1 /* 2 * Copyright (c) 1997, 2025, 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 "asm/macroAssembler.hpp" 26 #include "asm/macroAssembler.inline.hpp" 27 #include "ci/ciReplay.hpp" 28 #include "classfile/javaClasses.hpp" 29 #include "code/exceptionHandlerTable.hpp" 30 #include "code/nmethod.hpp" 31 #include "compiler/compilationFailureInfo.hpp" 32 #include "compiler/compilationMemoryStatistic.hpp" 33 #include "compiler/compileBroker.hpp" 34 #include "compiler/compileLog.hpp" 35 #include "compiler/compilerOracle.hpp" 36 #include "compiler/compiler_globals.hpp" 37 #include "compiler/disassembler.hpp" 38 #include "compiler/oopMap.hpp" 39 #include "gc/shared/barrierSet.hpp" 40 #include "gc/shared/c2/barrierSetC2.hpp" 41 #include "jfr/jfrEvents.hpp" 42 #include "jvm_io.h" 43 #include "memory/allocation.hpp" 44 #include "memory/resourceArea.hpp" 45 #include "opto/addnode.hpp" 46 #include "opto/block.hpp" 47 #include "opto/c2compiler.hpp" 48 #include "opto/callGenerator.hpp" 49 #include "opto/callnode.hpp" 50 #include "opto/castnode.hpp" 51 #include "opto/cfgnode.hpp" 52 #include "opto/chaitin.hpp" 53 #include "opto/compile.hpp" 54 #include "opto/connode.hpp" 55 #include "opto/convertnode.hpp" 56 #include "opto/divnode.hpp" 57 #include "opto/escape.hpp" 58 #include "opto/idealGraphPrinter.hpp" 59 #include "opto/locknode.hpp" 60 #include "opto/loopnode.hpp" 61 #include "opto/machnode.hpp" 62 #include "opto/macro.hpp" 63 #include "opto/matcher.hpp" 64 #include "opto/mathexactnode.hpp" 65 #include "opto/memnode.hpp" 66 #include "opto/mulnode.hpp" 67 #include "opto/narrowptrnode.hpp" 68 #include "opto/node.hpp" 69 #include "opto/opcodes.hpp" 70 #include "opto/output.hpp" 71 #include "opto/parse.hpp" 72 #include "opto/phaseX.hpp" 73 #include "opto/rootnode.hpp" 74 #include "opto/runtime.hpp" 75 #include "opto/stringopts.hpp" 76 #include "opto/type.hpp" 77 #include "opto/vector.hpp" 78 #include "opto/vectornode.hpp" 79 #include "runtime/globals_extension.hpp" 80 #include "runtime/sharedRuntime.hpp" 81 #include "runtime/signature.hpp" 82 #include "runtime/stubRoutines.hpp" 83 #include "runtime/timer.hpp" 84 #include "utilities/align.hpp" 85 #include "utilities/copy.hpp" 86 #include "utilities/macros.hpp" 87 #include "utilities/resourceHash.hpp" 88 89 // -------------------- Compile::mach_constant_base_node ----------------------- 90 // Constant table base node singleton. 91 MachConstantBaseNode* Compile::mach_constant_base_node() { 92 if (_mach_constant_base_node == nullptr) { 93 _mach_constant_base_node = new MachConstantBaseNode(); 94 _mach_constant_base_node->add_req(C->root()); 95 } 96 return _mach_constant_base_node; 97 } 98 99 100 /// Support for intrinsics. 101 102 // Return the index at which m must be inserted (or already exists). 103 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 104 class IntrinsicDescPair { 105 private: 106 ciMethod* _m; 107 bool _is_virtual; 108 public: 109 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 110 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 111 ciMethod* m= elt->method(); 112 ciMethod* key_m = key->_m; 113 if (key_m < m) return -1; 114 else if (key_m > m) return 1; 115 else { 116 bool is_virtual = elt->is_virtual(); 117 bool key_virtual = key->_is_virtual; 118 if (key_virtual < is_virtual) return -1; 119 else if (key_virtual > is_virtual) return 1; 120 else return 0; 121 } 122 } 123 }; 124 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 125 #ifdef ASSERT 126 for (int i = 1; i < _intrinsics.length(); i++) { 127 CallGenerator* cg1 = _intrinsics.at(i-1); 128 CallGenerator* cg2 = _intrinsics.at(i); 129 assert(cg1->method() != cg2->method() 130 ? cg1->method() < cg2->method() 131 : cg1->is_virtual() < cg2->is_virtual(), 132 "compiler intrinsics list must stay sorted"); 133 } 134 #endif 135 IntrinsicDescPair pair(m, is_virtual); 136 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 137 } 138 139 void Compile::register_intrinsic(CallGenerator* cg) { 140 bool found = false; 141 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 142 assert(!found, "registering twice"); 143 _intrinsics.insert_before(index, cg); 144 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 145 } 146 147 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 148 assert(m->is_loaded(), "don't try this on unloaded methods"); 149 if (_intrinsics.length() > 0) { 150 bool found = false; 151 int index = intrinsic_insertion_index(m, is_virtual, found); 152 if (found) { 153 return _intrinsics.at(index); 154 } 155 } 156 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 157 if (m->intrinsic_id() != vmIntrinsics::_none && 158 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 159 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 160 if (cg != nullptr) { 161 // Save it for next time: 162 register_intrinsic(cg); 163 return cg; 164 } else { 165 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 166 } 167 } 168 return nullptr; 169 } 170 171 // Compile::make_vm_intrinsic is defined in library_call.cpp. 172 173 #ifndef PRODUCT 174 // statistics gathering... 175 176 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0}; 177 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0}; 178 179 inline int as_int(vmIntrinsics::ID id) { 180 return vmIntrinsics::as_int(id); 181 } 182 183 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 184 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 185 int oflags = _intrinsic_hist_flags[as_int(id)]; 186 assert(flags != 0, "what happened?"); 187 if (is_virtual) { 188 flags |= _intrinsic_virtual; 189 } 190 bool changed = (flags != oflags); 191 if ((flags & _intrinsic_worked) != 0) { 192 juint count = (_intrinsic_hist_count[as_int(id)] += 1); 193 if (count == 1) { 194 changed = true; // first time 195 } 196 // increment the overall count also: 197 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1; 198 } 199 if (changed) { 200 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 201 // Something changed about the intrinsic's virtuality. 202 if ((flags & _intrinsic_virtual) != 0) { 203 // This is the first use of this intrinsic as a virtual call. 204 if (oflags != 0) { 205 // We already saw it as a non-virtual, so note both cases. 206 flags |= _intrinsic_both; 207 } 208 } else if ((oflags & _intrinsic_both) == 0) { 209 // This is the first use of this intrinsic as a non-virtual 210 flags |= _intrinsic_both; 211 } 212 } 213 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags); 214 } 215 // update the overall flags also: 216 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags; 217 return changed; 218 } 219 220 static char* format_flags(int flags, char* buf) { 221 buf[0] = 0; 222 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 223 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 224 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 225 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 226 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 227 if (buf[0] == 0) strcat(buf, ","); 228 assert(buf[0] == ',', "must be"); 229 return &buf[1]; 230 } 231 232 void Compile::print_intrinsic_statistics() { 233 char flagsbuf[100]; 234 ttyLocker ttyl; 235 if (xtty != nullptr) xtty->head("statistics type='intrinsic'"); 236 tty->print_cr("Compiler intrinsic usage:"); 237 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)]; 238 if (total == 0) total = 1; // avoid div0 in case of no successes 239 #define PRINT_STAT_LINE(name, c, f) \ 240 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 241 for (auto id : EnumRange<vmIntrinsicID>{}) { 242 int flags = _intrinsic_hist_flags[as_int(id)]; 243 juint count = _intrinsic_hist_count[as_int(id)]; 244 if ((flags | count) != 0) { 245 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 246 } 247 } 248 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf)); 249 if (xtty != nullptr) xtty->tail("statistics"); 250 } 251 252 void Compile::print_statistics() { 253 { ttyLocker ttyl; 254 if (xtty != nullptr) xtty->head("statistics type='opto'"); 255 Parse::print_statistics(); 256 PhaseStringOpts::print_statistics(); 257 PhaseCCP::print_statistics(); 258 PhaseRegAlloc::print_statistics(); 259 PhaseOutput::print_statistics(); 260 PhasePeephole::print_statistics(); 261 PhaseIdealLoop::print_statistics(); 262 ConnectionGraph::print_statistics(); 263 PhaseMacroExpand::print_statistics(); 264 if (xtty != nullptr) xtty->tail("statistics"); 265 } 266 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) { 267 // put this under its own <statistics> element. 268 print_intrinsic_statistics(); 269 } 270 } 271 #endif //PRODUCT 272 273 void Compile::gvn_replace_by(Node* n, Node* nn) { 274 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 275 Node* use = n->last_out(i); 276 bool is_in_table = initial_gvn()->hash_delete(use); 277 uint uses_found = 0; 278 for (uint j = 0; j < use->len(); j++) { 279 if (use->in(j) == n) { 280 if (j < use->req()) 281 use->set_req(j, nn); 282 else 283 use->set_prec(j, nn); 284 uses_found++; 285 } 286 } 287 if (is_in_table) { 288 // reinsert into table 289 initial_gvn()->hash_find_insert(use); 290 } 291 record_for_igvn(use); 292 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist); 293 i -= uses_found; // we deleted 1 or more copies of this edge 294 } 295 } 296 297 298 // Identify all nodes that are reachable from below, useful. 299 // Use breadth-first pass that records state in a Unique_Node_List, 300 // recursive traversal is slower. 301 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 302 int estimated_worklist_size = live_nodes(); 303 useful.map( estimated_worklist_size, nullptr ); // preallocate space 304 305 // Initialize worklist 306 if (root() != nullptr) { useful.push(root()); } 307 // If 'top' is cached, declare it useful to preserve cached node 308 if (cached_top_node()) { useful.push(cached_top_node()); } 309 310 // Push all useful nodes onto the list, breadthfirst 311 for( uint next = 0; next < useful.size(); ++next ) { 312 assert( next < unique(), "Unique useful nodes < total nodes"); 313 Node *n = useful.at(next); 314 uint max = n->len(); 315 for( uint i = 0; i < max; ++i ) { 316 Node *m = n->in(i); 317 if (not_a_node(m)) continue; 318 useful.push(m); 319 } 320 } 321 } 322 323 // Update dead_node_list with any missing dead nodes using useful 324 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 325 void Compile::update_dead_node_list(Unique_Node_List &useful) { 326 uint max_idx = unique(); 327 VectorSet& useful_node_set = useful.member_set(); 328 329 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 330 // If node with index node_idx is not in useful set, 331 // mark it as dead in dead node list. 332 if (!useful_node_set.test(node_idx)) { 333 record_dead_node(node_idx); 334 } 335 } 336 } 337 338 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 339 int shift = 0; 340 for (int i = 0; i < inlines->length(); i++) { 341 CallGenerator* cg = inlines->at(i); 342 if (useful.member(cg->call_node())) { 343 if (shift > 0) { 344 inlines->at_put(i - shift, cg); 345 } 346 } else { 347 shift++; // skip over the dead element 348 } 349 } 350 if (shift > 0) { 351 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array 352 } 353 } 354 355 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) { 356 assert(dead != nullptr && dead->is_Call(), "sanity"); 357 int found = 0; 358 for (int i = 0; i < inlines->length(); i++) { 359 if (inlines->at(i)->call_node() == dead) { 360 inlines->remove_at(i); 361 found++; 362 NOT_DEBUG( break; ) // elements are unique, so exit early 363 } 364 } 365 assert(found <= 1, "not unique"); 366 } 367 368 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)> 369 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) { 370 for (int i = node_list.length() - 1; i >= 0; i--) { 371 N* node = node_list.at(i); 372 if (!useful.member(node)) { 373 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 374 } 375 } 376 } 377 378 void Compile::remove_useless_node(Node* dead) { 379 remove_modified_node(dead); 380 381 // Constant node that has no out-edges and has only one in-edge from 382 // root is usually dead. However, sometimes reshaping walk makes 383 // it reachable by adding use edges. So, we will NOT count Con nodes 384 // as dead to be conservative about the dead node count at any 385 // given time. 386 if (!dead->is_Con()) { 387 record_dead_node(dead->_idx); 388 } 389 if (dead->is_macro()) { 390 remove_macro_node(dead); 391 } 392 if (dead->is_expensive()) { 393 remove_expensive_node(dead); 394 } 395 if (dead->is_OpaqueTemplateAssertionPredicate()) { 396 remove_template_assertion_predicate_opaq(dead); 397 } 398 if (dead->is_ParsePredicate()) { 399 remove_parse_predicate(dead->as_ParsePredicate()); 400 } 401 if (dead->for_post_loop_opts_igvn()) { 402 remove_from_post_loop_opts_igvn(dead); 403 } 404 if (dead->is_Call()) { 405 remove_useless_late_inlines( &_late_inlines, dead); 406 remove_useless_late_inlines( &_string_late_inlines, dead); 407 remove_useless_late_inlines( &_boxing_late_inlines, dead); 408 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead); 409 410 if (dead->is_CallStaticJava()) { 411 remove_unstable_if_trap(dead->as_CallStaticJava(), false); 412 } 413 } 414 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 415 bs->unregister_potential_barrier_node(dead); 416 } 417 418 // Disconnect all useless nodes by disconnecting those at the boundary. 419 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) { 420 uint next = 0; 421 while (next < useful.size()) { 422 Node *n = useful.at(next++); 423 if (n->is_SafePoint()) { 424 // We're done with a parsing phase. Replaced nodes are not valid 425 // beyond that point. 426 n->as_SafePoint()->delete_replaced_nodes(); 427 } 428 // Use raw traversal of out edges since this code removes out edges 429 int max = n->outcnt(); 430 for (int j = 0; j < max; ++j) { 431 Node* child = n->raw_out(j); 432 if (!useful.member(child)) { 433 assert(!child->is_top() || child != top(), 434 "If top is cached in Compile object it is in useful list"); 435 // Only need to remove this out-edge to the useless node 436 n->raw_del_out(j); 437 --j; 438 --max; 439 } 440 } 441 if (n->outcnt() == 1 && n->has_special_unique_user()) { 442 assert(useful.member(n->unique_out()), "do not push a useless node"); 443 worklist.push(n->unique_out()); 444 } 445 } 446 447 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes 448 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes 449 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes 450 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes 451 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass 452 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps 453 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes 454 #ifdef ASSERT 455 if (_modified_nodes != nullptr) { 456 _modified_nodes->remove_useless_nodes(useful.member_set()); 457 } 458 #endif 459 460 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 461 bs->eliminate_useless_gc_barriers(useful, this); 462 // clean up the late inline lists 463 remove_useless_late_inlines( &_late_inlines, useful); 464 remove_useless_late_inlines( &_string_late_inlines, useful); 465 remove_useless_late_inlines( &_boxing_late_inlines, useful); 466 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful); 467 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 468 } 469 470 // ============================================================================ 471 //------------------------------CompileWrapper--------------------------------- 472 class CompileWrapper : public StackObj { 473 Compile *const _compile; 474 public: 475 CompileWrapper(Compile* compile); 476 477 ~CompileWrapper(); 478 }; 479 480 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 481 // the Compile* pointer is stored in the current ciEnv: 482 ciEnv* env = compile->env(); 483 assert(env == ciEnv::current(), "must already be a ciEnv active"); 484 assert(env->compiler_data() == nullptr, "compile already active?"); 485 env->set_compiler_data(compile); 486 assert(compile == Compile::current(), "sanity"); 487 488 compile->set_type_dict(nullptr); 489 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 490 compile->clone_map().set_clone_idx(0); 491 compile->set_type_last_size(0); 492 compile->set_last_tf(nullptr, nullptr); 493 compile->set_indexSet_arena(nullptr); 494 compile->set_indexSet_free_block_list(nullptr); 495 compile->init_type_arena(); 496 Type::Initialize(compile); 497 _compile->begin_method(); 498 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 499 } 500 CompileWrapper::~CompileWrapper() { 501 // simulate crash during compilation 502 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned"); 503 504 _compile->end_method(); 505 _compile->env()->set_compiler_data(nullptr); 506 } 507 508 509 //----------------------------print_compile_messages--------------------------- 510 void Compile::print_compile_messages() { 511 #ifndef PRODUCT 512 // Check if recompiling 513 if (!subsume_loads() && PrintOpto) { 514 // Recompiling without allowing machine instructions to subsume loads 515 tty->print_cr("*********************************************************"); 516 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 517 tty->print_cr("*********************************************************"); 518 } 519 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) { 520 // Recompiling without escape analysis 521 tty->print_cr("*********************************************************"); 522 tty->print_cr("** Bailout: Recompile without escape analysis **"); 523 tty->print_cr("*********************************************************"); 524 } 525 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) { 526 // Recompiling without iterative escape analysis 527 tty->print_cr("*********************************************************"); 528 tty->print_cr("** Bailout: Recompile without iterative escape analysis**"); 529 tty->print_cr("*********************************************************"); 530 } 531 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) { 532 // Recompiling without reducing allocation merges 533 tty->print_cr("*********************************************************"); 534 tty->print_cr("** Bailout: Recompile without reduce allocation merges **"); 535 tty->print_cr("*********************************************************"); 536 } 537 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) { 538 // Recompiling without boxing elimination 539 tty->print_cr("*********************************************************"); 540 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 541 tty->print_cr("*********************************************************"); 542 } 543 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) { 544 // Recompiling without locks coarsening 545 tty->print_cr("*********************************************************"); 546 tty->print_cr("** Bailout: Recompile without locks coarsening **"); 547 tty->print_cr("*********************************************************"); 548 } 549 if (env()->break_at_compile()) { 550 // Open the debugger when compiling this method. 551 tty->print("### Breaking when compiling: "); 552 method()->print_short_name(); 553 tty->cr(); 554 BREAKPOINT; 555 } 556 557 if( PrintOpto ) { 558 if (is_osr_compilation()) { 559 tty->print("[OSR]%3d", _compile_id); 560 } else { 561 tty->print("%3d", _compile_id); 562 } 563 } 564 #endif 565 } 566 567 #ifndef PRODUCT 568 void Compile::print_ideal_ir(const char* phase_name) { 569 // keep the following output all in one block 570 // This output goes directly to the tty, not the compiler log. 571 // To enable tools to match it up with the compilation activity, 572 // be sure to tag this tty output with the compile ID. 573 574 // Node dumping can cause a safepoint, which can break the tty lock. 575 // Buffer all node dumps, so that all safepoints happen before we lock. 576 ResourceMark rm; 577 stringStream ss; 578 579 if (_output == nullptr) { 580 ss.print_cr("AFTER: %s", phase_name); 581 // Print out all nodes in ascending order of index. 582 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss); 583 } else { 584 // Dump the node blockwise if we have a scheduling 585 _output->print_scheduling(&ss); 586 } 587 588 // Check that the lock is not broken by a safepoint. 589 NoSafepointVerifier nsv; 590 ttyLocker ttyl; 591 if (xtty != nullptr) { 592 xtty->head("ideal compile_id='%d'%s compile_phase='%s'", 593 compile_id(), 594 is_osr_compilation() ? " compile_kind='osr'" : "", 595 phase_name); 596 } 597 598 tty->print("%s", ss.as_string()); 599 600 if (xtty != nullptr) { 601 xtty->tail("ideal"); 602 } 603 } 604 #endif 605 606 // ============================================================================ 607 //------------------------------Compile standard------------------------------- 608 609 // Compile a method. entry_bci is -1 for normal compilations and indicates 610 // the continuation bci for on stack replacement. 