1 /* 2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "asm/macroAssembler.hpp" 26 #include "asm/macroAssembler.inline.hpp" 27 #include "ci/ciReplay.hpp" 28 #include "classfile/javaClasses.hpp" 29 #include "code/aotCodeCache.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/compilerDefinitions.hpp" 38 #include "compiler/compilerOracle.hpp" 39 #include "compiler/disassembler.hpp" 40 #include "compiler/oopMap.hpp" 41 #include "gc/shared/barrierSet.hpp" 42 #include "gc/shared/c2/barrierSetC2.hpp" 43 #include "jfr/jfrEvents.hpp" 44 #include "jvm_io.h" 45 #include "memory/allocation.hpp" 46 #include "memory/arena.hpp" 47 #include "memory/resourceArea.hpp" 48 #include "opto/addnode.hpp" 49 #include "opto/block.hpp" 50 #include "opto/c2compiler.hpp" 51 #include "opto/callGenerator.hpp" 52 #include "opto/callnode.hpp" 53 #include "opto/castnode.hpp" 54 #include "opto/cfgnode.hpp" 55 #include "opto/chaitin.hpp" 56 #include "opto/compile.hpp" 57 #include "opto/connode.hpp" 58 #include "opto/convertnode.hpp" 59 #include "opto/divnode.hpp" 60 #include "opto/escape.hpp" 61 #include "opto/idealGraphPrinter.hpp" 62 #include "opto/locknode.hpp" 63 #include "opto/loopnode.hpp" 64 #include "opto/machnode.hpp" 65 #include "opto/macro.hpp" 66 #include "opto/matcher.hpp" 67 #include "opto/mathexactnode.hpp" 68 #include "opto/memnode.hpp" 69 #include "opto/mulnode.hpp" 70 #include "opto/narrowptrnode.hpp" 71 #include "opto/node.hpp" 72 #include "opto/opaquenode.hpp" 73 #include "opto/opcodes.hpp" 74 #include "opto/output.hpp" 75 #include "opto/parse.hpp" 76 #include "opto/phaseX.hpp" 77 #include "opto/rootnode.hpp" 78 #include "opto/runtime.hpp" 79 #include "opto/stringopts.hpp" 80 #include "opto/type.hpp" 81 #include "opto/vector.hpp" 82 #include "opto/vectornode.hpp" 83 #include "runtime/globals_extension.hpp" 84 #include "runtime/sharedRuntime.hpp" 85 #include "runtime/signature.hpp" 86 #include "runtime/stubRoutines.hpp" 87 #include "runtime/timer.hpp" 88 #include "utilities/align.hpp" 89 #include "utilities/copy.hpp" 90 #include "utilities/macros.hpp" 91 #include "utilities/resourceHash.hpp" 92 93 // -------------------- Compile::mach_constant_base_node ----------------------- 94 // Constant table base node singleton. 95 MachConstantBaseNode* Compile::mach_constant_base_node() { 96 if (_mach_constant_base_node == nullptr) { 97 _mach_constant_base_node = new MachConstantBaseNode(); 98 _mach_constant_base_node->add_req(C->root()); 99 } 100 return _mach_constant_base_node; 101 } 102 103 104 /// Support for intrinsics. 105 106 // Return the index at which m must be inserted (or already exists). 107 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 108 class IntrinsicDescPair { 109 private: 110 ciMethod* _m; 111 bool _is_virtual; 112 public: 113 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 114 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 115 ciMethod* m= elt->method(); 116 ciMethod* key_m = key->_m; 117 if (key_m < m) return -1; 118 else if (key_m > m) return 1; 119 else { 120 bool is_virtual = elt->is_virtual(); 121 bool key_virtual = key->_is_virtual; 122 if (key_virtual < is_virtual) return -1; 123 else if (key_virtual > is_virtual) return 1; 124 else return 0; 125 } 126 } 127 }; 128 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 129 #ifdef ASSERT 130 for (int i = 1; i < _intrinsics.length(); i++) { 131 CallGenerator* cg1 = _intrinsics.at(i-1); 132 CallGenerator* cg2 = _intrinsics.at(i); 133 assert(cg1->method() != cg2->method() 134 ? cg1->method() < cg2->method() 135 : cg1->is_virtual() < cg2->is_virtual(), 136 "compiler intrinsics list must stay sorted"); 137 } 138 #endif 139 IntrinsicDescPair pair(m, is_virtual); 140 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 141 } 142 143 void Compile::register_intrinsic(CallGenerator* cg) { 144 bool found = false; 145 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 146 assert(!found, "registering twice"); 147 _intrinsics.insert_before(index, cg); 148 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 149 } 150 151 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 152 assert(m->is_loaded(), "don't try this on unloaded methods"); 153 if (_intrinsics.length() > 0) { 154 bool found = false; 155 int index = intrinsic_insertion_index(m, is_virtual, found); 156 if (found) { 157 return _intrinsics.at(index); 158 } 159 } 160 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 161 if (m->intrinsic_id() != vmIntrinsics::_none && 162 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 163 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 164 if (cg != nullptr) { 165 // Save it for next time: 166 register_intrinsic(cg); 167 return cg; 168 } else { 169 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 170 } 171 } 172 return nullptr; 173 } 174 175 // Compile::make_vm_intrinsic is defined in library_call.cpp. 176 177 #ifndef PRODUCT 178 // statistics gathering... 179 180 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0}; 181 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0}; 182 183 inline int as_int(vmIntrinsics::ID id) { 184 return vmIntrinsics::as_int(id); 185 } 186 187 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 188 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 189 int oflags = _intrinsic_hist_flags[as_int(id)]; 190 assert(flags != 0, "what happened?"); 191 if (is_virtual) { 192 flags |= _intrinsic_virtual; 193 } 194 bool changed = (flags != oflags); 195 if ((flags & _intrinsic_worked) != 0) { 196 juint count = (_intrinsic_hist_count[as_int(id)] += 1); 197 if (count == 1) { 198 changed = true; // first time 199 } 200 // increment the overall count also: 201 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1; 202 } 203 if (changed) { 204 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 205 // Something changed about the intrinsic's virtuality. 206 if ((flags & _intrinsic_virtual) != 0) { 207 // This is the first use of this intrinsic as a virtual call. 208 if (oflags != 0) { 209 // We already saw it as a non-virtual, so note both cases. 210 flags |= _intrinsic_both; 211 } 212 } else if ((oflags & _intrinsic_both) == 0) { 213 // This is the first use of this intrinsic as a non-virtual 214 flags |= _intrinsic_both; 215 } 216 } 217 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags); 218 } 219 // update the overall flags also: 220 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags; 221 return changed; 222 } 223 224 static char* format_flags(int flags, char* buf) { 225 buf[0] = 0; 226 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 227 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 228 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 229 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 230 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 231 if (buf[0] == 0) strcat(buf, ","); 232 assert(buf[0] == ',', "must be"); 233 return &buf[1]; 234 } 235 236 void Compile::print_intrinsic_statistics() { 237 char flagsbuf[100]; 238 ttyLocker ttyl; 239 if (xtty != nullptr) xtty->head("statistics type='intrinsic'"); 240 tty->print_cr("Compiler intrinsic usage:"); 241 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)]; 242 if (total == 0) total = 1; // avoid div0 in case of no successes 243 #define PRINT_STAT_LINE(name, c, f) \ 244 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 245 for (auto id : EnumRange<vmIntrinsicID>{}) { 246 int flags = _intrinsic_hist_flags[as_int(id)]; 247 juint count = _intrinsic_hist_count[as_int(id)]; 248 if ((flags | count) != 0) { 249 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 250 } 251 } 252 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf)); 253 if (xtty != nullptr) xtty->tail("statistics"); 254 } 255 256 void Compile::print_statistics() { 257 { ttyLocker ttyl; 258 if (xtty != nullptr) xtty->head("statistics type='opto'"); 259 Parse::print_statistics(); 260 PhaseStringOpts::print_statistics(); 261 PhaseCCP::print_statistics(); 262 PhaseRegAlloc::print_statistics(); 263 PhaseOutput::print_statistics(); 264 PhasePeephole::print_statistics(); 265 PhaseIdealLoop::print_statistics(); 266 ConnectionGraph::print_statistics(); 267 PhaseMacroExpand::print_statistics(); 268 if (xtty != nullptr) xtty->tail("statistics"); 269 } 270 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) { 271 // put this under its own <statistics> element. 272 print_intrinsic_statistics(); 273 } 274 } 275 #endif //PRODUCT 276 277 void Compile::gvn_replace_by(Node* n, Node* nn) { 278 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 279 Node* use = n->last_out(i); 280 bool is_in_table = initial_gvn()->hash_delete(use); 281 uint uses_found = 0; 282 for (uint j = 0; j < use->len(); j++) { 283 if (use->in(j) == n) { 284 if (j < use->req()) 285 use->set_req(j, nn); 286 else 287 use->set_prec(j, nn); 288 uses_found++; 289 } 290 } 291 if (is_in_table) { 292 // reinsert into table 293 initial_gvn()->hash_find_insert(use); 294 } 295 record_for_igvn(use); 296 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist); 297 i -= uses_found; // we deleted 1 or more copies of this edge 298 } 299 } 300 301 302 // Identify all nodes that are reachable from below, useful. 303 // Use breadth-first pass that records state in a Unique_Node_List, 304 // recursive traversal is slower. 305 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 306 int estimated_worklist_size = live_nodes(); 307 useful.map( estimated_worklist_size, nullptr ); // preallocate space 308 309 // Initialize worklist 310 if (root() != nullptr) { useful.push(root()); } 311 // If 'top' is cached, declare it useful to preserve cached node 312 if (cached_top_node()) { useful.push(cached_top_node()); } 313 314 // Push all useful nodes onto the list, breadthfirst 315 for( uint next = 0; next < useful.size(); ++next ) { 316 assert( next < unique(), "Unique useful nodes < total nodes"); 317 Node *n = useful.at(next); 318 uint max = n->len(); 319 for( uint i = 0; i < max; ++i ) { 320 Node *m = n->in(i); 321 if (not_a_node(m)) continue; 322 useful.push(m); 323 } 324 } 325 } 326 327 // Update dead_node_list with any missing dead nodes using useful 328 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 329 void Compile::update_dead_node_list(Unique_Node_List &useful) { 330 uint max_idx = unique(); 331 VectorSet& useful_node_set = useful.member_set(); 332 333 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 334 // If node with index node_idx is not in useful set, 335 // mark it as dead in dead node list. 336 if (!useful_node_set.test(node_idx)) { 337 record_dead_node(node_idx); 338 } 339 } 340 } 341 342 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 343 int shift = 0; 344 for (int i = 0; i < inlines->length(); i++) { 345 CallGenerator* cg = inlines->at(i); 346 if (useful.member(cg->call_node())) { 347 if (shift > 0) { 348 inlines->at_put(i - shift, cg); 349 } 350 } else { 351 shift++; // skip over the dead element 352 } 353 } 354 if (shift > 0) { 355 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array 356 } 357 } 358 359 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) { 360 assert(dead != nullptr && dead->is_Call(), "sanity"); 361 int found = 0; 362 for (int i = 0; i < inlines->length(); i++) { 363 if (inlines->at(i)->call_node() == dead) { 364 inlines->remove_at(i); 365 found++; 366 NOT_DEBUG( break; ) // elements are unique, so exit early 367 } 368 } 369 assert(found <= 1, "not unique"); 370 } 371 372 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)> 373 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) { 374 for (int i = node_list.length() - 1; i >= 0; i--) { 375 N* node = node_list.at(i); 376 if (!useful.member(node)) { 377 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 378 } 379 } 380 } 381 382 void Compile::remove_useless_node(Node* dead) { 383 remove_modified_node(dead); 384 385 // Constant node that has no out-edges and has only one in-edge from 386 // root is usually dead. However, sometimes reshaping walk makes 387 // it reachable by adding use edges. So, we will NOT count Con nodes 388 // as dead to be conservative about the dead node count at any 389 // given time. 390 if (!dead->is_Con()) { 391 record_dead_node(dead->_idx); 392 } 393 if (dead->is_macro()) { 394 remove_macro_node(dead); 395 } 396 if (dead->is_expensive()) { 397 remove_expensive_node(dead); 398 } 399 if (dead->is_OpaqueTemplateAssertionPredicate()) { 400 remove_template_assertion_predicate_opaque(dead->as_OpaqueTemplateAssertionPredicate()); 401 } 402 if (dead->is_ParsePredicate()) { 403 remove_parse_predicate(dead->as_ParsePredicate()); 404 } 405 if (dead->for_post_loop_opts_igvn()) { 406 remove_from_post_loop_opts_igvn(dead); 407 } 408 if (dead->for_merge_stores_igvn()) { 409 remove_from_merge_stores_igvn(dead); 410 } 411 if (dead->is_Call()) { 412 remove_useless_late_inlines( &_late_inlines, dead); 413 remove_useless_late_inlines( &_string_late_inlines, dead); 414 remove_useless_late_inlines( &_boxing_late_inlines, dead); 415 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead); 416 417 if (dead->is_CallStaticJava()) { 418 remove_unstable_if_trap(dead->as_CallStaticJava(), false); 419 } 420 } 421 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 422 bs->unregister_potential_barrier_node(dead); 423 } 424 425 // Disconnect all useless nodes by disconnecting those at the boundary. 426 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist, const Unique_Node_List* root_and_safepoints) { 427 uint next = 0; 428 while (next < useful.size()) { 429 Node *n = useful.at(next++); 430 if (n->is_SafePoint()) { 431 // We're done with a parsing phase. Replaced nodes are not valid 432 // beyond that point. 433 n->as_SafePoint()->delete_replaced_nodes(); 434 } 435 // Use raw traversal of out edges since this code removes out edges 436 int max = n->outcnt(); 437 for (int j = 0; j < max; ++j) { 438 Node* child = n->raw_out(j); 439 if (!useful.member(child)) { 440 assert(!child->is_top() || child != top(), 441 "If top is cached in Compile object it is in useful list"); 442 // Only need to remove this out-edge to the useless node 443 n->raw_del_out(j); 444 --j; 445 --max; 446 if (child->is_data_proj_of_pure_function(n)) { 447 worklist.push(n); 448 } 449 } 450 } 451 if (n->outcnt() == 1 && n->has_special_unique_user()) { 452 assert(useful.member(n->unique_out()), "do not push a useless node"); 453 worklist.push(n->unique_out()); 454 } 455 } 456 457 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes 458 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes 459 // Remove useless Template Assertion Predicate opaque nodes 460 remove_useless_nodes(_template_assertion_predicate_opaques, useful); 461 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes 462 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass 463 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass 464 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps 465 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes 466 #ifdef ASSERT 467 if (_modified_nodes != nullptr) { 468 _modified_nodes->remove_useless_nodes(useful.member_set()); 469 } 470 #endif 471 472 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 473 bs->eliminate_useless_gc_barriers(useful, this); 474 // clean up the late inline lists 475 remove_useless_late_inlines( &_late_inlines, useful); 476 remove_useless_late_inlines( &_string_late_inlines, useful); 477 remove_useless_late_inlines( &_boxing_late_inlines, useful); 478 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful); 479 DEBUG_ONLY(verify_graph_edges(true /*check for no_dead_code*/, root_and_safepoints);) 480 } 481 482 // ============================================================================ 483 //------------------------------CompileWrapper--------------------------------- 484 class CompileWrapper : public StackObj { 485 Compile *const _compile; 486 public: 487 CompileWrapper(Compile* compile); 488 489 ~CompileWrapper(); 490 }; 491 492 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 493 // the Compile* pointer is stored in the current ciEnv: 494 ciEnv* env = compile->env(); 495 assert(env == ciEnv::current(), "must already be a ciEnv active"); 496 assert(env->compiler_data() == nullptr, "compile already active?"); 497 env->set_compiler_data(compile); 498 assert(compile == Compile::current(), "sanity"); 499 500 compile->set_type_dict(nullptr); 501 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 502 compile->clone_map().set_clone_idx(0); 503 compile->set_type_last_size(0); 504 compile->set_last_tf(nullptr, nullptr); 505 compile->set_indexSet_arena(nullptr); 506 compile->set_indexSet_free_block_list(nullptr); 507 compile->init_type_arena(); 508 Type::Initialize(compile); 509 _compile->begin_method(); 510 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 511 } 512 CompileWrapper::~CompileWrapper() { 513 // simulate crash during compilation 514 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned"); 515 516 _compile->end_method(); 517 _compile->env()->set_compiler_data(nullptr); 518 } 519 520 521 //----------------------------print_compile_messages--------------------------- 522 void Compile::print_compile_messages() { 523 #ifndef PRODUCT 524 // Check if recompiling 525 if (!subsume_loads() && PrintOpto) { 526 // Recompiling without allowing machine instructions to subsume loads 527 tty->print_cr("*********************************************************"); 528 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 529 tty->print_cr("*********************************************************"); 530 } 531 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) { 532 // Recompiling without escape analysis 533 tty->print_cr("*********************************************************"); 534 tty->print_cr("** Bailout: Recompile without escape analysis **"); 535 tty->print_cr("*********************************************************"); 536 } 537 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) { 538 // Recompiling without iterative escape analysis 539 tty->print_cr("*********************************************************"); 540 tty->print_cr("** Bailout: Recompile without iterative escape analysis**"); 541 tty->print_cr("*********************************************************"); 542 } 543 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) { 544 // Recompiling without reducing allocation merges 545 tty->print_cr("*********************************************************"); 546 tty->print_cr("** Bailout: Recompile without reduce allocation merges **"); 547 tty->print_cr("*********************************************************"); 548 } 549 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) { 550 // Recompiling without boxing elimination 551 tty->print_cr("*********************************************************"); 552 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 553 tty->print_cr("*********************************************************"); 554 } 555 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) { 556 // Recompiling without locks coarsening 557 tty->print_cr("*********************************************************"); 558 tty->print_cr("** Bailout: Recompile without locks coarsening **"); 559 tty->print_cr("*********************************************************"); 560 } 561 if (env()->break_at_compile()) { 562 // Open the debugger when compiling this method. 563 tty->print("### Breaking when compiling: "); 564 method()->print_short_name(); 565 tty->cr(); 566 BREAKPOINT; 567 } 568 569 if( PrintOpto ) { 570 if (is_osr_compilation()) { 571 tty->print("[OSR]%3d", _compile_id); 572 } else { 573 tty->print("%3d", _compile_id); 574 } 575 } 576 #endif 577 } 578 579 #ifndef PRODUCT 580 void Compile::print_phase(const char* phase_name) { 581 tty->print_cr("%u.\t%s", ++_phase_counter, phase_name); 582 } 583 584 void Compile::print_ideal_ir(const char* phase_name) { 585 // keep the following output all in one block 586 // This output goes directly to the tty, not the compiler log. 587 // To enable tools to match it up with the compilation activity, 588 // be sure to tag this tty output with the compile ID. 589 590 // Node dumping can cause a safepoint, which can break the tty lock. 591 // Buffer all node dumps, so that all safepoints happen before we lock. 592 ResourceMark rm; 593 stringStream ss; 594 595 if (_output == nullptr) { 596 ss.print_cr("AFTER: %s", phase_name); 597 // Print out all nodes in ascending order of index. 598 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss); 599 } else { 600 // Dump the node blockwise if we have a scheduling 601 _output->print_scheduling(&ss); 602 } 603 604 // Check that the lock is not broken by a safepoint. 605 NoSafepointVerifier nsv; 606 ttyLocker ttyl; 607 if (xtty != nullptr) { 608 xtty->head("ideal compile_id='%d'%s compile_phase='%s'", 609 compile_id(), 610 is_osr_compilation() ? " compile_kind='osr'" : "", 611 phase_name); 612 } 613 614 tty->print("%s", ss.as_string()); 615 616 if (xtty != nullptr) { 617 xtty->tail("ideal"); 618 } 619 } 620 #endif 621 622 // ============================================================================ 623 //------------------------------Compile standard------------------------------- 624 625 // Compile a method. entry_bci is -1 for normal compilations and indicates 626 // the continuation bci for on stack replacement. 