1 /* 2 * Copyright (c) 1997, 2023, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/javaClasses.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compilationMemoryStatistic.hpp" 33 #include "compiler/compileBroker.hpp" 34 #include "compiler/compileLog.hpp" 35 #include "compiler/disassembler.hpp" 36 #include "compiler/oopMap.hpp" 37 #include "gc/shared/barrierSet.hpp" 38 #include "gc/shared/c2/barrierSetC2.hpp" 39 #include "jfr/jfrEvents.hpp" 40 #include "jvm_io.h" 41 #include "memory/allocation.hpp" 42 #include "memory/resourceArea.hpp" 43 #include "opto/addnode.hpp" 44 #include "opto/block.hpp" 45 #include "opto/c2compiler.hpp" 46 #include "opto/callGenerator.hpp" 47 #include "opto/callnode.hpp" 48 #include "opto/castnode.hpp" 49 #include "opto/cfgnode.hpp" 50 #include "opto/chaitin.hpp" 51 #include "opto/compile.hpp" 52 #include "opto/connode.hpp" 53 #include "opto/convertnode.hpp" 54 #include "opto/divnode.hpp" 55 #include "opto/escape.hpp" 56 #include "opto/idealGraphPrinter.hpp" 57 #include "opto/loopnode.hpp" 58 #include "opto/machnode.hpp" 59 #include "opto/macro.hpp" 60 #include "opto/matcher.hpp" 61 #include "opto/mathexactnode.hpp" 62 #include "opto/memnode.hpp" 63 #include "opto/mulnode.hpp" 64 #include "opto/narrowptrnode.hpp" 65 #include "opto/node.hpp" 66 #include "opto/opcodes.hpp" 67 #include "opto/output.hpp" 68 #include "opto/parse.hpp" 69 #include "opto/phaseX.hpp" 70 #include "opto/rootnode.hpp" 71 #include "opto/runtime.hpp" 72 #include "opto/stringopts.hpp" 73 #include "opto/type.hpp" 74 #include "opto/vector.hpp" 75 #include "opto/vectornode.hpp" 76 #include "runtime/globals_extension.hpp" 77 #include "runtime/sharedRuntime.hpp" 78 #include "runtime/signature.hpp" 79 #include "runtime/stubRoutines.hpp" 80 #include "runtime/timer.hpp" 81 #include "utilities/align.hpp" 82 #include "utilities/copy.hpp" 83 #include "utilities/macros.hpp" 84 #include "utilities/resourceHash.hpp" 85 86 // -------------------- Compile::mach_constant_base_node ----------------------- 87 // Constant table base node singleton. 88 MachConstantBaseNode* Compile::mach_constant_base_node() { 89 if (_mach_constant_base_node == nullptr) { 90 _mach_constant_base_node = new MachConstantBaseNode(); 91 _mach_constant_base_node->add_req(C->root()); 92 } 93 return _mach_constant_base_node; 94 } 95 96 97 /// Support for intrinsics. 98 99 // Return the index at which m must be inserted (or already exists). 100 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 101 class IntrinsicDescPair { 102 private: 103 ciMethod* _m; 104 bool _is_virtual; 105 public: 106 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 107 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 108 ciMethod* m= elt->method(); 109 ciMethod* key_m = key->_m; 110 if (key_m < m) return -1; 111 else if (key_m > m) return 1; 112 else { 113 bool is_virtual = elt->is_virtual(); 114 bool key_virtual = key->_is_virtual; 115 if (key_virtual < is_virtual) return -1; 116 else if (key_virtual > is_virtual) return 1; 117 else return 0; 118 } 119 } 120 }; 121 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 122 #ifdef ASSERT 123 for (int i = 1; i < _intrinsics.length(); i++) { 124 CallGenerator* cg1 = _intrinsics.at(i-1); 125 CallGenerator* cg2 = _intrinsics.at(i); 126 assert(cg1->method() != cg2->method() 127 ? cg1->method() < cg2->method() 128 : cg1->is_virtual() < cg2->is_virtual(), 129 "compiler intrinsics list must stay sorted"); 130 } 131 #endif 132 IntrinsicDescPair pair(m, is_virtual); 133 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 134 } 135 136 void Compile::register_intrinsic(CallGenerator* cg) { 137 bool found = false; 138 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 139 assert(!found, "registering twice"); 140 _intrinsics.insert_before(index, cg); 141 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 142 } 143 144 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 145 assert(m->is_loaded(), "don't try this on unloaded methods"); 146 if (_intrinsics.length() > 0) { 147 bool found = false; 148 int index = intrinsic_insertion_index(m, is_virtual, found); 149 if (found) { 150 return _intrinsics.at(index); 151 } 152 } 153 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 154 if (m->intrinsic_id() != vmIntrinsics::_none && 155 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 156 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 157 if (cg != nullptr) { 158 // Save it for next time: 159 register_intrinsic(cg); 160 return cg; 161 } else { 162 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 163 } 164 } 165 return nullptr; 166 } 167 168 // Compile::make_vm_intrinsic is defined in library_call.cpp. 169 170 #ifndef PRODUCT 171 // statistics gathering... 172 173 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0}; 174 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0}; 175 176 inline int as_int(vmIntrinsics::ID id) { 177 return vmIntrinsics::as_int(id); 178 } 179 180 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 181 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 182 int oflags = _intrinsic_hist_flags[as_int(id)]; 183 assert(flags != 0, "what happened?"); 184 if (is_virtual) { 185 flags |= _intrinsic_virtual; 186 } 187 bool changed = (flags != oflags); 188 if ((flags & _intrinsic_worked) != 0) { 189 juint count = (_intrinsic_hist_count[as_int(id)] += 1); 190 if (count == 1) { 191 changed = true; // first time 192 } 193 // increment the overall count also: 194 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1; 195 } 196 if (changed) { 197 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 198 // Something changed about the intrinsic's virtuality. 199 if ((flags & _intrinsic_virtual) != 0) { 200 // This is the first use of this intrinsic as a virtual call. 201 if (oflags != 0) { 202 // We already saw it as a non-virtual, so note both cases. 203 flags |= _intrinsic_both; 204 } 205 } else if ((oflags & _intrinsic_both) == 0) { 206 // This is the first use of this intrinsic as a non-virtual 207 flags |= _intrinsic_both; 208 } 209 } 210 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags); 211 } 212 // update the overall flags also: 213 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags; 214 return changed; 215 } 216 217 static char* format_flags(int flags, char* buf) { 218 buf[0] = 0; 219 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 220 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 221 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 222 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 223 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 224 if (buf[0] == 0) strcat(buf, ","); 225 assert(buf[0] == ',', "must be"); 226 return &buf[1]; 227 } 228 229 void Compile::print_intrinsic_statistics() { 230 char flagsbuf[100]; 231 ttyLocker ttyl; 232 if (xtty != nullptr) xtty->head("statistics type='intrinsic'"); 233 tty->print_cr("Compiler intrinsic usage:"); 234 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)]; 235 if (total == 0) total = 1; // avoid div0 in case of no successes 236 #define PRINT_STAT_LINE(name, c, f) \ 237 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 238 for (auto id : EnumRange<vmIntrinsicID>{}) { 239 int flags = _intrinsic_hist_flags[as_int(id)]; 240 juint count = _intrinsic_hist_count[as_int(id)]; 241 if ((flags | count) != 0) { 242 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 243 } 244 } 245 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf)); 246 if (xtty != nullptr) xtty->tail("statistics"); 247 } 248 249 void Compile::print_statistics() { 250 { ttyLocker ttyl; 251 if (xtty != nullptr) xtty->head("statistics type='opto'"); 252 Parse::print_statistics(); 253 PhaseStringOpts::print_statistics(); 254 PhaseCCP::print_statistics(); 255 PhaseRegAlloc::print_statistics(); 256 PhaseOutput::print_statistics(); 257 PhasePeephole::print_statistics(); 258 PhaseIdealLoop::print_statistics(); 259 ConnectionGraph::print_statistics(); 260 PhaseMacroExpand::print_statistics(); 261 if (xtty != nullptr) xtty->tail("statistics"); 262 } 263 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) { 264 // put this under its own <statistics> element. 265 print_intrinsic_statistics(); 266 } 267 } 268 #endif //PRODUCT 269 270 void Compile::gvn_replace_by(Node* n, Node* nn) { 271 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 272 Node* use = n->last_out(i); 273 bool is_in_table = initial_gvn()->hash_delete(use); 274 uint uses_found = 0; 275 for (uint j = 0; j < use->len(); j++) { 276 if (use->in(j) == n) { 277 if (j < use->req()) 278 use->set_req(j, nn); 279 else 280 use->set_prec(j, nn); 281 uses_found++; 282 } 283 } 284 if (is_in_table) { 285 // reinsert into table 286 initial_gvn()->hash_find_insert(use); 287 } 288 record_for_igvn(use); 289 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist); 290 i -= uses_found; // we deleted 1 or more copies of this edge 291 } 292 } 293 294 295 // Identify all nodes that are reachable from below, useful. 296 // Use breadth-first pass that records state in a Unique_Node_List, 297 // recursive traversal is slower. 298 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 299 int estimated_worklist_size = live_nodes(); 300 useful.map( estimated_worklist_size, nullptr ); // preallocate space 301 302 // Initialize worklist 303 if (root() != nullptr) { useful.push(root()); } 304 // If 'top' is cached, declare it useful to preserve cached node 305 if (cached_top_node()) { useful.push(cached_top_node()); } 306 307 // Push all useful nodes onto the list, breadthfirst 308 for( uint next = 0; next < useful.size(); ++next ) { 309 assert( next < unique(), "Unique useful nodes < total nodes"); 310 Node *n = useful.at(next); 311 uint max = n->len(); 312 for( uint i = 0; i < max; ++i ) { 313 Node *m = n->in(i); 314 if (not_a_node(m)) continue; 315 useful.push(m); 316 } 317 } 318 } 319 320 // Update dead_node_list with any missing dead nodes using useful 321 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 322 void Compile::update_dead_node_list(Unique_Node_List &useful) { 323 uint max_idx = unique(); 324 VectorSet& useful_node_set = useful.member_set(); 325 326 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 327 // If node with index node_idx is not in useful set, 328 // mark it as dead in dead node list. 329 if (!useful_node_set.test(node_idx)) { 330 record_dead_node(node_idx); 331 } 332 } 333 } 334 335 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 336 int shift = 0; 337 for (int i = 0; i < inlines->length(); i++) { 338 CallGenerator* cg = inlines->at(i); 339 if (useful.member(cg->call_node())) { 340 if (shift > 0) { 341 inlines->at_put(i - shift, cg); 342 } 343 } else { 344 shift++; // skip over the dead element 345 } 346 } 347 if (shift > 0) { 348 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array 349 } 350 } 351 352 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) { 353 assert(dead != nullptr && dead->is_Call(), "sanity"); 354 int found = 0; 355 for (int i = 0; i < inlines->length(); i++) { 356 if (inlines->at(i)->call_node() == dead) { 357 inlines->remove_at(i); 358 found++; 359 NOT_DEBUG( break; ) // elements are unique, so exit early 360 } 361 } 362 assert(found <= 1, "not unique"); 363 } 364 365 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)> 366 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) { 367 for (int i = node_list.length() - 1; i >= 0; i--) { 368 N* node = node_list.at(i); 369 if (!useful.member(node)) { 370 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 371 } 372 } 373 } 374 375 void Compile::remove_useless_node(Node* dead) { 376 remove_modified_node(dead); 377 378 // Constant node that has no out-edges and has only one in-edge from 379 // root is usually dead. However, sometimes reshaping walk makes 380 // it reachable by adding use edges. So, we will NOT count Con nodes 381 // as dead to be conservative about the dead node count at any 382 // given time. 383 if (!dead->is_Con()) { 384 record_dead_node(dead->_idx); 385 } 386 if (dead->is_macro()) { 387 remove_macro_node(dead); 388 } 389 if (dead->is_expensive()) { 390 remove_expensive_node(dead); 391 } 392 if (dead->Opcode() == Op_Opaque4) { 393 remove_template_assertion_predicate_opaq(dead); 394 } 395 if (dead->is_ParsePredicate()) { 396 remove_parse_predicate(dead->as_ParsePredicate()); 397 } 398 if (dead->for_post_loop_opts_igvn()) { 399 remove_from_post_loop_opts_igvn(dead); 400 } 401 if (dead->is_Call()) { 402 remove_useless_late_inlines( &_late_inlines, dead); 403 remove_useless_late_inlines( &_string_late_inlines, dead); 404 remove_useless_late_inlines( &_boxing_late_inlines, dead); 405 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead); 406 407 if (dead->is_CallStaticJava()) { 408 remove_unstable_if_trap(dead->as_CallStaticJava(), false); 409 } 410 } 411 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 412 bs->unregister_potential_barrier_node(dead); 413 } 414 415 // Disconnect all useless nodes by disconnecting those at the boundary. 416 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) { 417 uint next = 0; 418 while (next < useful.size()) { 419 Node *n = useful.at(next++); 420 if (n->is_SafePoint()) { 421 // We're done with a parsing phase. Replaced nodes are not valid 422 // beyond that point. 423 n->as_SafePoint()->delete_replaced_nodes(); 424 } 425 // Use raw traversal of out edges since this code removes out edges 426 int max = n->outcnt(); 427 for (int j = 0; j < max; ++j) { 428 Node* child = n->raw_out(j); 429 if (!useful.member(child)) { 430 assert(!child->is_top() || child != top(), 431 "If top is cached in Compile object it is in useful list"); 432 // Only need to remove this out-edge to the useless node 433 n->raw_del_out(j); 434 --j; 435 --max; 436 } 437 } 438 if (n->outcnt() == 1 && n->has_special_unique_user()) { 439 assert(useful.member(n->unique_out()), "do not push a useless node"); 440 worklist.push(n->unique_out()); 441 } 442 } 443 444 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes 445 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes 446 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes 447 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes 448 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass 449 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps 450 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes 451 #ifdef ASSERT 452 if (_modified_nodes != nullptr) { 453 _modified_nodes->remove_useless_nodes(useful.member_set()); 454 } 455 #endif 456 457 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 458 bs->eliminate_useless_gc_barriers(useful, this); 459 // clean up the late inline lists 460 remove_useless_late_inlines( &_late_inlines, useful); 461 remove_useless_late_inlines( &_string_late_inlines, useful); 462 remove_useless_late_inlines( &_boxing_late_inlines, useful); 463 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful); 464 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 465 } 466 467 // ============================================================================ 468 //------------------------------CompileWrapper--------------------------------- 469 class CompileWrapper : public StackObj { 470 Compile *const _compile; 471 public: 472 CompileWrapper(Compile* compile); 473 474 ~CompileWrapper(); 475 }; 476 477 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 478 // the Compile* pointer is stored in the current ciEnv: 479 ciEnv* env = compile->env(); 480 assert(env == ciEnv::current(), "must already be a ciEnv active"); 481 assert(env->compiler_data() == nullptr, "compile already active?"); 482 env->set_compiler_data(compile); 483 assert(compile == Compile::current(), "sanity"); 484 485 compile->set_type_dict(nullptr); 486 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 487 compile->clone_map().set_clone_idx(0); 488 compile->set_type_last_size(0); 489 compile->set_last_tf(nullptr, nullptr); 490 compile->set_indexSet_arena(nullptr); 491 compile->set_indexSet_free_block_list(nullptr); 492 compile->init_type_arena(); 493 Type::Initialize(compile); 494 _compile->begin_method(); 495 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 496 } 497 CompileWrapper::~CompileWrapper() { 498 // simulate crash during compilation 499 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned"); 500 501 _compile->end_method(); 502 _compile->env()->set_compiler_data(nullptr); 503 } 504 505 506 //----------------------------print_compile_messages--------------------------- 507 void Compile::print_compile_messages() { 508 #ifndef PRODUCT 509 // Check if recompiling 510 if (!subsume_loads() && PrintOpto) { 511 // Recompiling without allowing machine instructions to subsume loads 512 tty->print_cr("*********************************************************"); 513 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 514 tty->print_cr("*********************************************************"); 515 } 516 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) { 517 // Recompiling without escape analysis 518 tty->print_cr("*********************************************************"); 519 tty->print_cr("** Bailout: Recompile without escape analysis **"); 520 tty->print_cr("*********************************************************"); 521 } 522 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) { 523 // Recompiling without iterative escape analysis 524 tty->print_cr("*********************************************************"); 525 tty->print_cr("** Bailout: Recompile without iterative escape analysis**"); 526 tty->print_cr("*********************************************************"); 527 } 528 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) { 529 // Recompiling without reducing allocation merges 530 tty->print_cr("*********************************************************"); 531 tty->print_cr("** Bailout: Recompile without reduce allocation merges **"); 532 tty->print_cr("*********************************************************"); 533 } 534 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) { 535 // Recompiling without boxing elimination 536 tty->print_cr("*********************************************************"); 537 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 538 tty->print_cr("*********************************************************"); 539 } 540 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) { 541 // Recompiling without locks coarsening 542 tty->print_cr("*********************************************************"); 543 tty->print_cr("** Bailout: Recompile without locks coarsening **"); 544 tty->print_cr("*********************************************************"); 545 } 546 if (env()->break_at_compile()) { 547 // Open the debugger when compiling this method. 548 tty->print("### Breaking when compiling: "); 549 method()->print_short_name(); 550 tty->cr(); 551 BREAKPOINT; 552 } 553 554 if( PrintOpto ) { 555 if (is_osr_compilation()) { 556 tty->print("[OSR]%3d", _compile_id); 557 } else { 558 tty->print("%3d", _compile_id); 559 } 560 } 561 #endif 562 } 563 564 #ifndef PRODUCT 565 void Compile::print_ideal_ir(const char* phase_name) { 566 // keep the following output all in one block 567 // This output goes directly to the tty, not the compiler log. 568 // To enable tools to match it up with the compilation activity, 569 // be sure to tag this tty output with the compile ID. 570 571 // Node dumping can cause a safepoint, which can break the tty lock. 572 // Buffer all node dumps, so that all safepoints happen before we lock. 573 ResourceMark rm; 574 stringStream ss; 575 576 if (_output == nullptr) { 577 ss.print_cr("AFTER: %s", phase_name); 578 // Print out all nodes in ascending order of index. 579 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss); 580 } else { 581 // Dump the node blockwise if we have a scheduling 582 _output->print_scheduling(&ss); 583 } 584 585 // Check that the lock is not broken by a safepoint. 586 NoSafepointVerifier nsv; 587 ttyLocker ttyl; 588 if (xtty != nullptr) { 589 xtty->head("ideal compile_id='%d'%s compile_phase='%s'", 590 compile_id(), 591 is_osr_compilation() ? " compile_kind='osr'" : "", 592 phase_name); 593 } 594 595 tty->print("%s", ss.as_string()); 596 597 if (xtty != nullptr) { 598 xtty->tail("ideal"); 599 } 600 } 601 #endif 602 603 // ============================================================================ 604 //------------------------------Compile standard------------------------------- 605 606 // Compile a method. entry_bci is -1 for normal compilations and indicates 607 // the continuation bci for on stack replacement. 608 609 610 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci, 611 Options options, DirectiveSet* directive) 612 : Phase(Compiler), 613 _compile_id(ci_env->compile_id()), 614 _options(options), 615 _method(target), 616 _entry_bci(osr_bci), 617 _ilt(nullptr), 618 _stub_function(nullptr), 619 _stub_name(nullptr), 620 _stub_entry_point(nullptr), 621 _max_node_limit(MaxNodeLimit), 622 _post_loop_opts_phase(false), 623 _inlining_progress(false), 624 _inlining_incrementally(false), 625 _do_cleanup(false), 626 _has_reserved_stack_access(target->has_reserved_stack_access()), 627 #ifndef PRODUCT 628 _igv_idx(0), 629 _trace_opto_output(directive->TraceOptoOutputOption), 630 #endif 631 _has_method_handle_invokes(false), 632 _clinit_barrier_on_entry(false), 633 _stress_seed(0), 634 _comp_arena(mtCompiler), 635 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 636 _env(ci_env), 637 _directive(directive), 638 _log(ci_env->log()), 639 _failure_reason(nullptr), 640 _intrinsics (comp_arena(), 0, 0, nullptr), 641 _macro_nodes (comp_arena(), 8, 0, nullptr), 642 _parse_predicates (comp_arena(), 8, 0, nullptr), 643 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr), 644 _expensive_nodes (comp_arena(), 8, 0, nullptr), 645 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr), 646 _unstable_if_traps (comp_arena(), 8, 0, nullptr), 647 _coarsened_locks (comp_arena(), 8, 0, nullptr), 648 _congraph(nullptr), 649 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 650 _unique(0), 651 _dead_node_count(0), 652 _dead_node_list(comp_arena()), 653 _node_arena_one(mtCompiler, Arena::Tag::tag_node), 654 _node_arena_two(mtCompiler, Arena::Tag::tag_node), 655 _node_arena(&_node_arena_one), 656 _mach_constant_base_node(nullptr), 657 _Compile_types(mtCompiler), 658 _initial_gvn(nullptr), 659 _igvn_worklist(nullptr), 660 _types(nullptr), 661 _node_hash(nullptr), 662 _late_inlines(comp_arena(), 2, 0, nullptr), 663 _string_late_inlines(comp_arena(), 2, 0, nullptr), 664 _boxing_late_inlines(comp_arena(), 2, 0, nullptr), 665 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr), 666 _late_inlines_pos(0), 667 _number_of_mh_late_inlines(0), 668 _oom(false), 669 _print_inlining_stream(new (mtCompiler) stringStream()), 670 _print_inlining_list(nullptr), 671 _print_inlining_idx(0), 672 _print_inlining_output(nullptr), 673 _replay_inline_data(nullptr), 674 _java_calls(0), 675 _inner_loops(0), 676 _interpreter_frame_size(0), 677 _output(nullptr) 678 #ifndef PRODUCT 679 , _in_dump_cnt(0) 680 #endif 681 { 682 C = this; 683 CompileWrapper cw(this); 684 685 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 686 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false); 687 688 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 689 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 690 // We can always print a disassembly, either abstract (hex dump) or 691 // with the help of a suitable hsdis library. Thus, we should not 692 // couple print_assembly and print_opto_assembly controls. 