611 612 613 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci, 614 Options options, DirectiveSet* directive) 615 : Phase(Compiler), 616 _compile_id(ci_env->compile_id()), 617 _options(options), 618 _method(target), 619 _entry_bci(osr_bci), 620 _ilt(nullptr), 621 _stub_function(nullptr), 622 _stub_name(nullptr), 623 _stub_entry_point(nullptr), 624 _max_node_limit(MaxNodeLimit), 625 _post_loop_opts_phase(false), 626 _allow_macro_nodes(true), 627 _inlining_progress(false), 628 _inlining_incrementally(false), 629 _do_cleanup(false), 630 _has_reserved_stack_access(target->has_reserved_stack_access()), 631 #ifndef PRODUCT 632 _igv_idx(0), 633 _trace_opto_output(directive->TraceOptoOutputOption), 634 #endif 635 _has_method_handle_invokes(false), 636 _clinit_barrier_on_entry(false), 637 _stress_seed(0), 638 _comp_arena(mtCompiler), 639 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 640 _env(ci_env), 641 _directive(directive), 642 _log(ci_env->log()), 643 _first_failure_details(nullptr), 644 _intrinsics(comp_arena(), 0, 0, nullptr), 645 _macro_nodes(comp_arena(), 8, 0, nullptr), 646 _parse_predicates(comp_arena(), 8, 0, nullptr), 647 _template_assertion_predicate_opaqs(comp_arena(), 8, 0, nullptr), 648 _expensive_nodes(comp_arena(), 8, 0, nullptr), 649 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr), 650 _unstable_if_traps(comp_arena(), 8, 0, nullptr), 651 _coarsened_locks(comp_arena(), 8, 0, nullptr), 652 _congraph(nullptr), 653 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 654 _unique(0), 655 _dead_node_count(0), 656 _dead_node_list(comp_arena()), 657 _node_arena_one(mtCompiler, Arena::Tag::tag_node), 658 _node_arena_two(mtCompiler, Arena::Tag::tag_node), 659 _node_arena(&_node_arena_one), 660 _mach_constant_base_node(nullptr), 661 _Compile_types(mtCompiler), 662 _initial_gvn(nullptr), 663 _igvn_worklist(nullptr), 664 _types(nullptr), 665 _node_hash(nullptr), 666 _late_inlines(comp_arena(), 2, 0, nullptr), 667 _string_late_inlines(comp_arena(), 2, 0, nullptr), 668 _boxing_late_inlines(comp_arena(), 2, 0, nullptr), 669 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr), 670 _late_inlines_pos(0), 671 _number_of_mh_late_inlines(0), 672 _oom(false), 673 _replay_inline_data(nullptr), 674 _inline_printer(this), 675 _java_calls(0), 676 _inner_loops(0), 677 _interpreter_frame_size(0), 678 _output(nullptr) 679 #ifndef PRODUCT 680 , 681 _in_dump_cnt(0) 682 #endif 683 { 684 C = this; 685 CompileWrapper cw(this); 686 687 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 688 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false); 689 690 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 691 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 692 // We can always print a disassembly, either abstract (hex dump) or 693 // with the help of a suitable hsdis library. Thus, we should not 694 // couple print_assembly and print_opto_assembly controls. 695 // But: always print opto and regular assembly on compile command 'print'. 696 bool print_assembly = directive->PrintAssemblyOption; 697 set_print_assembly(print_opto_assembly || print_assembly); 698 #else 699 set_print_assembly(false); // must initialize. 700 #endif 701 702 #ifndef PRODUCT 703 set_parsed_irreducible_loop(false); 704 #endif 705 706 if (directive->ReplayInlineOption) { 707 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 708 } 709 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 710 set_print_intrinsics(directive->PrintIntrinsicsOption); 711 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 712 713 if (ProfileTraps) { 714 // Make sure the method being compiled gets its own MDO, 715 // so we can at least track the decompile_count(). 716 method()->ensure_method_data(); 717 } 718 719 if (StressLCM || StressGCM || StressIGVN || StressCCP || 720 StressIncrementalInlining || StressMacroExpansion || StressUnstableIfTraps || StressBailout) { 721 initialize_stress_seed(directive); 722 } 723 724 Init(/*do_aliasing=*/ true); 725 726 print_compile_messages(); 727 728 _ilt = InlineTree::build_inline_tree_root(); 729 730 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 731 assert(num_alias_types() >= AliasIdxRaw, ""); 732 733 #define MINIMUM_NODE_HASH 1023 734 735 // GVN that will be run immediately on new nodes 736 uint estimated_size = method()->code_size()*4+64; 737 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 738 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 739 _types = new (comp_arena()) Type_Array(comp_arena()); 740 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size); 741 PhaseGVN gvn; 742 set_initial_gvn(&gvn); 743 744 { // Scope for timing the parser 745 TracePhase tp(_t_parser); 746 747 // Put top into the hash table ASAP. 748 initial_gvn()->transform(top()); 749 750 // Set up tf(), start(), and find a CallGenerator. 751 CallGenerator* cg = nullptr; 752 if (is_osr_compilation()) { 753 const TypeTuple *domain = StartOSRNode::osr_domain(); 754 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 755 init_tf(TypeFunc::make(domain, range)); 756 StartNode* s = new StartOSRNode(root(), domain); 757 initial_gvn()->set_type_bottom(s); 758 verify_start(s); 759 cg = CallGenerator::for_osr(method(), entry_bci()); 760 } else { 761 // Normal case. 762 init_tf(TypeFunc::make(method())); 763 StartNode* s = new StartNode(root(), tf()->domain()); 764 initial_gvn()->set_type_bottom(s); 765 verify_start(s); 766 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 767 // With java.lang.ref.reference.get() we must go through the 768 // intrinsic - even when get() is the root 769 // method of the compile - so that, if necessary, the value in 770 // the referent field of the reference object gets recorded by 771 // the pre-barrier code. 772 cg = find_intrinsic(method(), false); 773 } 774 if (cg == nullptr) { 775 float past_uses = method()->interpreter_invocation_count(); 776 float expected_uses = past_uses; 777 cg = CallGenerator::for_inline(method(), expected_uses); 778 } 779 } 780 if (failing()) return; 781 if (cg == nullptr) { 782 const char* reason = InlineTree::check_can_parse(method()); 783 assert(reason != nullptr, "expect reason for parse failure"); 784 stringStream ss; 785 ss.print("cannot parse method: %s", reason); 786 record_method_not_compilable(ss.as_string()); 787 return; 788 } 789 790 gvn.set_type(root(), root()->bottom_type()); 791 792 JVMState* jvms = build_start_state(start(), tf()); 793 if ((jvms = cg->generate(jvms)) == nullptr) { 794 assert(failure_reason() != nullptr, "expect reason for parse failure"); 795 stringStream ss; 796 ss.print("method parse failed: %s", failure_reason()); 797 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true)); 798 return; 799 } 800 GraphKit kit(jvms); 801 802 if (!kit.stopped()) { 803 // Accept return values, and transfer control we know not where. 804 // This is done by a special, unique ReturnNode bound to root. 805 return_values(kit.jvms()); 806 } 807 808 if (kit.has_exceptions()) { 809 // Any exceptions that escape from this call must be rethrown 810 // to whatever caller is dynamically above us on the stack. 811 // This is done by a special, unique RethrowNode bound to root. 812 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 813 } 814 815 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 816 817 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 818 inline_string_calls(true); 819 } 820 821 if (failing()) return; 822 823 // Remove clutter produced by parsing. 824 if (!failing()) { 825 ResourceMark rm; 826 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 827 } 828 } 829 830 // Note: Large methods are capped off in do_one_bytecode(). 831 if (failing()) return; 832 833 // After parsing, node notes are no longer automagic. 834 // They must be propagated by register_new_node_with_optimizer(), 835 // clone(), or the like. 836 set_default_node_notes(nullptr); 837 838 #ifndef PRODUCT 839 if (should_print_igv(1)) { 840 _igv_printer->print_inlining(); 841 } 842 #endif 843 844 if (failing()) return; 845 NOT_PRODUCT( verify_graph_edges(); ) 846 847 // Now optimize 848 Optimize(); 849 if (failing()) return; 850 NOT_PRODUCT( verify_graph_edges(); ) 851 852 #ifndef PRODUCT 853 if (should_print_ideal()) { 854 print_ideal_ir("print_ideal"); 855 } 856 #endif 857 858 #ifdef ASSERT 859 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 860 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 861 #endif 862 863 // Dump compilation data to replay it. 864 if (directive->DumpReplayOption) { 865 env()->dump_replay_data(_compile_id); 866 } 867 if (directive->DumpInlineOption && (ilt() != nullptr)) { 868 env()->dump_inline_data(_compile_id); 869 } 870 871 // Now that we know the size of all the monitors we can add a fixed slot 872 // for the original deopt pc. 873 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size); 874 set_fixed_slots(next_slot); 875 876 // Compute when to use implicit null checks. Used by matching trap based 877 // nodes and NullCheck optimization. 878 set_allowed_deopt_reasons(); 879 880 // Now generate code 881 Code_Gen(); 882 } 883 884 //------------------------------Compile---------------------------------------- 885 // Compile a runtime stub 886 Compile::Compile(ciEnv* ci_env, 887 TypeFunc_generator generator, 888 address stub_function, 889 const char* stub_name, 890 int is_fancy_jump, 891 bool pass_tls, 892 bool return_pc, 893 DirectiveSet* directive) 894 : Phase(Compiler), 895 _compile_id(0), 896 _options(Options::for_runtime_stub()), 897 _method(nullptr), 898 _entry_bci(InvocationEntryBci), 899 _stub_function(stub_function), 900 _stub_name(stub_name), 901 _stub_entry_point(nullptr), 902 _max_node_limit(MaxNodeLimit), 903 _post_loop_opts_phase(false), 904 _allow_macro_nodes(true), 905 _inlining_progress(false), 906 _inlining_incrementally(false), 907 _has_reserved_stack_access(false), 908 #ifndef PRODUCT 909 _igv_idx(0), 910 _trace_opto_output(directive->TraceOptoOutputOption), 911 #endif 912 _has_method_handle_invokes(false), 913 _clinit_barrier_on_entry(false), 914 _stress_seed(0), 915 _comp_arena(mtCompiler), 916 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 917 _env(ci_env), 918 _directive(directive), 919 _log(ci_env->log()), 920 _first_failure_details(nullptr), 921 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr), 922 _congraph(nullptr), 923 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 924 _unique(0), 925 _dead_node_count(0), 926 _dead_node_list(comp_arena()), 927 _node_arena_one(mtCompiler), 928 _node_arena_two(mtCompiler), 929 _node_arena(&_node_arena_one), 930 _mach_constant_base_node(nullptr), 931 _Compile_types(mtCompiler), 932 _initial_gvn(nullptr), 933 _igvn_worklist(nullptr), 934 _types(nullptr), 935 _node_hash(nullptr), 936 _number_of_mh_late_inlines(0), 937 _oom(false), 938 _replay_inline_data(nullptr), 939 _inline_printer(this), 940 _java_calls(0), 941 _inner_loops(0), 942 _interpreter_frame_size(0), 943 _output(nullptr), 944 #ifndef PRODUCT 945 _in_dump_cnt(0), 946 #endif 947 _allowed_reasons(0) { 948 C = this; 949 950 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false); 951 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false); 952 953 #ifndef PRODUCT 954 set_print_assembly(PrintFrameConverterAssembly); 955 set_parsed_irreducible_loop(false); 956 #else 957 set_print_assembly(false); // Must initialize. 958 #endif 959 set_has_irreducible_loop(false); // no loops 960 961 CompileWrapper cw(this); 962 Init(/*do_aliasing=*/ false); 963 init_tf((*generator)()); 964 965 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 966 _types = new (comp_arena()) Type_Array(comp_arena()); 967 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255); 968 969 if (StressLCM || StressGCM || StressBailout) { 970 initialize_stress_seed(directive); 971 } 972 973 { 974 PhaseGVN gvn; 975 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 976 gvn.transform(top()); 977 978 GraphKit kit; 979 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 980 } 981 982 NOT_PRODUCT( verify_graph_edges(); ) 983 984 Code_Gen(); 985 } 986 987 Compile::~Compile() { 988 delete _first_failure_details; 989 }; 990 991 //------------------------------Init------------------------------------------- 992 // Prepare for a single compilation 993 void Compile::Init(bool aliasing) { 994 _do_aliasing = aliasing; 995 _unique = 0; 996 _regalloc = nullptr; 997 998 _tf = nullptr; // filled in later 999 _top = nullptr; // cached later 1000 _matcher = nullptr; // filled in later 1001 _cfg = nullptr; // filled in later 1002 1003 IA32_ONLY( set_24_bit_selection_and_mode(true, false); ) 1004 1005 _node_note_array = nullptr; 1006 _default_node_notes = nullptr; 1007 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize() 1008 1009 _immutable_memory = nullptr; // filled in at first inquiry 1010 1011 #ifdef ASSERT 1012 _phase_optimize_finished = false; 1013 _phase_verify_ideal_loop = false; 1014 _exception_backedge = false; 1015 _type_verify = nullptr; 1016 #endif 1017 1018 // Globally visible Nodes 1019 // First set TOP to null to give safe behavior during creation of RootNode 1020 set_cached_top_node(nullptr); 1021 set_root(new RootNode()); 1022 // Now that you have a Root to point to, create the real TOP 1023 set_cached_top_node( new ConNode(Type::TOP) ); 1024 set_recent_alloc(nullptr, nullptr); 1025 1026 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1027 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1028 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1029 env()->set_dependencies(new Dependencies(env())); 1030 1031 _fixed_slots = 0; 1032 set_has_split_ifs(false); 1033 set_has_loops(false); // first approximation 1034 set_has_stringbuilder(false); 1035 set_has_boxed_value(false); 1036 _trap_can_recompile = false; // no traps emitted yet 1037 _major_progress = true; // start out assuming good things will happen 1038 set_has_unsafe_access(false); 1039 set_max_vector_size(0); 1040 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1041 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1042 set_decompile_count(0); 1043 1044 #ifndef PRODUCT 1045 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter)); 1046 #endif 1047 1048 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1049 _loop_opts_cnt = LoopOptsCount; 1050 set_do_inlining(Inline); 1051 set_max_inline_size(MaxInlineSize); 1052 set_freq_inline_size(FreqInlineSize); 1053 set_do_scheduling(OptoScheduling); 1054 1055 set_do_vector_loop(false); 1056 set_has_monitors(false); 1057 set_has_scoped_access(false); 1058 1059 if (AllowVectorizeOnDemand) { 1060 if (has_method() && _directive->VectorizeOption) { 1061 set_do_vector_loop(true); 1062 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());}) 1063 } else if (has_method() && method()->name() != nullptr && 1064 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1065 set_do_vector_loop(true); 1066 } 1067 } 1068 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1069 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());}) 1070 1071 _max_node_limit = _directive->MaxNodeLimitOption; 1072 1073 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1074 set_clinit_barrier_on_entry(true); 1075 } 1076 if (debug_info()->recording_non_safepoints()) { 1077 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1078 (comp_arena(), 8, 0, nullptr)); 1079 set_default_node_notes(Node_Notes::make(this)); 1080 } 1081 1082 const int grow_ats = 16; 1083 _max_alias_types = grow_ats; 1084 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1085 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1086 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1087 { 1088 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1089 } 1090 // Initialize the first few types. 1091 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr); 1092 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1093 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1094 _num_alias_types = AliasIdxRaw+1; 1095 // Zero out the alias type cache. 1096 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1097 // A null adr_type hits in the cache right away. Preload the right answer. 1098 probe_alias_cache(nullptr)->_index = AliasIdxTop; 1099 } 1100 1101 #ifdef ASSERT 1102 // Verify that the current StartNode is valid. 1103 void Compile::verify_start(StartNode* s) const { 1104 assert(failing_internal() || s == start(), "should be StartNode"); 1105 } 1106 #endif 1107 1108 /** 1109 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1110 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1111 * the ideal graph. 1112 */ 1113 StartNode* Compile::start() const { 1114 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason()); 1115 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1116 Node* start = root()->fast_out(i); 1117 if (start->is_Start()) { 1118 return start->as_Start(); 1119 } 1120 } 1121 fatal("Did not find Start node!"); 1122 return nullptr; 1123 } 1124 1125 //-------------------------------immutable_memory------------------------------------- 1126 // Access immutable memory 1127 Node* Compile::immutable_memory() { 1128 if (_immutable_memory != nullptr) { 1129 return _immutable_memory; 1130 } 1131 StartNode* s = start(); 1132 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1133 Node *p = s->fast_out(i); 1134 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1135 _immutable_memory = p; 1136 return _immutable_memory; 1137 } 1138 } 1139 ShouldNotReachHere(); 1140 return nullptr; 1141 } 1142 1143 //----------------------set_cached_top_node------------------------------------ 1144 // Install the cached top node, and make sure Node::is_top works correctly. 1145 void Compile::set_cached_top_node(Node* tn) { 1146 if (tn != nullptr) verify_top(tn); 1147 Node* old_top = _top; 1148 _top = tn; 1149 // Calling Node::setup_is_top allows the nodes the chance to adjust 1150 // their _out arrays. 1151 if (_top != nullptr) _top->setup_is_top(); 1152 if (old_top != nullptr) old_top->setup_is_top(); 1153 assert(_top == nullptr || top()->is_top(), ""); 1154 } 1155 1156 #ifdef ASSERT 1157 uint Compile::count_live_nodes_by_graph_walk() { 1158 Unique_Node_List useful(comp_arena()); 1159 // Get useful node list by walking the graph. 1160 identify_useful_nodes(useful); 1161 return useful.size(); 1162 } 1163 1164 void Compile::print_missing_nodes() { 1165 1166 // Return if CompileLog is null and PrintIdealNodeCount is false. 1167 if ((_log == nullptr) && (! PrintIdealNodeCount)) { 1168 return; 1169 } 1170 1171 // This is an expensive function. It is executed only when the user 1172 // specifies VerifyIdealNodeCount option or otherwise knows the 1173 // additional work that needs to be done to identify reachable nodes 1174 // by walking the flow graph and find the missing ones using 1175 // _dead_node_list. 1176 1177 Unique_Node_List useful(comp_arena()); 1178 // Get useful node list by walking the graph. 1179 identify_useful_nodes(useful); 1180 1181 uint l_nodes = C->live_nodes(); 1182 uint l_nodes_by_walk = useful.size(); 1183 1184 if (l_nodes != l_nodes_by_walk) { 1185 if (_log != nullptr) { 1186 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1187 _log->stamp(); 1188 _log->end_head(); 1189 } 1190 VectorSet& useful_member_set = useful.member_set(); 1191 int last_idx = l_nodes_by_walk; 1192 for (int i = 0; i < last_idx; i++) { 1193 if (useful_member_set.test(i)) { 1194 if (_dead_node_list.test(i)) { 1195 if (_log != nullptr) { 1196 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1197 } 1198 if (PrintIdealNodeCount) { 1199 // Print the log message to tty 1200 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1201 useful.at(i)->dump(); 1202 } 1203 } 1204 } 1205 else if (! _dead_node_list.test(i)) { 1206 if (_log != nullptr) { 1207 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1208 } 1209 if (PrintIdealNodeCount) { 1210 // Print the log message to tty 1211 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1212 } 1213 } 1214 } 1215 if (_log != nullptr) { 1216 _log->tail("mismatched_nodes"); 1217 } 1218 } 1219 } 1220 void Compile::record_modified_node(Node* n) { 1221 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) { 1222 _modified_nodes->push(n); 1223 } 1224 } 1225 1226 void Compile::remove_modified_node(Node* n) { 1227 if (_modified_nodes != nullptr) { 1228 _modified_nodes->remove(n); 1229 } 1230 } 1231 #endif 1232 1233 #ifndef PRODUCT 1234 void Compile::verify_top(Node* tn) const { 1235 if (tn != nullptr) { 1236 assert(tn->is_Con(), "top node must be a constant"); 1237 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1238 assert(tn->in(0) != nullptr, "must have live top node"); 1239 } 1240 } 1241 #endif 1242 1243 1244 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1245 1246 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1247 guarantee(arr != nullptr, ""); 1248 int num_blocks = arr->length(); 1249 if (grow_by < num_blocks) grow_by = num_blocks; 1250 int num_notes = grow_by * _node_notes_block_size; 1251 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1252 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1253 while (num_notes > 0) { 1254 arr->append(notes); 1255 notes += _node_notes_block_size; 1256 num_notes -= _node_notes_block_size; 1257 } 1258 assert(num_notes == 0, "exact multiple, please"); 1259 } 1260 1261 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1262 if (source == nullptr || dest == nullptr) return false; 1263 1264 if (dest->is_Con()) 1265 return false; // Do not push debug info onto constants. 1266 1267 #ifdef ASSERT 1268 // Leave a bread crumb trail pointing to the original node: 1269 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) { 1270 dest->set_debug_orig(source); 1271 } 1272 #endif 1273 1274 if (node_note_array() == nullptr) 1275 return false; // Not collecting any notes now. 1276 1277 // This is a copy onto a pre-existing node, which may already have notes. 1278 // If both nodes have notes, do not overwrite any pre-existing notes. 1279 Node_Notes* source_notes = node_notes_at(source->_idx); 1280 if (source_notes == nullptr || source_notes->is_clear()) return false; 1281 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1282 if (dest_notes == nullptr || dest_notes->is_clear()) { 1283 return set_node_notes_at(dest->_idx, source_notes); 1284 } 1285 1286 Node_Notes merged_notes = (*source_notes); 1287 // The order of operations here ensures that dest notes will win... 1288 merged_notes.update_from(dest_notes); 1289 return set_node_notes_at(dest->_idx, &merged_notes); 1290 } 1291 1292 1293 //--------------------------allow_range_check_smearing------------------------- 1294 // Gating condition for coalescing similar range checks. 1295 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1296 // single covering check that is at least as strong as any of them. 1297 // If the optimization succeeds, the simplified (strengthened) range check 1298 // will always succeed. If it fails, we will deopt, and then give up 1299 // on the optimization. 1300 bool Compile::allow_range_check_smearing() const { 1301 // If this method has already thrown a range-check, 1302 // assume it was because we already tried range smearing 1303 // and it failed. 