627 628 629 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci, 630 Options options, DirectiveSet* directive) 631 : Phase(Compiler), 632 _compile_id(ci_env->compile_id()), 633 _options(options), 634 _method(target), 635 _entry_bci(osr_bci), 636 _ilt(nullptr), 637 _stub_function(nullptr), 638 _stub_name(nullptr), 639 _stub_id(-1), 640 _stub_entry_point(nullptr), 641 _max_node_limit(MaxNodeLimit), 642 _post_loop_opts_phase(false), 643 _merge_stores_phase(false), 644 _allow_macro_nodes(true), 645 _inlining_progress(false), 646 _inlining_incrementally(false), 647 _do_cleanup(false), 648 _has_reserved_stack_access(target->has_reserved_stack_access()), 649 #ifndef PRODUCT 650 _igv_idx(0), 651 _trace_opto_output(directive->TraceOptoOutputOption), 652 #endif 653 _has_method_handle_invokes(false), 654 _clinit_barrier_on_entry(false), 655 _stress_seed(0), 656 _comp_arena(mtCompiler, Arena::Tag::tag_comp), 657 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 658 _env(ci_env), 659 _directive(directive), 660 _log(ci_env->log()), 661 _first_failure_details(nullptr), 662 _intrinsics(comp_arena(), 0, 0, nullptr), 663 _macro_nodes(comp_arena(), 8, 0, nullptr), 664 _parse_predicates(comp_arena(), 8, 0, nullptr), 665 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr), 666 _expensive_nodes(comp_arena(), 8, 0, nullptr), 667 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr), 668 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr), 669 _unstable_if_traps(comp_arena(), 8, 0, nullptr), 670 _coarsened_locks(comp_arena(), 8, 0, nullptr), 671 _congraph(nullptr), 672 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 673 _unique(0), 674 _dead_node_count(0), 675 _dead_node_list(comp_arena()), 676 _node_arena_one(mtCompiler, Arena::Tag::tag_node), 677 _node_arena_two(mtCompiler, Arena::Tag::tag_node), 678 _node_arena(&_node_arena_one), 679 _mach_constant_base_node(nullptr), 680 _Compile_types(mtCompiler, Arena::Tag::tag_type), 681 _initial_gvn(nullptr), 682 _igvn_worklist(nullptr), 683 _types(nullptr), 684 _node_hash(nullptr), 685 _late_inlines(comp_arena(), 2, 0, nullptr), 686 _string_late_inlines(comp_arena(), 2, 0, nullptr), 687 _boxing_late_inlines(comp_arena(), 2, 0, nullptr), 688 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr), 689 _late_inlines_pos(0), 690 _number_of_mh_late_inlines(0), 691 _oom(false), 692 _replay_inline_data(nullptr), 693 _inline_printer(this), 694 _java_calls(0), 695 _inner_loops(0), 696 _interpreter_frame_size(0), 697 _output(nullptr) 698 #ifndef PRODUCT 699 , 700 _in_dump_cnt(0) 701 #endif 702 { 703 C = this; 704 CompileWrapper cw(this); 705 706 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 707 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false); 708 709 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 710 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 711 // We can always print a disassembly, either abstract (hex dump) or 712 // with the help of a suitable hsdis library. Thus, we should not 713 // couple print_assembly and print_opto_assembly controls. 714 // But: always print opto and regular assembly on compile command 'print'. 715 bool print_assembly = directive->PrintAssemblyOption; 716 set_print_assembly(print_opto_assembly || print_assembly); 717 #else 718 set_print_assembly(false); // must initialize. 719 #endif 720 721 #ifndef PRODUCT 722 set_parsed_irreducible_loop(false); 723 #endif 724 725 if (directive->ReplayInlineOption) { 726 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 727 } 728 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 729 set_print_intrinsics(directive->PrintIntrinsicsOption); 730 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 731 732 if (ProfileTraps) { 733 // Make sure the method being compiled gets its own MDO, 734 // so we can at least track the decompile_count(). 735 method()->ensure_method_data(); 736 } 737 738 if (StressLCM || StressGCM || StressIGVN || StressCCP || 739 StressIncrementalInlining || StressMacroExpansion || 740 StressMacroElimination || StressUnstableIfTraps || 741 StressBailout || StressLoopPeeling) { 742 initialize_stress_seed(directive); 743 } 744 745 Init(/*do_aliasing=*/ true); 746 747 print_compile_messages(); 748 749 _ilt = InlineTree::build_inline_tree_root(); 750 751 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 752 assert(num_alias_types() >= AliasIdxRaw, ""); 753 754 #define MINIMUM_NODE_HASH 1023 755 756 // GVN that will be run immediately on new nodes 757 uint estimated_size = method()->code_size()*4+64; 758 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 759 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 760 _types = new (comp_arena()) Type_Array(comp_arena()); 761 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size); 762 PhaseGVN gvn; 763 set_initial_gvn(&gvn); 764 765 { // Scope for timing the parser 766 TracePhase tp(_t_parser); 767 768 // Put top into the hash table ASAP. 769 initial_gvn()->transform(top()); 770 771 // Set up tf(), start(), and find a CallGenerator. 772 CallGenerator* cg = nullptr; 773 if (is_osr_compilation()) { 774 const TypeTuple *domain = StartOSRNode::osr_domain(); 775 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 776 init_tf(TypeFunc::make(domain, range)); 777 StartNode* s = new StartOSRNode(root(), domain); 778 initial_gvn()->set_type_bottom(s); 779 verify_start(s); 780 cg = CallGenerator::for_osr(method(), entry_bci()); 781 } else { 782 // Normal case. 783 init_tf(TypeFunc::make(method())); 784 StartNode* s = new StartNode(root(), tf()->domain()); 785 initial_gvn()->set_type_bottom(s); 786 verify_start(s); 787 float past_uses = method()->interpreter_invocation_count(); 788 float expected_uses = past_uses; 789 cg = CallGenerator::for_inline(method(), expected_uses); 790 } 791 if (failing()) return; 792 if (cg == nullptr) { 793 const char* reason = InlineTree::check_can_parse(method()); 794 assert(reason != nullptr, "expect reason for parse failure"); 795 stringStream ss; 796 ss.print("cannot parse method: %s", reason); 797 record_method_not_compilable(ss.as_string()); 798 return; 799 } 800 801 gvn.set_type(root(), root()->bottom_type()); 802 803 JVMState* jvms = build_start_state(start(), tf()); 804 if ((jvms = cg->generate(jvms)) == nullptr) { 805 assert(failure_reason() != nullptr, "expect reason for parse failure"); 806 stringStream ss; 807 ss.print("method parse failed: %s", failure_reason()); 808 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true)); 809 return; 810 } 811 GraphKit kit(jvms); 812 813 if (!kit.stopped()) { 814 // Accept return values, and transfer control we know not where. 815 // This is done by a special, unique ReturnNode bound to root. 816 return_values(kit.jvms()); 817 } 818 819 if (kit.has_exceptions()) { 820 // Any exceptions that escape from this call must be rethrown 821 // to whatever caller is dynamically above us on the stack. 822 // This is done by a special, unique RethrowNode bound to root. 823 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 824 } 825 826 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 827 828 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 829 inline_string_calls(true); 830 } 831 832 if (failing()) return; 833 834 // Remove clutter produced by parsing. 835 if (!failing()) { 836 ResourceMark rm; 837 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 838 } 839 } 840 841 // Note: Large methods are capped off in do_one_bytecode(). 842 if (failing()) return; 843 844 // After parsing, node notes are no longer automagic. 845 // They must be propagated by register_new_node_with_optimizer(), 846 // clone(), or the like. 847 set_default_node_notes(nullptr); 848 849 #ifndef PRODUCT 850 if (should_print_igv(1)) { 851 _igv_printer->print_inlining(); 852 } 853 #endif 854 855 if (failing()) return; 856 NOT_PRODUCT( verify_graph_edges(); ) 857 858 // Now optimize 859 Optimize(); 860 if (failing()) return; 861 NOT_PRODUCT( verify_graph_edges(); ) 862 863 #ifndef PRODUCT 864 if (should_print_ideal()) { 865 print_ideal_ir("print_ideal"); 866 } 867 #endif 868 869 #ifdef ASSERT 870 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 871 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 872 #endif 873 874 // Dump compilation data to replay it. 875 if (directive->DumpReplayOption) { 876 env()->dump_replay_data(_compile_id); 877 } 878 if (directive->DumpInlineOption && (ilt() != nullptr)) { 879 env()->dump_inline_data(_compile_id); 880 } 881 882 // Now that we know the size of all the monitors we can add a fixed slot 883 // for the original deopt pc. 884 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size); 885 set_fixed_slots(next_slot); 886 887 // Compute when to use implicit null checks. Used by matching trap based 888 // nodes and NullCheck optimization. 889 set_allowed_deopt_reasons(); 890 891 // Now generate code 892 Code_Gen(); 893 } 894 895 //------------------------------Compile---------------------------------------- 896 // Compile a runtime stub 897 Compile::Compile(ciEnv* ci_env, 898 TypeFunc_generator generator, 899 address stub_function, 900 const char* stub_name, 901 int stub_id, 902 int is_fancy_jump, 903 bool pass_tls, 904 bool return_pc, 905 DirectiveSet* directive) 906 : Phase(Compiler), 907 _compile_id(0), 908 _options(Options::for_runtime_stub()), 909 _method(nullptr), 910 _entry_bci(InvocationEntryBci), 911 _stub_function(stub_function), 912 _stub_name(stub_name), 913 _stub_id(stub_id), 914 _stub_entry_point(nullptr), 915 _max_node_limit(MaxNodeLimit), 916 _post_loop_opts_phase(false), 917 _merge_stores_phase(false), 918 _allow_macro_nodes(true), 919 _inlining_progress(false), 920 _inlining_incrementally(false), 921 _has_reserved_stack_access(false), 922 #ifndef PRODUCT 923 _igv_idx(0), 924 _trace_opto_output(directive->TraceOptoOutputOption), 925 #endif 926 _has_method_handle_invokes(false), 927 _clinit_barrier_on_entry(false), 928 _stress_seed(0), 929 _comp_arena(mtCompiler, Arena::Tag::tag_comp), 930 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 931 _env(ci_env), 932 _directive(directive), 933 _log(ci_env->log()), 934 _first_failure_details(nullptr), 935 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr), 936 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr), 937 _congraph(nullptr), 938 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 939 _unique(0), 940 _dead_node_count(0), 941 _dead_node_list(comp_arena()), 942 _node_arena_one(mtCompiler, Arena::Tag::tag_node), 943 _node_arena_two(mtCompiler, Arena::Tag::tag_node), 944 _node_arena(&_node_arena_one), 945 _mach_constant_base_node(nullptr), 946 _Compile_types(mtCompiler, Arena::Tag::tag_type), 947 _initial_gvn(nullptr), 948 _igvn_worklist(nullptr), 949 _types(nullptr), 950 _node_hash(nullptr), 951 _number_of_mh_late_inlines(0), 952 _oom(false), 953 _replay_inline_data(nullptr), 954 _inline_printer(this), 955 _java_calls(0), 956 _inner_loops(0), 957 _interpreter_frame_size(0), 958 _output(nullptr), 959 #ifndef PRODUCT 960 _in_dump_cnt(0), 961 #endif 962 _allowed_reasons(0) { 963 C = this; 964 965 // try to reuse an existing stub 966 { 967 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, _stub_id, stub_name); 968 if (blob != nullptr) { 969 RuntimeStub* rs = blob->as_runtime_stub(); 970 _stub_entry_point = rs->entry_point(); 971 return; 972 } 973 } 974 975 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false); 976 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false); 977 978 #ifndef PRODUCT 979 set_print_assembly(PrintFrameConverterAssembly); 980 set_parsed_irreducible_loop(false); 981 #else 982 set_print_assembly(false); // Must initialize. 983 #endif 984 set_has_irreducible_loop(false); // no loops 985 986 CompileWrapper cw(this); 987 Init(/*do_aliasing=*/ false); 988 init_tf((*generator)()); 989 990 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 991 _types = new (comp_arena()) Type_Array(comp_arena()); 992 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255); 993 994 if (StressLCM || StressGCM || StressBailout) { 995 initialize_stress_seed(directive); 996 } 997 998 { 999 PhaseGVN gvn; 1000 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1001 gvn.transform(top()); 1002 1003 GraphKit kit; 1004 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1005 } 1006 1007 NOT_PRODUCT( verify_graph_edges(); ) 1008 1009 Code_Gen(); 1010 } 1011 1012 Compile::~Compile() { 1013 delete _first_failure_details; 1014 }; 1015 1016 //------------------------------Init------------------------------------------- 1017 // Prepare for a single compilation 1018 void Compile::Init(bool aliasing) { 1019 _do_aliasing = aliasing; 1020 _unique = 0; 1021 _regalloc = nullptr; 1022 1023 _tf = nullptr; // filled in later 1024 _top = nullptr; // cached later 1025 _matcher = nullptr; // filled in later 1026 _cfg = nullptr; // filled in later 1027 1028 _node_note_array = nullptr; 1029 _default_node_notes = nullptr; 1030 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize() 1031 1032 _immutable_memory = nullptr; // filled in at first inquiry 1033 1034 #ifdef ASSERT 1035 _phase_optimize_finished = false; 1036 _phase_verify_ideal_loop = false; 1037 _exception_backedge = false; 1038 _type_verify = nullptr; 1039 #endif 1040 1041 // Globally visible Nodes 1042 // First set TOP to null to give safe behavior during creation of RootNode 1043 set_cached_top_node(nullptr); 1044 set_root(new RootNode()); 1045 // Now that you have a Root to point to, create the real TOP 1046 set_cached_top_node( new ConNode(Type::TOP) ); 1047 set_recent_alloc(nullptr, nullptr); 1048 1049 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1050 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1051 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1052 env()->set_dependencies(new Dependencies(env())); 1053 1054 _fixed_slots = 0; 1055 set_has_split_ifs(false); 1056 set_has_loops(false); // first approximation 1057 set_has_stringbuilder(false); 1058 set_has_boxed_value(false); 1059 _trap_can_recompile = false; // no traps emitted yet 1060 _major_progress = true; // start out assuming good things will happen 1061 set_has_unsafe_access(false); 1062 set_max_vector_size(0); 1063 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1064 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1065 set_decompile_count(0); 1066 1067 #ifndef PRODUCT 1068 _phase_counter = 0; 1069 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter)); 1070 #endif 1071 1072 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1073 _loop_opts_cnt = LoopOptsCount; 1074 set_do_inlining(Inline); 1075 set_max_inline_size(MaxInlineSize); 1076 set_freq_inline_size(FreqInlineSize); 1077 set_do_scheduling(OptoScheduling); 1078 1079 set_do_vector_loop(false); 1080 set_has_monitors(false); 1081 set_has_scoped_access(false); 1082 1083 if (AllowVectorizeOnDemand) { 1084 if (has_method() && _directive->VectorizeOption) { 1085 set_do_vector_loop(true); 1086 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());}) 1087 } else if (has_method() && method()->name() != nullptr && 1088 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1089 set_do_vector_loop(true); 1090 } 1091 } 1092 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1093 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());}) 1094 1095 _max_node_limit = _directive->MaxNodeLimitOption; 1096 1097 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1098 set_clinit_barrier_on_entry(true); 1099 } 1100 if (debug_info()->recording_non_safepoints()) { 1101 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1102 (comp_arena(), 8, 0, nullptr)); 1103 set_default_node_notes(Node_Notes::make(this)); 1104 } 1105 1106 const int grow_ats = 16; 1107 _max_alias_types = grow_ats; 1108 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1109 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1110 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1111 { 1112 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1113 } 1114 // Initialize the first few types. 1115 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr); 1116 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1117 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1118 _num_alias_types = AliasIdxRaw+1; 1119 // Zero out the alias type cache. 1120 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1121 // A null adr_type hits in the cache right away. Preload the right answer. 1122 probe_alias_cache(nullptr)->_index = AliasIdxTop; 1123 } 1124 1125 #ifdef ASSERT 1126 // Verify that the current StartNode is valid. 1127 void Compile::verify_start(StartNode* s) const { 1128 assert(failing_internal() || s == start(), "should be StartNode"); 1129 } 1130 #endif 1131 1132 /** 1133 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1134 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1135 * the ideal graph. 1136 */ 1137 StartNode* Compile::start() const { 1138 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason()); 1139 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1140 Node* start = root()->fast_out(i); 1141 if (start->is_Start()) { 1142 return start->as_Start(); 1143 } 1144 } 1145 fatal("Did not find Start node!"); 1146 return nullptr; 1147 } 1148 1149 //-------------------------------immutable_memory------------------------------------- 1150 // Access immutable memory 1151 Node* Compile::immutable_memory() { 1152 if (_immutable_memory != nullptr) { 1153 return _immutable_memory; 1154 } 1155 StartNode* s = start(); 1156 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1157 Node *p = s->fast_out(i); 1158 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1159 _immutable_memory = p; 1160 return _immutable_memory; 1161 } 1162 } 1163 ShouldNotReachHere(); 1164 return nullptr; 1165 } 1166 1167 //----------------------set_cached_top_node------------------------------------ 1168 // Install the cached top node, and make sure Node::is_top works correctly. 1169 void Compile::set_cached_top_node(Node* tn) { 1170 if (tn != nullptr) verify_top(tn); 1171 Node* old_top = _top; 1172 _top = tn; 1173 // Calling Node::setup_is_top allows the nodes the chance to adjust 1174 // their _out arrays. 1175 if (_top != nullptr) _top->setup_is_top(); 1176 if (old_top != nullptr) old_top->setup_is_top(); 1177 assert(_top == nullptr || top()->is_top(), ""); 1178 } 1179 1180 #ifdef ASSERT 1181 uint Compile::count_live_nodes_by_graph_walk() { 1182 Unique_Node_List useful(comp_arena()); 1183 // Get useful node list by walking the graph. 1184 identify_useful_nodes(useful); 1185 return useful.size(); 1186 } 1187 1188 void Compile::print_missing_nodes() { 1189 1190 // Return if CompileLog is null and PrintIdealNodeCount is false. 1191 if ((_log == nullptr) && (! PrintIdealNodeCount)) { 1192 return; 1193 } 1194 1195 // This is an expensive function. It is executed only when the user 1196 // specifies VerifyIdealNodeCount option or otherwise knows the 1197 // additional work that needs to be done to identify reachable nodes 1198 // by walking the flow graph and find the missing ones using 1199 // _dead_node_list. 1200 1201 Unique_Node_List useful(comp_arena()); 1202 // Get useful node list by walking the graph. 1203 identify_useful_nodes(useful); 1204 1205 uint l_nodes = C->live_nodes(); 1206 uint l_nodes_by_walk = useful.size(); 1207 1208 if (l_nodes != l_nodes_by_walk) { 1209 if (_log != nullptr) { 1210 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1211 _log->stamp(); 1212 _log->end_head(); 1213 } 1214 VectorSet& useful_member_set = useful.member_set(); 1215 int last_idx = l_nodes_by_walk; 1216 for (int i = 0; i < last_idx; i++) { 1217 if (useful_member_set.test(i)) { 1218 if (_dead_node_list.test(i)) { 1219 if (_log != nullptr) { 1220 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1221 } 1222 if (PrintIdealNodeCount) { 1223 // Print the log message to tty 1224 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1225 useful.at(i)->dump(); 1226 } 1227 } 1228 } 1229 else if (! _dead_node_list.test(i)) { 1230 if (_log != nullptr) { 1231 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1232 } 1233 if (PrintIdealNodeCount) { 1234 // Print the log message to tty 1235 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1236 } 1237 } 1238 } 1239 if (_log != nullptr) { 1240 _log->tail("mismatched_nodes"); 1241 } 1242 } 1243 } 1244 void Compile::record_modified_node(Node* n) { 1245 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) { 1246 _modified_nodes->push(n); 1247 } 1248 } 1249 1250 void Compile::remove_modified_node(Node* n) { 1251 if (_modified_nodes != nullptr) { 1252 _modified_nodes->remove(n); 1253 } 1254 } 1255 #endif 1256 1257 #ifndef PRODUCT 1258 void Compile::verify_top(Node* tn) const { 1259 if (tn != nullptr) { 1260 assert(tn->is_Con(), "top node must be a constant"); 1261 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1262 assert(tn->in(0) != nullptr, "must have live top node"); 1263 } 1264 } 1265 #endif 1266 1267 1268 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1269 1270 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1271 guarantee(arr != nullptr, ""); 1272 int num_blocks = arr->length(); 1273 if (grow_by < num_blocks) grow_by = num_blocks; 1274 int num_notes = grow_by * _node_notes_block_size; 1275 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1276 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1277 while (num_notes > 0) { 1278 arr->append(notes); 1279 notes += _node_notes_block_size; 1280 num_notes -= _node_notes_block_size; 1281 } 1282 assert(num_notes == 0, "exact multiple, please"); 1283 } 1284 1285 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1286 if (source == nullptr || dest == nullptr) return false; 1287 1288 if (dest->is_Con()) 1289 return false; // Do not push debug info onto constants. 1290 1291 #ifdef ASSERT 1292 // Leave a bread crumb trail pointing to the original node: 1293 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) { 1294 dest->set_debug_orig(source); 1295 } 1296 #endif 1297 1298 if (node_note_array() == nullptr) 1299 return false; // Not collecting any notes now. 1300 1301 // This is a copy onto a pre-existing node, which may already have notes. 1302 // If both nodes have notes, do not overwrite any pre-existing notes. 1303 Node_Notes* source_notes = node_notes_at(source->_idx); 1304 if (source_notes == nullptr || source_notes->is_clear()) return false; 1305 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1306 if (dest_notes == nullptr || dest_notes->is_clear()) { 1307 return set_node_notes_at(dest->_idx, source_notes); 1308 } 1309 1310 Node_Notes merged_notes = (*source_notes); 1311 // The order of operations here ensures that dest notes will win... 1312 merged_notes.update_from(dest_notes); 1313 return set_node_notes_at(dest->_idx, &merged_notes); 1314 } 1315 1316 1317 //--------------------------allow_range_check_smearing------------------------- 1318 // Gating condition for coalescing similar range checks. 