1 /* 2 * Copyright (c) 2005, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "c1/c1_CFGPrinter.hpp" 27 #include "c1/c1_CodeStubs.hpp" 28 #include "c1/c1_Compilation.hpp" 29 #include "c1/c1_FrameMap.hpp" 30 #include "c1/c1_IR.hpp" 31 #include "c1/c1_LIRGenerator.hpp" 32 #include "c1/c1_LinearScan.hpp" 33 #include "c1/c1_ValueStack.hpp" 34 #include "utilities/bitMap.inline.hpp" 35 #ifdef TARGET_ARCH_x86 36 # include "vmreg_x86.inline.hpp" 37 #endif 38 #ifdef TARGET_ARCH_sparc 39 # include "vmreg_sparc.inline.hpp" 40 #endif 41 #ifdef TARGET_ARCH_zero 42 # include "vmreg_zero.inline.hpp" 43 #endif 44 #ifdef TARGET_ARCH_arm 45 # include "vmreg_arm.inline.hpp" 46 #endif 47 #ifdef TARGET_ARCH_ppc 48 # include "vmreg_ppc.inline.hpp" 49 #endif 50 51 52 #ifndef PRODUCT 53 54 static LinearScanStatistic _stat_before_alloc; 55 static LinearScanStatistic _stat_after_asign; 56 static LinearScanStatistic _stat_final; 57 58 static LinearScanTimers _total_timer; 59 60 // helper macro for short definition of timer 61 #define TIME_LINEAR_SCAN(timer_name) TraceTime _block_timer("", _total_timer.timer(LinearScanTimers::timer_name), TimeLinearScan || TimeEachLinearScan, Verbose); 62 63 // helper macro for short definition of trace-output inside code 64 #define TRACE_LINEAR_SCAN(level, code) \ 65 if (TraceLinearScanLevel >= level) { \ 66 code; \ 67 } 68 69 #else 70 71 #define TIME_LINEAR_SCAN(timer_name) 72 #define TRACE_LINEAR_SCAN(level, code) 73 74 #endif 75 76 // Map BasicType to spill size in 32-bit words, matching VMReg's notion of words 77 #ifdef _LP64 78 static int type2spill_size[T_CONFLICT+1]={ -1, 0, 0, 0, 1, 1, 1, 2, 1, 1, 1, 2, 2, 2, 0, 2, 1, 2, 1, -1}; 79 #else 80 static int type2spill_size[T_CONFLICT+1]={ -1, 0, 0, 0, 1, 1, 1, 2, 1, 1, 1, 2, 1, 1, 0, 1, -1, 1, 1, -1}; 81 #endif 82 83 84 // Implementation of LinearScan 85 86 LinearScan::LinearScan(IR* ir, LIRGenerator* gen, FrameMap* frame_map) 87 : _compilation(ir->compilation()) 88 , _ir(ir) 89 , _gen(gen) 90 , _frame_map(frame_map) 91 , _num_virtual_regs(gen->max_virtual_register_number()) 92 , _has_fpu_registers(false) 93 , _num_calls(-1) 94 , _max_spills(0) 95 , _unused_spill_slot(-1) 96 , _intervals(0) // initialized later with correct length 97 , _new_intervals_from_allocation(new IntervalList()) 98 , _sorted_intervals(NULL) 99 , _needs_full_resort(false) 100 , _lir_ops(0) // initialized later with correct length 101 , _block_of_op(0) // initialized later with correct length 102 , _has_info(0) 103 , _has_call(0) 104 , _scope_value_cache(0) // initialized later with correct length 105 , _interval_in_loop(0, 0) // initialized later with correct length 106 , _cached_blocks(*ir->linear_scan_order()) 107 #ifdef X86 108 , _fpu_stack_allocator(NULL) 109 #endif 110 { 111 assert(this->ir() != NULL, "check if valid"); 112 assert(this->compilation() != NULL, "check if valid"); 113 assert(this->gen() != NULL, "check if valid"); 114 assert(this->frame_map() != NULL, "check if valid"); 115 } 116 117 118 // ********** functions for converting LIR-Operands to register numbers 119 // 120 // Emulate a flat register file comprising physical integer registers, 121 // physical floating-point registers and virtual registers, in that order. 122 // Virtual registers already have appropriate numbers, since V0 is 123 // the number of physical registers. 124 // Returns -1 for hi word if opr is a single word operand. 125 // 126 // Note: the inverse operation (calculating an operand for register numbers) 127 // is done in calc_operand_for_interval() 128 129 int LinearScan::reg_num(LIR_Opr opr) { 130 assert(opr->is_register(), "should not call this otherwise"); 131 132 if (opr->is_virtual_register()) { 133 assert(opr->vreg_number() >= nof_regs, "found a virtual register with a fixed-register number"); 134 return opr->vreg_number(); 135 } else if (opr->is_single_cpu()) { 136 return opr->cpu_regnr(); 137 } else if (opr->is_double_cpu()) { 138 return opr->cpu_regnrLo(); 139 #ifdef X86 140 } else if (opr->is_single_xmm()) { 141 return opr->fpu_regnr() + pd_first_xmm_reg; 142 } else if (opr->is_double_xmm()) { 143 return opr->fpu_regnrLo() + pd_first_xmm_reg; 144 #endif 145 } else if (opr->is_single_fpu()) { 146 return opr->fpu_regnr() + pd_first_fpu_reg; 147 } else if (opr->is_double_fpu()) { 148 return opr->fpu_regnrLo() + pd_first_fpu_reg; 149 } else { 150 ShouldNotReachHere(); 151 return -1; 152 } 153 } 154 155 int LinearScan::reg_numHi(LIR_Opr opr) { 156 assert(opr->is_register(), "should not call this otherwise"); 157 158 if (opr->is_virtual_register()) { 159 return -1; 160 } else if (opr->is_single_cpu()) { 161 return -1; 162 } else if (opr->is_double_cpu()) { 163 return opr->cpu_regnrHi(); 164 #ifdef X86 165 } else if (opr->is_single_xmm()) { 166 return -1; 167 } else if (opr->is_double_xmm()) { 168 return -1; 169 #endif 170 } else if (opr->is_single_fpu()) { 171 return -1; 172 } else if (opr->is_double_fpu()) { 173 return opr->fpu_regnrHi() + pd_first_fpu_reg; 174 } else { 175 ShouldNotReachHere(); 176 return -1; 177 } 178 } 179 180 181 // ********** functions for classification of intervals 182 183 bool LinearScan::is_precolored_interval(const Interval* i) { 184 return i->reg_num() < LinearScan::nof_regs; 185 } 186 187 bool LinearScan::is_virtual_interval(const Interval* i) { 188 return i->reg_num() >= LIR_OprDesc::vreg_base; 189 } 190 191 bool LinearScan::is_precolored_cpu_interval(const Interval* i) { 192 return i->reg_num() < LinearScan::nof_cpu_regs; 193 } 194 195 bool LinearScan::is_virtual_cpu_interval(const Interval* i) { 196 #if defined(__SOFTFP__) || defined(E500V2) 197 return i->reg_num() >= LIR_OprDesc::vreg_base; 198 #else 199 return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() != T_FLOAT && i->type() != T_DOUBLE); 200 #endif // __SOFTFP__ or E500V2 201 } 202 203 bool LinearScan::is_precolored_fpu_interval(const Interval* i) { 204 return i->reg_num() >= LinearScan::nof_cpu_regs && i->reg_num() < LinearScan::nof_regs; 205 } 206 207 bool LinearScan::is_virtual_fpu_interval(const Interval* i) { 208 #if defined(__SOFTFP__) || defined(E500V2) 209 return false; 210 #else 211 return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() == T_FLOAT || i->type() == T_DOUBLE); 212 #endif // __SOFTFP__ or E500V2 213 } 214 215 bool LinearScan::is_in_fpu_register(const Interval* i) { 216 // fixed intervals not needed for FPU stack allocation 217 return i->reg_num() >= nof_regs && pd_first_fpu_reg <= i->assigned_reg() && i->assigned_reg() <= pd_last_fpu_reg; 218 } 219 220 bool LinearScan::is_oop_interval(const Interval* i) { 221 // fixed intervals never contain oops 222 return i->reg_num() >= nof_regs && i->type() == T_OBJECT; 223 } 224 225 226 // ********** General helper functions 227 228 // compute next unused stack index that can be used for spilling 229 int LinearScan::allocate_spill_slot(bool double_word) { 230 int spill_slot; 231 if (double_word) { 232 if ((_max_spills & 1) == 1) { 233 // alignment of double-word values 234 // the hole because of the alignment is filled with the next single-word value 235 assert(_unused_spill_slot == -1, "wasting a spill slot"); 236 _unused_spill_slot = _max_spills; 237 _max_spills++; 238 } 239 spill_slot = _max_spills; 240 _max_spills += 2; 241 242 } else if (_unused_spill_slot != -1) { 243 // re-use hole that was the result of a previous double-word alignment 244 spill_slot = _unused_spill_slot; 245 _unused_spill_slot = -1; 246 247 } else { 248 spill_slot = _max_spills; 249 _max_spills++; 250 } 251 252 int result = spill_slot + LinearScan::nof_regs + frame_map()->argcount(); 253 254 // the class OopMapValue uses only 11 bits for storing the name of the 255 // oop location. So a stack slot bigger than 2^11 leads to an overflow 256 // that is not reported in product builds. Prevent this by checking the 257 // spill slot here (altough this value and the later used location name 258 // are slightly different) 259 if (result > 2000) { 260 bailout("too many stack slots used"); 261 } 262 263 return result; 264 } 265 266 void LinearScan::assign_spill_slot(Interval* it) { 267 // assign the canonical spill slot of the parent (if a part of the interval 268 // is already spilled) or allocate a new spill slot 269 if (it->canonical_spill_slot() >= 0) { 270 it->assign_reg(it->canonical_spill_slot()); 271 } else { 272 int spill = allocate_spill_slot(type2spill_size[it->type()] == 2); 273 it->set_canonical_spill_slot(spill); 274 it->assign_reg(spill); 275 } 276 } 277 278 void LinearScan::propagate_spill_slots() { 279 if (!frame_map()->finalize_frame(max_spills())) { 280 bailout("frame too large"); 281 } 282 } 283 284 // create a new interval with a predefined reg_num 285 // (only used for parent intervals that are created during the building phase) 286 Interval* LinearScan::create_interval(int reg_num) { 287 assert(_intervals.at(reg_num) == NULL, "overwriting exisiting interval"); 288 289 Interval* interval = new Interval(reg_num); 290 _intervals.at_put(reg_num, interval); 291 292 // assign register number for precolored intervals 293 if (reg_num < LIR_OprDesc::vreg_base) { 294 interval->assign_reg(reg_num); 295 } 296 return interval; 297 } 298 299 // assign a new reg_num to the interval and append it to the list of intervals 300 // (only used for child intervals that are created during register allocation) 301 void LinearScan::append_interval(Interval* it) { 302 it->set_reg_num(_intervals.length()); 303 _intervals.append(it); 304 _new_intervals_from_allocation->append(it); 305 } 306 307 // copy the vreg-flags if an interval is split 308 void LinearScan::copy_register_flags(Interval* from, Interval* to) { 309 if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::byte_reg)) { 310 gen()->set_vreg_flag(to->reg_num(), LIRGenerator::byte_reg); 311 } 312 if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::callee_saved)) { 313 gen()->set_vreg_flag(to->reg_num(), LIRGenerator::callee_saved); 314 } 315 316 // Note: do not copy the must_start_in_memory flag because it is not necessary for child 317 // intervals (only the very beginning of the interval must be in memory) 318 } 319 320 321 // ********** spill move optimization 322 // eliminate moves from register to stack if stack slot is known to be correct 323 324 // called during building of intervals 325 void LinearScan::change_spill_definition_pos(Interval* interval, int def_pos) { 326 assert(interval->is_split_parent(), "can only be called for split parents"); 327 328 switch (interval->spill_state()) { 329 case noDefinitionFound: 330 assert(interval->spill_definition_pos() == -1, "must no be set before"); 331 interval->set_spill_definition_pos(def_pos); 332 interval->set_spill_state(oneDefinitionFound); 333 break; 334 335 case oneDefinitionFound: 336 assert(def_pos <= interval->spill_definition_pos(), "positions are processed in reverse order when intervals are created"); 337 if (def_pos < interval->spill_definition_pos() - 2) { 338 // second definition found, so no spill optimization possible for this interval 339 interval->set_spill_state(noOptimization); 340 } else { 341 // two consecutive definitions (because of two-operand LIR form) 342 assert(block_of_op_with_id(def_pos) == block_of_op_with_id(interval->spill_definition_pos()), "block must be equal"); 343 } 344 break; 345 346 case noOptimization: 347 // nothing to do 348 break; 349 350 default: 351 assert(false, "other states not allowed at this time"); 352 } 353 } 354 355 // called during register allocation 356 void LinearScan::change_spill_state(Interval* interval, int spill_pos) { 357 switch (interval->spill_state()) { 358 case oneDefinitionFound: { 359 int def_loop_depth = block_of_op_with_id(interval->spill_definition_pos())->loop_depth(); 360 int spill_loop_depth = block_of_op_with_id(spill_pos)->loop_depth(); 361 362 if (def_loop_depth < spill_loop_depth) { 363 // the loop depth of the spilling position is higher then the loop depth 364 // at the definition of the interval -> move write to memory out of loop 365 // by storing at definitin of the interval 366 interval->set_spill_state(storeAtDefinition); 367 } else { 368 // the interval is currently spilled only once, so for now there is no 369 // reason to store the interval at the definition 370 interval->set_spill_state(oneMoveInserted); 371 } 372 break; 373 } 374 375 case oneMoveInserted: { 376 // the interval is spilled more then once, so it is better to store it to 377 // memory at the definition 378 interval->set_spill_state(storeAtDefinition); 379 break; 380 } 381 382 case storeAtDefinition: 383 case startInMemory: 384 case noOptimization: 385 case noDefinitionFound: 386 // nothing to do 387 break; 388 389 default: 390 assert(false, "other states not allowed at this time"); 391 } 392 } 393 394 395 bool LinearScan::must_store_at_definition(const Interval* i) { 396 return i->is_split_parent() && i->spill_state() == storeAtDefinition; 397 } 398 399 // called once before asignment of register numbers 400 void LinearScan::eliminate_spill_moves() { 401 TIME_LINEAR_SCAN(timer_eliminate_spill_moves); 402 TRACE_LINEAR_SCAN(3, tty->print_cr("***** Eliminating unnecessary spill moves")); 403 404 // collect all intervals that must be stored after their definion. 405 // the list is sorted by Interval::spill_definition_pos 406 Interval* interval; 407 Interval* temp_list; 408 create_unhandled_lists(&interval, &temp_list, must_store_at_definition, NULL); 409 410 #ifdef ASSERT 411 Interval* prev = NULL; 412 Interval* temp = interval; 413 while (temp != Interval::end()) { 414 assert(temp->spill_definition_pos() > 0, "invalid spill definition pos"); 415 if (prev != NULL) { 416 assert(temp->from() >= prev->from(), "intervals not sorted"); 417 assert(temp->spill_definition_pos() >= prev->spill_definition_pos(), "when intervals are sorted by from, then they must also be sorted by spill_definition_pos"); 418 } 419 420 assert(temp->canonical_spill_slot() >= LinearScan::nof_regs, "interval has no spill slot assigned"); 421 assert(temp->spill_definition_pos() >= temp->from(), "invalid order"); 422 assert(temp->spill_definition_pos() <= temp->from() + 2, "only intervals defined once at their start-pos can be optimized"); 423 424 TRACE_LINEAR_SCAN(4, tty->print_cr("interval %d (from %d to %d) must be stored at %d", temp->reg_num(), temp->from(), temp->to(), temp->spill_definition_pos())); 425 426 temp = temp->next(); 427 } 428 #endif 429 430 LIR_InsertionBuffer insertion_buffer; 431 int num_blocks = block_count(); 432 for (int i = 0; i < num_blocks; i++) { 433 BlockBegin* block = block_at(i); 434 LIR_OpList* instructions = block->lir()->instructions_list(); 435 int num_inst = instructions->length(); 436 bool has_new = false; 437 438 // iterate all instructions of the block. skip the first because it is always a label 439 for (int j = 1; j < num_inst; j++) { 440 LIR_Op* op = instructions->at(j); 441 int op_id = op->id(); 442 443 if (op_id == -1) { 444 // remove move from register to stack if the stack slot is guaranteed to be correct. 445 // only moves that have been inserted by LinearScan can be removed. 446 assert(op->code() == lir_move, "only moves can have a op_id of -1"); 447 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); 448 assert(op->as_Op1()->result_opr()->is_virtual(), "LinearScan inserts only moves to virtual registers"); 449 450 LIR_Op1* op1 = (LIR_Op1*)op; 451 Interval* interval = interval_at(op1->result_opr()->vreg_number()); 452 453 if (interval->assigned_reg() >= LinearScan::nof_regs && interval->always_in_memory()) { 454 // move target is a stack slot that is always correct, so eliminate instruction 455 TRACE_LINEAR_SCAN(4, tty->print_cr("eliminating move from interval %d to %d", op1->in_opr()->vreg_number(), op1->result_opr()->vreg_number())); 456 instructions->at_put(j, NULL); // NULL-instructions are deleted by assign_reg_num 457 } 458 459 } else { 460 // insert move from register to stack just after the beginning of the interval 461 assert(interval == Interval::end() || interval->spill_definition_pos() >= op_id, "invalid order"); 462 assert(interval == Interval::end() || (interval->is_split_parent() && interval->spill_state() == storeAtDefinition), "invalid interval"); 463 464 while (interval != Interval::end() && interval->spill_definition_pos() == op_id) { 465 if (!has_new) { 466 // prepare insertion buffer (appended when all instructions of the block are processed) 467 insertion_buffer.init(block->lir()); 468 has_new = true; 469 } 470 471 LIR_Opr from_opr = operand_for_interval(interval); 472 LIR_Opr to_opr = canonical_spill_opr(interval); 473 assert(from_opr->is_fixed_cpu() || from_opr->is_fixed_fpu(), "from operand must be a register"); 474 assert(to_opr->is_stack(), "to operand must be a stack slot"); 475 476 insertion_buffer.move(j, from_opr, to_opr); 477 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting move after definition of interval %d to stack slot %d at op_id %d", interval->reg_num(), interval->canonical_spill_slot() - LinearScan::nof_regs, op_id)); 478 479 interval = interval->next(); 480 } 481 } 482 } // end of instruction iteration 483 484 if (has_new) { 485 block->lir()->append(&insertion_buffer); 486 } 487 } // end of block iteration 488 489 assert(interval == Interval::end(), "missed an interval"); 490 } 491 492 493 // ********** Phase 1: number all instructions in all blocks 494 // Compute depth-first and linear scan block orders, and number LIR_Op nodes for linear scan. 495 496 void LinearScan::number_instructions() { 497 { 498 // dummy-timer to measure the cost of the timer itself 499 // (this time is then subtracted from all other timers to get the real value) 500 TIME_LINEAR_SCAN(timer_do_nothing); 501 } 502 TIME_LINEAR_SCAN(timer_number_instructions); 503 504 // Assign IDs to LIR nodes and build a mapping, lir_ops, from ID to LIR_Op node. 505 int num_blocks = block_count(); 506 int num_instructions = 0; 507 int i; 508 for (i = 0; i < num_blocks; i++) { 509 num_instructions += block_at(i)->lir()->instructions_list()->length(); 510 } 511 512 // initialize with correct length 513 _lir_ops = LIR_OpArray(num_instructions); 514 _block_of_op = BlockBeginArray(num_instructions); 515 516 int op_id = 0; 517 int idx = 0; 518 519 for (i = 0; i < num_blocks; i++) { 520 BlockBegin* block = block_at(i); 521 block->set_first_lir_instruction_id(op_id); 522 LIR_OpList* instructions = block->lir()->instructions_list(); 523 524 int num_inst = instructions->length(); 525 for (int j = 0; j < num_inst; j++) { 526 LIR_Op* op = instructions->at(j); 527 op->set_id(op_id); 528 529 _lir_ops.at_put(idx, op); 530 _block_of_op.at_put(idx, block); 531 assert(lir_op_with_id(op_id) == op, "must match"); 532 533 idx++; 534 op_id += 2; // numbering of lir_ops by two 535 } 536 block->set_last_lir_instruction_id(op_id - 2); 537 } 538 assert(idx == num_instructions, "must match"); 539 assert(idx * 2 == op_id, "must match"); 540 541 _has_call = BitMap(num_instructions); _has_call.clear(); 542 _has_info = BitMap(num_instructions); _has_info.clear(); 543 } 544 545 546 // ********** Phase 2: compute local live sets separately for each block 547 // (sets live_gen and live_kill for each block) 548 549 void LinearScan::set_live_gen_kill(Value value, LIR_Op* op, BitMap& live_gen, BitMap& live_kill) { 550 LIR_Opr opr = value->operand(); 551 Constant* con = value->as_Constant(); 552 553 // check some asumptions about debug information 554 assert(!value->type()->is_illegal(), "if this local is used by the interpreter it shouldn't be of indeterminate type"); 555 assert(con == NULL || opr->is_virtual() || opr->is_constant() || opr->is_illegal(), "asumption: Constant instructions have only constant operands"); 556 assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands"); 557 558 if ((con == NULL || con->is_pinned()) && opr->is_register()) { 559 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 560 int reg = opr->vreg_number(); 561 if (!live_kill.at(reg)) { 562 live_gen.set_bit(reg); 563 TRACE_LINEAR_SCAN(4, tty->print_cr(" Setting live_gen for value %c%d, LIR op_id %d, register number %d", value->type()->tchar(), value->id(), op->id(), reg)); 564 } 565 } 566 } 567 568 569 void LinearScan::compute_local_live_sets() { 570 TIME_LINEAR_SCAN(timer_compute_local_live_sets); 571 572 int num_blocks = block_count(); 573 int live_size = live_set_size(); 574 bool local_has_fpu_registers = false; 575 int local_num_calls = 0; 576 LIR_OpVisitState visitor; 577 578 BitMap2D local_interval_in_loop = BitMap2D(_num_virtual_regs, num_loops()); 579 local_interval_in_loop.clear(); 580 581 // iterate all blocks 582 for (int i = 0; i < num_blocks; i++) { 583 BlockBegin* block = block_at(i); 584 585 BitMap live_gen(live_size); live_gen.clear(); 586 BitMap live_kill(live_size); live_kill.clear(); 587 588 if (block->is_set(BlockBegin::exception_entry_flag)) { 589 // Phi functions at the begin of an exception handler are 590 // implicitly defined (= killed) at the beginning of the block. 591 for_each_phi_fun(block, phi, 592 live_kill.set_bit(phi->operand()->vreg_number()) 593 ); 594 } 595 596 LIR_OpList* instructions = block->lir()->instructions_list(); 597 int num_inst = instructions->length(); 598 599 // iterate all instructions of the block. skip the first because it is always a label 600 assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label"); 601 for (int j = 1; j < num_inst; j++) { 602 LIR_Op* op = instructions->at(j); 603 604 // visit operation to collect all operands 605 visitor.visit(op); 606 607 if (visitor.has_call()) { 608 _has_call.set_bit(op->id() >> 1); 609 local_num_calls++; 610 } 611 if (visitor.info_count() > 0) { 612 _has_info.set_bit(op->id() >> 1); 613 } 614 615 // iterate input operands of instruction 616 int k, n, reg; 617 n = visitor.opr_count(LIR_OpVisitState::inputMode); 618 for (k = 0; k < n; k++) { 619 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k); 620 assert(opr->is_register(), "visitor should only return register operands"); 621 622 if (opr->is_virtual_register()) { 623 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 624 reg = opr->vreg_number(); 625 if (!live_kill.at(reg)) { 626 live_gen.set_bit(reg); 627 TRACE_LINEAR_SCAN(4, tty->print_cr(" Setting live_gen for register %d at instruction %d", reg, op->id())); 628 } 629 if (block->loop_index() >= 0) { 630 local_interval_in_loop.set_bit(reg, block->loop_index()); 631 } 632 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); 633 } 634 635 #ifdef ASSERT 636 // fixed intervals are never live at block boundaries, so 637 // they need not be processed in live sets. 638 // this is checked by these assertions to be sure about it. 639 // the entry block may have incoming values in registers, which is ok. 640 if (!opr->is_virtual_register() && block != ir()->start()) { 641 reg = reg_num(opr); 642 if (is_processed_reg_num(reg)) { 643 assert(live_kill.at(reg), "using fixed register that is not defined in this block"); 644 } 645 reg = reg_numHi(opr); 646 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 647 assert(live_kill.at(reg), "using fixed register that is not defined in this block"); 648 } 649 } 650 #endif 651 } 652 653 // Add uses of live locals from interpreter's point of view for proper debug information generation 654 n = visitor.info_count(); 655 for (k = 0; k < n; k++) { 656 CodeEmitInfo* info = visitor.info_at(k); 657 ValueStack* stack = info->stack(); 658 for_each_state_value(stack, value, 659 set_live_gen_kill(value, op, live_gen, live_kill) 660 ); 661 } 662 663 // iterate temp operands of instruction 664 n = visitor.opr_count(LIR_OpVisitState::tempMode); 665 for (k = 0; k < n; k++) { 666 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k); 667 assert(opr->is_register(), "visitor should only return register operands"); 668 669 if (opr->is_virtual_register()) { 670 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 671 reg = opr->vreg_number(); 672 live_kill.set_bit(reg); 673 if (block->loop_index() >= 0) { 674 local_interval_in_loop.set_bit(reg, block->loop_index()); 675 } 676 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); 677 } 678 679 #ifdef ASSERT 680 // fixed intervals are never live at block boundaries, so 681 // they need not be processed in live sets 682 // process them only in debug mode so that this can be checked 683 if (!opr->is_virtual_register()) { 684 reg = reg_num(opr); 685 if (is_processed_reg_num(reg)) { 686 live_kill.set_bit(reg_num(opr)); 687 } 688 reg = reg_numHi(opr); 689 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 690 live_kill.set_bit(reg); 691 } 692 } 693 #endif 694 } 695 696 // iterate output operands of instruction 697 n = visitor.opr_count(LIR_OpVisitState::outputMode); 698 for (k = 0; k < n; k++) { 699 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k); 700 assert(opr->is_register(), "visitor should only return register operands"); 701 702 if (opr->is_virtual_register()) { 703 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 704 reg = opr->vreg_number(); 705 live_kill.set_bit(reg); 706 if (block->loop_index() >= 0) { 707 local_interval_in_loop.set_bit(reg, block->loop_index()); 708 } 709 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); 710 } 711 712 #ifdef ASSERT 713 // fixed intervals are never live at block boundaries, so 714 // they need not be processed in live sets 715 // process them only in debug mode so that this can be checked 716 if (!opr->is_virtual_register()) { 717 reg = reg_num(opr); 718 if (is_processed_reg_num(reg)) { 719 live_kill.set_bit(reg_num(opr)); 720 } 721 reg = reg_numHi(opr); 722 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 723 live_kill.set_bit(reg); 724 } 725 } 726 #endif 727 } 728 } // end of instruction iteration 729 730 block->set_live_gen (live_gen); 731 block->set_live_kill(live_kill); 732 block->set_live_in (BitMap(live_size)); block->live_in().clear(); 733 block->set_live_out (BitMap(live_size)); block->live_out().clear(); 734 735 TRACE_LINEAR_SCAN(4, tty->print("live_gen B%d ", block->block_id()); print_bitmap(block->live_gen())); 736 TRACE_LINEAR_SCAN(4, tty->print("live_kill B%d ", block->block_id()); print_bitmap(block->live_kill())); 737 } // end of block iteration 738 739 // propagate local calculated information into LinearScan object 740 _has_fpu_registers = local_has_fpu_registers; 741 compilation()->set_has_fpu_code(local_has_fpu_registers); 742 743 _num_calls = local_num_calls; 744 _interval_in_loop = local_interval_in_loop; 745 } 746 747 748 // ********** Phase 3: perform a backward dataflow analysis to compute global live sets 749 // (sets live_in and live_out for each block) 750 751 void LinearScan::compute_global_live_sets() { 752 TIME_LINEAR_SCAN(timer_compute_global_live_sets); 753 754 int num_blocks = block_count(); 755 bool change_occurred; 756 bool change_occurred_in_block; 757 int iteration_count = 0; 758 BitMap live_out(live_set_size()); live_out.clear(); // scratch set for calculations 759 760 // Perform a backward dataflow analysis to compute live_out and live_in for each block. 761 // The loop is executed until a fixpoint is reached (no changes in an iteration) 762 // Exception handlers must be processed because not all live values are 763 // present in the state array, e.g. because of global value numbering 764 do { 765 change_occurred = false; 766 767 // iterate all blocks in reverse order 768 for (int i = num_blocks - 1; i >= 0; i--) { 769 BlockBegin* block = block_at(i); 770 771 change_occurred_in_block = false; 772 773 // live_out(block) is the union of live_in(sux), for successors sux of block 774 int n = block->number_of_sux(); 775 int e = block->number_of_exception_handlers(); 776 if (n + e > 0) { 777 // block has successors 778 if (n > 0) { 779 live_out.set_from(block->sux_at(0)->live_in()); 780 for (int j = 1; j < n; j++) { 781 live_out.set_union(block->sux_at(j)->live_in()); 782 } 783 } else { 784 live_out.clear(); 785 } 786 for (int j = 0; j < e; j++) { 787 live_out.set_union(block->exception_handler_at(j)->live_in()); 788 } 789 790 if (!block->live_out().is_same(live_out)) { 791 // A change occurred. Swap the old and new live out sets to avoid copying. 792 BitMap temp = block->live_out(); 793 block->set_live_out(live_out); 794 live_out = temp; 795 796 change_occurred = true; 797 change_occurred_in_block = true; 798 } 799 } 800 801 if (iteration_count == 0 || change_occurred_in_block) { 802 // live_in(block) is the union of live_gen(block) with (live_out(block) & !live_kill(block)) 803 // note: live_in has to be computed only in first iteration or if live_out has changed! 804 BitMap live_in = block->live_in(); 805 live_in.set_from(block->live_out()); 806 live_in.set_difference(block->live_kill()); 807 live_in.set_union(block->live_gen()); 808 } 809 810 #ifndef PRODUCT 811 if (TraceLinearScanLevel >= 4) { 812 char c = ' '; 813 if (iteration_count == 0 || change_occurred_in_block) { 814 c = '*'; 815 } 816 tty->print("(%d) live_in%c B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_in()); 817 tty->print("(%d) live_out%c B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_out()); 818 } 819 #endif 820 } 821 iteration_count++; 822 823 if (change_occurred && iteration_count > 50) { 824 BAILOUT("too many iterations in compute_global_live_sets"); 825 } 826 } while (change_occurred); 827 828 829 #ifdef ASSERT 830 // check that fixed intervals are not live at block boundaries 831 // (live set must be empty at fixed intervals) 832 for (int i = 0; i < num_blocks; i++) { 833 BlockBegin* block = block_at(i); 834 for (int j = 0; j < LIR_OprDesc::vreg_base; j++) { 835 assert(block->live_in().at(j) == false, "live_in set of fixed register must be empty"); 836 assert(block->live_out().at(j) == false, "live_out set of fixed register must be empty"); 837 assert(block->live_gen().at(j) == false, "live_gen set of fixed register must be empty"); 838 } 839 } 840 #endif 841 842 // check that the live_in set of the first block is empty 843 BitMap live_in_args(ir()->start()->live_in().size()); 844 live_in_args.clear(); 845 if (!ir()->start()->live_in().is_same(live_in_args)) { 846 #ifdef ASSERT 847 tty->print_cr("Error: live_in set of first block must be empty (when this fails, virtual registers are used before they are defined)"); 848 tty->print_cr("affected registers:"); 849 print_bitmap(ir()->start()->live_in()); 850 851 // print some additional information to simplify debugging 852 for (unsigned int i = 0; i < ir()->start()->live_in().size(); i++) { 853 if (ir()->start()->live_in().at(i)) { 854 Instruction* instr = gen()->instruction_for_vreg(i); 855 tty->print_cr("* vreg %d (HIR instruction %c%d)", i, instr == NULL ? ' ' : instr->type()->tchar(), instr == NULL ? 