end with the same instruction 5945 while (true) { 5946 LIR_Op* op = instruction_at(0); 5947 for (i = 1; i < num_preds; i++) { 5948 if (operations_different(op, instruction_at(i))) { 5949 // these instructions are different and cannot be optimized -> 5950 // no further optimization possible 5951 return; 5952 } 5953 } 5954 5955 TRACE_LINEAR_SCAN(4, tty->print("found instruction that is equal in all %d predecessors: ", num_preds); op->print()); 5956 5957 // insert the instruction at the beginning of the current block 5958 block->lir()->insert_before(1, op); 5959 5960 // delete the instruction at the end of all predecessors 5961 for (i = 0; i < num_preds; i++) { 5962 remove_cur_instruction(i, true); 5963 } 5964 } 5965 } 5966 5967 5968 void EdgeMoveOptimizer::optimize_moves_at_block_begin(BlockBegin* block) { 5969 TRACE_LINEAR_SCAN(4, tty->print_cr("optimization moves at begin of block B%d", block->block_id())); 5970 5971 init_instructions(); 5972 int num_sux = block->number_of_sux(); 5973 5974 LIR_OpList* cur_instructions = block->lir()->instructions_list(); 5975 5976 assert(num_sux == 2, "method should not be called otherwise"); 5977 assert(cur_instructions->last()->code() == lir_branch, "block with successor must end with branch"); 5978 assert(cur_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch"); 5979 assert(cur_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch"); 5980 5981 if (cur_instructions->last()->info() != NULL) { 5982 // can no optimize instructions when debug info is needed 5983 return; 5984 } 5985 5986 LIR_Op* branch = cur_instructions->at(cur_instructions->length() - 2); 5987 if (branch->info() != NULL || (branch->code() != lir_branch && branch->code() != lir_cond_float_branch)) { 5988 // not a valid case for optimization 5989 // currently, only blocks that end with two branches (conditional branch followed 5990 // by unconditional branch) are optimized 5991 return; 5992 } 5993 5994 // now it is guaranteed that the block ends with two branch instructions. 5995 // the instructions are inserted at the end of the block before these two branches 5996 int insert_idx = cur_instructions->length() - 2; 5997 5998 int i; 5999 #ifdef ASSERT 6000 for (i = insert_idx - 1; i >= 0; i--) { 6001 LIR_Op* op = cur_instructions->at(i); 6002 if ((op->code() == lir_branch || op->code() == lir_cond_float_branch) && ((LIR_OpBranch*)op)->block() != NULL) { 6003 assert(false, "block with two successors can have only two branch instructions"); 6004 } 6005 } 6006 #endif 6007 6008 // setup a list with the lir-instructions of all successors 6009 for (i = 0; i < num_sux; i++) { 6010 BlockBegin* sux = block->sux_at(i); 6011 LIR_OpList* sux_instructions = sux->lir()->instructions_list(); 6012 6013 assert(sux_instructions->at(0)->code() == lir_label, "block must start with label"); 6014 6015 if (sux->number_of_preds() != 1) { 6016 // this can happen with switch-statements where multiple edges are between 6017 // the same blocks. 6018 return; 6019 } 6020 assert(sux->pred_at(0) == block, "invalid control flow"); 6021 assert(!sux->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed"); 6022 6023 // ignore the label at the beginning of the block 6024 append_instructions(sux_instructions, 1); 6025 } 6026 6027 // process lir-instructions while all successors begin with the same instruction 6028 while (true) { 6029 LIR_Op* op = instruction_at(0); 6030 for (i = 1; i < num_sux; i++) { 6031 if (operations_different(op, instruction_at(i))) { 6032 // these instructions are different and cannot be optimized -> 6033 // no further optimization possible 6034 return; 6035 } 6036 } 6037 6038 TRACE_LINEAR_SCAN(4, tty->print("----- found instruction that is equal in all %d successors: ", num_sux); op->print()); 6039 6040 // insert instruction at end of current block 6041 block->lir()->insert_before(insert_idx, op); 6042 insert_idx++; 6043 6044 // delete the instructions at the beginning of all successors 6045 for (i = 0; i < num_sux; i++) { 6046 remove_cur_instruction(i, false); 6047 } 6048 } 6049 } 6050 6051 6052 // Implementation of ControlFlowOptimizer 6053 6054 ControlFlowOptimizer::ControlFlowOptimizer() : 6055 _original_preds(4) 6056 { 6057 } 6058 6059 void ControlFlowOptimizer::optimize(BlockList* code) { 6060 ControlFlowOptimizer optimizer = ControlFlowOptimizer(); 6061 6062 // push the OSR entry block to the end so that we're not jumping over it. 6063 BlockBegin* osr_entry = code->at(0)->end()->as_Base()->osr_entry(); 6064 if (osr_entry) { 6065 int index = osr_entry->linear_scan_number(); 6066 assert(code->at(index) == osr_entry, "wrong index"); 6067 code->remove_at(index); 6068 code->append(osr_entry); 6069 } 6070 6071 optimizer.reorder_short_loops(code); 6072 optimizer.delete_empty_blocks(code); 6073 optimizer.delete_unnecessary_jumps(code); 6074 optimizer.delete_jumps_to_return(code); 6075 } 6076 6077 void ControlFlowOptimizer::reorder_short_loop(BlockList* code, BlockBegin* header_block, int header_idx) { 6078 int i = header_idx + 1; 6079 int max_end = MIN2(header_idx + ShortLoopSize, code->length()); 6080 while (i < max_end && code->at(i)->loop_depth() >= header_block->loop_depth()) { 6081 i++; 6082 } 6083 6084 if (i == code->length() || code->at(i)->loop_depth() < header_block->loop_depth()) { 6085 int end_idx = i - 1; 6086 BlockBegin* end_block = code->at(end_idx); 6087 6088 if (end_block->number_of_sux() == 1 && end_block->sux_at(0) == header_block) { 6089 // short loop from header_idx to end_idx found -> reorder blocks such that 6090 // the header_block is the last block instead of the first block of the loop 6091 TRACE_LINEAR_SCAN(1, tty->print_cr("Reordering short loop: length %d, header B%d, end B%d", 6092 end_idx - header_idx + 1, 6093 header_block->block_id(), end_block->block_id())); 6094 6095 for (int j = header_idx; j < end_idx; j++) { 6096 code->at_put(j, code->at(j + 1)); 6097 } 6098 code->at_put(end_idx, header_block); 6099 6100 // correct the flags so that any loop alignment occurs in the right place. 6101 assert(code->at(end_idx)->is_set(BlockBegin::backward_branch_target_flag), "must be backward branch target"); 6102 code->at(end_idx)->clear(BlockBegin::backward_branch_target_flag); 6103 code->at(header_idx)->set(BlockBegin::backward_branch_target_flag); 6104 } 6105 } 6106 } 6107 6108 void ControlFlowOptimizer::reorder_short_loops(BlockList* code) { 6109 for (int i = code->length() - 1; i >= 0; i--) { 6110 BlockBegin* block = code->at(i); 6111 6112 if (block->is_set(BlockBegin::linear_scan_loop_header_flag)) { 6113 reorder_short_loop(code, block, i); 6114 } 6115 } 6116 6117 DEBUG_ONLY(verify(code)); 6118 } 6119 6120 // only blocks with exactly one successor can be deleted. Such blocks 6121 // must always end with an unconditional branch to this successor 6122 bool ControlFlowOptimizer::can_delete_block(BlockBegin* block) { 6123 if (block->number_of_sux() != 1 || block->number_of_exception_handlers() != 0 || block->is_entry_block()) { 6124 return false; 6125 } 6126 6127 LIR_OpList* instructions = block->lir()->instructions_list(); 6128 6129 assert(instructions->length() >= 2, "block must have label and branch"); 6130 assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label"); 6131 