1319 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1320 // single covering check that is at least as strong as any of them. 1321 // If the optimization succeeds, the simplified (strengthened) range check 1322 // will always succeed. If it fails, we will deopt, and then give up 1323 // on the optimization. 1324 bool Compile::allow_range_check_smearing() const { 1325 // If this method has already thrown a range-check, 1326 // assume it was because we already tried range smearing 1327 // and it failed. 1328 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1329 return !already_trapped; 1330 } 1331 1332 1333 //------------------------------flatten_alias_type----------------------------- 1334 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1335 assert(do_aliasing(), "Aliasing should be enabled"); 1336 int offset = tj->offset(); 1337 TypePtr::PTR ptr = tj->ptr(); 1338 1339 // Known instance (scalarizable allocation) alias only with itself. 1340 bool is_known_inst = tj->isa_oopptr() != nullptr && 1341 tj->is_oopptr()->is_known_instance(); 1342 1343 // Process weird unsafe references. 1344 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1345 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops"); 1346 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1347 tj = TypeOopPtr::BOTTOM; 1348 ptr = tj->ptr(); 1349 offset = tj->offset(); 1350 } 1351 1352 // Array pointers need some flattening 1353 const TypeAryPtr* ta = tj->isa_aryptr(); 1354 if (ta && ta->is_stable()) { 1355 // Erase stability property for alias analysis. 1356 tj = ta = ta->cast_to_stable(false); 1357 } 1358 if( ta && is_known_inst ) { 1359 if ( offset != Type::OffsetBot && 1360 offset > arrayOopDesc::length_offset_in_bytes() ) { 1361 offset = Type::OffsetBot; // Flatten constant access into array body only 1362 tj = ta = ta-> 1363 remove_speculative()-> 1364 cast_to_ptr_type(ptr)-> 1365 with_offset(offset); 1366 } 1367 } else if (ta) { 1368 // For arrays indexed by constant indices, we flatten the alias 1369 // space to include all of the array body. Only the header, klass 1370 // and array length can be accessed un-aliased. 1371 if( offset != Type::OffsetBot ) { 1372 if( ta->const_oop() ) { // MethodData* or Method* 1373 offset = Type::OffsetBot; // Flatten constant access into array body 1374 tj = ta = ta-> 1375 remove_speculative()-> 1376 cast_to_ptr_type(ptr)-> 1377 cast_to_exactness(false)-> 1378 with_offset(offset); 1379 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1380 // range is OK as-is. 1381 tj = ta = TypeAryPtr::RANGE; 1382 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1383 tj = TypeInstPtr::KLASS; // all klass loads look alike 1384 ta = TypeAryPtr::RANGE; // generic ignored junk 1385 ptr = TypePtr::BotPTR; 1386 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1387 tj = TypeInstPtr::MARK; 1388 ta = TypeAryPtr::RANGE; // generic ignored junk 1389 ptr = TypePtr::BotPTR; 1390 } else { // Random constant offset into array body 1391 offset = Type::OffsetBot; // Flatten constant access into array body 1392 tj = ta = ta-> 1393 remove_speculative()-> 1394 cast_to_ptr_type(ptr)-> 1395 cast_to_exactness(false)-> 1396 with_offset(offset); 1397 } 1398 } 1399 // Arrays of fixed size alias with arrays of unknown size. 1400 if (ta->size() != TypeInt::POS) { 1401 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1402 tj = ta = ta-> 1403 remove_speculative()-> 1404 cast_to_ptr_type(ptr)-> 1405 with_ary(tary)-> 1406 cast_to_exactness(false); 1407 } 1408 // Arrays of known objects become arrays of unknown objects. 1409 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1410 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1411 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1412 } 1413 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1414 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1415 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1416 } 1417 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1418 // cannot be distinguished by bytecode alone. 1419 if (ta->elem() == TypeInt::BOOL) { 1420 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1421 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1422 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1423 } 1424 // During the 2nd round of IterGVN, NotNull castings are removed. 1425 // Make sure the Bottom and NotNull variants alias the same. 1426 // Also, make sure exact and non-exact variants alias the same. 1427 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) { 1428 tj = ta = ta-> 1429 remove_speculative()-> 1430 cast_to_ptr_type(TypePtr::BotPTR)-> 1431 cast_to_exactness(false)-> 1432 with_offset(offset); 1433 } 1434 } 1435 1436 // Oop pointers need some flattening 1437 const TypeInstPtr *to = tj->isa_instptr(); 1438 if (to && to != TypeOopPtr::BOTTOM) { 1439 ciInstanceKlass* ik = to->instance_klass(); 1440 if( ptr == TypePtr::Constant ) { 1441 if (ik != ciEnv::current()->Class_klass() || 1442 offset < ik->layout_helper_size_in_bytes()) { 1443 // No constant oop pointers (such as Strings); they alias with 1444 // unknown strings. 1445 assert(!is_known_inst, "not scalarizable allocation"); 1446 tj = to = to-> 1447 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1448 remove_speculative()-> 1449 cast_to_ptr_type(TypePtr::BotPTR)-> 1450 cast_to_exactness(false); 1451 } 1452 } else if( is_known_inst ) { 1453 tj = to; // Keep NotNull and klass_is_exact for instance type 1454 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1455 // During the 2nd round of IterGVN, NotNull castings are removed. 1456 // Make sure the Bottom and NotNull variants alias the same. 1457 // Also, make sure exact and non-exact variants alias the same. 1458 tj = to = to-> 1459 remove_speculative()-> 1460 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1461 cast_to_ptr_type(TypePtr::BotPTR)-> 1462 cast_to_exactness(false); 1463 } 1464 if (to->speculative() != nullptr) { 1465 tj = to = to->remove_speculative(); 1466 } 1467 // Canonicalize the holder of this field 1468 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1469 // First handle header references such as a LoadKlassNode, even if the 1470 // object's klass is unloaded at compile time (4965979). 1471 if (!is_known_inst) { // Do it only for non-instance types 1472 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset); 1473 } 1474 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) { 1475 // Static fields are in the space above the normal instance 1476 // fields in the java.lang.Class instance. 1477 if (ik != ciEnv::current()->Class_klass()) { 1478 to = nullptr; 1479 tj = TypeOopPtr::BOTTOM; 1480 offset = tj->offset(); 1481 } 1482 } else { 1483 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset); 1484 assert(offset < canonical_holder->layout_helper_size_in_bytes(), ""); 1485 assert(tj->offset() == offset, "no change to offset expected"); 1486 bool xk = to->klass_is_exact(); 1487 int instance_id = to->instance_id(); 1488 1489 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder 1490 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only 1491 // its interfaces are included. 1492 if (xk && ik->equals(canonical_holder)) { 1493 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type"); 1494 } else { 1495 assert(xk || !is_known_inst, "Known instance should be exact type"); 1496 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id); 1497 } 1498 } 1499 } 1500 1501 // Klass pointers to object array klasses need some flattening 1502 const TypeKlassPtr *tk = tj->isa_klassptr(); 1503 if( tk ) { 1504 // If we are referencing a field within a Klass, we need 1505 // to assume the worst case of an Object. Both exact and 1506 // inexact types must flatten to the same alias class so 1507 // use NotNull as the PTR. 1508 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1509 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, 1510 env()->Object_klass(), 1511 offset); 1512 } 1513 1514 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) { 1515 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass()); 1516 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs 1517 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset); 1518 } else { 1519 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset); 1520 } 1521 } 1522 1523 // Check for precise loads from the primary supertype array and force them 1524 // to the supertype cache alias index. Check for generic array loads from 1525 // the primary supertype array and also force them to the supertype cache 1526 // alias index. Since the same load can reach both, we need to merge 1527 // these 2 disparate memories into the same alias class. Since the 1528 // primary supertype array is read-only, there's no chance of confusion 1529 // where we bypass an array load and an array store. 1530 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1531 if (offset == Type::OffsetBot || 1532 (offset >= primary_supers_offset && 1533 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1534 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1535 offset = in_bytes(Klass::secondary_super_cache_offset()); 1536 tj = tk = tk->with_offset(offset); 1537 } 1538 } 1539 1540 // Flatten all Raw pointers together. 1541 if (tj->base() == Type::RawPtr) 1542 tj = TypeRawPtr::BOTTOM; 1543 1544 if (tj->base() == Type::AnyPtr) 1545 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1546 1547 offset = tj->offset(); 1548 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1549 1550 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1551 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1552 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1553 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1554 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1555 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1556 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1557 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1558 assert( tj->ptr() != TypePtr::TopPTR && 1559 tj->ptr() != TypePtr::AnyNull && 1560 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1561 // assert( tj->ptr() != TypePtr::Constant || 1562 // tj->base() == Type::RawPtr || 1563 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1564 1565 return tj; 1566 } 1567 1568 void Compile::AliasType::Init(int i, const TypePtr* at) { 1569 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1570 _index = i; 1571 _adr_type = at; 1572 _field = nullptr; 1573 _element = nullptr; 1574 _is_rewritable = true; // default 1575 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr; 1576 if (atoop != nullptr && atoop->is_known_instance()) { 1577 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1578 _general_index = Compile::current()->get_alias_index(gt); 1579 } else { 1580 _general_index = 0; 1581 } 1582 } 1583 1584 BasicType Compile::AliasType::basic_type() const { 1585 if (element() != nullptr) { 1586 const Type* element = adr_type()->is_aryptr()->elem(); 1587 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1588 } if (field() != nullptr) { 1589 return field()->layout_type(); 1590 } else { 1591 return T_ILLEGAL; // unknown 1592 } 1593 } 1594 1595 //---------------------------------print_on------------------------------------ 1596 #ifndef PRODUCT 1597 void Compile::AliasType::print_on(outputStream* st) { 1598 if (index() < 10) 1599 st->print("@ <%d> ", index()); 1600 else st->print("@ <%d>", index()); 1601 st->print(is_rewritable() ? " " : " RO"); 1602 int offset = adr_type()->offset(); 1603 if (offset == Type::OffsetBot) 1604 st->print(" +any"); 1605 else st->print(" +%-3d", offset); 1606 st->print(" in "); 1607 adr_type()->dump_on(st); 1608 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1609 if (field() != nullptr && tjp) { 1610 if (tjp->is_instptr()->instance_klass() != field()->holder() || 1611 tjp->offset() != field()->offset_in_bytes()) { 1612 st->print(" != "); 1613 field()->print(); 1614 st->print(" ***"); 1615 } 1616 } 1617 } 1618 1619 void print_alias_types() { 1620 Compile* C = Compile::current(); 1621 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1622 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1623 C->alias_type(idx)->print_on(tty); 1624 tty->cr(); 1625 } 1626 } 1627 #endif 1628 1629 1630 //----------------------------probe_alias_cache-------------------------------- 1631 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1632 intptr_t key = (intptr_t) adr_type; 1633 key ^= key >> logAliasCacheSize; 1634 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1635 } 1636 1637 1638 //-----------------------------grow_alias_types-------------------------------- 1639 void Compile::grow_alias_types() { 1640 const int old_ats = _max_alias_types; // how many before? 1641 const int new_ats = old_ats; // how many more? 1642 const int grow_ats = old_ats+new_ats; // how many now? 1643 _max_alias_types = grow_ats; 1644 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1645 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1646 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1647 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1648 } 1649 1650 1651 //--------------------------------find_alias_type------------------------------ 1652 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1653 if (!do_aliasing()) { 1654 return alias_type(AliasIdxBot); 1655 } 1656 1657 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1658 if (ace->_adr_type == adr_type) { 1659 return alias_type(ace->_index); 1660 } 1661 1662 // Handle special cases. 1663 if (adr_type == nullptr) return alias_type(AliasIdxTop); 1664 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1665 1666 // Do it the slow way. 1667 const TypePtr* flat = flatten_alias_type(adr_type); 1668 1669 #ifdef ASSERT 1670 { 1671 ResourceMark rm; 1672 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1673 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1674 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1675 Type::str(adr_type)); 1676 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1677 const TypeOopPtr* foop = flat->is_oopptr(); 1678 // Scalarizable allocations have exact klass always. 1679 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1680 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1681 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1682 Type::str(foop), Type::str(xoop)); 1683 } 1684 } 1685 #endif 1686 1687 int idx = AliasIdxTop; 1688 for (int i = 0; i < num_alias_types(); i++) { 1689 if (alias_type(i)->adr_type() == flat) { 1690 idx = i; 1691 break; 1692 } 1693 } 1694 1695 if (idx == AliasIdxTop) { 1696 if (no_create) return nullptr; 1697 // Grow the array if necessary. 1698 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1699 // Add a new alias type. 1700 idx = _num_alias_types++; 1701 _alias_types[idx]->Init(idx, flat); 1702 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1703 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1704 if (flat->isa_instptr()) { 1705 if (flat->offset() == java_lang_Class::klass_offset() 1706 && flat->is_instptr()->instance_klass() == env()->Class_klass()) 1707 alias_type(idx)->set_rewritable(false); 1708 } 1709 if (flat->isa_aryptr()) { 1710 #ifdef ASSERT 1711 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1712 // (T_BYTE has the weakest alignment and size restrictions...) 1713 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1714 #endif 1715 if (flat->offset() == TypePtr::OffsetBot) { 1716 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1717 } 1718 } 1719 if (flat->isa_klassptr()) { 1720 if (UseCompactObjectHeaders) { 1721 if (flat->offset() == in_bytes(Klass::prototype_header_offset())) 1722 alias_type(idx)->set_rewritable(false); 1723 } 1724 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1725 alias_type(idx)->set_rewritable(false); 1726 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1727 alias_type(idx)->set_rewritable(false); 1728 if (flat->offset() == in_bytes(Klass::misc_flags_offset())) 1729 alias_type(idx)->set_rewritable(false); 1730 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1731 alias_type(idx)->set_rewritable(false); 1732 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1733 alias_type(idx)->set_rewritable(false); 1734 } 1735 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1736 // but the base pointer type is not distinctive enough to identify 1737 // references into JavaThread.) 1738 1739 // Check for final fields. 1740 const TypeInstPtr* tinst = flat->isa_instptr(); 1741 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1742 ciField* field; 1743 if (tinst->const_oop() != nullptr && 1744 tinst->instance_klass() == ciEnv::current()->Class_klass() && 1745 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) { 1746 // static field 1747 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1748 field = k->get_field_by_offset(tinst->offset(), true); 1749 } else { 1750 ciInstanceKlass *k = tinst->instance_klass(); 1751 field = k->get_field_by_offset(tinst->offset(), false); 1752 } 1753 assert(field == nullptr || 1754 original_field == nullptr || 1755 (field->holder() == original_field->holder() && 1756 field->offset_in_bytes() == original_field->offset_in_bytes() && 1757 field->is_static() == original_field->is_static()), "wrong field?"); 1758 // Set field() and is_rewritable() attributes. 1759 if (field != nullptr) alias_type(idx)->set_field(field); 1760 } 1761 } 1762 1763 // Fill the cache for next time. 1764 ace->_adr_type = adr_type; 1765 ace->_index = idx; 1766 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1767 1768 // Might as well try to fill the cache for the flattened version, too. 1769 AliasCacheEntry* face = probe_alias_cache(flat); 1770 if (face->_adr_type == nullptr) { 1771 face->_adr_type = flat; 1772 face->_index = idx; 1773 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1774 } 1775 1776 return alias_type(idx); 1777 } 1778 1779 1780 Compile::AliasType* Compile::alias_type(ciField* field) { 1781 const TypeOopPtr* t; 1782 if (field->is_static()) 1783 t = TypeInstPtr::make(field->holder()->java_mirror()); 1784 else 1785 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1786 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1787 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1788 return atp; 1789 } 1790 1791 1792 //------------------------------have_alias_type-------------------------------- 1793 bool Compile::have_alias_type(const TypePtr* adr_type) { 1794 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1795 if (ace->_adr_type == adr_type) { 1796 return true; 1797 } 1798 1799 // Handle special cases. 1800 if (adr_type == nullptr) return true; 1801 if (adr_type == TypePtr::BOTTOM) return true; 1802 1803 return find_alias_type(adr_type, true, nullptr) != nullptr; 1804 } 1805 1806 //-----------------------------must_alias-------------------------------------- 1807 // True if all values of the given address type are in the given alias category. 1808 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1809 if (alias_idx == AliasIdxBot) return true; // the universal category 1810 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP 1811 if (alias_idx == AliasIdxTop) return false; // the empty category 1812 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1813 1814 // the only remaining possible overlap is identity 1815 int adr_idx = get_alias_index(adr_type); 1816 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1817 assert(adr_idx == alias_idx || 1818 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1819 && adr_type != TypeOopPtr::BOTTOM), 1820 "should not be testing for overlap with an unsafe pointer"); 1821 return adr_idx == alias_idx; 1822 } 1823 1824 //------------------------------can_alias-------------------------------------- 1825 // True if any values of the given address type are in the given alias category. 1826 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1827 if (alias_idx == AliasIdxTop) return false; // the empty category 1828 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP 1829 // Known instance doesn't alias with bottom memory 1830 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1831 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1832 1833 // the only remaining possible overlap is identity 1834 int adr_idx = get_alias_index(adr_type); 1835 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1836 return adr_idx == alias_idx; 1837 } 1838 1839 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their 1840 // uncommon traps if no Runtime Predicates were created from the Parse Predicates. 1841 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) { 1842 if (parse_predicate_count() == 0) { 1843 return; 1844 } 1845 for (int i = 0; i < parse_predicate_count(); i++) { 1846 ParsePredicateNode* parse_predicate = _parse_predicates.at(i); 1847 parse_predicate->mark_useless(igvn); 1848 } 1849 _parse_predicates.clear(); 1850 } 1851 1852 void Compile::record_for_post_loop_opts_igvn(Node* n) { 1853 if (!n->for_post_loop_opts_igvn()) { 1854 assert(!_for_post_loop_igvn.contains(n), "duplicate"); 1855 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1856 _for_post_loop_igvn.append(n); 1857 } 1858 } 1859 1860 void Compile::remove_from_post_loop_opts_igvn(Node* n) { 1861 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1862 _for_post_loop_igvn.remove(n); 1863 } 1864 1865 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) { 1866 // Verify that all previous optimizations produced a valid graph 1867 // at least to this point, even if no loop optimizations were done. 1868 PhaseIdealLoop::verify(igvn); 1869 1870 if (has_loops() || _loop_opts_cnt > 0) { 1871 print_method(PHASE_AFTER_LOOP_OPTS, 2); 1872 } 1873 C->set_post_loop_opts_phase(); // no more loop opts allowed 1874 1875 assert(!C->major_progress(), "not cleared"); 1876 1877 if (_for_post_loop_igvn.length() > 0) { 1878 while (_for_post_loop_igvn.length() > 0) { 1879 Node* n = _for_post_loop_igvn.pop(); 1880 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1881 igvn._worklist.push(n); 1882 } 1883 igvn.optimize(); 1884 if (failing()) return; 1885 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed"); 1886 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now"); 1887 1888 // Sometimes IGVN sets major progress (e.g., when processing loop nodes). 