1 /* 2 * Copyright (c) 1997, 2023, 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 #ifdef ASSERT 957 _type_verify_symmetry = true; 958 _phase_optimize_finished = false; 959 _exception_backedge = false; 960 #endif 961 962 // Globally visible Nodes 963 // First set TOP to NULL to give safe behavior during creation of RootNode 964 set_cached_top_node(NULL); 965 set_root(new RootNode()); 966 // Now that you have a Root to point to, create the real TOP 967 set_cached_top_node( new ConNode(Type::TOP) ); 968 set_recent_alloc(NULL, NULL); 969 970 // Create Debug Information Recorder to record scopes, oopmaps, etc. 971 env()->set_oop_recorder(new OopRecorder(env()->arena())); 972 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 973 env()->set_dependencies(new Dependencies(env())); 974 975 _fixed_slots = 0; 976 set_has_split_ifs(false); 977 set_has_loops(false); // first approximation 978 set_has_stringbuilder(false); 979 set_has_boxed_value(false); 980 _trap_can_recompile = false; // no traps emitted yet 981 _major_progress = true; // start out assuming good things will happen 982 set_has_unsafe_access(false); 983 set_max_vector_size(0); 984 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 985 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 986 set_decompile_count(0); 987 988 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 989 _loop_opts_cnt = LoopOptsCount; 990 set_do_inlining(Inline); 991 set_max_inline_size(MaxInlineSize); 992 set_freq_inline_size(FreqInlineSize); 993 set_do_scheduling(OptoScheduling); 994 995 set_do_vector_loop(false); 996 997 if (AllowVectorizeOnDemand) { 998 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 999 set_do_vector_loop(true); 1000 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());}) 1001 } else if (has_method() && method()->name() != 0 && 1002 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1003 set_do_vector_loop(true); 1004 } 1005 } 1006 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1007 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());}) 1008 1009 set_age_code(has_method() && method()->profile_aging()); 1010 set_rtm_state(NoRTM); // No RTM lock eliding by default 1011 _max_node_limit = _directive->MaxNodeLimitOption; 1012 1013 #if INCLUDE_RTM_OPT 1014 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1015 int rtm_state = method()->method_data()->rtm_state(); 1016 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) { 1017 // Don't generate RTM lock eliding code. 1018 set_rtm_state(NoRTM); 1019 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1020 // Generate RTM lock eliding code without abort ratio calculation code. 1021 set_rtm_state(UseRTM); 1022 } else if (UseRTMDeopt) { 1023 // Generate RTM lock eliding code and include abort ratio calculation 1024 // code if UseRTMDeopt is on. 1025 set_rtm_state(ProfileRTM); 1026 } 1027 } 1028 #endif 1029 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1030 set_clinit_barrier_on_entry(true); 1031 } 1032 if (debug_info()->recording_non_safepoints()) { 1033 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1034 (comp_arena(), 8, 0, NULL)); 1035 set_default_node_notes(Node_Notes::make(this)); 1036 } 1037 1038 // // -- Initialize types before each compile -- 1039 // // Update cached type information 1040 // if( _method && _method->constants() ) 1041 // Type::update_loaded_types(_method, _method->constants()); 1042 1043 // Init alias_type map. 1044 if (!_do_escape_analysis && aliaslevel == 3) 1045 aliaslevel = 2; // No unique types without escape analysis 1046 _AliasLevel = aliaslevel; 1047 const int grow_ats = 16; 1048 _max_alias_types = grow_ats; 1049 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1050 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1051 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1052 { 1053 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1054 } 1055 // Initialize the first few types. 1056 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1057 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1058 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1059 _num_alias_types = AliasIdxRaw+1; 1060 // Zero out the alias type cache. 1061 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1062 // A NULL adr_type hits in the cache right away. Preload the right answer. 1063 probe_alias_cache(NULL)->_index = AliasIdxTop; 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 (UseCompactObjectHeaders) { 1651 if (flat->offset() == in_bytes(Klass::prototype_header_offset())) 1652 alias_type(idx)->set_rewritable(false); 1653 } 1654 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1655 alias_type(idx)->set_rewritable(false); 1656 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1657 alias_type(idx)->set_rewritable(false); 1658 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1659 alias_type(idx)->set_rewritable(false); 1660 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1661 alias_type(idx)->set_rewritable(false); 1662 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1663 alias_type(idx)->set_rewritable(false); 1664 } 1665 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1666 // but the base pointer type is not distinctive enough to identify 1667 // references into JavaThread.) 1668 1669 // Check for final fields. 1670 const TypeInstPtr* tinst = flat->isa_instptr(); 1671 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1672 ciField* field; 1673 if (tinst->const_oop() != NULL && 1674 tinst->klass() == ciEnv::current()->Class_klass() && 1675 tinst->offset() >= (tinst->klass()->as_instance_klass()->layout_helper_size_in_bytes())) { 1676 // static field 1677 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1678 field = k->get_field_by_offset(tinst->offset(), true); 1679 } else { 1680 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1681 field = k->get_field_by_offset(tinst->offset(), false); 1682 } 1683 assert(field == NULL || 1684 original_field == NULL || 1685 (field->holder() == original_field->holder() && 1686 field->offset() == original_field->offset() && 1687 field->is_static() == original_field->is_static()), "wrong field?"); 1688 // Set field() and is_rewritable() attributes. 1689 if (field != NULL) alias_type(idx)->set_field(field); 1690 } 1691 } 1692 1693 // Fill the cache for next time. 1694 ace->_adr_type = adr_type; 1695 ace->_index = idx; 1696 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1697 1698 // Might as well try to fill the cache for the flattened version, too. 1699 AliasCacheEntry* face = probe_alias_cache(flat); 1700 if (face->_adr_type == NULL) { 1701 face->_adr_type = flat; 1702 face->_index = idx; 1703 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1704 } 1705 1706 return alias_type(idx); 1707 } 1708 1709 1710 Compile::AliasType* Compile::alias_type(ciField* field) { 1711 const TypeOopPtr* t; 1712 if (field->is_static()) 1713 t = TypeInstPtr::make(field->holder()->java_mirror()); 1714 else 1715 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1716 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1717 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1718 return atp; 1719 } 1720 1721 1722 //------------------------------have_alias_type-------------------------------- 1723 bool Compile::have_alias_type(const TypePtr* adr_type) { 1724 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1725 if (ace->_adr_type == adr_type) { 1726 return true; 1727 } 1728 1729 // Handle special cases. 1730 if (adr_type == NULL) return true; 1731 if (adr_type == TypePtr::BOTTOM) return true; 1732 1733 return find_alias_type(adr_type, true, NULL) != NULL; 1734 } 1735 1736 //-----------------------------must_alias-------------------------------------- 1737 // True if all values of the given address type are in the given alias category. 1738 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1739 if (alias_idx == AliasIdxBot) return true; // the universal category 1740 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1741 if (alias_idx == AliasIdxTop) return false; // the empty category 1742 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1743 1744 // the only remaining possible overlap is identity 1745 int adr_idx = get_alias_index(adr_type); 1746 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1747 assert(adr_idx == alias_idx || 1748 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1749 && adr_type != TypeOopPtr::BOTTOM), 1750 "should not be testing for overlap with an unsafe pointer"); 1751 return adr_idx == alias_idx; 1752 } 1753 1754 //------------------------------can_alias-------------------------------------- 1755 // True if any values of the given address type are in the given alias category. 1756 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1757 if (alias_idx == AliasIdxTop) return false; // the empty category 1758 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1759 // Known instance doesn't alias with bottom memory 1760 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1761 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1762 1763 // the only remaining possible overlap is identity 1764 int adr_idx = get_alias_index(adr_type); 1765 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1766 return adr_idx == alias_idx; 1767 } 1768 1769 //---------------------cleanup_loop_predicates----------------------- 1770 // Remove the opaque nodes that protect the predicates so that all unused 1771 // checks and uncommon_traps will be eliminated from the ideal graph 1772 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1773 if (predicate_count()==0) return; 1774 for (int i = predicate_count(); i > 0; i--) { 1775 Node * n = predicate_opaque1_node(i-1); 1776 assert(n->Opcode() == Op_Opaque1, "must be"); 1777 igvn.replace_node(n, n->in(1)); 1778 } 1779 assert(predicate_count()==0, "should be clean!"); 1780 } 1781 1782 void Compile::record_for_post_loop_opts_igvn(Node* n) { 1783 if (!n->for_post_loop_opts_igvn()) { 1784 assert(!_for_post_loop_igvn.contains(n), "duplicate"); 1785 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1786 _for_post_loop_igvn.append(n); 1787 } 1788 } 1789 1790 void Compile::remove_from_post_loop_opts_igvn(Node* n) { 1791 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1792 _for_post_loop_igvn.remove(n); 1793 } 1794 1795 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) { 1796 // Verify that all previous optimizations produced a valid graph 1797 // at least to this point, even if no loop optimizations were done. 1798 PhaseIdealLoop::verify(igvn); 1799 1800 C->set_post_loop_opts_phase(); // no more loop opts allowed 1801 1802 assert(!C->major_progress(), "not cleared"); 1803 1804 if (_for_post_loop_igvn.length() > 0) { 1805 while (_for_post_loop_igvn.length() > 0) { 1806 Node* n = _for_post_loop_igvn.pop(); 1807 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1808 igvn._worklist.push(n); 1809 } 1810 igvn.optimize(); 1811 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed"); 1812 1813 // Sometimes IGVN sets major progress (e.g., when processing loop nodes). 1814 if (C->major_progress()) { 1815 C->clear_major_progress(); // ensure that major progress is now clear 1816 } 1817 } 1818 } 1819 1820 // StringOpts and late inlining of string methods 1821 void Compile::inline_string_calls(bool parse_time) { 1822 { 1823 // remove useless nodes to make the usage analysis simpler 1824 ResourceMark rm; 1825 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1826 } 1827 1828 { 1829 ResourceMark rm; 1830 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1831 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1832 print_method(PHASE_AFTER_STRINGOPTS, 3); 1833 } 1834 1835 // now inline anything that we skipped the first time around 1836 if (!parse_time) { 1837 _late_inlines_pos = _late_inlines.length(); 1838 } 1839 1840 while (_string_late_inlines.length() > 0) { 1841 CallGenerator* cg = _string_late_inlines.pop(); 1842 cg->do_late_inline(); 1843 if (failing()) return; 1844 } 1845 _string_late_inlines.trunc_to(0); 1846 } 1847 1848 // Late inlining of boxing methods 1849 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1850 if (_boxing_late_inlines.length() > 0) { 1851 assert(has_boxed_value(), "inconsistent"); 1852 1853 PhaseGVN* gvn = initial_gvn(); 1854 set_inlining_incrementally(true); 1855 1856 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1857 for_igvn()->clear(); 1858 gvn->replace_with(&igvn); 1859 1860 _late_inlines_pos = _late_inlines.length(); 1861 1862 while (_boxing_late_inlines.length() > 0) { 1863 CallGenerator* cg = _boxing_late_inlines.pop(); 1864 cg->do_late_inline(); 1865 if (failing()) return; 1866 } 1867 _boxing_late_inlines.trunc_to(0); 1868 1869 inline_incrementally_cleanup(igvn); 1870 1871 set_inlining_incrementally(false); 1872 } 1873 } 1874 1875 bool Compile::inline_incrementally_one() { 1876 assert(IncrementalInline, "incremental inlining should be on"); 1877 1878 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 1879 1880 set_inlining_progress(false); 1881 set_do_cleanup(false); 1882 1883 for (int i = 0; i < _late_inlines.length(); i++) { 1884 _late_inlines_pos = i+1; 1885 CallGenerator* cg = _late_inlines.at(i); 1886 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline(); 1887 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call 1888 cg->do_late_inline(); 1889 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed"); 1890 if (failing()) { 1891 return false; 1892 } else if (inlining_progress()) { 1893 _late_inlines_pos = i+1; // restore the position in case new elements were inserted 1894 print_method(PHASE_INCREMENTAL_INLINE_STEP, cg->call_node(), 3); 1895 break; // process one call site at a time 1896 } 1897 } else { 1898 // Ignore late inline direct calls when inlining is not allowed. 1899 // They are left in the late inline list when node budget is exhausted until the list is fully drained. 1900 } 1901 } 1902 // Remove processed elements. 