1 /* 2 * Copyright (c) 1997, 2021, 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 "jvm_io.h" 27 #include "asm/macroAssembler.hpp" 28 #include "asm/macroAssembler.inline.hpp" 29 #include "ci/ciReplay.hpp" 30 #include "classfile/javaClasses.hpp" 31 #include "code/exceptionHandlerTable.hpp" 32 #include "code/nmethod.hpp" 33 #include "compiler/compileBroker.hpp" 34 #include "compiler/compileLog.hpp" 35 #include "compiler/disassembler.hpp" 36 #include "compiler/oopMap.hpp" 37 #include "gc/shared/barrierSet.hpp" 38 #include "gc/shared/c2/barrierSetC2.hpp" 39 #include "jfr/jfrEvents.hpp" 40 #include "memory/resourceArea.hpp" 41 #include "opto/addnode.hpp" 42 #include "opto/block.hpp" 43 #include "opto/c2compiler.hpp" 44 #include "opto/callGenerator.hpp" 45 #include "opto/callnode.hpp" 46 #include "opto/castnode.hpp" 47 #include "opto/cfgnode.hpp" 48 #include "opto/chaitin.hpp" 49 #include "opto/compile.hpp" 50 #include "opto/connode.hpp" 51 #include "opto/convertnode.hpp" 52 #include "opto/divnode.hpp" 53 #include "opto/escape.hpp" 54 #include "opto/idealGraphPrinter.hpp" 55 #include "opto/loopnode.hpp" 56 #include "opto/machnode.hpp" 57 #include "opto/macro.hpp" 58 #include "opto/matcher.hpp" 59 #include "opto/mathexactnode.hpp" 60 #include "opto/memnode.hpp" 61 #include "opto/mulnode.hpp" 62 #include "opto/narrowptrnode.hpp" 63 #include "opto/node.hpp" 64 #include "opto/opcodes.hpp" 65 #include "opto/output.hpp" 66 #include "opto/parse.hpp" 67 #include "opto/phaseX.hpp" 68 #include "opto/rootnode.hpp" 69 #include "opto/runtime.hpp" 70 #include "opto/stringopts.hpp" 71 #include "opto/type.hpp" 72 #include "opto/vector.hpp" 73 #include "opto/vectornode.hpp" 74 #include "runtime/globals_extension.hpp" 75 #include "runtime/sharedRuntime.hpp" 76 #include "runtime/signature.hpp" 77 #include "runtime/stubRoutines.hpp" 78 #include "runtime/timer.hpp" 79 #include "utilities/align.hpp" 80 #include "utilities/copy.hpp" 81 #include "utilities/macros.hpp" 82 #include "utilities/resourceHash.hpp" 83 84 85 // -------------------- Compile::mach_constant_base_node ----------------------- 86 // Constant table base node singleton. 87 MachConstantBaseNode* Compile::mach_constant_base_node() { 88 if (_mach_constant_base_node == NULL) { 89 _mach_constant_base_node = new MachConstantBaseNode(); 90 _mach_constant_base_node->add_req(C->root()); 91 } 92 return _mach_constant_base_node; 93 } 94 95 96 /// Support for intrinsics. 97 98 // Return the index at which m must be inserted (or already exists). 99 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 100 class IntrinsicDescPair { 101 private: 102 ciMethod* _m; 103 bool _is_virtual; 104 public: 105 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 106 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 107 ciMethod* m= elt->method(); 108 ciMethod* key_m = key->_m; 109 if (key_m < m) return -1; 110 else if (key_m > m) return 1; 111 else { 112 bool is_virtual = elt->is_virtual(); 113 bool key_virtual = key->_is_virtual; 114 if (key_virtual < is_virtual) return -1; 115 else if (key_virtual > is_virtual) return 1; 116 else return 0; 117 } 118 } 119 }; 120 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 121 #ifdef ASSERT 122 for (int i = 1; i < _intrinsics.length(); i++) { 123 CallGenerator* cg1 = _intrinsics.at(i-1); 124 CallGenerator* cg2 = _intrinsics.at(i); 125 assert(cg1->method() != cg2->method() 126 ? cg1->method() < cg2->method() 127 : cg1->is_virtual() < cg2->is_virtual(), 128 "compiler intrinsics list must stay sorted"); 129 } 130 #endif 131 IntrinsicDescPair pair(m, is_virtual); 132 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 133 } 134 135 void Compile::register_intrinsic(CallGenerator* cg) { 136 bool found = false; 137 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 138 assert(!found, "registering twice"); 139 _intrinsics.insert_before(index, cg); 140 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 141 } 142 143 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 144 assert(m->is_loaded(), "don't try this on unloaded methods"); 145 if (_intrinsics.length() > 0) { 146 bool found = false; 147 int index = intrinsic_insertion_index(m, is_virtual, found); 148 if (found) { 149 return _intrinsics.at(index); 150 } 151 } 152 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 153 if (m->intrinsic_id() != vmIntrinsics::_none && 154 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 155 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 156 if (cg != NULL) { 157 // Save it for next time: 158 register_intrinsic(cg); 159 return cg; 160 } else { 161 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 162 } 163 } 164 return NULL; 165 } 166 167 // Compile::make_vm_intrinsic is defined in library_call.cpp. 168 169 #ifndef PRODUCT 170 // statistics gathering... 171 172 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0}; 173 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0}; 174 175 inline int as_int(vmIntrinsics::ID id) { 176 return vmIntrinsics::as_int(id); 177 } 178 179 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 180 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 181 int oflags = _intrinsic_hist_flags[as_int(id)]; 182 assert(flags != 0, "what happened?"); 183 if (is_virtual) { 184 flags |= _intrinsic_virtual; 185 } 186 bool changed = (flags != oflags); 187 if ((flags & _intrinsic_worked) != 0) { 188 juint count = (_intrinsic_hist_count[as_int(id)] += 1); 189 if (count == 1) { 190 changed = true; // first time 191 } 192 // increment the overall count also: 193 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1; 194 } 195 if (changed) { 196 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 197 // Something changed about the intrinsic's virtuality. 198 if ((flags & _intrinsic_virtual) != 0) { 199 // This is the first use of this intrinsic as a virtual call. 200 if (oflags != 0) { 201 // We already saw it as a non-virtual, so note both cases. 202 flags |= _intrinsic_both; 203 } 204 } else if ((oflags & _intrinsic_both) == 0) { 205 // This is the first use of this intrinsic as a non-virtual 206 flags |= _intrinsic_both; 207 } 208 } 209 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags); 210 } 211 // update the overall flags also: 212 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags; 213 return changed; 214 } 215 216 static char* format_flags(int flags, char* buf) { 217 buf[0] = 0; 218 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 219 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 220 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 221 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 222 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 223 if (buf[0] == 0) strcat(buf, ","); 224 assert(buf[0] == ',', "must be"); 225 return &buf[1]; 226 } 227 228 void Compile::print_intrinsic_statistics() { 229 char flagsbuf[100]; 230 ttyLocker ttyl; 231 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 232 tty->print_cr("Compiler intrinsic usage:"); 233 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)]; 234 if (total == 0) total = 1; // avoid div0 in case of no successes 235 #define PRINT_STAT_LINE(name, c, f) \ 236 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 237 for (auto id : EnumRange<vmIntrinsicID>{}) { 238 int flags = _intrinsic_hist_flags[as_int(id)]; 239 juint count = _intrinsic_hist_count[as_int(id)]; 240 if ((flags | count) != 0) { 241 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 242 } 243 } 244 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf)); 245 if (xtty != NULL) xtty->tail("statistics"); 246 } 247 248 void Compile::print_statistics() { 249 { ttyLocker ttyl; 250 if (xtty != NULL) xtty->head("statistics type='opto'"); 251 Parse::print_statistics(); 252 PhaseCCP::print_statistics(); 253 PhaseRegAlloc::print_statistics(); 254 PhaseOutput::print_statistics(); 255 PhasePeephole::print_statistics(); 256 PhaseIdealLoop::print_statistics(); 257 if (xtty != NULL) xtty->tail("statistics"); 258 } 259 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) { 260 // put this under its own <statistics> element. 261 print_intrinsic_statistics(); 262 } 263 } 264 #endif //PRODUCT 265 266 void Compile::gvn_replace_by(Node* n, Node* nn) { 267 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 268 Node* use = n->last_out(i); 269 bool is_in_table = initial_gvn()->hash_delete(use); 270 uint uses_found = 0; 271 for (uint j = 0; j < use->len(); j++) { 272 if (use->in(j) == n) { 273 if (j < use->req()) 274 use->set_req(j, nn); 275 else 276 use->set_prec(j, nn); 277 uses_found++; 278 } 279 } 280 if (is_in_table) { 281 // reinsert into table 282 initial_gvn()->hash_find_insert(use); 283 } 284 record_for_igvn(use); 285 i -= uses_found; // we deleted 1 or more copies of this edge 286 } 287 } 288 289 290 // Identify all nodes that are reachable from below, useful. 291 // Use breadth-first pass that records state in a Unique_Node_List, 292 // recursive traversal is slower. 293 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 294 int estimated_worklist_size = live_nodes(); 295 useful.map( estimated_worklist_size, NULL ); // preallocate space 296 297 // Initialize worklist 298 if (root() != NULL) { useful.push(root()); } 299 // If 'top' is cached, declare it useful to preserve cached node 300 if( cached_top_node() ) { useful.push(cached_top_node()); } 301 302 // Push all useful nodes onto the list, breadthfirst 303 for( uint next = 0; next < useful.size(); ++next ) { 304 assert( next < unique(), "Unique useful nodes < total nodes"); 305 Node *n = useful.at(next); 306 uint max = n->len(); 307 for( uint i = 0; i < max; ++i ) { 308 Node *m = n->in(i); 309 if (not_a_node(m)) continue; 310 useful.push(m); 311 } 312 } 313 } 314 315 // Update dead_node_list with any missing dead nodes using useful 316 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 317 void Compile::update_dead_node_list(Unique_Node_List &useful) { 318 uint max_idx = unique(); 319 VectorSet& useful_node_set = useful.member_set(); 320 321 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 322 // If node with index node_idx is not in useful set, 323 // mark it as dead in dead node list. 324 if (!useful_node_set.test(node_idx)) { 325 record_dead_node(node_idx); 326 } 327 } 328 } 329 330 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 331 int shift = 0; 332 for (int i = 0; i < inlines->length(); i++) { 333 CallGenerator* cg = inlines->at(i); 334 if (useful.member(cg->call_node())) { 335 if (shift > 0) { 336 inlines->at_put(i - shift, cg); 337 } 338 } else { 339 shift++; // skip over the dead element 340 } 341 } 342 if (shift > 0) { 343 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array 344 } 345 } 346 347 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) { 348 assert(dead != NULL && dead->is_Call(), "sanity"); 349 int found = 0; 350 for (int i = 0; i < inlines->length(); i++) { 351 if (inlines->at(i)->call_node() == dead) { 352 inlines->remove_at(i); 353 found++; 354 NOT_DEBUG( break; ) // elements are unique, so exit early 355 } 356 } 357 assert(found <= 1, "not unique"); 358 } 359 360 void Compile::remove_useless_nodes(GrowableArray<Node*>& node_list, Unique_Node_List& useful) { 361 for (int i = node_list.length() - 1; i >= 0; i--) { 362 Node* n = node_list.at(i); 363 if (!useful.member(n)) { 364 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 365 } 366 } 367 } 368 369 void Compile::remove_useless_node(Node* dead) { 370 remove_modified_node(dead); 371 372 // Constant node that has no out-edges and has only one in-edge from 373 // root is usually dead. However, sometimes reshaping walk makes 374 // it reachable by adding use edges. So, we will NOT count Con nodes 375 // as dead to be conservative about the dead node count at any 376 // given time. 377 if (!dead->is_Con()) { 378 record_dead_node(dead->_idx); 379 } 380 if (dead->is_macro()) { 381 remove_macro_node(dead); 382 } 383 if (dead->is_expensive()) { 384 remove_expensive_node(dead); 385 } 386 if (dead->Opcode() == Op_Opaque4) { 387 remove_skeleton_predicate_opaq(dead); 388 } 389 if (dead->for_post_loop_opts_igvn()) { 390 remove_from_post_loop_opts_igvn(dead); 391 } 392 if (dead->is_Call()) { 393 remove_useless_late_inlines( &_late_inlines, dead); 394 remove_useless_late_inlines( &_string_late_inlines, dead); 395 remove_useless_late_inlines( &_boxing_late_inlines, dead); 396 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead); 397 } 398 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 399 bs->unregister_potential_barrier_node(dead); 400 } 401 402 // Disconnect all useless nodes by disconnecting those at the boundary. 403 void Compile::disconnect_useless_nodes(Unique_Node_List &useful, Unique_Node_List* worklist) { 404 uint next = 0; 405 while (next < useful.size()) { 406 Node *n = useful.at(next++); 407 if (n->is_SafePoint()) { 408 // We're done with a parsing phase. Replaced nodes are not valid 409 // beyond that point. 410 n->as_SafePoint()->delete_replaced_nodes(); 411 } 412 // Use raw traversal of out edges since this code removes out edges 413 int max = n->outcnt(); 414 for (int j = 0; j < max; ++j) { 415 Node* child = n->raw_out(j); 416 if (!useful.member(child)) { 417 assert(!child->is_top() || child != top(), 418 "If top is cached in Compile object it is in useful list"); 419 // Only need to remove this out-edge to the useless node 420 n->raw_del_out(j); 421 --j; 422 --max; 423 } 424 } 425 if (n->outcnt() == 1 && n->has_special_unique_user()) { 426 worklist->push(n->unique_out()); 427 } 428 } 429 430 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes 431 remove_useless_nodes(_predicate_opaqs, useful); // remove useless predicate opaque nodes 432 remove_useless_nodes(_skeleton_predicate_opaqs, useful); 433 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes 434 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass 435 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes 436 #ifdef ASSERT 437 if (_modified_nodes != NULL) { 438 _modified_nodes->remove_useless_nodes(useful.member_set()); 439 } 440 #endif 441 442 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 443 bs->eliminate_useless_gc_barriers(useful, this); 444 // clean up the late inline lists 445 remove_useless_late_inlines( &_late_inlines, useful); 446 remove_useless_late_inlines( &_string_late_inlines, useful); 447 remove_useless_late_inlines( &_boxing_late_inlines, useful); 448 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful); 449 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 450 } 451 452 // ============================================================================ 453 //------------------------------CompileWrapper--------------------------------- 454 class CompileWrapper : public StackObj { 455 Compile *const _compile; 456 public: 457 CompileWrapper(Compile* compile); 458 459 ~CompileWrapper(); 460 }; 461 462 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 463 // the Compile* pointer is stored in the current ciEnv: 464 ciEnv* env = compile->env(); 465 assert(env == ciEnv::current(), "must already be a ciEnv active"); 466 assert(env->compiler_data() == NULL, "compile already active?"); 467 env->set_compiler_data(compile); 468 assert(compile == Compile::current(), "sanity"); 469 470 compile->set_type_dict(NULL); 471 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 472 compile->clone_map().set_clone_idx(0); 473 compile->set_type_last_size(0); 474 compile->set_last_tf(NULL, NULL); 475 compile->set_indexSet_arena(NULL); 476 compile->set_indexSet_free_block_list(NULL); 477 compile->init_type_arena(); 478 Type::Initialize(compile); 479 _compile->begin_method(); 480 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 481 } 482 CompileWrapper::~CompileWrapper() { 483 _compile->end_method(); 484 _compile->env()->set_compiler_data(NULL); 485 } 486 487 488 //----------------------------print_compile_messages--------------------------- 489 void Compile::print_compile_messages() { 490 #ifndef PRODUCT 491 // Check if recompiling 492 if (_subsume_loads == false && PrintOpto) { 493 // Recompiling without allowing machine instructions to subsume loads 494 tty->print_cr("*********************************************************"); 495 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 496 tty->print_cr("*********************************************************"); 497 } 498 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 499 // Recompiling without escape analysis 500 tty->print_cr("*********************************************************"); 501 tty->print_cr("** Bailout: Recompile without escape analysis **"); 502 tty->print_cr("*********************************************************"); 503 } 504 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 505 // Recompiling without boxing elimination 506 tty->print_cr("*********************************************************"); 507 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 508 tty->print_cr("*********************************************************"); 509 } 510 if ((_do_locks_coarsening != EliminateLocks) && PrintOpto) { 511 // Recompiling without locks coarsening 512 tty->print_cr("*********************************************************"); 513 tty->print_cr("** Bailout: Recompile without locks coarsening **"); 514 tty->print_cr("*********************************************************"); 515 } 516 if (env()->break_at_compile()) { 517 // Open the debugger when compiling this method. 518 tty->print("### Breaking when compiling: "); 519 method()->print_short_name(); 520 tty->cr(); 521 BREAKPOINT; 522 } 523 524 if( PrintOpto ) { 525 if (is_osr_compilation()) { 526 tty->print("[OSR]%3d", _compile_id); 527 } else { 528 tty->print("%3d", _compile_id); 529 } 530 } 531 #endif 532 } 533 534 // ============================================================================ 535 //------------------------------Compile standard------------------------------- 536 debug_only( int Compile::_debug_idx = 100000; ) 537 538 // Compile a method. entry_bci is -1 for normal compilations and indicates 539 // the continuation bci for on stack replacement. 540 541 542 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci, 543 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, 544 bool do_locks_coarsening, bool install_code, DirectiveSet* directive) 545 : Phase(Compiler), 546 _compile_id(ci_env->compile_id()), 547 _subsume_loads(subsume_loads), 548 _do_escape_analysis(do_escape_analysis), 549 _install_code(install_code), 550 _eliminate_boxing(eliminate_boxing), 551 _do_locks_coarsening(do_locks_coarsening), 552 _method(target), 553 _entry_bci(osr_bci), 554 _stub_function(NULL), 555 _stub_name(NULL), 556 _stub_entry_point(NULL), 557 _max_node_limit(MaxNodeLimit), 558 _post_loop_opts_phase(false), 559 _inlining_progress(false), 560 _inlining_incrementally(false), 561 _do_cleanup(false), 562 _has_reserved_stack_access(target->has_reserved_stack_access()), 563 #ifndef PRODUCT 564 _igv_idx(0), 565 _trace_opto_output(directive->TraceOptoOutputOption), 566 _print_ideal(directive->PrintIdealOption), 567 #endif 568 _has_method_handle_invokes(false), 569 _clinit_barrier_on_entry(false), 570 _stress_seed(0), 571 _comp_arena(mtCompiler), 572 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 573 _env(ci_env), 574 _directive(directive), 575 _log(ci_env->log()), 576 _failure_reason(NULL), 577 _intrinsics (comp_arena(), 0, 0, NULL), 578 _macro_nodes (comp_arena(), 8, 0, NULL), 579 _predicate_opaqs (comp_arena(), 8, 0, NULL), 580 _skeleton_predicate_opaqs (comp_arena(), 8, 0, NULL), 581 _expensive_nodes (comp_arena(), 8, 0, NULL), 582 _for_post_loop_igvn(comp_arena(), 8, 0, NULL), 583 _coarsened_locks (comp_arena(), 8, 0, NULL), 584 _congraph(NULL), 585 NOT_PRODUCT(_printer(NULL) COMMA) 586 _dead_node_list(comp_arena()), 587 _dead_node_count(0), 588 _node_arena(mtCompiler), 589 _old_arena(mtCompiler), 590 _mach_constant_base_node(NULL), 591 _Compile_types(mtCompiler), 592 _initial_gvn(NULL), 593 _for_igvn(NULL), 594 _late_inlines(comp_arena(), 2, 0, NULL), 595 _string_late_inlines(comp_arena(), 2, 0, NULL), 596 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 597 _vector_reboxing_late_inlines(comp_arena(), 2, 0, NULL), 598 _late_inlines_pos(0), 599 _number_of_mh_late_inlines(0), 600 _native_invokers(comp_arena(), 1, 0, NULL), 601 _print_inlining_stream(NULL), 602 _print_inlining_list(NULL), 603 _print_inlining_idx(0), 604 _print_inlining_output(NULL), 605 _replay_inline_data(NULL), 606 _java_calls(0), 607 _inner_loops(0), 608 _interpreter_frame_size(0) 609 #ifndef PRODUCT 610 , _in_dump_cnt(0) 611 #endif 612 { 613 C = this; 614 CompileWrapper cw(this); 615 616 if (CITimeVerbose) { 617 tty->print(" "); 618 target->holder()->name()->print(); 619 tty->print("."); 620 target->print_short_name(); 621 tty->print(" "); 622 } 623 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 624 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 625 626 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 627 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 628 // We can always print a disassembly, either abstract (hex dump) or 629 // with the help of a suitable hsdis library. Thus, we should not 630 // couple print_assembly and print_opto_assembly controls. 