1 /* 2 * Copyright (c) 1997, 2024, 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/compilationFailureInfo.hpp" 33 #include "compiler/compilationMemoryStatistic.hpp" 34 #include "compiler/compileBroker.hpp" 35 #include "compiler/compileLog.hpp" 36 #include "compiler/compiler_globals.hpp" 37 #include "compiler/disassembler.hpp" 38 #include "compiler/oopMap.hpp" 39 #include "gc/shared/barrierSet.hpp" 40 #include "gc/shared/c2/barrierSetC2.hpp" 41 #include "jfr/jfrEvents.hpp" 42 #include "jvm_io.h" 43 #include "memory/allocation.hpp" 44 #include "memory/resourceArea.hpp" 45 #include "opto/addnode.hpp" 46 #include "opto/block.hpp" 47 #include "opto/c2compiler.hpp" 48 #include "opto/callGenerator.hpp" 49 #include "opto/callnode.hpp" 50 #include "opto/castnode.hpp" 51 #include "opto/cfgnode.hpp" 52 #include "opto/chaitin.hpp" 53 #include "opto/compile.hpp" 54 #include "opto/connode.hpp" 55 #include "opto/convertnode.hpp" 56 #include "opto/divnode.hpp" 57 #include "opto/escape.hpp" 58 #include "opto/idealGraphPrinter.hpp" 59 #include "opto/loopnode.hpp" 60 #include "opto/machnode.hpp" 61 #include "opto/macro.hpp" 62 #include "opto/matcher.hpp" 63 #include "opto/mathexactnode.hpp" 64 #include "opto/memnode.hpp" 65 #include "opto/mulnode.hpp" 66 #include "opto/narrowptrnode.hpp" 67 #include "opto/node.hpp" 68 #include "opto/opcodes.hpp" 69 #include "opto/output.hpp" 70 #include "opto/parse.hpp" 71 #include "opto/phaseX.hpp" 72 #include "opto/rootnode.hpp" 73 #include "opto/runtime.hpp" 74 #include "opto/stringopts.hpp" 75 #include "opto/type.hpp" 76 #include "opto/vector.hpp" 77 #include "opto/vectornode.hpp" 78 #include "runtime/globals_extension.hpp" 79 #include "runtime/sharedRuntime.hpp" 80 #include "runtime/signature.hpp" 81 #include "runtime/stubRoutines.hpp" 82 #include "runtime/timer.hpp" 83 #include "utilities/align.hpp" 84 #include "utilities/copy.hpp" 85 #include "utilities/macros.hpp" 86 #include "utilities/resourceHash.hpp" 87 88 // -------------------- Compile::mach_constant_base_node ----------------------- 89 // Constant table base node singleton. 90 MachConstantBaseNode* Compile::mach_constant_base_node() { 91 if (_mach_constant_base_node == nullptr) { 92 _mach_constant_base_node = new MachConstantBaseNode(); 93 _mach_constant_base_node->add_req(C->root()); 94 } 95 return _mach_constant_base_node; 96 } 97 98 99 /// Support for intrinsics. 100 101 // Return the index at which m must be inserted (or already exists). 102 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 103 class IntrinsicDescPair { 104 private: 105 ciMethod* _m; 106 bool _is_virtual; 107 public: 108 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 109 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 110 ciMethod* m= elt->method(); 111 ciMethod* key_m = key->_m; 112 if (key_m < m) return -1; 113 else if (key_m > m) return 1; 114 else { 115 bool is_virtual = elt->is_virtual(); 116 bool key_virtual = key->_is_virtual; 117 if (key_virtual < is_virtual) return -1; 118 else if (key_virtual > is_virtual) return 1; 119 else return 0; 120 } 121 } 122 }; 123 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 124 #ifdef ASSERT 125 for (int i = 1; i < _intrinsics.length(); i++) { 126 CallGenerator* cg1 = _intrinsics.at(i-1); 127 CallGenerator* cg2 = _intrinsics.at(i); 128 assert(cg1->method() != cg2->method() 129 ? cg1->method() < cg2->method() 130 : cg1->is_virtual() < cg2->is_virtual(), 131 "compiler intrinsics list must stay sorted"); 132 } 133 #endif 134 IntrinsicDescPair pair(m, is_virtual); 135 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 136 } 137 138 void Compile::register_intrinsic(CallGenerator* cg) { 139 bool found = false; 140 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 141 assert(!found, "registering twice"); 142 _intrinsics.insert_before(index, cg); 143 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 144 } 145 146 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 147 assert(m->is_loaded(), "don't try this on unloaded methods"); 148 if (_intrinsics.length() > 0) { 149 bool found = false; 150 int index = intrinsic_insertion_index(m, is_virtual, found); 151 if (found) { 152 return _intrinsics.at(index); 153 } 154 } 155 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 156 if (m->intrinsic_id() != vmIntrinsics::_none && 157 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 158 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 159 if (cg != nullptr) { 160 // Save it for next time: 161 register_intrinsic(cg); 162 return cg; 163 } else { 164 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 165 } 166 } 167 return nullptr; 168 } 169 170 // Compile::make_vm_intrinsic is defined in library_call.cpp. 171 172 #ifndef PRODUCT 173 // statistics gathering... 174 175 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0}; 176 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0}; 177 178 inline int as_int(vmIntrinsics::ID id) { 179 return vmIntrinsics::as_int(id); 180 } 181 182 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 183 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 184 int oflags = _intrinsic_hist_flags[as_int(id)]; 185 assert(flags != 0, "what happened?"); 186 if (is_virtual) { 187 flags |= _intrinsic_virtual; 188 } 189 bool changed = (flags != oflags); 190 if ((flags & _intrinsic_worked) != 0) { 191 juint count = (_intrinsic_hist_count[as_int(id)] += 1); 192 if (count == 1) { 193 changed = true; // first time 194 } 195 // increment the overall count also: 196 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1; 197 } 198 if (changed) { 199 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 200 // Something changed about the intrinsic's virtuality. 201 if ((flags & _intrinsic_virtual) != 0) { 202 // This is the first use of this intrinsic as a virtual call. 203 if (oflags != 0) { 204 // We already saw it as a non-virtual, so note both cases. 205 flags |= _intrinsic_both; 206 } 207 } else if ((oflags & _intrinsic_both) == 0) { 208 // This is the first use of this intrinsic as a non-virtual 209 flags |= _intrinsic_both; 210 } 211 } 212 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags); 213 } 214 // update the overall flags also: 215 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags; 216 return changed; 217 } 218 219 static char* format_flags(int flags, char* buf) { 220 buf[0] = 0; 221 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 222 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 223 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 224 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 225 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 226 if (buf[0] == 0) strcat(buf, ","); 227 assert(buf[0] == ',', "must be"); 228 return &buf[1]; 229 } 230 231 void Compile::print_intrinsic_statistics() { 232 char flagsbuf[100]; 233 ttyLocker ttyl; 234 if (xtty != nullptr) xtty->head("statistics type='intrinsic'"); 235 tty->print_cr("Compiler intrinsic usage:"); 236 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)]; 237 if (total == 0) total = 1; // avoid div0 in case of no successes 238 #define PRINT_STAT_LINE(name, c, f) \ 239 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 240 for (auto id : EnumRange<vmIntrinsicID>{}) { 241 int flags = _intrinsic_hist_flags[as_int(id)]; 242 juint count = _intrinsic_hist_count[as_int(id)]; 243 if ((flags | count) != 0) { 244 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 245 } 246 } 247 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf)); 248 if (xtty != nullptr) xtty->tail("statistics"); 249 } 250 251 void Compile::print_statistics() { 252 { ttyLocker ttyl; 253 if (xtty != nullptr) xtty->head("statistics type='opto'"); 254 Parse::print_statistics(); 255 PhaseStringOpts::print_statistics(); 256 PhaseCCP::print_statistics(); 257 PhaseRegAlloc::print_statistics(); 258 PhaseOutput::print_statistics(); 259 PhasePeephole::print_statistics(); 260 PhaseIdealLoop::print_statistics(); 261 ConnectionGraph::print_statistics(); 262 PhaseMacroExpand::print_statistics(); 263 if (xtty != nullptr) xtty->tail("statistics"); 264 } 265 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) { 266 // put this under its own <statistics> element. 267 print_intrinsic_statistics(); 268 } 269 } 270 #endif //PRODUCT 271 272 void Compile::gvn_replace_by(Node* n, Node* nn) { 273 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 274 Node* use = n->last_out(i); 275 bool is_in_table = initial_gvn()->hash_delete(use); 276 uint uses_found = 0; 277 for (uint j = 0; j < use->len(); j++) { 278 if (use->in(j) == n) { 279 if (j < use->req()) 280 use->set_req(j, nn); 281 else 282 use->set_prec(j, nn); 283 uses_found++; 284 } 285 } 286 if (is_in_table) { 287 // reinsert into table 288 initial_gvn()->hash_find_insert(use); 289 } 290 record_for_igvn(use); 291 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist); 292 i -= uses_found; // we deleted 1 or more copies of this edge 293 } 294 } 295 296 297 // Identify all nodes that are reachable from below, useful. 298 // Use breadth-first pass that records state in a Unique_Node_List, 299 // recursive traversal is slower. 300 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 301 int estimated_worklist_size = live_nodes(); 302 useful.map( estimated_worklist_size, nullptr ); // preallocate space 303 304 // Initialize worklist 305 if (root() != nullptr) { useful.push(root()); } 306 // If 'top' is cached, declare it useful to preserve cached node 307 if (cached_top_node()) { useful.push(cached_top_node()); } 308 309 // Push all useful nodes onto the list, breadthfirst 310 for( uint next = 0; next < useful.size(); ++next ) { 311 assert( next < unique(), "Unique useful nodes < total nodes"); 312 Node *n = useful.at(next); 313 uint max = n->len(); 314 for( uint i = 0; i < max; ++i ) { 315 Node *m = n->in(i); 316 if (not_a_node(m)) continue; 317 useful.push(m); 318 } 319 } 320 } 321 322 // Update dead_node_list with any missing dead nodes using useful 323 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 324 void Compile::update_dead_node_list(Unique_Node_List &useful) { 325 uint max_idx = unique(); 326 VectorSet& useful_node_set = useful.member_set(); 327 328 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 329 // If node with index node_idx is not in useful set, 330 // mark it as dead in dead node list. 331 if (!useful_node_set.test(node_idx)) { 332 record_dead_node(node_idx); 333 } 334 } 335 } 336 337 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 338 int shift = 0; 339 for (int i = 0; i < inlines->length(); i++) { 340 CallGenerator* cg = inlines->at(i); 341 if (useful.member(cg->call_node())) { 342 if (shift > 0) { 343 inlines->at_put(i - shift, cg); 344 } 345 } else { 346 shift++; // skip over the dead element 347 } 348 } 349 if (shift > 0) { 350 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array 351 } 352 } 353 354 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) { 355 assert(dead != nullptr && dead->is_Call(), "sanity"); 356 int found = 0; 357 for (int i = 0; i < inlines->length(); i++) { 358 if (inlines->at(i)->call_node() == dead) { 359 inlines->remove_at(i); 360 found++; 361 NOT_DEBUG( break; ) // elements are unique, so exit early 362 } 363 } 364 assert(found <= 1, "not unique"); 365 } 366 367 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)> 368 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) { 369 for (int i = node_list.length() - 1; i >= 0; i--) { 370 N* node = node_list.at(i); 371 if (!useful.member(node)) { 372 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 373 } 374 } 375 } 376 377 void Compile::remove_useless_node(Node* dead) { 378 remove_modified_node(dead); 379 380 // Constant node that has no out-edges and has only one in-edge from 381 // root is usually dead. However, sometimes reshaping walk makes 382 // it reachable by adding use edges. So, we will NOT count Con nodes 383 // as dead to be conservative about the dead node count at any 384 // given time. 385 if (!dead->is_Con()) { 386 record_dead_node(dead->_idx); 387 } 388 if (dead->is_macro()) { 389 remove_macro_node(dead); 390 } 391 if (dead->is_expensive()) { 392 remove_expensive_node(dead); 393 } 394 if (dead->Opcode() == Op_Opaque4) { 395 remove_template_assertion_predicate_opaq(dead); 396 } 397 if (dead->is_ParsePredicate()) { 398 remove_parse_predicate(dead->as_ParsePredicate()); 399 } 400 if (dead->for_post_loop_opts_igvn()) { 401 remove_from_post_loop_opts_igvn(dead); 402 } 403 if (dead->is_Call()) { 404 remove_useless_late_inlines( &_late_inlines, dead); 405 remove_useless_late_inlines( &_string_late_inlines, dead); 406 remove_useless_late_inlines( &_boxing_late_inlines, dead); 407 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead); 408 409 if (dead->is_CallStaticJava()) { 410 remove_unstable_if_trap(dead->as_CallStaticJava(), false); 411 } 412 } 413 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 414 bs->unregister_potential_barrier_node(dead); 415 } 416 417 // Disconnect all useless nodes by disconnecting those at the boundary. 418 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) { 419 uint next = 0; 420 while (next < useful.size()) { 421 Node *n = useful.at(next++); 422 if (n->is_SafePoint()) { 423 // We're done with a parsing phase. Replaced nodes are not valid 424 // beyond that point. 425 n->as_SafePoint()->delete_replaced_nodes(); 426 } 427 // Use raw traversal of out edges since this code removes out edges 428 int max = n->outcnt(); 429 for (int j = 0; j < max; ++j) { 430 Node* child = n->raw_out(j); 431 if (!useful.member(child)) { 432 assert(!child->is_top() || child != top(), 433 "If top is cached in Compile object it is in useful list"); 434 // Only need to remove this out-edge to the useless node 435 n->raw_del_out(j); 436 --j; 437 --max; 438 } 439 } 440 if (n->outcnt() == 1 && n->has_special_unique_user()) { 441 assert(useful.member(n->unique_out()), "do not push a useless node"); 442 worklist.push(n->unique_out()); 443 } 444 } 445 446 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes 447 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes 448 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes 449 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes 450 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass 451 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps 452 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes 453 #ifdef ASSERT 454 if (_modified_nodes != nullptr) { 455 _modified_nodes->remove_useless_nodes(useful.member_set()); 456 } 457 #endif 458 459 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 460 bs->eliminate_useless_gc_barriers(useful, this); 461 // clean up the late inline lists 462 remove_useless_late_inlines( &_late_inlines, useful); 463 remove_useless_late_inlines( &_string_late_inlines, useful); 464 remove_useless_late_inlines( &_boxing_late_inlines, useful); 465 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful); 466 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 467 } 468 469 // ============================================================================ 470 //------------------------------CompileWrapper--------------------------------- 471 class CompileWrapper : public StackObj { 472 Compile *const _compile; 473 public: 474 CompileWrapper(Compile* compile); 475 476 ~CompileWrapper(); 477 }; 478 479 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 480 // the Compile* pointer is stored in the current ciEnv: 481 ciEnv* env = compile->env(); 482 assert(env == ciEnv::current(), "must already be a ciEnv active"); 483 assert(env->compiler_data() == nullptr, "compile already active?"); 484 env->set_compiler_data(compile); 485 assert(compile == Compile::current(), "sanity"); 486 487 compile->set_type_dict(nullptr); 488 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 489 compile->clone_map().set_clone_idx(0); 490 compile->set_type_last_size(0); 491 compile->set_last_tf(nullptr, nullptr); 492 compile->set_indexSet_arena(nullptr); 493 compile->set_indexSet_free_block_list(nullptr); 494 compile->init_type_arena(); 495 Type::Initialize(compile); 496 _compile->begin_method(); 497 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 498 } 499 CompileWrapper::~CompileWrapper() { 500 // simulate crash during compilation 501 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned"); 502 503 _compile->end_method(); 504 _compile->env()->set_compiler_data(nullptr); 505 } 506 507 508 //----------------------------print_compile_messages--------------------------- 509 void Compile::print_compile_messages() { 510 #ifndef PRODUCT 511 // Check if recompiling 512 if (!subsume_loads() && PrintOpto) { 513 // Recompiling without allowing machine instructions to subsume loads 514 tty->print_cr("*********************************************************"); 515 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 516 tty->print_cr("*********************************************************"); 517 } 518 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) { 519 // Recompiling without escape analysis 520 tty->print_cr("*********************************************************"); 521 tty->print_cr("** Bailout: Recompile without escape analysis **"); 522 tty->print_cr("*********************************************************"); 523 } 524 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) { 525 // Recompiling without iterative escape analysis 526 tty->print_cr("*********************************************************"); 527 tty->print_cr("** Bailout: Recompile without iterative escape analysis**"); 528 tty->print_cr("*********************************************************"); 529 } 530 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) { 531 // Recompiling without reducing allocation merges 532 tty->print_cr("*********************************************************"); 533 tty->print_cr("** Bailout: Recompile without reduce allocation merges **"); 534 tty->print_cr("*********************************************************"); 535 } 536 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) { 537 // Recompiling without boxing elimination 538 tty->print_cr("*********************************************************"); 539 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 540 tty->print_cr("*********************************************************"); 541 } 542 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) { 543 // Recompiling without locks coarsening 544 tty->print_cr("*********************************************************"); 545 tty->print_cr("** Bailout: Recompile without locks coarsening **"); 546 tty->print_cr("*********************************************************"); 547 } 548 if (env()->break_at_compile()) { 549 // Open the debugger when compiling this method. 550 tty->print("### Breaking when compiling: "); 551 method()->print_short_name(); 552 tty->cr(); 553 BREAKPOINT; 554 } 555 556 if( PrintOpto ) { 557 if (is_osr_compilation()) { 558 tty->print("[OSR]%3d", _compile_id); 559 } else { 560 tty->print("%3d", _compile_id); 561 } 562 } 563 #endif 564 } 565 566 #ifndef PRODUCT 567 void Compile::print_ideal_ir(const char* phase_name) { 568 // keep the following output all in one block 569 // This output goes directly to the tty, not the compiler log. 570 // To enable tools to match it up with the compilation activity, 571 // be sure to tag this tty output with the compile ID. 572 573 // Node dumping can cause a safepoint, which can break the tty lock. 574 // Buffer all node dumps, so that all safepoints happen before we lock. 575 ResourceMark rm; 576 stringStream ss; 577 578 if (_output == nullptr) { 579 ss.print_cr("AFTER: %s", phase_name); 580 // Print out all nodes in ascending order of index. 581 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss); 582 } else { 583 // Dump the node blockwise if we have a scheduling 584 _output->print_scheduling(&ss); 585 } 586 587 // Check that the lock is not broken by a safepoint. 588 NoSafepointVerifier nsv; 589 ttyLocker ttyl; 590 if (xtty != nullptr) { 591 xtty->head("ideal compile_id='%d'%s compile_phase='%s'", 592 compile_id(), 593 is_osr_compilation() ? " compile_kind='osr'" : "", 594 phase_name); 595 } 596 597 tty->print("%s", ss.as_string()); 598 599 if (xtty != nullptr) { 600 xtty->tail("ideal"); 601 } 602 } 603 #endif 604 605 // ============================================================================ 606 //------------------------------Compile standard------------------------------- 607 608 // Compile a method. entry_bci is -1 for normal compilations and indicates 609 // the continuation bci for on stack replacement. 610 611 612 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci, 613 Options options, DirectiveSet* directive) 614 : Phase(Compiler), 615 _compile_id(ci_env->compile_id()), 616 _options(options), 617 _method(target), 618 _entry_bci(osr_bci), 619 _ilt(nullptr), 620 _stub_function(nullptr), 621 _stub_name(nullptr), 622 _stub_entry_point(nullptr), 623 _max_node_limit(MaxNodeLimit), 624 _post_loop_opts_phase(false), 625 _inlining_progress(false), 626 _inlining_incrementally(false), 627 _do_cleanup(false), 628 _has_reserved_stack_access(target->has_reserved_stack_access()), 629 #ifndef PRODUCT 630 _igv_idx(0), 631 _trace_opto_output(directive->TraceOptoOutputOption), 632 #endif 633 _has_method_handle_invokes(false), 634 _clinit_barrier_on_entry(false), 635 _stress_seed(0), 636 _comp_arena(mtCompiler), 637 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 638 _env(ci_env), 639 _directive(directive), 640 _log(ci_env->log()), 641 _failure_reason(nullptr), 642 _first_failure_details(nullptr), 643 _intrinsics (comp_arena(), 0, 0, nullptr), 644 _macro_nodes (comp_arena(), 8, 0, nullptr), 645 _parse_predicates (comp_arena(), 8, 0, nullptr), 646 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr), 647 _expensive_nodes (comp_arena(), 8, 0, nullptr), 648 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr), 649 _unstable_if_traps (comp_arena(), 8, 0, nullptr), 650 _coarsened_locks (comp_arena(), 8, 0, nullptr), 651 _pea(nullptr), 652 _congraph(nullptr), 653 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 654 _unique(0), 655 _dead_node_count(0), 656 _dead_node_list(comp_arena()), 657 _node_arena_one(mtCompiler, Arena::Tag::tag_node), 658 _node_arena_two(mtCompiler, Arena::Tag::tag_node), 659 _node_arena(&_node_arena_one), 660 _mach_constant_base_node(nullptr), 661 _Compile_types(mtCompiler), 662 _initial_gvn(nullptr), 663 _igvn_worklist(nullptr), 664 _types(nullptr), 665 _node_hash(nullptr), 666 _late_inlines(comp_arena(), 2, 0, nullptr), 667 _string_late_inlines(comp_arena(), 2, 0, nullptr), 668 _boxing_late_inlines(comp_arena(), 2, 0, nullptr), 669 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr), 670 _late_inlines_pos(0), 671 _number_of_mh_late_inlines(0), 672 _oom(false), 673 _print_inlining_stream(new (mtCompiler) stringStream()), 674 _print_inlining_list(nullptr), 675 _print_inlining_idx(0), 676 _print_inlining_output(nullptr), 677 _replay_inline_data(nullptr), 678 _java_calls(0), 679 _inner_loops(0), 680 _interpreter_frame_size(0), 681 _output(nullptr) 682 #ifndef PRODUCT 683 , _in_dump_cnt(0) 684 #endif 685 { 686 C = this; 687 CompileWrapper cw(this); 688 689 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 690 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false); 691 692 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 693 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 694 // We can always print a disassembly, either abstract (hex dump) or 695 // with the help of a suitable hsdis library. Thus, we should not 696 // couple print_assembly and print_opto_assembly controls. 