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 xtty->print("%s", ss.as_string()); // print to tty would use xml escape encoding 594 xtty->tail("ideal"); 595 } else { 596 tty->print("%s", ss.as_string()); 597 } 598 } 599 #endif 600 601 // ============================================================================ 602 //------------------------------Compile standard------------------------------- 603 604 // Compile a method. entry_bci is -1 for normal compilations and indicates 605 // the continuation bci for on stack replacement. 606 607 608 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci, 609 Options options, DirectiveSet* directive) 610 : Phase(Compiler), 611 _compile_id(ci_env->compile_id()), 612 _options(options), 613 _method(target), 614 _entry_bci(osr_bci), 615 _ilt(nullptr), 616 _stub_function(nullptr), 617 _stub_name(nullptr), 618 _stub_entry_point(nullptr), 619 _max_node_limit(MaxNodeLimit), 620 _post_loop_opts_phase(false), 621 _inlining_progress(false), 622 _inlining_incrementally(false), 623 _do_cleanup(false), 624 _has_reserved_stack_access(target->has_reserved_stack_access()), 625 #ifndef PRODUCT 626 _igv_idx(0), 627 _trace_opto_output(directive->TraceOptoOutputOption), 628 #endif 629 _has_method_handle_invokes(false), 630 _clinit_barrier_on_entry(false), 631 _stress_seed(0), 632 _comp_arena(mtCompiler), 633 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 634 _env(ci_env), 635 _directive(directive), 636 _log(ci_env->log()), 637 _failure_reason(nullptr), 638 _intrinsics (comp_arena(), 0, 0, nullptr), 639 _macro_nodes (comp_arena(), 8, 0, nullptr), 640 _parse_predicates (comp_arena(), 8, 0, nullptr), 641 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr), 642 _expensive_nodes (comp_arena(), 8, 0, nullptr), 643 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr), 644 _unstable_if_traps (comp_arena(), 8, 0, nullptr), 645 _coarsened_locks (comp_arena(), 8, 0, nullptr), 646 _congraph(nullptr), 647 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 648 _unique(0), 649 _dead_node_count(0), 650 _dead_node_list(comp_arena()), 651 _node_arena_one(mtCompiler, Arena::Tag::tag_node), 652 _node_arena_two(mtCompiler, Arena::Tag::tag_node), 653 _node_arena(&_node_arena_one), 654 _mach_constant_base_node(nullptr), 655 _Compile_types(mtCompiler), 656 _initial_gvn(nullptr), 657 _igvn_worklist(nullptr), 658 _types(nullptr), 659 _node_hash(nullptr), 660 _late_inlines(comp_arena(), 2, 0, nullptr), 661 _string_late_inlines(comp_arena(), 2, 0, nullptr), 662 _boxing_late_inlines(comp_arena(), 2, 0, nullptr), 663 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr), 664 _late_inlines_pos(0), 665 _number_of_mh_late_inlines(0), 666 _oom(false), 667 _print_inlining_stream(new (mtCompiler) stringStream()), 668 _print_inlining_list(nullptr), 669 _print_inlining_idx(0), 670 _print_inlining_output(nullptr), 671 _replay_inline_data(nullptr), 672 _java_calls(0), 673 _inner_loops(0), 674 _interpreter_frame_size(0), 675 _output(nullptr) 676 #ifndef PRODUCT 677 , _in_dump_cnt(0) 678 #endif 679 { 680 C = this; 681 CompileWrapper cw(this); 682 683 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 684 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false); 685 686 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 687 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 688 // We can always print a disassembly, either abstract (hex dump) or 689 // with the help of a suitable hsdis library. Thus, we should not 690 // couple print_assembly and print_opto_assembly controls. 691 // But: always print opto and regular assembly on compile command 'print'. 692 bool print_assembly = directive->PrintAssemblyOption; 693 set_print_assembly(print_opto_assembly || print_assembly); 694 #else 695 set_print_assembly(false); // must initialize. 696 #endif 697 698 #ifndef PRODUCT 699 set_parsed_irreducible_loop(false); 700 #endif 701 702 if (directive->ReplayInlineOption) { 703 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 704 } 705 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 706 set_print_intrinsics(directive->PrintIntrinsicsOption); 707 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 708 709 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 710 // Make sure the method being compiled gets its own MDO, 711 // so we can at least track the decompile_count(). 712 // Need MDO to record RTM code generation state. 713 method()->ensure_method_data(); 714 } 715 716 Init(/*do_aliasing=*/ true); 717 718 print_compile_messages(); 719 720 _ilt = InlineTree::build_inline_tree_root(); 721 722 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 723 assert(num_alias_types() >= AliasIdxRaw, ""); 724 725 #define MINIMUM_NODE_HASH 1023 726 727 // GVN that will be run immediately on new nodes 728 uint estimated_size = method()->code_size()*4+64; 729 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 730 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 731 _types = new (comp_arena()) Type_Array(comp_arena()); 732 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size); 733 PhaseGVN gvn; 734 set_initial_gvn(&gvn); 735 736 print_inlining_init(); 737 { // Scope for timing the parser 738 TracePhase tp("parse", &timers[_t_parser]); 739 740 // Put top into the hash table ASAP. 741 initial_gvn()->transform_no_reclaim(top()); 742 743 // Set up tf(), start(), and find a CallGenerator. 744 CallGenerator* cg = nullptr; 745 if (is_osr_compilation()) { 746 const TypeTuple *domain = StartOSRNode::osr_domain(); 747 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 748 init_tf(TypeFunc::make(domain, range)); 749 StartNode* s = new StartOSRNode(root(), domain); 750 initial_gvn()->set_type_bottom(s); 751 init_start(s); 752 cg = CallGenerator::for_osr(method(), entry_bci()); 753 } else { 754 // Normal case. 755 init_tf(TypeFunc::make(method())); 756 StartNode* s = new StartNode(root(), tf()->domain()); 757 initial_gvn()->set_type_bottom(s); 758 init_start(s); 759 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 760 // With java.lang.ref.reference.get() we must go through the 761 // intrinsic - even when get() is the root 762 // method of the compile - so that, if necessary, the value in 763 // the referent field of the reference object gets recorded by 764 // the pre-barrier code. 765 cg = find_intrinsic(method(), false); 766 } 767 if (cg == nullptr) { 768 float past_uses = method()->interpreter_invocation_count(); 769 float expected_uses = past_uses; 770 cg = CallGenerator::for_inline(method(), expected_uses); 771 } 772 } 773 if (failing()) return; 774 if (cg == nullptr) { 775 const char* reason = InlineTree::check_can_parse(method()); 776 assert(reason != nullptr, "expect reason for parse failure"); 777 stringStream ss; 778 ss.print("cannot parse method: %s", reason); 779 record_method_not_compilable(ss.as_string()); 780 return; 781 } 782 783 gvn.set_type(root(), root()->bottom_type()); 784 785 JVMState* jvms = build_start_state(start(), tf()); 786 if ((jvms = cg->generate(jvms)) == nullptr) { 787 assert(failure_reason() != nullptr, "expect reason for parse failure"); 788 stringStream ss; 789 ss.print("method parse failed: %s", failure_reason()); 790 record_method_not_compilable(ss.as_string()); 791 return; 792 } 793 GraphKit kit(jvms); 794 795 if (!kit.stopped()) { 796 // Accept return values, and transfer control we know not where. 797 // This is done by a special, unique ReturnNode bound to root. 798 return_values(kit.jvms()); 799 } 800 801 if (kit.has_exceptions()) { 802 // Any exceptions that escape from this call must be rethrown 803 // to whatever caller is dynamically above us on the stack. 804 // This is done by a special, unique RethrowNode bound to root. 805 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 806 } 807 808 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 809 810 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 811 inline_string_calls(true); 812 } 813 814 if (failing()) return; 815 816 // Remove clutter produced by parsing. 817 if (!failing()) { 818 ResourceMark rm; 819 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 820 } 821 } 822 823 // Note: Large methods are capped off in do_one_bytecode(). 824 if (failing()) return; 825 826 // After parsing, node notes are no longer automagic. 827 // They must be propagated by register_new_node_with_optimizer(), 828 // clone(), or the like. 829 set_default_node_notes(nullptr); 830 831 #ifndef PRODUCT 832 if (should_print_igv(1)) { 833 _igv_printer->print_inlining(); 834 } 835 #endif 836 837 if (failing()) return; 838 NOT_PRODUCT( verify_graph_edges(); ) 839 840 // If any phase is randomized for stress testing, seed random number 841 // generation and log the seed for repeatability. 842 if (StressLCM || StressGCM || StressIGVN || StressCCP || StressIncrementalInlining) { 843 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) { 844 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds()); 845 FLAG_SET_ERGO(StressSeed, _stress_seed); 846 } else { 847 _stress_seed = StressSeed; 848 } 849 if (_log != nullptr) { 850 _log->elem("stress_test seed='%u'", _stress_seed); 851 } 852 } 853 854 // Now optimize 855 Optimize(); 856 if (failing()) return; 857 NOT_PRODUCT( verify_graph_edges(); ) 858 859 #ifndef PRODUCT 860 if (should_print_ideal()) { 861 print_ideal_ir("print_ideal"); 862 } 863 #endif 864 865 #ifdef ASSERT 866 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 867 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 868 #endif 869 870 // Dump compilation data to replay it. 871 if (directive->DumpReplayOption) { 872 env()->dump_replay_data(_compile_id); 873 } 874 if (directive->DumpInlineOption && (ilt() != nullptr)) { 875 env()->dump_inline_data(_compile_id); 876 } 877 878 // Now that we know the size of all the monitors we can add a fixed slot 879 // for the original deopt pc. 880 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size); 881 set_fixed_slots(next_slot); 882 883 // Compute when to use implicit null checks. Used by matching trap based 884 // nodes and NullCheck optimization. 885 set_allowed_deopt_reasons(); 886 887 // Now generate code 888 Code_Gen(); 889 } 890 891 //------------------------------Compile---------------------------------------- 892 // Compile a runtime stub 893 Compile::Compile( ciEnv* ci_env, 894 TypeFunc_generator generator, 895 address stub_function, 896 const char *stub_name, 897 int is_fancy_jump, 898 bool pass_tls, 899 bool return_pc, 900 DirectiveSet* directive) 901 : Phase(Compiler), 902 _compile_id(0), 903 _options(Options::for_runtime_stub()), 904 _method(nullptr), 905 _entry_bci(InvocationEntryBci), 906 _stub_function(stub_function), 907 _stub_name(stub_name), 908 _stub_entry_point(nullptr), 909 _max_node_limit(MaxNodeLimit), 910 _post_loop_opts_phase(false), 911 _inlining_progress(false), 912 _inlining_incrementally(false), 913 _has_reserved_stack_access(false), 914 #ifndef PRODUCT 915 _igv_idx(0), 916 _trace_opto_output(directive->TraceOptoOutputOption), 917 #endif 918 _has_method_handle_invokes(false), 919 _clinit_barrier_on_entry(false), 920 _stress_seed(0), 921 _comp_arena(mtCompiler), 922 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 923 _env(ci_env), 924 _directive(directive), 925 _log(ci_env->log()), 926 _failure_reason(nullptr), 927 _congraph(nullptr), 928 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 929 _unique(0), 930 _dead_node_count(0), 931 _dead_node_list(comp_arena()), 932 _node_arena_one(mtCompiler), 933 _node_arena_two(mtCompiler), 934 _node_arena(&_node_arena_one), 935 _mach_constant_base_node(nullptr), 936 _Compile_types(mtCompiler), 937 _initial_gvn(nullptr), 938 _igvn_worklist(nullptr), 939 _types(nullptr), 940 _node_hash(nullptr), 941 _number_of_mh_late_inlines(0), 942 _oom(false), 943 _print_inlining_stream(new (mtCompiler) stringStream()), 944 _print_inlining_list(nullptr), 945 _print_inlining_idx(0), 946 _print_inlining_output(nullptr), 947 _replay_inline_data(nullptr), 948 _java_calls(0), 949 _inner_loops(0), 950 _interpreter_frame_size(0), 951 _output(nullptr), 952 #ifndef PRODUCT 953 _in_dump_cnt(0), 954 #endif 955 _allowed_reasons(0) { 956 C = this; 957 958 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false); 959 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false); 960 961 #ifndef PRODUCT 962 set_print_assembly(PrintFrameConverterAssembly); 963 set_parsed_irreducible_loop(false); 964 #else 965 set_print_assembly(false); // Must initialize. 966 #endif 967 set_has_irreducible_loop(false); // no loops 968 969 CompileWrapper cw(this); 970 Init(/*do_aliasing=*/ false); 971 init_tf((*generator)()); 972 973 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 974 _types = new (comp_arena()) Type_Array(comp_arena()); 975 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255); 976 { 977 PhaseGVN gvn; 978 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 979 gvn.transform_no_reclaim(top()); 980 981 GraphKit kit; 982 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 983 } 984 985 NOT_PRODUCT( verify_graph_edges(); ) 986 987 Code_Gen(); 988 } 989 990 //------------------------------Init------------------------------------------- 991 // Prepare for a single compilation 992 void Compile::Init(bool aliasing) { 993 _do_aliasing = aliasing; 994 _unique = 0; 995 _regalloc = nullptr; 996 997 _tf = nullptr; // filled in later 998 _top = nullptr; // cached later 999 _matcher = nullptr; // filled in later 1000 _cfg = nullptr; // filled in later 1001 1002 IA32_ONLY( set_24_bit_selection_and_mode(true, false); ) 1003 1004 _node_note_array = nullptr; 1005 _default_node_notes = nullptr; 1006 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize() 1007 1008 _immutable_memory = nullptr; // filled in at first inquiry 1009 1010 #ifdef ASSERT 1011 _phase_optimize_finished = false; 1012 _exception_backedge = false; 1013 _type_verify = nullptr; 1014 #endif 1015 1016 // Globally visible Nodes 1017 // First set TOP to null to give safe behavior during creation of RootNode 1018 set_cached_top_node(nullptr); 1019 set_root(new RootNode()); 1020 // Now that you have a Root to point to, create the real TOP 1021 set_cached_top_node( new ConNode(Type::TOP) ); 1022 set_recent_alloc(nullptr, nullptr); 1023 1024 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1025 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1026 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1027 env()->set_dependencies(new Dependencies(env())); 1028 1029 _fixed_slots = 0; 1030 set_has_split_ifs(false); 1031 set_has_loops(false); // first approximation 1032 set_has_stringbuilder(false); 1033 set_has_boxed_value(false); 1034 _trap_can_recompile = false; // no traps emitted yet 1035 _major_progress = true; // start out assuming good things will happen 1036 set_has_unsafe_access(false); 1037 set_max_vector_size(0); 1038 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1039 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1040 set_decompile_count(0); 1041 1042 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1043 _loop_opts_cnt = LoopOptsCount; 1044 set_do_inlining(Inline); 1045 set_max_inline_size(MaxInlineSize); 1046 set_freq_inline_size(FreqInlineSize); 1047 set_do_scheduling(OptoScheduling); 1048 1049 set_do_vector_loop(false); 1050 set_has_monitors(false); 1051 1052 if (AllowVectorizeOnDemand) { 1053 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1054 set_do_vector_loop(true); 1055 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());}) 1056 } else if (has_method() && method()->name() != 0 && 1057 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1058 set_do_vector_loop(true); 1059 } 1060 } 1061 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1062 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());}) 1063 1064 set_rtm_state(NoRTM); // No RTM lock eliding by default 1065 _max_node_limit = _directive->MaxNodeLimitOption; 1066 1067 #if INCLUDE_RTM_OPT 1068 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != nullptr)) { 1069 int rtm_state = method()->method_data()->rtm_state(); 1070 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) { 1071 // Don't generate RTM lock eliding code. 1072 set_rtm_state(NoRTM); 1073 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1074 // Generate RTM lock eliding code without abort ratio calculation code. 1075 set_rtm_state(UseRTM); 1076 } else if (UseRTMDeopt) { 1077 // Generate RTM lock eliding code and include abort ratio calculation 1078 // code if UseRTMDeopt is on. 1079 set_rtm_state(ProfileRTM); 1080 } 1081 } 1082 #endif 1083 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1084 set_clinit_barrier_on_entry(true); 1085 } 1086 if (debug_info()->recording_non_safepoints()) { 1087 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1088 (comp_arena(), 8, 0, nullptr)); 1089 set_default_node_notes(Node_Notes::make(this)); 1090 } 1091 1092 const int grow_ats = 16; 1093 _max_alias_types = grow_ats; 1094 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1095 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1096 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1097 { 1098 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1099 } 1100 // Initialize the first few types. 1101 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr); 1102 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1103 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1104 _num_alias_types = AliasIdxRaw+1; 1105 // Zero out the alias type cache. 1106 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1107 // A null adr_type hits in the cache right away. Preload the right answer. 1108 probe_alias_cache(nullptr)->_index = AliasIdxTop; 1109 } 1110 1111 //---------------------------init_start---------------------------------------- 1112 // Install the StartNode on this compile object. 1113 void Compile::init_start(StartNode* s) { 1114 if (failing()) 1115 return; // already failing 1116 assert(s == start(), ""); 1117 } 1118 1119 /** 1120 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1121 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1122 * the ideal graph. 1123 */ 1124 StartNode* Compile::start() const { 1125 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1126 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1127 Node* start = root()->fast_out(i); 1128 if (start->is_Start()) { 1129 return start->as_Start(); 1130 } 1131 } 1132 fatal("Did not find Start node!"); 1133 return nullptr; 1134 } 1135 1136 //-------------------------------immutable_memory------------------------------------- 1137 // Access immutable memory 1138 Node* Compile::immutable_memory() { 1139 if (_immutable_memory != nullptr) { 1140 return _immutable_memory; 1141 } 1142 StartNode* s = start(); 1143 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1144 Node *p = s->fast_out(i); 1145 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1146 _immutable_memory = p; 1147 return _immutable_memory; 1148 } 1149 } 1150 ShouldNotReachHere(); 1151 return nullptr; 1152 } 1153 1154 //----------------------set_cached_top_node------------------------------------ 1155 // Install the cached top node, and make sure Node::is_top works correctly. 1156 void Compile::set_cached_top_node(Node* tn) { 1157 if (tn != nullptr) verify_top(tn); 1158 Node* old_top = _top; 1159 _top = tn; 1160 // Calling Node::setup_is_top allows the nodes the chance to adjust 1161 // their _out arrays. 1162 if (_top != nullptr) _top->setup_is_top(); 1163 if (old_top != nullptr) old_top->setup_is_top(); 1164 assert(_top == nullptr || top()->is_top(), ""); 1165 } 1166 1167 #ifdef ASSERT 1168 uint Compile::count_live_nodes_by_graph_walk() { 1169 Unique_Node_List useful(comp_arena()); 1170 // Get useful node list by walking the graph. 1171 identify_useful_nodes(useful); 1172 return useful.size(); 1173 } 1174 1175 void Compile::print_missing_nodes() { 1176 1177 // Return if CompileLog is null and PrintIdealNodeCount is false. 