1304 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1305 return !already_trapped; 1306 } 1307 1308 1309 //------------------------------flatten_alias_type----------------------------- 1310 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1311 assert(do_aliasing(), "Aliasing should be enabled"); 1312 int offset = tj->offset(); 1313 TypePtr::PTR ptr = tj->ptr(); 1314 1315 // Known instance (scalarizable allocation) alias only with itself. 1316 bool is_known_inst = tj->isa_oopptr() != nullptr && 1317 tj->is_oopptr()->is_known_instance(); 1318 1319 // Process weird unsafe references. 1320 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1321 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops"); 1322 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1323 tj = TypeOopPtr::BOTTOM; 1324 ptr = tj->ptr(); 1325 offset = tj->offset(); 1326 } 1327 1328 // Array pointers need some flattening 1329 const TypeAryPtr* ta = tj->isa_aryptr(); 1330 if (ta && ta->is_stable()) { 1331 // Erase stability property for alias analysis. 1332 tj = ta = ta->cast_to_stable(false); 1333 } 1334 if( ta && is_known_inst ) { 1335 if ( offset != Type::OffsetBot && 1336 offset > arrayOopDesc::length_offset_in_bytes() ) { 1337 offset = Type::OffsetBot; // Flatten constant access into array body only 1338 tj = ta = ta-> 1339 remove_speculative()-> 1340 cast_to_ptr_type(ptr)-> 1341 with_offset(offset); 1342 } 1343 } else if (ta) { 1344 // For arrays indexed by constant indices, we flatten the alias 1345 // space to include all of the array body. Only the header, klass 1346 // and array length can be accessed un-aliased. 1347 if( offset != Type::OffsetBot ) { 1348 if( ta->const_oop() ) { // MethodData* or Method* 1349 offset = Type::OffsetBot; // Flatten constant access into array body 1350 tj = ta = ta-> 1351 remove_speculative()-> 1352 cast_to_ptr_type(ptr)-> 1353 cast_to_exactness(false)-> 1354 with_offset(offset); 1355 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1356 // range is OK as-is. 1357 tj = ta = TypeAryPtr::RANGE; 1358 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1359 tj = TypeInstPtr::KLASS; // all klass loads look alike 1360 ta = TypeAryPtr::RANGE; // generic ignored junk 1361 ptr = TypePtr::BotPTR; 1362 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1363 tj = TypeInstPtr::MARK; 1364 ta = TypeAryPtr::RANGE; // generic ignored junk 1365 ptr = TypePtr::BotPTR; 1366 } else { // Random constant offset into array body 1367 offset = Type::OffsetBot; // Flatten constant access into array body 1368 tj = ta = ta-> 1369 remove_speculative()-> 1370 cast_to_ptr_type(ptr)-> 1371 cast_to_exactness(false)-> 1372 with_offset(offset); 1373 } 1374 } 1375 // Arrays of fixed size alias with arrays of unknown size. 1376 if (ta->size() != TypeInt::POS) { 1377 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1378 tj = ta = ta-> 1379 remove_speculative()-> 1380 cast_to_ptr_type(ptr)-> 1381 with_ary(tary)-> 1382 cast_to_exactness(false); 1383 } 1384 // Arrays of known objects become arrays of unknown objects. 1385 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1386 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1387 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1388 } 1389 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1390 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1391 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1392 } 1393 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1394 // cannot be distinguished by bytecode alone. 1395 if (ta->elem() == TypeInt::BOOL) { 1396 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1397 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1398 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1399 } 1400 // During the 2nd round of IterGVN, NotNull castings are removed. 1401 // Make sure the Bottom and NotNull variants alias the same. 1402 // Also, make sure exact and non-exact variants alias the same. 1403 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) { 1404 tj = ta = ta-> 1405 remove_speculative()-> 1406 cast_to_ptr_type(TypePtr::BotPTR)-> 1407 cast_to_exactness(false)-> 1408 with_offset(offset); 1409 } 1410 } 1411 1412 // Oop pointers need some flattening 1413 const TypeInstPtr *to = tj->isa_instptr(); 1414 if (to && to != TypeOopPtr::BOTTOM) { 1415 ciInstanceKlass* ik = to->instance_klass(); 1416 if( ptr == TypePtr::Constant ) { 1417 if (ik != ciEnv::current()->Class_klass() || 1418 offset < ik->layout_helper_size_in_bytes()) { 1419 // No constant oop pointers (such as Strings); they alias with 1420 // unknown strings. 1421 assert(!is_known_inst, "not scalarizable allocation"); 1422 tj = to = to-> 1423 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1424 remove_speculative()-> 1425 cast_to_ptr_type(TypePtr::BotPTR)-> 1426 cast_to_exactness(false); 1427 } 1428 } else if( is_known_inst ) { 1429 tj = to; // Keep NotNull and klass_is_exact for instance type 1430 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1431 // During the 2nd round of IterGVN, NotNull castings are removed. 1432 // Make sure the Bottom and NotNull variants alias the same. 1433 // Also, make sure exact and non-exact variants alias the same. 1434 tj = to = to-> 1435 remove_speculative()-> 1436 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1437 cast_to_ptr_type(TypePtr::BotPTR)-> 1438 cast_to_exactness(false); 1439 } 1440 if (to->speculative() != nullptr) { 1441 tj = to = to->remove_speculative(); 1442 } 1443 // Canonicalize the holder of this field 1444 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1445 // First handle header references such as a LoadKlassNode, even if the 1446 // object's klass is unloaded at compile time (4965979). 1447 if (!is_known_inst) { // Do it only for non-instance types 1448 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset); 1449 } 1450 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) { 1451 // Static fields are in the space above the normal instance 1452 // fields in the java.lang.Class instance. 1453 if (ik != ciEnv::current()->Class_klass()) { 1454 to = nullptr; 1455 tj = TypeOopPtr::BOTTOM; 1456 offset = tj->offset(); 1457 } 1458 } else { 1459 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset); 1460 assert(offset < canonical_holder->layout_helper_size_in_bytes(), ""); 1461 assert(tj->offset() == offset, "no change to offset expected"); 1462 bool xk = to->klass_is_exact(); 1463 int instance_id = to->instance_id(); 1464 1465 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder 1466 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only 1467 // its interfaces are included. 1468 if (xk && ik->equals(canonical_holder)) { 1469 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type"); 1470 } else { 1471 assert(xk || !is_known_inst, "Known instance should be exact type"); 1472 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id); 1473 } 1474 } 1475 } 1476 1477 // Klass pointers to object array klasses need some flattening 1478 const TypeKlassPtr *tk = tj->isa_klassptr(); 1479 if( tk ) { 1480 // If we are referencing a field within a Klass, we need 1481 // to assume the worst case of an Object. Both exact and 1482 // inexact types must flatten to the same alias class so 1483 // use NotNull as the PTR. 1484 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1485 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, 1486 env()->Object_klass(), 1487 offset); 1488 } 1489 1490 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) { 1491 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass()); 1492 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs 1493 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset); 1494 } else { 1495 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset); 1496 } 1497 } 1498 1499 // Check for precise loads from the primary supertype array and force them 1500 // to the supertype cache alias index. Check for generic array loads from 1501 // the primary supertype array and also force them to the supertype cache 1502 // alias index. Since the same load can reach both, we need to merge 1503 // these 2 disparate memories into the same alias class. Since the 1504 // primary supertype array is read-only, there's no chance of confusion 1505 // where we bypass an array load and an array store. 1506 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1507 if (offset == Type::OffsetBot || 1508 (offset >= primary_supers_offset && 1509 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1510 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1511 offset = in_bytes(Klass::secondary_super_cache_offset()); 1512 tj = tk = tk->with_offset(offset); 1513 } 1514 } 1515 1516 // Flatten all Raw pointers together. 1517 if (tj->base() == Type::RawPtr) 1518 tj = TypeRawPtr::BOTTOM; 1519 1520 if (tj->base() == Type::AnyPtr) 1521 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1522 1523 offset = tj->offset(); 1524 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1525 1526 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1527 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1528 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1529 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1530 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1531 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1532 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1533 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1534 assert( tj->ptr() != TypePtr::TopPTR && 1535 tj->ptr() != TypePtr::AnyNull && 1536 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1537 // assert( tj->ptr() != TypePtr::Constant || 1538 // tj->base() == Type::RawPtr || 1539 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1540 1541 return tj; 1542 } 1543 1544 void Compile::AliasType::Init(int i, const TypePtr* at) { 1545 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1546 _index = i; 1547 _adr_type = at; 1548 _field = nullptr; 1549 _element = nullptr; 1550 _is_rewritable = true; // default 1551 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr; 1552 if (atoop != nullptr && atoop->is_known_instance()) { 1553 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1554 _general_index = Compile::current()->get_alias_index(gt); 1555 } else { 1556 _general_index = 0; 1557 } 1558 } 1559 1560 BasicType Compile::AliasType::basic_type() const { 1561 if (element() != nullptr) { 1562 const Type* element = adr_type()->is_aryptr()->elem(); 1563 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1564 } if (field() != nullptr) { 1565 return field()->layout_type(); 1566 } else { 1567 return T_ILLEGAL; // unknown 1568 } 1569 } 1570 1571 //---------------------------------print_on------------------------------------ 1572 #ifndef PRODUCT 1573 void Compile::AliasType::print_on(outputStream* st) { 1574 if (index() < 10) 1575 st->print("@ <%d> ", index()); 1576 else st->print("@ <%d>", index()); 1577 st->print(is_rewritable() ? " " : " RO"); 1578 int offset = adr_type()->offset(); 1579 if (offset == Type::OffsetBot) 1580 st->print(" +any"); 1581 else st->print(" +%-3d", offset); 1582 st->print(" in "); 1583 adr_type()->dump_on(st); 1584 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1585 if (field() != nullptr && tjp) { 1586 if (tjp->is_instptr()->instance_klass() != field()->holder() || 1587 tjp->offset() != field()->offset_in_bytes()) { 1588 st->print(" != "); 1589 field()->print(); 1590 st->print(" ***"); 1591 } 1592 } 1593 } 1594 1595 void print_alias_types() { 1596 Compile* C = Compile::current(); 1597 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1598 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1599 C->alias_type(idx)->print_on(tty); 1600 tty->cr(); 1601 } 1602 } 1603 #endif 1604 1605 1606 //----------------------------probe_alias_cache-------------------------------- 1607 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1608 intptr_t key = (intptr_t) adr_type; 1609 key ^= key >> logAliasCacheSize; 1610 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1611 } 1612 1613 1614 //-----------------------------grow_alias_types-------------------------------- 1615 void Compile::grow_alias_types() { 1616 const int old_ats = _max_alias_types; // how many before? 1617 const int new_ats = old_ats; // how many more? 1618 const int grow_ats = old_ats+new_ats; // how many now? 1619 _max_alias_types = grow_ats; 1620 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1621 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1622 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1623 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1624 } 1625 1626 1627 //--------------------------------find_alias_type------------------------------ 1628 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1629 if (!do_aliasing()) { 1630 return alias_type(AliasIdxBot); 1631 } 1632 1633 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1634 if (ace->_adr_type == adr_type) { 1635 return alias_type(ace->_index); 1636 } 1637 1638 // Handle special cases. 1639 if (adr_type == nullptr) return alias_type(AliasIdxTop); 1640 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1641 1642 // Do it the slow way. 1643 const TypePtr* flat = flatten_alias_type(adr_type); 1644 1645 #ifdef ASSERT 1646 { 1647 ResourceMark rm; 1648 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1649 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1650 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1651 Type::str(adr_type)); 1652 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1653 const TypeOopPtr* foop = flat->is_oopptr(); 1654 // Scalarizable allocations have exact klass always. 1655 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1656 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1657 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1658 Type::str(foop), Type::str(xoop)); 1659 } 1660 } 1661 #endif 1662 1663 int idx = AliasIdxTop; 1664 for (int i = 0; i < num_alias_types(); i++) { 1665 if (alias_type(i)->adr_type() == flat) { 1666 idx = i; 1667 break; 1668 } 1669 } 1670 1671 if (idx == AliasIdxTop) { 1672 if (no_create) return nullptr; 1673 // Grow the array if necessary. 1674 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1675 // Add a new alias type. 1676 idx = _num_alias_types++; 1677 _alias_types[idx]->Init(idx, flat); 1678 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1679 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1680 if (flat->isa_instptr()) { 1681 if (flat->offset() == java_lang_Class::klass_offset() 1682 && flat->is_instptr()->instance_klass() == env()->Class_klass()) 1683 alias_type(idx)->set_rewritable(false); 1684 } 1685 if (flat->isa_aryptr()) { 1686 #ifdef ASSERT 1687 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1688 // (T_BYTE has the weakest alignment and size restrictions...) 1689 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1690 #endif 1691 if (flat->offset() == TypePtr::OffsetBot) { 1692 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1693 } 1694 } 1695 if (flat->isa_klassptr()) { 1696 if (UseCompactObjectHeaders) { 1697 if (flat->offset() == in_bytes(Klass::prototype_header_offset())) 1698 alias_type(idx)->set_rewritable(false); 1699 } 1700 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1701 alias_type(idx)->set_rewritable(false); 1702 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1703 alias_type(idx)->set_rewritable(false); 1704 if (flat->offset() == in_bytes(Klass::misc_flags_offset())) 1705 alias_type(idx)->set_rewritable(false); 1706 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1707 alias_type(idx)->set_rewritable(false); 1708 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1709 alias_type(idx)->set_rewritable(false); 1710 } 1711 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1712 // but the base pointer type is not distinctive enough to identify 1713 // references into JavaThread.) 1714 1715 // Check for final fields. 1716 const TypeInstPtr* tinst = flat->isa_instptr(); 1717 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1718 ciField* field; 1719 if (tinst->const_oop() != nullptr && 1720 tinst->instance_klass() == ciEnv::current()->Class_klass() && 1721 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) { 1722 // static field 1723 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1724 field = k->get_field_by_offset(tinst->offset(), true); 1725 } else { 1726 ciInstanceKlass *k = tinst->instance_klass(); 1727 field = k->get_field_by_offset(tinst->offset(), false); 1728 } 1729 assert(field == nullptr || 1730 original_field == nullptr || 1731 (field->holder() == original_field->holder() && 1732 field->offset_in_bytes() == original_field->offset_in_bytes() && 1733 field->is_static() == original_field->is_static()), "wrong field?"); 1734 // Set field() and is_rewritable() attributes. 1735 if (field != nullptr) alias_type(idx)->set_field(field); 1736 } 1737 } 1738 1739 // Fill the cache for next time. 1740 ace->_adr_type = adr_type; 1741 ace->_index = idx; 1742 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1743 1744 // Might as well try to fill the cache for the flattened version, too. 1745 AliasCacheEntry* face = probe_alias_cache(flat); 1746 if (face->_adr_type == nullptr) { 1747 face->_adr_type = flat; 1748 face->_index = idx; 1749 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1750 } 1751 1752 return alias_type(idx); 1753 } 1754 1755 1756 Compile::AliasType* Compile::alias_type(ciField* field) { 1757 const TypeOopPtr* t; 1758 if (field->is_static()) 1759 t = TypeInstPtr::make(field->holder()->java_mirror()); 1760 else 1761 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1762 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1763 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1764 return atp; 1765 } 1766 1767 1768 //------------------------------have_alias_type-------------------------------- 1769 bool Compile::have_alias_type(const TypePtr* adr_type) { 1770 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1771 if (ace->_adr_type == adr_type) { 1772 return true; 1773 } 1774 1775 // Handle special cases. 1776 if (adr_type == nullptr) return true; 1777 if (adr_type == TypePtr::BOTTOM) return true; 1778 1779 return find_alias_type(adr_type, true, nullptr) != nullptr; 1780 } 1781 1782 //-----------------------------must_alias-------------------------------------- 1783 // True if all values of the given address type are in the given alias category. 1784 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1785 if (alias_idx == AliasIdxBot) return true; // the universal category 1786 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP 1787 if (alias_idx == AliasIdxTop) return false; // the empty category 1788 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1789 1790 // the only remaining possible overlap is identity 1791 int adr_idx = get_alias_index(adr_type); 1792 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1793 assert(adr_idx == alias_idx || 1794 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1795 && adr_type != TypeOopPtr::BOTTOM), 1796 "should not be testing for overlap with an unsafe pointer"); 1797 return adr_idx == alias_idx; 1798 } 1799 1800 //------------------------------can_alias-------------------------------------- 1801 // True if any values of the given address type are in the given alias category. 1802 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1803 if (alias_idx == AliasIdxTop) return false; // the empty category 1804 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP 1805 // Known instance doesn't alias with bottom memory 1806 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1807 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1808 1809 // the only remaining possible overlap is identity 1810 int adr_idx = get_alias_index(adr_type); 1811 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1812 return adr_idx == alias_idx; 1813 } 1814 1815 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their 1816 // uncommon traps if no Runtime Predicates were created from the Parse Predicates. 1817 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) { 1818 if (parse_predicate_count() == 0) { 1819 return; 1820 } 1821 for (int i = 0; i < parse_predicate_count(); i++) { 1822 ParsePredicateNode* parse_predicate = _parse_predicates.at(i); 1823 parse_predicate->mark_useless(); 1824 igvn._worklist.push(parse_predicate); 1825 } 1826 _parse_predicates.clear(); 1827 } 1828 1829 void Compile::record_for_post_loop_opts_igvn(Node* n) { 1830 if (!n->for_post_loop_opts_igvn()) { 1831 assert(!_for_post_loop_igvn.contains(n), "duplicate"); 1832 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1833 _for_post_loop_igvn.append(n); 1834 } 1835 } 1836 1837 void Compile::remove_from_post_loop_opts_igvn(Node* n) { 1838 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1839 _for_post_loop_igvn.remove(n); 1840 } 1841 1842 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) { 1843 // Verify that all previous optimizations produced a valid graph 1844 // at least to this point, even if no loop optimizations were done. 1845 PhaseIdealLoop::verify(igvn); 1846 1847 C->set_post_loop_opts_phase(); // no more loop opts allowed 1848 1849 assert(!C->major_progress(), "not cleared"); 1850 1851 if (_for_post_loop_igvn.length() > 0) { 1852 while (_for_post_loop_igvn.length() > 0) { 1853 Node* n = _for_post_loop_igvn.pop(); 1854 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1855 igvn._worklist.push(n); 1856 } 1857 igvn.optimize(); 1858 if (failing()) return; 1859 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed"); 1860 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now"); 1861 1862 // Sometimes IGVN sets major progress (e.g., when processing loop nodes). 1863 if (C->major_progress()) { 1864 C->clear_major_progress(); // ensure that major progress is now clear 1865 } 1866 } 1867 } 1868 1869 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) { 1870 if (OptimizeUnstableIf) { 1871 _unstable_if_traps.append(trap); 1872 } 1873 } 1874 1875 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) { 1876 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) { 1877 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1878 Node* n = trap->uncommon_trap(); 1879 if (!useful.member(n)) { 1880 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 1881 } 1882 } 1883 } 1884 1885 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead 1886 // or fold-compares case. Return true if succeed or not found. 