631 // But: always print opto and regular assembly on compile command 'print'. 632 bool print_assembly = directive->PrintAssemblyOption; 633 set_print_assembly(print_opto_assembly || print_assembly); 634 #else 635 set_print_assembly(false); // must initialize. 636 #endif 637 638 #ifndef PRODUCT 639 set_parsed_irreducible_loop(false); 640 641 if (directive->ReplayInlineOption) { 642 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 643 } 644 #endif 645 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 646 set_print_intrinsics(directive->PrintIntrinsicsOption); 647 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 648 649 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 650 // Make sure the method being compiled gets its own MDO, 651 // so we can at least track the decompile_count(). 652 // Need MDO to record RTM code generation state. 653 method()->ensure_method_data(); 654 } 655 656 Init(::AliasLevel); 657 658 659 print_compile_messages(); 660 661 _ilt = InlineTree::build_inline_tree_root(); 662 663 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 664 assert(num_alias_types() >= AliasIdxRaw, ""); 665 666 #define MINIMUM_NODE_HASH 1023 667 // Node list that Iterative GVN will start with 668 Unique_Node_List for_igvn(comp_arena()); 669 set_for_igvn(&for_igvn); 670 671 // GVN that will be run immediately on new nodes 672 uint estimated_size = method()->code_size()*4+64; 673 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 674 PhaseGVN gvn(node_arena(), estimated_size); 675 set_initial_gvn(&gvn); 676 677 print_inlining_init(); 678 { // Scope for timing the parser 679 TracePhase tp("parse", &timers[_t_parser]); 680 681 // Put top into the hash table ASAP. 682 initial_gvn()->transform_no_reclaim(top()); 683 684 // Set up tf(), start(), and find a CallGenerator. 685 CallGenerator* cg = NULL; 686 if (is_osr_compilation()) { 687 const TypeTuple *domain = StartOSRNode::osr_domain(); 688 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 689 init_tf(TypeFunc::make(domain, range)); 690 StartNode* s = new StartOSRNode(root(), domain); 691 initial_gvn()->set_type_bottom(s); 692 init_start(s); 693 cg = CallGenerator::for_osr(method(), entry_bci()); 694 } else { 695 // Normal case. 696 init_tf(TypeFunc::make(method())); 697 StartNode* s = new StartNode(root(), tf()->domain()); 698 initial_gvn()->set_type_bottom(s); 699 init_start(s); 700 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 701 // With java.lang.ref.reference.get() we must go through the 702 // intrinsic - even when get() is the root 703 // method of the compile - so that, if necessary, the value in 704 // the referent field of the reference object gets recorded by 705 // the pre-barrier code. 706 cg = find_intrinsic(method(), false); 707 } 708 if (cg == NULL) { 709 float past_uses = method()->interpreter_invocation_count(); 710 float expected_uses = past_uses; 711 cg = CallGenerator::for_inline(method(), expected_uses); 712 } 713 } 714 if (failing()) return; 715 if (cg == NULL) { 716 record_method_not_compilable("cannot parse method"); 717 return; 718 } 719 JVMState* jvms = build_start_state(start(), tf()); 720 if ((jvms = cg->generate(jvms)) == NULL) { 721 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 722 record_method_not_compilable("method parse failed"); 723 } 724 return; 725 } 726 GraphKit kit(jvms); 727 728 if (!kit.stopped()) { 729 // Accept return values, and transfer control we know not where. 730 // This is done by a special, unique ReturnNode bound to root. 731 return_values(kit.jvms()); 732 } 733 734 if (kit.has_exceptions()) { 735 // Any exceptions that escape from this call must be rethrown 736 // to whatever caller is dynamically above us on the stack. 737 // This is done by a special, unique RethrowNode bound to root. 738 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 739 } 740 741 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 742 743 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 744 inline_string_calls(true); 745 } 746 747 if (failing()) return; 748 749 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 750 751 // Remove clutter produced by parsing. 752 if (!failing()) { 753 ResourceMark rm; 754 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 755 } 756 } 757 758 // Note: Large methods are capped off in do_one_bytecode(). 759 if (failing()) return; 760 761 // After parsing, node notes are no longer automagic. 762 // They must be propagated by register_new_node_with_optimizer(), 763 // clone(), or the like. 764 set_default_node_notes(NULL); 765 766 #ifndef PRODUCT 767 if (should_print(1)) { 768 _printer->print_inlining(); 769 } 770 #endif 771 772 if (failing()) return; 773 NOT_PRODUCT( verify_graph_edges(); ) 774 775 // If any phase is randomized for stress testing, seed random number 776 // generation and log the seed for repeatability. 777 if (StressLCM || StressGCM || StressIGVN || StressCCP) { 778 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && RepeatCompilation)) { 779 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds()); 780 FLAG_SET_ERGO(StressSeed, _stress_seed); 781 } else { 782 _stress_seed = StressSeed; 783 } 784 if (_log != NULL) { 785 _log->elem("stress_test seed='%u'", _stress_seed); 786 } 787 } 788 789 // Now optimize 790 Optimize(); 791 if (failing()) return; 792 NOT_PRODUCT( verify_graph_edges(); ) 793 794 #ifndef PRODUCT 795 if (print_ideal()) { 796 ttyLocker ttyl; // keep the following output all in one block 797 // This output goes directly to the tty, not the compiler log. 798 // To enable tools to match it up with the compilation activity, 799 // be sure to tag this tty output with the compile ID. 800 if (xtty != NULL) { 801 xtty->head("ideal compile_id='%d'%s", compile_id(), 802 is_osr_compilation() ? " compile_kind='osr'" : 803 ""); 804 } 805 root()->dump(9999); 806 if (xtty != NULL) { 807 xtty->tail("ideal"); 808 } 809 } 810 #endif 811 812 #ifdef ASSERT 813 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 814 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 815 #endif 816 817 // Dump compilation data to replay it. 818 if (directive->DumpReplayOption) { 819 env()->dump_replay_data(_compile_id); 820 } 821 if (directive->DumpInlineOption && (ilt() != NULL)) { 822 env()->dump_inline_data(_compile_id); 823 } 824 825 // Now that we know the size of all the monitors we can add a fixed slot 826 // for the original deopt pc. 827 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size); 828 set_fixed_slots(next_slot); 829 830 // Compute when to use implicit null checks. Used by matching trap based 831 // nodes and NullCheck optimization. 832 set_allowed_deopt_reasons(); 833 834 // Now generate code 835 Code_Gen(); 836 } 837 838 //------------------------------Compile---------------------------------------- 839 // Compile a runtime stub 840 Compile::Compile( ciEnv* ci_env, 841 TypeFunc_generator generator, 842 address stub_function, 843 const char *stub_name, 844 int is_fancy_jump, 845 bool pass_tls, 846 bool return_pc, 847 DirectiveSet* directive) 848 : Phase(Compiler), 849 _compile_id(0), 850 _subsume_loads(true), 851 _do_escape_analysis(false), 852 _install_code(true), 853 _eliminate_boxing(false), 854 _do_locks_coarsening(false), 855 _method(NULL), 856 _entry_bci(InvocationEntryBci), 857 _stub_function(stub_function), 858 _stub_name(stub_name), 859 _stub_entry_point(NULL), 860 _max_node_limit(MaxNodeLimit), 861 _post_loop_opts_phase(false), 862 _inlining_progress(false), 863 _inlining_incrementally(false), 864 _has_reserved_stack_access(false), 865 #ifndef PRODUCT 866 _igv_idx(0), 867 _trace_opto_output(directive->TraceOptoOutputOption), 868 _print_ideal(directive->PrintIdealOption), 869 #endif 870 _has_method_handle_invokes(false), 871 _clinit_barrier_on_entry(false), 872 _stress_seed(0), 873 _comp_arena(mtCompiler), 874 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 875 _env(ci_env), 876 _directive(directive), 877 _log(ci_env->log()), 878 _failure_reason(NULL), 879 _congraph(NULL), 880 NOT_PRODUCT(_printer(NULL) COMMA) 881 _dead_node_list(comp_arena()), 882 _dead_node_count(0), 883 _node_arena(mtCompiler), 884 _old_arena(mtCompiler), 885 _mach_constant_base_node(NULL), 886 _Compile_types(mtCompiler), 887 _initial_gvn(NULL), 888 _for_igvn(NULL), 889 _number_of_mh_late_inlines(0), 890 _native_invokers(), 891 _print_inlining_stream(NULL), 892 _print_inlining_list(NULL), 893 _print_inlining_idx(0), 894 _print_inlining_output(NULL), 895 _replay_inline_data(NULL), 896 _java_calls(0), 897 _inner_loops(0), 898 _interpreter_frame_size(0), 899 #ifndef PRODUCT 900 _in_dump_cnt(0), 901 #endif 902 _allowed_reasons(0) { 903 C = this; 904 905 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 906 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 907 908 #ifndef PRODUCT 909 set_print_assembly(PrintFrameConverterAssembly); 910 set_parsed_irreducible_loop(false); 911 #else 912 set_print_assembly(false); // Must initialize. 913 #endif 914 set_has_irreducible_loop(false); // no loops 915 916 CompileWrapper cw(this); 917 Init(/*AliasLevel=*/ 0); 918 init_tf((*generator)()); 919 920 { 921 // The following is a dummy for the sake of GraphKit::gen_stub 922 Unique_Node_List for_igvn(comp_arena()); 923 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 924 PhaseGVN gvn(Thread::current()->resource_area(),255); 925 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 926 gvn.transform_no_reclaim(top()); 927 928 GraphKit kit; 929 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 930 } 931 932 NOT_PRODUCT( verify_graph_edges(); ) 933 934 Code_Gen(); 935 } 936 937 //------------------------------Init------------------------------------------- 938 // Prepare for a single compilation 939 void Compile::Init(int aliaslevel) { 940 _unique = 0; 941 _regalloc = NULL; 942 943 _tf = NULL; // filled in later 944 _top = NULL; // cached later 945 _matcher = NULL; // filled in later 946 _cfg = NULL; // filled in later 947 948 IA32_ONLY( set_24_bit_selection_and_mode(true, false); ) 949 950 _node_note_array = NULL; 951 _default_node_notes = NULL; 952 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 953 954 _immutable_memory = NULL; // filled in at first inquiry 955 956 // Globally visible Nodes 957 // First set TOP to NULL to give safe behavior during creation of RootNode 958 set_cached_top_node(NULL); 959 set_root(new RootNode()); 960 // Now that you have a Root to point to, create the real TOP 961 set_cached_top_node( new ConNode(Type::TOP) ); 962 set_recent_alloc(NULL, NULL); 963 964 // Create Debug Information Recorder to record scopes, oopmaps, etc. 965 env()->set_oop_recorder(new OopRecorder(env()->arena())); 966 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 967 env()->set_dependencies(new Dependencies(env())); 968 969 _fixed_slots = 0; 970 set_has_split_ifs(false); 971 set_has_loops(false); // first approximation 972 set_has_stringbuilder(false); 973 set_has_boxed_value(false); 974 _trap_can_recompile = false; // no traps emitted yet 975 _major_progress = true; // start out assuming good things will happen 976 set_has_unsafe_access(false); 977 set_max_vector_size(0); 978 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 979 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 980 set_decompile_count(0); 981 982 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 983 _loop_opts_cnt = LoopOptsCount; 984 set_do_inlining(Inline); 985 set_max_inline_size(MaxInlineSize); 986 set_freq_inline_size(FreqInlineSize); 987 set_do_scheduling(OptoScheduling); 988 989 set_do_vector_loop(false); 990 991 if (AllowVectorizeOnDemand) { 992 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 993 set_do_vector_loop(true); 994 NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii());}) 995 } else if (has_method() && method()->name() != 0 && 996 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 997 set_do_vector_loop(true); 998 } 999 } 1000 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1001 NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n", method()->name()->as_quoted_ascii());}) 1002 1003 set_age_code(has_method() && method()->profile_aging()); 1004 set_rtm_state(NoRTM); // No RTM lock eliding by default 1005 _max_node_limit = _directive->MaxNodeLimitOption; 1006 1007 #if INCLUDE_RTM_OPT 1008 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1009 int rtm_state = method()->method_data()->rtm_state(); 1010 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) { 1011 // Don't generate RTM lock eliding code. 1012 set_rtm_state(NoRTM); 1013 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1014 // Generate RTM lock eliding code without abort ratio calculation code. 1015 set_rtm_state(UseRTM); 1016 } else if (UseRTMDeopt) { 1017 // Generate RTM lock eliding code and include abort ratio calculation 1018 // code if UseRTMDeopt is on. 1019 set_rtm_state(ProfileRTM); 1020 } 1021 } 1022 #endif 1023 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1024 set_clinit_barrier_on_entry(true); 1025 } 1026 if (debug_info()->recording_non_safepoints()) { 1027 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1028 (comp_arena(), 8, 0, NULL)); 1029 set_default_node_notes(Node_Notes::make(this)); 1030 } 1031 1032 // // -- Initialize types before each compile -- 1033 // // Update cached type information 1034 // if( _method && _method->constants() ) 1035 // Type::update_loaded_types(_method, _method->constants()); 1036 1037 // Init alias_type map. 1038 if (!_do_escape_analysis && aliaslevel == 3) 1039 aliaslevel = 2; // No unique types without escape analysis 1040 _AliasLevel = aliaslevel; 1041 const int grow_ats = 16; 1042 _max_alias_types = grow_ats; 1043 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1044 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1045 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1046 { 1047 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1048 } 1049 // Initialize the first few types. 1050 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1051 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1052 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1053 _num_alias_types = AliasIdxRaw+1; 1054 // Zero out the alias type cache. 1055 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1056 // A NULL adr_type hits in the cache right away. Preload the right answer. 1057 probe_alias_cache(NULL)->_index = AliasIdxTop; 1058 1059 #ifdef ASSERT 1060 _type_verify_symmetry = true; 1061 _phase_optimize_finished = false; 1062 _exception_backedge = false; 1063 #endif 1064 } 1065 1066 //---------------------------init_start---------------------------------------- 1067 // Install the StartNode on this compile object. 1068 void Compile::init_start(StartNode* s) { 1069 if (failing()) 1070 return; // already failing 1071 assert(s == start(), ""); 1072 } 1073 1074 /** 1075 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1076 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1077 * the ideal graph. 1078 */ 1079 StartNode* Compile::start() const { 1080 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1081 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1082 Node* start = root()->fast_out(i); 1083 if (start->is_Start()) { 1084 return start->as_Start(); 1085 } 1086 } 1087 fatal("Did not find Start node!"); 1088 return NULL; 1089 } 1090 1091 //-------------------------------immutable_memory------------------------------------- 1092 // Access immutable memory 1093 Node* Compile::immutable_memory() { 1094 if (_immutable_memory != NULL) { 1095 return _immutable_memory; 1096 } 1097 StartNode* s = start(); 1098 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1099 Node *p = s->fast_out(i); 1100 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1101 _immutable_memory = p; 1102 return _immutable_memory; 1103 } 1104 } 1105 ShouldNotReachHere(); 1106 return NULL; 1107 } 1108 1109 //----------------------set_cached_top_node------------------------------------ 1110 // Install the cached top node, and make sure Node::is_top works correctly. 1111 void Compile::set_cached_top_node(Node* tn) { 1112 if (tn != NULL) verify_top(tn); 1113 Node* old_top = _top; 1114 _top = tn; 1115 // Calling Node::setup_is_top allows the nodes the chance to adjust 1116 // their _out arrays. 1117 if (_top != NULL) _top->setup_is_top(); 1118 if (old_top != NULL) old_top->setup_is_top(); 1119 assert(_top == NULL || top()->is_top(), ""); 1120 } 1121 1122 #ifdef ASSERT 1123 uint Compile::count_live_nodes_by_graph_walk() { 1124 Unique_Node_List useful(comp_arena()); 1125 // Get useful node list by walking the graph. 1126 identify_useful_nodes(useful); 1127 return useful.size(); 1128 } 1129 1130 void Compile::print_missing_nodes() { 1131 1132 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1133 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1134 return; 1135 } 1136 1137 // This is an expensive function. It is executed only when the user 1138 // specifies VerifyIdealNodeCount option or otherwise knows the 1139 // additional work that needs to be done to identify reachable nodes 1140 // by walking the flow graph and find the missing ones using 1141 // _dead_node_list. 1142 1143 Unique_Node_List useful(comp_arena()); 1144 // Get useful node list by walking the graph. 1145 identify_useful_nodes(useful); 1146 1147 uint l_nodes = C->live_nodes(); 1148 uint l_nodes_by_walk = useful.size(); 1149 1150 if (l_nodes != l_nodes_by_walk) { 1151 if (_log != NULL) { 1152 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1153 _log->stamp(); 1154 _log->end_head(); 1155 } 1156 VectorSet& useful_member_set = useful.member_set(); 1157 int last_idx = l_nodes_by_walk; 1158 for (int i = 0; i < last_idx; i++) { 1159 if (useful_member_set.test(i)) { 1160 if (_dead_node_list.test(i)) { 1161 if (_log != NULL) { 1162 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1163 } 1164 if (PrintIdealNodeCount) { 1165 // Print the log message to tty 1166 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1167 useful.at(i)->dump(); 1168 } 1169 } 1170 } 1171 else if (! _dead_node_list.test(i)) { 1172 if (_log != NULL) { 1173 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1174 } 1175 if (PrintIdealNodeCount) { 1176 // Print the log message to tty 1177 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1178 } 1179 } 1180 } 1181 if (_log != NULL) { 1182 _log->tail("mismatched_nodes"); 1183 } 1184 } 1185 } 1186 void Compile::record_modified_node(Node* n) { 1187 if (_modified_nodes != NULL && !_inlining_incrementally && !n->is_Con()) { 1188 _modified_nodes->push(n); 1189 } 1190 } 1191 1192 void Compile::remove_modified_node(Node* n) { 1193 if (_modified_nodes != NULL) { 1194 _modified_nodes->remove(n); 1195 } 1196 } 1197 #endif 1198 1199 #ifndef PRODUCT 1200 void Compile::verify_top(Node* tn) const { 1201 if (tn != NULL) { 1202 assert(tn->is_Con(), "top node must be a constant"); 1203 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1204 assert(tn->in(0) != NULL, "must have live top node"); 1205 } 1206 } 1207 #endif 1208 1209 1210 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1211 1212 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1213 guarantee(arr != NULL, ""); 1214 int num_blocks = arr->length(); 1215 if (grow_by < num_blocks) grow_by = num_blocks; 1216 int num_notes = grow_by * _node_notes_block_size; 1217 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1218 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1219 while (num_notes > 0) { 1220 arr->append(notes); 1221 notes += _node_notes_block_size; 1222 num_notes -= _node_notes_block_size; 1223 } 1224 assert(num_notes == 0, "exact multiple, please"); 1225 } 1226 1227 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1228 if (source == NULL || dest == NULL) return false; 1229 1230 if (dest->is_Con()) 1231 return false; // Do not push debug info onto constants. 1232 1233 #ifdef ASSERT 1234 // Leave a bread crumb trail pointing to the original node: 1235 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1236 dest->set_debug_orig(source); 1237 } 1238 #endif 1239 1240 if (node_note_array() == NULL) 1241 return false; // Not collecting any notes now. 1242 1243 // This is a copy onto a pre-existing node, which may already have notes. 1244 // If both nodes have notes, do not overwrite any pre-existing notes. 