693 // But: always print opto and regular assembly on compile command 'print'. 694 bool print_assembly = directive->PrintAssemblyOption; 695 set_print_assembly(print_opto_assembly || print_assembly); 696 #else 697 set_print_assembly(false); // must initialize. 698 #endif 699 700 #ifndef PRODUCT 701 set_parsed_irreducible_loop(false); 702 #endif 703 704 if (directive->ReplayInlineOption) { 705 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 706 } 707 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 708 set_print_intrinsics(directive->PrintIntrinsicsOption); 709 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 710 711 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 712 // Make sure the method being compiled gets its own MDO, 713 // so we can at least track the decompile_count(). 714 // Need MDO to record RTM code generation state. 715 method()->ensure_method_data(); 716 } 717 718 Init(/*do_aliasing=*/ true); 719 720 print_compile_messages(); 721 722 _ilt = InlineTree::build_inline_tree_root(); 723 724 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 725 assert(num_alias_types() >= AliasIdxRaw, ""); 726 727 #define MINIMUM_NODE_HASH 1023 728 729 // GVN that will be run immediately on new nodes 730 uint estimated_size = method()->code_size()*4+64; 731 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 732 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 733 _types = new (comp_arena()) Type_Array(comp_arena()); 734 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size); 735 PhaseGVN gvn; 736 set_initial_gvn(&gvn); 737 738 print_inlining_init(); 739 { // Scope for timing the parser 740 TracePhase tp("parse", &timers[_t_parser]); 741 742 // Put top into the hash table ASAP. 743 initial_gvn()->transform_no_reclaim(top()); 744 745 // Set up tf(), start(), and find a CallGenerator. 746 CallGenerator* cg = nullptr; 747 if (is_osr_compilation()) { 748 const TypeTuple *domain = StartOSRNode::osr_domain(); 749 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 750 init_tf(TypeFunc::make(domain, range)); 751 StartNode* s = new StartOSRNode(root(), domain); 752 initial_gvn()->set_type_bottom(s); 753 init_start(s); 754 cg = CallGenerator::for_osr(method(), entry_bci()); 755 } else { 756 // Normal case. 757 init_tf(TypeFunc::make(method())); 758 StartNode* s = new StartNode(root(), tf()->domain()); 759 initial_gvn()->set_type_bottom(s); 760 init_start(s); 761 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 762 // With java.lang.ref.reference.get() we must go through the 763 // intrinsic - even when get() is the root 764 // method of the compile - so that, if necessary, the value in 765 // the referent field of the reference object gets recorded by 766 // the pre-barrier code. 767 cg = find_intrinsic(method(), false); 768 } 769 if (cg == nullptr) { 770 float past_uses = method()->interpreter_invocation_count(); 771 float expected_uses = past_uses; 772 cg = CallGenerator::for_inline(method(), expected_uses); 773 } 774 } 775 if (failing()) return; 776 if (cg == nullptr) { 777 const char* reason = InlineTree::check_can_parse(method()); 778 assert(reason != nullptr, "expect reason for parse failure"); 779 stringStream ss; 780 ss.print("cannot parse method: %s", reason); 781 record_method_not_compilable(ss.as_string()); 782 return; 783 } 784 785 gvn.set_type(root(), root()->bottom_type()); 786 787 JVMState* jvms = build_start_state(start(), tf()); 788 if ((jvms = cg->generate(jvms)) == nullptr) { 789 assert(failure_reason() != nullptr, "expect reason for parse failure"); 790 stringStream ss; 791 ss.print("method parse failed: %s", failure_reason()); 792 record_method_not_compilable(ss.as_string()); 793 return; 794 } 795 GraphKit kit(jvms); 796 797 if (!kit.stopped()) { 798 // Accept return values, and transfer control we know not where. 799 // This is done by a special, unique ReturnNode bound to root. 800 return_values(kit.jvms()); 801 } 802 803 if (kit.has_exceptions()) { 804 // Any exceptions that escape from this call must be rethrown 805 // to whatever caller is dynamically above us on the stack. 806 // This is done by a special, unique RethrowNode bound to root. 807 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 808 } 809 810 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 811 812 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 813 inline_string_calls(true); 814 } 815 816 if (failing()) return; 817 818 // Remove clutter produced by parsing. 819 if (!failing()) { 820 ResourceMark rm; 821 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 822 } 823 } 824 825 // Note: Large methods are capped off in do_one_bytecode(). 826 if (failing()) return; 827 828 // After parsing, node notes are no longer automagic. 829 // They must be propagated by register_new_node_with_optimizer(), 830 // clone(), or the like. 831 set_default_node_notes(nullptr); 832 833 #ifndef PRODUCT 834 if (should_print_igv(1)) { 835 _igv_printer->print_inlining(); 836 } 837 #endif 838 839 if (failing()) return; 840 NOT_PRODUCT( verify_graph_edges(); ) 841 842 // If any phase is randomized for stress testing, seed random number 843 // generation and log the seed for repeatability. 844 if (StressLCM || StressGCM || StressIGVN || StressCCP || StressIncrementalInlining) { 845 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) { 846 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds()); 847 FLAG_SET_ERGO(StressSeed, _stress_seed); 848 } else { 849 _stress_seed = StressSeed; 850 } 851 if (_log != nullptr) { 852 _log->elem("stress_test seed='%u'", _stress_seed); 853 } 854 } 855 856 // Now optimize 857 Optimize(); 858 if (failing()) return; 859 NOT_PRODUCT( verify_graph_edges(); ) 860 861 #ifndef PRODUCT 862 if (should_print_ideal()) { 863 print_ideal_ir("print_ideal"); 864 } 865 #endif 866 867 #ifdef ASSERT 868 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 869 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 870 #endif 871 872 // Dump compilation data to replay it. 873 if (directive->DumpReplayOption) { 874 env()->dump_replay_data(_compile_id); 875 } 876 if (directive->DumpInlineOption && (ilt() != nullptr)) { 877 env()->dump_inline_data(_compile_id); 878 } 879 880 // Now that we know the size of all the monitors we can add a fixed slot 881 // for the original deopt pc. 882 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size); 883 set_fixed_slots(next_slot); 884 885 // Compute when to use implicit null checks. Used by matching trap based 886 // nodes and NullCheck optimization. 887 set_allowed_deopt_reasons(); 888 889 // Now generate code 890 Code_Gen(); 891 } 892 893 //------------------------------Compile---------------------------------------- 894 // Compile a runtime stub 895 Compile::Compile( ciEnv* ci_env, 896 TypeFunc_generator generator, 897 address stub_function, 898 const char *stub_name, 899 int is_fancy_jump, 900 bool pass_tls, 901 bool return_pc, 902 DirectiveSet* directive) 903 : Phase(Compiler), 904 _compile_id(0), 905 _options(Options::for_runtime_stub()), 906 _method(nullptr), 907 _entry_bci(InvocationEntryBci), 908 _stub_function(stub_function), 909 _stub_name(stub_name), 910 _stub_entry_point(nullptr), 911 _max_node_limit(MaxNodeLimit), 912 _post_loop_opts_phase(false), 913 _inlining_progress(false), 914 _inlining_incrementally(false), 915 _has_reserved_stack_access(false), 916 #ifndef PRODUCT 917 _igv_idx(0), 918 _trace_opto_output(directive->TraceOptoOutputOption), 919 #endif 920 _has_method_handle_invokes(false), 921 _clinit_barrier_on_entry(false), 922 _stress_seed(0), 923 _comp_arena(mtCompiler), 924 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 925 _env(ci_env), 926 _directive(directive), 927 _log(ci_env->log()), 928 _failure_reason(nullptr), 929 _congraph(nullptr), 930 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 931 _unique(0), 932 _dead_node_count(0), 933 _dead_node_list(comp_arena()), 934 _node_arena_one(mtCompiler), 935 _node_arena_two(mtCompiler), 936 _node_arena(&_node_arena_one), 937 _mach_constant_base_node(nullptr), 938 _Compile_types(mtCompiler), 939 _initial_gvn(nullptr), 940 _igvn_worklist(nullptr), 941 _types(nullptr), 942 _node_hash(nullptr), 943 _number_of_mh_late_inlines(0), 944 _oom(false), 945 _print_inlining_stream(new (mtCompiler) stringStream()), 946 _print_inlining_list(nullptr), 947 _print_inlining_idx(0), 948 _print_inlining_output(nullptr), 949 _replay_inline_data(nullptr), 950 _java_calls(0), 951 _inner_loops(0), 952 _interpreter_frame_size(0), 953 _output(nullptr), 954 #ifndef PRODUCT 955 _in_dump_cnt(0), 956 #endif 957 _allowed_reasons(0) { 958 C = this; 959 960 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false); 961 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false); 962 963 #ifndef PRODUCT 964 set_print_assembly(PrintFrameConverterAssembly); 965 set_parsed_irreducible_loop(false); 966 #else 967 set_print_assembly(false); // Must initialize. 968 #endif 969 set_has_irreducible_loop(false); // no loops 970 971 CompileWrapper cw(this); 972 Init(/*do_aliasing=*/ false); 973 init_tf((*generator)()); 974 975 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 976 _types = new (comp_arena()) Type_Array(comp_arena()); 977 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255); 978 { 979 PhaseGVN gvn; 980 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 981 gvn.transform_no_reclaim(top()); 982 983 GraphKit kit; 984 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 985 } 986 987 NOT_PRODUCT( verify_graph_edges(); ) 988 989 Code_Gen(); 990 } 991 992 //------------------------------Init------------------------------------------- 993 // Prepare for a single compilation 994 void Compile::Init(bool aliasing) { 995 _do_aliasing = aliasing; 996 _unique = 0; 997 _regalloc = nullptr; 998 999 _tf = nullptr; // filled in later 1000 _top = nullptr; // cached later 1001 _matcher = nullptr; // filled in later 1002 _cfg = nullptr; // filled in later 1003 1004 IA32_ONLY( set_24_bit_selection_and_mode(true, false); ) 1005 1006 _node_note_array = nullptr; 1007 _default_node_notes = nullptr; 1008 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize() 1009 1010 _immutable_memory = nullptr; // filled in at first inquiry 1011 1012 #ifdef ASSERT 1013 _phase_optimize_finished = 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 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1045 _loop_opts_cnt = LoopOptsCount; 1046 set_do_inlining(Inline); 1047 set_max_inline_size(MaxInlineSize); 1048 set_freq_inline_size(FreqInlineSize); 1049 set_do_scheduling(OptoScheduling); 1050 1051 set_do_vector_loop(false); 1052 set_has_monitors(false); 1053 1054 if (AllowVectorizeOnDemand) { 1055 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1056 set_do_vector_loop(true); 1057 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());}) 1058 } else if (has_method() && method()->name() != 0 && 1059 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1060 set_do_vector_loop(true); 1061 } 1062 } 1063 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1064 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());}) 1065 1066 set_rtm_state(NoRTM); // No RTM lock eliding by default 1067 _max_node_limit = _directive->MaxNodeLimitOption; 1068 1069 #if INCLUDE_RTM_OPT 1070 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != nullptr)) { 1071 int rtm_state = method()->method_data()->rtm_state(); 1072 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) { 1073 // Don't generate RTM lock eliding code. 1074 set_rtm_state(NoRTM); 1075 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1076 // Generate RTM lock eliding code without abort ratio calculation code. 1077 set_rtm_state(UseRTM); 1078 } else if (UseRTMDeopt) { 1079 // Generate RTM lock eliding code and include abort ratio calculation 1080 // code if UseRTMDeopt is on. 1081 set_rtm_state(ProfileRTM); 1082 } 1083 } 1084 #endif 1085 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1086 set_clinit_barrier_on_entry(true); 1087 } 1088 if (debug_info()->recording_non_safepoints()) { 1089 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1090 (comp_arena(), 8, 0, nullptr)); 1091 set_default_node_notes(Node_Notes::make(this)); 1092 } 1093 1094 const int grow_ats = 16; 1095 _max_alias_types = grow_ats; 1096 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1097 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1098 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1099 { 1100 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1101 } 1102 // Initialize the first few types. 1103 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr); 1104 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1105 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1106 _num_alias_types = AliasIdxRaw+1; 1107 // Zero out the alias type cache. 1108 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1109 // A null adr_type hits in the cache right away. Preload the right answer. 1110 probe_alias_cache(nullptr)->_index = AliasIdxTop; 1111 } 1112 1113 //---------------------------init_start---------------------------------------- 1114 // Install the StartNode on this compile object. 1115 void Compile::init_start(StartNode* s) { 1116 if (failing()) 1117 return; // already failing 1118 assert(s == start(), ""); 1119 } 1120 1121 /** 1122 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1123 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1124 * the ideal graph. 1125 */ 1126 StartNode* Compile::start() const { 1127 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1128 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1129 Node* start = root()->fast_out(i); 1130 if (start->is_Start()) { 1131 return start->as_Start(); 1132 } 1133 } 1134 fatal("Did not find Start node!"); 1135 return nullptr; 1136 } 1137 1138 //-------------------------------immutable_memory------------------------------------- 1139 // Access immutable memory 1140 Node* Compile::immutable_memory() { 1141 if (_immutable_memory != nullptr) { 1142 return _immutable_memory; 1143 } 1144 StartNode* s = start(); 1145 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1146 Node *p = s->fast_out(i); 1147 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1148 _immutable_memory = p; 1149 return _immutable_memory; 1150 } 1151 } 1152 ShouldNotReachHere(); 1153 return nullptr; 1154 } 1155 1156 //----------------------set_cached_top_node------------------------------------ 1157 // Install the cached top node, and make sure Node::is_top works correctly. 1158 void Compile::set_cached_top_node(Node* tn) { 1159 if (tn != nullptr) verify_top(tn); 1160 Node* old_top = _top; 1161 _top = tn; 1162 // Calling Node::setup_is_top allows the nodes the chance to adjust 1163 // their _out arrays. 1164 if (_top != nullptr) _top->setup_is_top(); 1165 if (old_top != nullptr) old_top->setup_is_top(); 1166 assert(_top == nullptr || top()->is_top(), ""); 1167 } 1168 1169 #ifdef ASSERT 1170 uint Compile::count_live_nodes_by_graph_walk() { 1171 Unique_Node_List useful(comp_arena()); 1172 // Get useful node list by walking the graph. 1173 identify_useful_nodes(useful); 1174 return useful.size(); 1175 } 1176 1177 void Compile::print_missing_nodes() { 1178 1179 // Return if CompileLog is null and PrintIdealNodeCount is false. 1180 if ((_log == nullptr) && (! PrintIdealNodeCount)) { 1181 return; 1182 } 1183 1184 // This is an expensive function. It is executed only when the user 1185 // specifies VerifyIdealNodeCount option or otherwise knows the 1186 // additional work that needs to be done to identify reachable nodes 1187 // by walking the flow graph and find the missing ones using 1188 // _dead_node_list. 1189 1190 Unique_Node_List useful(comp_arena()); 1191 // Get useful node list by walking the graph. 1192 identify_useful_nodes(useful); 1193 1194 uint l_nodes = C->live_nodes(); 1195 uint l_nodes_by_walk = useful.size(); 1196 1197 if (l_nodes != l_nodes_by_walk) { 1198 if (_log != nullptr) { 1199 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1200 _log->stamp(); 1201 _log->end_head(); 1202 } 1203 VectorSet& useful_member_set = useful.member_set(); 1204 int last_idx = l_nodes_by_walk; 1205 for (int i = 0; i < last_idx; i++) { 1206 if (useful_member_set.test(i)) { 1207 if (_dead_node_list.test(i)) { 1208 if (_log != nullptr) { 1209 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1210 } 1211 if (PrintIdealNodeCount) { 1212 // Print the log message to tty 1213 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1214 useful.at(i)->dump(); 1215 } 1216 } 1217 } 1218 else if (! _dead_node_list.test(i)) { 1219 if (_log != nullptr) { 1220 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1221 } 1222 if (PrintIdealNodeCount) { 1223 // Print the log message to tty 1224 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1225 } 1226 } 1227 } 1228 if (_log != nullptr) { 1229 _log->tail("mismatched_nodes"); 1230 } 1231 } 1232 } 1233 void Compile::record_modified_node(Node* n) { 1234 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) { 1235 _modified_nodes->push(n); 1236 } 1237 } 1238 1239 void Compile::remove_modified_node(Node* n) { 1240 if (_modified_nodes != nullptr) { 1241 _modified_nodes->remove(n); 1242 } 1243 } 1244 #endif 1245 1246 #ifndef PRODUCT 1247 void Compile::verify_top(Node* tn) const { 1248 if (tn != nullptr) { 1249 assert(tn->is_Con(), "top node must be a constant"); 1250 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1251 assert(tn->in(0) != nullptr, "must have live top node"); 1252 } 1253 } 1254 #endif 1255 1256 1257 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1258 1259 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1260 guarantee(arr != nullptr, ""); 1261 int num_blocks = arr->length(); 1262 if (grow_by < num_blocks) grow_by = num_blocks; 1263 int num_notes = grow_by * _node_notes_block_size; 1264 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1265 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1266 while (num_notes > 0) { 1267 arr->append(notes); 1268 notes += _node_notes_block_size; 1269 num_notes -= _node_notes_block_size; 1270 } 1271 assert(num_notes == 0, "exact multiple, please"); 1272 } 1273 1274 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1275 if (source == nullptr || dest == nullptr) return false; 1276 1277 if (dest->is_Con()) 1278 return false; // Do not push debug info onto constants. 1279 1280 #ifdef ASSERT 1281 // Leave a bread crumb trail pointing to the original node: 1282 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) { 1283 dest->set_debug_orig(source); 1284 } 1285 #endif 1286 1287 if (node_note_array() == nullptr) 1288 return false; // Not collecting any notes now. 1289 1290 // This is a copy onto a pre-existing node, which may already have notes. 1291 // If both nodes have notes, do not overwrite any pre-existing notes. 1292 Node_Notes* source_notes = node_notes_at(source->_idx); 1293 if (source_notes == nullptr || source_notes->is_clear()) return false; 1294 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1295 if (dest_notes == nullptr || dest_notes->is_clear()) { 1296 return set_node_notes_at(dest->_idx, source_notes); 1297 } 1298 1299 Node_Notes merged_notes = (*source_notes); 1300 // The order of operations here ensures that dest notes will win... 1301 merged_notes.update_from(dest_notes); 1302 return set_node_notes_at(dest->_idx, &merged_notes); 1303 } 1304 1305 1306 //--------------------------allow_range_check_smearing------------------------- 1307 // Gating condition for coalescing similar range checks. 1308 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1309 // single covering check that is at least as strong as any of them. 1310 // If the optimization succeeds, the simplified (strengthened) range check 1311 // will always succeed. If it fails, we will deopt, and then give up 1312 // on the optimization. 1313 bool Compile::allow_range_check_smearing() const { 1314 // If this method has already thrown a range-check, 1315 // assume it was because we already tried range smearing 1316 // and it failed. 1317 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1318 return !already_trapped; 1319 } 1320 1321 1322 //------------------------------flatten_alias_type----------------------------- 1323 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1324 assert(do_aliasing(), "Aliasing should be enabled"); 1325 int offset = tj->offset(); 1326 TypePtr::PTR ptr = tj->ptr(); 1327 1328 // Known instance (scalarizable allocation) alias only with itself. 1329 bool is_known_inst = tj->isa_oopptr() != nullptr && 1330 tj->is_oopptr()->is_known_instance(); 1331 1332 // Process weird unsafe references. 