assert(instructions->last()->as_OpBranch() != NULL, "last instrcution must always be a branch"); 6132 assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "branch must be unconditional"); 6133 assert(instructions->last()->as_OpBranch()->block() == block->sux_at(0), "branch target must be the successor"); 6134 6135 // block must have exactly one successor 6136 6137 if (instructions->length() == 2 && instructions->last()->info() == NULL) { 6138 return true; 6139 } 6140 return false; 6141 } 6142 6143 // substitute branch targets in all branch-instructions of this blocks 6144 void ControlFlowOptimizer::substitute_branch_target(BlockBegin* block, BlockBegin* target_from, BlockBegin* target_to) { 6145 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())); 6146 6147 LIR_OpList* instructions = block->lir()->instructions_list(); 6148 6149 assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label"); 6150 for (int i = instructions->length() - 1; i >= 1; i--) { 6151 LIR_Op* op = instructions->at(i); 6152 6153 if (op->code() == lir_branch || op->code() == lir_cond_float_branch) { 6154 assert(op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch"); 6155 LIR_OpBranch* branch = (LIR_OpBranch*)op; 6156 6157 if (branch->block() == target_from) { 6158 branch->change_block(target_to); 6159 } 6160 if (branch->ublock() == target_from) { 6161 branch->change_ublock(target_to); 6162 } 6163 } 6164 } 6165 } 6166 6167 void ControlFlowOptimizer::delete_empty_blocks(BlockList* code) { 6168 int old_pos = 0; 6169 int new_pos = 0; 6170 int num_blocks = code->length(); 6171 6172 while (old_pos < num_blocks) { 6173 BlockBegin* block = code->at(old_pos); 6174 6175 if (can_delete_block(block)) { 6176 BlockBegin* new_target = block->sux_at(0); 6177 6178 // propagate backward branch target flag for correct code alignment 6179 if (block->is_set(BlockBegin::backward_branch_target_flag)) { 6180 new_target->set(BlockBegin::backward_branch_target_flag); 6181 } 6182 6183 // collect a list with all predecessors that contains each predecessor only once 6184 // the predecessors of cur are changed during the substitution, so a copy of the 6185 // predecessor list is necessary 6186 int j; 6187 _original_preds.clear(); 6188 for (j = block->number_of_preds() - 1; j >= 0; j--) { 6189 BlockBegin* pred = block->pred_at(j); 6190 if (_original_preds.index_of(pred) == -1) { 6191 _original_preds.append(pred); 6192 } 6193 } 6194 6195 for (j = _original_preds.length() - 1; j >= 0; j--) { 6196 BlockBegin* pred = _original_preds.at(j); 6197 substitute_branch_target(pred, block, new_target); 6198 pred->substitute_sux(block, new_target); 6199 } 6200 } else { 6201 // adjust position of this block in the block list if blocks before 6202 // have been deleted 6203 if (new_pos != old_pos) { 6204 code->at_put(new_pos, code->at(old_pos)); 6205 } 6206 new_pos++; 6207 } 6208 old_pos++; 6209 } 6210 code->truncate(new_pos); 6211 6212 DEBUG_ONLY(verify(code)); 6213 } 6214 6215 void ControlFlowOptimizer::delete_unnecessary_jumps(BlockList* code) { 6216 // skip the last block because there a branch is always necessary 6217 for (int i = code->length() - 2; i >= 0; i--) { 6218 BlockBegin* block = code->at(i); 6219 LIR_OpList* instructions = block->lir()->instructions_list(); 6220 6221 LIR_Op* last_op = instructions->last(); 6222 if (last_op->code() == lir_branch) { 6223 assert(last_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch"); 6224 LIR_OpBranch* last_branch = (LIR_OpBranch*)last_op; 6225 6226 assert(last_branch->block() != NULL, "last branch must always have a block as target"); 6227 assert(last_branch->label() == last_branch->block()->label(), "must