1903 _late_inlines.remove_till(_late_inlines_pos); 1904 _late_inlines_pos = 0; 1905 1906 assert(inlining_progress() || _late_inlines.length() == 0, "no progress"); 1907 1908 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 1909 1910 set_inlining_progress(false); 1911 set_do_cleanup(false); 1912 1913 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption; 1914 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup; 1915 } 1916 1917 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 1918 { 1919 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 1920 ResourceMark rm; 1921 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1922 } 1923 { 1924 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 1925 igvn = PhaseIterGVN(initial_gvn()); 1926 igvn.optimize(); 1927 } 1928 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3); 1929 } 1930 1931 // Perform incremental inlining until bound on number of live nodes is reached 1932 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1933 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 1934 1935 set_inlining_incrementally(true); 1936 uint low_live_nodes = 0; 1937 1938 while (_late_inlines.length() > 0) { 1939 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1940 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1941 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 1942 // PhaseIdealLoop is expensive so we only try it once we are 1943 // out of live nodes and we only try it again if the previous 1944 // helped got the number of nodes down significantly 1945 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 1946 if (failing()) return; 1947 low_live_nodes = live_nodes(); 1948 _major_progress = true; 1949 } 1950 1951 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1952 bool do_print_inlining = print_inlining() || print_intrinsics(); 1953 if (do_print_inlining || log() != NULL) { 1954 // Print inlining message for candidates that we couldn't inline for lack of space. 1955 for (int i = 0; i < _late_inlines.length(); i++) { 1956 CallGenerator* cg = _late_inlines.at(i); 1957 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 1958 if (do_print_inlining) { 1959 cg->print_inlining_late(msg); 1960 } 1961 log_late_inline_failure(cg, msg); 1962 } 1963 } 1964 break; // finish 1965 } 1966 } 1967 1968 for_igvn()->clear(); 1969 initial_gvn()->replace_with(&igvn); 1970 1971 while (inline_incrementally_one()) { 1972 assert(!failing(), "inconsistent"); 1973 } 1974 if (failing()) return; 1975 1976 inline_incrementally_cleanup(igvn); 1977 1978 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 1979 1980 if (failing()) return; 1981 1982 if (_late_inlines.length() == 0) { 1983 break; // no more progress 1984 } 1985 } 1986 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1987 1988 if (_string_late_inlines.length() > 0) { 1989 assert(has_stringbuilder(), "inconsistent"); 1990 for_igvn()->clear(); 1991 initial_gvn()->replace_with(&igvn); 1992 1993 inline_string_calls(false); 1994 1995 if (failing()) return; 1996 1997 inline_incrementally_cleanup(igvn); 1998 } 1999 2000 set_inlining_incrementally(false); 2001 } 2002 2003 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) { 2004 // "inlining_incrementally() == false" is used to signal that no inlining is allowed 2005 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details). 2006 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == NULL" 2007 // as if "inlining_incrementally() == true" were set. 2008 assert(inlining_incrementally() == false, "not allowed"); 2009 assert(_modified_nodes == NULL, "not allowed"); 2010 assert(_late_inlines.length() > 0, "sanity"); 2011 2012 while (_late_inlines.length() > 0) { 2013 for_igvn()->clear(); 2014 initial_gvn()->replace_with(&igvn); 2015 2016 while (inline_incrementally_one()) { 2017 assert(!failing(), "inconsistent"); 2018 } 2019 if (failing()) return; 2020 2021 inline_incrementally_cleanup(igvn); 2022 } 2023 } 2024 2025 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2026 if (_loop_opts_cnt > 0) { 2027 while (major_progress() && (_loop_opts_cnt > 0)) { 2028 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2029 PhaseIdealLoop::optimize(igvn, mode); 2030 _loop_opts_cnt--; 2031 if (failing()) return false; 2032 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2033 } 2034 } 2035 return true; 2036 } 2037 2038 // Remove edges from "root" to each SafePoint at a backward branch. 2039 // They were inserted during parsing (see add_safepoint()) to make 2040 // infinite loops without calls or exceptions visible to root, i.e., 2041 // useful. 2042 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2043 Node *r = root(); 2044 if (r != NULL) { 2045 for (uint i = r->req(); i < r->len(); ++i) { 2046 Node *n = r->in(i); 2047 if (n != NULL && n->is_SafePoint()) { 2048 r->rm_prec(i); 2049 if (n->outcnt() == 0) { 2050 igvn.remove_dead_node(n); 2051 } 2052 --i; 2053 } 2054 } 2055 // Parsing may have added top inputs to the root node (Path 2056 // leading to the Halt node proven dead). Make sure we get a 2057 // chance to clean them up. 2058 igvn._worklist.push(r); 2059 igvn.optimize(); 2060 } 2061 } 2062 2063 //------------------------------Optimize--------------------------------------- 2064 // Given a graph, optimize it. 2065 void Compile::Optimize() { 2066 TracePhase tp("optimizer", &timers[_t_optimizer]); 2067 2068 #ifndef PRODUCT 2069 if (env()->break_at_compile()) { 2070 BREAKPOINT; 2071 } 2072 2073 #endif 2074 2075 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2076 #ifdef ASSERT 2077 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2078 #endif 2079 2080 ResourceMark rm; 2081 2082 print_inlining_reinit(); 2083 2084 NOT_PRODUCT( verify_graph_edges(); ) 2085 2086 print_method(PHASE_AFTER_PARSING); 2087 2088 { 2089 // Iterative Global Value Numbering, including ideal transforms 2090 // Initialize IterGVN with types and values from parse-time GVN 2091 PhaseIterGVN igvn(initial_gvn()); 2092 #ifdef ASSERT 2093 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2094 #endif 2095 { 2096 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2097 igvn.optimize(); 2098 } 2099 2100 if (failing()) return; 2101 2102 print_method(PHASE_ITER_GVN1, 2); 2103 2104 inline_incrementally(igvn); 2105 2106 print_method(PHASE_INCREMENTAL_INLINE, 2); 2107 2108 if (failing()) return; 2109 2110 if (eliminate_boxing()) { 2111 // Inline valueOf() methods now. 2112 inline_boxing_calls(igvn); 2113 2114 if (AlwaysIncrementalInline) { 2115 inline_incrementally(igvn); 2116 } 2117 2118 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2119 2120 if (failing()) return; 2121 } 2122 2123 // Remove the speculative part of types and clean up the graph from 2124 // the extra CastPP nodes whose only purpose is to carry them. Do 2125 // that early so that optimizations are not disrupted by the extra 2126 // CastPP nodes. 2127 remove_speculative_types(igvn); 2128 2129 // No more new expensive nodes will be added to the list from here 2130 // so keep only the actual candidates for optimizations. 2131 cleanup_expensive_nodes(igvn); 2132 2133 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2134 if (EnableVectorSupport && has_vbox_nodes()) { 2135 TracePhase tp("", &timers[_t_vector]); 2136 PhaseVector pv(igvn); 2137 pv.optimize_vector_boxes(); 2138 2139 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2140 } 2141 assert(!has_vbox_nodes(), "sanity"); 2142 2143 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2144 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2145 initial_gvn()->replace_with(&igvn); 2146 for_igvn()->clear(); 2147 Unique_Node_List new_worklist(C->comp_arena()); 2148 { 2149 ResourceMark rm; 2150 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2151 } 2152 Unique_Node_List* save_for_igvn = for_igvn(); 2153 set_for_igvn(&new_worklist); 2154 igvn = PhaseIterGVN(initial_gvn()); 2155 igvn.optimize(); 2156 set_for_igvn(save_for_igvn); 2157 } 2158 2159 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop 2160 // safepoints 2161 remove_root_to_sfpts_edges(igvn); 2162 2163 // Perform escape analysis 2164 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2165 if (has_loops()) { 2166 // Cleanup graph (remove dead nodes). 2167 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2168 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2169 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2170 if (failing()) return; 2171 } 2172 ConnectionGraph::do_analysis(this, &igvn); 2173 2174 if (failing()) return; 2175 2176 // Optimize out fields loads from scalar replaceable allocations. 2177 igvn.optimize(); 2178 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2179 2180 if (failing()) return; 2181 2182 if (congraph() != NULL && macro_count() > 0) { 2183 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2184 PhaseMacroExpand mexp(igvn); 2185 mexp.eliminate_macro_nodes(); 2186 igvn.set_delay_transform(false); 2187 2188 igvn.optimize(); 2189 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2190 2191 if (failing()) return; 2192 } 2193 } 2194 2195 // Loop transforms on the ideal graph. Range Check Elimination, 2196 // peeling, unrolling, etc. 2197 2198 // Set loop opts counter 2199 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2200 { 2201 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2202 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2203 _loop_opts_cnt--; 2204 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2205 if (failing()) return; 2206 } 2207 // Loop opts pass if partial peeling occurred in previous pass 2208 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2209 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2210 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2211 _loop_opts_cnt--; 2212 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2213 if (failing()) return; 2214 } 2215 // Loop opts pass for loop-unrolling before CCP 2216 if(major_progress() && (_loop_opts_cnt > 0)) { 2217 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2218 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2219 _loop_opts_cnt--; 2220 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2221 } 2222 if (!failing()) { 2223 // Verify that last round of loop opts produced a valid graph 2224 PhaseIdealLoop::verify(igvn); 2225 } 2226 } 2227 if (failing()) return; 2228 2229 // Conditional Constant Propagation; 2230 PhaseCCP ccp( &igvn ); 2231 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2232 { 2233 TracePhase tp("ccp", &timers[_t_ccp]); 2234 ccp.do_transform(); 2235 } 2236 print_method(PHASE_CCP1, 2); 2237 2238 assert( true, "Break here to ccp.dump_old2new_map()"); 2239 2240 // Iterative Global Value Numbering, including ideal transforms 2241 { 2242 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2243 igvn = ccp; 2244 igvn.optimize(); 2245 } 2246 print_method(PHASE_ITER_GVN2, 2); 2247 2248 if (failing()) return; 2249 2250 // Loop transforms on the ideal graph. Range Check Elimination, 2251 // peeling, unrolling, etc. 2252 if (!optimize_loops(igvn, LoopOptsDefault)) { 2253 return; 2254 } 2255 2256 if (failing()) return; 2257 2258 C->clear_major_progress(); // ensure that major progress is now clear 2259 2260 process_for_post_loop_opts_igvn(igvn); 2261 2262 #ifdef ASSERT 2263 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2264 #endif 2265 2266 { 2267 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2268 PhaseMacroExpand mex(igvn); 2269 if (mex.expand_macro_nodes()) { 2270 assert(failing(), "must bail out w/ explicit message"); 2271 return; 2272 } 2273 print_method(PHASE_MACRO_EXPANSION, 2); 2274 } 2275 2276 { 2277 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2278 if (bs->expand_barriers(this, igvn)) { 2279 assert(failing(), "must bail out w/ explicit message"); 2280 return; 2281 } 2282 print_method(PHASE_BARRIER_EXPANSION, 2); 2283 } 2284 2285 if (C->max_vector_size() > 0) { 2286 C->optimize_logic_cones(igvn); 2287 igvn.optimize(); 2288 } 2289 2290 DEBUG_ONLY( _modified_nodes = NULL; ) 2291 2292 assert(igvn._worklist.size() == 0, "not empty"); 2293 2294 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty"); 2295 2296 if (_late_inlines.length() > 0) { 2297 // More opportunities to optimize virtual and MH calls. 2298 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option. 2299 process_late_inline_calls_no_inline(igvn); 2300 } 2301 } // (End scope of igvn; run destructor if necessary for asserts.) 2302 2303 check_no_dead_use(); 2304 2305 process_print_inlining(); 2306 2307 // A method with only infinite loops has no edges entering loops from root 2308 { 2309 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2310 if (final_graph_reshaping()) { 2311 assert(failing(), "must bail out w/ explicit message"); 2312 return; 2313 } 2314 } 2315 2316 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2317 DEBUG_ONLY(set_phase_optimize_finished();) 2318 } 2319 2320 #ifdef ASSERT 2321 void Compile::check_no_dead_use() const { 2322 ResourceMark rm; 2323 Unique_Node_List wq; 2324 wq.push(root()); 2325 for (uint i = 0; i < wq.size(); ++i) { 2326 Node* n = wq.at(i); 2327 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2328 Node* u = n->fast_out(j); 2329 if (u->outcnt() == 0 && !u->is_Con()) { 2330 u->dump(); 2331 fatal("no reachable node should have no use"); 2332 } 2333 wq.push(u); 2334 } 2335 } 2336 } 2337 #endif 2338 2339 void Compile::inline_vector_reboxing_calls() { 2340 if (C->_vector_reboxing_late_inlines.length() > 0) { 2341 _late_inlines_pos = C->_late_inlines.length(); 2342 while (_vector_reboxing_late_inlines.length() > 0) { 2343 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 2344 cg->do_late_inline(); 2345 if (failing()) return; 2346 print_method(PHASE_INLINE_VECTOR_REBOX, cg->call_node()); 2347 } 2348 _vector_reboxing_late_inlines.trunc_to(0); 2349 } 2350 } 2351 2352 bool Compile::has_vbox_nodes() { 2353 if (C->_vector_reboxing_late_inlines.length() > 0) { 2354 return true; 2355 } 2356 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 2357 Node * n = C->macro_node(macro_idx); 2358 assert(n->is_macro(), "only macro nodes expected here"); 2359 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 2360 return true; 2361 } 2362 } 2363 return false; 2364 } 2365 2366 //---------------------------- Bitwise operation packing optimization --------------------------- 2367 2368 static bool is_vector_unary_bitwise_op(Node* n) { 2369 return n->Opcode() == Op_XorV && 2370 VectorNode::is_vector_bitwise_not_pattern(n); 2371 } 2372 2373 static bool is_vector_binary_bitwise_op(Node* n) { 2374 switch (n->Opcode()) { 2375 case Op_AndV: 2376 case