1333 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1334 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops"); 1335 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1336 tj = TypeOopPtr::BOTTOM; 1337 ptr = tj->ptr(); 1338 offset = tj->offset(); 1339 } 1340 1341 // Array pointers need some flattening 1342 const TypeAryPtr* ta = tj->isa_aryptr(); 1343 if (ta && ta->is_stable()) { 1344 // Erase stability property for alias analysis. 1345 tj = ta = ta->cast_to_stable(false); 1346 } 1347 if( ta && is_known_inst ) { 1348 if ( offset != Type::OffsetBot && 1349 offset > arrayOopDesc::length_offset_in_bytes() ) { 1350 offset = Type::OffsetBot; // Flatten constant access into array body only 1351 tj = ta = ta-> 1352 remove_speculative()-> 1353 cast_to_ptr_type(ptr)-> 1354 with_offset(offset); 1355 } 1356 } else if (ta) { 1357 // For arrays indexed by constant indices, we flatten the alias 1358 // space to include all of the array body. Only the header, klass 1359 // and array length can be accessed un-aliased. 1360 if( offset != Type::OffsetBot ) { 1361 if( ta->const_oop() ) { // MethodData* or Method* 1362 offset = Type::OffsetBot; // Flatten constant access into array body 1363 tj = ta = ta-> 1364 remove_speculative()-> 1365 cast_to_ptr_type(ptr)-> 1366 cast_to_exactness(false)-> 1367 with_offset(offset); 1368 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1369 // range is OK as-is. 1370 tj = ta = TypeAryPtr::RANGE; 1371 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1372 tj = TypeInstPtr::KLASS; // all klass loads look alike 1373 ta = TypeAryPtr::RANGE; // generic ignored junk 1374 ptr = TypePtr::BotPTR; 1375 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1376 tj = TypeInstPtr::MARK; 1377 ta = TypeAryPtr::RANGE; // generic ignored junk 1378 ptr = TypePtr::BotPTR; 1379 } else { // Random constant offset into array body 1380 offset = Type::OffsetBot; // Flatten constant access into array body 1381 tj = ta = ta-> 1382 remove_speculative()-> 1383 cast_to_ptr_type(ptr)-> 1384 cast_to_exactness(false)-> 1385 with_offset(offset); 1386 } 1387 } 1388 // Arrays of fixed size alias with arrays of unknown size. 1389 if (ta->size() != TypeInt::POS) { 1390 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1391 tj = ta = ta-> 1392 remove_speculative()-> 1393 cast_to_ptr_type(ptr)-> 1394 with_ary(tary)-> 1395 cast_to_exactness(false); 1396 } 1397 // Arrays of known objects become arrays of unknown objects. 1398 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1399 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1400 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1401 } 1402 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1403 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1404 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1405 } 1406 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1407 // cannot be distinguished by bytecode alone. 1408 if (ta->elem() == TypeInt::BOOL) { 1409 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1410 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1411 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1412 } 1413 // During the 2nd round of IterGVN, NotNull castings are removed. 1414 // Make sure the Bottom and NotNull variants alias the same. 1415 // Also, make sure exact and non-exact variants alias the same. 1416 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) { 1417 tj = ta = ta-> 1418 remove_speculative()-> 1419 cast_to_ptr_type(TypePtr::BotPTR)-> 1420 cast_to_exactness(false)-> 1421 with_offset(offset); 1422 } 1423 } 1424 1425 // Oop pointers need some flattening 1426 const TypeInstPtr *to = tj->isa_instptr(); 1427 if (to && to != TypeOopPtr::BOTTOM) { 1428 ciInstanceKlass* ik = to->instance_klass(); 1429 if( ptr == TypePtr::Constant ) { 1430 if (ik != ciEnv::current()->Class_klass() || 1431 offset < ik->layout_helper_size_in_bytes()) { 1432 // No constant oop pointers (such as Strings); they alias with 1433 // unknown strings. 1434 assert(!is_known_inst, "not scalarizable allocation"); 1435 tj = to = to-> 1436 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1437 remove_speculative()-> 1438 cast_to_ptr_type(TypePtr::BotPTR)-> 1439 cast_to_exactness(false); 1440 } 1441 } else if( is_known_inst ) { 1442 tj = to; // Keep NotNull and klass_is_exact for instance type 1443 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1444 // During the 2nd round of IterGVN, NotNull castings are removed. 1445 // Make sure the Bottom and NotNull variants alias the same. 1446 // Also, make sure exact and non-exact variants alias the same. 1447 tj = to = to-> 1448 remove_speculative()-> 1449 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1450 cast_to_ptr_type(TypePtr::BotPTR)-> 1451 cast_to_exactness(false); 1452 } 1453 if (to->speculative() != nullptr) { 1454 tj = to = to->remove_speculative(); 1455 } 1456 // Canonicalize the holder of this field 1457 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1458 // First handle header references such as a LoadKlassNode, even if the 1459 // object's klass is unloaded at compile time (4965979). 1460 if (!is_known_inst) { // Do it only for non-instance types 1461 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset); 1462 } 1463 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) { 1464 // Static fields are in the space above the normal instance 1465 // fields in the java.lang.Class instance. 1466 if (ik != ciEnv::current()->Class_klass()) { 1467 to = nullptr; 1468 tj = TypeOopPtr::BOTTOM; 1469 offset = tj->offset(); 1470 } 1471 } else { 1472 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset); 1473 assert(offset < canonical_holder->layout_helper_size_in_bytes(), ""); 1474 if (!ik->equals(canonical_holder) || tj->offset() != offset) { 1475 if( is_known_inst ) { 1476 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id()); 1477 } else { 1478 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset); 1479 } 1480 } 1481 } 1482 } 1483 1484 // Klass pointers to object array klasses need some flattening 1485 const TypeKlassPtr *tk = tj->isa_klassptr(); 1486 if( tk ) { 1487 // If we are referencing a field within a Klass, we need 1488 // to assume the worst case of an Object. Both exact and 1489 // inexact types must flatten to the same alias class so 1490 // use NotNull as the PTR. 1491 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1492 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, 1493 env()->Object_klass(), 1494 offset); 1495 } 1496 1497 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) { 1498 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass()); 1499 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs 1500 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset); 1501 } else { 1502 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset); 1503 } 1504 } 1505 1506 // Check for precise loads from the primary supertype array and force them 1507 // to the supertype cache alias index. Check for generic array loads from 1508 // the primary supertype array and also force them to the supertype cache 1509 // alias index. Since the same load can reach both, we need to merge 1510 // these 2 disparate memories into the same alias class. Since the 1511 // primary supertype array is read-only, there's no chance of confusion 1512 // where we bypass an array load and an array store. 1513 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1514 if (offset == Type::OffsetBot || 1515 (offset >= primary_supers_offset && 1516 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1517 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1518 offset = in_bytes(Klass::secondary_super_cache_offset()); 1519 tj = tk = tk->with_offset(offset); 1520 } 1521 } 1522 1523 // Flatten all Raw pointers together. 1524 if (tj->base() == Type::RawPtr) 1525 tj = TypeRawPtr::BOTTOM; 1526 1527 if (tj->base() == Type::AnyPtr) 1528 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1529 1530 offset = tj->offset(); 1531 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1532 1533 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1534 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1535 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1536 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1537 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1538 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1539 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1540 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1541 assert( tj->ptr() != TypePtr::TopPTR && 1542 tj->ptr() != TypePtr::AnyNull && 1543 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1544 // assert( tj->ptr() != TypePtr::Constant || 1545 // tj->base() == Type::RawPtr || 1546 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1547 1548 return tj; 1549 } 1550 1551 void Compile::AliasType::Init(int i, const TypePtr* at) { 1552 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1553 _index = i; 1554 _adr_type = at; 1555 _field = nullptr; 1556 _element = nullptr; 1557 _is_rewritable = true; // default 1558 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr; 1559 if (atoop != nullptr && atoop->is_known_instance()) { 1560 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1561 _general_index = Compile::current()->get_alias_index(gt); 1562 } else { 1563 _general_index = 0; 1564 } 1565 } 1566 1567 BasicType Compile::AliasType::basic_type() const { 1568 if (element() != nullptr) { 1569 const Type* element = adr_type()->is_aryptr()->elem(); 1570 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1571 } if (field() != nullptr) { 1572 return field()->layout_type(); 1573 } else { 1574 return T_ILLEGAL; // unknown 1575 } 1576 } 1577 1578 //---------------------------------print_on------------------------------------ 1579 #ifndef PRODUCT 1580 void Compile::AliasType::print_on(outputStream* st) { 1581 if (index() < 10) 1582 st->print("@ <%d> ", index()); 1583 else st->print("@ <%d>", index()); 1584 st->print(is_rewritable() ? " " : " RO"); 1585 int offset = adr_type()->offset(); 1586 if (offset == Type::OffsetBot) 1587 st->print(" +any"); 1588 else st->print(" +%-3d", offset); 1589 st->print(" in "); 1590 adr_type()->dump_on(st); 1591 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1592 if (field() != nullptr && tjp) { 1593 if (tjp->is_instptr()->instance_klass() != field()->holder() || 1594 tjp->offset() != field()->offset_in_bytes()) { 1595 st->print(" != "); 1596 field()->print(); 1597 st->print(" ***"); 1598 } 1599 } 1600 } 1601 1602 void print_alias_types() { 1603 Compile* C = Compile::current(); 1604 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1605 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1606 C->alias_type(idx)->print_on(tty); 1607 tty->cr(); 1608 } 1609 } 1610 #endif 1611 1612 1613 //----------------------------probe_alias_cache-------------------------------- 1614 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1615 intptr_t key = (intptr_t) adr_type; 1616 key ^= key >> logAliasCacheSize; 1617 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1618 } 1619 1620 1621 //-----------------------------grow_alias_types-------------------------------- 1622 void Compile::grow_alias_types() { 1623 const int old_ats = _max_alias_types; // how many before? 1624 const int new_ats = old_ats; // how many more? 1625 const int grow_ats = old_ats+new_ats; // how many now? 1626 _max_alias_types = grow_ats; 1627 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1628 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1629 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1630 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1631 } 1632 1633 1634 //--------------------------------find_alias_type------------------------------ 1635 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1636 if (!do_aliasing()) { 1637 return alias_type(AliasIdxBot); 1638 } 1639 1640 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1641 if (ace->_adr_type == adr_type) { 1642 return alias_type(ace->_index); 1643 } 1644 1645 // Handle special cases. 1646 if (adr_type == nullptr) return alias_type(AliasIdxTop); 1647 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1648 1649 // Do it the slow way. 1650 const TypePtr* flat = flatten_alias_type(adr_type); 1651 1652 #ifdef ASSERT 1653 { 1654 ResourceMark rm; 1655 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1656 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1657 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1658 Type::str(adr_type)); 1659 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1660 const TypeOopPtr* foop = flat->is_oopptr(); 1661 // Scalarizable allocations have exact klass always. 1662 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1663 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1664 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1665 Type::str(foop), Type::str(xoop)); 1666 } 1667 } 1668 #endif 1669 1670 int idx = AliasIdxTop; 1671 for (int i = 0; i < num_alias_types(); i++) { 1672 if (alias_type(i)->adr_type() == flat) { 1673 idx = i; 1674 break; 1675 } 1676 } 1677 1678 if (idx == AliasIdxTop) { 1679 if (no_create) return nullptr; 1680 // Grow the array if necessary. 1681 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1682 // Add a new alias type. 1683 idx = _num_alias_types++; 1684 _alias_types[idx]->Init(idx, flat); 1685 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1686 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1687 if (flat->isa_instptr()) { 1688 if (flat->offset() == java_lang_Class::klass_offset() 1689 && flat->is_instptr()->instance_klass() == env()->Class_klass()) 1690 alias_type(idx)->set_rewritable(false); 1691 } 1692 if (flat->isa_aryptr()) { 1693 #ifdef ASSERT 1694 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1695 // (T_BYTE has the weakest alignment and size restrictions...) 1696 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1697 #endif 1698 if (flat->offset() == TypePtr::OffsetBot) { 1699 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1700 } 1701 } 1702 if (flat->isa_klassptr()) { 1703 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1704 alias_type(idx)->set_rewritable(false); 1705 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1706 alias_type(idx)->set_rewritable(false); 1707 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1708 alias_type(idx)->set_rewritable(false); 1709 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1710 alias_type(idx)->set_rewritable(false); 1711 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1712 alias_type(idx)->set_rewritable(false); 1713 } 1714 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1715 // but the base pointer type is not distinctive enough to identify 1716 // references into JavaThread.) 1717 1718 // Check for final fields. 1719 const TypeInstPtr* tinst = flat->isa_instptr(); 1720 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1721 ciField* field; 1722 if (tinst->const_oop() != nullptr && 1723 tinst->instance_klass() == ciEnv::current()->Class_klass() && 1724 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) { 1725 // static field 1726 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1727 field = k->get_field_by_offset(tinst->offset(), true); 1728 } else { 1729 ciInstanceKlass *k = tinst->instance_klass(); 1730 field = k->get_field_by_offset(tinst->offset(), false); 1731 } 1732 assert(field == nullptr || 1733 original_field == nullptr || 1734 (field->holder() == original_field->holder() && 1735 field->offset_in_bytes() == original_field->offset_in_bytes() && 1736 field->is_static() == original_field->is_static()), "wrong field?"); 1737 // Set field() and is_rewritable() attributes. 1738 if (field != nullptr) alias_type(idx)->set_field(field); 1739 } 1740 } 1741 1742 // Fill the cache for next time. 1743 ace->_adr_type = adr_type; 1744 ace->_index = idx; 1745 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1746 1747 // Might as well try to fill the cache for the flattened version, too. 1748 AliasCacheEntry* face = probe_alias_cache(flat); 1749 if (face->_adr_type == nullptr) { 1750 face->_adr_type = flat; 1751 face->_index = idx; 1752 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1753 } 1754 1755 return alias_type(idx); 1756 } 1757 1758 1759 Compile::AliasType* Compile::alias_type(ciField* field) { 1760 const TypeOopPtr* t; 1761 if (field->is_static()) 1762 t = TypeInstPtr::make(field->holder()->java_mirror()); 1763 else 1764 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1765 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1766 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1767 return atp; 1768 } 1769 1770 1771 //------------------------------have_alias_type-------------------------------- 1772 bool Compile::have_alias_type(const TypePtr* adr_type) { 1773 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1774 if (ace->_adr_type == adr_type) { 1775 return true; 1776 } 1777 1778 // Handle special cases. 1779 if (adr_type == nullptr) return true; 1780 if (adr_type == TypePtr::BOTTOM) return true; 1781 1782 return find_alias_type(adr_type, true, nullptr) != nullptr; 1783 } 1784 1785 //-----------------------------must_alias-------------------------------------- 1786 // True if all values of the given address type are in the given alias category. 1787 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1788 if (alias_idx == AliasIdxBot) return true; // the universal category 1789 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP 1790 if (alias_idx == AliasIdxTop) return false; // the empty category 1791 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1792 1793 // the only remaining possible overlap is identity 1794 int adr_idx = get_alias_index(adr_type); 1795 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1796 assert(adr_idx == alias_idx || 1797 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1798 && adr_type != TypeOopPtr::BOTTOM), 1799 "should not be testing for overlap with an unsafe pointer"); 1800 return adr_idx == alias_idx; 1801 } 1802 1803 //------------------------------can_alias-------------------------------------- 1804 // True if any values of the given address type are in the given alias category. 1805 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1806 if (alias_idx == AliasIdxTop) return false; // the empty category 1807 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP 1808 // Known instance doesn't alias with bottom memory 1809 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1810 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1811 1812 // the only remaining possible overlap is identity 1813 int adr_idx = get_alias_index(adr_type); 1814 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1815 return adr_idx == alias_idx; 1816 } 1817 1818 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their 1819 // uncommon traps if no Runtime Predicates were created from the Parse Predicates. 1820 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) { 1821 if (parse_predicate_count() == 0) { 1822 return; 1823 } 1824 for (int i = 0; i < parse_predicate_count(); i++) { 1825 ParsePredicateNode* parse_predicate = _parse_predicates.at(i); 1826 parse_predicate->mark_useless(); 1827 igvn._worklist.push(parse_predicate); 1828 } 1829 _parse_predicates.clear(); 1830 } 1831 1832 void Compile::record_for_post_loop_opts_igvn(Node* n) { 1833 if (!n->for_post_loop_opts_igvn()) { 1834 assert(!_for_post_loop_igvn.contains(n), "duplicate"); 1835 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1836 _for_post_loop_igvn.append(n); 1837 } 1838 } 1839 1840 void Compile::remove_from_post_loop_opts_igvn(Node* n) { 1841 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1842 _for_post_loop_igvn.remove(n); 1843 } 1844 1845 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) { 1846 // Verify that all previous optimizations produced a valid graph 1847 // at least to this point, even if no loop optimizations were done. 1848 PhaseIdealLoop::verify(igvn); 1849 1850 C->set_post_loop_opts_phase(); // no more loop opts allowed 1851 1852 assert(!C->major_progress(), "not cleared"); 1853 1854 if (_for_post_loop_igvn.length() > 0) { 1855 while (_for_post_loop_igvn.length() > 0) { 1856 Node* n = _for_post_loop_igvn.pop(); 1857 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1858 igvn._worklist.push(n); 1859 } 1860 igvn.optimize(); 1861 if (failing()) return; 1862 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed"); 1863 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now"); 1864 1865 // Sometimes IGVN sets major progress (e.g., when processing loop nodes). 1866 if (C->major_progress()) { 1867 C->clear_major_progress(); // ensure that major progress is now clear 1868 } 1869 } 1870 } 1871 1872 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) { 1873 if (OptimizeUnstableIf) { 1874 _unstable_if_traps.append(trap); 1875 } 1876 } 1877 1878 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) { 1879 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) { 1880 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1881 Node* n = trap->uncommon_trap(); 1882 if (!useful.member(n)) { 1883 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 1884 } 1885 } 1886 } 1887 1888 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead 1889 // or fold-compares case. Return true if succeed or not found. 1890 // 1891 // In rare cases, the found trap has been processed. It is too late to delete it. Return 1892 // false and ask fold-compares to yield. 1893 // 1894 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused 1895 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path 1896 // when deoptimization does happen. 1897 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) { 1898 for (int i = 0; i < _unstable_if_traps.length(); ++i) { 1899 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1900 if (trap->uncommon_trap() == unc) { 1901 if (yield && trap->modified()) { 1902 return false; 1903 } 1904 _unstable_if_traps.delete_at(i); 1905 break; 1906 } 1907 } 1908 return true; 1909 } 1910 1911 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path. 1912 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering. 