be equal"); 6228 6229 if (last_branch->info() == NULL) { 6230 if (last_branch->block() == code->at(i + 1)) { 6231 6232 TRACE_LINEAR_SCAN(3, tty->print_cr("Deleting unconditional branch at end of block B%d", block->block_id())); 6233 6234 // delete last branch instruction 6235 instructions->truncate(instructions->length() - 1); 6236 6237 } else { 6238 LIR_Op* prev_op = instructions->at(instructions->length() - 2); 6239 if (prev_op->code() == lir_branch || prev_op->code() == lir_cond_float_branch) { 6240 assert(prev_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch"); 6241 LIR_OpBranch* prev_branch = (LIR_OpBranch*)prev_op; 6242 6243 if (prev_branch->stub() == NULL) { 6244 6245 LIR_Op2* prev_cmp = NULL; 6246 6247 for(int j = instructions->length() - 3; j >= 0 && prev_cmp == NULL; j--) { 6248 prev_op = instructions->at(j); 6249 if (prev_op->code() == lir_cmp) { 6250 assert(prev_op->as_Op2() != NULL, "branch must be of type LIR_Op2"); 6251 prev_cmp = (LIR_Op2*)prev_op; 6252 assert(prev_branch->cond() == prev_cmp->condition(), "should be the same"); 6253 } 6254 } 6255 assert(prev_cmp != NULL, "should have found comp instruction for branch"); 6256 if (prev_branch->block() == code->at(i + 1) && prev_branch->info() == NULL) { 6257 6258 TRACE_LINEAR_SCAN(3, tty->print_cr("Negating conditional branch and deleting unconditional branch at end of block B%d", block->block_id())); 6259 6260 // eliminate a conditional branch to the immediate successor 6261 prev_branch->change_block(last_branch->block()); 6262 prev_branch->negate_cond(); 6263 prev_cmp->set_condition(prev_branch->cond()); 6264 instructions->truncate(instructions->length() - 1); 6265 } 6266 } 6267 } 6268 } 6269 } 6270 } 6271 } 6272 6273 DEBUG_ONLY(verify(code)); 6274 } 6275 6276 void ControlFlowOptimizer::delete_jumps_to_return(BlockList* code) { 6277 #ifdef ASSERT 6278 BitMap return_converted(BlockBegin::number_of_blocks()); 6279 return_converted.clear(); 6280 #endif 6281 6282 for (int i = code->length() - 1; i >= 0; i--) { 6283 BlockBegin* block = code->at(i); 6284 LIR_OpList* cur_instructions = block->lir()->instructions_list(); 6285 LIR_Op* cur_last_op = cur_instructions->last(); 6286 6287 assert(cur_instructions->at(0)->code() == lir_label, "first instruction must always be a label"); 6288 if (cur_instructions->length() == 2 && cur_last_op->code() == lir_return) { 6289 // the block contains only a label and a return 6290 // if a predecessor ends with an unconditional jump to this block, then the jump 6291 // can be replaced with a return instruction 6292 // 6293 // Note: the original block with only a return statement cannot be deleted completely 6294 // because the predecessors might have other (conditional) jumps to this block 6295 // -> this may lead to unnecesary return instructions in the final code 6296 6297 assert(cur_last_op->info() == NULL, "return instructions do not have debug information"); 6298 assert(block->number_of_sux() == 0 || 6299 (return_converted.at(block->block_id()) && block->number_of_sux() == 1), 6300 "blocks that end with return must not have successors"); 6301 6302 assert(cur_last_op->as_Op1() != NULL, "return must be LIR_Op1"); 6303 LIR_Opr return_opr = ((LIR_Op1*)cur_last_op)->in_opr(); 6304 6305 for (int j = block->number_of_preds() - 1; j >= 0; j--) { 6306 BlockBegin* pred = block->pred_at(j); 6307 LIR_OpList* pred_instructions = pred->lir()->instructions_list(); 6308 LIR_Op* pred_last_op = pred_instructions->last(); 6309 6310 if (pred_last_op->code() == lir_branch) { 6311 assert(pred_last_op->as_OpBranch() != NULL, "branch must be