1178 if ((_log == nullptr) && (! PrintIdealNodeCount)) { 1179 return; 1180 } 1181 1182 // This is an expensive function. It is executed only when the user 1183 // specifies VerifyIdealNodeCount option or otherwise knows the 1184 // additional work that needs to be done to identify reachable nodes 1185 // by walking the flow graph and find the missing ones using 1186 // _dead_node_list. 1187 1188 Unique_Node_List useful(comp_arena()); 1189 // Get useful node list by walking the graph. 1190 identify_useful_nodes(useful); 1191 1192 uint l_nodes = C->live_nodes(); 1193 uint l_nodes_by_walk = useful.size(); 1194 1195 if (l_nodes != l_nodes_by_walk) { 1196 if (_log != nullptr) { 1197 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1198 _log->stamp(); 1199 _log->end_head(); 1200 } 1201 VectorSet& useful_member_set = useful.member_set(); 1202 int last_idx = l_nodes_by_walk; 1203 for (int i = 0; i < last_idx; i++) { 1204 if (useful_member_set.test(i)) { 1205 if (_dead_node_list.test(i)) { 1206 if (_log != nullptr) { 1207 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1208 } 1209 if (PrintIdealNodeCount) { 1210 // Print the log message to tty 1211 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1212 useful.at(i)->dump(); 1213 } 1214 } 1215 } 1216 else if (! _dead_node_list.test(i)) { 1217 if (_log != nullptr) { 1218 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1219 } 1220 if (PrintIdealNodeCount) { 1221 // Print the log message to tty 1222 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1223 } 1224 } 1225 } 1226 if (_log != nullptr) { 1227 _log->tail("mismatched_nodes"); 1228 } 1229 } 1230 } 1231 void Compile::record_modified_node(Node* n) { 1232 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) { 1233 _modified_nodes->push(n); 1234 } 1235 } 1236 1237 void Compile::remove_modified_node(Node* n) { 1238 if (_modified_nodes != nullptr) { 1239 _modified_nodes->remove(n); 1240 } 1241 } 1242 #endif 1243 1244 #ifndef PRODUCT 1245 void Compile::verify_top(Node* tn) const { 1246 if (tn != nullptr) { 1247 assert(tn->is_Con(), "top node must be a constant"); 1248 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1249 assert(tn->in(0) != nullptr, "must have live top node"); 1250 } 1251 } 1252 #endif 1253 1254 1255 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1256 1257 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1258 guarantee(arr != nullptr, ""); 1259 int num_blocks = arr->length(); 1260 if (grow_by < num_blocks) grow_by = num_blocks; 1261 int num_notes = grow_by * _node_notes_block_size; 1262 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1263 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1264 while (num_notes > 0) { 1265 arr->append(notes); 1266 notes += _node_notes_block_size; 1267 num_notes -= _node_notes_block_size; 1268 } 1269 assert(num_notes == 0, "exact multiple, please"); 1270 } 1271 1272 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1273 if (source == nullptr || dest == nullptr) return false; 1274 1275 if (dest->is_Con()) 1276 return false; // Do not push debug info onto constants. 1277 1278 #ifdef ASSERT 1279 // Leave a bread crumb trail pointing to the original node: 1280 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) { 1281 dest->set_debug_orig(source); 1282 } 1283 #endif 1284 1285 if (node_note_array() == nullptr) 1286 return false; // Not collecting any notes now. 1287 1288 // This is a copy onto a pre-existing node, which may already have notes. 1289 // If both nodes have notes, do not overwrite any pre-existing notes. 1290 Node_Notes* source_notes = node_notes_at(source->_idx); 1291 if (source_notes == nullptr || source_notes->is_clear()) return false; 1292 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1293 if (dest_notes == nullptr || dest_notes->is_clear()) { 1294 return set_node_notes_at(dest->_idx, source_notes); 1295 } 1296 1297 Node_Notes merged_notes = (*source_notes); 1298 // The order of operations here ensures that dest notes will win... 1299 merged_notes.update_from(dest_notes); 1300 return set_node_notes_at(dest->_idx, &merged_notes); 1301 } 1302 1303 1304 //--------------------------allow_range_check_smearing------------------------- 1305 // Gating condition for coalescing similar range checks. 1306 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1307 // single covering check that is at least as strong as any of them. 1308 // If the optimization succeeds, the simplified (strengthened) range check 1309 // will always succeed. If it fails, we will deopt, and then give up 1310 // on the optimization. 1311 bool Compile::allow_range_check_smearing() const { 1312 // If this method has already thrown a range-check, 1313 // assume it was because we already tried range smearing 1314 // and it failed. 1315 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1316 return !already_trapped; 1317 } 1318 1319 1320 //------------------------------flatten_alias_type----------------------------- 1321 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1322 assert(do_aliasing(), "Aliasing should be enabled"); 1323 int offset = tj->offset(); 1324 TypePtr::PTR ptr = tj->ptr(); 1325 1326 // Known instance (scalarizable allocation) alias only with itself. 1327 bool is_known_inst = tj->isa_oopptr() != nullptr && 1328 tj->is_oopptr()->is_known_instance(); 1329 1330 // Process weird unsafe references. 1331 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1332 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops"); 1333 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1334 tj = TypeOopPtr::BOTTOM; 1335 ptr = tj->ptr(); 1336 offset = tj->offset(); 1337 } 1338 1339 // Array pointers need some flattening 1340 const TypeAryPtr* ta = tj->isa_aryptr(); 1341 if (ta && ta->is_stable()) { 1342 // Erase stability property for alias analysis. 1343 tj = ta = ta->cast_to_stable(false); 1344 } 1345 if( ta && is_known_inst ) { 1346 if ( offset != Type::OffsetBot && 1347 offset > arrayOopDesc::length_offset_in_bytes() ) { 1348 offset = Type::OffsetBot; // Flatten constant access into array body only 1349 tj = ta = ta-> 1350 remove_speculative()-> 1351 cast_to_ptr_type(ptr)-> 1352 with_offset(offset); 1353 } 1354 } else if (ta) { 1355 // For arrays indexed by constant indices, we flatten the alias 1356 // space to include all of the array body. Only the header, klass 1357 // and array length can be accessed un-aliased. 1358 if( offset != Type::OffsetBot ) { 1359 if( ta->const_oop() ) { // MethodData* or Method* 1360 offset = Type::OffsetBot; // Flatten constant access into array body 1361 tj = ta = ta-> 1362 remove_speculative()-> 1363 cast_to_ptr_type(ptr)-> 1364 cast_to_exactness(false)-> 1365 with_offset(offset); 1366 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1367 // range is OK as-is. 1368 tj = ta = TypeAryPtr::RANGE; 1369 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1370 tj = TypeInstPtr::KLASS; // all klass loads look alike 1371 ta = TypeAryPtr::RANGE; // generic ignored junk 1372 ptr = TypePtr::BotPTR; 1373 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1374 tj = TypeInstPtr::MARK; 1375 ta = TypeAryPtr::RANGE; // generic ignored junk 1376 ptr = TypePtr::BotPTR; 1377 } else { // Random constant offset into array body 1378 offset = Type::OffsetBot; // Flatten constant access into array body 1379 tj = ta = ta-> 1380 remove_speculative()-> 1381 cast_to_ptr_type(ptr)-> 1382 cast_to_exactness(false)-> 1383 with_offset(offset); 1384 } 1385 } 1386 // Arrays of fixed size alias with arrays of unknown size. 1387 if (ta->size() != TypeInt::POS) { 1388 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1389 tj = ta = ta-> 1390 remove_speculative()-> 1391 cast_to_ptr_type(ptr)-> 1392 with_ary(tary)-> 1393 cast_to_exactness(false); 1394 } 1395 // Arrays of known objects become arrays of unknown objects. 1396 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1397 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1398 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1399 } 1400 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1401 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1402 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1403 } 1404 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1405 // cannot be distinguished by bytecode alone. 1406 if (ta->elem() == TypeInt::BOOL) { 1407 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1408 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1409 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1410 } 1411 // During the 2nd round of IterGVN, NotNull castings are removed. 1412 // Make sure the Bottom and NotNull variants alias the same. 1413 // Also, make sure exact and non-exact variants alias the same. 1414 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) { 1415 tj = ta = ta-> 1416 remove_speculative()-> 1417 cast_to_ptr_type(TypePtr::BotPTR)-> 1418 cast_to_exactness(false)-> 1419 with_offset(offset); 1420 } 1421 } 1422 1423 // Oop pointers need some flattening 1424 const TypeInstPtr *to = tj->isa_instptr(); 1425 if (to && to != TypeOopPtr::BOTTOM) { 1426 ciInstanceKlass* ik = to->instance_klass(); 1427 if( ptr == TypePtr::Constant ) { 1428 if (ik != ciEnv::current()->Class_klass() || 1429 offset < ik->layout_helper_size_in_bytes()) { 1430 // No constant oop pointers (such as Strings); they alias with 1431 // unknown strings. 1432 assert(!is_known_inst, "not scalarizable allocation"); 1433 tj = to = to-> 1434 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1435 remove_speculative()-> 1436 cast_to_ptr_type(TypePtr::BotPTR)-> 1437 cast_to_exactness(false); 1438 } 1439 } else if( is_known_inst ) { 1440 tj = to; // Keep NotNull and klass_is_exact for instance type 1441 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1442 // During the 2nd round of IterGVN, NotNull castings are removed. 1443 // Make sure the Bottom and NotNull variants alias the same. 1444 // Also, make sure exact and non-exact variants alias the same. 1445 tj = to = to-> 1446 remove_speculative()-> 1447 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1448 cast_to_ptr_type(TypePtr::BotPTR)-> 1449 cast_to_exactness(false); 1450 } 1451 if (to->speculative() != nullptr) { 1452 tj = to = to->remove_speculative(); 1453 } 1454 // Canonicalize the holder of this field 1455 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1456 // First handle header references such as a LoadKlassNode, even if the 1457 // object's klass is unloaded at compile time (4965979). 1458 if (!is_known_inst) { // Do it only for non-instance types 1459 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset); 1460 } 1461 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) { 1462 // Static fields are in the space above the normal instance 1463 // fields in the java.lang.Class instance. 1464 if (ik != ciEnv::current()->Class_klass()) { 1465 to = nullptr; 1466 tj = TypeOopPtr::BOTTOM; 1467 offset = tj->offset(); 1468 } 1469 } else { 1470 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset); 1471 assert(offset < canonical_holder->layout_helper_size_in_bytes(), ""); 1472 if (!ik->equals(canonical_holder) || tj->offset() != offset) { 1473 if( is_known_inst ) { 1474 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id()); 1475 } else { 1476 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset); 1477 } 1478 } 1479 } 1480 } 1481 1482 // Klass pointers to object array klasses need some flattening 1483 const TypeKlassPtr *tk = tj->isa_klassptr(); 1484 if( tk ) { 1485 // If we are referencing a field within a Klass, we need 1486 // to assume the worst case of an Object. Both exact and 1487 // inexact types must flatten to the same alias class so 1488 // use NotNull as the PTR. 1489 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1490 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, 1491 env()->Object_klass(), 1492 offset); 1493 } 1494 1495 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) { 1496 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass()); 1497 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs 1498 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset); 1499 } else { 1500 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset); 1501 } 1502 } 1503 1504 // Check for precise loads from the primary supertype array and force them 1505 // to the supertype cache alias index. Check for generic array loads from 1506 // the primary supertype array and also force them to the supertype cache 1507 // alias index. Since the same load can reach both, we need to merge 1508 // these 2 disparate memories into the same alias class. Since the 1509 // primary supertype array is read-only, there's no chance of confusion 1510 // where we bypass an array load and an array store. 1511 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1512 if (offset == Type::OffsetBot || 1513 (offset >= primary_supers_offset && 1514 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1515 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1516 offset = in_bytes(Klass::secondary_super_cache_offset()); 1517 tj = tk = tk->with_offset(offset); 1518 } 1519 } 1520 1521 // Flatten all Raw pointers together. 1522 if (tj->base() == Type::RawPtr) 1523 tj = TypeRawPtr::BOTTOM; 1524 1525 if (tj->base() == Type::AnyPtr) 1526 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1527 1528 offset = tj->offset(); 1529 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1530 1531 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1532 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1533 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1534 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1535 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1536 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1537 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1538 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1539 assert( tj->ptr() != TypePtr::TopPTR && 1540 tj->ptr() != TypePtr::AnyNull && 1541 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1542 // assert( tj->ptr() != TypePtr::Constant || 1543 // tj->base() == Type::RawPtr || 1544 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1545 1546 return tj; 1547 } 1548 1549 void Compile::AliasType::Init(int i, const TypePtr* at) { 1550 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1551 _index = i; 1552 _adr_type = at; 1553 _field = nullptr; 1554 _element = nullptr; 1555 _is_rewritable = true; // default 1556 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr; 1557 if (atoop != nullptr && atoop->is_known_instance()) { 1558 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1559 _general_index = Compile::current()->get_alias_index(gt); 1560 } else { 1561 _general_index = 0; 1562 } 1563 } 1564 1565 BasicType Compile::AliasType::basic_type() const { 1566 if (element() != nullptr) { 1567 const Type* element = adr_type()->is_aryptr()->elem(); 1568 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1569 } if (field() != nullptr) { 1570 return field()->layout_type(); 1571 } else { 1572 return T_ILLEGAL; // unknown 1573 } 1574 } 1575 1576 //---------------------------------print_on------------------------------------ 1577 #ifndef PRODUCT 1578 void Compile::AliasType::print_on(outputStream* st) { 1579 if (index() < 10) 1580 st->print("@ <%d> ", index()); 1581 else st->print("@ <%d>", index()); 1582 st->print(is_rewritable() ? " " : " RO"); 1583 int offset = adr_type()->offset(); 1584 if (offset == Type::OffsetBot) 1585 st->print(" +any"); 1586 else st->print(" +%-3d", offset); 1587 st->print(" in "); 1588 adr_type()->dump_on(st); 1589 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1590 if (field() != nullptr && tjp) { 1591 if (tjp->is_instptr()->instance_klass() != field()->holder() || 1592 tjp->offset() != field()->offset_in_bytes()) { 1593 st->print(" != "); 1594 field()->print(); 1595 st->print(" ***"); 1596 } 1597 } 1598 } 1599 1600 void print_alias_types() { 1601 Compile* C = Compile::current(); 1602 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1603 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1604 C->alias_type(idx)->print_on(tty); 1605 tty->cr(); 1606 } 1607 } 1608 #endif 1609 1610 1611 //----------------------------probe_alias_cache-------------------------------- 1612 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1613 intptr_t key = (intptr_t) adr_type; 1614 key ^= key >> logAliasCacheSize; 1615 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1616 } 1617 1618 1619 //-----------------------------grow_alias_types-------------------------------- 1620 void Compile::grow_alias_types() { 1621 const int old_ats = _max_alias_types; // how many before? 1622 const int new_ats = old_ats; // how many more? 1623 const int grow_ats = old_ats+new_ats; // how many now? 1624 _max_alias_types = grow_ats; 1625 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1626 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1627 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1628 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1629 } 1630 1631 1632 //--------------------------------find_alias_type------------------------------ 1633 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1634 if (!do_aliasing()) { 1635 return alias_type(AliasIdxBot); 1636 } 1637 1638 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1639 if (ace->_adr_type == adr_type) { 1640 return alias_type(ace->_index); 1641 } 1642 1643 // Handle special cases. 1644 if (adr_type == nullptr) return alias_type(AliasIdxTop); 1645 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1646 1647 // Do it the slow way. 1648 const TypePtr* flat = flatten_alias_type(adr_type); 1649 1650 #ifdef ASSERT 1651 { 1652 ResourceMark rm; 1653 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1654 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1655 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1656 Type::str(adr_type)); 1657 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1658 const TypeOopPtr* foop = flat->is_oopptr(); 1659 // Scalarizable allocations have exact klass always. 1660 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1661 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1662 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1663 Type::str(foop), Type::str(xoop)); 1664 } 1665 } 1666 #endif 1667 1668 int idx = AliasIdxTop; 1669 for (int i = 0; i < num_alias_types(); i++) { 1670 if (alias_type(i)->adr_type() == flat) { 1671 idx = i; 1672 break; 1673 } 1674 } 1675 1676 if (idx == AliasIdxTop) { 1677 if (no_create) return nullptr; 1678 // Grow the array if necessary. 1679 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1680 // Add a new alias type. 1681 idx = _num_alias_types++; 1682 _alias_types[idx]->Init(idx, flat); 1683 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1684 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1685 if (flat->isa_instptr()) { 1686 if (flat->offset() == java_lang_Class::klass_offset() 1687 && flat->is_instptr()->instance_klass() == env()->Class_klass()) 1688 alias_type(idx)->set_rewritable(false); 1689 } 1690 if (flat->isa_aryptr()) { 1691 #ifdef ASSERT 1692 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1693 // (T_BYTE has the weakest alignment and size restrictions...) 1694 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1695 #endif 1696 if (flat->offset() == TypePtr::OffsetBot) { 1697 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1698 } 1699 } 1700 if (flat->isa_klassptr()) { 1701 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1702 alias_type(idx)->set_rewritable(false); 1703 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1704 alias_type(idx)->set_rewritable(false); 1705 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1706 alias_type(idx)->set_rewritable(false); 1707 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1708 alias_type(idx)->set_rewritable(false); 1709 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1710 alias_type(idx)->set_rewritable(false); 1711 } 1712 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1713 // but the base pointer type is not distinctive enough to identify 1714 // references into JavaThread.) 