Op_OrV: 2377 return true; 2378 2379 case Op_XorV: 2380 return !is_vector_unary_bitwise_op(n); 2381 2382 default: 2383 return false; 2384 } 2385 } 2386 2387 static bool is_vector_ternary_bitwise_op(Node* n) { 2388 return n->Opcode() == Op_MacroLogicV; 2389 } 2390 2391 static bool is_vector_bitwise_op(Node* n) { 2392 return is_vector_unary_bitwise_op(n) || 2393 is_vector_binary_bitwise_op(n) || 2394 is_vector_ternary_bitwise_op(n); 2395 } 2396 2397 static bool is_vector_bitwise_cone_root(Node* n) { 2398 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) { 2399 return false; 2400 } 2401 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2402 if (is_vector_bitwise_op(n->fast_out(i))) { 2403 return false; 2404 } 2405 } 2406 return true; 2407 } 2408 2409 static uint collect_unique_inputs(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2410 uint cnt = 0; 2411 if (is_vector_bitwise_op(n)) { 2412 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2413 for (uint i = 1; i < n->req(); i++) { 2414 Node* in = n->in(i); 2415 bool skip = VectorNode::is_all_ones_vector(in); 2416 if (!skip && !inputs.member(in)) { 2417 inputs.push(in); 2418 cnt++; 2419 } 2420 } 2421 assert(cnt <= 1, "not unary"); 2422 } else { 2423 uint last_req = n->req(); 2424 if (is_vector_ternary_bitwise_op(n)) { 2425 last_req = n->req() - 1; // skip last input 2426 } 2427 for (uint i = 1; i < last_req; i++) { 2428 Node* def = n->in(i); 2429 if (!inputs.member(def)) { 2430 inputs.push(def); 2431 cnt++; 2432 } 2433 } 2434 } 2435 partition.push(n); 2436 } else { // not a bitwise operations 2437 if (!inputs.member(n)) { 2438 inputs.push(n); 2439 cnt++; 2440 } 2441 } 2442 return cnt; 2443 } 2444 2445 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 2446 Unique_Node_List useful_nodes; 2447 C->identify_useful_nodes(useful_nodes); 2448 2449 for (uint i = 0; i < useful_nodes.size(); i++) { 2450 Node* n = useful_nodes.at(i); 2451 if (is_vector_bitwise_cone_root(n)) { 2452 list.push(n); 2453 } 2454 } 2455 } 2456 2457 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 2458 const TypeVect* vt, 2459 Unique_Node_List& partition, 2460 Unique_Node_List& inputs) { 2461 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 2462 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 2463 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 2464 2465 Node* in1 = inputs.at(0); 2466 Node* in2 = inputs.at(1); 2467 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 2468 2469 uint func = compute_truth_table(partition, inputs); 2470 return igvn.transform(MacroLogicVNode::make(igvn, in3, in2, in1, func, vt)); 2471 } 2472 2473 static uint extract_bit(uint func, uint pos) { 2474 return (func & (1 << pos)) >> pos; 2475 } 2476 2477 // 2478 // A macro logic node represents a truth table. It has 4 inputs, 2479 // First three inputs corresponds to 3 columns of a truth table 2480 // and fourth input captures the logic function. 2481 // 2482 // eg. fn = (in1 AND in2) OR in3; 2483 // 2484 // MacroNode(in1,in2,in3,fn) 2485 // 2486 // ----------------- 2487 // in1 in2 in3 fn 2488 // ----------------- 2489 // 0 0 0 0 2490 // 0 0 1 1 2491 // 0 1 0 0 2492 // 0 1 1 1 2493 // 1 0 0 0 2494 // 1 0 1 1 2495 // 1 1 0 1 2496 // 1 1 1 1 2497 // 2498 2499 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 2500 int res = 0; 2501 for (int i = 0; i < 8; i++) { 2502 int bit1 = extract_bit(in1, i); 2503 int bit2 = extract_bit(in2, i); 2504 int bit3 = extract_bit(in3, i); 2505 2506 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 2507 int func_bit = extract_bit(func, func_bit_pos); 2508 2509 res |= func_bit << i; 2510 } 2511 return res; 2512 } 2513 2514 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 2515 assert(n != NULL, ""); 2516 assert(eval_map.contains(n), "absent"); 2517 return *(eval_map.get(n)); 2518 } 2519 2520 static void eval_operands(Node* n, 2521 uint& func1, uint& func2, uint& func3, 2522 ResourceHashtable<Node*,uint>& eval_map) { 2523 assert(is_vector_bitwise_op(n), ""); 2524 2525 if (is_vector_unary_bitwise_op(n)) { 2526 Node* opnd = n->in(1); 2527 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 2528 opnd = n->in(2); 2529 } 2530 func1 = eval_operand(opnd, eval_map); 2531 } else if (is_vector_binary_bitwise_op(n)) { 2532 func1 = eval_operand(n->in(1), eval_map); 2533 func2 = eval_operand(n->in(2), eval_map); 2534 } else { 2535 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 2536 func1 = eval_operand(n->in(1), eval_map); 2537 func2 = eval_operand(n->in(2), eval_map); 2538 func3 = eval_operand(n->in(3), eval_map); 2539 } 2540 } 2541 2542 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 2543 assert(inputs.size() <= 3, "sanity"); 2544 ResourceMark rm; 2545 uint res = 0; 2546 ResourceHashtable<Node*,uint> eval_map; 2547 2548 // Populate precomputed functions for inputs. 2549 // Each input corresponds to one column of 3 input truth-table. 2550 uint input_funcs[] = { 0xAA, // (_, _, a) -> a 2551 0xCC, // (_, b, _) -> b 2552 0xF0 }; // (c, _, _) -> c 2553 for (uint i = 0; i < inputs.size(); i++) { 2554 eval_map.put(inputs.at(i), input_funcs[i]); 2555 } 2556 2557 for (uint i = 0; i < partition.size(); i++) { 2558 Node* n = partition.at(i); 2559 2560 uint func1 = 0, func2 = 0, func3 = 0; 2561 eval_operands(n, func1, func2, func3, eval_map); 2562 2563 switch (n->Opcode()) { 2564 case Op_OrV: 2565 assert(func3 == 0, "not binary"); 2566 res = func1 | func2; 2567 break; 2568 case Op_AndV: 2569 assert(func3 == 0, "not binary"); 2570 res = func1 & func2; 2571 break; 2572 case Op_XorV: 2573 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2574 assert(func2 == 0 && func3 == 0, "not unary"); 2575 res = (~func1) & 0xFF; 2576 } else { 2577 assert(func3 == 0, "not binary"); 2578 res = func1 ^ func2; 2579 } 2580 break; 2581 case Op_MacroLogicV: 2582 // Ordering of inputs may change during evaluation of sub-tree 2583 // containing MacroLogic node as a child node, thus a re-evaluation 2584 // makes sure that function is evaluated in context of current 2585 // inputs. 2586 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 2587 break; 2588 2589 default: assert(false, "not supported: %s", n->Name()); 2590 } 2591 assert(res <= 0xFF, "invalid"); 2592 eval_map.put(n, res); 2593 } 2594 return res; 2595 } 2596 2597 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2598 assert(partition.size() == 0, "not empty"); 2599 assert(inputs.size() == 0, "not empty"); 2600 if (is_vector_ternary_bitwise_op(n)) { 2601 return false; 2602 } 2603 2604 bool is_unary_op = is_vector_unary_bitwise_op(n); 2605 if (is_unary_op) { 2606 assert(collect_unique_inputs(n, partition, inputs) == 1, "not unary"); 2607 return false; // too few inputs 2608 } 2609 2610 assert(is_vector_binary_bitwise_op(n), "not binary"); 2611 Node* in1 = n->in(1); 2612 Node* in2 = n->in(2); 2613 2614 int in1_unique_inputs_cnt = collect_unique_inputs(in1, partition, inputs); 2615 int in2_unique_inputs_cnt = collect_unique_inputs(in2, partition, inputs); 2616 partition.push(n); 2617 2618 // Too many inputs? 2619 if (inputs.size() > 3) { 2620 partition.clear(); 2621 inputs.clear(); 2622 { // Recompute in2 inputs 2623 Unique_Node_List not_used; 2624 in2_unique_inputs_cnt = collect_unique_inputs(in2, not_used, not_used); 2625 } 2626 // Pick the node with minimum number of inputs. 2627 if (in1_unique_inputs_cnt >= 3 && in2_unique_inputs_cnt >= 3) { 2628 return false; // still too many inputs 2629 } 2630 // Recompute partition & inputs. 2631 Node* child = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in1 : in2); 2632 collect_unique_inputs(child, partition, inputs); 2633 2634 Node* other_input = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in2 : in1); 2635 inputs.push(other_input); 2636 2637 partition.push(n); 2638 } 2639 2640 return (partition.size() == 2 || partition.size() == 3) && 2641 (inputs.size() == 2 || inputs.size() == 3); 2642 } 2643 2644 2645 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 2646 assert(is_vector_bitwise_op(n), "not a root"); 2647 2648 visited.set(n->_idx); 2649 2650 // 1) Do a DFS walk over the logic cone. 2651 for (uint i = 1; i < n->req(); i++) { 2652 Node* in = n->in(i); 2653 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 2654 process_logic_cone_root(igvn, in, visited); 2655 } 2656 } 2657 2658 // 2) Bottom up traversal: Merge node[s] with 2659 // the parent to form macro logic node. 2660 Unique_Node_List partition; 2661 Unique_Node_List inputs; 2662 if (compute_logic_cone(n, partition, inputs)) { 2663 const TypeVect* vt = n->bottom_type()->is_vect(); 2664 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 2665 igvn.replace_node(n, macro_logic); 2666 } 2667 } 2668 2669 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 2670 ResourceMark rm; 2671 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 2672 Unique_Node_List list; 2673 collect_logic_cone_roots(list); 2674 2675 while (list.size() > 0) { 2676 Node* n = list.pop(); 2677 const TypeVect* vt = n->bottom_type()->is_vect(); 2678 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 2679 if (supported) { 2680 VectorSet visited(comp_arena()); 2681 process_logic_cone_root(igvn, n, visited); 2682 } 2683 } 2684 } 2685 } 2686 2687 //------------------------------Code_Gen--------------------------------------- 2688 // Given a graph, generate code for it 2689 void Compile::Code_Gen() { 2690 if (failing()) { 2691 return; 2692 } 2693 2694 // Perform instruction selection. You might think we could reclaim Matcher 2695 // memory PDQ, but actually the Matcher is used in generating spill code. 2696 // Internals of the Matcher (including some VectorSets) must remain live 2697 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2698 // set a bit in reclaimed memory. 2699 2700 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2701 // nodes. Mapping is only valid at the root of each matched subtree. 2702 NOT_PRODUCT( verify_graph_edges(); ) 2703 2704 Matcher matcher; 2705 _matcher = &matcher; 2706 { 2707 TracePhase tp("matcher", &timers[_t_matcher]); 2708 matcher.match(); 2709 if (failing()) { 2710 return; 2711 } 2712 } 2713 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2714 // nodes. Mapping is only valid at the root of each matched subtree. 2715 NOT_PRODUCT( verify_graph_edges(); ) 2716 2717 // If you have too many nodes, or if matching has failed, bail out 2718 check_node_count(0, "out of nodes matching instructions"); 2719 if (failing()) { 2720 return; 2721 } 2722 2723 print_method(PHASE_MATCHING, 2); 2724 2725 // Build a proper-looking CFG 2726 PhaseCFG cfg(node_arena(), root(), matcher); 2727 _cfg = &cfg; 2728 { 2729 TracePhase tp("scheduler", &timers[_t_scheduler]); 2730 bool success = cfg.do_global_code_motion(); 2731 if (!success) { 2732 return; 2733 } 2734 2735 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2736 NOT_PRODUCT( verify_graph_edges(); ) 2737 cfg.verify(); 2738 } 2739 2740 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2741 _regalloc = ®alloc; 2742 { 2743 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2744 // Perform register allocation. After Chaitin, use-def chains are 2745 // no longer accurate (at spill code) and so must be ignored. 2746 // Node->LRG->reg mappings are still accurate. 2747 _regalloc->Register_Allocate(); 2748 2749 // Bail out if the allocator builds too many nodes 2750 if (failing()) { 2751 return; 2752 } 2753 } 2754 2755 // Prior to register allocation we kept empty basic blocks in case the 2756 // the allocator needed a place to spill. After register allocation we 2757 // are not adding any new instructions. If any basic block is empty, we 2758 // can now safely remove it. 2759 { 2760 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2761 cfg.remove_empty_blocks(); 2762 if (do_freq_based_layout()) { 2763 PhaseBlockLayout layout(cfg); 2764 } else { 2765 cfg.set_loop_alignment(); 2766 } 2767 cfg.fixup_flow(); 2768 cfg.remove_unreachable_blocks(); 2769 cfg.verify_dominator_tree(); 2770 } 2771 2772 // Apply peephole optimizations 2773 if( OptoPeephole ) { 2774 TracePhase tp("peephole", &timers[_t_peephole]); 2775 PhasePeephole peep( _regalloc, cfg); 2776 peep.do_transform(); 2777 } 2778 2779 // Do late expand if CPU requires this. 2780 if (Matcher::require_postalloc_expand) { 2781 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2782 cfg.postalloc_expand(_regalloc); 2783 } 2784 2785 // Convert Nodes to instruction bits in a buffer 2786 { 2787 TracePhase tp("output", &timers[_t_output]); 2788 PhaseOutput output; 2789 output.Output(); 2790 if (failing()) return; 2791 output.install(); 2792 } 2793 2794 print_method(PHASE_FINAL_CODE); 2795 2796 // He's dead, Jim. 2797 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2798 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2799 } 2800 2801 //------------------------------Final_Reshape_Counts--------------------------- 2802 // This class defines counters to help identify when a method 2803 // may/must be executed using hardware with only 24-bit precision. 2804 struct Final_Reshape_Counts : public StackObj { 2805 int _call_count; // count non-inlined 'common' calls 2806 int _float_count; // count float ops requiring 24-bit precision 2807 int _double_count; // count double ops requiring more precision 2808 int _java_call_count; // count non-inlined 'java' calls 2809 int _inner_loop_count; // count loops which need alignment 2810 VectorSet _visited; // Visitation flags 2811 Node_List _tests; // Set of IfNodes & PCTableNodes 2812 2813 Final_Reshape_Counts() : 2814 _call_count(0), _float_count(0), _double_count(0), 2815 _java_call_count(0), _inner_loop_count(0) { } 2816 2817 void inc_call_count () { _call_count ++; } 2818 void inc_float_count () { _float_count ++; } 2819 void inc_double_count() { _double_count++; } 2820 void inc_java_call_count() { _java_call_count++; } 2821 void inc_inner_loop_count() { _inner_loop_count++; } 2822 2823 int get_call_count () const { return _call_count ; } 2824 int get_float_count () const { return _float_count ; } 2825 int get_double_count() const { return _double_count; } 2826 int get_java_call_count() const { return _java_call_count; } 2827 int get_inner_loop_count() const { return _inner_loop_count; } 2828 }; 2829 2830 // Eliminate trivially redundant StoreCMs and accumulate their 2831 // precedence edges. 