LIR_OpBranch"); 6312 LIR_OpBranch* pred_last_branch = (LIR_OpBranch*)pred_last_op; 6313 6314 if (pred_last_branch->block() == block && pred_last_branch->cond() == lir_cond_always && pred_last_branch->info() == NULL) { 6315 // replace the jump to a return with a direct return 6316 // Note: currently the edge between the blocks is not deleted 6317 pred_instructions->at_put(pred_instructions->length() - 1, new LIR_Op1(lir_return, return_opr)); 6318 #ifdef ASSERT 6319 return_converted.set_bit(pred->block_id()); 6320 #endif 6321 } 6322 } 6323 } 6324 } 6325 } 6326 } 6327 6328 6329 #ifdef ASSERT 6330 void ControlFlowOptimizer::verify(BlockList* code) { 6331 for (int i = 0; i < code->length(); i++) { 6332 BlockBegin* block = code->at(i); 6333 LIR_OpList* instructions = block->lir()->instructions_list(); 6334 6335 int j; 6336 for (j = 0; j < instructions->length(); j++) { 6337 LIR_OpBranch* op_branch = instructions->at(j)->as_OpBranch(); 6338 6339 if (op_branch != NULL) { 6340 assert(op_branch->block() == NULL || code->index_of(op_branch->block()) != -1, "branch target not valid"); 6341 assert(op_branch->ublock() == NULL || code->index_of(op_branch->ublock()) != -1, "branch target not valid"); 6342 } 6343 } 6344 6345 for (j = 0; j < block->number_of_sux() - 1; j++) { 6346 BlockBegin* sux = block->sux_at(j); 6347 assert(code->index_of(sux) != -1, "successor not valid"); 6348 } 6349 6350 for (j = 0; j < block->number_of_preds() - 1; j++) { 6351 BlockBegin* pred = block->pred_at(j); 6352 assert(code->index_of(pred) != -1, "successor not valid"); 6353 } 6354 } 6355 } 6356 #endif 6357 6358 6359 #ifndef PRODUCT 6360 6361 // Implementation of LinearStatistic 6362 6363 const char* LinearScanStatistic::counter_name(int counter_idx) { 6364 switch (counter_idx) { 6365 case counter_method: return "compiled methods"; 6366 case counter_fpu_method: return "methods using fpu"; 6367 case counter_loop_method: return "methods with loops"; 6368 case counter_exception_method:return "methods with xhandler"; 6369 6370 case counter_loop: return "loops"; 6371 case counter_block: return "blocks"; 6372 case counter_loop_block: return "blocks inside loop"; 6373 case counter_exception_block: return "exception handler entries"; 6374 case counter_interval: return "intervals"; 6375 case counter_fixed_interval: return "fixed intervals"; 6376 case counter_range: return "ranges"; 6377 case counter_fixed_range: return "fixed ranges"; 6378 case counter_use_pos: return "use positions"; 6379 case counter_fixed_use_pos: return "fixed use positions"; 6380 case counter_spill_slots: return "spill slots"; 6381 6382 // counter for classes of lir instructions 6383 case counter_instruction: return "total instructions"; 6384 case counter_label: return "labels"; 6385 case counter_entry: return "method entries"; 6386 case counter_return: return "method returns"; 6387 case counter_call: return "method calls"; 6388 case counter_move: return "moves"; 6389 case counter_cmp: return "compare"; 6390 case counter_cond_branch: return "conditional branches"; 6391 case counter_uncond_branch: return "unconditional branches"; 6392 case counter_stub_branch: return "branches to stub"; 6393 case counter_alu: return "artithmetic + logic"; 6394 case counter_alloc: return "allocations"; 6395 case counter_sync: return "synchronisation"; 6396 case counter_throw: return "throw"; 6397 case counter_unwind: return "unwind"; 6398 case counter_typecheck: return "type+null-checks"; 6399 case counter_fpu_stack: return "fpu-stack"; 6400 case counter_misc_inst: return "other instructions"; 6401 case counter_other_inst: return "misc. instructions"; 