1245 Node_Notes* source_notes = node_notes_at(source->_idx); 1246 if (source_notes == NULL || source_notes->is_clear()) return false; 1247 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1248 if (dest_notes == NULL || dest_notes->is_clear()) { 1249 return set_node_notes_at(dest->_idx, source_notes); 1250 } 1251 1252 Node_Notes merged_notes = (*source_notes); 1253 // The order of operations here ensures that dest notes will win... 1254 merged_notes.update_from(dest_notes); 1255 return set_node_notes_at(dest->_idx, &merged_notes); 1256 } 1257 1258 1259 //--------------------------allow_range_check_smearing------------------------- 1260 // Gating condition for coalescing similar range checks. 1261 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1262 // single covering check that is at least as strong as any of them. 1263 // If the optimization succeeds, the simplified (strengthened) range check 1264 // will always succeed. If it fails, we will deopt, and then give up 1265 // on the optimization. 1266 bool Compile::allow_range_check_smearing() const { 1267 // If this method has already thrown a range-check, 1268 // assume it was because we already tried range smearing 1269 // and it failed. 1270 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1271 return !already_trapped; 1272 } 1273 1274 1275 //------------------------------flatten_alias_type----------------------------- 1276 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1277 int offset = tj->offset(); 1278 TypePtr::PTR ptr = tj->ptr(); 1279 1280 // Known instance (scalarizable allocation) alias only with itself. 1281 bool is_known_inst = tj->isa_oopptr() != NULL && 1282 tj->is_oopptr()->is_known_instance(); 1283 1284 // Process weird unsafe references. 1285 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1286 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1287 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1288 tj = TypeOopPtr::BOTTOM; 1289 ptr = tj->ptr(); 1290 offset = tj->offset(); 1291 } 1292 1293 // Array pointers need some flattening 1294 const TypeAryPtr *ta = tj->isa_aryptr(); 1295 if (ta && ta->is_stable()) { 1296 // Erase stability property for alias analysis. 1297 tj = ta = ta->cast_to_stable(false); 1298 } 1299 if( ta && is_known_inst ) { 1300 if ( offset != Type::OffsetBot && 1301 offset > arrayOopDesc::length_offset_in_bytes() ) { 1302 offset = Type::OffsetBot; // Flatten constant access into array body only 1303 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1304 } 1305 } else if( ta && _AliasLevel >= 2 ) { 1306 // For arrays indexed by constant indices, we flatten the alias 1307 // space to include all of the array body. Only the header, klass 1308 // and array length can be accessed un-aliased. 1309 if( offset != Type::OffsetBot ) { 1310 if( ta->const_oop() ) { // MethodData* or Method* 1311 offset = Type::OffsetBot; // Flatten constant access into array body 1312 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1313 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1314 // range is OK as-is. 1315 tj = ta = TypeAryPtr::RANGE; 1316 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1317 tj = TypeInstPtr::KLASS; // all klass loads look alike 1318 ta = TypeAryPtr::RANGE; // generic ignored junk 1319 ptr = TypePtr::BotPTR; 1320 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1321 tj = TypeInstPtr::MARK; 1322 ta = TypeAryPtr::RANGE; // generic ignored junk 1323 ptr = TypePtr::BotPTR; 1324 } else { // Random constant offset into array body 1325 offset = Type::OffsetBot; // Flatten constant access into array body 1326 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1327 } 1328 } 1329 // Arrays of fixed size alias with arrays of unknown size. 1330 if (ta->size() != TypeInt::POS) { 1331 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1332 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1333 } 1334 // Arrays of known objects become arrays of unknown objects. 1335 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1336 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1337 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1338 } 1339 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1340 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1341 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1342 } 1343 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1344 // cannot be distinguished by bytecode alone. 1345 if (ta->elem() == TypeInt::BOOL) { 1346 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1347 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1348 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1349 } 1350 // During the 2nd round of IterGVN, NotNull castings are removed. 1351 // Make sure the Bottom and NotNull variants alias the same. 1352 // Also, make sure exact and non-exact variants alias the same. 1353 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1354 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1355 } 1356 } 1357 1358 // Oop pointers need some flattening 1359 const TypeInstPtr *to = tj->isa_instptr(); 1360 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1361 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1362 if( ptr == TypePtr::Constant ) { 1363 if (to->klass() != ciEnv::current()->Class_klass() || 1364 offset < k->layout_helper_size_in_bytes()) { 1365 // No constant oop pointers (such as Strings); they alias with 1366 // unknown strings. 1367 assert(!is_known_inst, "not scalarizable allocation"); 1368 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1369 } 1370 } else if( is_known_inst ) { 1371 tj = to; // Keep NotNull and klass_is_exact for instance type 1372 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1373 // During the 2nd round of IterGVN, NotNull castings are removed. 1374 // Make sure the Bottom and NotNull variants alias the same. 1375 // Also, make sure exact and non-exact variants alias the same. 1376 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1377 } 1378 if (to->speculative() != NULL) { 1379 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1380 } 1381 // Canonicalize the holder of this field 1382 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1383 // First handle header references such as a LoadKlassNode, even if the 1384 // object's klass is unloaded at compile time (4965979). 1385 if (!is_known_inst) { // Do it only for non-instance types 1386 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1387 } 1388 } else if (offset < 0 || offset >= k->layout_helper_size_in_bytes()) { 1389 // Static fields are in the space above the normal instance 1390 // fields in the java.lang.Class instance. 1391 if (to->klass() != ciEnv::current()->Class_klass()) { 1392 to = NULL; 1393 tj = TypeOopPtr::BOTTOM; 1394 offset = tj->offset(); 1395 } 1396 } else { 1397 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1398 assert(offset < canonical_holder->layout_helper_size_in_bytes(), ""); 1399 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1400 if( is_known_inst ) { 1401 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1402 } else { 1403 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1404 } 1405 } 1406 } 1407 } 1408 1409 // Klass pointers to object array klasses need some flattening 1410 const TypeKlassPtr *tk = tj->isa_klassptr(); 1411 if( tk ) { 1412 // If we are referencing a field within a Klass, we need 1413 // to assume the worst case of an Object. Both exact and 1414 // inexact types must flatten to the same alias class so 1415 // use NotNull as the PTR. 1416 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1417 1418 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1419 TypeKlassPtr::OBJECT->klass(), 1420 offset); 1421 } 1422 1423 ciKlass* klass = tk->klass(); 1424 if( klass->is_obj_array_klass() ) { 1425 ciKlass* k = TypeAryPtr::OOPS->klass(); 1426 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1427 k = TypeInstPtr::BOTTOM->klass(); 1428 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1429 } 1430 1431 // Check for precise loads from the primary supertype array and force them 1432 // to the supertype cache alias index. Check for generic array loads from 1433 // the primary supertype array and also force them to the supertype cache 1434 // alias index. Since the same load can reach both, we need to merge 1435 // these 2 disparate memories into the same alias class. Since the 1436 // primary supertype array is read-only, there's no chance of confusion 1437 // where we bypass an array load and an array store. 1438 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1439 if (offset == Type::OffsetBot || 1440 (offset >= primary_supers_offset && 1441 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1442 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1443 offset = in_bytes(Klass::secondary_super_cache_offset()); 1444 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1445 } 1446 } 1447 1448 // Flatten all Raw pointers together. 1449 if (tj->base() == Type::RawPtr) 1450 tj = TypeRawPtr::BOTTOM; 1451 1452 if (tj->base() == Type::AnyPtr) 1453 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1454 1455 // Flatten all to bottom for now 1456 switch( _AliasLevel ) { 1457 case 0: 1458 tj = TypePtr::BOTTOM; 1459 break; 1460 case 1: // Flatten to: oop, static, field or array 1461 switch (tj->base()) { 1462 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1463 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1464 case Type::AryPtr: // do not distinguish arrays at all 1465 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1466 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1467 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1468 default: ShouldNotReachHere(); 1469 } 1470 break; 1471 case 2: // No collapsing at level 2; keep all splits 1472 case 3: // No collapsing at level 3; keep all splits 1473 break; 1474 default: 1475 Unimplemented(); 1476 } 1477 1478 offset = tj->offset(); 1479 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1480 1481 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1482 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1483 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1484 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1485 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1486 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1487 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1488 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1489 assert( tj->ptr() != TypePtr::TopPTR && 1490 tj->ptr() != TypePtr::AnyNull && 1491 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1492 // assert( tj->ptr() != TypePtr::Constant || 1493 // tj->base() == Type::RawPtr || 1494 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1495 1496 return tj; 1497 } 1498 1499 void Compile::AliasType::Init(int i, const TypePtr* at) { 1500 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1501 _index = i; 1502 _adr_type = at; 1503 _field = NULL; 1504 _element = NULL; 1505 _is_rewritable = true; // default 1506 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1507 if (atoop != NULL && atoop->is_known_instance()) { 1508 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1509 _general_index = Compile::current()->get_alias_index(gt); 1510 } else { 1511 _general_index = 0; 1512 } 1513 } 1514 1515 BasicType Compile::AliasType::basic_type() const { 1516 if (element() != NULL) { 1517 const Type* element = adr_type()->is_aryptr()->elem(); 1518 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1519 } if (field() != NULL) { 1520 return field()->layout_type(); 1521 } else { 1522 return T_ILLEGAL; // unknown 1523 } 1524 } 1525 1526 //---------------------------------print_on------------------------------------ 1527 #ifndef PRODUCT 1528 void Compile::AliasType::print_on(outputStream* st) { 1529 if (index() < 10) 1530 st->print("@ <%d> ", index()); 1531 else st->print("@ <%d>", index()); 1532 st->print(is_rewritable() ? " " : " RO"); 1533 int offset = adr_type()->offset(); 1534 if (offset == Type::OffsetBot) 1535 st->print(" +any"); 1536 else st->print(" +%-3d", offset); 1537 st->print(" in "); 1538 adr_type()->dump_on(st); 1539 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1540 if (field() != NULL && tjp) { 1541 if (tjp->klass() != field()->holder() || 1542 tjp->offset() != field()->offset_in_bytes()) { 1543 st->print(" != "); 1544 field()->print(); 1545 st->print(" ***"); 1546 } 1547 } 1548 } 1549 1550 void print_alias_types() { 1551 Compile* C = Compile::current(); 1552 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1553 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1554 C->alias_type(idx)->print_on(tty); 1555 tty->cr(); 1556 } 1557 } 1558 #endif 1559 1560 1561 //----------------------------probe_alias_cache-------------------------------- 1562 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1563 intptr_t key = (intptr_t) adr_type; 1564 key ^= key >> logAliasCacheSize; 1565 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1566 } 1567 1568 1569 //-----------------------------grow_alias_types-------------------------------- 1570 void Compile::grow_alias_types() { 1571 const int old_ats = _max_alias_types; // how many before? 1572 const int new_ats = old_ats; // how many more? 1573 const int grow_ats = old_ats+new_ats; // how many now? 1574 _max_alias_types = grow_ats; 1575 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1576 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1577 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1578 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1579 } 1580 1581 1582 //--------------------------------find_alias_type------------------------------ 1583 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1584 if (_AliasLevel == 0) 1585 return alias_type(AliasIdxBot); 1586 1587 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1588 if (ace->_adr_type == adr_type) { 1589 return alias_type(ace->_index); 1590 } 1591 1592 // Handle special cases. 1593 if (adr_type == NULL) return alias_type(AliasIdxTop); 1594 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1595 1596 // Do it the slow way. 1597 const TypePtr* flat = flatten_alias_type(adr_type); 1598 1599 #ifdef ASSERT 1600 { 1601 ResourceMark rm; 1602 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1603 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1604 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1605 Type::str(adr_type)); 1606 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1607 const TypeOopPtr* foop = flat->is_oopptr(); 1608 // Scalarizable allocations have exact klass always. 1609 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1610 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1611 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1612 Type::str(foop), Type::str(xoop)); 1613 } 1614 } 1615 #endif 1616 1617 int idx = AliasIdxTop; 1618 for (int i = 0; i < num_alias_types(); i++) { 1619 if (alias_type(i)->adr_type() == flat) { 1620 idx = i; 1621 break; 1622 } 1623 } 1624 1625 if (idx == AliasIdxTop) { 1626 if (no_create) return NULL; 1627 // Grow the array if necessary. 1628 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1629 // Add a new alias type. 1630 idx = _num_alias_types++; 1631 _alias_types[idx]->Init(idx, flat); 1632 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1633 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1634 if (flat->isa_instptr()) { 1635 if (flat->offset() == java_lang_Class::klass_offset() 1636 && flat->is_instptr()->klass() == env()->Class_klass()) 1637 alias_type(idx)->set_rewritable(false); 1638 } 1639 if (flat->isa_aryptr()) { 1640 #ifdef ASSERT 1641 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1642 // (T_BYTE has the weakest alignment and size restrictions...) 1643 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1644 #endif 1645 if (flat->offset() == TypePtr::OffsetBot) { 1646 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1647 } 1648 } 1649 if (flat->isa_klassptr()) { 1650 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1651 alias_type(idx)->set_rewritable(false); 1652 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1653 alias_type(idx)->set_rewritable(false); 1654 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1655 alias_type(idx)->set_rewritable(false); 1656 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1657 alias_type(idx)->set_rewritable(false); 1658 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1659 alias_type(idx)->set_rewritable(false); 1660 } 1661 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1662 // but the base pointer type is not distinctive enough to identify 1663 // references into JavaThread.) 1664 1665 // Check for final fields. 1666 const TypeInstPtr* tinst = flat->isa_instptr(); 1667 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1668 ciField* field; 1669 if (tinst->const_oop() != NULL && 1670 tinst->klass() == ciEnv::current()->Class_klass() && 1671 tinst->offset() >= (tinst->klass()->as_instance_klass()->layout_helper_size_in_bytes())) { 1672 // static field 1673 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1674 field = k->get_field_by_offset(tinst->offset(), true); 1675 } else { 1676 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1677 field = k->get_field_by_offset(tinst->offset(), false); 1678 } 1679 assert(field == NULL || 1680 original_field == NULL || 1681 (field->holder() == original_field->holder() && 1682 field->offset() == original_field->offset() && 1683 field->is_static() == original_field->is_static()), "wrong field?"); 1684 // Set field() and is_rewritable() attributes. 1685 if (field != NULL) alias_type(idx)->set_field(field); 1686 } 1687 } 1688 1689 // Fill the cache for next time. 1690 ace->_adr_type = adr_type; 1691 ace->_index = idx; 1692 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1693 1694 // Might as well try to fill the cache for the flattened version, too. 1695 AliasCacheEntry* face = probe_alias_cache(flat); 1696 if (face->_adr_type == NULL) { 1697 face->_adr_type = flat; 1698 face->_index = idx; 1699 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1700 } 1701 1702 return alias_type(idx); 1703 } 1704 1705 1706 Compile::AliasType* Compile::alias_type(ciField* field) { 1707 const TypeOopPtr* t; 1708 if (field->is_static()) 1709 t = TypeInstPtr::make(field->holder()->java_mirror()); 1710 else 1711 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1712 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1713 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1714 return atp; 1715 } 1716 1717 1718 //------------------------------have_alias_type-------------------------------- 1719 bool Compile::have_alias_type(const TypePtr* adr_type) { 1720 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1721 if (ace->_adr_type == adr_type) { 1722 return true; 1723 } 1724 1725 // Handle special cases. 1726 if (adr_type == NULL) return true; 1727 if (adr_type == TypePtr::BOTTOM) return true; 1728 1729 return find_alias_type(adr_type, true, NULL) != NULL; 1730 } 1731 1732 //-----------------------------must_alias-------------------------------------- 1733 // True if all values of the given address type are in the given alias category. 1734 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1735 if (alias_idx == AliasIdxBot) return true; // the universal category 1736 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1737 if (alias_idx == AliasIdxTop) return false; // the empty category 1738 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1739 1740 // the only remaining possible overlap is identity 1741 int adr_idx = get_alias_index(adr_type); 1742 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1743 assert(adr_idx == alias_idx || 1744 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1745 && adr_type != TypeOopPtr::BOTTOM), 1746 "should not be testing for overlap with an unsafe pointer"); 1747 return adr_idx == alias_idx; 1748 } 1749 1750 //------------------------------can_alias-------------------------------------- 1751 // True if any values of the given address type are in the given alias category. 1752 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1753 if (alias_idx == AliasIdxTop) return false; // the empty category 1754 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1755 // Known instance doesn't alias with bottom memory 1756 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1757 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1758 1759 // the only remaining possible overlap is identity 1760 int adr_idx = get_alias_index(adr_type); 1761 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1762 return adr_idx == alias_idx; 1763 } 1764 1765 //---------------------cleanup_loop_predicates----------------------- 1766 // Remove the opaque nodes that protect the predicates so that all unused 1767 // checks and uncommon_traps will be eliminated from the ideal graph 1768 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1769 if (predicate_count()==0) return; 1770 for (int i = predicate_count(); i > 0; i--) { 1771 Node * n = predicate_opaque1_node(i-1); 1772 assert(n->Opcode() == Op_Opaque1, "must be"); 1773 igvn.replace_node(n, n->in(1)); 1774 } 1775 assert(predicate_count()==0, "should be clean!"); 1776 } 1777 1778 void Compile::record_for_post_loop_opts_igvn(Node* n) { 1779 if (!n->for_post_loop_opts_igvn()) { 1780 assert(!