1887 // 1888 // In rare cases, the found trap has been processed. It is too late to delete it. Return 1889 // false and ask fold-compares to yield. 1890 // 1891 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused 1892 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path 1893 // when deoptimization does happen. 1894 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) { 1895 for (int i = 0; i < _unstable_if_traps.length(); ++i) { 1896 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1897 if (trap->uncommon_trap() == unc) { 1898 if (yield && trap->modified()) { 1899 return false; 1900 } 1901 _unstable_if_traps.delete_at(i); 1902 break; 1903 } 1904 } 1905 return true; 1906 } 1907 1908 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path. 1909 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering. 1910 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) { 1911 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) { 1912 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1913 CallStaticJavaNode* unc = trap->uncommon_trap(); 1914 int next_bci = trap->next_bci(); 1915 bool modified = trap->modified(); 1916 1917 if (next_bci != -1 && !modified) { 1918 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!"); 1919 JVMState* jvms = unc->jvms(); 1920 ciMethod* method = jvms->method(); 1921 ciBytecodeStream iter(method); 1922 1923 iter.force_bci(jvms->bci()); 1924 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if"); 1925 Bytecodes::Code c = iter.cur_bc(); 1926 Node* lhs = nullptr; 1927 Node* rhs = nullptr; 1928 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) { 1929 lhs = unc->peek_operand(0); 1930 rhs = unc->peek_operand(1); 1931 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) { 1932 lhs = unc->peek_operand(0); 1933 } 1934 1935 ResourceMark rm; 1936 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci); 1937 assert(live_locals.is_valid(), "broken liveness info"); 1938 int len = (int)live_locals.size(); 1939 1940 for (int i = 0; i < len; i++) { 1941 Node* local = unc->local(jvms, i); 1942 // kill local using the liveness of next_bci. 1943 // give up when the local looks like an operand to secure reexecution. 1944 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) { 1945 uint idx = jvms->locoff() + i; 1946 #ifdef ASSERT 1947 if (PrintOpto && Verbose) { 1948 tty->print("[unstable_if] kill local#%d: ", idx); 1949 local->dump(); 1950 tty->cr(); 1951 } 1952 #endif 1953 igvn.replace_input_of(unc, idx, top()); 1954 modified = true; 1955 } 1956 } 1957 } 1958 1959 // keep the mondified trap for late query 1960 if (modified) { 1961 trap->set_modified(); 1962 } else { 1963 _unstable_if_traps.delete_at(i); 1964 } 1965 } 1966 igvn.optimize(); 1967 } 1968 1969 // StringOpts and late inlining of string methods 1970 void Compile::inline_string_calls(bool parse_time) { 1971 { 1972 // remove useless nodes to make the usage analysis simpler 1973 ResourceMark rm; 1974 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 1975 } 1976 1977 { 1978 ResourceMark rm; 1979 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1980 PhaseStringOpts pso(initial_gvn()); 1981 print_method(PHASE_AFTER_STRINGOPTS, 3); 1982 } 1983 1984 // now inline anything that we skipped the first time around 1985 if (!parse_time) { 1986 _late_inlines_pos = _late_inlines.length(); 1987 } 1988 1989 while (_string_late_inlines.length() > 0) { 1990 CallGenerator* cg = _string_late_inlines.pop(); 1991 cg->do_late_inline(); 1992 if (failing()) return; 1993 } 1994 _string_late_inlines.trunc_to(0); 1995 } 1996 1997 // Late inlining of boxing methods 1998 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1999 if (_boxing_late_inlines.length() > 0) { 2000 assert(has_boxed_value(), "inconsistent"); 2001 2002 set_inlining_incrementally(true); 2003 2004 igvn_worklist()->ensure_empty(); // should be done with igvn 2005 2006 _late_inlines_pos = _late_inlines.length(); 2007 2008 while (_boxing_late_inlines.length() > 0) { 2009 CallGenerator* cg = _boxing_late_inlines.pop(); 2010 cg->do_late_inline(); 2011 if (failing()) return; 2012 } 2013 _boxing_late_inlines.trunc_to(0); 2014 2015 inline_incrementally_cleanup(igvn); 2016 2017 set_inlining_incrementally(false); 2018 } 2019 } 2020 2021 bool Compile::inline_incrementally_one() { 2022 assert(IncrementalInline, "incremental inlining should be on"); 2023 2024 TracePhase tp(_t_incrInline_inline); 2025 2026 set_inlining_progress(false); 2027 set_do_cleanup(false); 2028 2029 for (int i = 0; i < _late_inlines.length(); i++) { 2030 _late_inlines_pos = i+1; 2031 CallGenerator* cg = _late_inlines.at(i); 2032 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline(); 2033 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call 2034 cg->do_late_inline(); 2035 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed"); 2036 if (failing()) { 2037 return false; 2038 } else if (inlining_progress()) { 2039 _late_inlines_pos = i+1; // restore the position in case new elements were inserted 2040 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node()); 2041 break; // process one call site at a time 2042 } 2043 } else { 2044 // Ignore late inline direct calls when inlining is not allowed. 2045 // They are left in the late inline list when node budget is exhausted until the list is fully drained. 2046 } 2047 } 2048 // Remove processed elements. 2049 _late_inlines.remove_till(_late_inlines_pos); 2050 _late_inlines_pos = 0; 2051 2052 assert(inlining_progress() || _late_inlines.length() == 0, "no progress"); 2053 2054 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2055 2056 set_inlining_progress(false); 2057 set_do_cleanup(false); 2058 2059 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption; 2060 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup; 2061 } 2062 2063 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2064 { 2065 TracePhase tp(_t_incrInline_pru); 2066 ResourceMark rm; 2067 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2068 } 2069 { 2070 TracePhase tp(_t_incrInline_igvn); 2071 igvn.reset_from_gvn(initial_gvn()); 2072 igvn.optimize(); 2073 if (failing()) return; 2074 } 2075 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3); 2076 } 2077 2078 // Perform incremental inlining until bound on number of live nodes is reached 2079 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2080 TracePhase tp(_t_incrInline); 2081 2082 set_inlining_incrementally(true); 2083 uint low_live_nodes = 0; 2084 2085 while (_late_inlines.length() > 0) { 2086 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2087 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2088 TracePhase tp(_t_incrInline_ideal); 2089 // PhaseIdealLoop is expensive so we only try it once we are 2090 // out of live nodes and we only try it again if the previous 2091 // helped got the number of nodes down significantly 2092 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2093 if (failing()) return; 2094 low_live_nodes = live_nodes(); 2095 _major_progress = true; 2096 } 2097 2098 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2099 bool do_print_inlining = print_inlining() || print_intrinsics(); 2100 if (do_print_inlining || log() != nullptr) { 2101 // Print inlining message for candidates that we couldn't inline for lack of space. 2102 for (int i = 0; i < _late_inlines.length(); i++) { 2103 CallGenerator* cg = _late_inlines.at(i); 2104 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 2105 if (do_print_inlining) { 2106 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg); 2107 } 2108 log_late_inline_failure(cg, msg); 2109 } 2110 } 2111 break; // finish 2112 } 2113 } 2114 2115 igvn_worklist()->ensure_empty(); // should be done with igvn 2116 2117 while (inline_incrementally_one()) { 2118 assert(!failing_internal() || failure_is_artificial(), "inconsistent"); 2119 } 2120 if (failing()) return; 2121 2122 inline_incrementally_cleanup(igvn); 2123 2124 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 2125 2126 if (failing()) return; 2127 2128 if (_late_inlines.length() == 0) { 2129 break; // no more progress 2130 } 2131 } 2132 2133 igvn_worklist()->ensure_empty(); // should be done with igvn 2134 2135 if (_string_late_inlines.length() > 0) { 2136 assert(has_stringbuilder(), "inconsistent"); 2137 2138 inline_string_calls(false); 2139 2140 if (failing()) return; 2141 2142 inline_incrementally_cleanup(igvn); 2143 } 2144 2145 set_inlining_incrementally(false); 2146 } 2147 2148 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) { 2149 // "inlining_incrementally() == false" is used to signal that no inlining is allowed 2150 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details). 2151 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr" 2152 // as if "inlining_incrementally() == true" were set. 2153 assert(inlining_incrementally() == false, "not allowed"); 2154 assert(_modified_nodes == nullptr, "not allowed"); 2155 assert(_late_inlines.length() > 0, "sanity"); 2156 2157 while (_late_inlines.length() > 0) { 2158 igvn_worklist()->ensure_empty(); // should be done with igvn 2159 2160 while (inline_incrementally_one()) { 2161 assert(!failing_internal() || failure_is_artificial(), "inconsistent"); 2162 } 2163 if (failing()) return; 2164 2165 inline_incrementally_cleanup(igvn); 2166 } 2167 } 2168 2169 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2170 if (_loop_opts_cnt > 0) { 2171 while (major_progress() && (_loop_opts_cnt > 0)) { 2172 TracePhase tp(_t_idealLoop); 2173 PhaseIdealLoop::optimize(igvn, mode); 2174 _loop_opts_cnt--; 2175 if (failing()) return false; 2176 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2177 } 2178 } 2179 return true; 2180 } 2181 2182 // Remove edges from "root" to each SafePoint at a backward branch. 2183 // They were inserted during parsing (see add_safepoint()) to make 2184 // infinite loops without calls or exceptions visible to root, i.e., 2185 // useful. 2186 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2187 Node *r = root(); 2188 if (r != nullptr) { 2189 for (uint i = r->req(); i < r->len(); ++i) { 2190 Node *n = r->in(i); 2191 if (n != nullptr && n->is_SafePoint()) { 2192 r->rm_prec(i); 2193 if (n->outcnt() == 0) { 2194 igvn.remove_dead_node(n); 2195 } 2196 --i; 2197 } 2198 } 2199 // Parsing may have added top inputs to the root node (Path 2200 // leading to the Halt node proven dead). Make sure we get a 2201 // chance to clean them up. 2202 igvn._worklist.push(r); 2203 igvn.optimize(); 2204 } 2205 } 2206 2207 //------------------------------Optimize--------------------------------------- 2208 // Given a graph, optimize it. 2209 void Compile::Optimize() { 2210 TracePhase tp(_t_optimizer); 2211 2212 #ifndef PRODUCT 2213 if (env()->break_at_compile()) { 2214 BREAKPOINT; 2215 } 2216 2217 #endif 2218 2219 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2220 #ifdef ASSERT 2221 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2222 #endif 2223 2224 ResourceMark rm; 2225 2226 NOT_PRODUCT( verify_graph_edges(); ) 2227 2228 print_method(PHASE_AFTER_PARSING, 1); 2229 2230 { 2231 // Iterative Global Value Numbering, including ideal transforms 2232 // Initialize IterGVN with types and values from parse-time GVN 2233 PhaseIterGVN igvn(initial_gvn()); 2234 #ifdef ASSERT 2235 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2236 #endif 2237 { 2238 TracePhase tp(_t_iterGVN); 2239 igvn.optimize(); 2240 } 2241 2242 if (failing()) return; 2243 2244 print_method(PHASE_ITER_GVN1, 2); 2245 2246 process_for_unstable_if_traps(igvn); 2247 2248 if (failing()) return; 2249 2250 inline_incrementally(igvn); 2251 2252 print_method(PHASE_INCREMENTAL_INLINE, 2); 2253 2254 if (failing()) return; 2255 2256 if (eliminate_boxing()) { 2257 // Inline valueOf() methods now. 2258 inline_boxing_calls(igvn); 2259 2260 if (failing()) return; 2261 2262 if (AlwaysIncrementalInline || StressIncrementalInlining) { 2263 inline_incrementally(igvn); 2264 } 2265 2266 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2267 2268 if (failing()) return; 2269 } 2270 2271 // Remove the speculative part of types and clean up the graph from 2272 // the extra CastPP nodes whose only purpose is to carry them. Do 2273 // that early so that optimizations are not disrupted by the extra 2274 // CastPP nodes. 2275 remove_speculative_types(igvn); 2276 2277 if (failing()) return; 2278 2279 // No more new expensive nodes will be added to the list from here 2280 // so keep only the actual candidates for optimizations. 2281 cleanup_expensive_nodes(igvn); 2282 2283 if (failing()) return; 2284 2285 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2286 if (EnableVectorSupport && has_vbox_nodes()) { 2287 TracePhase tp(_t_vector); 2288 PhaseVector pv(igvn); 2289 pv.optimize_vector_boxes(); 2290 if (failing()) return; 2291 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2292 } 2293 assert(!has_vbox_nodes(), "sanity"); 2294 2295 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2296 Compile::TracePhase tp(_t_renumberLive); 2297 igvn_worklist()->ensure_empty(); // should be done with igvn 2298 { 2299 ResourceMark rm; 2300 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist()); 2301 } 2302 igvn.reset_from_gvn(initial_gvn()); 2303 igvn.optimize(); 2304 if (failing()) return; 2305 } 2306 2307 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop 2308 // safepoints 2309 remove_root_to_sfpts_edges(igvn); 2310 2311 if (failing()) return; 2312 2313 // Perform escape analysis 2314 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) { 2315 if (has_loops()) { 2316 // Cleanup graph (remove dead nodes). 2317 TracePhase tp(_t_idealLoop); 2318 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2319 if (failing()) return; 2320 } 2321 bool progress; 2322 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2323 do { 2324 ConnectionGraph::do_analysis(this, &igvn); 2325 2326 if (failing()) return; 2327 2328 int mcount = macro_count(); // Record number of allocations and locks before IGVN 2329 2330 // Optimize out fields loads from scalar replaceable allocations. 2331 igvn.optimize(); 2332 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2333 2334 if (failing()) return; 2335 2336 if (congraph() != nullptr && macro_count() > 0) { 2337 TracePhase tp(_t_macroEliminate); 2338 PhaseMacroExpand mexp(igvn); 2339 mexp.eliminate_macro_nodes(); 2340 if (failing()) return; 2341 2342 igvn.set_delay_transform(false); 2343 igvn.optimize(); 2344 if (failing()) return; 2345 2346 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2347 } 2348 2349 ConnectionGraph::verify_ram_nodes(this, root()); 2350 if (failing()) return; 2351 2352 progress = do_iterative_escape_analysis() && 2353 (macro_count() < mcount) && 2354 ConnectionGraph::has_candidates(this); 2355 // Try again if candidates exist and made progress 2356 // by removing some allocations and/or locks. 2357 } while (progress); 2358 } 2359 2360 // Loop transforms on the ideal graph. Range Check Elimination, 2361 // peeling, unrolling, etc. 2362 2363 // Set loop opts counter 2364 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2365 { 2366 TracePhase tp(_t_idealLoop); 2367 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2368 _loop_opts_cnt--; 2369 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2370 if (failing()) return; 2371 } 2372 // Loop opts pass if partial peeling occurred in previous pass 2373 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2374 TracePhase tp(_t_idealLoop); 2375 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2376 _loop_opts_cnt--; 2377 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2378 if (failing()) return; 2379 } 2380 // Loop opts pass for loop-unrolling before CCP 2381 if(major_progress() && (_loop_opts_cnt > 0)) { 2382 TracePhase tp(_t_idealLoop); 2383 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2384 _loop_opts_cnt--; 2385 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2386 } 2387 if (!failing()) { 2388 // Verify that last round of loop opts produced a valid graph 2389 PhaseIdealLoop::verify(igvn); 2390 } 2391 } 2392 if (failing()) return; 2393 2394 // Conditional Constant Propagation; 2395 print_method(PHASE_BEFORE_CCP1, 2); 2396 PhaseCCP ccp( &igvn ); 2397 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2398 { 2399 TracePhase tp(_t_ccp); 2400 ccp.do_transform(); 2401 } 2402 print_method(PHASE_CCP1, 2); 2403 2404 assert( true, "Break here to ccp.dump_old2new_map()"); 2405 2406 // Iterative Global Value Numbering, including ideal transforms 2407 { 2408 TracePhase tp(_t_iterGVN2); 2409 igvn.reset_from_igvn(&ccp); 2410 igvn.optimize(); 2411 } 2412 print_method(PHASE_ITER_GVN2, 2); 2413 2414 if (failing()) return; 2415 2416 // Loop transforms on the ideal graph. Range Check Elimination, 2417 // peeling, unrolling, etc. 2418 if (!optimize_loops(igvn, LoopOptsDefault)) { 2419 return; 2420 } 2421 2422 if (failing()) return; 2423 2424 C->clear_major_progress(); // ensure that major progress is now clear 2425 2426 process_for_post_loop_opts_igvn(igvn); 2427 2428 if (failing()) return; 2429 2430 #ifdef ASSERT 2431 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2432 #endif 2433 2434 { 2435 TracePhase tp(_t_macroExpand); 2436 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3); 2437 PhaseMacroExpand mex(igvn); 2438 if (mex.expand_macro_nodes()) { 2439 assert(failing(), "must bail out w/ explicit message"); 2440 return; 2441 } 2442 print_method(PHASE_AFTER_MACRO_EXPANSION, 2); 2443 } 2444 2445 { 2446 TracePhase tp(_t_barrierExpand); 2447 if (bs->expand_barriers(this, igvn)) { 2448 assert(failing(), "must bail out w/ explicit message"); 2449 return; 2450 } 2451 print_method(PHASE_BARRIER_EXPANSION, 2); 2452 } 2453 2454 if (C->max_vector_size() > 0) { 2455 C->optimize_logic_cones(igvn); 2456 igvn.optimize(); 2457 if (failing()) return; 2458 } 2459 2460 DEBUG_ONLY( _modified_nodes = nullptr; ) 2461 2462 assert(igvn._worklist.size() == 0, "not empty"); 2463 2464 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty"); 2465 2466 if (_late_inlines.length() > 0) { 2467 // More opportunities to optimize virtual and MH calls. 2468 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option. 2469 process_late_inline_calls_no_inline(igvn); 2470 if (failing()) return; 2471 } 2472 } // (End scope of igvn; run destructor if necessary for asserts.) 2473 2474 check_no_dead_use(); 2475 2476 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have 2477 // to remove hashes to unlock nodes for modifications. 2478 C->node_hash()->clear(); 2479 2480 // A method with only infinite loops has no edges entering loops from root 2481 { 2482 TracePhase tp(_t_graphReshaping); 2483 if (final_graph_reshaping()) { 2484 assert(failing(), "must bail out w/ explicit message"); 2485 return; 2486 } 2487 } 2488 2489 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2490 DEBUG_ONLY(set_phase_optimize_finished();) 2491 } 2492 2493 #ifdef ASSERT 2494 void Compile::check_no_dead_use() const { 2495 ResourceMark rm; 2496 Unique_Node_List wq; 2497 wq.push(root()); 2498 for (uint i = 0; i < wq.size(); ++i) { 2499 Node* n = wq.at(i); 2500 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2501 Node* u = n->fast_out(j); 2502 if (u->outcnt() == 0 && !u->is_Con()) { 2503 u->dump(); 2504 fatal("no reachable node should have no use"); 2505 } 2506 wq.push(u); 2507 } 2508 } 2509 } 2510 #endif 2511 2512 void Compile::inline_vector_reboxing_calls() { 2513 if (C->_vector_reboxing_late_inlines.length() > 0) { 2514 _late_inlines_pos = C->_late_inlines.length(); 2515 while (_vector_reboxing_late_inlines.length() > 0) { 2516 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 2517 cg->do_late_inline(); 2518 if (failing()) return; 2519 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node()); 2520 } 2521 _vector_reboxing_late_inlines.trunc_to(0); 2522 } 2523 } 2524 2525 bool Compile::has_vbox_nodes() { 2526 if (C->_vector_reboxing_late_inlines.length() > 0) { 2527 return true; 2528 } 2529 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 2530 Node * n = C->macro_node(macro_idx); 2531 assert(n->is_macro(), "only macro nodes expected here"); 2532 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 2533 return true; 2534 } 2535 } 2536 return false; 2537 } 2538 2539 //---------------------------- Bitwise operation packing optimization --------------------------- 2540 2541 static bool is_vector_unary_bitwise_op(Node* n) { 2542 return n->Opcode() == Op_XorV && 2543 VectorNode::is_vector_bitwise_not_pattern(n); 2544 } 2545 2546 static bool is_vector_binary_bitwise_op(Node* n) { 2547 switch (n->Opcode()) { 2548 case Op_AndV: 2549 case Op_OrV: 2550 return true; 2551 2552 case Op_XorV: 2553 return !is_vector_unary_bitwise_op(n); 2554 2555 default: 2556 return false; 2557 } 2558 } 2559 2560 static bool is_vector_ternary_bitwise_op(Node* n) { 2561 return n->Opcode() == Op_MacroLogicV; 2562 } 2563 2564 static bool is_vector_bitwise_op(Node* n) { 2565 return is_vector_unary_bitwise_op(n) || 2566 is_vector_binary_bitwise_op(n) || 2567 is_vector_ternary_bitwise_op(n); 2568 } 2569 2570 static bool is_vector_bitwise_cone_root(Node* n) { 2571 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) { 2572 return false; 2573 } 2574 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2575 if (is_vector_bitwise_op(n->fast_out(i))) { 2576 return false; 2577 } 2578 } 2579 return true; 2580 } 2581 2582 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) { 2583 uint cnt = 0; 2584 if (is_vector_bitwise_op(n)) { 2585 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req(); 2586 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2587 for (uint i = 1; i < inp_cnt; i++) { 2588 Node* in = n->in(i); 2589 bool skip = VectorNode::is_all_ones_vector(in); 2590 if (!skip && !inputs.member(in)) { 2591 inputs.push(in); 2592 cnt++; 2593 } 2594 } 2595 assert(cnt <= 1, "not unary"); 2596 } else { 2597 uint last_req = inp_cnt; 2598 if (is_vector_ternary_bitwise_op(n)) { 2599 last_req = inp_cnt - 1; // skip last input 2600 } 2601 for (uint i = 1; i < last_req; i++) { 2602 Node* def = n->in(i); 2603 if (!inputs.member(def)) { 2604 inputs.push(def); 2605 cnt++; 2606 } 2607 } 2608 } 2609 } else { // not a bitwise operations 2610 if (!inputs.member(n)) { 2611 inputs.push(n); 2612 cnt++; 2613 } 2614 } 2615 return cnt; 2616 } 2617 2618 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 2619 Unique_Node_List useful_nodes; 2620 C->identify_useful_nodes(useful_nodes); 2621 2622 for (uint i = 0; i < useful_nodes.size(); i++) { 2623 Node* n = useful_nodes.at(i); 2624 if (is_vector_bitwise_cone_root(n)) { 2625 list.push(n); 2626 } 2627 } 2628 } 2629 2630 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 2631 const TypeVect* vt, 2632 Unique_Node_List& partition, 2633 Unique_Node_List& inputs) { 2634 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 2635 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 2636 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 2637 2638 Node* in1 = inputs.at(0); 2639 Node* in2 = inputs.at(1); 2640 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 2641 2642 uint func = compute_truth_table(partition, inputs); 2643 2644 Node* pn = partition.at(partition.size() - 1); 2645 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2646 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt)); 2647 } 2648 2649 static uint extract_bit(uint func, uint pos) { 2650 return (func & (1 << pos)) >> pos; 2651 } 2652 2653 // 2654 // A macro logic node represents a truth table. It has 4 inputs, 2655 // First three inputs corresponds to 3 columns of a truth table 2656 // and fourth input captures the logic function. 