2832 void Compile::eliminate_redundant_card_marks(Node* n) { 2833 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2834 if (n->in(MemNode::Address)->outcnt() > 1) { 2835 // There are multiple users of the same address so it might be 2836 // possible to eliminate some of the StoreCMs 2837 Node* mem = n->in(MemNode::Memory); 2838 Node* adr = n->in(MemNode::Address); 2839 Node* val = n->in(MemNode::ValueIn); 2840 Node* prev = n; 2841 bool done = false; 2842 // Walk the chain of StoreCMs eliminating ones that match. As 2843 // long as it's a chain of single users then the optimization is 2844 // safe. Eliminating partially redundant StoreCMs would require 2845 // cloning copies down the other paths. 2846 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2847 if (adr == mem->in(MemNode::Address) && 2848 val == mem->in(MemNode::ValueIn)) { 2849 // redundant StoreCM 2850 if (mem->req() > MemNode::OopStore) { 2851 // Hasn't been processed by this code yet. 2852 n->add_prec(mem->in(MemNode::OopStore)); 2853 } else { 2854 // Already converted to precedence edge 2855 for (uint i = mem->req(); i < mem->len(); i++) { 2856 // Accumulate any precedence edges 2857 if (mem->in(i) != NULL) { 2858 n->add_prec(mem->in(i)); 2859 } 2860 } 2861 // Everything above this point has been processed. 2862 done = true; 2863 } 2864 // Eliminate the previous StoreCM 2865 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2866 assert(mem->outcnt() == 0, "should be dead"); 2867 mem->disconnect_inputs(this); 2868 } else { 2869 prev = mem; 2870 } 2871 mem = prev->in(MemNode::Memory); 2872 } 2873 } 2874 } 2875 2876 //------------------------------final_graph_reshaping_impl---------------------- 2877 // Implement items 1-5 from final_graph_reshaping below. 2878 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 2879 2880 if ( n->outcnt() == 0 ) return; // dead node 2881 uint nop = n->Opcode(); 2882 2883 // Check for 2-input instruction with "last use" on right input. 2884 // Swap to left input. Implements item (2). 2885 if( n->req() == 3 && // two-input instruction 2886 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2887 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2888 n->in(2)->outcnt() == 1 &&// right use IS a last use 2889 !n->in(2)->is_Con() ) { // right use is not a constant 2890 // Check for commutative opcode 2891 switch( nop ) { 2892 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2893 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 2894 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 2895 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2896 case Op_AndL: case Op_XorL: case Op_OrL: 2897 case Op_AndI: case Op_XorI: case Op_OrI: { 2898 // Move "last use" input to left by swapping inputs 2899 n->swap_edges(1, 2); 2900 break; 2901 } 2902 default: 2903 break; 2904 } 2905 } 2906 2907 #ifdef ASSERT 2908 if( n->is_Mem() ) { 2909 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2910 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2911 // oop will be recorded in oop map if load crosses safepoint 2912 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2913 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2914 "raw memory operations should have control edge"); 2915 } 2916 if (n->is_MemBar()) { 2917 MemBarNode* mb = n->as_MemBar(); 2918 if (mb->trailing_store() || mb->trailing_load_store()) { 2919 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2920 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 2921 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2922 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2923 } else if (mb->leading()) { 2924 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2925 } 2926 } 2927 #endif 2928 // Count FPU ops and common calls, implements item (3) 2929 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes); 2930 if (!gc_handled) { 2931 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes); 2932 } 2933 2934 // Collect CFG split points 2935 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 2936 frc._tests.push(n); 2937 } 2938 } 2939 2940 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) { 2941 switch( nop ) { 2942 // Count all float operations that may use FPU 2943 case Op_AddF: 2944 case Op_SubF: 2945 case Op_MulF: 2946 case Op_DivF: 2947 case Op_NegF: 2948 case Op_ModF: 2949 case Op_ConvI2F: 2950 case Op_ConF: 2951 case Op_CmpF: 2952 case Op_CmpF3: 2953 case Op_StoreF: 2954 case Op_LoadF: 2955 // case Op_ConvL2F: // longs are split into 32-bit halves 2956 frc.inc_float_count(); 2957 break; 2958 2959 case Op_ConvF2D: 2960 case Op_ConvD2F: 2961 frc.inc_float_count(); 2962 frc.inc_double_count(); 2963 break; 2964 2965 // Count all double operations that may use FPU 2966 case Op_AddD: 2967 case Op_SubD: 2968 case Op_MulD: 2969 case Op_DivD: 2970 case Op_NegD: 2971 case Op_ModD: 2972 case Op_ConvI2D: 2973 case Op_ConvD2I: 2974 // case Op_ConvL2D: // handled by leaf call 2975 // case Op_ConvD2L: // handled by leaf call 2976 case Op_ConD: 2977 case Op_CmpD: 2978 case Op_CmpD3: 2979 case Op_StoreD: 2980 case Op_LoadD: 2981 case Op_LoadD_unaligned: 2982 frc.inc_double_count(); 2983 break; 2984 case Op_Opaque1: // Remove Opaque Nodes before matching 2985 case Op_Opaque2: // Remove Opaque Nodes before matching 2986 case Op_Opaque3: 2987 n->subsume_by(n->in(1), this); 2988 break; 2989 case Op_CallStaticJava: 2990 case Op_CallJava: 2991 case Op_CallDynamicJava: 2992 frc.inc_java_call_count(); // Count java call site; 2993 case Op_CallRuntime: 2994 case Op_CallLeaf: 2995 case Op_CallLeafVector: 2996 case Op_CallNative: 2997 case Op_CallLeafNoFP: { 2998 assert (n->is_Call(), ""); 2999 CallNode *call = n->as_Call(); 3000 // Count call sites where the FP mode bit would have to be flipped. 3001 // Do not count uncommon runtime calls: 3002 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3003 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3004 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3005 frc.inc_call_count(); // Count the call site 3006 } else { // See if uncommon argument is shared 3007 Node *n = call->in(TypeFunc::Parms); 3008 int nop = n->Opcode(); 3009 // Clone shared simple arguments to uncommon calls, item (1). 3010 if (n->outcnt() > 1 && 3011 !n->is_Proj() && 3012 nop != Op_CreateEx && 3013 nop != Op_CheckCastPP && 3014 nop != Op_DecodeN && 3015 nop != Op_DecodeNKlass && 3016 !n->is_Mem() && 3017 !n->is_Phi()) { 3018 Node *x = n->clone(); 3019 call->set_req(TypeFunc::Parms, x); 3020 } 3021 } 3022 break; 3023 } 3024 3025 case Op_StoreCM: 3026 { 3027 // Convert OopStore dependence into precedence edge 3028 Node* prec = n->in(MemNode::OopStore); 3029 n->del_req(MemNode::OopStore); 3030 n->add_prec(prec); 3031 eliminate_redundant_card_marks(n); 3032 } 3033 3034 // fall through 3035 3036 case Op_StoreB: 3037 case Op_StoreC: 3038 case Op_StorePConditional: 3039 case Op_StoreI: 3040 case Op_StoreL: 3041 case Op_StoreIConditional: 3042 case Op_StoreLConditional: 3043 case Op_CompareAndSwapB: 3044 case Op_CompareAndSwapS: 3045 case Op_CompareAndSwapI: 3046 case Op_CompareAndSwapL: 3047 case Op_CompareAndSwapP: 3048 case Op_CompareAndSwapN: 3049 case Op_WeakCompareAndSwapB: 3050 case Op_WeakCompareAndSwapS: 3051 case Op_WeakCompareAndSwapI: 3052 case Op_WeakCompareAndSwapL: 3053 case Op_WeakCompareAndSwapP: 3054 case Op_WeakCompareAndSwapN: 3055 case Op_CompareAndExchangeB: 3056 case Op_CompareAndExchangeS: 3057 case Op_CompareAndExchangeI: 3058 case Op_CompareAndExchangeL: 3059 case Op_CompareAndExchangeP: 3060 case Op_CompareAndExchangeN: 3061 case Op_GetAndAddS: 3062 case Op_GetAndAddB: 3063 case Op_GetAndAddI: 3064 case Op_GetAndAddL: 3065 case Op_GetAndSetS: 3066 case Op_GetAndSetB: 3067 case Op_GetAndSetI: 3068 case Op_GetAndSetL: 3069 case Op_GetAndSetP: 3070 case Op_GetAndSetN: 3071 case Op_StoreP: 3072 case Op_StoreN: 3073 case Op_StoreNKlass: 3074 case Op_LoadB: 3075 case Op_LoadUB: 3076 case Op_LoadUS: 3077 case Op_LoadI: 3078 case Op_LoadKlass: 3079 case Op_LoadNKlass: 3080 case Op_LoadL: 3081 case Op_LoadL_unaligned: 3082 case Op_LoadPLocked: 3083 case Op_LoadP: 3084 case Op_LoadN: 3085 case Op_LoadRange: 3086 case Op_LoadS: 3087 break; 3088 3089 case Op_AddP: { // Assert sane base pointers 3090 Node *addp = n->in(AddPNode::Address); 3091 assert( !addp->is_AddP() || 3092 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3093 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3094 "Base pointers must match (addp %u)", addp->_idx ); 3095 #ifdef _LP64 3096 if ((UseCompressedOops || UseCompressedClassPointers) && 3097 addp->Opcode() == Op_ConP && 3098 addp == n->in(AddPNode::Base) && 3099 n->in(AddPNode::Offset)->is_Con()) { 3100 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3101 // on the platform and on the compressed oops mode. 3102 // Use addressing with narrow klass to load with offset on x86. 3103 // Some platforms can use the constant pool to load ConP. 3104 // Do this transformation here since IGVN will convert ConN back to ConP. 3105 const Type* t = addp->bottom_type(); 3106 bool is_oop = t->isa_oopptr() != NULL; 3107 bool is_klass = t->isa_klassptr() != NULL; 3108 3109 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3110 (is_klass && Matcher::const_klass_prefer_decode())) { 3111 Node* nn = NULL; 3112 3113 int op = is_oop ? Op_ConN : Op_ConNKlass; 3114 3115 // Look for existing ConN node of the same exact type. 3116 Node* r = root(); 3117 uint cnt = r->outcnt(); 3118 for (uint i = 0; i < cnt; i++) { 3119 Node* m = r->raw_out(i); 3120 if (m!= NULL && m->Opcode() == op && 3121 m->bottom_type()->make_ptr() == t) { 3122 nn = m; 3123 break; 3124 } 3125 } 3126 if (nn != NULL) { 3127 // Decode a narrow oop to match address 3128 // [R12 + narrow_oop_reg<<3 + offset] 3129 if (is_oop) { 3130 nn = new DecodeNNode(nn, t); 3131 } else { 3132 nn = new DecodeNKlassNode(nn, t); 3133 } 3134 // Check for succeeding AddP which uses the same Base. 3135 // Otherwise we will run into the assertion above when visiting that guy. 3136 for (uint i = 0; i < n->outcnt(); ++i) { 3137 Node *out_i = n->raw_out(i); 3138 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3139 out_i->set_req(AddPNode::Base, nn); 3140 #ifdef ASSERT 3141 for (uint j = 0; j < out_i->outcnt(); ++j) { 3142 Node *out_j = out_i->raw_out(j); 3143 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3144 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3145 } 3146 #endif 3147 } 3148 } 3149 n->set_req(AddPNode::Base, nn); 3150 n->set_req(AddPNode::Address, nn); 3151 if (addp->outcnt() == 0) { 3152 addp->disconnect_inputs(this); 3153 } 3154 } 3155 } 3156 } 3157 #endif 3158 break; 3159 } 3160 3161 case Op_CastPP: { 3162 // Remove CastPP nodes to gain more freedom during scheduling but 3163 // keep the dependency they encode as control or precedence edges 3164 // (if control is set already) on memory operations. Some CastPP 3165 // nodes don't have a control (don't carry a dependency): skip 3166 // those. 3167 if (n->in(0) != NULL) { 3168 ResourceMark rm; 3169 Unique_Node_List wq; 3170 wq.push(n); 3171 for (uint next = 0; next < wq.size(); ++next) { 3172 Node *m = wq.at(next); 3173 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3174 Node* use = m->fast_out(i); 3175 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3176 use->ensure_control_or_add_prec(n->in(0)); 3177 } else { 3178 switch(use->Opcode()) { 3179 case Op_AddP: 3180 case Op_DecodeN: 3181 case Op_DecodeNKlass: 3182 case Op_CheckCastPP: 3183 case Op_CastPP: 3184 wq.push(use); 3185 break; 3186 } 3187 } 3188 } 3189 } 3190 } 3191 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3192 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3193 Node* in1 = n->in(1); 3194 const Type* t = n->bottom_type(); 3195 Node* new_in1 = in1->clone(); 3196 new_in1->as_DecodeN()->set_type(t); 3197 3198 if (!Matcher::narrow_oop_use_complex_address()) { 3199 // 3200 // x86, ARM and friends can handle 2 adds in addressing mode 3201 // and Matcher can fold a DecodeN node into address by using 3202 // a narrow oop directly and do implicit NULL check in address: 3203 // 3204 // [R12 + narrow_oop_reg<<3 + offset] 3205 // NullCheck narrow_oop_reg 3206 // 3207 // On other platforms (Sparc) we have to keep new DecodeN node and 3208 // use it to do implicit NULL check in address: 3209 // 3210 // decode_not_null narrow_oop_reg, base_reg 3211 // [base_reg + offset] 3212 // NullCheck base_reg 3213 // 3214 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3215 // to keep the information to which NULL check the new DecodeN node 3216 // corresponds to use it as value in implicit_null_check(). 3217 // 3218 new_in1->set_req(0, n->in(0)); 3219 } 3220 3221 n->subsume_by(new_in1, this); 3222 if (in1->outcnt() == 0) { 3223 in1->disconnect_inputs(this); 3224 } 3225 } else { 3226 n->subsume_by(n->in(1), this); 3227 if (n->outcnt() == 0) { 3228 n->disconnect_inputs(this); 3229 } 3230 } 3231 break; 3232 } 3233 #ifdef _LP64 3234 case Op_CmpP: 3235 // Do this transformation here to preserve CmpPNode::sub() and 3236 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3237 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3238 Node* in1 = n->in(1); 3239 Node* in2 = n->in(2); 3240 if (!in1->is_DecodeNarrowPtr()) { 3241 in2 = in1; 3242 in1 = n->in(2); 3243 } 3244 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3245 3246 Node* new_in2 = NULL; 3247 if (in2->is_DecodeNarrowPtr()) { 3248 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3249 new_in2 = in2->in(1); 3250 } else if (in2->Opcode() == Op_ConP) { 3251 const Type* t = in2->bottom_type(); 3252 if (t == TypePtr::NULL_PTR) { 3253 assert(in1->is_DecodeN(), "compare klass to null?"); 3254 // Don't convert CmpP null check into CmpN if compressed 3255 // oops implicit null check is not generated. 3256 // This will allow to generate normal oop implicit null check. 