1889 if (C->major_progress()) { 1890 C->clear_major_progress(); // ensure that major progress is now clear 1891 } 1892 } 1893 } 1894 1895 void Compile::record_for_merge_stores_igvn(Node* n) { 1896 if (!n->for_merge_stores_igvn()) { 1897 assert(!_for_merge_stores_igvn.contains(n), "duplicate"); 1898 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn); 1899 _for_merge_stores_igvn.append(n); 1900 } 1901 } 1902 1903 void Compile::remove_from_merge_stores_igvn(Node* n) { 1904 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn); 1905 _for_merge_stores_igvn.remove(n); 1906 } 1907 1908 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during 1909 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between 1910 // the stores, and we merge the wrong sequence of stores. 1911 // Example: 1912 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI 1913 // Apply MergeStores: 1914 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI 1915 // Remove more RangeChecks: 1916 // StoreI [ StoreL ] StoreI 1917 // But now it would have been better to do this instead: 1918 // [ StoreL ] [ StoreL ] 1919 // 1920 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round, 1921 // since we never unset _merge_stores_phase. 1922 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) { 1923 C->set_merge_stores_phase(); 1924 1925 if (_for_merge_stores_igvn.length() > 0) { 1926 while (_for_merge_stores_igvn.length() > 0) { 1927 Node* n = _for_merge_stores_igvn.pop(); 1928 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn); 1929 igvn._worklist.push(n); 1930 } 1931 igvn.optimize(); 1932 if (failing()) return; 1933 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed"); 1934 print_method(PHASE_AFTER_MERGE_STORES, 3); 1935 } 1936 } 1937 1938 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) { 1939 if (OptimizeUnstableIf) { 1940 _unstable_if_traps.append(trap); 1941 } 1942 } 1943 1944 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) { 1945 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) { 1946 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1947 Node* n = trap->uncommon_trap(); 1948 if (!useful.member(n)) { 1949 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 1950 } 1951 } 1952 } 1953 1954 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead 1955 // or fold-compares case. Return true if succeed or not found. 1956 // 1957 // In rare cases, the found trap has been processed. It is too late to delete it. Return 1958 // false and ask fold-compares to yield. 1959 // 1960 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused 1961 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path 1962 // when deoptimization does happen. 1963 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) { 1964 for (int i = 0; i < _unstable_if_traps.length(); ++i) { 1965 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1966 if (trap->uncommon_trap() == unc) { 1967 if (yield && trap->modified()) { 1968 return false; 1969 } 1970 _unstable_if_traps.delete_at(i); 1971 break; 1972 } 1973 } 1974 return true; 1975 } 1976 1977 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path. 1978 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering. 1979 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) { 1980 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) { 1981 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1982 CallStaticJavaNode* unc = trap->uncommon_trap(); 1983 int next_bci = trap->next_bci(); 1984 bool modified = trap->modified(); 1985 1986 if (next_bci != -1 && !modified) { 1987 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!"); 1988 JVMState* jvms = unc->jvms(); 1989 ciMethod* method = jvms->method(); 1990 ciBytecodeStream iter(method); 1991 1992 iter.force_bci(jvms->bci()); 1993 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if"); 1994 Bytecodes::Code c = iter.cur_bc(); 1995 Node* lhs = nullptr; 1996 Node* rhs = nullptr; 1997 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) { 1998 lhs = unc->peek_operand(0); 1999 rhs = unc->peek_operand(1); 2000 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) { 2001 lhs = unc->peek_operand(0); 2002 } 2003 2004 ResourceMark rm; 2005 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci); 2006 assert(live_locals.is_valid(), "broken liveness info"); 2007 int len = (int)live_locals.size(); 2008 2009 for (int i = 0; i < len; i++) { 2010 Node* local = unc->local(jvms, i); 2011 // kill local using the liveness of next_bci. 2012 // give up when the local looks like an operand to secure reexecution. 2013 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) { 2014 uint idx = jvms->locoff() + i; 2015 #ifdef ASSERT 2016 if (PrintOpto && Verbose) { 2017 tty->print("[unstable_if] kill local#%d: ", idx); 2018 local->dump(); 2019 tty->cr(); 2020 } 2021 #endif 2022 igvn.replace_input_of(unc, idx, top()); 2023 modified = true; 2024 } 2025 } 2026 } 2027 2028 // keep the mondified trap for late query 2029 if (modified) { 2030 trap->set_modified(); 2031 } else { 2032 _unstable_if_traps.delete_at(i); 2033 } 2034 } 2035 igvn.optimize(); 2036 } 2037 2038 // StringOpts and late inlining of string methods 2039 void Compile::inline_string_calls(bool parse_time) { 2040 { 2041 // remove useless nodes to make the usage analysis simpler 2042 ResourceMark rm; 2043 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2044 } 2045 2046 { 2047 ResourceMark rm; 2048 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2049 PhaseStringOpts pso(initial_gvn()); 2050 print_method(PHASE_AFTER_STRINGOPTS, 3); 2051 } 2052 2053 // now inline anything that we skipped the first time around 2054 if (!parse_time) { 2055 _late_inlines_pos = _late_inlines.length(); 2056 } 2057 2058 while (_string_late_inlines.length() > 0) { 2059 CallGenerator* cg = _string_late_inlines.pop(); 2060 cg->do_late_inline(); 2061 if (failing()) return; 2062 } 2063 _string_late_inlines.trunc_to(0); 2064 } 2065 2066 // Late inlining of boxing methods 2067 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2068 if (_boxing_late_inlines.length() > 0) { 2069 assert(has_boxed_value(), "inconsistent"); 2070 2071 set_inlining_incrementally(true); 2072 2073 igvn_worklist()->ensure_empty(); // should be done with igvn 2074 2075 _late_inlines_pos = _late_inlines.length(); 2076 2077 while (_boxing_late_inlines.length() > 0) { 2078 CallGenerator* cg = _boxing_late_inlines.pop(); 2079 cg->do_late_inline(); 2080 if (failing()) return; 2081 } 2082 _boxing_late_inlines.trunc_to(0); 2083 2084 inline_incrementally_cleanup(igvn); 2085 2086 set_inlining_incrementally(false); 2087 } 2088 } 2089 2090 bool Compile::inline_incrementally_one() { 2091 assert(IncrementalInline, "incremental inlining should be on"); 2092 2093 TracePhase tp(_t_incrInline_inline); 2094 2095 set_inlining_progress(false); 2096 set_do_cleanup(false); 2097 2098 for (int i = 0; i < _late_inlines.length(); i++) { 2099 _late_inlines_pos = i+1; 2100 CallGenerator* cg = _late_inlines.at(i); 2101 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node()); 2102 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline(); 2103 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call 2104 cg->do_late_inline(); 2105 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed"); 2106 if (failing()) { 2107 return false; 2108 } else if (inlining_progress()) { 2109 _late_inlines_pos = i+1; // restore the position in case new elements were inserted 2110 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node()); 2111 break; // process one call site at a time 2112 } else { 2113 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node()); 2114 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) { 2115 // Avoid potential infinite loop if node already in the IGVN list 2116 assert(false, "scheduled for IGVN during inlining attempt"); 2117 } else { 2118 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt 2119 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass"); 2120 cg->call_node()->set_generator(cg); 2121 } 2122 } 2123 } else { 2124 // Ignore late inline direct calls when inlining is not allowed. 2125 // They are left in the late inline list when node budget is exhausted until the list is fully drained. 2126 } 2127 } 2128 // Remove processed elements. 2129 _late_inlines.remove_till(_late_inlines_pos); 2130 _late_inlines_pos = 0; 2131 2132 assert(inlining_progress() || _late_inlines.length() == 0, "no progress"); 2133 2134 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2135 2136 set_inlining_progress(false); 2137 set_do_cleanup(false); 2138 2139 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption; 2140 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup; 2141 } 2142 2143 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2144 { 2145 TracePhase tp(_t_incrInline_pru); 2146 ResourceMark rm; 2147 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2148 } 2149 { 2150 TracePhase tp(_t_incrInline_igvn); 2151 igvn.reset_from_gvn(initial_gvn()); 2152 igvn.optimize(); 2153 if (failing()) return; 2154 } 2155 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3); 2156 } 2157 2158 // Perform incremental inlining until bound on number of live nodes is reached 2159 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2160 TracePhase tp(_t_incrInline); 2161 2162 set_inlining_incrementally(true); 2163 uint low_live_nodes = 0; 2164 2165 while (_late_inlines.length() > 0) { 2166 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2167 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2168 TracePhase tp(_t_incrInline_ideal); 2169 // PhaseIdealLoop is expensive so we only try it once we are 2170 // out of live nodes and we only try it again if the previous 2171 // helped got the number of nodes down significantly 2172 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2173 if (failing()) return; 2174 low_live_nodes = live_nodes(); 2175 _major_progress = true; 2176 } 2177 2178 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2179 bool do_print_inlining = print_inlining() || print_intrinsics(); 2180 if (do_print_inlining || log() != nullptr) { 2181 // Print inlining message for candidates that we couldn't inline for lack of space. 2182 for (int i = 0; i < _late_inlines.length(); i++) { 2183 CallGenerator* cg = _late_inlines.at(i); 2184 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 2185 if (do_print_inlining) { 2186 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg); 2187 } 2188 log_late_inline_failure(cg, msg); 2189 } 2190 } 2191 break; // finish 2192 } 2193 } 2194 2195 igvn_worklist()->ensure_empty(); // should be done with igvn 2196 2197 while (inline_incrementally_one()) { 2198 assert(!failing_internal() || failure_is_artificial(), "inconsistent"); 2199 } 2200 if (failing()) return; 2201 2202 inline_incrementally_cleanup(igvn); 2203 2204 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 2205 2206 if (failing()) return; 2207 2208 if (_late_inlines.length() == 0) { 2209 break; // no more progress 2210 } 2211 } 2212 2213 igvn_worklist()->ensure_empty(); // should be done with igvn 2214 2215 if (_string_late_inlines.length() > 0) { 2216 assert(has_stringbuilder(), "inconsistent"); 2217 2218 inline_string_calls(false); 2219 2220 if (failing()) return; 2221 2222 inline_incrementally_cleanup(igvn); 2223 } 2224 2225 set_inlining_incrementally(false); 2226 } 2227 2228 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) { 2229 // "inlining_incrementally() == false" is used to signal that no inlining is allowed 2230 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details). 2231 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr" 2232 // as if "inlining_incrementally() == true" were set. 2233 assert(inlining_incrementally() == false, "not allowed"); 2234 assert(_modified_nodes == nullptr, "not allowed"); 2235 assert(_late_inlines.length() > 0, "sanity"); 2236 2237 while (_late_inlines.length() > 0) { 2238 igvn_worklist()->ensure_empty(); // should be done with igvn 2239 2240 while (inline_incrementally_one()) { 2241 assert(!failing_internal() || failure_is_artificial(), "inconsistent"); 2242 } 2243 if (failing()) return; 2244 2245 inline_incrementally_cleanup(igvn); 2246 } 2247 } 2248 2249 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2250 if (_loop_opts_cnt > 0) { 2251 while (major_progress() && (_loop_opts_cnt > 0)) { 2252 TracePhase tp(_t_idealLoop); 2253 PhaseIdealLoop::optimize(igvn, mode); 2254 _loop_opts_cnt--; 2255 if (failing()) return false; 2256 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2257 } 2258 } 2259 return true; 2260 } 2261 2262 // Remove edges from "root" to each SafePoint at a backward branch. 2263 // They were inserted during parsing (see add_safepoint()) to make 2264 // infinite loops without calls or exceptions visible to root, i.e., 2265 // useful. 2266 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2267 Node *r = root(); 2268 if (r != nullptr) { 2269 for (uint i = r->req(); i < r->len(); ++i) { 2270 Node *n = r->in(i); 2271 if (n != nullptr && n->is_SafePoint()) { 2272 r->rm_prec(i); 2273 if (n->outcnt() == 0) { 2274 igvn.remove_dead_node(n); 2275 } 2276 --i; 2277 } 2278 } 2279 // Parsing may have added top inputs to the root node (Path 2280 // leading to the Halt node proven dead). Make sure we get a 2281 // chance to clean them up. 2282 igvn._worklist.push(r); 2283 igvn.optimize(); 2284 } 2285 } 2286 2287 //------------------------------Optimize--------------------------------------- 2288 // Given a graph, optimize it. 2289 void Compile::Optimize() { 2290 TracePhase tp(_t_optimizer); 2291 2292 #ifndef PRODUCT 2293 if (env()->break_at_compile()) { 2294 BREAKPOINT; 2295 } 2296 2297 #endif 2298 2299 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2300 #ifdef ASSERT 2301 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2302 #endif 2303 2304 ResourceMark rm; 2305 2306 NOT_PRODUCT( verify_graph_edges(); ) 2307 2308 print_method(PHASE_AFTER_PARSING, 1); 2309 2310 { 2311 // Iterative Global Value Numbering, including ideal transforms 2312 // Initialize IterGVN with types and values from parse-time GVN 2313 PhaseIterGVN igvn(initial_gvn()); 2314 #ifdef ASSERT 2315 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2316 #endif 2317 { 2318 TracePhase tp(_t_iterGVN); 2319 igvn.optimize(); 2320 } 2321 2322 if (failing()) return; 2323 2324 print_method(PHASE_ITER_GVN1, 2); 2325 2326 process_for_unstable_if_traps(igvn); 2327 2328 if (failing()) return; 2329 2330 inline_incrementally(igvn); 2331 2332 print_method(PHASE_INCREMENTAL_INLINE, 2); 2333 2334 if (failing()) return; 2335 2336 if (eliminate_boxing()) { 2337 // Inline valueOf() methods now. 2338 inline_boxing_calls(igvn); 2339 2340 if (failing()) return; 2341 2342 if (AlwaysIncrementalInline || StressIncrementalInlining) { 2343 inline_incrementally(igvn); 2344 } 2345 2346 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2347 2348 if (failing()) return; 2349 } 2350 2351 // Remove the speculative part of types and clean up the graph from 2352 // the extra CastPP nodes whose only purpose is to carry them. Do 2353 // that early so that optimizations are not disrupted by the extra 2354 // CastPP nodes. 2355 remove_speculative_types(igvn); 2356 2357 if (failing()) return; 2358 2359 // No more new expensive nodes will be added to the list from here 2360 // so keep only the actual candidates for optimizations. 2361 cleanup_expensive_nodes(igvn); 2362 2363 if (failing()) return; 2364 2365 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2366 if (EnableVectorSupport && has_vbox_nodes()) { 2367 TracePhase tp(_t_vector); 2368 PhaseVector pv(igvn); 2369 pv.optimize_vector_boxes(); 2370 if (failing()) return; 2371 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2372 } 2373 assert(!has_vbox_nodes(), "sanity"); 2374 2375 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2376 Compile::TracePhase tp(_t_renumberLive); 2377 igvn_worklist()->ensure_empty(); // should be done with igvn 2378 { 2379 ResourceMark rm; 2380 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist()); 2381 } 2382 igvn.reset_from_gvn(initial_gvn()); 2383 igvn.optimize(); 2384 if (failing()) return; 2385 } 2386 2387 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop 2388 // safepoints 2389 remove_root_to_sfpts_edges(igvn); 2390 2391 if (failing()) return; 2392 2393 if (has_loops()) { 2394 print_method(PHASE_BEFORE_LOOP_OPTS, 2); 2395 } 2396 2397 // Perform escape analysis 2398 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) { 2399 if (has_loops()) { 2400 // Cleanup graph (remove dead nodes). 2401 TracePhase tp(_t_idealLoop); 2402 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2403 if (failing()) return; 2404 } 2405 bool progress; 2406 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2407 do { 2408 ConnectionGraph::do_analysis(this, &igvn); 2409 2410 if (failing()) return; 2411 2412 int mcount = macro_count(); // Record number of allocations and locks before IGVN 2413 2414 // Optimize out fields loads from scalar replaceable allocations. 2415 igvn.optimize(); 2416 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2417 2418 if (failing()) return; 2419 2420 if (congraph() != nullptr && macro_count() > 0) { 2421 TracePhase tp(_t_macroEliminate); 2422 PhaseMacroExpand mexp(igvn); 2423 mexp.eliminate_macro_nodes(); 2424 if (failing()) return; 2425 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2); 2426 2427 igvn.set_delay_transform(false); 2428 igvn.optimize(); 2429 if (failing()) return; 2430 2431 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2432 } 2433 2434 ConnectionGraph::verify_ram_nodes(this, root()); 2435 if (failing()) return; 2436 2437 progress = do_iterative_escape_analysis() && 2438 (macro_count() < mcount) && 2439 ConnectionGraph::has_candidates(this); 2440 // Try again if candidates exist and made progress 2441 // by removing some allocations and/or locks. 2442 } while (progress); 2443 } 2444 2445 // Loop transforms on the ideal graph. Range Check Elimination, 2446 // peeling, unrolling, etc. 2447 2448 // Set loop opts counter 2449 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2450 { 2451 TracePhase tp(_t_idealLoop); 2452 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2453 _loop_opts_cnt--; 2454 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2455 if (failing()) return; 2456 } 2457 // Loop opts pass if partial peeling occurred in previous pass 2458 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2459 TracePhase tp(_t_idealLoop); 2460 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2461 _loop_opts_cnt--; 2462 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2463 if (failing()) return; 2464 } 2465 // Loop opts pass for loop-unrolling before CCP 2466 if(major_progress() && (_loop_opts_cnt > 0)) { 2467 TracePhase tp(_t_idealLoop); 2468 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2469 _loop_opts_cnt--; 2470 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2471 } 2472 if (!failing()) { 2473 // Verify that last round of loop opts produced a valid graph 2474 PhaseIdealLoop::verify(igvn); 2475 } 2476 } 2477 if (failing()) return; 2478 2479 // Conditional Constant Propagation; 2480 print_method(PHASE_BEFORE_CCP1, 2); 2481 PhaseCCP ccp( &igvn ); 2482 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2483 { 2484 TracePhase tp(_t_ccp); 2485 ccp.do_transform(); 2486 } 2487 print_method(PHASE_CCP1, 2); 2488 2489 assert( true, "Break here to ccp.dump_old2new_map()"); 2490 2491 // Iterative Global Value Numbering, including ideal transforms 2492 { 2493 TracePhase tp(_t_iterGVN2); 2494 igvn.reset_from_igvn(&ccp); 2495 igvn.optimize(); 2496 } 2497 print_method(PHASE_ITER_GVN2, 2); 2498 2499 if (failing()) return; 2500 2501 // Loop transforms on the ideal graph. Range Check Elimination, 2502 // peeling, unrolling, etc. 2503 if (!optimize_loops(igvn, LoopOptsDefault)) { 2504 return; 2505 } 2506 2507 if (failing()) return; 2508 2509 C->clear_major_progress(); // ensure that major progress is now clear 2510 2511 process_for_post_loop_opts_igvn(igvn); 2512 2513 process_for_merge_stores_igvn(igvn); 2514 2515 if (failing()) return; 2516 2517 #ifdef ASSERT 2518 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2519 #endif 2520 2521 { 2522 TracePhase tp(_t_macroExpand); 2523 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3); 2524 PhaseMacroExpand mex(igvn); 2525 // Do not allow new macro nodes once we start to eliminate and expand 2526 C->reset_allow_macro_nodes(); 2527 // Last attempt to eliminate macro nodes before expand 2528 mex.eliminate_macro_nodes(); 2529 if (failing()) { 2530 return; 2531 } 2532 mex.eliminate_opaque_looplimit_macro_nodes(); 2533 if (failing()) { 2534 return; 2535 } 2536 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2); 2537 if (mex.expand_macro_nodes()) { 2538 assert(failing(), "must bail out w/ explicit message"); 2539 return; 2540 } 2541 print_method(PHASE_AFTER_MACRO_EXPANSION, 2); 2542 } 2543 2544 { 2545 TracePhase tp(_t_barrierExpand); 2546 if (bs->expand_barriers(this, igvn)) { 2547 assert(failing(), "must bail out w/ explicit message"); 2548 return; 2549 } 2550 print_method(PHASE_BARRIER_EXPANSION, 2); 2551 } 2552 2553 if (C->max_vector_size() > 0) { 2554 C->optimize_logic_cones(igvn); 2555 igvn.optimize(); 2556 if (failing()) return; 2557 } 2558 2559 DEBUG_ONLY( _modified_nodes = nullptr; ) 2560 2561 assert(igvn._worklist.size() == 0, "not empty"); 2562 2563 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty"); 2564 2565 if (_late_inlines.length() > 0) { 2566 // More opportunities to optimize virtual and MH calls. 2567 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option. 2568 process_late_inline_calls_no_inline(igvn); 2569 if (failing()) return; 2570 } 2571 } // (End scope of igvn; run destructor if necessary for asserts.) 2572 2573 check_no_dead_use(); 2574 2575 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have 2576 // to remove hashes to unlock nodes for modifications. 