1715 1716 // Check for final fields. 1717 const TypeInstPtr* tinst = flat->isa_instptr(); 1718 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1719 ciField* field; 1720 if (tinst->const_oop() != nullptr && 1721 tinst->instance_klass() == ciEnv::current()->Class_klass() && 1722 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) { 1723 // static field 1724 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1725 field = k->get_field_by_offset(tinst->offset(), true); 1726 } else { 1727 ciInstanceKlass *k = tinst->instance_klass(); 1728 field = k->get_field_by_offset(tinst->offset(), false); 1729 } 1730 assert(field == nullptr || 1731 original_field == nullptr || 1732 (field->holder() == original_field->holder() && 1733 field->offset_in_bytes() == original_field->offset_in_bytes() && 1734 field->is_static() == original_field->is_static()), "wrong field?"); 1735 // Set field() and is_rewritable() attributes. 1736 if (field != nullptr) alias_type(idx)->set_field(field); 1737 } 1738 } 1739 1740 // Fill the cache for next time. 1741 ace->_adr_type = adr_type; 1742 ace->_index = idx; 1743 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1744 1745 // Might as well try to fill the cache for the flattened version, too. 1746 AliasCacheEntry* face = probe_alias_cache(flat); 1747 if (face->_adr_type == nullptr) { 1748 face->_adr_type = flat; 1749 face->_index = idx; 1750 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1751 } 1752 1753 return alias_type(idx); 1754 } 1755 1756 1757 Compile::AliasType* Compile::alias_type(ciField* field) { 1758 const TypeOopPtr* t; 1759 if (field->is_static()) 1760 t = TypeInstPtr::make(field->holder()->java_mirror()); 1761 else 1762 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1763 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1764 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1765 return atp; 1766 } 1767 1768 1769 //------------------------------have_alias_type-------------------------------- 1770 bool Compile::have_alias_type(const TypePtr* adr_type) { 1771 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1772 if (ace->_adr_type == adr_type) { 1773 return true; 1774 } 1775 1776 // Handle special cases. 1777 if (adr_type == nullptr) return true; 1778 if (adr_type == TypePtr::BOTTOM) return true; 1779 1780 return find_alias_type(adr_type, true, nullptr) != nullptr; 1781 } 1782 1783 //-----------------------------must_alias-------------------------------------- 1784 // True if all values of the given address type are in the given alias category. 1785 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1786 if (alias_idx == AliasIdxBot) return true; // the universal category 1787 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP 1788 if (alias_idx == AliasIdxTop) return false; // the empty category 1789 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1790 1791 // the only remaining possible overlap is identity 1792 int adr_idx = get_alias_index(adr_type); 1793 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1794 assert(adr_idx == alias_idx || 1795 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1796 && adr_type != TypeOopPtr::BOTTOM), 1797 "should not be testing for overlap with an unsafe pointer"); 1798 return adr_idx == alias_idx; 1799 } 1800 1801 //------------------------------can_alias-------------------------------------- 1802 // True if any values of the given address type are in the given alias category. 1803 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1804 if (alias_idx == AliasIdxTop) return false; // the empty category 1805 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP 1806 // Known instance doesn't alias with bottom memory 1807 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1808 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1809 1810 // the only remaining possible overlap is identity 1811 int adr_idx = get_alias_index(adr_type); 1812 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1813 return adr_idx == alias_idx; 1814 } 1815 1816 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their 1817 // uncommon traps if no Runtime Predicates were created from the Parse Predicates. 1818 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) { 1819 if (parse_predicate_count() == 0) { 1820 return; 1821 } 1822 for (int i = 0; i < parse_predicate_count(); i++) { 1823 ParsePredicateNode* parse_predicate = _parse_predicates.at(i); 1824 parse_predicate->mark_useless(); 1825 igvn._worklist.push(parse_predicate); 1826 } 1827 _parse_predicates.clear(); 1828 } 1829 1830 void Compile::record_for_post_loop_opts_igvn(Node* n) { 1831 if (!n->for_post_loop_opts_igvn()) { 1832 assert(!_for_post_loop_igvn.contains(n), "duplicate"); 1833 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1834 _for_post_loop_igvn.append(n); 1835 } 1836 } 1837 1838 void Compile::remove_from_post_loop_opts_igvn(Node* n) { 1839 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1840 _for_post_loop_igvn.remove(n); 1841 } 1842 1843 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) { 1844 // Verify that all previous optimizations produced a valid graph 1845 // at least to this point, even if no loop optimizations were done. 1846 PhaseIdealLoop::verify(igvn); 1847 1848 C->set_post_loop_opts_phase(); // no more loop opts allowed 1849 1850 assert(!C->major_progress(), "not cleared"); 1851 1852 if (_for_post_loop_igvn.length() > 0) { 1853 while (_for_post_loop_igvn.length() > 0) { 1854 Node* n = _for_post_loop_igvn.pop(); 1855 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1856 igvn._worklist.push(n); 1857 } 1858 igvn.optimize(); 1859 if (failing()) return; 1860 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed"); 1861 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now"); 1862 1863 // Sometimes IGVN sets major progress (e.g., when processing loop nodes). 1864 if (C->major_progress()) { 1865 C->clear_major_progress(); // ensure that major progress is now clear 1866 } 1867 } 1868 } 1869 1870 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) { 1871 if (OptimizeUnstableIf) { 1872 _unstable_if_traps.append(trap); 1873 } 1874 } 1875 1876 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) { 1877 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) { 1878 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1879 Node* n = trap->uncommon_trap(); 1880 if (!useful.member(n)) { 1881 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 1882 } 1883 } 1884 } 1885 1886 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead 1887 // or fold-compares case. Return true if succeed or not found. 1888 // 1889 // In rare cases, the found trap has been processed. It is too late to delete it. Return 1890 // false and ask fold-compares to yield. 1891 // 1892 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused 1893 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path 1894 // when deoptimization does happen. 1895 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) { 1896 for (int i = 0; i < _unstable_if_traps.length(); ++i) { 1897 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1898 if (trap->uncommon_trap() == unc) { 1899 if (yield && trap->modified()) { 1900 return false; 1901 } 1902 _unstable_if_traps.delete_at(i); 1903 break; 1904 } 1905 } 1906 return true; 1907 } 1908 1909 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path. 1910 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering. 1911 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) { 1912 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) { 1913 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1914 CallStaticJavaNode* unc = trap->uncommon_trap(); 1915 int next_bci = trap->next_bci(); 1916 bool modified = trap->modified(); 1917 1918 if (next_bci != -1 && !modified) { 1919 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!"); 1920 JVMState* jvms = unc->jvms(); 1921 ciMethod* method = jvms->method(); 1922 ciBytecodeStream iter(method); 1923 1924 iter.force_bci(jvms->bci()); 1925 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if"); 1926 Bytecodes::Code c = iter.cur_bc(); 1927 Node* lhs = nullptr; 1928 Node* rhs = nullptr; 1929 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) { 1930 lhs = unc->peek_operand(0); 1931 rhs = unc->peek_operand(1); 1932 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) { 1933 lhs = unc->peek_operand(0); 1934 } 1935 1936 ResourceMark rm; 1937 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci); 1938 assert(live_locals.is_valid(), "broken liveness info"); 1939 int len = (int)live_locals.size(); 1940 1941 for (int i = 0; i < len; i++) { 1942 Node* local = unc->local(jvms, i); 1943 // kill local using the liveness of next_bci. 1944 // give up when the local looks like an operand to secure reexecution. 1945 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) { 1946 uint idx = jvms->locoff() + i; 1947 #ifdef ASSERT 1948 if (PrintOpto && Verbose) { 1949 tty->print("[unstable_if] kill local#%d: ", idx); 1950 local->dump(); 1951 tty->cr(); 1952 } 1953 #endif 1954 igvn.replace_input_of(unc, idx, top()); 1955 modified = true; 1956 } 1957 } 1958 } 1959 1960 // keep the mondified trap for late query 1961 if (modified) { 1962 trap->set_modified(); 1963 } else { 1964 _unstable_if_traps.delete_at(i); 1965 } 1966 } 1967 igvn.optimize(); 1968 } 1969 1970 // StringOpts and late inlining of string methods 1971 void Compile::inline_string_calls(bool parse_time) { 1972 { 1973 // remove useless nodes to make the usage analysis simpler 1974 ResourceMark rm; 1975 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 1976 } 1977 1978 { 1979 ResourceMark rm; 1980 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1981 PhaseStringOpts pso(initial_gvn()); 1982 print_method(PHASE_AFTER_STRINGOPTS, 3); 1983 } 1984 1985 // now inline anything that we skipped the first time around 1986 if (!parse_time) { 1987 _late_inlines_pos = _late_inlines.length(); 1988 } 1989 1990 while (_string_late_inlines.length() > 0) { 1991 CallGenerator* cg = _string_late_inlines.pop(); 1992 cg->do_late_inline(); 1993 if (failing()) return; 1994 } 1995 _string_late_inlines.trunc_to(0); 1996 } 1997 1998 // Late inlining of boxing methods 1999 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2000 if (_boxing_late_inlines.length() > 0) { 2001 assert(has_boxed_value(), "inconsistent"); 2002 2003 set_inlining_incrementally(true); 2004 2005 igvn_worklist()->ensure_empty(); // should be done with igvn 2006 2007 _late_inlines_pos = _late_inlines.length(); 2008 2009 while (_boxing_late_inlines.length() > 0) { 2010 CallGenerator* cg = _boxing_late_inlines.pop(); 2011 cg->do_late_inline(); 2012 if (failing()) return; 2013 } 2014 _boxing_late_inlines.trunc_to(0); 2015 2016 inline_incrementally_cleanup(igvn); 2017 2018 set_inlining_incrementally(false); 2019 } 2020 } 2021 2022 bool Compile::inline_incrementally_one() { 2023 assert(IncrementalInline, "incremental inlining should be on"); 2024 2025 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2026 2027 set_inlining_progress(false); 2028 set_do_cleanup(false); 2029 2030 for (int i = 0; i < _late_inlines.length(); i++) { 2031 _late_inlines_pos = i+1; 2032 CallGenerator* cg = _late_inlines.at(i); 2033 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline(); 2034 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call 2035 cg->do_late_inline(); 2036 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed"); 2037 if (failing()) { 2038 return false; 2039 } else if (inlining_progress()) { 2040 _late_inlines_pos = i+1; // restore the position in case new elements were inserted 2041 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node()); 2042 break; // process one call site at a time 2043 } 2044 } else { 2045 // Ignore late inline direct calls when inlining is not allowed. 2046 // They are left in the late inline list when node budget is exhausted until the list is fully drained. 2047 } 2048 } 2049 // Remove processed elements. 2050 _late_inlines.remove_till(_late_inlines_pos); 2051 _late_inlines_pos = 0; 2052 2053 assert(inlining_progress() || _late_inlines.length() == 0, "no progress"); 2054 2055 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2056 2057 set_inlining_progress(false); 2058 set_do_cleanup(false); 2059 2060 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption; 2061 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup; 2062 } 2063 2064 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2065 { 2066 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2067 ResourceMark rm; 2068 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2069 } 2070 { 2071 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2072 igvn.reset_from_gvn(initial_gvn()); 2073 igvn.optimize(); 2074 if (failing()) return; 2075 } 2076 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3); 2077 } 2078 2079 // Perform incremental inlining until bound on number of live nodes is reached 2080 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2081 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2082 2083 set_inlining_incrementally(true); 2084 uint low_live_nodes = 0; 2085 2086 while (_late_inlines.length() > 0) { 2087 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2088 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2089 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2090 // PhaseIdealLoop is expensive so we only try it once we are 2091 // out of live nodes and we only try it again if the previous 2092 // helped got the number of nodes down significantly 2093 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2094 if (failing()) return; 2095 low_live_nodes = live_nodes(); 2096 _major_progress = true; 2097 } 2098 2099 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2100 bool do_print_inlining = print_inlining() || print_intrinsics(); 2101 if (do_print_inlining || log() != nullptr) { 2102 // Print inlining message for candidates that we couldn't inline for lack of space. 2103 for (int i = 0; i < _late_inlines.length(); i++) { 2104 CallGenerator* cg = _late_inlines.at(i); 2105 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 2106 if (do_print_inlining) { 2107 cg->print_inlining_late(InliningResult::FAILURE, msg); 2108 } 2109 log_late_inline_failure(cg, msg); 2110 } 2111 } 2112 break; // finish 2113 } 2114 } 2115 2116 igvn_worklist()->ensure_empty(); // should be done with igvn 2117 2118 while (inline_incrementally_one()) { 2119 assert(!failing(), "inconsistent"); 2120 } 2121 if (failing()) return; 2122 2123 inline_incrementally_cleanup(igvn); 2124 2125 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 2126 2127 if (failing()) return; 2128 2129 if (_late_inlines.length() == 0) { 2130 break; // no more progress 2131 } 2132 } 2133 2134 igvn_worklist()->ensure_empty(); // should be done with igvn 2135 2136 if (_string_late_inlines.length() > 0) { 2137 assert(has_stringbuilder(), "inconsistent"); 2138 2139 inline_string_calls(false); 2140 2141 if (failing()) return; 2142 2143 inline_incrementally_cleanup(igvn); 2144 } 2145 2146 set_inlining_incrementally(false); 2147 } 2148 2149 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) { 2150 // "inlining_incrementally() == false" is used to signal that no inlining is allowed 2151 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details). 2152 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr" 2153 // as if "inlining_incrementally() == true" were set. 2154 assert(inlining_incrementally() == false, "not allowed"); 2155 assert(_modified_nodes == nullptr, "not allowed"); 2156 assert(_late_inlines.length() > 0, "sanity"); 2157 2158 while (_late_inlines.length() > 0) { 2159 igvn_worklist()->ensure_empty(); // should be done with igvn 2160 2161 while (inline_incrementally_one()) { 2162 assert(!failing(), "inconsistent"); 2163 } 2164 if (failing()) return; 2165 2166 inline_incrementally_cleanup(igvn); 2167 } 2168 } 2169 2170 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2171 if (_loop_opts_cnt > 0) { 2172 while (major_progress() && (_loop_opts_cnt > 0)) { 2173 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2174 PhaseIdealLoop::optimize(igvn, mode); 2175 _loop_opts_cnt--; 2176 if (failing()) return false; 2177 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2178 } 2179 } 2180 return true; 2181 } 2182 2183 // Remove edges from "root" to each SafePoint at a backward branch. 2184 // They were inserted during parsing (see add_safepoint()) to make 2185 // infinite loops without calls or exceptions visible to root, i.e., 2186 // useful. 2187 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2188 Node *r = root(); 2189 if (r != nullptr) { 2190 for (uint i = r->req(); i < r->len(); ++i) { 2191 Node *n = r->in(i); 2192 if (n != nullptr && n->is_SafePoint()) { 2193 r->rm_prec(i); 2194 if (n->outcnt() == 0) { 2195 igvn.remove_dead_node(n); 2196 } 2197 --i; 2198 } 2199 } 2200 // Parsing may have added top inputs to the root node (Path 2201 // leading to the Halt node proven dead). Make sure we get a 2202 // chance to clean them up. 2203 igvn._worklist.push(r); 2204 igvn.optimize(); 2205 } 2206 } 2207 2208 //------------------------------Optimize--------------------------------------- 2209 // Given a graph, optimize it. 2210 void Compile::Optimize() { 2211 TracePhase tp("optimizer", &timers[_t_optimizer]); 2212 2213 #ifndef PRODUCT 2214 if (env()->break_at_compile()) { 2215 BREAKPOINT; 2216 } 2217 2218 #endif 2219 2220 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2221 #ifdef ASSERT 2222 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2223 #endif 2224 2225 ResourceMark rm; 2226 2227 print_inlining_reinit(); 2228 2229 NOT_PRODUCT( verify_graph_edges(); ) 2230 2231 print_method(PHASE_AFTER_PARSING, 1); 2232 2233 { 2234 // Iterative Global Value Numbering, including ideal transforms 2235 // Initialize IterGVN with types and values from parse-time GVN 2236 PhaseIterGVN igvn(initial_gvn()); 2237 #ifdef ASSERT 2238 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2239 #endif 2240 { 2241 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2242 igvn.optimize(); 2243 } 2244 2245 if (failing()) return; 2246 2247 print_method(PHASE_ITER_GVN1, 2); 2248 2249 process_for_unstable_if_traps(igvn); 2250 2251 if (failing()) return; 2252 2253 inline_incrementally(igvn); 2254 2255 print_method(PHASE_INCREMENTAL_INLINE, 2); 2256 2257 if (failing()) return; 2258 2259 if (eliminate_boxing()) { 2260 // Inline valueOf() methods now. 2261 inline_boxing_calls(igvn); 2262 2263 if (failing()) return; 2264 2265 if (AlwaysIncrementalInline || StressIncrementalInlining) { 2266 inline_incrementally(igvn); 2267 } 2268 2269 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2270 2271 if (failing()) return; 2272 } 2273 2274 // Remove the speculative part of types and clean up the graph from 2275 // the extra CastPP nodes whose only purpose is to carry them. Do 2276 // that early so that optimizations are not disrupted by the extra 2277 // CastPP nodes. 2278 remove_speculative_types(igvn); 2279 2280 if (failing()) return; 2281 2282 // No more new expensive nodes will be added to the list from here 2283 // so keep only the actual candidates for optimizations. 