697 // But: always print opto and regular assembly on compile command 'print'. 698 bool print_assembly = directive->PrintAssemblyOption; 699 set_print_assembly(print_opto_assembly || print_assembly); 700 #else 701 set_print_assembly(false); // must initialize. 702 #endif 703 704 #ifndef PRODUCT 705 set_parsed_irreducible_loop(false); 706 #endif 707 708 if (directive->ReplayInlineOption) { 709 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 710 } 711 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 712 set_print_intrinsics(directive->PrintIntrinsicsOption); 713 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 714 715 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 716 // Make sure the method being compiled gets its own MDO, 717 // so we can at least track the decompile_count(). 718 // Need MDO to record RTM code generation state. 719 method()->ensure_method_data(); 720 } 721 722 Init(/*do_aliasing=*/ true); 723 724 print_compile_messages(); 725 726 _ilt = InlineTree::build_inline_tree_root(); 727 728 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 729 assert(num_alias_types() >= AliasIdxRaw, ""); 730 731 #define MINIMUM_NODE_HASH 1023 732 733 // GVN that will be run immediately on new nodes 734 uint estimated_size = method()->code_size()*4+64; 735 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 736 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 737 _types = new (comp_arena()) Type_Array(comp_arena()); 738 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size); 739 PhaseGVN gvn; 740 set_initial_gvn(&gvn); 741 742 if (DoPartialEscapeAnalysis) { 743 _pea = new (mtCompiler) PartialEscapeAnalysis(comp_arena()); 744 } 745 746 print_inlining_init(); 747 { // Scope for timing the parser 748 TracePhase tp("parse", &timers[_t_parser]); 749 750 // Put top into the hash table ASAP. 751 initial_gvn()->transform(top()); 752 753 // Set up tf(), start(), and find a CallGenerator. 754 CallGenerator* cg = nullptr; 755 if (is_osr_compilation()) { 756 const TypeTuple *domain = StartOSRNode::osr_domain(); 757 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 758 init_tf(TypeFunc::make(domain, range)); 759 StartNode* s = new StartOSRNode(root(), domain); 760 initial_gvn()->set_type_bottom(s); 761 init_start(s); 762 cg = CallGenerator::for_osr(method(), entry_bci()); 763 } else { 764 // Normal case. 765 init_tf(TypeFunc::make(method())); 766 StartNode* s = new StartNode(root(), tf()->domain()); 767 initial_gvn()->set_type_bottom(s); 768 init_start(s); 769 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 770 // With java.lang.ref.reference.get() we must go through the 771 // intrinsic - even when get() is the root 772 // method of the compile - so that, if necessary, the value in 773 // the referent field of the reference object gets recorded by 774 // the pre-barrier code. 775 cg = find_intrinsic(method(), false); 776 } 777 if (cg == nullptr) { 778 float past_uses = method()->interpreter_invocation_count(); 779 float expected_uses = past_uses; 780 cg = CallGenerator::for_inline(method(), expected_uses); 781 } 782 } 783 if (failing()) return; 784 if (cg == nullptr) { 785 const char* reason = InlineTree::check_can_parse(method()); 786 assert(reason != nullptr, "expect reason for parse failure"); 787 stringStream ss; 788 ss.print("cannot parse method: %s", reason); 789 record_method_not_compilable(ss.as_string()); 790 return; 791 } 792 793 gvn.set_type(root(), root()->bottom_type()); 794 795 JVMState* jvms = build_start_state(start(), tf()); 796 if ((jvms = cg->generate(jvms)) == nullptr) { 797 assert(failure_reason() != nullptr, "expect reason for parse failure"); 798 stringStream ss; 799 ss.print("method parse failed: %s", failure_reason()); 800 record_method_not_compilable(ss.as_string()); 801 return; 802 } 803 GraphKit kit(jvms); 804 805 if (!kit.stopped()) { 806 // Accept return values, and transfer control we know not where. 807 // This is done by a special, unique ReturnNode bound to root. 808 return_values(kit.jvms()); 809 } 810 811 if (kit.has_exceptions()) { 812 // Any exceptions that escape from this call must be rethrown 813 // to whatever caller is dynamically above us on the stack. 814 // This is done by a special, unique RethrowNode bound to root. 815 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 816 } 817 818 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 819 820 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 821 inline_string_calls(true); 822 } 823 824 if (failing()) return; 825 826 // Remove clutter produced by parsing. 827 if (!failing()) { 828 ResourceMark rm; 829 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 830 } 831 } 832 833 // Note: Large methods are capped off in do_one_bytecode(). 834 if (failing()) return; 835 836 // After parsing, node notes are no longer automagic. 837 // They must be propagated by register_new_node_with_optimizer(), 838 // clone(), or the like. 839 set_default_node_notes(nullptr); 840 841 #ifndef PRODUCT 842 if (should_print_igv(1)) { 843 _igv_printer->print_inlining(); 844 } 845 #endif 846 847 if (failing()) return; 848 NOT_PRODUCT( verify_graph_edges(); ) 849 850 // If any phase is randomized for stress testing, seed random number 851 // generation and log the seed for repeatability. 852 if (StressLCM || StressGCM || StressIGVN || StressCCP || StressIncrementalInlining) { 853 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) { 854 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds()); 855 FLAG_SET_ERGO(StressSeed, _stress_seed); 856 } else { 857 _stress_seed = StressSeed; 858 } 859 if (_log != nullptr) { 860 _log->elem("stress_test seed='%u'", _stress_seed); 861 } 862 } 863 864 // Now optimize 865 Optimize(); 866 if (failing()) return; 867 NOT_PRODUCT( verify_graph_edges(); ) 868 869 #ifndef PRODUCT 870 if (should_print_ideal()) { 871 print_ideal_ir("print_ideal"); 872 } 873 #endif 874 875 #ifdef ASSERT 876 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 877 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 878 #endif 879 880 // Dump compilation data to replay it. 881 if (directive->DumpReplayOption) { 882 env()->dump_replay_data(_compile_id); 883 } 884 if (directive->DumpInlineOption && (ilt() != nullptr)) { 885 env()->dump_inline_data(_compile_id); 886 } 887 888 // Now that we know the size of all the monitors we can add a fixed slot 889 // for the original deopt pc. 890 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size); 891 set_fixed_slots(next_slot); 892 893 // Compute when to use implicit null checks. Used by matching trap based 894 // nodes and NullCheck optimization. 895 set_allowed_deopt_reasons(); 896 897 // Now generate code 898 Code_Gen(); 899 } 900 901 //------------------------------Compile---------------------------------------- 902 // Compile a runtime stub 903 Compile::Compile( ciEnv* ci_env, 904 TypeFunc_generator generator, 905 address stub_function, 906 const char *stub_name, 907 int is_fancy_jump, 908 bool pass_tls, 909 bool return_pc, 910 DirectiveSet* directive) 911 : Phase(Compiler), 912 _compile_id(0), 913 _options(Options::for_runtime_stub()), 914 _method(nullptr), 915 _entry_bci(InvocationEntryBci), 916 _stub_function(stub_function), 917 _stub_name(stub_name), 918 _stub_entry_point(nullptr), 919 _max_node_limit(MaxNodeLimit), 920 _post_loop_opts_phase(false), 921 _inlining_progress(false), 922 _inlining_incrementally(false), 923 _has_reserved_stack_access(false), 924 #ifndef PRODUCT 925 _igv_idx(0), 926 _trace_opto_output(directive->TraceOptoOutputOption), 927 #endif 928 _has_method_handle_invokes(false), 929 _clinit_barrier_on_entry(false), 930 _stress_seed(0), 931 _comp_arena(mtCompiler), 932 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 933 _env(ci_env), 934 _directive(directive), 935 _log(ci_env->log()), 936 _failure_reason(nullptr), 937 _first_failure_details(nullptr), 938 _congraph(nullptr), 939 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 940 _unique(0), 941 _dead_node_count(0), 942 _dead_node_list(comp_arena()), 943 _node_arena_one(mtCompiler), 944 _node_arena_two(mtCompiler), 945 _node_arena(&_node_arena_one), 946 _mach_constant_base_node(nullptr), 947 _Compile_types(mtCompiler), 948 _initial_gvn(nullptr), 949 _igvn_worklist(nullptr), 950 _types(nullptr), 951 _node_hash(nullptr), 952 _number_of_mh_late_inlines(0), 953 _oom(false), 954 _print_inlining_stream(new (mtCompiler) stringStream()), 955 _print_inlining_list(nullptr), 956 _print_inlining_idx(0), 957 _print_inlining_output(nullptr), 958 _replay_inline_data(nullptr), 959 _java_calls(0), 960 _inner_loops(0), 961 _interpreter_frame_size(0), 962 _output(nullptr), 963 #ifndef PRODUCT 964 _in_dump_cnt(0), 965 #endif 966 _allowed_reasons(0) { 967 C = this; 968 969 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false); 970 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false); 971 972 #ifndef PRODUCT 973 set_print_assembly(PrintFrameConverterAssembly); 974 set_parsed_irreducible_loop(false); 975 #else 976 set_print_assembly(false); // Must initialize. 977 #endif 978 set_has_irreducible_loop(false); // no loops 979 980 CompileWrapper cw(this); 981 Init(/*do_aliasing=*/ false); 982 init_tf((*generator)()); 983 984 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 985 _types = new (comp_arena()) Type_Array(comp_arena()); 986 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255); 987 { 988 PhaseGVN gvn; 989 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 990 gvn.transform(top()); 991 992 GraphKit kit; 993 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 994 } 995 996 NOT_PRODUCT( verify_graph_edges(); ) 997 998 Code_Gen(); 999 } 1000 1001 Compile::~Compile() { 1002 delete _print_inlining_stream; 1003 delete _first_failure_details; 1004 }; 1005 1006 //------------------------------Init------------------------------------------- 1007 // Prepare for a single compilation 1008 void Compile::Init(bool aliasing) { 1009 _do_aliasing = aliasing; 1010 _unique = 0; 1011 _regalloc = nullptr; 1012 1013 _tf = nullptr; // filled in later 1014 _top = nullptr; // cached later 1015 _matcher = nullptr; // filled in later 1016 _cfg = nullptr; // filled in later 1017 1018 IA32_ONLY( set_24_bit_selection_and_mode(true, false); ) 1019 1020 _node_note_array = nullptr; 1021 _default_node_notes = nullptr; 1022 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize() 1023 1024 _immutable_memory = nullptr; // filled in at first inquiry 1025 1026 #ifdef ASSERT 1027 _phase_optimize_finished = false; 1028 _exception_backedge = false; 1029 _type_verify = nullptr; 1030 #endif 1031 1032 // Globally visible Nodes 1033 // First set TOP to null to give safe behavior during creation of RootNode 1034 set_cached_top_node(nullptr); 1035 set_root(new RootNode()); 1036 // Now that you have a Root to point to, create the real TOP 1037 set_cached_top_node( new ConNode(Type::TOP) ); 1038 set_recent_alloc(nullptr, nullptr); 1039 1040 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1041 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1042 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1043 env()->set_dependencies(new Dependencies(env())); 1044 1045 _fixed_slots = 0; 1046 set_has_split_ifs(false); 1047 set_has_loops(false); // first approximation 1048 set_has_stringbuilder(false); 1049 set_has_boxed_value(false); 1050 _trap_can_recompile = false; // no traps emitted yet 1051 _major_progress = true; // start out assuming good things will happen 1052 set_has_unsafe_access(false); 1053 set_max_vector_size(0); 1054 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1055 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1056 set_decompile_count(0); 1057 1058 #ifndef PRODUCT 1059 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter)); 1060 #endif 1061 1062 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1063 _loop_opts_cnt = LoopOptsCount; 1064 set_do_inlining(Inline); 1065 set_max_inline_size(MaxInlineSize); 1066 set_freq_inline_size(FreqInlineSize); 1067 set_do_scheduling(OptoScheduling); 1068 1069 set_do_vector_loop(false); 1070 set_has_monitors(false); 1071 1072 if (AllowVectorizeOnDemand) { 1073 if (has_method() && _directive->VectorizeOption) { 1074 set_do_vector_loop(true); 1075 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());}) 1076 } else if (has_method() && method()->name() != 0 && 1077 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1078 set_do_vector_loop(true); 1079 } 1080 } 1081 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1082 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());}) 1083 1084 set_rtm_state(NoRTM); // No RTM lock eliding by default 1085 _max_node_limit = _directive->MaxNodeLimitOption; 1086 1087 #if INCLUDE_RTM_OPT 1088 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != nullptr)) { 1089 int rtm_state = method()->method_data()->rtm_state(); 1090 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) { 1091 // Don't generate RTM lock eliding code. 1092 set_rtm_state(NoRTM); 1093 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1094 // Generate RTM lock eliding code without abort ratio calculation code. 1095 set_rtm_state(UseRTM); 1096 } else if (UseRTMDeopt) { 1097 // Generate RTM lock eliding code and include abort ratio calculation 1098 // code if UseRTMDeopt is on. 1099 set_rtm_state(ProfileRTM); 1100 } 1101 } 1102 #endif 1103 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1104 set_clinit_barrier_on_entry(true); 1105 } 1106 if (debug_info()->recording_non_safepoints()) { 1107 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1108 (comp_arena(), 8, 0, nullptr)); 1109 set_default_node_notes(Node_Notes::make(this)); 1110 } 1111 1112 const int grow_ats = 16; 1113 _max_alias_types = grow_ats; 1114 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1115 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1116 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1117 { 1118 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1119 } 1120 // Initialize the first few types. 1121 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr); 1122 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1123 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1124 _num_alias_types = AliasIdxRaw+1; 1125 // Zero out the alias type cache. 1126 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1127 // A null adr_type hits in the cache right away. Preload the right answer. 1128 probe_alias_cache(nullptr)->_index = AliasIdxTop; 1129 } 1130 1131 //---------------------------init_start---------------------------------------- 1132 // Install the StartNode on this compile object. 1133 void Compile::init_start(StartNode* s) { 1134 if (failing()) 1135 return; // already failing 1136 assert(s == start(), ""); 1137 } 1138 1139 /** 1140 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1141 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1142 * the ideal graph. 1143 */ 1144 StartNode* Compile::start() const { 1145 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1146 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1147 Node* start = root()->fast_out(i); 1148 if (start->is_Start()) { 1149 return start->as_Start(); 1150 } 1151 } 1152 fatal("Did not find Start node!"); 1153 return nullptr; 1154 } 1155 1156 //-------------------------------immutable_memory------------------------------------- 1157 // Access immutable memory 1158 Node* Compile::immutable_memory() { 1159 if (_immutable_memory != nullptr) { 1160 return _immutable_memory; 1161 } 1162 StartNode* s = start(); 1163 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1164 Node *p = s->fast_out(i); 1165 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1166 _immutable_memory = p; 1167 return _immutable_memory; 1168 } 1169 } 1170 ShouldNotReachHere(); 1171 return nullptr; 1172 } 1173 1174 //----------------------set_cached_top_node------------------------------------ 1175 // Install the cached top node, and make sure Node::is_top works correctly. 1176 void Compile::set_cached_top_node(Node* tn) { 1177 if (tn != nullptr) verify_top(tn); 1178 Node* old_top = _top; 1179 _top = tn; 1180 // Calling Node::setup_is_top allows the nodes the chance to adjust 1181 // their _out arrays. 1182 if (_top != nullptr) _top->setup_is_top(); 1183 if (old_top != nullptr) old_top->setup_is_top(); 1184 assert(_top == nullptr || top()->is_top(), ""); 1185 } 1186 1187 #ifdef ASSERT 1188 uint Compile::count_live_nodes_by_graph_walk() { 1189 Unique_Node_List useful(comp_arena()); 1190 // Get useful node list by walking the graph. 1191 identify_useful_nodes(useful); 1192 return useful.size(); 1193 } 1194 1195 void Compile::print_missing_nodes() { 1196 1197 // Return if CompileLog is null and PrintIdealNodeCount is false. 1198 if ((_log == nullptr) && (! PrintIdealNodeCount)) { 1199 return; 1200 } 1201 1202 // This is an expensive function. It is executed only when the user 1203 // specifies VerifyIdealNodeCount option or otherwise knows the 1204 // additional work that needs to be done to identify reachable nodes 1205 // by walking the flow graph and find the missing ones using 1206 // _dead_node_list. 1207 1208 Unique_Node_List useful(comp_arena()); 1209 // Get useful node list by walking the graph. 1210 identify_useful_nodes(useful); 1211 1212 uint l_nodes = C->live_nodes(); 1213 uint l_nodes_by_walk = useful.size(); 1214 1215 if (l_nodes != l_nodes_by_walk) { 1216 if (_log != nullptr) { 1217 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1218 _log->stamp(); 1219 _log->end_head(); 1220 } 1221 VectorSet& useful_member_set = useful.member_set(); 1222 int last_idx = l_nodes_by_walk; 1223 for (int i = 0; i < last_idx; i++) { 1224 if (useful_member_set.test(i)) { 1225 if (_dead_node_list.test(i)) { 1226 if (_log != nullptr) { 1227 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1228 } 1229 if (PrintIdealNodeCount) { 1230 // Print the log message to tty 1231 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1232 useful.at(i)->dump(); 1233 } 1234 } 1235 } 1236 else if (! _dead_node_list.test(i)) { 1237 if (_log != nullptr) { 1238 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1239 } 1240 if (PrintIdealNodeCount) { 1241 // Print the log message to tty 1242 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1243 } 1244 } 1245 } 1246 if (_log != nullptr) { 1247 _log->tail("mismatched_nodes"); 1248 } 1249 } 1250 } 1251 void Compile::record_modified_node(Node* n) { 1252 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) { 1253 _modified_nodes->push(n); 1254 } 1255 } 1256 1257 void Compile::remove_modified_node(Node* n) { 1258 if (_modified_nodes != nullptr) { 1259 _modified_nodes->remove(n); 1260 } 1261 } 1262 #endif 1263 1264 #ifndef PRODUCT 1265 void Compile::verify_top(Node* tn) const { 1266 if (tn != nullptr) { 1267 assert(tn->is_Con(), "top node must be a constant"); 1268 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1269 assert(tn->in(0) != nullptr, "must have live top node"); 1270 } 1271 } 1272 #endif 1273 1274 1275 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1276 1277 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1278 guarantee(arr != nullptr, ""); 1279 int num_blocks = arr->length(); 1280 if (grow_by < num_blocks) grow_by = num_blocks; 1281 int num_notes = grow_by * _node_notes_block_size; 1282 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1283 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1284 while (num_notes > 0) { 1285 arr->append(notes); 1286 notes += _node_notes_block_size; 1287 num_notes -= _node_notes_block_size; 1288 } 1289 assert(num_notes == 0, "exact multiple, please"); 1290 } 1291 1292 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1293 if (source == nullptr || dest == nullptr) return false; 1294 1295 if (dest->is_Con()) 1296 return false; // Do not push debug info onto constants. 1297 1298 #ifdef ASSERT 1299 // Leave a bread crumb trail pointing to the original node: 1300 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) { 1301 dest->set_debug_orig(source); 1302 } 1303 #endif 1304 1305 if (node_note_array() == nullptr) 1306 return false; // Not collecting any notes now. 1307 1308 // This is a copy onto a pre-existing node, which may already have notes. 1309 // If both nodes have notes, do not overwrite any pre-existing notes. 1310 Node_Notes* source_notes = node_notes_at(source->_idx); 1311 if (source_notes == nullptr || source_notes->is_clear()) return false; 1312 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1313 if (dest_notes == nullptr || dest_notes->is_clear()) { 1314 return set_node_notes_at(dest->_idx, source_notes); 1315 } 1316 1317 Node_Notes merged_notes = (*source_notes); 1318 // The order of operations here ensures that dest notes will win... 1319 merged_notes.update_from(dest_notes); 1320 return set_node_notes_at(dest->_idx, &merged_notes); 1321 } 1322 1323 1324 //--------------------------allow_range_check_smearing------------------------- 1325 // Gating condition for coalescing similar range checks. 1326 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1327 // single covering check that is at least as strong as any of them. 1328 // If the optimization succeeds, the simplified (strengthened) range check 1329 // will always succeed. If it fails, we will deopt, and then give up 1330 // on the optimization. 1331 bool Compile::allow_range_check_smearing() const { 1332 // If this method has already thrown a range-check, 1333 // assume it was because we already tried range smearing 1334 // and it failed. 