1 /* 2 * Copyright (c) 1999, 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_Compilation.hpp" 27 #include "c1/c1_FrameMap.hpp" 28 #include "c1/c1_GraphBuilder.hpp" 29 #include "c1/c1_IR.hpp" 30 #include "c1/c1_InstructionPrinter.hpp" 31 #include "c1/c1_Optimizer.hpp" 32 #include "utilities/bitMap.inline.hpp" 33 34 35 // Implementation of XHandlers 36 // 37 // Note: This code could eventually go away if we are 38 // just using the ciExceptionHandlerStream. 39 40 XHandlers::XHandlers(ciMethod* method) : _list(method->exception_table_length()) { 41 ciExceptionHandlerStream s(method); 42 while (!s.is_done()) { 43 _list.append(new XHandler(s.handler())); 44 s.next(); 45 } 46 assert(s.count() == method->exception_table_length(), "exception table lengths inconsistent"); 47 } 48 49 // deep copy of all XHandler contained in list 50 XHandlers::XHandlers(XHandlers* other) : 51 _list(other->length()) 52 { 53 for (int i = 0; i < other->length(); i++) { 54 _list.append(new XHandler(other->handler_at(i))); 55 } 56 } 57 58 // Returns whether a particular exception type can be caught. Also 59 // returns true if klass is unloaded or any exception handler 60 // classes are unloaded. type_is_exact indicates whether the throw 61 // is known to be exactly that class or it might throw a subtype. 62 bool XHandlers::could_catch(ciInstanceKlass* klass, bool type_is_exact) const { 63 // the type is unknown so be conservative 64 if (!klass->is_loaded()) { 65 return true; 66 } 67 68 for (int i = 0; i < length(); i++) { 69 XHandler* handler = handler_at(i); 70 if (handler->is_catch_all()) { 71 // catch of ANY 72 return true; 73 } 74 ciInstanceKlass* handler_klass = handler->catch_klass(); 75 // if it's unknown it might be catchable 76 if (!handler_klass->is_loaded()) { 77 return true; 78 } 79 // if the throw type is definitely a subtype of the catch type 80 // then it can be caught. 81 if (klass->is_subtype_of(handler_klass)) { 82 return true; 83 } 84 if (!type_is_exact) { 85 // If the type isn't exactly known then it can also be caught by 86 // catch statements where the inexact type is a subtype of the 87 // catch type. 88 // given: foo extends bar extends Exception 89 // throw bar can be caught by catch foo, catch bar, and catch 90 // Exception, however it can't be caught by any handlers without 91 // bar in its type hierarchy. 92 if (handler_klass->is_subtype_of(klass)) { 93 return true; 94 } 95 } 96 } 97 98 return false; 99 } 100 101 102 bool XHandlers::equals(XHandlers* others) const { 103 if (others == NULL) return false; 104 if (length() != others->length()) return false; 105 106 for (int i = 0; i < length(); i++) { 107 if (!handler_at(i)->equals(others->handler_at(i))) return false; 108 } 109 return true; 110 } 111 112 bool XHandler::equals(XHandler* other) const { 113 assert(entry_pco() != -1 && other->entry_pco() != -1, "must have entry_pco"); 114 115 if (entry_pco() != other->entry_pco()) return false; 116 if (scope_count() != other->scope_count()) return false; 117 if (_desc != other->_desc) return false; 118 119 assert(entry_block() == other->entry_block(), "entry_block must be equal when entry_pco is equal"); 120 return true; 121 } 122 123 124 // Implementation of IRScope 125 BlockBegin* IRScope::build_graph(Compilation* compilation, int osr_bci) { 126 GraphBuilder gm(compilation, this); 127 NOT_PRODUCT(if (PrintValueNumbering && Verbose) gm.print_stats()); 128 if (compilation->bailed_out()) return NULL; 129 return gm.start(); 130 } 131 132 133 IRScope::IRScope(Compilation* compilation, IRScope* caller, int caller_bci, ciMethod* method, int osr_bci, bool create_graph) 134 : _callees(2) 135 , _compilation(compilation) 136 , _requires_phi_function(method->max_locals()) 137 { 138 _caller = caller; 139 _level = caller == NULL ? 0 : caller->level() + 1; 140 _method = method; 141 _xhandlers = new XHandlers(method); 142 _number_of_locks = 0; 143 _monitor_pairing_ok = method->has_balanced_monitors(); 144 _wrote_final = false; 145 _start = NULL; 146 147 if (osr_bci == -1) { 148 _requires_phi_function.clear(); 149 } else { 150 // selective creation of phi functions is not possibel in osr-methods 151 _requires_phi_function.set_range(0, method->max_locals()); 152 } 153 154 assert(method->holder()->is_loaded() , "method holder must be loaded"); 155 156 // build graph if monitor pairing is ok 157 if (create_graph && monitor_pairing_ok()) _start = build_graph(compilation, osr_bci); 158 } 159 160 161 int IRScope::max_stack() const { 162 int my_max = method()->max_stack(); 163 int callee_max = 0; 164 for (int i = 0; i < number_of_callees(); i++) { 165 callee_max = MAX2(callee_max, callee_no(i)->max_stack()); 166 } 167 return my_max + callee_max; 168 } 169 170 171 bool IRScopeDebugInfo::should_reexecute() { 172 ciMethod* cur_method = scope()->method(); 173 int cur_bci = bci(); 174 if (cur_method != NULL && cur_bci != SynchronizationEntryBCI) { 175 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 176 return Interpreter::bytecode_should_reexecute(code); 177 } else 178 return false; 179 } 180 181 182 // Implementation of CodeEmitInfo 183 184 // Stack must be NON-null 185 CodeEmitInfo::CodeEmitInfo(ValueStack* stack, XHandlers* exception_handlers, bool deoptimize_on_exception) 186 : _scope(stack->scope()) 187 , _scope_debug_info(NULL) 188 , _oop_map(NULL) 189 , _stack(stack) 190 , _exception_handlers(exception_handlers) 191 , _is_method_handle_invoke(false) 192 , _deoptimize_on_exception(deoptimize_on_exception) { 193 assert(_stack != NULL, "must be non null"); 194 } 195 196 197 CodeEmitInfo::CodeEmitInfo(CodeEmitInfo* info, ValueStack* stack) 198 : _scope(info->_scope) 199 , _exception_handlers(NULL) 200 , _scope_debug_info(NULL) 201 , _oop_map(NULL) 202 , _stack(stack == NULL ? info->_stack : stack) 203 , _is_method_handle_invoke(info->_is_method_handle_invoke) 204 , _deoptimize_on_exception(info->_deoptimize_on_exception) { 205 206 // deep copy of exception handlers 207 if (info->_exception_handlers != NULL) { 208 _exception_handlers = new XHandlers(info->_exception_handlers); 209 } 210 } 211 212 213 void CodeEmitInfo::record_debug_info(DebugInformationRecorder* recorder, int pc_offset) { 214 // record the safepoint before recording the debug info for