3257 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3258 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3259 // 3260 // This transformation together with CastPP transformation above 3261 // will generated code for implicit NULL checks for compressed oops. 3262 // 3263 // The original code after Optimize() 3264 // 3265 // LoadN memory, narrow_oop_reg 3266 // decode narrow_oop_reg, base_reg 3267 // CmpP base_reg, NULL 3268 // CastPP base_reg // NotNull 3269 // Load [base_reg + offset], val_reg 3270 // 3271 // after these transformations will be 3272 // 3273 // LoadN memory, narrow_oop_reg 3274 // CmpN narrow_oop_reg, NULL 3275 // decode_not_null narrow_oop_reg, base_reg 3276 // Load [base_reg + offset], val_reg 3277 // 3278 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3279 // since narrow oops can be used in debug info now (see the code in 3280 // final_graph_reshaping_walk()). 3281 // 3282 // At the end the code will be matched to 3283 // on x86: 3284 // 3285 // Load_narrow_oop memory, narrow_oop_reg 3286 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3287 // NullCheck narrow_oop_reg 3288 // 3289 // and on sparc: 3290 // 3291 // Load_narrow_oop memory, narrow_oop_reg 3292 // decode_not_null narrow_oop_reg, base_reg 3293 // Load [base_reg + offset], val_reg 3294 // NullCheck base_reg 3295 // 3296 } else if (t->isa_oopptr()) { 3297 new_in2 = ConNode::make(t->make_narrowoop()); 3298 } else if (t->isa_klassptr()) { 3299 new_in2 = ConNode::make(t->make_narrowklass()); 3300 } 3301 } 3302 if (new_in2 != NULL) { 3303 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3304 n->subsume_by(cmpN, this); 3305 if (in1->outcnt() == 0) { 3306 in1->disconnect_inputs(this); 3307 } 3308 if (in2->outcnt() == 0) { 3309 in2->disconnect_inputs(this); 3310 } 3311 } 3312 } 3313 break; 3314 3315 case Op_DecodeN: 3316 case Op_DecodeNKlass: 3317 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3318 // DecodeN could be pinned when it can't be fold into 3319 // an address expression, see the code for Op_CastPP above. 3320 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3321 break; 3322 3323 case Op_EncodeP: 3324 case Op_EncodePKlass: { 3325 Node* in1 = n->in(1); 3326 if (in1->is_DecodeNarrowPtr()) { 3327 n->subsume_by(in1->in(1), this); 3328 } else if (in1->Opcode() == Op_ConP) { 3329 const Type* t = in1->bottom_type(); 3330 if (t == TypePtr::NULL_PTR) { 3331 assert(t->isa_oopptr(), "null klass?"); 3332 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3333 } else if (t->isa_oopptr()) { 3334 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3335 } else if (t->isa_klassptr()) { 3336 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3337 } 3338 } 3339 if (in1->outcnt() == 0) { 3340 in1->disconnect_inputs(this); 3341 } 3342 break; 3343 } 3344 3345 case Op_Proj: { 3346 if (OptimizeStringConcat || IncrementalInline) { 3347 ProjNode* proj = n->as_Proj(); 3348 if (proj->_is_io_use) { 3349 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, ""); 3350 // Separate projections were used for the exception path which 3351 // are normally removed by a late inline. If it wasn't inlined 3352 // then they will hang around and should just be replaced with 3353 // the original one. Merge them. 3354 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/); 3355 if (non_io_proj != NULL) { 3356 proj->subsume_by(non_io_proj , this); 3357 } 3358 } 3359 } 3360 break; 3361 } 3362 3363 case Op_Phi: 3364 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3365 // The EncodeP optimization may create Phi with the same edges 3366 // for all paths. It is not handled well by Register Allocator. 3367 Node* unique_in = n->in(1); 3368 assert(unique_in != NULL, ""); 3369 uint cnt = n->req(); 3370 for (uint i = 2; i < cnt; i++) { 3371 Node* m = n->in(i); 3372 assert(m != NULL, ""); 3373 if (unique_in != m) 3374 unique_in = NULL; 3375 } 3376 if (unique_in != NULL) { 3377 n->subsume_by(unique_in, this); 3378 } 3379 } 3380 break; 3381 3382 #endif 3383 3384 #ifdef ASSERT 3385 case Op_CastII: 3386 // Verify that all range check dependent CastII nodes were removed. 3387 if (n->isa_CastII()->has_range_check()) { 3388 n->dump(3); 3389 assert(false, "Range check dependent CastII node was not removed"); 3390 } 3391 break; 3392 #endif 3393 3394 case Op_ModI: 3395 if (UseDivMod) { 3396 // Check if a%b and a/b both exist 3397 Node* d = n->find_similar(Op_DivI); 3398 if (d) { 3399 // Replace them with a fused divmod if supported 3400 if (Matcher::has_match_rule(Op_DivModI)) { 3401 DivModINode* divmod = DivModINode::make(n); 3402 d->subsume_by(divmod->div_proj(), this); 3403 n->subsume_by(divmod->mod_proj(), this); 3404 } else { 3405 // replace a%b with a-((a/b)*b) 3406 Node* mult = new MulINode(d, d->in(2)); 3407 Node* sub = new SubINode(d->in(1), mult); 3408 n->subsume_by(sub, this); 3409 } 3410 } 3411 } 3412 break; 3413 3414 case Op_ModL: 3415 if (UseDivMod) { 3416 // Check if a%b and a/b both exist 3417 Node* d = n->find_similar(Op_DivL); 3418 if (d) { 3419 // Replace them with a fused divmod if supported 3420 if (Matcher::has_match_rule(Op_DivModL)) { 3421 DivModLNode* divmod = DivModLNode::make(n); 3422 d->subsume_by(divmod->div_proj(), this); 3423 n->subsume_by(divmod->mod_proj(), this); 3424 } else { 3425 // replace a%b with a-((a/b)*b) 3426 Node* mult = new MulLNode(d, d->in(2)); 3427 Node* sub = new SubLNode(d->in(1), mult); 3428 n->subsume_by(sub, this); 3429 } 3430 } 3431 } 3432 break; 3433 3434 case Op_LoadVector: 3435 case Op_StoreVector: 3436 case Op_LoadVectorGather: 3437 case Op_StoreVectorScatter: 3438 case Op_VectorCmpMasked: 3439 case Op_VectorMaskGen: 3440 case Op_LoadVectorMasked: 3441 case Op_StoreVectorMasked: 3442 break; 3443 3444 case Op_AddReductionVI: 3445 case Op_AddReductionVL: 3446 case Op_AddReductionVF: 3447 case Op_AddReductionVD: 3448 case Op_MulReductionVI: 3449 case Op_MulReductionVL: 3450 case Op_MulReductionVF: 3451 case Op_MulReductionVD: 3452 case Op_MinReductionV: 3453 case Op_MaxReductionV: 3454 case Op_AndReductionV: 3455 case Op_OrReductionV: 3456 case Op_XorReductionV: 3457 break; 3458 3459 case Op_PackB: 3460 case Op_PackS: 3461 case Op_PackI: 3462 case Op_PackF: 3463 case Op_PackL: 3464 case Op_PackD: 3465 if (n->req()-1 > 2) { 3466 // Replace many operand PackNodes with a binary tree for matching 3467 PackNode* p = (PackNode*) n; 3468 Node* btp = p->binary_tree_pack(1, n->req()); 3469 n->subsume_by(btp, this); 3470 } 3471 break; 3472 case Op_Loop: 3473 assert(!n->as_Loop()->is_transformed_long_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop"); 3474 case Op_CountedLoop: 3475 case Op_LongCountedLoop: 3476 case Op_OuterStripMinedLoop: 3477 if (n->as_Loop()->is_inner_loop()) { 3478 frc.inc_inner_loop_count(); 3479 } 3480 n->as_Loop()->verify_strip_mined(0); 3481 break; 3482 case Op_LShiftI: 3483 case Op_RShiftI: 3484 case Op_URShiftI: 3485 case Op_LShiftL: 3486 case Op_RShiftL: 3487 case Op_URShiftL: 3488 if (Matcher::need_masked_shift_count) { 3489 // The cpu's shift instructions don't restrict the count to the 3490 // lower 5/6 bits. We need to do the masking ourselves. 3491 Node* in2 = n->in(2); 3492 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3493 const TypeInt* t = in2->find_int_type(); 3494 if (t != NULL && t->is_con()) { 3495 juint shift = t->get_con(); 3496 if (shift > mask) { // Unsigned cmp 3497 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3498 } 3499 } else { 3500 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3501 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3502 n->set_req(2, shift); 3503 } 3504 } 3505 if (in2->outcnt() == 0) { // Remove dead node 3506 in2->disconnect_inputs(this); 3507 } 3508 } 3509 break; 3510 case Op_MemBarStoreStore: 3511 case Op_MemBarRelease: 3512 // Break the link with AllocateNode: it is no longer useful and 3513 // confuses register allocation. 3514 if (n->req() > MemBarNode::Precedent) { 3515 n->set_req(MemBarNode::Precedent, top()); 3516 } 3517 break; 3518 case Op_MemBarAcquire: { 3519 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3520 // At parse time, the trailing MemBarAcquire for a volatile load 3521 // is created with an edge to the load. After optimizations, 3522 // that input may be a chain of Phis. If those phis have no 3523 // other use, then the MemBarAcquire keeps them alive and 3524 // register allocation can be confused. 3525 dead_nodes.push(n->in(MemBarNode::Precedent)); 3526 n->set_req(MemBarNode::Precedent, top()); 3527 } 3528 break; 3529 } 3530 case Op_Blackhole: 3531 break; 3532 case Op_RangeCheck: { 3533 RangeCheckNode* rc = n->as_RangeCheck(); 3534 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3535 n->subsume_by(iff, this); 3536 frc._tests.push(iff); 3537 break; 3538 } 3539 case Op_ConvI2L: { 3540 if (!Matcher::convi2l_type_required) { 3541 // Code generation on some platforms doesn't need accurate 3542 // ConvI2L types. Widening the type can help remove redundant 3543 // address computations. 3544 n->as_Type()->set_type(TypeLong::INT); 3545 ResourceMark rm; 3546 Unique_Node_List wq; 3547 wq.push(n); 3548 for (uint next = 0; next < wq.size(); next++) { 3549 Node *m = wq.at(next); 3550 3551 for(;;) { 3552 // Loop over all nodes with identical inputs edges as m 3553 Node* k = m->find_similar(m->Opcode()); 3554 if (k == NULL) { 3555 break; 3556 } 3557 // Push their uses so we get a chance to remove node made 3558 // redundant 3559 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3560 Node* u = k->fast_out(i); 3561 if (u->Opcode() == Op_LShiftL || 3562 u->Opcode() == Op_AddL || 3563 u->Opcode() == Op_SubL || 3564 u->Opcode() == Op_AddP) { 3565 wq.push(u); 3566 } 3567 } 3568 // Replace all nodes with identical edges as m with m 3569 k->subsume_by(m, this); 3570 } 3571 } 3572 } 3573 break; 3574 } 3575 case Op_CmpUL: { 3576 if (!Matcher::has_match_rule(Op_CmpUL)) { 3577 // No support for unsigned long comparisons 3578 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3579 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3580 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3581 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3582 Node* andl = new AndLNode(orl, remove_sign_mask); 3583 Node* cmp = new CmpLNode(andl, n->in(2)); 3584 n->subsume_by(cmp, this); 3585 } 3586 break; 3587 } 3588 default: 3589 assert(!n->is_Call(), ""); 3590 assert(!n->is_Mem(), ""); 3591 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3592 break; 3593 } 3594 } 3595 3596 //------------------------------final_graph_reshaping_walk--------------------- 3597 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3598 // requires that the walk visits a node's inputs before visiting the node. 3599 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3600 Unique_Node_List sfpt; 3601 3602 frc._visited.set(root->_idx); // first, mark node as visited 3603 uint cnt = root->req(); 3604 Node *n = root; 3605 uint i = 0; 3606 while (true) { 3607 if (i < cnt) { 3608 // Place all non-visited non-null inputs onto stack 3609 Node* m = n->in(i); 3610 ++i; 3611 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3612 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3613 // compute worst case interpreter size in case of a deoptimization 3614 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3615 3616 sfpt.push(m); 3617 } 3618 cnt = m->req(); 3619 nstack.push(n, i); // put on stack parent and next input's index 3620 n = m; 3621 i = 0; 3622 } 3623 } else { 3624 // Now do post-visit work 3625 final_graph_reshaping_impl(n, frc, dead_nodes); 3626 if (nstack.is_empty()) 3627 break; // finished 3628 n = nstack.node(); // Get node from stack 3629 cnt = n->req(); 3630 i = nstack.index(); 3631 nstack.pop(); // Shift to the next node on stack 3632 } 3633 } 3634 3635 // Skip next transformation if compressed oops are not used. 3636 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3637 (!UseCompressedOops && !UseCompressedClassPointers)) 3638 return; 3639 3640 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3641 // It could be done for an uncommon traps or any safepoints/calls 3642 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3643 while (sfpt.size() > 0) { 3644 n = sfpt.pop(); 3645 JVMState *jvms = n->as_SafePoint()->jvms(); 3646 assert(jvms != NULL, "sanity"); 3647 int start = jvms->debug_start(); 3648 int end = n->req(); 3649 bool is_uncommon = (n->is_CallStaticJava() && 3650 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3651 for (int j = start; j < end; j++) { 3652 Node* in = n->in(j); 3653 if (in->is_DecodeNarrowPtr()) { 3654 bool safe_to_skip = true; 3655 if (!is_uncommon ) { 3656 // Is it safe to skip? 3657 for (uint i = 0; i < in->outcnt(); i++) { 3658 Node* u = in->raw_out(i); 3659 if (!u->is_SafePoint() || 3660 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3661 safe_to_skip = false; 3662 } 3663 } 3664 } 3665 if (safe_to_skip) { 3666 n->set_req(j, in->in(1)); 3667 } 3668 if (in->outcnt() == 0) { 3669 in->disconnect_inputs(this); 3670 } 3671 } 3672 } 3673 } 3674 } 3675 3676 //------------------------------final_graph_reshaping-------------------------- 3677 // Final Graph Reshaping. 3678 // 3679 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3680 // and not commoned up and forced early. Must come after regular 3681 // optimizations to avoid GVN undoing the cloning. Clone constant 3682 // inputs to Loop Phis; these will be split by the allocator anyways. 3683 // Remove Opaque nodes. 3684 // (2) Move last-uses by commutative operations to the left input to encourage 3685 // Intel update-in-place two-address operations and better register usage 3686 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3687 // calls canonicalizing them back. 3688 // (3) Count the number of double-precision FP ops, single-precision FP ops 3689 // and call sites. On Intel, we can get correct rounding either by 3690 // forcing singles to memory (requires extra stores and loads after each 3691 // FP bytecode) or we can set a rounding mode bit (requires setting and 3692 // clearing the mode bit around call sites). The mode bit is only used 3693 // if the relative frequency of single FP ops to calls is low enough. 3694 // This is a key transform for SPEC mpeg_audio. 3695 // (4) Detect infinite loops; blobs of code reachable from above but not 3696 // below. Several of the Code_Gen algorithms fail on such code shapes, 3697 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3698 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3699 // Detection is by looking for IfNodes where only 1 projection is 3700 // reachable from below or CatchNodes missing some targets. 