1335 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1336 return !already_trapped; 1337 } 1338 1339 1340 //------------------------------flatten_alias_type----------------------------- 1341 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1342 assert(do_aliasing(), "Aliasing should be enabled"); 1343 int offset = tj->offset(); 1344 TypePtr::PTR ptr = tj->ptr(); 1345 1346 // Known instance (scalarizable allocation) alias only with itself. 1347 bool is_known_inst = tj->isa_oopptr() != nullptr && 1348 tj->is_oopptr()->is_known_instance(); 1349 1350 // Process weird unsafe references. 1351 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1352 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops"); 1353 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1354 tj = TypeOopPtr::BOTTOM; 1355 ptr = tj->ptr(); 1356 offset = tj->offset(); 1357 } 1358 1359 // Array pointers need some flattening 1360 const TypeAryPtr* ta = tj->isa_aryptr(); 1361 if (ta && ta->is_stable()) { 1362 // Erase stability property for alias analysis. 1363 tj = ta = ta->cast_to_stable(false); 1364 } 1365 if( ta && is_known_inst ) { 1366 if ( offset != Type::OffsetBot && 1367 offset > arrayOopDesc::length_offset_in_bytes() ) { 1368 offset = Type::OffsetBot; // Flatten constant access into array body only 1369 tj = ta = ta-> 1370 remove_speculative()-> 1371 cast_to_ptr_type(ptr)-> 1372 with_offset(offset); 1373 } 1374 } else if (ta) { 1375 // For arrays indexed by constant indices, we flatten the alias 1376 // space to include all of the array body. Only the header, klass 1377 // and array length can be accessed un-aliased. 1378 if( offset != Type::OffsetBot ) { 1379 if( ta->const_oop() ) { // MethodData* or Method* 1380 offset = Type::OffsetBot; // Flatten constant access into array body 1381 tj = ta = ta-> 1382 remove_speculative()-> 1383 cast_to_ptr_type(ptr)-> 1384 cast_to_exactness(false)-> 1385 with_offset(offset); 1386 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1387 // range is OK as-is. 1388 tj = ta = TypeAryPtr::RANGE; 1389 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1390 tj = TypeInstPtr::KLASS; // all klass loads look alike 1391 ta = TypeAryPtr::RANGE; // generic ignored junk 1392 ptr = TypePtr::BotPTR; 1393 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1394 tj = TypeInstPtr::MARK; 1395 ta = TypeAryPtr::RANGE; // generic ignored junk 1396 ptr = TypePtr::BotPTR; 1397 } else { // Random constant offset into array body 1398 offset = Type::OffsetBot; // Flatten constant access into array body 1399 tj = ta = ta-> 1400 remove_speculative()-> 1401 cast_to_ptr_type(ptr)-> 1402 cast_to_exactness(false)-> 1403 with_offset(offset); 1404 } 1405 } 1406 // Arrays of fixed size alias with arrays of unknown size. 1407 if (ta->size() != TypeInt::POS) { 1408 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1409 tj = ta = ta-> 1410 remove_speculative()-> 1411 cast_to_ptr_type(ptr)-> 1412 with_ary(tary)-> 1413 cast_to_exactness(false); 1414 } 1415 // Arrays of known objects become arrays of unknown objects. 1416 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1417 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1418 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1419 } 1420 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1421 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1422 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1423 } 1424 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1425 // cannot be distinguished by bytecode alone. 1426 if (ta->elem() == TypeInt::BOOL) { 1427 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1428 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1429 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1430 } 1431 // During the 2nd round of IterGVN, NotNull castings are removed. 1432 // Make sure the Bottom and NotNull variants alias the same. 1433 // Also, make sure exact and non-exact variants alias the same. 1434 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) { 1435 tj = ta = ta-> 1436 remove_speculative()-> 1437 cast_to_ptr_type(TypePtr::BotPTR)-> 1438 cast_to_exactness(false)-> 1439 with_offset(offset); 1440 } 1441 } 1442 1443 // Oop pointers need some flattening 1444 const TypeInstPtr *to = tj->isa_instptr(); 1445 if (to && to != TypeOopPtr::BOTTOM) { 1446 ciInstanceKlass* ik = to->instance_klass(); 1447 if( ptr == TypePtr::Constant ) { 1448 if (ik != ciEnv::current()->Class_klass() || 1449 offset < ik->layout_helper_size_in_bytes()) { 1450 // No constant oop pointers (such as Strings); they alias with 1451 // unknown strings. 1452 assert(!is_known_inst, "not scalarizable allocation"); 1453 tj = to = to-> 1454 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1455 remove_speculative()-> 1456 cast_to_ptr_type(TypePtr::BotPTR)-> 1457 cast_to_exactness(false); 1458 } 1459 } else if( is_known_inst ) { 1460 tj = to; // Keep NotNull and klass_is_exact for instance type 1461 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1462 // During the 2nd round of IterGVN, NotNull castings are removed. 1463 // Make sure the Bottom and NotNull variants alias the same. 1464 // Also, make sure exact and non-exact variants alias the same. 1465 tj = to = to-> 1466 remove_speculative()-> 1467 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1468 cast_to_ptr_type(TypePtr::BotPTR)-> 1469 cast_to_exactness(false); 1470 } 1471 if (to->speculative() != nullptr) { 1472 tj = to = to->remove_speculative(); 1473 } 1474 // Canonicalize the holder of this field 1475 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1476 // First handle header references such as a LoadKlassNode, even if the 1477 // object's klass is unloaded at compile time (4965979). 1478 if (!is_known_inst) { // Do it only for non-instance types 1479 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset); 1480 } 1481 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) { 1482 // Static fields are in the space above the normal instance 1483 // fields in the java.lang.Class instance. 1484 if (ik != ciEnv::current()->Class_klass()) { 1485 to = nullptr; 1486 tj = TypeOopPtr::BOTTOM; 1487 offset = tj->offset(); 1488 } 1489 } else { 1490 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset); 1491 assert(offset < canonical_holder->layout_helper_size_in_bytes(), ""); 1492 if (!ik->equals(canonical_holder) || tj->offset() != offset) { 1493 if( is_known_inst ) { 1494 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id()); 1495 } else { 1496 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset); 1497 } 1498 } 1499 } 1500 } 1501 1502 // Klass pointers to object array klasses need some flattening 1503 const TypeKlassPtr *tk = tj->isa_klassptr(); 1504 if( tk ) { 1505 // If we are referencing a field within a Klass, we need 1506 // to assume the worst case of an Object. Both exact and 1507 // inexact types must flatten to the same alias class so 1508 // use NotNull as the PTR. 1509 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1510 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, 1511 env()->Object_klass(), 1512 offset); 1513 } 1514 1515 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) { 1516 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass()); 1517 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs 1518 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset); 1519 } else { 1520 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset); 1521 } 1522 } 1523 1524 // Check for precise loads from the primary supertype array and force them 1525 // to the supertype cache alias index. Check for generic array loads from 1526 // the primary supertype array and also force them to the supertype cache 1527 // alias index. Since the same load can reach both, we need to merge 1528 // these 2 disparate memories into the same alias class. Since the 1529 // primary supertype array is read-only, there's no chance of confusion 1530 // where we bypass an array load and an array store. 1531 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1532 if (offset == Type::OffsetBot || 1533 (offset >= primary_supers_offset && 1534 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1535 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1536 offset = in_bytes(Klass::secondary_super_cache_offset()); 1537 tj = tk = tk->with_offset(offset); 1538 } 1539 } 1540 1541 // Flatten all Raw pointers together. 1542 if (tj->base() == Type::RawPtr) 1543 tj = TypeRawPtr::BOTTOM; 1544 1545 if (tj->base() == Type::AnyPtr) 1546 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1547 1548 offset = tj->offset(); 1549 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1550 1551 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1552 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1553 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1554 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1555 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1556 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1557 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1558 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1559 assert( tj->ptr() != TypePtr::TopPTR && 1560 tj->ptr() != TypePtr::AnyNull && 1561 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1562 // assert( tj->ptr() != TypePtr::Constant || 1563 // tj->base() == Type::RawPtr || 1564 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1565 1566 return tj; 1567 } 1568 1569 void Compile::AliasType::Init(int i, const TypePtr* at) { 1570 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1571 _index = i; 1572 _adr_type = at; 1573 _field = nullptr; 1574 _element = nullptr; 1575 _is_rewritable = true; // default 1576 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr; 1577 if (atoop != nullptr && atoop->is_known_instance()) { 1578 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1579 _general_index = Compile::current()->get_alias_index(gt); 1580 } else { 1581 _general_index = 0; 1582 } 1583 } 1584 1585 BasicType Compile::AliasType::basic_type() const { 1586 if (element() != nullptr) { 1587 const Type* element = adr_type()->is_aryptr()->elem(); 1588 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1589 } if (field() != nullptr) { 1590 return field()->layout_type(); 1591 } else { 1592 return T_ILLEGAL; // unknown 1593 } 1594 } 1595 1596 //---------------------------------print_on------------------------------------ 1597 #ifndef PRODUCT 1598 void Compile::AliasType::print_on(outputStream* st) { 1599 if (index() < 10) 1600 st->print("@ <%d> ", index()); 1601 else st->print("@ <%d>", index()); 1602 st->print(is_rewritable() ? " " : " RO"); 1603 int offset = adr_type()->offset(); 1604 if (offset == Type::OffsetBot) 1605 st->print(" +any"); 1606 else st->print(" +%-3d", offset); 1607 st->print(" in "); 1608 adr_type()->dump_on(st); 1609 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1610 if (field() != nullptr && tjp) { 1611 if (tjp->is_instptr()->instance_klass() != field()->holder() || 1612 tjp->offset() != field()->offset_in_bytes()) { 1613 st->print(" != "); 1614 field()->print(); 1615 st->print(" ***"); 1616 } 1617 } 1618 } 1619 1620 void print_alias_types() { 1621 Compile* C = Compile::current(); 1622 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1623 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1624 C->alias_type(idx)->print_on(tty); 1625 tty->cr(); 1626 } 1627 } 1628 #endif 1629 1630 1631 //----------------------------probe_alias_cache-------------------------------- 1632 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1633 intptr_t key = (intptr_t) adr_type; 1634 key ^= key >> logAliasCacheSize; 1635 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1636 } 1637 1638 1639 //-----------------------------grow_alias_types-------------------------------- 1640 void Compile::grow_alias_types() { 1641 const int old_ats = _max_alias_types; // how many before? 1642 const int new_ats = old_ats; // how many more? 1643 const int grow_ats = old_ats+new_ats; // how many now? 1644 _max_alias_types = grow_ats; 1645 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1646 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1647 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1648 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1649 } 1650 1651 1652 //--------------------------------find_alias_type------------------------------ 1653 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1654 if (!do_aliasing()) { 1655 return alias_type(AliasIdxBot); 1656 } 1657 1658 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1659 if (ace->_adr_type == adr_type) { 1660 return alias_type(ace->_index); 1661 } 1662 1663 // Handle special cases. 1664 if (adr_type == nullptr) return alias_type(AliasIdxTop); 1665 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1666 1667 // Do it the slow way. 1668 const TypePtr* flat = flatten_alias_type(adr_type); 1669 1670 #ifdef ASSERT 1671 { 1672 ResourceMark rm; 1673 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1674 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1675 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1676 Type::str(adr_type)); 1677 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1678 const TypeOopPtr* foop = flat->is_oopptr(); 1679 // Scalarizable allocations have exact klass always. 1680 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1681 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1682 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1683 Type::str(foop), Type::str(xoop)); 1684 } 1685 } 1686 #endif 1687 1688 int idx = AliasIdxTop; 1689 for (int i = 0; i < num_alias_types(); i++) { 1690 if (alias_type(i)->adr_type() == flat) { 1691 idx = i; 1692 break; 1693 } 1694 } 1695 1696 if (idx == AliasIdxTop) { 1697 if (no_create) return nullptr; 1698 // Grow the array if necessary. 1699 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1700 // Add a new alias type. 1701 idx = _num_alias_types++; 1702 _alias_types[idx]->Init(idx, flat); 1703 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1704 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1705 if (flat->isa_instptr()) { 1706 if (flat->offset() == java_lang_Class::klass_offset() 1707 && flat->is_instptr()->instance_klass() == env()->Class_klass()) 1708 alias_type(idx)->set_rewritable(false); 1709 } 1710 if (flat->isa_aryptr()) { 1711 #ifdef ASSERT 1712 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1713 // (T_BYTE has the weakest alignment and size restrictions...) 1714 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1715 #endif 1716 if (flat->offset() == TypePtr::OffsetBot) { 1717 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1718 } 1719 } 1720 if (flat->isa_klassptr()) { 1721 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1722 alias_type(idx)->set_rewritable(false); 1723 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1724 alias_type(idx)->set_rewritable(false); 1725 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1726 alias_type(idx)->set_rewritable(false); 1727 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1728 alias_type(idx)->set_rewritable(false); 1729 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1730 alias_type(idx)->set_rewritable(false); 1731 } 1732 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1733 // but the base pointer type is not distinctive enough to identify 1734 // references into JavaThread.) 1735 1736 // Check for final fields. 1737 const TypeInstPtr* tinst = flat->isa_instptr(); 1738 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1739 ciField* field; 1740 if (tinst->const_oop() != nullptr && 1741 tinst->instance_klass() == ciEnv::current()->Class_klass() && 1742 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) { 1743 // static field 1744 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1745 field = k->get_field_by_offset(tinst->offset(), true); 1746 } else { 1747 ciInstanceKlass *k = tinst->instance_klass(); 1748 field = k->get_field_by_offset(tinst->offset(), false); 1749 } 1750 assert(field == nullptr || 1751 original_field == nullptr || 1752 (field->holder() == original_field->holder() && 1753 field->offset_in_bytes() == original_field->offset_in_bytes() && 1754 field->is_static() == original_field->is_static()), "wrong field?"); 1755 // Set field() and is_rewritable() attributes. 1756 if (field != nullptr) alias_type(idx)->set_field(field); 1757 } 1758 } 1759 1760 // Fill the cache for next time. 1761 ace->_adr_type = adr_type; 1762 ace->_index = idx; 1763 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1764 1765 // Might as well try to fill the cache for the flattened version, too. 1766 AliasCacheEntry* face = probe_alias_cache(flat); 1767 if (face->_adr_type == nullptr) { 1768 face->_adr_type = flat; 1769 face->_index = idx; 1770 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1771 } 1772 1773 return alias_type(idx); 1774 } 1775 1776 1777 Compile::AliasType* Compile::alias_type(ciField* field) { 1778 const TypeOopPtr* t; 1779 if (field->is_static()) 1780 t = TypeInstPtr::make(field->holder()->java_mirror()); 1781 else 1782 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1783 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1784 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1785 return atp; 1786 } 1787 1788 1789 //------------------------------have_alias_type-------------------------------- 1790 bool Compile::have_alias_type(const TypePtr* adr_type) { 1791 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1792 if (ace->_adr_type == adr_type) { 1793 return true; 1794 } 1795 1796 // Handle special cases. 1797 if (adr_type == nullptr) return true; 1798 if (adr_type == TypePtr::BOTTOM) return true; 1799 1800 return find_alias_type(adr_type, true, nullptr) != nullptr; 1801 } 1802 1803 //-----------------------------must_alias-------------------------------------- 1804 // True if all values of the given address type are in the given alias category. 1805 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1806 if (alias_idx == AliasIdxBot) return true; // the universal category 1807 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP 1808 if (alias_idx == AliasIdxTop) return false; // the empty category 1809 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1810 1811 // the only remaining possible overlap is identity 1812 int adr_idx = get_alias_index(adr_type); 1813 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1814 assert(adr_idx == alias_idx || 1815 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1816 && adr_type != TypeOopPtr::BOTTOM), 1817 "should not be testing for overlap with an unsafe pointer"); 1818 return adr_idx == alias_idx; 1819 } 1820 1821 //------------------------------can_alias-------------------------------------- 1822 // True if any values of the given address type are in the given alias category. 1823 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1824 if (alias_idx == AliasIdxTop) return false; // the empty category 1825 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP 1826 // Known instance doesn't alias with bottom memory 1827 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1828 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1829 1830 // the only remaining possible overlap is identity 1831 int adr_idx = get_alias_index(adr_type); 1832 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1833 return adr_idx == alias_idx; 1834 } 1835 1836 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their 1837 // uncommon traps if no Runtime Predicates were created from the Parse Predicates. 1838 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) { 1839 if (parse_predicate_count() == 0) { 1840 return; 1841 } 1842 for (int i = 0; i < parse_predicate_count(); i++) { 1843 ParsePredicateNode* parse_predicate = _parse_predicates.at(i); 1844 parse_predicate->mark_useless(); 1845 igvn._worklist.push(parse_predicate); 1846 } 1847 _parse_predicates.clear(); 1848 } 1849 1850 void Compile::record_for_post_loop_opts_igvn(Node* n) { 1851 if (!n->for_post_loop_opts_igvn()) { 1852 assert(!_for_post_loop_igvn.contains(n), "duplicate"); 1853 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1854 _for_post_loop_igvn.append(n); 1855 } 1856 } 1857 1858 void Compile::remove_from_post_loop_opts_igvn(Node* n) { 1859 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1860 _for_post_loop_igvn.remove(n); 1861 } 1862 1863 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) { 1864 // Verify that all previous optimizations produced a valid graph 1865 // at least to this point, even if no loop optimizations were done. 1866 PhaseIdealLoop::verify(igvn); 1867 1868 C->set_post_loop_opts_phase(); // no more loop opts allowed 1869 1870 assert(!C->major_progress(), "not cleared"); 1871 1872 if (_for_post_loop_igvn.length() > 0) { 1873 while (_for_post_loop_igvn.length() > 0) { 1874 Node* n = _for_post_loop_igvn.pop(); 1875 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1876 igvn._worklist.push(n); 1877 } 1878 igvn.optimize(); 1879 if (failing()) return; 1880 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed"); 1881 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now"); 1882 1883 // Sometimes IGVN sets major progress (e.g., when processing loop nodes). 1884 if (C->major_progress()) { 1885 C->clear_major_progress(); // ensure that major progress is now clear 1886 } 1887 } 1888 } 1889 1890 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) { 1891 if (OptimizeUnstableIf) { 1892 _unstable_if_traps.append(trap); 1893 } 1894 } 1895 1896 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) { 1897 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) { 1898 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1899 Node* n = trap->uncommon_trap(); 1900 if (!useful.member(n)) { 1901 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 1902 } 1903 } 1904 } 1905 1906 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead 1907 // or fold-compares case. Return true if succeed or not found. 1908 // 1909 // In rare cases, the found trap has been processed. It is too late to delete it. Return 1910 // false and ask fold-compares to yield. 1911 // 1912 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused 1913 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path 1914 // when deoptimization does happen. 1915 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) { 1916 for (int i = 0; i < _unstable_if_traps.length(); ++i) { 1917 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1918 if (trap->uncommon_trap() == unc) { 1919 if (yield && trap->modified()) { 1920 return false; 1921 } 1922 _unstable_if_traps.delete_at(i); 1923 break; 1924 } 1925 } 1926 return true; 1927 } 1928 1929 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path. 1930 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering. 