0 : instr->id()); 856 857 for (int j = 0; j < num_blocks; j++) { 858 BlockBegin* block = block_at(j); 859 if (block->live_gen().at(i)) { 860 tty->print_cr(" used in block B%d", block->block_id()); 861 } 862 if (block->live_kill().at(i)) { 863 tty->print_cr(" defined in block B%d", block->block_id()); 864 } 865 } 866 } 867 } 868 869 #endif 870 // when this fails, virtual registers are used before they are defined. 871 assert(false, "live_in set of first block must be empty"); 872 // bailout of if this occurs in product mode. 873 bailout("live_in set of first block not empty"); 874 } 875 } 876 877 878 // ********** Phase 4: build intervals 879 // (fills the list _intervals) 880 881 void LinearScan::add_use(Value value, int from, int to, IntervalUseKind use_kind) { 882 assert(!value->type()->is_illegal(), "if this value is used by the interpreter it shouldn't be of indeterminate type"); 883 LIR_Opr opr = value->operand(); 884 Constant* con = value->as_Constant(); 885 886 if ((con == NULL || con->is_pinned()) && opr->is_register()) { 887 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 888 add_use(opr, from, to, use_kind); 889 } 890 } 891 892 893 void LinearScan::add_def(LIR_Opr opr, int def_pos, IntervalUseKind use_kind) { 894 TRACE_LINEAR_SCAN(2, tty->print(" def "); opr->print(tty); tty->print_cr(" def_pos %d (%d)", def_pos, use_kind)); 895 assert(opr->is_register(), "should not be called otherwise"); 896 897 if (opr->is_virtual_register()) { 898 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 899 add_def(opr->vreg_number(), def_pos, use_kind, opr->type_register()); 900 901 } else { 902 int reg = reg_num(opr); 903 if (is_processed_reg_num(reg)) { 904 add_def(reg, def_pos, use_kind, opr->type_register()); 905 } 906 reg = reg_numHi(opr); 907 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 908 add_def(reg, def_pos, use_kind, opr->type_register()); 909 } 910 } 911 } 912 913 void LinearScan::add_use(LIR_Opr opr, int from, int to, IntervalUseKind use_kind) { 914 TRACE_LINEAR_SCAN(2, tty->print(" use "); opr->print(tty); tty->print_cr(" from %d to %d (%d)", from, to, use_kind)); 915 assert(opr->is_register(), "should not be called otherwise"); 916 917 if (opr->is_virtual_register()) { 918 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 919 add_use(opr->vreg_number(), from, to, use_kind, opr->type_register()); 920 921 } else { 922 int reg = reg_num(opr); 923 if (is_processed_reg_num(reg)) { 924 add_use(reg, from, to, use_kind, opr->type_register()); 925 } 926 reg = reg_numHi(opr); 927 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 928 add_use(reg, from, to, use_kind, opr->type_register()); 929 } 930 } 931 } 932 933 void LinearScan::add_temp(LIR_Opr opr, int temp_pos, IntervalUseKind use_kind) { 934 TRACE_LINEAR_SCAN(2, tty->print(" temp "); opr->print(tty); tty->print_cr(" temp_pos %d (%d)", temp_pos, use_kind)); 935 assert(opr->is_register(), "should not be called otherwise"); 936 937 if (opr->is_virtual_register()) { 938 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); 939 add_temp(opr->vreg_number(), temp_pos, use_kind, opr->type_register()); 940 941 } else { 942 int reg = reg_num(opr); 943 if (is_processed_reg_num(reg)) { 944 add_temp(reg, temp_pos, use_kind, opr->type_register()); 945 } 946 reg = reg_numHi(opr); 947 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { 948 add_temp(reg, temp_pos, use_kind, opr->type_register()); 949 } 950 } 951 } 952 953 954 void LinearScan::add_def(int reg_num, int def_pos, IntervalUseKind use_kind, BasicType type) { 955 Interval* interval = interval_at(reg_num); 956 if (interval != NULL) { 957 assert(interval->reg_num() == reg_num, "wrong interval"); 958 959 if (type != T_ILLEGAL) { 960 interval->set_type(type); 961 } 962 963 Range* r = interval->first(); 964 if (r->from() <= def_pos) { 965 // Update the starting point (when a range is first created for a use, its 966 // start is the beginning of the current block until a def is encountered.) 967 r->set_from(def_pos); 968 interval->add_use_pos(def_pos, use_kind); 969 970 } else { 971 // Dead value - make vacuous interval 972 // also add use_kind for dead intervals 973 interval->add_range(def_pos, def_pos + 1); 974 interval->add_use_pos(def_pos, use_kind); 975 TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: def of reg %d at %d occurs without use", reg_num, def_pos)); 976 } 977 978 } else { 979 // Dead value - make vacuous interval 980 // also add use_kind for dead intervals 981 interval = create_interval(reg_num); 982 if (type != T_ILLEGAL) { 983 interval->set_type(type); 984 } 985 986 interval->add_range(def_pos, def_pos + 1); 987 interval->add_use_pos(def_pos, use_kind); 988 TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: dead value %d at %d in live intervals", reg_num, def_pos)); 989 } 990 991 change_spill_definition_pos(interval, def_pos); 992 if (use_kind == noUse && interval->spill_state() <= startInMemory) { 993 // detection of method-parameters and roundfp-results 994 // TODO: move this directly to position where use-kind is computed 995 interval->set_spill_state(startInMemory); 996 } 997 } 998 999 void LinearScan::add_use(int reg_num, int from, int to, IntervalUseKind use_kind, BasicType type) { 1000 Interval* interval = interval_at(reg_num); 1001 if (interval == NULL) { 1002 interval = create_interval(reg_num); 1003 } 1004 assert(interval->reg_num() == reg_num, "wrong interval"); 1005 1006 if (type != T_ILLEGAL) { 1007 interval->set_type(type); 1008 } 1009 1010 interval->add_range(from, to); 1011 interval->add_use_pos(to, use_kind); 1012 } 1013 1014 void LinearScan::add_temp(int reg_num, int temp_pos, IntervalUseKind use_kind, BasicType type) { 1015 Interval* interval = interval_at(reg_num); 1016 if (interval == NULL) { 1017 interval = create_interval(reg_num); 1018 } 1019 assert(interval->reg_num() == reg_num, "wrong interval"); 1020 1021 if (type != T_ILLEGAL) { 1022 interval->set_type(type); 1023 } 1024 1025 interval->add_range(temp_pos, temp_pos + 1); 1026 interval->add_use_pos(temp_pos, use_kind); 1027 } 1028 1029 1030 // the results of this functions are used for optimizing spilling and reloading 1031 // if the functions return shouldHaveRegister and the interval is spilled, 1032 // it is not reloaded to a register. 1033 IntervalUseKind LinearScan::use_kind_of_output_operand(LIR_Op* op, LIR_Opr opr) { 1034 if (op->code() == lir_move) { 1035 assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1"); 1036 LIR_Op1* move = (LIR_Op1*)op; 1037 LIR_Opr res = move->result_opr(); 1038 bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory); 1039 1040 if (result_in_memory) { 1041 // Begin of an interval with must_start_in_memory set. 1042 // This interval will always get a stack slot first, so return noUse. 1043 return noUse; 1044 1045 } else if (move->in_opr()->is_stack()) { 1046 // method argument (condition must be equal to handle_method_arguments) 1047 return noUse; 1048 1049 } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) { 1050 // Move from register to register 1051 if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) { 1052 // special handling of phi-function moves inside osr-entry blocks 1053 // input operand must have a register instead of output operand (leads to better register allocation) 1054 return shouldHaveRegister; 1055 } 1056 } 1057 } 1058 1059 if (opr->is_virtual() && 1060 gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::must_start_in_memory)) { 1061 // result is a stack-slot, so prevent immediate reloading 1062 return noUse; 1063 } 1064 1065 // all other operands require a register 1066 return mustHaveRegister; 1067 } 1068 1069 IntervalUseKind LinearScan::use_kind_of_input_operand(LIR_Op* op, LIR_Opr opr) { 1070 if (op->code() == lir_move) { 1071 assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1"); 1072 LIR_Op1* move = (LIR_Op1*)op; 1073 LIR_Opr res = move->result_opr(); 1074 bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory); 1075 1076 if (result_in_memory) { 1077 // Move to an interval with must_start_in_memory set. 1078 // To avoid moves from stack to stack (not allowed) force the input operand to a register 1079 return mustHaveRegister; 1080 1081 } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) { 1082 // Move from register to register 1083 if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) { 1084 // special handling of phi-function moves inside osr-entry blocks 1085 // input operand must have a register instead of output operand (leads to better register allocation) 1086 return mustHaveRegister; 1087 } 1088 1089 // The input operand is not forced to a register (moves from stack to register are allowed), 1090 // but it is faster if the input operand is in a register 1091 return shouldHaveRegister; 1092 } 1093 } 1094 1095 1096 #ifdef X86 1097 if (op->code() == lir_cmove) { 1098 // conditional moves can handle stack operands 1099 assert(op->result_opr()->is_register(), "result must always be in a register"); 1100 return shouldHaveRegister; 1101 } 1102 1103 // optimizations for second input operand of arithmehtic operations on Intel 1104 // this operand is allowed to be on the stack in some cases 1105 BasicType opr_type = opr->type_register(); 1106 if (opr_type == T_FLOAT || opr_type == T_DOUBLE) { 1107 if ((UseSSE == 1 && opr_type == T_FLOAT) || UseSSE >= 2) { 1108 // SSE float instruction (T_DOUBLE only supported with SSE2) 1109 switch (op->code()) { 1110 case lir_cmp: 1111 case lir_add: 1112 case lir_sub: 1113 case lir_mul: 1114 case lir_div: 1115 { 1116 assert(op->as_Op2() != NULL, "must be LIR_Op2"); 1117 LIR_Op2* op2 = (LIR_Op2*)op; 1118 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { 1119 assert((op2->result_opr()->is_register() || op->code() == lir_cmp) && op2->in_opr1()->is_register(), "cannot mark second operand as stack if others are not in register"); 1120 return shouldHaveRegister; 1121 } 1122 } 1123 } 1124 } else { 1125 // FPU stack float instruction 1126 switch (op->code()) { 1127 case lir_add: 1128 case lir_sub: 1129 case lir_mul: 1130 case lir_div: 1131 { 1132 assert(op->as_Op2() != NULL, "must be LIR_Op2"); 1133 LIR_Op2* op2 = (LIR_Op2*)op; 1134 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { 1135 assert((op2->result_opr()->is_register() || op->code() == lir_cmp) && op2->in_opr1()->is_register(), "cannot mark second operand as stack if others are not in register"); 1136 return shouldHaveRegister; 1137 } 1138 } 1139 } 1140 } 1141 // We want to sometimes use logical operations on pointers, in particular in GC barriers. 1142 // Since 64bit logical operations do not current support operands on stack, we have to make sure 1143 // T_OBJECT doesn't get spilled along with T_LONG. 1144 } else if (opr_type != T_LONG LP64_ONLY(&& opr_type != T_OBJECT)) { 1145 // integer instruction (note: long operands must always be in register) 1146 switch (op->code()) { 1147 case lir_cmp: 1148 case lir_add: 1149 case lir_sub: 1150 case lir_logic_and: 1151 case lir_logic_or: 1152 case lir_logic_xor: 1153 { 1154 assert(op->as_Op2() != NULL, "must be LIR_Op2"); 1155 LIR_Op2* op2 = (LIR_Op2*)op; 1156 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { 1157 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"); 1158 return shouldHaveRegister; 1159 } 1160 } 1161 } 1162 } 1163 #endif // X86 1164 1165 // all other operands require a register 1166 return mustHaveRegister; 1167 } 1168 1169 1170 void LinearScan::handle_method_arguments(LIR_Op* op) { 1171 // special handling for method arguments (moves from stack to virtual register): 1172 // the interval gets no register assigned, but the stack slot. 1173 // it is split before the first use by the register allocator. 1174 1175 if (op->code() == lir_move) { 1176 assert(op->as_Op1() != NULL, "must be LIR_Op1"); 1177 LIR_Op1* move = (LIR_Op1*)op; 1178 1179 if (move->in_opr()->is_stack()) { 1180 #ifdef ASSERT 1181 int arg_size = compilation()->method()->arg_size(); 1182 LIR_Opr o = move->in_opr(); 1183 if (o->is_single_stack()) { 1184 assert(o->single_stack_ix() >= 0 && o->single_stack_ix() < arg_size, "out of range"); 1185 } else if (o->is_double_stack()) { 1186 assert(o->double_stack_ix() >= 0 && o->double_stack_ix() < arg_size, "out of range"); 1187 } else { 1188 ShouldNotReachHere(); 1189 } 1190 1191 assert(move->id() > 0, "invalid id"); 1192 assert(block_of_op_with_id(move->id())->number_of_preds() == 0, "move from stack must be in first block"); 1193 assert(move->result_opr()->is_virtual(), "result of move must be a virtual register"); 1194 1195 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()))); 1196 #endif 1197 1198 Interval* interval = interval_at(reg_num(move->result_opr())); 1199 1200 int stack_slot = LinearScan::nof_regs + (move->in_opr()->is_single_stack() ? move->in_opr()->single_stack_ix() : move->in_opr()->double_stack_ix()); 1201 interval->set_canonical_spill_slot(stack_slot); 1202 interval->assign_reg(stack_slot); 1203 } 1204 } 1205 } 1206 1207 void LinearScan::handle_doubleword_moves(LIR_Op* op) { 1208 // special handling for doubleword move from memory to register: 1209 // in this case the registers of the input address and the result 1210 // registers must not overlap -> add a temp range for the input registers 1211 if (op->code() == lir_move) { 1212 assert(op->as_Op1() != NULL, "must be LIR_Op1"); 1213 LIR_Op1* move = (LIR_Op1*)op; 1214 1215 if (move->result_opr()->is_double_cpu() && move->in_opr()->is_pointer()) { 1216 LIR_Address* address = move->in_opr()->as_address_ptr(); 1217 if (address != NULL) { 1218 if (address->base()->is_valid()) { 1219 add_temp(address->base(), op->id(), noUse); 1220 } 1221 if (address->index()->is_valid()) { 1222 add_temp(address->index(), op->id(), noUse); 1223 } 1224 } 1225 } 1226 } 1227 } 1228 1229 void LinearScan::add_register_hints(LIR_Op* op) { 1230 switch (op->code()) { 1231 case lir_move: // fall through 1232 case lir_convert: { 1233 assert(op->as_Op1() != NULL, "lir_move, lir_convert must be LIR_Op1"); 1234 LIR_Op1* move = (LIR_Op1*)op; 1235 1236 LIR_Opr move_from = move->in_opr(); 1237 LIR_Opr move_to = move->result_opr(); 1238 1239 if (move_to->is_register() && move_from->is_register()) { 1240 Interval* from = interval_at(reg_num(move_from)); 1241 Interval* to = interval_at(reg_num(move_to)); 1242 if (from != NULL && to != NULL) { 1243 to->set_register_hint(from); 1244 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())); 1245 } 1246 } 1247 break; 1248 } 1249 case lir_cmove: { 1250 assert(op->as_Op2() != NULL, "lir_cmove must be LIR_Op2"); 1251 LIR_Op2* cmove = (LIR_Op2*)op; 1252 1253 LIR_Opr move_from = cmove->in_opr1(); 1254 LIR_Opr move_to = cmove->result_opr(); 1255 1256 if (move_to->is_register() && move_from->is_register()) { 1257 Interval* from = interval_at(reg_num(move_from)); 1258 Interval* to = interval_at(reg_num(move_to)); 1259 if (from != NULL && to != NULL) { 1260 to->set_register_hint(from); 1261 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())); 1262 } 1263 } 1264 break; 1265 } 1266 } 1267 } 1268 1269 1270 void LinearScan::build_intervals() { 1271 TIME_LINEAR_SCAN(timer_build_intervals); 1272 1273 // initialize interval list with expected number of intervals 1274 // (32 is added to have some space for split children without having to resize the list) 1275 _intervals = IntervalList(num_virtual_regs() + 32); 1276 // initialize all slots that are used by build_intervals 1277 _intervals.at_put_grow(num_virtual_regs() - 1, NULL, NULL); 1278 1279 // create a list with all caller-save registers (cpu, fpu, xmm) 1280 // when an instruction is a call, a temp range is created for all these registers 1281 int num_caller_save_registers = 0; 1282 int caller_save_registers[LinearScan::nof_regs]; 1283 1284 int i; 1285 for (i = 0; i < FrameMap::nof_caller_save_cpu_regs(); i++) { 1286 LIR_Opr opr = FrameMap::caller_save_cpu_reg_at(i); 1287 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands"); 1288 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register"); 1289 caller_save_registers[num_caller_save_registers++] = reg_num(opr); 1290 } 1291 1292 // temp ranges for fpu registers are only created when the method has 1293 // virtual fpu operands. Otherwise no allocation for fpu registers is 1294 // perfomed and so the temp ranges would be useless 1295 if (has_fpu_registers()) { 1296 #ifdef X86 1297 if (UseSSE < 2) { 1298 #endif 1299 for (i = 0; i < FrameMap::nof_caller_save_fpu_regs; i++) { 1300 LIR_Opr opr = FrameMap::caller_save_fpu_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 #ifdef X86 1306 } 1307 if (UseSSE > 0) { 1308 for (i = 0; i < FrameMap::nof_caller_save_xmm_regs; i++) { 1309 LIR_Opr opr = FrameMap::caller_save_xmm_reg_at(i); 1310 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands"); 1311 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register"); 1312 caller_save_registers[num_caller_save_registers++] = reg_num(opr); 1313 } 1314 } 1315 #endif 1316 } 1317 assert(num_caller_save_registers <= LinearScan::nof_regs, "out of bounds"); 1318 1319 1320 LIR_OpVisitState visitor; 1321 1322 // iterate all blocks in reverse order 1323 for (i = block_count() - 1; i >= 0; i--) { 1324 BlockBegin* block = block_at(i); 1325 LIR_OpList* instructions = block->lir()->instructions_list(); 1326 int block_from = block->first_lir_instruction_id(); 1327 int block_to = block->last_lir_instruction_id(); 1328 1329 assert(block_from == instructions->at(0)->id(), "must be"); 1330 assert(block_to == instructions->at(instructions->length() - 1)->id(), "must be"); 1331 1332 // Update intervals for registers live at the end of this block; 1333 BitMap live = block->live_out(); 1334 int size = (int)live.size(); 1335 for (int number = (int)live.get_next_one_offset(0, size); number < size; number = (int)live.get_next_one_offset(number + 1, size)) { 1336 assert(live.at(number), "should not stop here otherwise"); 1337 assert(number >= LIR_OprDesc::vreg_base, "fixed intervals must not be live on block bounds"); 1338 TRACE_LINEAR_SCAN(2, tty->print_cr("live in %d to %d", number, block_to + 2)); 1339 1340 add_use(number, block_from, block_to + 2, noUse, T_ILLEGAL); 1341 1342 // add special use positions for loop-end blocks when the 1343 // interval is used anywhere inside this loop. It's possible 1344 // that the block was part of a non-natural loop, so it might 1345 // have an invalid loop index. 1346 if (block->is_set(BlockBegin::linear_scan_loop_end_flag) && 1347 block->loop_index() != -1 && 1348 is_interval_in_loop(number, block->loop_index())) { 1349 interval_at(number)->add_use_pos(block_to + 1, loopEndMarker); 1350 } 1351 } 1352 1353 // iterate all instructions of the block in reverse order. 1354 // skip the first instruction because it is always a label 1355 // definitions of intervals are processed before uses 1356 assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label"); 1357 for (int j = instructions->length() - 1; j >= 1; j--) { 1358 LIR_Op* op = instructions->at(j); 1359 int op_id = op->id(); 1360 1361 // visit operation to collect all operands 1362 visitor.visit(op); 1363 1364 // add a temp range for each register if operation destroys caller-save registers 1365 if (visitor.has_call()) { 1366 for (int k = 0; k < num_caller_save_registers; k++) { 1367 add_temp(caller_save_registers[k], op_id, noUse, T_ILLEGAL); 1368 } 1369 TRACE_LINEAR_SCAN(4, tty->print_cr("operation destroys all caller-save registers")); 1370 } 1371 1372 // Add any platform dependent temps 1373 pd_add_temps(op); 1374 1375 // visit definitions (output and temp operands) 1376 int k, n; 1377 n = visitor.opr_count(LIR_OpVisitState::outputMode); 1378 for (k = 0; k < n; k++) { 1379 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k); 1380 assert(opr->is_register(), "visitor should only return register operands"); 1381 add_def(opr, op_id, use_kind_of_output_operand(op, opr)); 1382 } 1383 1384 n = visitor.opr_count(LIR_OpVisitState::tempMode); 1385 for (k = 0; k < n; k++) { 1386 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k); 1387 assert(opr->is_register(), "visitor should only return register operands"); 1388 add_temp(opr, op_id, mustHaveRegister); 1389 } 1390 1391 // visit uses (input operands) 1392 n = visitor.opr_count(LIR_OpVisitState::inputMode); 1393 for (k = 0; k < n; k++) { 1394 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k); 1395 assert(opr->is_register(), "visitor should only return register operands"); 1396 add_use(opr, block_from, op_id, use_kind_of_input_operand(op, opr)); 1397 } 1398 1399 // Add uses of live locals from interpreter's point of view for proper 1400 // debug information generation 1401 // Treat these operands as temp values (if the life range is extended 1402 // to a call site, the value would be in a register at the call otherwise) 1403 n = visitor.info_count(); 1404 for (k = 0; k < n; k++) { 1405 CodeEmitInfo* info = visitor.info_at(k); 1406 ValueStack* stack = info->stack(); 1407 for_each_state_value(stack, value, 1408 add_use(value, block_from, op_id + 1, noUse); 1409 ); 1410 } 1411 1412 // special steps for some instructions (especially moves) 1413 handle_method_arguments(op); 1414 handle_doubleword_moves(op); 1415 add_register_hints(op); 1416 1417 } // end of instruction iteration 1418 } // end of block iteration 1419 1420 1421 // add the range [0, 1[ to all fixed intervals 1422 // -> the register allocator need not handle unhandled fixed intervals 1423 for (int n = 0; n < LinearScan::nof_regs; n++) { 1424 Interval* interval = interval_at(n); 1425 if (interval != NULL) { 1426 interval->add_range(0, 1); 1427 } 1428 } 1429 } 1430 1431 1432 // ********** Phase 5: actual register allocation 1433 1434 int LinearScan::interval_cmp(Interval** a, Interval** b) { 1435 if (*a != NULL) { 1436 if (*b != NULL) { 1437 return (*a)->from() - (*b)->from(); 1438 } else { 1439 return -1; 1440 } 1441 } else { 1442 if (*b != NULL) { 1443 return 1; 1444 } else { 1445 return 0; 1446 } 1447 } 1448 } 1449 1450 #ifndef PRODUCT 1451 bool LinearScan::is_sorted(IntervalArray* intervals) { 1452 int from = -1; 1453 int i, j; 1454 for (i = 0; i < intervals->length(); i ++) { 1455 Interval* it = intervals->at(i); 1456 if (it != NULL) { 1457 if (from > it->from()) { 1458 assert(false, ""); 1459 return false; 1460 } 1461 from = it->from(); 1462 } 1463 } 1464 1465 // check in both directions if sorted list and unsorted list contain same intervals 1466 for (i = 0; i < interval_count(); i++) { 1467 if (interval_at(i) != NULL) { 1468 int num_found = 0; 1469 for (j = 0; j < intervals->length(); j++) { 1470 if (interval_at(i) == intervals->at(j)) { 1471 num_found++; 1472 } 1473 } 1474 assert(num_found == 1, "lists do not contain same intervals"); 1475 } 1476 } 1477 for (j = 0; j < intervals->length(); j++) { 1478 int num_found = 0; 1479 for (i = 0; i < interval_count(); i++) { 1480 if (interval_at(i) == intervals->at(j)) { 1481 num_found++; 1482 } 1483 } 1484 assert(num_found == 1, "lists do not contain same intervals"); 1485 } 1486 1487 return true; 1488 } 1489 #endif 1490 1491 void LinearScan::add_to_list(Interval** first, Interval** prev, Interval* interval) { 1492 if (*prev != NULL) { 1493 (*prev)->set_next(interval); 1494 } else { 1495 *first = interval; 1496 } 1497 *prev = interval; 1498 } 1499 1500 void LinearScan::create_unhandled_lists(Interval** list1, Interval** list2, bool (is_list1)(const Interval* i), bool (is_list2)(const Interval* i)) { 1501 assert(is_sorted(_sorted_intervals), "interval list is not sorted"); 1502 1503 *list1 = *list2 = Interval::end(); 1504 1505 Interval* list1_prev = NULL; 1506 Interval* list2_prev = NULL; 1507 Interval* v; 1508 1509 const int n = _sorted_intervals->length(); 1510 for (int i = 0; i < n; i++) { 1511 v = _sorted_intervals->at(i); 1512 if (v == NULL) continue; 1513 1514 if (is_list1(v)) { 1515 add_to_list(list1, &list1_prev, v); 1516 } else if (is_list2 == NULL || is_list2(v)) { 1517 add_to_list(list2, &list2_prev, v); 1518 } 1519 } 1520 1521 if (list1_prev != NULL) list1_prev->set_next(Interval::end()); 1522 if (list2_prev != NULL) list2_prev->set_next(Interval::end()); 1523 1524 assert(list1_prev == NULL || list1_prev->next() == Interval::end(), "linear list ends not with sentinel"); 1525 assert(list2_prev == NULL || list2_prev->next() == Interval::end(), "linear list ends not with sentinel"); 1526 } 1527 1528 1529 void LinearScan::sort_intervals_before_allocation() { 1530 TIME_LINEAR_SCAN(timer_sort_intervals_before); 1531 1532 if (_needs_full_resort) { 1533 // There is no known reason why this should occur but just in case... 1534 assert(false, "should never occur"); 1535 // Re-sort existing interval list because an Interval::from() has changed 1536 _sorted_intervals->sort(interval_cmp); 1537 _needs_full_resort = false; 1538 } 1539 1540 IntervalList* unsorted_list = &_intervals; 1541 int unsorted_len = unsorted_list->length(); 1542 int sorted_len = 0; 1543 int unsorted_idx; 1544 int sorted_idx = 0; 1545 int sorted_from_max = -1; 1546 1547 // calc number of items for sorted list (sorted list must not contain NULL values) 1548 for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) { 1549 if (unsorted_list->at(unsorted_idx) != NULL) { 1550 sorted_len++; 1551 } 1552 } 1553 IntervalArray* sorted_list = new IntervalArray(sorted_len); 1554 1555 // special sorting algorithm: the original interval-list is almost sorted, 1556 // only some intervals are swapped. So this is much faster than a complete QuickSort 1557 for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) { 1558 Interval* cur_interval = unsorted_list->at(unsorted_idx); 1559 1560 if (cur_interval != NULL) { 1561 int cur_from = cur_interval->from(); 1562 1563 if (sorted_from_max <= cur_from) { 1564 sorted_list->at_put(sorted_idx++, cur_interval); 1565 sorted_from_max = cur_interval->from(); 1566 } else { 1567 // the asumption that the intervals are already sorted failed, 1568 // so this interval must be sorted in manually 1569 int j; 1570 for (j = sorted_idx - 1; j >= 0 && cur_from < sorted_list->at(j)->from(); j--) { 1571 sorted_list->at_put(j + 1, sorted_list->at(j)); 1572 } 1573 sorted_list->at_put(j + 1, cur_interval); 1574 sorted_idx++; 1575 } 1576 } 1577 } 1578 _sorted_intervals = sorted_list; 1579 assert(is_sorted(_sorted_intervals), "intervals unsorted"); 1580 } 1581 1582 void LinearScan::sort_intervals_after_allocation() { 1583 TIME_LINEAR_SCAN(timer_sort_intervals_after); 1584 1585 if (_needs_full_resort) { 1586 // Re-sort existing interval list because an Interval::from() has changed 1587 _sorted_intervals->sort(interval_cmp); 1588 _needs_full_resort = false; 1589 } 1590 1591 IntervalArray* old_list = _sorted_intervals; 1592 IntervalList* new_list = _new_intervals_from_allocation; 1593 int old_len = old_list->length(); 1594 int new_len = new_list->length(); 1595 1596 if (new_len == 0) { 1597 // no intervals have been added during allocation, so sorted list is already up to date 1598 assert(is_sorted(_sorted_intervals), "intervals unsorted"); 1599 return; 1600 } 1601 1602 // conventional sort-algorithm for new intervals 1603 new_list->sort(interval_cmp); 1604 1605 // merge old and new list (both already sorted) into one combined list 1606 IntervalArray* combined_list = new IntervalArray(old_len + new_len); 1607 int old_idx = 0; 1608 int new_idx = 0; 1609 1610 while (old_idx + new_idx < old_len + new_len) { 1611 if (new_idx >= new_len || (old_idx < old_len && old_list->at(old_idx)->from() <= new_list->at(new_idx)->from())) { 1612 combined_list->at_put(old_idx + new_idx, old_list->at(old_idx)); 1613 old_idx++; 1614 } else { 1615 combined_list->at_put(old_idx + new_idx, new_list->at(new_idx)); 1616 new_idx++; 1617 } 1618 } 1619 1620 _sorted_intervals = combined_list; 1621 assert(is_sorted(_sorted_intervals), "intervals unsorted"); 1622 } 1623 1624 1625 void LinearScan::allocate_registers() { 1626 TIME_LINEAR_SCAN(timer_allocate_registers); 1627 1628 Interval* precolored_cpu_intervals, *not_precolored_cpu_intervals; 1629 Interval* precolored_fpu_intervals, *not_precolored_fpu_intervals; 1630 1631 create_unhandled_lists(&precolored_cpu_intervals, ¬_precolored_cpu_intervals, is_precolored_cpu_interval, is_virtual_cpu_interval); 1632 if (has_fpu_registers()) { 1633 create_unhandled_lists(&precolored_fpu_intervals, ¬_precolored_fpu_intervals, is_precolored_fpu_interval, is_virtual_fpu_interval); 1634 #ifdef ASSERT 1635 } else { 1636 // fpu register allocation is omitted because no virtual fpu registers are present 1637 // just check this again... 