3701 // (5) Assert for insane oop offsets in debug mode. 3702 3703 bool Compile::final_graph_reshaping() { 3704 // an infinite loop may have been eliminated by the optimizer, 3705 // in which case the graph will be empty. 3706 if (root()->req() == 1) { 3707 record_method_not_compilable("trivial infinite loop"); 3708 return true; 3709 } 3710 3711 // Expensive nodes have their control input set to prevent the GVN 3712 // from freely commoning them. There's no GVN beyond this point so 3713 // no need to keep the control input. We want the expensive nodes to 3714 // be freely moved to the least frequent code path by gcm. 3715 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3716 for (int i = 0; i < expensive_count(); i++) { 3717 _expensive_nodes.at(i)->set_req(0, NULL); 3718 } 3719 3720 Final_Reshape_Counts frc; 3721 3722 // Visit everybody reachable! 3723 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3724 Node_Stack nstack(live_nodes() >> 1); 3725 Unique_Node_List dead_nodes; 3726 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes); 3727 3728 // Check for unreachable (from below) code (i.e., infinite loops). 3729 for( uint i = 0; i < frc._tests.size(); i++ ) { 3730 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3731 // Get number of CFG targets. 3732 // Note that PCTables include exception targets after calls. 3733 uint required_outcnt = n->required_outcnt(); 3734 if (n->outcnt() != required_outcnt) { 3735 // Check for a few special cases. Rethrow Nodes never take the 3736 // 'fall-thru' path, so expected kids is 1 less. 3737 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3738 if (n->in(0)->in(0)->is_Call()) { 3739 CallNode* call = n->in(0)->in(0)->as_Call(); 3740 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3741 required_outcnt--; // Rethrow always has 1 less kid 3742 } else if (call->req() > TypeFunc::Parms && 3743 call->is_CallDynamicJava()) { 3744 // Check for null receiver. In such case, the optimizer has 3745 // detected that the virtual call will always result in a null 3746 // pointer exception. The fall-through projection of this CatchNode 3747 // will not be populated. 3748 Node* arg0 = call->in(TypeFunc::Parms); 3749 if (arg0->is_Type() && 3750 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3751 required_outcnt--; 3752 } 3753 } else if (call->entry_point() == OptoRuntime::new_array_Java() || 3754 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 3755 // Check for illegal array length. In such case, the optimizer has 3756 // detected that the allocation attempt will always result in an 3757 // exception. There is no fall-through projection of this CatchNode . 3758 assert(call->is_CallStaticJava(), "static call expected"); 3759 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 3760 uint valid_length_test_input = call->req() - 1; 3761 Node* valid_length_test = call->in(valid_length_test_input); 3762 call->del_req(valid_length_test_input); 3763 if (valid_length_test->find_int_con(1) == 0) { 3764 required_outcnt--; 3765 } 3766 dead_nodes.push(valid_length_test); 3767 assert(n->outcnt() == required_outcnt, "malformed control flow"); 3768 continue; 3769 } 3770 } 3771 } 3772 // Recheck with a better notion of 'required_outcnt' 3773 if (n->outcnt() != required_outcnt) { 3774 record_method_not_compilable("malformed control flow"); 3775 return true; // Not all targets reachable! 3776 } 3777 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) { 3778 CallNode* call = n->in(0)->in(0)->as_Call(); 3779 if (call->entry_point() == OptoRuntime::new_array_Java() || 3780 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 3781 assert(call->is_CallStaticJava(), "static call expected"); 3782 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 3783 uint valid_length_test_input = call->req() - 1; 3784 dead_nodes.push(call->in(valid_length_test_input)); 3785 call->del_req(valid_length_test_input); // valid length test useless now 3786 } 3787 } 3788 // Check that I actually visited all kids. Unreached kids 3789 // must be infinite loops. 3790 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3791 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3792 record_method_not_compilable("infinite loop"); 3793 return true; // Found unvisited kid; must be unreach 3794 } 3795 3796 // Here so verification code in final_graph_reshaping_walk() 3797 // always see an OuterStripMinedLoopEnd 3798 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) { 3799 IfNode* init_iff = n->as_If(); 3800 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3801 n->subsume_by(iff, this); 3802 } 3803 } 3804 3805 while (dead_nodes.size() > 0) { 3806 Node* m = dead_nodes.pop(); 3807 if (m->outcnt() == 0 && m != top()) { 3808 for (uint j = 0; j < m->req(); j++) { 3809 Node* in = m->in(j); 3810 if (in != NULL) { 3811 dead_nodes.push(in); 3812 } 3813 } 3814 m->disconnect_inputs(this); 3815 } 3816 } 3817 3818 #ifdef IA32 3819 // If original bytecodes contained a mixture of floats and doubles 3820 // check if the optimizer has made it homogenous, item (3). 3821 if (UseSSE == 0 && 3822 frc.get_float_count() > 32 && 3823 frc.get_double_count() == 0 && 3824 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3825 set_24_bit_selection_and_mode(false, true); 3826 } 3827 #endif // IA32 3828 3829 set_java_calls(frc.get_java_call_count()); 3830 set_inner_loops(frc.get_inner_loop_count()); 3831 3832 // No infinite loops, no reason to bail out. 3833 return false; 3834 } 3835 3836 //-----------------------------too_many_traps---------------------------------- 3837 // Report if there are too many traps at the current method and bci. 3838 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3839 bool Compile::too_many_traps(ciMethod* method, 3840 int bci, 3841 Deoptimization::DeoptReason reason) { 3842 ciMethodData* md = method->method_data(); 3843 if (md->is_empty()) { 3844 // Assume the trap has not occurred, or that it occurred only 3845 // because of a transient condition during start-up in the interpreter. 3846 return false; 3847 } 3848 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3849 if (md->has_trap_at(bci, m, reason) != 0) { 3850 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3851 // Also, if there are multiple reasons, or if there is no per-BCI record, 3852 // assume the worst. 3853 if (log()) 3854 log()->elem("observe trap='%s' count='%d'", 3855 Deoptimization::trap_reason_name(reason), 3856 md->trap_count(reason)); 3857 return true; 3858 } else { 3859 // Ignore method/bci and see if there have been too many globally. 3860 return too_many_traps(reason, md); 3861 } 3862 } 3863 3864 // Less-accurate variant which does not require a method and bci. 3865 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3866 ciMethodData* logmd) { 3867 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3868 // Too many traps globally. 3869 // Note that we use cumulative trap_count, not just md->trap_count. 3870 if (log()) { 3871 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3872 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3873 Deoptimization::trap_reason_name(reason), 3874 mcount, trap_count(reason)); 3875 } 3876 return true; 3877 } else { 3878 // The coast is clear. 3879 return false; 3880 } 3881 } 3882 3883 //--------------------------too_many_recompiles-------------------------------- 3884 // Report if there are too many recompiles at the current method and bci. 3885 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3886 // Is not eager to return true, since this will cause the compiler to use 3887 // Action_none for a trap point, to avoid too many recompilations. 3888 bool Compile::too_many_recompiles(ciMethod* method, 3889 int bci, 3890 Deoptimization::DeoptReason reason) { 3891 ciMethodData* md = method->method_data(); 3892 if (md->is_empty()) { 3893 // Assume the trap has not occurred, or that it occurred only 3894 // because of a transient condition during start-up in the interpreter. 3895 return false; 3896 } 3897 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3898 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3899 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3900 Deoptimization::DeoptReason per_bc_reason 3901 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3902 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3903 if ((per_bc_reason == Deoptimization::Reason_none 3904 || md->has_trap_at(bci, m, reason) != 0) 3905 // The trap frequency measure we care about is the recompile count: 3906 && md->trap_recompiled_at(bci, m) 3907 && md->overflow_recompile_count() >= bc_cutoff) { 3908 // Do not emit a trap here if it has already caused recompilations. 3909 // Also, if there are multiple reasons, or if there is no per-BCI record, 3910 // assume the worst. 3911 if (log()) 3912 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3913 Deoptimization::trap_reason_name(reason), 3914 md->trap_count(reason), 3915 md->overflow_recompile_count()); 3916 return true; 3917 } else if (trap_count(reason) != 0 3918 && decompile_count() >= m_cutoff) { 3919 // Too many recompiles globally, and we have seen this sort of trap. 3920 // Use cumulative decompile_count, not just md->decompile_count. 3921 if (log()) 3922 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3923 Deoptimization::trap_reason_name(reason), 3924 md->trap_count(reason), trap_count(reason), 3925 md->decompile_count(), decompile_count()); 3926 return true; 3927 } else { 3928 // The coast is clear. 3929 return false; 3930 } 3931 } 3932 3933 // Compute when not to trap. Used by matching trap based nodes and 3934 // NullCheck optimization. 3935 void Compile::set_allowed_deopt_reasons() { 3936 _allowed_reasons = 0; 3937 if (is_method_compilation()) { 3938 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3939 assert(rs < BitsPerInt, "recode bit map"); 3940 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3941 _allowed_reasons |= nth_bit(rs); 3942 } 3943 } 3944 } 3945 } 3946 3947 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 3948 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 3949 } 3950 3951 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 3952 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 3953 } 3954 3955 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 3956 if (holder->is_initialized()) { 3957 return false; 3958 } 3959 if (holder->is_being_initialized()) { 3960 if (accessing_method->holder() == holder) { 3961 // Access inside a class. The barrier can be elided when access happens in <clinit>, 3962 // <init>, or a static method. In all those cases, there was an initialization 3963 // barrier on the holder klass passed. 3964 if (accessing_method->is_static_initializer() || 3965 accessing_method->is_object_initializer() || 3966 accessing_method->is_static()) { 3967 return false; 3968 } 3969 } else if (accessing_method->holder()->is_subclass_of(holder)) { 3970 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 3971 // In case of <init> or a static method, the barrier is on the subclass is not enough: 3972 // child class can become fully initialized while its parent class is still being initialized. 3973 if (accessing_method->is_static_initializer()) { 3974 return false; 3975 } 3976 } 3977 ciMethod* root = method(); // the root method of compilation 3978 if (root != accessing_method) { 3979 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 3980 } 3981 } 3982 return true; 3983 } 3984 3985 #ifndef PRODUCT 3986 //------------------------------verify_graph_edges--------------------------- 3987 // Walk the Graph and verify that there is a one-to-one correspondence 3988 // between Use-Def edges and Def-Use edges in the graph. 3989 void Compile::verify_graph_edges(bool no_dead_code) { 3990 if (VerifyGraphEdges) { 3991 Unique_Node_List visited; 3992 // Call recursive graph walk to check edges 3993 _root->verify_edges(visited); 3994 if (no_dead_code) { 3995 // Now make sure that no visited node is used by an unvisited node. 3996 bool dead_nodes = false; 3997 Unique_Node_List checked; 3998 while (visited.size() > 0) { 3999 Node* n = visited.pop(); 4000 checked.push(n); 4001 for (uint i = 0; i < n->outcnt(); i++) { 4002 Node* use = n->raw_out(i); 4003 if (checked.member(use)) continue; // already checked 4004 if (visited.member(use)) continue; // already in the graph 4005 if (use->is_Con()) continue; // a dead ConNode is OK 4006 // At this point, we have found a dead node which is DU-reachable. 4007 if (!dead_nodes) { 4008 tty->print_cr("*** Dead nodes reachable via DU edges:"); 4009 dead_nodes = true; 4010 } 4011 use->dump(2); 4012 tty->print_cr("---"); 4013 checked.push(use); // No repeats; pretend it is now checked. 4014 } 4015 } 4016 assert(!dead_nodes, "using nodes must be reachable from root"); 4017 } 4018 } 4019 } 4020 #endif 4021 4022 // The Compile object keeps track of failure reasons separately from the ciEnv. 4023 // This is required because there is not quite a 1-1 relation between the 4024 // ciEnv and its compilation task and the Compile object. Note that one 4025 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4026 // to backtrack and retry without subsuming loads. Other than this backtracking 4027 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4028 // by the logic in C2Compiler. 4029 void Compile::record_failure(const char* reason) { 4030 if (log() != NULL) { 4031 log()->elem("failure reason='%s' phase='compile'", reason); 4032 } 4033 if (_failure_reason == NULL) { 4034 // Record the first failure reason. 4035 _failure_reason = reason; 4036 } 4037 4038 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4039 C->print_method(PHASE_FAILURE); 4040 } 4041 _root = NULL; // flush the graph, too 4042 } 4043 4044 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 4045 : TraceTime(name, accumulator, CITime, CITimeVerbose), 4046 _phase_name(name), _dolog(CITimeVerbose) 4047 { 4048 if (_dolog) { 4049 C = Compile::current(); 4050 _log = C->log(); 4051 } else { 4052 C = NULL; 4053 _log = NULL; 4054 } 4055 if (_log != NULL) { 4056 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4057 _log->stamp(); 4058 _log->end_head(); 4059 } 4060 } 4061 4062 Compile::TracePhase::~TracePhase() { 4063 4064 C = Compile::current(); 4065 if (_dolog) { 4066 _log = C->log(); 4067 } else { 4068 _log = NULL; 4069 } 4070 4071 #ifdef ASSERT 4072 if (PrintIdealNodeCount) { 4073 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4074 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 4075 } 4076 4077 if (VerifyIdealNodeCount) { 4078 Compile::current()->print_missing_nodes(); 4079 } 4080 #endif 4081 4082 if (_log != NULL) { 4083 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4084 } 4085 } 4086 4087 //----------------------------static_subtype_check----------------------------- 4088 // Shortcut important common cases when superklass is exact: 4089 // (0) superklass is java.lang.Object (can occur in reflective code) 4090 // (1) subklass is already limited to a subtype of superklass => always ok 4091 // (2) subklass does not overlap with superklass => always fail 4092 // (3) superklass has NO subtypes and we can check with a simple compare. 