_for_post_loop_igvn.contains(n), "duplicate"); 1781 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1782 _for_post_loop_igvn.append(n); 1783 } 1784 } 1785 1786 void Compile::remove_from_post_loop_opts_igvn(Node* n) { 1787 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1788 _for_post_loop_igvn.remove(n); 1789 } 1790 1791 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) { 1792 // Verify that all previous optimizations produced a valid graph 1793 // at least to this point, even if no loop optimizations were done. 1794 PhaseIdealLoop::verify(igvn); 1795 1796 C->set_post_loop_opts_phase(); // no more loop opts allowed 1797 1798 assert(!C->major_progress(), "not cleared"); 1799 1800 if (_for_post_loop_igvn.length() > 0) { 1801 while (_for_post_loop_igvn.length() > 0) { 1802 Node* n = _for_post_loop_igvn.pop(); 1803 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1804 igvn._worklist.push(n); 1805 } 1806 igvn.optimize(); 1807 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed"); 1808 1809 // Sometimes IGVN sets major progress (e.g., when processing loop nodes). 1810 if (C->major_progress()) { 1811 C->clear_major_progress(); // ensure that major progress is now clear 1812 } 1813 } 1814 } 1815 1816 // StringOpts and late inlining of string methods 1817 void Compile::inline_string_calls(bool parse_time) { 1818 { 1819 // remove useless nodes to make the usage analysis simpler 1820 ResourceMark rm; 1821 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1822 } 1823 1824 { 1825 ResourceMark rm; 1826 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1827 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1828 print_method(PHASE_AFTER_STRINGOPTS, 3); 1829 } 1830 1831 // now inline anything that we skipped the first time around 1832 if (!parse_time) { 1833 _late_inlines_pos = _late_inlines.length(); 1834 } 1835 1836 while (_string_late_inlines.length() > 0) { 1837 CallGenerator* cg = _string_late_inlines.pop(); 1838 cg->do_late_inline(); 1839 if (failing()) return; 1840 } 1841 _string_late_inlines.trunc_to(0); 1842 } 1843 1844 // Late inlining of boxing methods 1845 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1846 if (_boxing_late_inlines.length() > 0) { 1847 assert(has_boxed_value(), "inconsistent"); 1848 1849 PhaseGVN* gvn = initial_gvn(); 1850 set_inlining_incrementally(true); 1851 1852 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1853 for_igvn()->clear(); 1854 gvn->replace_with(&igvn); 1855 1856 _late_inlines_pos = _late_inlines.length(); 1857 1858 while (_boxing_late_inlines.length() > 0) { 1859 CallGenerator* cg = _boxing_late_inlines.pop(); 1860 cg->do_late_inline(); 1861 if (failing()) return; 1862 } 1863 _boxing_late_inlines.trunc_to(0); 1864 1865 inline_incrementally_cleanup(igvn); 1866 1867 set_inlining_incrementally(false); 1868 } 1869 } 1870 1871 bool Compile::inline_incrementally_one() { 1872 assert(IncrementalInline, "incremental inlining should be on"); 1873 1874 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 1875 1876 set_inlining_progress(false); 1877 set_do_cleanup(false); 1878 1879 for (int i = 0; i < _late_inlines.length(); i++) { 1880 _late_inlines_pos = i+1; 1881 CallGenerator* cg = _late_inlines.at(i); 1882 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline(); 1883 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call 1884 cg->do_late_inline(); 1885 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed"); 1886 if (failing()) { 1887 return false; 1888 } else if (inlining_progress()) { 1889 _late_inlines_pos = i+1; // restore the position in case new elements were inserted 1890 print_method(PHASE_INCREMENTAL_INLINE_STEP, cg->call_node(), 3); 1891 break; // process one call site at a time 1892 } 1893 } else { 1894 // Ignore late inline direct calls when inlining is not allowed. 1895 // They are left in the late inline list when node budget is exhausted until the list is fully drained. 1896 } 1897 } 1898 // Remove processed elements. 1899 _late_inlines.remove_till(_late_inlines_pos); 1900 _late_inlines_pos = 0; 1901 1902 assert(inlining_progress() || _late_inlines.length() == 0, "no progress"); 1903 1904 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 1905 1906 set_inlining_progress(false); 1907 set_do_cleanup(false); 1908 1909 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption; 1910 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup; 1911 } 1912 1913 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 1914 { 1915 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 1916 ResourceMark rm; 1917 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1918 } 1919 { 1920 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 1921 igvn = PhaseIterGVN(initial_gvn()); 1922 igvn.optimize(); 1923 } 1924 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3); 1925 } 1926 1927 // Perform incremental inlining until bound on number of live nodes is reached 1928 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1929 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 1930 1931 set_inlining_incrementally(true); 1932 uint low_live_nodes = 0; 1933 1934 while (_late_inlines.length() > 0) { 1935 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1936 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1937 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 1938 // PhaseIdealLoop is expensive so we only try it once we are 1939 // out of live nodes and we only try it again if the previous 1940 // helped got the number of nodes down significantly 1941 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 1942 if (failing()) return; 1943 low_live_nodes = live_nodes(); 1944 _major_progress = true; 1945 } 1946 1947 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1948 bool do_print_inlining = print_inlining() || print_intrinsics(); 1949 if (do_print_inlining || log() != NULL) { 1950 // Print inlining message for candidates that we couldn't inline for lack of space. 1951 for (int i = 0; i < _late_inlines.length(); i++) { 1952 CallGenerator* cg = _late_inlines.at(i); 1953 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 1954 if (do_print_inlining) { 1955 cg->print_inlining_late(msg); 1956 } 1957 log_late_inline_failure(cg, msg); 1958 } 1959 } 1960 break; // finish 1961 } 1962 } 1963 1964 for_igvn()->clear(); 1965 initial_gvn()->replace_with(&igvn); 1966 1967 while (inline_incrementally_one()) { 1968 assert(!failing(), "inconsistent"); 1969 } 1970 if (failing()) return; 1971 1972 inline_incrementally_cleanup(igvn); 1973 1974 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 1975 1976 if (failing()) return; 1977 1978 if (_late_inlines.length() == 0) { 1979 break; // no more progress 1980 } 1981 } 1982 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1983 1984 if (_string_late_inlines.length() > 0) { 1985 assert(has_stringbuilder(), "inconsistent"); 1986 for_igvn()->clear(); 1987 initial_gvn()->replace_with(&igvn); 1988 1989 inline_string_calls(false); 1990 1991 if (failing()) return; 1992 1993 inline_incrementally_cleanup(igvn); 1994 } 1995 1996 set_inlining_incrementally(false); 1997 } 1998 1999 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) { 2000 // "inlining_incrementally() == false" is used to signal that no inlining is allowed 2001 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details). 2002 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == NULL" 2003 // as if "inlining_incrementally() == true" were set. 2004 assert(inlining_incrementally() == false, "not allowed"); 2005 assert(_modified_nodes == NULL, "not allowed"); 2006 assert(_late_inlines.length() > 0, "sanity"); 2007 2008 while (_late_inlines.length() > 0) { 2009 for_igvn()->clear(); 2010 initial_gvn()->replace_with(&igvn); 2011 2012 while (inline_incrementally_one()) { 2013 assert(!failing(), "inconsistent"); 2014 } 2015 if (failing()) return; 2016 2017 inline_incrementally_cleanup(igvn); 2018 } 2019 } 2020 2021 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2022 if (_loop_opts_cnt > 0) { 2023 while (major_progress() && (_loop_opts_cnt > 0)) { 2024 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2025 PhaseIdealLoop::optimize(igvn, mode); 2026 _loop_opts_cnt--; 2027 if (failing()) return false; 2028 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2029 } 2030 } 2031 return true; 2032 } 2033 2034 // Remove edges from "root" to each SafePoint at a backward branch. 2035 // They were inserted during parsing (see add_safepoint()) to make 2036 // infinite loops without calls or exceptions visible to root, i.e., 2037 // useful. 2038 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2039 Node *r = root(); 2040 if (r != NULL) { 2041 for (uint i = r->req(); i < r->len(); ++i) { 2042 Node *n = r->in(i); 2043 if (n != NULL && n->is_SafePoint()) { 2044 r->rm_prec(i); 2045 if (n->outcnt() == 0) { 2046 igvn.remove_dead_node(n); 2047 } 2048 --i; 2049 } 2050 } 2051 // Parsing may have added top inputs to the root node (Path 2052 // leading to the Halt node proven dead). Make sure we get a 2053 // chance to clean them up. 2054 igvn._worklist.push(r); 2055 igvn.optimize(); 2056 } 2057 } 2058 2059 //------------------------------Optimize--------------------------------------- 2060 // Given a graph, optimize it. 2061 void Compile::Optimize() { 2062 TracePhase tp("optimizer", &timers[_t_optimizer]); 2063 2064 #ifndef PRODUCT 2065 if (env()->break_at_compile()) { 2066 BREAKPOINT; 2067 } 2068 2069 #endif 2070 2071 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2072 #ifdef ASSERT 2073 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2074 #endif 2075 2076 ResourceMark rm; 2077 2078 print_inlining_reinit(); 2079 2080 NOT_PRODUCT( verify_graph_edges(); ) 2081 2082 print_method(PHASE_AFTER_PARSING); 2083 2084 { 2085 // Iterative Global Value Numbering, including ideal transforms 2086 // Initialize IterGVN with types and values from parse-time GVN 2087 PhaseIterGVN igvn(initial_gvn()); 2088 #ifdef ASSERT 2089 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2090 #endif 2091 { 2092 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2093 igvn.optimize(); 2094 } 2095 2096 if (failing()) return; 2097 2098 print_method(PHASE_ITER_GVN1, 2); 2099 2100 inline_incrementally(igvn); 2101 2102 print_method(PHASE_INCREMENTAL_INLINE, 2); 2103 2104 if (failing()) return; 2105 2106 if (eliminate_boxing()) { 2107 // Inline valueOf() methods now. 2108 inline_boxing_calls(igvn); 2109 2110 if (AlwaysIncrementalInline) { 2111 inline_incrementally(igvn); 2112 } 2113 2114 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2115 2116 if (failing()) return; 2117 } 2118 2119 // Remove the speculative part of types and clean up the graph from 2120 // the extra CastPP nodes whose only purpose is to carry them. Do 2121 // that early so that optimizations are not disrupted by the extra 2122 // CastPP nodes. 2123 remove_speculative_types(igvn); 2124 2125 // No more new expensive nodes will be added to the list from here 2126 // so keep only the actual candidates for optimizations. 2127 cleanup_expensive_nodes(igvn); 2128 2129 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2130 if (EnableVectorSupport && has_vbox_nodes()) { 2131 TracePhase tp("", &timers[_t_vector]); 2132 PhaseVector pv(igvn); 2133 pv.optimize_vector_boxes(); 2134 2135 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2136 } 2137 assert(!has_vbox_nodes(), "sanity"); 2138 2139 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2140 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2141 initial_gvn()->replace_with(&igvn); 2142 for_igvn()->clear(); 2143 Unique_Node_List new_worklist(C->comp_arena()); 2144 { 2145 ResourceMark rm; 2146 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2147 } 2148 Unique_Node_List* save_for_igvn = for_igvn(); 2149 set_for_igvn(&new_worklist); 2150 igvn = PhaseIterGVN(initial_gvn()); 2151 igvn.optimize(); 2152 set_for_igvn(save_for_igvn); 2153 } 2154 2155 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop 2156 // safepoints 2157 remove_root_to_sfpts_edges(igvn); 2158 2159 // Perform escape analysis 2160 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2161 if (has_loops()) { 2162 // Cleanup graph (remove dead nodes). 2163 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2164 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2165 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2166 if (failing()) return; 2167 } 2168 ConnectionGraph::do_analysis(this, &igvn); 2169 2170 if (failing()) return; 2171 2172 // Optimize out fields loads from scalar replaceable allocations. 2173 igvn.optimize(); 2174 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2175 2176 if (failing()) return; 2177 2178 if (congraph() != NULL && macro_count() > 0) { 2179 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2180 PhaseMacroExpand mexp(igvn); 2181 mexp.eliminate_macro_nodes(); 2182 igvn.set_delay_transform(false); 2183 2184 igvn.optimize(); 2185 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2186 2187 if (failing()) return; 2188 } 2189 } 2190 2191 // Loop transforms on the ideal graph. Range Check Elimination, 2192 // peeling, unrolling, etc. 2193 2194 // Set loop opts counter 2195 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2196 { 2197 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2198 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2199 _loop_opts_cnt--; 2200 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2201 if (failing()) return; 2202 } 2203 // Loop opts pass if partial peeling occurred in previous pass 2204 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2205 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2206 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2207 _loop_opts_cnt--; 2208 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2209 if (failing()) return; 2210 } 2211 // Loop opts pass for loop-unrolling before CCP 2212 if(major_progress() && (_loop_opts_cnt > 0)) { 2213 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2214 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2215 _loop_opts_cnt--; 2216 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2217 } 2218 if (!failing()) { 2219 // Verify that last round of loop opts produced a valid graph 2220 PhaseIdealLoop::verify(igvn); 2221 } 2222 } 2223 if (failing()) return; 2224 2225 // Conditional Constant Propagation; 2226 PhaseCCP ccp( &igvn ); 2227 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2228 { 2229 TracePhase tp("ccp", &timers[_t_ccp]); 2230 ccp.do_transform(); 2231 } 2232 print_method(PHASE_CCP1, 2); 2233 2234 assert( true, "Break here to ccp.dump_old2new_map()"); 2235 2236 // Iterative Global Value Numbering, including ideal transforms 2237 { 2238 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2239 igvn = ccp; 2240 igvn.optimize(); 2241 } 2242 print_method(PHASE_ITER_GVN2, 2); 2243 2244 if (failing()) return; 2245 2246 // Loop transforms on the ideal graph. Range Check Elimination, 2247 // peeling, unrolling, etc. 2248 if (!optimize_loops(igvn, LoopOptsDefault)) { 2249 return; 2250 } 2251 2252 if (failing()) return; 2253 2254 C->clear_major_progress(); // ensure that major progress is now clear 2255 2256 process_for_post_loop_opts_igvn(igvn); 2257 2258 #ifdef ASSERT 2259 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2260 #endif 2261 2262 { 2263 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2264 PhaseMacroExpand mex(igvn); 2265 if (mex.expand_macro_nodes()) { 2266 assert(failing(), "must bail out w/ explicit message"); 2267 return; 2268 } 2269 print_method(PHASE_MACRO_EXPANSION, 2); 2270 } 2271 2272 { 2273 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2274 if (bs->expand_barriers(this, igvn)) { 2275 assert(failing(), "must bail out w/ explicit message"); 2276 return; 2277 } 2278 print_method(PHASE_BARRIER_EXPANSION, 2); 2279 } 2280 2281 if (C->max_vector_size() > 0) { 2282 C->optimize_logic_cones(igvn); 2283 igvn.optimize(); 2284 } 2285 2286 DEBUG_ONLY( _modified_nodes = NULL; ) 2287 2288 assert(igvn._worklist.size() == 0, "not empty"); 2289 2290 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty"); 2291 2292 if (_late_inlines.length() > 0) { 2293 // More opportunities to optimize virtual and MH calls. 2294 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option. 2295 process_late_inline_calls_no_inline(igvn); 2296 } 2297 } // (End scope of igvn; run destructor if necessary for asserts.) 2298 2299 check_no_dead_use(); 2300 2301 process_print_inlining(); 2302 2303 // A method with only infinite loops has no edges entering loops from root 2304 { 2305 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2306 if (final_graph_reshaping()) { 2307 assert(failing(), "must bail out w/ explicit message"); 2308 return; 2309 } 2310 } 2311 2312 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2313 DEBUG_ONLY(set_phase_optimize_finished();) 2314 } 2315 2316 #ifdef ASSERT 2317 void Compile::check_no_dead_use() const { 2318 ResourceMark rm; 2319 Unique_Node_List wq; 2320 wq.push(root()); 2321 for (uint i = 0; i < wq.size(); ++i) { 2322 Node* n = wq.at(i); 2323 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2324 Node* u = n->fast_out(j); 2325 if (u->outcnt() == 0 && !u->is_Con()) { 2326 u->dump(); 2327 fatal("no reachable node should have no use"); 2328 } 2329 wq.push(u); 2330 } 2331 } 2332 } 2333 #endif 2334 2335 void Compile::inline_vector_reboxing_calls() { 2336 if (C->_vector_reboxing_late_inlines.length() > 0) { 2337 _late_inlines_pos = C->_late_inlines.length(); 2338 while (_vector_reboxing_late_inlines.length() > 0) { 2339 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 2340 cg->do_late_inline(); 2341 if (failing()) return; 2342 print_method(PHASE_INLINE_VECTOR_REBOX, cg->call_node()); 2343 } 2344 _vector_reboxing_late_inlines.trunc_to(0); 2345 } 2346 } 2347 2348 bool Compile::has_vbox_nodes() { 2349 if (C->_vector_reboxing_late_inlines.length() > 0) { 2350 return true; 2351 } 2352 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 2353 Node * n = C->macro_node(macro_idx); 2354 assert(n->is_macro(), "only macro nodes expected here"); 2355 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 2356 return true; 2357 } 2358 } 2359 return false; 2360 } 2361 2362 //---------------------------- Bitwise operation packing optimization --------------------------- 2363 2364 static bool is_vector_unary_bitwise_op(Node* n) { 2365 return n->Opcode() == Op_XorV && 2366 VectorNode::is_vector_bitwise_not_pattern(n); 2367 } 2368 2369 static bool is_vector_binary_bitwise_op(Node* n) { 2370 switch (n->Opcode()) { 2371 case Op_AndV: 2372 case Op_OrV: 2373 return true; 2374 2375 case Op_XorV: 2376 return !is_vector_unary_bitwise_op(n); 2377 2378 default: 2379 return false; 2380 } 2381 } 2382 2383 static bool is_vector_ternary_bitwise_op(Node* n) { 2384 return n->Opcode() == Op_MacroLogicV; 2385 } 2386 2387 static bool is_vector_bitwise_op(Node* n) { 2388 return is_vector_unary_bitwise_op(n) || 2389 is_vector_binary_bitwise_op(n) || 2390 is_vector_ternary_bitwise_op(n); 2391 } 2392 2393 static bool is_vector_bitwise_cone_root(Node* n) { 2394 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) { 2395 return false; 2396 } 2397 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2398 if (is_vector_bitwise_op(n->fast_out(i))) { 2399 return false; 2400 } 2401 } 2402 return true; 2403 } 2404 2405 static uint collect_unique_inputs(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2406 uint cnt = 0; 2407 if (is_vector_bitwise_op(n)) { 2408 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2409 for (uint i = 1; i < n->req(); i++) { 2410 Node* in = n->in(i); 2411 bool skip = VectorNode::is_all_ones_vector(in); 2412 if (!skip && !inputs.member(in)) { 2413 inputs.push(in); 2414 cnt++; 2415 } 2416 } 2417 assert(cnt <= 1, "not unary"); 2418 } else { 2419 uint last_req = n->req(); 2420 if (is_vector_ternary_bitwise_op(n)) { 2421 last_req = n->req() - 1; // skip last input 2422 } 2423 for (uint i = 1; i < last_req; i++) { 2424 Node* def = n->in(i); 2425 if (!inputs.member(def)) { 2426 inputs.push(def); 2427 cnt++; 2428 } 2429 } 2430 } 2431 partition.push(n); 2432 } else { // not a bitwise operations 2433 if (!inputs.member(n)) { 2434 inputs.push(n); 2435 cnt++; 2436 } 2437 } 2438 return cnt; 2439 } 2440 2441 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 2442 Unique_Node_List useful_nodes; 2443 C->identify_useful_nodes(useful_nodes); 2444 2445 for (uint i = 0; i < useful_nodes.size(); i++) { 2446 Node* n = useful_nodes.at(i); 2447 if (is_vector_bitwise_cone_root(n)) { 2448 list.push(n); 2449 } 2450 } 2451 } 2452 2453 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 2454 const TypeVect* vt, 2455 Unique_Node_List& partition, 2456 Unique_Node_List& inputs) { 2457 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 2458 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 2459 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 2460 2461 Node* in1 = inputs.at(0); 2462 Node* in2 = inputs.at(1); 2463 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 2464 2465 uint func = compute_truth_table(partition, inputs); 2466 return igvn.transform(MacroLogicVNode::make(igvn, in3, in2, in1, func, vt)); 2467 } 2468 2469 static uint extract_bit(uint func, uint pos) { 2470 return (func & (1 << pos)) >> pos; 2471 } 2472 2473 // 2474 // A macro logic node represents a truth table. It has 4 inputs, 2475 // First three inputs corresponds to 3 columns of a truth table 2476 // and fourth input captures the logic function. 