1638 create_unhandled_lists(&precolored_fpu_intervals, ¬_precolored_fpu_intervals, is_precolored_fpu_interval, is_virtual_fpu_interval); 1639 assert(not_precolored_fpu_intervals == Interval::end(), "missed an uncolored fpu interval"); 1640 #endif 1641 } 1642 1643 // allocate cpu registers 1644 LinearScanWalker cpu_lsw(this, precolored_cpu_intervals, not_precolored_cpu_intervals); 1645 cpu_lsw.walk(); 1646 cpu_lsw.finish_allocation(); 1647 1648 if (has_fpu_registers()) { 1649 // allocate fpu registers 1650 LinearScanWalker fpu_lsw(this, precolored_fpu_intervals, not_precolored_fpu_intervals); 1651 fpu_lsw.walk(); 1652 fpu_lsw.finish_allocation(); 1653 } 1654 } 1655 1656 1657 // ********** Phase 6: resolve data flow 1658 // (insert moves at edges between blocks if intervals have been split) 1659 1660 // wrapper for Interval::split_child_at_op_id that performs a bailout in product mode 1661 // instead of returning NULL 1662 Interval* LinearScan::split_child_at_op_id(Interval* interval, int op_id, LIR_OpVisitState::OprMode mode) { 1663 Interval* result = interval->split_child_at_op_id(op_id, mode); 1664 if (result != NULL) { 1665 return result; 1666 } 1667 1668 assert(false, "must find an interval, but do a clean bailout in product mode"); 1669 result = new Interval(LIR_OprDesc::vreg_base); 1670 result->assign_reg(0); 1671 result->set_type(T_INT); 1672 BAILOUT_("LinearScan: interval is NULL", result); 1673 } 1674 1675 1676 Interval* LinearScan::interval_at_block_begin(BlockBegin* block, int reg_num) { 1677 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds"); 1678 assert(interval_at(reg_num) != NULL, "no interval found"); 1679 1680 return split_child_at_op_id(interval_at(reg_num), block->first_lir_instruction_id(), LIR_OpVisitState::outputMode); 1681 } 1682 1683 Interval* LinearScan::interval_at_block_end(BlockBegin* block, int reg_num) { 1684 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds"); 1685 assert(interval_at(reg_num) != NULL, "no interval found"); 1686 1687 return split_child_at_op_id(interval_at(reg_num), block->last_lir_instruction_id() + 1, LIR_OpVisitState::outputMode); 1688 } 1689 1690 Interval* LinearScan::interval_at_op_id(int reg_num, int op_id) { 1691 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds"); 1692 assert(interval_at(reg_num) != NULL, "no interval found"); 1693 1694 return split_child_at_op_id(interval_at(reg_num), op_id, LIR_OpVisitState::inputMode); 1695 } 1696 1697 1698 void LinearScan::resolve_collect_mappings(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) { 1699 DEBUG_ONLY(move_resolver.check_empty()); 1700 1701 const int num_regs = num_virtual_regs(); 1702 const int size = live_set_size(); 1703 const BitMap live_at_edge = to_block->live_in(); 1704 1705 // visit all registers where the live_at_edge bit is set 1706 for (int r = (int)live_at_edge.get_next_one_offset(0, size); r < size; r = (int)live_at_edge.get_next_one_offset(r + 1, size)) { 1707 assert(r < num_regs, "live information set for not exisiting interval"); 1708 assert(from_block->live_out().at(r) && to_block->live_in().at(r), "interval not live at this edge"); 1709 1710 Interval* from_interval = interval_at_block_end(from_block, r); 1711 Interval* to_interval = interval_at_block_begin(to_block, r); 1712 1713 if (from_interval != to_interval && (from_interval->assigned_reg() != to_interval->assigned_reg() || from_interval->assigned_regHi() != to_interval->assigned_regHi())) { 1714 // need to insert move instruction 1715 move_resolver.add_mapping(from_interval, to_interval); 1716 } 1717 } 1718 } 1719 1720 1721 void LinearScan::resolve_find_insert_pos(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) { 1722 if (from_block->number_of_sux() <= 1) { 1723 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at end of from_block B%d", from_block->block_id())); 1724 1725 LIR_OpList* instructions = from_block->lir()->instructions_list(); 1726 LIR_OpBranch* branch = instructions->last()->as_OpBranch(); 1727 if (branch != NULL) { 1728 // insert moves before branch 1729 assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump"); 1730 move_resolver.set_insert_position(from_block->lir(), instructions->length() - 2); 1731 } else { 1732 move_resolver.set_insert_position(from_block->lir(), instructions->length() - 1); 1733 } 1734 1735 } else { 1736 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at beginning of to_block B%d", to_block->block_id())); 1737 #ifdef ASSERT 1738 assert(from_block->lir()->instructions_list()->at(0)->as_OpLabel() != NULL, "block does not start with a label"); 1739 1740 // because the number of predecessor edges matches the number of 1741 // successor edges, blocks which are reached by switch statements 1742 // may have be more than one predecessor but it will be guaranteed 1743 // that all predecessors will be the same. 1744 for (int i = 0; i < to_block->number_of_preds(); i++) { 1745 assert(from_block == to_block->pred_at(i), "all critical edges must be broken"); 1746 } 1747 #endif 1748 1749 move_resolver.set_insert_position(to_block->lir(), 0); 1750 } 1751 } 1752 1753 1754 // insert necessary moves (spilling or reloading) at edges between blocks if interval has been split 1755 void LinearScan::resolve_data_flow() { 1756 TIME_LINEAR_SCAN(timer_resolve_data_flow); 1757 1758 int num_blocks = block_count(); 1759 MoveResolver move_resolver(this); 1760 BitMap block_completed(num_blocks); block_completed.clear(); 1761 BitMap already_resolved(num_blocks); already_resolved.clear(); 1762 1763 int i; 1764 for (i = 0; i < num_blocks; i++) { 1765 BlockBegin* block = block_at(i); 1766 1767 // check if block has only one predecessor and only one successor 1768 if (block->number_of_preds() == 1 && block->number_of_sux() == 1 && block->number_of_exception_handlers() == 0) { 1769 LIR_OpList* instructions = block->lir()->instructions_list(); 1770 assert(instructions->at(0)->code() == lir_label, "block must start with label"); 1771 assert(instructions->last()->code() == lir_branch, "block with successors must end with branch"); 1772 assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block with successor must end with unconditional branch"); 1773 1774 // check if block is empty (only label and branch) 1775 if (instructions->length() == 2) { 1776 BlockBegin* pred = block->pred_at(0); 1777 BlockBegin* sux = block->sux_at(0); 1778 1779 // prevent optimization of two consecutive blocks 1780 if (!block_completed.at(pred->linear_scan_number()) && !block_completed.at(sux->linear_scan_number())) { 1781 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())); 1782 block_completed.set_bit(block->linear_scan_number()); 1783 1784 // directly resolve between pred and sux (without looking at the empty block between) 1785 resolve_collect_mappings(pred, sux, move_resolver); 1786 if (move_resolver.has_mappings()) { 1787 move_resolver.set_insert_position(block->lir(), 0); 1788 move_resolver.resolve_and_append_moves(); 1789 } 1790 } 1791 } 1792 } 1793 } 1794 1795 1796 for (i = 0; i < num_blocks; i++) { 1797 if (!block_completed.at(i)) { 1798 BlockBegin* from_block = block_at(i); 1799 already_resolved.set_from(block_completed); 1800 1801 int num_sux = from_block->number_of_sux(); 1802 for (int s = 0; s < num_sux; s++) { 1803 BlockBegin* to_block = from_block->sux_at(s); 1804 1805 // check for duplicate edges between the same blocks (can happen with switch blocks) 1806 if (!already_resolved.at(to_block->linear_scan_number())) { 1807 TRACE_LINEAR_SCAN(3, tty->print_cr("**** processing edge between B%d and B%d", from_block->block_id(), to_block->block_id())); 1808 already_resolved.set_bit(to_block->linear_scan_number()); 1809 1810 // collect all intervals that have been split between from_block and to_block 1811 resolve_collect_mappings(from_block, to_block, move_resolver); 1812 if (move_resolver.has_mappings()) { 1813 resolve_find_insert_pos(from_block, to_block, move_resolver); 1814 move_resolver.resolve_and_append_moves(); 1815 } 1816 } 1817 } 1818 } 1819 } 1820 } 1821 1822 1823 void LinearScan::resolve_exception_entry(BlockBegin* block, int reg_num, MoveResolver &move_resolver) { 1824 if (interval_at(reg_num) == NULL) { 1825 // if a phi function is never used, no interval is created -> ignore this 1826 return; 1827 } 1828 1829 Interval* interval = interval_at_block_begin(block, reg_num); 1830 int reg = interval->assigned_reg(); 1831 int regHi = interval->assigned_regHi(); 1832 1833 if ((reg < nof_regs && interval->always_in_memory()) || 1834 (use_fpu_stack_allocation() && reg >= pd_first_fpu_reg && reg <= pd_last_fpu_reg)) { 1835 // the interval is split to get a short range that is located on the stack 1836 // in the following two cases: 1837 // * the interval started in memory (e.g. method parameter), but is currently in a register 1838 // this is an optimization for exception handling that reduces the number of moves that 1839 // are necessary for resolving the states when an exception uses this exception handler 1840 // * the interval would be on the fpu stack at the begin of the exception handler 1841 // this is not allowed because of the complicated fpu stack handling on Intel 1842 1843 // range that will be spilled to memory 1844 int from_op_id = block->first_lir_instruction_id(); 1845 int to_op_id = from_op_id + 1; // short live range of length 1 1846 assert(interval->from() <= from_op_id && interval->to() >= to_op_id, 1847 "no split allowed between exception entry and first instruction"); 1848 1849 if (interval->from() != from_op_id) { 1850 // the part before from_op_id is unchanged 1851 interval = interval->split(from_op_id); 1852 interval->assign_reg(reg, regHi); 1853 append_interval(interval); 1854 } else { 1855 _needs_full_resort = true; 1856 } 1857 assert(interval->from() == from_op_id, "must be true now"); 1858 1859 Interval* spilled_part = interval; 1860 if (interval->to() != to_op_id) { 1861 // the part after to_op_id is unchanged 1862 spilled_part = interval->split_from_start(to_op_id); 1863 append_interval(spilled_part); 1864 move_resolver.add_mapping(spilled_part, interval); 1865 } 1866 assign_spill_slot(spilled_part); 1867 1868 assert(spilled_part->from() == from_op_id && spilled_part->to() == to_op_id, "just checking"); 1869 } 1870 } 1871 1872 void LinearScan::resolve_exception_entry(BlockBegin* block, MoveResolver &move_resolver) { 1873 assert(block->is_set(BlockBegin::exception_entry_flag), "should not call otherwise"); 1874 DEBUG_ONLY(move_resolver.check_empty()); 1875 1876 // visit all registers where the live_in bit is set 1877 int size = live_set_size(); 1878 for (int r = (int)block->live_in().get_next_one_offset(0, size); r < size; r = (int)block->live_in().get_next_one_offset(r + 1, size)) { 1879 resolve_exception_entry(block, r, move_resolver); 1880 } 1881 1882 // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately 1883 for_each_phi_fun(block, phi, 1884 resolve_exception_entry(block, phi->operand()->vreg_number(), move_resolver) 1885 ); 1886 1887 if (move_resolver.has_mappings()) { 1888 // insert moves after first instruction 1889 move_resolver.set_insert_position(block->lir(), 0); 1890 move_resolver.resolve_and_append_moves(); 1891 } 1892 } 1893 1894 1895 void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, int reg_num, Phi* phi, MoveResolver &move_resolver) { 1896 if (interval_at(reg_num) == NULL) { 1897 // if a phi function is never used, no interval is created -> ignore this 1898 return; 1899 } 1900 1901 // the computation of to_interval is equal to resolve_collect_mappings, 1902 // but from_interval is more complicated because of phi functions 1903 BlockBegin* to_block = handler->entry_block(); 1904 Interval* to_interval = interval_at_block_begin(to_block, reg_num); 1905 1906 if (phi != NULL) { 1907 // phi function of the exception entry block 1908 // no moves are created for this phi function in the LIR_Generator, so the 1909 // interval at the throwing instruction must be searched using the operands 1910 // of the phi function 1911 Value from_value = phi->operand_at(handler->phi_operand()); 1912 1913 // with phi functions it can happen that the same from_value is used in 1914 // multiple mappings, so notify move-resolver that this is allowed 1915 move_resolver.set_multiple_reads_allowed(); 1916 1917 Constant* con = from_value->as_Constant(); 1918 if (con != NULL && !con->is_pinned()) { 1919 // unpinned constants may have no register, so add mapping from constant to interval 1920 move_resolver.add_mapping(LIR_OprFact::value_type(con->type()), to_interval); 1921 } else { 1922 // search split child at the throwing op_id 1923 Interval* from_interval = interval_at_op_id(from_value->operand()->vreg_number(), throwing_op_id); 1924 move_resolver.add_mapping(from_interval, to_interval); 1925 } 1926 1927 } else { 1928 // no phi function, so use reg_num also for from_interval 1929 // search split child at the throwing op_id 1930 Interval* from_interval = interval_at_op_id(reg_num, throwing_op_id); 1931 if (from_interval != to_interval) { 1932 // optimization to reduce number of moves: when to_interval is on stack and 1933 // the stack slot is known to be always correct, then no move is necessary 1934 if (!from_interval->always_in_memory() || from_interval->canonical_spill_slot() != to_interval->assigned_reg()) { 1935 move_resolver.add_mapping(from_interval, to_interval); 1936 } 1937 } 1938 } 1939 } 1940 1941 void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, MoveResolver &move_resolver) { 1942 TRACE_LINEAR_SCAN(4, tty->print_cr("resolving exception handler B%d: throwing_op_id=%d", handler->entry_block()->block_id(), throwing_op_id)); 1943 1944 DEBUG_ONLY(move_resolver.check_empty()); 1945 assert(handler->lir_op_id() == -1, "already processed this xhandler"); 1946 DEBUG_ONLY(handler->set_lir_op_id(throwing_op_id)); 1947 assert(handler->entry_code() == NULL, "code already present"); 1948 1949 // visit all registers where the live_in bit is set 1950 BlockBegin* block = handler->entry_block(); 1951 int size = live_set_size(); 1952 for (int r = (int)block->live_in().get_next_one_offset(0, size); r < size; r = (int)block->live_in().get_next_one_offset(r + 1, size)) { 1953 resolve_exception_edge(handler, throwing_op_id, r, NULL, move_resolver); 1954 } 1955 1956 // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately 1957 for_each_phi_fun(block, phi, 1958 resolve_exception_edge(handler, throwing_op_id, phi->operand()->vreg_number(), phi, move_resolver) 1959 ); 1960 1961 if (move_resolver.has_mappings()) { 1962 LIR_List* entry_code = new LIR_List(compilation()); 1963 move_resolver.set_insert_position(entry_code, 0); 1964 move_resolver.resolve_and_append_moves(); 1965 1966 entry_code->jump(handler->entry_block()); 1967 handler->set_entry_code(entry_code); 1968 } 1969 } 1970 1971 1972 void LinearScan::resolve_exception_handlers() { 1973 MoveResolver move_resolver(this); 1974 LIR_OpVisitState visitor; 1975 int num_blocks = block_count(); 1976 1977 int i; 1978 for (i = 0; i < num_blocks; i++) { 1979 BlockBegin* block = block_at(i); 1980 if (block->is_set(BlockBegin::exception_entry_flag)) { 1981 resolve_exception_entry(block, move_resolver); 1982 } 1983 } 1984 1985 for (i = 0; i < num_blocks; i++) { 1986 BlockBegin* block = block_at(i); 1987 LIR_List* ops = block->lir(); 1988 int num_ops = ops->length(); 1989 1990 // iterate all instructions of the block. skip the first because it is always a label 1991 assert(visitor.no_operands(ops->at(0)), "first operation must always be a label"); 1992 for (int j = 1; j < num_ops; j++) { 1993 LIR_Op* op = ops->at(j); 1994 int op_id = op->id(); 1995 1996 if (op_id != -1 && has_info(op_id)) { 1997 // visit operation to collect all operands 1998 visitor.visit(op); 1999 assert(visitor.info_count() > 0, "should not visit otherwise"); 2000 2001 XHandlers* xhandlers = visitor.all_xhandler(); 2002 int n = xhandlers->length(); 2003 for (int k = 0; k < n; k++) { 2004 resolve_exception_edge(xhandlers->handler_at(k), op_id, move_resolver); 2005 } 2006 2007 #ifdef ASSERT 2008 } else { 2009 visitor.visit(op); 2010 assert(visitor.all_xhandler()->length() == 0, "missed exception handler"); 2011 #endif 2012 } 2013 } 2014 } 2015 } 2016 2017 2018 // ********** Phase 7: assign register numbers back to LIR 2019 // (includes computation of debug information and oop maps) 2020 2021 VMReg LinearScan::vm_reg_for_interval(Interval* interval) { 2022 VMReg reg = interval->cached_vm_reg(); 2023 if (!reg->is_valid() ) { 2024 reg = vm_reg_for_operand(operand_for_interval(interval)); 2025 interval->set_cached_vm_reg(reg); 2026 } 2027 assert(reg == vm_reg_for_operand(operand_for_interval(interval)), "wrong cached value"); 2028 return reg; 2029 } 2030 2031 VMReg LinearScan::vm_reg_for_operand(LIR_Opr opr) { 2032 assert(opr->is_oop(), "currently only implemented for oop operands"); 2033 return frame_map()->regname(opr); 2034 } 2035 2036 2037 LIR_Opr LinearScan::operand_for_interval(Interval* interval) { 2038 LIR_Opr opr = interval->cached_opr(); 2039 if (opr->is_illegal()) { 2040 opr = calc_operand_for_interval(interval); 2041 interval->set_cached_opr(opr); 2042 } 2043 2044 assert(opr == calc_operand_for_interval(interval), "wrong cached value"); 2045 return opr; 2046 } 2047 2048 LIR_Opr LinearScan::calc_operand_for_interval(const Interval* interval) { 2049 int assigned_reg = interval->assigned_reg(); 2050 BasicType type = interval->type(); 2051 2052 if (assigned_reg >= nof_regs) { 2053 // stack slot 2054 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2055 return LIR_OprFact::stack(assigned_reg - nof_regs, type); 2056 2057 } else { 2058 // register 2059 switch (type) { 2060 case T_OBJECT: { 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_oop(assigned_reg); 2064 } 2065 2066 case T_ADDRESS: { 2067 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2068 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2069 return LIR_OprFact::single_cpu_address(assigned_reg); 2070 } 2071 2072 case T_METADATA: { 2073 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2074 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2075 return LIR_OprFact::single_cpu_metadata(assigned_reg); 2076 } 2077 2078 #ifdef __SOFTFP__ 2079 case T_FLOAT: // fall through 2080 #endif // __SOFTFP__ 2081 case T_INT: { 2082 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2083 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2084 return LIR_OprFact::single_cpu(assigned_reg); 2085 } 2086 2087 #ifdef __SOFTFP__ 2088 case T_DOUBLE: // fall through 2089 #endif // __SOFTFP__ 2090 case T_LONG: { 2091 int assigned_regHi = interval->assigned_regHi(); 2092 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); 2093 assert(num_physical_regs(T_LONG) == 1 || 2094 (assigned_regHi >= pd_first_cpu_reg && assigned_regHi <= pd_last_cpu_reg), "no cpu register"); 2095 2096 assert(assigned_reg != assigned_regHi, "invalid allocation"); 2097 assert(num_physical_regs(T_LONG) == 1 || assigned_reg < assigned_regHi, 2098 "register numbers must be sorted (ensure that e.g. a move from eax,ebx to ebx,eax can not occur)"); 2099 assert((assigned_regHi != any_reg) ^ (num_physical_regs(T_LONG) == 1), "must be match"); 2100 if (requires_adjacent_regs(T_LONG)) { 2101 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == assigned_regHi, "must be sequential and even"); 2102 } 2103 2104 #ifdef _LP64 2105 return LIR_OprFact::double_cpu(assigned_reg, assigned_reg); 2106 #else 2107 #if defined(SPARC) || defined(PPC) 2108 return LIR_OprFact::double_cpu(assigned_regHi, assigned_reg); 2109 #else 2110 return LIR_OprFact::double_cpu(assigned_reg, assigned_regHi); 2111 #endif // SPARC 2112 #endif // LP64 2113 } 2114 2115 #ifndef __SOFTFP__ 2116 case T_FLOAT: { 2117 #ifdef X86 2118 if (UseSSE >= 1) { 2119 assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= pd_last_xmm_reg, "no xmm register"); 2120 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2121 return LIR_OprFact::single_xmm(assigned_reg - pd_first_xmm_reg); 2122 } 2123 #endif 2124 2125 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register"); 2126 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); 2127 return LIR_OprFact::single_fpu(assigned_reg - pd_first_fpu_reg); 2128 } 2129 2130 case T_DOUBLE: { 2131 #ifdef X86 2132 if (UseSSE >= 2) { 2133 assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= pd_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(ARM) 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 // Check if this is a patch site. 2386 bool is_patch_info = false; 2387 if (op->code() == lir_move) { 2388 assert(!is_call_site, "move must not be a call site"); 2389 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); 2390 LIR_Op1* move = (LIR_Op1*)op; 2391 2392 is_patch_info = move->patch_code() != lir_patch_none; 2393 } 2394 2395 // Iterate through active intervals 2396 for (Interval* interval = iw->active_first(fixedKind); interval != Interval::end(); interval = interval->next()) { 2397 int assigned_reg = interval->assigned_reg(); 2398 2399 assert(interval->current_from() <= op->id() && op->id() <= interval->current_to(), "interval should not be active otherwise"); 2400 assert(interval->assigned_regHi() == any_reg, "oop must be single word"); 2401 assert(interval->reg_num() >= LIR_OprDesc::vreg_base, "fixed interval found"); 2402 2403 // Check if this range covers the instruction. Intervals that 2404 // start or end at the current operation are not included in the 2405 // oop map, except in the case of patching moves. For patching 2406 // moves, any intervals which end at this instruction are included 2407 // in the oop map since we may safepoint while doing the patch 2408 // before we've consumed the inputs. 2409 if (is_patch_info || op->id() < interval->current_to()) { 2410 2411 // caller-save registers must not be included into oop-maps at calls 2412 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"); 2413 2414 VMReg name = vm_reg_for_interval(interval); 2415 set_oop(map, name); 2416 2417 // Spill optimization: when the stack value is guaranteed to be always correct, 2418 // then it must be added to the oop map even if the interval is currently in a register 2419 if (interval->always_in_memory() && 2420 op->id() > interval->spill_definition_pos() && 2421 interval->assigned_reg() != interval->canonical_spill_slot()) { 2422 assert(interval->spill_definition_pos() > 0, "position not set correctly"); 2423 assert(interval->canonical_spill_slot() >= LinearScan::nof_regs, "no spill slot assigned"); 2424 assert(interval->assigned_reg() < LinearScan::nof_regs, "interval is on stack, so stack slot is registered twice"); 2425 2426 set_oop(map, frame_map()->slot_regname(interval->canonical_spill_slot() - LinearScan::nof_regs)); 2427 } 2428 } 2429 } 2430 2431 // add oops from lock stack 2432 assert(info->stack() != NULL, "CodeEmitInfo must always have a stack"); 2433 int locks_count = info->stack()->total_locks_size(); 2434 for (int i = 0; i < locks_count; i++) { 2435 set_oop(map, frame_map()->monitor_object_regname(i)); 2436 } 2437 2438 return map; 2439 } 2440 2441 2442 void LinearScan::compute_oop_map(IntervalWalker* iw, const LIR_OpVisitState &visitor, LIR_Op* op) { 2443 assert(visitor.info_count() > 0, "no oop map needed"); 2444 2445 // compute oop_map only for first CodeEmitInfo 2446 // because it is (in most cases) equal for all other infos of the same operation 2447 CodeEmitInfo* first_info = visitor.info_at(0); 2448 OopMap* first_oop_map = compute_oop_map(iw, op, first_info, visitor.has_call()); 2449 2450 for (int i = 0; i < visitor.info_count(); i++) { 2451 CodeEmitInfo* info = visitor.info_at(i); 2452 OopMap* oop_map = first_oop_map; 2453 2454 if (info->stack()->locks_size() != first_info->stack()->locks_size()) { 2455 // this info has a different number of locks then the precomputed oop map 2456 // (possible for lock and unlock instructions) -> compute oop map with 2457 // correct lock information 2458 oop_map = compute_oop_map(iw, op, info, visitor.has_call()); 2459 } 2460 2461 if (info->_oop_map == NULL) { 2462 info->_oop_map = oop_map; 2463 } else { 2464 // a CodeEmitInfo can not be shared between different LIR-instructions 2465 // because interval splitting can occur anywhere between two instructions 2466 // and so the oop maps must be different 2467 // -> check if the already set oop_map is exactly the one calculated for this operation 2468 assert(info->_oop_map == oop_map, "same CodeEmitInfo used for multiple LIR instructions"); 2469 } 2470 } 2471 } 2472 2473 2474 // frequently used constants 2475 // Allocate them with new so they are never destroyed (otherwise, a 2476 // forced exit could destroy these objects while they are still in 2477 // use). 2478 ConstantOopWriteValue* LinearScan::_oop_null_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantOopWriteValue(NULL); 2479 ConstantIntValue* LinearScan::_int_m1_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(-1); 2480 ConstantIntValue* LinearScan::_int_0_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(0); 2481 ConstantIntValue* LinearScan::_int_1_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(1); 2482 ConstantIntValue* LinearScan::_int_2_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(2); 2483 LocationValue* _illegal_value = new (ResourceObj::C_HEAP, mtCompiler) LocationValue(Location()); 2484 2485 void LinearScan::init_compute_debug_info() { 2486 // cache for frequently used scope values 2487 // (cpu registers and stack slots) 2488 _scope_value_cache = ScopeValueArray((LinearScan::nof_cpu_regs + frame_map()->argcount() + max_spills()) * 2, NULL); 2489 } 2490 2491 MonitorValue* LinearScan::location_for_monitor_index(int monitor_index) { 2492 Location loc; 2493 if (!frame_map()->location_for_monitor_object(monitor_index, &loc)) { 2494 bailout("too large frame"); 2495 } 2496 ScopeValue* object_scope_value = new LocationValue(loc); 2497 2498 if (!frame_map()->location_for_monitor_lock(monitor_index, &loc)) { 2499 bailout("too large frame"); 2500 } 2501 return new MonitorValue(object_scope_value, loc); 2502 } 2503 2504 LocationValue* LinearScan::location_for_name(int name, Location::Type loc_type) { 2505 Location loc; 2506 if (!frame_map()->locations_for_slot(name, loc_type, &loc)) { 2507 bailout("too large frame"); 2508 } 2509 return new LocationValue(loc); 2510 } 2511 2512 2513 int LinearScan::append_scope_value_for_constant(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) { 2514 assert(opr->is_constant(), "should not be called otherwise"); 2515 2516 LIR_Const* c = opr->as_constant_ptr(); 2517 BasicType t = c->type(); 2518 switch (t) { 2519 case T_OBJECT: { 2520 jobject value = c->as_jobject(); 2521 if (value == NULL) { 2522 scope_values->append(_oop_null_scope_value); 2523 } else { 2524 scope_values->append(new ConstantOopWriteValue(c->as_jobject())); 2525 } 2526 return 1; 2527 } 2528 2529 case T_INT: // fall through 2530 case T_FLOAT: { 2531 int value = c->as_jint_bits(); 2532 switch (value) { 2533 case -1: scope_values->append(_int_m1_scope_value); break; 2534 case 0: scope_values->append(_int_0_scope_value); break; 2535 case 1: scope_values->append(_int_1_scope_value); break; 2536 case 2: scope_values->append(_int_2_scope_value); break; 2537 default: scope_values->append(new ConstantIntValue(c->as_jint_bits())); break; 2538 } 2539 return 1; 2540 } 2541 2542 case T_LONG: // fall through 2543 case T_DOUBLE: { 2544 #ifdef _LP64 2545 scope_values->append(_int_0_scope_value); 2546 scope_values->append(new ConstantLongValue(c->as_jlong_bits())); 2547 #else 2548 if (hi_word_offset_in_bytes > lo_word_offset_in_bytes) { 2549 scope_values->append(new ConstantIntValue(c->as_jint_hi_bits())); 2550 scope_values->append(new ConstantIntValue(c->as_jint_lo_bits())); 2551 } else { 2552 scope_values->append(new ConstantIntValue(c->as_jint_lo_bits())); 2553 scope_values->append(new ConstantIntValue(c->as_jint_hi_bits())); 2554 } 2555 #endif 2556 return 2; 2557 } 2558 2559 case T_ADDRESS: { 2560 #ifdef _LP64 2561 scope_values->append(new ConstantLongValue(c->as_jint())); 2562 #else 2563 scope_values->append(new ConstantIntValue(c->as_jint())); 2564 #endif 2565 return 1; 2566 } 2567 2568 default: 2569 ShouldNotReachHere(); 2570 return -1; 2571 } 2572 } 2573 2574 int LinearScan::append_scope_value_for_operand(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) { 2575 if (opr->is_single_stack()) { 2576 int stack_idx = opr->single_stack_ix(); 2577 bool is_oop = opr->is_oop_register(); 2578 int cache_idx = (stack_idx + LinearScan::nof_cpu_regs) * 2 + (is_oop ? 1 : 0); 2579 2580 ScopeValue* sv = _scope_value_cache.at(cache_idx); 2581 if (sv == NULL) { 2582 Location::Type loc_type = is_oop ? Location::oop : Location::normal; 2583 sv = location_for_name(stack_idx, loc_type); 2584 _scope_value_cache.at_put(cache_idx, sv); 2585 } 2586 2587 // check if cached value is correct 2588 DEBUG_ONLY(assert_equal(sv, location_for_name(stack_idx, is_oop ? Location::oop : Location::normal))); 2589 2590 scope_values->append(sv); 2591 return 1; 2592 2593 } else if (opr->is_single_cpu()) { 2594 bool is_oop = opr->is_oop_register(); 2595 int cache_idx = opr->cpu_regnr() * 2 + (is_oop ? 1 : 0); 2596 Location::Type int_loc_type = NOT_LP64(Location::normal) LP64_ONLY(Location::int_in_long); 2597 2598 ScopeValue* sv = _scope_value_cache.at(cache_idx); 2599 if (sv == NULL) { 2600 Location::Type loc_type = is_oop ? Location::oop : int_loc_type; 2601 VMReg rname = frame_map()->regname(opr); 2602 sv = new LocationValue(Location::new_reg_loc(loc_type, rname)); 2603 _scope_value_cache.at_put(cache_idx, sv); 2604 } 2605 2606 // check if cached value is correct 2607 DEBUG_ONLY(assert_equal(sv, new LocationValue(Location::new_reg_loc(is_oop ? Location::oop : int_loc_type, frame_map()->regname(opr))))); 2608 2609 scope_values->append(sv); 2610 return 1; 2611 2612 #ifdef X86 2613 } else if (opr->is_single_xmm()) { 2614 VMReg rname = opr->as_xmm_float_reg()->as_VMReg(); 2615 LocationValue* sv = new LocationValue(Location::new_reg_loc(Location::normal, rname)); 2616 2617 scope_values->append(sv); 2618 return 1; 2619 #endif 2620 2621 } else if (opr->is_single_fpu()) { 2622 #ifdef X86 2623 // the exact location of fpu stack values is only known 2624 // during fpu stack allocation, so the stack allocator object 2625 // must be present 2626 assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)"); 2627 assert(_fpu_stack_allocator != NULL, "must be present"); 2628 opr = _fpu_stack_allocator->to_fpu_stack(opr); 2629 #endif 2630 2631 Location::Type loc_type = float_saved_as_double ? Location::float_in_dbl : Location::normal; 2632 VMReg rname = frame_map()->fpu_regname(opr->fpu_regnr()); 2633 #ifndef __SOFTFP__ 2634 #ifndef VM_LITTLE_ENDIAN 2635 if (! float_saved_as_double) { 2636 // On big endian system, we may have an issue if float registers use only 2637 // the low half of the (same) double registers. 2638 // Both the float and the double could have the same regnr but would correspond 2639 // to two different addresses once saved. 