4093 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4094 if (StressReflectiveCode) { 4095 return SSC_full_test; // Let caller generate the general case. 4096 } 4097 4098 if (superk == env()->Object_klass()) { 4099 return SSC_always_true; // (0) this test cannot fail 4100 } 4101 4102 ciType* superelem = superk; 4103 ciType* subelem = subk; 4104 if (superelem->is_array_klass()) { 4105 superelem = superelem->as_array_klass()->base_element_type(); 4106 } 4107 if (subelem->is_array_klass()) { 4108 subelem = subelem->as_array_klass()->base_element_type(); 4109 } 4110 4111 if (!subk->is_interface()) { // cannot trust static interface types yet 4112 if (subk->is_subtype_of(superk)) { 4113 return SSC_always_true; // (1) false path dead; no dynamic test needed 4114 } 4115 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4116 !(subelem->is_klass() && subelem->as_klass()->is_interface()) && 4117 !superk->is_subtype_of(subk)) { 4118 return SSC_always_false; // (2) true path dead; no dynamic test needed 4119 } 4120 } 4121 4122 // If casting to an instance klass, it must have no subtypes 4123 if (superk->is_interface()) { 4124 // Cannot trust interfaces yet. 4125 // %%% S.B. superk->nof_implementors() == 1 4126 } else if (superelem->is_instance_klass()) { 4127 ciInstanceKlass* ik = superelem->as_instance_klass(); 4128 if (!ik->has_subklass() && !ik->is_interface()) { 4129 if (!ik->is_final()) { 4130 // Add a dependency if there is a chance of a later subclass. 4131 dependencies()->assert_leaf_type(ik); 4132 } 4133 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4134 } 4135 } else { 4136 // A primitive array type has no subtypes. 4137 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4138 } 4139 4140 return SSC_full_test; 4141 } 4142 4143 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4144 #ifdef _LP64 4145 // The scaled index operand to AddP must be a clean 64-bit value. 4146 // Java allows a 32-bit int to be incremented to a negative 4147 // value, which appears in a 64-bit register as a large 4148 // positive number. Using that large positive number as an 4149 // operand in pointer arithmetic has bad consequences. 4150 // On the other hand, 32-bit overflow is rare, and the possibility 4151 // can often be excluded, if we annotate the ConvI2L node with 4152 // a type assertion that its value is known to be a small positive 4153 // number. (The prior range check has ensured this.) 4154 // This assertion is used by ConvI2LNode::Ideal. 4155 int index_max = max_jint - 1; // array size is max_jint, index is one less 4156 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4157 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4158 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4159 #endif 4160 return idx; 4161 } 4162 4163 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4164 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 4165 if (ctrl != NULL) { 4166 // Express control dependency by a CastII node with a narrow type. 4167 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */); 4168 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4169 // node from floating above the range check during loop optimizations. Otherwise, the 4170 // ConvI2L node may be eliminated independently of the range check, causing the data path 4171 // to become TOP while the control path is still there (although it's unreachable). 4172 value->set_req(0, ctrl); 4173 value = phase->transform(value); 4174 } 4175 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4176 return phase->transform(new ConvI2LNode(value, ltype)); 4177 } 4178 4179 void Compile::print_inlining_stream_free() { 4180 if (_print_inlining_stream != NULL) { 4181 _print_inlining_stream->~stringStream(); 4182 _print_inlining_stream = NULL; 4183 } 4184 } 4185 4186 // The message about the current inlining is accumulated in 4187 // _print_inlining_stream and transfered into the _print_inlining_list 4188 // once we know whether inlining succeeds or not. For regular 4189 // inlining, messages are appended to the buffer pointed by 4190 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4191 // a new buffer is added after _print_inlining_idx in the list. This 4192 // way we can update the inlining message for late inlining call site 4193 // when the inlining is attempted again. 4194 void Compile::print_inlining_init() { 4195 if (print_inlining() || print_intrinsics()) { 4196 // print_inlining_init is actually called several times. 4197 print_inlining_stream_free(); 4198 _print_inlining_stream = new stringStream(); 4199 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer()); 4200 } 4201 } 4202 4203 void Compile::print_inlining_reinit() { 4204 if (print_inlining() || print_intrinsics()) { 4205 print_inlining_stream_free(); 4206 // Re allocate buffer when we change ResourceMark 4207 _print_inlining_stream = new stringStream(); 4208 } 4209 } 4210 4211 void Compile::print_inlining_reset() { 4212 _print_inlining_stream->reset(); 4213 } 4214 4215 void Compile::print_inlining_commit() { 4216 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4217 // Transfer the message from _print_inlining_stream to the current 4218 // _print_inlining_list buffer and clear _print_inlining_stream. 4219 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 4220 print_inlining_reset(); 4221 } 4222 4223 void Compile::print_inlining_push() { 4224 // Add new buffer to the _print_inlining_list at current position 4225 _print_inlining_idx++; 4226 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer()); 4227 } 4228 4229 Compile::PrintInliningBuffer* Compile::print_inlining_current() { 4230 return _print_inlining_list->at(_print_inlining_idx); 4231 } 4232 4233 void Compile::print_inlining_update(CallGenerator* cg) { 4234 if (print_inlining() || print_intrinsics()) { 4235 if (cg->is_late_inline()) { 4236 if (print_inlining_current()->cg() != cg && 4237 (print_inlining_current()->cg() != NULL || 4238 print_inlining_current()->ss()->size() != 0)) { 4239 print_inlining_push(); 4240 } 4241 print_inlining_commit(); 4242 print_inlining_current()->set_cg(cg); 4243 } else { 4244 if (print_inlining_current()->cg() != NULL) { 4245 print_inlining_push(); 4246 } 4247 print_inlining_commit(); 4248 } 4249 } 4250 } 4251 4252 void Compile::print_inlining_move_to(CallGenerator* cg) { 4253 // We resume inlining at a late inlining call site. Locate the 4254 // corresponding inlining buffer so that we can update it. 4255 if (print_inlining() || print_intrinsics()) { 4256 for (int i = 0; i < _print_inlining_list->length(); i++) { 4257 if (_print_inlining_list->at(i)->cg() == cg) { 4258 _print_inlining_idx = i; 4259 return; 4260 } 4261 } 4262 ShouldNotReachHere(); 4263 } 4264 } 4265 4266 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4267 if (print_inlining() || print_intrinsics()) { 4268 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4269 assert(print_inlining_current()->cg() == cg, "wrong entry"); 4270 // replace message with new message 4271 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer()); 4272 print_inlining_commit(); 4273 print_inlining_current()->set_cg(cg); 4274 } 4275 } 4276 4277 void Compile::print_inlining_assert_ready() { 4278 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4279 } 4280 4281 void Compile::process_print_inlining() { 4282 assert(_late_inlines.length() == 0, "not drained yet"); 4283 if (print_inlining() || print_intrinsics()) { 4284 ResourceMark rm; 4285 stringStream ss; 4286 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once."); 4287 for (int i = 0; i < _print_inlining_list->length(); i++) { 4288 PrintInliningBuffer* pib = _print_inlining_list->at(i); 4289 ss.print("%s", pib->ss()->as_string()); 4290 delete pib; 4291 DEBUG_ONLY(_print_inlining_list->at_put(i, NULL)); 4292 } 4293 // Reset _print_inlining_list, it only contains destructed objects. 4294 // It is on the arena, so it will be freed when the arena is reset. 4295 _print_inlining_list = NULL; 4296 // _print_inlining_stream won't be used anymore, either. 4297 print_inlining_stream_free(); 4298 size_t end = ss.size(); 4299 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4300 strncpy(_print_inlining_output, ss.base(), end+1); 4301 _print_inlining_output[end] = 0; 4302 } 4303 } 4304 4305 void Compile::dump_print_inlining() { 4306 if (_print_inlining_output != NULL) { 4307 tty->print_raw(_print_inlining_output); 4308 } 4309 } 4310 4311 void Compile::log_late_inline(CallGenerator* cg) { 4312 if (log() != NULL) { 4313 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4314 cg->unique_id()); 4315 JVMState* p = cg->call_node()->jvms(); 4316 while (p != NULL) { 4317 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4318 p = p->caller(); 4319 } 4320 log()->tail("late_inline"); 4321 } 4322 } 4323 4324 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4325 log_late_inline(cg); 4326 if (log() != NULL) { 4327 log()->inline_fail(msg); 4328 } 4329 } 4330 4331 void Compile::log_inline_id(CallGenerator* cg) { 4332 if (log() != NULL) { 4333 // The LogCompilation tool needs a unique way to identify late 4334 // inline call sites. This id must be unique for this call site in 4335 // this compilation. Try to have it unique across compilations as 4336 // well because it can be convenient when grepping through the log 4337 // file. 4338 // Distinguish OSR compilations from others in case CICountOSR is 4339 // on. 4340 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4341 cg->set_unique_id(id); 4342 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4343 } 4344 } 4345 4346 void Compile::log_inline_failure(const char* msg) { 4347 if (C->log() != NULL) { 4348 C->log()->inline_fail(msg); 4349 } 4350 } 4351 4352 4353 // Dump inlining replay data to the stream. 4354 // Don't change thread state and acquire any locks. 4355 void Compile::dump_inline_data(outputStream* out) { 4356 InlineTree* inl_tree = ilt(); 4357 if (inl_tree != NULL) { 4358 out->print(" inline %d", inl_tree->count()); 4359 inl_tree->dump_replay_data(out); 4360 } 4361 } 4362 4363 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4364 if (n1->Opcode() < n2->Opcode()) return -1; 4365 else if (n1->Opcode() > n2->Opcode()) return 1; 4366 4367 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4368 for (uint i = 1; i < n1->req(); i++) { 4369 if (n1->in(i) < n2->in(i)) return -1; 4370 else if (n1->in(i) > n2->in(i)) return 1; 4371 } 4372 4373 return 0; 4374 } 4375 4376 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4377 Node* n1 = *n1p; 4378 Node* n2 = *n2p; 4379 4380 return cmp_expensive_nodes(n1, n2); 4381 } 4382 4383 void Compile::sort_expensive_nodes() { 4384 if (!expensive_nodes_sorted()) { 4385 _expensive_nodes.sort(cmp_expensive_nodes); 4386 } 4387 } 4388 4389 bool Compile::expensive_nodes_sorted() const { 4390 for (int i = 1; i < _expensive_nodes.length(); i++) { 4391 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) { 4392 return false; 4393 } 4394 } 4395 return true; 4396 } 4397 4398 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4399 if (_expensive_nodes.length() == 0) { 4400 return false; 4401 } 4402 4403 assert(OptimizeExpensiveOps, "optimization off?"); 4404 4405 // Take this opportunity to remove dead nodes from the list 4406 int j = 0; 4407 for (int i = 0; i < _expensive_nodes.length(); i++) { 4408 Node* n = _expensive_nodes.at(i); 4409 if (!n->is_unreachable(igvn)) { 4410 assert(n->is_expensive(), "should be expensive"); 4411 _expensive_nodes.at_put(j, n); 4412 j++; 4413 } 4414 } 4415 _expensive_nodes.trunc_to(j); 4416 4417 // Then sort the list so that similar nodes are next to each other 4418 // and check for at least two nodes of identical kind with same data 4419 // inputs. 4420 sort_expensive_nodes(); 4421 4422 for (int i = 0; i < _expensive_nodes.length()-1; i++) { 4423 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) { 4424 return true; 4425 } 4426 } 4427 4428 return false; 4429 } 4430 4431 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4432 if (_expensive_nodes.length() == 0) { 4433 return; 4434 } 4435 4436 assert(OptimizeExpensiveOps, "optimization off?"); 4437 4438 // Sort to bring similar nodes next to each other and clear the 4439 // control input of nodes for which there's only a single copy. 4440 sort_expensive_nodes(); 4441 4442 int j = 0; 4443 int identical = 0; 4444 int i = 0; 4445 bool modified = false; 4446 for (; i < _expensive_nodes.length()-1; i++) { 4447 assert(j <= i, "can't write beyond current index"); 4448 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) { 4449 identical++; 4450 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4451 continue; 4452 } 4453 if (identical > 0) { 4454 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4455 identical = 0; 4456 } else { 4457 Node* n = _expensive_nodes.at(i); 4458 igvn.replace_input_of(n, 0, NULL); 4459 igvn.hash_insert(n); 4460 modified = true; 4461 } 4462 } 4463 if (identical > 0) { 4464 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4465 } else if (_expensive_nodes.length() >= 1) { 4466 Node* n = _expensive_nodes.at(i); 4467 igvn.replace_input_of(n, 0, NULL); 4468 igvn.hash_insert(n); 4469 modified = true; 4470 } 4471 _expensive_nodes.trunc_to(j); 4472 if (modified) { 4473 igvn.optimize(); 4474 } 4475 } 4476 4477 void Compile::add_expensive_node(Node * n) { 4478 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list"); 4479 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4480 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4481 if (OptimizeExpensiveOps) { 4482 _expensive_nodes.append(n); 4483 } else { 4484 // Clear control input and let IGVN optimize expensive nodes if 4485 // OptimizeExpensiveOps is off. 4486 n->set_req(0, NULL); 4487 } 4488 } 4489 4490 /** 4491 * Track coarsened Lock and Unlock nodes. 