2577 C->node_hash()->clear(); 2578 2579 // A method with only infinite loops has no edges entering loops from root 2580 { 2581 TracePhase tp(_t_graphReshaping); 2582 if (final_graph_reshaping()) { 2583 assert(failing(), "must bail out w/ explicit message"); 2584 return; 2585 } 2586 } 2587 2588 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2589 DEBUG_ONLY(set_phase_optimize_finished();) 2590 } 2591 2592 #ifdef ASSERT 2593 void Compile::check_no_dead_use() const { 2594 ResourceMark rm; 2595 Unique_Node_List wq; 2596 wq.push(root()); 2597 for (uint i = 0; i < wq.size(); ++i) { 2598 Node* n = wq.at(i); 2599 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2600 Node* u = n->fast_out(j); 2601 if (u->outcnt() == 0 && !u->is_Con()) { 2602 u->dump(); 2603 fatal("no reachable node should have no use"); 2604 } 2605 wq.push(u); 2606 } 2607 } 2608 } 2609 #endif 2610 2611 void Compile::inline_vector_reboxing_calls() { 2612 if (C->_vector_reboxing_late_inlines.length() > 0) { 2613 _late_inlines_pos = C->_late_inlines.length(); 2614 while (_vector_reboxing_late_inlines.length() > 0) { 2615 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 2616 cg->do_late_inline(); 2617 if (failing()) return; 2618 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node()); 2619 } 2620 _vector_reboxing_late_inlines.trunc_to(0); 2621 } 2622 } 2623 2624 bool Compile::has_vbox_nodes() { 2625 if (C->_vector_reboxing_late_inlines.length() > 0) { 2626 return true; 2627 } 2628 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 2629 Node * n = C->macro_node(macro_idx); 2630 assert(n->is_macro(), "only macro nodes expected here"); 2631 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 2632 return true; 2633 } 2634 } 2635 return false; 2636 } 2637 2638 //---------------------------- Bitwise operation packing optimization --------------------------- 2639 2640 static bool is_vector_unary_bitwise_op(Node* n) { 2641 return n->Opcode() == Op_XorV && 2642 VectorNode::is_vector_bitwise_not_pattern(n); 2643 } 2644 2645 static bool is_vector_binary_bitwise_op(Node* n) { 2646 switch (n->Opcode()) { 2647 case Op_AndV: 2648 case Op_OrV: 2649 return true; 2650 2651 case Op_XorV: 2652 return !is_vector_unary_bitwise_op(n); 2653 2654 default: 2655 return false; 2656 } 2657 } 2658 2659 static bool is_vector_ternary_bitwise_op(Node* n) { 2660 return n->Opcode() == Op_MacroLogicV; 2661 } 2662 2663 static bool is_vector_bitwise_op(Node* n) { 2664 return is_vector_unary_bitwise_op(n) || 2665 is_vector_binary_bitwise_op(n) || 2666 is_vector_ternary_bitwise_op(n); 2667 } 2668 2669 static bool is_vector_bitwise_cone_root(Node* n) { 2670 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) { 2671 return false; 2672 } 2673 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2674 if (is_vector_bitwise_op(n->fast_out(i))) { 2675 return false; 2676 } 2677 } 2678 return true; 2679 } 2680 2681 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) { 2682 uint cnt = 0; 2683 if (is_vector_bitwise_op(n)) { 2684 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req(); 2685 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2686 for (uint i = 1; i < inp_cnt; i++) { 2687 Node* in = n->in(i); 2688 bool skip = VectorNode::is_all_ones_vector(in); 2689 if (!skip && !inputs.member(in)) { 2690 inputs.push(in); 2691 cnt++; 2692 } 2693 } 2694 assert(cnt <= 1, "not unary"); 2695 } else { 2696 uint last_req = inp_cnt; 2697 if (is_vector_ternary_bitwise_op(n)) { 2698 last_req = inp_cnt - 1; // skip last input 2699 } 2700 for (uint i = 1; i < last_req; i++) { 2701 Node* def = n->in(i); 2702 if (!inputs.member(def)) { 2703 inputs.push(def); 2704 cnt++; 2705 } 2706 } 2707 } 2708 } else { // not a bitwise operations 2709 if (!inputs.member(n)) { 2710 inputs.push(n); 2711 cnt++; 2712 } 2713 } 2714 return cnt; 2715 } 2716 2717 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 2718 Unique_Node_List useful_nodes; 2719 C->identify_useful_nodes(useful_nodes); 2720 2721 for (uint i = 0; i < useful_nodes.size(); i++) { 2722 Node* n = useful_nodes.at(i); 2723 if (is_vector_bitwise_cone_root(n)) { 2724 list.push(n); 2725 } 2726 } 2727 } 2728 2729 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 2730 const TypeVect* vt, 2731 Unique_Node_List& partition, 2732 Unique_Node_List& inputs) { 2733 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 2734 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 2735 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 2736 2737 Node* in1 = inputs.at(0); 2738 Node* in2 = inputs.at(1); 2739 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 2740 2741 uint func = compute_truth_table(partition, inputs); 2742 2743 Node* pn = partition.at(partition.size() - 1); 2744 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2745 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt)); 2746 } 2747 2748 static uint extract_bit(uint func, uint pos) { 2749 return (func & (1 << pos)) >> pos; 2750 } 2751 2752 // 2753 // A macro logic node represents a truth table. It has 4 inputs, 2754 // First three inputs corresponds to 3 columns of a truth table 2755 // and fourth input captures the logic function. 2756 // 2757 // eg. fn = (in1 AND in2) OR in3; 2758 // 2759 // MacroNode(in1,in2,in3,fn) 2760 // 2761 // ----------------- 2762 // in1 in2 in3 fn 2763 // ----------------- 2764 // 0 0 0 0 2765 // 0 0 1 1 2766 // 0 1 0 0 2767 // 0 1 1 1 2768 // 1 0 0 0 2769 // 1 0 1 1 2770 // 1 1 0 1 2771 // 1 1 1 1 2772 // 2773 2774 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 2775 int res = 0; 2776 for (int i = 0; i < 8; i++) { 2777 int bit1 = extract_bit(in1, i); 2778 int bit2 = extract_bit(in2, i); 2779 int bit3 = extract_bit(in3, i); 2780 2781 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 2782 int func_bit = extract_bit(func, func_bit_pos); 2783 2784 res |= func_bit << i; 2785 } 2786 return res; 2787 } 2788 2789 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 2790 assert(n != nullptr, ""); 2791 assert(eval_map.contains(n), "absent"); 2792 return *(eval_map.get(n)); 2793 } 2794 2795 static void eval_operands(Node* n, 2796 uint& func1, uint& func2, uint& func3, 2797 ResourceHashtable<Node*,uint>& eval_map) { 2798 assert(is_vector_bitwise_op(n), ""); 2799 2800 if (is_vector_unary_bitwise_op(n)) { 2801 Node* opnd = n->in(1); 2802 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 2803 opnd = n->in(2); 2804 } 2805 func1 = eval_operand(opnd, eval_map); 2806 } else if (is_vector_binary_bitwise_op(n)) { 2807 func1 = eval_operand(n->in(1), eval_map); 2808 func2 = eval_operand(n->in(2), eval_map); 2809 } else { 2810 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 2811 func1 = eval_operand(n->in(1), eval_map); 2812 func2 = eval_operand(n->in(2), eval_map); 2813 func3 = eval_operand(n->in(3), eval_map); 2814 } 2815 } 2816 2817 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 2818 assert(inputs.size() <= 3, "sanity"); 2819 ResourceMark rm; 2820 uint res = 0; 2821 ResourceHashtable<Node*,uint> eval_map; 2822 2823 // Populate precomputed functions for inputs. 2824 // Each input corresponds to one column of 3 input truth-table. 2825 uint input_funcs[] = { 0xAA, // (_, _, c) -> c 2826 0xCC, // (_, b, _) -> b 2827 0xF0 }; // (a, _, _) -> a 2828 for (uint i = 0; i < inputs.size(); i++) { 2829 eval_map.put(inputs.at(i), input_funcs[2-i]); 2830 } 2831 2832 for (uint i = 0; i < partition.size(); i++) { 2833 Node* n = partition.at(i); 2834 2835 uint func1 = 0, func2 = 0, func3 = 0; 2836 eval_operands(n, func1, func2, func3, eval_map); 2837 2838 switch (n->Opcode()) { 2839 case Op_OrV: 2840 assert(func3 == 0, "not binary"); 2841 res = func1 | func2; 2842 break; 2843 case Op_AndV: 2844 assert(func3 == 0, "not binary"); 2845 res = func1 & func2; 2846 break; 2847 case Op_XorV: 2848 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2849 assert(func2 == 0 && func3 == 0, "not unary"); 2850 res = (~func1) & 0xFF; 2851 } else { 2852 assert(func3 == 0, "not binary"); 2853 res = func1 ^ func2; 2854 } 2855 break; 2856 case Op_MacroLogicV: 2857 // Ordering of inputs may change during evaluation of sub-tree 2858 // containing MacroLogic node as a child node, thus a re-evaluation 2859 // makes sure that function is evaluated in context of current 2860 // inputs. 2861 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 2862 break; 2863 2864 default: assert(false, "not supported: %s", n->Name()); 2865 } 2866 assert(res <= 0xFF, "invalid"); 2867 eval_map.put(n, res); 2868 } 2869 return res; 2870 } 2871 2872 // Criteria under which nodes gets packed into a macro logic node:- 2873 // 1) Parent and both child nodes are all unmasked or masked with 2874 // same predicates. 2875 // 2) Masked parent can be packed with left child if it is predicated 2876 // and both have same predicates. 2877 // 3) Masked parent can be packed with right child if its un-predicated 2878 // or has matching predication condition. 2879 // 4) An unmasked parent can be packed with an unmasked child. 2880 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2881 assert(partition.size() == 0, "not empty"); 2882 assert(inputs.size() == 0, "not empty"); 2883 if (is_vector_ternary_bitwise_op(n)) { 2884 return false; 2885 } 2886 2887 bool is_unary_op = is_vector_unary_bitwise_op(n); 2888 if (is_unary_op) { 2889 assert(collect_unique_inputs(n, inputs) == 1, "not unary"); 2890 return false; // too few inputs 2891 } 2892 2893 bool pack_left_child = true; 2894 bool pack_right_child = true; 2895 2896 bool left_child_LOP = is_vector_bitwise_op(n->in(1)); 2897 bool right_child_LOP = is_vector_bitwise_op(n->in(2)); 2898 2899 int left_child_input_cnt = 0; 2900 int right_child_input_cnt = 0; 2901 2902 bool parent_is_predicated = n->is_predicated_vector(); 2903 bool left_child_predicated = n->in(1)->is_predicated_vector(); 2904 bool right_child_predicated = n->in(2)->is_predicated_vector(); 2905 2906 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr; 2907 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2908 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2909 2910 do { 2911 if (pack_left_child && left_child_LOP && 2912 ((!parent_is_predicated && !left_child_predicated) || 2913 ((parent_is_predicated && left_child_predicated && 2914 parent_pred == left_child_pred)))) { 2915 partition.push(n->in(1)); 2916 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs); 2917 } else { 2918 inputs.push(n->in(1)); 2919 left_child_input_cnt = 1; 2920 } 2921 2922 if (pack_right_child && right_child_LOP && 2923 (!right_child_predicated || 2924 (right_child_predicated && parent_is_predicated && 2925 parent_pred == right_child_pred))) { 2926 partition.push(n->in(2)); 2927 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs); 2928 } else { 2929 inputs.push(n->in(2)); 2930 right_child_input_cnt = 1; 2931 } 2932 2933 if (inputs.size() > 3) { 2934 assert(partition.size() > 0, ""); 2935 inputs.clear(); 2936 partition.clear(); 2937 if (left_child_input_cnt > right_child_input_cnt) { 2938 pack_left_child = false; 2939 } else { 2940 pack_right_child = false; 2941 } 2942 } else { 2943 break; 2944 } 2945 } while(true); 2946 2947 if(partition.size()) { 2948 partition.push(n); 2949 } 2950 2951 return (partition.size() == 2 || partition.size() == 3) && 2952 (inputs.size() == 2 || inputs.size() == 3); 2953 } 2954 2955 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 2956 assert(is_vector_bitwise_op(n), "not a root"); 2957 2958 visited.set(n->_idx); 2959 2960 // 1) Do a DFS walk over the logic cone. 2961 for (uint i = 1; i < n->req(); i++) { 2962 Node* in = n->in(i); 2963 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 2964 process_logic_cone_root(igvn, in, visited); 2965 } 2966 } 2967 2968 // 2) Bottom up traversal: Merge node[s] with 2969 // the parent to form macro logic node. 2970 Unique_Node_List partition; 2971 Unique_Node_List inputs; 2972 if (compute_logic_cone(n, partition, inputs)) { 2973 const TypeVect* vt = n->bottom_type()->is_vect(); 2974 Node* pn = partition.at(partition.size() - 1); 2975 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2976 if (mask == nullptr || 2977 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) { 2978 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 2979 VectorNode::trace_new_vector(macro_logic, "MacroLogic"); 2980 igvn.replace_node(n, macro_logic); 2981 } 2982 } 2983 } 2984 2985 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 2986 ResourceMark rm; 2987 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 2988 Unique_Node_List list; 2989 collect_logic_cone_roots(list); 2990 2991 while (list.size() > 0) { 2992 Node* n = list.pop(); 2993 const TypeVect* vt = n->bottom_type()->is_vect(); 2994 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 2995 if (supported) { 2996 VectorSet visited(comp_arena()); 2997 process_logic_cone_root(igvn, n, visited); 2998 } 2999 } 3000 } 3001 } 3002 3003 //------------------------------Code_Gen--------------------------------------- 3004 // Given a graph, generate code for it 3005 void Compile::Code_Gen() { 3006 if (failing()) { 3007 return; 3008 } 3009 3010 // Perform instruction selection. You might think we could reclaim Matcher 3011 // memory PDQ, but actually the Matcher is used in generating spill code. 3012 // Internals of the Matcher (including some VectorSets) must remain live 3013 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 3014 // set a bit in reclaimed memory. 3015 3016 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 3017 // nodes. Mapping is only valid at the root of each matched subtree. 3018 NOT_PRODUCT( verify_graph_edges(); ) 3019 3020 Matcher matcher; 3021 _matcher = &matcher; 3022 { 3023 TracePhase tp(_t_matcher); 3024 matcher.match(); 3025 if (failing()) { 3026 return; 3027 } 3028 } 3029 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 3030 // nodes. Mapping is only valid at the root of each matched subtree. 3031 NOT_PRODUCT( verify_graph_edges(); ) 3032 3033 // If you have too many nodes, or if matching has failed, bail out 3034 check_node_count(0, "out of nodes matching instructions"); 3035 if (failing()) { 3036 return; 3037 } 3038 3039 print_method(PHASE_MATCHING, 2); 3040 3041 // Build a proper-looking CFG 3042 PhaseCFG cfg(node_arena(), root(), matcher); 3043 if (failing()) { 3044 return; 3045 } 3046 _cfg = &cfg; 3047 { 3048 TracePhase tp(_t_scheduler); 3049 bool success = cfg.do_global_code_motion(); 3050 if (!success) { 3051 return; 3052 } 3053 3054 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 3055 NOT_PRODUCT( verify_graph_edges(); ) 3056 cfg.verify(); 3057 if (failing()) { 3058 return; 3059 } 3060 } 3061 3062 PhaseChaitin regalloc(unique(), cfg, matcher, false); 3063 _regalloc = ®alloc; 3064 { 3065 TracePhase tp(_t_registerAllocation); 3066 // Perform register allocation. After Chaitin, use-def chains are 3067 // no longer accurate (at spill code) and so must be ignored. 3068 // Node->LRG->reg mappings are still accurate. 3069 _regalloc->Register_Allocate(); 3070 3071 // Bail out if the allocator builds too many nodes 3072 if (failing()) { 3073 return; 3074 } 3075 3076 print_method(PHASE_REGISTER_ALLOCATION, 2); 3077 } 3078 3079 // Prior to register allocation we kept empty basic blocks in case the 3080 // the allocator needed a place to spill. After register allocation we 3081 // are not adding any new instructions. If any basic block is empty, we 3082 // can now safely remove it. 3083 { 3084 TracePhase tp(_t_blockOrdering); 3085 cfg.remove_empty_blocks(); 3086 if (do_freq_based_layout()) { 3087 PhaseBlockLayout layout(cfg); 3088 } else { 3089 cfg.set_loop_alignment(); 3090 } 3091 cfg.fixup_flow(); 3092 cfg.remove_unreachable_blocks(); 3093 cfg.verify_dominator_tree(); 3094 print_method(PHASE_BLOCK_ORDERING, 3); 3095 } 3096 3097 // Apply peephole optimizations 3098 if( OptoPeephole ) { 3099 TracePhase tp(_t_peephole); 3100 PhasePeephole peep( _regalloc, cfg); 3101 peep.do_transform(); 3102 print_method(PHASE_PEEPHOLE, 3); 3103 } 3104 3105 // Do late expand if CPU requires this. 3106 if (Matcher::require_postalloc_expand) { 3107 TracePhase tp(_t_postalloc_expand); 3108 cfg.postalloc_expand(_regalloc); 3109 print_method(PHASE_POSTALLOC_EXPAND, 3); 3110 } 3111 3112 #ifdef ASSERT 3113 { 3114 CompilationMemoryStatistic::do_test_allocations(); 3115 if (failing()) return; 3116 } 3117 #endif 3118 3119 // Convert Nodes to instruction bits in a buffer 3120 { 3121 TracePhase tp(_t_output); 3122 PhaseOutput output; 3123 output.Output(); 3124 if (failing()) return; 3125 output.install(); 3126 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here 3127 } 3128 3129 // He's dead, Jim. 3130 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 3131 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 3132 } 3133 3134 //------------------------------Final_Reshape_Counts--------------------------- 3135 // This class defines counters to help identify when a method 3136 // may/must be executed using hardware with only 24-bit precision. 3137 struct Final_Reshape_Counts : public StackObj { 3138 int _call_count; // count non-inlined 'common' calls 3139 int _float_count; // count float ops requiring 24-bit precision 3140 int _double_count; // count double ops requiring more precision 3141 int _java_call_count; // count non-inlined 'java' calls 3142 int _inner_loop_count; // count loops which need alignment 3143 VectorSet _visited; // Visitation flags 3144 Node_List _tests; // Set of IfNodes & PCTableNodes 3145 3146 Final_Reshape_Counts() : 3147 _call_count(0), _float_count(0), _double_count(0), 3148 _java_call_count(0), _inner_loop_count(0) { } 3149 3150 void inc_call_count () { _call_count ++; } 3151 void inc_float_count () { _float_count ++; } 3152 void inc_double_count() { _double_count++; } 3153 void inc_java_call_count() { _java_call_count++; } 3154 void inc_inner_loop_count() { _inner_loop_count++; } 3155 3156 int get_call_count () const { return _call_count ; } 3157 int get_float_count () const { return _float_count ; } 3158 int get_double_count() const { return _double_count; } 3159 int get_java_call_count() const { return _java_call_count; } 3160 int get_inner_loop_count() const { return _inner_loop_count; } 3161 }; 3162 3163 //------------------------------final_graph_reshaping_impl---------------------- 3164 // Implement items 1-5 from final_graph_reshaping below. 3165 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3166 3167 if ( n->outcnt() == 0 ) return; // dead node 3168 uint nop = n->Opcode(); 3169 3170 // Check for 2-input instruction with "last use" on right input. 3171 // Swap to left input. Implements item (2). 3172 if( n->req() == 3 && // two-input instruction 3173 n->in(1)->outcnt() > 1 && // left use is NOT a last use 3174 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 3175 n->in(2)->outcnt() == 1 &&// right use IS a last use 3176 !n->in(2)->is_Con() ) { // right use is not a constant 3177 // Check for commutative opcode 3178 switch( nop ) { 3179 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 3180 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 3181 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 3182 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 3183 case Op_AndL: case Op_XorL: case Op_OrL: 3184 case Op_AndI: case Op_XorI: case Op_OrI: { 3185 // Move "last use" input to left by swapping inputs 3186 n->swap_edges(1, 2); 3187 break; 3188 } 3189 default: 3190 break; 3191 } 3192 } 3193 3194 #ifdef ASSERT 3195 if( n->is_Mem() ) { 3196 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 3197 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw || 3198 // oop will be recorded in oop map if load crosses safepoint 3199 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 3200 LoadNode::is_immutable_value(n->in(MemNode::Address)))), 3201 "raw memory operations should have control edge"); 3202 } 3203 if (n->is_MemBar()) { 3204 MemBarNode* mb = n->as_MemBar(); 3205 if (mb->trailing_store() || mb->trailing_load_store()) { 3206 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 3207 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 3208 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 3209 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 3210 } else if (mb->leading()) { 3211 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 3212 } 3213 } 3214 #endif 3215 // Count FPU ops and common calls, implements item (3) 3216 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes); 3217 if (!gc_handled) { 3218 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes); 3219 } 3220 3221 // Collect CFG split points 3222 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3223 frc._tests.push(n); 3224 } 3225 } 3226 3227 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) { 3228 if (!UseDivMod) { 3229 return; 3230 } 3231 3232 // Check if "a % b" and "a / b" both exist 3233 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned)); 3234 if (d == nullptr) { 3235 return; 3236 } 3237 3238 // Replace them with a fused divmod if supported 3239 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) { 3240 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned); 3241 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero. 3242 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the 3243 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency 3244 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the 3245 // DivMod node so the dependency is not lost. 3246 divmod->add_prec_from(n); 3247 divmod->add_prec_from(d); 3248 d->subsume_by(divmod->div_proj(), this); 3249 n->subsume_by(divmod->mod_proj(), this); 3250 } else { 3251 // Replace "a % b" with "a - ((a / b) * b)" 3252 Node* mult = MulNode::make(d, d->in(2), bt); 3253 Node* sub = SubNode::make(d->in(1), mult, bt); 3254 n->subsume_by(sub, this); 3255 } 3256 } 3257 3258 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) { 3259 switch( nop ) { 3260 // Count all float operations that may use FPU 3261 case Op_AddF: 3262 case Op_SubF: 3263 case Op_MulF: 3264 case Op_DivF: 3265 case Op_NegF: 3266 case Op_ModF: 3267 case Op_ConvI2F: 3268 case Op_ConF: 3269 case Op_CmpF: 3270 case Op_CmpF3: 3271 case Op_StoreF: 3272 case Op_LoadF: 3273 // case Op_ConvL2F: // longs are split into 32-bit halves 3274 frc.inc_float_count(); 3275 break; 3276 3277 case Op_ConvF2D: 3278 case Op_ConvD2F: 3279 frc.inc_float_count(); 3280 frc.inc_double_count(); 3281 break; 3282 3283 // Count all double operations that may use FPU 3284 case Op_AddD: 3285 case Op_SubD: 3286 case Op_MulD: 3287 case Op_DivD: 3288 case Op_NegD: 3289 case Op_ModD: 3290 case Op_ConvI2D: 3291 case Op_ConvD2I: 3292 // case Op_ConvL2D: // handled by leaf call 3293 // case Op_ConvD2L: // handled by leaf call 3294 case Op_ConD: 3295 case Op_CmpD: 3296 case Op_CmpD3: 3297 case Op_StoreD: 3298 case Op_LoadD: 3299 case Op_LoadD_unaligned: 3300 frc.inc_double_count(); 3301 break; 3302 case Op_Opaque1: // Remove Opaque Nodes before matching 3303 n->subsume_by(n->in(1), this); 3304 break; 3305 case Op_CallStaticJava: 3306 case Op_CallJava: 3307 case Op_CallDynamicJava: 3308 frc.inc_java_call_count(); // Count java call site; 3309 case Op_CallRuntime: 3310 case Op_CallLeaf: 3311 case Op_CallLeafVector: 3312 case Op_CallLeafNoFP: { 3313 assert (n->is_Call(), ""); 3314 CallNode *call = n->as_Call(); 3315 // Count call sites where the FP mode bit would have to be flipped. 