1931 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) { 1932 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) { 1933 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1934 CallStaticJavaNode* unc = trap->uncommon_trap(); 1935 int next_bci = trap->next_bci(); 1936 bool modified = trap->modified(); 1937 1938 if (next_bci != -1 && !modified) { 1939 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!"); 1940 JVMState* jvms = unc->jvms(); 1941 ciMethod* method = jvms->method(); 1942 ciBytecodeStream iter(method); 1943 1944 iter.force_bci(jvms->bci()); 1945 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if"); 1946 Bytecodes::Code c = iter.cur_bc(); 1947 Node* lhs = nullptr; 1948 Node* rhs = nullptr; 1949 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) { 1950 lhs = unc->peek_operand(0); 1951 rhs = unc->peek_operand(1); 1952 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) { 1953 lhs = unc->peek_operand(0); 1954 } 1955 1956 ResourceMark rm; 1957 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci); 1958 assert(live_locals.is_valid(), "broken liveness info"); 1959 int len = (int)live_locals.size(); 1960 1961 for (int i = 0; i < len; i++) { 1962 Node* local = unc->local(jvms, i); 1963 // kill local using the liveness of next_bci. 1964 // give up when the local looks like an operand to secure reexecution. 1965 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) { 1966 uint idx = jvms->locoff() + i; 1967 #ifdef ASSERT 1968 if (PrintOpto && Verbose) { 1969 tty->print("[unstable_if] kill local#%d: ", idx); 1970 local->dump(); 1971 tty->cr(); 1972 } 1973 #endif 1974 igvn.replace_input_of(unc, idx, top()); 1975 modified = true; 1976 } 1977 } 1978 } 1979 1980 // keep the mondified trap for late query 1981 if (modified) { 1982 trap->set_modified(); 1983 } else { 1984 _unstable_if_traps.delete_at(i); 1985 } 1986 } 1987 igvn.optimize(); 1988 } 1989 1990 // StringOpts and late inlining of string methods 1991 void Compile::inline_string_calls(bool parse_time) { 1992 { 1993 // remove useless nodes to make the usage analysis simpler 1994 ResourceMark rm; 1995 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 1996 } 1997 1998 { 1999 ResourceMark rm; 2000 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2001 PhaseStringOpts pso(initial_gvn()); 2002 print_method(PHASE_AFTER_STRINGOPTS, 3); 2003 } 2004 2005 // now inline anything that we skipped the first time around 2006 if (!parse_time) { 2007 _late_inlines_pos = _late_inlines.length(); 2008 } 2009 2010 while (_string_late_inlines.length() > 0) { 2011 CallGenerator* cg = _string_late_inlines.pop(); 2012 cg->do_late_inline(); 2013 if (failing()) return; 2014 } 2015 _string_late_inlines.trunc_to(0); 2016 } 2017 2018 // Late inlining of boxing methods 2019 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2020 if (_boxing_late_inlines.length() > 0) { 2021 assert(has_boxed_value(), "inconsistent"); 2022 2023 set_inlining_incrementally(true); 2024 2025 igvn_worklist()->ensure_empty(); // should be done with igvn 2026 2027 _late_inlines_pos = _late_inlines.length(); 2028 2029 while (_boxing_late_inlines.length() > 0) { 2030 CallGenerator* cg = _boxing_late_inlines.pop(); 2031 cg->do_late_inline(); 2032 if (failing()) return; 2033 } 2034 _boxing_late_inlines.trunc_to(0); 2035 2036 inline_incrementally_cleanup(igvn); 2037 2038 set_inlining_incrementally(false); 2039 } 2040 } 2041 2042 bool Compile::inline_incrementally_one() { 2043 assert(IncrementalInline, "incremental inlining should be on"); 2044 2045 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2046 2047 set_inlining_progress(false); 2048 set_do_cleanup(false); 2049 2050 for (int i = 0; i < _late_inlines.length(); i++) { 2051 _late_inlines_pos = i+1; 2052 CallGenerator* cg = _late_inlines.at(i); 2053 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline(); 2054 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call 2055 cg->do_late_inline(); 2056 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed"); 2057 if (failing()) { 2058 return false; 2059 } else if (inlining_progress()) { 2060 _late_inlines_pos = i+1; // restore the position in case new elements were inserted 2061 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node()); 2062 break; // process one call site at a time 2063 } 2064 } else { 2065 // Ignore late inline direct calls when inlining is not allowed. 2066 // They are left in the late inline list when node budget is exhausted until the list is fully drained. 2067 } 2068 } 2069 // Remove processed elements. 2070 _late_inlines.remove_till(_late_inlines_pos); 2071 _late_inlines_pos = 0; 2072 2073 assert(inlining_progress() || _late_inlines.length() == 0, "no progress"); 2074 2075 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2076 2077 set_inlining_progress(false); 2078 set_do_cleanup(false); 2079 2080 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption; 2081 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup; 2082 } 2083 2084 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2085 { 2086 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2087 ResourceMark rm; 2088 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2089 } 2090 { 2091 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2092 igvn.reset_from_gvn(initial_gvn()); 2093 igvn.optimize(); 2094 if (failing()) return; 2095 } 2096 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3); 2097 } 2098 2099 // Perform incremental inlining until bound on number of live nodes is reached 2100 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2101 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2102 2103 set_inlining_incrementally(true); 2104 uint low_live_nodes = 0; 2105 2106 while (_late_inlines.length() > 0) { 2107 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2108 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2109 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2110 // PhaseIdealLoop is expensive so we only try it once we are 2111 // out of live nodes and we only try it again if the previous 2112 // helped got the number of nodes down significantly 2113 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2114 if (failing()) return; 2115 low_live_nodes = live_nodes(); 2116 _major_progress = true; 2117 } 2118 2119 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2120 bool do_print_inlining = print_inlining() || print_intrinsics(); 2121 if (do_print_inlining || log() != nullptr) { 2122 // Print inlining message for candidates that we couldn't inline for lack of space. 2123 for (int i = 0; i < _late_inlines.length(); i++) { 2124 CallGenerator* cg = _late_inlines.at(i); 2125 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 2126 if (do_print_inlining) { 2127 cg->print_inlining_late(InliningResult::FAILURE, msg); 2128 } 2129 log_late_inline_failure(cg, msg); 2130 } 2131 } 2132 break; // finish 2133 } 2134 } 2135 2136 igvn_worklist()->ensure_empty(); // should be done with igvn 2137 2138 while (inline_incrementally_one()) { 2139 assert(!failing(), "inconsistent"); 2140 } 2141 if (failing()) return; 2142 2143 inline_incrementally_cleanup(igvn); 2144 2145 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 2146 2147 if (failing()) return; 2148 2149 if (_late_inlines.length() == 0) { 2150 break; // no more progress 2151 } 2152 } 2153 2154 igvn_worklist()->ensure_empty(); // should be done with igvn 2155 2156 if (_string_late_inlines.length() > 0) { 2157 assert(has_stringbuilder(), "inconsistent"); 2158 2159 inline_string_calls(false); 2160 2161 if (failing()) return; 2162 2163 inline_incrementally_cleanup(igvn); 2164 } 2165 2166 set_inlining_incrementally(false); 2167 } 2168 2169 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) { 2170 // "inlining_incrementally() == false" is used to signal that no inlining is allowed 2171 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details). 2172 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr" 2173 // as if "inlining_incrementally() == true" were set. 2174 assert(inlining_incrementally() == false, "not allowed"); 2175 assert(_modified_nodes == nullptr, "not allowed"); 2176 assert(_late_inlines.length() > 0, "sanity"); 2177 2178 while (_late_inlines.length() > 0) { 2179 igvn_worklist()->ensure_empty(); // should be done with igvn 2180 2181 while (inline_incrementally_one()) { 2182 assert(!failing(), "inconsistent"); 2183 } 2184 if (failing()) return; 2185 2186 inline_incrementally_cleanup(igvn); 2187 } 2188 } 2189 2190 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2191 if (_loop_opts_cnt > 0) { 2192 while (major_progress() && (_loop_opts_cnt > 0)) { 2193 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2194 PhaseIdealLoop::optimize(igvn, mode); 2195 _loop_opts_cnt--; 2196 if (failing()) return false; 2197 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2198 } 2199 } 2200 return true; 2201 } 2202 2203 // Remove edges from "root" to each SafePoint at a backward branch. 2204 // They were inserted during parsing (see add_safepoint()) to make 2205 // infinite loops without calls or exceptions visible to root, i.e., 2206 // useful. 2207 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2208 Node *r = root(); 2209 if (r != nullptr) { 2210 for (uint i = r->req(); i < r->len(); ++i) { 2211 Node *n = r->in(i); 2212 if (n != nullptr && n->is_SafePoint()) { 2213 r->rm_prec(i); 2214 if (n->outcnt() == 0) { 2215 igvn.remove_dead_node(n); 2216 } 2217 --i; 2218 } 2219 } 2220 // Parsing may have added top inputs to the root node (Path 2221 // leading to the Halt node proven dead). Make sure we get a 2222 // chance to clean them up. 2223 igvn._worklist.push(r); 2224 igvn.optimize(); 2225 } 2226 } 2227 2228 //------------------------------Optimize--------------------------------------- 2229 // Given a graph, optimize it. 2230 void Compile::Optimize() { 2231 TracePhase tp("optimizer", &timers[_t_optimizer]); 2232 2233 #ifndef PRODUCT 2234 if (env()->break_at_compile()) { 2235 BREAKPOINT; 2236 } 2237 2238 #endif 2239 2240 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2241 #ifdef ASSERT 2242 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2243 #endif 2244 2245 ResourceMark rm; 2246 2247 print_inlining_reinit(); 2248 2249 NOT_PRODUCT( verify_graph_edges(); ) 2250 2251 print_method(PHASE_AFTER_PARSING, 1); 2252 2253 { 2254 // Iterative Global Value Numbering, including ideal transforms 2255 // Initialize IterGVN with types and values from parse-time GVN 2256 PhaseIterGVN igvn(initial_gvn()); 2257 #ifdef ASSERT 2258 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2259 #endif 2260 { 2261 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2262 igvn.optimize(); 2263 } 2264 2265 if (failing()) return; 2266 2267 print_method(PHASE_ITER_GVN1, 2); 2268 2269 process_for_unstable_if_traps(igvn); 2270 2271 if (failing()) return; 2272 2273 inline_incrementally(igvn); 2274 2275 print_method(PHASE_INCREMENTAL_INLINE, 2); 2276 2277 if (failing()) return; 2278 2279 if (eliminate_boxing()) { 2280 // Inline valueOf() methods now. 2281 inline_boxing_calls(igvn); 2282 2283 if (failing()) return; 2284 2285 if (AlwaysIncrementalInline || StressIncrementalInlining) { 2286 inline_incrementally(igvn); 2287 } 2288 2289 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2290 2291 if (failing()) return; 2292 } 2293 2294 // Remove the speculative part of types and clean up the graph from 2295 // the extra CastPP nodes whose only purpose is to carry them. Do 2296 // that early so that optimizations are not disrupted by the extra 2297 // CastPP nodes. 2298 remove_speculative_types(igvn); 2299 2300 if (failing()) return; 2301 2302 // No more new expensive nodes will be added to the list from here 2303 // so keep only the actual candidates for optimizations. 2304 cleanup_expensive_nodes(igvn); 2305 2306 if (failing()) return; 2307 2308 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2309 if (EnableVectorSupport && has_vbox_nodes()) { 2310 TracePhase tp("", &timers[_t_vector]); 2311 PhaseVector pv(igvn); 2312 pv.optimize_vector_boxes(); 2313 if (failing()) return; 2314 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2315 } 2316 assert(!has_vbox_nodes(), "sanity"); 2317 2318 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2319 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2320 igvn_worklist()->ensure_empty(); // should be done with igvn 2321 { 2322 ResourceMark rm; 2323 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist()); 2324 } 2325 igvn.reset_from_gvn(initial_gvn()); 2326 igvn.optimize(); 2327 if (failing()) return; 2328 } 2329 2330 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop 2331 // safepoints 2332 remove_root_to_sfpts_edges(igvn); 2333 2334 if (failing()) return; 2335 2336 // Perform escape analysis 2337 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) { 2338 if (has_loops()) { 2339 // Cleanup graph (remove dead nodes). 2340 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2341 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2342 if (failing()) return; 2343 } 2344 bool progress; 2345 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2346 do { 2347 ConnectionGraph::do_analysis(this, &igvn); 2348 2349 if (failing()) return; 2350 2351 int mcount = macro_count(); // Record number of allocations and locks before IGVN 2352 2353 // Optimize out fields loads from scalar replaceable allocations. 2354 igvn.optimize(); 2355 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2356 2357 if (failing()) return; 2358 2359 if (congraph() != nullptr && macro_count() > 0) { 2360 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2361 PhaseMacroExpand mexp(igvn); 2362 mexp.eliminate_macro_nodes(); 2363 if (failing()) return; 2364 2365 igvn.set_delay_transform(false); 2366 igvn.optimize(); 2367 if (failing()) return; 2368 2369 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2370 } 2371 2372 ConnectionGraph::verify_ram_nodes(this, root()); 2373 if (failing()) return; 2374 2375 progress = do_iterative_escape_analysis() && 2376 (macro_count() < mcount) && 2377 ConnectionGraph::has_candidates(this); 2378 // Try again if candidates exist and made progress 2379 // by removing some allocations and/or locks. 2380 } while (progress); 2381 } 2382 2383 // Loop transforms on the ideal graph. Range Check Elimination, 2384 // peeling, unrolling, etc. 2385 2386 // Set loop opts counter 2387 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2388 { 2389 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2390 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2391 _loop_opts_cnt--; 2392 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2393 if (failing()) return; 2394 } 2395 // Loop opts pass if partial peeling occurred in previous pass 2396 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2397 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2398 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2399 _loop_opts_cnt--; 2400 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2401 if (failing()) return; 2402 } 2403 // Loop opts pass for loop-unrolling before CCP 2404 if(major_progress() && (_loop_opts_cnt > 0)) { 2405 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2406 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2407 _loop_opts_cnt--; 2408 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2409 } 2410 if (!failing()) { 2411 // Verify that last round of loop opts produced a valid graph 2412 PhaseIdealLoop::verify(igvn); 2413 } 2414 } 2415 if (failing()) return; 2416 2417 // Conditional Constant Propagation; 2418 print_method(PHASE_BEFORE_CCP1, 2); 2419 PhaseCCP ccp( &igvn ); 2420 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2421 { 2422 TracePhase tp("ccp", &timers[_t_ccp]); 2423 ccp.do_transform(); 2424 } 2425 print_method(PHASE_CCP1, 2); 2426 2427 assert( true, "Break here to ccp.dump_old2new_map()"); 2428 2429 // Iterative Global Value Numbering, including ideal transforms 2430 { 2431 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2432 igvn.reset_from_igvn(&ccp); 2433 igvn.optimize(); 2434 } 2435 print_method(PHASE_ITER_GVN2, 2); 2436 2437 if (failing()) return; 2438 2439 // Loop transforms on the ideal graph. Range Check Elimination, 2440 // peeling, unrolling, etc. 2441 if (!optimize_loops(igvn, LoopOptsDefault)) { 2442 return; 2443 } 2444 2445 if (failing()) return; 2446 2447 C->clear_major_progress(); // ensure that major progress is now clear 2448 2449 process_for_post_loop_opts_igvn(igvn); 2450 2451 if (failing()) return; 2452 2453 #ifdef ASSERT 2454 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2455 #endif 2456 2457 { 2458 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2459 PhaseMacroExpand mex(igvn); 2460 if (mex.expand_macro_nodes()) { 2461 assert(failing(), "must bail out w/ explicit message"); 2462 return; 2463 } 2464 print_method(PHASE_MACRO_EXPANSION, 2); 2465 } 2466 2467 { 2468 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2469 if (bs->expand_barriers(this, igvn)) { 2470 assert(failing(), "must bail out w/ explicit message"); 2471 return; 2472 } 2473 print_method(PHASE_BARRIER_EXPANSION, 2); 2474 } 2475 2476 if (C->max_vector_size() > 0) { 2477 C->optimize_logic_cones(igvn); 2478 igvn.optimize(); 2479 if (failing()) return; 2480 } 2481 2482 DEBUG_ONLY( _modified_nodes = nullptr; ) 2483 2484 assert(igvn._worklist.size() == 0, "not empty"); 2485 2486 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty"); 2487 2488 if (_late_inlines.length() > 0) { 2489 // More opportunities to optimize virtual and MH calls. 2490 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option. 2491 process_late_inline_calls_no_inline(igvn); 2492 if (failing()) return; 2493 } 2494 } // (End scope of igvn; run destructor if necessary for asserts.) 2495 2496 check_no_dead_use(); 2497 2498 process_print_inlining(); 2499 2500 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have 2501 // to remove hashes to unlock nodes for modifications. 2502 C->node_hash()->clear(); 2503 2504 // A method with only infinite loops has no edges entering loops from root 2505 { 2506 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2507 if (final_graph_reshaping()) { 2508 assert(failing(), "must bail out w/ explicit message"); 2509 return; 2510 } 2511 } 2512 2513 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2514 DEBUG_ONLY(set_phase_optimize_finished();) 2515 } 2516 2517 #ifdef ASSERT 2518 void Compile::check_no_dead_use() const { 2519 ResourceMark rm; 2520 Unique_Node_List wq; 2521 wq.push(root()); 2522 for (uint i = 0; i < wq.size(); ++i) { 2523 Node* n = wq.at(i); 2524 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2525 Node* u = n->fast_out(j); 2526 if (u->outcnt() == 0 && !u->is_Con()) { 2527 u->dump(); 2528 fatal("no reachable node should have no use"); 2529 } 2530 wq.push(u); 2531 } 2532 } 2533 } 2534 #endif 2535 2536 void Compile::inline_vector_reboxing_calls() { 2537 if (C->_vector_reboxing_late_inlines.length() > 0) { 2538 _late_inlines_pos = C->_late_inlines.length(); 2539 while (_vector_reboxing_late_inlines.length() > 0) { 2540 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 2541 cg->do_late_inline(); 2542 if (failing()) return; 2543 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node()); 2544 } 2545 _vector_reboxing_late_inlines.trunc_to(0); 2546 } 2547 } 2548 2549 bool Compile::has_vbox_nodes() { 2550 if (C->_vector_reboxing_late_inlines.length() > 0) { 2551 return true; 2552 } 2553 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 2554 Node * n = C->macro_node(macro_idx); 2555 assert(n->is_macro(), "only macro nodes expected here"); 2556 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 2557 return true; 2558 } 2559 } 2560 return false; 2561 } 2562 2563 //---------------------------- Bitwise operation packing optimization --------------------------- 2564 2565 static bool is_vector_unary_bitwise_op(Node* n) { 2566 return n->Opcode() == Op_XorV && 2567 VectorNode::is_vector_bitwise_not_pattern(n); 2568 } 2569 2570 static bool is_vector_binary_bitwise_op(Node* n) { 2571 switch (n->Opcode()) { 2572 case Op_AndV: 2573 case Op_OrV: 2574 return true; 2575 2576 case Op_XorV: 2577 return !is_vector_unary_bitwise_op(n); 2578 2579 default: 2580 return false; 2581 } 2582 } 2583 2584 static bool is_vector_ternary_bitwise_op(Node* n) { 2585 return n->Opcode() == Op_MacroLogicV; 2586 } 2587 2588 static bool is_vector_bitwise_op(Node* n) { 2589 return is_vector_unary_bitwise_op(n) || 2590 is_vector_binary_bitwise_op(n) || 2591 is_vector_ternary_bitwise_op(n); 2592 } 2593 2594 static bool is_vector_bitwise_cone_root(Node* n) { 2595 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) { 2596 return false; 2597 } 2598 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2599 if (is_vector_bitwise_op(n->fast_out(i))) { 2600 return false; 2601 } 2602 } 2603 return true; 2604 } 2605 2606 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) { 2607 uint cnt = 0; 2608 if (is_vector_bitwise_op(n)) { 2609 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req(); 2610 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2611 for (uint i = 1; i < inp_cnt; i++) { 2612 Node* in = n->in(i); 2613 bool skip = VectorNode::is_all_ones_vector(in); 2614 if (!skip && !inputs.member(in)) { 2615 inputs.push(in); 2616 cnt++; 2617 } 2618 } 2619 assert(cnt <= 1, "not unary"); 2620 } else { 2621 uint last_req = inp_cnt; 2622 if (is_vector_ternary_bitwise_op(n)) { 2623 last_req = inp_cnt - 1; // skip last input 2624 } 2625 for (uint i = 1; i < last_req; i++) { 2626 Node* def = n->in(i); 2627 if (!inputs.member(def)) { 2628 inputs.push(def); 2629 cnt++; 2630 } 2631 } 2632 } 2633 } else { // not a bitwise operations 2634 if (!inputs.member(n)) { 2635 inputs.push(n); 2636 cnt++; 2637 } 2638 } 2639 return cnt; 2640 } 2641 2642 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 2643 Unique_Node_List useful_nodes; 2644 C->identify_useful_nodes(useful_nodes); 2645 2646 for (uint i = 0; i < useful_nodes.size(); i++) { 2647 Node* n = useful_nodes.at(i); 2648 if (is_vector_bitwise_cone_root(n)) { 2649 list.push(n); 2650 } 2651 } 2652 } 2653 2654 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 2655 const TypeVect* vt, 2656 Unique_Node_List& partition, 2657 Unique_Node_List& inputs) { 2658 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 2659 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 2660 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 2661 2662 Node* in1 = inputs.at(0); 2663 Node* in2 = inputs.at(1); 2664 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 2665 2666 uint func = compute_truth_table(partition, inputs); 2667 2668 Node* pn = partition.at(partition.size() - 1); 2669 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2670 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt)); 2671 } 2672 2673 static uint extract_bit(uint func, uint pos) { 2674 return (func & (1 << pos)) >> pos; 2675 } 2676 2677 // 2678 // A macro logic node represents a truth table. It has 4 inputs, 2679 // First three inputs corresponds to 3 columns of a truth table 2680 // and fourth input captures the logic function. 