4492 */ 4493 4494 class Lock_List : public Node_List { 4495 uint _origin_cnt; 4496 public: 4497 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {} 4498 uint origin_cnt() const { return _origin_cnt; } 4499 }; 4500 4501 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) { 4502 int length = locks.length(); 4503 if (length > 0) { 4504 // Have to keep this list until locks elimination during Macro nodes elimination. 4505 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length); 4506 for (int i = 0; i < length; i++) { 4507 AbstractLockNode* lock = locks.at(i); 4508 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx); 4509 locks_list->push(lock); 4510 } 4511 _coarsened_locks.append(locks_list); 4512 } 4513 } 4514 4515 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) { 4516 int count = coarsened_count(); 4517 for (int i = 0; i < count; i++) { 4518 Node_List* locks_list = _coarsened_locks.at(i); 4519 for (uint j = 0; j < locks_list->size(); j++) { 4520 Node* lock = locks_list->at(j); 4521 assert(lock->is_AbstractLock(), "sanity"); 4522 if (!useful.member(lock)) { 4523 locks_list->yank(lock); 4524 } 4525 } 4526 } 4527 } 4528 4529 void Compile::remove_coarsened_lock(Node* n) { 4530 if (n->is_AbstractLock()) { 4531 int count = coarsened_count(); 4532 for (int i = 0; i < count; i++) { 4533 Node_List* locks_list = _coarsened_locks.at(i); 4534 locks_list->yank(n); 4535 } 4536 } 4537 } 4538 4539 bool Compile::coarsened_locks_consistent() { 4540 int count = coarsened_count(); 4541 for (int i = 0; i < count; i++) { 4542 bool unbalanced = false; 4543 bool modified = false; // track locks kind modifications 4544 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i); 4545 uint size = locks_list->size(); 4546 if (size == 0) { 4547 unbalanced = false; // All locks were eliminated - good 4548 } else if (size != locks_list->origin_cnt()) { 4549 unbalanced = true; // Some locks were removed from list 4550 } else { 4551 for (uint j = 0; j < size; j++) { 4552 Node* lock = locks_list->at(j); 4553 // All nodes in group should have the same state (modified or not) 4554 if (!lock->as_AbstractLock()->is_coarsened()) { 4555 if (j == 0) { 4556 // first on list was modified, the rest should be too for consistency 4557 modified = true; 4558 } else if (!modified) { 4559 // this lock was modified but previous locks on the list were not 4560 unbalanced = true; 4561 break; 4562 } 4563 } else if (modified) { 4564 // previous locks on list were modified but not this lock 4565 unbalanced = true; 4566 break; 4567 } 4568 } 4569 } 4570 if (unbalanced) { 4571 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified 4572 #ifdef ASSERT 4573 if (PrintEliminateLocks) { 4574 tty->print_cr("=== unbalanced coarsened locks ==="); 4575 for (uint l = 0; l < size; l++) { 4576 locks_list->at(l)->dump(); 4577 } 4578 } 4579 #endif 4580 record_failure(C2Compiler::retry_no_locks_coarsening()); 4581 return false; 4582 } 4583 } 4584 return true; 4585 } 4586 4587 /** 4588 * Remove the speculative part of types and clean up the graph 4589 */ 4590 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4591 if (UseTypeSpeculation) { 4592 Unique_Node_List worklist; 4593 worklist.push(root()); 4594 int modified = 0; 4595 // Go over all type nodes that carry a speculative type, drop the 4596 // speculative part of the type and enqueue the node for an igvn 4597 // which may optimize it out. 4598 for (uint next = 0; next < worklist.size(); ++next) { 4599 Node *n = worklist.at(next); 4600 if (n->is_Type()) { 4601 TypeNode* tn = n->as_Type(); 4602 const Type* t = tn->type(); 4603 const Type* t_no_spec = t->remove_speculative(); 4604 if (t_no_spec != t) { 4605 bool in_hash = igvn.hash_delete(n); 4606 assert(in_hash, "node should be in igvn hash table"); 4607 tn->set_type(t_no_spec); 4608 igvn.hash_insert(n); 4609 igvn._worklist.push(n); // give it a chance to go away 4610 modified++; 4611 } 4612 } 4613 // Iterate over outs - endless loops is unreachable from below 4614 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4615 Node *m = n->fast_out(i); 4616 if (not_a_node(m)) { 4617 continue; 4618 } 4619 worklist.push(m); 4620 } 4621 } 4622 // Drop the speculative part of all types in the igvn's type table 4623 igvn.remove_speculative_types(); 4624 if (modified > 0) { 4625 igvn.optimize(); 4626 } 4627 #ifdef ASSERT 4628 // Verify that after the IGVN is over no speculative type has resurfaced 4629 worklist.clear(); 4630 worklist.push(root()); 4631 for (uint next = 0; next < worklist.size(); ++next) { 4632 Node *n = worklist.at(next); 4633 const Type* t = igvn.type_or_null(n); 4634 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4635 if (n->is_Type()) { 4636 t = n->as_Type()->type(); 4637 assert(t == t->remove_speculative(), "no more speculative types"); 4638 } 4639 // Iterate over outs - endless loops is unreachable from below 4640 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4641 Node *m = n->fast_out(i); 4642 if (not_a_node(m)) { 4643 continue; 4644 } 4645 worklist.push(m); 4646 } 4647 } 4648 igvn.check_no_speculative_types(); 4649 #endif 4650 } 4651 } 4652 4653 // Auxiliary methods to support randomized stressing/fuzzing. 4654 4655 int Compile::random() { 4656 _stress_seed = os::next_random(_stress_seed); 4657 return static_cast<int>(_stress_seed); 4658 } 4659 4660 // This method can be called the arbitrary number of times, with current count 4661 // as the argument. The logic allows selecting a single candidate from the 4662 // running list of candidates as follows: 4663 // int count = 0; 4664 // Cand* selected = null; 4665 // while(cand = cand->next()) { 4666 // if (randomized_select(++count)) { 4667 // selected = cand; 4668 // } 4669 // } 4670 // 4671 // Including count equalizes the chances any candidate is "selected". 4672 // This is useful when we don't have the complete list of candidates to choose 4673 // from uniformly. In this case, we need to adjust the randomicity of the 4674 // selection, or else we will end up biasing the selection towards the latter 4675 // candidates. 4676 // 4677 // Quick back-envelope calculation shows that for the list of n candidates 4678 // the equal probability for the candidate to persist as "best" can be 4679 // achieved by replacing it with "next" k-th candidate with the probability 4680 // of 1/k. It can be easily shown that by the end of the run, the 4681 // probability for any candidate is converged to 1/n, thus giving the 4682 // uniform distribution among all the candidates. 4683 // 4684 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4685 #define RANDOMIZED_DOMAIN_POW 29 4686 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4687 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4688 bool Compile::randomized_select(int count) { 4689 assert(count > 0, "only positive"); 4690 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4691 } 4692 4693 CloneMap& Compile::clone_map() { return _clone_map; } 4694 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4695 4696 void NodeCloneInfo::dump() const { 4697 tty->print(" {%d:%d} ", idx(), gen()); 4698 } 4699 4700 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4701 uint64_t val = value(old->_idx); 4702 NodeCloneInfo cio(val); 4703 assert(val != 0, "old node should be in the map"); 4704 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4705 insert(nnn->_idx, cin.get()); 4706 #ifndef PRODUCT 4707 if (is_debug()) { 4708 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4709 } 4710 #endif 4711 } 4712 4713 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4714 NodeCloneInfo cio(value(old->_idx)); 4715 if (cio.get() == 0) { 4716 cio.set(old->_idx, 0); 4717 insert(old->_idx, cio.get()); 4718 #ifndef PRODUCT 4719 if (is_debug()) { 4720 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4721 } 4722 #endif 4723 } 4724 clone(old, nnn, gen); 4725 } 4726 4727 int CloneMap::max_gen() const { 4728 int g = 0; 4729 DictI di(_dict); 4730 for(; di.test(); ++di) { 4731 int t = gen(di._key); 4732 if (g < t) { 4733 g = t; 4734 #ifndef PRODUCT 4735 if (is_debug()) { 4736 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4737 } 4738 #endif 4739 } 4740 } 4741 return g; 4742 } 4743 4744 void CloneMap::dump(node_idx_t key) const { 4745 uint64_t val = value(key); 4746 if (val != 0) { 4747 NodeCloneInfo ni(val); 4748 ni.dump(); 4749 } 4750 } 4751 4752 // Move Allocate nodes to the start of the list 4753 void Compile::sort_macro_nodes() { 4754 int count = macro_count(); 4755 int allocates = 0; 4756 for (int i = 0; i < count; i++) { 4757 Node* n = macro_node(i); 4758 if (n->is_Allocate()) { 4759 if (i != allocates) { 4760 Node* tmp = macro_node(allocates); 4761 _macro_nodes.at_put(allocates, n); 4762 _macro_nodes.at_put(i, tmp); 4763 } 4764 allocates++; 4765 } 4766 } 4767 } 4768 4769 void Compile::print_method(CompilerPhaseType cpt, const char *name, int level) { 4770 EventCompilerPhase event; 4771 if (event.should_commit()) { 4772 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 4773 } 4774 #ifndef PRODUCT 4775 if (should_print(level)) { 4776 _printer->print_method(name, level); 4777 } 4778 #endif 4779 C->_latest_stage_start_counter.stamp(); 4780 } 4781 4782 void Compile::print_method(CompilerPhaseType cpt, int level, int idx) { 4783 char output[1024]; 4784 #ifndef PRODUCT 4785 if (idx != 0) { 4786 jio_snprintf(output, sizeof(output), "%s:%d", CompilerPhaseTypeHelper::to_string(cpt), idx); 4787 } else { 4788 jio_snprintf(output, sizeof(output), "%s", CompilerPhaseTypeHelper::to_string(cpt)); 4789 } 4790 #endif 4791 print_method(cpt, output, level); 4792 } 4793 4794 void Compile::print_method(CompilerPhaseType cpt, Node* n, int level) { 4795 ResourceMark rm; 4796 stringStream ss; 4797 ss.print_raw(CompilerPhaseTypeHelper::to_string(cpt)); 4798 if (n != NULL) { 4799 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]); 4800 } else { 4801 ss.print_raw(": NULL"); 4802 } 4803 C->print_method(cpt, ss.as_string(), level); 4804 } 4805 4806 void Compile::end_method(int level) { 4807 EventCompilerPhase event; 4808 if (event.should_commit()) { 4809 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, level); 4810 } 4811 4812 #ifndef PRODUCT 4813 if (_method != NULL && should_print(level)) { 4814 _printer->end_method(); 4815 } 4816 #endif 4817 } 4818 4819 4820 #ifndef PRODUCT 4821 IdealGraphPrinter* Compile::_debug_file_printer = NULL; 4822 IdealGraphPrinter* Compile::_debug_network_printer = NULL; 4823 4824 // Called from debugger. Prints method to the default file with the default phase name. 4825 // This works regardless of any Ideal Graph Visualizer flags set or not. 4826 void igv_print() { 4827 Compile::current()->igv_print_method_to_file(); 4828 } 4829 4830 // Same as igv_print() above but with a specified phase name. 4831 void igv_print(const char* phase_name) { 4832 Compile::current()->igv_print_method_to_file(phase_name); 4833 } 4834 4835 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 4836 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 4837 // This works regardless of any Ideal Graph Visualizer flags set or not. 4838 void igv_print(bool network) { 4839 if (network) { 4840 Compile::current()->igv_print_method_to_network(); 4841 } else { 4842 Compile::current()->igv_print_method_to_file(); 4843 } 4844 } 4845 4846 // Same as igv_print(bool network) above but with a specified phase name. 4847 void igv_print(bool network, const char* phase_name) { 4848 if (network) { 4849 Compile::current()->igv_print_method_to_network(phase_name); 4850 } else { 4851 Compile::current()->igv_print_method_to_file(phase_name); 4852 } 4853 } 4854 4855 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 4856 void igv_print_default() { 4857 Compile::current()->print_method(PHASE_DEBUG, 0); 4858 } 4859 4860 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 4861 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 4862 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 4863 void igv_append() { 4864 Compile::current()->igv_print_method_to_file("Debug", true); 4865 } 4866 4867 // Same as igv_append() above but with a specified phase name. 4868 void igv_append(const char* phase_name) { 4869 Compile::current()->igv_print_method_to_file(phase_name, true); 4870 } 4871 4872 void Compile::igv_print_method_to_file(const char* phase_name, bool append) { 4873 const char* file_name = "custom_debug.xml"; 4874 if (_debug_file_printer == NULL) { 4875 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 4876 } else { 4877 _debug_file_printer->update_compiled_method(C->method()); 4878 } 4879 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 4880 _debug_file_printer->print(phase_name, (Node*)C->root()); 4881 } 4882 4883 void Compile::igv_print_method_to_network(const char* phase_name) { 4884 if (_debug_network_printer == NULL) { 4885 _debug_network_printer = new IdealGraphPrinter(C); 4886 } else { 4887 _debug_network_printer->update_compiled_method(C->method()); 4888 } 4889 tty->print_cr("Method printed over network stream to IGV"); 4890 _debug_network_printer->print(phase_name, (Node*)C->root()); 4891 } 4892 #endif 4893 4894 void Compile::add_native_invoker(RuntimeStub* stub) { 4895 _native_invokers.append(stub); 4896 } 4897 4898 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 4899 if (type != NULL && phase->type(value)->higher_equal(type)) { 4900 return value; 4901 } 4902 Node* result = NULL; 4903 if (bt == T_BYTE) { 4904 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 4905 result = new RShiftINode(result, phase->intcon(24)); 4906 } else if (bt == T_BOOLEAN) { 4907 result = new AndINode(value, phase->intcon(0xFF)); 4908 } else if (bt == T_CHAR) { 4909 result = new AndINode(value,phase->intcon(0xFFFF)); 4910 } else { 4911 assert(bt == T_SHORT, "unexpected narrow type"); 4912 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 4913 result = new RShiftINode(result, phase->intcon(16)); 4914 } 4915 if (transform_res) { 4916 result = phase->transform(result); 4917 } 4918 return result; 4919 } 4920