1913 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) { 1914 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) { 1915 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1916 CallStaticJavaNode* unc = trap->uncommon_trap(); 1917 int next_bci = trap->next_bci(); 1918 bool modified = trap->modified(); 1919 1920 if (next_bci != -1 && !modified) { 1921 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!"); 1922 JVMState* jvms = unc->jvms(); 1923 ciMethod* method = jvms->method(); 1924 ciBytecodeStream iter(method); 1925 1926 iter.force_bci(jvms->bci()); 1927 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if"); 1928 Bytecodes::Code c = iter.cur_bc(); 1929 Node* lhs = nullptr; 1930 Node* rhs = nullptr; 1931 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) { 1932 lhs = unc->peek_operand(0); 1933 rhs = unc->peek_operand(1); 1934 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) { 1935 lhs = unc->peek_operand(0); 1936 } 1937 1938 ResourceMark rm; 1939 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci); 1940 assert(live_locals.is_valid(), "broken liveness info"); 1941 int len = (int)live_locals.size(); 1942 1943 for (int i = 0; i < len; i++) { 1944 Node* local = unc->local(jvms, i); 1945 // kill local using the liveness of next_bci. 1946 // give up when the local looks like an operand to secure reexecution. 1947 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) { 1948 uint idx = jvms->locoff() + i; 1949 #ifdef ASSERT 1950 if (PrintOpto && Verbose) { 1951 tty->print("[unstable_if] kill local#%d: ", idx); 1952 local->dump(); 1953 tty->cr(); 1954 } 1955 #endif 1956 igvn.replace_input_of(unc, idx, top()); 1957 modified = true; 1958 } 1959 } 1960 } 1961 1962 // keep the mondified trap for late query 1963 if (modified) { 1964 trap->set_modified(); 1965 } else { 1966 _unstable_if_traps.delete_at(i); 1967 } 1968 } 1969 igvn.optimize(); 1970 } 1971 1972 // StringOpts and late inlining of string methods 1973 void Compile::inline_string_calls(bool parse_time) { 1974 { 1975 // remove useless nodes to make the usage analysis simpler 1976 ResourceMark rm; 1977 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 1978 } 1979 1980 { 1981 ResourceMark rm; 1982 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1983 PhaseStringOpts pso(initial_gvn()); 1984 print_method(PHASE_AFTER_STRINGOPTS, 3); 1985 } 1986 1987 // now inline anything that we skipped the first time around 1988 if (!parse_time) { 1989 _late_inlines_pos = _late_inlines.length(); 1990 } 1991 1992 while (_string_late_inlines.length() > 0) { 1993 CallGenerator* cg = _string_late_inlines.pop(); 1994 cg->do_late_inline(); 1995 if (failing()) return; 1996 } 1997 _string_late_inlines.trunc_to(0); 1998 } 1999 2000 // Late inlining of boxing methods 2001 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2002 if (_boxing_late_inlines.length() > 0) { 2003 assert(has_boxed_value(), "inconsistent"); 2004 2005 set_inlining_incrementally(true); 2006 2007 igvn_worklist()->ensure_empty(); // should be done with igvn 2008 2009 _late_inlines_pos = _late_inlines.length(); 2010 2011 while (_boxing_late_inlines.length() > 0) { 2012 CallGenerator* cg = _boxing_late_inlines.pop(); 2013 cg->do_late_inline(); 2014 if (failing()) return; 2015 } 2016 _boxing_late_inlines.trunc_to(0); 2017 2018 inline_incrementally_cleanup(igvn); 2019 2020 set_inlining_incrementally(false); 2021 } 2022 } 2023 2024 bool Compile::inline_incrementally_one() { 2025 assert(IncrementalInline, "incremental inlining should be on"); 2026 2027 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2028 2029 set_inlining_progress(false); 2030 set_do_cleanup(false); 2031 2032 for (int i = 0; i < _late_inlines.length(); i++) { 2033 _late_inlines_pos = i+1; 2034 CallGenerator* cg = _late_inlines.at(i); 2035 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline(); 2036 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call 2037 cg->do_late_inline(); 2038 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed"); 2039 if (failing()) { 2040 return false; 2041 } else if (inlining_progress()) { 2042 _late_inlines_pos = i+1; // restore the position in case new elements were inserted 2043 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node()); 2044 break; // process one call site at a time 2045 } 2046 } else { 2047 // Ignore late inline direct calls when inlining is not allowed. 2048 // They are left in the late inline list when node budget is exhausted until the list is fully drained. 2049 } 2050 } 2051 // Remove processed elements. 2052 _late_inlines.remove_till(_late_inlines_pos); 2053 _late_inlines_pos = 0; 2054 2055 assert(inlining_progress() || _late_inlines.length() == 0, "no progress"); 2056 2057 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2058 2059 set_inlining_progress(false); 2060 set_do_cleanup(false); 2061 2062 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption; 2063 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup; 2064 } 2065 2066 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2067 { 2068 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2069 ResourceMark rm; 2070 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2071 } 2072 { 2073 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2074 igvn.reset_from_gvn(initial_gvn()); 2075 igvn.optimize(); 2076 if (failing()) return; 2077 } 2078 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3); 2079 } 2080 2081 // Perform incremental inlining until bound on number of live nodes is reached 2082 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2083 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2084 2085 set_inlining_incrementally(true); 2086 uint low_live_nodes = 0; 2087 2088 while (_late_inlines.length() > 0) { 2089 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2090 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2091 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2092 // PhaseIdealLoop is expensive so we only try it once we are 2093 // out of live nodes and we only try it again if the previous 2094 // helped got the number of nodes down significantly 2095 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2096 if (failing()) return; 2097 low_live_nodes = live_nodes(); 2098 _major_progress = true; 2099 } 2100 2101 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2102 bool do_print_inlining = print_inlining() || print_intrinsics(); 2103 if (do_print_inlining || log() != nullptr) { 2104 // Print inlining message for candidates that we couldn't inline for lack of space. 2105 for (int i = 0; i < _late_inlines.length(); i++) { 2106 CallGenerator* cg = _late_inlines.at(i); 2107 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 2108 if (do_print_inlining) { 2109 cg->print_inlining_late(InliningResult::FAILURE, msg); 2110 } 2111 log_late_inline_failure(cg, msg); 2112 } 2113 } 2114 break; // finish 2115 } 2116 } 2117 2118 igvn_worklist()->ensure_empty(); // should be done with igvn 2119 2120 while (inline_incrementally_one()) { 2121 assert(!failing(), "inconsistent"); 2122 } 2123 if (failing()) return; 2124 2125 inline_incrementally_cleanup(igvn); 2126 2127 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 2128 2129 if (failing()) return; 2130 2131 if (_late_inlines.length() == 0) { 2132 break; // no more progress 2133 } 2134 } 2135 2136 igvn_worklist()->ensure_empty(); // should be done with igvn 2137 2138 if (_string_late_inlines.length() > 0) { 2139 assert(has_stringbuilder(), "inconsistent"); 2140 2141 inline_string_calls(false); 2142 2143 if (failing()) return; 2144 2145 inline_incrementally_cleanup(igvn); 2146 } 2147 2148 set_inlining_incrementally(false); 2149 } 2150 2151 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) { 2152 // "inlining_incrementally() == false" is used to signal that no inlining is allowed 2153 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details). 2154 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr" 2155 // as if "inlining_incrementally() == true" were set. 2156 assert(inlining_incrementally() == false, "not allowed"); 2157 assert(_modified_nodes == nullptr, "not allowed"); 2158 assert(_late_inlines.length() > 0, "sanity"); 2159 2160 while (_late_inlines.length() > 0) { 2161 igvn_worklist()->ensure_empty(); // should be done with igvn 2162 2163 while (inline_incrementally_one()) { 2164 assert(!failing(), "inconsistent"); 2165 } 2166 if (failing()) return; 2167 2168 inline_incrementally_cleanup(igvn); 2169 } 2170 } 2171 2172 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2173 if (_loop_opts_cnt > 0) { 2174 while (major_progress() && (_loop_opts_cnt > 0)) { 2175 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2176 PhaseIdealLoop::optimize(igvn, mode); 2177 _loop_opts_cnt--; 2178 if (failing()) return false; 2179 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2180 } 2181 } 2182 return true; 2183 } 2184 2185 // Remove edges from "root" to each SafePoint at a backward branch. 2186 // They were inserted during parsing (see add_safepoint()) to make 2187 // infinite loops without calls or exceptions visible to root, i.e., 2188 // useful. 2189 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2190 Node *r = root(); 2191 if (r != nullptr) { 2192 for (uint i = r->req(); i < r->len(); ++i) { 2193 Node *n = r->in(i); 2194 if (n != nullptr && n->is_SafePoint()) { 2195 r->rm_prec(i); 2196 if (n->outcnt() == 0) { 2197 igvn.remove_dead_node(n); 2198 } 2199 --i; 2200 } 2201 } 2202 // Parsing may have added top inputs to the root node (Path 2203 // leading to the Halt node proven dead). Make sure we get a 2204 // chance to clean them up. 2205 igvn._worklist.push(r); 2206 igvn.optimize(); 2207 } 2208 } 2209 2210 //------------------------------Optimize--------------------------------------- 2211 // Given a graph, optimize it. 2212 void Compile::Optimize() { 2213 TracePhase tp("optimizer", &timers[_t_optimizer]); 2214 2215 #ifndef PRODUCT 2216 if (env()->break_at_compile()) { 2217 BREAKPOINT; 2218 } 2219 2220 #endif 2221 2222 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2223 #ifdef ASSERT 2224 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2225 #endif 2226 2227 ResourceMark rm; 2228 2229 print_inlining_reinit(); 2230 2231 NOT_PRODUCT( verify_graph_edges(); ) 2232 2233 print_method(PHASE_AFTER_PARSING, 1); 2234 2235 { 2236 // Iterative Global Value Numbering, including ideal transforms 2237 // Initialize IterGVN with types and values from parse-time GVN 2238 PhaseIterGVN igvn(initial_gvn()); 2239 #ifdef ASSERT 2240 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2241 #endif 2242 { 2243 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2244 igvn.optimize(); 2245 } 2246 2247 if (failing()) return; 2248 2249 print_method(PHASE_ITER_GVN1, 2); 2250 2251 process_for_unstable_if_traps(igvn); 2252 2253 if (failing()) return; 2254 2255 inline_incrementally(igvn); 2256 2257 print_method(PHASE_INCREMENTAL_INLINE, 2); 2258 2259 if (failing()) return; 2260 2261 if (eliminate_boxing()) { 2262 // Inline valueOf() methods now. 2263 inline_boxing_calls(igvn); 2264 2265 if (failing()) return; 2266 2267 if (AlwaysIncrementalInline || StressIncrementalInlining) { 2268 inline_incrementally(igvn); 2269 } 2270 2271 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2272 2273 if (failing()) return; 2274 } 2275 2276 // Remove the speculative part of types and clean up the graph from 2277 // the extra CastPP nodes whose only purpose is to carry them. Do 2278 // that early so that optimizations are not disrupted by the extra 2279 // CastPP nodes. 2280 remove_speculative_types(igvn); 2281 2282 if (failing()) return; 2283 2284 // No more new expensive nodes will be added to the list from here 2285 // so keep only the actual candidates for optimizations. 2286 cleanup_expensive_nodes(igvn); 2287 2288 if (failing()) return; 2289 2290 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2291 if (EnableVectorSupport && has_vbox_nodes()) { 2292 TracePhase tp("", &timers[_t_vector]); 2293 PhaseVector pv(igvn); 2294 pv.optimize_vector_boxes(); 2295 if (failing()) return; 2296 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2297 } 2298 assert(!has_vbox_nodes(), "sanity"); 2299 2300 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2301 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2302 igvn_worklist()->ensure_empty(); // should be done with igvn 2303 { 2304 ResourceMark rm; 2305 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist()); 2306 } 2307 igvn.reset_from_gvn(initial_gvn()); 2308 igvn.optimize(); 2309 if (failing()) return; 2310 } 2311 2312 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop 2313 // safepoints 2314 remove_root_to_sfpts_edges(igvn); 2315 2316 if (failing()) return; 2317 2318 // Perform escape analysis 2319 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) { 2320 if (has_loops()) { 2321 // Cleanup graph (remove dead nodes). 2322 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2323 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2324 if (failing()) return; 2325 } 2326 bool progress; 2327 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2328 do { 2329 ConnectionGraph::do_analysis(this, &igvn); 2330 2331 if (failing()) return; 2332 2333 int mcount = macro_count(); // Record number of allocations and locks before IGVN 2334 2335 // Optimize out fields loads from scalar replaceable allocations. 2336 igvn.optimize(); 2337 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2338 2339 if (failing()) return; 2340 2341 if (congraph() != nullptr && macro_count() > 0) { 2342 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2343 PhaseMacroExpand mexp(igvn); 2344 mexp.eliminate_macro_nodes(); 2345 if (failing()) return; 2346 2347 igvn.set_delay_transform(false); 2348 igvn.optimize(); 2349 if (failing()) return; 2350 2351 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2352 } 2353 2354 ConnectionGraph::verify_ram_nodes(this, root()); 2355 if (failing()) return; 2356 2357 progress = do_iterative_escape_analysis() && 2358 (macro_count() < mcount) && 2359 ConnectionGraph::has_candidates(this); 2360 // Try again if candidates exist and made progress 2361 // by removing some allocations and/or locks. 2362 } while (progress); 2363 } 2364 2365 // Loop transforms on the ideal graph. Range Check Elimination, 2366 // peeling, unrolling, etc. 2367 2368 // Set loop opts counter 2369 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2370 { 2371 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2372 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2373 _loop_opts_cnt--; 2374 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2375 if (failing()) return; 2376 } 2377 // Loop opts pass if partial peeling occurred in previous pass 2378 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2379 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2380 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2381 _loop_opts_cnt--; 2382 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2383 if (failing()) return; 2384 } 2385 // Loop opts pass for loop-unrolling before CCP 2386 if(major_progress() && (_loop_opts_cnt > 0)) { 2387 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2388 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2389 _loop_opts_cnt--; 2390 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2391 } 2392 if (!failing()) { 2393 // Verify that last round of loop opts produced a valid graph 2394 PhaseIdealLoop::verify(igvn); 2395 } 2396 } 2397 if (failing()) return; 2398 2399 // Conditional Constant Propagation; 2400 PhaseCCP ccp( &igvn ); 2401 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2402 { 2403 TracePhase tp("ccp", &timers[_t_ccp]); 2404 ccp.do_transform(); 2405 } 2406 print_method(PHASE_CCP1, 2); 2407 2408 assert( true, "Break here to ccp.dump_old2new_map()"); 2409 2410 // Iterative Global Value Numbering, including ideal transforms 2411 { 2412 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2413 igvn.reset_from_igvn(&ccp); 2414 igvn.optimize(); 2415 } 2416 print_method(PHASE_ITER_GVN2, 2); 2417 2418 if (failing()) return; 2419 2420 // Loop transforms on the ideal graph. Range Check Elimination, 2421 // peeling, unrolling, etc. 2422 if (!optimize_loops(igvn, LoopOptsDefault)) { 2423 return; 2424 } 2425 2426 if (failing()) return; 2427 2428 C->clear_major_progress(); // ensure that major progress is now clear 2429 2430 process_for_post_loop_opts_igvn(igvn); 2431 2432 if (failing()) return; 2433 2434 #ifdef ASSERT 2435 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2436 #endif 2437 2438 { 2439 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2440 PhaseMacroExpand mex(igvn); 2441 if (mex.expand_macro_nodes()) { 2442 assert(failing(), "must bail out w/ explicit message"); 2443 return; 2444 } 2445 print_method(PHASE_MACRO_EXPANSION, 2); 2446 } 2447 2448 { 2449 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2450 if (bs->expand_barriers(this, igvn)) { 2451 assert(failing(), "must bail out w/ explicit message"); 2452 return; 2453 } 2454 print_method(PHASE_BARRIER_EXPANSION, 2); 2455 } 2456 2457 if (C->max_vector_size() > 0) { 2458 C->optimize_logic_cones(igvn); 2459 igvn.optimize(); 2460 if (failing()) return; 2461 } 2462 2463 DEBUG_ONLY( _modified_nodes = nullptr; ) 2464 2465 assert(igvn._worklist.size() == 0, "not empty"); 2466 2467 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty"); 2468 2469 if (_late_inlines.length() > 0) { 2470 // More opportunities to optimize virtual and MH calls. 2471 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option. 2472 process_late_inline_calls_no_inline(igvn); 2473 if (failing()) return; 2474 } 2475 } // (End scope of igvn; run destructor if necessary for asserts.) 2476 2477 check_no_dead_use(); 2478 2479 process_print_inlining(); 2480 2481 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have 2482 // to remove hashes to unlock nodes for modifications. 2483 C->node_hash()->clear(); 2484 2485 // A method with only infinite loops has no edges entering loops from root 2486 { 2487 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2488 if (final_graph_reshaping()) { 2489 assert(failing(), "must bail out w/ explicit message"); 2490 return; 2491 } 2492 } 2493 2494 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2495 DEBUG_ONLY(set_phase_optimize_finished();) 2496 } 2497 2498 #ifdef ASSERT 2499 void Compile::check_no_dead_use() const { 2500 ResourceMark rm; 2501 Unique_Node_List wq; 2502 wq.push(root()); 2503 for (uint i = 0; i < wq.size(); ++i) { 2504 Node* n = wq.at(i); 2505 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2506 Node* u = n->fast_out(j); 2507 if (u->outcnt() == 0 && !u->is_Con()) { 2508 u->dump(); 2509 fatal("no reachable node should have no use"); 2510 } 2511 wq.push(u); 2512 } 2513 } 2514 } 2515 #endif 2516 2517 void Compile::inline_vector_reboxing_calls() { 2518 if (C->_vector_reboxing_late_inlines.length() > 0) { 2519 _late_inlines_pos = C->_late_inlines.length(); 2520 while (_vector_reboxing_late_inlines.length() > 0) { 2521 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 2522 cg->do_late_inline(); 2523 if (failing()) return; 2524 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node()); 2525 } 2526 _vector_reboxing_late_inlines.trunc_to(0); 2527 } 2528 } 2529 2530 bool Compile::has_vbox_nodes() { 2531 if (C->_vector_reboxing_late_inlines.length() > 0) { 2532 return true; 2533 } 2534 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 2535 Node * n = C->macro_node(macro_idx); 2536 assert(n->is_macro(), "only macro nodes expected here"); 2537 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 2538 return true; 2539 } 2540 } 2541 return false; 2542 } 2543 2544 //---------------------------- Bitwise operation packing optimization --------------------------- 2545 2546 static bool is_vector_unary_bitwise_op(Node* n) { 2547 return n->Opcode() == Op_XorV && 2548 VectorNode::is_vector_bitwise_not_pattern(n); 2549 } 2550 2551 static bool is_vector_binary_bitwise_op(Node* n) { 2552 switch (n->Opcode()) { 2553 case Op_AndV: 2554 case Op_OrV: 2555 return true; 2556 2557 case Op_XorV: 2558 return !is_vector_unary_bitwise_op(n); 2559 2560 default: 2561 return false; 2562 } 2563 } 2564 2565 static bool is_vector_ternary_bitwise_op(Node* n) { 2566 return n->Opcode() == Op_MacroLogicV; 2567 } 2568 2569 static bool is_vector_bitwise_op(Node* n) { 2570 return is_vector_unary_bitwise_op(n) || 2571 is_vector_binary_bitwise_op(n) || 2572 is_vector_ternary_bitwise_op(n); 2573 } 2574 2575 static bool is_vector_bitwise_cone_root(Node* n) { 2576 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) { 2577 return false; 2578 } 2579 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2580 if (is_vector_bitwise_op(n->fast_out(i))) { 2581 return false; 2582 } 2583 } 2584 return true; 2585 } 2586 2587 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) { 2588 uint cnt = 0; 2589 if (is_vector_bitwise_op(n)) { 2590 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req(); 2591 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2592 for (uint i = 1; i < inp_cnt; i++) { 2593 Node* in = n->in(i); 2594 bool skip = VectorNode::is_all_ones_vector(in); 2595 if (!skip && !inputs.member(in)) { 2596 inputs.push(in); 2597 cnt++; 2598 } 2599 } 2600 assert(cnt <= 1, "not unary"); 2601 } else { 2602 uint last_req = inp_cnt; 2603 if (is_vector_ternary_bitwise_op(n)) { 2604 last_req = inp_cnt - 1; // skip last input 2605 } 2606 for (uint i = 1; i < last_req; i++) { 2607 Node* def = n->in(i); 2608 if (!inputs.member(def)) { 2609 inputs.push(def); 2610 cnt++; 2611 } 2612 } 2613 } 2614 } else { // not a bitwise operations 2615 if (!inputs.member(n)) { 2616 inputs.push(n); 2617 cnt++; 2618 } 2619 } 2620 return cnt; 2621 } 2622 2623 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 2624 Unique_Node_List useful_nodes; 2625 C->identify_useful_nodes(useful_nodes); 2626 2627 for (uint i = 0; i < useful_nodes.size(); i++) { 2628 Node* n = useful_nodes.at(i); 2629 if (is_vector_bitwise_cone_root(n)) { 2630 list.push(n); 2631 } 2632 } 2633 } 2634 2635 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 2636 const TypeVect* vt, 2637 Unique_Node_List& partition, 2638 Unique_Node_List& inputs) { 2639 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 2640 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 2641 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 2642 2643 Node* in1 = inputs.at(0); 2644 Node* in2 = inputs.at(1); 2645 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 2646 2647 uint func = compute_truth_table(partition, inputs); 2648 2649 Node* pn = partition.at(partition.size() - 1); 2650 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2651 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt)); 2652 } 2653 2654 static uint extract_bit(uint func, uint pos) { 2655 return (func & (1 << pos)) >> pos; 2656 } 2657 2658 // 2659 // A macro logic node represents a truth table. It has 4 inputs, 2660 // First three inputs corresponds to 3 columns of a truth table 2661 // and fourth input captures the logic function. 