2284 cleanup_expensive_nodes(igvn); 2285 2286 if (failing()) return; 2287 2288 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2289 if (EnableVectorSupport && has_vbox_nodes()) { 2290 TracePhase tp("", &timers[_t_vector]); 2291 PhaseVector pv(igvn); 2292 pv.optimize_vector_boxes(); 2293 if (failing()) return; 2294 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2295 } 2296 assert(!has_vbox_nodes(), "sanity"); 2297 2298 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2299 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2300 igvn_worklist()->ensure_empty(); // should be done with igvn 2301 { 2302 ResourceMark rm; 2303 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist()); 2304 } 2305 igvn.reset_from_gvn(initial_gvn()); 2306 igvn.optimize(); 2307 if (failing()) return; 2308 } 2309 2310 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop 2311 // safepoints 2312 remove_root_to_sfpts_edges(igvn); 2313 2314 if (failing()) return; 2315 2316 // Perform escape analysis 2317 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) { 2318 if (has_loops()) { 2319 // Cleanup graph (remove dead nodes). 2320 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2321 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2322 if (failing()) return; 2323 } 2324 bool progress; 2325 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2326 do { 2327 ConnectionGraph::do_analysis(this, &igvn); 2328 2329 if (failing()) return; 2330 2331 int mcount = macro_count(); // Record number of allocations and locks before IGVN 2332 2333 // Optimize out fields loads from scalar replaceable allocations. 2334 igvn.optimize(); 2335 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2336 2337 if (failing()) return; 2338 2339 if (congraph() != nullptr && macro_count() > 0) { 2340 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2341 PhaseMacroExpand mexp(igvn); 2342 mexp.eliminate_macro_nodes(); 2343 if (failing()) return; 2344 2345 igvn.set_delay_transform(false); 2346 igvn.optimize(); 2347 if (failing()) return; 2348 2349 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2350 } 2351 2352 ConnectionGraph::verify_ram_nodes(this, root()); 2353 if (failing()) return; 2354 2355 progress = do_iterative_escape_analysis() && 2356 (macro_count() < mcount) && 2357 ConnectionGraph::has_candidates(this); 2358 // Try again if candidates exist and made progress 2359 // by removing some allocations and/or locks. 2360 } while (progress); 2361 } 2362 2363 // Loop transforms on the ideal graph. Range Check Elimination, 2364 // peeling, unrolling, etc. 2365 2366 // Set loop opts counter 2367 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2368 { 2369 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2370 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2371 _loop_opts_cnt--; 2372 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2373 if (failing()) return; 2374 } 2375 // Loop opts pass if partial peeling occurred in previous pass 2376 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2377 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2378 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2379 _loop_opts_cnt--; 2380 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2381 if (failing()) return; 2382 } 2383 // Loop opts pass for loop-unrolling before CCP 2384 if(major_progress() && (_loop_opts_cnt > 0)) { 2385 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2386 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2387 _loop_opts_cnt--; 2388 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2389 } 2390 if (!failing()) { 2391 // Verify that last round of loop opts produced a valid graph 2392 PhaseIdealLoop::verify(igvn); 2393 } 2394 } 2395 if (failing()) return; 2396 2397 // Conditional Constant Propagation; 2398 PhaseCCP ccp( &igvn ); 2399 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2400 { 2401 TracePhase tp("ccp", &timers[_t_ccp]); 2402 ccp.do_transform(); 2403 } 2404 print_method(PHASE_CCP1, 2); 2405 2406 assert( true, "Break here to ccp.dump_old2new_map()"); 2407 2408 // Iterative Global Value Numbering, including ideal transforms 2409 { 2410 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2411 igvn.reset_from_igvn(&ccp); 2412 igvn.optimize(); 2413 } 2414 print_method(PHASE_ITER_GVN2, 2); 2415 2416 if (failing()) return; 2417 2418 // Loop transforms on the ideal graph. Range Check Elimination, 2419 // peeling, unrolling, etc. 2420 if (!optimize_loops(igvn, LoopOptsDefault)) { 2421 return; 2422 } 2423 2424 if (failing()) return; 2425 2426 C->clear_major_progress(); // ensure that major progress is now clear 2427 2428 process_for_post_loop_opts_igvn(igvn); 2429 2430 if (failing()) return; 2431 2432 #ifdef ASSERT 2433 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2434 #endif 2435 2436 { 2437 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2438 PhaseMacroExpand mex(igvn); 2439 if (mex.expand_macro_nodes()) { 2440 assert(failing(), "must bail out w/ explicit message"); 2441 return; 2442 } 2443 print_method(PHASE_MACRO_EXPANSION, 2); 2444 } 2445 2446 { 2447 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2448 if (bs->expand_barriers(this, igvn)) { 2449 assert(failing(), "must bail out w/ explicit message"); 2450 return; 2451 } 2452 print_method(PHASE_BARRIER_EXPANSION, 2); 2453 } 2454 2455 if (C->max_vector_size() > 0) { 2456 C->optimize_logic_cones(igvn); 2457 igvn.optimize(); 2458 if (failing()) return; 2459 } 2460 2461 DEBUG_ONLY( _modified_nodes = nullptr; ) 2462 2463 assert(igvn._worklist.size() == 0, "not empty"); 2464 2465 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty"); 2466 2467 if (_late_inlines.length() > 0) { 2468 // More opportunities to optimize virtual and MH calls. 2469 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option. 2470 process_late_inline_calls_no_inline(igvn); 2471 if (failing()) return; 2472 } 2473 } // (End scope of igvn; run destructor if necessary for asserts.) 2474 2475 check_no_dead_use(); 2476 2477 process_print_inlining(); 2478 2479 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have 2480 // to remove hashes to unlock nodes for modifications. 2481 C->node_hash()->clear(); 2482 2483 // A method with only infinite loops has no edges entering loops from root 2484 { 2485 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2486 if (final_graph_reshaping()) { 2487 assert(failing(), "must bail out w/ explicit message"); 2488 return; 2489 } 2490 } 2491 2492 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2493 DEBUG_ONLY(set_phase_optimize_finished();) 2494 } 2495 2496 #ifdef ASSERT 2497 void Compile::check_no_dead_use() const { 2498 ResourceMark rm; 2499 Unique_Node_List wq; 2500 wq.push(root()); 2501 for (uint i = 0; i < wq.size(); ++i) { 2502 Node* n = wq.at(i); 2503 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2504 Node* u = n->fast_out(j); 2505 if (u->outcnt() == 0 && !u->is_Con()) { 2506 u->dump(); 2507 fatal("no reachable node should have no use"); 2508 } 2509 wq.push(u); 2510 } 2511 } 2512 } 2513 #endif 2514 2515 void Compile::inline_vector_reboxing_calls() { 2516 if (C->_vector_reboxing_late_inlines.length() > 0) { 2517 _late_inlines_pos = C->_late_inlines.length(); 2518 while (_vector_reboxing_late_inlines.length() > 0) { 2519 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 2520 cg->do_late_inline(); 2521 if (failing()) return; 2522 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node()); 2523 } 2524 _vector_reboxing_late_inlines.trunc_to(0); 2525 } 2526 } 2527 2528 bool Compile::has_vbox_nodes() { 2529 if (C->_vector_reboxing_late_inlines.length() > 0) { 2530 return true; 2531 } 2532 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 2533 Node * n = C->macro_node(macro_idx); 2534 assert(n->is_macro(), "only macro nodes expected here"); 2535 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 2536 return true; 2537 } 2538 } 2539 return false; 2540 } 2541 2542 //---------------------------- Bitwise operation packing optimization --------------------------- 2543 2544 static bool is_vector_unary_bitwise_op(Node* n) { 2545 return n->Opcode() == Op_XorV && 2546 VectorNode::is_vector_bitwise_not_pattern(n); 2547 } 2548 2549 static bool is_vector_binary_bitwise_op(Node* n) { 2550 switch (n->Opcode()) { 2551 case Op_AndV: 2552 case Op_OrV: 2553 return true; 2554 2555 case Op_XorV: 2556 return !is_vector_unary_bitwise_op(n); 2557 2558 default: 2559 return false; 2560 } 2561 } 2562 2563 static bool is_vector_ternary_bitwise_op(Node* n) { 2564 return n->Opcode() == Op_MacroLogicV; 2565 } 2566 2567 static bool is_vector_bitwise_op(Node* n) { 2568 return is_vector_unary_bitwise_op(n) || 2569 is_vector_binary_bitwise_op(n) || 2570 is_vector_ternary_bitwise_op(n); 2571 } 2572 2573 static bool is_vector_bitwise_cone_root(Node* n) { 2574 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) { 2575 return false; 2576 } 2577 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2578 if (is_vector_bitwise_op(n->fast_out(i))) { 2579 return false; 2580 } 2581 } 2582 return true; 2583 } 2584 2585 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) { 2586 uint cnt = 0; 2587 if (is_vector_bitwise_op(n)) { 2588 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req(); 2589 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2590 for (uint i = 1; i < inp_cnt; i++) { 2591 Node* in = n->in(i); 2592 bool skip = VectorNode::is_all_ones_vector(in); 2593 if (!skip && !inputs.member(in)) { 2594 inputs.push(in); 2595 cnt++; 2596 } 2597 } 2598 assert(cnt <= 1, "not unary"); 2599 } else { 2600 uint last_req = inp_cnt; 2601 if (is_vector_ternary_bitwise_op(n)) { 2602 last_req = inp_cnt - 1; // skip last input 2603 } 2604 for (uint i = 1; i < last_req; i++) { 2605 Node* def = n->in(i); 2606 if (!inputs.member(def)) { 2607 inputs.push(def); 2608 cnt++; 2609 } 2610 } 2611 } 2612 } else { // not a bitwise operations 2613 if (!inputs.member(n)) { 2614 inputs.push(n); 2615 cnt++; 2616 } 2617 } 2618 return cnt; 2619 } 2620 2621 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 2622 Unique_Node_List useful_nodes; 2623 C->identify_useful_nodes(useful_nodes); 2624 2625 for (uint i = 0; i < useful_nodes.size(); i++) { 2626 Node* n = useful_nodes.at(i); 2627 if (is_vector_bitwise_cone_root(n)) { 2628 list.push(n); 2629 } 2630 } 2631 } 2632 2633 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 2634 const TypeVect* vt, 2635 Unique_Node_List& partition, 2636 Unique_Node_List& inputs) { 2637 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 2638 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 2639 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 2640 2641 Node* in1 = inputs.at(0); 2642 Node* in2 = inputs.at(1); 2643 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 2644 2645 uint func = compute_truth_table(partition, inputs); 2646 2647 Node* pn = partition.at(partition.size() - 1); 2648 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2649 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt)); 2650 } 2651 2652 static uint extract_bit(uint func, uint pos) { 2653 return (func & (1 << pos)) >> pos; 2654 } 2655 2656 // 2657 // A macro logic node represents a truth table. It has 4 inputs, 2658 // First three inputs corresponds to 3 columns of a truth table 2659 // and fourth input captures the logic function. 2660 // 2661 // eg. fn = (in1 AND in2) OR in3; 2662 // 2663 // MacroNode(in1,in2,in3,fn) 2664 // 2665 // ----------------- 2666 // in1 in2 in3 fn 2667 // ----------------- 2668 // 0 0 0 0 2669 // 0 0 1 1 2670 // 0 1 0 0 2671 // 0 1 1 1 2672 // 1 0 0 0 2673 // 1 0 1 1 2674 // 1 1 0 1 2675 // 1 1 1 1 2676 // 2677 2678 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 2679 int res = 0; 2680 for (int i = 0; i < 8; i++) { 2681 int bit1 = extract_bit(in1, i); 2682 int bit2 = extract_bit(in2, i); 2683 int bit3 = extract_bit(in3, i); 2684 2685 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 2686 int func_bit = extract_bit(func, func_bit_pos); 2687 2688 res |= func_bit << i; 2689 } 2690 return res; 2691 } 2692 2693 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 2694 assert(n != nullptr, ""); 2695 assert(eval_map.contains(n), "absent"); 2696 return *(eval_map.get(n)); 2697 } 2698 2699 static void eval_operands(Node* n, 2700 uint& func1, uint& func2, uint& func3, 2701 ResourceHashtable<Node*,uint>& eval_map) { 2702 assert(is_vector_bitwise_op(n), ""); 2703 2704 if (is_vector_unary_bitwise_op(n)) { 2705 Node* opnd = n->in(1); 2706 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 2707 opnd = n->in(2); 2708 } 2709 func1 = eval_operand(opnd, eval_map); 2710 } else if (is_vector_binary_bitwise_op(n)) { 2711 func1 = eval_operand(n->in(1), eval_map); 2712 func2 = eval_operand(n->in(2), eval_map); 2713 } else { 2714 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 2715 func1 = eval_operand(n->in(1), eval_map); 2716 func2 = eval_operand(n->in(2), eval_map); 2717 func3 = eval_operand(n->in(3), eval_map); 2718 } 2719 } 2720 2721 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 2722 assert(inputs.size() <= 3, "sanity"); 2723 ResourceMark rm; 2724 uint res = 0; 2725 ResourceHashtable<Node*,uint> eval_map; 2726 2727 // Populate precomputed functions for inputs. 2728 // Each input corresponds to one column of 3 input truth-table. 2729 uint input_funcs[] = { 0xAA, // (_, _, c) -> c 2730 0xCC, // (_, b, _) -> b 2731 0xF0 }; // (a, _, _) -> a 2732 for (uint i = 0; i < inputs.size(); i++) { 2733 eval_map.put(inputs.at(i), input_funcs[2-i]); 2734 } 2735 2736 for (uint i = 0; i < partition.size(); i++) { 2737 Node* n = partition.at(i); 2738 2739 uint func1 = 0, func2 = 0, func3 = 0; 2740 eval_operands(n, func1, func2, func3, eval_map); 2741 2742 switch (n->Opcode()) { 2743 case Op_OrV: 2744 assert(func3 == 0, "not binary"); 2745 res = func1 | func2; 2746 break; 2747 case Op_AndV: 2748 assert(func3 == 0, "not binary"); 2749 res = func1 & func2; 2750 break; 2751 case Op_XorV: 2752 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2753 assert(func2 == 0 && func3 == 0, "not unary"); 2754 res = (~func1) & 0xFF; 2755 } else { 2756 assert(func3 == 0, "not binary"); 2757 res = func1 ^ func2; 2758 } 2759 break; 2760 case Op_MacroLogicV: 2761 // Ordering of inputs may change during evaluation of sub-tree 2762 // containing MacroLogic node as a child node, thus a re-evaluation 2763 // makes sure that function is evaluated in context of current 2764 // inputs. 2765 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 2766 break; 2767 2768 default: assert(false, "not supported: %s", n->Name()); 2769 } 2770 assert(res <= 0xFF, "invalid"); 2771 eval_map.put(n, res); 2772 } 2773 return res; 2774 } 2775 2776 // Criteria under which nodes gets packed into a macro logic node:- 2777 // 1) Parent and both child nodes are all unmasked or masked with 2778 // same predicates. 2779 // 2) Masked parent can be packed with left child if it is predicated 2780 // and both have same predicates. 2781 // 3) Masked parent can be packed with right child if its un-predicated 2782 // or has matching predication condition. 2783 // 4) An unmasked parent can be packed with an unmasked child. 2784 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2785 assert(partition.size() == 0, "not empty"); 2786 assert(inputs.size() == 0, "not empty"); 2787 if (is_vector_ternary_bitwise_op(n)) { 2788 return false; 2789 } 2790 2791 bool is_unary_op = is_vector_unary_bitwise_op(n); 2792 if (is_unary_op) { 2793 assert(collect_unique_inputs(n, inputs) == 1, "not unary"); 2794 return false; // too few inputs 2795 } 2796 2797 bool pack_left_child = true; 2798 bool pack_right_child = true; 2799 2800 bool left_child_LOP = is_vector_bitwise_op(n->in(1)); 2801 bool right_child_LOP = is_vector_bitwise_op(n->in(2)); 2802 2803 int left_child_input_cnt = 0; 2804 int right_child_input_cnt = 0; 2805 2806 bool parent_is_predicated = n->is_predicated_vector(); 2807 bool left_child_predicated = n->in(1)->is_predicated_vector(); 2808 bool right_child_predicated = n->in(2)->is_predicated_vector(); 2809 2810 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr; 2811 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2812 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2813 2814 do { 2815 if (pack_left_child && left_child_LOP && 2816 ((!parent_is_predicated && !left_child_predicated) || 2817 ((parent_is_predicated && left_child_predicated && 2818 parent_pred == left_child_pred)))) { 2819 partition.push(n->in(1)); 2820 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs); 2821 } else { 2822 inputs.push(n->in(1)); 2823 left_child_input_cnt = 1; 2824 } 2825 2826 if (pack_right_child && right_child_LOP && 2827 (!right_child_predicated || 2828 (right_child_predicated && parent_is_predicated && 2829 parent_pred == right_child_pred))) { 2830 partition.push(n->in(2)); 2831 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs); 2832 } else { 2833 inputs.push(n->in(2)); 2834 right_child_input_cnt = 1; 2835 } 2836 2837 if (inputs.size() > 3) { 2838 assert(partition.size() > 0, ""); 2839 inputs.clear(); 2840 partition.clear(); 2841 if (left_child_input_cnt > right_child_input_cnt) { 2842 pack_left_child = false; 2843 } else { 2844 pack_right_child = false; 2845 } 2846 } else { 2847 break; 2848 } 2849 } while(true); 2850 2851 if(partition.size()) { 2852 partition.push(n); 2853 } 2854 2855 return (partition.size() == 2 || partition.size() == 3) && 2856 (inputs.size() == 2 || inputs.size() == 3); 2857 } 2858 2859 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 2860 assert(is_vector_bitwise_op(n), "not a root"); 2861 2862 visited.set(n->_idx); 2863 2864 // 1) Do a DFS walk over the logic cone. 2865 for (uint i = 1; i < n->req(); i++) { 2866 Node* in = n->in(i); 2867 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 2868 process_logic_cone_root(igvn, in, visited); 2869 } 2870 } 2871 2872 // 2) Bottom up traversal: Merge node[s] with 2873 // the parent to form macro logic node. 2874 Unique_Node_List partition; 2875 Unique_Node_List inputs; 2876 if (compute_logic_cone(n, partition, inputs)) { 2877 const TypeVect* vt = n->bottom_type()->is_vect(); 2878 Node* pn = partition.at(partition.size() - 1); 2879 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2880 if (mask == nullptr || 2881 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) { 2882 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 2883 VectorNode::trace_new_vector(macro_logic, "MacroLogic"); 2884 igvn.replace_node(n, macro_logic); 2885 } 2886 } 2887 } 2888 2889 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 2890 ResourceMark rm; 2891 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 2892 Unique_Node_List list; 2893 collect_logic_cone_roots(list); 2894 2895 while (list.size() > 0) { 2896 Node* n = list.pop(); 2897 const TypeVect* vt = n->bottom_type()->is_vect(); 2898 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 2899 if (supported) { 2900 VectorSet visited(comp_arena()); 2901 process_logic_cone_root(igvn, n, visited); 2902 } 2903 } 2904 } 2905 } 2906 2907 //------------------------------Code_Gen--------------------------------------- 2908 // Given a graph, generate code for it 2909 void Compile::Code_Gen() { 2910 if (failing()) { 2911 return; 2912 } 2913 2914 // Perform instruction selection. You might think we could reclaim Matcher 2915 // memory PDQ, but actually the Matcher is used in generating spill code. 2916 // Internals of the Matcher (including some VectorSets) must remain live 2917 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2918 // set a bit in reclaimed memory. 2919 2920 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2921 // nodes. Mapping is only valid at the root of each matched subtree. 2922 NOT_PRODUCT( verify_graph_edges(); ) 2923 2924 Matcher matcher; 2925 _matcher = &matcher; 2926 { 2927 TracePhase tp("matcher", &timers[_t_matcher]); 2928 matcher.match(); 2929 if (failing()) { 2930 return; 2931 } 2932 } 2933 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2934 // nodes. Mapping is only valid at the root of each matched subtree. 