2681 // 2682 // eg. fn = (in1 AND in2) OR in3; 2683 // 2684 // MacroNode(in1,in2,in3,fn) 2685 // 2686 // ----------------- 2687 // in1 in2 in3 fn 2688 // ----------------- 2689 // 0 0 0 0 2690 // 0 0 1 1 2691 // 0 1 0 0 2692 // 0 1 1 1 2693 // 1 0 0 0 2694 // 1 0 1 1 2695 // 1 1 0 1 2696 // 1 1 1 1 2697 // 2698 2699 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 2700 int res = 0; 2701 for (int i = 0; i < 8; i++) { 2702 int bit1 = extract_bit(in1, i); 2703 int bit2 = extract_bit(in2, i); 2704 int bit3 = extract_bit(in3, i); 2705 2706 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 2707 int func_bit = extract_bit(func, func_bit_pos); 2708 2709 res |= func_bit << i; 2710 } 2711 return res; 2712 } 2713 2714 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 2715 assert(n != nullptr, ""); 2716 assert(eval_map.contains(n), "absent"); 2717 return *(eval_map.get(n)); 2718 } 2719 2720 static void eval_operands(Node* n, 2721 uint& func1, uint& func2, uint& func3, 2722 ResourceHashtable<Node*,uint>& eval_map) { 2723 assert(is_vector_bitwise_op(n), ""); 2724 2725 if (is_vector_unary_bitwise_op(n)) { 2726 Node* opnd = n->in(1); 2727 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 2728 opnd = n->in(2); 2729 } 2730 func1 = eval_operand(opnd, eval_map); 2731 } else if (is_vector_binary_bitwise_op(n)) { 2732 func1 = eval_operand(n->in(1), eval_map); 2733 func2 = eval_operand(n->in(2), eval_map); 2734 } else { 2735 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 2736 func1 = eval_operand(n->in(1), eval_map); 2737 func2 = eval_operand(n->in(2), eval_map); 2738 func3 = eval_operand(n->in(3), eval_map); 2739 } 2740 } 2741 2742 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 2743 assert(inputs.size() <= 3, "sanity"); 2744 ResourceMark rm; 2745 uint res = 0; 2746 ResourceHashtable<Node*,uint> eval_map; 2747 2748 // Populate precomputed functions for inputs. 2749 // Each input corresponds to one column of 3 input truth-table. 2750 uint input_funcs[] = { 0xAA, // (_, _, c) -> c 2751 0xCC, // (_, b, _) -> b 2752 0xF0 }; // (a, _, _) -> a 2753 for (uint i = 0; i < inputs.size(); i++) { 2754 eval_map.put(inputs.at(i), input_funcs[2-i]); 2755 } 2756 2757 for (uint i = 0; i < partition.size(); i++) { 2758 Node* n = partition.at(i); 2759 2760 uint func1 = 0, func2 = 0, func3 = 0; 2761 eval_operands(n, func1, func2, func3, eval_map); 2762 2763 switch (n->Opcode()) { 2764 case Op_OrV: 2765 assert(func3 == 0, "not binary"); 2766 res = func1 | func2; 2767 break; 2768 case Op_AndV: 2769 assert(func3 == 0, "not binary"); 2770 res = func1 & func2; 2771 break; 2772 case Op_XorV: 2773 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2774 assert(func2 == 0 && func3 == 0, "not unary"); 2775 res = (~func1) & 0xFF; 2776 } else { 2777 assert(func3 == 0, "not binary"); 2778 res = func1 ^ func2; 2779 } 2780 break; 2781 case Op_MacroLogicV: 2782 // Ordering of inputs may change during evaluation of sub-tree 2783 // containing MacroLogic node as a child node, thus a re-evaluation 2784 // makes sure that function is evaluated in context of current 2785 // inputs. 2786 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 2787 break; 2788 2789 default: assert(false, "not supported: %s", n->Name()); 2790 } 2791 assert(res <= 0xFF, "invalid"); 2792 eval_map.put(n, res); 2793 } 2794 return res; 2795 } 2796 2797 // Criteria under which nodes gets packed into a macro logic node:- 2798 // 1) Parent and both child nodes are all unmasked or masked with 2799 // same predicates. 2800 // 2) Masked parent can be packed with left child if it is predicated 2801 // and both have same predicates. 2802 // 3) Masked parent can be packed with right child if its un-predicated 2803 // or has matching predication condition. 2804 // 4) An unmasked parent can be packed with an unmasked child. 2805 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2806 assert(partition.size() == 0, "not empty"); 2807 assert(inputs.size() == 0, "not empty"); 2808 if (is_vector_ternary_bitwise_op(n)) { 2809 return false; 2810 } 2811 2812 bool is_unary_op = is_vector_unary_bitwise_op(n); 2813 if (is_unary_op) { 2814 assert(collect_unique_inputs(n, inputs) == 1, "not unary"); 2815 return false; // too few inputs 2816 } 2817 2818 bool pack_left_child = true; 2819 bool pack_right_child = true; 2820 2821 bool left_child_LOP = is_vector_bitwise_op(n->in(1)); 2822 bool right_child_LOP = is_vector_bitwise_op(n->in(2)); 2823 2824 int left_child_input_cnt = 0; 2825 int right_child_input_cnt = 0; 2826 2827 bool parent_is_predicated = n->is_predicated_vector(); 2828 bool left_child_predicated = n->in(1)->is_predicated_vector(); 2829 bool right_child_predicated = n->in(2)->is_predicated_vector(); 2830 2831 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr; 2832 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2833 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2834 2835 do { 2836 if (pack_left_child && left_child_LOP && 2837 ((!parent_is_predicated && !left_child_predicated) || 2838 ((parent_is_predicated && left_child_predicated && 2839 parent_pred == left_child_pred)))) { 2840 partition.push(n->in(1)); 2841 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs); 2842 } else { 2843 inputs.push(n->in(1)); 2844 left_child_input_cnt = 1; 2845 } 2846 2847 if (pack_right_child && right_child_LOP && 2848 (!right_child_predicated || 2849 (right_child_predicated && parent_is_predicated && 2850 parent_pred == right_child_pred))) { 2851 partition.push(n->in(2)); 2852 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs); 2853 } else { 2854 inputs.push(n->in(2)); 2855 right_child_input_cnt = 1; 2856 } 2857 2858 if (inputs.size() > 3) { 2859 assert(partition.size() > 0, ""); 2860 inputs.clear(); 2861 partition.clear(); 2862 if (left_child_input_cnt > right_child_input_cnt) { 2863 pack_left_child = false; 2864 } else { 2865 pack_right_child = false; 2866 } 2867 } else { 2868 break; 2869 } 2870 } while(true); 2871 2872 if(partition.size()) { 2873 partition.push(n); 2874 } 2875 2876 return (partition.size() == 2 || partition.size() == 3) && 2877 (inputs.size() == 2 || inputs.size() == 3); 2878 } 2879 2880 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 2881 assert(is_vector_bitwise_op(n), "not a root"); 2882 2883 visited.set(n->_idx); 2884 2885 // 1) Do a DFS walk over the logic cone. 2886 for (uint i = 1; i < n->req(); i++) { 2887 Node* in = n->in(i); 2888 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 2889 process_logic_cone_root(igvn, in, visited); 2890 } 2891 } 2892 2893 // 2) Bottom up traversal: Merge node[s] with 2894 // the parent to form macro logic node. 2895 Unique_Node_List partition; 2896 Unique_Node_List inputs; 2897 if (compute_logic_cone(n, partition, inputs)) { 2898 const TypeVect* vt = n->bottom_type()->is_vect(); 2899 Node* pn = partition.at(partition.size() - 1); 2900 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2901 if (mask == nullptr || 2902 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) { 2903 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 2904 VectorNode::trace_new_vector(macro_logic, "MacroLogic"); 2905 igvn.replace_node(n, macro_logic); 2906 } 2907 } 2908 } 2909 2910 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 2911 ResourceMark rm; 2912 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 2913 Unique_Node_List list; 2914 collect_logic_cone_roots(list); 2915 2916 while (list.size() > 0) { 2917 Node* n = list.pop(); 2918 const TypeVect* vt = n->bottom_type()->is_vect(); 2919 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 2920 if (supported) { 2921 VectorSet visited(comp_arena()); 2922 process_logic_cone_root(igvn, n, visited); 2923 } 2924 } 2925 } 2926 } 2927 2928 //------------------------------Code_Gen--------------------------------------- 2929 // Given a graph, generate code for it 2930 void Compile::Code_Gen() { 2931 if (failing()) { 2932 return; 2933 } 2934 2935 // Perform instruction selection. You might think we could reclaim Matcher 2936 // memory PDQ, but actually the Matcher is used in generating spill code. 2937 // Internals of the Matcher (including some VectorSets) must remain live 2938 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2939 // set a bit in reclaimed memory. 2940 2941 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2942 // nodes. Mapping is only valid at the root of each matched subtree. 2943 NOT_PRODUCT( verify_graph_edges(); ) 2944 2945 Matcher matcher; 2946 _matcher = &matcher; 2947 { 2948 TracePhase tp("matcher", &timers[_t_matcher]); 2949 matcher.match(); 2950 if (failing()) { 2951 return; 2952 } 2953 } 2954 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2955 // nodes. Mapping is only valid at the root of each matched subtree. 2956 NOT_PRODUCT( verify_graph_edges(); ) 2957 2958 // If you have too many nodes, or if matching has failed, bail out 2959 check_node_count(0, "out of nodes matching instructions"); 2960 if (failing()) { 2961 return; 2962 } 2963 2964 print_method(PHASE_MATCHING, 2); 2965 2966 // Build a proper-looking CFG 2967 PhaseCFG cfg(node_arena(), root(), matcher); 2968 _cfg = &cfg; 2969 { 2970 TracePhase tp("scheduler", &timers[_t_scheduler]); 2971 bool success = cfg.do_global_code_motion(); 2972 if (!success) { 2973 return; 2974 } 2975 2976 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2977 NOT_PRODUCT( verify_graph_edges(); ) 2978 cfg.verify(); 2979 } 2980 2981 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2982 _regalloc = ®alloc; 2983 { 2984 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2985 // Perform register allocation. After Chaitin, use-def chains are 2986 // no longer accurate (at spill code) and so must be ignored. 2987 // Node->LRG->reg mappings are still accurate. 2988 _regalloc->Register_Allocate(); 2989 2990 // Bail out if the allocator builds too many nodes 2991 if (failing()) { 2992 return; 2993 } 2994 2995 print_method(PHASE_REGISTER_ALLOCATION, 2); 2996 } 2997 2998 // Prior to register allocation we kept empty basic blocks in case the 2999 // the allocator needed a place to spill. After register allocation we 3000 // are not adding any new instructions. If any basic block is empty, we 3001 // can now safely remove it. 3002 { 3003 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 3004 cfg.remove_empty_blocks(); 3005 if (do_freq_based_layout()) { 3006 PhaseBlockLayout layout(cfg); 3007 } else { 3008 cfg.set_loop_alignment(); 3009 } 3010 cfg.fixup_flow(); 3011 cfg.remove_unreachable_blocks(); 3012 cfg.verify_dominator_tree(); 3013 print_method(PHASE_BLOCK_ORDERING, 3); 3014 } 3015 3016 // Apply peephole optimizations 3017 if( OptoPeephole ) { 3018 TracePhase tp("peephole", &timers[_t_peephole]); 3019 PhasePeephole peep( _regalloc, cfg); 3020 peep.do_transform(); 3021 print_method(PHASE_PEEPHOLE, 3); 3022 } 3023 3024 // Do late expand if CPU requires this. 3025 if (Matcher::require_postalloc_expand) { 3026 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 3027 cfg.postalloc_expand(_regalloc); 3028 print_method(PHASE_POSTALLOC_EXPAND, 3); 3029 } 3030 3031 // Convert Nodes to instruction bits in a buffer 3032 { 3033 TracePhase tp("output", &timers[_t_output]); 3034 PhaseOutput output; 3035 output.Output(); 3036 if (failing()) return; 3037 output.install(); 3038 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here 3039 } 3040 3041 // He's dead, Jim. 3042 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 3043 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 3044 } 3045 3046 //------------------------------Final_Reshape_Counts--------------------------- 3047 // This class defines counters to help identify when a method 3048 // may/must be executed using hardware with only 24-bit precision. 3049 struct Final_Reshape_Counts : public StackObj { 3050 int _call_count; // count non-inlined 'common' calls 3051 int _float_count; // count float ops requiring 24-bit precision 3052 int _double_count; // count double ops requiring more precision 3053 int _java_call_count; // count non-inlined 'java' calls 3054 int _inner_loop_count; // count loops which need alignment 3055 VectorSet _visited; // Visitation flags 3056 Node_List _tests; // Set of IfNodes & PCTableNodes 3057 3058 Final_Reshape_Counts() : 3059 _call_count(0), _float_count(0), _double_count(0), 3060 _java_call_count(0), _inner_loop_count(0) { } 3061 3062 void inc_call_count () { _call_count ++; } 3063 void inc_float_count () { _float_count ++; } 3064 void inc_double_count() { _double_count++; } 3065 void inc_java_call_count() { _java_call_count++; } 3066 void inc_inner_loop_count() { _inner_loop_count++; } 3067 3068 int get_call_count () const { return _call_count ; } 3069 int get_float_count () const { return _float_count ; } 3070 int get_double_count() const { return _double_count; } 3071 int get_java_call_count() const { return _java_call_count; } 3072 int get_inner_loop_count() const { return _inner_loop_count; } 3073 }; 3074 3075 // Eliminate trivially redundant StoreCMs and accumulate their 3076 // precedence edges. 3077 void Compile::eliminate_redundant_card_marks(Node* n) { 3078 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 3079 if (n->in(MemNode::Address)->outcnt() > 1) { 3080 // There are multiple users of the same address so it might be 3081 // possible to eliminate some of the StoreCMs 3082 Node* mem = n->in(MemNode::Memory); 3083 Node* adr = n->in(MemNode::Address); 3084 Node* val = n->in(MemNode::ValueIn); 3085 Node* prev = n; 3086 bool done = false; 3087 // Walk the chain of StoreCMs eliminating ones that match. As 3088 // long as it's a chain of single users then the optimization is 3089 // safe. Eliminating partially redundant StoreCMs would require 3090 // cloning copies down the other paths. 3091 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 3092 if (adr == mem->in(MemNode::Address) && 3093 val == mem->in(MemNode::ValueIn)) { 3094 // redundant StoreCM 3095 if (mem->req() > MemNode::OopStore) { 3096 // Hasn't been processed by this code yet. 3097 n->add_prec(mem->in(MemNode::OopStore)); 3098 } else { 3099 // Already converted to precedence edge 3100 for (uint i = mem->req(); i < mem->len(); i++) { 3101 // Accumulate any precedence edges 3102 if (mem->in(i) != nullptr) { 3103 n->add_prec(mem->in(i)); 3104 } 3105 } 3106 // Everything above this point has been processed. 3107 done = true; 3108 } 3109 // Eliminate the previous StoreCM 3110 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 3111 assert(mem->outcnt() == 0, "should be dead"); 3112 mem->disconnect_inputs(this); 3113 } else { 3114 prev = mem; 3115 } 3116 mem = prev->in(MemNode::Memory); 3117 } 3118 } 3119 } 3120 3121 //------------------------------final_graph_reshaping_impl---------------------- 3122 // Implement items 1-5 from final_graph_reshaping below. 3123 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3124 3125 if ( n->outcnt() == 0 ) return; // dead node 3126 uint nop = n->Opcode(); 3127 3128 // Check for 2-input instruction with "last use" on right input. 3129 // Swap to left input. Implements item (2). 3130 if( n->req() == 3 && // two-input instruction 3131 n->in(1)->outcnt() > 1 && // left use is NOT a last use 3132 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 3133 n->in(2)->outcnt() == 1 &&// right use IS a last use 3134 !n->in(2)->is_Con() ) { // right use is not a constant 3135 // Check for commutative opcode 3136 switch( nop ) { 3137 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 3138 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 3139 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 3140 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 3141 case Op_AndL: case Op_XorL: case Op_OrL: 3142 case Op_AndI: case Op_XorI: case Op_OrI: { 3143 // Move "last use" input to left by swapping inputs 3144 n->swap_edges(1, 2); 3145 break; 3146 } 3147 default: 3148 break; 3149 } 3150 } 3151 3152 #ifdef ASSERT 3153 if( n->is_Mem() ) { 3154 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 3155 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw || 3156 // oop will be recorded in oop map if load crosses safepoint 3157 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 3158 LoadNode::is_immutable_value(n->in(MemNode::Address)))), 3159 "raw memory operations should have control edge"); 3160 } 3161 if (n->is_MemBar()) { 3162 MemBarNode* mb = n->as_MemBar(); 3163 if (mb->trailing_store() || mb->trailing_load_store()) { 3164 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 3165 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 3166 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 3167 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 3168 } else if (mb->leading()) { 3169 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 3170 } 3171 } 3172 #endif 3173 // Count FPU ops and common calls, implements item (3) 3174 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes); 3175 if (!gc_handled) { 3176 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes); 3177 } 3178 3179 // Collect CFG split points 3180 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3181 frc._tests.push(n); 3182 } 3183 } 3184 3185 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) { 3186 switch( nop ) { 3187 // Count all float operations that may use FPU 3188 case Op_AddF: 3189 case Op_SubF: 3190 case Op_MulF: 3191 case Op_DivF: 3192 case Op_NegF: 3193 case Op_ModF: 3194 case Op_ConvI2F: 3195 case Op_ConF: 3196 case Op_CmpF: 3197 case Op_CmpF3: 3198 case Op_StoreF: 3199 case Op_LoadF: 3200 // case Op_ConvL2F: // longs are split into 32-bit halves 3201 frc.inc_float_count(); 3202 break; 3203 3204 case Op_ConvF2D: 3205 case Op_ConvD2F: 3206 frc.inc_float_count(); 3207 frc.inc_double_count(); 3208 break; 3209 3210 // Count all double operations that may use FPU 3211 case Op_AddD: 3212 case Op_SubD: 3213 case Op_MulD: 3214 case Op_DivD: 3215 case Op_NegD: 3216 case Op_ModD: 3217 case Op_ConvI2D: 3218 case Op_ConvD2I: 3219 // case Op_ConvL2D: // handled by leaf call 3220 // case Op_ConvD2L: // handled by leaf call 3221 case Op_ConD: 3222 case Op_CmpD: 3223 case Op_CmpD3: 3224 case Op_StoreD: 3225 case Op_LoadD: 3226 case Op_LoadD_unaligned: 3227 frc.inc_double_count(); 3228 break; 3229 case Op_Opaque1: // Remove Opaque Nodes before matching 3230 case Op_Opaque3: 3231 n->subsume_by(n->in(1), this); 3232 break; 3233 case Op_CallStaticJava: 3234 case Op_CallJava: 3235 case Op_CallDynamicJava: 3236 frc.inc_java_call_count(); // Count java call site; 3237 case Op_CallRuntime: 3238 case Op_CallLeaf: 3239 case Op_CallLeafVector: 3240 case Op_CallLeafNoFP: { 3241 assert (n->is_Call(), ""); 3242 CallNode *call = n->as_Call(); 3243 // Count call sites where the FP mode bit would have to be flipped. 3244 // Do not count uncommon runtime calls: 3245 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3246 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3247 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3248 frc.inc_call_count(); // Count the call site 3249 } else { // See if uncommon argument is shared 3250 Node *n = call->in(TypeFunc::Parms); 3251 int nop = n->Opcode(); 3252 // Clone shared simple arguments to uncommon calls, item (1). 3253 if (n->outcnt() > 1 && 3254 !n->is_Proj() && 3255 nop != Op_CreateEx && 3256 nop != Op_CheckCastPP && 3257 nop != Op_DecodeN && 3258 nop != Op_DecodeNKlass && 3259 !n->is_Mem() && 3260 !n->is_Phi()) { 3261 Node *x = n->clone(); 3262 call->set_req(TypeFunc::Parms, x); 3263 } 3264 } 3265 break; 3266 } 3267 3268 case Op_StoreCM: 3269 { 3270 // Convert OopStore dependence into precedence edge 3271 Node* prec = n->in(MemNode::OopStore); 3272 n->del_req(MemNode::OopStore); 3273 n->add_prec(prec); 3274 eliminate_redundant_card_marks(n); 3275 } 3276 3277 // fall through 3278 3279 case Op_StoreB: 3280 case Op_StoreC: 3281 case Op_StoreI: 3282 case Op_StoreL: 3283 case Op_CompareAndSwapB: 3284 case Op_CompareAndSwapS: 3285 case Op_CompareAndSwapI: 3286 case Op_CompareAndSwapL: 3287 case Op_CompareAndSwapP: 3288 case Op_CompareAndSwapN: 3289 case Op_WeakCompareAndSwapB: 3290 case Op_WeakCompareAndSwapS: 3291 case Op_WeakCompareAndSwapI: 3292 case Op_WeakCompareAndSwapL: 3293 case Op_WeakCompareAndSwapP: 3294 case Op_WeakCompareAndSwapN: 3295 case Op_CompareAndExchangeB: 3296 case Op_CompareAndExchangeS: 3297 case Op_CompareAndExchangeI: 3298 case Op_CompareAndExchangeL: 3299 case Op_CompareAndExchangeP: 3300 case Op_CompareAndExchangeN: 3301 case Op_GetAndAddS: 3302 case Op_GetAndAddB: 3303 case Op_GetAndAddI: 3304 case Op_GetAndAddL: 3305 case Op_GetAndSetS: 3306 case Op_GetAndSetB: 3307 case Op_GetAndSetI: 3308 case Op_GetAndSetL: 3309 case Op_GetAndSetP: 3310 case Op_GetAndSetN: 3311 case Op_StoreP: 3312 case Op_StoreN: 3313 case Op_StoreNKlass: 3314 case Op_LoadB: 3315 case Op_LoadUB: 3316 case Op_LoadUS: 3317 case Op_LoadI: 3318 case Op_LoadKlass: 3319 case Op_LoadNKlass: 3320 case Op_LoadL: 3321 case Op_LoadL_unaligned: 3322 case Op_LoadP: 3323 case Op_LoadN: 3324 case Op_LoadRange: 3325 case Op_LoadS: 3326 break; 3327 3328 case Op_AddP: { // Assert sane base pointers 3329 Node *addp = n->in(AddPNode::Address); 3330 assert( !addp->is_AddP() || 3331 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3332 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3333 "Base pointers must match (addp %u)", addp->_idx ); 3334 #ifdef _LP64 3335 if ((UseCompressedOops || UseCompressedClassPointers) && 3336 addp->Opcode() == Op_ConP && 3337 addp == n->in(AddPNode::Base) && 3338 n->in(AddPNode::Offset)->is_Con()) { 3339 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3340 // on the platform and on the compressed oops mode. 3341 // Use addressing with narrow klass to load with offset on x86. 