2657 // 2658 // eg. fn = (in1 AND in2) OR in3; 2659 // 2660 // MacroNode(in1,in2,in3,fn) 2661 // 2662 // ----------------- 2663 // in1 in2 in3 fn 2664 // ----------------- 2665 // 0 0 0 0 2666 // 0 0 1 1 2667 // 0 1 0 0 2668 // 0 1 1 1 2669 // 1 0 0 0 2670 // 1 0 1 1 2671 // 1 1 0 1 2672 // 1 1 1 1 2673 // 2674 2675 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 2676 int res = 0; 2677 for (int i = 0; i < 8; i++) { 2678 int bit1 = extract_bit(in1, i); 2679 int bit2 = extract_bit(in2, i); 2680 int bit3 = extract_bit(in3, i); 2681 2682 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 2683 int func_bit = extract_bit(func, func_bit_pos); 2684 2685 res |= func_bit << i; 2686 } 2687 return res; 2688 } 2689 2690 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 2691 assert(n != nullptr, ""); 2692 assert(eval_map.contains(n), "absent"); 2693 return *(eval_map.get(n)); 2694 } 2695 2696 static void eval_operands(Node* n, 2697 uint& func1, uint& func2, uint& func3, 2698 ResourceHashtable<Node*,uint>& eval_map) { 2699 assert(is_vector_bitwise_op(n), ""); 2700 2701 if (is_vector_unary_bitwise_op(n)) { 2702 Node* opnd = n->in(1); 2703 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 2704 opnd = n->in(2); 2705 } 2706 func1 = eval_operand(opnd, eval_map); 2707 } else if (is_vector_binary_bitwise_op(n)) { 2708 func1 = eval_operand(n->in(1), eval_map); 2709 func2 = eval_operand(n->in(2), eval_map); 2710 } else { 2711 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 2712 func1 = eval_operand(n->in(1), eval_map); 2713 func2 = eval_operand(n->in(2), eval_map); 2714 func3 = eval_operand(n->in(3), eval_map); 2715 } 2716 } 2717 2718 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 2719 assert(inputs.size() <= 3, "sanity"); 2720 ResourceMark rm; 2721 uint res = 0; 2722 ResourceHashtable<Node*,uint> eval_map; 2723 2724 // Populate precomputed functions for inputs. 2725 // Each input corresponds to one column of 3 input truth-table. 2726 uint input_funcs[] = { 0xAA, // (_, _, c) -> c 2727 0xCC, // (_, b, _) -> b 2728 0xF0 }; // (a, _, _) -> a 2729 for (uint i = 0; i < inputs.size(); i++) { 2730 eval_map.put(inputs.at(i), input_funcs[2-i]); 2731 } 2732 2733 for (uint i = 0; i < partition.size(); i++) { 2734 Node* n = partition.at(i); 2735 2736 uint func1 = 0, func2 = 0, func3 = 0; 2737 eval_operands(n, func1, func2, func3, eval_map); 2738 2739 switch (n->Opcode()) { 2740 case Op_OrV: 2741 assert(func3 == 0, "not binary"); 2742 res = func1 | func2; 2743 break; 2744 case Op_AndV: 2745 assert(func3 == 0, "not binary"); 2746 res = func1 & func2; 2747 break; 2748 case Op_XorV: 2749 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2750 assert(func2 == 0 && func3 == 0, "not unary"); 2751 res = (~func1) & 0xFF; 2752 } else { 2753 assert(func3 == 0, "not binary"); 2754 res = func1 ^ func2; 2755 } 2756 break; 2757 case Op_MacroLogicV: 2758 // Ordering of inputs may change during evaluation of sub-tree 2759 // containing MacroLogic node as a child node, thus a re-evaluation 2760 // makes sure that function is evaluated in context of current 2761 // inputs. 2762 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 2763 break; 2764 2765 default: assert(false, "not supported: %s", n->Name()); 2766 } 2767 assert(res <= 0xFF, "invalid"); 2768 eval_map.put(n, res); 2769 } 2770 return res; 2771 } 2772 2773 // Criteria under which nodes gets packed into a macro logic node:- 2774 // 1) Parent and both child nodes are all unmasked or masked with 2775 // same predicates. 2776 // 2) Masked parent can be packed with left child if it is predicated 2777 // and both have same predicates. 2778 // 3) Masked parent can be packed with right child if its un-predicated 2779 // or has matching predication condition. 2780 // 4) An unmasked parent can be packed with an unmasked child. 2781 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2782 assert(partition.size() == 0, "not empty"); 2783 assert(inputs.size() == 0, "not empty"); 2784 if (is_vector_ternary_bitwise_op(n)) { 2785 return false; 2786 } 2787 2788 bool is_unary_op = is_vector_unary_bitwise_op(n); 2789 if (is_unary_op) { 2790 assert(collect_unique_inputs(n, inputs) == 1, "not unary"); 2791 return false; // too few inputs 2792 } 2793 2794 bool pack_left_child = true; 2795 bool pack_right_child = true; 2796 2797 bool left_child_LOP = is_vector_bitwise_op(n->in(1)); 2798 bool right_child_LOP = is_vector_bitwise_op(n->in(2)); 2799 2800 int left_child_input_cnt = 0; 2801 int right_child_input_cnt = 0; 2802 2803 bool parent_is_predicated = n->is_predicated_vector(); 2804 bool left_child_predicated = n->in(1)->is_predicated_vector(); 2805 bool right_child_predicated = n->in(2)->is_predicated_vector(); 2806 2807 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr; 2808 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2809 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2810 2811 do { 2812 if (pack_left_child && left_child_LOP && 2813 ((!parent_is_predicated && !left_child_predicated) || 2814 ((parent_is_predicated && left_child_predicated && 2815 parent_pred == left_child_pred)))) { 2816 partition.push(n->in(1)); 2817 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs); 2818 } else { 2819 inputs.push(n->in(1)); 2820 left_child_input_cnt = 1; 2821 } 2822 2823 if (pack_right_child && right_child_LOP && 2824 (!right_child_predicated || 2825 (right_child_predicated && parent_is_predicated && 2826 parent_pred == right_child_pred))) { 2827 partition.push(n->in(2)); 2828 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs); 2829 } else { 2830 inputs.push(n->in(2)); 2831 right_child_input_cnt = 1; 2832 } 2833 2834 if (inputs.size() > 3) { 2835 assert(partition.size() > 0, ""); 2836 inputs.clear(); 2837 partition.clear(); 2838 if (left_child_input_cnt > right_child_input_cnt) { 2839 pack_left_child = false; 2840 } else { 2841 pack_right_child = false; 2842 } 2843 } else { 2844 break; 2845 } 2846 } while(true); 2847 2848 if(partition.size()) { 2849 partition.push(n); 2850 } 2851 2852 return (partition.size() == 2 || partition.size() == 3) && 2853 (inputs.size() == 2 || inputs.size() == 3); 2854 } 2855 2856 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 2857 assert(is_vector_bitwise_op(n), "not a root"); 2858 2859 visited.set(n->_idx); 2860 2861 // 1) Do a DFS walk over the logic cone. 2862 for (uint i = 1; i < n->req(); i++) { 2863 Node* in = n->in(i); 2864 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 2865 process_logic_cone_root(igvn, in, visited); 2866 } 2867 } 2868 2869 // 2) Bottom up traversal: Merge node[s] with 2870 // the parent to form macro logic node. 2871 Unique_Node_List partition; 2872 Unique_Node_List inputs; 2873 if (compute_logic_cone(n, partition, inputs)) { 2874 const TypeVect* vt = n->bottom_type()->is_vect(); 2875 Node* pn = partition.at(partition.size() - 1); 2876 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2877 if (mask == nullptr || 2878 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) { 2879 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 2880 VectorNode::trace_new_vector(macro_logic, "MacroLogic"); 2881 igvn.replace_node(n, macro_logic); 2882 } 2883 } 2884 } 2885 2886 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 2887 ResourceMark rm; 2888 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 2889 Unique_Node_List list; 2890 collect_logic_cone_roots(list); 2891 2892 while (list.size() > 0) { 2893 Node* n = list.pop(); 2894 const TypeVect* vt = n->bottom_type()->is_vect(); 2895 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 2896 if (supported) { 2897 VectorSet visited(comp_arena()); 2898 process_logic_cone_root(igvn, n, visited); 2899 } 2900 } 2901 } 2902 } 2903 2904 //------------------------------Code_Gen--------------------------------------- 2905 // Given a graph, generate code for it 2906 void Compile::Code_Gen() { 2907 if (failing()) { 2908 return; 2909 } 2910 2911 // Perform instruction selection. You might think we could reclaim Matcher 2912 // memory PDQ, but actually the Matcher is used in generating spill code. 2913 // Internals of the Matcher (including some VectorSets) must remain live 2914 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2915 // set a bit in reclaimed memory. 2916 2917 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2918 // nodes. Mapping is only valid at the root of each matched subtree. 2919 NOT_PRODUCT( verify_graph_edges(); ) 2920 2921 Matcher matcher; 2922 _matcher = &matcher; 2923 { 2924 TracePhase tp(_t_matcher); 2925 matcher.match(); 2926 if (failing()) { 2927 return; 2928 } 2929 } 2930 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2931 // nodes. Mapping is only valid at the root of each matched subtree. 2932 NOT_PRODUCT( verify_graph_edges(); ) 2933 2934 // If you have too many nodes, or if matching has failed, bail out 2935 check_node_count(0, "out of nodes matching instructions"); 2936 if (failing()) { 2937 return; 2938 } 2939 2940 print_method(PHASE_MATCHING, 2); 2941 2942 // Build a proper-looking CFG 2943 PhaseCFG cfg(node_arena(), root(), matcher); 2944 if (failing()) { 2945 return; 2946 } 2947 _cfg = &cfg; 2948 { 2949 TracePhase tp(_t_scheduler); 2950 bool success = cfg.do_global_code_motion(); 2951 if (!success) { 2952 return; 2953 } 2954 2955 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2956 NOT_PRODUCT( verify_graph_edges(); ) 2957 cfg.verify(); 2958 if (failing()) { 2959 return; 2960 } 2961 } 2962 2963 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2964 _regalloc = ®alloc; 2965 { 2966 TracePhase tp(_t_registerAllocation); 2967 // Perform register allocation. After Chaitin, use-def chains are 2968 // no longer accurate (at spill code) and so must be ignored. 2969 // Node->LRG->reg mappings are still accurate. 2970 _regalloc->Register_Allocate(); 2971 2972 // Bail out if the allocator builds too many nodes 2973 if (failing()) { 2974 return; 2975 } 2976 2977 print_method(PHASE_REGISTER_ALLOCATION, 2); 2978 } 2979 2980 // Prior to register allocation we kept empty basic blocks in case the 2981 // the allocator needed a place to spill. After register allocation we 2982 // are not adding any new instructions. If any basic block is empty, we 2983 // can now safely remove it. 2984 { 2985 TracePhase tp(_t_blockOrdering); 2986 cfg.remove_empty_blocks(); 2987 if (do_freq_based_layout()) { 2988 PhaseBlockLayout layout(cfg); 2989 } else { 2990 cfg.set_loop_alignment(); 2991 } 2992 cfg.fixup_flow(); 2993 cfg.remove_unreachable_blocks(); 2994 cfg.verify_dominator_tree(); 2995 print_method(PHASE_BLOCK_ORDERING, 3); 2996 } 2997 2998 // Apply peephole optimizations 2999 if( OptoPeephole ) { 3000 TracePhase tp(_t_peephole); 3001 PhasePeephole peep( _regalloc, cfg); 3002 peep.do_transform(); 3003 print_method(PHASE_PEEPHOLE, 3); 3004 } 3005 3006 // Do late expand if CPU requires this. 3007 if (Matcher::require_postalloc_expand) { 3008 TracePhase tp(_t_postalloc_expand); 3009 cfg.postalloc_expand(_regalloc); 3010 print_method(PHASE_POSTALLOC_EXPAND, 3); 3011 } 3012 3013 // Convert Nodes to instruction bits in a buffer 3014 { 3015 TracePhase tp(_t_output); 3016 PhaseOutput output; 3017 output.Output(); 3018 if (failing()) return; 3019 output.install(); 3020 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here 3021 } 3022 3023 // He's dead, Jim. 3024 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 3025 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 3026 } 3027 3028 //------------------------------Final_Reshape_Counts--------------------------- 3029 // This class defines counters to help identify when a method 3030 // may/must be executed using hardware with only 24-bit precision. 3031 struct Final_Reshape_Counts : public StackObj { 3032 int _call_count; // count non-inlined 'common' calls 3033 int _float_count; // count float ops requiring 24-bit precision 3034 int _double_count; // count double ops requiring more precision 3035 int _java_call_count; // count non-inlined 'java' calls 3036 int _inner_loop_count; // count loops which need alignment 3037 VectorSet _visited; // Visitation flags 3038 Node_List _tests; // Set of IfNodes & PCTableNodes 3039 3040 Final_Reshape_Counts() : 3041 _call_count(0), _float_count(0), _double_count(0), 3042 _java_call_count(0), _inner_loop_count(0) { } 3043 3044 void inc_call_count () { _call_count ++; } 3045 void inc_float_count () { _float_count ++; } 3046 void inc_double_count() { _double_count++; } 3047 void inc_java_call_count() { _java_call_count++; } 3048 void inc_inner_loop_count() { _inner_loop_count++; } 3049 3050 int get_call_count () const { return _call_count ; } 3051 int get_float_count () const { return _float_count ; } 3052 int get_double_count() const { return _double_count; } 3053 int get_java_call_count() const { return _java_call_count; } 3054 int get_inner_loop_count() const { return _inner_loop_count; } 3055 }; 3056 3057 //------------------------------final_graph_reshaping_impl---------------------- 3058 // Implement items 1-5 from final_graph_reshaping below. 3059 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3060 3061 if ( n->outcnt() == 0 ) return; // dead node 3062 uint nop = n->Opcode(); 3063 3064 // Check for 2-input instruction with "last use" on right input. 3065 // Swap to left input. Implements item (2). 3066 if( n->req() == 3 && // two-input instruction 3067 n->in(1)->outcnt() > 1 && // left use is NOT a last use 3068 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 3069 n->in(2)->outcnt() == 1 &&// right use IS a last use 3070 !n->in(2)->is_Con() ) { // right use is not a constant 3071 // Check for commutative opcode 3072 switch( nop ) { 3073 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 3074 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 3075 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 3076 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 3077 case Op_AndL: case Op_XorL: case Op_OrL: 3078 case Op_AndI: case Op_XorI: case Op_OrI: { 3079 // Move "last use" input to left by swapping inputs 3080 n->swap_edges(1, 2); 3081 break; 3082 } 3083 default: 3084 break; 3085 } 3086 } 3087 3088 #ifdef ASSERT 3089 if( n->is_Mem() ) { 3090 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 3091 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw || 3092 // oop will be recorded in oop map if load crosses safepoint 3093 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 3094 LoadNode::is_immutable_value(n->in(MemNode::Address)))), 3095 "raw memory operations should have control edge"); 3096 } 3097 if (n->is_MemBar()) { 3098 MemBarNode* mb = n->as_MemBar(); 3099 if (mb->trailing_store() || mb->trailing_load_store()) { 3100 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 3101 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 3102 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 3103 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 3104 } else if (mb->leading()) { 3105 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 3106 } 3107 } 3108 #endif 3109 // Count FPU ops and common calls, implements item (3) 3110 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes); 3111 if (!gc_handled) { 3112 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes); 3113 } 3114 3115 // Collect CFG split points 3116 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3117 frc._tests.push(n); 3118 } 3119 } 3120 3121 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) { 3122 if (!UseDivMod) { 3123 return; 3124 } 3125 3126 // Check if "a % b" and "a / b" both exist 3127 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned)); 3128 if (d == nullptr) { 3129 return; 3130 } 3131 3132 // Replace them with a fused divmod if supported 3133 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) { 3134 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned); 3135 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero. 3136 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the 3137 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency 3138 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the 3139 // DivMod node so the dependency is not lost. 3140 divmod->add_prec_from(n); 3141 divmod->add_prec_from(d); 3142 d->subsume_by(divmod->div_proj(), this); 3143 n->subsume_by(divmod->mod_proj(), this); 3144 } else { 3145 // Replace "a % b" with "a - ((a / b) * b)" 3146 Node* mult = MulNode::make(d, d->in(2), bt); 3147 Node* sub = SubNode::make(d->in(1), mult, bt); 3148 n->subsume_by(sub, this); 3149 } 3150 } 3151 3152 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) { 3153 switch( nop ) { 3154 // Count all float operations that may use FPU 3155 case Op_AddF: 3156 case Op_SubF: 3157 case Op_MulF: 3158 case Op_DivF: 3159 case Op_NegF: 3160 case Op_ModF: 3161 case Op_ConvI2F: 3162 case Op_ConF: 3163 case Op_CmpF: 3164 case Op_CmpF3: 3165 case Op_StoreF: 3166 case Op_LoadF: 3167 // case Op_ConvL2F: // longs are split into 32-bit halves 3168 frc.inc_float_count(); 3169 break; 3170 3171 case Op_ConvF2D: 3172 case Op_ConvD2F: 3173 frc.inc_float_count(); 3174 frc.inc_double_count(); 3175 break; 3176 3177 // Count all double operations that may use FPU 3178 case Op_AddD: 3179 case Op_SubD: 3180 case Op_MulD: 3181 case Op_DivD: 3182 case Op_NegD: 3183 case Op_ModD: 3184 case Op_ConvI2D: 3185 case Op_ConvD2I: 3186 // case Op_ConvL2D: // handled by leaf call 3187 // case Op_ConvD2L: // handled by leaf call 3188 case Op_ConD: 3189 case Op_CmpD: 3190 case Op_CmpD3: 3191 case Op_StoreD: 3192 case Op_LoadD: 3193 case Op_LoadD_unaligned: 3194 frc.inc_double_count(); 3195 break; 3196 case Op_Opaque1: // Remove Opaque Nodes before matching 3197 n->subsume_by(n->in(1), this); 3198 break; 3199 case Op_CallStaticJava: 3200 case Op_CallJava: 3201 case Op_CallDynamicJava: 3202 frc.inc_java_call_count(); // Count java call site; 3203 case Op_CallRuntime: 3204 case Op_CallLeaf: 3205 case Op_CallLeafVector: 3206 case Op_CallLeafNoFP: { 3207 assert (n->is_Call(), ""); 3208 CallNode *call = n->as_Call(); 3209 // Count call sites where the FP mode bit would have to be flipped. 3210 // Do not count uncommon runtime calls: 3211 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3212 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3213 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3214 frc.inc_call_count(); // Count the call site 3215 } else { // See if uncommon argument is shared 3216 Node *n = call->in(TypeFunc::Parms); 3217 int nop = n->Opcode(); 3218 // Clone shared simple arguments to uncommon calls, item (1). 3219 if (n->outcnt() > 1 && 3220 !n->is_Proj() && 3221 nop != Op_CreateEx && 3222 nop != Op_CheckCastPP && 3223 nop != Op_DecodeN && 3224 nop != Op_DecodeNKlass && 3225 !n->is_Mem() && 3226 !n->is_Phi()) { 3227 Node *x = n->clone(); 3228 call->set_req(TypeFunc::Parms, x); 3229 } 3230 } 3231 break; 3232 } 3233 case Op_StoreB: 3234 case Op_StoreC: 3235 case Op_StoreI: 3236 case Op_StoreL: 3237 case Op_CompareAndSwapB: 3238 case Op_CompareAndSwapS: 3239 case Op_CompareAndSwapI: 3240 case Op_CompareAndSwapL: 3241 case Op_CompareAndSwapP: 3242 case Op_CompareAndSwapN: 3243 case Op_WeakCompareAndSwapB: 3244 case Op_WeakCompareAndSwapS: 3245 case Op_WeakCompareAndSwapI: 3246 case Op_WeakCompareAndSwapL: 3247 case Op_WeakCompareAndSwapP: 3248 case Op_WeakCompareAndSwapN: 3249 case Op_CompareAndExchangeB: 3250 case Op_CompareAndExchangeS: 3251 case Op_CompareAndExchangeI: 3252 case Op_CompareAndExchangeL: 3253 case Op_CompareAndExchangeP: 3254 case Op_CompareAndExchangeN: 3255 case Op_GetAndAddS: 3256 case Op_GetAndAddB: 3257 case Op_GetAndAddI: 3258 case Op_GetAndAddL: 3259 case Op_GetAndSetS: 3260 case Op_GetAndSetB: 3261 case Op_GetAndSetI: 3262 case Op_GetAndSetL: 3263 case Op_GetAndSetP: 3264 case Op_GetAndSetN: 3265 case Op_StoreP: 3266 case Op_StoreN: 3267 case Op_StoreNKlass: 3268 case Op_LoadB: 3269 case Op_LoadUB: 3270 case Op_LoadUS: 3271 case Op_LoadI: 3272 case Op_LoadKlass: 3273 case Op_LoadNKlass: 3274 case Op_LoadL: 3275 case Op_LoadL_unaligned: 3276 case Op_LoadP: 3277 case Op_LoadN: 3278 case Op_LoadRange: 3279 case Op_LoadS: 3280 break; 3281 3282 case Op_AddP: { // Assert sane base pointers 3283 Node *addp = n->in(AddPNode::Address); 3284 assert( !addp->is_AddP() || 3285 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3286 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3287 "Base pointers must match (addp %u)", addp->_idx ); 3288 #ifdef _LP64 3289 if ((UseCompressedOops || UseCompressedClassPointers) && 3290 addp->Opcode() == Op_ConP && 3291 addp == n->in(AddPNode::Base) && 3292 n->in(AddPNode::Offset)->is_Con()) { 3293 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3294 // on the platform and on the compressed oops mode. 3295 // Use addressing with narrow klass to load with offset on x86. 