2935 NOT_PRODUCT( verify_graph_edges(); ) 2936 2937 // If you have too many nodes, or if matching has failed, bail out 2938 check_node_count(0, "out of nodes matching instructions"); 2939 if (failing()) { 2940 return; 2941 } 2942 2943 print_method(PHASE_MATCHING, 2); 2944 2945 // Build a proper-looking CFG 2946 PhaseCFG cfg(node_arena(), root(), matcher); 2947 _cfg = &cfg; 2948 { 2949 TracePhase tp("scheduler", &timers[_t_scheduler]); 2950 bool success = cfg.do_global_code_motion(); 2951 if (!success) { 2952 return; 2953 } 2954 2955 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2956 NOT_PRODUCT( verify_graph_edges(); ) 2957 cfg.verify(); 2958 } 2959 2960 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2961 _regalloc = ®alloc; 2962 { 2963 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2964 // Perform register allocation. After Chaitin, use-def chains are 2965 // no longer accurate (at spill code) and so must be ignored. 2966 // Node->LRG->reg mappings are still accurate. 2967 _regalloc->Register_Allocate(); 2968 2969 // Bail out if the allocator builds too many nodes 2970 if (failing()) { 2971 return; 2972 } 2973 } 2974 2975 // Prior to register allocation we kept empty basic blocks in case the 2976 // the allocator needed a place to spill. After register allocation we 2977 // are not adding any new instructions. If any basic block is empty, we 2978 // can now safely remove it. 2979 { 2980 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2981 cfg.remove_empty_blocks(); 2982 if (do_freq_based_layout()) { 2983 PhaseBlockLayout layout(cfg); 2984 } else { 2985 cfg.set_loop_alignment(); 2986 } 2987 cfg.fixup_flow(); 2988 cfg.remove_unreachable_blocks(); 2989 cfg.verify_dominator_tree(); 2990 } 2991 2992 // Apply peephole optimizations 2993 if( OptoPeephole ) { 2994 TracePhase tp("peephole", &timers[_t_peephole]); 2995 PhasePeephole peep( _regalloc, cfg); 2996 peep.do_transform(); 2997 } 2998 2999 // Do late expand if CPU requires this. 3000 if (Matcher::require_postalloc_expand) { 3001 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 3002 cfg.postalloc_expand(_regalloc); 3003 } 3004 3005 // Convert Nodes to instruction bits in a buffer 3006 { 3007 TracePhase tp("output", &timers[_t_output]); 3008 PhaseOutput output; 3009 output.Output(); 3010 if (failing()) return; 3011 output.install(); 3012 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here 3013 } 3014 3015 // He's dead, Jim. 3016 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 3017 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 3018 } 3019 3020 //------------------------------Final_Reshape_Counts--------------------------- 3021 // This class defines counters to help identify when a method 3022 // may/must be executed using hardware with only 24-bit precision. 3023 struct Final_Reshape_Counts : public StackObj { 3024 int _call_count; // count non-inlined 'common' calls 3025 int _float_count; // count float ops requiring 24-bit precision 3026 int _double_count; // count double ops requiring more precision 3027 int _java_call_count; // count non-inlined 'java' calls 3028 int _inner_loop_count; // count loops which need alignment 3029 VectorSet _visited; // Visitation flags 3030 Node_List _tests; // Set of IfNodes & PCTableNodes 3031 3032 Final_Reshape_Counts() : 3033 _call_count(0), _float_count(0), _double_count(0), 3034 _java_call_count(0), _inner_loop_count(0) { } 3035 3036 void inc_call_count () { _call_count ++; } 3037 void inc_float_count () { _float_count ++; } 3038 void inc_double_count() { _double_count++; } 3039 void inc_java_call_count() { _java_call_count++; } 3040 void inc_inner_loop_count() { _inner_loop_count++; } 3041 3042 int get_call_count () const { return _call_count ; } 3043 int get_float_count () const { return _float_count ; } 3044 int get_double_count() const { return _double_count; } 3045 int get_java_call_count() const { return _java_call_count; } 3046 int get_inner_loop_count() const { return _inner_loop_count; } 3047 }; 3048 3049 // Eliminate trivially redundant StoreCMs and accumulate their 3050 // precedence edges. 3051 void Compile::eliminate_redundant_card_marks(Node* n) { 3052 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 3053 if (n->in(MemNode::Address)->outcnt() > 1) { 3054 // There are multiple users of the same address so it might be 3055 // possible to eliminate some of the StoreCMs 3056 Node* mem = n->in(MemNode::Memory); 3057 Node* adr = n->in(MemNode::Address); 3058 Node* val = n->in(MemNode::ValueIn); 3059 Node* prev = n; 3060 bool done = false; 3061 // Walk the chain of StoreCMs eliminating ones that match. As 3062 // long as it's a chain of single users then the optimization is 3063 // safe. Eliminating partially redundant StoreCMs would require 3064 // cloning copies down the other paths. 3065 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 3066 if (adr == mem->in(MemNode::Address) && 3067 val == mem->in(MemNode::ValueIn)) { 3068 // redundant StoreCM 3069 if (mem->req() > MemNode::OopStore) { 3070 // Hasn't been processed by this code yet. 3071 n->add_prec(mem->in(MemNode::OopStore)); 3072 } else { 3073 // Already converted to precedence edge 3074 for (uint i = mem->req(); i < mem->len(); i++) { 3075 // Accumulate any precedence edges 3076 if (mem->in(i) != nullptr) { 3077 n->add_prec(mem->in(i)); 3078 } 3079 } 3080 // Everything above this point has been processed. 3081 done = true; 3082 } 3083 // Eliminate the previous StoreCM 3084 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 3085 assert(mem->outcnt() == 0, "should be dead"); 3086 mem->disconnect_inputs(this); 3087 } else { 3088 prev = mem; 3089 } 3090 mem = prev->in(MemNode::Memory); 3091 } 3092 } 3093 } 3094 3095 //------------------------------final_graph_reshaping_impl---------------------- 3096 // Implement items 1-5 from final_graph_reshaping below. 3097 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3098 3099 if ( n->outcnt() == 0 ) return; // dead node 3100 uint nop = n->Opcode(); 3101 3102 // Check for 2-input instruction with "last use" on right input. 3103 // Swap to left input. Implements item (2). 3104 if( n->req() == 3 && // two-input instruction 3105 n->in(1)->outcnt() > 1 && // left use is NOT a last use 3106 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 3107 n->in(2)->outcnt() == 1 &&// right use IS a last use 3108 !n->in(2)->is_Con() ) { // right use is not a constant 3109 // Check for commutative opcode 3110 switch( nop ) { 3111 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 3112 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 3113 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 3114 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 3115 case Op_AndL: case Op_XorL: case Op_OrL: 3116 case Op_AndI: case Op_XorI: case Op_OrI: { 3117 // Move "last use" input to left by swapping inputs 3118 n->swap_edges(1, 2); 3119 break; 3120 } 3121 default: 3122 break; 3123 } 3124 } 3125 3126 #ifdef ASSERT 3127 if( n->is_Mem() ) { 3128 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 3129 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw || 3130 // oop will be recorded in oop map if load crosses safepoint 3131 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 3132 LoadNode::is_immutable_value(n->in(MemNode::Address))), 3133 "raw memory operations should have control edge"); 3134 } 3135 if (n->is_MemBar()) { 3136 MemBarNode* mb = n->as_MemBar(); 3137 if (mb->trailing_store() || mb->trailing_load_store()) { 3138 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 3139 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 3140 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 3141 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 3142 } else if (mb->leading()) { 3143 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 3144 } 3145 } 3146 #endif 3147 // Count FPU ops and common calls, implements item (3) 3148 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes); 3149 if (!gc_handled) { 3150 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes); 3151 } 3152 3153 // Collect CFG split points 3154 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3155 frc._tests.push(n); 3156 } 3157 } 3158 3159 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) { 3160 switch( nop ) { 3161 // Count all float operations that may use FPU 3162 case Op_AddF: 3163 case Op_SubF: 3164 case Op_MulF: 3165 case Op_DivF: 3166 case Op_NegF: 3167 case Op_ModF: 3168 case Op_ConvI2F: 3169 case Op_ConF: 3170 case Op_CmpF: 3171 case Op_CmpF3: 3172 case Op_StoreF: 3173 case Op_LoadF: 3174 // case Op_ConvL2F: // longs are split into 32-bit halves 3175 frc.inc_float_count(); 3176 break; 3177 3178 case Op_ConvF2D: 3179 case Op_ConvD2F: 3180 frc.inc_float_count(); 3181 frc.inc_double_count(); 3182 break; 3183 3184 // Count all double operations that may use FPU 3185 case Op_AddD: 3186 case Op_SubD: 3187 case Op_MulD: 3188 case Op_DivD: 3189 case Op_NegD: 3190 case Op_ModD: 3191 case Op_ConvI2D: 3192 case Op_ConvD2I: 3193 // case Op_ConvL2D: // handled by leaf call 3194 // case Op_ConvD2L: // handled by leaf call 3195 case Op_ConD: 3196 case Op_CmpD: 3197 case Op_CmpD3: 3198 case Op_StoreD: 3199 case Op_LoadD: 3200 case Op_LoadD_unaligned: 3201 frc.inc_double_count(); 3202 break; 3203 case Op_Opaque1: // Remove Opaque Nodes before matching 3204 case Op_Opaque3: 3205 n->subsume_by(n->in(1), this); 3206 break; 3207 case Op_CallStaticJava: 3208 case Op_CallJava: 3209 case Op_CallDynamicJava: 3210 frc.inc_java_call_count(); // Count java call site; 3211 case Op_CallRuntime: 3212 case Op_CallLeaf: 3213 case Op_CallLeafVector: 3214 case Op_CallLeafNoFP: { 3215 assert (n->is_Call(), ""); 3216 CallNode *call = n->as_Call(); 3217 // Count call sites where the FP mode bit would have to be flipped. 3218 // Do not count uncommon runtime calls: 3219 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3220 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3221 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3222 frc.inc_call_count(); // Count the call site 3223 } else { // See if uncommon argument is shared 3224 Node *n = call->in(TypeFunc::Parms); 3225 int nop = n->Opcode(); 3226 // Clone shared simple arguments to uncommon calls, item (1). 3227 if (n->outcnt() > 1 && 3228 !n->is_Proj() && 3229 nop != Op_CreateEx && 3230 nop != Op_CheckCastPP && 3231 nop != Op_DecodeN && 3232 nop != Op_DecodeNKlass && 3233 !n->is_Mem() && 3234 !n->is_Phi()) { 3235 Node *x = n->clone(); 3236 call->set_req(TypeFunc::Parms, x); 3237 } 3238 } 3239 break; 3240 } 3241 3242 case Op_StoreCM: 3243 { 3244 // Convert OopStore dependence into precedence edge 3245 Node* prec = n->in(MemNode::OopStore); 3246 n->del_req(MemNode::OopStore); 3247 n->add_prec(prec); 3248 eliminate_redundant_card_marks(n); 3249 } 3250 3251 // fall through 3252 3253 case Op_StoreB: 3254 case Op_StoreC: 3255 case Op_StoreI: 3256 case Op_StoreL: 3257 case Op_CompareAndSwapB: 3258 case Op_CompareAndSwapS: 3259 case Op_CompareAndSwapI: 3260 case Op_CompareAndSwapL: 3261 case Op_CompareAndSwapP: 3262 case Op_CompareAndSwapN: 3263 case Op_WeakCompareAndSwapB: 3264 case Op_WeakCompareAndSwapS: 3265 case Op_WeakCompareAndSwapI: 3266 case Op_WeakCompareAndSwapL: 3267 case Op_WeakCompareAndSwapP: 3268 case Op_WeakCompareAndSwapN: 3269 case Op_CompareAndExchangeB: 3270 case Op_CompareAndExchangeS: 3271 case Op_CompareAndExchangeI: 3272 case Op_CompareAndExchangeL: 3273 case Op_CompareAndExchangeP: 3274 case Op_CompareAndExchangeN: 3275 case Op_GetAndAddS: 3276 case Op_GetAndAddB: 3277 case Op_GetAndAddI: 3278 case Op_GetAndAddL: 3279 case Op_GetAndSetS: 3280 case Op_GetAndSetB: 3281 case Op_GetAndSetI: 3282 case Op_GetAndSetL: 3283 case Op_GetAndSetP: 3284 case Op_GetAndSetN: 3285 case Op_StoreP: 3286 case Op_StoreN: 3287 case Op_StoreNKlass: 3288 case Op_LoadB: 3289 case Op_LoadUB: 3290 case Op_LoadUS: 3291 case Op_LoadI: 3292 case Op_LoadKlass: 3293 case Op_LoadNKlass: 3294 case Op_LoadL: 3295 case Op_LoadL_unaligned: 3296 case Op_LoadP: 3297 case Op_LoadN: 3298 case Op_LoadRange: 3299 case Op_LoadS: 3300 break; 3301 3302 case Op_AddP: { // Assert sane base pointers 3303 Node *addp = n->in(AddPNode::Address); 3304 assert( !addp->is_AddP() || 3305 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3306 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3307 "Base pointers must match (addp %u)", addp->_idx ); 3308 #ifdef _LP64 3309 if ((UseCompressedOops || UseCompressedClassPointers) && 3310 addp->Opcode() == Op_ConP && 3311 addp == n->in(AddPNode::Base) && 3312 n->in(AddPNode::Offset)->is_Con()) { 3313 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3314 // on the platform and on the compressed oops mode. 3315 // Use addressing with narrow klass to load with offset on x86. 3316 // Some platforms can use the constant pool to load ConP. 3317 // Do this transformation here since IGVN will convert ConN back to ConP. 3318 const Type* t = addp->bottom_type(); 3319 bool is_oop = t->isa_oopptr() != nullptr; 3320 bool is_klass = t->isa_klassptr() != nullptr; 3321 3322 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3323 (is_klass && Matcher::const_klass_prefer_decode())) { 3324 Node* nn = nullptr; 3325 3326 int op = is_oop ? Op_ConN : Op_ConNKlass; 3327 3328 // Look for existing ConN node of the same exact type. 3329 Node* r = root(); 3330 uint cnt = r->outcnt(); 3331 for (uint i = 0; i < cnt; i++) { 3332 Node* m = r->raw_out(i); 3333 if (m!= nullptr && m->Opcode() == op && 3334 m->bottom_type()->make_ptr() == t) { 3335 nn = m; 3336 break; 3337 } 3338 } 3339 if (nn != nullptr) { 3340 // Decode a narrow oop to match address 3341 // [R12 + narrow_oop_reg<<3 + offset] 3342 if (is_oop) { 3343 nn = new DecodeNNode(nn, t); 3344 } else { 3345 nn = new DecodeNKlassNode(nn, t); 3346 } 3347 // Check for succeeding AddP which uses the same Base. 3348 // Otherwise we will run into the assertion above when visiting that guy. 3349 for (uint i = 0; i < n->outcnt(); ++i) { 3350 Node *out_i = n->raw_out(i); 3351 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3352 out_i->set_req(AddPNode::Base, nn); 3353 #ifdef ASSERT 3354 for (uint j = 0; j < out_i->outcnt(); ++j) { 3355 Node *out_j = out_i->raw_out(j); 3356 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3357 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3358 } 3359 #endif 3360 } 3361 } 3362 n->set_req(AddPNode::Base, nn); 3363 n->set_req(AddPNode::Address, nn); 3364 if (addp->outcnt() == 0) { 3365 addp->disconnect_inputs(this); 3366 } 3367 } 3368 } 3369 } 3370 #endif 3371 break; 3372 } 3373 3374 case Op_CastPP: { 3375 // Remove CastPP nodes to gain more freedom during scheduling but 3376 // keep the dependency they encode as control or precedence edges 3377 // (if control is set already) on memory operations. Some CastPP 3378 // nodes don't have a control (don't carry a dependency): skip 3379 // those. 3380 if (n->in(0) != nullptr) { 3381 ResourceMark rm; 3382 Unique_Node_List wq; 3383 wq.push(n); 3384 for (uint next = 0; next < wq.size(); ++next) { 3385 Node *m = wq.at(next); 3386 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3387 Node* use = m->fast_out(i); 3388 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3389 use->ensure_control_or_add_prec(n->in(0)); 3390 } else { 3391 switch(use->Opcode()) { 3392 case Op_AddP: 3393 case Op_DecodeN: 3394 case Op_DecodeNKlass: 3395 case Op_CheckCastPP: 3396 case Op_CastPP: 3397 wq.push(use); 3398 break; 3399 } 3400 } 3401 } 3402 } 3403 } 3404 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3405 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3406 Node* in1 = n->in(1); 3407 const Type* t = n->bottom_type(); 3408 Node* new_in1 = in1->clone(); 3409 new_in1->as_DecodeN()->set_type(t); 3410 3411 if (!Matcher::narrow_oop_use_complex_address()) { 3412 // 3413 // x86, ARM and friends can handle 2 adds in addressing mode 3414 // and Matcher can fold a DecodeN node into address by using 3415 // a narrow oop directly and do implicit null check in address: 3416 // 3417 // [R12 + narrow_oop_reg<<3 + offset] 3418 // NullCheck narrow_oop_reg 3419 // 3420 // On other platforms (Sparc) we have to keep new DecodeN node and 3421 // use it to do implicit null check in address: 3422 // 3423 // decode_not_null narrow_oop_reg, base_reg 3424 // [base_reg + offset] 3425 // NullCheck base_reg 3426 // 3427 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3428 // to keep the information to which null check the new DecodeN node 3429 // corresponds to use it as value in implicit_null_check(). 3430 // 3431 new_in1->set_req(0, n->in(0)); 3432 } 3433 3434 n->subsume_by(new_in1, this); 3435 if (in1->outcnt() == 0) { 3436 in1->disconnect_inputs(this); 3437 } 3438 } else { 3439 n->subsume_by(n->in(1), this); 3440 if (n->outcnt() == 0) { 3441 n->disconnect_inputs(this); 3442 } 3443 } 3444 break; 3445 } 3446 #ifdef _LP64 3447 case Op_CmpP: 3448 // Do this transformation here to preserve CmpPNode::sub() and 3449 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3450 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3451 Node* in1 = n->in(1); 3452 Node* in2 = n->in(2); 3453 if (!in1->is_DecodeNarrowPtr()) { 3454 in2 = in1; 3455 in1 = n->in(2); 3456 } 3457 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3458 3459 Node* new_in2 = nullptr; 3460 if (in2->is_DecodeNarrowPtr()) { 3461 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3462 new_in2 = in2->in(1); 3463 } else if (in2->Opcode() == Op_ConP) { 3464 const Type* t = in2->bottom_type(); 3465 if (t == TypePtr::NULL_PTR) { 3466 assert(in1->is_DecodeN(), "compare klass to null?"); 3467 // Don't convert CmpP null check into CmpN if compressed 3468 // oops implicit null check is not generated. 3469 // This will allow to generate normal oop implicit null check. 3470 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3471 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3472 // 3473 // This transformation together with CastPP transformation above 3474 // will generated code for implicit null checks for compressed oops. 3475 // 3476 // The original code after Optimize() 3477 // 3478 // LoadN memory, narrow_oop_reg 3479 // decode narrow_oop_reg, base_reg 3480 // CmpP base_reg, nullptr 3481 // CastPP base_reg // NotNull 3482 // Load [base_reg + offset], val_reg 3483 // 3484 // after these transformations will be 3485 // 3486 // LoadN memory, narrow_oop_reg 3487 // CmpN narrow_oop_reg, nullptr 3488 // decode_not_null narrow_oop_reg, base_reg 3489 // Load [base_reg + offset], val_reg 3490 // 3491 // and the uncommon path (== nullptr) will use narrow_oop_reg directly 3492 // since narrow oops can be used in debug info now (see the code in 3493 // final_graph_reshaping_walk()). 3494 // 3495 // At the end the code will be matched to 3496 // on x86: 3497 // 3498 // Load_narrow_oop memory, narrow_oop_reg 3499 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3500 // NullCheck narrow_oop_reg 3501 // 3502 // and on sparc: 3503 // 3504 // Load_narrow_oop memory, narrow_oop_reg 3505 // decode_not_null narrow_oop_reg, base_reg 3506 // Load [base_reg + offset], val_reg 3507 // NullCheck base_reg 3508 // 3509 } else if (t->isa_oopptr()) { 3510 new_in2 = ConNode::make(t->make_narrowoop()); 3511 } else if (t->isa_klassptr()) { 3512 new_in2 = ConNode::make(t->make_narrowklass()); 3513 } 3514 } 3515 if (new_in2 != nullptr) { 3516 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3517 n->subsume_by(cmpN, this); 3518 if (in1->outcnt() == 0) { 3519 in1->disconnect_inputs(this); 3520 } 3521 if (in2->outcnt() == 0) { 3522 in2->disconnect_inputs(this); 3523 } 3524 } 3525 } 3526 break; 3527 3528 case Op_DecodeN: 3529 case Op_DecodeNKlass: 3530 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3531 // DecodeN could be pinned when it can't be fold into 3532 // an address expression, see the code for Op_CastPP above. 