6402 6403 // counter for different types of moves 6404 case counter_move_total: return "total moves"; 6405 case counter_move_reg_reg: return "register->register"; 6406 case counter_move_reg_stack: return "register->stack"; 6407 case counter_move_stack_reg: return "stack->register"; 6408 case counter_move_stack_stack:return "stack->stack"; 6409 case counter_move_reg_mem: return "register->memory"; 6410 case counter_move_mem_reg: return "memory->register"; 6411 case counter_move_const_any: return "constant->any"; 6412 6413 case blank_line_1: return ""; 6414 case blank_line_2: return ""; 6415 6416 default: ShouldNotReachHere(); return ""; 6417 } 6418 } 6419 6420 LinearScanStatistic::Counter LinearScanStatistic::base_counter(int counter_idx) { 6421 if (counter_idx == counter_fpu_method || counter_idx == counter_loop_method || counter_idx == counter_exception_method) { 6422 return counter_method; 6423 } else if (counter_idx == counter_loop_block || counter_idx == counter_exception_block) { 6424 return counter_block; 6425 } else if (counter_idx >= counter_instruction && counter_idx <= counter_other_inst) { 6426 return counter_instruction; 6427 } else if (counter_idx >= counter_move_total && counter_idx <= counter_move_const_any) { 6428 return counter_move_total; 6429 } 6430 return invalid_counter; 6431 } 6432 6433 LinearScanStatistic::LinearScanStatistic() { 6434 for (int i = 0; i < number_of_counters; i++) { 6435 _counters_sum[i] = 0; 6436 _counters_max[i] = -1; 6437 } 6438 6439 } 6440 6441 // add the method-local numbers to the total sum 6442 void LinearScanStatistic::sum_up(LinearScanStatistic &method_statistic) { 6443 for (int i = 0; i < number_of_counters; i++) { 6444 _counters_sum[i] += method_statistic._counters_sum[i]; 6445 _counters_max[i] = MAX2(_counters_max[i], method_statistic._counters_sum[i]); 6446 } 6447 } 6448 6449 void LinearScanStatistic::print(const char* title) { 6450 if (CountLinearScan || TraceLinearScanLevel > 0) { 6451 tty->cr(); 6452 tty->print_cr("***** LinearScan statistic - %s *****", title); 6453 6454 for (int i = 0; i < number_of_counters; i++) { 6455 if (_counters_sum[i] > 0 || _counters_max[i] >= 0) { 6456 tty->print("%25s: %8d", counter_name(i), _counters_sum[i]); 6457 6458 if (base_counter(i) != invalid_counter) { 6459 tty->print(" (%5.1f%%) ", _counters_sum[i] * 100.0 / _counters_sum[base_counter(i)]); 6460 } else { 6461 tty->print(" "); 6462 } 6463 6464 if (_counters_max[i] >= 0) { 6465 tty->print("%8d", _counters_max[i]); 6466 } 6467 } 6468 tty->cr(); 6469 } 6470 } 6471 } 6472 6473 void LinearScanStatistic::collect(LinearScan* allocator) { 6474 inc_counter(counter_method); 6475 if (allocator->has_fpu_registers()) { 6476 inc_counter(counter_fpu_method); 6477 } 6478 if (allocator->num_loops() > 0) { 6479 inc_counter(counter_loop_method); 6480 } 6481 inc_counter(counter_loop, allocator->num_loops()); 6482 inc_counter(counter_spill_slots, allocator->max_spills()); 6483 6484 int i; 6485 for (i = 0; i < allocator->interval_count(); i++) { 6486 Interval* cur = allocator->interval_at(i); 6487 6488 if (cur != NULL) { 6489 inc_counter(counter_interval); 6490 inc_counter(counter_use_pos, cur->num_use_positions()); 6491 if (LinearScan::is_precolored_interval(cur)) { 6492 inc_counter(counter_fixed_interval); 6493 inc_counter(counter_fixed_use_pos, cur->num_use_positions()); 6494 } 6495 6496 Range* range = cur->first(); 6497 while (range != Range::end()) { 6498 inc_counter(counter_range); 6499 if (LinearScan::is_precolored_interval(cur)) { 6500 inc_counter(counter_fixed_range); 6501 } 6502 range = range->next(); 6503 } 6504 } 6505 } 6506 