2662 // 2663 // eg. fn = (in1 AND in2) OR in3; 2664 // 2665 // MacroNode(in1,in2,in3,fn) 2666 // 2667 // ----------------- 2668 // in1 in2 in3 fn 2669 // ----------------- 2670 // 0 0 0 0 2671 // 0 0 1 1 2672 // 0 1 0 0 2673 // 0 1 1 1 2674 // 1 0 0 0 2675 // 1 0 1 1 2676 // 1 1 0 1 2677 // 1 1 1 1 2678 // 2679 2680 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 2681 int res = 0; 2682 for (int i = 0; i < 8; i++) { 2683 int bit1 = extract_bit(in1, i); 2684 int bit2 = extract_bit(in2, i); 2685 int bit3 = extract_bit(in3, i); 2686 2687 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 2688 int func_bit = extract_bit(func, func_bit_pos); 2689 2690 res |= func_bit << i; 2691 } 2692 return res; 2693 } 2694 2695 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 2696 assert(n != nullptr, ""); 2697 assert(eval_map.contains(n), "absent"); 2698 return *(eval_map.get(n)); 2699 } 2700 2701 static void eval_operands(Node* n, 2702 uint& func1, uint& func2, uint& func3, 2703 ResourceHashtable<Node*,uint>& eval_map) { 2704 assert(is_vector_bitwise_op(n), ""); 2705 2706 if (is_vector_unary_bitwise_op(n)) { 2707 Node* opnd = n->in(1); 2708 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 2709 opnd = n->in(2); 2710 } 2711 func1 = eval_operand(opnd, eval_map); 2712 } else if (is_vector_binary_bitwise_op(n)) { 2713 func1 = eval_operand(n->in(1), eval_map); 2714 func2 = eval_operand(n->in(2), eval_map); 2715 } else { 2716 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 2717 func1 = eval_operand(n->in(1), eval_map); 2718 func2 = eval_operand(n->in(2), eval_map); 2719 func3 = eval_operand(n->in(3), eval_map); 2720 } 2721 } 2722 2723 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 2724 assert(inputs.size() <= 3, "sanity"); 2725 ResourceMark rm; 2726 uint res = 0; 2727 ResourceHashtable<Node*,uint> eval_map; 2728 2729 // Populate precomputed functions for inputs. 2730 // Each input corresponds to one column of 3 input truth-table. 2731 uint input_funcs[] = { 0xAA, // (_, _, c) -> c 2732 0xCC, // (_, b, _) -> b 2733 0xF0 }; // (a, _, _) -> a 2734 for (uint i = 0; i < inputs.size(); i++) { 2735 eval_map.put(inputs.at(i), input_funcs[2-i]); 2736 } 2737 2738 for (uint i = 0; i < partition.size(); i++) { 2739 Node* n = partition.at(i); 2740 2741 uint func1 = 0, func2 = 0, func3 = 0; 2742 eval_operands(n, func1, func2, func3, eval_map); 2743 2744 switch (n->Opcode()) { 2745 case Op_OrV: 2746 assert(func3 == 0, "not binary"); 2747 res = func1 | func2; 2748 break; 2749 case Op_AndV: 2750 assert(func3 == 0, "not binary"); 2751 res = func1 & func2; 2752 break; 2753 case Op_XorV: 2754 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2755 assert(func2 == 0 && func3 == 0, "not unary"); 2756 res = (~func1) & 0xFF; 2757 } else { 2758 assert(func3 == 0, "not binary"); 2759 res = func1 ^ func2; 2760 } 2761 break; 2762 case Op_MacroLogicV: 2763 // Ordering of inputs may change during evaluation of sub-tree 2764 // containing MacroLogic node as a child node, thus a re-evaluation 2765 // makes sure that function is evaluated in context of current 2766 // inputs. 2767 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 2768 break; 2769 2770 default: assert(false, "not supported: %s", n->Name()); 2771 } 2772 assert(res <= 0xFF, "invalid"); 2773 eval_map.put(n, res); 2774 } 2775 return res; 2776 } 2777 2778 // Criteria under which nodes gets packed into a macro logic node:- 2779 // 1) Parent and both child nodes are all unmasked or masked with 2780 // same predicates. 2781 // 2) Masked parent can be packed with left child if it is predicated 2782 // and both have same predicates. 2783 // 3) Masked parent can be packed with right child if its un-predicated 2784 // or has matching predication condition. 2785 // 4) An unmasked parent can be packed with an unmasked child. 2786 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2787 assert(partition.size() == 0, "not empty"); 2788 assert(inputs.size() == 0, "not empty"); 2789 if (is_vector_ternary_bitwise_op(n)) { 2790 return false; 2791 } 2792 2793 bool is_unary_op = is_vector_unary_bitwise_op(n); 2794 if (is_unary_op) { 2795 assert(collect_unique_inputs(n, inputs) == 1, "not unary"); 2796 return false; // too few inputs 2797 } 2798 2799 bool pack_left_child = true; 2800 bool pack_right_child = true; 2801 2802 bool left_child_LOP = is_vector_bitwise_op(n->in(1)); 2803 bool right_child_LOP = is_vector_bitwise_op(n->in(2)); 2804 2805 int left_child_input_cnt = 0; 2806 int right_child_input_cnt = 0; 2807 2808 bool parent_is_predicated = n->is_predicated_vector(); 2809 bool left_child_predicated = n->in(1)->is_predicated_vector(); 2810 bool right_child_predicated = n->in(2)->is_predicated_vector(); 2811 2812 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr; 2813 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2814 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2815 2816 do { 2817 if (pack_left_child && left_child_LOP && 2818 ((!parent_is_predicated && !left_child_predicated) || 2819 ((parent_is_predicated && left_child_predicated && 2820 parent_pred == left_child_pred)))) { 2821 partition.push(n->in(1)); 2822 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs); 2823 } else { 2824 inputs.push(n->in(1)); 2825 left_child_input_cnt = 1; 2826 } 2827 2828 if (pack_right_child && right_child_LOP && 2829 (!right_child_predicated || 2830 (right_child_predicated && parent_is_predicated && 2831 parent_pred == right_child_pred))) { 2832 partition.push(n->in(2)); 2833 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs); 2834 } else { 2835 inputs.push(n->in(2)); 2836 right_child_input_cnt = 1; 2837 } 2838 2839 if (inputs.size() > 3) { 2840 assert(partition.size() > 0, ""); 2841 inputs.clear(); 2842 partition.clear(); 2843 if (left_child_input_cnt > right_child_input_cnt) { 2844 pack_left_child = false; 2845 } else { 2846 pack_right_child = false; 2847 } 2848 } else { 2849 break; 2850 } 2851 } while(true); 2852 2853 if(partition.size()) { 2854 partition.push(n); 2855 } 2856 2857 return (partition.size() == 2 || partition.size() == 3) && 2858 (inputs.size() == 2 || inputs.size() == 3); 2859 } 2860 2861 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 2862 assert(is_vector_bitwise_op(n), "not a root"); 2863 2864 visited.set(n->_idx); 2865 2866 // 1) Do a DFS walk over the logic cone. 2867 for (uint i = 1; i < n->req(); i++) { 2868 Node* in = n->in(i); 2869 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 2870 process_logic_cone_root(igvn, in, visited); 2871 } 2872 } 2873 2874 // 2) Bottom up traversal: Merge node[s] with 2875 // the parent to form macro logic node. 2876 Unique_Node_List partition; 2877 Unique_Node_List inputs; 2878 if (compute_logic_cone(n, partition, inputs)) { 2879 const TypeVect* vt = n->bottom_type()->is_vect(); 2880 Node* pn = partition.at(partition.size() - 1); 2881 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2882 if (mask == nullptr || 2883 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) { 2884 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 2885 VectorNode::trace_new_vector(macro_logic, "MacroLogic"); 2886 igvn.replace_node(n, macro_logic); 2887 } 2888 } 2889 } 2890 2891 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 2892 ResourceMark rm; 2893 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 2894 Unique_Node_List list; 2895 collect_logic_cone_roots(list); 2896 2897 while (list.size() > 0) { 2898 Node* n = list.pop(); 2899 const TypeVect* vt = n->bottom_type()->is_vect(); 2900 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 2901 if (supported) { 2902 VectorSet visited(comp_arena()); 2903 process_logic_cone_root(igvn, n, visited); 2904 } 2905 } 2906 } 2907 } 2908 2909 //------------------------------Code_Gen--------------------------------------- 2910 // Given a graph, generate code for it 2911 void Compile::Code_Gen() { 2912 if (failing()) { 2913 return; 2914 } 2915 2916 // Perform instruction selection. You might think we could reclaim Matcher 2917 // memory PDQ, but actually the Matcher is used in generating spill code. 2918 // Internals of the Matcher (including some VectorSets) must remain live 2919 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2920 // set a bit in reclaimed memory. 2921 2922 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2923 // nodes. Mapping is only valid at the root of each matched subtree. 2924 NOT_PRODUCT( verify_graph_edges(); ) 2925 2926 Matcher matcher; 2927 _matcher = &matcher; 2928 { 2929 TracePhase tp("matcher", &timers[_t_matcher]); 2930 matcher.match(); 2931 if (failing()) { 2932 return; 2933 } 2934 } 2935 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2936 // nodes. Mapping is only valid at the root of each matched subtree. 2937 NOT_PRODUCT( verify_graph_edges(); ) 2938 2939 // If you have too many nodes, or if matching has failed, bail out 2940 check_node_count(0, "out of nodes matching instructions"); 2941 if (failing()) { 2942 return; 2943 } 2944 2945 print_method(PHASE_MATCHING, 2); 2946 2947 // Build a proper-looking CFG 2948 PhaseCFG cfg(node_arena(), root(), matcher); 2949 _cfg = &cfg; 2950 { 2951 TracePhase tp("scheduler", &timers[_t_scheduler]); 2952 bool success = cfg.do_global_code_motion(); 2953 if (!success) { 2954 return; 2955 } 2956 2957 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2958 NOT_PRODUCT( verify_graph_edges(); ) 2959 cfg.verify(); 2960 } 2961 2962 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2963 _regalloc = ®alloc; 2964 { 2965 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2966 // Perform register allocation. After Chaitin, use-def chains are 2967 // no longer accurate (at spill code) and so must be ignored. 2968 // Node->LRG->reg mappings are still accurate. 2969 _regalloc->Register_Allocate(); 2970 2971 // Bail out if the allocator builds too many nodes 2972 if (failing()) { 2973 return; 2974 } 2975 } 2976 2977 // Prior to register allocation we kept empty basic blocks in case the 2978 // the allocator needed a place to spill. After register allocation we 2979 // are not adding any new instructions. If any basic block is empty, we 2980 // can now safely remove it. 2981 { 2982 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2983 cfg.remove_empty_blocks(); 2984 if (do_freq_based_layout()) { 2985 PhaseBlockLayout layout(cfg); 2986 } else { 2987 cfg.set_loop_alignment(); 2988 } 2989 cfg.fixup_flow(); 2990 cfg.remove_unreachable_blocks(); 2991 cfg.verify_dominator_tree(); 2992 } 2993 2994 // Apply peephole optimizations 2995 if( OptoPeephole ) { 2996 TracePhase tp("peephole", &timers[_t_peephole]); 2997 PhasePeephole peep( _regalloc, cfg); 2998 peep.do_transform(); 2999 } 3000 3001 // Do late expand if CPU requires this. 3002 if (Matcher::require_postalloc_expand) { 3003 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 3004 cfg.postalloc_expand(_regalloc); 3005 } 3006 3007 // Convert Nodes to instruction bits in a buffer 3008 { 3009 TracePhase tp("output", &timers[_t_output]); 3010 PhaseOutput output; 3011 output.Output(); 3012 if (failing()) return; 3013 output.install(); 3014 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here 3015 } 3016 3017 // He's dead, Jim. 3018 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 3019 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 3020 } 3021 3022 //------------------------------Final_Reshape_Counts--------------------------- 3023 // This class defines counters to help identify when a method 3024 // may/must be executed using hardware with only 24-bit precision. 3025 struct Final_Reshape_Counts : public StackObj { 3026 int _call_count; // count non-inlined 'common' calls 3027 int _float_count; // count float ops requiring 24-bit precision 3028 int _double_count; // count double ops requiring more precision 3029 int _java_call_count; // count non-inlined 'java' calls 3030 int _inner_loop_count; // count loops which need alignment 3031 VectorSet _visited; // Visitation flags 3032 Node_List _tests; // Set of IfNodes & PCTableNodes 3033 3034 Final_Reshape_Counts() : 3035 _call_count(0), _float_count(0), _double_count(0), 3036 _java_call_count(0), _inner_loop_count(0) { } 3037 3038 void inc_call_count () { _call_count ++; } 3039 void inc_float_count () { _float_count ++; } 3040 void inc_double_count() { _double_count++; } 3041 void inc_java_call_count() { _java_call_count++; } 3042 void inc_inner_loop_count() { _inner_loop_count++; } 3043 3044 int get_call_count () const { return _call_count ; } 3045 int get_float_count () const { return _float_count ; } 3046 int get_double_count() const { return _double_count; } 3047 int get_java_call_count() const { return _java_call_count; } 3048 int get_inner_loop_count() const { return _inner_loop_count; } 3049 }; 3050 3051 // Eliminate trivially redundant StoreCMs and accumulate their 3052 // precedence edges. 3053 void Compile::eliminate_redundant_card_marks(Node* n) { 3054 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 3055 if (n->in(MemNode::Address)->outcnt() > 1) { 3056 // There are multiple users of the same address so it might be 3057 // possible to eliminate some of the StoreCMs 3058 Node* mem = n->in(MemNode::Memory); 3059 Node* adr = n->in(MemNode::Address); 3060 Node* val = n->in(MemNode::ValueIn); 3061 Node* prev = n; 3062 bool done = false; 3063 // Walk the chain of StoreCMs eliminating ones that match. As 3064 // long as it's a chain of single users then the optimization is 3065 // safe. Eliminating partially redundant StoreCMs would require 3066 // cloning copies down the other paths. 3067 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 3068 if (adr == mem->in(MemNode::Address) && 3069 val == mem->in(MemNode::ValueIn)) { 3070 // redundant StoreCM 3071 if (mem->req() > MemNode::OopStore) { 3072 // Hasn't been processed by this code yet. 3073 n->add_prec(mem->in(MemNode::OopStore)); 3074 } else { 3075 // Already converted to precedence edge 3076 for (uint i = mem->req(); i < mem->len(); i++) { 3077 // Accumulate any precedence edges 3078 if (mem->in(i) != nullptr) { 3079 n->add_prec(mem->in(i)); 3080 } 3081 } 3082 // Everything above this point has been processed. 3083 done = true; 3084 } 3085 // Eliminate the previous StoreCM 3086 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 3087 assert(mem->outcnt() == 0, "should be dead"); 3088 mem->disconnect_inputs(this); 3089 } else { 3090 prev = mem; 3091 } 3092 mem = prev->in(MemNode::Memory); 3093 } 3094 } 3095 } 3096 3097 //------------------------------final_graph_reshaping_impl---------------------- 3098 // Implement items 1-5 from final_graph_reshaping below. 3099 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3100 3101 if ( n->outcnt() == 0 ) return; // dead node 3102 uint nop = n->Opcode(); 3103 3104 // Check for 2-input instruction with "last use" on right input. 3105 // Swap to left input. Implements item (2). 3106 if( n->req() == 3 && // two-input instruction 3107 n->in(1)->outcnt() > 1 && // left use is NOT a last use 3108 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 3109 n->in(2)->outcnt() == 1 &&// right use IS a last use 3110 !n->in(2)->is_Con() ) { // right use is not a constant 3111 // Check for commutative opcode 3112 switch( nop ) { 3113 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 3114 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 3115 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 3116 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 3117 case Op_AndL: case Op_XorL: case Op_OrL: 3118 case Op_AndI: case Op_XorI: case Op_OrI: { 3119 // Move "last use" input to left by swapping inputs 3120 n->swap_edges(1, 2); 3121 break; 3122 } 3123 default: 3124 break; 3125 } 3126 } 3127 3128 #ifdef ASSERT 3129 if( n->is_Mem() ) { 3130 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 3131 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw || 3132 // oop will be recorded in oop map if load crosses safepoint 3133 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 3134 LoadNode::is_immutable_value(n->in(MemNode::Address))), 3135 "raw memory operations should have control edge"); 3136 } 3137 if (n->is_MemBar()) { 3138 MemBarNode* mb = n->as_MemBar(); 3139 if (mb->trailing_store() || mb->trailing_load_store()) { 3140 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 3141 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 3142 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 3143 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 3144 } else if (mb->leading()) { 3145 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 3146 } 3147 } 3148 #endif 3149 // Count FPU ops and common calls, implements item (3) 3150 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes); 3151 if (!gc_handled) { 3152 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes); 3153 } 3154 3155 // Collect CFG split points 3156 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3157 frc._tests.push(n); 3158 } 3159 } 3160 3161 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) { 3162 switch( nop ) { 3163 // Count all float operations that may use FPU 3164 case Op_AddF: 3165 case Op_SubF: 3166 case Op_MulF: 3167 case Op_DivF: 3168 case Op_NegF: 3169 case Op_ModF: 3170 case Op_ConvI2F: 3171 case Op_ConF: 3172 case Op_CmpF: 3173 case Op_CmpF3: 3174 case Op_StoreF: 3175 case Op_LoadF: 3176 // case Op_ConvL2F: // longs are split into 32-bit halves 3177 frc.inc_float_count(); 3178 break; 3179 3180 case Op_ConvF2D: 3181 case Op_ConvD2F: 3182 frc.inc_float_count(); 3183 frc.inc_double_count(); 3184 break; 3185 3186 // Count all double operations that may use FPU 3187 case Op_AddD: 3188 case Op_SubD: 3189 case Op_MulD: 3190 case Op_DivD: 3191 case Op_NegD: 3192 case Op_ModD: 3193 case Op_ConvI2D: 3194 case Op_ConvD2I: 3195 // case Op_ConvL2D: // handled by leaf call 3196 // case Op_ConvD2L: // handled by leaf call 3197 case Op_ConD: 3198 case Op_CmpD: 3199 case Op_CmpD3: 3200 case Op_StoreD: 3201 case Op_LoadD: 3202 case Op_LoadD_unaligned: 3203 frc.inc_double_count(); 3204 break; 3205 case Op_Opaque1: // Remove Opaque Nodes before matching 3206 case Op_Opaque3: 3207 n->subsume_by(n->in(1), this); 3208 break; 3209 case Op_CallStaticJava: 3210 case Op_CallJava: 3211 case Op_CallDynamicJava: 3212 frc.inc_java_call_count(); // Count java call site; 3213 case Op_CallRuntime: 3214 case Op_CallLeaf: 3215 case Op_CallLeafVector: 3216 case Op_CallLeafNoFP: { 3217 assert (n->is_Call(), ""); 3218 CallNode *call = n->as_Call(); 3219 // Count call sites where the FP mode bit would have to be flipped. 3220 // Do not count uncommon runtime calls: 3221 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3222 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3223 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3224 frc.inc_call_count(); // Count the call site 3225 } else { // See if uncommon argument is shared 3226 Node *n = call->in(TypeFunc::Parms); 3227 int nop = n->Opcode(); 3228 // Clone shared simple arguments to uncommon calls, item (1). 3229 if (n->outcnt() > 1 && 3230 !n->is_Proj() && 3231 nop != Op_CreateEx && 3232 nop != Op_CheckCastPP && 3233 nop != Op_DecodeN && 3234 nop != Op_DecodeNKlass && 3235 !n->is_Mem() && 3236 !n->is_Phi()) { 3237 Node *x = n->clone(); 3238 call->set_req(TypeFunc::Parms, x); 3239 } 3240 } 3241 break; 3242 } 3243 3244 case Op_StoreCM: 3245 { 3246 // Convert OopStore dependence into precedence edge 3247 Node* prec = n->in(MemNode::OopStore); 3248 n->del_req(MemNode::OopStore); 3249 n->add_prec(prec); 3250 eliminate_redundant_card_marks(n); 3251 } 3252 3253 // fall through 3254 3255 case Op_StoreB: 3256 case Op_StoreC: 3257 case Op_StoreI: 3258 case Op_StoreL: 3259 case Op_CompareAndSwapB: 3260 case Op_CompareAndSwapS: 3261 case Op_CompareAndSwapI: 3262 case Op_CompareAndSwapL: 3263 case Op_CompareAndSwapP: 3264 case Op_CompareAndSwapN: 3265 case Op_WeakCompareAndSwapB: 3266 case Op_WeakCompareAndSwapS: 3267 case Op_WeakCompareAndSwapI: 3268 case Op_WeakCompareAndSwapL: 3269 case Op_WeakCompareAndSwapP: 3270 case Op_WeakCompareAndSwapN: 3271 case Op_CompareAndExchangeB: 3272 case Op_CompareAndExchangeS: 3273 case Op_CompareAndExchangeI: 3274 case Op_CompareAndExchangeL: 3275 case Op_CompareAndExchangeP: 3276 case Op_CompareAndExchangeN: 3277 case Op_GetAndAddS: 3278 case Op_GetAndAddB: 3279 case Op_GetAndAddI: 3280 case Op_GetAndAddL: 3281 case Op_GetAndSetS: 3282 case Op_GetAndSetB: 3283 case Op_GetAndSetI: 3284 case Op_GetAndSetL: 3285 case Op_GetAndSetP: 3286 case Op_GetAndSetN: 3287 case Op_StoreP: 3288 case Op_StoreN: 3289 case Op_StoreNKlass: 3290 case Op_LoadB: 3291 case Op_LoadUB: 3292 case Op_LoadUS: 3293 case Op_LoadI: 3294 case Op_LoadKlass: 3295 case Op_LoadNKlass: 3296 case Op_LoadL: 3297 case Op_LoadL_unaligned: 3298 case Op_LoadP: 3299 case Op_LoadN: 3300 case Op_LoadRange: 3301 case Op_LoadS: 3302 break; 3303 3304 case Op_AddP: { // Assert sane base pointers 3305 Node *addp = n->in(AddPNode::Address); 3306 assert( !addp->is_AddP() || 3307 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3308 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3309 "Base pointers must match (addp %u)", addp->_idx ); 3310 #ifdef _LP64 3311 if ((UseCompressedOops || UseCompressedClassPointers) && 3312 addp->Opcode() == Op_ConP && 3313 addp == n->in(AddPNode::Base) && 3314 n->in(AddPNode::Offset)->is_Con()) { 3315 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3316 // on the platform and on the compressed oops mode. 3317 // Use addressing with narrow klass to load with offset on x86. 