3296 // Some platforms can use the constant pool to load ConP. 3297 // Do this transformation here since IGVN will convert ConN back to ConP. 3298 const Type* t = addp->bottom_type(); 3299 bool is_oop = t->isa_oopptr() != nullptr; 3300 bool is_klass = t->isa_klassptr() != nullptr; 3301 3302 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) || 3303 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() && 3304 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) { 3305 Node* nn = nullptr; 3306 3307 int op = is_oop ? Op_ConN : Op_ConNKlass; 3308 3309 // Look for existing ConN node of the same exact type. 3310 Node* r = root(); 3311 uint cnt = r->outcnt(); 3312 for (uint i = 0; i < cnt; i++) { 3313 Node* m = r->raw_out(i); 3314 if (m!= nullptr && m->Opcode() == op && 3315 m->bottom_type()->make_ptr() == t) { 3316 nn = m; 3317 break; 3318 } 3319 } 3320 if (nn != nullptr) { 3321 // Decode a narrow oop to match address 3322 // [R12 + narrow_oop_reg<<3 + offset] 3323 if (is_oop) { 3324 nn = new DecodeNNode(nn, t); 3325 } else { 3326 nn = new DecodeNKlassNode(nn, t); 3327 } 3328 // Check for succeeding AddP which uses the same Base. 3329 // Otherwise we will run into the assertion above when visiting that guy. 3330 for (uint i = 0; i < n->outcnt(); ++i) { 3331 Node *out_i = n->raw_out(i); 3332 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3333 out_i->set_req(AddPNode::Base, nn); 3334 #ifdef ASSERT 3335 for (uint j = 0; j < out_i->outcnt(); ++j) { 3336 Node *out_j = out_i->raw_out(j); 3337 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3338 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3339 } 3340 #endif 3341 } 3342 } 3343 n->set_req(AddPNode::Base, nn); 3344 n->set_req(AddPNode::Address, nn); 3345 if (addp->outcnt() == 0) { 3346 addp->disconnect_inputs(this); 3347 } 3348 } 3349 } 3350 } 3351 #endif 3352 break; 3353 } 3354 3355 case Op_CastPP: { 3356 // Remove CastPP nodes to gain more freedom during scheduling but 3357 // keep the dependency they encode as control or precedence edges 3358 // (if control is set already) on memory operations. Some CastPP 3359 // nodes don't have a control (don't carry a dependency): skip 3360 // those. 3361 if (n->in(0) != nullptr) { 3362 ResourceMark rm; 3363 Unique_Node_List wq; 3364 wq.push(n); 3365 for (uint next = 0; next < wq.size(); ++next) { 3366 Node *m = wq.at(next); 3367 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3368 Node* use = m->fast_out(i); 3369 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3370 use->ensure_control_or_add_prec(n->in(0)); 3371 } else { 3372 switch(use->Opcode()) { 3373 case Op_AddP: 3374 case Op_DecodeN: 3375 case Op_DecodeNKlass: 3376 case Op_CheckCastPP: 3377 case Op_CastPP: 3378 wq.push(use); 3379 break; 3380 } 3381 } 3382 } 3383 } 3384 } 3385 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3386 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3387 Node* in1 = n->in(1); 3388 const Type* t = n->bottom_type(); 3389 Node* new_in1 = in1->clone(); 3390 new_in1->as_DecodeN()->set_type(t); 3391 3392 if (!Matcher::narrow_oop_use_complex_address()) { 3393 // 3394 // x86, ARM and friends can handle 2 adds in addressing mode 3395 // and Matcher can fold a DecodeN node into address by using 3396 // a narrow oop directly and do implicit null check in address: 3397 // 3398 // [R12 + narrow_oop_reg<<3 + offset] 3399 // NullCheck narrow_oop_reg 3400 // 3401 // On other platforms (Sparc) we have to keep new DecodeN node and 3402 // use it to do implicit null check in address: 3403 // 3404 // decode_not_null narrow_oop_reg, base_reg 3405 // [base_reg + offset] 3406 // NullCheck base_reg 3407 // 3408 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3409 // to keep the information to which null check the new DecodeN node 3410 // corresponds to use it as value in implicit_null_check(). 3411 // 3412 new_in1->set_req(0, n->in(0)); 3413 } 3414 3415 n->subsume_by(new_in1, this); 3416 if (in1->outcnt() == 0) { 3417 in1->disconnect_inputs(this); 3418 } 3419 } else { 3420 n->subsume_by(n->in(1), this); 3421 if (n->outcnt() == 0) { 3422 n->disconnect_inputs(this); 3423 } 3424 } 3425 break; 3426 } 3427 case Op_CastII: { 3428 n->as_CastII()->remove_range_check_cast(this); 3429 break; 3430 } 3431 #ifdef _LP64 3432 case Op_CmpP: 3433 // Do this transformation here to preserve CmpPNode::sub() and 3434 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3435 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3436 Node* in1 = n->in(1); 3437 Node* in2 = n->in(2); 3438 if (!in1->is_DecodeNarrowPtr()) { 3439 in2 = in1; 3440 in1 = n->in(2); 3441 } 3442 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3443 3444 Node* new_in2 = nullptr; 3445 if (in2->is_DecodeNarrowPtr()) { 3446 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3447 new_in2 = in2->in(1); 3448 } else if (in2->Opcode() == Op_ConP) { 3449 const Type* t = in2->bottom_type(); 3450 if (t == TypePtr::NULL_PTR) { 3451 assert(in1->is_DecodeN(), "compare klass to null?"); 3452 // Don't convert CmpP null check into CmpN if compressed 3453 // oops implicit null check is not generated. 3454 // This will allow to generate normal oop implicit null check. 3455 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3456 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3457 // 3458 // This transformation together with CastPP transformation above 3459 // will generated code for implicit null checks for compressed oops. 3460 // 3461 // The original code after Optimize() 3462 // 3463 // LoadN memory, narrow_oop_reg 3464 // decode narrow_oop_reg, base_reg 3465 // CmpP base_reg, nullptr 3466 // CastPP base_reg // NotNull 3467 // Load [base_reg + offset], val_reg 3468 // 3469 // after these transformations will be 3470 // 3471 // LoadN memory, narrow_oop_reg 3472 // CmpN narrow_oop_reg, nullptr 3473 // decode_not_null narrow_oop_reg, base_reg 3474 // Load [base_reg + offset], val_reg 3475 // 3476 // and the uncommon path (== nullptr) will use narrow_oop_reg directly 3477 // since narrow oops can be used in debug info now (see the code in 3478 // final_graph_reshaping_walk()). 3479 // 3480 // At the end the code will be matched to 3481 // on x86: 3482 // 3483 // Load_narrow_oop memory, narrow_oop_reg 3484 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3485 // NullCheck narrow_oop_reg 3486 // 3487 // and on sparc: 3488 // 3489 // Load_narrow_oop memory, narrow_oop_reg 3490 // decode_not_null narrow_oop_reg, base_reg 3491 // Load [base_reg + offset], val_reg 3492 // NullCheck base_reg 3493 // 3494 } else if (t->isa_oopptr()) { 3495 new_in2 = ConNode::make(t->make_narrowoop()); 3496 } else if (t->isa_klassptr()) { 3497 new_in2 = ConNode::make(t->make_narrowklass()); 3498 } 3499 } 3500 if (new_in2 != nullptr) { 3501 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3502 n->subsume_by(cmpN, this); 3503 if (in1->outcnt() == 0) { 3504 in1->disconnect_inputs(this); 3505 } 3506 if (in2->outcnt() == 0) { 3507 in2->disconnect_inputs(this); 3508 } 3509 } 3510 } 3511 break; 3512 3513 case Op_DecodeN: 3514 case Op_DecodeNKlass: 3515 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3516 // DecodeN could be pinned when it can't be fold into 3517 // an address expression, see the code for Op_CastPP above. 3518 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3519 break; 3520 3521 case Op_EncodeP: 3522 case Op_EncodePKlass: { 3523 Node* in1 = n->in(1); 3524 if (in1->is_DecodeNarrowPtr()) { 3525 n->subsume_by(in1->in(1), this); 3526 } else if (in1->Opcode() == Op_ConP) { 3527 const Type* t = in1->bottom_type(); 3528 if (t == TypePtr::NULL_PTR) { 3529 assert(t->isa_oopptr(), "null klass?"); 3530 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3531 } else if (t->isa_oopptr()) { 3532 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3533 } else if (t->isa_klassptr()) { 3534 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3535 } 3536 } 3537 if (in1->outcnt() == 0) { 3538 in1->disconnect_inputs(this); 3539 } 3540 break; 3541 } 3542 3543 case Op_Proj: { 3544 if (OptimizeStringConcat || IncrementalInline) { 3545 ProjNode* proj = n->as_Proj(); 3546 if (proj->_is_io_use) { 3547 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, ""); 3548 // Separate projections were used for the exception path which 3549 // are normally removed by a late inline. If it wasn't inlined 3550 // then they will hang around and should just be replaced with 3551 // the original one. Merge them. 3552 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/); 3553 if (non_io_proj != nullptr) { 3554 proj->subsume_by(non_io_proj , this); 3555 } 3556 } 3557 } 3558 break; 3559 } 3560 3561 case Op_Phi: 3562 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3563 // The EncodeP optimization may create Phi with the same edges 3564 // for all paths. It is not handled well by Register Allocator. 3565 Node* unique_in = n->in(1); 3566 assert(unique_in != nullptr, ""); 3567 uint cnt = n->req(); 3568 for (uint i = 2; i < cnt; i++) { 3569 Node* m = n->in(i); 3570 assert(m != nullptr, ""); 3571 if (unique_in != m) 3572 unique_in = nullptr; 3573 } 3574 if (unique_in != nullptr) { 3575 n->subsume_by(unique_in, this); 3576 } 3577 } 3578 break; 3579 3580 #endif 3581 3582 case Op_ModI: 3583 handle_div_mod_op(n, T_INT, false); 3584 break; 3585 3586 case Op_ModL: 3587 handle_div_mod_op(n, T_LONG, false); 3588 break; 3589 3590 case Op_UModI: 3591 handle_div_mod_op(n, T_INT, true); 3592 break; 3593 3594 case Op_UModL: 3595 handle_div_mod_op(n, T_LONG, true); 3596 break; 3597 3598 case Op_LoadVector: 3599 case Op_StoreVector: 3600 #ifdef ASSERT 3601 // Add VerifyVectorAlignment node between adr and load / store. 3602 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) { 3603 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() : 3604 n->as_StoreVector()->must_verify_alignment(); 3605 if (must_verify_alignment) { 3606 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() : 3607 n->as_StoreVector()->memory_size(); 3608 // The memory access should be aligned to the vector width in bytes. 3609 // However, the underlying array is possibly less well aligned, but at least 3610 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in 3611 // a loop we can expect at least the following alignment: 3612 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes); 3613 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range"); 3614 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2"); 3615 // Create mask from alignment. e.g. 0b1000 -> 0b0111 3616 jlong mask = guaranteed_alignment - 1; 3617 Node* mask_con = ConLNode::make(mask); 3618 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con); 3619 n->set_req(MemNode::Address, va); 3620 } 3621 } 3622 #endif 3623 break; 3624 3625 case Op_LoadVectorGather: 3626 case Op_StoreVectorScatter: 3627 case Op_LoadVectorGatherMasked: 3628 case Op_StoreVectorScatterMasked: 3629 case Op_VectorCmpMasked: 3630 case Op_VectorMaskGen: 3631 case Op_LoadVectorMasked: 3632 case Op_StoreVectorMasked: 3633 break; 3634 3635 case Op_AddReductionVI: 3636 case Op_AddReductionVL: 3637 case Op_AddReductionVF: 3638 case Op_AddReductionVD: 3639 case Op_MulReductionVI: 3640 case Op_MulReductionVL: 3641 case Op_MulReductionVF: 3642 case Op_MulReductionVD: 3643 case Op_MinReductionV: 3644 case Op_MaxReductionV: 3645 case Op_AndReductionV: 3646 case Op_OrReductionV: 3647 case Op_XorReductionV: 3648 break; 3649 3650 case Op_PackB: 3651 case Op_PackS: 3652 case Op_PackI: 3653 case Op_PackF: 3654 case Op_PackL: 3655 case Op_PackD: 3656 if (n->req()-1 > 2) { 3657 // Replace many operand PackNodes with a binary tree for matching 3658 PackNode* p = (PackNode*) n; 3659 Node* btp = p->binary_tree_pack(1, n->req()); 3660 n->subsume_by(btp, this); 3661 } 3662 break; 3663 case Op_Loop: 3664 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop"); 3665 case Op_CountedLoop: 3666 case Op_LongCountedLoop: 3667 case Op_OuterStripMinedLoop: 3668 if (n->as_Loop()->is_inner_loop()) { 3669 frc.inc_inner_loop_count(); 3670 } 3671 n->as_Loop()->verify_strip_mined(0); 3672 break; 3673 case Op_LShiftI: 3674 case Op_RShiftI: 3675 case Op_URShiftI: 3676 case Op_LShiftL: 3677 case Op_RShiftL: 3678 case Op_URShiftL: 3679 if (Matcher::need_masked_shift_count) { 3680 // The cpu's shift instructions don't restrict the count to the 3681 // lower 5/6 bits. We need to do the masking ourselves. 3682 Node* in2 = n->in(2); 3683 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3684 const TypeInt* t = in2->find_int_type(); 3685 if (t != nullptr && t->is_con()) { 3686 juint shift = t->get_con(); 3687 if (shift > mask) { // Unsigned cmp 3688 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3689 } 3690 } else { 3691 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) { 3692 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3693 n->set_req(2, shift); 3694 } 3695 } 3696 if (in2->outcnt() == 0) { // Remove dead node 3697 in2->disconnect_inputs(this); 3698 } 3699 } 3700 break; 3701 case Op_MemBarStoreStore: 3702 case Op_MemBarRelease: 3703 // Break the link with AllocateNode: it is no longer useful and 3704 // confuses register allocation. 3705 if (n->req() > MemBarNode::Precedent) { 3706 n->set_req(MemBarNode::Precedent, top()); 3707 } 3708 break; 3709 case Op_MemBarAcquire: { 3710 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3711 // At parse time, the trailing MemBarAcquire for a volatile load 3712 // is created with an edge to the load. After optimizations, 3713 // that input may be a chain of Phis. If those phis have no 3714 // other use, then the MemBarAcquire keeps them alive and 3715 // register allocation can be confused. 3716 dead_nodes.push(n->in(MemBarNode::Precedent)); 3717 n->set_req(MemBarNode::Precedent, top()); 3718 } 3719 break; 3720 } 3721 case Op_Blackhole: 3722 break; 3723 case Op_RangeCheck: { 3724 RangeCheckNode* rc = n->as_RangeCheck(); 3725 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3726 n->subsume_by(iff, this); 3727 frc._tests.push(iff); 3728 break; 3729 } 3730 case Op_ConvI2L: { 3731 if (!Matcher::convi2l_type_required) { 3732 // Code generation on some platforms doesn't need accurate 3733 // ConvI2L types. Widening the type can help remove redundant 3734 // address computations. 3735 n->as_Type()->set_type(TypeLong::INT); 3736 ResourceMark rm; 3737 Unique_Node_List wq; 3738 wq.push(n); 3739 for (uint next = 0; next < wq.size(); next++) { 3740 Node *m = wq.at(next); 3741 3742 for(;;) { 3743 // Loop over all nodes with identical inputs edges as m 3744 Node* k = m->find_similar(m->Opcode()); 3745 if (k == nullptr) { 3746 break; 3747 } 3748 // Push their uses so we get a chance to remove node made 3749 // redundant 3750 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3751 Node* u = k->fast_out(i); 3752 if (u->Opcode() == Op_LShiftL || 3753 u->Opcode() == Op_AddL || 3754 u->Opcode() == Op_SubL || 3755 u->Opcode() == Op_AddP) { 3756 wq.push(u); 3757 } 3758 } 3759 // Replace all nodes with identical edges as m with m 3760 k->subsume_by(m, this); 3761 } 3762 } 3763 } 3764 break; 3765 } 3766 case Op_CmpUL: { 3767 if (!Matcher::has_match_rule(Op_CmpUL)) { 3768 // No support for unsigned long comparisons 3769 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3770 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3771 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3772 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3773 Node* andl = new AndLNode(orl, remove_sign_mask); 3774 Node* cmp = new CmpLNode(andl, n->in(2)); 3775 n->subsume_by(cmp, this); 3776 } 3777 break; 3778 } 3779 #ifdef ASSERT 3780 case Op_ConNKlass: { 3781 const TypePtr* tp = n->as_Type()->type()->make_ptr(); 3782 ciKlass* klass = tp->is_klassptr()->exact_klass(); 3783 assert(klass->is_in_encoding_range(), "klass cannot be compressed"); 3784 break; 3785 } 3786 #endif 3787 default: 3788 assert(!n->is_Call(), ""); 3789 assert(!n->is_Mem(), ""); 3790 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3791 break; 3792 } 3793 } 3794 3795 //------------------------------final_graph_reshaping_walk--------------------- 3796 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3797 // requires that the walk visits a node's inputs before visiting the node. 3798 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3799 Unique_Node_List sfpt; 3800 3801 frc._visited.set(root->_idx); // first, mark node as visited 3802 uint cnt = root->req(); 3803 Node *n = root; 3804 uint i = 0; 3805 while (true) { 3806 if (i < cnt) { 3807 // Place all non-visited non-null inputs onto stack 3808 Node* m = n->in(i); 3809 ++i; 3810 if (m != nullptr && !frc._visited.test_set(m->_idx)) { 3811 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) { 3812 // compute worst case interpreter size in case of a deoptimization 3813 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3814 3815 sfpt.push(m); 3816 } 3817 cnt = m->req(); 3818 nstack.push(n, i); // put on stack parent and next input's index 3819 n = m; 3820 i = 0; 3821 } 3822 } else { 3823 // Now do post-visit work 3824 final_graph_reshaping_impl(n, frc, dead_nodes); 3825 if (nstack.is_empty()) 3826 break; // finished 3827 n = nstack.node(); // Get node from stack 3828 cnt = n->req(); 3829 i = nstack.index(); 3830 nstack.pop(); // Shift to the next node on stack 3831 } 3832 } 3833 3834 // Skip next transformation if compressed oops are not used. 3835 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3836 (!UseCompressedOops && !UseCompressedClassPointers)) 3837 return; 3838 3839 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3840 // It could be done for an uncommon traps or any safepoints/calls 3841 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3842 while (sfpt.size() > 0) { 3843 n = sfpt.pop(); 3844 JVMState *jvms = n->as_SafePoint()->jvms(); 3845 assert(jvms != nullptr, "sanity"); 3846 int start = jvms->debug_start(); 3847 int end = n->req(); 3848 bool is_uncommon = (n->is_CallStaticJava() && 3849 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3850 for (int j = start; j < end; j++) { 3851 Node* in = n->in(j); 3852 if (in->is_DecodeNarrowPtr()) { 3853 bool safe_to_skip = true; 3854 if (!is_uncommon ) { 3855 // Is it safe to skip? 3856 for (uint i = 0; i < in->outcnt(); i++) { 3857 Node* u = in->raw_out(i); 3858 if (!u->is_SafePoint() || 3859 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3860 safe_to_skip = false; 3861 } 3862 } 3863 } 3864 if (safe_to_skip) { 3865 n->set_req(j, in->in(1)); 3866 } 3867 if (in->outcnt() == 0) { 3868 in->disconnect_inputs(this); 3869 } 3870 } 3871 } 3872 } 3873 } 3874 3875 //------------------------------final_graph_reshaping-------------------------- 3876 // Final Graph Reshaping. 3877 // 3878 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3879 // and not commoned up and forced early. Must come after regular 3880 // optimizations to avoid GVN undoing the cloning. Clone constant 3881 // inputs to Loop Phis; these will be split by the allocator anyways. 3882 // Remove Opaque nodes. 3883 // (2) Move last-uses by commutative operations to the left input to encourage 3884 // Intel update-in-place two-address operations and better register usage 3885 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3886 // calls canonicalizing them back. 3887 // (3) Count the number of double-precision FP ops, single-precision FP ops 3888 // and call sites. On Intel, we can get correct rounding either by 3889 // forcing singles to memory (requires extra stores and loads after each 3890 // FP bytecode) or we can set a rounding mode bit (requires setting and 3891 // clearing the mode bit around call sites). The mode bit is only used 3892 // if the relative frequency of single FP ops to calls is low enough. 3893 // This is a key transform for SPEC mpeg_audio. 3894 // (4) Detect infinite loops; blobs of code reachable from above but not 3895 // below. Several of the Code_Gen algorithms fail on such code shapes, 3896 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3897 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3898 // Detection is by looking for IfNodes where only 1 projection is 3899 // reachable from below or CatchNodes missing some targets. 3900 // (5) Assert for insane oop offsets in debug mode. 3901 3902 bool Compile::final_graph_reshaping() { 3903 // an infinite loop may have been eliminated by the optimizer, 3904 // in which case the graph will be empty. 3905 if (root()->req() == 1) { 3906 // Do not compile method that is only a trivial infinite loop, 3907 // since the content of the loop may have been eliminated. 3908 record_method_not_compilable("trivial infinite loop"); 3909 return true; 3910 } 3911 3912 // Expensive nodes have their control input set to prevent the GVN 3913 // from freely commoning them. There's no GVN beyond this point so 3914 // no need to keep the control input. We want the expensive nodes to 3915 // be freely moved to the least frequent code path by gcm. 3916 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3917 for (int i = 0; i < expensive_count(); i++) { 3918 _expensive_nodes.at(i)->set_req(0, nullptr); 3919 } 3920 3921 Final_Reshape_Counts frc; 3922 3923 // Visit everybody reachable! 3924 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3925 Node_Stack nstack(live_nodes() >> 1); 3926 Unique_Node_List dead_nodes; 3927 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes); 3928 3929 // Check for unreachable (from below) code (i.e., infinite loops). 3930 for( uint i = 0; i < frc._tests.size(); i++ ) { 3931 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3932 // Get number of CFG targets. 