2640 2641 // get next safely (no assertion checks) 2642 VMReg next = VMRegImpl::as_VMReg(1+rname->value()); 2643 if (next->is_reg() && 2644 (next->as_FloatRegister() == rname->as_FloatRegister())) { 2645 // the back-end does use the same numbering for the double and the float 2646 rname = next; // VMReg for the low bits, e.g. the real VMReg for the float 2647 } 2648 } 2649 #endif 2650 #endif 2651 LocationValue* sv = new LocationValue(Location::new_reg_loc(loc_type, rname)); 2652 2653 scope_values->append(sv); 2654 return 1; 2655 2656 } else { 2657 // double-size operands 2658 2659 ScopeValue* first; 2660 ScopeValue* second; 2661 2662 if (opr->is_double_stack()) { 2663 #ifdef _LP64 2664 Location loc1; 2665 Location::Type loc_type = opr->type() == T_LONG ? Location::lng : Location::dbl; 2666 if (!frame_map()->locations_for_slot(opr->double_stack_ix(), loc_type, &loc1, NULL)) { 2667 bailout("too large frame"); 2668 } 2669 // Does this reverse on x86 vs. sparc? 2670 first = new LocationValue(loc1); 2671 second = _int_0_scope_value; 2672 #else 2673 Location loc1, loc2; 2674 if (!frame_map()->locations_for_slot(opr->double_stack_ix(), Location::normal, &loc1, &loc2)) { 2675 bailout("too large frame"); 2676 } 2677 first = new LocationValue(loc1); 2678 second = new LocationValue(loc2); 2679 #endif // _LP64 2680 2681 } else if (opr->is_double_cpu()) { 2682 #ifdef _LP64 2683 VMReg rname_first = opr->as_register_lo()->as_VMReg(); 2684 first = new LocationValue(Location::new_reg_loc(Location::lng, rname_first)); 2685 second = _int_0_scope_value; 2686 #else 2687 VMReg rname_first = opr->as_register_lo()->as_VMReg(); 2688 VMReg rname_second = opr->as_register_hi()->as_VMReg(); 2689 2690 if (hi_word_offset_in_bytes < lo_word_offset_in_bytes) { 2691 // lo/hi and swapped relative to first and second, so swap them 2692 VMReg tmp = rname_first; 2693 rname_first = rname_second; 2694 rname_second = tmp; 2695 } 2696 2697 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); 2698 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); 2699 #endif //_LP64 2700 2701 2702 #ifdef X86 2703 } else if (opr->is_double_xmm()) { 2704 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation"); 2705 VMReg rname_first = opr->as_xmm_double_reg()->as_VMReg(); 2706 # ifdef _LP64 2707 first = new LocationValue(Location::new_reg_loc(Location::dbl, rname_first)); 2708 second = _int_0_scope_value; 2709 # else 2710 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); 2711 // %%% This is probably a waste but we'll keep things as they were for now 2712 if (true) { 2713 VMReg rname_second = rname_first->next(); 2714 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); 2715 } 2716 # endif 2717 #endif 2718 2719 } else if (opr->is_double_fpu()) { 2720 // On SPARC, fpu_regnrLo/fpu_regnrHi represents the two halves of 2721 // the double as float registers in the native ordering. On X86, 2722 // fpu_regnrLo is a FPU stack slot whose VMReg represents 2723 // the low-order word of the double and fpu_regnrLo + 1 is the 2724 // name for the other half. *first and *second must represent the 2725 // least and most significant words, respectively. 2726 2727 #ifdef X86 2728 // the exact location of fpu stack values is only known 2729 // during fpu stack allocation, so the stack allocator object 2730 // must be present 2731 assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)"); 2732 assert(_fpu_stack_allocator != NULL, "must be present"); 2733 opr = _fpu_stack_allocator->to_fpu_stack(opr); 2734 2735 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation (only fpu_regnrLo is used)"); 2736 #endif 2737 #ifdef SPARC 2738 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi() + 1, "assumed in calculation (only fpu_regnrHi is used)"); 2739 #endif 2740 #ifdef ARM 2741 assert(opr->fpu_regnrHi() == opr->fpu_regnrLo() + 1, "assumed in calculation (only fpu_regnrLo is used)"); 2742 #endif 2743 #ifdef PPC 2744 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation (only fpu_regnrHi is used)"); 2745 #endif 2746 2747 #ifdef VM_LITTLE_ENDIAN 2748 VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrLo()); 2749 #else 2750 VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrHi()); 2751 #endif 2752 2753 #ifdef _LP64 2754 first = new LocationValue(Location::new_reg_loc(Location::dbl, rname_first)); 2755 second = _int_0_scope_value; 2756 #else 2757 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); 2758 // %%% This is probably a waste but we'll keep things as they were for now 2759 if (true) { 2760 VMReg rname_second = rname_first->next(); 2761 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); 2762 } 2763 #endif 2764 2765 } else { 2766 ShouldNotReachHere(); 2767 first = NULL; 2768 second = NULL; 2769 } 2770 2771 assert(first != NULL && second != NULL, "must be set"); 2772 // The convention the interpreter uses is that the second local 2773 // holds the first raw word of the native double representation. 2774 // This is actually reasonable, since locals and stack arrays 2775 // grow downwards in all implementations. 2776 // (If, on some machine, the interpreter's Java locals or stack 2777 // were to grow upwards, the embedded doubles would be word-swapped.) 2778 scope_values->append(second); 2779 scope_values->append(first); 2780 return 2; 2781 } 2782 } 2783 2784 2785 int LinearScan::append_scope_value(int op_id, Value value, GrowableArray<ScopeValue*>* scope_values) { 2786 if (value != NULL) { 2787 LIR_Opr opr = value->operand(); 2788 Constant* con = value->as_Constant(); 2789 2790 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)"); 2791 assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands"); 2792 2793 if (con != NULL && !con->is_pinned() && !opr->is_constant()) { 2794 // Unpinned constants may have a virtual operand for a part of the lifetime 2795 // or may be illegal when it was optimized away, 2796 // so always use a constant operand 2797 opr = LIR_OprFact::value_type(con->type()); 2798 } 2799 assert(opr->is_virtual() || opr->is_constant(), "other cases not allowed here"); 2800 2801 if (opr->is_virtual()) { 2802 LIR_OpVisitState::OprMode mode = LIR_OpVisitState::inputMode; 2803 2804 BlockBegin* block = block_of_op_with_id(op_id); 2805 if (block->number_of_sux() == 1 && op_id == block->last_lir_instruction_id()) { 2806 // generating debug information for the last instruction of a block. 2807 // if this instruction is a branch, spill moves are inserted before this branch 2808 // and so the wrong operand would be returned (spill moves at block boundaries are not 2809 // considered in the live ranges of intervals) 2810 // Solution: use the first op_id of the branch target block instead. 2811 if (block->lir()->instructions_list()->last()->as_OpBranch() != NULL) { 2812 if (block->live_out().at(opr->vreg_number())) { 2813 op_id = block->sux_at(0)->first_lir_instruction_id(); 2814 mode = LIR_OpVisitState::outputMode; 2815 } 2816 } 2817 } 2818 2819 // Get current location of operand 2820 // The operand must be live because debug information is considered when building the intervals 2821 // if the interval is not live, color_lir_opr will cause an assertion failure 2822 opr = color_lir_opr(opr, op_id, mode); 2823 assert(!has_call(op_id) || opr->is_stack() || !is_caller_save(reg_num(opr)), "can not have caller-save register operands at calls"); 2824 2825 // Append to ScopeValue array 2826 return append_scope_value_for_operand(opr, scope_values); 2827 2828 } else { 2829 assert(value->as_Constant() != NULL, "all other instructions have only virtual operands"); 2830 assert(opr->is_constant(), "operand must be constant"); 2831 2832 return append_scope_value_for_constant(opr, scope_values); 2833 } 2834 } else { 2835 // append a dummy value because real value not needed 2836 scope_values->append(_illegal_value); 2837 return 1; 2838 } 2839 } 2840 2841 2842 IRScopeDebugInfo* LinearScan::compute_debug_info_for_scope(int op_id, IRScope* cur_scope, ValueStack* cur_state, ValueStack* innermost_state) { 2843 IRScopeDebugInfo* caller_debug_info = NULL; 2844 2845 ValueStack* caller_state = cur_state->caller_state(); 2846 if (caller_state != NULL) { 2847 // process recursively to compute outermost scope first 2848 caller_debug_info = compute_debug_info_for_scope(op_id, cur_scope->caller(), caller_state, innermost_state); 2849 } 2850 2851 // initialize these to null. 2852 // If we don't need deopt info or there are no locals, expressions or monitors, 2853 // then these get recorded as no information and avoids the allocation of 0 length arrays. 2854 GrowableArray<ScopeValue*>* locals = NULL; 2855 GrowableArray<ScopeValue*>* expressions = NULL; 2856 GrowableArray<MonitorValue*>* monitors = NULL; 2857 2858 // describe local variable values 2859 int nof_locals = cur_state->locals_size(); 2860 if (nof_locals > 0) { 2861 locals = new GrowableArray<ScopeValue*>(nof_locals); 2862 2863 int pos = 0; 2864 while (pos < nof_locals) { 2865 assert(pos < cur_state->locals_size(), "why not?"); 2866 2867 Value local = cur_state->local_at(pos); 2868 pos += append_scope_value(op_id, local, locals); 2869 2870 assert(locals->length() == pos, "must match"); 2871 } 2872 assert(locals->length() == cur_scope->method()->max_locals(), "wrong number of locals"); 2873 assert(locals->length() == cur_state->locals_size(), "wrong number of locals"); 2874 } else if (cur_scope->method()->max_locals() > 0) { 2875 assert(cur_state->kind() == ValueStack::EmptyExceptionState, "should be"); 2876 nof_locals = cur_scope->method()->max_locals(); 2877 locals = new GrowableArray<ScopeValue*>(nof_locals); 2878 for(int i = 0; i < nof_locals; i++) { 2879 locals->append(_illegal_value); 2880 } 2881 } 2882 2883 // describe expression stack 2884 int nof_stack = cur_state->stack_size(); 2885 if (nof_stack > 0) { 2886 expressions = new GrowableArray<ScopeValue*>(nof_stack); 2887 2888 int pos = 0; 2889 while (pos < nof_stack) { 2890 Value expression = cur_state->stack_at_inc(pos); 2891 append_scope_value(op_id, expression, expressions); 2892 2893 assert(expressions->length() == pos, "must match"); 2894 } 2895 assert(expressions->length() == cur_state->stack_size(), "wrong number of stack entries"); 2896 } 2897 2898 // describe monitors 2899 int nof_locks = cur_state->locks_size(); 2900 if (nof_locks > 0) { 2901 int lock_offset = cur_state->caller_state() != NULL ? cur_state->caller_state()->total_locks_size() : 0; 2902 monitors = new GrowableArray<MonitorValue*>(nof_locks); 2903 for (int i = 0; i < nof_locks; i++) { 2904 monitors->append(location_for_monitor_index(lock_offset + i)); 2905 } 2906 } 2907 2908 return new IRScopeDebugInfo(cur_scope, cur_state->bci(), locals, expressions, monitors, caller_debug_info); 2909 } 2910 2911 2912 void LinearScan::compute_debug_info(CodeEmitInfo* info, int op_id) { 2913 TRACE_LINEAR_SCAN(3, tty->print_cr("creating debug information at op_id %d", op_id)); 2914 2915 IRScope* innermost_scope = info->scope(); 2916 ValueStack* innermost_state = info->stack(); 2917 2918 assert(innermost_scope != NULL && innermost_state != NULL, "why is it missing?"); 2919 2920 DEBUG_ONLY(check_stack_depth(info, innermost_state->stack_size())); 2921 2922 if (info->_scope_debug_info == NULL) { 2923 // compute debug information 2924 info->_scope_debug_info = compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state); 2925 } else { 2926 // debug information already set. Check that it is correct from the current point of view 2927 DEBUG_ONLY(assert_equal(info->_scope_debug_info, compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state))); 2928 } 2929 } 2930 2931 2932 void LinearScan::assign_reg_num(LIR_OpList* instructions, IntervalWalker* iw) { 2933 LIR_OpVisitState visitor; 2934 int num_inst = instructions->length(); 2935 bool has_dead = false; 2936 2937 for (int j = 0; j < num_inst; j++) { 2938 LIR_Op* op = instructions->at(j); 2939 if (op == NULL) { // this can happen when spill-moves are removed in eliminate_spill_moves 2940 has_dead = true; 2941 continue; 2942 } 2943 int op_id = op->id(); 2944 2945 // visit instruction to get list of operands 2946 visitor.visit(op); 2947 2948 // iterate all modes of the visitor and process all virtual operands 2949 for_each_visitor_mode(mode) { 2950 int n = visitor.opr_count(mode); 2951 for (int k = 0; k < n; k++) { 2952 LIR_Opr opr = visitor.opr_at(mode, k); 2953 if (opr->is_virtual_register()) { 2954 visitor.set_opr_at(mode, k, color_lir_opr(opr, op_id, mode)); 2955 } 2956 } 2957 } 2958 2959 if (visitor.info_count() > 0) { 2960 // exception handling 2961 if (compilation()->has_exception_handlers()) { 2962 XHandlers* xhandlers = visitor.all_xhandler(); 2963 int n = xhandlers->length(); 2964 for (int k = 0; k < n; k++) { 2965 XHandler* handler = xhandlers->handler_at(k); 2966 if (handler->entry_code() != NULL) { 2967 assign_reg_num(handler->entry_code()->instructions_list(), NULL); 2968 } 2969 } 2970 } else { 2971 assert(visitor.all_xhandler()->length() == 0, "missed exception handler"); 2972 } 2973 2974 // compute oop map 2975 assert(iw != NULL, "needed for compute_oop_map"); 2976 compute_oop_map(iw, visitor, op); 2977 2978 // compute debug information 2979 if (!use_fpu_stack_allocation()) { 2980 // compute debug information if fpu stack allocation is not needed. 2981 // when fpu stack allocation is needed, the debug information can not 2982 // be computed here because the exact location of fpu operands is not known 2983 // -> debug information is created inside the fpu stack allocator 2984 int n = visitor.info_count(); 2985 for (int k = 0; k < n; k++) { 2986 compute_debug_info(visitor.info_at(k), op_id); 2987 } 2988 } 2989 } 2990 2991 #ifdef ASSERT 2992 // make sure we haven't made the op invalid. 2993 op->verify(); 2994 #endif 2995 2996 // remove useless moves 2997 if (op->code() == lir_move) { 2998 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); 2999 LIR_Op1* move = (LIR_Op1*)op; 3000 LIR_Opr src = move->in_opr(); 3001 LIR_Opr dst = move->result_opr(); 3002 if (dst == src || 3003 !dst->is_pointer() && !src->is_pointer() && 3004 src->is_same_register(dst)) { 3005 instructions->at_put(j, NULL); 3006 has_dead = true; 3007 } 3008 } 3009 } 3010 3011 if (has_dead) { 3012 // iterate all instructions of the block and remove all null-values. 3013 int insert_point = 0; 3014 for (int j = 0; j < num_inst; j++) { 3015 LIR_Op* op = instructions->at(j); 3016 if (op != NULL) { 3017 if (insert_point != j) { 3018 instructions->at_put(insert_point, op); 3019 } 3020 insert_point++; 3021 } 3022 } 3023 instructions->truncate(insert_point); 3024 } 3025 } 3026 3027 void LinearScan::assign_reg_num() { 3028 TIME_LINEAR_SCAN(timer_assign_reg_num); 3029 3030 init_compute_debug_info(); 3031 IntervalWalker* iw = init_compute_oop_maps(); 3032 3033 int num_blocks = block_count(); 3034 for (int i = 0; i < num_blocks; i++) { 3035 BlockBegin* block = block_at(i); 3036 assign_reg_num(block->lir()->instructions_list(), iw); 3037 } 3038 } 3039 3040 3041 void LinearScan::do_linear_scan() { 3042 NOT_PRODUCT(_total_timer.begin_method()); 3043 3044 number_instructions(); 3045 3046 NOT_PRODUCT(print_lir(1, "Before Register Allocation")); 3047 3048 compute_local_live_sets(); 3049 compute_global_live_sets(); 3050 CHECK_BAILOUT(); 3051 3052 build_intervals(); 3053 CHECK_BAILOUT(); 3054 sort_intervals_before_allocation(); 3055 3056 NOT_PRODUCT(print_intervals("Before Register Allocation")); 3057 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_before_alloc)); 3058 3059 allocate_registers(); 3060 CHECK_BAILOUT(); 3061 3062 resolve_data_flow(); 3063 if (compilation()->has_exception_handlers()) { 3064 resolve_exception_handlers(); 3065 } 3066 // fill in number of spill slots into frame_map 3067 propagate_spill_slots(); 3068 CHECK_BAILOUT(); 3069 3070 NOT_PRODUCT(print_intervals("After Register Allocation")); 3071 NOT_PRODUCT(print_lir(2, "LIR after register allocation:")); 3072 3073 sort_intervals_after_allocation(); 3074 3075 DEBUG_ONLY(verify()); 3076 3077 eliminate_spill_moves(); 3078 assign_reg_num(); 3079 CHECK_BAILOUT(); 3080 3081 NOT_PRODUCT(print_lir(2, "LIR after assignment of register numbers:")); 3082 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_after_asign)); 3083 3084 { TIME_LINEAR_SCAN(timer_allocate_fpu_stack); 3085 3086 if (use_fpu_stack_allocation()) { 3087 allocate_fpu_stack(); // Only has effect on Intel 3088 NOT_PRODUCT(print_lir(2, "LIR after FPU stack allocation:")); 3089 } 3090 } 3091 3092 { TIME_LINEAR_SCAN(timer_optimize_lir); 3093 3094 EdgeMoveOptimizer::optimize(ir()->code()); 3095 ControlFlowOptimizer::optimize(ir()->code()); 3096 // check that cfg is still correct after optimizations 3097 ir()->verify(); 3098 } 3099 3100 NOT_PRODUCT(print_lir(1, "Before Code Generation", false)); 3101 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_final)); 3102 NOT_PRODUCT(_total_timer.end_method(this)); 3103 } 3104 3105 3106 // ********** Printing functions 3107 3108 #ifndef PRODUCT 3109 3110 void LinearScan::print_timers(double total) { 3111 _total_timer.print(total); 3112 } 3113 3114 void LinearScan::print_statistics() { 3115 _stat_before_alloc.print("before allocation"); 3116 _stat_after_asign.print("after assignment of register"); 3117 _stat_final.print("after optimization"); 3118 } 3119 3120 void LinearScan::print_bitmap(BitMap& b) { 3121 for (unsigned int i = 0; i < b.size(); i++) { 3122 if (b.at(i)) tty->print("%d ", i); 3123 } 3124 tty->cr(); 3125 } 3126 3127 void LinearScan::print_intervals(const char* label) { 3128 if (TraceLinearScanLevel >= 1) { 3129 int i; 3130 tty->cr(); 3131 tty->print_cr("%s", label); 3132 3133 for (i = 0; i < interval_count(); i++) { 3134 Interval* interval = interval_at(i); 3135 if (interval != NULL) { 3136 interval->print(); 3137 } 3138 } 3139 3140 tty->cr(); 3141 tty->print_cr("--- Basic Blocks ---"); 3142 for (i = 0; i < block_count(); i++) { 3143 BlockBegin* block = block_at(i); 3144 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()); 3145 } 3146 tty->cr(); 3147 tty->cr(); 3148 } 3149 3150 if (PrintCFGToFile) { 3151 CFGPrinter::print_intervals(&_intervals, label); 3152 } 3153 } 3154 3155 void LinearScan::print_lir(int level, const char* label, bool hir_valid) { 3156 if (TraceLinearScanLevel >= level) { 3157 tty->cr(); 3158 tty->print_cr("%s", label); 3159 print_LIR(ir()->linear_scan_order()); 3160 tty->cr(); 3161 } 3162 3163 if (level == 1 && PrintCFGToFile) { 3164 CFGPrinter::print_cfg(ir()->linear_scan_order(), label, hir_valid, true); 3165 } 3166 } 3167 3168 #endif //PRODUCT 3169 3170 3171 // ********** verification functions for allocation 3172 // (check that all intervals have a correct register and that no registers are overwritten) 3173 #ifdef ASSERT 3174 3175 void LinearScan::verify() { 3176 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying intervals ******************************************")); 3177 verify_intervals(); 3178 3179 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that no oops are in fixed intervals ****************")); 3180 verify_no_oops_in_fixed_intervals(); 3181 3182 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that unpinned constants are not alive across block boundaries")); 3183 verify_constants(); 3184 3185 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying register allocation ********************************")); 3186 verify_registers(); 3187 3188 TRACE_LINEAR_SCAN(2, tty->print_cr("********* no errors found **********************************************")); 3189 } 3190 3191 void LinearScan::verify_intervals() { 3192 int len = interval_count(); 3193 bool has_error = false; 3194 3195 for (int i = 0; i < len; i++) { 3196 Interval* i1 = interval_at(i); 3197 if (i1 == NULL) continue; 3198 3199 i1->check_split_children(); 3200 3201 if (i1->reg_num() != i) { 3202 tty->print_cr("Interval %d is on position %d in list", i1->reg_num(), i); i1->print(); tty->cr(); 3203 has_error = true; 3204 } 3205 3206 if (i1->reg_num() >= LIR_OprDesc::vreg_base && i1->type() == T_ILLEGAL) { 3207 tty->print_cr("Interval %d has no type assigned", i1->reg_num()); i1->print(); tty->cr(); 3208 has_error = true; 3209 } 3210 3211 if (i1->assigned_reg() == any_reg) { 3212 tty->print_cr("Interval %d has no register assigned", i1->reg_num()); i1->print(); tty->cr(); 3213 has_error = true; 3214 } 3215 3216 if (i1->assigned_reg() == i1->assigned_regHi()) { 3217 tty->print_cr("Interval %d: low and high register equal", i1->reg_num()); i1->print(); tty->cr(); 3218 has_error = true; 3219 } 3220 3221 if (!is_processed_reg_num(i1->assigned_reg())) { 3222 tty->print_cr("Can not have an Interval for an ignored register"); i1->print(); tty->cr(); 3223 has_error = true; 3224 } 3225 3226 if (i1->first() == Range::end()) { 3227 tty->print_cr("Interval %d has no Range", i1->reg_num()); i1->print(); tty->cr(); 3228 has_error = true; 3229 } 3230 3231 for (Range* r = i1->first(); r != Range::end(); r = r->next()) { 3232 if (r->from() >= r->to()) { 3233 tty->print_cr("Interval %d has zero length range", i1->reg_num()); i1->print(); tty->cr(); 3234 has_error = true; 3235 } 3236 } 3237 3238 for (int j = i + 1; j < len; j++) { 3239 Interval* i2 = interval_at(j); 3240 if (i2 == NULL) continue; 3241 3242 // special intervals that are created in MoveResolver 3243 // -> ignore them because the range information has no meaning there 3244 if (i1->from() == 1 && i1->to() == 2) continue; 3245 if (i2->from() == 1 && i2->to() == 2) continue; 3246 3247 int r1 = i1->assigned_reg(); 3248 int r1Hi = i1->assigned_regHi(); 3249 int r2 = i2->assigned_reg(); 3250 int r2Hi = i2->assigned_regHi(); 3251 if (i1->intersects(i2) && (r1 == r2 || r1 == r2Hi || (r1Hi != any_reg && (r1Hi == r2 || r1Hi == r2Hi)))) { 3252 tty->print_cr("Intervals %d and %d overlap and have the same register assigned", i1->reg_num(), i2->reg_num()); 3253 i1->print(); tty->cr(); 3254 i2->print(); tty->cr(); 3255 has_error = true; 3256 } 3257 } 3258 } 3259 3260 assert(has_error == false, "register allocation invalid"); 3261 } 3262 3263 3264 void LinearScan::verify_no_oops_in_fixed_intervals() { 3265 Interval* fixed_intervals; 3266 Interval* other_intervals; 3267 create_unhandled_lists(&fixed_intervals, &other_intervals, is_precolored_cpu_interval, NULL); 3268 3269 // to ensure a walking until the last instruction id, add a dummy interval 3270 // with a high operation id 3271 other_intervals = new Interval(any_reg); 3272 other_intervals->add_range(max_jint - 2, max_jint - 1); 3273 IntervalWalker* iw = new IntervalWalker(this, fixed_intervals, other_intervals); 3274 3275 LIR_OpVisitState visitor; 3276 for (int i = 0; i < block_count(); i++) { 3277 BlockBegin* block = block_at(i); 3278 3279 LIR_OpList* instructions = block->lir()->instructions_list(); 3280 3281 for (int j = 0; j < instructions->length(); j++) { 3282 LIR_Op* op = instructions->at(j); 3283 int op_id = op->id(); 3284 3285 visitor.visit(op); 3286 3287 if (visitor.info_count() > 0) { 3288 iw->walk_before(op->id()); 3289 bool check_live = true; 3290 if (op->code() == lir_move) { 3291 LIR_Op1* move = (LIR_Op1*)op; 3292 check_live = (move->patch_code() == lir_patch_none); 3293 } 3294 LIR_OpBranch* branch = op->as_OpBranch(); 3295 if (branch != NULL && branch->stub() != NULL && branch->stub()->is_exception_throw_stub()) { 3296 // Don't bother checking the stub in this case since the 3297 // exception stub will never return to normal control flow. 3298 check_live = false; 3299 } 3300 3301 // Make sure none of the fixed registers is live across an 3302 // oopmap since we can't handle that correctly. 3303 if (check_live) { 3304 for (Interval* interval = iw->active_first(fixedKind); 3305 interval != Interval::end(); 3306 interval = interval->next()) { 3307 if (interval->current_to() > op->id() + 1) { 3308 // This interval is live out of this op so make sure 3309 // that this interval represents some value that's 3310 // referenced by this op either as an input or output. 3311 bool ok = false; 3312 for_each_visitor_mode(mode) { 3313 int n = visitor.opr_count(mode); 3314 for (int k = 0; k < n; k++) { 3315 LIR_Opr opr = visitor.opr_at(mode, k); 3316 if (opr->is_fixed_cpu()) { 3317 if (interval_at(reg_num(opr)) == interval) { 3318 ok = true; 3319 break; 3320 } 3321 int hi = reg_numHi(opr); 3322 if (hi != -1 && interval_at(hi) == interval) { 3323 ok = true; 3324 break; 3325 } 3326 } 3327 } 3328 } 3329 assert(ok, "fixed intervals should never be live across an oopmap point"); 3330 } 3331 } 3332 } 3333 } 3334 3335 // oop-maps at calls do not contain registers, so check is not needed 3336 if (!visitor.has_call()) { 3337 3338 for_each_visitor_mode(mode) { 3339 int n = visitor.opr_count(mode); 3340 for (int k = 0; k < n; k++) { 3341 LIR_Opr opr = visitor.opr_at(mode, k); 3342 3343 if (opr->is_fixed_cpu() && opr->is_oop()) { 3344 // operand is a non-virtual cpu register and contains an oop 3345 TRACE_LINEAR_SCAN(4, op->print_on(tty); tty->print("checking operand "); opr->print(); tty->cr()); 3346 3347 Interval* interval = interval_at(reg_num(opr)); 3348 assert(interval != NULL, "no interval"); 3349 3350 if (mode == LIR_OpVisitState::inputMode) { 3351 if (interval->to() >= op_id + 1) { 3352 assert(interval->to() < op_id + 2 || 3353 interval->has_hole_between(op_id, op_id + 2), 3354 "oop input operand live after instruction"); 3355 } 3356 } else if (mode == LIR_OpVisitState::outputMode) { 3357 if (interval->from() <= op_id - 1) { 3358 assert(interval->has_hole_between(op_id - 1, op_id), 3359 "oop input operand live after instruction"); 3360 } 3361 } 3362 } 3363 } 3364 } 3365 } 3366 } 3367 } 3368 } 3369 3370 3371 void LinearScan::verify_constants() { 3372 int num_regs = num_virtual_regs(); 3373 int size = live_set_size(); 3374 int num_blocks = block_count(); 3375 3376 for (int i = 0; i < num_blocks; i++) { 3377 BlockBegin* block = block_at(i); 3378 BitMap live_at_edge = block->live_in(); 3379 3380 // visit all registers where the live_at_edge bit is set 3381 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)) { 3382 TRACE_LINEAR_SCAN(4, tty->print("checking interval %d of block B%d", r, block->block_id())); 3383 3384 Value value = gen()->instruction_for_vreg(r); 3385 3386 assert(value != NULL, "all intervals live across block boundaries must have Value"); 3387 assert(value->operand()->is_register() && value->operand()->is_virtual(), "value must have virtual operand"); 3388 assert(value->operand()->vreg_number() == r, "register number must match"); 3389 // TKR assert(value->as_Constant() == NULL || value->is_pinned(), "only pinned constants can be alive accross block boundaries"); 3390 } 3391 } 3392 } 3393 3394 3395 class RegisterVerifier: public StackObj { 3396 private: 3397 LinearScan* _allocator; 3398 BlockList _work_list; // all blocks that must be processed 3399 IntervalsList _saved_states; // saved information of previous check 3400 3401 // simplified access to methods of LinearScan 3402 Compilation* compilation() const { return _allocator->compilation(); } 3403 Interval* interval_at(int reg_num) const { return _allocator->interval_at(reg_num); } 3404 int reg_num(LIR_Opr opr) const { return _allocator->reg_num(opr); } 3405 3406 // currently, only registers are processed 3407 int state_size() { return LinearScan::nof_regs; } 3408 3409 // accessors 3410 IntervalList* state_for_block(BlockBegin* block) { return _saved_states.at(block->block_id()); } 3411 void set_state_for_block(BlockBegin* block, IntervalList* saved_state) { _saved_states.at_put(block->block_id(), saved_state); } 3412 void add_to_work_list(BlockBegin* block) { if (!_work_list.contains(block)) _work_list.append(block); } 3413 3414 // helper functions 3415 IntervalList* copy(IntervalList* input_state); 3416 void state_put(IntervalList* input_state, int reg, Interval* interval); 3417 bool check_state(IntervalList* input_state, int reg, Interval* interval); 3418 3419 void process_block(BlockBegin* block); 3420 void process_xhandler(XHandler* xhandler, IntervalList* input_state); 3421 void process_successor(BlockBegin* block, IntervalList* input_state); 3422 void process_operations(LIR_List* ops, IntervalList* input_state); 3423 3424 public: 3425 RegisterVerifier(LinearScan* allocator) 3426 : _allocator(allocator) 3427 , _work_list(16) 3428 , _saved_states(BlockBegin::number_of_blocks(), NULL) 3429 { } 3430 3431 void verify(BlockBegin* start); 3432 }; 3433 3434 3435 // entry function from LinearScan that starts the verification 3436 void LinearScan::verify_registers() { 3437 RegisterVerifier verifier(this); 3438 verifier.verify(block_at(0)); 3439 } 3440 3441 3442 void RegisterVerifier::verify(BlockBegin* start) { 3443 // setup input registers (method arguments) for first block 3444 IntervalList* input_state = new IntervalList(state_size(), NULL); 3445 CallingConvention* args = compilation()->frame_map()->incoming_arguments(); 3446 for (int n = 0; n < args->length(); n++) { 3447 LIR_Opr opr = args->at(n); 3448 if (opr->is_register()) { 3449 Interval* interval = interval_at(reg_num(opr)); 3450 3451 if (interval->assigned_reg() < state_size()) { 3452 input_state->at_put(interval->assigned_reg(), interval); 3453 } 3454 if (interval->assigned_regHi() != LinearScan::any_reg && interval->assigned_regHi() < state_size()) { 3455 input_state->at_put(interval->assigned_regHi(), interval); 3456 } 3457 } 3458 } 3459 3460 set_state_for_block(start, input_state); 3461 add_to_work_list(start); 3462 3463 // main loop for verification 3464 do { 3465 BlockBegin* block = _work_list.at(0); 3466 _work_list.remove_at(0); 3467 3468 process_block(block); 3469 } while (!