2477 // 2478 // eg. fn = (in1 AND in2) OR in3; 2479 // 2480 // MacroNode(in1,in2,in3,fn) 2481 // 2482 // ----------------- 2483 // in1 in2 in3 fn 2484 // ----------------- 2485 // 0 0 0 0 2486 // 0 0 1 1 2487 // 0 1 0 0 2488 // 0 1 1 1 2489 // 1 0 0 0 2490 // 1 0 1 1 2491 // 1 1 0 1 2492 // 1 1 1 1 2493 // 2494 2495 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 2496 int res = 0; 2497 for (int i = 0; i < 8; i++) { 2498 int bit1 = extract_bit(in1, i); 2499 int bit2 = extract_bit(in2, i); 2500 int bit3 = extract_bit(in3, i); 2501 2502 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 2503 int func_bit = extract_bit(func, func_bit_pos); 2504 2505 res |= func_bit << i; 2506 } 2507 return res; 2508 } 2509 2510 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 2511 assert(n != NULL, ""); 2512 assert(eval_map.contains(n), "absent"); 2513 return *(eval_map.get(n)); 2514 } 2515 2516 static void eval_operands(Node* n, 2517 uint& func1, uint& func2, uint& func3, 2518 ResourceHashtable<Node*,uint>& eval_map) { 2519 assert(is_vector_bitwise_op(n), ""); 2520 2521 if (is_vector_unary_bitwise_op(n)) { 2522 Node* opnd = n->in(1); 2523 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 2524 opnd = n->in(2); 2525 } 2526 func1 = eval_operand(opnd, eval_map); 2527 } else if (is_vector_binary_bitwise_op(n)) { 2528 func1 = eval_operand(n->in(1), eval_map); 2529 func2 = eval_operand(n->in(2), eval_map); 2530 } else { 2531 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 2532 func1 = eval_operand(n->in(1), eval_map); 2533 func2 = eval_operand(n->in(2), eval_map); 2534 func3 = eval_operand(n->in(3), eval_map); 2535 } 2536 } 2537 2538 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 2539 assert(inputs.size() <= 3, "sanity"); 2540 ResourceMark rm; 2541 uint res = 0; 2542 ResourceHashtable<Node*,uint> eval_map; 2543 2544 // Populate precomputed functions for inputs. 2545 // Each input corresponds to one column of 3 input truth-table. 2546 uint input_funcs[] = { 0xAA, // (_, _, a) -> a 2547 0xCC, // (_, b, _) -> b 2548 0xF0 }; // (c, _, _) -> c 2549 for (uint i = 0; i < inputs.size(); i++) { 2550 eval_map.put(inputs.at(i), input_funcs[i]); 2551 } 2552 2553 for (uint i = 0; i < partition.size(); i++) { 2554 Node* n = partition.at(i); 2555 2556 uint func1 = 0, func2 = 0, func3 = 0; 2557 eval_operands(n, func1, func2, func3, eval_map); 2558 2559 switch (n->Opcode()) { 2560 case Op_OrV: 2561 assert(func3 == 0, "not binary"); 2562 res = func1 | func2; 2563 break; 2564 case Op_AndV: 2565 assert(func3 == 0, "not binary"); 2566 res = func1 & func2; 2567 break; 2568 case Op_XorV: 2569 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2570 assert(func2 == 0 && func3 == 0, "not unary"); 2571 res = (~func1) & 0xFF; 2572 } else { 2573 assert(func3 == 0, "not binary"); 2574 res = func1 ^ func2; 2575 } 2576 break; 2577 case Op_MacroLogicV: 2578 // Ordering of inputs may change during evaluation of sub-tree 2579 // containing MacroLogic node as a child node, thus a re-evaluation 2580 // makes sure that function is evaluated in context of current 2581 // inputs. 2582 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 2583 break; 2584 2585 default: assert(false, "not supported: %s", n->Name()); 2586 } 2587 assert(res <= 0xFF, "invalid"); 2588 eval_map.put(n, res); 2589 } 2590 return res; 2591 } 2592 2593 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2594 assert(partition.size() == 0, "not empty"); 2595 assert(inputs.size() == 0, "not empty"); 2596 if (is_vector_ternary_bitwise_op(n)) { 2597 return false; 2598 } 2599 2600 bool is_unary_op = is_vector_unary_bitwise_op(n); 2601 if (is_unary_op) { 2602 assert(collect_unique_inputs(n, partition, inputs) == 1, "not unary"); 2603 return false; // too few inputs 2604 } 2605 2606 assert(is_vector_binary_bitwise_op(n), "not binary"); 2607 Node* in1 = n->in(1); 2608 Node* in2 = n->in(2); 2609 2610 int in1_unique_inputs_cnt = collect_unique_inputs(in1, partition, inputs); 2611 int in2_unique_inputs_cnt = collect_unique_inputs(in2, partition, inputs); 2612 partition.push(n); 2613 2614 // Too many inputs? 2615 if (inputs.size() > 3) { 2616 partition.clear(); 2617 inputs.clear(); 2618 { // Recompute in2 inputs 2619 Unique_Node_List not_used; 2620 in2_unique_inputs_cnt = collect_unique_inputs(in2, not_used, not_used); 2621 } 2622 // Pick the node with minimum number of inputs. 2623 if (in1_unique_inputs_cnt >= 3 && in2_unique_inputs_cnt >= 3) { 2624 return false; // still too many inputs 2625 } 2626 // Recompute partition & inputs. 2627 Node* child = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in1 : in2); 2628 collect_unique_inputs(child, partition, inputs); 2629 2630 Node* other_input = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in2 : in1); 2631 inputs.push(other_input); 2632 2633 partition.push(n); 2634 } 2635 2636 return (partition.size() == 2 || partition.size() == 3) && 2637 (inputs.size() == 2 || inputs.size() == 3); 2638 } 2639 2640 2641 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 2642 assert(is_vector_bitwise_op(n), "not a root"); 2643 2644 visited.set(n->_idx); 2645 2646 // 1) Do a DFS walk over the logic cone. 2647 for (uint i = 1; i < n->req(); i++) { 2648 Node* in = n->in(i); 2649 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 2650 process_logic_cone_root(igvn, in, visited); 2651 } 2652 } 2653 2654 // 2) Bottom up traversal: Merge node[s] with 2655 // the parent to form macro logic node. 2656 Unique_Node_List partition; 2657 Unique_Node_List inputs; 2658 if (compute_logic_cone(n, partition, inputs)) { 2659 const TypeVect* vt = n->bottom_type()->is_vect(); 2660 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 2661 igvn.replace_node(n, macro_logic); 2662 } 2663 } 2664 2665 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 2666 ResourceMark rm; 2667 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 2668 Unique_Node_List list; 2669 collect_logic_cone_roots(list); 2670 2671 while (list.size() > 0) { 2672 Node* n = list.pop(); 2673 const TypeVect* vt = n->bottom_type()->is_vect(); 2674 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 2675 if (supported) { 2676 VectorSet visited(comp_arena()); 2677 process_logic_cone_root(igvn, n, visited); 2678 } 2679 } 2680 } 2681 } 2682 2683 //------------------------------Code_Gen--------------------------------------- 2684 // Given a graph, generate code for it 2685 void Compile::Code_Gen() { 2686 if (failing()) { 2687 return; 2688 } 2689 2690 // Perform instruction selection. You might think we could reclaim Matcher 2691 // memory PDQ, but actually the Matcher is used in generating spill code. 2692 // Internals of the Matcher (including some VectorSets) must remain live 2693 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2694 // set a bit in reclaimed memory. 2695 2696 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2697 // nodes. Mapping is only valid at the root of each matched subtree. 2698 NOT_PRODUCT( verify_graph_edges(); ) 2699 2700 Matcher matcher; 2701 _matcher = &matcher; 2702 { 2703 TracePhase tp("matcher", &timers[_t_matcher]); 2704 matcher.match(); 2705 if (failing()) { 2706 return; 2707 } 2708 } 2709 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2710 // nodes. Mapping is only valid at the root of each matched subtree. 2711 NOT_PRODUCT( verify_graph_edges(); ) 2712 2713 // If you have too many nodes, or if matching has failed, bail out 2714 check_node_count(0, "out of nodes matching instructions"); 2715 if (failing()) { 2716 return; 2717 } 2718 2719 print_method(PHASE_MATCHING, 2); 2720 2721 // Build a proper-looking CFG 2722 PhaseCFG cfg(node_arena(), root(), matcher); 2723 _cfg = &cfg; 2724 { 2725 TracePhase tp("scheduler", &timers[_t_scheduler]); 2726 bool success = cfg.do_global_code_motion(); 2727 if (!success) { 2728 return; 2729 } 2730 2731 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2732 NOT_PRODUCT( verify_graph_edges(); ) 2733 cfg.verify(); 2734 } 2735 2736 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2737 _regalloc = ®alloc; 2738 { 2739 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2740 // Perform register allocation. After Chaitin, use-def chains are 2741 // no longer accurate (at spill code) and so must be ignored. 2742 // Node->LRG->reg mappings are still accurate. 2743 _regalloc->Register_Allocate(); 2744 2745 // Bail out if the allocator builds too many nodes 2746 if (failing()) { 2747 return; 2748 } 2749 } 2750 2751 // Prior to register allocation we kept empty basic blocks in case the 2752 // the allocator needed a place to spill. After register allocation we 2753 // are not adding any new instructions. If any basic block is empty, we 2754 // can now safely remove it. 2755 { 2756 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2757 cfg.remove_empty_blocks(); 2758 if (do_freq_based_layout()) { 2759 PhaseBlockLayout layout(cfg); 2760 } else { 2761 cfg.set_loop_alignment(); 2762 } 2763 cfg.fixup_flow(); 2764 cfg.remove_unreachable_blocks(); 2765 cfg.verify_dominator_tree(); 2766 } 2767 2768 // Apply peephole optimizations 2769 if( OptoPeephole ) { 2770 TracePhase tp("peephole", &timers[_t_peephole]); 2771 PhasePeephole peep( _regalloc, cfg); 2772 peep.do_transform(); 2773 } 2774 2775 // Do late expand if CPU requires this. 2776 if (Matcher::require_postalloc_expand) { 2777 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2778 cfg.postalloc_expand(_regalloc); 2779 } 2780 2781 // Convert Nodes to instruction bits in a buffer 2782 { 2783 TracePhase tp("output", &timers[_t_output]); 2784 PhaseOutput output; 2785 output.Output(); 2786 if (failing()) return; 2787 output.install(); 2788 } 2789 2790 print_method(PHASE_FINAL_CODE); 2791 2792 // He's dead, Jim. 2793 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2794 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2795 } 2796 2797 //------------------------------Final_Reshape_Counts--------------------------- 2798 // This class defines counters to help identify when a method 2799 // may/must be executed using hardware with only 24-bit precision. 2800 struct Final_Reshape_Counts : public StackObj { 2801 int _call_count; // count non-inlined 'common' calls 2802 int _float_count; // count float ops requiring 24-bit precision 2803 int _double_count; // count double ops requiring more precision 2804 int _java_call_count; // count non-inlined 'java' calls 2805 int _inner_loop_count; // count loops which need alignment 2806 VectorSet _visited; // Visitation flags 2807 Node_List _tests; // Set of IfNodes & PCTableNodes 2808 2809 Final_Reshape_Counts() : 2810 _call_count(0), _float_count(0), _double_count(0), 2811 _java_call_count(0), _inner_loop_count(0) { } 2812 2813 void inc_call_count () { _call_count ++; } 2814 void inc_float_count () { _float_count ++; } 2815 void inc_double_count() { _double_count++; } 2816 void inc_java_call_count() { _java_call_count++; } 2817 void inc_inner_loop_count() { _inner_loop_count++; } 2818 2819 int get_call_count () const { return _call_count ; } 2820 int get_float_count () const { return _float_count ; } 2821 int get_double_count() const { return _double_count; } 2822 int get_java_call_count() const { return _java_call_count; } 2823 int get_inner_loop_count() const { return _inner_loop_count; } 2824 }; 2825 2826 // Eliminate trivially redundant StoreCMs and accumulate their 2827 // precedence edges. 2828 void Compile::eliminate_redundant_card_marks(Node* n) { 2829 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2830 if (n->in(MemNode::Address)->outcnt() > 1) { 2831 // There are multiple users of the same address so it might be 2832 // possible to eliminate some of the StoreCMs 2833 Node* mem = n->in(MemNode::Memory); 2834 Node* adr = n->in(MemNode::Address); 2835 Node* val = n->in(MemNode::ValueIn); 2836 Node* prev = n; 2837 bool done = false; 2838 // Walk the chain of StoreCMs eliminating ones that match. As 2839 // long as it's a chain of single users then the optimization is 2840 // safe. Eliminating partially redundant StoreCMs would require 2841 // cloning copies down the other paths. 2842 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2843 if (adr == mem->in(MemNode::Address) && 2844 val == mem->in(MemNode::ValueIn)) { 2845 // redundant StoreCM 2846 if (mem->req() > MemNode::OopStore) { 2847 // Hasn't been processed by this code yet. 2848 n->add_prec(mem->in(MemNode::OopStore)); 2849 } else { 2850 // Already converted to precedence edge 2851 for (uint i = mem->req(); i < mem->len(); i++) { 2852 // Accumulate any precedence edges 2853 if (mem->in(i) != NULL) { 2854 n->add_prec(mem->in(i)); 2855 } 2856 } 2857 // Everything above this point has been processed. 2858 done = true; 2859 } 2860 // Eliminate the previous StoreCM 2861 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2862 assert(mem->outcnt() == 0, "should be dead"); 2863 mem->disconnect_inputs(this); 2864 } else { 2865 prev = mem; 2866 } 2867 mem = prev->in(MemNode::Memory); 2868 } 2869 } 2870 } 2871 2872 //------------------------------final_graph_reshaping_impl---------------------- 2873 // Implement items 1-5 from final_graph_reshaping below. 2874 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2875 2876 if ( n->outcnt() == 0 ) return; // dead node 2877 uint nop = n->Opcode(); 2878 2879 // Check for 2-input instruction with "last use" on right input. 2880 // Swap to left input. Implements item (2). 2881 if( n->req() == 3 && // two-input instruction 2882 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2883 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2884 n->in(2)->outcnt() == 1 &&// right use IS a last use 2885 !n->in(2)->is_Con() ) { // right use is not a constant 2886 // Check for commutative opcode 2887 switch( nop ) { 2888 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2889 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 2890 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 2891 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2892 case Op_AndL: case Op_XorL: case Op_OrL: 2893 case Op_AndI: case Op_XorI: case Op_OrI: { 2894 // Move "last use" input to left by swapping inputs 2895 n->swap_edges(1, 2); 2896 break; 2897 } 2898 default: 2899 break; 2900 } 2901 } 2902 2903 #ifdef ASSERT 2904 if( n->is_Mem() ) { 2905 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2906 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2907 // oop will be recorded in oop map if load crosses safepoint 2908 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2909 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2910 "raw memory operations should have control edge"); 2911 } 2912 if (n->is_MemBar()) { 2913 MemBarNode* mb = n->as_MemBar(); 2914 if (mb->trailing_store() || mb->trailing_load_store()) { 2915 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2916 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 2917 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2918 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2919 } else if (mb->leading()) { 2920 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2921 } 2922 } 2923 #endif 2924 // Count FPU ops and common calls, implements item (3) 2925 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop); 2926 if (!gc_handled) { 2927 final_graph_reshaping_main_switch(n, frc, nop); 2928 } 2929 2930 // Collect CFG split points 2931 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 2932 frc._tests.push(n); 2933 } 2934 } 2935 2936 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) { 2937 switch( nop ) { 2938 // Count all float operations that may use FPU 2939 case Op_AddF: 2940 case Op_SubF: 2941 case Op_MulF: 2942 case Op_DivF: 2943 case Op_NegF: 2944 case Op_ModF: 2945 case Op_ConvI2F: 2946 case Op_ConF: 2947 case Op_CmpF: 2948 case Op_CmpF3: 2949 case Op_StoreF: 2950 case Op_LoadF: 2951 // case Op_ConvL2F: // longs are split into 32-bit halves 2952 frc.inc_float_count(); 2953 break; 2954 2955 case Op_ConvF2D: 2956 case Op_ConvD2F: 2957 frc.inc_float_count(); 2958 frc.inc_double_count(); 2959 break; 2960 2961 // Count all double operations that may use FPU 2962 case Op_AddD: 2963 case Op_SubD: 2964 case Op_MulD: 2965 case Op_DivD: 2966 case Op_NegD: 2967 case Op_ModD: 2968 case Op_ConvI2D: 2969 case Op_ConvD2I: 2970 // case Op_ConvL2D: // handled by leaf call 2971 // case Op_ConvD2L: // handled by leaf call 2972 case Op_ConD: 2973 case Op_CmpD: 2974 case Op_CmpD3: 2975 case Op_StoreD: 2976 case Op_LoadD: 2977 case Op_LoadD_unaligned: 2978 frc.inc_double_count(); 2979 break; 2980 case Op_Opaque1: // Remove Opaque Nodes before matching 2981 case Op_Opaque2: // Remove Opaque Nodes before matching 2982 case Op_Opaque3: 2983 n->subsume_by(n->in(1), this); 2984 break; 2985 case Op_CallStaticJava: 2986 case Op_CallJava: 2987 case Op_CallDynamicJava: 2988 frc.inc_java_call_count(); // Count java call site; 2989 case Op_CallRuntime: 2990 case Op_CallLeaf: 2991 case Op_CallLeafVector: 2992 case Op_CallNative: 2993 case Op_CallLeafNoFP: { 2994 assert (n->is_Call(), ""); 2995 CallNode *call = n->as_Call(); 2996 // Count call sites where the FP mode bit would have to be flipped. 2997 // Do not count uncommon runtime calls: 2998 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2999 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3000 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3001 frc.inc_call_count(); // Count the call site 3002 } else { // See if uncommon argument is shared 3003 Node *n = call->in(TypeFunc::Parms); 3004 int nop = n->Opcode(); 3005 // Clone shared simple arguments to uncommon calls, item (1). 3006 if (n->outcnt() > 1 && 3007 !n->is_Proj() && 3008 nop != Op_CreateEx && 3009 nop != Op_CheckCastPP && 3010 nop != Op_DecodeN && 3011 nop != Op_DecodeNKlass && 3012 !n->is_Mem() && 3013 !n->is_Phi()) { 3014 Node *x = n->clone(); 3015 call->set_req(TypeFunc::Parms, x); 3016 } 3017 } 3018 break; 3019 } 3020 3021 case Op_StoreCM: 3022 { 3023 // Convert OopStore dependence into precedence edge 3024 Node* prec = n->in(MemNode::OopStore); 3025 n->del_req(MemNode::OopStore); 3026 n->add_prec(prec); 3027 eliminate_redundant_card_marks(n); 3028 } 3029 3030 // fall through 3031 3032 case Op_StoreB: 3033 case Op_StoreC: 3034 case Op_StorePConditional: 3035 case Op_StoreI: 3036 case Op_StoreL: 3037 case Op_StoreIConditional: 3038 case Op_StoreLConditional: 3039 case Op_CompareAndSwapB: 3040 case Op_CompareAndSwapS: 3041 case Op_CompareAndSwapI: 3042 case Op_CompareAndSwapL: 3043 case Op_CompareAndSwapP: 3044 case Op_CompareAndSwapN: 3045 case Op_WeakCompareAndSwapB: 3046 case Op_WeakCompareAndSwapS: 3047 case Op_WeakCompareAndSwapI: 3048 case Op_WeakCompareAndSwapL: 3049 case Op_WeakCompareAndSwapP: 3050 case Op_WeakCompareAndSwapN: 3051 case Op_CompareAndExchangeB: 3052 case Op_CompareAndExchangeS: 3053 case Op_CompareAndExchangeI: 3054 case Op_CompareAndExchangeL: 3055 case Op_CompareAndExchangeP: 3056 case Op_CompareAndExchangeN: 3057 case Op_GetAndAddS: 3058 case Op_GetAndAddB: 3059 case Op_GetAndAddI: 3060 case Op_GetAndAddL: 3061 case Op_GetAndSetS: 3062 case Op_GetAndSetB: 3063 case Op_GetAndSetI: 3064 case Op_GetAndSetL: 3065 case Op_GetAndSetP: 3066 case Op_GetAndSetN: 3067 case Op_StoreP: 3068 case Op_StoreN: 3069 case Op_StoreNKlass: 3070 case Op_LoadB: 3071 case Op_LoadUB: 3072 case Op_LoadUS: 3073 case Op_LoadI: 3074 case Op_LoadKlass: 3075 case Op_LoadNKlass: 3076 case Op_LoadL: 3077 case Op_LoadL_unaligned: 3078 case Op_LoadPLocked: 3079 case Op_LoadP: 3080 case Op_LoadN: 3081 case Op_LoadRange: 3082 case Op_LoadS: 3083 break; 3084 3085 case Op_AddP: { // Assert sane base pointers 3086 Node *addp = n->in(AddPNode::Address); 3087 assert( !addp->is_AddP() || 3088 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3089 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3090 "Base pointers must match (addp %u)", addp->_idx ); 3091 #ifdef _LP64 3092 if ((UseCompressedOops || UseCompressedClassPointers) && 3093 addp->Opcode() == Op_ConP && 3094 addp == n->in(AddPNode::Base) && 3095 n->in(AddPNode::Offset)->is_Con()) { 3096 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3097 // on the platform and on the compressed oops mode. 3098 // Use addressing with narrow klass to load with offset on x86. 