3342 // Some platforms can use the constant pool to load ConP. 3343 // Do this transformation here since IGVN will convert ConN back to ConP. 3344 const Type* t = addp->bottom_type(); 3345 bool is_oop = t->isa_oopptr() != nullptr; 3346 bool is_klass = t->isa_klassptr() != nullptr; 3347 3348 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3349 (is_klass && Matcher::const_klass_prefer_decode())) { 3350 Node* nn = nullptr; 3351 3352 int op = is_oop ? Op_ConN : Op_ConNKlass; 3353 3354 // Look for existing ConN node of the same exact type. 3355 Node* r = root(); 3356 uint cnt = r->outcnt(); 3357 for (uint i = 0; i < cnt; i++) { 3358 Node* m = r->raw_out(i); 3359 if (m!= nullptr && m->Opcode() == op && 3360 m->bottom_type()->make_ptr() == t) { 3361 nn = m; 3362 break; 3363 } 3364 } 3365 if (nn != nullptr) { 3366 // Decode a narrow oop to match address 3367 // [R12 + narrow_oop_reg<<3 + offset] 3368 if (is_oop) { 3369 nn = new DecodeNNode(nn, t); 3370 } else { 3371 nn = new DecodeNKlassNode(nn, t); 3372 } 3373 // Check for succeeding AddP which uses the same Base. 3374 // Otherwise we will run into the assertion above when visiting that guy. 3375 for (uint i = 0; i < n->outcnt(); ++i) { 3376 Node *out_i = n->raw_out(i); 3377 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3378 out_i->set_req(AddPNode::Base, nn); 3379 #ifdef ASSERT 3380 for (uint j = 0; j < out_i->outcnt(); ++j) { 3381 Node *out_j = out_i->raw_out(j); 3382 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3383 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3384 } 3385 #endif 3386 } 3387 } 3388 n->set_req(AddPNode::Base, nn); 3389 n->set_req(AddPNode::Address, nn); 3390 if (addp->outcnt() == 0) { 3391 addp->disconnect_inputs(this); 3392 } 3393 } 3394 } 3395 } 3396 #endif 3397 break; 3398 } 3399 3400 case Op_CastPP: { 3401 // Remove CastPP nodes to gain more freedom during scheduling but 3402 // keep the dependency they encode as control or precedence edges 3403 // (if control is set already) on memory operations. Some CastPP 3404 // nodes don't have a control (don't carry a dependency): skip 3405 // those. 3406 if (n->in(0) != nullptr) { 3407 ResourceMark rm; 3408 Unique_Node_List wq; 3409 wq.push(n); 3410 for (uint next = 0; next < wq.size(); ++next) { 3411 Node *m = wq.at(next); 3412 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3413 Node* use = m->fast_out(i); 3414 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3415 use->ensure_control_or_add_prec(n->in(0)); 3416 } else { 3417 switch(use->Opcode()) { 3418 case Op_AddP: 3419 case Op_DecodeN: 3420 case Op_DecodeNKlass: 3421 case Op_CheckCastPP: 3422 case Op_CastPP: 3423 wq.push(use); 3424 break; 3425 } 3426 } 3427 } 3428 } 3429 } 3430 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3431 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3432 Node* in1 = n->in(1); 3433 const Type* t = n->bottom_type(); 3434 Node* new_in1 = in1->clone(); 3435 new_in1->as_DecodeN()->set_type(t); 3436 3437 if (!Matcher::narrow_oop_use_complex_address()) { 3438 // 3439 // x86, ARM and friends can handle 2 adds in addressing mode 3440 // and Matcher can fold a DecodeN node into address by using 3441 // a narrow oop directly and do implicit null check in address: 3442 // 3443 // [R12 + narrow_oop_reg<<3 + offset] 3444 // NullCheck narrow_oop_reg 3445 // 3446 // On other platforms (Sparc) we have to keep new DecodeN node and 3447 // use it to do implicit null check in address: 3448 // 3449 // decode_not_null narrow_oop_reg, base_reg 3450 // [base_reg + offset] 3451 // NullCheck base_reg 3452 // 3453 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3454 // to keep the information to which null check the new DecodeN node 3455 // corresponds to use it as value in implicit_null_check(). 3456 // 3457 new_in1->set_req(0, n->in(0)); 3458 } 3459 3460 n->subsume_by(new_in1, this); 3461 if (in1->outcnt() == 0) { 3462 in1->disconnect_inputs(this); 3463 } 3464 } else { 3465 n->subsume_by(n->in(1), this); 3466 if (n->outcnt() == 0) { 3467 n->disconnect_inputs(this); 3468 } 3469 } 3470 break; 3471 } 3472 #ifdef _LP64 3473 case Op_CmpP: 3474 // Do this transformation here to preserve CmpPNode::sub() and 3475 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3476 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3477 Node* in1 = n->in(1); 3478 Node* in2 = n->in(2); 3479 if (!in1->is_DecodeNarrowPtr()) { 3480 in2 = in1; 3481 in1 = n->in(2); 3482 } 3483 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3484 3485 Node* new_in2 = nullptr; 3486 if (in2->is_DecodeNarrowPtr()) { 3487 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3488 new_in2 = in2->in(1); 3489 } else if (in2->Opcode() == Op_ConP) { 3490 const Type* t = in2->bottom_type(); 3491 if (t == TypePtr::NULL_PTR) { 3492 assert(in1->is_DecodeN(), "compare klass to null?"); 3493 // Don't convert CmpP null check into CmpN if compressed 3494 // oops implicit null check is not generated. 3495 // This will allow to generate normal oop implicit null check. 3496 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3497 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3498 // 3499 // This transformation together with CastPP transformation above 3500 // will generated code for implicit null checks for compressed oops. 3501 // 3502 // The original code after Optimize() 3503 // 3504 // LoadN memory, narrow_oop_reg 3505 // decode narrow_oop_reg, base_reg 3506 // CmpP base_reg, nullptr 3507 // CastPP base_reg // NotNull 3508 // Load [base_reg + offset], val_reg 3509 // 3510 // after these transformations will be 3511 // 3512 // LoadN memory, narrow_oop_reg 3513 // CmpN narrow_oop_reg, nullptr 3514 // decode_not_null narrow_oop_reg, base_reg 3515 // Load [base_reg + offset], val_reg 3516 // 3517 // and the uncommon path (== nullptr) will use narrow_oop_reg directly 3518 // since narrow oops can be used in debug info now (see the code in 3519 // final_graph_reshaping_walk()). 3520 // 3521 // At the end the code will be matched to 3522 // on x86: 3523 // 3524 // Load_narrow_oop memory, narrow_oop_reg 3525 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3526 // NullCheck narrow_oop_reg 3527 // 3528 // and on sparc: 3529 // 3530 // Load_narrow_oop memory, narrow_oop_reg 3531 // decode_not_null narrow_oop_reg, base_reg 3532 // Load [base_reg + offset], val_reg 3533 // NullCheck base_reg 3534 // 3535 } else if (t->isa_oopptr()) { 3536 new_in2 = ConNode::make(t->make_narrowoop()); 3537 } else if (t->isa_klassptr()) { 3538 new_in2 = ConNode::make(t->make_narrowklass()); 3539 } 3540 } 3541 if (new_in2 != nullptr) { 3542 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3543 n->subsume_by(cmpN, this); 3544 if (in1->outcnt() == 0) { 3545 in1->disconnect_inputs(this); 3546 } 3547 if (in2->outcnt() == 0) { 3548 in2->disconnect_inputs(this); 3549 } 3550 } 3551 } 3552 break; 3553 3554 case Op_DecodeN: 3555 case Op_DecodeNKlass: 3556 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3557 // DecodeN could be pinned when it can't be fold into 3558 // an address expression, see the code for Op_CastPP above. 3559 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3560 break; 3561 3562 case Op_EncodeP: 3563 case Op_EncodePKlass: { 3564 Node* in1 = n->in(1); 3565 if (in1->is_DecodeNarrowPtr()) { 3566 n->subsume_by(in1->in(1), this); 3567 } else if (in1->Opcode() == Op_ConP) { 3568 const Type* t = in1->bottom_type(); 3569 if (t == TypePtr::NULL_PTR) { 3570 assert(t->isa_oopptr(), "null klass?"); 3571 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3572 } else if (t->isa_oopptr()) { 3573 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3574 } else if (t->isa_klassptr()) { 3575 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3576 } 3577 } 3578 if (in1->outcnt() == 0) { 3579 in1->disconnect_inputs(this); 3580 } 3581 break; 3582 } 3583 3584 case Op_Proj: { 3585 if (OptimizeStringConcat || IncrementalInline) { 3586 ProjNode* proj = n->as_Proj(); 3587 if (proj->_is_io_use) { 3588 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, ""); 3589 // Separate projections were used for the exception path which 3590 // are normally removed by a late inline. If it wasn't inlined 3591 // then they will hang around and should just be replaced with 3592 // the original one. Merge them. 3593 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/); 3594 if (non_io_proj != nullptr) { 3595 proj->subsume_by(non_io_proj , this); 3596 } 3597 } 3598 } 3599 break; 3600 } 3601 3602 case Op_Phi: 3603 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3604 // The EncodeP optimization may create Phi with the same edges 3605 // for all paths. It is not handled well by Register Allocator. 3606 Node* unique_in = n->in(1); 3607 assert(unique_in != nullptr, ""); 3608 uint cnt = n->req(); 3609 for (uint i = 2; i < cnt; i++) { 3610 Node* m = n->in(i); 3611 assert(m != nullptr, ""); 3612 if (unique_in != m) 3613 unique_in = nullptr; 3614 } 3615 if (unique_in != nullptr) { 3616 n->subsume_by(unique_in, this); 3617 } 3618 } 3619 break; 3620 3621 #endif 3622 3623 #ifdef ASSERT 3624 case Op_CastII: 3625 // Verify that all range check dependent CastII nodes were removed. 3626 if (n->isa_CastII()->has_range_check()) { 3627 n->dump(3); 3628 assert(false, "Range check dependent CastII node was not removed"); 3629 } 3630 break; 3631 #endif 3632 3633 case Op_ModI: 3634 if (UseDivMod) { 3635 // Check if a%b and a/b both exist 3636 Node* d = n->find_similar(Op_DivI); 3637 if (d) { 3638 // Replace them with a fused divmod if supported 3639 if (Matcher::has_match_rule(Op_DivModI)) { 3640 DivModINode* divmod = DivModINode::make(n); 3641 d->subsume_by(divmod->div_proj(), this); 3642 n->subsume_by(divmod->mod_proj(), this); 3643 } else { 3644 // replace a%b with a-((a/b)*b) 3645 Node* mult = new MulINode(d, d->in(2)); 3646 Node* sub = new SubINode(d->in(1), mult); 3647 n->subsume_by(sub, this); 3648 } 3649 } 3650 } 3651 break; 3652 3653 case Op_ModL: 3654 if (UseDivMod) { 3655 // Check if a%b and a/b both exist 3656 Node* d = n->find_similar(Op_DivL); 3657 if (d) { 3658 // Replace them with a fused divmod if supported 3659 if (Matcher::has_match_rule(Op_DivModL)) { 3660 DivModLNode* divmod = DivModLNode::make(n); 3661 d->subsume_by(divmod->div_proj(), this); 3662 n->subsume_by(divmod->mod_proj(), this); 3663 } else { 3664 // replace a%b with a-((a/b)*b) 3665 Node* mult = new MulLNode(d, d->in(2)); 3666 Node* sub = new SubLNode(d->in(1), mult); 3667 n->subsume_by(sub, this); 3668 } 3669 } 3670 } 3671 break; 3672 3673 case Op_UModI: 3674 if (UseDivMod) { 3675 // Check if a%b and a/b both exist 3676 Node* d = n->find_similar(Op_UDivI); 3677 if (d) { 3678 // Replace them with a fused unsigned divmod if supported 3679 if (Matcher::has_match_rule(Op_UDivModI)) { 3680 UDivModINode* divmod = UDivModINode::make(n); 3681 d->subsume_by(divmod->div_proj(), this); 3682 n->subsume_by(divmod->mod_proj(), this); 3683 } else { 3684 // replace a%b with a-((a/b)*b) 3685 Node* mult = new MulINode(d, d->in(2)); 3686 Node* sub = new SubINode(d->in(1), mult); 3687 n->subsume_by(sub, this); 3688 } 3689 } 3690 } 3691 break; 3692 3693 case Op_UModL: 3694 if (UseDivMod) { 3695 // Check if a%b and a/b both exist 3696 Node* d = n->find_similar(Op_UDivL); 3697 if (d) { 3698 // Replace them with a fused unsigned divmod if supported 3699 if (Matcher::has_match_rule(Op_UDivModL)) { 3700 UDivModLNode* divmod = UDivModLNode::make(n); 3701 d->subsume_by(divmod->div_proj(), this); 3702 n->subsume_by(divmod->mod_proj(), this); 3703 } else { 3704 // replace a%b with a-((a/b)*b) 3705 Node* mult = new MulLNode(d, d->in(2)); 3706 Node* sub = new SubLNode(d->in(1), mult); 3707 n->subsume_by(sub, this); 3708 } 3709 } 3710 } 3711 break; 3712 3713 case Op_LoadVector: 3714 case Op_StoreVector: 3715 #ifdef ASSERT 3716 // Add VerifyVectorAlignment node between adr and load / store. 3717 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) { 3718 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() : 3719 n->as_StoreVector()->must_verify_alignment(); 3720 if (must_verify_alignment) { 3721 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() : 3722 n->as_StoreVector()->memory_size(); 3723 // The memory access should be aligned to the vector width in bytes. 3724 // However, the underlying array is possibly less well aligned, but at least 3725 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in 3726 // a loop we can expect at least the following alignment: 3727 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes); 3728 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range"); 3729 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2"); 3730 // Create mask from alignment. e.g. 0b1000 -> 0b0111 3731 jlong mask = guaranteed_alignment - 1; 3732 Node* mask_con = ConLNode::make(mask); 3733 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con); 3734 n->set_req(MemNode::Address, va); 3735 } 3736 } 3737 #endif 3738 break; 3739 3740 case Op_LoadVectorGather: 3741 case Op_StoreVectorScatter: 3742 case Op_LoadVectorGatherMasked: 3743 case Op_StoreVectorScatterMasked: 3744 case Op_VectorCmpMasked: 3745 case Op_VectorMaskGen: 3746 case Op_LoadVectorMasked: 3747 case Op_StoreVectorMasked: 3748 break; 3749 3750 case Op_AddReductionVI: 3751 case Op_AddReductionVL: 3752 case Op_AddReductionVF: 3753 case Op_AddReductionVD: 3754 case Op_MulReductionVI: 3755 case Op_MulReductionVL: 3756 case Op_MulReductionVF: 3757 case Op_MulReductionVD: 3758 case Op_MinReductionV: 3759 case Op_MaxReductionV: 3760 case Op_AndReductionV: 3761 case Op_OrReductionV: 3762 case Op_XorReductionV: 3763 break; 3764 3765 case Op_PackB: 3766 case Op_PackS: 3767 case Op_PackI: 3768 case Op_PackF: 3769 case Op_PackL: 3770 case Op_PackD: 3771 if (n->req()-1 > 2) { 3772 // Replace many operand PackNodes with a binary tree for matching 3773 PackNode* p = (PackNode*) n; 3774 Node* btp = p->binary_tree_pack(1, n->req()); 3775 n->subsume_by(btp, this); 3776 } 3777 break; 3778 case Op_Loop: 3779 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop"); 3780 case Op_CountedLoop: 3781 case Op_LongCountedLoop: 3782 case Op_OuterStripMinedLoop: 3783 if (n->as_Loop()->is_inner_loop()) { 3784 frc.inc_inner_loop_count(); 3785 } 3786 n->as_Loop()->verify_strip_mined(0); 3787 break; 3788 case Op_LShiftI: 3789 case Op_RShiftI: 3790 case Op_URShiftI: 3791 case Op_LShiftL: 3792 case Op_RShiftL: 3793 case Op_URShiftL: 3794 if (Matcher::need_masked_shift_count) { 3795 // The cpu's shift instructions don't restrict the count to the 3796 // lower 5/6 bits. We need to do the masking ourselves. 3797 Node* in2 = n->in(2); 3798 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3799 const TypeInt* t = in2->find_int_type(); 3800 if (t != nullptr && t->is_con()) { 3801 juint shift = t->get_con(); 3802 if (shift > mask) { // Unsigned cmp 3803 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3804 } 3805 } else { 3806 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) { 3807 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3808 n->set_req(2, shift); 3809 } 3810 } 3811 if (in2->outcnt() == 0) { // Remove dead node 3812 in2->disconnect_inputs(this); 3813 } 3814 } 3815 break; 3816 case Op_MemBarStoreStore: 3817 case Op_MemBarRelease: 3818 // Break the link with AllocateNode: it is no longer useful and 3819 // confuses register allocation. 3820 if (n->req() > MemBarNode::Precedent) { 3821 n->set_req(MemBarNode::Precedent, top()); 3822 } 3823 break; 3824 case Op_MemBarAcquire: { 3825 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3826 // At parse time, the trailing MemBarAcquire for a volatile load 3827 // is created with an edge to the load. After optimizations, 3828 // that input may be a chain of Phis. If those phis have no 3829 // other use, then the MemBarAcquire keeps them alive and 3830 // register allocation can be confused. 3831 dead_nodes.push(n->in(MemBarNode::Precedent)); 3832 n->set_req(MemBarNode::Precedent, top()); 3833 } 3834 break; 3835 } 3836 case Op_Blackhole: 3837 break; 3838 case Op_RangeCheck: { 3839 RangeCheckNode* rc = n->as_RangeCheck(); 3840 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3841 n->subsume_by(iff, this); 3842 frc._tests.push(iff); 3843 break; 3844 } 3845 case Op_ConvI2L: { 3846 if (!Matcher::convi2l_type_required) { 3847 // Code generation on some platforms doesn't need accurate 3848 // ConvI2L types. Widening the type can help remove redundant 3849 // address computations. 3850 n->as_Type()->set_type(TypeLong::INT); 3851 ResourceMark rm; 3852 Unique_Node_List wq; 3853 wq.push(n); 3854 for (uint next = 0; next < wq.size(); next++) { 3855 Node *m = wq.at(next); 3856 3857 for(;;) { 3858 // Loop over all nodes with identical inputs edges as m 3859 Node* k = m->find_similar(m->Opcode()); 3860 if (k == nullptr) { 3861 break; 3862 } 3863 // Push their uses so we get a chance to remove node made 3864 // redundant 3865 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3866 Node* u = k->fast_out(i); 3867 if (u->Opcode() == Op_LShiftL || 3868 u->Opcode() == Op_AddL || 3869 u->Opcode() == Op_SubL || 3870 u->Opcode() == Op_AddP) { 3871 wq.push(u); 3872 } 3873 } 3874 // Replace all nodes with identical edges as m with m 3875 k->subsume_by(m, this); 3876 } 3877 } 3878 } 3879 break; 3880 } 3881 case Op_CmpUL: { 3882 if (!Matcher::has_match_rule(Op_CmpUL)) { 3883 // No support for unsigned long comparisons 3884 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3885 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3886 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3887 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3888 Node* andl = new AndLNode(orl, remove_sign_mask); 3889 Node* cmp = new CmpLNode(andl, n->in(2)); 3890 n->subsume_by(cmp, this); 3891 } 3892 break; 3893 } 3894 default: 3895 assert(!n->is_Call(), ""); 3896 assert(!n->is_Mem(), ""); 3897 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3898 break; 3899 } 3900 } 3901 3902 //------------------------------final_graph_reshaping_walk--------------------- 3903 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3904 // requires that the walk visits a node's inputs before visiting the node. 3905 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3906 Unique_Node_List sfpt; 3907 3908 frc._visited.set(root->_idx); // first, mark node as visited 3909 uint cnt = root->req(); 3910 Node *n = root; 3911 uint i = 0; 3912 while (true) { 3913 if (i < cnt) { 3914 // Place all non-visited non-null inputs onto stack 3915 Node* m = n->in(i); 3916 ++i; 3917 if (m != nullptr && !frc._visited.test_set(m->_idx)) { 3918 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) { 3919 // compute worst case interpreter size in case of a deoptimization 3920 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3921 3922 sfpt.push(m); 3923 } 3924 cnt = m->req(); 3925 nstack.push(n, i); // put on stack parent and next input's index 3926 n = m; 3927 i = 0; 3928 } 3929 } else { 3930 // Now do post-visit work 3931 final_graph_reshaping_impl(n, frc, dead_nodes); 3932 if (nstack.is_empty()) 3933 break; // finished 3934 n = nstack.node(); // Get node from stack 3935 cnt = n->req(); 3936 i = nstack.index(); 3937 nstack.pop(); // Shift to the next node on stack 3938 } 3939 } 3940 3941 // Skip next transformation if compressed oops are not used. 3942 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3943 (!UseCompressedOops && !UseCompressedClassPointers)) 3944 return; 3945 3946 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3947 // It could be done for an uncommon traps or any safepoints/calls 3948 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3949 while (sfpt.size() > 0) { 3950 n = sfpt.pop(); 3951 JVMState *jvms = n->as_SafePoint()->jvms(); 3952 assert(jvms != nullptr, "sanity"); 3953 int start = jvms->debug_start(); 3954 int end = n->req(); 3955 bool is_uncommon = (n->is_CallStaticJava() && 3956 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3957 for (int j = start; j < end; j++) { 3958 Node* in = n->in(j); 3959 if (in->is_DecodeNarrowPtr()) { 3960 bool safe_to_skip = true; 3961 if (!is_uncommon ) { 3962 // Is it safe to skip? 3963 for (uint i = 0; i < in->outcnt(); i++) { 3964 Node* u = in->raw_out(i); 3965 if (!u->is_SafePoint() || 3966 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3967 safe_to_skip = false; 3968 } 3969 } 3970 } 3971 if (safe_to_skip) { 3972 n->set_req(j, in->in(1)); 3973 } 3974 if (in->outcnt() == 0) { 3975 in->disconnect_inputs(this); 3976 } 3977 } 3978 } 3979 } 3980 } 3981 3982 //------------------------------final_graph_reshaping-------------------------- 3983 // Final Graph Reshaping. 3984 // 3985 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3986 // and not commoned up and forced early. Must come after regular 3987 // optimizations to avoid GVN undoing the cloning. Clone constant 3988 // inputs to Loop Phis; these will be split by the allocator anyways. 3989 // Remove Opaque nodes. 3990 // (2) Move last-uses by commutative operations to the left input to encourage 3991 // Intel update-in-place two-address operations and better register usage 3992 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3993 // calls canonicalizing them back. 3994 // (3) Count the number of double-precision FP ops, single-precision FP ops 3995 // and call sites. On Intel, we can get correct rounding either by 3996 // forcing singles to memory (requires extra stores and loads after each 3997 // FP bytecode) or we can set a rounding mode bit (requires setting and 3998 // clearing the mode bit around call sites). The mode bit is only used 3999 // if the relative frequency of single FP ops to calls is low enough. 4000 // This is a key transform for SPEC mpeg_audio. 4001 // (4) Detect infinite loops; blobs of code reachable from above but not 4002 // below. Several of the Code_Gen algorithms fail on such code shapes, 4003 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 4004 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 4005 // Detection is by looking for IfNodes where only 1 projection is 4006 // reachable from below or CatchNodes missing some targets. 4007 // (5) Assert for insane oop offsets in debug mode. 4008 4009 bool Compile::final_graph_reshaping() { 4010 // an infinite loop may have been eliminated by the optimizer, 4011 // in which case the graph will be empty. 4012 if (root()->req() == 1) { 4013 // Do not compile method that is only a trivial infinite loop, 4014 // since the content of the loop may have been eliminated. 