3318 // Some platforms can use the constant pool to load ConP. 3319 // Do this transformation here since IGVN will convert ConN back to ConP. 3320 const Type* t = addp->bottom_type(); 3321 bool is_oop = t->isa_oopptr() != nullptr; 3322 bool is_klass = t->isa_klassptr() != nullptr; 3323 3324 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3325 (is_klass && Matcher::const_klass_prefer_decode())) { 3326 Node* nn = nullptr; 3327 3328 int op = is_oop ? Op_ConN : Op_ConNKlass; 3329 3330 // Look for existing ConN node of the same exact type. 3331 Node* r = root(); 3332 uint cnt = r->outcnt(); 3333 for (uint i = 0; i < cnt; i++) { 3334 Node* m = r->raw_out(i); 3335 if (m!= nullptr && m->Opcode() == op && 3336 m->bottom_type()->make_ptr() == t) { 3337 nn = m; 3338 break; 3339 } 3340 } 3341 if (nn != nullptr) { 3342 // Decode a narrow oop to match address 3343 // [R12 + narrow_oop_reg<<3 + offset] 3344 if (is_oop) { 3345 nn = new DecodeNNode(nn, t); 3346 } else { 3347 nn = new DecodeNKlassNode(nn, t); 3348 } 3349 // Check for succeeding AddP which uses the same Base. 3350 // Otherwise we will run into the assertion above when visiting that guy. 3351 for (uint i = 0; i < n->outcnt(); ++i) { 3352 Node *out_i = n->raw_out(i); 3353 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3354 out_i->set_req(AddPNode::Base, nn); 3355 #ifdef ASSERT 3356 for (uint j = 0; j < out_i->outcnt(); ++j) { 3357 Node *out_j = out_i->raw_out(j); 3358 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3359 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3360 } 3361 #endif 3362 } 3363 } 3364 n->set_req(AddPNode::Base, nn); 3365 n->set_req(AddPNode::Address, nn); 3366 if (addp->outcnt() == 0) { 3367 addp->disconnect_inputs(this); 3368 } 3369 } 3370 } 3371 } 3372 #endif 3373 break; 3374 } 3375 3376 case Op_CastPP: { 3377 // Remove CastPP nodes to gain more freedom during scheduling but 3378 // keep the dependency they encode as control or precedence edges 3379 // (if control is set already) on memory operations. Some CastPP 3380 // nodes don't have a control (don't carry a dependency): skip 3381 // those. 3382 if (n->in(0) != nullptr) { 3383 ResourceMark rm; 3384 Unique_Node_List wq; 3385 wq.push(n); 3386 for (uint next = 0; next < wq.size(); ++next) { 3387 Node *m = wq.at(next); 3388 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3389 Node* use = m->fast_out(i); 3390 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3391 use->ensure_control_or_add_prec(n->in(0)); 3392 } else { 3393 switch(use->Opcode()) { 3394 case Op_AddP: 3395 case Op_DecodeN: 3396 case Op_DecodeNKlass: 3397 case Op_CheckCastPP: 3398 case Op_CastPP: 3399 wq.push(use); 3400 break; 3401 } 3402 } 3403 } 3404 } 3405 } 3406 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3407 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3408 Node* in1 = n->in(1); 3409 const Type* t = n->bottom_type(); 3410 Node* new_in1 = in1->clone(); 3411 new_in1->as_DecodeN()->set_type(t); 3412 3413 if (!Matcher::narrow_oop_use_complex_address()) { 3414 // 3415 // x86, ARM and friends can handle 2 adds in addressing mode 3416 // and Matcher can fold a DecodeN node into address by using 3417 // a narrow oop directly and do implicit null check in address: 3418 // 3419 // [R12 + narrow_oop_reg<<3 + offset] 3420 // NullCheck narrow_oop_reg 3421 // 3422 // On other platforms (Sparc) we have to keep new DecodeN node and 3423 // use it to do implicit null check in address: 3424 // 3425 // decode_not_null narrow_oop_reg, base_reg 3426 // [base_reg + offset] 3427 // NullCheck base_reg 3428 // 3429 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3430 // to keep the information to which null check the new DecodeN node 3431 // corresponds to use it as value in implicit_null_check(). 3432 // 3433 new_in1->set_req(0, n->in(0)); 3434 } 3435 3436 n->subsume_by(new_in1, this); 3437 if (in1->outcnt() == 0) { 3438 in1->disconnect_inputs(this); 3439 } 3440 } else { 3441 n->subsume_by(n->in(1), this); 3442 if (n->outcnt() == 0) { 3443 n->disconnect_inputs(this); 3444 } 3445 } 3446 break; 3447 } 3448 #ifdef _LP64 3449 case Op_CmpP: 3450 // Do this transformation here to preserve CmpPNode::sub() and 3451 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3452 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3453 Node* in1 = n->in(1); 3454 Node* in2 = n->in(2); 3455 if (!in1->is_DecodeNarrowPtr()) { 3456 in2 = in1; 3457 in1 = n->in(2); 3458 } 3459 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3460 3461 Node* new_in2 = nullptr; 3462 if (in2->is_DecodeNarrowPtr()) { 3463 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3464 new_in2 = in2->in(1); 3465 } else if (in2->Opcode() == Op_ConP) { 3466 const Type* t = in2->bottom_type(); 3467 if (t == TypePtr::NULL_PTR) { 3468 assert(in1->is_DecodeN(), "compare klass to null?"); 3469 // Don't convert CmpP null check into CmpN if compressed 3470 // oops implicit null check is not generated. 3471 // This will allow to generate normal oop implicit null check. 3472 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3473 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3474 // 3475 // This transformation together with CastPP transformation above 3476 // will generated code for implicit null checks for compressed oops. 3477 // 3478 // The original code after Optimize() 3479 // 3480 // LoadN memory, narrow_oop_reg 3481 // decode narrow_oop_reg, base_reg 3482 // CmpP base_reg, nullptr 3483 // CastPP base_reg // NotNull 3484 // Load [base_reg + offset], val_reg 3485 // 3486 // after these transformations will be 3487 // 3488 // LoadN memory, narrow_oop_reg 3489 // CmpN narrow_oop_reg, nullptr 3490 // decode_not_null narrow_oop_reg, base_reg 3491 // Load [base_reg + offset], val_reg 3492 // 3493 // and the uncommon path (== nullptr) will use narrow_oop_reg directly 3494 // since narrow oops can be used in debug info now (see the code in 3495 // final_graph_reshaping_walk()). 3496 // 3497 // At the end the code will be matched to 3498 // on x86: 3499 // 3500 // Load_narrow_oop memory, narrow_oop_reg 3501 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3502 // NullCheck narrow_oop_reg 3503 // 3504 // and on sparc: 3505 // 3506 // Load_narrow_oop memory, narrow_oop_reg 3507 // decode_not_null narrow_oop_reg, base_reg 3508 // Load [base_reg + offset], val_reg 3509 // NullCheck base_reg 3510 // 3511 } else if (t->isa_oopptr()) { 3512 new_in2 = ConNode::make(t->make_narrowoop()); 3513 } else if (t->isa_klassptr()) { 3514 new_in2 = ConNode::make(t->make_narrowklass()); 3515 } 3516 } 3517 if (new_in2 != nullptr) { 3518 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3519 n->subsume_by(cmpN, this); 3520 if (in1->outcnt() == 0) { 3521 in1->disconnect_inputs(this); 3522 } 3523 if (in2->outcnt() == 0) { 3524 in2->disconnect_inputs(this); 3525 } 3526 } 3527 } 3528 break; 3529 3530 case Op_DecodeN: 3531 case Op_DecodeNKlass: 3532 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3533 // DecodeN could be pinned when it can't be fold into 3534 // an address expression, see the code for Op_CastPP above. 3535 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3536 break; 3537 3538 case Op_EncodeP: 3539 case Op_EncodePKlass: { 3540 Node* in1 = n->in(1); 3541 if (in1->is_DecodeNarrowPtr()) { 3542 n->subsume_by(in1->in(1), this); 3543 } else if (in1->Opcode() == Op_ConP) { 3544 const Type* t = in1->bottom_type(); 3545 if (t == TypePtr::NULL_PTR) { 3546 assert(t->isa_oopptr(), "null klass?"); 3547 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3548 } else if (t->isa_oopptr()) { 3549 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3550 } else if (t->isa_klassptr()) { 3551 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3552 } 3553 } 3554 if (in1->outcnt() == 0) { 3555 in1->disconnect_inputs(this); 3556 } 3557 break; 3558 } 3559 3560 case Op_Proj: { 3561 if (OptimizeStringConcat || IncrementalInline) { 3562 ProjNode* proj = n->as_Proj(); 3563 if (proj->_is_io_use) { 3564 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, ""); 3565 // Separate projections were used for the exception path which 3566 // are normally removed by a late inline. If it wasn't inlined 3567 // then they will hang around and should just be replaced with 3568 // the original one. Merge them. 3569 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/); 3570 if (non_io_proj != nullptr) { 3571 proj->subsume_by(non_io_proj , this); 3572 } 3573 } 3574 } 3575 break; 3576 } 3577 3578 case Op_Phi: 3579 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3580 // The EncodeP optimization may create Phi with the same edges 3581 // for all paths. It is not handled well by Register Allocator. 3582 Node* unique_in = n->in(1); 3583 assert(unique_in != nullptr, ""); 3584 uint cnt = n->req(); 3585 for (uint i = 2; i < cnt; i++) { 3586 Node* m = n->in(i); 3587 assert(m != nullptr, ""); 3588 if (unique_in != m) 3589 unique_in = nullptr; 3590 } 3591 if (unique_in != nullptr) { 3592 n->subsume_by(unique_in, this); 3593 } 3594 } 3595 break; 3596 3597 #endif 3598 3599 #ifdef ASSERT 3600 case Op_CastII: 3601 // Verify that all range check dependent CastII nodes were removed. 3602 if (n->isa_CastII()->has_range_check()) { 3603 n->dump(3); 3604 assert(false, "Range check dependent CastII node was not removed"); 3605 } 3606 break; 3607 #endif 3608 3609 case Op_ModI: 3610 if (UseDivMod) { 3611 // Check if a%b and a/b both exist 3612 Node* d = n->find_similar(Op_DivI); 3613 if (d) { 3614 // Replace them with a fused divmod if supported 3615 if (Matcher::has_match_rule(Op_DivModI)) { 3616 DivModINode* divmod = DivModINode::make(n); 3617 d->subsume_by(divmod->div_proj(), this); 3618 n->subsume_by(divmod->mod_proj(), this); 3619 } else { 3620 // replace a%b with a-((a/b)*b) 3621 Node* mult = new MulINode(d, d->in(2)); 3622 Node* sub = new SubINode(d->in(1), mult); 3623 n->subsume_by(sub, this); 3624 } 3625 } 3626 } 3627 break; 3628 3629 case Op_ModL: 3630 if (UseDivMod) { 3631 // Check if a%b and a/b both exist 3632 Node* d = n->find_similar(Op_DivL); 3633 if (d) { 3634 // Replace them with a fused divmod if supported 3635 if (Matcher::has_match_rule(Op_DivModL)) { 3636 DivModLNode* divmod = DivModLNode::make(n); 3637 d->subsume_by(divmod->div_proj(), this); 3638 n->subsume_by(divmod->mod_proj(), this); 3639 } else { 3640 // replace a%b with a-((a/b)*b) 3641 Node* mult = new MulLNode(d, d->in(2)); 3642 Node* sub = new SubLNode(d->in(1), mult); 3643 n->subsume_by(sub, this); 3644 } 3645 } 3646 } 3647 break; 3648 3649 case Op_UModI: 3650 if (UseDivMod) { 3651 // Check if a%b and a/b both exist 3652 Node* d = n->find_similar(Op_UDivI); 3653 if (d) { 3654 // Replace them with a fused unsigned divmod if supported 3655 if (Matcher::has_match_rule(Op_UDivModI)) { 3656 UDivModINode* divmod = UDivModINode::make(n); 3657 d->subsume_by(divmod->div_proj(), this); 3658 n->subsume_by(divmod->mod_proj(), this); 3659 } else { 3660 // replace a%b with a-((a/b)*b) 3661 Node* mult = new MulINode(d, d->in(2)); 3662 Node* sub = new SubINode(d->in(1), mult); 3663 n->subsume_by(sub, this); 3664 } 3665 } 3666 } 3667 break; 3668 3669 case Op_UModL: 3670 if (UseDivMod) { 3671 // Check if a%b and a/b both exist 3672 Node* d = n->find_similar(Op_UDivL); 3673 if (d) { 3674 // Replace them with a fused unsigned divmod if supported 3675 if (Matcher::has_match_rule(Op_UDivModL)) { 3676 UDivModLNode* divmod = UDivModLNode::make(n); 3677 d->subsume_by(divmod->div_proj(), this); 3678 n->subsume_by(divmod->mod_proj(), this); 3679 } else { 3680 // replace a%b with a-((a/b)*b) 3681 Node* mult = new MulLNode(d, d->in(2)); 3682 Node* sub = new SubLNode(d->in(1), mult); 3683 n->subsume_by(sub, this); 3684 } 3685 } 3686 } 3687 break; 3688 3689 case Op_LoadVector: 3690 case Op_StoreVector: 3691 case Op_LoadVectorGather: 3692 case Op_StoreVectorScatter: 3693 case Op_LoadVectorGatherMasked: 3694 case Op_StoreVectorScatterMasked: 3695 case Op_VectorCmpMasked: 3696 case Op_VectorMaskGen: 3697 case Op_LoadVectorMasked: 3698 case Op_StoreVectorMasked: 3699 break; 3700 3701 case Op_AddReductionVI: 3702 case Op_AddReductionVL: 3703 case Op_AddReductionVF: 3704 case Op_AddReductionVD: 3705 case Op_MulReductionVI: 3706 case Op_MulReductionVL: 3707 case Op_MulReductionVF: 3708 case Op_MulReductionVD: 3709 case Op_MinReductionV: 3710 case Op_MaxReductionV: 3711 case Op_AndReductionV: 3712 case Op_OrReductionV: 3713 case Op_XorReductionV: 3714 break; 3715 3716 case Op_PackB: 3717 case Op_PackS: 3718 case Op_PackI: 3719 case Op_PackF: 3720 case Op_PackL: 3721 case Op_PackD: 3722 if (n->req()-1 > 2) { 3723 // Replace many operand PackNodes with a binary tree for matching 3724 PackNode* p = (PackNode*) n; 3725 Node* btp = p->binary_tree_pack(1, n->req()); 3726 n->subsume_by(btp, this); 3727 } 3728 break; 3729 case Op_Loop: 3730 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop"); 3731 case Op_CountedLoop: 3732 case Op_LongCountedLoop: 3733 case Op_OuterStripMinedLoop: 3734 if (n->as_Loop()->is_inner_loop()) { 3735 frc.inc_inner_loop_count(); 3736 } 3737 n->as_Loop()->verify_strip_mined(0); 3738 break; 3739 case Op_LShiftI: 3740 case Op_RShiftI: 3741 case Op_URShiftI: 3742 case Op_LShiftL: 3743 case Op_RShiftL: 3744 case Op_URShiftL: 3745 if (Matcher::need_masked_shift_count) { 3746 // The cpu's shift instructions don't restrict the count to the 3747 // lower 5/6 bits. We need to do the masking ourselves. 3748 Node* in2 = n->in(2); 3749 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3750 const TypeInt* t = in2->find_int_type(); 3751 if (t != nullptr && t->is_con()) { 3752 juint shift = t->get_con(); 3753 if (shift > mask) { // Unsigned cmp 3754 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3755 } 3756 } else { 3757 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) { 3758 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3759 n->set_req(2, shift); 3760 } 3761 } 3762 if (in2->outcnt() == 0) { // Remove dead node 3763 in2->disconnect_inputs(this); 3764 } 3765 } 3766 break; 3767 case Op_MemBarStoreStore: 3768 case Op_MemBarRelease: 3769 // Break the link with AllocateNode: it is no longer useful and 3770 // confuses register allocation. 3771 if (n->req() > MemBarNode::Precedent) { 3772 n->set_req(MemBarNode::Precedent, top()); 3773 } 3774 break; 3775 case Op_MemBarAcquire: { 3776 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3777 // At parse time, the trailing MemBarAcquire for a volatile load 3778 // is created with an edge to the load. After optimizations, 3779 // that input may be a chain of Phis. If those phis have no 3780 // other use, then the MemBarAcquire keeps them alive and 3781 // register allocation can be confused. 3782 dead_nodes.push(n->in(MemBarNode::Precedent)); 3783 n->set_req(MemBarNode::Precedent, top()); 3784 } 3785 break; 3786 } 3787 case Op_Blackhole: 3788 break; 3789 case Op_RangeCheck: { 3790 RangeCheckNode* rc = n->as_RangeCheck(); 3791 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3792 n->subsume_by(iff, this); 3793 frc._tests.push(iff); 3794 break; 3795 } 3796 case Op_ConvI2L: { 3797 if (!Matcher::convi2l_type_required) { 3798 // Code generation on some platforms doesn't need accurate 3799 // ConvI2L types. Widening the type can help remove redundant 3800 // address computations. 3801 n->as_Type()->set_type(TypeLong::INT); 3802 ResourceMark rm; 3803 Unique_Node_List wq; 3804 wq.push(n); 3805 for (uint next = 0; next < wq.size(); next++) { 3806 Node *m = wq.at(next); 3807 3808 for(;;) { 3809 // Loop over all nodes with identical inputs edges as m 3810 Node* k = m->find_similar(m->Opcode()); 3811 if (k == nullptr) { 3812 break; 3813 } 3814 // Push their uses so we get a chance to remove node made 3815 // redundant 3816 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3817 Node* u = k->fast_out(i); 3818 if (u->Opcode() == Op_LShiftL || 3819 u->Opcode() == Op_AddL || 3820 u->Opcode() == Op_SubL || 3821 u->Opcode() == Op_AddP) { 3822 wq.push(u); 3823 } 3824 } 3825 // Replace all nodes with identical edges as m with m 3826 k->subsume_by(m, this); 3827 } 3828 } 3829 } 3830 break; 3831 } 3832 case Op_CmpUL: { 3833 if (!Matcher::has_match_rule(Op_CmpUL)) { 3834 // No support for unsigned long comparisons 3835 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3836 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3837 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3838 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3839 Node* andl = new AndLNode(orl, remove_sign_mask); 3840 Node* cmp = new CmpLNode(andl, n->in(2)); 3841 n->subsume_by(cmp, this); 3842 } 3843 break; 3844 } 3845 default: 3846 assert(!n->is_Call(), ""); 3847 assert(!n->is_Mem(), ""); 3848 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3849 break; 3850 } 3851 } 3852 3853 //------------------------------final_graph_reshaping_walk--------------------- 3854 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3855 // requires that the walk visits a node's inputs before visiting the node. 3856 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3857 Unique_Node_List sfpt; 3858 3859 frc._visited.set(root->_idx); // first, mark node as visited 3860 uint cnt = root->req(); 3861 Node *n = root; 3862 uint i = 0; 3863 while (true) { 3864 if (i < cnt) { 3865 // Place all non-visited non-null inputs onto stack 3866 Node* m = n->in(i); 3867 ++i; 3868 if (m != nullptr && !frc._visited.test_set(m->_idx)) { 3869 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) { 3870 // compute worst case interpreter size in case of a deoptimization 3871 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3872 3873 sfpt.push(m); 3874 } 3875 cnt = m->req(); 3876 nstack.push(n, i); // put on stack parent and next input's index 3877 n = m; 3878 i = 0; 3879 } 3880 } else { 3881 // Now do post-visit work 3882 final_graph_reshaping_impl(n, frc, dead_nodes); 3883 if (nstack.is_empty()) 3884 break; // finished 3885 n = nstack.node(); // Get node from stack 3886 cnt = n->req(); 3887 i = nstack.index(); 3888 nstack.pop(); // Shift to the next node on stack 3889 } 3890 } 3891 3892 // Skip next transformation if compressed oops are not used. 3893 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3894 (!UseCompressedOops && !UseCompressedClassPointers)) 3895 return; 3896 3897 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3898 // It could be done for an uncommon traps or any safepoints/calls 3899 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3900 while (sfpt.size() > 0) { 3901 n = sfpt.pop(); 3902 JVMState *jvms = n->as_SafePoint()->jvms(); 3903 assert(jvms != nullptr, "sanity"); 3904 int start = jvms->debug_start(); 3905 int end = n->req(); 3906 bool is_uncommon = (n->is_CallStaticJava() && 3907 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3908 for (int j = start; j < end; j++) { 3909 Node* in = n->in(j); 3910 if (in->is_DecodeNarrowPtr()) { 3911 bool safe_to_skip = true; 3912 if (!is_uncommon ) { 3913 // Is it safe to skip? 3914 for (uint i = 0; i < in->outcnt(); i++) { 3915 Node* u = in->raw_out(i); 3916 if (!u->is_SafePoint() || 3917 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3918 safe_to_skip = false; 3919 } 3920 } 3921 } 3922 if (safe_to_skip) { 3923 n->set_req(j, in->in(1)); 3924 } 3925 if (in->outcnt() == 0) { 3926 in->disconnect_inputs(this); 3927 } 3928 } 3929 } 3930 } 3931 } 3932 3933 //------------------------------final_graph_reshaping-------------------------- 3934 // Final Graph Reshaping. 3935 // 3936 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3937 // and not commoned up and forced early. Must come after regular 3938 // optimizations to avoid GVN undoing the cloning. Clone constant 3939 // inputs to Loop Phis; these will be split by the allocator anyways. 3940 // Remove Opaque nodes. 3941 // (2) Move last-uses by commutative operations to the left input to encourage 3942 // Intel update-in-place two-address operations and better register usage 3943 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3944 // calls canonicalizing them back. 3945 // (3) Count the number of double-precision FP ops, single-precision FP ops 3946 // and call sites. On Intel, we can get correct rounding either by 3947 // forcing singles to memory (requires extra stores and loads after each 3948 // FP bytecode) or we can set a rounding mode bit (requires setting and 3949 // clearing the mode bit around call sites). The mode bit is only used 3950 // if the relative frequency of single FP ops to calls is low enough. 3951 // This is a key transform for SPEC mpeg_audio. 3952 // (4) Detect infinite loops; blobs of code reachable from above but not 3953 // below. Several of the Code_Gen algorithms fail on such code shapes, 3954 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3955 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3956 // Detection is by looking for IfNodes where only 1 projection is 3957 // reachable from below or CatchNodes missing some targets. 3958 // (5) Assert for insane oop offsets in debug mode. 3959 3960 bool Compile::final_graph_reshaping() { 3961 // an infinite loop may have been eliminated by the optimizer, 3962 // in which case the graph will be empty. 3963 if (root()->req() == 1) { 3964 // Do not compile method that is only a trivial infinite loop, 3965 // since the content of the loop may have been eliminated. 3966 record_method_not_compilable("trivial infinite loop"); 3967 return true; 3968 } 3969 3970 // Expensive nodes have their control input set to prevent the GVN 3971 // from freely commoning them. There's no GVN beyond this point so 3972 // no need to keep the control input. We want the expensive nodes to 3973 // be freely moved to the least frequent code path by gcm. 3974 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3975 for (int i = 0; i < expensive_count(); i++) { 3976 _expensive_nodes.at(i)->set_req(0, nullptr); 3977 } 3978 3979 Final_Reshape_Counts frc; 3980 3981 // Visit everybody reachable! 3982 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3983 Node_Stack nstack(live_nodes() >> 1); 3984 Unique_Node_List dead_nodes; 3985 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes); 3986 3987 // Check for unreachable (from below) code (i.e., infinite loops). 