3099 // Some platforms can use the constant pool to load ConP. 3100 // Do this transformation here since IGVN will convert ConN back to ConP. 3101 const Type* t = addp->bottom_type(); 3102 bool is_oop = t->isa_oopptr() != NULL; 3103 bool is_klass = t->isa_klassptr() != NULL; 3104 3105 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3106 (is_klass && Matcher::const_klass_prefer_decode())) { 3107 Node* nn = NULL; 3108 3109 int op = is_oop ? Op_ConN : Op_ConNKlass; 3110 3111 // Look for existing ConN node of the same exact type. 3112 Node* r = root(); 3113 uint cnt = r->outcnt(); 3114 for (uint i = 0; i < cnt; i++) { 3115 Node* m = r->raw_out(i); 3116 if (m!= NULL && m->Opcode() == op && 3117 m->bottom_type()->make_ptr() == t) { 3118 nn = m; 3119 break; 3120 } 3121 } 3122 if (nn != NULL) { 3123 // Decode a narrow oop to match address 3124 // [R12 + narrow_oop_reg<<3 + offset] 3125 if (is_oop) { 3126 nn = new DecodeNNode(nn, t); 3127 } else { 3128 nn = new DecodeNKlassNode(nn, t); 3129 } 3130 // Check for succeeding AddP which uses the same Base. 3131 // Otherwise we will run into the assertion above when visiting that guy. 3132 for (uint i = 0; i < n->outcnt(); ++i) { 3133 Node *out_i = n->raw_out(i); 3134 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3135 out_i->set_req(AddPNode::Base, nn); 3136 #ifdef ASSERT 3137 for (uint j = 0; j < out_i->outcnt(); ++j) { 3138 Node *out_j = out_i->raw_out(j); 3139 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3140 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3141 } 3142 #endif 3143 } 3144 } 3145 n->set_req(AddPNode::Base, nn); 3146 n->set_req(AddPNode::Address, nn); 3147 if (addp->outcnt() == 0) { 3148 addp->disconnect_inputs(this); 3149 } 3150 } 3151 } 3152 } 3153 #endif 3154 break; 3155 } 3156 3157 case Op_CastPP: { 3158 // Remove CastPP nodes to gain more freedom during scheduling but 3159 // keep the dependency they encode as control or precedence edges 3160 // (if control is set already) on memory operations. Some CastPP 3161 // nodes don't have a control (don't carry a dependency): skip 3162 // those. 3163 if (n->in(0) != NULL) { 3164 ResourceMark rm; 3165 Unique_Node_List wq; 3166 wq.push(n); 3167 for (uint next = 0; next < wq.size(); ++next) { 3168 Node *m = wq.at(next); 3169 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3170 Node* use = m->fast_out(i); 3171 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3172 use->ensure_control_or_add_prec(n->in(0)); 3173 } else { 3174 switch(use->Opcode()) { 3175 case Op_AddP: 3176 case Op_DecodeN: 3177 case Op_DecodeNKlass: 3178 case Op_CheckCastPP: 3179 case Op_CastPP: 3180 wq.push(use); 3181 break; 3182 } 3183 } 3184 } 3185 } 3186 } 3187 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3188 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3189 Node* in1 = n->in(1); 3190 const Type* t = n->bottom_type(); 3191 Node* new_in1 = in1->clone(); 3192 new_in1->as_DecodeN()->set_type(t); 3193 3194 if (!Matcher::narrow_oop_use_complex_address()) { 3195 // 3196 // x86, ARM and friends can handle 2 adds in addressing mode 3197 // and Matcher can fold a DecodeN node into address by using 3198 // a narrow oop directly and do implicit NULL check in address: 3199 // 3200 // [R12 + narrow_oop_reg<<3 + offset] 3201 // NullCheck narrow_oop_reg 3202 // 3203 // On other platforms (Sparc) we have to keep new DecodeN node and 3204 // use it to do implicit NULL check in address: 3205 // 3206 // decode_not_null narrow_oop_reg, base_reg 3207 // [base_reg + offset] 3208 // NullCheck base_reg 3209 // 3210 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3211 // to keep the information to which NULL check the new DecodeN node 3212 // corresponds to use it as value in implicit_null_check(). 3213 // 3214 new_in1->set_req(0, n->in(0)); 3215 } 3216 3217 n->subsume_by(new_in1, this); 3218 if (in1->outcnt() == 0) { 3219 in1->disconnect_inputs(this); 3220 } 3221 } else { 3222 n->subsume_by(n->in(1), this); 3223 if (n->outcnt() == 0) { 3224 n->disconnect_inputs(this); 3225 } 3226 } 3227 break; 3228 } 3229 #ifdef _LP64 3230 case Op_CmpP: 3231 // Do this transformation here to preserve CmpPNode::sub() and 3232 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3233 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3234 Node* in1 = n->in(1); 3235 Node* in2 = n->in(2); 3236 if (!in1->is_DecodeNarrowPtr()) { 3237 in2 = in1; 3238 in1 = n->in(2); 3239 } 3240 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3241 3242 Node* new_in2 = NULL; 3243 if (in2->is_DecodeNarrowPtr()) { 3244 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3245 new_in2 = in2->in(1); 3246 } else if (in2->Opcode() == Op_ConP) { 3247 const Type* t = in2->bottom_type(); 3248 if (t == TypePtr::NULL_PTR) { 3249 assert(in1->is_DecodeN(), "compare klass to null?"); 3250 // Don't convert CmpP null check into CmpN if compressed 3251 // oops implicit null check is not generated. 3252 // This will allow to generate normal oop implicit null check. 3253 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3254 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3255 // 3256 // This transformation together with CastPP transformation above 3257 // will generated code for implicit NULL checks for compressed oops. 3258 // 3259 // The original code after Optimize() 3260 // 3261 // LoadN memory, narrow_oop_reg 3262 // decode narrow_oop_reg, base_reg 3263 // CmpP base_reg, NULL 3264 // CastPP base_reg // NotNull 3265 // Load [base_reg + offset], val_reg 3266 // 3267 // after these transformations will be 3268 // 3269 // LoadN memory, narrow_oop_reg 3270 // CmpN narrow_oop_reg, NULL 3271 // decode_not_null narrow_oop_reg, base_reg 3272 // Load [base_reg + offset], val_reg 3273 // 3274 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3275 // since narrow oops can be used in debug info now (see the code in 3276 // final_graph_reshaping_walk()). 3277 // 3278 // At the end the code will be matched to 3279 // on x86: 3280 // 3281 // Load_narrow_oop memory, narrow_oop_reg 3282 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3283 // NullCheck narrow_oop_reg 3284 // 3285 // and on sparc: 3286 // 3287 // Load_narrow_oop memory, narrow_oop_reg 3288 // decode_not_null narrow_oop_reg, base_reg 3289 // Load [base_reg + offset], val_reg 3290 // NullCheck base_reg 3291 // 3292 } else if (t->isa_oopptr()) { 3293 new_in2 = ConNode::make(t->make_narrowoop()); 3294 } else if (t->isa_klassptr()) { 3295 new_in2 = ConNode::make(t->make_narrowklass()); 3296 } 3297 } 3298 if (new_in2 != NULL) { 3299 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3300 n->subsume_by(cmpN, this); 3301 if (in1->outcnt() == 0) { 3302 in1->disconnect_inputs(this); 3303 } 3304 if (in2->outcnt() == 0) { 3305 in2->disconnect_inputs(this); 3306 } 3307 } 3308 } 3309 break; 3310 3311 case Op_DecodeN: 3312 case Op_DecodeNKlass: 3313 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3314 // DecodeN could be pinned when it can't be fold into 3315 // an address expression, see the code for Op_CastPP above. 3316 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3317 break; 3318 3319 case Op_EncodeP: 3320 case Op_EncodePKlass: { 3321 Node* in1 = n->in(1); 3322 if (in1->is_DecodeNarrowPtr()) { 3323 n->subsume_by(in1->in(1), this); 3324 } else if (in1->Opcode() == Op_ConP) { 3325 const Type* t = in1->bottom_type(); 3326 if (t == TypePtr::NULL_PTR) { 3327 assert(t->isa_oopptr(), "null klass?"); 3328 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3329 } else if (t->isa_oopptr()) { 3330 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3331 } else if (t->isa_klassptr()) { 3332 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3333 } 3334 } 3335 if (in1->outcnt() == 0) { 3336 in1->disconnect_inputs(this); 3337 } 3338 break; 3339 } 3340 3341 case Op_Proj: { 3342 if (OptimizeStringConcat || IncrementalInline) { 3343 ProjNode* proj = n->as_Proj(); 3344 if (proj->_is_io_use) { 3345 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, ""); 3346 // Separate projections were used for the exception path which 3347 // are normally removed by a late inline. If it wasn't inlined 3348 // then they will hang around and should just be replaced with 3349 // the original one. Merge them. 3350 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/); 3351 if (non_io_proj != NULL) { 3352 proj->subsume_by(non_io_proj , this); 3353 } 3354 } 3355 } 3356 break; 3357 } 3358 3359 case Op_Phi: 3360 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3361 // The EncodeP optimization may create Phi with the same edges 3362 // for all paths. It is not handled well by Register Allocator. 3363 Node* unique_in = n->in(1); 3364 assert(unique_in != NULL, ""); 3365 uint cnt = n->req(); 3366 for (uint i = 2; i < cnt; i++) { 3367 Node* m = n->in(i); 3368 assert(m != NULL, ""); 3369 if (unique_in != m) 3370 unique_in = NULL; 3371 } 3372 if (unique_in != NULL) { 3373 n->subsume_by(unique_in, this); 3374 } 3375 } 3376 break; 3377 3378 #endif 3379 3380 #ifdef ASSERT 3381 case Op_CastII: 3382 // Verify that all range check dependent CastII nodes were removed. 3383 if (n->isa_CastII()->has_range_check()) { 3384 n->dump(3); 3385 assert(false, "Range check dependent CastII node was not removed"); 3386 } 3387 break; 3388 #endif 3389 3390 case Op_ModI: 3391 if (UseDivMod) { 3392 // Check if a%b and a/b both exist 3393 Node* d = n->find_similar(Op_DivI); 3394 if (d) { 3395 // Replace them with a fused divmod if supported 3396 if (Matcher::has_match_rule(Op_DivModI)) { 3397 DivModINode* divmod = DivModINode::make(n); 3398 d->subsume_by(divmod->div_proj(), this); 3399 n->subsume_by(divmod->mod_proj(), this); 3400 } else { 3401 // replace a%b with a-((a/b)*b) 3402 Node* mult = new MulINode(d, d->in(2)); 3403 Node* sub = new SubINode(d->in(1), mult); 3404 n->subsume_by(sub, this); 3405 } 3406 } 3407 } 3408 break; 3409 3410 case Op_ModL: 3411 if (UseDivMod) { 3412 // Check if a%b and a/b both exist 3413 Node* d = n->find_similar(Op_DivL); 3414 if (d) { 3415 // Replace them with a fused divmod if supported 3416 if (Matcher::has_match_rule(Op_DivModL)) { 3417 DivModLNode* divmod = DivModLNode::make(n); 3418 d->subsume_by(divmod->div_proj(), this); 3419 n->subsume_by(divmod->mod_proj(), this); 3420 } else { 3421 // replace a%b with a-((a/b)*b) 3422 Node* mult = new MulLNode(d, d->in(2)); 3423 Node* sub = new SubLNode(d->in(1), mult); 3424 n->subsume_by(sub, this); 3425 } 3426 } 3427 } 3428 break; 3429 3430 case Op_LoadVector: 3431 case Op_StoreVector: 3432 case Op_LoadVectorGather: 3433 case Op_StoreVectorScatter: 3434 case Op_VectorCmpMasked: 3435 case Op_VectorMaskGen: 3436 case Op_LoadVectorMasked: 3437 case Op_StoreVectorMasked: 3438 break; 3439 3440 case Op_AddReductionVI: 3441 case Op_AddReductionVL: 3442 case Op_AddReductionVF: 3443 case Op_AddReductionVD: 3444 case Op_MulReductionVI: 3445 case Op_MulReductionVL: 3446 case Op_MulReductionVF: 3447 case Op_MulReductionVD: 3448 case Op_MinReductionV: 3449 case Op_MaxReductionV: 3450 case Op_AndReductionV: 3451 case Op_OrReductionV: 3452 case Op_XorReductionV: 3453 break; 3454 3455 case Op_PackB: 3456 case Op_PackS: 3457 case Op_PackI: 3458 case Op_PackF: 3459 case Op_PackL: 3460 case Op_PackD: 3461 if (n->req()-1 > 2) { 3462 // Replace many operand PackNodes with a binary tree for matching 3463 PackNode* p = (PackNode*) n; 3464 Node* btp = p->binary_tree_pack(1, n->req()); 3465 n->subsume_by(btp, this); 3466 } 3467 break; 3468 case Op_Loop: 3469 assert(!n->as_Loop()->is_transformed_long_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop"); 3470 case Op_CountedLoop: 3471 case Op_LongCountedLoop: 3472 case Op_OuterStripMinedLoop: 3473 if (n->as_Loop()->is_inner_loop()) { 3474 frc.inc_inner_loop_count(); 3475 } 3476 n->as_Loop()->verify_strip_mined(0); 3477 break; 3478 case Op_LShiftI: 3479 case Op_RShiftI: 3480 case Op_URShiftI: 3481 case Op_LShiftL: 3482 case Op_RShiftL: 3483 case Op_URShiftL: 3484 if (Matcher::need_masked_shift_count) { 3485 // The cpu's shift instructions don't restrict the count to the 3486 // lower 5/6 bits. We need to do the masking ourselves. 3487 Node* in2 = n->in(2); 3488 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3489 const TypeInt* t = in2->find_int_type(); 3490 if (t != NULL && t->is_con()) { 3491 juint shift = t->get_con(); 3492 if (shift > mask) { // Unsigned cmp 3493 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3494 } 3495 } else { 3496 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3497 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3498 n->set_req(2, shift); 3499 } 3500 } 3501 if (in2->outcnt() == 0) { // Remove dead node 3502 in2->disconnect_inputs(this); 3503 } 3504 } 3505 break; 3506 case Op_MemBarStoreStore: 3507 case Op_MemBarRelease: 3508 // Break the link with AllocateNode: it is no longer useful and 3509 // confuses register allocation. 3510 if (n->req() > MemBarNode::Precedent) { 3511 n->set_req(MemBarNode::Precedent, top()); 3512 } 3513 break; 3514 case Op_MemBarAcquire: { 3515 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3516 // At parse time, the trailing MemBarAcquire for a volatile load 3517 // is created with an edge to the load. After optimizations, 3518 // that input may be a chain of Phis. If those phis have no 3519 // other use, then the MemBarAcquire keeps them alive and 3520 // register allocation can be confused. 3521 ResourceMark rm; 3522 Unique_Node_List wq; 3523 wq.push(n->in(MemBarNode::Precedent)); 3524 n->set_req(MemBarNode::Precedent, top()); 3525 while (wq.size() > 0) { 3526 Node* m = wq.pop(); 3527 if (m->outcnt() == 0 && m != top()) { 3528 for (uint j = 0; j < m->req(); j++) { 3529 Node* in = m->in(j); 3530 if (in != NULL) { 3531 wq.push(in); 3532 } 3533 } 3534 m->disconnect_inputs(this); 3535 } 3536 } 3537 } 3538 break; 3539 } 3540 case Op_Blackhole: 3541 break; 3542 case Op_RangeCheck: { 3543 RangeCheckNode* rc = n->as_RangeCheck(); 3544 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3545 n->subsume_by(iff, this); 3546 frc._tests.push(iff); 3547 break; 3548 } 3549 case Op_ConvI2L: { 3550 if (!Matcher::convi2l_type_required) { 3551 // Code generation on some platforms doesn't need accurate 3552 // ConvI2L types. Widening the type can help remove redundant 3553 // address computations. 3554 n->as_Type()->set_type(TypeLong::INT); 3555 ResourceMark rm; 3556 Unique_Node_List wq; 3557 wq.push(n); 3558 for (uint next = 0; next < wq.size(); next++) { 3559 Node *m = wq.at(next); 3560 3561 for(;;) { 3562 // Loop over all nodes with identical inputs edges as m 3563 Node* k = m->find_similar(m->Opcode()); 3564 if (k == NULL) { 3565 break; 3566 } 3567 // Push their uses so we get a chance to remove node made 3568 // redundant 3569 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3570 Node* u = k->fast_out(i); 3571 if (u->Opcode() == Op_LShiftL || 3572 u->Opcode() == Op_AddL || 3573 u->Opcode() == Op_SubL || 3574 u->Opcode() == Op_AddP) { 3575 wq.push(u); 3576 } 3577 } 3578 // Replace all nodes with identical edges as m with m 3579 k->subsume_by(m, this); 3580 } 3581 } 3582 } 3583 break; 3584 } 3585 case Op_CmpUL: { 3586 if (!Matcher::has_match_rule(Op_CmpUL)) { 3587 // No support for unsigned long comparisons 3588 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3589 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3590 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3591 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3592 Node* andl = new AndLNode(orl, remove_sign_mask); 3593 Node* cmp = new CmpLNode(andl, n->in(2)); 3594 n->subsume_by(cmp, this); 3595 } 3596 break; 3597 } 3598 default: 3599 assert(!n->is_Call(), ""); 3600 assert(!n->is_Mem(), ""); 3601 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3602 break; 3603 } 3604 } 3605 3606 //------------------------------final_graph_reshaping_walk--------------------- 3607 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3608 // requires that the walk visits a node's inputs before visiting the node. 3609 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3610 Unique_Node_List sfpt; 3611 3612 frc._visited.set(root->_idx); // first, mark node as visited 3613 uint cnt = root->req(); 3614 Node *n = root; 3615 uint i = 0; 3616 while (true) { 3617 if (i < cnt) { 3618 // Place all non-visited non-null inputs onto stack 3619 Node* m = n->in(i); 3620 ++i; 3621 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3622 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3623 // compute worst case interpreter size in case of a deoptimization 3624 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3625 3626 sfpt.push(m); 3627 } 3628 cnt = m->req(); 3629 nstack.push(n, i); // put on stack parent and next input's index 3630 n = m; 3631 i = 0; 3632 } 3633 } else { 3634 // Now do post-visit work 3635 final_graph_reshaping_impl( n, frc ); 3636 if (nstack.is_empty()) 3637 break; // finished 3638 n = nstack.node(); // Get node from stack 3639 cnt = n->req(); 3640 i = nstack.index(); 3641 nstack.pop(); // Shift to the next node on stack 3642 } 3643 } 3644 3645 // Skip next transformation if compressed oops are not used. 3646 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3647 (!UseCompressedOops && !UseCompressedClassPointers)) 3648 return; 3649 3650 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3651 // It could be done for an uncommon traps or any safepoints/calls 3652 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3653 while (sfpt.size() > 0) { 3654 n = sfpt.pop(); 3655 JVMState *jvms = n->as_SafePoint()->jvms(); 3656 assert(jvms != NULL, "sanity"); 3657 int start = jvms->debug_start(); 3658 int end = n->req(); 3659 bool is_uncommon = (n->is_CallStaticJava() && 3660 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3661 for (int j = start; j < end; j++) { 3662 Node* in = n->in(j); 3663 if (in->is_DecodeNarrowPtr()) { 3664 bool safe_to_skip = true; 3665 if (!is_uncommon ) { 3666 // Is it safe to skip? 3667 for (uint i = 0; i < in->outcnt(); i++) { 3668 Node* u = in->raw_out(i); 3669 if (!u->is_SafePoint() || 3670 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3671 safe_to_skip = false; 3672 } 3673 } 3674 } 3675 if (safe_to_skip) { 3676 n->set_req(j, in->in(1)); 3677 } 3678 if (in->outcnt() == 0) { 3679 in->disconnect_inputs(this); 3680 } 3681 } 3682 } 3683 } 3684 } 3685 3686 //------------------------------final_graph_reshaping-------------------------- 3687 // Final Graph Reshaping. 3688 // 3689 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3690 // and not commoned up and forced early. Must come after regular 3691 // optimizations to avoid GVN undoing the cloning. Clone constant 3692 // inputs to Loop Phis; these will be split by the allocator anyways. 3693 // Remove Opaque nodes. 3694 // (2) Move last-uses by commutative operations to the left input to encourage 3695 // Intel update-in-place two-address operations and better register usage 3696 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3697 // calls canonicalizing them back. 3698 // (3) Count the number of double-precision FP ops, single-precision FP ops 3699 // and call sites. On Intel, we can get correct rounding either by 3700 // forcing singles to memory (requires extra stores and loads after each 3701 // FP bytecode) or we can set a rounding mode bit (requires setting and 3702 // clearing the mode bit around call sites). The mode bit is only used 3703 // if the relative frequency of single FP ops to calls is low enough. 3704 // This is a key transform for SPEC mpeg_audio. 3705 // (4) Detect infinite loops; blobs of code reachable from above but not 3706 // below. Several of the Code_Gen algorithms fail on such code shapes, 3707 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3708 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3709 // Detection is by looking for IfNodes where only 1 projection is 3710 // reachable from below or CatchNodes missing some targets. 