enclosing scopes 215 recorder->add_safepoint(pc_offset, _oop_map->deep_copy()); 216 _scope_debug_info->record_debug_info(recorder, pc_offset, true/*topmost*/, _is_method_handle_invoke); 217 recorder->end_safepoint(pc_offset); 218 } 219 220 221 void CodeEmitInfo::record_non_safepoint_debug_info(DebugInformationRecorder* recorder, int pc_offset) { 222 if (recorder->recording_non_safepoints()) { 223 recorder->add_non_safepoint(pc_offset); 224 record_scope_non_safepoint_debug_info(recorder, pc_offset, scope(), stack()); 225 recorder->end_non_safepoint(pc_offset); 226 } 227 } 228 229 230 void CodeEmitInfo::record_scope_non_safepoint_debug_info(DebugInformationRecorder* recorder, int pc_offset, IRScope *scope, ValueStack *state) { 231 if (scope->caller() != NULL) { 232 record_scope_non_safepoint_debug_info(recorder, pc_offset, scope->caller(), state->caller_state()); 233 } 234 recorder->describe_scope(pc_offset, scope->method(), state->bci(), false); 235 } 236 237 238 void CodeEmitInfo::add_register_oop(LIR_Opr opr) { 239 assert(_oop_map != NULL, "oop map must already exist"); 240 assert(opr->is_single_cpu(), "should not call otherwise"); 241 242 VMReg name = frame_map()->regname(opr); 243 _oop_map->set_oop(name); 244 } 245 246 // Mirror the stack size calculation in the deopt code 247 // How much stack space would we need at this point in the program in 248 // case of deoptimization? 249 int CodeEmitInfo::interpreter_frame_size() const { 250 ValueStack* state = _stack; 251 int size = 0; 252 int callee_parameters = 0; 253 int callee_locals = 0; 254 int extra_args = state->scope()->method()->max_stack() - state->stack_size(); 255 256 while (state != NULL) { 257 int locks = state->locks_size(); 258 int temps = state->stack_size(); 259 bool is_top_frame = (state == _stack); 260 ciMethod* method = state->scope()->method(); 261 262 int frame_size = BytesPerWord * Interpreter::size_activation(method->max_stack(), 263 temps + callee_parameters, 264 extra_args, 265 locks, 266 callee_parameters, 267 callee_locals, 268 is_top_frame); 269 size += frame_size; 270 271 callee_parameters = method->size_of_parameters(); 272 callee_locals = method->max_locals(); 273 extra_args = 0; 274 state = state->caller_state(); 275 } 276 return size + Deoptimization::last_frame_adjust(0, callee_locals) * BytesPerWord; 277 } 278 279 // Implementation of IR 280 281 IR::IR(Compilation* compilation, ciMethod* method, int osr_bci) : 282 _locals_size(in_WordSize(-1)) 283 , _num_loops(0) { 284 // setup IR fields 285 _compilation = compilation; 286 _top_scope = new IRScope(compilation, NULL, -1, method, osr_bci, true); 287 _code = NULL; 288 } 289 290 291 void IR::optimize_blocks() { 292 Optimizer opt(this); 293 if (!compilation()->profile_branches()) { 294 if (DoCEE) { 295 opt.eliminate_conditional_expressions(); 296 #ifndef PRODUCT 297 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after CEE"); print(true); } 298 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after CEE"); print(false); } 299 #endif 300 } 301 if (EliminateBlocks) { 302 opt.eliminate_blocks(); 303 #ifndef PRODUCT 304 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after block elimination"); print(true); } 305 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after block elimination"); print(false); } 306 #endif 307 } 308 } 309 } 310 311 void IR::eliminate_null_checks() { 312 Optimizer opt(this); 313 if (EliminateNullChecks) { 314 opt.eliminate_null_checks(); 315 #ifndef PRODUCT 316 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after null check elimination"); print(true); } 317 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after null check elimination"); print(false); } 318 #endif 319 } 320 } 321 322 323 static int sort_pairs(BlockPair** a, BlockPair** b) { 324 if ((*a)->from() == (*b)->from()) { 325 return (*a)->to()->block_id() - (*b)->to()->block_id(); 326 } else { 327 return (*a)->from()->block_id() - (*b)->from()->block_id(); 328 } 329 } 330 331 332 class CriticalEdgeFinder: public BlockClosure { 333 BlockPairList blocks; 334 IR* _ir; 335 336 public: 337 CriticalEdgeFinder(IR* ir): _ir(ir) {} 338 void block_do(BlockBegin* bb) { 339 BlockEnd* be = bb->end(); 340 int nos = be->number_of_sux(); 341 if (nos >= 2) { 342 for (int i = 0; i < nos; i++) { 343 BlockBegin* sux = be->sux_at(i); 344 if (sux->number_of_preds() >= 2) { 345 blocks.append(new BlockPair(bb, sux)); 346 } 347 } 348 } 349 } 350 351 void split_edges() { 352 BlockPair* last_pair = NULL; 353 blocks.sort(sort_pairs); 354 for (int i = 0; i < blocks.length(); i++) { 355 BlockPair* pair = blocks.at(i); 356 if (last_pair != NULL && pair->is_same(last_pair)) continue; 357 BlockBegin* from = pair->from(); 358 BlockBegin* to = pair->to(); 359 BlockBegin* split = from->insert_block_between(to); 360 #ifndef PRODUCT 361 if ((PrintIR || PrintIR1) && Verbose) { 362 tty->print_cr("Split critical edge B%d -> B%d (new block B%d)", 363 from->block_id(), to->block_id(), split->block_id()); 364 } 365 #endif 366 last_pair = pair; 367 } 368 } 369 }; 370 371 void IR::split_critical_edges() { 372 CriticalEdgeFinder cef(this); 373 374 iterate_preorder(&cef); 375 cef.split_edges(); 376 } 377 378 379 class UseCountComputer: public ValueVisitor, BlockClosure { 380 private: 381 void visit(Value* n) { 382 // Local instructions and Phis for expression stack values at the 383 // start of basic blocks are not added to the instruction list 384 if (!(*n)->is_linked() && (*n)->can_be_linked()) { 385 assert(false, "a node was not appended to the graph"); 386 Compilation::current()->bailout("a node was not appended to the graph"); 387 } 388 // use n's input if not visited before 389 if (!(*n)->is_pinned() && !