3316 // Do not count uncommon runtime calls: 3317 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3318 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3319 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3320 frc.inc_call_count(); // Count the call site 3321 } else { // See if uncommon argument is shared 3322 Node *n = call->in(TypeFunc::Parms); 3323 int nop = n->Opcode(); 3324 // Clone shared simple arguments to uncommon calls, item (1). 3325 if (n->outcnt() > 1 && 3326 !n->is_Proj() && 3327 nop != Op_CreateEx && 3328 nop != Op_CheckCastPP && 3329 nop != Op_DecodeN && 3330 nop != Op_DecodeNKlass && 3331 !n->is_Mem() && 3332 !n->is_Phi()) { 3333 Node *x = n->clone(); 3334 call->set_req(TypeFunc::Parms, x); 3335 } 3336 } 3337 break; 3338 } 3339 case Op_StoreB: 3340 case Op_StoreC: 3341 case Op_StoreI: 3342 case Op_StoreL: 3343 case Op_CompareAndSwapB: 3344 case Op_CompareAndSwapS: 3345 case Op_CompareAndSwapI: 3346 case Op_CompareAndSwapL: 3347 case Op_CompareAndSwapP: 3348 case Op_CompareAndSwapN: 3349 case Op_WeakCompareAndSwapB: 3350 case Op_WeakCompareAndSwapS: 3351 case Op_WeakCompareAndSwapI: 3352 case Op_WeakCompareAndSwapL: 3353 case Op_WeakCompareAndSwapP: 3354 case Op_WeakCompareAndSwapN: 3355 case Op_CompareAndExchangeB: 3356 case Op_CompareAndExchangeS: 3357 case Op_CompareAndExchangeI: 3358 case Op_CompareAndExchangeL: 3359 case Op_CompareAndExchangeP: 3360 case Op_CompareAndExchangeN: 3361 case Op_GetAndAddS: 3362 case Op_GetAndAddB: 3363 case Op_GetAndAddI: 3364 case Op_GetAndAddL: 3365 case Op_GetAndSetS: 3366 case Op_GetAndSetB: 3367 case Op_GetAndSetI: 3368 case Op_GetAndSetL: 3369 case Op_GetAndSetP: 3370 case Op_GetAndSetN: 3371 case Op_StoreP: 3372 case Op_StoreN: 3373 case Op_StoreNKlass: 3374 case Op_LoadB: 3375 case Op_LoadUB: 3376 case Op_LoadUS: 3377 case Op_LoadI: 3378 case Op_LoadKlass: 3379 case Op_LoadNKlass: 3380 case Op_LoadL: 3381 case Op_LoadL_unaligned: 3382 case Op_LoadP: 3383 case Op_LoadN: 3384 case Op_LoadRange: 3385 case Op_LoadS: 3386 break; 3387 3388 case Op_AddP: { // Assert sane base pointers 3389 Node *addp = n->in(AddPNode::Address); 3390 assert( !addp->is_AddP() || 3391 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3392 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3393 "Base pointers must match (addp %u)", addp->_idx ); 3394 #ifdef _LP64 3395 if ((UseCompressedOops || UseCompressedClassPointers) && 3396 addp->Opcode() == Op_ConP && 3397 addp == n->in(AddPNode::Base) && 3398 n->in(AddPNode::Offset)->is_Con()) { 3399 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3400 // on the platform and on the compressed oops mode. 3401 // Use addressing with narrow klass to load with offset on x86. 3402 // Some platforms can use the constant pool to load ConP. 3403 // Do this transformation here since IGVN will convert ConN back to ConP. 3404 const Type* t = addp->bottom_type(); 3405 bool is_oop = t->isa_oopptr() != nullptr; 3406 bool is_klass = t->isa_klassptr() != nullptr; 3407 3408 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) || 3409 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() && 3410 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) { 3411 Node* nn = nullptr; 3412 3413 int op = is_oop ? Op_ConN : Op_ConNKlass; 3414 3415 // Look for existing ConN node of the same exact type. 3416 Node* r = root(); 3417 uint cnt = r->outcnt(); 3418 for (uint i = 0; i < cnt; i++) { 3419 Node* m = r->raw_out(i); 3420 if (m!= nullptr && m->Opcode() == op && 3421 m->bottom_type()->make_ptr() == t) { 3422 nn = m; 3423 break; 3424 } 3425 } 3426 if (nn != nullptr) { 3427 // Decode a narrow oop to match address 3428 // [R12 + narrow_oop_reg<<3 + offset] 3429 if (is_oop) { 3430 nn = new DecodeNNode(nn, t); 3431 } else { 3432 nn = new DecodeNKlassNode(nn, t); 3433 } 3434 // Check for succeeding AddP which uses the same Base. 3435 // Otherwise we will run into the assertion above when visiting that guy. 3436 for (uint i = 0; i < n->outcnt(); ++i) { 3437 Node *out_i = n->raw_out(i); 3438 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3439 out_i->set_req(AddPNode::Base, nn); 3440 #ifdef ASSERT 3441 for (uint j = 0; j < out_i->outcnt(); ++j) { 3442 Node *out_j = out_i->raw_out(j); 3443 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3444 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3445 } 3446 #endif 3447 } 3448 } 3449 n->set_req(AddPNode::Base, nn); 3450 n->set_req(AddPNode::Address, nn); 3451 if (addp->outcnt() == 0) { 3452 addp->disconnect_inputs(this); 3453 } 3454 } 3455 } 3456 } 3457 #endif 3458 break; 3459 } 3460 3461 case Op_CastPP: { 3462 // Remove CastPP nodes to gain more freedom during scheduling but 3463 // keep the dependency they encode as control or precedence edges 3464 // (if control is set already) on memory operations. Some CastPP 3465 // nodes don't have a control (don't carry a dependency): skip 3466 // those. 3467 if (n->in(0) != nullptr) { 3468 ResourceMark rm; 3469 Unique_Node_List wq; 3470 wq.push(n); 3471 for (uint next = 0; next < wq.size(); ++next) { 3472 Node *m = wq.at(next); 3473 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3474 Node* use = m->fast_out(i); 3475 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3476 use->ensure_control_or_add_prec(n->in(0)); 3477 } else { 3478 switch(use->Opcode()) { 3479 case Op_AddP: 3480 case Op_DecodeN: 3481 case Op_DecodeNKlass: 3482 case Op_CheckCastPP: 3483 case Op_CastPP: 3484 wq.push(use); 3485 break; 3486 } 3487 } 3488 } 3489 } 3490 } 3491 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3492 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3493 Node* in1 = n->in(1); 3494 const Type* t = n->bottom_type(); 3495 Node* new_in1 = in1->clone(); 3496 new_in1->as_DecodeN()->set_type(t); 3497 3498 if (!Matcher::narrow_oop_use_complex_address()) { 3499 // 3500 // x86, ARM and friends can handle 2 adds in addressing mode 3501 // and Matcher can fold a DecodeN node into address by using 3502 // a narrow oop directly and do implicit null check in address: 3503 // 3504 // [R12 + narrow_oop_reg<<3 + offset] 3505 // NullCheck narrow_oop_reg 3506 // 3507 // On other platforms (Sparc) we have to keep new DecodeN node and 3508 // use it to do implicit null check in address: 3509 // 3510 // decode_not_null narrow_oop_reg, base_reg 3511 // [base_reg + offset] 3512 // NullCheck base_reg 3513 // 3514 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3515 // to keep the information to which null check the new DecodeN node 3516 // corresponds to use it as value in implicit_null_check(). 3517 // 3518 new_in1->set_req(0, n->in(0)); 3519 } 3520 3521 n->subsume_by(new_in1, this); 3522 if (in1->outcnt() == 0) { 3523 in1->disconnect_inputs(this); 3524 } 3525 } else { 3526 n->subsume_by(n->in(1), this); 3527 if (n->outcnt() == 0) { 3528 n->disconnect_inputs(this); 3529 } 3530 } 3531 break; 3532 } 3533 case Op_CastII: { 3534 n->as_CastII()->remove_range_check_cast(this); 3535 break; 3536 } 3537 #ifdef _LP64 3538 case Op_CmpP: 3539 // Do this transformation here to preserve CmpPNode::sub() and 3540 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3541 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3542 Node* in1 = n->in(1); 3543 Node* in2 = n->in(2); 3544 if (!in1->is_DecodeNarrowPtr()) { 3545 in2 = in1; 3546 in1 = n->in(2); 3547 } 3548 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3549 3550 Node* new_in2 = nullptr; 3551 if (in2->is_DecodeNarrowPtr()) { 3552 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3553 new_in2 = in2->in(1); 3554 } else if (in2->Opcode() == Op_ConP) { 3555 const Type* t = in2->bottom_type(); 3556 if (t == TypePtr::NULL_PTR) { 3557 assert(in1->is_DecodeN(), "compare klass to null?"); 3558 // Don't convert CmpP null check into CmpN if compressed 3559 // oops implicit null check is not generated. 3560 // This will allow to generate normal oop implicit null check. 3561 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3562 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3563 // 3564 // This transformation together with CastPP transformation above 3565 // will generated code for implicit null checks for compressed oops. 3566 // 3567 // The original code after Optimize() 3568 // 3569 // LoadN memory, narrow_oop_reg 3570 // decode narrow_oop_reg, base_reg 3571 // CmpP base_reg, nullptr 3572 // CastPP base_reg // NotNull 3573 // Load [base_reg + offset], val_reg 3574 // 3575 // after these transformations will be 3576 // 3577 // LoadN memory, narrow_oop_reg 3578 // CmpN narrow_oop_reg, nullptr 3579 // decode_not_null narrow_oop_reg, base_reg 3580 // Load [base_reg + offset], val_reg 3581 // 3582 // and the uncommon path (== nullptr) will use narrow_oop_reg directly 3583 // since narrow oops can be used in debug info now (see the code in 3584 // final_graph_reshaping_walk()). 3585 // 3586 // At the end the code will be matched to 3587 // on x86: 3588 // 3589 // Load_narrow_oop memory, narrow_oop_reg 3590 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3591 // NullCheck narrow_oop_reg 3592 // 3593 // and on sparc: 3594 // 3595 // Load_narrow_oop memory, narrow_oop_reg 3596 // decode_not_null narrow_oop_reg, base_reg 3597 // Load [base_reg + offset], val_reg 3598 // NullCheck base_reg 3599 // 3600 } else if (t->isa_oopptr()) { 3601 new_in2 = ConNode::make(t->make_narrowoop()); 3602 } else if (t->isa_klassptr()) { 3603 new_in2 = ConNode::make(t->make_narrowklass()); 3604 } 3605 } 3606 if (new_in2 != nullptr) { 3607 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3608 n->subsume_by(cmpN, this); 3609 if (in1->outcnt() == 0) { 3610 in1->disconnect_inputs(this); 3611 } 3612 if (in2->outcnt() == 0) { 3613 in2->disconnect_inputs(this); 3614 } 3615 } 3616 } 3617 break; 3618 3619 case Op_DecodeN: 3620 case Op_DecodeNKlass: 3621 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3622 // DecodeN could be pinned when it can't be fold into 3623 // an address expression, see the code for Op_CastPP above. 3624 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3625 break; 3626 3627 case Op_EncodeP: 3628 case Op_EncodePKlass: { 3629 Node* in1 = n->in(1); 3630 if (in1->is_DecodeNarrowPtr()) { 3631 n->subsume_by(in1->in(1), this); 3632 } else if (in1->Opcode() == Op_ConP) { 3633 const Type* t = in1->bottom_type(); 3634 if (t == TypePtr::NULL_PTR) { 3635 assert(t->isa_oopptr(), "null klass?"); 3636 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3637 } else if (t->isa_oopptr()) { 3638 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3639 } else if (t->isa_klassptr()) { 3640 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3641 } 3642 } 3643 if (in1->outcnt() == 0) { 3644 in1->disconnect_inputs(this); 3645 } 3646 break; 3647 } 3648 3649 case Op_Proj: { 3650 if (OptimizeStringConcat || IncrementalInline) { 3651 ProjNode* proj = n->as_Proj(); 3652 if (proj->_is_io_use) { 3653 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, ""); 3654 // Separate projections were used for the exception path which 3655 // are normally removed by a late inline. If it wasn't inlined 3656 // then they will hang around and should just be replaced with 3657 // the original one. Merge them. 3658 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/); 3659 if (non_io_proj != nullptr) { 3660 proj->subsume_by(non_io_proj , this); 3661 } 3662 } 3663 } 3664 break; 3665 } 3666 3667 case Op_Phi: 3668 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3669 // The EncodeP optimization may create Phi with the same edges 3670 // for all paths. It is not handled well by Register Allocator. 3671 Node* unique_in = n->in(1); 3672 assert(unique_in != nullptr, ""); 3673 uint cnt = n->req(); 3674 for (uint i = 2; i < cnt; i++) { 3675 Node* m = n->in(i); 3676 assert(m != nullptr, ""); 3677 if (unique_in != m) 3678 unique_in = nullptr; 3679 } 3680 if (unique_in != nullptr) { 3681 n->subsume_by(unique_in, this); 3682 } 3683 } 3684 break; 3685 3686 #endif 3687 3688 case Op_ModI: 3689 handle_div_mod_op(n, T_INT, false); 3690 break; 3691 3692 case Op_ModL: 3693 handle_div_mod_op(n, T_LONG, false); 3694 break; 3695 3696 case Op_UModI: 3697 handle_div_mod_op(n, T_INT, true); 3698 break; 3699 3700 case Op_UModL: 3701 handle_div_mod_op(n, T_LONG, true); 3702 break; 3703 3704 case Op_LoadVector: 3705 case Op_StoreVector: 3706 #ifdef ASSERT 3707 // Add VerifyVectorAlignment node between adr and load / store. 3708 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) { 3709 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() : 3710 n->as_StoreVector()->must_verify_alignment(); 3711 if (must_verify_alignment) { 3712 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() : 3713 n->as_StoreVector()->memory_size(); 3714 // The memory access should be aligned to the vector width in bytes. 3715 // However, the underlying array is possibly less well aligned, but at least 3716 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in 3717 // a loop we can expect at least the following alignment: 3718 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes); 3719 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range"); 3720 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2"); 3721 // Create mask from alignment. e.g. 0b1000 -> 0b0111 3722 jlong mask = guaranteed_alignment - 1; 3723 Node* mask_con = ConLNode::make(mask); 3724 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con); 3725 n->set_req(MemNode::Address, va); 3726 } 3727 } 3728 #endif 3729 break; 3730 3731 case Op_LoadVectorGather: 3732 case Op_StoreVectorScatter: 3733 case Op_LoadVectorGatherMasked: 3734 case Op_StoreVectorScatterMasked: 3735 case Op_VectorCmpMasked: 3736 case Op_VectorMaskGen: 3737 case Op_LoadVectorMasked: 3738 case Op_StoreVectorMasked: 3739 break; 3740 3741 case Op_AddReductionVI: 3742 case Op_AddReductionVL: 3743 case Op_AddReductionVF: 3744 case Op_AddReductionVD: 3745 case Op_MulReductionVI: 3746 case Op_MulReductionVL: 3747 case Op_MulReductionVF: 3748 case Op_MulReductionVD: 3749 case Op_MinReductionV: 3750 case Op_MaxReductionV: 3751 case Op_AndReductionV: 3752 case Op_OrReductionV: 3753 case Op_XorReductionV: 3754 break; 3755 3756 case Op_PackB: 3757 case Op_PackS: 3758 case Op_PackI: 3759 case Op_PackF: 3760 case Op_PackL: 3761 case Op_PackD: 3762 if (n->req()-1 > 2) { 3763 // Replace many operand PackNodes with a binary tree for matching 3764 PackNode* p = (PackNode*) n; 3765 Node* btp = p->binary_tree_pack(1, n->req()); 3766 n->subsume_by(btp, this); 3767 } 3768 break; 3769 case Op_Loop: 3770 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop"); 3771 case Op_CountedLoop: 3772 case Op_LongCountedLoop: 3773 case Op_OuterStripMinedLoop: 3774 if (n->as_Loop()->is_inner_loop()) { 3775 frc.inc_inner_loop_count(); 3776 } 3777 n->as_Loop()->verify_strip_mined(0); 3778 break; 3779 case Op_LShiftI: 3780 case Op_RShiftI: 3781 case Op_URShiftI: 3782 case Op_LShiftL: 3783 case Op_RShiftL: 3784 case Op_URShiftL: 3785 if (Matcher::need_masked_shift_count) { 3786 // The cpu's shift instructions don't restrict the count to the 3787 // lower 5/6 bits. We need to do the masking ourselves. 3788 Node* in2 = n->in(2); 3789 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3790 const TypeInt* t = in2->find_int_type(); 3791 if (t != nullptr && t->is_con()) { 3792 juint shift = t->get_con(); 3793 if (shift > mask) { // Unsigned cmp 3794 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3795 } 3796 } else { 3797 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) { 3798 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3799 n->set_req(2, shift); 3800 } 3801 } 3802 if (in2->outcnt() == 0) { // Remove dead node 3803 in2->disconnect_inputs(this); 3804 } 3805 } 3806 break; 3807 case Op_MemBarStoreStore: 3808 case Op_MemBarRelease: 3809 // Break the link with AllocateNode: it is no longer useful and 3810 // confuses register allocation. 3811 if (n->req() > MemBarNode::Precedent) { 3812 n->set_req(MemBarNode::Precedent, top()); 3813 } 3814 break; 3815 case Op_MemBarAcquire: { 3816 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3817 // At parse time, the trailing MemBarAcquire for a volatile load 3818 // is created with an edge to the load. After optimizations, 3819 // that input may be a chain of Phis. If those phis have no 3820 // other use, then the MemBarAcquire keeps them alive and 3821 // register allocation can be confused. 3822 dead_nodes.push(n->in(MemBarNode::Precedent)); 3823 n->set_req(MemBarNode::Precedent, top()); 3824 } 3825 break; 3826 } 3827 case Op_Blackhole: 3828 break; 3829 case Op_RangeCheck: { 3830 RangeCheckNode* rc = n->as_RangeCheck(); 3831 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3832 n->subsume_by(iff, this); 3833 frc._tests.push(iff); 3834 break; 3835 } 3836 case Op_ConvI2L: { 3837 if (!Matcher::convi2l_type_required) { 3838 // Code generation on some platforms doesn't need accurate 3839 // ConvI2L types. Widening the type can help remove redundant 3840 // address computations. 3841 n->as_Type()->set_type(TypeLong::INT); 3842 ResourceMark rm; 3843 Unique_Node_List wq; 3844 wq.push(n); 3845 for (uint next = 0; next < wq.size(); next++) { 3846 Node *m = wq.at(next); 3847 3848 for(;;) { 3849 // Loop over all nodes with identical inputs edges as m 3850 Node* k = m->find_similar(m->Opcode()); 3851 if (k == nullptr) { 3852 break; 3853 } 3854 // Push their uses so we get a chance to remove node made 3855 // redundant 3856 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3857 Node* u = k->fast_out(i); 3858 if (u->Opcode() == Op_LShiftL || 3859 u->Opcode() == Op_AddL || 3860 u->Opcode() == Op_SubL || 3861 u->Opcode() == Op_AddP) { 3862 wq.push(u); 3863 } 3864 } 3865 // Replace all nodes with identical edges as m with m 3866 k->subsume_by(m, this); 3867 } 3868 } 3869 } 3870 break; 3871 } 3872 case Op_CmpUL: { 3873 if (!Matcher::has_match_rule(Op_CmpUL)) { 3874 // No support for unsigned long comparisons 3875 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3876 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3877 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3878 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3879 Node* andl = new AndLNode(orl, remove_sign_mask); 3880 Node* cmp = new CmpLNode(andl, n->in(2)); 3881 n->subsume_by(cmp, this); 3882 } 3883 break; 3884 } 3885 #ifdef ASSERT 3886 case Op_ConNKlass: { 3887 const TypePtr* tp = n->as_Type()->type()->make_ptr(); 3888 ciKlass* klass = tp->is_klassptr()->exact_klass(); 3889 assert(klass->is_in_encoding_range(), "klass cannot be compressed"); 3890 break; 3891 } 3892 #endif 3893 default: 3894 assert(!n->is_Call(), ""); 3895 assert(!n->is_Mem(), ""); 3896 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3897 break; 3898 } 3899 } 3900 3901 //------------------------------final_graph_reshaping_walk--------------------- 3902 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3903 // requires that the walk visits a node's inputs before visiting the node. 3904 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3905 Unique_Node_List sfpt; 3906 3907 frc._visited.set(root->_idx); // first, mark node as visited 3908 uint cnt = root->req(); 3909 Node *n = root; 3910 uint i = 0; 3911 while (true) { 3912 if (i < cnt) { 3913 // Place all non-visited non-null inputs onto stack 3914 Node* m = n->in(i); 3915 ++i; 3916 if (m != nullptr && !frc._visited.test_set(m->_idx)) { 3917 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) { 3918 // compute worst case interpreter size in case of a deoptimization 3919 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3920 3921 sfpt.push(m); 3922 } 3923 cnt = m->req(); 3924 nstack.push(n, i); // put on stack parent and next input's index 3925 n = m; 3926 i = 0; 3927 } 3928 } else { 3929 // Now do post-visit work 3930 final_graph_reshaping_impl(n, frc, dead_nodes); 3931 if (nstack.is_empty()) 3932 break; // finished 3933 n = nstack.node(); // Get node from stack 3934 cnt = n->req(); 3935 i = nstack.index(); 3936 nstack.pop(); // Shift to the next node on stack 3937 } 3938 } 3939 3940 // Skip next transformation if compressed oops are not used. 3941 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3942 (!UseCompressedOops && !UseCompressedClassPointers)) 3943 return; 3944 3945 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3946 // It could be done for an uncommon traps or any safepoints/calls 3947 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3948 while (sfpt.size() > 0) { 3949 n = sfpt.pop(); 3950 JVMState *jvms = n->as_SafePoint()->jvms(); 3951 assert(jvms != nullptr, "sanity"); 3952 int start = jvms->debug_start(); 3953 int end = n->req(); 3954 bool is_uncommon = (n->is_CallStaticJava() && 3955 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3956 for (int j = start; j < end; j++) { 3957 Node* in = n->in(j); 3958 if (in->is_DecodeNarrowPtr()) { 3959 bool safe_to_skip = true; 3960 if (!is_uncommon ) { 3961 // Is it safe to skip? 3962 for (uint i = 0; i < in->outcnt(); i++) { 3963 Node* u = in->raw_out(i); 3964 if (!u->is_SafePoint() || 3965 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3966 safe_to_skip = false; 3967 } 3968 } 3969 } 3970 if (safe_to_skip) { 3971 n->set_req(j, in->in(1)); 3972 } 3973 if (in->outcnt() == 0) { 3974 in->disconnect_inputs(this); 3975 } 3976 } 3977 } 3978 } 3979 } 3980 3981 //------------------------------final_graph_reshaping-------------------------- 3982 // Final Graph Reshaping. 