3933 // Note that PCTables include exception targets after calls. 3934 uint required_outcnt = n->required_outcnt(); 3935 if (n->outcnt() != required_outcnt) { 3936 // Check for a few special cases. Rethrow Nodes never take the 3937 // 'fall-thru' path, so expected kids is 1 less. 3938 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3939 if (n->in(0)->in(0)->is_Call()) { 3940 CallNode* call = n->in(0)->in(0)->as_Call(); 3941 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3942 required_outcnt--; // Rethrow always has 1 less kid 3943 } else if (call->req() > TypeFunc::Parms && 3944 call->is_CallDynamicJava()) { 3945 // Check for null receiver. In such case, the optimizer has 3946 // detected that the virtual call will always result in a null 3947 // pointer exception. The fall-through projection of this CatchNode 3948 // will not be populated. 3949 Node* arg0 = call->in(TypeFunc::Parms); 3950 if (arg0->is_Type() && 3951 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3952 required_outcnt--; 3953 } 3954 } else if (call->entry_point() == OptoRuntime::new_array_Java() || 3955 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 3956 // Check for illegal array length. In such case, the optimizer has 3957 // detected that the allocation attempt will always result in an 3958 // exception. There is no fall-through projection of this CatchNode . 3959 assert(call->is_CallStaticJava(), "static call expected"); 3960 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 3961 uint valid_length_test_input = call->req() - 1; 3962 Node* valid_length_test = call->in(valid_length_test_input); 3963 call->del_req(valid_length_test_input); 3964 if (valid_length_test->find_int_con(1) == 0) { 3965 required_outcnt--; 3966 } 3967 dead_nodes.push(valid_length_test); 3968 assert(n->outcnt() == required_outcnt, "malformed control flow"); 3969 continue; 3970 } 3971 } 3972 } 3973 3974 // Recheck with a better notion of 'required_outcnt' 3975 if (n->outcnt() != required_outcnt) { 3976 record_method_not_compilable("malformed control flow"); 3977 return true; // Not all targets reachable! 3978 } 3979 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) { 3980 CallNode* call = n->in(0)->in(0)->as_Call(); 3981 if (call->entry_point() == OptoRuntime::new_array_Java() || 3982 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 3983 assert(call->is_CallStaticJava(), "static call expected"); 3984 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 3985 uint valid_length_test_input = call->req() - 1; 3986 dead_nodes.push(call->in(valid_length_test_input)); 3987 call->del_req(valid_length_test_input); // valid length test useless now 3988 } 3989 } 3990 // Check that I actually visited all kids. Unreached kids 3991 // must be infinite loops. 3992 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3993 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3994 record_method_not_compilable("infinite loop"); 3995 return true; // Found unvisited kid; must be unreach 3996 } 3997 3998 // Here so verification code in final_graph_reshaping_walk() 3999 // always see an OuterStripMinedLoopEnd 4000 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) { 4001 IfNode* init_iff = n->as_If(); 4002 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 4003 n->subsume_by(iff, this); 4004 } 4005 } 4006 4007 while (dead_nodes.size() > 0) { 4008 Node* m = dead_nodes.pop(); 4009 if (m->outcnt() == 0 && m != top()) { 4010 for (uint j = 0; j < m->req(); j++) { 4011 Node* in = m->in(j); 4012 if (in != nullptr) { 4013 dead_nodes.push(in); 4014 } 4015 } 4016 m->disconnect_inputs(this); 4017 } 4018 } 4019 4020 #ifdef IA32 4021 // If original bytecodes contained a mixture of floats and doubles 4022 // check if the optimizer has made it homogeneous, item (3). 4023 if (UseSSE == 0 && 4024 frc.get_float_count() > 32 && 4025 frc.get_double_count() == 0 && 4026 (10 * frc.get_call_count() < frc.get_float_count()) ) { 4027 set_24_bit_selection_and_mode(false, true); 4028 } 4029 #endif // IA32 4030 4031 set_java_calls(frc.get_java_call_count()); 4032 set_inner_loops(frc.get_inner_loop_count()); 4033 4034 // No infinite loops, no reason to bail out. 4035 return false; 4036 } 4037 4038 //-----------------------------too_many_traps---------------------------------- 4039 // Report if there are too many traps at the current method and bci. 4040 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 4041 bool Compile::too_many_traps(ciMethod* method, 4042 int bci, 4043 Deoptimization::DeoptReason reason) { 4044 ciMethodData* md = method->method_data(); 4045 if (md->is_empty()) { 4046 // Assume the trap has not occurred, or that it occurred only 4047 // because of a transient condition during start-up in the interpreter. 4048 return false; 4049 } 4050 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4051 if (md->has_trap_at(bci, m, reason) != 0) { 4052 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 4053 // Also, if there are multiple reasons, or if there is no per-BCI record, 4054 // assume the worst. 4055 if (log()) 4056 log()->elem("observe trap='%s' count='%d'", 4057 Deoptimization::trap_reason_name(reason), 4058 md->trap_count(reason)); 4059 return true; 4060 } else { 4061 // Ignore method/bci and see if there have been too many globally. 4062 return too_many_traps(reason, md); 4063 } 4064 } 4065 4066 // Less-accurate variant which does not require a method and bci. 4067 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 4068 ciMethodData* logmd) { 4069 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 4070 // Too many traps globally. 4071 // Note that we use cumulative trap_count, not just md->trap_count. 4072 if (log()) { 4073 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason); 4074 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 4075 Deoptimization::trap_reason_name(reason), 4076 mcount, trap_count(reason)); 4077 } 4078 return true; 4079 } else { 4080 // The coast is clear. 4081 return false; 4082 } 4083 } 4084 4085 //--------------------------too_many_recompiles-------------------------------- 4086 // Report if there are too many recompiles at the current method and bci. 4087 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 4088 // Is not eager to return true, since this will cause the compiler to use 4089 // Action_none for a trap point, to avoid too many recompilations. 4090 bool Compile::too_many_recompiles(ciMethod* method, 4091 int bci, 4092 Deoptimization::DeoptReason reason) { 4093 ciMethodData* md = method->method_data(); 4094 if (md->is_empty()) { 4095 // Assume the trap has not occurred, or that it occurred only 4096 // because of a transient condition during start-up in the interpreter. 4097 return false; 4098 } 4099 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 4100 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 4101 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 4102 Deoptimization::DeoptReason per_bc_reason 4103 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 4104 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4105 if ((per_bc_reason == Deoptimization::Reason_none 4106 || md->has_trap_at(bci, m, reason) != 0) 4107 // The trap frequency measure we care about is the recompile count: 4108 && md->trap_recompiled_at(bci, m) 4109 && md->overflow_recompile_count() >= bc_cutoff) { 4110 // Do not emit a trap here if it has already caused recompilations. 4111 // Also, if there are multiple reasons, or if there is no per-BCI record, 4112 // assume the worst. 4113 if (log()) 4114 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 4115 Deoptimization::trap_reason_name(reason), 4116 md->trap_count(reason), 4117 md->overflow_recompile_count()); 4118 return true; 4119 } else if (trap_count(reason) != 0 4120 && decompile_count() >= m_cutoff) { 4121 // Too many recompiles globally, and we have seen this sort of trap. 4122 // Use cumulative decompile_count, not just md->decompile_count. 4123 if (log()) 4124 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 4125 Deoptimization::trap_reason_name(reason), 4126 md->trap_count(reason), trap_count(reason), 4127 md->decompile_count(), decompile_count()); 4128 return true; 4129 } else { 4130 // The coast is clear. 4131 return false; 4132 } 4133 } 4134 4135 // Compute when not to trap. Used by matching trap based nodes and 4136 // NullCheck optimization. 4137 void Compile::set_allowed_deopt_reasons() { 4138 _allowed_reasons = 0; 4139 if (is_method_compilation()) { 4140 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 4141 assert(rs < BitsPerInt, "recode bit map"); 4142 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 4143 _allowed_reasons |= nth_bit(rs); 4144 } 4145 } 4146 } 4147 } 4148 4149 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 4150 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 4151 } 4152 4153 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 4154 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 4155 } 4156 4157 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 4158 if (holder->is_initialized()) { 4159 return false; 4160 } 4161 if (holder->is_being_initialized()) { 4162 if (accessing_method->holder() == holder) { 4163 // Access inside a class. The barrier can be elided when access happens in <clinit>, 4164 // <init>, or a static method. In all those cases, there was an initialization 4165 // barrier on the holder klass passed. 4166 if (accessing_method->is_static_initializer() || 4167 accessing_method->is_object_initializer() || 4168 accessing_method->is_static()) { 4169 return false; 4170 } 4171 } else if (accessing_method->holder()->is_subclass_of(holder)) { 4172 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 4173 // In case of <init> or a static method, the barrier is on the subclass is not enough: 4174 // child class can become fully initialized while its parent class is still being initialized. 4175 if (accessing_method->is_static_initializer()) { 4176 return false; 4177 } 4178 } 4179 ciMethod* root = method(); // the root method of compilation 4180 if (root != accessing_method) { 4181 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 4182 } 4183 } 4184 return true; 4185 } 4186 4187 #ifndef PRODUCT 4188 //------------------------------verify_bidirectional_edges--------------------- 4189 // For each input edge to a node (ie - for each Use-Def edge), verify that 4190 // there is a corresponding Def-Use edge. 4191 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) { 4192 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc 4193 uint stack_size = live_nodes() >> 4; 4194 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize)); 4195 nstack.push(_root); 4196 4197 while (nstack.size() > 0) { 4198 Node* n = nstack.pop(); 4199 if (visited.member(n)) { 4200 continue; 4201 } 4202 visited.push(n); 4203 4204 // Walk over all input edges, checking for correspondence 4205 uint length = n->len(); 4206 for (uint i = 0; i < length; i++) { 4207 Node* in = n->in(i); 4208 if (in != nullptr && !visited.member(in)) { 4209 nstack.push(in); // Put it on stack 4210 } 4211 if (in != nullptr && !in->is_top()) { 4212 // Count instances of `next` 4213 int cnt = 0; 4214 for (uint idx = 0; idx < in->_outcnt; idx++) { 4215 if (in->_out[idx] == n) { 4216 cnt++; 4217 } 4218 } 4219 assert(cnt > 0, "Failed to find Def-Use edge."); 4220 // Check for duplicate edges 4221 // walk the input array downcounting the input edges to n 4222 for (uint j = 0; j < length; j++) { 4223 if (n->in(j) == in) { 4224 cnt--; 4225 } 4226 } 4227 assert(cnt == 0, "Mismatched edge count."); 4228 } else if (in == nullptr) { 4229 assert(i == 0 || i >= n->req() || 4230 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() || 4231 (n->is_Unlock() && i == (n->req() - 1)) || 4232 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion 4233 "only region, phi, arraycopy, unlock or membar nodes have null data edges"); 4234 } else { 4235 assert(in->is_top(), "sanity"); 4236 // Nothing to check. 4237 } 4238 } 4239 } 4240 } 4241 4242 //------------------------------verify_graph_edges--------------------------- 4243 // Walk the Graph and verify that there is a one-to-one correspondence 4244 // between Use-Def edges and Def-Use edges in the graph. 4245 void Compile::verify_graph_edges(bool no_dead_code) { 4246 if (VerifyGraphEdges) { 4247 Unique_Node_List visited; 4248 4249 // Call graph walk to check edges 4250 verify_bidirectional_edges(visited); 4251 if (no_dead_code) { 4252 // Now make sure that no visited node is used by an unvisited node. 4253 bool dead_nodes = false; 4254 Unique_Node_List checked; 4255 while (visited.size() > 0) { 4256 Node* n = visited.pop(); 4257 checked.push(n); 4258 for (uint i = 0; i < n->outcnt(); i++) { 4259 Node* use = n->raw_out(i); 4260 if (checked.member(use)) continue; // already checked 4261 if (visited.member(use)) continue; // already in the graph 4262 if (use->is_Con()) continue; // a dead ConNode is OK 4263 // At this point, we have found a dead node which is DU-reachable. 4264 if (!dead_nodes) { 4265 tty->print_cr("*** Dead nodes reachable via DU edges:"); 4266 dead_nodes = true; 4267 } 4268 use->dump(2); 4269 tty->print_cr("---"); 4270 checked.push(use); // No repeats; pretend it is now checked. 4271 } 4272 } 4273 assert(!dead_nodes, "using nodes must be reachable from root"); 4274 } 4275 } 4276 } 4277 #endif 4278 4279 // The Compile object keeps track of failure reasons separately from the ciEnv. 4280 // This is required because there is not quite a 1-1 relation between the 4281 // ciEnv and its compilation task and the Compile object. Note that one 4282 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4283 // to backtrack and retry without subsuming loads. Other than this backtracking 4284 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4285 // by the logic in C2Compiler. 4286 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) { 4287 if (log() != nullptr) { 4288 log()->elem("failure reason='%s' phase='compile'", reason); 4289 } 4290 if (_failure_reason.get() == nullptr) { 4291 // Record the first failure reason. 4292 _failure_reason.set(reason); 4293 if (CaptureBailoutInformation) { 4294 _first_failure_details = new CompilationFailureInfo(reason); 4295 } 4296 } else { 4297 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure."); 4298 } 4299 4300 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4301 C->print_method(PHASE_FAILURE, 1); 4302 } 4303 _root = nullptr; // flush the graph, too 4304 } 4305 4306 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id) 4307 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose), 4308 _compile(Compile::current()), 4309 _log(nullptr), 4310 _dolog(CITimeVerbose) 4311 { 4312 assert(_compile != nullptr, "sanity check"); 4313 if (_dolog) { 4314 _log = _compile->log(); 4315 } 4316 if (_log != nullptr) { 4317 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes()); 4318 _log->stamp(); 4319 _log->end_head(); 4320 } 4321 } 4322 4323 Compile::TracePhase::TracePhase(PhaseTraceId id) 4324 : TracePhase(Phase::get_phase_trace_id_text(id), id) {} 4325 4326 Compile::TracePhase::~TracePhase() { 4327 if (_compile->failing_internal()) { 4328 if (_log != nullptr) { 4329 _log->done("phase"); 4330 } 4331 return; // timing code, not stressing bailouts. 4332 } 4333 #ifdef ASSERT 4334 if (PrintIdealNodeCount) { 4335 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4336 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk()); 4337 } 4338 4339 if (VerifyIdealNodeCount) { 4340 _compile->print_missing_nodes(); 4341 } 4342 #endif 4343 4344 if (_log != nullptr) { 4345 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes()); 4346 } 4347 } 4348 4349 //----------------------------static_subtype_check----------------------------- 4350 // Shortcut important common cases when superklass is exact: 4351 // (0) superklass is java.lang.Object (can occur in reflective code) 4352 // (1) subklass is already limited to a subtype of superklass => always ok 4353 // (2) subklass does not overlap with superklass => always fail 4354 // (3) superklass has NO subtypes and we can check with a simple compare. 4355 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) { 4356 if (skip) { 4357 return SSC_full_test; // Let caller generate the general case. 4358 } 4359 4360 if (subk->is_java_subtype_of(superk)) { 4361 return SSC_always_true; // (0) and (1) this test cannot fail 4362 } 4363 4364 if (!subk->maybe_java_subtype_of(superk)) { 4365 return SSC_always_false; // (2) true path dead; no dynamic test needed 4366 } 4367 4368 const Type* superelem = superk; 4369 if (superk->isa_aryklassptr()) { 4370 int ignored; 4371 superelem = superk->is_aryklassptr()->base_element_type(ignored); 4372 } 4373 4374 if (superelem->isa_instklassptr()) { 4375 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass(); 4376 if (!ik->has_subklass()) { 4377 if (!ik->is_final()) { 4378 // Add a dependency if there is a chance of a later subclass. 4379 dependencies()->assert_leaf_type(ik); 4380 } 4381 if (!superk->maybe_java_subtype_of(subk)) { 4382 return SSC_always_false; 4383 } 4384 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4385 } 4386 } else { 4387 // A primitive array type has no subtypes. 4388 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4389 } 4390 4391 return SSC_full_test; 4392 } 4393 4394 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4395 #ifdef _LP64 4396 // The scaled index operand to AddP must be a clean 64-bit value. 4397 // Java allows a 32-bit int to be incremented to a negative 4398 // value, which appears in a 64-bit register as a large 4399 // positive number. Using that large positive number as an 4400 // operand in pointer arithmetic has bad consequences. 4401 // On the other hand, 32-bit overflow is rare, and the possibility 4402 // can often be excluded, if we annotate the ConvI2L node with 4403 // a type assertion that its value is known to be a small positive 4404 // number. (The prior range check has ensured this.) 4405 // This assertion is used by ConvI2LNode::Ideal. 4406 int index_max = max_jint - 1; // array size is max_jint, index is one less 4407 if (sizetype != nullptr) index_max = sizetype->_hi - 1; 4408 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4409 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4410 #endif 4411 return idx; 4412 } 4413 4414 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4415 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 4416 if (ctrl != nullptr) { 4417 // Express control dependency by a CastII node with a narrow type. 4418 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4419 // node from floating above the range check during loop optimizations. Otherwise, the 4420 // ConvI2L node may be eliminated independently of the range check, causing the data path 4421 // to become TOP while the control path is still there (although it's unreachable). 4422 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */); 4423 value = phase->transform(value); 4424 } 4425 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4426 return phase->transform(new ConvI2LNode(value, ltype)); 4427 } 4428 4429 void Compile::dump_print_inlining() { 4430 inline_printer()->print_on(tty); 4431 } 4432 4433 void Compile::log_late_inline(CallGenerator* cg) { 4434 if (log() != nullptr) { 4435 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4436 cg->unique_id()); 4437 JVMState* p = cg->call_node()->jvms(); 4438 while (p != nullptr) { 4439 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4440 p = p->caller(); 4441 } 4442 log()->tail("late_inline"); 4443 } 4444 } 4445 4446 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4447 log_late_inline(cg); 4448 if (log() != nullptr) { 4449 log()->inline_fail(msg); 4450 } 4451 } 4452 4453 void Compile::log_inline_id(CallGenerator* cg) { 4454 if (log() != nullptr) { 4455 // The LogCompilation tool needs a unique way to identify late 4456 // inline call sites. This id must be unique for this call site in 4457 // this compilation. Try to have it unique across compilations as 4458 // well because it can be convenient when grepping through the log 4459 // file. 4460 // Distinguish OSR compilations from others in case CICountOSR is 4461 // on. 4462 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4463 cg->set_unique_id(id); 4464 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4465 } 4466 } 4467 4468 void Compile::log_inline_failure(const char* msg) { 4469 if (C->log() != nullptr) { 4470 C->log()->inline_fail(msg); 4471 } 4472 } 4473 4474 4475 // Dump inlining replay data to the stream. 4476 // Don't change thread state and acquire any locks. 4477 void Compile::dump_inline_data(outputStream* out) { 4478 InlineTree* inl_tree = ilt(); 4479 if (inl_tree != nullptr) { 4480 out->print(" inline %d", inl_tree->count()); 4481 inl_tree->dump_replay_data(out); 4482 } 4483 } 4484 4485 void Compile::dump_inline_data_reduced(outputStream* out) { 4486 assert(ReplayReduce, ""); 4487 4488 InlineTree* inl_tree = ilt(); 4489 if (inl_tree == nullptr) { 4490 return; 4491 } 4492 // Enable iterative replay file reduction 4493 // Output "compile" lines for depth 1 subtrees, 4494 // simulating that those trees were compiled 4495 // instead of inlined. 