6507 bool has_xhandlers = false; 6508 // Note: only count blocks that are in code-emit order 6509 for (i = 0; i < allocator->ir()->code()->length(); i++) { 6510 BlockBegin* cur = allocator->ir()->code()->at(i); 6511 6512 inc_counter(counter_block); 6513 if (cur->loop_depth() > 0) { 6514 inc_counter(counter_loop_block); 6515 } 6516 if (cur->is_set(BlockBegin::exception_entry_flag)) { 6517 inc_counter(counter_exception_block); 6518 has_xhandlers = true; 6519 } 6520 6521 LIR_OpList* instructions = cur->lir()->instructions_list(); 6522 for (int j = 0; j < instructions->length(); j++) { 6523 LIR_Op* op = instructions->at(j); 6524 6525 inc_counter(counter_instruction); 6526 6527 switch (op->code()) { 6528 case lir_label: inc_counter(counter_label); break; 6529 case lir_std_entry: 6530 case lir_osr_entry: inc_counter(counter_entry); break; 6531 case lir_return: inc_counter(counter_return); break; 6532 6533 case lir_rtcall: 6534 case lir_static_call: 6535 case lir_optvirtual_call: 6536 case lir_virtual_call: inc_counter(counter_call); break; 6537 6538 case lir_move: { 6539 inc_counter(counter_move); 6540 inc_counter(counter_move_total); 6541 6542 LIR_Opr in = op->as_Op1()->in_opr(); 6543 LIR_Opr res = op->as_Op1()->result_opr(); 6544 if (in->is_register()) { 6545 if (res->is_register()) { 6546 inc_counter(counter_move_reg_reg); 6547 } else if (res->is_stack()) { 6548 inc_counter(counter_move_reg_stack); 6549 } else if (res->is_address()) { 6550 inc_counter(counter_move_reg_mem); 6551 } else { 6552 ShouldNotReachHere(); 6553 } 6554 } else if (in->is_stack()) { 6555 if (res->is_register()) { 6556 inc_counter(counter_move_stack_reg); 6557 } else { 6558 inc_counter(counter_move_stack_stack); 6559 } 6560 } else if (in->is_address()) { 6561 assert(res->is_register(), "must be"); 6562 inc_counter(counter_move_mem_reg); 6563 } else if (in->is_constant()) { 6564 inc_counter(counter_move_const_any); 6565 } else { 6566 ShouldNotReachHere(); 6567 } 6568 break; 6569 } 6570 6571 case lir_cmp: inc_counter(counter_cmp); break; 6572 6573 case lir_branch: 6574 case lir_cond_float_branch: { 6575 LIR_OpBranch* branch = op->as_OpBranch(); 6576 if (branch->block() == NULL) { 6577 inc_counter(counter_stub_branch); 6578 } else if (branch->cond() == lir_cond_always) { 6579 inc_counter(counter_uncond_branch); 6580 } else { 6581 inc_counter(counter_cond_branch); 6582 } 6583 break; 6584 } 6585 6586 case lir_neg: 6587 case lir_add: 6588 case lir_sub: 6589 case lir_mul: 6590 case lir_mul_strictfp: 6591 case lir_div: 6592 case lir_div_strictfp: 6593 case lir_rem: 6594 case lir_sqrt: 6595 case lir_sin: 6596 case lir_cos: 6597 case lir_abs: 6598 case lir_log10: 6599 case lir_log: 6600 case lir_pow: 6601 case lir_exp: 6602 case lir_logic_and: 6603 case lir_logic_or: 6604 case lir_logic_xor: 6605 case lir_shl: 6606 case lir_shr: 6607 case lir_ushr: inc_counter(counter_alu); break; 6608 6609 case lir_alloc_object: 6610 case lir_alloc_array: inc_counter(counter_alloc); break; 6611 6612 case lir_monaddr: 6613 case lir_lock: 6614 case lir_unlock: inc_counter(counter_sync); break; 6615 6616 case lir_throw: inc_counter(counter_throw); break; 6617 6618 case lir_unwind: inc_counter(counter_unwind); break; 6619 6620 case lir_null_check: 6621 case lir_leal: 6622 case lir_instanceof: 6623 case lir_checkcast: 6624 case lir_store_check: inc_counter(counter_typecheck); break; 6625 6626 case lir_fpop_raw: 6627 case lir_fxch: 6628 case lir_fld: inc_counter(counter_fpu_stack); break; 6629 6630 case lir_nop: 6631 case lir_push: 6632 case lir_pop: 6633 case lir_convert: 6634 case lir_roundfp: 6635 case lir_cmove: inc_counter(counter_misc_inst); break; 6636 6637 default: inc_counter(counter_other_inst); break; 6638 } 6639 } 6640 } 6641 6642 if (has_xhandlers) { 6643 inc_counter(counter_exception_method); 6644 } 6645 } 6646 6647 void LinearScanStatistic::compute(LinearScan* allocator, LinearScanStatistic &global_statistic) { 6648 if (CountLinearScan || TraceLinearScanLevel > 0) { 6649 6650 LinearScanStatistic local_statistic = LinearScanStatistic(); 6651 6652 local_statistic.collect(allocator); 6653 global_statistic.sum_up(local_statistic); 6654 6655 if (TraceLinearScanLevel > 2) { 6656 local_statistic.print("current local statistic"); 6657 } 6658 } 6659 } 6660 6661 6662 // Implementation of LinearTimers 6663 6664 LinearScanTimers::LinearScanTimers() { 6665 for (int i = 0; i < number_of_timers; i++) { 6666 timer(i)->reset(); 6667 } 6668 } 6669 6670 const char* LinearScanTimers::timer_name(int idx) { 6671 switch (idx) { 6672 case timer_do_nothing: return "Nothing (Time Check)"; 6673 case timer_number_instructions: return "Number Instructions"; 6674 case timer_compute_local_live_sets: return "Local Live Sets"; 6675 case timer_compute_global_live_sets: return "Global Live Sets"; 6676 case timer_build_intervals: return "Build Intervals"; 6677 case timer_sort_intervals_before: return "Sort Intervals Before"; 6678 case timer_allocate_registers: return "Allocate Registers"; 6679 case timer_resolve_data_flow: return "Resolve Data Flow"; 6680 case timer_sort_intervals_after: return "Sort Intervals After"; 6681 case timer_eliminate_spill_moves: return "Spill optimization"; 6682 case timer_assign_reg_num: return "Assign Reg Num"; 6683 case timer_allocate_fpu_stack: return "Allocate FPU Stack"; 6684 case timer_optimize_lir: return "Optimize LIR"; 6685 default: ShouldNotReachHere(); return ""; 6686 } 6687 } 6688 6689 void LinearScanTimers::begin_method() { 6690 if (TimeEachLinearScan) { 6691 // reset all timers to measure only current method 6692 for (int i = 0; i < number_of_timers; i++) { 6693 timer(i)->reset(); 6694 } 6695 } 6696 } 6697 6698 void LinearScanTimers::end_method(LinearScan* allocator) { 6699 if (TimeEachLinearScan) { 6700 6701 double c = timer(timer_do_nothing)->seconds(); 6702 double total = 0; 6703 for (int i = 1; i < number_of_timers; i++) { 6704 total += timer(i)->seconds() - c; 6705 } 6706 6707 if (total >= 0.0005) { 6708 // print all information in one line for automatic processing 6709 tty->print("@"); allocator->compilation()->method()->print_name(); 6710 6711 tty->print("@ %d ", allocator->compilation()->method()->code_size()); 6712 tty->print("@ %d ", allocator->block_at(allocator->block_count() - 1)->last_lir_instruction_id() / 2); 6713 tty->print("@ %d ", allocator->block_count()); 6714 tty->print("@ %d ", allocator->num_virtual_regs()); 6715 tty->print("@ %d ", allocator->interval_count()); 6716 tty->print("@ %d ", allocator->_num_calls); 6717 tty->print("@ %d ", allocator->num_loops()); 6718 6719 tty->print("@ %6.6f ", total); 6720 for (int i = 1; i < number_of_timers; i++) { 6721 tty->print("@ %4.1f ", ((timer(i)->seconds() - c) / total) * 100); 6722 } 6723 tty->cr(); 6724 } 6725 } 6726 } 6727 6728 void LinearScanTimers::print(double total_time) { 6729 if (TimeLinearScan) { 6730 // correction value: sum of dummy-timer that only measures the time that 6731 // is necesary to start and stop itself 6732 double c = timer(timer_do_nothing)->seconds(); 6733 6734 for (int i = 0; i < number_of_timers; i++) { 6735 double t = timer(i)->seconds(); 6736 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); 6737 } 6738 } 6739 } 6740 6741 #endif // #ifndef PRODUCT