4015 record_method_not_compilable("trivial infinite loop"); 4016 return true; 4017 } 4018 4019 // Expensive nodes have their control input set to prevent the GVN 4020 // from freely commoning them. There's no GVN beyond this point so 4021 // no need to keep the control input. We want the expensive nodes to 4022 // be freely moved to the least frequent code path by gcm. 4023 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 4024 for (int i = 0; i < expensive_count(); i++) { 4025 _expensive_nodes.at(i)->set_req(0, nullptr); 4026 } 4027 4028 Final_Reshape_Counts frc; 4029 4030 // Visit everybody reachable! 4031 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 4032 Node_Stack nstack(live_nodes() >> 1); 4033 Unique_Node_List dead_nodes; 4034 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes); 4035 4036 // Check for unreachable (from below) code (i.e., infinite loops). 4037 for( uint i = 0; i < frc._tests.size(); i++ ) { 4038 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 4039 // Get number of CFG targets. 4040 // Note that PCTables include exception targets after calls. 4041 uint required_outcnt = n->required_outcnt(); 4042 if (n->outcnt() != required_outcnt) { 4043 // Check for a few special cases. Rethrow Nodes never take the 4044 // 'fall-thru' path, so expected kids is 1 less. 4045 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 4046 if (n->in(0)->in(0)->is_Call()) { 4047 CallNode* call = n->in(0)->in(0)->as_Call(); 4048 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 4049 required_outcnt--; // Rethrow always has 1 less kid 4050 } else if (call->req() > TypeFunc::Parms && 4051 call->is_CallDynamicJava()) { 4052 // Check for null receiver. In such case, the optimizer has 4053 // detected that the virtual call will always result in a null 4054 // pointer exception. The fall-through projection of this CatchNode 4055 // will not be populated. 4056 Node* arg0 = call->in(TypeFunc::Parms); 4057 if (arg0->is_Type() && 4058 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 4059 required_outcnt--; 4060 } 4061 } else if (call->entry_point() == OptoRuntime::new_array_Java() || 4062 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4063 // Check for illegal array length. In such case, the optimizer has 4064 // detected that the allocation attempt will always result in an 4065 // exception. There is no fall-through projection of this CatchNode . 4066 assert(call->is_CallStaticJava(), "static call expected"); 4067 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4068 uint valid_length_test_input = call->req() - 1; 4069 Node* valid_length_test = call->in(valid_length_test_input); 4070 call->del_req(valid_length_test_input); 4071 if (valid_length_test->find_int_con(1) == 0) { 4072 required_outcnt--; 4073 } 4074 dead_nodes.push(valid_length_test); 4075 assert(n->outcnt() == required_outcnt, "malformed control flow"); 4076 continue; 4077 } 4078 } 4079 } 4080 4081 // Recheck with a better notion of 'required_outcnt' 4082 if (n->outcnt() != required_outcnt) { 4083 record_method_not_compilable("malformed control flow"); 4084 return true; // Not all targets reachable! 4085 } 4086 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) { 4087 CallNode* call = n->in(0)->in(0)->as_Call(); 4088 if (call->entry_point() == OptoRuntime::new_array_Java() || 4089 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4090 assert(call->is_CallStaticJava(), "static call expected"); 4091 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4092 uint valid_length_test_input = call->req() - 1; 4093 dead_nodes.push(call->in(valid_length_test_input)); 4094 call->del_req(valid_length_test_input); // valid length test useless now 4095 } 4096 } 4097 // Check that I actually visited all kids. Unreached kids 4098 // must be infinite loops. 4099 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 4100 if (!frc._visited.test(n->fast_out(j)->_idx)) { 4101 record_method_not_compilable("infinite loop"); 4102 return true; // Found unvisited kid; must be unreach 4103 } 4104 4105 // Here so verification code in final_graph_reshaping_walk() 4106 // always see an OuterStripMinedLoopEnd 4107 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) { 4108 IfNode* init_iff = n->as_If(); 4109 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 4110 n->subsume_by(iff, this); 4111 } 4112 } 4113 4114 while (dead_nodes.size() > 0) { 4115 Node* m = dead_nodes.pop(); 4116 if (m->outcnt() == 0 && m != top()) { 4117 for (uint j = 0; j < m->req(); j++) { 4118 Node* in = m->in(j); 4119 if (in != nullptr) { 4120 dead_nodes.push(in); 4121 } 4122 } 4123 m->disconnect_inputs(this); 4124 } 4125 } 4126 4127 #ifdef IA32 4128 // If original bytecodes contained a mixture of floats and doubles 4129 // check if the optimizer has made it homogeneous, item (3). 4130 if (UseSSE == 0 && 4131 frc.get_float_count() > 32 && 4132 frc.get_double_count() == 0 && 4133 (10 * frc.get_call_count() < frc.get_float_count()) ) { 4134 set_24_bit_selection_and_mode(false, true); 4135 } 4136 #endif // IA32 4137 4138 set_java_calls(frc.get_java_call_count()); 4139 set_inner_loops(frc.get_inner_loop_count()); 4140 4141 // No infinite loops, no reason to bail out. 4142 return false; 4143 } 4144 4145 //-----------------------------too_many_traps---------------------------------- 4146 // Report if there are too many traps at the current method and bci. 4147 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 4148 bool Compile::too_many_traps(ciMethod* method, 4149 int bci, 4150 Deoptimization::DeoptReason reason) { 4151 ciMethodData* md = method->method_data(); 4152 if (md->is_empty()) { 4153 // Assume the trap has not occurred, or that it occurred only 4154 // because of a transient condition during start-up in the interpreter. 4155 return false; 4156 } 4157 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4158 if (md->has_trap_at(bci, m, reason) != 0) { 4159 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 4160 // Also, if there are multiple reasons, or if there is no per-BCI record, 4161 // assume the worst. 4162 if (log()) 4163 log()->elem("observe trap='%s' count='%d'", 4164 Deoptimization::trap_reason_name(reason), 4165 md->trap_count(reason)); 4166 return true; 4167 } else { 4168 // Ignore method/bci and see if there have been too many globally. 4169 return too_many_traps(reason, md); 4170 } 4171 } 4172 4173 // Less-accurate variant which does not require a method and bci. 4174 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 4175 ciMethodData* logmd) { 4176 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 4177 // Too many traps globally. 4178 // Note that we use cumulative trap_count, not just md->trap_count. 4179 if (log()) { 4180 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason); 4181 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 4182 Deoptimization::trap_reason_name(reason), 4183 mcount, trap_count(reason)); 4184 } 4185 return true; 4186 } else { 4187 // The coast is clear. 4188 return false; 4189 } 4190 } 4191 4192 //--------------------------too_many_recompiles-------------------------------- 4193 // Report if there are too many recompiles at the current method and bci. 4194 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 4195 // Is not eager to return true, since this will cause the compiler to use 4196 // Action_none for a trap point, to avoid too many recompilations. 4197 bool Compile::too_many_recompiles(ciMethod* method, 4198 int bci, 4199 Deoptimization::DeoptReason reason) { 4200 ciMethodData* md = method->method_data(); 4201 if (md->is_empty()) { 4202 // Assume the trap has not occurred, or that it occurred only 4203 // because of a transient condition during start-up in the interpreter. 4204 return false; 4205 } 4206 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 4207 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 4208 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 4209 Deoptimization::DeoptReason per_bc_reason 4210 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 4211 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4212 if ((per_bc_reason == Deoptimization::Reason_none 4213 || md->has_trap_at(bci, m, reason) != 0) 4214 // The trap frequency measure we care about is the recompile count: 4215 && md->trap_recompiled_at(bci, m) 4216 && md->overflow_recompile_count() >= bc_cutoff) { 4217 // Do not emit a trap here if it has already caused recompilations. 4218 // Also, if there are multiple reasons, or if there is no per-BCI record, 4219 // assume the worst. 4220 if (log()) 4221 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 4222 Deoptimization::trap_reason_name(reason), 4223 md->trap_count(reason), 4224 md->overflow_recompile_count()); 4225 return true; 4226 } else if (trap_count(reason) != 0 4227 && decompile_count() >= m_cutoff) { 4228 // Too many recompiles globally, and we have seen this sort of trap. 4229 // Use cumulative decompile_count, not just md->decompile_count. 4230 if (log()) 4231 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 4232 Deoptimization::trap_reason_name(reason), 4233 md->trap_count(reason), trap_count(reason), 4234 md->decompile_count(), decompile_count()); 4235 return true; 4236 } else { 4237 // The coast is clear. 4238 return false; 4239 } 4240 } 4241 4242 // Compute when not to trap. Used by matching trap based nodes and 4243 // NullCheck optimization. 4244 void Compile::set_allowed_deopt_reasons() { 4245 _allowed_reasons = 0; 4246 if (is_method_compilation()) { 4247 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 4248 assert(rs < BitsPerInt, "recode bit map"); 4249 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 4250 _allowed_reasons |= nth_bit(rs); 4251 } 4252 } 4253 } 4254 } 4255 4256 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 4257 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 4258 } 4259 4260 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 4261 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 4262 } 4263 4264 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 4265 if (holder->is_initialized()) { 4266 return false; 4267 } 4268 if (holder->is_being_initialized()) { 4269 if (accessing_method->holder() == holder) { 4270 // Access inside a class. The barrier can be elided when access happens in <clinit>, 4271 // <init>, or a static method. In all those cases, there was an initialization 4272 // barrier on the holder klass passed. 4273 if (accessing_method->is_static_initializer() || 4274 accessing_method->is_object_initializer() || 4275 accessing_method->is_static()) { 4276 return false; 4277 } 4278 } else if (accessing_method->holder()->is_subclass_of(holder)) { 4279 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 4280 // In case of <init> or a static method, the barrier is on the subclass is not enough: 4281 // child class can become fully initialized while its parent class is still being initialized. 4282 if (accessing_method->is_static_initializer()) { 4283 return false; 4284 } 4285 } 4286 ciMethod* root = method(); // the root method of compilation 4287 if (root != accessing_method) { 4288 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 4289 } 4290 } 4291 return true; 4292 } 4293 4294 #ifndef PRODUCT 4295 //------------------------------verify_bidirectional_edges--------------------- 4296 // For each input edge to a node (ie - for each Use-Def edge), verify that 4297 // there is a corresponding Def-Use edge. 4298 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) { 4299 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc 4300 uint stack_size = live_nodes() >> 4; 4301 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize)); 4302 nstack.push(_root); 4303 4304 while (nstack.size() > 0) { 4305 Node* n = nstack.pop(); 4306 if (visited.member(n)) { 4307 continue; 4308 } 4309 visited.push(n); 4310 4311 // Walk over all input edges, checking for correspondence 4312 uint length = n->len(); 4313 for (uint i = 0; i < length; i++) { 4314 Node* in = n->in(i); 4315 if (in != nullptr && !visited.member(in)) { 4316 nstack.push(in); // Put it on stack 4317 } 4318 if (in != nullptr && !in->is_top()) { 4319 // Count instances of `next` 4320 int cnt = 0; 4321 for (uint idx = 0; idx < in->_outcnt; idx++) { 4322 if (in->_out[idx] == n) { 4323 cnt++; 4324 } 4325 } 4326 assert(cnt > 0, "Failed to find Def-Use edge."); 4327 // Check for duplicate edges 4328 // walk the input array downcounting the input edges to n 4329 for (uint j = 0; j < length; j++) { 4330 if (n->in(j) == in) { 4331 cnt--; 4332 } 4333 } 4334 assert(cnt == 0, "Mismatched edge count."); 4335 } else if (in == nullptr) { 4336 assert(i == 0 || i >= n->req() || 4337 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() || 4338 (n->is_Unlock() && i == (n->req() - 1)) || 4339 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion 4340 "only region, phi, arraycopy, unlock or membar nodes have null data edges"); 4341 } else { 4342 assert(in->is_top(), "sanity"); 4343 // Nothing to check. 4344 } 4345 } 4346 } 4347 } 4348 4349 //------------------------------verify_graph_edges--------------------------- 4350 // Walk the Graph and verify that there is a one-to-one correspondence 4351 // between Use-Def edges and Def-Use edges in the graph. 4352 void Compile::verify_graph_edges(bool no_dead_code) { 4353 if (VerifyGraphEdges) { 4354 Unique_Node_List visited; 4355 4356 // Call graph walk to check edges 4357 verify_bidirectional_edges(visited); 4358 if (no_dead_code) { 4359 // Now make sure that no visited node is used by an unvisited node. 4360 bool dead_nodes = false; 4361 Unique_Node_List checked; 4362 while (visited.size() > 0) { 4363 Node* n = visited.pop(); 4364 checked.push(n); 4365 for (uint i = 0; i < n->outcnt(); i++) { 4366 Node* use = n->raw_out(i); 4367 if (checked.member(use)) continue; // already checked 4368 if (visited.member(use)) continue; // already in the graph 4369 if (use->is_Con()) continue; // a dead ConNode is OK 4370 // At this point, we have found a dead node which is DU-reachable. 4371 if (!dead_nodes) { 4372 tty->print_cr("*** Dead nodes reachable via DU edges:"); 4373 dead_nodes = true; 4374 } 4375 use->dump(2); 4376 tty->print_cr("---"); 4377 checked.push(use); // No repeats; pretend it is now checked. 4378 } 4379 } 4380 assert(!dead_nodes, "using nodes must be reachable from root"); 4381 } 4382 } 4383 } 4384 #endif 4385 4386 // The Compile object keeps track of failure reasons separately from the ciEnv. 4387 // This is required because there is not quite a 1-1 relation between the 4388 // ciEnv and its compilation task and the Compile object. Note that one 4389 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4390 // to backtrack and retry without subsuming loads. Other than this backtracking 4391 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4392 // by the logic in C2Compiler. 4393 void Compile::record_failure(const char* reason) { 4394 if (log() != nullptr) { 4395 log()->elem("failure reason='%s' phase='compile'", reason); 4396 } 4397 if (_failure_reason == nullptr) { 4398 // Record the first failure reason. 4399 _failure_reason = reason; 4400 if (CaptureBailoutInformation) { 4401 _first_failure_details = new CompilationFailureInfo(reason); 4402 } 4403 } 4404 4405 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4406 C->print_method(PHASE_FAILURE, 1); 4407 } 4408 _root = nullptr; // flush the graph, too 4409 } 4410 4411 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 4412 : TraceTime(name, accumulator, CITime, CITimeVerbose), 4413 _compile(Compile::current()), 4414 _log(nullptr), 4415 _phase_name(name), 4416 _dolog(CITimeVerbose) 4417 { 4418 assert(_compile != nullptr, "sanity check"); 4419 if (_dolog) { 4420 _log = _compile->log(); 4421 } 4422 if (_log != nullptr) { 4423 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes()); 4424 _log->stamp(); 4425 _log->end_head(); 4426 } 4427 } 4428 4429 Compile::TracePhase::~TracePhase() { 4430 if (_compile->failing()) return; 4431 #ifdef ASSERT 4432 if (PrintIdealNodeCount) { 4433 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4434 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk()); 4435 } 4436 4437 if (VerifyIdealNodeCount) { 4438 _compile->print_missing_nodes(); 4439 } 4440 #endif 4441 4442 if (_log != nullptr) { 4443 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes()); 4444 } 4445 } 4446 4447 //----------------------------static_subtype_check----------------------------- 4448 // Shortcut important common cases when superklass is exact: 4449 // (0) superklass is java.lang.Object (can occur in reflective code) 4450 // (1) subklass is already limited to a subtype of superklass => always ok 4451 // (2) subklass does not overlap with superklass => always fail 4452 // (3) superklass has NO subtypes and we can check with a simple compare. 4453 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) { 4454 if (skip) { 4455 return SSC_full_test; // Let caller generate the general case. 4456 } 4457 4458 if (subk->is_java_subtype_of(superk)) { 4459 return SSC_always_true; // (0) and (1) this test cannot fail 4460 } 4461 4462 if (!subk->maybe_java_subtype_of(superk)) { 4463 return SSC_always_false; // (2) true path dead; no dynamic test needed 4464 } 4465 4466 const Type* superelem = superk; 4467 if (superk->isa_aryklassptr()) { 4468 int ignored; 4469 superelem = superk->is_aryklassptr()->base_element_type(ignored); 4470 } 4471 4472 if (superelem->isa_instklassptr()) { 4473 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass(); 4474 if (!ik->has_subklass()) { 4475 if (!ik->is_final()) { 4476 // Add a dependency if there is a chance of a later subclass. 4477 dependencies()->assert_leaf_type(ik); 4478 } 4479 if (!superk->maybe_java_subtype_of(subk)) { 4480 return SSC_always_false; 4481 } 4482 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4483 } 4484 } else { 4485 // A primitive array type has no subtypes. 4486 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4487 } 4488 4489 return SSC_full_test; 4490 } 4491 4492 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4493 #ifdef _LP64 4494 // The scaled index operand to AddP must be a clean 64-bit value. 4495 // Java allows a 32-bit int to be incremented to a negative 4496 // value, which appears in a 64-bit register as a large 4497 // positive number. Using that large positive number as an 4498 // operand in pointer arithmetic has bad consequences. 4499 // On the other hand, 32-bit overflow is rare, and the possibility 4500 // can often be excluded, if we annotate the ConvI2L node with 4501 // a type assertion that its value is known to be a small positive 4502 // number. (The prior range check has ensured this.) 4503 // This assertion is used by ConvI2LNode::Ideal. 4504 int index_max = max_jint - 1; // array size is max_jint, index is one less 4505 if (sizetype != nullptr) index_max = sizetype->_hi - 1; 4506 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4507 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4508 #endif 4509 return idx; 4510 } 4511 4512 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4513 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 4514 if (ctrl != nullptr) { 4515 // Express control dependency by a CastII node with a narrow type. 4516 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4517 // node from floating above the range check during loop optimizations. Otherwise, the 4518 // ConvI2L node may be eliminated independently of the range check, causing the data path 4519 // to become TOP while the control path is still there (although it's unreachable). 4520 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */); 4521 value = phase->transform(value); 4522 } 4523 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4524 return phase->transform(new ConvI2LNode(value, ltype)); 4525 } 4526 4527 // The message about the current inlining is accumulated in 4528 // _print_inlining_stream and transferred into the _print_inlining_list 4529 // once we know whether inlining succeeds or not. For regular 4530 // inlining, messages are appended to the buffer pointed by 4531 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4532 // a new buffer is added after _print_inlining_idx in the list. This 4533 // way we can update the inlining message for late inlining call site 4534 // when the inlining is attempted again. 4535 void Compile::print_inlining_init() { 4536 if (print_inlining() || print_intrinsics()) { 4537 // print_inlining_init is actually called several times. 4538 print_inlining_reset(); 4539 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer()); 4540 } 4541 } 4542 4543 void Compile::print_inlining_reinit() { 4544 if (print_inlining() || print_intrinsics()) { 4545 print_inlining_reset(); 4546 } 4547 } 4548 4549 void Compile::print_inlining_reset() { 4550 _print_inlining_stream->reset(); 4551 } 4552 4553 void Compile::print_inlining_commit() { 4554 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4555 // Transfer the message from _print_inlining_stream to the current 4556 // _print_inlining_list buffer and clear _print_inlining_stream. 4557 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 4558 print_inlining_reset(); 4559 } 4560 4561 void Compile::print_inlining_push() { 4562 // Add new buffer to the _print_inlining_list at current position 4563 _print_inlining_idx++; 4564 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer()); 4565 } 4566 4567 Compile::PrintInliningBuffer* Compile::print_inlining_current() { 4568 return _print_inlining_list->at(_print_inlining_idx); 4569 } 4570 4571 void Compile::print_inlining_update(CallGenerator* cg) { 4572 if (print_inlining() || print_intrinsics()) { 4573 if (cg->is_late_inline()) { 4574 if (print_inlining_current()->cg() != cg && 4575 (print_inlining_current()->cg() != nullptr || 4576 print_inlining_current()->ss()->size() != 0)) { 4577 print_inlining_push(); 4578 } 4579 print_inlining_commit(); 4580 print_inlining_current()->set_cg(cg); 4581 } else { 4582 if (print_inlining_current()->cg() != nullptr) { 4583 print_inlining_push(); 4584 } 4585 print_inlining_commit(); 4586 } 4587 } 4588 } 4589 4590 void Compile::print_inlining_move_to(CallGenerator* cg) { 4591 // We resume inlining at a late inlining call site. Locate the 4592 // corresponding inlining buffer so that we can update it. 4593 if (print_inlining() || print_intrinsics()) { 4594 for (int i = 0; i < _print_inlining_list->length(); i++) { 4595 if (_print_inlining_list->at(i)->cg() == cg) { 4596 _print_inlining_idx = i; 4597 return; 4598 } 4599 } 4600 ShouldNotReachHere(); 4601 } 4602 } 4603 4604 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4605 if (print_inlining() || print_intrinsics()) { 4606 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4607 assert(print_inlining_current()->cg() == cg, "wrong entry"); 4608 // replace message with new message 4609 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer()); 4610 print_inlining_commit(); 4611 print_inlining_current()->set_cg(cg); 4612 } 4613 } 4614 4615 void Compile::print_inlining_assert_ready() { 4616 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data"); 4617 } 4618 4619 void Compile::process_print_inlining() { 4620 assert(_late_inlines.length() == 0, "not drained yet"); 4621 if (print_inlining() || print_intrinsics()) { 4622 ResourceMark rm; 4623 stringStream ss; 4624 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once."); 4625 for (int i = 0; i < _print_inlining_list->length(); i++) { 4626 PrintInliningBuffer* pib = _print_inlining_list->at(i); 4627 ss.print("%s", pib->ss()->freeze()); 4628 delete pib; 4629 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr)); 4630 } 4631 // Reset _print_inlining_list, it only contains destructed objects. 