3988 for( uint i = 0; i < frc._tests.size(); i++ ) { 3989 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3990 // Get number of CFG targets. 3991 // Note that PCTables include exception targets after calls. 3992 uint required_outcnt = n->required_outcnt(); 3993 if (n->outcnt() != required_outcnt) { 3994 // Check for a few special cases. Rethrow Nodes never take the 3995 // 'fall-thru' path, so expected kids is 1 less. 3996 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3997 if (n->in(0)->in(0)->is_Call()) { 3998 CallNode* call = n->in(0)->in(0)->as_Call(); 3999 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 4000 required_outcnt--; // Rethrow always has 1 less kid 4001 } else if (call->req() > TypeFunc::Parms && 4002 call->is_CallDynamicJava()) { 4003 // Check for null receiver. In such case, the optimizer has 4004 // detected that the virtual call will always result in a null 4005 // pointer exception. The fall-through projection of this CatchNode 4006 // will not be populated. 4007 Node* arg0 = call->in(TypeFunc::Parms); 4008 if (arg0->is_Type() && 4009 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 4010 required_outcnt--; 4011 } 4012 } else if (call->entry_point() == OptoRuntime::new_array_Java() || 4013 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4014 // Check for illegal array length. In such case, the optimizer has 4015 // detected that the allocation attempt will always result in an 4016 // exception. There is no fall-through projection of this CatchNode . 4017 assert(call->is_CallStaticJava(), "static call expected"); 4018 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4019 uint valid_length_test_input = call->req() - 1; 4020 Node* valid_length_test = call->in(valid_length_test_input); 4021 call->del_req(valid_length_test_input); 4022 if (valid_length_test->find_int_con(1) == 0) { 4023 required_outcnt--; 4024 } 4025 dead_nodes.push(valid_length_test); 4026 assert(n->outcnt() == required_outcnt, "malformed control flow"); 4027 continue; 4028 } 4029 } 4030 } 4031 4032 // Recheck with a better notion of 'required_outcnt' 4033 if (n->outcnt() != required_outcnt) { 4034 record_method_not_compilable("malformed control flow"); 4035 return true; // Not all targets reachable! 4036 } 4037 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) { 4038 CallNode* call = n->in(0)->in(0)->as_Call(); 4039 if (call->entry_point() == OptoRuntime::new_array_Java() || 4040 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4041 assert(call->is_CallStaticJava(), "static call expected"); 4042 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4043 uint valid_length_test_input = call->req() - 1; 4044 dead_nodes.push(call->in(valid_length_test_input)); 4045 call->del_req(valid_length_test_input); // valid length test useless now 4046 } 4047 } 4048 // Check that I actually visited all kids. Unreached kids 4049 // must be infinite loops. 4050 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 4051 if (!frc._visited.test(n->fast_out(j)->_idx)) { 4052 record_method_not_compilable("infinite loop"); 4053 return true; // Found unvisited kid; must be unreach 4054 } 4055 4056 // Here so verification code in final_graph_reshaping_walk() 4057 // always see an OuterStripMinedLoopEnd 4058 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) { 4059 IfNode* init_iff = n->as_If(); 4060 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 4061 n->subsume_by(iff, this); 4062 } 4063 } 4064 4065 while (dead_nodes.size() > 0) { 4066 Node* m = dead_nodes.pop(); 4067 if (m->outcnt() == 0 && m != top()) { 4068 for (uint j = 0; j < m->req(); j++) { 4069 Node* in = m->in(j); 4070 if (in != nullptr) { 4071 dead_nodes.push(in); 4072 } 4073 } 4074 m->disconnect_inputs(this); 4075 } 4076 } 4077 4078 #ifdef IA32 4079 // If original bytecodes contained a mixture of floats and doubles 4080 // check if the optimizer has made it homogeneous, item (3). 4081 if (UseSSE == 0 && 4082 frc.get_float_count() > 32 && 4083 frc.get_double_count() == 0 && 4084 (10 * frc.get_call_count() < frc.get_float_count()) ) { 4085 set_24_bit_selection_and_mode(false, true); 4086 } 4087 #endif // IA32 4088 4089 set_java_calls(frc.get_java_call_count()); 4090 set_inner_loops(frc.get_inner_loop_count()); 4091 4092 // No infinite loops, no reason to bail out. 4093 return false; 4094 } 4095 4096 //-----------------------------too_many_traps---------------------------------- 4097 // Report if there are too many traps at the current method and bci. 4098 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 4099 bool Compile::too_many_traps(ciMethod* method, 4100 int bci, 4101 Deoptimization::DeoptReason reason) { 4102 ciMethodData* md = method->method_data(); 4103 if (md->is_empty()) { 4104 // Assume the trap has not occurred, or that it occurred only 4105 // because of a transient condition during start-up in the interpreter. 4106 return false; 4107 } 4108 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4109 if (md->has_trap_at(bci, m, reason) != 0) { 4110 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 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' count='%d'", 4115 Deoptimization::trap_reason_name(reason), 4116 md->trap_count(reason)); 4117 return true; 4118 } else { 4119 // Ignore method/bci and see if there have been too many globally. 4120 return too_many_traps(reason, md); 4121 } 4122 } 4123 4124 // Less-accurate variant which does not require a method and bci. 4125 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 4126 ciMethodData* logmd) { 4127 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 4128 // Too many traps globally. 4129 // Note that we use cumulative trap_count, not just md->trap_count. 4130 if (log()) { 4131 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason); 4132 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 4133 Deoptimization::trap_reason_name(reason), 4134 mcount, trap_count(reason)); 4135 } 4136 return true; 4137 } else { 4138 // The coast is clear. 4139 return false; 4140 } 4141 } 4142 4143 //--------------------------too_many_recompiles-------------------------------- 4144 // Report if there are too many recompiles at the current method and bci. 4145 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 4146 // Is not eager to return true, since this will cause the compiler to use 4147 // Action_none for a trap point, to avoid too many recompilations. 4148 bool Compile::too_many_recompiles(ciMethod* method, 4149 int bci, 4150 Deoptimization::DeoptReason reason) { 4151 ciMethodData* md = method->method_data(); 4152 if (md->is_empty()) { 4153 // Assume the trap has not occurred, or that it occurred only 4154 // because of a transient condition during start-up in the interpreter. 4155 return false; 4156 } 4157 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 4158 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 4159 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 4160 Deoptimization::DeoptReason per_bc_reason 4161 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 4162 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4163 if ((per_bc_reason == Deoptimization::Reason_none 4164 || md->has_trap_at(bci, m, reason) != 0) 4165 // The trap frequency measure we care about is the recompile count: 4166 && md->trap_recompiled_at(bci, m) 4167 && md->overflow_recompile_count() >= bc_cutoff) { 4168 // Do not emit a trap here if it has already caused recompilations. 4169 // Also, if there are multiple reasons, or if there is no per-BCI record, 4170 // assume the worst. 4171 if (log()) 4172 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 4173 Deoptimization::trap_reason_name(reason), 4174 md->trap_count(reason), 4175 md->overflow_recompile_count()); 4176 return true; 4177 } else if (trap_count(reason) != 0 4178 && decompile_count() >= m_cutoff) { 4179 // Too many recompiles globally, and we have seen this sort of trap. 4180 // Use cumulative decompile_count, not just md->decompile_count. 4181 if (log()) 4182 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 4183 Deoptimization::trap_reason_name(reason), 4184 md->trap_count(reason), trap_count(reason), 4185 md->decompile_count(), decompile_count()); 4186 return true; 4187 } else { 4188 // The coast is clear. 4189 return false; 4190 } 4191 } 4192 4193 // Compute when not to trap. Used by matching trap based nodes and 4194 // NullCheck optimization. 4195 void Compile::set_allowed_deopt_reasons() { 4196 _allowed_reasons = 0; 4197 if (is_method_compilation()) { 4198 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 4199 assert(rs < BitsPerInt, "recode bit map"); 4200 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 4201 _allowed_reasons |= nth_bit(rs); 4202 } 4203 } 4204 } 4205 } 4206 4207 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 4208 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 4209 } 4210 4211 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 4212 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 4213 } 4214 4215 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 4216 if (holder->is_initialized()) { 4217 return false; 4218 } 4219 if (holder->is_being_initialized()) { 4220 if (accessing_method->holder() == holder) { 4221 // Access inside a class. The barrier can be elided when access happens in <clinit>, 4222 // <init>, or a static method. In all those cases, there was an initialization 4223 // barrier on the holder klass passed. 4224 if (accessing_method->is_static_initializer() || 4225 accessing_method->is_object_initializer() || 4226 accessing_method->is_static()) { 4227 return false; 4228 } 4229 } else if (accessing_method->holder()->is_subclass_of(holder)) { 4230 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 4231 // In case of <init> or a static method, the barrier is on the subclass is not enough: 4232 // child class can become fully initialized while its parent class is still being initialized. 4233 if (accessing_method->is_static_initializer()) { 4234 return false; 4235 } 4236 } 4237 ciMethod* root = method(); // the root method of compilation 4238 if (root != accessing_method) { 4239 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 4240 } 4241 } 4242 return true; 4243 } 4244 4245 #ifndef PRODUCT 4246 //------------------------------verify_bidirectional_edges--------------------- 4247 // For each input edge to a node (ie - for each Use-Def edge), verify that 4248 // there is a corresponding Def-Use edge. 4249 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) { 4250 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc 4251 uint stack_size = live_nodes() >> 4; 4252 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize)); 4253 nstack.push(_root); 4254 4255 while (nstack.size() > 0) { 4256 Node* n = nstack.pop(); 4257 if (visited.member(n)) { 4258 continue; 4259 } 4260 visited.push(n); 4261 4262 // Walk over all input edges, checking for correspondence 4263 uint length = n->len(); 4264 for (uint i = 0; i < length; i++) { 4265 Node* in = n->in(i); 4266 if (in != nullptr && !visited.member(in)) { 4267 nstack.push(in); // Put it on stack 4268 } 4269 if (in != nullptr && !in->is_top()) { 4270 // Count instances of `next` 4271 int cnt = 0; 4272 for (uint idx = 0; idx < in->_outcnt; idx++) { 4273 if (in->_out[idx] == n) { 4274 cnt++; 4275 } 4276 } 4277 assert(cnt > 0, "Failed to find Def-Use edge."); 4278 // Check for duplicate edges 4279 // walk the input array downcounting the input edges to n 4280 for (uint j = 0; j < length; j++) { 4281 if (n->in(j) == in) { 4282 cnt--; 4283 } 4284 } 4285 assert(cnt == 0, "Mismatched edge count."); 4286 } else if (in == nullptr) { 4287 assert(i == 0 || i >= n->req() || 4288 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() || 4289 (n->is_Unlock() && i == (n->req() - 1)) || 4290 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion 4291 "only region, phi, arraycopy, unlock or membar nodes have null data edges"); 4292 } else { 4293 assert(in->is_top(), "sanity"); 4294 // Nothing to check. 4295 } 4296 } 4297 } 4298 } 4299 4300 //------------------------------verify_graph_edges--------------------------- 4301 // Walk the Graph and verify that there is a one-to-one correspondence 4302 // between Use-Def edges and Def-Use edges in the graph. 4303 void Compile::verify_graph_edges(bool no_dead_code) { 4304 if (VerifyGraphEdges) { 4305 Unique_Node_List visited; 4306 4307 // Call graph walk to check edges 4308 verify_bidirectional_edges(visited); 4309 if (no_dead_code) { 4310 // Now make sure that no visited node is used by an unvisited node. 4311 bool dead_nodes = false; 4312 Unique_Node_List checked; 4313 while (visited.size() > 0) { 4314 Node* n = visited.pop(); 4315 checked.push(n); 4316 for (uint i = 0; i < n->outcnt(); i++) { 4317 Node* use = n->raw_out(i); 4318 if (checked.member(use)) continue; // already checked 4319 if (visited.member(use)) continue; // already in the graph 4320 if (use->is_Con()) continue; // a dead ConNode is OK 4321 // At this point, we have found a dead node which is DU-reachable. 4322 if (!dead_nodes) { 4323 tty->print_cr("*** Dead nodes reachable via DU edges:"); 4324 dead_nodes = true; 4325 } 4326 use->dump(2); 4327 tty->print_cr("---"); 4328 checked.push(use); // No repeats; pretend it is now checked. 4329 } 4330 } 4331 assert(!dead_nodes, "using nodes must be reachable from root"); 4332 } 4333 } 4334 } 4335 #endif 4336 4337 // The Compile object keeps track of failure reasons separately from the ciEnv. 4338 // This is required because there is not quite a 1-1 relation between the 4339 // ciEnv and its compilation task and the Compile object. Note that one 4340 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4341 // to backtrack and retry without subsuming loads. Other than this backtracking 4342 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4343 // by the logic in C2Compiler. 4344 void Compile::record_failure(const char* reason) { 4345 if (log() != nullptr) { 4346 log()->elem("failure reason='%s' phase='compile'", reason); 4347 } 4348 if (_failure_reason == nullptr) { 4349 // Record the first failure reason. 4350 _failure_reason = reason; 4351 } 4352 4353 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4354 C->print_method(PHASE_FAILURE, 1); 4355 } 4356 _root = nullptr; // flush the graph, too 4357 } 4358 4359 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 4360 : TraceTime(name, accumulator, CITime, CITimeVerbose), 4361 _compile(Compile::current()), 4362 _log(nullptr), 4363 _phase_name(name), 4364 _dolog(CITimeVerbose) 4365 { 4366 assert(_compile != nullptr, "sanity check"); 4367 if (_dolog) { 4368 _log = _compile->log(); 4369 } 4370 if (_log != nullptr) { 4371 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes()); 4372 _log->stamp(); 4373 _log->end_head(); 4374 } 4375 } 4376 4377 Compile::TracePhase::~TracePhase() { 4378 if (_compile->failing()) return; 4379 #ifdef ASSERT 4380 if (PrintIdealNodeCount) { 4381 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4382 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk()); 4383 } 4384 4385 if (VerifyIdealNodeCount) { 4386 _compile->print_missing_nodes(); 4387 } 4388 #endif 4389 4390 if (_log != nullptr) { 4391 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes()); 4392 } 4393 } 4394 4395 //----------------------------static_subtype_check----------------------------- 4396 // Shortcut important common cases when superklass is exact: 4397 // (0) superklass is java.lang.Object (can occur in reflective code) 4398 // (1) subklass is already limited to a subtype of superklass => always ok 4399 // (2) subklass does not overlap with superklass => always fail 4400 // (3) superklass has NO subtypes and we can check with a simple compare. 4401 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) { 4402 if (skip) { 4403 return SSC_full_test; // Let caller generate the general case. 4404 } 4405 4406 if (subk->is_java_subtype_of(superk)) { 4407 return SSC_always_true; // (0) and (1) this test cannot fail 4408 } 4409 4410 if (!subk->maybe_java_subtype_of(superk)) { 4411 return SSC_always_false; // (2) true path dead; no dynamic test needed 4412 } 4413 4414 const Type* superelem = superk; 4415 if (superk->isa_aryklassptr()) { 4416 int ignored; 4417 superelem = superk->is_aryklassptr()->base_element_type(ignored); 4418 } 4419 4420 if (superelem->isa_instklassptr()) { 4421 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass(); 4422 if (!ik->has_subklass()) { 4423 if (!ik->is_final()) { 4424 // Add a dependency if there is a chance of a later subclass. 4425 dependencies()->assert_leaf_type(ik); 4426 } 4427 if (!superk->maybe_java_subtype_of(subk)) { 4428 return SSC_always_false; 4429 } 4430 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4431 } 4432 } else { 4433 // A primitive array type has no subtypes. 4434 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4435 } 4436 4437 return SSC_full_test; 4438 } 4439 4440 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4441 #ifdef _LP64 4442 // The scaled index operand to AddP must be a clean 64-bit value. 4443 // Java allows a 32-bit int to be incremented to a negative 4444 // value, which appears in a 64-bit register as a large 4445 // positive number. Using that large positive number as an 4446 // operand in pointer arithmetic has bad consequences. 4447 // On the other hand, 32-bit overflow is rare, and the possibility 4448 // can often be excluded, if we annotate the ConvI2L node with 4449 // a type assertion that its value is known to be a small positive 4450 // number. (The prior range check has ensured this.) 4451 // This assertion is used by ConvI2LNode::Ideal. 4452 int index_max = max_jint - 1; // array size is max_jint, index is one less 4453 if (sizetype != nullptr) index_max = sizetype->_hi - 1; 4454 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4455 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4456 #endif 4457 return idx; 4458 } 4459 4460 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4461 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 4462 if (ctrl != nullptr) { 4463 // Express control dependency by a CastII node with a narrow type. 4464 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */); 4465 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4466 // node from floating above the range check during loop optimizations. Otherwise, the 4467 // ConvI2L node may be eliminated independently of the range check, causing the data path 4468 // to become TOP while the control path is still there (although it's unreachable). 4469 value->set_req(0, ctrl); 4470 value = phase->transform(value); 4471 } 4472 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4473 return phase->transform(new ConvI2LNode(value, ltype)); 4474 } 4475 4476 // The message about the current inlining is accumulated in 4477 // _print_inlining_stream and transferred into the _print_inlining_list 4478 // once we know whether inlining succeeds or not. For regular 4479 // inlining, messages are appended to the buffer pointed by 4480 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4481 // a new buffer is added after _print_inlining_idx in the list. This 4482 // way we can update the inlining message for late inlining call site 4483 // when the inlining is attempted again. 4484 void Compile::print_inlining_init() { 4485 if (print_inlining() || print_intrinsics()) { 4486 // print_inlining_init is actually called several times. 4487 print_inlining_reset(); 4488 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer()); 4489 } 4490 } 4491 4492 void Compile::print_inlining_reinit() { 4493 if (print_inlining() || print_intrinsics()) { 4494 print_inlining_reset(); 4495 } 4496 } 4497 4498 void Compile::print_inlining_reset() { 4499 _print_inlining_stream->reset(); 4500 } 4501 4502 void Compile::print_inlining_commit() { 4503 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4504 // Transfer the message from _print_inlining_stream to the current 4505 // _print_inlining_list buffer and clear _print_inlining_stream. 4506 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 4507 print_inlining_reset(); 4508 } 4509 4510 void Compile::print_inlining_push() { 4511 // Add new buffer to the _print_inlining_list at current position 4512 _print_inlining_idx++; 4513 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer()); 4514 } 4515 4516 Compile::PrintInliningBuffer* Compile::print_inlining_current() { 4517 return _print_inlining_list->at(_print_inlining_idx); 4518 } 4519 4520 void Compile::print_inlining_update(CallGenerator* cg) { 4521 if (print_inlining() || print_intrinsics()) { 4522 if (cg->is_late_inline()) { 4523 if (print_inlining_current()->cg() != cg && 4524 (print_inlining_current()->cg() != nullptr || 4525 print_inlining_current()->ss()->size() != 0)) { 4526 print_inlining_push(); 4527 } 4528 print_inlining_commit(); 4529 print_inlining_current()->set_cg(cg); 4530 } else { 4531 if (print_inlining_current()->cg() != nullptr) { 4532 print_inlining_push(); 4533 } 4534 print_inlining_commit(); 4535 } 4536 } 4537 } 4538 4539 void Compile::print_inlining_move_to(CallGenerator* cg) { 4540 // We resume inlining at a late inlining call site. Locate the 4541 // corresponding inlining buffer so that we can update it. 4542 if (print_inlining() || print_intrinsics()) { 4543 for (int i = 0; i < _print_inlining_list->length(); i++) { 4544 if (_print_inlining_list->at(i)->cg() == cg) { 4545 _print_inlining_idx = i; 4546 return; 4547 } 4548 } 4549 ShouldNotReachHere(); 4550 } 4551 } 4552 4553 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4554 if (print_inlining() || print_intrinsics()) { 4555 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4556 assert(print_inlining_current()->cg() == cg, "wrong entry"); 4557 // replace message with new message 4558 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer()); 4559 print_inlining_commit(); 4560 print_inlining_current()->set_cg(cg); 4561 } 4562 } 4563 4564 void Compile::print_inlining_assert_ready() { 4565 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data"); 4566 } 4567 4568 void Compile::process_print_inlining() { 4569 assert(_late_inlines.length() == 0, "not drained yet"); 4570 if (print_inlining() || print_intrinsics()) { 4571 ResourceMark rm; 4572 stringStream ss; 4573 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once."); 4574 for (int i = 0; i < _print_inlining_list->length(); i++) { 4575 PrintInliningBuffer* pib = _print_inlining_list->at(i); 4576 ss.print("%s", pib->ss()->freeze()); 4577 delete pib; 4578 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr)); 4579 } 4580 // Reset _print_inlining_list, it only contains destructed objects. 