3533 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3534 break; 3535 3536 case Op_EncodeP: 3537 case Op_EncodePKlass: { 3538 Node* in1 = n->in(1); 3539 if (in1->is_DecodeNarrowPtr()) { 3540 n->subsume_by(in1->in(1), this); 3541 } else if (in1->Opcode() == Op_ConP) { 3542 const Type* t = in1->bottom_type(); 3543 if (t == TypePtr::NULL_PTR) { 3544 assert(t->isa_oopptr(), "null klass?"); 3545 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3546 } else if (t->isa_oopptr()) { 3547 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3548 } else if (t->isa_klassptr()) { 3549 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3550 } 3551 } 3552 if (in1->outcnt() == 0) { 3553 in1->disconnect_inputs(this); 3554 } 3555 break; 3556 } 3557 3558 case Op_Proj: { 3559 if (OptimizeStringConcat || IncrementalInline) { 3560 ProjNode* proj = n->as_Proj(); 3561 if (proj->_is_io_use) { 3562 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, ""); 3563 // Separate projections were used for the exception path which 3564 // are normally removed by a late inline. If it wasn't inlined 3565 // then they will hang around and should just be replaced with 3566 // the original one. Merge them. 3567 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/); 3568 if (non_io_proj != nullptr) { 3569 proj->subsume_by(non_io_proj , this); 3570 } 3571 } 3572 } 3573 break; 3574 } 3575 3576 case Op_Phi: 3577 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3578 // The EncodeP optimization may create Phi with the same edges 3579 // for all paths. It is not handled well by Register Allocator. 3580 Node* unique_in = n->in(1); 3581 assert(unique_in != nullptr, ""); 3582 uint cnt = n->req(); 3583 for (uint i = 2; i < cnt; i++) { 3584 Node* m = n->in(i); 3585 assert(m != nullptr, ""); 3586 if (unique_in != m) 3587 unique_in = nullptr; 3588 } 3589 if (unique_in != nullptr) { 3590 n->subsume_by(unique_in, this); 3591 } 3592 } 3593 break; 3594 3595 #endif 3596 3597 #ifdef ASSERT 3598 case Op_CastII: 3599 // Verify that all range check dependent CastII nodes were removed. 3600 if (n->isa_CastII()->has_range_check()) { 3601 n->dump(3); 3602 assert(false, "Range check dependent CastII node was not removed"); 3603 } 3604 break; 3605 #endif 3606 3607 case Op_ModI: 3608 if (UseDivMod) { 3609 // Check if a%b and a/b both exist 3610 Node* d = n->find_similar(Op_DivI); 3611 if (d) { 3612 // Replace them with a fused divmod if supported 3613 if (Matcher::has_match_rule(Op_DivModI)) { 3614 DivModINode* divmod = DivModINode::make(n); 3615 d->subsume_by(divmod->div_proj(), this); 3616 n->subsume_by(divmod->mod_proj(), this); 3617 } else { 3618 // replace a%b with a-((a/b)*b) 3619 Node* mult = new MulINode(d, d->in(2)); 3620 Node* sub = new SubINode(d->in(1), mult); 3621 n->subsume_by(sub, this); 3622 } 3623 } 3624 } 3625 break; 3626 3627 case Op_ModL: 3628 if (UseDivMod) { 3629 // Check if a%b and a/b both exist 3630 Node* d = n->find_similar(Op_DivL); 3631 if (d) { 3632 // Replace them with a fused divmod if supported 3633 if (Matcher::has_match_rule(Op_DivModL)) { 3634 DivModLNode* divmod = DivModLNode::make(n); 3635 d->subsume_by(divmod->div_proj(), this); 3636 n->subsume_by(divmod->mod_proj(), this); 3637 } else { 3638 // replace a%b with a-((a/b)*b) 3639 Node* mult = new MulLNode(d, d->in(2)); 3640 Node* sub = new SubLNode(d->in(1), mult); 3641 n->subsume_by(sub, this); 3642 } 3643 } 3644 } 3645 break; 3646 3647 case Op_UModI: 3648 if (UseDivMod) { 3649 // Check if a%b and a/b both exist 3650 Node* d = n->find_similar(Op_UDivI); 3651 if (d) { 3652 // Replace them with a fused unsigned divmod if supported 3653 if (Matcher::has_match_rule(Op_UDivModI)) { 3654 UDivModINode* divmod = UDivModINode::make(n); 3655 d->subsume_by(divmod->div_proj(), this); 3656 n->subsume_by(divmod->mod_proj(), this); 3657 } else { 3658 // replace a%b with a-((a/b)*b) 3659 Node* mult = new MulINode(d, d->in(2)); 3660 Node* sub = new SubINode(d->in(1), mult); 3661 n->subsume_by(sub, this); 3662 } 3663 } 3664 } 3665 break; 3666 3667 case Op_UModL: 3668 if (UseDivMod) { 3669 // Check if a%b and a/b both exist 3670 Node* d = n->find_similar(Op_UDivL); 3671 if (d) { 3672 // Replace them with a fused unsigned divmod if supported 3673 if (Matcher::has_match_rule(Op_UDivModL)) { 3674 UDivModLNode* divmod = UDivModLNode::make(n); 3675 d->subsume_by(divmod->div_proj(), this); 3676 n->subsume_by(divmod->mod_proj(), this); 3677 } else { 3678 // replace a%b with a-((a/b)*b) 3679 Node* mult = new MulLNode(d, d->in(2)); 3680 Node* sub = new SubLNode(d->in(1), mult); 3681 n->subsume_by(sub, this); 3682 } 3683 } 3684 } 3685 break; 3686 3687 case Op_LoadVector: 3688 case Op_StoreVector: 3689 case Op_LoadVectorGather: 3690 case Op_StoreVectorScatter: 3691 case Op_LoadVectorGatherMasked: 3692 case Op_StoreVectorScatterMasked: 3693 case Op_VectorCmpMasked: 3694 case Op_VectorMaskGen: 3695 case Op_LoadVectorMasked: 3696 case Op_StoreVectorMasked: 3697 break; 3698 3699 case Op_AddReductionVI: 3700 case Op_AddReductionVL: 3701 case Op_AddReductionVF: 3702 case Op_AddReductionVD: 3703 case Op_MulReductionVI: 3704 case Op_MulReductionVL: 3705 case Op_MulReductionVF: 3706 case Op_MulReductionVD: 3707 case Op_MinReductionV: 3708 case Op_MaxReductionV: 3709 case Op_AndReductionV: 3710 case Op_OrReductionV: 3711 case Op_XorReductionV: 3712 break; 3713 3714 case Op_PackB: 3715 case Op_PackS: 3716 case Op_PackI: 3717 case Op_PackF: 3718 case Op_PackL: 3719 case Op_PackD: 3720 if (n->req()-1 > 2) { 3721 // Replace many operand PackNodes with a binary tree for matching 3722 PackNode* p = (PackNode*) n; 3723 Node* btp = p->binary_tree_pack(1, n->req()); 3724 n->subsume_by(btp, this); 3725 } 3726 break; 3727 case Op_Loop: 3728 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop"); 3729 case Op_CountedLoop: 3730 case Op_LongCountedLoop: 3731 case Op_OuterStripMinedLoop: 3732 if (n->as_Loop()->is_inner_loop()) { 3733 frc.inc_inner_loop_count(); 3734 } 3735 n->as_Loop()->verify_strip_mined(0); 3736 break; 3737 case Op_LShiftI: 3738 case Op_RShiftI: 3739 case Op_URShiftI: 3740 case Op_LShiftL: 3741 case Op_RShiftL: 3742 case Op_URShiftL: 3743 if (Matcher::need_masked_shift_count) { 3744 // The cpu's shift instructions don't restrict the count to the 3745 // lower 5/6 bits. We need to do the masking ourselves. 3746 Node* in2 = n->in(2); 3747 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3748 const TypeInt* t = in2->find_int_type(); 3749 if (t != nullptr && t->is_con()) { 3750 juint shift = t->get_con(); 3751 if (shift > mask) { // Unsigned cmp 3752 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3753 } 3754 } else { 3755 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) { 3756 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3757 n->set_req(2, shift); 3758 } 3759 } 3760 if (in2->outcnt() == 0) { // Remove dead node 3761 in2->disconnect_inputs(this); 3762 } 3763 } 3764 break; 3765 case Op_MemBarStoreStore: 3766 case Op_MemBarRelease: 3767 // Break the link with AllocateNode: it is no longer useful and 3768 // confuses register allocation. 3769 if (n->req() > MemBarNode::Precedent) { 3770 n->set_req(MemBarNode::Precedent, top()); 3771 } 3772 break; 3773 case Op_MemBarAcquire: { 3774 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3775 // At parse time, the trailing MemBarAcquire for a volatile load 3776 // is created with an edge to the load. After optimizations, 3777 // that input may be a chain of Phis. If those phis have no 3778 // other use, then the MemBarAcquire keeps them alive and 3779 // register allocation can be confused. 3780 dead_nodes.push(n->in(MemBarNode::Precedent)); 3781 n->set_req(MemBarNode::Precedent, top()); 3782 } 3783 break; 3784 } 3785 case Op_Blackhole: 3786 break; 3787 case Op_RangeCheck: { 3788 RangeCheckNode* rc = n->as_RangeCheck(); 3789 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3790 n->subsume_by(iff, this); 3791 frc._tests.push(iff); 3792 break; 3793 } 3794 case Op_ConvI2L: { 3795 if (!Matcher::convi2l_type_required) { 3796 // Code generation on some platforms doesn't need accurate 3797 // ConvI2L types. Widening the type can help remove redundant 3798 // address computations. 3799 n->as_Type()->set_type(TypeLong::INT); 3800 ResourceMark rm; 3801 Unique_Node_List wq; 3802 wq.push(n); 3803 for (uint next = 0; next < wq.size(); next++) { 3804 Node *m = wq.at(next); 3805 3806 for(;;) { 3807 // Loop over all nodes with identical inputs edges as m 3808 Node* k = m->find_similar(m->Opcode()); 3809 if (k == nullptr) { 3810 break; 3811 } 3812 // Push their uses so we get a chance to remove node made 3813 // redundant 3814 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3815 Node* u = k->fast_out(i); 3816 if (u->Opcode() == Op_LShiftL || 3817 u->Opcode() == Op_AddL || 3818 u->Opcode() == Op_SubL || 3819 u->Opcode() == Op_AddP) { 3820 wq.push(u); 3821 } 3822 } 3823 // Replace all nodes with identical edges as m with m 3824 k->subsume_by(m, this); 3825 } 3826 } 3827 } 3828 break; 3829 } 3830 case Op_CmpUL: { 3831 if (!Matcher::has_match_rule(Op_CmpUL)) { 3832 // No support for unsigned long comparisons 3833 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3834 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3835 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3836 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3837 Node* andl = new AndLNode(orl, remove_sign_mask); 3838 Node* cmp = new CmpLNode(andl, n->in(2)); 3839 n->subsume_by(cmp, this); 3840 } 3841 break; 3842 } 3843 default: 3844 assert(!n->is_Call(), ""); 3845 assert(!n->is_Mem(), ""); 3846 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3847 break; 3848 } 3849 } 3850 3851 //------------------------------final_graph_reshaping_walk--------------------- 3852 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3853 // requires that the walk visits a node's inputs before visiting the node. 3854 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3855 Unique_Node_List sfpt; 3856 3857 frc._visited.set(root->_idx); // first, mark node as visited 3858 uint cnt = root->req(); 3859 Node *n = root; 3860 uint i = 0; 3861 while (true) { 3862 if (i < cnt) { 3863 // Place all non-visited non-null inputs onto stack 3864 Node* m = n->in(i); 3865 ++i; 3866 if (m != nullptr && !frc._visited.test_set(m->_idx)) { 3867 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) { 3868 // compute worst case interpreter size in case of a deoptimization 3869 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3870 3871 sfpt.push(m); 3872 } 3873 cnt = m->req(); 3874 nstack.push(n, i); // put on stack parent and next input's index 3875 n = m; 3876 i = 0; 3877 } 3878 } else { 3879 // Now do post-visit work 3880 final_graph_reshaping_impl(n, frc, dead_nodes); 3881 if (nstack.is_empty()) 3882 break; // finished 3883 n = nstack.node(); // Get node from stack 3884 cnt = n->req(); 3885 i = nstack.index(); 3886 nstack.pop(); // Shift to the next node on stack 3887 } 3888 } 3889 3890 // Skip next transformation if compressed oops are not used. 3891 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3892 (!UseCompressedOops && !UseCompressedClassPointers)) 3893 return; 3894 3895 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3896 // It could be done for an uncommon traps or any safepoints/calls 3897 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3898 while (sfpt.size() > 0) { 3899 n = sfpt.pop(); 3900 JVMState *jvms = n->as_SafePoint()->jvms(); 3901 assert(jvms != nullptr, "sanity"); 3902 int start = jvms->debug_start(); 3903 int end = n->req(); 3904 bool is_uncommon = (n->is_CallStaticJava() && 3905 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3906 for (int j = start; j < end; j++) { 3907 Node* in = n->in(j); 3908 if (in->is_DecodeNarrowPtr()) { 3909 bool safe_to_skip = true; 3910 if (!is_uncommon ) { 3911 // Is it safe to skip? 3912 for (uint i = 0; i < in->outcnt(); i++) { 3913 Node* u = in->raw_out(i); 3914 if (!u->is_SafePoint() || 3915 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3916 safe_to_skip = false; 3917 } 3918 } 3919 } 3920 if (safe_to_skip) { 3921 n->set_req(j, in->in(1)); 3922 } 3923 if (in->outcnt() == 0) { 3924 in->disconnect_inputs(this); 3925 } 3926 } 3927 } 3928 } 3929 } 3930 3931 //------------------------------final_graph_reshaping-------------------------- 3932 // Final Graph Reshaping. 3933 // 3934 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3935 // and not commoned up and forced early. Must come after regular 3936 // optimizations to avoid GVN undoing the cloning. Clone constant 3937 // inputs to Loop Phis; these will be split by the allocator anyways. 3938 // Remove Opaque nodes. 3939 // (2) Move last-uses by commutative operations to the left input to encourage 3940 // Intel update-in-place two-address operations and better register usage 3941 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3942 // calls canonicalizing them back. 3943 // (3) Count the number of double-precision FP ops, single-precision FP ops 3944 // and call sites. On Intel, we can get correct rounding either by 3945 // forcing singles to memory (requires extra stores and loads after each 3946 // FP bytecode) or we can set a rounding mode bit (requires setting and 3947 // clearing the mode bit around call sites). The mode bit is only used 3948 // if the relative frequency of single FP ops to calls is low enough. 3949 // This is a key transform for SPEC mpeg_audio. 3950 // (4) Detect infinite loops; blobs of code reachable from above but not 3951 // below. Several of the Code_Gen algorithms fail on such code shapes, 3952 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3953 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3954 // Detection is by looking for IfNodes where only 1 projection is 3955 // reachable from below or CatchNodes missing some targets. 3956 // (5) Assert for insane oop offsets in debug mode. 3957 3958 bool Compile::final_graph_reshaping() { 3959 // an infinite loop may have been eliminated by the optimizer, 3960 // in which case the graph will be empty. 3961 if (root()->req() == 1) { 3962 // Do not compile method that is only a trivial infinite loop, 3963 // since the content of the loop may have been eliminated. 3964 record_method_not_compilable("trivial infinite loop"); 3965 return true; 3966 } 3967 3968 // Expensive nodes have their control input set to prevent the GVN 3969 // from freely commoning them. There's no GVN beyond this point so 3970 // no need to keep the control input. We want the expensive nodes to 3971 // be freely moved to the least frequent code path by gcm. 3972 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3973 for (int i = 0; i < expensive_count(); i++) { 3974 _expensive_nodes.at(i)->set_req(0, nullptr); 3975 } 3976 3977 Final_Reshape_Counts frc; 3978 3979 // Visit everybody reachable! 3980 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3981 Node_Stack nstack(live_nodes() >> 1); 3982 Unique_Node_List dead_nodes; 3983 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes); 3984 3985 // Check for unreachable (from below) code (i.e., infinite loops). 3986 for( uint i = 0; i < frc._tests.size(); i++ ) { 3987 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3988 // Get number of CFG targets. 3989 // Note that PCTables include exception targets after calls. 3990 uint required_outcnt = n->required_outcnt(); 3991 if (n->outcnt() != required_outcnt) { 3992 // Check for a few special cases. Rethrow Nodes never take the 3993 // 'fall-thru' path, so expected kids is 1 less. 3994 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3995 if (n->in(0)->in(0)->is_Call()) { 3996 CallNode* call = n->in(0)->in(0)->as_Call(); 3997 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3998 required_outcnt--; // Rethrow always has 1 less kid 3999 } else if (call->req() > TypeFunc::Parms && 4000 call->is_CallDynamicJava()) { 4001 // Check for null receiver. In such case, the optimizer has 4002 // detected that the virtual call will always result in a null 4003 // pointer exception. The fall-through projection of this CatchNode 4004 // will not be populated. 4005 Node* arg0 = call->in(TypeFunc::Parms); 4006 if (arg0->is_Type() && 4007 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 4008 required_outcnt--; 4009 } 4010 } else if (call->entry_point() == OptoRuntime::new_array_Java() || 4011 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4012 // Check for illegal array length. In such case, the optimizer has 4013 // detected that the allocation attempt will always result in an 4014 // exception. There is no fall-through projection of this CatchNode . 4015 assert(call->is_CallStaticJava(), "static call expected"); 4016 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4017 uint valid_length_test_input = call->req() - 1; 4018 Node* valid_length_test = call->in(valid_length_test_input); 4019 call->del_req(valid_length_test_input); 4020 if (valid_length_test->find_int_con(1) == 0) { 4021 required_outcnt--; 4022 } 4023 dead_nodes.push(valid_length_test); 4024 assert(n->outcnt() == required_outcnt, "malformed control flow"); 4025 continue; 4026 } 4027 } 4028 } 4029 4030 // Recheck with a better notion of 'required_outcnt' 4031 if (n->outcnt() != required_outcnt) { 4032 record_method_not_compilable("malformed control flow"); 4033 return true; // Not all targets reachable! 4034 } 4035 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) { 4036 CallNode* call = n->in(0)->in(0)->as_Call(); 4037 if (call->entry_point() == OptoRuntime::new_array_Java() || 4038 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4039 assert(call->is_CallStaticJava(), "static call expected"); 4040 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4041 uint valid_length_test_input = call->req() - 1; 4042 dead_nodes.push(call->in(valid_length_test_input)); 4043 call->del_req(valid_length_test_input); // valid length test useless now 4044 } 4045 } 4046 // Check that I actually visited all kids. Unreached kids 4047 // must be infinite loops. 4048 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 4049 if (!frc._visited.test(n->fast_out(j)->_idx)) { 4050 record_method_not_compilable("infinite loop"); 4051 return true; // Found unvisited kid; must be unreach 4052 } 4053 4054 // Here so verification code in final_graph_reshaping_walk() 4055 // always see an OuterStripMinedLoopEnd 4056 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) { 4057 IfNode* init_iff = n->as_If(); 4058 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 4059 n->subsume_by(iff, this); 4060 } 4061 } 4062 4063 while (dead_nodes.size() > 0) { 4064 Node* m = dead_nodes.pop(); 4065 if (m->outcnt() == 0 && m != top()) { 4066 for (uint j = 0; j < m->req(); j++) { 4067 Node* in = m->in(j); 4068 if (in != nullptr) { 4069 dead_nodes.push(in); 4070 } 4071 } 4072 m->disconnect_inputs(this); 4073 } 4074 } 4075 4076 #ifdef IA32 4077 // If original bytecodes contained a mixture of floats and doubles 4078 // check if the optimizer has made it homogeneous, item (3). 4079 if (UseSSE == 0 && 4080 frc.get_float_count() > 32 && 4081 frc.get_double_count() == 0 && 4082 (10 * frc.get_call_count() < frc.get_float_count()) ) { 4083 set_24_bit_selection_and_mode(false, true); 4084 } 4085 #endif // IA32 4086 4087 set_java_calls(frc.get_java_call_count()); 4088 set_inner_loops(frc.get_inner_loop_count()); 4089 4090 // No infinite loops, no reason to bail out. 4091 return false; 4092 } 4093 4094 //-----------------------------too_many_traps---------------------------------- 4095 // Report if there are too many traps at the current method and bci. 4096 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 4097 bool Compile::too_many_traps(ciMethod* method, 4098 int bci, 4099 Deoptimization::DeoptReason reason) { 4100 ciMethodData* md = method->method_data(); 4101 if (md->is_empty()) { 4102 // Assume the trap has not occurred, or that it occurred only 4103 // because of a transient condition during start-up in the interpreter. 