_work_list.is_empty()); 3470 } 3471 3472 void RegisterVerifier::process_block(BlockBegin* block) { 3473 TRACE_LINEAR_SCAN(2, tty->cr(); tty->print_cr("process_block B%d", block->block_id())); 3474 3475 // must copy state because it is modified 3476 IntervalList* input_state = copy(state_for_block(block)); 3477 3478 if (TraceLinearScanLevel >= 4) { 3479 tty->print_cr("Input-State of intervals:"); 3480 tty->print(" "); 3481 for (int i = 0; i < state_size(); i++) { 3482 if (input_state->at(i) != NULL) { 3483 tty->print(" %4d", input_state->at(i)->reg_num()); 3484 } else { 3485 tty->print(" __"); 3486 } 3487 } 3488 tty->cr(); 3489 tty->cr(); 3490 } 3491 3492 // process all operations of the block 3493 process_operations(block->lir(), input_state); 3494 3495 // iterate all successors 3496 for (int i = 0; i < block->number_of_sux(); i++) { 3497 process_successor(block->sux_at(i), input_state); 3498 } 3499 } 3500 3501 void RegisterVerifier::process_xhandler(XHandler* xhandler, IntervalList* input_state) { 3502 TRACE_LINEAR_SCAN(2, tty->print_cr("process_xhandler B%d", xhandler->entry_block()->block_id())); 3503 3504 // must copy state because it is modified 3505 input_state = copy(input_state); 3506 3507 if (xhandler->entry_code() != NULL) { 3508 process_operations(xhandler->entry_code(), input_state); 3509 } 3510 process_successor(xhandler->entry_block(), input_state); 3511 } 3512 3513 void RegisterVerifier::process_successor(BlockBegin* block, IntervalList* input_state) { 3514 IntervalList* saved_state = state_for_block(block); 3515 3516 if (saved_state != NULL) { 3517 // this block was already processed before. 3518 // check if new input_state is consistent with saved_state 3519 3520 bool saved_state_correct = true; 3521 for (int i = 0; i < state_size(); i++) { 3522 if (input_state->at(i) != saved_state->at(i)) { 3523 // current input_state and previous saved_state assume a different 3524 // interval in this register -> assume that this register is invalid 3525 if (saved_state->at(i) != NULL) { 3526 // invalidate old calculation only if it assumed that 3527 // register was valid. when the register was already invalid, 3528 // then the old calculation was correct. 3529 saved_state_correct = false; 3530 saved_state->at_put(i, NULL); 3531 3532 TRACE_LINEAR_SCAN(4, tty->print_cr("process_successor B%d: invalidating slot %d", block->block_id(), i)); 3533 } 3534 } 3535 } 3536 3537 if (saved_state_correct) { 3538 // already processed block with correct input_state 3539 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: previous visit already correct", block->block_id())); 3540 } else { 3541 // must re-visit this block 3542 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: must re-visit because input state changed", block->block_id())); 3543 add_to_work_list(block); 3544 } 3545 3546 } else { 3547 // block was not processed before, so set initial input_state 3548 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: initial visit", block->block_id())); 3549 3550 set_state_for_block(block, copy(input_state)); 3551 add_to_work_list(block); 3552 } 3553 } 3554 3555 3556 IntervalList* RegisterVerifier::copy(IntervalList* input_state) { 3557 IntervalList* copy_state = new IntervalList(input_state->length()); 3558 copy_state->push_all(input_state); 3559 return copy_state; 3560 } 3561 3562 void RegisterVerifier::state_put(IntervalList* input_state, int reg, Interval* interval) { 3563 if (reg != LinearScan::any_reg && reg < state_size()) { 3564 if (interval != NULL) { 3565 TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = %d", reg, interval->reg_num())); 3566 } else if (input_state->at(reg) != NULL) { 3567 TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = NULL", reg)); 3568 } 3569 3570 input_state->at_put(reg, interval); 3571 } 3572 } 3573 3574 bool RegisterVerifier::check_state(IntervalList* input_state, int reg, Interval* interval) { 3575 if (reg != LinearScan::any_reg && reg < state_size()) { 3576 if (input_state->at(reg) != interval) { 3577 tty->print_cr("!! Error in register allocation: register %d does not contain interval %d", reg, interval->reg_num()); 3578 return true; 3579 } 3580 } 3581 return false; 3582 } 3583 3584 void RegisterVerifier::process_operations(LIR_List* ops, IntervalList* input_state) { 3585 // visit all instructions of the block 3586 LIR_OpVisitState visitor; 3587 bool has_error = false; 3588 3589 for (int i = 0; i < ops->length(); i++) { 3590 LIR_Op* op = ops->at(i); 3591 visitor.visit(op); 3592 3593 TRACE_LINEAR_SCAN(4, op->print_on(tty)); 3594 3595 // check if input operands are correct 3596 int j; 3597 int n = visitor.opr_count(LIR_OpVisitState::inputMode); 3598 for (j = 0; j < n; j++) { 3599 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, j); 3600 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { 3601 Interval* interval = interval_at(reg_num(opr)); 3602 if (op->id() != -1) { 3603 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::inputMode); 3604 } 3605 3606 has_error |= check_state(input_state, interval->assigned_reg(), interval->split_parent()); 3607 has_error |= check_state(input_state, interval->assigned_regHi(), interval->split_parent()); 3608 3609 // When an operand is marked with is_last_use, then the fpu stack allocator 3610 // removes the register from the fpu stack -> the register contains no value 3611 if (opr->is_last_use()) { 3612 state_put(input_state, interval->assigned_reg(), NULL); 3613 state_put(input_state, interval->assigned_regHi(), NULL); 3614 } 3615 } 3616 } 3617 3618 // invalidate all caller save registers at calls 3619 if (visitor.has_call()) { 3620 for (j = 0; j < FrameMap::nof_caller_save_cpu_regs(); j++) { 3621 state_put(input_state, reg_num(FrameMap::caller_save_cpu_reg_at(j)), NULL); 3622 } 3623 for (j = 0; j < FrameMap::nof_caller_save_fpu_regs; j++) { 3624 state_put(input_state, reg_num(FrameMap::caller_save_fpu_reg_at(j)), NULL); 3625 } 3626 3627 #ifdef X86 3628 for (j = 0; j < FrameMap::nof_caller_save_xmm_regs; j++) { 3629 state_put(input_state, reg_num(FrameMap::caller_save_xmm_reg_at(j)), NULL); 3630 } 3631 #endif 3632 } 3633 3634 // process xhandler before output and temp operands 3635 XHandlers* xhandlers = visitor.all_xhandler(); 3636 n = xhandlers->length(); 3637 for (int k = 0; k < n; k++) { 3638 process_xhandler(xhandlers->handler_at(k), input_state); 3639 } 3640 3641 // set temp operands (some operations use temp operands also as output operands, so can't set them NULL) 3642 n = visitor.opr_count(LIR_OpVisitState::tempMode); 3643 for (j = 0; j < n; j++) { 3644 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, j); 3645 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { 3646 Interval* interval = interval_at(reg_num(opr)); 3647 if (op->id() != -1) { 3648 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::tempMode); 3649 } 3650 3651 state_put(input_state, interval->assigned_reg(), interval->split_parent()); 3652 state_put(input_state, interval->assigned_regHi(), interval->split_parent()); 3653 } 3654 } 3655 3656 // set output operands 3657 n = visitor.opr_count(LIR_OpVisitState::outputMode); 3658 for (j = 0; j < n; j++) { 3659 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, j); 3660 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { 3661 Interval* interval = interval_at(reg_num(opr)); 3662 if (op->id() != -1) { 3663 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::outputMode); 3664 } 3665 3666 state_put(input_state, interval->assigned_reg(), interval->split_parent()); 3667 state_put(input_state, interval->assigned_regHi(), interval->split_parent()); 3668 } 3669 } 3670 } 3671 assert(has_error == false, "Error in register allocation"); 3672 } 3673 3674 #endif // ASSERT 3675 3676 3677 3678 // **** Implementation of MoveResolver ****************************** 3679 3680 MoveResolver::MoveResolver(LinearScan* allocator) : 3681 _allocator(allocator), 3682 _multiple_reads_allowed(false), 3683 _mapping_from(8), 3684 _mapping_from_opr(8), 3685 _mapping_to(8), 3686 _insert_list(NULL), 3687 _insert_idx(-1), 3688 _insertion_buffer() 3689 { 3690 for (int i = 0; i < LinearScan::nof_regs; i++) { 3691 _register_blocked[i] = 0; 3692 } 3693 DEBUG_ONLY(check_empty()); 3694 } 3695 3696 3697 #ifdef ASSERT 3698 3699 void MoveResolver::check_empty() { 3700 assert(_mapping_from.length() == 0 && _mapping_from_opr.length() == 0 && _mapping_to.length() == 0, "list must be empty before and after processing"); 3701 for (int i = 0; i < LinearScan::nof_regs; i++) { 3702 assert(register_blocked(i) == 0, "register map must be empty before and after processing"); 3703 } 3704 assert(_multiple_reads_allowed == false, "must have default value"); 3705 } 3706 3707 void MoveResolver::verify_before_resolve() { 3708 assert(_mapping_from.length() == _mapping_from_opr.length(), "length must be equal"); 3709 assert(_mapping_from.length() == _mapping_to.length(), "length must be equal"); 3710 assert(_insert_list != NULL && _insert_idx != -1, "insert position not set"); 3711 3712 int i, j; 3713 if (!_multiple_reads_allowed) { 3714 for (i = 0; i < _mapping_from.length(); i++) { 3715 for (j = i + 1; j < _mapping_from.length(); j++) { 3716 assert(_mapping_from.at(i) == NULL || _mapping_from.at(i) != _mapping_from.at(j), "cannot read from same interval twice"); 3717 } 3718 } 3719 } 3720 3721 for (i = 0; i < _mapping_to.length(); i++) { 3722 for (j = i + 1; j < _mapping_to.length(); j++) { 3723 assert(_mapping_to.at(i) != _mapping_to.at(j), "cannot write to same interval twice"); 3724 } 3725 } 3726 3727 3728 BitMap used_regs(LinearScan::nof_regs + allocator()->frame_map()->argcount() + allocator()->max_spills()); 3729 used_regs.clear(); 3730 if (!_multiple_reads_allowed) { 3731 for (i = 0; i < _mapping_from.length(); i++) { 3732 Interval* it = _mapping_from.at(i); 3733 if (it != NULL) { 3734 assert(!used_regs.at(it->assigned_reg()), "cannot read from same register twice"); 3735 used_regs.set_bit(it->assigned_reg()); 3736 3737 if (it->assigned_regHi() != LinearScan::any_reg) { 3738 assert(!used_regs.at(it->assigned_regHi()), "cannot read from same register twice"); 3739 used_regs.set_bit(it->assigned_regHi()); 3740 } 3741 } 3742 } 3743 } 3744 3745 used_regs.clear(); 3746 for (i = 0; i < _mapping_to.length(); i++) { 3747 Interval* it = _mapping_to.at(i); 3748 assert(!used_regs.at(it->assigned_reg()), "cannot write to same register twice"); 3749 used_regs.set_bit(it->assigned_reg()); 3750 3751 if (it->assigned_regHi() != LinearScan::any_reg) { 3752 assert(!used_regs.at(it->assigned_regHi()), "cannot write to same register twice"); 3753 used_regs.set_bit(it->assigned_regHi()); 3754 } 3755 } 3756 3757 used_regs.clear(); 3758 for (i = 0; i < _mapping_from.length(); i++) { 3759 Interval* it = _mapping_from.at(i); 3760 if (it != NULL && it->assigned_reg() >= LinearScan::nof_regs) { 3761 used_regs.set_bit(it->assigned_reg()); 3762 } 3763 } 3764 for (i = 0; i < _mapping_to.length(); i++) { 3765 Interval* it = _mapping_to.at(i); 3766 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"); 3767 } 3768 } 3769 3770 #endif // ASSERT 3771 3772 3773 // mark assigned_reg and assigned_regHi of the interval as blocked 3774 void MoveResolver::block_registers(Interval* it) { 3775 int reg = it->assigned_reg(); 3776 if (reg < LinearScan::nof_regs) { 3777 assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used"); 3778 set_register_blocked(reg, 1); 3779 } 3780 reg = it->assigned_regHi(); 3781 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { 3782 assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used"); 3783 set_register_blocked(reg, 1); 3784 } 3785 } 3786 3787 // mark assigned_reg and assigned_regHi of the interval as unblocked 3788 void MoveResolver::unblock_registers(Interval* it) { 3789 int reg = it->assigned_reg(); 3790 if (reg < LinearScan::nof_regs) { 3791 assert(register_blocked(reg) > 0, "register already marked as unused"); 3792 set_register_blocked(reg, -1); 3793 } 3794 reg = it->assigned_regHi(); 3795 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { 3796 assert(register_blocked(reg) > 0, "register already marked as unused"); 3797 set_register_blocked(reg, -1); 3798 } 3799 } 3800 3801 // check if assigned_reg and assigned_regHi of the to-interval are not blocked (or only blocked by from) 3802 bool MoveResolver::save_to_process_move(Interval* from, Interval* to) { 3803 int from_reg = -1; 3804 int from_regHi = -1; 3805 if (from != NULL) { 3806 from_reg = from->assigned_reg(); 3807 from_regHi = from->assigned_regHi(); 3808 } 3809 3810 int reg = to->assigned_reg(); 3811 if (reg < LinearScan::nof_regs) { 3812 if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) { 3813 return false; 3814 } 3815 } 3816 reg = to->assigned_regHi(); 3817 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { 3818 if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) { 3819 return false; 3820 } 3821 } 3822 3823 return true; 3824 } 3825 3826 3827 void MoveResolver::create_insertion_buffer(LIR_List* list) { 3828 assert(!_insertion_buffer.initialized(), "overwriting existing buffer"); 3829 _insertion_buffer.init(list); 3830 } 3831 3832 void MoveResolver::append_insertion_buffer() { 3833 if (_insertion_buffer.initialized()) { 3834 _insertion_buffer.lir_list()->append(&_insertion_buffer); 3835 } 3836 assert(!_insertion_buffer.initialized(), "must be uninitialized now"); 3837 3838 _insert_list = NULL; 3839 _insert_idx = -1; 3840 } 3841 3842 void MoveResolver::insert_move(Interval* from_interval, Interval* to_interval) { 3843 assert(from_interval->reg_num() != to_interval->reg_num(), "from and to interval equal"); 3844 assert(from_interval->type() == to_interval->type(), "move between different types"); 3845 assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first"); 3846 assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer"); 3847 3848 LIR_Opr from_opr = LIR_OprFact::virtual_register(from_interval->reg_num(), from_interval->type()); 3849 LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type()); 3850 3851 if (!_multiple_reads_allowed) { 3852 // the last_use flag is an optimization for FPU stack allocation. When the same 3853 // input interval is used in more than one move, then it is too difficult to determine 3854 // if this move is really the last use. 3855 from_opr = from_opr->make_last_use(); 3856 } 3857 _insertion_buffer.move(_insert_idx, from_opr, to_opr); 3858 3859 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())); 3860 } 3861 3862 void MoveResolver::insert_move(LIR_Opr from_opr, Interval* to_interval) { 3863 assert(from_opr->type() == to_interval->type(), "move between different types"); 3864 assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first"); 3865 assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer"); 3866 3867 LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type()); 3868 _insertion_buffer.move(_insert_idx, from_opr, to_opr); 3869 3870 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())); 3871 } 3872 3873 3874 void MoveResolver::resolve_mappings() { 3875 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)); 3876 DEBUG_ONLY(verify_before_resolve()); 3877 3878 // Block all registers that are used as input operands of a move. 3879 // When a register is blocked, no move to this register is emitted. 3880 // This is necessary for detecting cycles in moves. 3881 int i; 3882 for (i = _mapping_from.length() - 1; i >= 0; i--) { 3883 Interval* from_interval = _mapping_from.at(i); 3884 if (from_interval != NULL) { 3885 block_registers(from_interval); 3886 } 3887 } 3888 3889 int spill_candidate = -1; 3890 while (_mapping_from.length() > 0) { 3891 bool processed_interval = false; 3892 3893 for (i = _mapping_from.length() - 1; i >= 0; i--) { 3894 Interval* from_interval = _mapping_from.at(i); 3895 Interval* to_interval = _mapping_to.at(i); 3896 3897 if (save_to_process_move(from_interval, to_interval)) { 3898 // this inverval can be processed because target is free 3899 if (from_interval != NULL) { 3900 insert_move(from_interval, to_interval); 3901 unblock_registers(from_interval); 3902 } else { 3903 insert_move(_mapping_from_opr.at(i), to_interval); 3904 } 3905 _mapping_from.remove_at(i); 3906 _mapping_from_opr.remove_at(i); 3907 _mapping_to.remove_at(i); 3908 3909 processed_interval = true; 3910 } else if (from_interval != NULL && from_interval->assigned_reg() < LinearScan::nof_regs) { 3911 // this interval cannot be processed now because target is not free 3912 // it starts in a register, so it is a possible candidate for spilling 3913 spill_candidate = i; 3914 } 3915 } 3916 3917 if (!processed_interval) { 3918 // no move could be processed because there is a cycle in the move list 3919 // (e.g. r1 -> r2, r2 -> r1), so one interval must be spilled to memory 3920 assert(spill_candidate != -1, "no interval in register for spilling found"); 3921 3922 // create a new spill interval and assign a stack slot to it 3923 Interval* from_interval = _mapping_from.at(spill_candidate); 3924 Interval* spill_interval = new Interval(-1); 3925 spill_interval->set_type(from_interval->type()); 3926 3927 // add a dummy range because real position is difficult to calculate 3928 // Note: this range is a special case when the integrity of the allocation is checked 3929 spill_interval->add_range(1, 2); 3930 3931 // do not allocate a new spill slot for temporary interval, but 3932 // use spill slot assigned to from_interval. Otherwise moves from 3933 // one stack slot to another can happen (not allowed by LIR_Assembler 3934 int spill_slot = from_interval->canonical_spill_slot(); 3935 if (spill_slot < 0) { 3936 spill_slot = allocator()->allocate_spill_slot(type2spill_size[spill_interval->type()] == 2); 3937 from_interval->set_canonical_spill_slot(spill_slot); 3938 } 3939 spill_interval->assign_reg(spill_slot); 3940 allocator()->append_interval(spill_interval); 3941 3942 TRACE_LINEAR_SCAN(4, tty->print_cr("created new Interval %d for spilling", spill_interval->reg_num())); 3943 3944 // insert a move from register to stack and update the mapping 3945 insert_move(from_interval, spill_interval); 3946 _mapping_from.at_put(spill_candidate, spill_interval); 3947 unblock_registers(from_interval); 3948 } 3949 } 3950 3951 // reset to default value 3952 _multiple_reads_allowed = false; 3953 3954 // check that all intervals have been processed 3955 DEBUG_ONLY(check_empty()); 3956 } 3957 3958 3959 void MoveResolver::set_insert_position(LIR_List* insert_list, int insert_idx) { 3960 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)); 3961 assert(_insert_list == NULL && _insert_idx == -1, "use move_insert_position instead of set_insert_position when data already set"); 3962 3963 create_insertion_buffer(insert_list); 3964 _insert_list = insert_list; 3965 _insert_idx = insert_idx; 3966 } 3967 3968 void MoveResolver::move_insert_position(LIR_List* insert_list, int insert_idx) { 3969 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)); 3970 3971 if (_insert_list != NULL && (insert_list != _insert_list || insert_idx != _insert_idx)) { 3972 // insert position changed -> resolve current mappings 3973 resolve_mappings(); 3974 } 3975 3976 if (insert_list != _insert_list) { 3977 // block changed -> append insertion_buffer because it is 3978 // bound to a specific block and create a new insertion_buffer 3979 append_insertion_buffer(); 3980 create_insertion_buffer(insert_list); 3981 } 3982 3983 _insert_list = insert_list; 3984 _insert_idx = insert_idx; 3985 } 3986 3987 void MoveResolver::add_mapping(Interval* from_interval, Interval* to_interval) { 3988 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())); 3989 3990 _mapping_from.append(from_interval); 3991 _mapping_from_opr.append(LIR_OprFact::illegalOpr); 3992 _mapping_to.append(to_interval); 3993 } 3994 3995 3996 void MoveResolver::add_mapping(LIR_Opr from_opr, Interval* to_interval) { 3997 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())); 3998 assert(from_opr->is_constant(), "only for constants"); 3999 4000 _mapping_from.append(NULL); 4001 _mapping_from_opr.append(from_opr); 4002 _mapping_to.append(to_interval); 4003 } 4004 4005 void MoveResolver::resolve_and_append_moves() { 4006 if (has_mappings()) { 4007 resolve_mappings(); 4008 } 4009 append_insertion_buffer(); 4010 } 4011 4012 4013 4014 // **** Implementation of Range ************************************* 4015 4016 Range::Range(int from, int to, Range* next) : 4017 _from(from), 4018 _to(to), 4019 _next(next) 4020 { 4021 } 4022 4023 // initialize sentinel 4024 Range* Range::_end = NULL; 4025 void Range::initialize(Arena* arena) { 4026 _end = new (arena) Range(max_jint, max_jint, NULL); 4027 } 4028 4029 int Range::intersects_at(Range* r2) const { 4030 const Range* r1 = this; 4031 4032 assert(r1 != NULL && r2 != NULL, "null ranges not allowed"); 4033 assert(r1 != _end && r2 != _end, "empty ranges not allowed"); 4034 4035 do { 4036 if (r1->from() < r2->from()) { 4037 if (r1->to() <= r2->from()) { 4038 r1 = r1->next(); if (r1 == _end) return -1; 4039 } else { 4040 return r2->from(); 4041 } 4042 } else if (r2->from() < r1->from()) { 4043 if (r2->to() <= r1->from()) { 4044 r2 = r2->next(); if (r2 == _end) return -1; 4045 } else { 4046 return r1->from(); 4047 } 4048 } else { // r1->from() == r2->from() 4049 if (r1->from() == r1->to()) { 4050 r1 = r1->next(); if (r1 == _end) return -1; 4051 } else if (r2->from() == r2->to()) { 4052 r2 = r2->next(); if (r2 == _end) return -1; 4053 } else { 4054 return r1->from(); 4055 } 4056 } 4057 } while (true); 4058 } 4059 4060 #ifndef PRODUCT 4061 void Range::print(outputStream* out) const { 4062 out->print("[%d, %d[ ", _from, _to); 4063 } 4064 #endif 4065 4066 4067 4068 // **** Implementation of Interval ********************************** 4069 4070 // initialize sentinel 4071 Interval* Interval::_end = NULL; 4072 void Interval::initialize(Arena* arena) { 4073 Range::initialize(arena); 4074 _end = new (arena) Interval(-1); 4075 } 4076 4077 Interval::Interval(int reg_num) : 4078 _reg_num(reg_num), 4079 _type(T_ILLEGAL), 4080 _first(Range::end()), 4081 _use_pos_and_kinds(12), 4082 _current(Range::end()), 4083 _next(_end), 4084 _state(invalidState), 4085 _assigned_reg(LinearScan::any_reg), 4086 _assigned_regHi(LinearScan::any_reg), 4087 _cached_to(-1), 4088 _cached_opr(LIR_OprFact::illegalOpr), 4089 _cached_vm_reg(VMRegImpl::Bad()), 4090 _split_children(0), 4091 _canonical_spill_slot(-1), 4092 _insert_move_when_activated(false), 4093 _register_hint(NULL), 4094 _spill_state(noDefinitionFound), 4095 _spill_definition_pos(-1) 4096 { 4097 _split_parent = this; 4098 _current_split_child = this; 4099 } 4100 4101 int Interval::calc_to() { 4102 assert(_first != Range::end(), "interval has no range"); 4103 4104 Range* r = _first; 4105 while (r->next() != Range::end()) { 4106 r = r->next(); 4107 } 4108 return r->to(); 4109 } 4110 4111 4112 #ifdef ASSERT 4113 // consistency check of split-children 4114 void Interval::check_split_children() { 4115 if (_split_children.length() > 0) { 4116 assert(is_split_parent(), "only split parents can have children"); 4117 4118 for (int i = 0; i < _split_children.length(); i++) { 4119 Interval* i1 = _split_children.at(i); 4120 4121 assert(i1->split_parent() == this, "not a split child of this interval"); 4122 assert(i1->type() == type(), "must be equal for all split children"); 4123 assert(i1->canonical_spill_slot() == canonical_spill_slot(), "must be equal for all split children"); 4124 4125 for (int j = i + 1; j < _split_children.length(); j++) { 4126 Interval* i2 = _split_children.at(j); 4127 4128 assert(i1->reg_num() != i2->reg_num(), "same register number"); 4129 4130 if (i1->from() < i2->from()) { 4131 assert(i1->to() <= i2->from() && i1->to() < i2->to(), "intervals overlapping"); 4132 } else { 4133 assert(i2->from() < i1->from(), "intervals start at same op_id"); 4134 assert(i2->to() <= i1->from() && i2->to() < i1->to(), "intervals overlapping"); 4135 } 4136 } 4137 } 4138 } 4139 } 4140 #endif // ASSERT 4141 4142 Interval* Interval::register_hint(bool search_split_child) const { 4143 if (!search_split_child) { 4144 return _register_hint; 4145 } 4146 4147 if (_register_hint != NULL) { 4148 assert(_register_hint->is_split_parent(), "ony split parents are valid hint registers"); 4149 4150 if (_register_hint->assigned_reg() >= 0 && _register_hint->assigned_reg() < LinearScan::nof_regs) { 4151 return _register_hint; 4152 4153 } else if (_register_hint->_split_children.length() > 0) { 4154 // search the first split child that has a register assigned 4155 int len = _register_hint->_split_children.length(); 4156 for (int i = 0; i < len; i++) { 4157 Interval* cur = _register_hint->_split_children.at(i); 4158 4159 if (cur->assigned_reg() >= 0 && cur->assigned_reg() < LinearScan::nof_regs) { 4160 return cur; 4161 } 4162 } 4163 } 4164 } 4165 4166 // no hint interval found that has a register assigned 4167 return NULL; 4168 } 4169 4170 4171 Interval* Interval::split_child_at_op_id(int op_id, LIR_OpVisitState::OprMode mode) { 4172 assert(is_split_parent(), "can only be called for split parents"); 4173 assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)"); 4174 4175 Interval* result; 4176 if (_split_children.length() == 0) { 4177 result = this; 4178 } else { 4179 result = NULL; 4180 int len = _split_children.length(); 4181 4182 // in outputMode, the end of the interval (op_id == cur->to()) is not valid 4183 int to_offset = (mode == LIR_OpVisitState::outputMode ? 0 : 1); 4184 4185 int i; 4186 for (i = 0; i < len; i++) { 4187 Interval* cur = _split_children.at(i); 4188 if (cur->from() <= op_id && op_id < cur->to() + to_offset) { 4189 if (i > 0) { 4190 // exchange current split child to start of list (faster access for next call) 4191 _split_children.at_put(i, _split_children.at(0)); 4192 _split_children.at_put(0, cur); 4193 } 4194 4195 // interval found 4196 result = cur; 4197 break; 4198 } 4199 } 4200 4201 #ifdef ASSERT 4202 for (i = 0; i < len; i++) { 4203 Interval* tmp = _split_children.at(i); 4204 if (tmp != result && tmp->from() <= op_id && op_id < tmp->to() + to_offset) { 4205 tty->print_cr("two valid result intervals found for op_id %d: %d and %d", op_id, result->reg_num(), tmp->reg_num()); 4206 result->print(); 4207 tmp->print(); 4208 assert(false, "two valid result intervals found"); 4209 } 4210 } 4211 #endif 4212 } 4213 4214 assert(result != NULL, "no matching interval found"); 4215 assert(result->covers(op_id, mode), "op_id not covered by interval"); 4216 4217 return result; 4218 } 4219 4220 4221 // returns the last split child that ends before the given op_id 4222 Interval* Interval::split_child_before_op_id(int op_id) { 4223 assert(op_id >= 0, "invalid op_id"); 4224 4225 Interval* parent = split_parent(); 4226 Interval* result = NULL; 4227 4228 int len = parent->_split_children.length(); 4229 assert(len > 0, "no split children available"); 4230 4231 for (int i = len - 1; i >= 0; i--) { 4232 Interval* cur = parent->_split_children.at(i); 4233 if (cur->to() <= op_id && (result == NULL || result->to() < cur->to())) { 4234 result = cur; 4235 } 4236 } 4237 4238 assert(result != NULL, "no split child found"); 4239 return result; 4240 } 4241 4242 4243 // checks if op_id is covered by any split child 4244 bool Interval::split_child_covers(int op_id, LIR_OpVisitState::OprMode mode) { 4245 assert(is_split_parent(), "can only be called for split parents"); 4246 assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)"); 4247 4248 if (_split_children.length() == 0) { 4249 // simple case if interval was not split 4250 return covers(op_id, mode); 4251 4252 } else { 4253 // extended case: check all split children 4254 int len = _split_children.length(); 4255 for (int i = 0; i < len; i++) { 4256 Interval* cur = _split_children.at(i); 4257 if (cur->covers(op_id, mode)) { 4258 return true; 4259 } 4260 } 4261 return false; 4262 } 4263 } 4264 4265 4266 // Note: use positions are sorted descending -> first use has highest index 4267 int Interval::first_usage(IntervalUseKind min_use_kind) const { 4268 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4269 4270 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4271 if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) { 4272 return _use_pos_and_kinds.at(i); 4273 } 4274 } 4275 return max_jint; 4276 } 4277 4278 int Interval::next_usage(IntervalUseKind min_use_kind, int from) const { 4279 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4280 4281 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4282 if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) >= min_use_kind) { 4283 return _use_pos_and_kinds.at(i); 4284 } 4285 } 4286 return max_jint; 4287 } 4288 4289 int Interval::next_usage_exact(IntervalUseKind exact_use_kind, int from) const { 4290 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4291 4292 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4293 if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) == exact_use_kind) { 4294 return _use_pos_and_kinds.at(i); 4295 } 4296 } 4297 return max_jint; 4298 } 4299 4300 int Interval::previous_usage(IntervalUseKind min_use_kind, int from) const { 4301 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); 4302 4303 int prev = 0; 4304 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4305 if (_use_pos_and_kinds.at(i) > from) { 4306 return prev; 4307 } 4308 if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) { 4309 prev = _use_pos_and_kinds.at(i); 4310 } 4311 } 4312 return prev; 4313 } 4314 4315 void Interval::add_use_pos(int pos, IntervalUseKind use_kind) { 4316 assert(covers(pos, LIR_OpVisitState::inputMode), "use position not covered by live range"); 4317 4318 // do not add use positions for precolored intervals because 4319 // they are never used 4320 if (use_kind != noUse && reg_num() >= LIR_OprDesc::vreg_base) { 4321 #ifdef ASSERT 4322 assert(_use_pos_and_kinds.length() % 2 == 0, "must be"); 4323 for (int i = 0; i < _use_pos_and_kinds.length(); i += 2) { 4324 assert(pos <= _use_pos_and_kinds.at(i), "already added a use-position with lower position"); 4325 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4326 if (i > 0) { 4327 assert(_use_pos_and_kinds.at(i) < _use_pos_and_kinds.at(i - 2), "not sorted descending"); 4328 } 4329 } 4330 #endif 4331 4332 // Note: add_use is called in descending order, so list gets sorted 4333 // automatically by just appending new use positions 4334 int len = _use_pos_and_kinds.length(); 4335 if (len == 0 || _use_pos_and_kinds.at(len - 2) > pos) { 4336 _use_pos_and_kinds.append(pos); 4337 _use_pos_and_kinds.append(use_kind); 4338 } else if (_use_pos_and_kinds.at(len - 1) < use_kind) { 4339 assert(_use_pos_and_kinds.at(len - 2) == pos, "list not sorted correctly"); 4340 _use_pos_and_kinds.at_put(len - 1, use_kind); 4341 } 4342 } 4343 } 4344 4345 void Interval::add_range(int from, int to) { 4346 assert(from < to, "invalid range"); 4347 assert(first() == Range::end() || to < first()->next()->from(), "not inserting at begin of interval"); 4348 assert(from <= first()->to(), "not inserting at begin of interval"); 4349 4350 if (first()->from() <= to) { 4351 // join intersecting ranges 4352 first()->set_from(MIN2(from, first()->from())); 4353 first()->set_to (MAX2(to, first()->to())); 4354 } else { 4355 // insert new range 4356 _first = new Range(from, to, first()); 4357 } 4358 } 4359 4360 Interval* Interval::new_split_child() { 4361 // allocate new interval 4362 Interval* result = new Interval(-1); 4363 result->set_type(type()); 4364 4365 Interval* parent = split_parent(); 4366 result->_split_parent = parent; 4367 result->set_register_hint(parent); 4368 4369 // insert new interval in children-list of parent 4370 if (parent->_split_children.length() == 0) { 4371 assert(is_split_parent(), "list must be initialized at first split"); 4372 4373 parent->_split_children = IntervalList(4); 4374 parent->_split_children.append(this); 4375 } 4376 parent->_split_children.append(result); 4377 4378 return result; 4379 } 4380 4381 // split this interval at the specified position and return 4382 // the remainder as a new interval. 