3983 // 3984 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3985 // and not commoned up and forced early. Must come after regular 3986 // optimizations to avoid GVN undoing the cloning. Clone constant 3987 // inputs to Loop Phis; these will be split by the allocator anyways. 3988 // Remove Opaque nodes. 3989 // (2) Move last-uses by commutative operations to the left input to encourage 3990 // Intel update-in-place two-address operations and better register usage 3991 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3992 // calls canonicalizing them back. 3993 // (3) Count the number of double-precision FP ops, single-precision FP ops 3994 // and call sites. On Intel, we can get correct rounding either by 3995 // forcing singles to memory (requires extra stores and loads after each 3996 // FP bytecode) or we can set a rounding mode bit (requires setting and 3997 // clearing the mode bit around call sites). The mode bit is only used 3998 // if the relative frequency of single FP ops to calls is low enough. 3999 // This is a key transform for SPEC mpeg_audio. 4000 // (4) Detect infinite loops; blobs of code reachable from above but not 4001 // below. Several of the Code_Gen algorithms fail on such code shapes, 4002 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 4003 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 4004 // Detection is by looking for IfNodes where only 1 projection is 4005 // reachable from below or CatchNodes missing some targets. 4006 // (5) Assert for insane oop offsets in debug mode. 4007 4008 bool Compile::final_graph_reshaping() { 4009 // an infinite loop may have been eliminated by the optimizer, 4010 // in which case the graph will be empty. 4011 if (root()->req() == 1) { 4012 // Do not compile method that is only a trivial infinite loop, 4013 // since the content of the loop may have been eliminated. 4014 record_method_not_compilable("trivial infinite loop"); 4015 return true; 4016 } 4017 4018 // Expensive nodes have their control input set to prevent the GVN 4019 // from freely commoning them. There's no GVN beyond this point so 4020 // no need to keep the control input. We want the expensive nodes to 4021 // be freely moved to the least frequent code path by gcm. 4022 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 4023 for (int i = 0; i < expensive_count(); i++) { 4024 _expensive_nodes.at(i)->set_req(0, nullptr); 4025 } 4026 4027 Final_Reshape_Counts frc; 4028 4029 // Visit everybody reachable! 4030 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 4031 Node_Stack nstack(live_nodes() >> 1); 4032 Unique_Node_List dead_nodes; 4033 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes); 4034 4035 // Check for unreachable (from below) code (i.e., infinite loops). 4036 for( uint i = 0; i < frc._tests.size(); i++ ) { 4037 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 4038 // Get number of CFG targets. 4039 // Note that PCTables include exception targets after calls. 4040 uint required_outcnt = n->required_outcnt(); 4041 if (n->outcnt() != required_outcnt) { 4042 // Check for a few special cases. Rethrow Nodes never take the 4043 // 'fall-thru' path, so expected kids is 1 less. 4044 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 4045 if (n->in(0)->in(0)->is_Call()) { 4046 CallNode* call = n->in(0)->in(0)->as_Call(); 4047 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 4048 required_outcnt--; // Rethrow always has 1 less kid 4049 } else if (call->req() > TypeFunc::Parms && 4050 call->is_CallDynamicJava()) { 4051 // Check for null receiver. In such case, the optimizer has 4052 // detected that the virtual call will always result in a null 4053 // pointer exception. The fall-through projection of this CatchNode 4054 // will not be populated. 4055 Node* arg0 = call->in(TypeFunc::Parms); 4056 if (arg0->is_Type() && 4057 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 4058 required_outcnt--; 4059 } 4060 } else if (call->entry_point() == OptoRuntime::new_array_Java() || 4061 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4062 // Check for illegal array length. In such case, the optimizer has 4063 // detected that the allocation attempt will always result in an 4064 // exception. There is no fall-through projection of this CatchNode . 4065 assert(call->is_CallStaticJava(), "static call expected"); 4066 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4067 uint valid_length_test_input = call->req() - 1; 4068 Node* valid_length_test = call->in(valid_length_test_input); 4069 call->del_req(valid_length_test_input); 4070 if (valid_length_test->find_int_con(1) == 0) { 4071 required_outcnt--; 4072 } 4073 dead_nodes.push(valid_length_test); 4074 assert(n->outcnt() == required_outcnt, "malformed control flow"); 4075 continue; 4076 } 4077 } 4078 } 4079 4080 // Recheck with a better notion of 'required_outcnt' 4081 if (n->outcnt() != required_outcnt) { 4082 record_method_not_compilable("malformed control flow"); 4083 return true; // Not all targets reachable! 4084 } 4085 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) { 4086 CallNode* call = n->in(0)->in(0)->as_Call(); 4087 if (call->entry_point() == OptoRuntime::new_array_Java() || 4088 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4089 assert(call->is_CallStaticJava(), "static call expected"); 4090 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4091 uint valid_length_test_input = call->req() - 1; 4092 dead_nodes.push(call->in(valid_length_test_input)); 4093 call->del_req(valid_length_test_input); // valid length test useless now 4094 } 4095 } 4096 // Check that I actually visited all kids. Unreached kids 4097 // must be infinite loops. 4098 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 4099 if (!frc._visited.test(n->fast_out(j)->_idx)) { 4100 record_method_not_compilable("infinite loop"); 4101 return true; // Found unvisited kid; must be unreach 4102 } 4103 4104 // Here so verification code in final_graph_reshaping_walk() 4105 // always see an OuterStripMinedLoopEnd 4106 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) { 4107 IfNode* init_iff = n->as_If(); 4108 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 4109 n->subsume_by(iff, this); 4110 } 4111 } 4112 4113 while (dead_nodes.size() > 0) { 4114 Node* m = dead_nodes.pop(); 4115 if (m->outcnt() == 0 && m != top()) { 4116 for (uint j = 0; j < m->req(); j++) { 4117 Node* in = m->in(j); 4118 if (in != nullptr) { 4119 dead_nodes.push(in); 4120 } 4121 } 4122 m->disconnect_inputs(this); 4123 } 4124 } 4125 4126 set_java_calls(frc.get_java_call_count()); 4127 set_inner_loops(frc.get_inner_loop_count()); 4128 4129 // No infinite loops, no reason to bail out. 4130 return false; 4131 } 4132 4133 //-----------------------------too_many_traps---------------------------------- 4134 // Report if there are too many traps at the current method and bci. 4135 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 4136 bool Compile::too_many_traps(ciMethod* method, 4137 int bci, 4138 Deoptimization::DeoptReason reason) { 4139 ciMethodData* md = method->method_data(); 4140 if (md->is_empty()) { 4141 // Assume the trap has not occurred, or that it occurred only 4142 // because of a transient condition during start-up in the interpreter. 4143 return false; 4144 } 4145 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4146 if (md->has_trap_at(bci, m, reason) != 0) { 4147 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 4148 // Also, if there are multiple reasons, or if there is no per-BCI record, 4149 // assume the worst. 4150 if (log()) 4151 log()->elem("observe trap='%s' count='%d'", 4152 Deoptimization::trap_reason_name(reason), 4153 md->trap_count(reason)); 4154 return true; 4155 } else { 4156 // Ignore method/bci and see if there have been too many globally. 4157 return too_many_traps(reason, md); 4158 } 4159 } 4160 4161 // Less-accurate variant which does not require a method and bci. 4162 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 4163 ciMethodData* logmd) { 4164 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 4165 // Too many traps globally. 4166 // Note that we use cumulative trap_count, not just md->trap_count. 4167 if (log()) { 4168 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason); 4169 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 4170 Deoptimization::trap_reason_name(reason), 4171 mcount, trap_count(reason)); 4172 } 4173 return true; 4174 } else { 4175 // The coast is clear. 4176 return false; 4177 } 4178 } 4179 4180 //--------------------------too_many_recompiles-------------------------------- 4181 // Report if there are too many recompiles at the current method and bci. 4182 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 4183 // Is not eager to return true, since this will cause the compiler to use 4184 // Action_none for a trap point, to avoid too many recompilations. 4185 bool Compile::too_many_recompiles(ciMethod* method, 4186 int bci, 4187 Deoptimization::DeoptReason reason) { 4188 ciMethodData* md = method->method_data(); 4189 if (md->is_empty()) { 4190 // Assume the trap has not occurred, or that it occurred only 4191 // because of a transient condition during start-up in the interpreter. 4192 return false; 4193 } 4194 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 4195 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 4196 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 4197 Deoptimization::DeoptReason per_bc_reason 4198 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 4199 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4200 if ((per_bc_reason == Deoptimization::Reason_none 4201 || md->has_trap_at(bci, m, reason) != 0) 4202 // The trap frequency measure we care about is the recompile count: 4203 && md->trap_recompiled_at(bci, m) 4204 && md->overflow_recompile_count() >= bc_cutoff) { 4205 // Do not emit a trap here if it has already caused recompilations. 4206 // Also, if there are multiple reasons, or if there is no per-BCI record, 4207 // assume the worst. 4208 if (log()) 4209 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 4210 Deoptimization::trap_reason_name(reason), 4211 md->trap_count(reason), 4212 md->overflow_recompile_count()); 4213 return true; 4214 } else if (trap_count(reason) != 0 4215 && decompile_count() >= m_cutoff) { 4216 // Too many recompiles globally, and we have seen this sort of trap. 4217 // Use cumulative decompile_count, not just md->decompile_count. 4218 if (log()) 4219 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 4220 Deoptimization::trap_reason_name(reason), 4221 md->trap_count(reason), trap_count(reason), 4222 md->decompile_count(), decompile_count()); 4223 return true; 4224 } else { 4225 // The coast is clear. 4226 return false; 4227 } 4228 } 4229 4230 // Compute when not to trap. Used by matching trap based nodes and 4231 // NullCheck optimization. 4232 void Compile::set_allowed_deopt_reasons() { 4233 _allowed_reasons = 0; 4234 if (is_method_compilation()) { 4235 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 4236 assert(rs < BitsPerInt, "recode bit map"); 4237 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 4238 _allowed_reasons |= nth_bit(rs); 4239 } 4240 } 4241 } 4242 } 4243 4244 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 4245 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 4246 } 4247 4248 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 4249 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 4250 } 4251 4252 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 4253 if (holder->is_initialized()) { 4254 return false; 4255 } 4256 if (holder->is_being_initialized()) { 4257 if (accessing_method->holder() == holder) { 4258 // Access inside a class. The barrier can be elided when access happens in <clinit>, 4259 // <init>, or a static method. In all those cases, there was an initialization 4260 // barrier on the holder klass passed. 4261 if (accessing_method->is_static_initializer() || 4262 accessing_method->is_object_initializer() || 4263 accessing_method->is_static()) { 4264 return false; 4265 } 4266 } else if (accessing_method->holder()->is_subclass_of(holder)) { 4267 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 4268 // In case of <init> or a static method, the barrier is on the subclass is not enough: 4269 // child class can become fully initialized while its parent class is still being initialized. 4270 if (accessing_method->is_static_initializer()) { 4271 return false; 4272 } 4273 } 4274 ciMethod* root = method(); // the root method of compilation 4275 if (root != accessing_method) { 4276 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 4277 } 4278 } 4279 return true; 4280 } 4281 4282 #ifndef PRODUCT 4283 //------------------------------verify_bidirectional_edges--------------------- 4284 // For each input edge to a node (ie - for each Use-Def edge), verify that 4285 // there is a corresponding Def-Use edge. 4286 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const { 4287 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc 4288 uint stack_size = live_nodes() >> 4; 4289 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize)); 4290 if (root_and_safepoints != nullptr) { 4291 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints"); 4292 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) { 4293 Node* root_or_safepoint = root_and_safepoints->at(i); 4294 // If the node is a safepoint, let's check if it still has a control input 4295 // Lack of control input signifies that this node was killed by CCP or 4296 // recursively by remove_globally_dead_node and it shouldn't be a starting 4297 // point. 4298 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) { 4299 nstack.push(root_or_safepoint); 4300 } 4301 } 4302 } else { 4303 nstack.push(_root); 4304 } 4305 4306 while (nstack.size() > 0) { 4307 Node* n = nstack.pop(); 4308 if (visited.member(n)) { 4309 continue; 4310 } 4311 visited.push(n); 4312 4313 // Walk over all input edges, checking for correspondence 4314 uint length = n->len(); 4315 for (uint i = 0; i < length; i++) { 4316 Node* in = n->in(i); 4317 if (in != nullptr && !visited.member(in)) { 4318 nstack.push(in); // Put it on stack 4319 } 4320 if (in != nullptr && !in->is_top()) { 4321 // Count instances of `next` 4322 int cnt = 0; 4323 for (uint idx = 0; idx < in->_outcnt; idx++) { 4324 if (in->_out[idx] == n) { 4325 cnt++; 4326 } 4327 } 4328 assert(cnt > 0, "Failed to find Def-Use edge."); 4329 // Check for duplicate edges 4330 // walk the input array downcounting the input edges to n 4331 for (uint j = 0; j < length; j++) { 4332 if (n->in(j) == in) { 4333 cnt--; 4334 } 4335 } 4336 assert(cnt == 0, "Mismatched edge count."); 4337 } else if (in == nullptr) { 4338 assert(i == 0 || i >= n->req() || 4339 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() || 4340 (n->is_Unlock() && i == (n->req() - 1)) || 4341 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion 4342 "only region, phi, arraycopy, unlock or membar nodes have null data edges"); 4343 } else { 4344 assert(in->is_top(), "sanity"); 4345 // Nothing to check. 4346 } 4347 } 4348 } 4349 } 4350 4351 //------------------------------verify_graph_edges--------------------------- 4352 // Walk the Graph and verify that there is a one-to-one correspondence 4353 // between Use-Def edges and Def-Use edges in the graph. 4354 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const { 4355 if (VerifyGraphEdges) { 4356 Unique_Node_List visited; 4357 4358 // Call graph walk to check edges 4359 verify_bidirectional_edges(visited, root_and_safepoints); 4360 if (no_dead_code) { 4361 // Now make sure that no visited node is used by an unvisited node. 4362 bool dead_nodes = false; 4363 Unique_Node_List checked; 4364 while (visited.size() > 0) { 4365 Node* n = visited.pop(); 4366 checked.push(n); 4367 for (uint i = 0; i < n->outcnt(); i++) { 4368 Node* use = n->raw_out(i); 4369 if (checked.member(use)) continue; // already checked 4370 if (visited.member(use)) continue; // already in the graph 4371 if (use->is_Con()) continue; // a dead ConNode is OK 4372 // At this point, we have found a dead node which is DU-reachable. 4373 if (!dead_nodes) { 4374 tty->print_cr("*** Dead nodes reachable via DU edges:"); 4375 dead_nodes = true; 4376 } 4377 use->dump(2); 4378 tty->print_cr("---"); 4379 checked.push(use); // No repeats; pretend it is now checked. 4380 } 4381 } 4382 assert(!dead_nodes, "using nodes must be reachable from root"); 4383 } 4384 } 4385 } 4386 #endif 4387 4388 // The Compile object keeps track of failure reasons separately from the ciEnv. 4389 // This is required because there is not quite a 1-1 relation between the 4390 // ciEnv and its compilation task and the Compile object. Note that one 4391 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4392 // to backtrack and retry without subsuming loads. Other than this backtracking 4393 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4394 // by the logic in C2Compiler. 4395 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) { 4396 if (log() != nullptr) { 4397 log()->elem("failure reason='%s' phase='compile'", reason); 4398 } 4399 if (_failure_reason.get() == nullptr) { 4400 // Record the first failure reason. 4401 _failure_reason.set(reason); 4402 if (CaptureBailoutInformation) { 4403 _first_failure_details = new CompilationFailureInfo(reason); 4404 } 4405 } else { 4406 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure."); 4407 } 4408 4409 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4410 C->print_method(PHASE_FAILURE, 1); 4411 } 4412 _root = nullptr; // flush the graph, too 4413 } 4414 4415 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id) 4416 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose), 4417 _compile(Compile::current()), 4418 _log(nullptr), 4419 _dolog(CITimeVerbose) 4420 { 4421 assert(_compile != nullptr, "sanity check"); 4422 assert(id != PhaseTraceId::_t_none, "Don't use none"); 4423 if (_dolog) { 4424 _log = _compile->log(); 4425 } 4426 if (_log != nullptr) { 4427 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes()); 4428 _log->stamp(); 4429 _log->end_head(); 4430 } 4431 4432 // Inform memory statistic, if enabled 4433 if (CompilationMemoryStatistic::enabled()) { 4434 CompilationMemoryStatistic::on_phase_start((int)id, name); 4435 } 4436 } 4437 4438 Compile::TracePhase::TracePhase(PhaseTraceId id) 4439 : TracePhase(Phase::get_phase_trace_id_text(id), id) {} 4440 4441 Compile::TracePhase::~TracePhase() { 4442 4443 // Inform memory statistic, if enabled 4444 if (CompilationMemoryStatistic::enabled()) { 4445 CompilationMemoryStatistic::on_phase_end(); 4446 } 4447 4448 if (_compile->failing_internal()) { 4449 if (_log != nullptr) { 4450 _log->done("phase"); 4451 } 4452 return; // timing code, not stressing bailouts. 4453 } 4454 #ifdef ASSERT 4455 if (PrintIdealNodeCount) { 4456 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4457 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk()); 4458 } 4459 4460 if (VerifyIdealNodeCount) { 4461 _compile->print_missing_nodes(); 4462 } 4463 #endif 4464 4465 if (_log != nullptr) { 4466 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes()); 4467 } 4468 } 4469 4470 //----------------------------static_subtype_check----------------------------- 4471 // Shortcut important common cases when superklass is exact: 4472 // (0) superklass is java.lang.Object (can occur in reflective code) 4473 // (1) subklass is already limited to a subtype of superklass => always ok 4474 // (2) subklass does not overlap with superklass => always fail 4475 // (3) superklass has NO subtypes and we can check with a simple compare. 4476 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) { 4477 if (skip) { 4478 return SSC_full_test; // Let caller generate the general case. 4479 } 4480 4481 if (subk->is_java_subtype_of(superk)) { 4482 return SSC_always_true; // (0) and (1) this test cannot fail 4483 } 4484 4485 if (!subk->maybe_java_subtype_of(superk)) { 4486 return SSC_always_false; // (2) true path dead; no dynamic test needed 4487 } 4488 4489 const Type* superelem = superk; 4490 if (superk->isa_aryklassptr()) { 4491 int ignored; 4492 superelem = superk->is_aryklassptr()->base_element_type(ignored); 4493 } 4494 4495 if (superelem->isa_instklassptr()) { 4496 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass(); 4497 if (!ik->has_subklass()) { 4498 if (!ik->is_final()) { 4499 // Add a dependency if there is a chance of a later subclass. 4500 dependencies()->assert_leaf_type(ik); 4501 } 4502 if (!superk->maybe_java_subtype_of(subk)) { 4503 return SSC_always_false; 4504 } 4505 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4506 } 4507 } else { 4508 // A primitive array type has no subtypes. 4509 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4510 } 4511 4512 return SSC_full_test; 4513 } 4514 4515 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4516 #ifdef _LP64 4517 // The scaled index operand to AddP must be a clean 64-bit value. 4518 // Java allows a 32-bit int to be incremented to a negative 4519 // value, which appears in a 64-bit register as a large 4520 // positive number. Using that large positive number as an 4521 // operand in pointer arithmetic has bad consequences. 4522 // On the other hand, 32-bit overflow is rare, and the possibility 4523 // can often be excluded, if we annotate the ConvI2L node with 4524 // a type assertion that its value is known to be a small positive 4525 // number. (The prior range check has ensured this.) 4526 // This assertion is used by ConvI2LNode::Ideal. 