(*n)->has_uses()) { 390 // note: a) if the instruction is pinned, it will be handled by compute_use_count 391 // b) if the instruction has uses, it was touched before 392 // => in both cases we don't need to update n's values 393 uses_do(n); 394 } 395 // use n 396 (*n)->_use_count++; 397 } 398 399 Values* worklist; 400 int depth; 401 enum { 402 max_recurse_depth = 20 403 }; 404 405 void uses_do(Value* n) { 406 depth++; 407 if (depth > max_recurse_depth) { 408 // don't allow the traversal to recurse too deeply 409 worklist->push(*n); 410 } else { 411 (*n)->input_values_do(this); 412 // special handling for some instructions 413 if ((*n)->as_BlockEnd() != NULL) { 414 // note on BlockEnd: 415 // must 'use' the stack only if the method doesn't 416 // terminate, however, in those cases stack is empty 417 (*n)->state_values_do(this); 418 } 419 } 420 depth--; 421 } 422 423 void block_do(BlockBegin* b) { 424 depth = 0; 425 // process all pinned nodes as the roots of expression trees 426 for (Instruction* n = b; n != NULL; n = n->next()) { 427 if (n->is_pinned()) uses_do(&n); 428 } 429 assert(depth == 0, "should have counted back down"); 430 431 // now process any unpinned nodes which recursed too deeply 432 while (worklist->length() > 0) { 433 Value t = worklist->pop(); 434 if (!t->is_pinned()) { 435 // compute the use count 436 uses_do(&t); 437 438 // pin the instruction so that LIRGenerator doesn't recurse 439 // too deeply during it's evaluation. 440 t->pin(); 441 } 442 } 443 assert(depth == 0, "should have counted back down"); 444 } 445 446 UseCountComputer() { 447 worklist = new Values(); 448 depth = 0; 449 } 450 451 public: 452 static void compute(BlockList* blocks) { 453 UseCountComputer ucc; 454 blocks->iterate_backward(&ucc); 455 } 456 }; 457 458 459 // helper macro for short definition of trace-output inside code 460 #ifndef PRODUCT 461 #define TRACE_LINEAR_SCAN(level, code) \ 462 if (TraceLinearScanLevel >= level) { \ 463 code; \ 464 } 465 #else 466 #define TRACE_LINEAR_SCAN(level, code) 467 #endif 468 469 class ComputeLinearScanOrder : public StackObj { 470 private: 471 int _max_block_id; // the highest block_id of a block 472 int _num_blocks; // total number of blocks (smaller than _max_block_id) 473 int _num_loops; // total number of loops 474 bool _iterative_dominators;// method requires iterative computation of dominatiors 475 476 BlockList* _linear_scan_order; // the resulting list of blocks in correct order 477 478 BitMap _visited_blocks; // used for recursive processing of blocks 479 BitMap _active_blocks; // used for recursive processing of blocks 480 BitMap _dominator_blocks; // temproary BitMap used for computation of dominator 481 intArray _forward_branches; // number of incoming forward branches for each block 482 BlockList _loop_end_blocks; // list of all loop end blocks collected during count_edges 483 BitMap2D _loop_map; // two-dimensional bit set: a bit is set if a block is contained in a loop 484 BlockList _work_list; // temporary list (used in mark_loops and compute_order) 485 BlockList _loop_headers; 486 487 Compilation* _compilation; 488 489 // accessors for _visited_blocks and _active_blocks 490 void init_visited() { _active_blocks.clear(); _visited_blocks.clear(); } 491 bool is_visited(BlockBegin* b) const { return _visited_blocks.at(b->block_id()); } 492 bool is_active(BlockBegin* b) const { return _active_blocks.at(b->block_id()); } 493 void set_visited(BlockBegin* b) { assert(!is_visited(b), "already set"); _visited_blocks.set_bit(b->block_id()); } 494 void set_active(BlockBegin* b) { assert(!is_active(b), "already set"); _active_blocks.set_bit(b->block_id()); } 495 void clear_active(BlockBegin* b) { assert(is_active(b), "not already"); _active_blocks.clear_bit(b->block_id()); } 496 497 // accessors for _forward_branches 498 void inc_forward_branches(BlockBegin* b) { _forward_branches.at_put(b->block_id(), _forward_branches.at(b->block_id()) + 1); } 499 int dec_forward_branches(BlockBegin* b) { _forward_branches.at_put(b->block_id(), _forward_branches.at(b->block_id()) - 1); return _forward_branches.at(b->block_id()); } 500 501 // accessors for _loop_map 502 bool is_block_in_loop (int loop_idx, BlockBegin* b) const { return _loop_map.at(loop_idx, b->block_id()); } 503 void set_block_in_loop (int loop_idx, BlockBegin* b) { _loop_map.set_bit(loop_idx, b->block_id()); } 504 void clear_block_in_loop(int loop_idx, int block_id) { _loop_map.clear_bit(loop_idx, block_id); } 505 506 // count edges between blocks 507 void count_edges(BlockBegin* cur, BlockBegin* parent); 508 509 // loop detection 510 void mark_loops(); 511 void clear_non_natural_loops(BlockBegin* start_block); 512 void assign_loop_depth(BlockBegin* start_block); 513 514 // computation of final block order 515 BlockBegin* common_dominator(BlockBegin* a, BlockBegin* b); 516 void compute_dominator(BlockBegin* cur, BlockBegin* parent); 517 int compute_weight(BlockBegin* cur); 518 bool ready_for_processing(BlockBegin* cur); 519 void sort_into_work_list(BlockBegin* b); 520 void append_block(BlockBegin* cur); 521 void compute_order(BlockBegin* start_block); 522 523 // fixup of dominators for non-natural loops 524 bool compute_dominators_iter(); 525 void compute_dominators(); 526 527 // debug functions 528 NOT_PRODUCT(void print_blocks();) 529 DEBUG_ONLY(void verify();) 530 531 Compilation* compilation() const { return _compilation; } 532 public: 533 ComputeLinearScanOrder(Compilation* c, BlockBegin* start_block); 534 535 // accessors for final result 536 BlockList* linear_scan_order() const { return _linear_scan_order; } 537 int num_loops() const { return _num_loops; } 538 }; 539 540 541 ComputeLinearScanOrder::ComputeLinearScanOrder(Compilation* c, BlockBegin* start_block) : 542 _max_block_id(BlockBegin::number_of_blocks()), 543 _num_blocks(0), 544 _num_loops(0), 545 _iterative_dominators(false), 546 _visited_blocks(_max_block_id), 547 _active_blocks(_max_block_id), 548 _dominator_blocks(_max_block_id), 549 _forward_branches(_max_block_id, 0), 550 _loop_end_blocks(8), 551 _work_list(8), 552 _linear_scan_order(NULL), // initialized later with correct size 553 _loop_map(0, 0), // initialized later with correct size 554 _compilation(c) 555 { 556 TRACE_LINEAR_SCAN(2, tty->print_cr("***** computing linear-scan block order")); 557 558 init_visited(); 559 count_edges(start_block, NULL); 560 561 if (compilation()->is_profiling()) { 562 ciMethod *method = compilation()->method(); 563 if (!method->is_accessor()) { 564 ciMethodData* md = method->method_data_or_null(); 565 assert(md != NULL, "Sanity"); 566 md->set_compilation_stats(_num_loops, _num_blocks); 567 } 568 } 569 570 if (_num_loops > 0) { 571 mark_loops(); 572 clear_non_natural_loops(start_block); 573 assign_loop_depth(start_block); 574 } 575 