4383 // 4384 // when an interval is split, a bi-directional link is established between the original interval 4385 // (the split parent) and the intervals that are split off this interval (the split children) 4386 // When a split child is split again, the new created interval is also a direct child 4387 // of the original parent (there is no tree of split children stored, but a flat list) 4388 // All split children are spilled to the same stack slot (stored in _canonical_spill_slot) 4389 // 4390 // Note: The new interval has no valid reg_num 4391 Interval* Interval::split(int split_pos) { 4392 assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals"); 4393 4394 // allocate new interval 4395 Interval* result = new_split_child(); 4396 4397 // split the ranges 4398 Range* prev = NULL; 4399 Range* cur = _first; 4400 while (cur != Range::end() && cur->to() <= split_pos) { 4401 prev = cur; 4402 cur = cur->next(); 4403 } 4404 assert(cur != Range::end(), "split interval after end of last range"); 4405 4406 if (cur->from() < split_pos) { 4407 result->_first = new Range(split_pos, cur->to(), cur->next()); 4408 cur->set_to(split_pos); 4409 cur->set_next(Range::end()); 4410 4411 } else { 4412 assert(prev != NULL, "split before start of first range"); 4413 result->_first = cur; 4414 prev->set_next(Range::end()); 4415 } 4416 result->_current = result->_first; 4417 _cached_to = -1; // clear cached value 4418 4419 // split list of use positions 4420 int total_len = _use_pos_and_kinds.length(); 4421 int start_idx = total_len - 2; 4422 while (start_idx >= 0 && _use_pos_and_kinds.at(start_idx) < split_pos) { 4423 start_idx -= 2; 4424 } 4425 4426 intStack new_use_pos_and_kinds(total_len - start_idx); 4427 int i; 4428 for (i = start_idx + 2; i < total_len; i++) { 4429 new_use_pos_and_kinds.append(_use_pos_and_kinds.at(i)); 4430 } 4431 4432 _use_pos_and_kinds.truncate(start_idx + 2); 4433 result->_use_pos_and_kinds = _use_pos_and_kinds; 4434 _use_pos_and_kinds = new_use_pos_and_kinds; 4435 4436 #ifdef ASSERT 4437 assert(_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos"); 4438 assert(result->_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos"); 4439 assert(_use_pos_and_kinds.length() + result->_use_pos_and_kinds.length() == total_len, "missed some entries"); 4440 4441 for (i = 0; i < _use_pos_and_kinds.length(); i += 2) { 4442 assert(_use_pos_and_kinds.at(i) < split_pos, "must be"); 4443 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4444 } 4445 for (i = 0; i < result->_use_pos_and_kinds.length(); i += 2) { 4446 assert(result->_use_pos_and_kinds.at(i) >= split_pos, "must be"); 4447 assert(result->_use_pos_and_kinds.at(i + 1) >= firstValidKind && result->_use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4448 } 4449 #endif 4450 4451 return result; 4452 } 4453 4454 // split this interval at the specified position and return 4455 // the head as a new interval (the original interval is the tail) 4456 // 4457 // Currently, only the first range can be split, and the new interval 4458 // must not have split positions 4459 Interval* Interval::split_from_start(int split_pos) { 4460 assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals"); 4461 assert(split_pos > from() && split_pos < to(), "can only split inside interval"); 4462 assert(split_pos > _first->from() && split_pos <= _first->to(), "can only split inside first range"); 4463 assert(first_usage(noUse) > split_pos, "can not split when use positions are present"); 4464 4465 // allocate new interval 4466 Interval* result = new_split_child(); 4467 4468 // the new created interval has only one range (checked by assertion above), 4469 // so the splitting of the ranges is very simple 4470 result->add_range(_first->from(), split_pos); 4471 4472 if (split_pos == _first->to()) { 4473 assert(_first->next() != Range::end(), "must not be at end"); 4474 _first = _first->next(); 4475 } else { 4476 _first->set_from(split_pos); 4477 } 4478 4479 return result; 4480 } 4481 4482 4483 // returns true if the op_id is inside the interval 4484 bool Interval::covers(int op_id, LIR_OpVisitState::OprMode mode) const { 4485 Range* cur = _first; 4486 4487 while (cur != Range::end() && cur->to() < op_id) { 4488 cur = cur->next(); 4489 } 4490 if (cur != Range::end()) { 4491 assert(cur->to() != cur->next()->from(), "ranges not separated"); 4492 4493 if (mode == LIR_OpVisitState::outputMode) { 4494 return cur->from() <= op_id && op_id < cur->to(); 4495 } else { 4496 return cur->from() <= op_id && op_id <= cur->to(); 4497 } 4498 } 4499 return false; 4500 } 4501 4502 // returns true if the interval has any hole between hole_from and hole_to 4503 // (even if the hole has only the length 1) 4504 bool Interval::has_hole_between(int hole_from, int hole_to) { 4505 assert(hole_from < hole_to, "check"); 4506 assert(from() <= hole_from && hole_to <= to(), "index out of interval"); 4507 4508 Range* cur = _first; 4509 while (cur != Range::end()) { 4510 assert(cur->to() < cur->next()->from(), "no space between ranges"); 4511 4512 // hole-range starts before this range -> hole 4513 if (hole_from < cur->from()) { 4514 return true; 4515 4516 // hole-range completely inside this range -> no hole 4517 } else if (hole_to <= cur->to()) { 4518 return false; 4519 4520 // overlapping of hole-range with this range -> hole 4521 } else if (hole_from <= cur->to()) { 4522 return true; 4523 } 4524 4525 cur = cur->next(); 4526 } 4527 4528 return false; 4529 } 4530 4531 4532 #ifndef PRODUCT 4533 void Interval::print(outputStream* out) const { 4534 const char* SpillState2Name[] = { "no definition", "no spill store", "one spill store", "store at definition", "start in memory", "no optimization" }; 4535 const char* UseKind2Name[] = { "N", "L", "S", "M" }; 4536 4537 const char* type_name; 4538 LIR_Opr opr = LIR_OprFact::illegal(); 4539 if (reg_num() < LIR_OprDesc::vreg_base) { 4540 type_name = "fixed"; 4541 // need a temporary operand for fixed intervals because type() cannot be called 4542 if (assigned_reg() >= pd_first_cpu_reg && assigned_reg() <= pd_last_cpu_reg) { 4543 opr = LIR_OprFact::single_cpu(assigned_reg()); 4544 } else if (assigned_reg() >= pd_first_fpu_reg && assigned_reg() <= pd_last_fpu_reg) { 4545 opr = LIR_OprFact::single_fpu(assigned_reg() - pd_first_fpu_reg); 4546 #ifdef X86 4547 } else if (assigned_reg() >= pd_first_xmm_reg && assigned_reg() <= pd_last_xmm_reg) { 4548 opr = LIR_OprFact::single_xmm(assigned_reg() - pd_first_xmm_reg); 4549 #endif 4550 } else { 4551 ShouldNotReachHere(); 4552 } 4553 } else { 4554 type_name = type2name(type()); 4555 if (assigned_reg() != -1 && 4556 (LinearScan::num_physical_regs(type()) == 1 || assigned_regHi() != -1)) { 4557 opr = LinearScan::calc_operand_for_interval(this); 4558 } 4559 } 4560 4561 out->print("%d %s ", reg_num(), type_name); 4562 if (opr->is_valid()) { 4563 out->print("\""); 4564 opr->print(out); 4565 out->print("\" "); 4566 } 4567 out->print("%d %d ", split_parent()->reg_num(), (register_hint(false) != NULL ? register_hint(false)->reg_num() : -1)); 4568 4569 // print ranges 4570 Range* cur = _first; 4571 while (cur != Range::end()) { 4572 cur->print(out); 4573 cur = cur->next(); 4574 assert(cur != NULL, "range list not closed with range sentinel"); 4575 } 4576 4577 // print use positions 4578 int prev = 0; 4579 assert(_use_pos_and_kinds.length() % 2 == 0, "must be"); 4580 for (int i =_use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { 4581 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); 4582 assert(prev < _use_pos_and_kinds.at(i), "use positions not sorted"); 4583 4584 out->print("%d %s ", _use_pos_and_kinds.at(i), UseKind2Name[_use_pos_and_kinds.at(i + 1)]); 4585 prev = _use_pos_and_kinds.at(i); 4586 } 4587 4588 out->print(" \"%s\"", SpillState2Name[spill_state()]); 4589 out->cr(); 4590 } 4591 #endif 4592 4593 4594 4595 // **** Implementation of IntervalWalker **************************** 4596 4597 IntervalWalker::IntervalWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first) 4598 : _compilation(allocator->compilation()) 4599 , _allocator(allocator) 4600 { 4601 _unhandled_first[fixedKind] = unhandled_fixed_first; 4602 _unhandled_first[anyKind] = unhandled_any_first; 4603 _active_first[fixedKind] = Interval::end(); 4604 _inactive_first[fixedKind] = Interval::end(); 4605 _active_first[anyKind] = Interval::end(); 4606 _inactive_first[anyKind] = Interval::end(); 4607 _current_position = -1; 4608 _current = NULL; 4609 next_interval(); 4610 } 4611 4612 4613 // append interval at top of list 4614 void IntervalWalker::append_unsorted(Interval** list, Interval* interval) { 4615 interval->set_next(*list); *list = interval; 4616 } 4617 4618 4619 // append interval in order of current range from() 4620 void IntervalWalker::append_sorted(Interval** list, Interval* interval) { 4621 Interval* prev = NULL; 4622 Interval* cur = *list; 4623 while (cur->current_from() < interval->current_from()) { 4624 prev = cur; cur = cur->next(); 4625 } 4626 if (prev == NULL) { 4627 *list = interval; 4628 } else { 4629 prev->set_next(interval); 4630 } 4631 interval->set_next(cur); 4632 } 4633 4634 void IntervalWalker::append_to_unhandled(Interval** list, Interval* interval) { 4635 assert(interval->from() >= current()->current_from(), "cannot append new interval before current walk position"); 4636 4637 Interval* prev = NULL; 4638 Interval* cur = *list; 4639 while (cur->from() < interval->from() || (cur->from() == interval->from() && cur->first_usage(noUse) < interval->first_usage(noUse))) { 4640 prev = cur; cur = cur->next(); 4641 } 4642 if (prev == NULL) { 4643 *list = interval; 4644 } else { 4645 prev->set_next(interval); 4646 } 4647 interval->set_next(cur); 4648 } 4649 4650 4651 inline bool IntervalWalker::remove_from_list(Interval** list, Interval* i) { 4652 while (*list != Interval::end() && *list != i) { 4653 list = (*list)->next_addr(); 4654 } 4655 if (*list != Interval::end()) { 4656 assert(*list == i, "check"); 4657 *list = (*list)->next(); 4658 return true; 4659 } else { 4660 return false; 4661 } 4662 } 4663 4664 void IntervalWalker::remove_from_list(Interval* i) { 4665 bool deleted; 4666 4667 if (i->state() == activeState) { 4668 deleted = remove_from_list(active_first_addr(anyKind), i); 4669 } else { 4670 assert(i->state() == inactiveState, "invalid state"); 4671 deleted = remove_from_list(inactive_first_addr(anyKind), i); 4672 } 4673 4674 assert(deleted, "interval has not been found in list"); 4675 } 4676 4677 4678 void IntervalWalker::walk_to(IntervalState state, int from) { 4679 assert (state == activeState || state == inactiveState, "wrong state"); 4680 for_each_interval_kind(kind) { 4681 Interval** prev = state == activeState ? active_first_addr(kind) : inactive_first_addr(kind); 4682 Interval* next = *prev; 4683 while (next->current_from() <= from) { 4684 Interval* cur = next; 4685 next = cur->next(); 4686 4687 bool range_has_changed = false; 4688 while (cur->current_to() <= from) { 4689 cur->next_range(); 4690 range_has_changed = true; 4691 } 4692 4693 // also handle move from inactive list to active list 4694 range_has_changed = range_has_changed || (state == inactiveState && cur->current_from() <= from); 4695 4696 if (range_has_changed) { 4697 // remove cur from list 4698 *prev = next; 4699 if (cur->current_at_end()) { 4700 // move to handled state (not maintained as a list) 4701 cur->set_state(handledState); 4702 interval_moved(cur, kind, state, handledState); 4703 } else if (cur->current_from() <= from){ 4704 // sort into active list 4705 append_sorted(active_first_addr(kind), cur); 4706 cur->set_state(activeState); 4707 if (*prev == cur) { 4708 assert(state == activeState, "check"); 4709 prev = cur->next_addr(); 4710 } 4711 interval_moved(cur, kind, state, activeState); 4712 } else { 4713 // sort into inactive list 4714 append_sorted(inactive_first_addr(kind), cur); 4715 cur->set_state(inactiveState); 4716 if (*prev == cur) { 4717 assert(state == inactiveState, "check"); 4718 prev = cur->next_addr(); 4719 } 4720 interval_moved(cur, kind, state, inactiveState); 4721 } 4722 } else { 4723 prev = cur->next_addr(); 4724 continue; 4725 } 4726 } 4727 } 4728 } 4729 4730 4731 void IntervalWalker::next_interval() { 4732 IntervalKind kind; 4733 Interval* any = _unhandled_first[anyKind]; 4734 Interval* fixed = _unhandled_first[fixedKind]; 4735 4736 if (any != Interval::end()) { 4737 // intervals may start at same position -> prefer fixed interval 4738 kind = fixed != Interval::end() && fixed->from() <= any->from() ? fixedKind : anyKind; 4739 4740 assert (kind == fixedKind && fixed->from() <= any->from() || 4741 kind == anyKind && any->from() <= fixed->from(), "wrong interval!!!"); 4742 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"); 4743 4744 } else if (fixed != Interval::end()) { 4745 kind = fixedKind; 4746 } else { 4747 _current = NULL; return; 4748 } 4749 _current_kind = kind; 4750 _current = _unhandled_first[kind]; 4751 _unhandled_first[kind] = _current->next(); 4752 _current->set_next(Interval::end()); 4753 _current->rewind_range(); 4754 } 4755 4756 4757 void IntervalWalker::walk_to(int lir_op_id) { 4758 assert(_current_position <= lir_op_id, "can not walk backwards"); 4759 while (current() != NULL) { 4760 bool is_active = current()->from() <= lir_op_id; 4761 int id = is_active ? current()->from() : lir_op_id; 4762 4763 TRACE_LINEAR_SCAN(2, if (_current_position < id) { tty->cr(); tty->print_cr("walk_to(%d) **************************************************************", id); }) 4764 4765 // set _current_position prior to call of walk_to 4766 _current_position = id; 4767 4768 // call walk_to even if _current_position == id 4769 walk_to(activeState, id); 4770 walk_to(inactiveState, id); 4771 4772 if (is_active) { 4773 current()->set_state(activeState); 4774 if (activate_current()) { 4775 append_sorted(active_first_addr(current_kind()), current()); 4776 interval_moved(current(), current_kind(), unhandledState, activeState); 4777 } 4778 4779 next_interval(); 4780 } else { 4781 return; 4782 } 4783 } 4784 } 4785 4786 void IntervalWalker::interval_moved(Interval* interval, IntervalKind kind, IntervalState from, IntervalState to) { 4787 #ifndef PRODUCT 4788 if (TraceLinearScanLevel >= 4) { 4789 #define print_state(state) \ 4790 switch(state) {\ 4791 case unhandledState: tty->print("unhandled"); break;\ 4792 case activeState: tty->print("active"); break;\ 4793 case inactiveState: tty->print("inactive"); break;\ 4794 case handledState: tty->print("handled"); break;\ 4795 default: ShouldNotReachHere(); \ 4796 } 4797 4798 print_state(from); tty->print(" to "); print_state(to); 4799 tty->fill_to(23); 4800 interval->print(); 4801 4802 #undef print_state 4803 } 4804 #endif 4805 } 4806 4807 4808 4809 // **** Implementation of LinearScanWalker ************************** 4810 4811 LinearScanWalker::LinearScanWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first) 4812 : IntervalWalker(allocator, unhandled_fixed_first, unhandled_any_first) 4813 , _move_resolver(allocator) 4814 { 4815 for (int i = 0; i < LinearScan::nof_regs; i++) { 4816 _spill_intervals[i] = new IntervalList(2); 4817 } 4818 } 4819 4820 4821 inline void LinearScanWalker::init_use_lists(bool only_process_use_pos) { 4822 for (int i = _first_reg; i <= _last_reg; i++) { 4823 _use_pos[i] = max_jint; 4824 4825 if (!only_process_use_pos) { 4826 _block_pos[i] = max_jint; 4827 _spill_intervals[i]->clear(); 4828 } 4829 } 4830 } 4831 4832 inline void LinearScanWalker::exclude_from_use(int reg) { 4833 assert(reg < LinearScan::nof_regs, "interval must have a register assigned (stack slots not allowed)"); 4834 if (reg >= _first_reg && reg <= _last_reg) { 4835 _use_pos[reg] = 0; 4836 } 4837 } 4838 inline void LinearScanWalker::exclude_from_use(Interval* i) { 4839 assert(i->assigned_reg() != any_reg, "interval has no register assigned"); 4840 4841 exclude_from_use(i->assigned_reg()); 4842 exclude_from_use(i->assigned_regHi()); 4843 } 4844 4845 inline void LinearScanWalker::set_use_pos(int reg, Interval* i, int use_pos, bool only_process_use_pos) { 4846 assert(use_pos != 0, "must use exclude_from_use to set use_pos to 0"); 4847 4848 if (reg >= _first_reg && reg <= _last_reg) { 4849 if (_use_pos[reg] > use_pos) { 4850 _use_pos[reg] = use_pos; 4851 } 4852 if (!only_process_use_pos) { 4853 _spill_intervals[reg]->append(i); 4854 } 4855 } 4856 } 4857 inline void LinearScanWalker::set_use_pos(Interval* i, int use_pos, bool only_process_use_pos) { 4858 assert(i->assigned_reg() != any_reg, "interval has no register assigned"); 4859 if (use_pos != -1) { 4860 set_use_pos(i->assigned_reg(), i, use_pos, only_process_use_pos); 4861 set_use_pos(i->assigned_regHi(), i, use_pos, only_process_use_pos); 4862 } 4863 } 4864 4865 inline void LinearScanWalker::set_block_pos(int reg, Interval* i, int block_pos) { 4866 if (reg >= _first_reg && reg <= _last_reg) { 4867 if (_block_pos[reg] > block_pos) { 4868 _block_pos[reg] = block_pos; 4869 } 4870 if (_use_pos[reg] > block_pos) { 4871 _use_pos[reg] = block_pos; 4872 } 4873 } 4874 } 4875 inline void LinearScanWalker::set_block_pos(Interval* i, int block_pos) { 4876 assert(i->assigned_reg() != any_reg, "interval has no register assigned"); 4877 if (block_pos != -1) { 4878 set_block_pos(i->assigned_reg(), i, block_pos); 4879 set_block_pos(i->assigned_regHi(), i, block_pos); 4880 } 4881 } 4882 4883 4884 void LinearScanWalker::free_exclude_active_fixed() { 4885 Interval* list = active_first(fixedKind); 4886 while (list != Interval::end()) { 4887 assert(list->assigned_reg() < LinearScan::nof_regs, "active interval must have a register assigned"); 4888 exclude_from_use(list); 4889 list = list->next(); 4890 } 4891 } 4892 4893 void LinearScanWalker::free_exclude_active_any() { 4894 Interval* list = active_first(anyKind); 4895 while (list != Interval::end()) { 4896 exclude_from_use(list); 4897 list = list->next(); 4898 } 4899 } 4900 4901 void LinearScanWalker::free_collect_inactive_fixed(Interval* cur) { 4902 Interval* list = inactive_first(fixedKind); 4903 while (list != Interval::end()) { 4904 if (cur->to() <= list->current_from()) { 4905 assert(list->current_intersects_at(cur) == -1, "must not intersect"); 4906 set_use_pos(list, list->current_from(), true); 4907 } else { 4908 set_use_pos(list, list->current_intersects_at(cur), true); 4909 } 4910 list = list->next(); 4911 } 4912 } 4913 4914 void LinearScanWalker::free_collect_inactive_any(Interval* cur) { 4915 Interval* list = inactive_first(anyKind); 4916 while (list != Interval::end()) { 4917 set_use_pos(list, list->current_intersects_at(cur), true); 4918 list = list->next(); 4919 } 4920 } 4921 4922 void LinearScanWalker::free_collect_unhandled(IntervalKind kind, Interval* cur) { 4923 Interval* list = unhandled_first(kind); 4924 while (list != Interval::end()) { 4925 set_use_pos(list, list->intersects_at(cur), true); 4926 if (kind == fixedKind && cur->to() <= list->from()) { 4927 set_use_pos(list, list->from(), true); 4928 } 4929 list = list->next(); 4930 } 4931 } 4932 4933 void LinearScanWalker::spill_exclude_active_fixed() { 4934 Interval* list = active_first(fixedKind); 4935 while (list != Interval::end()) { 4936 exclude_from_use(list); 4937 list = list->next(); 4938 } 4939 } 4940 4941 void LinearScanWalker::spill_block_unhandled_fixed(Interval* cur) { 4942 Interval* list = unhandled_first(fixedKind); 4943 while (list != Interval::end()) { 4944 set_block_pos(list, list->intersects_at(cur)); 4945 list = list->next(); 4946 } 4947 } 4948 4949 void LinearScanWalker::spill_block_inactive_fixed(Interval* cur) { 4950 Interval* list = inactive_first(fixedKind); 4951 while (list != Interval::end()) { 4952 if (cur->to() > list->current_from()) { 4953 set_block_pos(list, list->current_intersects_at(cur)); 4954 } else { 4955 assert(list->current_intersects_at(cur) == -1, "invalid optimization: intervals intersect"); 4956 } 4957 4958 list = list->next(); 4959 } 4960 } 4961 4962 void LinearScanWalker::spill_collect_active_any() { 4963 Interval* list = active_first(anyKind); 4964 while (list != Interval::end()) { 4965 set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false); 4966 list = list->next(); 4967 } 4968 } 4969 4970 void LinearScanWalker::spill_collect_inactive_any(Interval* cur) { 4971 Interval* list = inactive_first(anyKind); 4972 while (list != Interval::end()) { 4973 if (list->current_intersects(cur)) { 4974 set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false); 4975 } 4976 list = list->next(); 4977 } 4978 } 4979 4980 4981 void LinearScanWalker::insert_move(int op_id, Interval* src_it, Interval* dst_it) { 4982 // output all moves here. When source and target are equal, the move is 4983 // optimized away later in assign_reg_nums 4984 4985 op_id = (op_id + 1) & ~1; 4986 BlockBegin* op_block = allocator()->block_of_op_with_id(op_id); 4987 assert(op_id > 0 && allocator()->block_of_op_with_id(op_id - 2) == op_block, "cannot insert move at block boundary"); 4988 4989 // calculate index of instruction inside instruction list of current block 4990 // the minimal index (for a block with no spill moves) can be calculated because the 4991 // numbering of instructions is known. 4992 // When the block already contains spill moves, the index must be increased until the 4993 // correct index is reached. 4994 LIR_OpList* list = op_block->lir()->instructions_list(); 4995 int index = (op_id - list->at(0)->id()) / 2; 4996 assert(list->at(index)->id() <= op_id, "error in calculation"); 4997 4998 while (list->at(index)->id() != op_id) { 4999 index++; 5000 assert(0 <= index && index < list->length(), "index out of bounds"); 5001 } 5002 assert(1 <= index && index < list->length(), "index out of bounds"); 5003 assert(list->at(index)->id() == op_id, "error in calculation"); 5004 5005 // insert new instruction before instruction at position index 5006 _move_resolver.move_insert_position(op_block->lir(), index - 1); 5007 _move_resolver.add_mapping(src_it, dst_it); 5008 } 5009 5010 5011 int LinearScanWalker::find_optimal_split_pos(BlockBegin* min_block, BlockBegin* max_block, int max_split_pos) { 5012 int from_block_nr = min_block->linear_scan_number(); 5013 int to_block_nr = max_block->linear_scan_number(); 5014 5015 assert(0 <= from_block_nr && from_block_nr < block_count(), "out of range"); 5016 assert(0 <= to_block_nr && to_block_nr < block_count(), "out of range"); 5017 assert(from_block_nr < to_block_nr, "must cross block boundary"); 5018 5019 // Try to split at end of max_block. If this would be after 5020 // max_split_pos, then use the begin of max_block 5021 int optimal_split_pos = max_block->last_lir_instruction_id() + 2; 5022 if (optimal_split_pos > max_split_pos) { 5023 optimal_split_pos = max_block->first_lir_instruction_id(); 5024 } 5025 5026 int min_loop_depth = max_block->loop_depth(); 5027 for (int i = to_block_nr - 1; i >= from_block_nr; i--) { 5028 BlockBegin* cur = block_at(i); 5029 5030 if (cur->loop_depth() < min_loop_depth) { 5031 // block with lower loop-depth found -> split at the end of this block 5032 min_loop_depth = cur->loop_depth(); 5033 optimal_split_pos = cur->last_lir_instruction_id() + 2; 5034 } 5035 } 5036 assert(optimal_split_pos > allocator()->max_lir_op_id() || allocator()->is_block_begin(optimal_split_pos), "algorithm must move split pos to block boundary"); 5037 5038 return optimal_split_pos; 5039 } 5040 5041 5042 int LinearScanWalker::find_optimal_split_pos(Interval* it, int min_split_pos, int max_split_pos, bool do_loop_optimization) { 5043 int optimal_split_pos = -1; 5044 if (min_split_pos == max_split_pos) { 5045 // trivial case, no optimization of split position possible 5046 TRACE_LINEAR_SCAN(4, tty->print_cr(" min-pos and max-pos are equal, no optimization possible")); 5047 optimal_split_pos = min_split_pos; 5048 5049 } else { 5050 assert(min_split_pos < max_split_pos, "must be true then"); 5051 assert(min_split_pos > 0, "cannot access min_split_pos - 1 otherwise"); 5052 5053 // reason for using min_split_pos - 1: when the minimal split pos is exactly at the 5054 // beginning of a block, then min_split_pos is also a possible split position. 5055 // Use the block before as min_block, because then min_block->last_lir_instruction_id() + 2 == min_split_pos 5056 BlockBegin* min_block = allocator()->block_of_op_with_id(min_split_pos - 1); 5057 5058 // reason for using max_split_pos - 1: otherwise there would be an assertion failure 5059 // when an interval ends at the end of the last block of the method 5060 // (in this case, max_split_pos == allocator()->max_lir_op_id() + 2, and there is no 5061 // block at this op_id) 5062 BlockBegin* max_block = allocator()->block_of_op_with_id(max_split_pos - 1); 5063 5064 assert(min_block->linear_scan_number() <= max_block->linear_scan_number(), "invalid order"); 5065 if (min_block == max_block) { 5066 // split position cannot be moved to block boundary, so split as late as possible 5067 TRACE_LINEAR_SCAN(4, tty->print_cr(" cannot move split pos to block boundary because min_pos and max_pos are in same block")); 5068 optimal_split_pos = max_split_pos; 5069 5070 } else if (it->has_hole_between(max_split_pos - 1, max_split_pos) && !allocator()->is_block_begin(max_split_pos)) { 5071 // Do not move split position if the interval has a hole before max_split_pos. 5072 // Intervals resulting from Phi-Functions have more than one definition (marked 5073 // as mustHaveRegister) with a hole before each definition. When the register is needed 5074 // for the second definition, an earlier reloading is unnecessary. 5075 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has hole just before max_split_pos, so splitting at max_split_pos")); 5076 optimal_split_pos = max_split_pos; 5077 5078 } else { 5079 // seach optimal block boundary between min_split_pos and max_split_pos 5080 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())); 5081 5082 if (do_loop_optimization) { 5083 // Loop optimization: if a loop-end marker is found between min- and max-position, 5084 // then split before this loop 5085 int loop_end_pos = it->next_usage_exact(loopEndMarker, min_block->last_lir_instruction_id() + 2); 5086 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization: loop end found at pos %d", loop_end_pos)); 5087 5088 assert(loop_end_pos > min_split_pos, "invalid order"); 5089 if (loop_end_pos < max_split_pos) { 5090 // loop-end marker found between min- and max-position 5091 // if it is not the end marker for the same loop as the min-position, then move 5092 // the max-position to this loop block. 