3711 // (5) Assert for insane oop offsets in debug mode. 3712 3713 bool Compile::final_graph_reshaping() { 3714 // an infinite loop may have been eliminated by the optimizer, 3715 // in which case the graph will be empty. 3716 if (root()->req() == 1) { 3717 record_method_not_compilable("trivial infinite loop"); 3718 return true; 3719 } 3720 3721 // Expensive nodes have their control input set to prevent the GVN 3722 // from freely commoning them. There's no GVN beyond this point so 3723 // no need to keep the control input. We want the expensive nodes to 3724 // be freely moved to the least frequent code path by gcm. 3725 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3726 for (int i = 0; i < expensive_count(); i++) { 3727 _expensive_nodes.at(i)->set_req(0, NULL); 3728 } 3729 3730 Final_Reshape_Counts frc; 3731 3732 // Visit everybody reachable! 3733 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3734 Node_Stack nstack(live_nodes() >> 1); 3735 final_graph_reshaping_walk(nstack, root(), frc); 3736 3737 // Check for unreachable (from below) code (i.e., infinite loops). 3738 for( uint i = 0; i < frc._tests.size(); i++ ) { 3739 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3740 // Get number of CFG targets. 3741 // Note that PCTables include exception targets after calls. 3742 uint required_outcnt = n->required_outcnt(); 3743 if (n->outcnt() != required_outcnt) { 3744 // Check for a few special cases. Rethrow Nodes never take the 3745 // 'fall-thru' path, so expected kids is 1 less. 3746 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3747 if (n->in(0)->in(0)->is_Call()) { 3748 CallNode *call = n->in(0)->in(0)->as_Call(); 3749 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3750 required_outcnt--; // Rethrow always has 1 less kid 3751 } else if (call->req() > TypeFunc::Parms && 3752 call->is_CallDynamicJava()) { 3753 // Check for null receiver. In such case, the optimizer has 3754 // detected that the virtual call will always result in a null 3755 // pointer exception. The fall-through projection of this CatchNode 3756 // will not be populated. 3757 Node *arg0 = call->in(TypeFunc::Parms); 3758 if (arg0->is_Type() && 3759 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3760 required_outcnt--; 3761 } 3762 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3763 call->req() > TypeFunc::Parms+1 && 3764 call->is_CallStaticJava()) { 3765 // Check for negative array length. In such case, the optimizer has 3766 // detected that the allocation attempt will always result in an 3767 // exception. There is no fall-through projection of this CatchNode . 3768 Node *arg1 = call->in(TypeFunc::Parms+1); 3769 if (arg1->is_Type() && 3770 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3771 required_outcnt--; 3772 } 3773 } 3774 } 3775 } 3776 // Recheck with a better notion of 'required_outcnt' 3777 if (n->outcnt() != required_outcnt) { 3778 record_method_not_compilable("malformed control flow"); 3779 return true; // Not all targets reachable! 3780 } 3781 } 3782 // Check that I actually visited all kids. Unreached kids 3783 // must be infinite loops. 3784 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3785 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3786 record_method_not_compilable("infinite loop"); 3787 return true; // Found unvisited kid; must be unreach 3788 } 3789 3790 // Here so verification code in final_graph_reshaping_walk() 3791 // always see an OuterStripMinedLoopEnd 3792 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) { 3793 IfNode* init_iff = n->as_If(); 3794 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3795 n->subsume_by(iff, this); 3796 } 3797 } 3798 3799 #ifdef IA32 3800 // If original bytecodes contained a mixture of floats and doubles 3801 // check if the optimizer has made it homogenous, item (3). 3802 if (UseSSE == 0 && 3803 frc.get_float_count() > 32 && 3804 frc.get_double_count() == 0 && 3805 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3806 set_24_bit_selection_and_mode(false, true); 3807 } 3808 #endif // IA32 3809 3810 set_java_calls(frc.get_java_call_count()); 3811 set_inner_loops(frc.get_inner_loop_count()); 3812 3813 // No infinite loops, no reason to bail out. 3814 return false; 3815 } 3816 3817 //-----------------------------too_many_traps---------------------------------- 3818 // Report if there are too many traps at the current method and bci. 3819 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3820 bool Compile::too_many_traps(ciMethod* method, 3821 int bci, 3822 Deoptimization::DeoptReason reason) { 3823 ciMethodData* md = method->method_data(); 3824 if (md->is_empty()) { 3825 // Assume the trap has not occurred, or that it occurred only 3826 // because of a transient condition during start-up in the interpreter. 3827 return false; 3828 } 3829 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3830 if (md->has_trap_at(bci, m, reason) != 0) { 3831 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3832 // Also, if there are multiple reasons, or if there is no per-BCI record, 3833 // assume the worst. 3834 if (log()) 3835 log()->elem("observe trap='%s' count='%d'", 3836 Deoptimization::trap_reason_name(reason), 3837 md->trap_count(reason)); 3838 return true; 3839 } else { 3840 // Ignore method/bci and see if there have been too many globally. 3841 return too_many_traps(reason, md); 3842 } 3843 } 3844 3845 // Less-accurate variant which does not require a method and bci. 3846 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3847 ciMethodData* logmd) { 3848 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3849 // Too many traps globally. 3850 // Note that we use cumulative trap_count, not just md->trap_count. 3851 if (log()) { 3852 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3853 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3854 Deoptimization::trap_reason_name(reason), 3855 mcount, trap_count(reason)); 3856 } 3857 return true; 3858 } else { 3859 // The coast is clear. 3860 return false; 3861 } 3862 } 3863 3864 //--------------------------too_many_recompiles-------------------------------- 3865 // Report if there are too many recompiles at the current method and bci. 3866 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3867 // Is not eager to return true, since this will cause the compiler to use 3868 // Action_none for a trap point, to avoid too many recompilations. 3869 bool Compile::too_many_recompiles(ciMethod* method, 3870 int bci, 3871 Deoptimization::DeoptReason reason) { 3872 ciMethodData* md = method->method_data(); 3873 if (md->is_empty()) { 3874 // Assume the trap has not occurred, or that it occurred only 3875 // because of a transient condition during start-up in the interpreter. 3876 return false; 3877 } 3878 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3879 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3880 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3881 Deoptimization::DeoptReason per_bc_reason 3882 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3883 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3884 if ((per_bc_reason == Deoptimization::Reason_none 3885 || md->has_trap_at(bci, m, reason) != 0) 3886 // The trap frequency measure we care about is the recompile count: 3887 && md->trap_recompiled_at(bci, m) 3888 && md->overflow_recompile_count() >= bc_cutoff) { 3889 // Do not emit a trap here if it has already caused recompilations. 3890 // Also, if there are multiple reasons, or if there is no per-BCI record, 3891 // assume the worst. 3892 if (log()) 3893 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3894 Deoptimization::trap_reason_name(reason), 3895 md->trap_count(reason), 3896 md->overflow_recompile_count()); 3897 return true; 3898 } else if (trap_count(reason) != 0 3899 && decompile_count() >= m_cutoff) { 3900 // Too many recompiles globally, and we have seen this sort of trap. 3901 // Use cumulative decompile_count, not just md->decompile_count. 3902 if (log()) 3903 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3904 Deoptimization::trap_reason_name(reason), 3905 md->trap_count(reason), trap_count(reason), 3906 md->decompile_count(), decompile_count()); 3907 return true; 3908 } else { 3909 // The coast is clear. 3910 return false; 3911 } 3912 } 3913 3914 // Compute when not to trap. Used by matching trap based nodes and 3915 // NullCheck optimization. 3916 void Compile::set_allowed_deopt_reasons() { 3917 _allowed_reasons = 0; 3918 if (is_method_compilation()) { 3919 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3920 assert(rs < BitsPerInt, "recode bit map"); 3921 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3922 _allowed_reasons |= nth_bit(rs); 3923 } 3924 } 3925 } 3926 } 3927 3928 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 3929 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 3930 } 3931 3932 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 3933 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 3934 } 3935 3936 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 3937 if (holder->is_initialized()) { 3938 return false; 3939 } 3940 if (holder->is_being_initialized()) { 3941 if (accessing_method->holder() == holder) { 3942 // Access inside a class. The barrier can be elided when access happens in <clinit>, 3943 // <init>, or a static method. In all those cases, there was an initialization 3944 // barrier on the holder klass passed. 3945 if (accessing_method->is_static_initializer() || 3946 accessing_method->is_object_initializer() || 3947 accessing_method->is_static()) { 3948 return false; 3949 } 3950 } else if (accessing_method->holder()->is_subclass_of(holder)) { 3951 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 3952 // In case of <init> or a static method, the barrier is on the subclass is not enough: 3953 // child class can become fully initialized while its parent class is still being initialized. 3954 if (accessing_method->is_static_initializer()) { 3955 return false; 3956 } 3957 } 3958 ciMethod* root = method(); // the root method of compilation 3959 if (root != accessing_method) { 3960 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 3961 } 3962 } 3963 return true; 3964 } 3965 3966 #ifndef PRODUCT 3967 //------------------------------verify_graph_edges--------------------------- 3968 // Walk the Graph and verify that there is a one-to-one correspondence 3969 // between Use-Def edges and Def-Use edges in the graph. 3970 void Compile::verify_graph_edges(bool no_dead_code) { 3971 if (VerifyGraphEdges) { 3972 Unique_Node_List visited; 3973 // Call recursive graph walk to check edges 3974 _root->verify_edges(visited); 3975 if (no_dead_code) { 3976 // Now make sure that no visited node is used by an unvisited node. 3977 bool dead_nodes = false; 3978 Unique_Node_List checked; 3979 while (visited.size() > 0) { 3980 Node* n = visited.pop(); 3981 checked.push(n); 3982 for (uint i = 0; i < n->outcnt(); i++) { 3983 Node* use = n->raw_out(i); 3984 if (checked.member(use)) continue; // already checked 3985 if (visited.member(use)) continue; // already in the graph 3986 if (use->is_Con()) continue; // a dead ConNode is OK 3987 // At this point, we have found a dead node which is DU-reachable. 3988 if (!dead_nodes) { 3989 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3990 dead_nodes = true; 3991 } 3992 use->dump(2); 3993 tty->print_cr("---"); 3994 checked.push(use); // No repeats; pretend it is now checked. 3995 } 3996 } 3997 assert(!dead_nodes, "using nodes must be reachable from root"); 3998 } 3999 } 4000 } 4001 #endif 4002 4003 // The Compile object keeps track of failure reasons separately from the ciEnv. 4004 // This is required because there is not quite a 1-1 relation between the 4005 // ciEnv and its compilation task and the Compile object. Note that one 4006 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4007 // to backtrack and retry without subsuming loads. Other than this backtracking 4008 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4009 // by the logic in C2Compiler. 4010 void Compile::record_failure(const char* reason) { 4011 if (log() != NULL) { 4012 log()->elem("failure reason='%s' phase='compile'", reason); 4013 } 4014 if (_failure_reason == NULL) { 4015 // Record the first failure reason. 4016 _failure_reason = reason; 4017 } 4018 4019 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4020 C->print_method(PHASE_FAILURE); 4021 } 4022 _root = NULL; // flush the graph, too 4023 } 4024 4025 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 4026 : TraceTime(name, accumulator, CITime, CITimeVerbose), 4027 _phase_name(name), _dolog(CITimeVerbose) 4028 { 4029 if (_dolog) { 4030 C = Compile::current(); 4031 _log = C->log(); 4032 } else { 4033 C = NULL; 4034 _log = NULL; 4035 } 4036 if (_log != NULL) { 4037 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4038 _log->stamp(); 4039 _log->end_head(); 4040 } 4041 } 4042 4043 Compile::TracePhase::~TracePhase() { 4044 4045 C = Compile::current(); 4046 if (_dolog) { 4047 _log = C->log(); 4048 } else { 4049 _log = NULL; 4050 } 4051 4052 #ifdef ASSERT 4053 if (PrintIdealNodeCount) { 4054 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4055 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 4056 } 4057 4058 if (VerifyIdealNodeCount) { 4059 Compile::current()->print_missing_nodes(); 4060 } 4061 #endif 4062 4063 if (_log != NULL) { 4064 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4065 } 4066 } 4067 4068 //----------------------------static_subtype_check----------------------------- 4069 // Shortcut important common cases when superklass is exact: 4070 // (0) superklass is java.lang.Object (can occur in reflective code) 4071 // (1) subklass is already limited to a subtype of superklass => always ok 4072 // (2) subklass does not overlap with superklass => always fail 4073 // (3) superklass has NO subtypes and we can check with a simple compare. 4074 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4075 if (StressReflectiveCode) { 4076 return SSC_full_test; // Let caller generate the general case. 4077 } 4078 4079 if (superk == env()->Object_klass()) { 4080 return SSC_always_true; // (0) this test cannot fail 4081 } 4082 4083 ciType* superelem = superk; 4084 ciType* subelem = subk; 4085 if (superelem->is_array_klass()) { 4086 superelem = superelem->as_array_klass()->base_element_type(); 4087 } 4088 if (subelem->is_array_klass()) { 4089 subelem = subelem->as_array_klass()->base_element_type(); 4090 } 4091 4092 if (!subk->is_interface()) { // cannot trust static interface types yet 4093 if (subk->is_subtype_of(superk)) { 4094 return SSC_always_true; // (1) false path dead; no dynamic test needed 4095 } 4096 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4097 !(subelem->is_klass() && subelem->as_klass()->is_interface()) && 4098 !superk->is_subtype_of(subk)) { 4099 return SSC_always_false; // (2) true path dead; no dynamic test needed 4100 } 4101 } 4102 4103 // If casting to an instance klass, it must have no subtypes 4104 if (superk->is_interface()) { 4105 // Cannot trust interfaces yet. 4106 // %%% S.B. superk->nof_implementors() == 1 4107 } else if (superelem->is_instance_klass()) { 4108 ciInstanceKlass* ik = superelem->as_instance_klass(); 4109 if (!ik->has_subklass() && !ik->is_interface()) { 4110 if (!ik->is_final()) { 4111 // Add a dependency if there is a chance of a later subclass. 4112 dependencies()->assert_leaf_type(ik); 4113 } 4114 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4115 } 4116 } else { 4117 // A primitive array type has no subtypes. 4118 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4119 } 4120 4121 return SSC_full_test; 4122 } 4123 4124 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4125 #ifdef _LP64 4126 // The scaled index operand to AddP must be a clean 64-bit value. 4127 // Java allows a 32-bit int to be incremented to a negative 4128 // value, which appears in a 64-bit register as a large 4129 // positive number. Using that large positive number as an 4130 // operand in pointer arithmetic has bad consequences. 4131 // On the other hand, 32-bit overflow is rare, and the possibility 4132 // can often be excluded, if we annotate the ConvI2L node with 4133 // a type assertion that its value is known to be a small positive 4134 // number. (The prior range check has ensured this.) 4135 // This assertion is used by ConvI2LNode::Ideal. 4136 int index_max = max_jint - 1; // array size is max_jint, index is one less 4137 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4138 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4139 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4140 #endif 4141 return idx; 4142 } 4143 4144 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4145 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 4146 if (ctrl != NULL) { 4147 // Express control dependency by a CastII node with a narrow type. 4148 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */); 4149 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4150 // node from floating above the range check during loop optimizations. Otherwise, the 4151 // ConvI2L node may be eliminated independently of the range check, causing the data path 4152 // to become TOP while the control path is still there (although it's unreachable). 4153 value->set_req(0, ctrl); 4154 value = phase->transform(value); 4155 } 4156 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4157 return phase->transform(new ConvI2LNode(value, ltype)); 4158 } 4159 4160 void Compile::print_inlining_stream_free() { 4161 if (_print_inlining_stream != NULL) { 4162 _print_inlining_stream->~stringStream(); 4163 _print_inlining_stream = NULL; 4164 } 4165 } 4166 4167 // The message about the current inlining is accumulated in 4168 // _print_inlining_stream and transfered into the _print_inlining_list 4169 // once we know whether inlining succeeds or not. For regular 4170 // inlining, messages are appended to the buffer pointed by 4171 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4172 // a new buffer is added after _print_inlining_idx in the list. This 4173 // way we can update the inlining message for late inlining call site 4174 // when the inlining is attempted again. 4175 void Compile::print_inlining_init() { 4176 if (print_inlining() || print_intrinsics()) { 4177 // print_inlining_init is actually called several times. 4178 print_inlining_stream_free(); 4179 _print_inlining_stream = new stringStream(); 4180 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer()); 4181 } 4182 } 4183 4184 void Compile::print_inlining_reinit() { 4185 if (print_inlining() || print_intrinsics()) { 4186 print_inlining_stream_free(); 4187 // Re allocate buffer when we change ResourceMark 4188 _print_inlining_stream = new stringStream(); 4189 } 4190 } 4191 4192 void Compile::print_inlining_reset() { 4193 _print_inlining_stream->reset(); 4194 } 4195 4196 void Compile::print_inlining_commit() { 4197 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4198 // Transfer the message from _print_inlining_stream to the current 4199 // _print_inlining_list buffer and clear _print_inlining_stream. 4200 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 4201 print_inlining_reset(); 4202 } 4203 4204 void Compile::print_inlining_push() { 4205 // Add new buffer to the _print_inlining_list at current position 4206 _print_inlining_idx++; 4207 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer()); 4208 } 4209 4210 Compile::PrintInliningBuffer* Compile::print_inlining_current() { 4211 return _print_inlining_list->at(_print_inlining_idx); 4212 } 4213 4214 void Compile::print_inlining_update(CallGenerator* cg) { 4215 if (print_inlining() || print_intrinsics()) { 4216 if (cg->is_late_inline()) { 4217 if (print_inlining_current()->cg() != cg && 4218 (print_inlining_current()->cg() != NULL || 4219 print_inlining_current()->ss()->size() != 0)) { 4220 print_inlining_push(); 4221 } 4222 print_inlining_commit(); 4223 print_inlining_current()->set_cg(cg); 4224 } else { 4225 if (print_inlining_current()->cg() != NULL) { 4226 print_inlining_push(); 4227 } 4228 print_inlining_commit(); 4229 } 4230 } 4231 } 4232 4233 void Compile::print_inlining_move_to(CallGenerator* cg) { 4234 // We resume inlining at a late inlining call site. Locate the 4235 // corresponding inlining buffer so that we can update it. 4236 if (print_inlining() || print_intrinsics()) { 4237 for (int i = 0; i < _print_inlining_list->length(); i++) { 4238 if (_print_inlining_list->at(i)->cg() == cg) { 4239 _print_inlining_idx = i; 4240 return; 4241 } 4242 } 4243 ShouldNotReachHere(); 4244 } 4245 } 4246 4247 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4248 if (print_inlining() || print_intrinsics()) { 4249 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4250 assert(print_inlining_current()->cg() == cg, "wrong entry"); 4251 // replace message with new message 4252 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer()); 4253 print_inlining_commit(); 4254 print_inlining_current()->set_cg(cg); 4255 } 4256 } 4257 4258 void Compile::print_inlining_assert_ready() { 4259 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4260 } 4261 4262 void Compile::process_print_inlining() { 4263 assert(_late_inlines.length() == 0, "not drained yet"); 4264 if (print_inlining() || print_intrinsics()) { 4265 ResourceMark rm; 4266 stringStream ss; 4267 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once."); 4268 for (int i = 0; i < _print_inlining_list->length(); i++) { 4269 PrintInliningBuffer* pib = _print_inlining_list->at(i); 4270 ss.print("%s", pib->ss()->as_string()); 4271 delete pib; 4272 DEBUG_ONLY(_print_inlining_list->at_put(i, NULL)); 4273 } 4274 // Reset _print_inlining_list, it only contains destructed objects. 