576 compute_order(start_block); 577 compute_dominators(); 578 579 NOT_PRODUCT(print_blocks()); 580 DEBUG_ONLY(verify()); 581 } 582 583 584 // Traverse the CFG: 585 // * count total number of blocks 586 // * count all incoming edges and backward incoming edges 587 // * number loop header blocks 588 // * create a list with all loop end blocks 589 void ComputeLinearScanOrder::count_edges(BlockBegin* cur, BlockBegin* parent) { 590 TRACE_LINEAR_SCAN(3, tty->print_cr("Enter count_edges for block B%d coming from B%d", cur->block_id(), parent != NULL ? parent->block_id() : -1)); 591 assert(cur->dominator() == NULL, "dominator already initialized"); 592 593 if (is_active(cur)) { 594 TRACE_LINEAR_SCAN(3, tty->print_cr("backward branch")); 595 assert(is_visited(cur), "block must be visisted when block is active"); 596 assert(parent != NULL, "must have parent"); 597 598 cur->set(BlockBegin::linear_scan_loop_header_flag); 599 cur->set(BlockBegin::backward_branch_target_flag); 600 601 parent->set(BlockBegin::linear_scan_loop_end_flag); 602 603 // When a loop header is also the start of an exception handler, then the backward branch is 604 // an exception edge. Because such edges are usually critical edges which cannot be split, the 605 // loop must be excluded here from processing. 606 if (cur->is_set(BlockBegin::exception_entry_flag)) { 607 // Make sure that dominators are correct in this weird situation 608 _iterative_dominators = true; 609 return; 610 } 611 assert(parent->number_of_sux() == 1 && parent->sux_at(0) == cur, 612 "loop end blocks must have one successor (critical edges are split)"); 613 614 _loop_end_blocks.append(parent); 615 return; 616 } 617 618 // increment number of incoming forward branches 619 inc_forward_branches(cur); 620 621 if (is_visited(cur)) { 622 TRACE_LINEAR_SCAN(3, tty->print_cr("block already visited")); 623 return; 624 } 625 626 _num_blocks++; 627 set_visited(cur); 628 set_active(cur); 629 630 // recursive call for all successors 631 int i; 632 for (i = cur->number_of_sux() - 1; i >= 0; i--) { 633 count_edges(cur->sux_at(i), cur); 634 } 635 for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) { 636 count_edges(cur->exception_handler_at(i), cur); 637 } 638 639 clear_active(cur); 640 641 // Each loop has a unique number. 642 // When multiple loops are nested, assign_loop_depth assumes that the 643 // innermost loop has the lowest number. This is guaranteed by setting 644 // the loop number after the recursive calls for the successors above 645 // have returned. 646 if (cur->is_set(BlockBegin::linear_scan_loop_header_flag)) { 647 assert(cur->loop_index() == -1, "cannot set loop-index twice"); 648 TRACE_LINEAR_SCAN(3, tty->print_cr("Block B%d is loop header of loop %d", cur->block_id(), _num_loops)); 649 650 cur->set_loop_index(_num_loops); 651 _loop_headers.append(cur); 652 _num_loops++; 653 } 654 655 TRACE_LINEAR_SCAN(3, tty->print_cr("Finished count_edges for block B%d", cur->block_id())); 656 } 657 658 659 void ComputeLinearScanOrder::mark_loops() { 660 TRACE_LINEAR_SCAN(3, tty->print_cr("----- marking loops")); 661 662 _loop_map = BitMap2D(_num_loops, _max_block_id); 663 _loop_map.clear(); 664 665 for (int i = _loop_end_blocks.length() - 1; i >= 0; i--) { 666 BlockBegin* loop_end = _loop_end_blocks.at(i); 667 BlockBegin* loop_start = loop_end->sux_at(0); 668 int loop_idx = loop_start->loop_index(); 669 670 TRACE_LINEAR_SCAN(3, tty->print_cr("Processing loop from B%d to B%d (loop %d):", loop_start->block_id(), loop_end->block_id(), loop_idx)); 671 assert(loop_end->is_set(BlockBegin::linear_scan_loop_end_flag), "loop end flag must be set"); 672 assert(loop_end->number_of_sux() == 1, "incorrect number of successors"); 673 assert(loop_start->is_set(BlockBegin::linear_scan_loop_header_flag), "loop header flag must be set"); 674 assert(loop_idx >= 0 && loop_idx < _num_loops, "loop index not set"); 675 assert(_work_list.is_empty(), "work list must be empty before processing"); 676 677 // add the end-block of the loop to the working list 678 _work_list.push(loop_end); 679 set_block_in_loop(loop_idx, loop_end); 680 do { 681 BlockBegin* cur = _work_list.pop(); 682 683 TRACE_LINEAR_SCAN(3, tty->print_cr(" processing B%d", cur->block_id())); 684 assert(is_block_in_loop(loop_idx, cur), "bit in loop map must be set when block is in work list"); 685 686 // recursive processing of all predecessors ends when start block of loop is reached 687 if (cur != loop_start && !cur->is_set(BlockBegin::osr_entry_flag)) { 688 for (int j = cur->number_of_preds() - 1; j >= 0; j--) { 689 BlockBegin* pred = cur->pred_at(j); 690 691 if (!is_block_in_loop(loop_idx, pred) /*&& !pred->is_set(BlockBeginosr_entry_flag)*/) { 692 // this predecessor has not been processed yet, so add it to work list 693 TRACE_LINEAR_SCAN(3, tty->print_cr(" pushing B%d", pred->block_id())); 694 _work_list.push(pred); 695 set_block_in_loop(loop_idx, pred); 696 } 697 } 698 } 699 } while (!_work_list.is_empty()); 700 } 701 } 702 703 704 // check for non-natural loops (loops where the loop header does not dominate 705 // all other loop blocks = loops with mulitple entries). 706 // such loops are ignored 707 void ComputeLinearScanOrder::clear_non_natural_loops(BlockBegin* start_block) { 708 for (int i = _num_loops - 1; i >= 0; i--) { 709 if (is_block_in_loop(i, start_block)) { 710 // loop i contains the entry block of the method 711 // -> this is not a natural loop, so ignore it 712 TRACE_LINEAR_SCAN(2, tty->print_cr("Loop %d is non-natural, so it is ignored", i)); 713 714 BlockBegin *loop_header = _loop_headers.at(i); 715 assert(loop_header->is_set(BlockBegin::linear_scan_loop_header_flag), "Must be loop header"); 716 717 for (int j = 0; j < loop_header->number_of_preds(); j++) { 718 BlockBegin *pred = loop_header->pred_at(j); 719 pred->clear(BlockBegin::linear_scan_loop_end_flag); 720 } 721 722 loop_header->clear(BlockBegin::linear_scan_loop_header_flag); 723 724 for (int block_id = _max_block_id - 1; block_id >= 0; block_id--) { 725 clear_block_in_loop(i, block_id); 726 } 727 _iterative_dominators = true; 728 } 729 } 730 } 731 732 void ComputeLinearScanOrder::assign_loop_depth(BlockBegin* start_block) { 733 TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing loop-depth and weight")); 734 init_visited(); 735 736 assert(_work_list.is_empty(), "work list must be empty before processing"); 737 _work_list.append(start_block); 738 739 do { 740 BlockBegin* cur = _work_list.pop(); 741 742 if (!is_visited(cur)) { 743 set_visited(cur); 744 TRACE_LINEAR_SCAN(4, tty->print_cr("Computing loop depth for block B%d", cur->block_id())); 745 746 // compute loop-depth and loop-index for the block 747 assert(cur->loop_depth() == 0, "cannot set loop-depth twice"); 748 int i; 749 int loop_depth = 0; 750 int min_loop_idx = -1; 751 for (i = _num_loops - 1; i >= 0; i--) { 752 if (is_block_in_loop(i, cur)) { 753 loop_depth++; 754 min_loop_idx = i; 755 } 756 } 757 cur->set_loop_depth(loop_depth); 758 cur->set_loop_index(min_loop_idx); 759 760 // append all unvisited successors to work list 761 for (i = cur->number_of_sux() - 1; i >= 0; i--) { 762 _work_list.append(cur->sux_at(i)); 763 } 764 for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) { 765 _work_list.append(cur->exception_handler_at(i)); 766 } 767 } 768 } while (!_work_list.is_empty()); 769 } 770 771 772 BlockBegin* ComputeLinearScanOrder::common_dominator(BlockBegin* a, BlockBegin* b) { 773 assert(a != NULL && b != NULL, "must have input blocks"); 774 775 _dominator_blocks.clear(); 776 while (a != NULL) { 777 _dominator_blocks.set_bit(a->block_id()); 778 assert(a->dominator() != NULL || a == _linear_scan_order->at(0), "dominator must be initialized"); 779 a = a->dominator(); 780 } 781 while (b != NULL && !_dominator_blocks.at(b->block_id())) { 782 assert(b->dominator() != NULL || b == _linear_scan_order->at(0), "dominator must be initialized"); 783 b = b->dominator(); 784 } 785 786 assert(b != NULL, "could not find dominator"); 787 return b; 788 } 789 790 void ComputeLinearScanOrder::compute_dominator(BlockBegin* cur, BlockBegin* parent) { 791 if (cur->dominator() == NULL) { 792 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: initializing dominator of B%d to B%d", cur->block_id(), parent->block_id())); 793 cur->set_dominator(parent); 794 795 } else if (!(cur->is_set(BlockBegin::linear_scan_loop_header_flag) && parent->is_set(BlockBegin::linear_scan_loop_end_flag))) { 796 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: computing dominator of B%d: common dominator of B%d and B%d is B%d", cur->block_id(), parent->block_id(), cur->dominator()->block_id(), common_dominator(cur->dominator(), parent)->block_id())); 797 // Does not hold for exception blocks 798 assert(cur->number_of_preds() > 1 || cur->is_set(BlockBegin::exception_entry_flag), ""); 799 cur->set_dominator(common_dominator(cur->dominator(), parent)); 800 } 801 802 // Additional edge to xhandler of all our successors 803 // range check elimination needs that the state at the end of a 804 // block be valid in every block it dominates so cur must dominate 805 // the exception handlers of its successors. 806 int num_cur_xhandler = cur->number_of_exception_handlers(); 807 for (int j = 0; j < num_cur_xhandler; j++) { 808 BlockBegin* xhandler = cur->exception_handler_at(j); 809 compute_dominator(xhandler, parent); 810 } 811 } 812 813 814 int ComputeLinearScanOrder::compute_weight(BlockBegin* cur) { 815 BlockBegin* single_sux = NULL; 816 if (cur->number_of_sux() == 1) { 817 single_sux = cur->sux_at(0); 818 } 819 820 // limit loop-depth to 15 bit (only for security reason, it will never be so big) 821 int weight = (cur->loop_depth() & 0x7FFF) << 16; 822 823 // general macro for short definition of weight flags 824 // the first instance of INC_WEIGHT_IF has the highest priority 825 int cur_bit = 15; 826 #define INC_WEIGHT_IF(condition) if ((condition)) { weight |= (1 << cur_bit); } cur_bit--; 827 828 // this is necessery for the (very rare) case that two successing blocks have 829 // the same loop depth, but a different loop index (can happen for endless loops 830 // with exception handlers) 831 INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_header_flag)); 832 833 // loop end blocks (blocks that end with a backward branch) are added 834 // after all other blocks of the loop. 835 INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_end_flag)); 836 837 // critical edge split blocks are prefered because than they have a bigger 838 // proability to be completely empty 839 INC_WEIGHT_IF(cur->is_set(BlockBegin::critical_edge_split_flag)); 840 841 // exceptions should not be thrown in normal control flow, so these blocks 842 // are added as late as possible 843 INC_WEIGHT_IF(cur->end()->as_Throw() == NULL && (single_sux == NULL || single_sux->end()->as_Throw() == NULL)); 844 INC_WEIGHT_IF(cur->end()->as_Return() == NULL && (single_sux == NULL || single_sux->end()->as_Return() == NULL)); 845 846 // exceptions handlers are added as late as possible 847 INC_WEIGHT_IF(!cur->is_set(BlockBegin::exception_entry_flag)); 848 849 // guarantee that weight is > 0 850 weight |= 1; 851 852 #undef INC_WEIGHT_IF 853 assert(cur_bit >= 0, "too many flags"); 854 assert(weight > 0, "weight cannot become negative"); 855 856 return weight; 857 } 858 859 bool ComputeLinearScanOrder::ready_for_processing(BlockBegin* cur) { 860 // Discount the edge just traveled. 861 // When the number drops to zero, all forward branches were processed 862 if (dec_forward_branches(cur) != 0) { 863 return false; 864 } 865 866 assert(_linear_scan_order->index_of(cur) == -1, "block already processed (block can be ready only once)"); 867 assert(_work_list.index_of(cur) == -1, "block already in work-list (block can be ready only once)"); 868 return true; 869 } 870 871 void ComputeLinearScanOrder::sort_into_work_list(BlockBegin* cur) { 872 assert(_work_list.index_of(cur) == -1, "block already in work list"); 873 874 int cur_weight = compute_weight(cur); 875 876 // the linear_scan_number is used to cache the weight of a block 877 cur->set_linear_scan_number(cur_weight); 878 879 #ifndef PRODUCT 880 if (StressLinearScan) { 881 _work_list.insert_before(0, cur); 882 return; 883 } 884 #endif 885 886 _work_list.append(NULL); // provide space for new element 887 888 int insert_idx = _work_list.length() - 1; 889 while (insert_idx > 0 && _work_list.at(insert_idx - 1)->linear_scan_number() > cur_weight) { 890 _work_list.at_put(insert_idx, _work_list.at(insert_idx - 1)); 891 insert_idx--; 892 } 893 _work_list.at_put(insert_idx, cur); 894 895 TRACE_LINEAR_SCAN(3, tty->print_cr("Sorted B%d into worklist. new worklist:", cur->block_id())); 896 TRACE_LINEAR_SCAN(3, for (int i = 0; i < _work_list.length(); i++) tty->print_cr("%8d B%2d weight:%6x", i, _work_list.at(i)->block_id(), _work_list.at(i)->linear_scan_number())); 897 898 #ifdef ASSERT 899 for (int i = 0; i < _work_list.length(); i++) { 900 assert(_work_list.at(i)->linear_scan_number() > 0, "weight not set"); 901 assert(i == 0 || _work_list.at(i - 1)->linear_scan_number() <= _work_list.at(i)->linear_scan_number(), "incorrect order in worklist"); 902 } 903 #endif 904 } 905 906 void ComputeLinearScanOrder::append_block(BlockBegin* cur) { 907 TRACE_LINEAR_SCAN(3, tty->print_cr("appending block B%d (weight 0x%6x) to linear-scan order", cur->block_id(), cur->linear_scan_number())); 908 assert(_linear_scan_order->index_of(cur) == -1, "cannot add the same block twice"); 909 910 // currently, the linear scan order and code emit order are equal. 911 // therefore the linear_scan_number and the weight of a block must also 912 // be equal. 913 cur->set_linear_scan_number(_linear_scan_order->length()); 914 _linear_scan_order->append(cur); 915 } 916 917 void ComputeLinearScanOrder::compute_order(BlockBegin* start_block) { 918 TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing final block order")); 919 920 // the start block is always the first block in the linear scan order 921 _linear_scan_order = new BlockList(_num_blocks); 922 append_block(start_block); 923 924 assert(start_block->end()->as_Base() != NULL, "start block must end with Base-instruction"); 925 BlockBegin* std_entry = ((Base*)start_block->end())->std_entry(); 926 BlockBegin* osr_entry = ((Base*)start_block->end())->osr_entry(); 927 928 BlockBegin* sux_of_osr_entry = NULL; 929 if (osr_entry != NULL) { 930 // special handling for osr entry: 931 // ignore the edge between the osr entry and its successor for processing 932 // the osr entry block is added manually below 933 assert(osr_entry->number_of_sux() == 1, "osr entry must have exactly one successor"); 934 assert(osr_entry->sux_at(0)->number_of_preds() >= 2, "sucessor of osr entry must have two predecessors (otherwise it is not present in normal control flow"); 935 936 sux_of_osr_entry = osr_entry->sux_at(0); 937 dec_forward_branches(sux_of_osr_entry); 938 939 compute_dominator(osr_entry, start_block); 940 _iterative_dominators = true; 941 } 942 compute_dominator(std_entry, start_block); 943 944 // start processing with standard entry block 945 assert(_work_list.is_empty(), "list must be empty before processing"); 946 947 if (ready_for_processing(std_entry)) { 948 sort_into_work_list(std_entry); 949 } else { 950 assert(false, "the std_entry must be ready for processing (otherwise, the method has no start block)"); 951 } 952 953 do { 954 BlockBegin* cur = _work_list.pop(); 955 956 if (cur == sux_of_osr_entry) { 957 // the osr entry block is ignored in normal processing, it is never added to the 958 // work list. Instead, it is added as late as possible manually here. 959 append_block(osr_entry); 960 compute_dominator(cur, osr_entry); 961 } 962 append_block(cur); 963 964 int i; 965 int num_sux = cur->number_of_sux(); 966 // changed loop order to get "intuitive" order of if- and else-blocks 967 for (i = 0; i < num_sux; i++) { 968 BlockBegin* sux = cur->sux_at(i); 969 compute_dominator(sux, cur); 970 if (ready_for_processing(sux)) { 971 sort_into_work_list(sux); 972 } 973 } 974 num_sux = cur->number_of_exception_handlers(); 975 for (i = 0; i < num_sux; i++) { 976 BlockBegin* sux = cur->exception_handler_at(i); 977 if (ready_for_processing(sux)) { 978 sort_into_work_list(sux); 979 } 980 } 981 } while (_work_list.length() > 0); 982 } 983 984 985 bool ComputeLinearScanOrder::compute_dominators_iter() { 986 bool changed = false; 987 int num_blocks = _linear_scan_order->length(); 988 989 assert(_linear_scan_order->at(0)->dominator() == NULL, "must not have dominator"); 990 assert(_linear_scan_order->at(0)->number_of_preds() == 0, "must not have predecessors"); 991 for (int i = 1; i < num_blocks; i++) { 992 BlockBegin* block = _linear_scan_order->at(i); 993 994 BlockBegin* dominator = block->pred_at(0); 995 int num_preds = block->number_of_preds(); 996 997 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: Processing B%d", block->block_id())); 998 999 for (int j = 0; j < num_preds; j++) { 1000 1001 BlockBegin *pred = block->pred_at(j); 1002 TRACE_LINEAR_SCAN(4, tty->print_cr(" DOM: Subrocessing B%d", pred->block_id())); 1003 1004 if (block->is_set(BlockBegin::exception_entry_flag)) { 1005 dominator = common_dominator(dominator, pred); 1006 int num_pred_preds = pred->number_of_preds(); 1007 for (int k = 0; k < num_pred_preds; k++) { 1008 dominator = common_dominator(dominator, pred->pred_at(k)); 1009 } 1010 } else { 1011 dominator = common_dominator(dominator, pred); 1012 } 1013 } 1014 1015 if (dominator != block->dominator()) { 1016 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: updating dominator of B%d from B%d to B%d", block->block_id(), block->dominator()->block_id(), dominator->block_id())); 1017 1018 block->set_dominator(dominator); 1019 changed = true; 1020 } 1021 } 1022 return changed; 1023 } 1024 1025 void ComputeLinearScanOrder::compute_dominators() { 1026 TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing dominators (iterative computation reqired: %d)", _iterative_dominators)); 1027 1028 // iterative computation of dominators is only required for methods with non-natural loops 1029 // and OSR-methods. For all other methods, the dominators computed when generating the 1030 // linear scan block order are correct. 