5093 // Desired result: uses tagged as shouldHaveRegister inside a loop cause a reloading 5094 // of the interval (normally, only mustHaveRegister causes a reloading) 5095 BlockBegin* loop_block = allocator()->block_of_op_with_id(loop_end_pos); 5096 5097 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())); 5098 assert(loop_block != min_block, "loop_block and min_block must be different because block boundary is needed between"); 5099 5100 optimal_split_pos = find_optimal_split_pos(min_block, loop_block, loop_block->last_lir_instruction_id() + 2); 5101 if (optimal_split_pos == loop_block->last_lir_instruction_id() + 2) { 5102 optimal_split_pos = -1; 5103 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization not necessary")); 5104 } else { 5105 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization successful")); 5106 } 5107 } 5108 } 5109 5110 if (optimal_split_pos == -1) { 5111 // not calculated by loop optimization 5112 optimal_split_pos = find_optimal_split_pos(min_block, max_block, max_split_pos); 5113 } 5114 } 5115 } 5116 TRACE_LINEAR_SCAN(4, tty->print_cr(" optimal split position: %d", optimal_split_pos)); 5117 5118 return optimal_split_pos; 5119 } 5120 5121 5122 /* 5123 split an interval at the optimal position between min_split_pos and 5124 max_split_pos in two parts: 5125 1) the left part has already a location assigned 5126 2) the right part is sorted into to the unhandled-list 5127 */ 5128 void LinearScanWalker::split_before_usage(Interval* it, int min_split_pos, int max_split_pos) { 5129 TRACE_LINEAR_SCAN(2, tty->print ("----- splitting interval: "); it->print()); 5130 TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d", min_split_pos, max_split_pos)); 5131 5132 assert(it->from() < min_split_pos, "cannot split at start of interval"); 5133 assert(current_position() < min_split_pos, "cannot split before current position"); 5134 assert(min_split_pos <= max_split_pos, "invalid order"); 5135 assert(max_split_pos <= it->to(), "cannot split after end of interval"); 5136 5137 int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, true); 5138 5139 assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range"); 5140 assert(optimal_split_pos <= it->to(), "cannot split after end of interval"); 5141 assert(optimal_split_pos > it->from(), "cannot split at start of interval"); 5142 5143 if (optimal_split_pos == it->to() && it->next_usage(mustHaveRegister, min_split_pos) == max_jint) { 5144 // the split position would be just before the end of the interval 5145 // -> no split at all necessary 5146 TRACE_LINEAR_SCAN(4, tty->print_cr(" no split necessary because optimal split position is at end of interval")); 5147 return; 5148 } 5149 5150 // must calculate this before the actual split is performed and before split position is moved to odd op_id 5151 bool move_necessary = !allocator()->is_block_begin(optimal_split_pos) && !it->has_hole_between(optimal_split_pos - 1, optimal_split_pos); 5152 5153 if (!allocator()->is_block_begin(optimal_split_pos)) { 5154 // move position before actual instruction (odd op_id) 5155 optimal_split_pos = (optimal_split_pos - 1) | 1; 5156 } 5157 5158 TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d", optimal_split_pos)); 5159 assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary"); 5160 assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary"); 5161 5162 Interval* split_part = it->split(optimal_split_pos); 5163 5164 allocator()->append_interval(split_part); 5165 allocator()->copy_register_flags(it, split_part); 5166 split_part->set_insert_move_when_activated(move_necessary); 5167 append_to_unhandled(unhandled_first_addr(anyKind), split_part); 5168 5169 TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts (insert_move_when_activated: %d)", move_necessary)); 5170 TRACE_LINEAR_SCAN(2, tty->print (" "); it->print()); 5171 TRACE_LINEAR_SCAN(2, tty->print (" "); split_part->print()); 5172 } 5173 5174 /* 5175 split an interval at the optimal position between min_split_pos and 5176 max_split_pos in two parts: 5177 1) the left part has already a location assigned 5178 2) the right part is always on the stack and therefore ignored in further processing 5179 */ 5180 void LinearScanWalker::split_for_spilling(Interval* it) { 5181 // calculate allowed range of splitting position 5182 int max_split_pos = current_position(); 5183 int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, max_split_pos) + 1, it->from()); 5184 5185 TRACE_LINEAR_SCAN(2, tty->print ("----- splitting and spilling interval: "); it->print()); 5186 TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d", min_split_pos, max_split_pos)); 5187 5188 assert(it->state() == activeState, "why spill interval that is not active?"); 5189 assert(it->from() <= min_split_pos, "cannot split before start of interval"); 5190 assert(min_split_pos <= max_split_pos, "invalid order"); 5191 assert(max_split_pos < it->to(), "cannot split at end end of interval"); 5192 assert(current_position() < it->to(), "interval must not end before current position"); 5193 5194 if (min_split_pos == it->from()) { 5195 // the whole interval is never used, so spill it entirely to memory 5196 TRACE_LINEAR_SCAN(2, tty->print_cr(" spilling entire interval because split pos is at beginning of interval")); 5197 assert(it->first_usage(shouldHaveRegister) > current_position(), "interval must not have use position before current_position"); 5198 5199 allocator()->assign_spill_slot(it); 5200 allocator()->change_spill_state(it, min_split_pos); 5201 5202 // Also kick parent intervals out of register to memory when they have no use 5203 // position. This avoids short interval in register surrounded by intervals in 5204 // memory -> avoid useless moves from memory to register and back 5205 Interval* parent = it; 5206 while (parent != NULL && parent->is_split_child()) { 5207 parent = parent->split_child_before_op_id(parent->from()); 5208 5209 if (parent->assigned_reg() < LinearScan::nof_regs) { 5210 if (parent->first_usage(shouldHaveRegister) == max_jint) { 5211 // parent is never used, so kick it out of its assigned register 5212 TRACE_LINEAR_SCAN(4, tty->print_cr(" kicking out interval %d out of its register because it is never used", parent->reg_num())); 5213 allocator()->assign_spill_slot(parent); 5214 } else { 5215 // do not go further back because the register is actually used by the interval 5216 parent = NULL; 5217 } 5218 } 5219 } 5220 5221 } else { 5222 // search optimal split pos, split interval and spill only the right hand part 5223 int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, false); 5224 5225 assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range"); 5226 assert(optimal_split_pos < it->to(), "cannot split at end of interval"); 5227 assert(optimal_split_pos >= it->from(), "cannot split before start of interval"); 5228 5229 if (!allocator()->is_block_begin(optimal_split_pos)) { 5230 // move position before actual instruction (odd op_id) 5231 optimal_split_pos = (optimal_split_pos - 1) | 1; 5232 } 5233 5234 TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d", optimal_split_pos)); 5235 assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary"); 5236 assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary"); 5237 5238 Interval* spilled_part = it->split(optimal_split_pos); 5239 allocator()->append_interval(spilled_part); 5240 allocator()->assign_spill_slot(spilled_part); 5241 allocator()->change_spill_state(spilled_part, optimal_split_pos); 5242 5243 if (!allocator()->is_block_begin(optimal_split_pos)) { 5244 TRACE_LINEAR_SCAN(4, tty->print_cr(" inserting move from interval %d to %d", it->reg_num(), spilled_part->reg_num())); 5245 insert_move(optimal_split_pos, it, spilled_part); 5246 } 5247 5248 // the current_split_child is needed later when moves are inserted for reloading 5249 assert(spilled_part->current_split_child() == it, "overwriting wrong current_split_child"); 5250 spilled_part->make_current_split_child(); 5251 5252 TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts")); 5253 TRACE_LINEAR_SCAN(2, tty->print (" "); it->print()); 5254 TRACE_LINEAR_SCAN(2, tty->print (" "); spilled_part->print()); 5255 } 5256 } 5257 5258 5259 void LinearScanWalker::split_stack_interval(Interval* it) { 5260 int min_split_pos = current_position() + 1; 5261 int max_split_pos = MIN2(it->first_usage(shouldHaveRegister), it->to()); 5262 5263 split_before_usage(it, min_split_pos, max_split_pos); 5264 } 5265 5266 void LinearScanWalker::split_when_partial_register_available(Interval* it, int register_available_until) { 5267 int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, register_available_until), it->from() + 1); 5268 int max_split_pos = register_available_until; 5269 5270 split_before_usage(it, min_split_pos, max_split_pos); 5271 } 5272 5273 void LinearScanWalker::split_and_spill_interval(Interval* it) { 5274 assert(it->state() == activeState || it->state() == inactiveState, "other states not allowed"); 5275 5276 int current_pos = current_position(); 5277 if (it->state() == inactiveState) { 5278 // the interval is currently inactive, so no spill slot is needed for now. 5279 // when the split part is activated, the interval has a new chance to get a register, 5280 // so in the best case no stack slot is necessary 5281 assert(it->has_hole_between(current_pos - 1, current_pos + 1), "interval can not be inactive otherwise"); 5282 split_before_usage(it, current_pos + 1, current_pos + 1); 5283 5284 } else { 5285 // search the position where the interval must have a register and split 5286 // at the optimal position before. 5287 // The new created part is added to the unhandled list and will get a register 5288 // when it is activated 5289 int min_split_pos = current_pos + 1; 5290 int max_split_pos = MIN2(it->next_usage(mustHaveRegister, min_split_pos), it->to()); 5291 5292 split_before_usage(it, min_split_pos, max_split_pos); 5293 5294 assert(it->next_usage(mustHaveRegister, current_pos) == max_jint, "the remaining part is spilled to stack and therefore has no register"); 5295 split_for_spilling(it); 5296 } 5297 } 5298 5299 5300 int LinearScanWalker::find_free_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) { 5301 int min_full_reg = any_reg; 5302 int max_partial_reg = any_reg; 5303 5304 for (int i = _first_reg; i <= _last_reg; i++) { 5305 if (i == ignore_reg) { 5306 // this register must be ignored 5307 5308 } else if (_use_pos[i] >= interval_to) { 5309 // this register is free for the full interval 5310 if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) { 5311 min_full_reg = i; 5312 } 5313 } else if (_use_pos[i] > reg_needed_until) { 5314 // this register is at least free until reg_needed_until 5315 if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) { 5316 max_partial_reg = i; 5317 } 5318 } 5319 } 5320 5321 if (min_full_reg != any_reg) { 5322 return min_full_reg; 5323 } else if (max_partial_reg != any_reg) { 5324 *need_split = true; 5325 return max_partial_reg; 5326 } else { 5327 return any_reg; 5328 } 5329 } 5330 5331 int LinearScanWalker::find_free_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) { 5332 assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm"); 5333 5334 int min_full_reg = any_reg; 5335 int max_partial_reg = any_reg; 5336 5337 for (int i = _first_reg; i < _last_reg; i+=2) { 5338 if (_use_pos[i] >= interval_to && _use_pos[i + 1] >= interval_to) { 5339 // this register is free for the full interval 5340 if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) { 5341 min_full_reg = i; 5342 } 5343 } else if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) { 5344 // this register is at least free until reg_needed_until 5345 if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) { 5346 max_partial_reg = i; 5347 } 5348 } 5349 } 5350 5351 if (min_full_reg != any_reg) { 5352 return min_full_reg; 5353 } else if (max_partial_reg != any_reg) { 5354 *need_split = true; 5355 return max_partial_reg; 5356 } else { 5357 return any_reg; 5358 } 5359 } 5360 5361 5362 bool LinearScanWalker::alloc_free_reg(Interval* cur) { 5363 TRACE_LINEAR_SCAN(2, tty->print("trying to find free register for "); cur->print()); 5364 5365 init_use_lists(true); 5366 free_exclude_active_fixed(); 5367 free_exclude_active_any(); 5368 free_collect_inactive_fixed(cur); 5369 free_collect_inactive_any(cur); 5370 // free_collect_unhandled(fixedKind, cur); 5371 assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0"); 5372 5373 // _use_pos contains the start of the next interval that has this register assigned 5374 // (either as a fixed register or a normal allocated register in the past) 5375 // only intervals overlapping with cur are processed, non-overlapping invervals can be ignored safely 5376 TRACE_LINEAR_SCAN(4, tty->print_cr(" state of registers:")); 5377 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])); 5378 5379 int hint_reg, hint_regHi; 5380 Interval* register_hint = cur->register_hint(); 5381 if (register_hint != NULL) { 5382 hint_reg = register_hint->assigned_reg(); 5383 hint_regHi = register_hint->assigned_regHi(); 5384 5385 if (allocator()->is_precolored_cpu_interval(register_hint)) { 5386 assert(hint_reg != any_reg && hint_regHi == any_reg, "must be for fixed intervals"); 5387 hint_regHi = hint_reg + 1; // connect e.g. eax-edx 5388 } 5389 TRACE_LINEAR_SCAN(4, tty->print(" hint registers %d, %d from interval ", hint_reg, hint_regHi); register_hint->print()); 5390 5391 } else { 5392 hint_reg = any_reg; 5393 hint_regHi = any_reg; 5394 } 5395 assert(hint_reg == any_reg || hint_reg != hint_regHi, "hint reg and regHi equal"); 5396 assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned to interval"); 5397 5398 // the register must be free at least until this position 5399 int reg_needed_until = cur->from() + 1; 5400 int interval_to = cur->to(); 5401 5402 bool need_split = false; 5403 int split_pos = -1; 5404 int reg = any_reg; 5405 int regHi = any_reg; 5406 5407 if (_adjacent_regs) { 5408 reg = find_free_double_reg(reg_needed_until, interval_to, hint_reg, &need_split); 5409 regHi = reg + 1; 5410 if (reg == any_reg) { 5411 return false; 5412 } 5413 split_pos = MIN2(_use_pos[reg], _use_pos[regHi]); 5414 5415 } else { 5416 reg = find_free_reg(reg_needed_until, interval_to, hint_reg, any_reg, &need_split); 5417 if (reg == any_reg) { 5418 return false; 5419 } 5420 split_pos = _use_pos[reg]; 5421 5422 if (_num_phys_regs == 2) { 5423 regHi = find_free_reg(reg_needed_until, interval_to, hint_regHi, reg, &need_split); 5424 5425 if (_use_pos[reg] < interval_to && regHi == any_reg) { 5426 // do not split interval if only one register can be assigned until the split pos 5427 // (when one register is found for the whole interval, split&spill is only 5428 // performed for the hi register) 5429 return false; 5430 5431 } else if (regHi != any_reg) { 5432 split_pos = MIN2(split_pos, _use_pos[regHi]); 5433 5434 // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax 5435 if (reg > regHi) { 5436 int temp = reg; 5437 reg = regHi; 5438 regHi = temp; 5439 } 5440 } 5441 } 5442 } 5443 5444 cur->assign_reg(reg, regHi); 5445 TRACE_LINEAR_SCAN(2, tty->print_cr("selected register %d, %d", reg, regHi)); 5446 5447 assert(split_pos > 0, "invalid split_pos"); 5448 if (need_split) { 5449 // register not available for full interval, so split it 5450 split_when_partial_register_available(cur, split_pos); 5451 } 5452 5453 // only return true if interval is completely assigned 5454 return _num_phys_regs == 1 || regHi != any_reg; 5455 } 5456 5457 5458 int LinearScanWalker::find_locked_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) { 5459 int max_reg = any_reg; 5460 5461 for (int i = _first_reg; i <= _last_reg; i++) { 5462 if (i == ignore_reg) { 5463 // this register must be ignored 5464 5465 } else if (_use_pos[i] > reg_needed_until) { 5466 if (max_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_reg] && max_reg != hint_reg)) { 5467 max_reg = i; 5468 } 5469 } 5470 } 5471 5472 if (max_reg != any_reg && _block_pos[max_reg] <= interval_to) { 5473 *need_split = true; 5474 } 5475 5476 return max_reg; 5477 } 5478 5479 int LinearScanWalker::find_locked_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) { 5480 assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm"); 5481 5482 int max_reg = any_reg; 5483 5484 for (int i = _first_reg; i < _last_reg; i+=2) { 5485 if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) { 5486 if (max_reg == any_reg || _use_pos[i] > _use_pos[max_reg]) { 5487 max_reg = i; 5488 } 5489 } 5490 } 5491 5492 if (_block_pos[max_reg] <= interval_to || _block_pos[max_reg + 1] <= interval_to) { 5493 *need_split = true; 5494 } 5495 5496 return max_reg; 5497 } 5498 5499 void LinearScanWalker::split_and_spill_intersecting_intervals(int reg, int regHi) { 5500 assert(reg != any_reg, "no register assigned"); 5501 5502 for (int i = 0; i < _spill_intervals[reg]->length(); i++) { 5503 Interval* it = _spill_intervals[reg]->at(i); 5504 remove_from_list(it); 5505 split_and_spill_interval(it); 5506 } 5507 5508 if (regHi != any_reg) { 5509 IntervalList* processed = _spill_intervals[reg]; 5510 for (int i = 0; i < _spill_intervals[regHi]->length(); i++) { 5511 Interval* it = _spill_intervals[regHi]->at(i); 5512 if (processed->index_of(it) == -1) { 5513 remove_from_list(it); 5514 split_and_spill_interval(it); 5515 } 5516 } 5517 } 5518 } 5519 5520 5521 // Split an Interval and spill it to memory so that cur can be placed in a register 5522 void LinearScanWalker::alloc_locked_reg(Interval* cur) { 5523 TRACE_LINEAR_SCAN(2, tty->print("need to split and spill to get register for "); cur->print()); 5524 5525 // collect current usage of registers 5526 init_use_lists(false); 5527 spill_exclude_active_fixed(); 5528 // spill_block_unhandled_fixed(cur); 5529 assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0"); 5530 spill_block_inactive_fixed(cur); 5531 spill_collect_active_any(); 5532 spill_collect_inactive_any(cur); 5533 5534 #ifndef PRODUCT 5535 if (TraceLinearScanLevel >= 4) { 5536 tty->print_cr(" state of registers:"); 5537 for (int i = _first_reg; i <= _last_reg; i++) { 5538 tty->print(" reg %d: use_pos: %d, block_pos: %d, intervals: ", i, _use_pos[i], _block_pos[i]); 5539 for (int j = 0; j < _spill_intervals[i]->length(); j++) { 5540 tty->print("%d ", _spill_intervals[i]->at(j)->reg_num()); 5541 } 5542 tty->cr(); 5543 } 5544 } 5545 #endif 5546 5547 // the register must be free at least until this position 5548 int reg_needed_until = MIN2(cur->first_usage(mustHaveRegister), cur->from() + 1); 5549 int interval_to = cur->to(); 5550 assert (reg_needed_until > 0 && reg_needed_until < max_jint, "interval has no use"); 5551 5552 int split_pos = 0; 5553 int use_pos = 0; 5554 bool need_split = false; 5555 int reg, regHi; 5556 5557 if (_adjacent_regs) { 5558 reg = find_locked_double_reg(reg_needed_until, interval_to, any_reg, &need_split); 5559 regHi = reg + 1; 5560 5561 if (reg != any_reg) { 5562 use_pos = MIN2(_use_pos[reg], _use_pos[regHi]); 5563 split_pos = MIN2(_block_pos[reg], _block_pos[regHi]); 5564 } 5565 } else { 5566 reg = find_locked_reg(reg_needed_until, interval_to, any_reg, cur->assigned_reg(), &need_split); 5567 regHi = any_reg; 5568 5569 if (reg != any_reg) { 5570 use_pos = _use_pos[reg]; 5571 split_pos = _block_pos[reg]; 5572 5573 if (_num_phys_regs == 2) { 5574 if (cur->assigned_reg() != any_reg) { 5575 regHi = reg; 5576 reg = cur->assigned_reg(); 5577 } else { 5578 regHi = find_locked_reg(reg_needed_until, interval_to, any_reg, reg, &need_split); 5579 if (regHi != any_reg) { 5580 use_pos = MIN2(use_pos, _use_pos[regHi]); 5581 split_pos = MIN2(split_pos, _block_pos[regHi]); 5582 } 5583 } 5584 5585 if (regHi != any_reg && reg > regHi) { 5586 // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax 5587 int temp = reg; 5588 reg = regHi; 5589 regHi = temp; 5590 } 5591 } 5592 } 5593 } 5594 5595 if (reg == any_reg || (_num_phys_regs == 2 && regHi == any_reg) || use_pos <= cur->first_usage(mustHaveRegister)) { 5596 // the first use of cur is later than the spilling position -> spill cur 5597 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)); 5598 5599 if (cur->first_usage(mustHaveRegister) <= cur->from() + 1) { 5600 assert(false, "cannot spill interval that is used in first instruction (possible reason: no register found)"); 5601 // assign a reasonable register and do a bailout in product mode to avoid errors 5602 allocator()->assign_spill_slot(cur); 5603 BAILOUT("LinearScan: no register found"); 5604 } 5605 5606 split_and_spill_interval(cur); 5607 } else { 5608 TRACE_LINEAR_SCAN(4, tty->print_cr("decided to use register %d, %d", reg, regHi)); 5609 assert(reg != any_reg && (_num_phys_regs == 1 || regHi != any_reg), "no register found"); 5610 assert(split_pos > 0, "invalid split_pos"); 5611 assert(need_split == false || split_pos > cur->from(), "splitting interval at from"); 5612 5613 cur->assign_reg(reg, regHi); 5614 if (need_split) { 5615 // register not available for full interval, so split it 5616 split_when_partial_register_available(cur, split_pos); 5617 } 5618 5619 // perform splitting and spilling for all affected intervalls 5620 split_and_spill_intersecting_intervals(reg, regHi); 5621 } 5622 } 5623 5624 bool LinearScanWalker::no_allocation_possible(Interval* cur) { 5625 #ifdef X86 5626 // fast calculation of intervals that can never get a register because the 5627 // the next instruction is a call that blocks all registers 5628 // Note: this does not work if callee-saved registers are available (e.g. on Sparc) 5629 5630 // check if this interval is the result of a split operation 5631 // (an interval got a register until this position) 5632 int pos = cur->from(); 5633 if ((pos & 1) == 1) { 5634 // the current instruction is a call that blocks all registers 5635 if (pos < allocator()->max_lir_op_id() && allocator()->has_call(pos + 1)) { 5636 TRACE_LINEAR_SCAN(4, tty->print_cr(" free register cannot be available because all registers blocked by following call")); 5637 5638 // safety check that there is really no register available 5639 assert(alloc_free_reg(cur) == false, "found a register for this interval"); 5640 return true; 5641 } 5642 5643 } 5644 #endif 5645 return false; 5646 } 5647 5648 void LinearScanWalker::init_vars_for_alloc(Interval* cur) { 5649 BasicType type = cur->type(); 5650 _num_phys_regs = LinearScan::num_physical_regs(type); 5651 _adjacent_regs = LinearScan::requires_adjacent_regs(type); 5652 5653 if (pd_init_regs_for_alloc(cur)) { 5654 // the appropriate register range was selected. 5655 } else if (type == T_FLOAT || type == T_DOUBLE) { 5656 _first_reg = pd_first_fpu_reg; 5657 _last_reg = pd_last_fpu_reg; 5658 } else { 5659 _first_reg = pd_first_cpu_reg; 5660 _last_reg = FrameMap::last_cpu_reg(); 5661 } 5662 5663 assert(0 <= _first_reg && _first_reg < LinearScan::nof_regs, "out of range"); 5664 assert(0 <= _last_reg && _last_reg < LinearScan::nof_regs, "out of range"); 5665 } 5666 5667 5668 bool LinearScanWalker::is_move(LIR_Op* op, Interval* from, Interval* to) { 5669 if (op->code() != lir_move) { 5670 return false; 5671 } 5672 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); 5673 5674 LIR_Opr in = ((LIR_Op1*)op)->in_opr(); 5675 LIR_Opr res = ((LIR_Op1*)op)->result_opr(); 5676 return in->is_virtual() && res->is_virtual() && in->vreg_number() == from->reg_num() && res->vreg_number() == to->reg_num(); 5677 } 5678 5679 // optimization (especially for phi functions of nested loops): 5680 // assign same spill slot to non-intersecting intervals 5681 void LinearScanWalker::combine_spilled_intervals(Interval* cur) { 5682 if (cur->is_split_child()) { 5683 // optimization is only suitable for split parents 5684 return; 5685 } 5686 5687 Interval* register_hint = cur->register_hint(false); 5688 if (register_hint == NULL) { 5689 // cur is not the target of a move, otherwise register_hint would be set 5690 return; 5691 } 5692 assert(register_hint->is_split_parent(), "register hint must be split parent"); 5693 5694 if (cur->spill_state() != noOptimization || register_hint->spill_state() != noOptimization) { 5695 // combining the stack slots for intervals where spill move optimization is applied 5696 // is not benefitial and would cause problems 5697 return; 5698 } 5699 5700 int begin_pos = cur->from(); 5701 int end_pos = cur->to(); 5702 if (end_pos > allocator()->max_lir_op_id() || (begin_pos & 1) != 0 || (end_pos & 1) != 0) { 5703 // safety check that lir_op_with_id is allowed 5704 return; 5705 } 5706 5707 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)) { 5708 // cur and register_hint are not connected with two moves 5709 return; 5710 } 5711 5712 Interval* begin_hint = register_hint->split_child_at_op_id(begin_pos, LIR_OpVisitState::inputMode); 5713 Interval* end_hint = register_hint->split_child_at_op_id(end_pos, LIR_OpVisitState::outputMode); 5714 if (begin_hint == end_hint || begin_hint->to() != begin_pos || end_hint->from() != end_pos) { 5715 // register_hint must be split, otherwise the re-writing of use positions does not work 5716 return; 5717 } 5718 5719 assert(begin_hint->assigned_reg() != any_reg, "must have register assigned"); 5720 assert(end_hint->assigned_reg() == any_reg, "must not have register assigned"); 5721 assert(cur->first_usage(mustHaveRegister) == begin_pos, "must have use position at begin of interval because of move"); 5722 assert(end_hint->first_usage(mustHaveRegister) == end_pos, "must have use position at begin of interval because of move"); 5723 5724 if (begin_hint->assigned_reg() < LinearScan::nof_regs) { 5725 // register_hint is not spilled at begin_pos, so it would not be benefitial to immediately spill cur 5726 return; 5727 } 5728 assert(register_hint->canonical_spill_slot() != -1, "must be set when part of interval was spilled"); 5729 5730 // modify intervals such that cur gets the same stack slot as register_hint 5731 // delete use positions to prevent the intervals to get a register at beginning 5732 cur->set_canonical_spill_slot(register_hint->canonical_spill_slot()); 5733 cur->remove_first_use_pos(); 5734 end_hint->remove_first_use_pos(); 5735 } 5736 5737 5738 // allocate a physical register or memory location to an interval 5739 bool LinearScanWalker::activate_current() { 5740 Interval* cur = current(); 5741 bool result = true; 5742 5743 TRACE_LINEAR_SCAN(2, tty->print ("+++++ activating interval "); cur->print()); 5744 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())); 5745 5746 if (cur->assigned_reg() >= LinearScan::nof_regs) { 5747 // activating an interval that has a stack slot assigned -> split it at first use position 5748 // used for method parameters 5749 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has spill slot assigned (method parameter) -> split it before first use")); 5750 5751 split_stack_interval(cur); 5752 result = false; 5753 5754 } else if (allocator()->gen()->is_vreg_flag_set(cur->reg_num(), LIRGenerator::must_start_in_memory)) { 5755 // activating an interval that must start in a stack slot, but may get a register later 5756 // used for lir_roundfp: rounding is done by store to stack and reload later 5757 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval must start in stack slot -> split it before first use")); 5758 assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned"); 5759 5760 allocator()->assign_spill_slot(cur); 5761 split_stack_interval(cur); 5762 result = false; 5763 5764 } else if (cur->assigned_reg() == any_reg) { 5765 // interval has not assigned register -> normal allocation 5766 // (this is the normal case for most intervals) 5767 TRACE_LINEAR_SCAN(4, tty->print_cr(" normal allocation of register")); 5768 5769 // assign same spill slot to non-intersecting intervals 5770 combine_spilled_intervals(cur); 5771 5772 init_vars_for_alloc(cur); 5773 if (no_allocation_possible(cur) || !alloc_free_reg(cur)) { 5774 // no empty register available. 5775 // split and spill another interval so that this interval gets a register 5776 alloc_locked_reg(cur); 5777 } 5778 5779 // spilled intervals need not be move to active-list 5780 if (cur->assigned_reg() >= LinearScan::nof_regs) { 5781 result = false; 5782 } 5783 } 5784 5785 // load spilled values that become active from stack slot to register 5786 if (cur->insert_move_when_activated()) { 5787 assert(cur->is_split_child(), "must be"); 5788 assert(cur->current_split_child() != NULL, "must be"); 5789 assert(cur->current_split_child()->reg_num() != cur->reg_num(), "cannot insert move between same interval"); 5790 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())); 5791 5792 insert_move(cur->from(), cur->current_split_child(), cur); 5793 } 5794 cur->make_current_split_child(); 5795 5796 return result; // true = interval is moved to active list 5797 } 5798 5799 5800 // Implementation of EdgeMoveOptimizer 5801 5802 EdgeMoveOptimizer::EdgeMoveOptimizer() : 5803 _edge_instructions(4), 5804 _edge_instructions_idx(4) 5805 { 5806 } 5807 5808 void EdgeMoveOptimizer::optimize(BlockList* code) { 5809 EdgeMoveOptimizer optimizer = EdgeMoveOptimizer(); 5810 5811 // ignore the first block in the list (index 0 is not processed) 5812 for (int i = code->length() - 1; i >= 1; i--) { 5813 BlockBegin* block = code->at(i); 5814 5815 if (block->number_of_preds() > 1 && !block->is_set(BlockBegin::exception_entry_flag)) { 5816 optimizer.optimize_moves_at_block_end(block); 5817 } 5818 if (block->number_of_sux() == 2) { 5819 optimizer.optimize_moves_at_block_begin(block); 5820 } 5821 } 5822 } 5823 5824 5825 // clear all internal data structures 5826 void EdgeMoveOptimizer::init_instructions() { 5827 _edge_instructions.clear(); 5828 _edge_instructions_idx.clear(); 5829 } 5830 5831 // append a lir-instruction-list and the index of the current operation in to the list 5832 void EdgeMoveOptimizer::append_instructions(LIR_OpList* instructions, int instructions_idx) { 5833 _edge_instructions.append(instructions); 5834 _edge_instructions_idx.append(instructions_idx); 5835 } 5836 5837 // return the current operation of the given edge (predecessor or successor) 5838 LIR_Op* EdgeMoveOptimizer::instruction_at(int edge) { 5839 LIR_OpList* instructions = _edge_instructions.at(edge); 5840 int idx = _edge_instructions_idx.at(edge); 5841 5842 if (idx < instructions->length()) { 5843 return instructions->at(idx); 5844 } else { 5845 return NULL; 5846 } 5847 } 5848 5849 // removes the current operation of the given edge (predecessor or successor) 5850 void EdgeMoveOptimizer::remove_cur_instruction(int edge, bool decrement_index) { 5851 LIR_OpList* instructions = _edge_instructions.at(edge); 5852 int idx = _edge_instructions_idx.at(edge); 5853 instructions->remove_at(idx); 5854 5855 if (decrement_index) { 5856 _edge_instructions_idx.at_put(edge, idx - 1); 5857 } 5858 } 5859 5860 5861 bool EdgeMoveOptimizer::operations_different(LIR_Op* op1, LIR_Op* op2) { 5862 if (op1 == NULL || op2 == NULL) { 5863 // at least one block is already empty -> no optimization possible 5864 return true; 5865 } 5866 5867 if (op1->code() == lir_move && op2->code() == lir_move) { 5868 assert(op1->as_Op1() != NULL, "move must be LIR_Op1"); 5869 assert(op2->as_Op1() != NULL, "move must be LIR_Op1"); 5870 LIR_Op1* move1 = (LIR_Op1*)op1; 5871 LIR_Op1* move2 = (LIR_Op1*)op2; 5872 if (move1->info() == move2->info() && move1->in_opr() == move2->in_opr() && move1->result_opr() == move2->result_opr()) { 5873 // these moves are exactly equal and can be optimized 5874 return false; 5875 } 5876 5877 } else if (op1->code() == lir_fxch && op2->code() == lir_fxch) { 5878 assert(op1->as_Op1() != NULL, "fxch must be LIR_Op1"); 5879 assert(op2->as_Op1() != NULL, "fxch must be LIR_Op1"); 5880 LIR_Op1* fxch1 = (LIR_Op1*)op1; 5881 LIR_Op1* fxch2 = (LIR_Op1*)op2; 5882 if (fxch1->in_opr()->as_jint() == fxch2->in_opr()->as_jint()) { 5883 // equal FPU stack operations can be optimized 5884 return false; 5885 } 5886 5887 } else if (op1->code() == lir_fpop_raw && op2->code() == lir_fpop_raw) { 5888 // equal FPU stack operations can be optimized 5889 return false; 5890 } 5891 5892 // no optimization possible 5893 return true; 5894 } 5895 5896 void EdgeMoveOptimizer::optimize_moves_at_block_end(BlockBegin* block) { 5897 TRACE_LINEAR_SCAN(4, tty->print_cr("optimizing moves at end of block B%d", block->block_id())); 5898 5899 if (block->is_predecessor(block)) { 5900 // currently we can't handle this correctly. 