4581 // It is on the arena, so it will be freed when the arena is reset. 4582 _print_inlining_list = nullptr; 4583 // _print_inlining_stream won't be used anymore, either. 4584 print_inlining_reset(); 4585 size_t end = ss.size(); 4586 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4587 strncpy(_print_inlining_output, ss.freeze(), end+1); 4588 _print_inlining_output[end] = 0; 4589 } 4590 } 4591 4592 void Compile::dump_print_inlining() { 4593 if (_print_inlining_output != nullptr) { 4594 tty->print_raw(_print_inlining_output); 4595 } 4596 } 4597 4598 void Compile::log_late_inline(CallGenerator* cg) { 4599 if (log() != nullptr) { 4600 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4601 cg->unique_id()); 4602 JVMState* p = cg->call_node()->jvms(); 4603 while (p != nullptr) { 4604 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4605 p = p->caller(); 4606 } 4607 log()->tail("late_inline"); 4608 } 4609 } 4610 4611 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4612 log_late_inline(cg); 4613 if (log() != nullptr) { 4614 log()->inline_fail(msg); 4615 } 4616 } 4617 4618 void Compile::log_inline_id(CallGenerator* cg) { 4619 if (log() != nullptr) { 4620 // The LogCompilation tool needs a unique way to identify late 4621 // inline call sites. This id must be unique for this call site in 4622 // this compilation. Try to have it unique across compilations as 4623 // well because it can be convenient when grepping through the log 4624 // file. 4625 // Distinguish OSR compilations from others in case CICountOSR is 4626 // on. 4627 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4628 cg->set_unique_id(id); 4629 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4630 } 4631 } 4632 4633 void Compile::log_inline_failure(const char* msg) { 4634 if (C->log() != nullptr) { 4635 C->log()->inline_fail(msg); 4636 } 4637 } 4638 4639 4640 // Dump inlining replay data to the stream. 4641 // Don't change thread state and acquire any locks. 4642 void Compile::dump_inline_data(outputStream* out) { 4643 InlineTree* inl_tree = ilt(); 4644 if (inl_tree != nullptr) { 4645 out->print(" inline %d", inl_tree->count()); 4646 inl_tree->dump_replay_data(out); 4647 } 4648 } 4649 4650 void Compile::dump_inline_data_reduced(outputStream* out) { 4651 assert(ReplayReduce, ""); 4652 4653 InlineTree* inl_tree = ilt(); 4654 if (inl_tree == nullptr) { 4655 return; 4656 } 4657 // Enable iterative replay file reduction 4658 // Output "compile" lines for depth 1 subtrees, 4659 // simulating that those trees were compiled 4660 // instead of inlined. 4661 for (int i = 0; i < inl_tree->subtrees().length(); ++i) { 4662 InlineTree* sub = inl_tree->subtrees().at(i); 4663 if (sub->inline_level() != 1) { 4664 continue; 4665 } 4666 4667 ciMethod* method = sub->method(); 4668 int entry_bci = -1; 4669 int comp_level = env()->task()->comp_level(); 4670 out->print("compile "); 4671 method->dump_name_as_ascii(out); 4672 out->print(" %d %d", entry_bci, comp_level); 4673 out->print(" inline %d", sub->count()); 4674 sub->dump_replay_data(out, -1); 4675 out->cr(); 4676 } 4677 } 4678 4679 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4680 if (n1->Opcode() < n2->Opcode()) return -1; 4681 else if (n1->Opcode() > n2->Opcode()) return 1; 4682 4683 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4684 for (uint i = 1; i < n1->req(); i++) { 4685 if (n1->in(i) < n2->in(i)) return -1; 4686 else if (n1->in(i) > n2->in(i)) return 1; 4687 } 4688 4689 return 0; 4690 } 4691 4692 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4693 Node* n1 = *n1p; 4694 Node* n2 = *n2p; 4695 4696 return cmp_expensive_nodes(n1, n2); 4697 } 4698 4699 void Compile::sort_expensive_nodes() { 4700 if (!expensive_nodes_sorted()) { 4701 _expensive_nodes.sort(cmp_expensive_nodes); 4702 } 4703 } 4704 4705 bool Compile::expensive_nodes_sorted() const { 4706 for (int i = 1; i < _expensive_nodes.length(); i++) { 4707 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) { 4708 return false; 4709 } 4710 } 4711 return true; 4712 } 4713 4714 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4715 if (_expensive_nodes.length() == 0) { 4716 return false; 4717 } 4718 4719 assert(OptimizeExpensiveOps, "optimization off?"); 4720 4721 // Take this opportunity to remove dead nodes from the list 4722 int j = 0; 4723 for (int i = 0; i < _expensive_nodes.length(); i++) { 4724 Node* n = _expensive_nodes.at(i); 4725 if (!n->is_unreachable(igvn)) { 4726 assert(n->is_expensive(), "should be expensive"); 4727 _expensive_nodes.at_put(j, n); 4728 j++; 4729 } 4730 } 4731 _expensive_nodes.trunc_to(j); 4732 4733 // Then sort the list so that similar nodes are next to each other 4734 // and check for at least two nodes of identical kind with same data 4735 // inputs. 4736 sort_expensive_nodes(); 4737 4738 for (int i = 0; i < _expensive_nodes.length()-1; i++) { 4739 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) { 4740 return true; 4741 } 4742 } 4743 4744 return false; 4745 } 4746 4747 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4748 if (_expensive_nodes.length() == 0) { 4749 return; 4750 } 4751 4752 assert(OptimizeExpensiveOps, "optimization off?"); 4753 4754 // Sort to bring similar nodes next to each other and clear the 4755 // control input of nodes for which there's only a single copy. 4756 sort_expensive_nodes(); 4757 4758 int j = 0; 4759 int identical = 0; 4760 int i = 0; 4761 bool modified = false; 4762 for (; i < _expensive_nodes.length()-1; i++) { 4763 assert(j <= i, "can't write beyond current index"); 4764 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) { 4765 identical++; 4766 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4767 continue; 4768 } 4769 if (identical > 0) { 4770 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4771 identical = 0; 4772 } else { 4773 Node* n = _expensive_nodes.at(i); 4774 igvn.replace_input_of(n, 0, nullptr); 4775 igvn.hash_insert(n); 4776 modified = true; 4777 } 4778 } 4779 if (identical > 0) { 4780 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4781 } else if (_expensive_nodes.length() >= 1) { 4782 Node* n = _expensive_nodes.at(i); 4783 igvn.replace_input_of(n, 0, nullptr); 4784 igvn.hash_insert(n); 4785 modified = true; 4786 } 4787 _expensive_nodes.trunc_to(j); 4788 if (modified) { 4789 igvn.optimize(); 4790 } 4791 } 4792 4793 void Compile::add_expensive_node(Node * n) { 4794 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list"); 4795 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4796 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4797 if (OptimizeExpensiveOps) { 4798 _expensive_nodes.append(n); 4799 } else { 4800 // Clear control input and let IGVN optimize expensive nodes if 4801 // OptimizeExpensiveOps is off. 4802 n->set_req(0, nullptr); 4803 } 4804 } 4805 4806 /** 4807 * Track coarsened Lock and Unlock nodes. 4808 */ 4809 4810 class Lock_List : public Node_List { 4811 uint _origin_cnt; 4812 public: 4813 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {} 4814 uint origin_cnt() const { return _origin_cnt; } 4815 }; 4816 4817 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) { 4818 int length = locks.length(); 4819 if (length > 0) { 4820 // Have to keep this list until locks elimination during Macro nodes elimination. 4821 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length); 4822 for (int i = 0; i < length; i++) { 4823 AbstractLockNode* lock = locks.at(i); 4824 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx); 4825 locks_list->push(lock); 4826 } 4827 _coarsened_locks.append(locks_list); 4828 } 4829 } 4830 4831 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) { 4832 int count = coarsened_count(); 4833 for (int i = 0; i < count; i++) { 4834 Node_List* locks_list = _coarsened_locks.at(i); 4835 for (uint j = 0; j < locks_list->size(); j++) { 4836 Node* lock = locks_list->at(j); 4837 assert(lock->is_AbstractLock(), "sanity"); 4838 if (!useful.member(lock)) { 4839 locks_list->yank(lock); 4840 } 4841 } 4842 } 4843 } 4844 4845 void Compile::remove_coarsened_lock(Node* n) { 4846 if (n->is_AbstractLock()) { 4847 int count = coarsened_count(); 4848 for (int i = 0; i < count; i++) { 4849 Node_List* locks_list = _coarsened_locks.at(i); 4850 locks_list->yank(n); 4851 } 4852 } 4853 } 4854 4855 bool Compile::coarsened_locks_consistent() { 4856 int count = coarsened_count(); 4857 for (int i = 0; i < count; i++) { 4858 bool unbalanced = false; 4859 bool modified = false; // track locks kind modifications 4860 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i); 4861 uint size = locks_list->size(); 4862 if (size == 0) { 4863 unbalanced = false; // All locks were eliminated - good 4864 } else if (size != locks_list->origin_cnt()) { 4865 unbalanced = true; // Some locks were removed from list 4866 } else { 4867 for (uint j = 0; j < size; j++) { 4868 Node* lock = locks_list->at(j); 4869 // All nodes in group should have the same state (modified or not) 4870 if (!lock->as_AbstractLock()->is_coarsened()) { 4871 if (j == 0) { 4872 // first on list was modified, the rest should be too for consistency 4873 modified = true; 4874 } else if (!modified) { 4875 // this lock was modified but previous locks on the list were not 4876 unbalanced = true; 4877 break; 4878 } 4879 } else if (modified) { 4880 // previous locks on list were modified but not this lock 4881 unbalanced = true; 4882 break; 4883 } 4884 } 4885 } 4886 if (unbalanced) { 4887 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified 4888 #ifdef ASSERT 4889 if (PrintEliminateLocks) { 4890 tty->print_cr("=== unbalanced coarsened locks ==="); 4891 for (uint l = 0; l < size; l++) { 4892 locks_list->at(l)->dump(); 4893 } 4894 } 4895 #endif 4896 record_failure(C2Compiler::retry_no_locks_coarsening()); 4897 return false; 4898 } 4899 } 4900 return true; 4901 } 4902 4903 /** 4904 * Remove the speculative part of types and clean up the graph 4905 */ 4906 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4907 if (UseTypeSpeculation) { 4908 Unique_Node_List worklist; 4909 worklist.push(root()); 4910 int modified = 0; 4911 // Go over all type nodes that carry a speculative type, drop the 4912 // speculative part of the type and enqueue the node for an igvn 4913 // which may optimize it out. 4914 for (uint next = 0; next < worklist.size(); ++next) { 4915 Node *n = worklist.at(next); 4916 if (n->is_Type()) { 4917 TypeNode* tn = n->as_Type(); 4918 const Type* t = tn->type(); 4919 const Type* t_no_spec = t->remove_speculative(); 4920 if (t_no_spec != t) { 4921 bool in_hash = igvn.hash_delete(n); 4922 #ifdef ASSERT 4923 if (!in_hash) { 4924 tty->print_cr("current graph:"); 4925 n->dump_bfs(MaxNodeLimit, nullptr, "S$"); 4926 tty->cr(); 4927 tty->print_cr("erroneous node:"); 4928 n->dump(); 4929 assert(false, "node should be in igvn hash table"); 4930 } 4931 #endif 4932 tn->set_type(t_no_spec); 4933 igvn.hash_insert(n); 4934 igvn._worklist.push(n); // give it a chance to go away 4935 modified++; 4936 } 4937 } 4938 // Iterate over outs - endless loops is unreachable from below 4939 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4940 Node *m = n->fast_out(i); 4941 if (not_a_node(m)) { 4942 continue; 4943 } 4944 worklist.push(m); 4945 } 4946 } 4947 // Drop the speculative part of all types in the igvn's type table 4948 igvn.remove_speculative_types(); 4949 if (modified > 0) { 4950 igvn.optimize(); 4951 if (failing()) return; 4952 } 4953 #ifdef ASSERT 4954 // Verify that after the IGVN is over no speculative type has resurfaced 4955 worklist.clear(); 4956 worklist.push(root()); 4957 for (uint next = 0; next < worklist.size(); ++next) { 4958 Node *n = worklist.at(next); 4959 const Type* t = igvn.type_or_null(n); 4960 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types"); 4961 if (n->is_Type()) { 4962 t = n->as_Type()->type(); 4963 assert(t == t->remove_speculative(), "no more speculative types"); 4964 } 4965 // Iterate over outs - endless loops is unreachable from below 4966 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4967 Node *m = n->fast_out(i); 4968 if (not_a_node(m)) { 4969 continue; 4970 } 4971 worklist.push(m); 4972 } 4973 } 4974 igvn.check_no_speculative_types(); 4975 #endif 4976 } 4977 } 4978 4979 // Auxiliary methods to support randomized stressing/fuzzing. 4980 4981 int Compile::random() { 4982 _stress_seed = os::next_random(_stress_seed); 4983 return static_cast<int>(_stress_seed); 4984 } 4985 4986 // This method can be called the arbitrary number of times, with current count 4987 // as the argument. The logic allows selecting a single candidate from the 4988 // running list of candidates as follows: 4989 // int count = 0; 4990 // Cand* selected = null; 4991 // while(cand = cand->next()) { 4992 // if (randomized_select(++count)) { 4993 // selected = cand; 4994 // } 4995 // } 4996 // 4997 // Including count equalizes the chances any candidate is "selected". 4998 // This is useful when we don't have the complete list of candidates to choose 4999 // from uniformly. In this case, we need to adjust the randomicity of the 5000 // selection, or else we will end up biasing the selection towards the latter 5001 // candidates. 5002 // 5003 // Quick back-envelope calculation shows that for the list of n candidates 5004 // the equal probability for the candidate to persist as "best" can be 5005 // achieved by replacing it with "next" k-th candidate with the probability 5006 // of 1/k. It can be easily shown that by the end of the run, the 5007 // probability for any candidate is converged to 1/n, thus giving the 5008 // uniform distribution among all the candidates. 5009 // 5010 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 5011 #define RANDOMIZED_DOMAIN_POW 29 5012 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 5013 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 5014 bool Compile::randomized_select(int count) { 5015 assert(count > 0, "only positive"); 5016 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 5017 } 5018 5019 CloneMap& Compile::clone_map() { return _clone_map; } 5020 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 5021 5022 void NodeCloneInfo::dump_on(outputStream* st) const { 5023 st->print(" {%d:%d} ", idx(), gen()); 5024 } 5025 5026 void CloneMap::clone(Node* old, Node* nnn, int gen) { 5027 uint64_t val = value(old->_idx); 5028 NodeCloneInfo cio(val); 5029 assert(val != 0, "old node should be in the map"); 5030 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 5031 insert(nnn->_idx, cin.get()); 5032 #ifndef PRODUCT 5033 if (is_debug()) { 5034 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 5035 } 5036 #endif 5037 } 5038 5039 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 5040 NodeCloneInfo cio(value(old->_idx)); 5041 if (cio.get() == 0) { 5042 cio.set(old->_idx, 0); 5043 insert(old->_idx, cio.get()); 5044 #ifndef PRODUCT 5045 if (is_debug()) { 5046 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 5047 } 5048 #endif 5049 } 5050 clone(old, nnn, gen); 5051 } 5052 5053 int CloneMap::max_gen() const { 5054 int g = 0; 5055 DictI di(_dict); 5056 for(; di.test(); ++di) { 5057 int t = gen(di._key); 5058 if (g < t) { 5059 g = t; 5060 #ifndef PRODUCT 5061 if (is_debug()) { 5062 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 5063 } 5064 #endif 5065 } 5066 } 5067 return g; 5068 } 5069 5070 void CloneMap::dump(node_idx_t key, outputStream* st) const { 5071 uint64_t val = value(key); 5072 if (val != 0) { 5073 NodeCloneInfo ni(val); 5074 ni.dump_on(st); 5075 } 5076 } 5077 5078 // Move Allocate nodes to the start of the list 5079 void Compile::sort_macro_nodes() { 5080 int count = macro_count(); 5081 int allocates = 0; 5082 for (int i = 0; i < count; i++) { 5083 Node* n = macro_node(i); 5084 if (n->is_Allocate()) { 5085 if (i != allocates) { 5086 Node* tmp = macro_node(allocates); 5087 _macro_nodes.at_put(allocates, n); 5088 _macro_nodes.at_put(i, tmp); 5089 } 5090 allocates++; 5091 } 5092 } 5093 } 5094 5095 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) { 5096 if (failing()) { return; } 5097 EventCompilerPhase event; 5098 if (event.should_commit()) { 5099 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 5100 } 5101 #ifndef PRODUCT 5102 ResourceMark rm; 5103 stringStream ss; 5104 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt)); 5105 if (n != nullptr) { 5106 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]); 5107 } 5108 5109 const char* name = ss.as_string(); 5110 if (should_print_igv(level)) { 5111 _igv_printer->print_method(name, level); 5112 } 5113 if (should_print_phase(cpt)) { 5114 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt)); 5115 } 5116 #endif 5117 C->_latest_stage_start_counter.stamp(); 5118 } 5119 5120 // Only used from CompileWrapper 5121 void Compile::begin_method() { 5122 #ifndef PRODUCT 5123 if (_method != nullptr && should_print_igv(1)) { 5124 _igv_printer->begin_method(); 5125 } 5126 #endif 5127 C->_latest_stage_start_counter.stamp(); 5128 } 5129 5130 // Only used from CompileWrapper 5131 void Compile::end_method() { 5132 EventCompilerPhase event; 5133 if (event.should_commit()) { 5134 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1); 5135 } 5136 5137 #ifndef PRODUCT 5138 if (_method != nullptr && should_print_igv(1)) { 5139 _igv_printer->end_method(); 5140 } 5141 #endif 5142 } 5143 5144 bool Compile::should_print_phase(CompilerPhaseType cpt) { 5145 #ifndef PRODUCT 5146 if (_directive->should_print_phase(cpt)) { 5147 return true; 5148 } 5149 #endif 5150 return false; 5151 } 5152 5153 bool Compile::should_print_igv(const int level) { 5154 #ifndef PRODUCT 5155 if (PrintIdealGraphLevel < 0) { // disabled by the user 5156 return false; 5157 } 5158 5159 bool need = directive()->IGVPrintLevelOption >= level; 5160 if (need && !_igv_printer) { 5161 _igv_printer = IdealGraphPrinter::printer(); 5162 _igv_printer->set_compile(this); 5163 } 5164 return need; 5165 #else 5166 return false; 5167 #endif 5168 } 5169 5170 #ifndef PRODUCT 5171 IdealGraphPrinter* Compile::_debug_file_printer = nullptr; 5172 IdealGraphPrinter* Compile::_debug_network_printer = nullptr; 5173 5174 // Called from debugger. Prints method to the default file with the default phase name. 5175 // This works regardless of any Ideal Graph Visualizer flags set or not. 5176 void igv_print() { 5177 Compile::current()->igv_print_method_to_file(); 5178 } 5179 5180 // Same as igv_print() above but with a specified phase name. 5181 void igv_print(const char* phase_name) { 5182 Compile::current()->igv_print_method_to_file(phase_name); 5183 } 5184 5185 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 5186 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 5187 // This works regardless of any Ideal Graph Visualizer flags set or not. 5188 void igv_print(bool network) { 5189 if (network) { 5190 Compile::current()->igv_print_method_to_network(); 5191 } else { 5192 Compile::current()->igv_print_method_to_file(); 5193 } 5194 } 5195 5196 // Same as igv_print(bool network) above but with a specified phase name. 5197 void igv_print(bool network, const char* phase_name) { 5198 if (network) { 5199 Compile::current()->igv_print_method_to_network(phase_name); 5200 } else { 5201 Compile::current()->igv_print_method_to_file(phase_name); 5202 } 5203 } 5204 5205 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 5206 void igv_print_default() { 5207 Compile::current()->print_method(PHASE_DEBUG, 0); 5208 } 5209 5210 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 5211 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 5212 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 5213 void igv_append() { 5214 Compile::current()->igv_print_method_to_file("Debug", true); 5215 } 5216 5217 // Same as igv_append() above but with a specified phase name. 5218 void igv_append(const char* phase_name) { 5219 Compile::current()->igv_print_method_to_file(phase_name, true); 5220 } 5221 5222 void Compile::igv_print_method_to_file(const char* phase_name, bool append) { 5223 const char* file_name = "custom_debug.xml"; 5224 if (_debug_file_printer == nullptr) { 5225 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 5226 } else { 5227 _debug_file_printer->update_compiled_method(C->method()); 5228 } 5229 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 5230 _debug_file_printer->print(phase_name, (Node*)C->root()); 5231 } 5232 5233 void Compile::igv_print_method_to_network(const char* phase_name) { 5234 if (_debug_network_printer == nullptr) { 5235 _debug_network_printer = new IdealGraphPrinter(C); 5236 } else { 5237 _debug_network_printer->update_compiled_method(C->method()); 5238 } 5239 tty->print_cr("Method printed over network stream to IGV"); 5240 _debug_network_printer->print(phase_name, (Node*)C->root()); 5241 } 5242 #endif 5243 5244 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 5245 if (type != nullptr && phase->type(value)->higher_equal(type)) { 5246 return value; 5247 } 5248 Node* result = nullptr; 5249 if (bt == T_BYTE) { 5250 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 5251 result = new RShiftINode(result, phase->intcon(24)); 5252 } else if (bt == T_BOOLEAN) { 5253 result = new AndINode(value, phase->intcon(0xFF)); 5254 } else if (bt == T_CHAR) { 5255 result = new AndINode(value,phase->intcon(0xFFFF)); 5256 } else { 5257 assert(bt == T_SHORT, "unexpected narrow type"); 5258 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 5259 result = new RShiftINode(result, phase->intcon(16)); 5260 } 5261 if (transform_res) { 5262 result = phase->transform(result); 5263 } 5264 return result; 5265 } 5266 5267 void Compile::record_method_not_compilable_oom() { 5268 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit()); 5269 }