4104 return false; 4105 } 4106 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4107 if (md->has_trap_at(bci, m, reason) != 0) { 4108 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 4109 // Also, if there are multiple reasons, or if there is no per-BCI record, 4110 // assume the worst. 4111 if (log()) 4112 log()->elem("observe trap='%s' count='%d'", 4113 Deoptimization::trap_reason_name(reason), 4114 md->trap_count(reason)); 4115 return true; 4116 } else { 4117 // Ignore method/bci and see if there have been too many globally. 4118 return too_many_traps(reason, md); 4119 } 4120 } 4121 4122 // Less-accurate variant which does not require a method and bci. 4123 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 4124 ciMethodData* logmd) { 4125 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 4126 // Too many traps globally. 4127 // Note that we use cumulative trap_count, not just md->trap_count. 4128 if (log()) { 4129 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason); 4130 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 4131 Deoptimization::trap_reason_name(reason), 4132 mcount, trap_count(reason)); 4133 } 4134 return true; 4135 } else { 4136 // The coast is clear. 4137 return false; 4138 } 4139 } 4140 4141 //--------------------------too_many_recompiles-------------------------------- 4142 // Report if there are too many recompiles at the current method and bci. 4143 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 4144 // Is not eager to return true, since this will cause the compiler to use 4145 // Action_none for a trap point, to avoid too many recompilations. 4146 bool Compile::too_many_recompiles(ciMethod* method, 4147 int bci, 4148 Deoptimization::DeoptReason reason) { 4149 ciMethodData* md = method->method_data(); 4150 if (md->is_empty()) { 4151 // Assume the trap has not occurred, or that it occurred only 4152 // because of a transient condition during start-up in the interpreter. 4153 return false; 4154 } 4155 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 4156 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 4157 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 4158 Deoptimization::DeoptReason per_bc_reason 4159 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 4160 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4161 if ((per_bc_reason == Deoptimization::Reason_none 4162 || md->has_trap_at(bci, m, reason) != 0) 4163 // The trap frequency measure we care about is the recompile count: 4164 && md->trap_recompiled_at(bci, m) 4165 && md->overflow_recompile_count() >= bc_cutoff) { 4166 // Do not emit a trap here if it has already caused recompilations. 4167 // Also, if there are multiple reasons, or if there is no per-BCI record, 4168 // assume the worst. 4169 if (log()) 4170 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 4171 Deoptimization::trap_reason_name(reason), 4172 md->trap_count(reason), 4173 md->overflow_recompile_count()); 4174 return true; 4175 } else if (trap_count(reason) != 0 4176 && decompile_count() >= m_cutoff) { 4177 // Too many recompiles globally, and we have seen this sort of trap. 4178 // Use cumulative decompile_count, not just md->decompile_count. 4179 if (log()) 4180 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 4181 Deoptimization::trap_reason_name(reason), 4182 md->trap_count(reason), trap_count(reason), 4183 md->decompile_count(), decompile_count()); 4184 return true; 4185 } else { 4186 // The coast is clear. 4187 return false; 4188 } 4189 } 4190 4191 // Compute when not to trap. Used by matching trap based nodes and 4192 // NullCheck optimization. 4193 void Compile::set_allowed_deopt_reasons() { 4194 _allowed_reasons = 0; 4195 if (is_method_compilation()) { 4196 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 4197 assert(rs < BitsPerInt, "recode bit map"); 4198 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 4199 _allowed_reasons |= nth_bit(rs); 4200 } 4201 } 4202 } 4203 } 4204 4205 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 4206 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 4207 } 4208 4209 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 4210 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 4211 } 4212 4213 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 4214 if (holder->is_initialized()) { 4215 return false; 4216 } 4217 if (holder->is_being_initialized()) { 4218 if (accessing_method->holder() == holder) { 4219 // Access inside a class. The barrier can be elided when access happens in <clinit>, 4220 // <init>, or a static method. In all those cases, there was an initialization 4221 // barrier on the holder klass passed. 4222 if (accessing_method->is_static_initializer() || 4223 accessing_method->is_object_initializer() || 4224 accessing_method->is_static()) { 4225 return false; 4226 } 4227 } else if (accessing_method->holder()->is_subclass_of(holder)) { 4228 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 4229 // In case of <init> or a static method, the barrier is on the subclass is not enough: 4230 // child class can become fully initialized while its parent class is still being initialized. 4231 if (accessing_method->is_static_initializer()) { 4232 return false; 4233 } 4234 } 4235 ciMethod* root = method(); // the root method of compilation 4236 if (root != accessing_method) { 4237 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 4238 } 4239 } 4240 return true; 4241 } 4242 4243 #ifndef PRODUCT 4244 //------------------------------verify_bidirectional_edges--------------------- 4245 // For each input edge to a node (ie - for each Use-Def edge), verify that 4246 // there is a corresponding Def-Use edge. 4247 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) { 4248 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc 4249 uint stack_size = live_nodes() >> 4; 4250 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize)); 4251 nstack.push(_root); 4252 4253 while (nstack.size() > 0) { 4254 Node* n = nstack.pop(); 4255 if (visited.member(n)) { 4256 continue; 4257 } 4258 visited.push(n); 4259 4260 // Walk over all input edges, checking for correspondence 4261 uint length = n->len(); 4262 for (uint i = 0; i < length; i++) { 4263 Node* in = n->in(i); 4264 if (in != nullptr && !visited.member(in)) { 4265 nstack.push(in); // Put it on stack 4266 } 4267 if (in != nullptr && !in->is_top()) { 4268 // Count instances of `next` 4269 int cnt = 0; 4270 for (uint idx = 0; idx < in->_outcnt; idx++) { 4271 if (in->_out[idx] == n) { 4272 cnt++; 4273 } 4274 } 4275 assert(cnt > 0, "Failed to find Def-Use edge."); 4276 // Check for duplicate edges 4277 // walk the input array downcounting the input edges to n 4278 for (uint j = 0; j < length; j++) { 4279 if (n->in(j) == in) { 4280 cnt--; 4281 } 4282 } 4283 assert(cnt == 0, "Mismatched edge count."); 4284 } else if (in == nullptr) { 4285 assert(i == 0 || i >= n->req() || 4286 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() || 4287 (n->is_Unlock() && i == (n->req() - 1)) || 4288 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion 4289 "only region, phi, arraycopy, unlock or membar nodes have null data edges"); 4290 } else { 4291 assert(in->is_top(), "sanity"); 4292 // Nothing to check. 4293 } 4294 } 4295 } 4296 } 4297 4298 //------------------------------verify_graph_edges--------------------------- 4299 // Walk the Graph and verify that there is a one-to-one correspondence 4300 // between Use-Def edges and Def-Use edges in the graph. 4301 void Compile::verify_graph_edges(bool no_dead_code) { 4302 if (VerifyGraphEdges) { 4303 Unique_Node_List visited; 4304 4305 // Call graph walk to check edges 4306 verify_bidirectional_edges(visited); 4307 if (no_dead_code) { 4308 // Now make sure that no visited node is used by an unvisited node. 4309 bool dead_nodes = false; 4310 Unique_Node_List checked; 4311 while (visited.size() > 0) { 4312 Node* n = visited.pop(); 4313 checked.push(n); 4314 for (uint i = 0; i < n->outcnt(); i++) { 4315 Node* use = n->raw_out(i); 4316 if (checked.member(use)) continue; // already checked 4317 if (visited.member(use)) continue; // already in the graph 4318 if (use->is_Con()) continue; // a dead ConNode is OK 4319 // At this point, we have found a dead node which is DU-reachable. 4320 if (!dead_nodes) { 4321 tty->print_cr("*** Dead nodes reachable via DU edges:"); 4322 dead_nodes = true; 4323 } 4324 use->dump(2); 4325 tty->print_cr("---"); 4326 checked.push(use); // No repeats; pretend it is now checked. 4327 } 4328 } 4329 assert(!dead_nodes, "using nodes must be reachable from root"); 4330 } 4331 } 4332 } 4333 #endif 4334 4335 // The Compile object keeps track of failure reasons separately from the ciEnv. 4336 // This is required because there is not quite a 1-1 relation between the 4337 // ciEnv and its compilation task and the Compile object. Note that one 4338 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4339 // to backtrack and retry without subsuming loads. Other than this backtracking 4340 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4341 // by the logic in C2Compiler. 4342 void Compile::record_failure(const char* reason) { 4343 if (log() != nullptr) { 4344 log()->elem("failure reason='%s' phase='compile'", reason); 4345 } 4346 if (_failure_reason == nullptr) { 4347 // Record the first failure reason. 4348 _failure_reason = reason; 4349 } 4350 4351 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4352 C->print_method(PHASE_FAILURE, 1); 4353 } 4354 _root = nullptr; // flush the graph, too 4355 } 4356 4357 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 4358 : TraceTime(name, accumulator, CITime, CITimeVerbose), 4359 _compile(Compile::current()), 4360 _log(nullptr), 4361 _phase_name(name), 4362 _dolog(CITimeVerbose) 4363 { 4364 assert(_compile != nullptr, "sanity check"); 4365 if (_dolog) { 4366 _log = _compile->log(); 4367 } 4368 if (_log != nullptr) { 4369 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes()); 4370 _log->stamp(); 4371 _log->end_head(); 4372 } 4373 } 4374 4375 Compile::TracePhase::~TracePhase() { 4376 if (_compile->failing()) return; 4377 #ifdef ASSERT 4378 if (PrintIdealNodeCount) { 4379 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4380 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk()); 4381 } 4382 4383 if (VerifyIdealNodeCount) { 4384 _compile->print_missing_nodes(); 4385 } 4386 #endif 4387 4388 if (_log != nullptr) { 4389 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes()); 4390 } 4391 } 4392 4393 //----------------------------static_subtype_check----------------------------- 4394 // Shortcut important common cases when superklass is exact: 4395 // (0) superklass is java.lang.Object (can occur in reflective code) 4396 // (1) subklass is already limited to a subtype of superklass => always ok 4397 // (2) subklass does not overlap with superklass => always fail 4398 // (3) superklass has NO subtypes and we can check with a simple compare. 4399 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) { 4400 if (skip) { 4401 return SSC_full_test; // Let caller generate the general case. 4402 } 4403 4404 if (subk->is_java_subtype_of(superk)) { 4405 return SSC_always_true; // (0) and (1) this test cannot fail 4406 } 4407 4408 if (!subk->maybe_java_subtype_of(superk)) { 4409 return SSC_always_false; // (2) true path dead; no dynamic test needed 4410 } 4411 4412 const Type* superelem = superk; 4413 if (superk->isa_aryklassptr()) { 4414 int ignored; 4415 superelem = superk->is_aryklassptr()->base_element_type(ignored); 4416 } 4417 4418 if (superelem->isa_instklassptr()) { 4419 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass(); 4420 if (!ik->has_subklass()) { 4421 if (!ik->is_final()) { 4422 // Add a dependency if there is a chance of a later subclass. 4423 dependencies()->assert_leaf_type(ik); 4424 } 4425 if (!superk->maybe_java_subtype_of(subk)) { 4426 return SSC_always_false; 4427 } 4428 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4429 } 4430 } else { 4431 // A primitive array type has no subtypes. 4432 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4433 } 4434 4435 return SSC_full_test; 4436 } 4437 4438 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4439 #ifdef _LP64 4440 // The scaled index operand to AddP must be a clean 64-bit value. 4441 // Java allows a 32-bit int to be incremented to a negative 4442 // value, which appears in a 64-bit register as a large 4443 // positive number. Using that large positive number as an 4444 // operand in pointer arithmetic has bad consequences. 4445 // On the other hand, 32-bit overflow is rare, and the possibility 4446 // can often be excluded, if we annotate the ConvI2L node with 4447 // a type assertion that its value is known to be a small positive 4448 // number. (The prior range check has ensured this.) 4449 // This assertion is used by ConvI2LNode::Ideal. 4450 int index_max = max_jint - 1; // array size is max_jint, index is one less 4451 if (sizetype != nullptr) index_max = sizetype->_hi - 1; 4452 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4453 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4454 #endif 4455 return idx; 4456 } 4457 4458 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4459 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 4460 if (ctrl != nullptr) { 4461 // Express control dependency by a CastII node with a narrow type. 4462 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */); 4463 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4464 // node from floating above the range check during loop optimizations. Otherwise, the 4465 // ConvI2L node may be eliminated independently of the range check, causing the data path 4466 // to become TOP while the control path is still there (although it's unreachable). 4467 value->set_req(0, ctrl); 4468 value = phase->transform(value); 4469 } 4470 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4471 return phase->transform(new ConvI2LNode(value, ltype)); 4472 } 4473 4474 // The message about the current inlining is accumulated in 4475 // _print_inlining_stream and transferred into the _print_inlining_list 4476 // once we know whether inlining succeeds or not. For regular 4477 // inlining, messages are appended to the buffer pointed by 4478 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4479 // a new buffer is added after _print_inlining_idx in the list. This 4480 // way we can update the inlining message for late inlining call site 4481 // when the inlining is attempted again. 4482 void Compile::print_inlining_init() { 4483 if (print_inlining() || print_intrinsics()) { 4484 // print_inlining_init is actually called several times. 4485 print_inlining_reset(); 4486 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer()); 4487 } 4488 } 4489 4490 void Compile::print_inlining_reinit() { 4491 if (print_inlining() || print_intrinsics()) { 4492 print_inlining_reset(); 4493 } 4494 } 4495 4496 void Compile::print_inlining_reset() { 4497 _print_inlining_stream->reset(); 4498 } 4499 4500 void Compile::print_inlining_commit() { 4501 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4502 // Transfer the message from _print_inlining_stream to the current 4503 // _print_inlining_list buffer and clear _print_inlining_stream. 4504 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 4505 print_inlining_reset(); 4506 } 4507 4508 void Compile::print_inlining_push() { 4509 // Add new buffer to the _print_inlining_list at current position 4510 _print_inlining_idx++; 4511 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer()); 4512 } 4513 4514 Compile::PrintInliningBuffer* Compile::print_inlining_current() { 4515 return _print_inlining_list->at(_print_inlining_idx); 4516 } 4517 4518 void Compile::print_inlining_update(CallGenerator* cg) { 4519 if (print_inlining() || print_intrinsics()) { 4520 if (cg->is_late_inline()) { 4521 if (print_inlining_current()->cg() != cg && 4522 (print_inlining_current()->cg() != nullptr || 4523 print_inlining_current()->ss()->size() != 0)) { 4524 print_inlining_push(); 4525 } 4526 print_inlining_commit(); 4527 print_inlining_current()->set_cg(cg); 4528 } else { 4529 if (print_inlining_current()->cg() != nullptr) { 4530 print_inlining_push(); 4531 } 4532 print_inlining_commit(); 4533 } 4534 } 4535 } 4536 4537 void Compile::print_inlining_move_to(CallGenerator* cg) { 4538 // We resume inlining at a late inlining call site. Locate the 4539 // corresponding inlining buffer so that we can update it. 4540 if (print_inlining() || print_intrinsics()) { 4541 for (int i = 0; i < _print_inlining_list->length(); i++) { 4542 if (_print_inlining_list->at(i)->cg() == cg) { 4543 _print_inlining_idx = i; 4544 return; 4545 } 4546 } 4547 ShouldNotReachHere(); 4548 } 4549 } 4550 4551 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4552 if (print_inlining() || print_intrinsics()) { 4553 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4554 assert(print_inlining_current()->cg() == cg, "wrong entry"); 4555 // replace message with new message 4556 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer()); 4557 print_inlining_commit(); 4558 print_inlining_current()->set_cg(cg); 4559 } 4560 } 4561 4562 void Compile::print_inlining_assert_ready() { 4563 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data"); 4564 } 4565 4566 void Compile::process_print_inlining() { 4567 assert(_late_inlines.length() == 0, "not drained yet"); 4568 if (print_inlining() || print_intrinsics()) { 4569 ResourceMark rm; 4570 stringStream ss; 4571 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once."); 4572 for (int i = 0; i < _print_inlining_list->length(); i++) { 4573 PrintInliningBuffer* pib = _print_inlining_list->at(i); 4574 ss.print("%s", pib->ss()->freeze()); 4575 delete pib; 4576 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr)); 4577 } 4578 // Reset _print_inlining_list, it only contains destructed objects. 4579 // It is on the arena, so it will be freed when the arena is reset. 4580 _print_inlining_list = nullptr; 4581 // _print_inlining_stream won't be used anymore, either. 4582 print_inlining_reset(); 4583 size_t end = ss.size(); 4584 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4585 strncpy(_print_inlining_output, ss.freeze(), end+1); 4586 _print_inlining_output[end] = 0; 4587 } 4588 } 4589 4590 void Compile::dump_print_inlining() { 4591 if (_print_inlining_output != nullptr) { 4592 tty->print_raw(_print_inlining_output); 4593 } 4594 } 4595 4596 void Compile::log_late_inline(CallGenerator* cg) { 4597 if (log() != nullptr) { 4598 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4599 cg->unique_id()); 4600 JVMState* p = cg->call_node()->jvms(); 4601 while (p != nullptr) { 4602 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4603 p = p->caller(); 4604 } 4605 log()->tail("late_inline"); 4606 } 4607 } 4608 4609 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4610 log_late_inline(cg); 4611 if (log() != nullptr) { 4612 log()->inline_fail(msg); 4613 } 4614 } 4615 4616 void Compile::log_inline_id(CallGenerator* cg) { 4617 if (log() != nullptr) { 4618 // The LogCompilation tool needs a unique way to identify late 4619 // inline call sites. This id must be unique for this call site in 4620 // this compilation. Try to have it unique across compilations as 4621 // well because it can be convenient when grepping through the log 4622 // file. 4623 // Distinguish OSR compilations from others in case CICountOSR is 4624 // on. 4625 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4626 cg->set_unique_id(id); 4627 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4628 } 4629 } 4630 4631 void Compile::log_inline_failure(const char* msg) { 4632 if (C->log() != nullptr) { 4633 C->log()->inline_fail(msg); 4634 } 4635 } 4636 4637 4638 // Dump inlining replay data to the stream. 