5901 return; 5902 } 5903 5904 init_instructions(); 5905 int num_preds = block->number_of_preds(); 5906 assert(num_preds > 1, "do not call otherwise"); 5907 assert(!block->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed"); 5908 5909 // setup a list with the lir-instructions of all predecessors 5910 int i; 5911 for (i = 0; i < num_preds; i++) { 5912 BlockBegin* pred = block->pred_at(i); 5913 LIR_OpList* pred_instructions = pred->lir()->instructions_list(); 5914 5915 if (pred->number_of_sux() != 1) { 5916 // this can happen with switch-statements where multiple edges are between 5917 // the same blocks. 5918 return; 5919 } 5920 5921 assert(pred->number_of_sux() == 1, "can handle only one successor"); 5922 assert(pred->sux_at(0) == block, "invalid control flow"); 5923 assert(pred_instructions->last()->code() == lir_branch, "block with successor must end with branch"); 5924 assert(pred_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch"); 5925 assert(pred_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch"); 5926 5927 if (pred_instructions->last()->info() != NULL) { 5928 // can not optimize instructions when debug info is needed 5929 return; 5930 } 5931 5932 // ignore the unconditional branch at the end of the block 5933 append_instructions(pred_instructions, pred_instructions->length() - 2); 5934 } 5935 5936 5937 // process lir-instructions while all predecessors end with the same instruction 5938 while (true) { 5939 LIR_Op* op = instruction_at(0); 5940 for (i = 1; i < num_preds; i++) { 5941 if (operations_different(op, instruction_at(i))) { 5942 // these instructions are different and cannot be optimized -> 5943 // no further optimization possible 5944 return; 5945 } 5946 } 5947 5948 TRACE_LINEAR_SCAN(4, tty->print("found instruction that is equal in all %d predecessors: ", num_preds); op->print()); 5949 5950 // insert the instruction at the beginning of the current block 5951 block->lir()->insert_before(1, op); 5952 5953 // delete the instruction at the end of all predecessors 5954 for (i = 0; i < num_preds; i++) { 5955 remove_cur_instruction(i, true); 5956 } 5957 } 5958 } 5959 5960 5961 void EdgeMoveOptimizer::optimize_moves_at_block_begin(BlockBegin* block) { 5962 TRACE_LINEAR_SCAN(4, tty->print_cr("optimization moves at begin of block B%d", block->block_id())); 5963 5964 init_instructions(); 5965 int num_sux = block->number_of_sux(); 5966 5967 LIR_OpList* cur_instructions = block->lir()->instructions_list(); 5968 5969 assert(num_sux == 2, "method should not be called otherwise"); 5970 assert(cur_instructions->last()->code() == lir_branch, "block with successor must end with branch"); 5971 assert(cur_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch"); 5972 assert(cur_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch"); 5973 5974 if (cur_instructions->last()->info() != NULL) { 5975 // can no optimize instructions when debug info is needed 5976 return; 5977 } 5978 5979 LIR_Op* branch = cur_instructions->at(cur_instructions->length() - 2); 5980 if (branch->info() != NULL || (branch->code() != lir_branch && branch->code() != lir_cond_float_branch)) { 5981 // not a valid case for optimization 5982 // currently, only blocks that end with two branches (conditional branch followed 5983 // by unconditional branch) are optimized 5984 return; 5985 } 5986 5987 // now it is guaranteed that the block ends with two branch instructions. 5988 // the instructions are inserted at the end of the block before these two branches 5989 int insert_idx = cur_instructions->length() - 2; 5990 5991 int i; 5992 #ifdef ASSERT 5993 for (i = insert_idx - 1; i >= 0; i--) { 5994 LIR_Op* op = cur_instructions->at(i); 5995 if ((op->code() == lir_branch || op->code() == lir_cond_float_branch) && ((LIR_OpBranch*)op)->block() != NULL) { 5996 assert(false, "block with two successors can have only two branch instructions"); 5997 } 5998 } 5999 #endif 6000 6001 // setup a list with the lir-instructions of all successors 6002 for (i = 0; i < num_sux; i++) { 6003 BlockBegin* sux = block->sux_at(i); 6004 LIR_OpList* sux_instructions = sux->lir()->instructions_list(); 6005 6006 assert(sux_instructions->at(0)->code() == lir_label, "block must start with label"); 6007 6008 if (sux->number_of_preds() != 1) { 6009 // this can happen with switch-statements where multiple edges are between 6010 // the same blocks. 6011 return; 6012 } 6013 assert(sux->pred_at(0) == block, "invalid control flow"); 6014 assert(!sux->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed"); 6015 6016 // ignore the label at the beginning of the block 6017 append_instructions(sux_instructions, 1); 6018 } 6019 6020 // process lir-instructions while all successors begin with the same instruction 6021 while (true) { 6022 LIR_Op* op = instruction_at(0); 6023 for (i = 1; i < num_sux; i++) { 6024 if (operations_different(op, instruction_at(i))) { 6025 // these instructions are different and cannot be optimized -> 6026 // no further optimization possible 6027 return; 6028 } 6029 } 6030 6031 TRACE_LINEAR_SCAN(4, tty->print("----- found instruction that is equal in all %d successors: ", num_sux); op->print()); 6032 6033 // insert instruction at end of current block 6034 block->lir()->insert_before(insert_idx, op); 6035 insert_idx++; 6036 6037 // delete the instructions at the beginning of all successors 6038 for (i = 0; i < num_sux; i++) { 6039 remove_cur_instruction(i, false); 6040 } 6041 } 6042 } 6043 6044 6045 // Implementation of ControlFlowOptimizer 6046 6047 ControlFlowOptimizer::ControlFlowOptimizer() : 6048 _original_preds(4) 6049 { 6050 } 6051 6052 void ControlFlowOptimizer::optimize(BlockList* code) { 6053 ControlFlowOptimizer optimizer = ControlFlowOptimizer(); 6054 6055 // push the OSR entry block to the end so that we're not jumping over it. 6056 BlockBegin* osr_entry = code->at(0)->end()->as_Base()->osr_entry(); 6057 if (osr_entry) { 6058 int index = osr_entry->linear_scan_number(); 6059 assert(code->at(index) == osr_entry, "wrong index"); 6060 code->remove_at(index); 6061 code->append(osr_entry); 6062 } 6063 6064 optimizer.reorder_short_loops(code); 6065 optimizer.delete_empty_blocks(code); 6066 optimizer.delete_unnecessary_jumps(code); 6067 optimizer.delete_jumps_to_return(code); 6068 } 6069 6070 void ControlFlowOptimizer::reorder_short_loop(BlockList* code, BlockBegin* header_block, int header_idx) { 6071 int i = header_idx + 1; 6072 int max_end = MIN2(header_idx + ShortLoopSize, code->length()); 6073 while (i < max_end && code->at(i)->loop_depth() >= header_block->loop_depth()) { 6074 i++; 6075 } 6076 6077 if (i == code->length() || code->at(i)->loop_depth() < header_block->loop_depth()) { 6078 int end_idx = i - 1; 6079 BlockBegin* end_block = code->at(end_idx); 6080 6081 if (end_block->number_of_sux() == 1 && end_block->sux_at(0) == header_block) { 6082 // short loop from header_idx to end_idx found -> reorder blocks such that 6083 // the header_block is the last block instead of the first block of the loop 6084 TRACE_LINEAR_SCAN(1, tty->print_cr("Reordering short loop: length %d, header B%d, end B%d", 6085 end_idx - header_idx + 1, 6086 header_block->block_id(), end_block->block_id())); 6087 6088 for (int j = header_idx; j < end_idx; j++) { 6089 code->at_put(j, code->at(j + 1)); 6090 } 6091 code->at_put(end_idx, header_block); 6092 6093 // correct the flags so that any loop alignment occurs in the right place. 6094 assert(code->at(end_idx)->is_set(BlockBegin::backward_branch_target_flag), "must be backward branch target"); 6095 code->at(end_idx)->clear(BlockBegin::backward_branch_target_flag); 6096 code->at(header_idx)->set(BlockBegin::backward_branch_target_flag); 6097 } 6098 } 6099 } 6100 6101 void ControlFlowOptimizer::reorder_short_loops(BlockList* code) { 6102 for (int i = code->length() - 1; i >= 0; i--) { 6103 BlockBegin* block = code->at(i); 6104 6105 if (block->is_set(BlockBegin::linear_scan_loop_header_flag)) { 6106 reorder_short_loop(code, block, i); 6107 } 6108 } 6109 6110 DEBUG_ONLY(verify(code)); 6111 } 6112 6113 // only blocks with exactly one successor can be deleted. Such blocks 6114 // must always end with an unconditional branch to this successor 6115 bool ControlFlowOptimizer::can_delete_block(BlockBegin* block) { 6116 if (block->number_of_sux() != 1 || block->number_of_exception_handlers() != 0 || block->is_entry_block()) { 6117 return false; 6118 } 6119 6120 LIR_OpList* instructions = block->lir()->instructions_list(); 6121 6122 assert(instructions->length() >= 2, "block must have label and branch"); 6123 assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label"); 6124 assert(instructions->last()->as_OpBranch() != NULL, "last instrcution must always be a branch"); 6125 assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "branch must be unconditional"); 6126 assert(instructions->last()->as_OpBranch()->block() == block->sux_at(0), "branch target must be the successor"); 6127 6128 // block must have exactly one successor 6129 6130 if (instructions->length() == 2 && instructions->last()->info() == NULL) { 6131 return true; 6132 } 6133 return false; 6134 } 6135 6136 // substitute branch targets in all branch-instructions of this blocks 6137 void ControlFlowOptimizer::substitute_branch_target(BlockBegin* block, BlockBegin* target_from, BlockBegin* target_to) { 6138 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())); 6139 6140 LIR_OpList* instructions = block->lir()->instructions_list(); 6141 6142 assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label"); 6143 for (int i = instructions->length() - 1; i >= 1; i--) { 6144 LIR_Op* op = instructions->at(i); 6145 6146 if (op->code() == lir_branch || op->code() == lir_cond_float_branch) { 6147 assert(op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch"); 6148 LIR_OpBranch* branch = (LIR_OpBranch*)op; 6149 6150 if (branch->block() == target_from) { 6151 branch->change_block(target_to); 6152 } 6153 if (branch->ublock() == target_from) { 6154 branch->change_ublock(target_to); 6155 } 6156 } 6157 } 6158 } 6159 6160 void ControlFlowOptimizer::delete_empty_blocks(BlockList* code) { 6161 int old_pos = 0; 6162 int new_pos = 0; 6163 int num_blocks = code->length(); 6164 6165 while (old_pos < num_blocks) { 6166 BlockBegin* block = code->at(old_pos); 6167 6168 if (can_delete_block(block)) { 6169 BlockBegin* new_target = block->sux_at(0); 6170 6171 // propagate backward branch target flag for correct code alignment 6172 if (block->is_set(BlockBegin::backward_branch_target_flag)) { 6173 new_target->set(BlockBegin::backward_branch_target_flag); 6174 } 6175 6176 // collect a list with all predecessors that contains each predecessor only once 6177 // the predecessors of cur are changed during the substitution, so a copy of the 6178 // predecessor list is necessary 6179 int j; 6180 _original_preds.clear(); 6181 for (j = block->number_of_preds() - 1; j >= 0; j--) { 6182 BlockBegin* pred = block->pred_at(j); 6183 if (_original_preds.index_of(pred) == -1) { 6184 _original_preds.append(pred); 6185 } 6186 } 6187 6188 for (j = _original_preds.length() - 1; j >= 0; j--) { 6189 BlockBegin* pred = _original_preds.at(j); 6190 substitute_branch_target(pred, block, new_target); 6191 pred->substitute_sux(block, new_target); 6192 } 6193 } else { 6194 // adjust position of this block in the block list if blocks before 6195 // have been deleted 6196 if (new_pos != old_pos) { 6197 code->at_put(new_pos, code->at(old_pos)); 6198 } 6199 new_pos++; 6200 } 6201 old_pos++; 6202 } 6203 code->truncate(new_pos); 6204 6205 DEBUG_ONLY(verify(code)); 6206 } 6207 6208 void ControlFlowOptimizer::delete_unnecessary_jumps(BlockList* code) { 6209 // skip the last block because there a branch is always necessary 6210 for (int i = code->length() - 2; i >= 0; i--) { 6211 BlockBegin* block = code->at(i); 6212 LIR_OpList* instructions = block->lir()->instructions_list(); 6213 6214 LIR_Op* last_op = instructions->last(); 6215 if (last_op->code() == lir_branch) { 6216 assert(last_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch"); 6217 LIR_OpBranch* last_branch = (LIR_OpBranch*)last_op; 6218 6219 assert(last_branch->block() != NULL, "last branch must always have a block as target"); 6220 assert(last_branch->label() == last_branch->block()->label(), "must be equal"); 6221 6222 if (last_branch->info() == NULL) { 6223 if (last_branch->block() == code->at(i + 1)) { 6224 6225 TRACE_LINEAR_SCAN(3, tty->print_cr("Deleting unconditional branch at end of block B%d", block->block_id())); 6226 6227 // delete last branch instruction 6228 instructions->truncate(instructions->length() - 1); 6229 6230 } else { 6231 LIR_Op* prev_op = instructions->at(instructions->length() - 2); 6232 if (prev_op->code() == lir_branch || prev_op->code() == lir_cond_float_branch) { 6233 assert(prev_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch"); 6234 LIR_OpBranch* prev_branch = (LIR_OpBranch*)prev_op; 6235 6236 if (prev_branch->stub() == NULL) { 6237 6238 LIR_Op2* prev_cmp = NULL; 6239 6240 for(int j = instructions->length() - 3; j >= 0 && prev_cmp == NULL; j--) { 6241 prev_op = instructions->at(j); 6242 if (prev_op->code() == lir_cmp) { 6243 assert(prev_op->as_Op2() != NULL, "branch must be of type LIR_Op2"); 6244 prev_cmp = (LIR_Op2*)prev_op; 6245 assert(prev_branch->cond() == prev_cmp->condition(), "should be the same"); 6246 } 6247 } 6248 assert(prev_cmp != NULL, "should have found comp instruction for branch"); 6249 if (prev_branch->block() == code->at(i + 1) && prev_branch->info() == NULL) { 6250 6251 TRACE_LINEAR_SCAN(3, tty->print_cr("Negating conditional branch and deleting unconditional branch at end of block B%d", block->block_id())); 6252 6253 // eliminate a conditional branch to the immediate successor 6254 prev_branch->change_block(last_branch->block()); 6255 prev_branch->negate_cond(); 6256 prev_cmp->set_condition(prev_branch->cond()); 6257 instructions->truncate(instructions->length() - 1); 6258 } 6259 } 6260 } 6261 } 6262 } 6263 } 6264 } 6265 6266 DEBUG_ONLY(verify(code)); 6267 } 6268 6269 void ControlFlowOptimizer::delete_jumps_to_return(BlockList* code) { 6270 #ifdef ASSERT 6271 BitMap return_converted(BlockBegin::number_of_blocks()); 6272 return_converted.clear(); 6273 #endif 6274 6275 for (int i = code->length() - 1; i >= 0; i--) { 6276 BlockBegin* block = code->at(i); 6277 LIR_OpList* cur_instructions = block->lir()->instructions_list(); 6278 LIR_Op* cur_last_op = cur_instructions->last(); 6279 6280 assert(cur_instructions->at(0)->code() == lir_label, "first instruction must always be a label"); 6281 if (cur_instructions->length() == 2 && cur_last_op->code() == lir_return) { 6282 // the block contains only a label and a return 6283 // if a predecessor ends with an unconditional jump to this block, then the jump 6284 // can be replaced with a return instruction 6285 // 6286 // Note: the original block with only a return statement cannot be deleted completely 6287 // because the predecessors might have other (conditional) jumps to this block 6288 // -> this may lead to unnecesary return instructions in the final code 6289 6290 assert(cur_last_op->info() == NULL, "return instructions do not have debug information"); 6291 assert(block->number_of_sux() == 0 || 6292 (return_converted.at(block->block_id()) && block->number_of_sux() == 1), 6293 "blocks that end with return must not have successors"); 6294 6295 assert(cur_last_op->as_Op1() != NULL, "return must be LIR_Op1"); 6296 LIR_Opr return_opr = ((LIR_Op1*)cur_last_op)->in_opr(); 6297 6298 for (int j = block->number_of_preds() - 1; j >= 0; j--) { 6299 BlockBegin* pred = block->pred_at(j); 6300 LIR_OpList* pred_instructions = pred->lir()->instructions_list(); 6301 LIR_Op* pred_last_op = pred_instructions->last(); 6302 6303 if (pred_last_op->code() == lir_branch) { 6304 assert(pred_last_op->as_OpBranch() != NULL, "branch must be LIR_OpBranch"); 6305 LIR_OpBranch* pred_last_branch = (LIR_OpBranch*)pred_last_op; 6306 6307 if (pred_last_branch->block() == block && pred_last_branch->cond() == lir_cond_always && pred_last_branch->info() == NULL) { 6308 // replace the jump to a return with a direct return 6309 // Note: currently the edge between the blocks is not deleted 6310 pred_instructions->at_put(pred_instructions->length() - 1, new LIR_Op1(lir_return, return_opr)); 6311 #ifdef ASSERT 6312 return_converted.set_bit(pred->block_id()); 6313 #endif 6314 } 6315 } 6316 } 6317 } 6318 } 6319 } 6320 6321 6322 #ifdef ASSERT 6323 void ControlFlowOptimizer::verify(BlockList* code) { 6324 for (int i = 0; i < code->length(); i++) { 6325 BlockBegin* block = code->at(i); 6326 LIR_OpList* instructions = block->lir()->instructions_list(); 6327 6328 int j; 6329 for (j = 0; j < instructions->length(); j++) { 6330 LIR_OpBranch* op_branch = instructions->at(j)->as_OpBranch(); 6331 6332 if (op_branch != NULL) { 6333 assert(op_branch->block() == NULL || code->index_of(op_branch->block()) != -1, "branch target not valid"); 6334 assert(op_branch->ublock() == NULL || code->index_of(op_branch->ublock()) != -1, "branch target not valid"); 6335 } 6336 } 6337 6338 for (j = 0; j < block->number_of_sux() - 1; j++) { 6339 BlockBegin* sux = block->sux_at(j); 6340 assert(code->index_of(sux) != -1, "successor not valid"); 6341 } 6342 6343 for (j = 0; j < block->number_of_preds() - 1; j++) { 6344 BlockBegin* pred = block->pred_at(j); 6345 assert(code->index_of(pred) != -1, "successor not valid"); 6346 } 6347 } 6348 } 6349 #endif 6350 6351 6352 #ifndef PRODUCT 6353 6354 // Implementation of LinearStatistic 6355 6356 const char* LinearScanStatistic::counter_name(int counter_idx) { 6357 switch (counter_idx) { 6358 case counter_method: return "compiled methods"; 6359 case counter_fpu_method: return "methods using fpu"; 6360 case counter_loop_method: return "methods with loops"; 6361 case counter_exception_method:return "methods with xhandler"; 6362 6363 case counter_loop: return "loops"; 6364 case counter_block: return "blocks"; 6365 case counter_loop_block: return "blocks inside loop"; 6366 case counter_exception_block: return "exception handler entries"; 6367 case counter_interval: return "intervals"; 6368 case counter_fixed_interval: return "fixed intervals"; 6369 case counter_range: return "ranges"; 6370 case counter_fixed_range: return "fixed ranges"; 6371 case counter_use_pos: return "use positions"; 6372 case counter_fixed_use_pos: return "fixed use positions"; 6373 case counter_spill_slots: return "spill slots"; 6374 6375 // counter for classes of lir instructions 6376 case counter_instruction: return "total instructions"; 6377 case counter_label: return "labels"; 6378 case counter_entry: return "method entries"; 6379 case counter_return: return "method returns"; 6380 case counter_call: return "method calls"; 6381 case counter_move: return "moves"; 6382 case counter_cmp: return "compare"; 6383 case counter_cond_branch: return "conditional branches"; 6384 case counter_uncond_branch: return "unconditional branches"; 6385 case counter_stub_branch: return "branches to stub"; 6386 case counter_alu: return "artithmetic + logic"; 6387 case counter_alloc: return "allocations"; 6388 case counter_sync: return "synchronisation"; 6389 case counter_throw: return "throw"; 6390 case counter_unwind: return "unwind"; 6391 case counter_typecheck: return "type+null-checks"; 6392 case counter_fpu_stack: return "fpu-stack"; 6393 case counter_misc_inst: return "other instructions"; 6394 case counter_other_inst: return "misc. instructions"; 6395 6396 // counter for different types of moves 6397 case counter_move_total: return "total moves"; 6398 case counter_move_reg_reg: return "register->register"; 6399 case counter_move_reg_stack: return "register->stack"; 6400 case counter_move_stack_reg: return "stack->register"; 6401 case counter_move_stack_stack:return "stack->stack"; 6402 case counter_move_reg_mem: return "register->memory"; 6403 case counter_move_mem_reg: return "memory->register"; 6404 case counter_move_const_any: return "constant->any"; 6405 6406 case blank_line_1: return ""; 6407 case blank_line_2: return ""; 6408 6409 default: ShouldNotReachHere(); return ""; 6410 } 6411 } 6412 6413 LinearScanStatistic::Counter LinearScanStatistic::base_counter(int counter_idx) { 6414 if (counter_idx == counter_fpu_method || counter_idx == counter_loop_method || counter_idx == counter_exception_method) { 6415 return counter_method; 6416 } else if (counter_idx == counter_loop_block || counter_idx == counter_exception_block) { 6417 return counter_block; 6418 } else if (counter_idx >= counter_instruction && counter_idx <= counter_other_inst) { 6419 return counter_instruction; 6420 } else if (counter_idx >= counter_move_total && counter_idx <= counter_move_const_any) { 6421 return counter_move_total; 6422 } 6423 return invalid_counter; 6424 } 6425 6426 LinearScanStatistic::LinearScanStatistic() { 6427 for (int i = 0; i < number_of_counters; i++) { 6428 _counters_sum[i] = 0; 6429 _counters_max[i] = -1; 6430 } 6431 6432 } 6433 6434 // add the method-local numbers to the total sum 6435 void LinearScanStatistic::sum_up(LinearScanStatistic &method_statistic) { 6436 for (int i = 0; i < number_of_counters; i++) { 6437 _counters_sum[i] += method_statistic._counters_sum[i]; 6438 _counters_max[i] = MAX2(_counters_max[i], method_statistic._counters_sum[i]); 6439 } 6440 } 6441 6442 void LinearScanStatistic::print(const char* title) { 6443 if (CountLinearScan || TraceLinearScanLevel > 0) { 6444 tty->cr(); 6445 tty->print_cr("***** LinearScan statistic - %s *****", title); 6446 6447 for (int i = 0; i < number_of_counters; i++) { 6448 if (_counters_sum[i] > 0 || _counters_max[i] >= 0) { 6449 tty->print("%25s: %8d", counter_name(i), _counters_sum[i]); 6450 6451 if (base_counter(i) != invalid_counter) { 6452 tty->print(" (%5.1f%%) ", _counters_sum[i] * 100.0 / _counters_sum[base_counter(i)]); 6453 } else { 6454 tty->print(" "); 6455 } 6456 6457 if (_counters_max[i] >= 0) { 6458 tty->print("%8d", _counters_max[i]); 6459 } 6460 } 6461 tty->cr(); 6462 } 6463 } 6464 } 6465 6466 void LinearScanStatistic::collect(LinearScan* allocator) { 6467 inc_counter(counter_method); 6468 if (allocator->has_fpu_registers()) { 6469 inc_counter(counter_fpu_method); 6470 } 6471 if (allocator->num_loops() > 0) { 6472 inc_counter(counter_loop_method); 6473 } 6474 inc_counter(counter_loop, allocator->num_loops()); 6475 inc_counter(counter_spill_slots, allocator->max_spills()); 6476 6477 int i; 6478 for (i = 0; i < allocator->interval_count(); i++) { 6479 Interval* cur = allocator->interval_at(i); 6480 6481 if (cur != NULL) { 6482 inc_counter(counter_interval); 6483 inc_counter(counter_use_pos, cur->num_use_positions()); 6484 if (LinearScan::is_precolored_interval(cur)) { 6485 inc_counter(counter_fixed_interval); 6486 inc_counter(counter_fixed_use_pos, cur->num_use_positions()); 6487 } 6488 6489 Range* range = cur->first(); 6490 while (range != Range::end()) { 6491 inc_counter(counter_range); 6492 if (LinearScan::is_precolored_interval(cur)) { 6493 inc_counter(counter_fixed_range); 6494 } 6495 range = range->next(); 6496 } 6497 } 6498 } 6499 6500 bool has_xhandlers = false; 6501 // Note: only count blocks that are in code-emit order 6502 for (i = 0; i < allocator->ir()->code()->length(); i++) { 6503 BlockBegin* cur = allocator->ir()->code()->at(i); 6504 6505 inc_counter(counter_block); 6506 if (cur->loop_depth() > 0) { 6507 inc_counter(counter_loop_block); 6508 } 6509 if (cur->is_set(BlockBegin::exception_entry_flag)) { 6510 inc_counter(counter_exception_block); 6511 has_xhandlers = true; 6512 } 6513 6514 LIR_OpList* instructions = cur->lir()->instructions_list(); 6515 for (int j = 0; j < instructions->length(); j++) { 6516 LIR_Op* op = instructions->at(j); 6517 6518 inc_counter(counter_instruction); 6519 6520 switch (op->code()) { 6521 case lir_label: inc_counter(counter_label); break; 6522 case lir_std_entry: 6523 case lir_osr_entry: inc_counter(counter_entry); break; 6524 case lir_return: inc_counter(counter_return); break; 6525 6526 case lir_rtcall: 6527 case lir_static_call: 6528 case lir_optvirtual_call: 6529 case lir_virtual_call: inc_counter(counter_call); break; 6530 6531 case lir_move: { 6532 inc_counter(counter_move); 6533 inc_counter(counter_move_total); 6534 6535 LIR_Opr in = op->as_Op1()->in_opr(); 6536 LIR_Opr res = op->as_Op1()->result_opr(); 6537 if (in->is_register()) { 6538 if (res->is_register()) { 6539 inc_counter(counter_move_reg_reg); 6540 } else if (res->is_stack()) { 6541 inc_counter(counter_move_reg_stack); 6542 } else if (res->is_address()) { 6543 inc_counter(counter_move_reg_mem); 6544 } else { 6545 ShouldNotReachHere(); 6546 } 6547 } else if (in->is_stack()) { 6548 if (res->is_register()) { 6549 inc_counter(counter_move_stack_reg); 6550 } else { 6551 inc_counter(counter_move_stack_stack); 6552 } 6553 } else if (in->is_address()) { 6554 assert(res->is_register(), "must be"); 6555 inc_counter(counter_move_mem_reg); 6556 } else if (in->is_constant()) { 6557 inc_counter(counter_move_const_any); 6558 } else { 6559 ShouldNotReachHere(); 6560 } 6561 break; 6562 } 6563 6564 case lir_cmp: inc_counter(counter_cmp); break; 6565 6566 case lir_branch: 6567 case lir_cond_float_branch: { 6568 LIR_OpBranch* branch = op->as_OpBranch(); 6569 if (branch->block() == NULL) { 6570 inc_counter(counter_stub_branch); 6571 } else if (branch->cond() == lir_cond_always) { 6572 inc_counter(counter_uncond_branch); 6573 } else { 6574 inc_counter(counter_cond_branch); 6575 } 6576 break; 6577 } 6578 6579 case lir_neg: 6580 case lir_add: 6581 case lir_sub: 6582 case lir_mul: 6583 case lir_mul_strictfp: 6584 case lir_div: 6585 case lir_div_strictfp: 6586 case lir_rem: 6587 case lir_sqrt: 6588 case lir_sin: 6589 case lir_cos: 6590 case lir_abs: 6591 case lir_log10: 6592 case lir_log: 6593 case lir_pow: 6594 case lir_exp: 6595 case lir_logic_and: 6596 case lir_logic_or: 6597 case lir_logic_xor: 6598 case lir_shl: 6599 case lir_shr: 6600 case lir_ushr: inc_counter(counter_alu); break; 6601 6602 case lir_alloc_object: 6603 case lir_alloc_array: inc_counter(counter_alloc); break; 6604 6605 case lir_monaddr: 6606 case lir_lock: 6607 case lir_unlock: inc_counter(counter_sync); break; 6608 6609 case lir_throw: inc_counter(counter_throw); break; 6610 6611 case lir_unwind: inc_counter(counter_unwind); break; 6612 6613 case lir_null_check: 6614 case lir_leal: 6615 case lir_instanceof: 6616 case lir_checkcast: 6617 case lir_store_check: inc_counter(counter_typecheck); break; 6618 6619 case lir_fpop_raw: 6620 case lir_fxch: 6621 case lir_fld: inc_counter(counter_fpu_stack); break; 6622 6623 case lir_nop: 6624 case lir_push: 6625 case lir_pop: 6626 case lir_convert: 6627 case lir_roundfp: 6628 case lir_cmove: inc_counter(counter_misc_inst); break; 6629 6630 default: inc_counter(counter_other_inst); break; 6631 } 6632 } 6633 } 6634 6635 if (has_xhandlers) { 6636 inc_counter(counter_exception_method); 6637 } 6638 } 6639 6640 void LinearScanStatistic::compute(LinearScan* allocator, LinearScanStatistic &global_statistic) { 6641 if (CountLinearScan || TraceLinearScanLevel > 0) { 6642 6643 LinearScanStatistic local_statistic = LinearScanStatistic(); 6644 6645 local_statistic.collect(allocator); 6646 global_statistic.sum_up(local_statistic); 6647 6648 if (TraceLinearScanLevel > 2) { 6649 local_statistic.print("current local statistic"); 6650 } 6651 } 6652 } 6653 6654 6655 // Implementation of LinearTimers 6656 6657 LinearScanTimers::LinearScanTimers() { 6658 for (int i = 0; i < number_of_timers; i++) { 6659 timer(i)->reset(); 6660 } 6661 } 6662 6663 const char* LinearScanTimers::timer_name(int idx) { 6664 switch (idx) { 6665 case timer_do_nothing: return "Nothing (Time Check)"; 6666 case timer_number_instructions: return "Number Instructions"; 6667 case timer_compute_local_live_sets: return "Local Live Sets"; 6668 case timer_compute_global_live_sets: return "Global Live Sets"; 6669 case timer_build_intervals: return "Build Intervals"; 6670 case timer_sort_intervals_before: return "Sort Intervals Before"; 6671 case timer_allocate_registers: return "Allocate Registers"; 6672 case timer_resolve_data_flow: return "Resolve Data Flow"; 6673 case timer_sort_intervals_after: return "Sort Intervals After"; 6674 case timer_eliminate_spill_moves: return "Spill optimization"; 6675 case timer_assign_reg_num: return "Assign Reg Num"; 6676 case timer_allocate_fpu_stack: return "Allocate FPU Stack"; 6677 case timer_optimize_lir: return "Optimize LIR"; 6678 default: ShouldNotReachHere(); return ""; 6679 } 6680 } 6681 6682 void LinearScanTimers::begin_method() { 6683 if (TimeEachLinearScan) { 6684 // reset all timers to measure only current method 6685 for (int i = 0; i < number_of_timers; i++) { 6686 timer(i)->reset(); 6687 } 6688 } 6689 } 6690 6691 void LinearScanTimers::end_method(LinearScan* allocator) { 6692 if (TimeEachLinearScan) { 6693 6694 double c = timer(timer_do_nothing)->seconds(); 6695 double total = 0; 6696 for (int i = 1; i < number_of_timers; i++) { 6697 total += timer(i)->seconds() - c; 6698 } 6699 6700 if (total >= 0.0005) { 6701 // print all information in one line for automatic processing 6702 tty->print("@"); allocator->compilation()->method()->print_name(); 6703 6704 tty->print("@ %d ", allocator->compilation()->method()->code_size()); 6705 tty->print("@ %d ", allocator->block_at(allocator->block_count() - 1)->last_lir_instruction_id() / 2); 6706 tty->print("@ %d ", allocator->block_count()); 6707 tty->print("@ %d ", allocator->num_virtual_regs()); 6708 tty->print("@ %d ", allocator->interval_count()); 6709 tty->print("@ %d ", allocator->_num_calls); 6710 tty->print("@ %d ", allocator->num_loops()); 6711 6712 tty->print("@ %6.6f ", total); 6713 for (int i = 1; i < number_of_timers; i++) { 6714 tty->print("@ %4.1f ", ((timer(i)->seconds() - c) / total) * 100); 6715 } 6716 tty->cr(); 6717 } 6718 } 6719 } 6720 6721 void LinearScanTimers::print(double total_time) { 6722 if (TimeLinearScan) { 6723 // correction value: sum of dummy-timer that only measures the time that 6724 // is necesary to start and stop itself 6725 double c = timer(timer_do_nothing)->seconds(); 6726 6727 for (int i = 0; i < number_of_timers; i++) { 6728 double t = timer(i)->seconds(); 6729 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); 6730 } 6731 } 6732 } 6733 6734 #endif // #ifndef PRODUCT