4632 // It is on the arena, so it will be freed when the arena is reset. 4633 _print_inlining_list = nullptr; 4634 // _print_inlining_stream won't be used anymore, either. 4635 print_inlining_reset(); 4636 size_t end = ss.size(); 4637 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4638 strncpy(_print_inlining_output, ss.freeze(), end+1); 4639 _print_inlining_output[end] = 0; 4640 } 4641 } 4642 4643 void Compile::dump_print_inlining() { 4644 if (_print_inlining_output != nullptr) { 4645 tty->print_raw(_print_inlining_output); 4646 } 4647 } 4648 4649 void Compile::log_late_inline(CallGenerator* cg) { 4650 if (log() != nullptr) { 4651 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4652 cg->unique_id()); 4653 JVMState* p = cg->call_node()->jvms(); 4654 while (p != nullptr) { 4655 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4656 p = p->caller(); 4657 } 4658 log()->tail("late_inline"); 4659 } 4660 } 4661 4662 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4663 log_late_inline(cg); 4664 if (log() != nullptr) { 4665 log()->inline_fail(msg); 4666 } 4667 } 4668 4669 void Compile::log_inline_id(CallGenerator* cg) { 4670 if (log() != nullptr) { 4671 // The LogCompilation tool needs a unique way to identify late 4672 // inline call sites. This id must be unique for this call site in 4673 // this compilation. Try to have it unique across compilations as 4674 // well because it can be convenient when grepping through the log 4675 // file. 4676 // Distinguish OSR compilations from others in case CICountOSR is 4677 // on. 4678 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4679 cg->set_unique_id(id); 4680 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4681 } 4682 } 4683 4684 void Compile::log_inline_failure(const char* msg) { 4685 if (C->log() != nullptr) { 4686 C->log()->inline_fail(msg); 4687 } 4688 } 4689 4690 4691 // Dump inlining replay data to the stream. 4692 // Don't change thread state and acquire any locks. 4693 void Compile::dump_inline_data(outputStream* out) { 4694 InlineTree* inl_tree = ilt(); 4695 if (inl_tree != nullptr) { 4696 out->print(" inline %d", inl_tree->count()); 4697 inl_tree->dump_replay_data(out); 4698 } 4699 } 4700 4701 void Compile::dump_inline_data_reduced(outputStream* out) { 4702 assert(ReplayReduce, ""); 4703 4704 InlineTree* inl_tree = ilt(); 4705 if (inl_tree == nullptr) { 4706 return; 4707 } 4708 // Enable iterative replay file reduction 4709 // Output "compile" lines for depth 1 subtrees, 4710 // simulating that those trees were compiled 4711 // instead of inlined. 4712 for (int i = 0; i < inl_tree->subtrees().length(); ++i) { 4713 InlineTree* sub = inl_tree->subtrees().at(i); 4714 if (sub->inline_level() != 1) { 4715 continue; 4716 } 4717 4718 ciMethod* method = sub->method(); 4719 int entry_bci = -1; 4720 int comp_level = env()->task()->comp_level(); 4721 out->print("compile "); 4722 method->dump_name_as_ascii(out); 4723 out->print(" %d %d", entry_bci, comp_level); 4724 out->print(" inline %d", sub->count()); 4725 sub->dump_replay_data(out, -1); 4726 out->cr(); 4727 } 4728 } 4729 4730 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4731 if (n1->Opcode() < n2->Opcode()) return -1; 4732 else if (n1->Opcode() > n2->Opcode()) return 1; 4733 4734 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4735 for (uint i = 1; i < n1->req(); i++) { 4736 if (n1->in(i) < n2->in(i)) return -1; 4737 else if (n1->in(i) > n2->in(i)) return 1; 4738 } 4739 4740 return 0; 4741 } 4742 4743 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4744 Node* n1 = *n1p; 4745 Node* n2 = *n2p; 4746 4747 return cmp_expensive_nodes(n1, n2); 4748 } 4749 4750 void Compile::sort_expensive_nodes() { 4751 if (!expensive_nodes_sorted()) { 4752 _expensive_nodes.sort(cmp_expensive_nodes); 4753 } 4754 } 4755 4756 bool Compile::expensive_nodes_sorted() const { 4757 for (int i = 1; i < _expensive_nodes.length(); i++) { 4758 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) { 4759 return false; 4760 } 4761 } 4762 return true; 4763 } 4764 4765 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4766 if (_expensive_nodes.length() == 0) { 4767 return false; 4768 } 4769 4770 assert(OptimizeExpensiveOps, "optimization off?"); 4771 4772 // Take this opportunity to remove dead nodes from the list 4773 int j = 0; 4774 for (int i = 0; i < _expensive_nodes.length(); i++) { 4775 Node* n = _expensive_nodes.at(i); 4776 if (!n->is_unreachable(igvn)) { 4777 assert(n->is_expensive(), "should be expensive"); 4778 _expensive_nodes.at_put(j, n); 4779 j++; 4780 } 4781 } 4782 _expensive_nodes.trunc_to(j); 4783 4784 // Then sort the list so that similar nodes are next to each other 4785 // and check for at least two nodes of identical kind with same data 4786 // inputs. 4787 sort_expensive_nodes(); 4788 4789 for (int i = 0; i < _expensive_nodes.length()-1; i++) { 4790 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) { 4791 return true; 4792 } 4793 } 4794 4795 return false; 4796 } 4797 4798 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4799 if (_expensive_nodes.length() == 0) { 4800 return; 4801 } 4802 4803 assert(OptimizeExpensiveOps, "optimization off?"); 4804 4805 // Sort to bring similar nodes next to each other and clear the 4806 // control input of nodes for which there's only a single copy. 4807 sort_expensive_nodes(); 4808 4809 int j = 0; 4810 int identical = 0; 4811 int i = 0; 4812 bool modified = false; 4813 for (; i < _expensive_nodes.length()-1; i++) { 4814 assert(j <= i, "can't write beyond current index"); 4815 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) { 4816 identical++; 4817 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4818 continue; 4819 } 4820 if (identical > 0) { 4821 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4822 identical = 0; 4823 } else { 4824 Node* n = _expensive_nodes.at(i); 4825 igvn.replace_input_of(n, 0, nullptr); 4826 igvn.hash_insert(n); 4827 modified = true; 4828 } 4829 } 4830 if (identical > 0) { 4831 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4832 } else if (_expensive_nodes.length() >= 1) { 4833 Node* n = _expensive_nodes.at(i); 4834 igvn.replace_input_of(n, 0, nullptr); 4835 igvn.hash_insert(n); 4836 modified = true; 4837 } 4838 _expensive_nodes.trunc_to(j); 4839 if (modified) { 4840 igvn.optimize(); 4841 } 4842 } 4843 4844 void Compile::add_expensive_node(Node * n) { 4845 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list"); 4846 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4847 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4848 if (OptimizeExpensiveOps) { 4849 _expensive_nodes.append(n); 4850 } else { 4851 // Clear control input and let IGVN optimize expensive nodes if 4852 // OptimizeExpensiveOps is off. 4853 n->set_req(0, nullptr); 4854 } 4855 } 4856 4857 /** 4858 * Track coarsened Lock and Unlock nodes. 4859 */ 4860 4861 class Lock_List : public Node_List { 4862 uint _origin_cnt; 4863 public: 4864 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {} 4865 uint origin_cnt() const { return _origin_cnt; } 4866 }; 4867 4868 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) { 4869 int length = locks.length(); 4870 if (length > 0) { 4871 // Have to keep this list until locks elimination during Macro nodes elimination. 4872 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length); 4873 for (int i = 0; i < length; i++) { 4874 AbstractLockNode* lock = locks.at(i); 4875 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx); 4876 locks_list->push(lock); 4877 } 4878 _coarsened_locks.append(locks_list); 4879 } 4880 } 4881 4882 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) { 4883 int count = coarsened_count(); 4884 for (int i = 0; i < count; i++) { 4885 Node_List* locks_list = _coarsened_locks.at(i); 4886 for (uint j = 0; j < locks_list->size(); j++) { 4887 Node* lock = locks_list->at(j); 4888 assert(lock->is_AbstractLock(), "sanity"); 4889 if (!useful.member(lock)) { 4890 locks_list->yank(lock); 4891 } 4892 } 4893 } 4894 } 4895 4896 void Compile::remove_coarsened_lock(Node* n) { 4897 if (n->is_AbstractLock()) { 4898 int count = coarsened_count(); 4899 for (int i = 0; i < count; i++) { 4900 Node_List* locks_list = _coarsened_locks.at(i); 4901 locks_list->yank(n); 4902 } 4903 } 4904 } 4905 4906 bool Compile::coarsened_locks_consistent() { 4907 int count = coarsened_count(); 4908 for (int i = 0; i < count; i++) { 4909 bool unbalanced = false; 4910 bool modified = false; // track locks kind modifications 4911 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i); 4912 uint size = locks_list->size(); 4913 if (size == 0) { 4914 unbalanced = false; // All locks were eliminated - good 4915 } else if (size != locks_list->origin_cnt()) { 4916 unbalanced = true; // Some locks were removed from list 4917 } else { 4918 for (uint j = 0; j < size; j++) { 4919 Node* lock = locks_list->at(j); 4920 // All nodes in group should have the same state (modified or not) 4921 if (!lock->as_AbstractLock()->is_coarsened()) { 4922 if (j == 0) { 4923 // first on list was modified, the rest should be too for consistency 4924 modified = true; 4925 } else if (!modified) { 4926 // this lock was modified but previous locks on the list were not 4927 unbalanced = true; 4928 break; 4929 } 4930 } else if (modified) { 4931 // previous locks on list were modified but not this lock 4932 unbalanced = true; 4933 break; 4934 } 4935 } 4936 } 4937 if (unbalanced) { 4938 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified 4939 #ifdef ASSERT 4940 if (PrintEliminateLocks) { 4941 tty->print_cr("=== unbalanced coarsened locks ==="); 4942 for (uint l = 0; l < size; l++) { 4943 locks_list->at(l)->dump(); 4944 } 4945 } 4946 #endif 4947 record_failure(C2Compiler::retry_no_locks_coarsening()); 4948 return false; 4949 } 4950 } 4951 return true; 4952 } 4953 4954 /** 4955 * Remove the speculative part of types and clean up the graph 4956 */ 4957 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4958 if (UseTypeSpeculation) { 4959 Unique_Node_List worklist; 4960 worklist.push(root()); 4961 int modified = 0; 4962 // Go over all type nodes that carry a speculative type, drop the 4963 // speculative part of the type and enqueue the node for an igvn 4964 // which may optimize it out. 4965 for (uint next = 0; next < worklist.size(); ++next) { 4966 Node *n = worklist.at(next); 4967 if (n->is_Type()) { 4968 TypeNode* tn = n->as_Type(); 4969 const Type* t = tn->type(); 4970 const Type* t_no_spec = t->remove_speculative(); 4971 if (t_no_spec != t) { 4972 bool in_hash = igvn.hash_delete(n); 4973 #ifdef ASSERT 4974 if (!in_hash) { 4975 tty->print_cr("current graph:"); 4976 n->dump_bfs(MaxNodeLimit, nullptr, "S$"); 4977 tty->cr(); 4978 tty->print_cr("erroneous node:"); 4979 n->dump(); 4980 assert(false, "node should be in igvn hash table"); 4981 } 4982 #endif 4983 tn->set_type(t_no_spec); 4984 igvn.hash_insert(n); 4985 igvn._worklist.push(n); // give it a chance to go away 4986 modified++; 4987 } 4988 } 4989 // Iterate over outs - endless loops is unreachable from below 4990 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4991 Node *m = n->fast_out(i); 4992 if (not_a_node(m)) { 4993 continue; 4994 } 4995 worklist.push(m); 4996 } 4997 } 4998 // Drop the speculative part of all types in the igvn's type table 4999 igvn.remove_speculative_types(); 5000 if (modified > 0) { 5001 igvn.optimize(); 5002 if (failing()) return; 5003 } 5004 #ifdef ASSERT 5005 // Verify that after the IGVN is over no speculative type has resurfaced 5006 worklist.clear(); 5007 worklist.push(root()); 5008 for (uint next = 0; next < worklist.size(); ++next) { 5009 Node *n = worklist.at(next); 5010 const Type* t = igvn.type_or_null(n); 5011 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types"); 5012 if (n->is_Type()) { 5013 t = n->as_Type()->type(); 5014 assert(t == t->remove_speculative(), "no more speculative types"); 5015 } 5016 // Iterate over outs - endless loops is unreachable from below 5017 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 5018 Node *m = n->fast_out(i); 5019 if (not_a_node(m)) { 5020 continue; 5021 } 5022 worklist.push(m); 5023 } 5024 } 5025 igvn.check_no_speculative_types(); 5026 #endif 5027 } 5028 } 5029 5030 // Auxiliary methods to support randomized stressing/fuzzing. 5031 5032 int Compile::random() { 5033 _stress_seed = os::next_random(_stress_seed); 5034 return static_cast<int>(_stress_seed); 5035 } 5036 5037 // This method can be called the arbitrary number of times, with current count 5038 // as the argument. The logic allows selecting a single candidate from the 5039 // running list of candidates as follows: 5040 // int count = 0; 5041 // Cand* selected = null; 5042 // while(cand = cand->next()) { 5043 // if (randomized_select(++count)) { 5044 // selected = cand; 5045 // } 5046 // } 5047 // 5048 // Including count equalizes the chances any candidate is "selected". 5049 // This is useful when we don't have the complete list of candidates to choose 5050 // from uniformly. In this case, we need to adjust the randomicity of the 5051 // selection, or else we will end up biasing the selection towards the latter 5052 // candidates. 5053 // 5054 // Quick back-envelope calculation shows that for the list of n candidates 5055 // the equal probability for the candidate to persist as "best" can be 5056 // achieved by replacing it with "next" k-th candidate with the probability 5057 // of 1/k. It can be easily shown that by the end of the run, the 5058 // probability for any candidate is converged to 1/n, thus giving the 5059 // uniform distribution among all the candidates. 5060 // 5061 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 5062 #define RANDOMIZED_DOMAIN_POW 29 5063 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 5064 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 5065 bool Compile::randomized_select(int count) { 5066 assert(count > 0, "only positive"); 5067 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 5068 } 5069 5070 CloneMap& Compile::clone_map() { return _clone_map; } 5071 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 5072 5073 void NodeCloneInfo::dump_on(outputStream* st) const { 5074 st->print(" {%d:%d} ", idx(), gen()); 5075 } 5076 5077 void CloneMap::clone(Node* old, Node* nnn, int gen) { 5078 uint64_t val = value(old->_idx); 5079 NodeCloneInfo cio(val); 5080 assert(val != 0, "old node should be in the map"); 5081 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 5082 insert(nnn->_idx, cin.get()); 5083 #ifndef PRODUCT 5084 if (is_debug()) { 5085 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 5086 } 5087 #endif 5088 } 5089 5090 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 5091 NodeCloneInfo cio(value(old->_idx)); 5092 if (cio.get() == 0) { 5093 cio.set(old->_idx, 0); 5094 insert(old->_idx, cio.get()); 5095 #ifndef PRODUCT 5096 if (is_debug()) { 5097 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 5098 } 5099 #endif 5100 } 5101 clone(old, nnn, gen); 5102 } 5103 5104 int CloneMap::max_gen() const { 5105 int g = 0; 5106 DictI di(_dict); 5107 for(; di.test(); ++di) { 5108 int t = gen(di._key); 5109 if (g < t) { 5110 g = t; 5111 #ifndef PRODUCT 5112 if (is_debug()) { 5113 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 5114 } 5115 #endif 5116 } 5117 } 5118 return g; 5119 } 5120 5121 void CloneMap::dump(node_idx_t key, outputStream* st) const { 5122 uint64_t val = value(key); 5123 if (val != 0) { 5124 NodeCloneInfo ni(val); 5125 ni.dump_on(st); 5126 } 5127 } 5128 5129 // Move Allocate nodes to the start of the list 5130 void Compile::sort_macro_nodes() { 5131 int count = macro_count(); 5132 int allocates = 0; 5133 for (int i = 0; i < count; i++) { 5134 Node* n = macro_node(i); 5135 if (n->is_Allocate()) { 5136 if (i != allocates) { 5137 Node* tmp = macro_node(allocates); 5138 _macro_nodes.at_put(allocates, n); 5139 _macro_nodes.at_put(i, tmp); 5140 } 5141 allocates++; 5142 } 5143 } 5144 } 5145 5146 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) { 5147 if (failing()) { return; } 5148 EventCompilerPhase event; 5149 if (event.should_commit()) { 5150 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 5151 } 5152 #ifndef PRODUCT 5153 ResourceMark rm; 5154 stringStream ss; 5155 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt)); 5156 int iter = ++_igv_phase_iter[cpt]; 5157 if (iter > 1) { 5158 ss.print(" %d", iter); 5159 } 5160 if (n != nullptr) { 5161 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]); 5162 } 5163 5164 const char* name = ss.as_string(); 5165 if (should_print_igv(level)) { 5166 _igv_printer->print_method(name, level); 5167 } 5168 if (should_print_phase(cpt)) { 5169 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt)); 5170 } 5171 #endif 5172 C->_latest_stage_start_counter.stamp(); 5173 } 5174 5175 // Only used from CompileWrapper 5176 void Compile::begin_method() { 5177 #ifndef PRODUCT 5178 if (_method != nullptr && should_print_igv(1)) { 5179 _igv_printer->begin_method(); 5180 } 5181 #endif 5182 C->_latest_stage_start_counter.stamp(); 5183 } 5184 5185 // Only used from CompileWrapper 5186 void Compile::end_method() { 5187 EventCompilerPhase event; 5188 if (event.should_commit()) { 5189 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1); 5190 } 5191 5192 #ifndef PRODUCT 5193 if (_method != nullptr && should_print_igv(1)) { 5194 _igv_printer->end_method(); 5195 } 5196 #endif 5197 } 5198 5199 bool Compile::should_print_phase(CompilerPhaseType cpt) { 5200 #ifndef PRODUCT 5201 if (_directive->should_print_phase(cpt)) { 5202 return true; 5203 } 5204 #endif 5205 return false; 5206 } 5207 5208 bool Compile::should_print_igv(const int level) { 5209 #ifndef PRODUCT 5210 if (PrintIdealGraphLevel < 0) { // disabled by the user 5211 return false; 5212 } 5213 5214 bool need = directive()->IGVPrintLevelOption >= level; 5215 if (need && !_igv_printer) { 5216 _igv_printer = IdealGraphPrinter::printer(); 5217 _igv_printer->set_compile(this); 5218 } 5219 return need; 5220 #else 5221 return false; 5222 #endif 5223 } 5224 5225 #ifndef PRODUCT 5226 IdealGraphPrinter* Compile::_debug_file_printer = nullptr; 5227 IdealGraphPrinter* Compile::_debug_network_printer = nullptr; 5228 5229 // Called from debugger. Prints method to the default file with the default phase name. 5230 // This works regardless of any Ideal Graph Visualizer flags set or not. 5231 void igv_print() { 5232 Compile::current()->igv_print_method_to_file(); 5233 } 5234 5235 // Same as igv_print() above but with a specified phase name. 5236 void igv_print(const char* phase_name) { 5237 Compile::current()->igv_print_method_to_file(phase_name); 5238 } 5239 5240 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 5241 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 5242 // This works regardless of any Ideal Graph Visualizer flags set or not. 5243 void igv_print(bool network) { 5244 if (network) { 5245 Compile::current()->igv_print_method_to_network(); 5246 } else { 5247 Compile::current()->igv_print_method_to_file(); 5248 } 5249 } 5250 5251 // Same as igv_print(bool network) above but with a specified phase name. 5252 void igv_print(bool network, const char* phase_name) { 5253 if (network) { 5254 Compile::current()->igv_print_method_to_network(phase_name); 5255 } else { 5256 Compile::current()->igv_print_method_to_file(phase_name); 5257 } 5258 } 5259 5260 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 5261 void igv_print_default() { 5262 Compile::current()->print_method(PHASE_DEBUG, 0); 5263 } 5264 5265 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 5266 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 5267 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 5268 void igv_append() { 5269 Compile::current()->igv_print_method_to_file("Debug", true); 5270 } 5271 5272 // Same as igv_append() above but with a specified phase name. 5273 void igv_append(const char* phase_name) { 5274 Compile::current()->igv_print_method_to_file(phase_name, true); 5275 } 5276 5277 void Compile::igv_print_method_to_file(const char* phase_name, bool append) { 5278 const char* file_name = "custom_debug.xml"; 5279 if (_debug_file_printer == nullptr) { 5280 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 5281 } else { 5282 _debug_file_printer->update_compiled_method(C->method()); 5283 } 5284 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 5285 _debug_file_printer->print(phase_name, (Node*)C->root()); 5286 } 5287 5288 void Compile::igv_print_method_to_network(const char* phase_name) { 5289 if (_debug_network_printer == nullptr) { 5290 _debug_network_printer = new IdealGraphPrinter(C); 5291 } else { 5292 _debug_network_printer->update_compiled_method(C->method()); 5293 } 5294 tty->print_cr("Method printed over network stream to IGV"); 5295 _debug_network_printer->print(phase_name, (Node*)C->root()); 5296 } 5297 #endif 5298 5299 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 5300 if (type != nullptr && phase->type(value)->higher_equal(type)) { 5301 return value; 5302 } 5303 Node* result = nullptr; 5304 if (bt == T_BYTE) { 5305 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 5306 result = new RShiftINode(result, phase->intcon(24)); 5307 } else if (bt == T_BOOLEAN) { 5308 result = new AndINode(value, phase->intcon(0xFF)); 5309 } else if (bt == T_CHAR) { 5310 result = new AndINode(value,phase->intcon(0xFFFF)); 5311 } else { 5312 assert(bt == T_SHORT, "unexpected narrow type"); 5313 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 5314 result = new RShiftINode(result, phase->intcon(16)); 5315 } 5316 if (transform_res) { 5317 result = phase->transform(result); 5318 } 5319 return result; 5320 } 5321 5322 void Compile::record_method_not_compilable_oom() { 5323 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit()); 5324 }