1031 if (_iterative_dominators) { 1032 do { 1033 TRACE_LINEAR_SCAN(1, tty->print_cr("DOM: next iteration of fix-point calculation")); 1034 } while (compute_dominators_iter()); 1035 } 1036 1037 // check that dominators are correct 1038 assert(!compute_dominators_iter(), "fix point not reached"); 1039 1040 // Add Blocks to dominates-Array 1041 int num_blocks = _linear_scan_order->length(); 1042 for (int i = 0; i < num_blocks; i++) { 1043 BlockBegin* block = _linear_scan_order->at(i); 1044 1045 BlockBegin *dom = block->dominator(); 1046 if (dom) { 1047 assert(dom->dominator_depth() != -1, "Dominator must have been visited before"); 1048 dom->dominates()->append(block); 1049 block->set_dominator_depth(dom->dominator_depth() + 1); 1050 } else { 1051 block->set_dominator_depth(0); 1052 } 1053 } 1054 } 1055 1056 1057 #ifndef PRODUCT 1058 void ComputeLinearScanOrder::print_blocks() { 1059 if (TraceLinearScanLevel >= 2) { 1060 tty->print_cr("----- loop information:"); 1061 for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) { 1062 BlockBegin* cur = _linear_scan_order->at(block_idx); 1063 1064 tty->print("%4d: B%2d: ", cur->linear_scan_number(), cur->block_id()); 1065 for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) { 1066 tty->print ("%d ", is_block_in_loop(loop_idx, cur)); 1067 } 1068 tty->print_cr(" -> loop_index: %2d, loop_depth: %2d", cur->loop_index(), cur->loop_depth()); 1069 } 1070 } 1071 1072 if (TraceLinearScanLevel >= 1) { 1073 tty->print_cr("----- linear-scan block order:"); 1074 for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) { 1075 BlockBegin* cur = _linear_scan_order->at(block_idx); 1076 tty->print("%4d: B%2d loop: %2d depth: %2d", cur->linear_scan_number(), cur->block_id(), cur->loop_index(), cur->loop_depth()); 1077 1078 tty->print(cur->is_set(BlockBegin::exception_entry_flag) ? " ex" : " "); 1079 tty->print(cur->is_set(BlockBegin::critical_edge_split_flag) ? " ce" : " "); 1080 tty->print(cur->is_set(BlockBegin::linear_scan_loop_header_flag) ? " lh" : " "); 1081 tty->print(cur->is_set(BlockBegin::linear_scan_loop_end_flag) ? " le" : " "); 1082 1083 if (cur->dominator() != NULL) { 1084 tty->print(" dom: B%d ", cur->dominator()->block_id()); 1085 } else { 1086 tty->print(" dom: NULL "); 1087 } 1088 1089 if (cur->number_of_preds() > 0) { 1090 tty->print(" preds: "); 1091 for (int j = 0; j < cur->number_of_preds(); j++) { 1092 BlockBegin* pred = cur->pred_at(j); 1093 tty->print("B%d ", pred->block_id()); 1094 } 1095 } 1096 if (cur->number_of_sux() > 0) { 1097 tty->print(" sux: "); 1098 for (int j = 0; j < cur->number_of_sux(); j++) { 1099 BlockBegin* sux = cur->sux_at(j); 1100 tty->print("B%d ", sux->block_id()); 1101 } 1102 } 1103 if (cur->number_of_exception_handlers() > 0) { 1104 tty->print(" ex: "); 1105 for (int j = 0; j < cur->number_of_exception_handlers(); j++) { 1106 BlockBegin* ex = cur->exception_handler_at(j); 1107 tty->print("B%d ", ex->block_id()); 1108 } 1109 } 1110 tty->cr(); 1111 } 1112 } 1113 } 1114 #endif 1115 1116 #ifdef ASSERT 1117 void ComputeLinearScanOrder::verify() { 1118 assert(_linear_scan_order->length() == _num_blocks, "wrong number of blocks in list"); 1119 1120 if (StressLinearScan) { 1121 // blocks are scrambled when StressLinearScan is used 1122 return; 1123 } 1124 1125 // check that all successors of a block have a higher linear-scan-number 1126 // and that all predecessors of a block have a lower linear-scan-number 1127 // (only backward branches of loops are ignored) 1128 int i; 1129 for (i = 0; i < _linear_scan_order->length(); i++) { 1130 BlockBegin* cur = _linear_scan_order->at(i); 1131 1132 assert(cur->linear_scan_number() == i, "incorrect linear_scan_number"); 1133 assert(cur->linear_scan_number() >= 0 && cur->linear_scan_number() == _linear_scan_order->index_of(cur), "incorrect linear_scan_number"); 1134 1135 int j; 1136 for (j = cur->number_of_sux() - 1; j >= 0; j--) { 1137 BlockBegin* sux = cur->sux_at(j); 1138 1139 assert(sux->linear_scan_number() >= 0 && sux->linear_scan_number() == _linear_scan_order->index_of(sux), "incorrect linear_scan_number"); 1140 if (!sux->is_set(BlockBegin::backward_branch_target_flag)) { 1141 assert(cur->linear_scan_number() < sux->linear_scan_number(), "invalid order"); 1142 } 1143 if (cur->loop_depth() == sux->loop_depth()) { 1144 assert(cur->loop_index() == sux->loop_index() || sux->is_set(BlockBegin::linear_scan_loop_header_flag), "successing blocks with same loop depth must have same loop index"); 1145 } 1146 } 1147 1148 for (j = cur->number_of_preds() - 1; j >= 0; j--) { 1149 BlockBegin* pred = cur->pred_at(j); 1150 1151 assert(pred->linear_scan_number() >= 0 && pred->linear_scan_number() == _linear_scan_order->index_of(pred), "incorrect linear_scan_number"); 1152 if (!cur->is_set(BlockBegin::backward_branch_target_flag)) { 1153 assert(cur->linear_scan_number() > pred->linear_scan_number(), "invalid order"); 1154 } 1155 if (cur->loop_depth() == pred->loop_depth()) { 1156 assert(cur->loop_index() == pred->loop_index() || cur->is_set(BlockBegin::linear_scan_loop_header_flag), "successing blocks with same loop depth must have same loop index"); 1157 } 1158 1159 assert(cur->dominator()->linear_scan_number() <= cur->pred_at(j)->linear_scan_number(), "dominator must be before predecessors"); 1160 } 1161 1162 // check dominator 1163 if (i == 0) { 1164 assert(cur->dominator() == NULL, "first block has no dominator"); 1165 } else { 1166 assert(cur->dominator() != NULL, "all but first block must have dominator"); 1167 } 1168 // Assertion does not hold for exception handlers 1169 assert(cur->number_of_preds() != 1 || cur->dominator() == cur->pred_at(0) || cur->is_set(BlockBegin::exception_entry_flag), "Single predecessor must also be dominator"); 1170 } 1171 1172 // check that all loops are continuous 1173 for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) { 1174 int block_idx = 0; 1175 assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "the first block must not be present in any loop"); 1176 1177 // skip blocks before the loop 1178 while (block_idx < _num_blocks && !is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) { 1179 block_idx++; 1180 } 1181 // skip blocks of loop 1182 while (block_idx < _num_blocks && is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) { 1183 block_idx++; 1184 } 1185 // after the first non-loop block, there must not be another loop-block 1186 while (block_idx < _num_blocks) { 1187 assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "loop not continuous in linear-scan order"); 1188 block_idx++; 1189 } 1190 } 1191 } 1192 #endif 1193 1194 1195 void IR::compute_code() { 1196 assert(is_valid(), "IR must be valid"); 1197 1198 ComputeLinearScanOrder compute_order(compilation(), start()); 1199 _num_loops = compute_order.num_loops(); 1200 _code = compute_order.linear_scan_order(); 1201 } 1202 1203 1204 void IR::compute_use_counts() { 1205 // make sure all values coming out of this block get evaluated. 