4639 // Don't change thread state and acquire any locks. 4640 void Compile::dump_inline_data(outputStream* out) { 4641 InlineTree* inl_tree = ilt(); 4642 if (inl_tree != nullptr) { 4643 out->print(" inline %d", inl_tree->count()); 4644 inl_tree->dump_replay_data(out); 4645 } 4646 } 4647 4648 void Compile::dump_inline_data_reduced(outputStream* out) { 4649 assert(ReplayReduce, ""); 4650 4651 InlineTree* inl_tree = ilt(); 4652 if (inl_tree == nullptr) { 4653 return; 4654 } 4655 // Enable iterative replay file reduction 4656 // Output "compile" lines for depth 1 subtrees, 4657 // simulating that those trees were compiled 4658 // instead of inlined. 4659 for (int i = 0; i < inl_tree->subtrees().length(); ++i) { 4660 InlineTree* sub = inl_tree->subtrees().at(i); 4661 if (sub->inline_level() != 1) { 4662 continue; 4663 } 4664 4665 ciMethod* method = sub->method(); 4666 int entry_bci = -1; 4667 int comp_level = env()->task()->comp_level(); 4668 out->print("compile "); 4669 method->dump_name_as_ascii(out); 4670 out->print(" %d %d", entry_bci, comp_level); 4671 out->print(" inline %d", sub->count()); 4672 sub->dump_replay_data(out, -1); 4673 out->cr(); 4674 } 4675 } 4676 4677 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4678 if (n1->Opcode() < n2->Opcode()) return -1; 4679 else if (n1->Opcode() > n2->Opcode()) return 1; 4680 4681 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4682 for (uint i = 1; i < n1->req(); i++) { 4683 if (n1->in(i) < n2->in(i)) return -1; 4684 else if (n1->in(i) > n2->in(i)) return 1; 4685 } 4686 4687 return 0; 4688 } 4689 4690 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4691 Node* n1 = *n1p; 4692 Node* n2 = *n2p; 4693 4694 return cmp_expensive_nodes(n1, n2); 4695 } 4696 4697 void Compile::sort_expensive_nodes() { 4698 if (!expensive_nodes_sorted()) { 4699 _expensive_nodes.sort(cmp_expensive_nodes); 4700 } 4701 } 4702 4703 bool Compile::expensive_nodes_sorted() const { 4704 for (int i = 1; i < _expensive_nodes.length(); i++) { 4705 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) { 4706 return false; 4707 } 4708 } 4709 return true; 4710 } 4711 4712 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4713 if (_expensive_nodes.length() == 0) { 4714 return false; 4715 } 4716 4717 assert(OptimizeExpensiveOps, "optimization off?"); 4718 4719 // Take this opportunity to remove dead nodes from the list 4720 int j = 0; 4721 for (int i = 0; i < _expensive_nodes.length(); i++) { 4722 Node* n = _expensive_nodes.at(i); 4723 if (!n->is_unreachable(igvn)) { 4724 assert(n->is_expensive(), "should be expensive"); 4725 _expensive_nodes.at_put(j, n); 4726 j++; 4727 } 4728 } 4729 _expensive_nodes.trunc_to(j); 4730 4731 // Then sort the list so that similar nodes are next to each other 4732 // and check for at least two nodes of identical kind with same data 4733 // inputs. 4734 sort_expensive_nodes(); 4735 4736 for (int i = 0; i < _expensive_nodes.length()-1; i++) { 4737 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) { 4738 return true; 4739 } 4740 } 4741 4742 return false; 4743 } 4744 4745 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4746 if (_expensive_nodes.length() == 0) { 4747 return; 4748 } 4749 4750 assert(OptimizeExpensiveOps, "optimization off?"); 4751 4752 // Sort to bring similar nodes next to each other and clear the 4753 // control input of nodes for which there's only a single copy. 4754 sort_expensive_nodes(); 4755 4756 int j = 0; 4757 int identical = 0; 4758 int i = 0; 4759 bool modified = false; 4760 for (; i < _expensive_nodes.length()-1; i++) { 4761 assert(j <= i, "can't write beyond current index"); 4762 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) { 4763 identical++; 4764 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4765 continue; 4766 } 4767 if (identical > 0) { 4768 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4769 identical = 0; 4770 } else { 4771 Node* n = _expensive_nodes.at(i); 4772 igvn.replace_input_of(n, 0, nullptr); 4773 igvn.hash_insert(n); 4774 modified = true; 4775 } 4776 } 4777 if (identical > 0) { 4778 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4779 } else if (_expensive_nodes.length() >= 1) { 4780 Node* n = _expensive_nodes.at(i); 4781 igvn.replace_input_of(n, 0, nullptr); 4782 igvn.hash_insert(n); 4783 modified = true; 4784 } 4785 _expensive_nodes.trunc_to(j); 4786 if (modified) { 4787 igvn.optimize(); 4788 } 4789 } 4790 4791 void Compile::add_expensive_node(Node * n) { 4792 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list"); 4793 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4794 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4795 if (OptimizeExpensiveOps) { 4796 _expensive_nodes.append(n); 4797 } else { 4798 // Clear control input and let IGVN optimize expensive nodes if 4799 // OptimizeExpensiveOps is off. 4800 n->set_req(0, nullptr); 4801 } 4802 } 4803 4804 /** 4805 * Track coarsened Lock and Unlock nodes. 4806 */ 4807 4808 class Lock_List : public Node_List { 4809 uint _origin_cnt; 4810 public: 4811 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {} 4812 uint origin_cnt() const { return _origin_cnt; } 4813 }; 4814 4815 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) { 4816 int length = locks.length(); 4817 if (length > 0) { 4818 // Have to keep this list until locks elimination during Macro nodes elimination. 4819 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length); 4820 for (int i = 0; i < length; i++) { 4821 AbstractLockNode* lock = locks.at(i); 4822 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx); 4823 locks_list->push(lock); 4824 } 4825 _coarsened_locks.append(locks_list); 4826 } 4827 } 4828 4829 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) { 4830 int count = coarsened_count(); 4831 for (int i = 0; i < count; i++) { 4832 Node_List* locks_list = _coarsened_locks.at(i); 4833 for (uint j = 0; j < locks_list->size(); j++) { 4834 Node* lock = locks_list->at(j); 4835 assert(lock->is_AbstractLock(), "sanity"); 4836 if (!useful.member(lock)) { 4837 locks_list->yank(lock); 4838 } 4839 } 4840 } 4841 } 4842 4843 void Compile::remove_coarsened_lock(Node* n) { 4844 if (n->is_AbstractLock()) { 4845 int count = coarsened_count(); 4846 for (int i = 0; i < count; i++) { 4847 Node_List* locks_list = _coarsened_locks.at(i); 4848 locks_list->yank(n); 4849 } 4850 } 4851 } 4852 4853 bool Compile::coarsened_locks_consistent() { 4854 int count = coarsened_count(); 4855 for (int i = 0; i < count; i++) { 4856 bool unbalanced = false; 4857 bool modified = false; // track locks kind modifications 4858 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i); 4859 uint size = locks_list->size(); 4860 if (size == 0) { 4861 unbalanced = false; // All locks were eliminated - good 4862 } else if (size != locks_list->origin_cnt()) { 4863 unbalanced = true; // Some locks were removed from list 4864 } else { 4865 for (uint j = 0; j < size; j++) { 4866 Node* lock = locks_list->at(j); 4867 // All nodes in group should have the same state (modified or not) 4868 if (!lock->as_AbstractLock()->is_coarsened()) { 4869 if (j == 0) { 4870 // first on list was modified, the rest should be too for consistency 4871 modified = true; 4872 } else if (!modified) { 4873 // this lock was modified but previous locks on the list were not 4874 unbalanced = true; 4875 break; 4876 } 4877 } else if (modified) { 4878 // previous locks on list were modified but not this lock 4879 unbalanced = true; 4880 break; 4881 } 4882 } 4883 } 4884 if (unbalanced) { 4885 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified 4886 #ifdef ASSERT 4887 if (PrintEliminateLocks) { 4888 tty->print_cr("=== unbalanced coarsened locks ==="); 4889 for (uint l = 0; l < size; l++) { 4890 locks_list->at(l)->dump(); 4891 } 4892 } 4893 #endif 4894 record_failure(C2Compiler::retry_no_locks_coarsening()); 4895 return false; 4896 } 4897 } 4898 return true; 4899 } 4900 4901 /** 4902 * Remove the speculative part of types and clean up the graph 4903 */ 4904 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4905 if (UseTypeSpeculation) { 4906 Unique_Node_List worklist; 4907 worklist.push(root()); 4908 int modified = 0; 4909 // Go over all type nodes that carry a speculative type, drop the 4910 // speculative part of the type and enqueue the node for an igvn 4911 // which may optimize it out. 4912 for (uint next = 0; next < worklist.size(); ++next) { 4913 Node *n = worklist.at(next); 4914 if (n->is_Type()) { 4915 TypeNode* tn = n->as_Type(); 4916 const Type* t = tn->type(); 4917 const Type* t_no_spec = t->remove_speculative(); 4918 if (t_no_spec != t) { 4919 bool in_hash = igvn.hash_delete(n); 4920 #ifdef ASSERT 4921 if (!in_hash) { 4922 tty->print_cr("current graph:"); 4923 n->dump_bfs(MaxNodeLimit, nullptr, "S$"); 4924 tty->cr(); 4925 tty->print_cr("erroneous node:"); 4926 n->dump(); 4927 assert(false, "node should be in igvn hash table"); 4928 } 4929 #endif 4930 tn->set_type(t_no_spec); 4931 igvn.hash_insert(n); 4932 igvn._worklist.push(n); // give it a chance to go away 4933 modified++; 4934 } 4935 } 4936 // Iterate over outs - endless loops is unreachable from below 4937 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4938 Node *m = n->fast_out(i); 4939 if (not_a_node(m)) { 4940 continue; 4941 } 4942 worklist.push(m); 4943 } 4944 } 4945 // Drop the speculative part of all types in the igvn's type table 4946 igvn.remove_speculative_types(); 4947 if (modified > 0) { 4948 igvn.optimize(); 4949 if (failing()) return; 4950 } 4951 #ifdef ASSERT 4952 // Verify that after the IGVN is over no speculative type has resurfaced 4953 worklist.clear(); 4954 worklist.push(root()); 4955 for (uint next = 0; next < worklist.size(); ++next) { 4956 Node *n = worklist.at(next); 4957 const Type* t = igvn.type_or_null(n); 4958 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types"); 4959 if (n->is_Type()) { 4960 t = n->as_Type()->type(); 4961 assert(t == t->remove_speculative(), "no more speculative types"); 4962 } 4963 // Iterate over outs - endless loops is unreachable from below 4964 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4965 Node *m = n->fast_out(i); 4966 if (not_a_node(m)) { 4967 continue; 4968 } 4969 worklist.push(m); 4970 } 4971 } 4972 igvn.check_no_speculative_types(); 4973 #endif 4974 } 4975 } 4976 4977 // Auxiliary methods to support randomized stressing/fuzzing. 4978 4979 int Compile::random() { 4980 _stress_seed = os::next_random(_stress_seed); 4981 return static_cast<int>(_stress_seed); 4982 } 4983 4984 // This method can be called the arbitrary number of times, with current count 4985 // as the argument. The logic allows selecting a single candidate from the 4986 // running list of candidates as follows: 4987 // int count = 0; 4988 // Cand* selected = null; 4989 // while(cand = cand->next()) { 4990 // if (randomized_select(++count)) { 4991 // selected = cand; 4992 // } 4993 // } 4994 // 4995 // Including count equalizes the chances any candidate is "selected". 4996 // This is useful when we don't have the complete list of candidates to choose 4997 // from uniformly. In this case, we need to adjust the randomicity of the 4998 // selection, or else we will end up biasing the selection towards the latter 4999 // candidates. 5000 // 5001 // Quick back-envelope calculation shows that for the list of n candidates 5002 // the equal probability for the candidate to persist as "best" can be 5003 // achieved by replacing it with "next" k-th candidate with the probability 5004 // of 1/k. It can be easily shown that by the end of the run, the 5005 // probability for any candidate is converged to 1/n, thus giving the 5006 // uniform distribution among all the candidates. 5007 // 5008 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 5009 #define RANDOMIZED_DOMAIN_POW 29 5010 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 5011 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 5012 bool Compile::randomized_select(int count) { 5013 assert(count > 0, "only positive"); 5014 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 5015 } 5016 5017 CloneMap& Compile::clone_map() { return _clone_map; } 5018 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 5019 5020 void NodeCloneInfo::dump_on(outputStream* st) const { 5021 st->print(" {%d:%d} ", idx(), gen()); 5022 } 5023 5024 void CloneMap::clone(Node* old, Node* nnn, int gen) { 5025 uint64_t val = value(old->_idx); 5026 NodeCloneInfo cio(val); 5027 assert(val != 0, "old node should be in the map"); 5028 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 5029 insert(nnn->_idx, cin.get()); 5030 #ifndef PRODUCT 5031 if (is_debug()) { 5032 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 5033 } 5034 #endif 5035 } 5036 5037 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 5038 NodeCloneInfo cio(value(old->_idx)); 5039 if (cio.get() == 0) { 5040 cio.set(old->_idx, 0); 5041 insert(old->_idx, cio.get()); 5042 #ifndef PRODUCT 5043 if (is_debug()) { 5044 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 5045 } 5046 #endif 5047 } 5048 clone(old, nnn, gen); 5049 } 5050 5051 int CloneMap::max_gen() const { 5052 int g = 0; 5053 DictI di(_dict); 5054 for(; di.test(); ++di) { 5055 int t = gen(di._key); 5056 if (g < t) { 5057 g = t; 5058 #ifndef PRODUCT 5059 if (is_debug()) { 5060 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 5061 } 5062 #endif 5063 } 5064 } 5065 return g; 5066 } 5067 5068 void CloneMap::dump(node_idx_t key, outputStream* st) const { 5069 uint64_t val = value(key); 5070 if (val != 0) { 5071 NodeCloneInfo ni(val); 5072 ni.dump_on(st); 5073 } 5074 } 5075 5076 // Move Allocate nodes to the start of the list 5077 void Compile::sort_macro_nodes() { 5078 int count = macro_count(); 5079 int allocates = 0; 5080 for (int i = 0; i < count; i++) { 5081 Node* n = macro_node(i); 5082 if (n->is_Allocate()) { 5083 if (i != allocates) { 5084 Node* tmp = macro_node(allocates); 5085 _macro_nodes.at_put(allocates, n); 5086 _macro_nodes.at_put(i, tmp); 5087 } 5088 allocates++; 5089 } 5090 } 5091 } 5092 5093 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) { 5094 if (failing()) { return; } 5095 EventCompilerPhase event; 5096 if (event.should_commit()) { 5097 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 5098 } 5099 #ifndef PRODUCT 5100 ResourceMark rm; 5101 stringStream ss; 5102 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt)); 5103 if (n != nullptr) { 5104 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]); 5105 } 5106 5107 const char* name = ss.as_string(); 5108 if (should_print_igv(level)) { 5109 _igv_printer->print_method(name, level); 5110 } 5111 if (should_print_phase(cpt)) { 5112 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt)); 5113 } 5114 #endif 5115 C->_latest_stage_start_counter.stamp(); 5116 } 5117 5118 // Only used from CompileWrapper 5119 void Compile::begin_method() { 5120 #ifndef PRODUCT 5121 if (_method != nullptr && should_print_igv(1)) { 5122 _igv_printer->begin_method(); 5123 } 5124 #endif 5125 C->_latest_stage_start_counter.stamp(); 5126 } 5127 5128 // Only used from CompileWrapper 5129 void Compile::end_method() { 5130 EventCompilerPhase event; 5131 if (event.should_commit()) { 5132 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1); 5133 } 5134 5135 #ifndef PRODUCT 5136 if (_method != nullptr && should_print_igv(1)) { 5137 _igv_printer->end_method(); 5138 } 5139 #endif 5140 } 5141 5142 bool Compile::should_print_phase(CompilerPhaseType cpt) { 5143 #ifndef PRODUCT 5144 if ((_directive->ideal_phase_mask() & CompilerPhaseTypeHelper::to_bitmask(cpt)) != 0) { 5145 return true; 5146 } 5147 #endif 5148 return false; 5149 } 5150 5151 bool Compile::should_print_igv(const int level) { 5152 #ifndef PRODUCT 5153 if (PrintIdealGraphLevel < 0) { // disabled by the user 5154 return false; 5155 } 5156 5157 bool need = directive()->IGVPrintLevelOption >= level; 5158 if (need && !_igv_printer) { 5159 _igv_printer = IdealGraphPrinter::printer(); 5160 _igv_printer->set_compile(this); 5161 } 5162 return need; 5163 #else 5164 return false; 5165 #endif 5166 } 5167 5168 #ifndef PRODUCT 5169 IdealGraphPrinter* Compile::_debug_file_printer = nullptr; 5170 IdealGraphPrinter* Compile::_debug_network_printer = nullptr; 5171 5172 // Called from debugger. Prints method to the default file with the default phase name. 5173 // This works regardless of any Ideal Graph Visualizer flags set or not. 5174 void igv_print() { 5175 Compile::current()->igv_print_method_to_file(); 5176 } 5177 5178 // Same as igv_print() above but with a specified phase name. 5179 void igv_print(const char* phase_name) { 5180 Compile::current()->igv_print_method_to_file(phase_name); 5181 } 5182 5183 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 5184 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 5185 // This works regardless of any Ideal Graph Visualizer flags set or not. 5186 void igv_print(bool network) { 5187 if (network) { 5188 Compile::current()->igv_print_method_to_network(); 5189 } else { 5190 Compile::current()->igv_print_method_to_file(); 5191 } 5192 } 5193 5194 // Same as igv_print(bool network) above but with a specified phase name. 5195 void igv_print(bool network, const char* phase_name) { 5196 if (network) { 5197 Compile::current()->igv_print_method_to_network(phase_name); 5198 } else { 5199 Compile::current()->igv_print_method_to_file(phase_name); 5200 } 5201 } 5202 5203 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 5204 void igv_print_default() { 5205 Compile::current()->print_method(PHASE_DEBUG, 0); 5206 } 5207 5208 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 5209 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 5210 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 5211 void igv_append() { 5212 Compile::current()->igv_print_method_to_file("Debug", true); 5213 } 5214 5215 // Same as igv_append() above but with a specified phase name. 5216 void igv_append(const char* phase_name) { 5217 Compile::current()->igv_print_method_to_file(phase_name, true); 5218 } 5219 5220 void Compile::igv_print_method_to_file(const char* phase_name, bool append) { 5221 const char* file_name = "custom_debug.xml"; 5222 if (_debug_file_printer == nullptr) { 5223 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 5224 } else { 5225 _debug_file_printer->update_compiled_method(C->method()); 5226 } 5227 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 5228 _debug_file_printer->print(phase_name, (Node*)C->root()); 5229 } 5230 5231 void Compile::igv_print_method_to_network(const char* phase_name) { 5232 if (_debug_network_printer == nullptr) { 5233 _debug_network_printer = new IdealGraphPrinter(C); 5234 } else { 5235 _debug_network_printer->update_compiled_method(C->method()); 5236 } 5237 tty->print_cr("Method printed over network stream to IGV"); 5238 _debug_network_printer->print(phase_name, (Node*)C->root()); 5239 } 5240 #endif 5241 5242 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 5243 if (type != nullptr && phase->type(value)->higher_equal(type)) { 5244 return value; 5245 } 5246 Node* result = nullptr; 5247 if (bt == T_BYTE) { 5248 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 5249 result = new RShiftINode(result, phase->intcon(24)); 5250 } else if (bt == T_BOOLEAN) { 5251 result = new AndINode(value, phase->intcon(0xFF)); 5252 } else if (bt == T_CHAR) { 5253 result = new AndINode(value,phase->intcon(0xFFFF)); 5254 } else { 5255 assert(bt == T_SHORT, "unexpected narrow type"); 5256 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 5257 result = new RShiftINode(result, phase->intcon(16)); 5258 } 5259 if (transform_res) { 5260 result = phase->transform(result); 5261 } 5262 return result; 5263 } 5264 5265 void Compile::record_method_not_compilable_oom() { 5266 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit()); 5267 }