4275 // It is on the arena, so it will be freed when the arena is reset. 4276 _print_inlining_list = NULL; 4277 // _print_inlining_stream won't be used anymore, either. 4278 print_inlining_stream_free(); 4279 size_t end = ss.size(); 4280 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4281 strncpy(_print_inlining_output, ss.base(), end+1); 4282 _print_inlining_output[end] = 0; 4283 } 4284 } 4285 4286 void Compile::dump_print_inlining() { 4287 if (_print_inlining_output != NULL) { 4288 tty->print_raw(_print_inlining_output); 4289 } 4290 } 4291 4292 void Compile::log_late_inline(CallGenerator* cg) { 4293 if (log() != NULL) { 4294 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4295 cg->unique_id()); 4296 JVMState* p = cg->call_node()->jvms(); 4297 while (p != NULL) { 4298 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4299 p = p->caller(); 4300 } 4301 log()->tail("late_inline"); 4302 } 4303 } 4304 4305 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4306 log_late_inline(cg); 4307 if (log() != NULL) { 4308 log()->inline_fail(msg); 4309 } 4310 } 4311 4312 void Compile::log_inline_id(CallGenerator* cg) { 4313 if (log() != NULL) { 4314 // The LogCompilation tool needs a unique way to identify late 4315 // inline call sites. This id must be unique for this call site in 4316 // this compilation. Try to have it unique across compilations as 4317 // well because it can be convenient when grepping through the log 4318 // file. 4319 // Distinguish OSR compilations from others in case CICountOSR is 4320 // on. 4321 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4322 cg->set_unique_id(id); 4323 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4324 } 4325 } 4326 4327 void Compile::log_inline_failure(const char* msg) { 4328 if (C->log() != NULL) { 4329 C->log()->inline_fail(msg); 4330 } 4331 } 4332 4333 4334 // Dump inlining replay data to the stream. 4335 // Don't change thread state and acquire any locks. 4336 void Compile::dump_inline_data(outputStream* out) { 4337 InlineTree* inl_tree = ilt(); 4338 if (inl_tree != NULL) { 4339 out->print(" inline %d", inl_tree->count()); 4340 inl_tree->dump_replay_data(out); 4341 } 4342 } 4343 4344 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4345 if (n1->Opcode() < n2->Opcode()) return -1; 4346 else if (n1->Opcode() > n2->Opcode()) return 1; 4347 4348 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4349 for (uint i = 1; i < n1->req(); i++) { 4350 if (n1->in(i) < n2->in(i)) return -1; 4351 else if (n1->in(i) > n2->in(i)) return 1; 4352 } 4353 4354 return 0; 4355 } 4356 4357 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4358 Node* n1 = *n1p; 4359 Node* n2 = *n2p; 4360 4361 return cmp_expensive_nodes(n1, n2); 4362 } 4363 4364 void Compile::sort_expensive_nodes() { 4365 if (!expensive_nodes_sorted()) { 4366 _expensive_nodes.sort(cmp_expensive_nodes); 4367 } 4368 } 4369 4370 bool Compile::expensive_nodes_sorted() const { 4371 for (int i = 1; i < _expensive_nodes.length(); i++) { 4372 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) { 4373 return false; 4374 } 4375 } 4376 return true; 4377 } 4378 4379 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4380 if (_expensive_nodes.length() == 0) { 4381 return false; 4382 } 4383 4384 assert(OptimizeExpensiveOps, "optimization off?"); 4385 4386 // Take this opportunity to remove dead nodes from the list 4387 int j = 0; 4388 for (int i = 0; i < _expensive_nodes.length(); i++) { 4389 Node* n = _expensive_nodes.at(i); 4390 if (!n->is_unreachable(igvn)) { 4391 assert(n->is_expensive(), "should be expensive"); 4392 _expensive_nodes.at_put(j, n); 4393 j++; 4394 } 4395 } 4396 _expensive_nodes.trunc_to(j); 4397 4398 // Then sort the list so that similar nodes are next to each other 4399 // and check for at least two nodes of identical kind with same data 4400 // inputs. 4401 sort_expensive_nodes(); 4402 4403 for (int i = 0; i < _expensive_nodes.length()-1; i++) { 4404 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) { 4405 return true; 4406 } 4407 } 4408 4409 return false; 4410 } 4411 4412 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4413 if (_expensive_nodes.length() == 0) { 4414 return; 4415 } 4416 4417 assert(OptimizeExpensiveOps, "optimization off?"); 4418 4419 // Sort to bring similar nodes next to each other and clear the 4420 // control input of nodes for which there's only a single copy. 4421 sort_expensive_nodes(); 4422 4423 int j = 0; 4424 int identical = 0; 4425 int i = 0; 4426 bool modified = false; 4427 for (; i < _expensive_nodes.length()-1; i++) { 4428 assert(j <= i, "can't write beyond current index"); 4429 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) { 4430 identical++; 4431 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4432 continue; 4433 } 4434 if (identical > 0) { 4435 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4436 identical = 0; 4437 } else { 4438 Node* n = _expensive_nodes.at(i); 4439 igvn.replace_input_of(n, 0, NULL); 4440 igvn.hash_insert(n); 4441 modified = true; 4442 } 4443 } 4444 if (identical > 0) { 4445 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4446 } else if (_expensive_nodes.length() >= 1) { 4447 Node* n = _expensive_nodes.at(i); 4448 igvn.replace_input_of(n, 0, NULL); 4449 igvn.hash_insert(n); 4450 modified = true; 4451 } 4452 _expensive_nodes.trunc_to(j); 4453 if (modified) { 4454 igvn.optimize(); 4455 } 4456 } 4457 4458 void Compile::add_expensive_node(Node * n) { 4459 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list"); 4460 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4461 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4462 if (OptimizeExpensiveOps) { 4463 _expensive_nodes.append(n); 4464 } else { 4465 // Clear control input and let IGVN optimize expensive nodes if 4466 // OptimizeExpensiveOps is off. 4467 n->set_req(0, NULL); 4468 } 4469 } 4470 4471 /** 4472 * Track coarsened Lock and Unlock nodes. 4473 */ 4474 4475 class Lock_List : public Node_List { 4476 uint _origin_cnt; 4477 public: 4478 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {} 4479 uint origin_cnt() const { return _origin_cnt; } 4480 }; 4481 4482 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) { 4483 int length = locks.length(); 4484 if (length > 0) { 4485 // Have to keep this list until locks elimination during Macro nodes elimination. 4486 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length); 4487 for (int i = 0; i < length; i++) { 4488 AbstractLockNode* lock = locks.at(i); 4489 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx); 4490 locks_list->push(lock); 4491 } 4492 _coarsened_locks.append(locks_list); 4493 } 4494 } 4495 4496 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) { 4497 int count = coarsened_count(); 4498 for (int i = 0; i < count; i++) { 4499 Node_List* locks_list = _coarsened_locks.at(i); 4500 for (uint j = 0; j < locks_list->size(); j++) { 4501 Node* lock = locks_list->at(j); 4502 assert(lock->is_AbstractLock(), "sanity"); 4503 if (!useful.member(lock)) { 4504 locks_list->yank(lock); 4505 } 4506 } 4507 } 4508 } 4509 4510 void Compile::remove_coarsened_lock(Node* n) { 4511 if (n->is_AbstractLock()) { 4512 int count = coarsened_count(); 4513 for (int i = 0; i < count; i++) { 4514 Node_List* locks_list = _coarsened_locks.at(i); 4515 locks_list->yank(n); 4516 } 4517 } 4518 } 4519 4520 bool Compile::coarsened_locks_consistent() { 4521 int count = coarsened_count(); 4522 for (int i = 0; i < count; i++) { 4523 bool unbalanced = false; 4524 bool modified = false; // track locks kind modifications 4525 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i); 4526 uint size = locks_list->size(); 4527 if (size == 0) { 4528 unbalanced = false; // All locks were eliminated - good 4529 } else if (size != locks_list->origin_cnt()) { 4530 unbalanced = true; // Some locks were removed from list 4531 } else { 4532 for (uint j = 0; j < size; j++) { 4533 Node* lock = locks_list->at(j); 4534 // All nodes in group should have the same state (modified or not) 4535 if (!lock->as_AbstractLock()->is_coarsened()) { 4536 if (j == 0) { 4537 // first on list was modified, the rest should be too for consistency 4538 modified = true; 4539 } else if (!modified) { 4540 // this lock was modified but previous locks on the list were not 4541 unbalanced = true; 4542 break; 4543 } 4544 } else if (modified) { 4545 // previous locks on list were modified but not this lock 4546 unbalanced = true; 4547 break; 4548 } 4549 } 4550 } 4551 if (unbalanced) { 4552 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified 4553 #ifdef ASSERT 4554 if (PrintEliminateLocks) { 4555 tty->print_cr("=== unbalanced coarsened locks ==="); 4556 for (uint l = 0; l < size; l++) { 4557 locks_list->at(l)->dump(); 4558 } 4559 } 4560 #endif 4561 record_failure(C2Compiler::retry_no_locks_coarsening()); 4562 return false; 4563 } 4564 } 4565 return true; 4566 } 4567 4568 /** 4569 * Remove the speculative part of types and clean up the graph 4570 */ 4571 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4572 if (UseTypeSpeculation) { 4573 Unique_Node_List worklist; 4574 worklist.push(root()); 4575 int modified = 0; 4576 // Go over all type nodes that carry a speculative type, drop the 4577 // speculative part of the type and enqueue the node for an igvn 4578 // which may optimize it out. 4579 for (uint next = 0; next < worklist.size(); ++next) { 4580 Node *n = worklist.at(next); 4581 if (n->is_Type()) { 4582 TypeNode* tn = n->as_Type(); 4583 const Type* t = tn->type(); 4584 const Type* t_no_spec = t->remove_speculative(); 4585 if (t_no_spec != t) { 4586 bool in_hash = igvn.hash_delete(n); 4587 assert(in_hash, "node should be in igvn hash table"); 4588 tn->set_type(t_no_spec); 4589 igvn.hash_insert(n); 4590 igvn._worklist.push(n); // give it a chance to go away 4591 modified++; 4592 } 4593 } 4594 // Iterate over outs - endless loops is unreachable from below 4595 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4596 Node *m = n->fast_out(i); 4597 if (not_a_node(m)) { 4598 continue; 4599 } 4600 worklist.push(m); 4601 } 4602 } 4603 // Drop the speculative part of all types in the igvn's type table 4604 igvn.remove_speculative_types(); 4605 if (modified > 0) { 4606 igvn.optimize(); 4607 } 4608 #ifdef ASSERT 4609 // Verify that after the IGVN is over no speculative type has resurfaced 4610 worklist.clear(); 4611 worklist.push(root()); 4612 for (uint next = 0; next < worklist.size(); ++next) { 4613 Node *n = worklist.at(next); 4614 const Type* t = igvn.type_or_null(n); 4615 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4616 if (n->is_Type()) { 4617 t = n->as_Type()->type(); 4618 assert(t == t->remove_speculative(), "no more speculative types"); 4619 } 4620 // Iterate over outs - endless loops is unreachable from below 4621 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4622 Node *m = n->fast_out(i); 4623 if (not_a_node(m)) { 4624 continue; 4625 } 4626 worklist.push(m); 4627 } 4628 } 4629 igvn.check_no_speculative_types(); 4630 #endif 4631 } 4632 } 4633 4634 // Auxiliary methods to support randomized stressing/fuzzing. 4635 4636 int Compile::random() { 4637 _stress_seed = os::next_random(_stress_seed); 4638 return static_cast<int>(_stress_seed); 4639 } 4640 4641 // This method can be called the arbitrary number of times, with current count 4642 // as the argument. The logic allows selecting a single candidate from the 4643 // running list of candidates as follows: 4644 // int count = 0; 4645 // Cand* selected = null; 4646 // while(cand = cand->next()) { 4647 // if (randomized_select(++count)) { 4648 // selected = cand; 4649 // } 4650 // } 4651 // 4652 // Including count equalizes the chances any candidate is "selected". 4653 // This is useful when we don't have the complete list of candidates to choose 4654 // from uniformly. In this case, we need to adjust the randomicity of the 4655 // selection, or else we will end up biasing the selection towards the latter 4656 // candidates. 4657 // 4658 // Quick back-envelope calculation shows that for the list of n candidates 4659 // the equal probability for the candidate to persist as "best" can be 4660 // achieved by replacing it with "next" k-th candidate with the probability 4661 // of 1/k. It can be easily shown that by the end of the run, the 4662 // probability for any candidate is converged to 1/n, thus giving the 4663 // uniform distribution among all the candidates. 4664 // 4665 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4666 #define RANDOMIZED_DOMAIN_POW 29 4667 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4668 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4669 bool Compile::randomized_select(int count) { 4670 assert(count > 0, "only positive"); 4671 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4672 } 4673 4674 CloneMap& Compile::clone_map() { return _clone_map; } 4675 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4676 4677 void NodeCloneInfo::dump() const { 4678 tty->print(" {%d:%d} ", idx(), gen()); 4679 } 4680 4681 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4682 uint64_t val = value(old->_idx); 4683 NodeCloneInfo cio(val); 4684 assert(val != 0, "old node should be in the map"); 4685 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4686 insert(nnn->_idx, cin.get()); 4687 #ifndef PRODUCT 4688 if (is_debug()) { 4689 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4690 } 4691 #endif 4692 } 4693 4694 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4695 NodeCloneInfo cio(value(old->_idx)); 4696 if (cio.get() == 0) { 4697 cio.set(old->_idx, 0); 4698 insert(old->_idx, cio.get()); 4699 #ifndef PRODUCT 4700 if (is_debug()) { 4701 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4702 } 4703 #endif 4704 } 4705 clone(old, nnn, gen); 4706 } 4707 4708 int CloneMap::max_gen() const { 4709 int g = 0; 4710 DictI di(_dict); 4711 for(; di.test(); ++di) { 4712 int t = gen(di._key); 4713 if (g < t) { 4714 g = t; 4715 #ifndef PRODUCT 4716 if (is_debug()) { 4717 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4718 } 4719 #endif 4720 } 4721 } 4722 return g; 4723 } 4724 4725 void CloneMap::dump(node_idx_t key) const { 4726 uint64_t val = value(key); 4727 if (val != 0) { 4728 NodeCloneInfo ni(val); 4729 ni.dump(); 4730 } 4731 } 4732 4733 // Move Allocate nodes to the start of the list 4734 void Compile::sort_macro_nodes() { 4735 int count = macro_count(); 4736 int allocates = 0; 4737 for (int i = 0; i < count; i++) { 4738 Node* n = macro_node(i); 4739 if (n->is_Allocate()) { 4740 if (i != allocates) { 4741 Node* tmp = macro_node(allocates); 4742 _macro_nodes.at_put(allocates, n); 4743 _macro_nodes.at_put(i, tmp); 4744 } 4745 allocates++; 4746 } 4747 } 4748 } 4749 4750 void Compile::print_method(CompilerPhaseType cpt, const char *name, int level) { 4751 EventCompilerPhase event; 4752 if (event.should_commit()) { 4753 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 4754 } 4755 #ifndef PRODUCT 4756 if (should_print(level)) { 4757 _printer->print_method(name, level); 4758 } 4759 #endif 4760 C->_latest_stage_start_counter.stamp(); 4761 } 4762 4763 void Compile::print_method(CompilerPhaseType cpt, int level, int idx) { 4764 char output[1024]; 4765 #ifndef PRODUCT 4766 if (idx != 0) { 4767 jio_snprintf(output, sizeof(output), "%s:%d", CompilerPhaseTypeHelper::to_string(cpt), idx); 4768 } else { 4769 jio_snprintf(output, sizeof(output), "%s", CompilerPhaseTypeHelper::to_string(cpt)); 4770 } 4771 #endif 4772 print_method(cpt, output, level); 4773 } 4774 4775 void Compile::print_method(CompilerPhaseType cpt, Node* n, int level) { 4776 ResourceMark rm; 4777 stringStream ss; 4778 ss.print_raw(CompilerPhaseTypeHelper::to_string(cpt)); 4779 if (n != NULL) { 4780 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]); 4781 } else { 4782 ss.print_raw(": NULL"); 4783 } 4784 C->print_method(cpt, ss.as_string(), level); 4785 } 4786 4787 void Compile::end_method(int level) { 4788 EventCompilerPhase event; 4789 if (event.should_commit()) { 4790 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, level); 4791 } 4792 4793 #ifndef PRODUCT 4794 if (_method != NULL && should_print(level)) { 4795 _printer->end_method(); 4796 } 4797 #endif 4798 } 4799 4800 4801 #ifndef PRODUCT 4802 IdealGraphPrinter* Compile::_debug_file_printer = NULL; 4803 IdealGraphPrinter* Compile::_debug_network_printer = NULL; 4804 4805 // Called from debugger. Prints method to the default file with the default phase name. 4806 // This works regardless of any Ideal Graph Visualizer flags set or not. 4807 void igv_print() { 4808 Compile::current()->igv_print_method_to_file(); 4809 } 4810 4811 // Same as igv_print() above but with a specified phase name. 4812 void igv_print(const char* phase_name) { 4813 Compile::current()->igv_print_method_to_file(phase_name); 4814 } 4815 4816 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 4817 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 4818 // This works regardless of any Ideal Graph Visualizer flags set or not. 4819 void igv_print(bool network) { 4820 if (network) { 4821 Compile::current()->igv_print_method_to_network(); 4822 } else { 4823 Compile::current()->igv_print_method_to_file(); 4824 } 4825 } 4826 4827 // Same as igv_print(bool network) above but with a specified phase name. 4828 void igv_print(bool network, const char* phase_name) { 4829 if (network) { 4830 Compile::current()->igv_print_method_to_network(phase_name); 4831 } else { 4832 Compile::current()->igv_print_method_to_file(phase_name); 4833 } 4834 } 4835 4836 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 4837 void igv_print_default() { 4838 Compile::current()->print_method(PHASE_DEBUG, 0); 4839 } 4840 4841 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 4842 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 4843 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 4844 void igv_append() { 4845 Compile::current()->igv_print_method_to_file("Debug", true); 4846 } 4847 4848 // Same as igv_append() above but with a specified phase name. 4849 void igv_append(const char* phase_name) { 4850 Compile::current()->igv_print_method_to_file(phase_name, true); 4851 } 4852 4853 void Compile::igv_print_method_to_file(const char* phase_name, bool append) { 4854 const char* file_name = "custom_debug.xml"; 4855 if (_debug_file_printer == NULL) { 4856 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 4857 } else { 4858 _debug_file_printer->update_compiled_method(C->method()); 4859 } 4860 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 4861 _debug_file_printer->print(phase_name, (Node*)C->root()); 4862 } 4863 4864 void Compile::igv_print_method_to_network(const char* phase_name) { 4865 if (_debug_network_printer == NULL) { 4866 _debug_network_printer = new IdealGraphPrinter(C); 4867 } else { 4868 _debug_network_printer->update_compiled_method(C->method()); 4869 } 4870 tty->print_cr("Method printed over network stream to IGV"); 4871 _debug_network_printer->print(phase_name, (Node*)C->root()); 4872 } 4873 #endif 4874 4875 void Compile::add_native_invoker(RuntimeStub* stub) { 4876 _native_invokers.append(stub); 4877 } 4878 4879 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 4880 if (type != NULL && phase->type(value)->higher_equal(type)) { 4881 return value; 4882 } 4883 Node* result = NULL; 4884 if (bt == T_BYTE) { 4885 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 4886 result = new RShiftINode(result, phase->intcon(24)); 4887 } else if (bt == T_BOOLEAN) { 4888 result = new AndINode(value, phase->intcon(0xFF)); 4889 } else if (bt == T_CHAR) { 4890 result = new AndINode(value,phase->intcon(0xFFFF)); 4891 } else { 4892 assert(bt == T_SHORT, "unexpected narrow type"); 4893 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 4894 result = new RShiftINode(result, phase->intcon(16)); 4895 } 4896 if (transform_res) { 4897 result = phase->transform(result); 4898 } 4899 return result; 4900 } 4901