4496 for (int i = 0; i < inl_tree->subtrees().length(); ++i) { 4497 InlineTree* sub = inl_tree->subtrees().at(i); 4498 if (sub->inline_level() != 1) { 4499 continue; 4500 } 4501 4502 ciMethod* method = sub->method(); 4503 int entry_bci = -1; 4504 int comp_level = env()->task()->comp_level(); 4505 out->print("compile "); 4506 method->dump_name_as_ascii(out); 4507 out->print(" %d %d", entry_bci, comp_level); 4508 out->print(" inline %d", sub->count()); 4509 sub->dump_replay_data(out, -1); 4510 out->cr(); 4511 } 4512 } 4513 4514 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4515 if (n1->Opcode() < n2->Opcode()) return -1; 4516 else if (n1->Opcode() > n2->Opcode()) return 1; 4517 4518 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4519 for (uint i = 1; i < n1->req(); i++) { 4520 if (n1->in(i) < n2->in(i)) return -1; 4521 else if (n1->in(i) > n2->in(i)) return 1; 4522 } 4523 4524 return 0; 4525 } 4526 4527 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4528 Node* n1 = *n1p; 4529 Node* n2 = *n2p; 4530 4531 return cmp_expensive_nodes(n1, n2); 4532 } 4533 4534 void Compile::sort_expensive_nodes() { 4535 if (!expensive_nodes_sorted()) { 4536 _expensive_nodes.sort(cmp_expensive_nodes); 4537 } 4538 } 4539 4540 bool Compile::expensive_nodes_sorted() const { 4541 for (int i = 1; i < _expensive_nodes.length(); i++) { 4542 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) { 4543 return false; 4544 } 4545 } 4546 return true; 4547 } 4548 4549 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4550 if (_expensive_nodes.length() == 0) { 4551 return false; 4552 } 4553 4554 assert(OptimizeExpensiveOps, "optimization off?"); 4555 4556 // Take this opportunity to remove dead nodes from the list 4557 int j = 0; 4558 for (int i = 0; i < _expensive_nodes.length(); i++) { 4559 Node* n = _expensive_nodes.at(i); 4560 if (!n->is_unreachable(igvn)) { 4561 assert(n->is_expensive(), "should be expensive"); 4562 _expensive_nodes.at_put(j, n); 4563 j++; 4564 } 4565 } 4566 _expensive_nodes.trunc_to(j); 4567 4568 // Then sort the list so that similar nodes are next to each other 4569 // and check for at least two nodes of identical kind with same data 4570 // inputs. 4571 sort_expensive_nodes(); 4572 4573 for (int i = 0; i < _expensive_nodes.length()-1; i++) { 4574 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) { 4575 return true; 4576 } 4577 } 4578 4579 return false; 4580 } 4581 4582 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4583 if (_expensive_nodes.length() == 0) { 4584 return; 4585 } 4586 4587 assert(OptimizeExpensiveOps, "optimization off?"); 4588 4589 // Sort to bring similar nodes next to each other and clear the 4590 // control input of nodes for which there's only a single copy. 4591 sort_expensive_nodes(); 4592 4593 int j = 0; 4594 int identical = 0; 4595 int i = 0; 4596 bool modified = false; 4597 for (; i < _expensive_nodes.length()-1; i++) { 4598 assert(j <= i, "can't write beyond current index"); 4599 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) { 4600 identical++; 4601 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4602 continue; 4603 } 4604 if (identical > 0) { 4605 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4606 identical = 0; 4607 } else { 4608 Node* n = _expensive_nodes.at(i); 4609 igvn.replace_input_of(n, 0, nullptr); 4610 igvn.hash_insert(n); 4611 modified = true; 4612 } 4613 } 4614 if (identical > 0) { 4615 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4616 } else if (_expensive_nodes.length() >= 1) { 4617 Node* n = _expensive_nodes.at(i); 4618 igvn.replace_input_of(n, 0, nullptr); 4619 igvn.hash_insert(n); 4620 modified = true; 4621 } 4622 _expensive_nodes.trunc_to(j); 4623 if (modified) { 4624 igvn.optimize(); 4625 } 4626 } 4627 4628 void Compile::add_expensive_node(Node * n) { 4629 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list"); 4630 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4631 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4632 if (OptimizeExpensiveOps) { 4633 _expensive_nodes.append(n); 4634 } else { 4635 // Clear control input and let IGVN optimize expensive nodes if 4636 // OptimizeExpensiveOps is off. 4637 n->set_req(0, nullptr); 4638 } 4639 } 4640 4641 /** 4642 * Track coarsened Lock and Unlock nodes. 4643 */ 4644 4645 class Lock_List : public Node_List { 4646 uint _origin_cnt; 4647 public: 4648 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {} 4649 uint origin_cnt() const { return _origin_cnt; } 4650 }; 4651 4652 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) { 4653 int length = locks.length(); 4654 if (length > 0) { 4655 // Have to keep this list until locks elimination during Macro nodes elimination. 4656 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length); 4657 AbstractLockNode* alock = locks.at(0); 4658 BoxLockNode* box = alock->box_node()->as_BoxLock(); 4659 for (int i = 0; i < length; i++) { 4660 AbstractLockNode* lock = locks.at(i); 4661 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx); 4662 locks_list->push(lock); 4663 BoxLockNode* this_box = lock->box_node()->as_BoxLock(); 4664 if (this_box != box) { 4665 // Locking regions (BoxLock) could be Unbalanced here: 4666 // - its coarsened locks were eliminated in earlier 4667 // macro nodes elimination followed by loop unroll 4668 // - it is OSR locking region (no Lock node) 4669 // Preserve Unbalanced status in such cases. 4670 if (!this_box->is_unbalanced()) { 4671 this_box->set_coarsened(); 4672 } 4673 if (!box->is_unbalanced()) { 4674 box->set_coarsened(); 4675 } 4676 } 4677 } 4678 _coarsened_locks.append(locks_list); 4679 } 4680 } 4681 4682 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) { 4683 int count = coarsened_count(); 4684 for (int i = 0; i < count; i++) { 4685 Node_List* locks_list = _coarsened_locks.at(i); 4686 for (uint j = 0; j < locks_list->size(); j++) { 4687 Node* lock = locks_list->at(j); 4688 assert(lock->is_AbstractLock(), "sanity"); 4689 if (!useful.member(lock)) { 4690 locks_list->yank(lock); 4691 } 4692 } 4693 } 4694 } 4695 4696 void Compile::remove_coarsened_lock(Node* n) { 4697 if (n->is_AbstractLock()) { 4698 int count = coarsened_count(); 4699 for (int i = 0; i < count; i++) { 4700 Node_List* locks_list = _coarsened_locks.at(i); 4701 locks_list->yank(n); 4702 } 4703 } 4704 } 4705 4706 bool Compile::coarsened_locks_consistent() { 4707 int count = coarsened_count(); 4708 for (int i = 0; i < count; i++) { 4709 bool unbalanced = false; 4710 bool modified = false; // track locks kind modifications 4711 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i); 4712 uint size = locks_list->size(); 4713 if (size == 0) { 4714 unbalanced = false; // All locks were eliminated - good 4715 } else if (size != locks_list->origin_cnt()) { 4716 unbalanced = true; // Some locks were removed from list 4717 } else { 4718 for (uint j = 0; j < size; j++) { 4719 Node* lock = locks_list->at(j); 4720 // All nodes in group should have the same state (modified or not) 4721 if (!lock->as_AbstractLock()->is_coarsened()) { 4722 if (j == 0) { 4723 // first on list was modified, the rest should be too for consistency 4724 modified = true; 4725 } else if (!modified) { 4726 // this lock was modified but previous locks on the list were not 4727 unbalanced = true; 4728 break; 4729 } 4730 } else if (modified) { 4731 // previous locks on list were modified but not this lock 4732 unbalanced = true; 4733 break; 4734 } 4735 } 4736 } 4737 if (unbalanced) { 4738 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified 4739 #ifdef ASSERT 4740 if (PrintEliminateLocks) { 4741 tty->print_cr("=== unbalanced coarsened locks ==="); 4742 for (uint l = 0; l < size; l++) { 4743 locks_list->at(l)->dump(); 4744 } 4745 } 4746 #endif 4747 record_failure(C2Compiler::retry_no_locks_coarsening()); 4748 return false; 4749 } 4750 } 4751 return true; 4752 } 4753 4754 // Mark locking regions (identified by BoxLockNode) as unbalanced if 4755 // locks coarsening optimization removed Lock/Unlock nodes from them. 4756 // Such regions become unbalanced because coarsening only removes part 4757 // of Lock/Unlock nodes in region. As result we can't execute other 4758 // locks elimination optimizations which assume all code paths have 4759 // corresponding pair of Lock/Unlock nodes - they are balanced. 4760 void Compile::mark_unbalanced_boxes() const { 4761 int count = coarsened_count(); 4762 for (int i = 0; i < count; i++) { 4763 Node_List* locks_list = _coarsened_locks.at(i); 4764 uint size = locks_list->size(); 4765 if (size > 0) { 4766 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock(); 4767 BoxLockNode* box = alock->box_node()->as_BoxLock(); 4768 if (alock->is_coarsened()) { 4769 // coarsened_locks_consistent(), which is called before this method, verifies 4770 // that the rest of Lock/Unlock nodes on locks_list are also coarsened. 4771 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated"); 4772 for (uint j = 1; j < size; j++) { 4773 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here"); 4774 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock(); 4775 if (box != this_box) { 4776 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated"); 4777 box->set_unbalanced(); 4778 this_box->set_unbalanced(); 4779 } 4780 } 4781 } 4782 } 4783 } 4784 } 4785 4786 /** 4787 * Remove the speculative part of types and clean up the graph 4788 */ 4789 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4790 if (UseTypeSpeculation) { 4791 Unique_Node_List worklist; 4792 worklist.push(root()); 4793 int modified = 0; 4794 // Go over all type nodes that carry a speculative type, drop the 4795 // speculative part of the type and enqueue the node for an igvn 4796 // which may optimize it out. 4797 for (uint next = 0; next < worklist.size(); ++next) { 4798 Node *n = worklist.at(next); 4799 if (n->is_Type()) { 4800 TypeNode* tn = n->as_Type(); 4801 const Type* t = tn->type(); 4802 const Type* t_no_spec = t->remove_speculative(); 4803 if (t_no_spec != t) { 4804 bool in_hash = igvn.hash_delete(n); 4805 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table"); 4806 tn->set_type(t_no_spec); 4807 igvn.hash_insert(n); 4808 igvn._worklist.push(n); // give it a chance to go away 4809 modified++; 4810 } 4811 } 4812 // Iterate over outs - endless loops is unreachable from below 4813 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4814 Node *m = n->fast_out(i); 4815 if (not_a_node(m)) { 4816 continue; 4817 } 4818 worklist.push(m); 4819 } 4820 } 4821 // Drop the speculative part of all types in the igvn's type table 4822 igvn.remove_speculative_types(); 4823 if (modified > 0) { 4824 igvn.optimize(); 4825 if (failing()) return; 4826 } 4827 #ifdef ASSERT 4828 // Verify that after the IGVN is over no speculative type has resurfaced 4829 worklist.clear(); 4830 worklist.push(root()); 4831 for (uint next = 0; next < worklist.size(); ++next) { 4832 Node *n = worklist.at(next); 4833 const Type* t = igvn.type_or_null(n); 4834 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types"); 4835 if (n->is_Type()) { 4836 t = n->as_Type()->type(); 4837 assert(t == t->remove_speculative(), "no more speculative types"); 4838 } 4839 // Iterate over outs - endless loops is unreachable from below 4840 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4841 Node *m = n->fast_out(i); 4842 if (not_a_node(m)) { 4843 continue; 4844 } 4845 worklist.push(m); 4846 } 4847 } 4848 igvn.check_no_speculative_types(); 4849 #endif 4850 } 4851 } 4852 4853 // Auxiliary methods to support randomized stressing/fuzzing. 4854 4855 void Compile::initialize_stress_seed(const DirectiveSet* directive) { 4856 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) { 4857 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds()); 4858 FLAG_SET_ERGO(StressSeed, _stress_seed); 4859 } else { 4860 _stress_seed = StressSeed; 4861 } 4862 if (_log != nullptr) { 4863 _log->elem("stress_test seed='%u'", _stress_seed); 4864 } 4865 } 4866 4867 int Compile::random() { 4868 _stress_seed = os::next_random(_stress_seed); 4869 return static_cast<int>(_stress_seed); 4870 } 4871 4872 // This method can be called the arbitrary number of times, with current count 4873 // as the argument. The logic allows selecting a single candidate from the 4874 // running list of candidates as follows: 4875 // int count = 0; 4876 // Cand* selected = null; 4877 // while(cand = cand->next()) { 4878 // if (randomized_select(++count)) { 4879 // selected = cand; 4880 // } 4881 // } 4882 // 4883 // Including count equalizes the chances any candidate is "selected". 4884 // This is useful when we don't have the complete list of candidates to choose 4885 // from uniformly. In this case, we need to adjust the randomicity of the 4886 // selection, or else we will end up biasing the selection towards the latter 4887 // candidates. 4888 // 4889 // Quick back-envelope calculation shows that for the list of n candidates 4890 // the equal probability for the candidate to persist as "best" can be 4891 // achieved by replacing it with "next" k-th candidate with the probability 4892 // of 1/k. It can be easily shown that by the end of the run, the 4893 // probability for any candidate is converged to 1/n, thus giving the 4894 // uniform distribution among all the candidates. 4895 // 4896 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4897 #define RANDOMIZED_DOMAIN_POW 29 4898 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4899 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4900 bool Compile::randomized_select(int count) { 4901 assert(count > 0, "only positive"); 4902 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4903 } 4904 4905 #ifdef ASSERT 4906 // Failures are geometrically distributed with probability 1/StressBailoutMean. 4907 bool Compile::fail_randomly() { 4908 if ((random() % StressBailoutMean) != 0) { 4909 return false; 4910 } 4911 record_failure("StressBailout"); 4912 return true; 4913 } 4914 4915 bool Compile::failure_is_artificial() { 4916 return C->failure_reason_is("StressBailout"); 4917 } 4918 #endif 4919 4920 CloneMap& Compile::clone_map() { return _clone_map; } 4921 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4922 4923 void NodeCloneInfo::dump_on(outputStream* st) const { 4924 st->print(" {%d:%d} ", idx(), gen()); 4925 } 4926 4927 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4928 uint64_t val = value(old->_idx); 4929 NodeCloneInfo cio(val); 4930 assert(val != 0, "old node should be in the map"); 4931 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4932 insert(nnn->_idx, cin.get()); 4933 #ifndef PRODUCT 4934 if (is_debug()) { 4935 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4936 } 4937 #endif 4938 } 4939 4940 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4941 NodeCloneInfo cio(value(old->_idx)); 4942 if (cio.get() == 0) { 4943 cio.set(old->_idx, 0); 4944 insert(old->_idx, cio.get()); 4945 #ifndef PRODUCT 4946 if (is_debug()) { 4947 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4948 } 4949 #endif 4950 } 4951 clone(old, nnn, gen); 4952 } 4953 4954 int CloneMap::max_gen() const { 4955 int g = 0; 4956 DictI di(_dict); 4957 for(; di.test(); ++di) { 4958 int t = gen(di._key); 4959 if (g < t) { 4960 g = t; 4961 #ifndef PRODUCT 4962 if (is_debug()) { 4963 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4964 } 4965 #endif 4966 } 4967 } 4968 return g; 4969 } 4970 4971 void CloneMap::dump(node_idx_t key, outputStream* st) const { 4972 uint64_t val = value(key); 4973 if (val != 0) { 4974 NodeCloneInfo ni(val); 4975 ni.dump_on(st); 4976 } 4977 } 4978 4979 void Compile::shuffle_macro_nodes() { 4980 if (_macro_nodes.length() < 2) { 4981 return; 4982 } 4983 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) { 4984 uint j = C->random() % (i + 1); 4985 swap(_macro_nodes.at(i), _macro_nodes.at(j)); 4986 } 4987 } 4988 4989 // Move Allocate nodes to the start of the list 4990 void Compile::sort_macro_nodes() { 4991 int count = macro_count(); 4992 int allocates = 0; 4993 for (int i = 0; i < count; i++) { 4994 Node* n = macro_node(i); 4995 if (n->is_Allocate()) { 4996 if (i != allocates) { 4997 Node* tmp = macro_node(allocates); 4998 _macro_nodes.at_put(allocates, n); 4999 _macro_nodes.at_put(i, tmp); 5000 } 5001 allocates++; 5002 } 5003 } 5004 } 5005 5006 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) { 5007 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code. 5008 EventCompilerPhase event(UNTIMED); 5009 if (event.should_commit()) { 5010 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 5011 } 5012 #ifndef PRODUCT 5013 ResourceMark rm; 5014 stringStream ss; 5015 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt)); 5016 int iter = ++_igv_phase_iter[cpt]; 5017 if (iter > 1) { 5018 ss.print(" %d", iter); 5019 } 5020 if (n != nullptr) { 5021 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]); 5022 if (n->is_Call()) { 5023 CallNode* call = n->as_Call(); 5024 if (call->_name != nullptr) { 5025 // E.g. uncommon traps etc. 5026 ss.print(" - %s", call->_name); 5027 } else if (call->is_CallJava()) { 5028 CallJavaNode* call_java = call->as_CallJava(); 5029 if (call_java->method() != nullptr) { 5030 ss.print(" -"); 5031 call_java->method()->print_short_name(&ss); 5032 } 5033 } 5034 } 5035 } 5036 5037 const char* name = ss.as_string(); 5038 if (should_print_igv(level)) { 5039 _igv_printer->print_graph(name); 5040 } 5041 if (should_print_phase(cpt)) { 5042 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt)); 5043 } 5044 #endif 5045 C->_latest_stage_start_counter.stamp(); 5046 } 5047 5048 // Only used from CompileWrapper 5049 void Compile::begin_method() { 5050 #ifndef PRODUCT 5051 if (_method != nullptr && should_print_igv(1)) { 5052 _igv_printer->begin_method(); 5053 } 5054 #endif 5055 C->_latest_stage_start_counter.stamp(); 5056 } 5057 5058 // Only used from CompileWrapper 5059 void Compile::end_method() { 5060 EventCompilerPhase event(UNTIMED); 5061 if (event.should_commit()) { 5062 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1); 5063 } 5064 5065 #ifndef PRODUCT 5066 if (_method != nullptr && should_print_igv(1)) { 5067 _igv_printer->end_method(); 5068 } 5069 #endif 5070 } 5071 5072 bool Compile::should_print_phase(CompilerPhaseType cpt) { 5073 #ifndef PRODUCT 5074 if (_directive->should_print_phase(cpt)) { 5075 return true; 5076 } 5077 #endif 5078 return false; 5079 } 5080 5081 #ifndef PRODUCT 5082 void Compile::init_igv() { 5083 if (_igv_printer == nullptr) { 5084 _igv_printer = IdealGraphPrinter::printer(); 5085 _igv_printer->set_compile(this); 5086 } 5087 } 5088 #endif 5089 5090 bool Compile::should_print_igv(const int level) { 5091 #ifndef PRODUCT 5092 if (PrintIdealGraphLevel < 0) { // disabled by the user 5093 return false; 5094 } 5095 5096 bool need = directive()->IGVPrintLevelOption >= level; 5097 if (need) { 5098 Compile::init_igv(); 5099 } 5100 return need; 5101 #else 5102 return false; 5103 #endif 5104 } 5105 5106 #ifndef PRODUCT 5107 IdealGraphPrinter* Compile::_debug_file_printer = nullptr; 5108 IdealGraphPrinter* Compile::_debug_network_printer = nullptr; 5109 5110 // Called from debugger. Prints method to the default file with the default phase name. 5111 // This works regardless of any Ideal Graph Visualizer flags set or not. 5112 void igv_print() { 5113 Compile::current()->igv_print_method_to_file(); 5114 } 5115 5116 // Same as igv_print() above but with a specified phase name. 5117 void igv_print(const char* phase_name) { 5118 Compile::current()->igv_print_method_to_file(phase_name); 5119 } 5120 5121 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 5122 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 5123 // This works regardless of any Ideal Graph Visualizer flags set or not. 5124 void igv_print(bool network) { 5125 if (network) { 5126 Compile::current()->igv_print_method_to_network(); 5127 } else { 5128 Compile::current()->igv_print_method_to_file(); 5129 } 5130 } 5131 5132 // Same as igv_print(bool network) above but with a specified phase name. 5133 void igv_print(bool network, const char* phase_name) { 5134 if (network) { 5135 Compile::current()->igv_print_method_to_network(phase_name); 5136 } else { 5137 Compile::current()->igv_print_method_to_file(phase_name); 5138 } 5139 } 5140 5141 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 5142 void igv_print_default() { 5143 Compile::current()->print_method(PHASE_DEBUG, 0); 5144 } 5145 5146 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 5147 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 5148 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 5149 void igv_append() { 5150 Compile::current()->igv_print_method_to_file("Debug", true); 5151 } 5152 5153 // Same as igv_append() above but with a specified phase name. 5154 void igv_append(const char* phase_name) { 5155 Compile::current()->igv_print_method_to_file(phase_name, true); 5156 } 5157 5158 void Compile::igv_print_method_to_file(const char* phase_name, bool append) { 5159 const char* file_name = "custom_debug.xml"; 5160 if (_debug_file_printer == nullptr) { 5161 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 5162 } else { 5163 _debug_file_printer->update_compiled_method(C->method()); 5164 } 5165 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 5166 _debug_file_printer->print_graph(phase_name); 5167 } 5168 5169 void Compile::igv_print_method_to_network(const char* phase_name) { 5170 ResourceMark rm; 5171 GrowableArray<const Node*> empty_list; 5172 igv_print_graph_to_network(phase_name, (Node*) C->root(), empty_list); 5173 } 5174 5175 void Compile::igv_print_graph_to_network(const char* name, Node* node, GrowableArray<const Node*>& visible_nodes) { 5176 if (_debug_network_printer == nullptr) { 5177 _debug_network_printer = new IdealGraphPrinter(C); 5178 } else { 5179 _debug_network_printer->update_compiled_method(C->method()); 5180 } 5181 tty->print_cr("Method printed over network stream to IGV"); 5182 _debug_network_printer->print(name, C->root(), visible_nodes); 5183 } 5184 #endif 5185 5186 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 5187 if (type != nullptr && phase->type(value)->higher_equal(type)) { 5188 return value; 5189 } 5190 Node* result = nullptr; 5191 if (bt == T_BYTE) { 5192 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 5193 result = new RShiftINode(result, phase->intcon(24)); 5194 } else if (bt == T_BOOLEAN) { 5195 result = new AndINode(value, phase->intcon(0xFF)); 5196 } else if (bt == T_CHAR) { 5197 result = new AndINode(value,phase->intcon(0xFFFF)); 5198 } else { 5199 assert(bt == T_SHORT, "unexpected narrow type"); 5200 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 5201 result = new RShiftINode(result, phase->intcon(16)); 5202 } 5203 if (transform_res) { 5204 result = phase->transform(result); 5205 } 5206 return result; 5207 } 5208 5209 void Compile::record_method_not_compilable_oom() { 5210 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit()); 5211 }