1206 int num_blocks = _code->length(); 1207 for (int i = 0; i < num_blocks; i++) { 1208 _code->at(i)->end()->state()->pin_stack_for_linear_scan(); 1209 } 1210 1211 // compute use counts 1212 UseCountComputer::compute(_code); 1213 } 1214 1215 1216 void IR::iterate_preorder(BlockClosure* closure) { 1217 assert(is_valid(), "IR must be valid"); 1218 start()->iterate_preorder(closure); 1219 } 1220 1221 1222 void IR::iterate_postorder(BlockClosure* closure) { 1223 assert(is_valid(), "IR must be valid"); 1224 start()->iterate_postorder(closure); 1225 } 1226 1227 void IR::iterate_linear_scan_order(BlockClosure* closure) { 1228 linear_scan_order()->iterate_forward(closure); 1229 } 1230 1231 1232 #ifndef PRODUCT 1233 class BlockPrinter: public BlockClosure { 1234 private: 1235 InstructionPrinter* _ip; 1236 bool _cfg_only; 1237 bool _live_only; 1238 1239 public: 1240 BlockPrinter(InstructionPrinter* ip, bool cfg_only, bool live_only = false) { 1241 _ip = ip; 1242 _cfg_only = cfg_only; 1243 _live_only = live_only; 1244 } 1245 1246 virtual void block_do(BlockBegin* block) { 1247 if (_cfg_only) { 1248 _ip->print_instr(block); tty->cr(); 1249 } else { 1250 block->print_block(*_ip, _live_only); 1251 } 1252 } 1253 }; 1254 1255 1256 void IR::print(BlockBegin* start, bool cfg_only, bool live_only) { 1257 ttyLocker ttyl; 1258 InstructionPrinter ip(!cfg_only); 1259 BlockPrinter bp(&ip, cfg_only, live_only); 1260 start->iterate_preorder(&bp); 1261 tty->cr(); 1262 } 1263 1264 void IR::print(bool cfg_only, bool live_only) { 1265 if (is_valid()) { 1266 print(start(), cfg_only, live_only); 1267 } else { 1268 tty->print_cr("invalid IR"); 1269 } 1270 } 1271 1272 1273 define_array(BlockListArray, BlockList*) 1274 define_stack(BlockListList, BlockListArray) 1275 1276 class PredecessorValidator : public BlockClosure { 1277 private: 1278 BlockListList* _predecessors; 1279 BlockList* _blocks; 1280 1281 static int cmp(BlockBegin** a, BlockBegin** b) { 1282 return (*a)->block_id() - (*b)->block_id(); 1283 } 1284 1285 public: 1286 PredecessorValidator(IR* hir) { 1287 ResourceMark rm; 1288 _predecessors = new BlockListList(BlockBegin::number_of_blocks(), NULL); 1289 _blocks = new BlockList(); 1290 1291 int i; 1292 hir->start()->iterate_preorder(this); 1293 if (hir->code() != NULL) { 1294 assert(hir->code()->length() == _blocks->length(), "must match"); 1295 for (i = 0; i < _blocks->length(); i++) { 1296 assert(hir->code()->contains(_blocks->at(i)), "should be in both lists"); 1297 } 1298 } 1299 1300 for (i = 0; i < _blocks->length(); i++) { 1301 BlockBegin* block = _blocks->at(i); 1302 BlockList* preds = _predecessors->at(block->block_id()); 1303 if (preds == NULL) { 1304 assert(block->number_of_preds() == 0, "should be the same"); 1305 continue; 1306 } 1307 1308 // clone the pred list so we can mutate it 1309 BlockList* pred_copy = new BlockList(); 1310 int j; 1311 for (j = 0; j < block->number_of_preds(); j++) { 1312 pred_copy->append(block->pred_at(j)); 1313 } 1314 // sort them in the same order 1315 preds->sort(cmp); 1316 pred_copy->sort(cmp); 1317 int length = MIN2(preds->length(), block->number_of_preds()); 1318 for (j = 0; j < block->number_of_preds(); j++) { 1319 assert(preds->at(j) == pred_copy->at(j), "must match"); 1320 } 1321 1322 assert(preds->length() == block->number_of_preds(), "should be the same"); 1323 } 1324 } 1325 1326 virtual void block_do(BlockBegin* block) { 1327 _blocks->append(block); 1328 BlockEnd* be = block->end(); 1329 int n = be->number_of_sux(); 1330 int i; 1331 for (i = 0; i < n; i++) { 1332 BlockBegin* sux = be->sux_at(i); 1333 assert(!sux->is_set(BlockBegin::exception_entry_flag), "must not be xhandler"); 1334 1335 BlockList* preds = _predecessors->at_grow(sux->block_id(), NULL); 1336 if (preds == NULL) { 1337 preds = new BlockList(); 1338 _predecessors->at_put(sux->block_id(), preds); 1339 } 1340 preds->append(block); 1341 } 1342 1343 n = block->number_of_exception_handlers(); 1344 for (i = 0; i < n; i++) { 1345 BlockBegin* sux = block->exception_handler_at(i); 1346 assert(sux->is_set(BlockBegin::exception_entry_flag), "must be xhandler"); 1347 1348 BlockList* preds = _predecessors->at_grow(sux->block_id(), NULL); 1349 if (preds == NULL) { 1350 preds = new BlockList(); 1351 _predecessors->at_put(sux->block_id(), preds); 1352 } 1353 preds->append(block); 1354 } 1355 } 1356 }; 1357 1358 class VerifyBlockBeginField : public BlockClosure { 1359 1360 public: 1361 1362 virtual void block_do(BlockBegin *block) { 1363 for ( Instruction *cur = block; cur != NULL; cur = cur->next()) { 1364 assert(cur->block() == block, "Block begin is not correct"); 1365 } 1366 } 1367 }; 1368 1369 void IR::verify() { 1370 #ifdef ASSERT 1371 PredecessorValidator pv(this); 1372 VerifyBlockBeginField verifier; 1373 this->iterate_postorder(&verifier); 1374 #endif 1375 } 1376 1377 #endif // PRODUCT 1378 1379 void SubstitutionResolver::visit(Value* v) { 1380 Value v0 = *v; 1381 if (v0) { 1382 Value vs = v0->subst(); 1383 if (vs != v0) { 1384 *v = v0->subst(); 1385 } 1386 } 1387 } 1388 1389 #ifdef ASSERT 1390 class SubstitutionChecker: public ValueVisitor { 1391 void visit(Value* v) { 1392 Value v0 = *v; 1393 if (v0) { 1394 Value vs = v0->subst(); 1395 assert(vs == v0, "missed substitution"); 1396 } 1397 } 1398 }; 1399 #endif 1400 1401 1402 void SubstitutionResolver::block_do(BlockBegin* block) { 1403 Instruction* last = NULL; 1404 for (Instruction* n = block; n != NULL;) { 1405 n->values_do(this); 1406 // need to remove this instruction from the instruction stream 1407 if (n->subst() != n) { 1408 assert(last != NULL, "must have last"); 1409 last->set_next(n->next()); 1410 } else { 1411 last = n; 1412 } 1413 n = last->next(); 1414 } 1415 1416 #ifdef ASSERT 1417 SubstitutionChecker check_substitute; 1418 if (block->state()) block->state()->values_do(&check_substitute); 1419 block->block_values_do(&check_substitute); 1420 if (block->end() && block->end()->state()) block->end()->state()->values_do(&check_substitute); 1421 #endif 1422 }