4527 int index_max = max_jint - 1; // array size is max_jint, index is one less 4528 if (sizetype != nullptr && sizetype->_hi > 0) { 4529 index_max = sizetype->_hi - 1; 4530 } 4531 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4532 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4533 #endif 4534 return idx; 4535 } 4536 4537 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4538 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 4539 if (ctrl != nullptr) { 4540 // Express control dependency by a CastII node with a narrow type. 4541 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4542 // node from floating above the range check during loop optimizations. Otherwise, the 4543 // ConvI2L node may be eliminated independently of the range check, causing the data path 4544 // to become TOP while the control path is still there (although it's unreachable). 4545 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */); 4546 value = phase->transform(value); 4547 } 4548 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4549 return phase->transform(new ConvI2LNode(value, ltype)); 4550 } 4551 4552 void Compile::dump_print_inlining() { 4553 inline_printer()->print_on(tty); 4554 } 4555 4556 void Compile::log_late_inline(CallGenerator* cg) { 4557 if (log() != nullptr) { 4558 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4559 cg->unique_id()); 4560 JVMState* p = cg->call_node()->jvms(); 4561 while (p != nullptr) { 4562 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4563 p = p->caller(); 4564 } 4565 log()->tail("late_inline"); 4566 } 4567 } 4568 4569 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4570 log_late_inline(cg); 4571 if (log() != nullptr) { 4572 log()->inline_fail(msg); 4573 } 4574 } 4575 4576 void Compile::log_inline_id(CallGenerator* cg) { 4577 if (log() != nullptr) { 4578 // The LogCompilation tool needs a unique way to identify late 4579 // inline call sites. This id must be unique for this call site in 4580 // this compilation. Try to have it unique across compilations as 4581 // well because it can be convenient when grepping through the log 4582 // file. 4583 // Distinguish OSR compilations from others in case CICountOSR is 4584 // on. 4585 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4586 cg->set_unique_id(id); 4587 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4588 } 4589 } 4590 4591 void Compile::log_inline_failure(const char* msg) { 4592 if (C->log() != nullptr) { 4593 C->log()->inline_fail(msg); 4594 } 4595 } 4596 4597 4598 // Dump inlining replay data to the stream. 4599 // Don't change thread state and acquire any locks. 4600 void Compile::dump_inline_data(outputStream* out) { 4601 InlineTree* inl_tree = ilt(); 4602 if (inl_tree != nullptr) { 4603 out->print(" inline %d", inl_tree->count()); 4604 inl_tree->dump_replay_data(out); 4605 } 4606 } 4607 4608 void Compile::dump_inline_data_reduced(outputStream* out) { 4609 assert(ReplayReduce, ""); 4610 4611 InlineTree* inl_tree = ilt(); 4612 if (inl_tree == nullptr) { 4613 return; 4614 } 4615 // Enable iterative replay file reduction 4616 // Output "compile" lines for depth 1 subtrees, 4617 // simulating that those trees were compiled 4618 // instead of inlined. 4619 for (int i = 0; i < inl_tree->subtrees().length(); ++i) { 4620 InlineTree* sub = inl_tree->subtrees().at(i); 4621 if (sub->inline_level() != 1) { 4622 continue; 4623 } 4624 4625 ciMethod* method = sub->method(); 4626 int entry_bci = -1; 4627 int comp_level = env()->task()->comp_level(); 4628 out->print("compile "); 4629 method->dump_name_as_ascii(out); 4630 out->print(" %d %d", entry_bci, comp_level); 4631 out->print(" inline %d", sub->count()); 4632 sub->dump_replay_data(out, -1); 4633 out->cr(); 4634 } 4635 } 4636 4637 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4638 if (n1->Opcode() < n2->Opcode()) return -1; 4639 else if (n1->Opcode() > n2->Opcode()) return 1; 4640 4641 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4642 for (uint i = 1; i < n1->req(); i++) { 4643 if (n1->in(i) < n2->in(i)) return -1; 4644 else if (n1->in(i) > n2->in(i)) return 1; 4645 } 4646 4647 return 0; 4648 } 4649 4650 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4651 Node* n1 = *n1p; 4652 Node* n2 = *n2p; 4653 4654 return cmp_expensive_nodes(n1, n2); 4655 } 4656 4657 void Compile::sort_expensive_nodes() { 4658 if (!expensive_nodes_sorted()) { 4659 _expensive_nodes.sort(cmp_expensive_nodes); 4660 } 4661 } 4662 4663 bool Compile::expensive_nodes_sorted() const { 4664 for (int i = 1; i < _expensive_nodes.length(); i++) { 4665 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) { 4666 return false; 4667 } 4668 } 4669 return true; 4670 } 4671 4672 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4673 if (_expensive_nodes.length() == 0) { 4674 return false; 4675 } 4676 4677 assert(OptimizeExpensiveOps, "optimization off?"); 4678 4679 // Take this opportunity to remove dead nodes from the list 4680 int j = 0; 4681 for (int i = 0; i < _expensive_nodes.length(); i++) { 4682 Node* n = _expensive_nodes.at(i); 4683 if (!n->is_unreachable(igvn)) { 4684 assert(n->is_expensive(), "should be expensive"); 4685 _expensive_nodes.at_put(j, n); 4686 j++; 4687 } 4688 } 4689 _expensive_nodes.trunc_to(j); 4690 4691 // Then sort the list so that similar nodes are next to each other 4692 // and check for at least two nodes of identical kind with same data 4693 // inputs. 4694 sort_expensive_nodes(); 4695 4696 for (int i = 0; i < _expensive_nodes.length()-1; i++) { 4697 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) { 4698 return true; 4699 } 4700 } 4701 4702 return false; 4703 } 4704 4705 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4706 if (_expensive_nodes.length() == 0) { 4707 return; 4708 } 4709 4710 assert(OptimizeExpensiveOps, "optimization off?"); 4711 4712 // Sort to bring similar nodes next to each other and clear the 4713 // control input of nodes for which there's only a single copy. 4714 sort_expensive_nodes(); 4715 4716 int j = 0; 4717 int identical = 0; 4718 int i = 0; 4719 bool modified = false; 4720 for (; i < _expensive_nodes.length()-1; i++) { 4721 assert(j <= i, "can't write beyond current index"); 4722 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) { 4723 identical++; 4724 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4725 continue; 4726 } 4727 if (identical > 0) { 4728 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4729 identical = 0; 4730 } else { 4731 Node* n = _expensive_nodes.at(i); 4732 igvn.replace_input_of(n, 0, nullptr); 4733 igvn.hash_insert(n); 4734 modified = true; 4735 } 4736 } 4737 if (identical > 0) { 4738 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4739 } else if (_expensive_nodes.length() >= 1) { 4740 Node* n = _expensive_nodes.at(i); 4741 igvn.replace_input_of(n, 0, nullptr); 4742 igvn.hash_insert(n); 4743 modified = true; 4744 } 4745 _expensive_nodes.trunc_to(j); 4746 if (modified) { 4747 igvn.optimize(); 4748 } 4749 } 4750 4751 void Compile::add_expensive_node(Node * n) { 4752 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list"); 4753 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4754 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4755 if (OptimizeExpensiveOps) { 4756 _expensive_nodes.append(n); 4757 } else { 4758 // Clear control input and let IGVN optimize expensive nodes if 4759 // OptimizeExpensiveOps is off. 4760 n->set_req(0, nullptr); 4761 } 4762 } 4763 4764 /** 4765 * Track coarsened Lock and Unlock nodes. 4766 */ 4767 4768 class Lock_List : public Node_List { 4769 uint _origin_cnt; 4770 public: 4771 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {} 4772 uint origin_cnt() const { return _origin_cnt; } 4773 }; 4774 4775 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) { 4776 int length = locks.length(); 4777 if (length > 0) { 4778 // Have to keep this list until locks elimination during Macro nodes elimination. 4779 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length); 4780 AbstractLockNode* alock = locks.at(0); 4781 BoxLockNode* box = alock->box_node()->as_BoxLock(); 4782 for (int i = 0; i < length; i++) { 4783 AbstractLockNode* lock = locks.at(i); 4784 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx); 4785 locks_list->push(lock); 4786 BoxLockNode* this_box = lock->box_node()->as_BoxLock(); 4787 if (this_box != box) { 4788 // Locking regions (BoxLock) could be Unbalanced here: 4789 // - its coarsened locks were eliminated in earlier 4790 // macro nodes elimination followed by loop unroll 4791 // - it is OSR locking region (no Lock node) 4792 // Preserve Unbalanced status in such cases. 4793 if (!this_box->is_unbalanced()) { 4794 this_box->set_coarsened(); 4795 } 4796 if (!box->is_unbalanced()) { 4797 box->set_coarsened(); 4798 } 4799 } 4800 } 4801 _coarsened_locks.append(locks_list); 4802 } 4803 } 4804 4805 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) { 4806 int count = coarsened_count(); 4807 for (int i = 0; i < count; i++) { 4808 Node_List* locks_list = _coarsened_locks.at(i); 4809 for (uint j = 0; j < locks_list->size(); j++) { 4810 Node* lock = locks_list->at(j); 4811 assert(lock->is_AbstractLock(), "sanity"); 4812 if (!useful.member(lock)) { 4813 locks_list->yank(lock); 4814 } 4815 } 4816 } 4817 } 4818 4819 void Compile::remove_coarsened_lock(Node* n) { 4820 if (n->is_AbstractLock()) { 4821 int count = coarsened_count(); 4822 for (int i = 0; i < count; i++) { 4823 Node_List* locks_list = _coarsened_locks.at(i); 4824 locks_list->yank(n); 4825 } 4826 } 4827 } 4828 4829 bool Compile::coarsened_locks_consistent() { 4830 int count = coarsened_count(); 4831 for (int i = 0; i < count; i++) { 4832 bool unbalanced = false; 4833 bool modified = false; // track locks kind modifications 4834 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i); 4835 uint size = locks_list->size(); 4836 if (size == 0) { 4837 unbalanced = false; // All locks were eliminated - good 4838 } else if (size != locks_list->origin_cnt()) { 4839 unbalanced = true; // Some locks were removed from list 4840 } else { 4841 for (uint j = 0; j < size; j++) { 4842 Node* lock = locks_list->at(j); 4843 // All nodes in group should have the same state (modified or not) 4844 if (!lock->as_AbstractLock()->is_coarsened()) { 4845 if (j == 0) { 4846 // first on list was modified, the rest should be too for consistency 4847 modified = true; 4848 } else if (!modified) { 4849 // this lock was modified but previous locks on the list were not 4850 unbalanced = true; 4851 break; 4852 } 4853 } else if (modified) { 4854 // previous locks on list were modified but not this lock 4855 unbalanced = true; 4856 break; 4857 } 4858 } 4859 } 4860 if (unbalanced) { 4861 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified 4862 #ifdef ASSERT 4863 if (PrintEliminateLocks) { 4864 tty->print_cr("=== unbalanced coarsened locks ==="); 4865 for (uint l = 0; l < size; l++) { 4866 locks_list->at(l)->dump(); 4867 } 4868 } 4869 #endif 4870 record_failure(C2Compiler::retry_no_locks_coarsening()); 4871 return false; 4872 } 4873 } 4874 return true; 4875 } 4876 4877 // Mark locking regions (identified by BoxLockNode) as unbalanced if 4878 // locks coarsening optimization removed Lock/Unlock nodes from them. 4879 // Such regions become unbalanced because coarsening only removes part 4880 // of Lock/Unlock nodes in region. As result we can't execute other 4881 // locks elimination optimizations which assume all code paths have 4882 // corresponding pair of Lock/Unlock nodes - they are balanced. 4883 void Compile::mark_unbalanced_boxes() const { 4884 int count = coarsened_count(); 4885 for (int i = 0; i < count; i++) { 4886 Node_List* locks_list = _coarsened_locks.at(i); 4887 uint size = locks_list->size(); 4888 if (size > 0) { 4889 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock(); 4890 BoxLockNode* box = alock->box_node()->as_BoxLock(); 4891 if (alock->is_coarsened()) { 4892 // coarsened_locks_consistent(), which is called before this method, verifies 4893 // that the rest of Lock/Unlock nodes on locks_list are also coarsened. 4894 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated"); 4895 for (uint j = 1; j < size; j++) { 4896 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here"); 4897 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock(); 4898 if (box != this_box) { 4899 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated"); 4900 box->set_unbalanced(); 4901 this_box->set_unbalanced(); 4902 } 4903 } 4904 } 4905 } 4906 } 4907 } 4908 4909 /** 4910 * Remove the speculative part of types and clean up the graph 4911 */ 4912 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4913 if (UseTypeSpeculation) { 4914 Unique_Node_List worklist; 4915 worklist.push(root()); 4916 int modified = 0; 4917 // Go over all type nodes that carry a speculative type, drop the 4918 // speculative part of the type and enqueue the node for an igvn 4919 // which may optimize it out. 4920 for (uint next = 0; next < worklist.size(); ++next) { 4921 Node *n = worklist.at(next); 4922 if (n->is_Type()) { 4923 TypeNode* tn = n->as_Type(); 4924 const Type* t = tn->type(); 4925 const Type* t_no_spec = t->remove_speculative(); 4926 if (t_no_spec != t) { 4927 bool in_hash = igvn.hash_delete(n); 4928 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table"); 4929 tn->set_type(t_no_spec); 4930 igvn.hash_insert(n); 4931 igvn._worklist.push(n); // give it a chance to go away 4932 modified++; 4933 } 4934 } 4935 // Iterate over outs - endless loops is unreachable from below 4936 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4937 Node *m = n->fast_out(i); 4938 if (not_a_node(m)) { 4939 continue; 4940 } 4941 worklist.push(m); 4942 } 4943 } 4944 // Drop the speculative part of all types in the igvn's type table 4945 igvn.remove_speculative_types(); 4946 if (modified > 0) { 4947 igvn.optimize(); 4948 if (failing()) return; 4949 } 4950 #ifdef ASSERT 4951 // Verify that after the IGVN is over no speculative type has resurfaced 4952 worklist.clear(); 4953 worklist.push(root()); 4954 for (uint next = 0; next < worklist.size(); ++next) { 4955 Node *n = worklist.at(next); 4956 const Type* t = igvn.type_or_null(n); 4957 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types"); 4958 if (n->is_Type()) { 4959 t = n->as_Type()->type(); 4960 assert(t == t->remove_speculative(), "no more speculative types"); 4961 } 4962 // Iterate over outs - endless loops is unreachable from below 4963 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4964 Node *m = n->fast_out(i); 4965 if (not_a_node(m)) { 4966 continue; 4967 } 4968 worklist.push(m); 4969 } 4970 } 4971 igvn.check_no_speculative_types(); 4972 #endif 4973 } 4974 } 4975 4976 // Auxiliary methods to support randomized stressing/fuzzing. 4977 4978 void Compile::initialize_stress_seed(const DirectiveSet* directive) { 4979 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) { 4980 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds()); 4981 FLAG_SET_ERGO(StressSeed, _stress_seed); 4982 } else { 4983 _stress_seed = StressSeed; 4984 } 4985 if (_log != nullptr) { 4986 _log->elem("stress_test seed='%u'", _stress_seed); 4987 } 4988 } 4989 4990 int Compile::random() { 4991 _stress_seed = os::next_random(_stress_seed); 4992 return static_cast<int>(_stress_seed); 4993 } 4994 4995 // This method can be called the arbitrary number of times, with current count 4996 // as the argument. The logic allows selecting a single candidate from the 4997 // running list of candidates as follows: 4998 // int count = 0; 4999 // Cand* selected = null; 5000 // while(cand = cand->next()) { 5001 // if (randomized_select(++count)) { 5002 // selected = cand; 5003 // } 5004 // } 5005 // 5006 // Including count equalizes the chances any candidate is "selected". 5007 // This is useful when we don't have the complete list of candidates to choose 5008 // from uniformly. In this case, we need to adjust the randomicity of the 5009 // selection, or else we will end up biasing the selection towards the latter 5010 // candidates. 5011 // 5012 // Quick back-envelope calculation shows that for the list of n candidates 5013 // the equal probability for the candidate to persist as "best" can be 5014 // achieved by replacing it with "next" k-th candidate with the probability 5015 // of 1/k. It can be easily shown that by the end of the run, the 5016 // probability for any candidate is converged to 1/n, thus giving the 5017 // uniform distribution among all the candidates. 5018 // 5019 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 5020 #define RANDOMIZED_DOMAIN_POW 29 5021 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 5022 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 5023 bool Compile::randomized_select(int count) { 5024 assert(count > 0, "only positive"); 5025 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 5026 } 5027 5028 #ifdef ASSERT 5029 // Failures are geometrically distributed with probability 1/StressBailoutMean. 5030 bool Compile::fail_randomly() { 5031 if ((random() % StressBailoutMean) != 0) { 5032 return false; 5033 } 5034 record_failure("StressBailout"); 5035 return true; 5036 } 5037 5038 bool Compile::failure_is_artificial() { 5039 return C->failure_reason_is("StressBailout"); 5040 } 5041 #endif 5042 5043 CloneMap& Compile::clone_map() { return _clone_map; } 5044 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 5045 5046 void NodeCloneInfo::dump_on(outputStream* st) const { 5047 st->print(" {%d:%d} ", idx(), gen()); 5048 } 5049 5050 void CloneMap::clone(Node* old, Node* nnn, int gen) { 5051 uint64_t val = value(old->_idx); 5052 NodeCloneInfo cio(val); 5053 assert(val != 0, "old node should be in the map"); 5054 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 5055 insert(nnn->_idx, cin.get()); 5056 #ifndef PRODUCT 5057 if (is_debug()) { 5058 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 5059 } 5060 #endif 5061 } 5062 5063 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 5064 NodeCloneInfo cio(value(old->_idx)); 5065 if (cio.get() == 0) { 5066 cio.set(old->_idx, 0); 5067 insert(old->_idx, cio.get()); 5068 #ifndef PRODUCT 5069 if (is_debug()) { 5070 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 5071 } 5072 #endif 5073 } 5074 clone(old, nnn, gen); 5075 } 5076 5077 int CloneMap::max_gen() const { 5078 int g = 0; 5079 DictI di(_dict); 5080 for(; di.test(); ++di) { 5081 int t = gen(di._key); 5082 if (g < t) { 5083 g = t; 5084 #ifndef PRODUCT 5085 if (is_debug()) { 5086 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 5087 } 5088 #endif 5089 } 5090 } 5091 return g; 5092 } 5093 5094 void CloneMap::dump(node_idx_t key, outputStream* st) const { 5095 uint64_t val = value(key); 5096 if (val != 0) { 5097 NodeCloneInfo ni(val); 5098 ni.dump_on(st); 5099 } 5100 } 5101 5102 void Compile::shuffle_macro_nodes() { 5103 if (_macro_nodes.length() < 2) { 5104 return; 5105 } 5106 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) { 5107 uint j = C->random() % (i + 1); 5108 swap(_macro_nodes.at(i), _macro_nodes.at(j)); 5109 } 5110 } 5111 5112 // Move Allocate nodes to the start of the list 5113 void Compile::sort_macro_nodes() { 5114 int count = macro_count(); 5115 int allocates = 0; 5116 for (int i = 0; i < count; i++) { 5117 Node* n = macro_node(i); 5118 if (n->is_Allocate()) { 5119 if (i != allocates) { 5120 Node* tmp = macro_node(allocates); 5121 _macro_nodes.at_put(allocates, n); 5122 _macro_nodes.at_put(i, tmp); 5123 } 5124 allocates++; 5125 } 5126 } 5127 } 5128 5129 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) { 5130 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code. 5131 EventCompilerPhase event(UNTIMED); 5132 if (event.should_commit()) { 5133 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 5134 } 5135 #ifndef PRODUCT 5136 ResourceMark rm; 5137 stringStream ss; 5138 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt)); 5139 int iter = ++_igv_phase_iter[cpt]; 5140 if (iter > 1) { 5141 ss.print(" %d", iter); 5142 } 5143 if (n != nullptr) { 5144 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]); 5145 if (n->is_Call()) { 5146 CallNode* call = n->as_Call(); 5147 if (call->_name != nullptr) { 5148 // E.g. uncommon traps etc. 5149 ss.print(" - %s", call->_name); 5150 } else if (call->is_CallJava()) { 5151 CallJavaNode* call_java = call->as_CallJava(); 5152 if (call_java->method() != nullptr) { 5153 ss.print(" -"); 5154 call_java->method()->print_short_name(&ss); 5155 } 5156 } 5157 } 5158 } 5159 5160 const char* name = ss.as_string(); 5161 if (should_print_igv(level)) { 5162 _igv_printer->print_graph(name); 5163 } 5164 if (should_print_phase(level)) { 5165 print_phase(name); 5166 } 5167 if (should_print_ideal_phase(cpt)) { 5168 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt)); 5169 } 5170 #endif 5171 C->_latest_stage_start_counter.stamp(); 5172 } 5173 5174 // Only used from CompileWrapper 5175 void Compile::begin_method() { 5176 #ifndef PRODUCT 5177 if (_method != nullptr && should_print_igv(1)) { 5178 _igv_printer->begin_method(); 5179 } 5180 #endif 5181 C->_latest_stage_start_counter.stamp(); 5182 } 5183 5184 // Only used from CompileWrapper 5185 void Compile::end_method() { 5186 EventCompilerPhase event(UNTIMED); 5187 if (event.should_commit()) { 5188 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1); 5189 } 5190 5191 #ifndef PRODUCT 5192 if (_method != nullptr && should_print_igv(1)) { 5193 _igv_printer->end_method(); 5194 } 5195 #endif 5196 } 5197 5198 #ifndef PRODUCT 5199 bool Compile::should_print_phase(const int level) const { 5200 return PrintPhaseLevel > 0 && directive()->PhasePrintLevelOption >= level && 5201 _method != nullptr; // Do not print phases for stubs. 5202 } 5203 5204 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const { 5205 return _directive->should_print_ideal_phase(cpt); 5206 } 5207 5208 void Compile::init_igv() { 5209 if (_igv_printer == nullptr) { 5210 _igv_printer = IdealGraphPrinter::printer(); 5211 _igv_printer->set_compile(this); 5212 } 5213 } 5214 5215 bool Compile::should_print_igv(const int level) { 5216 PRODUCT_RETURN_(return false;); 5217 5218 if (PrintIdealGraphLevel < 0) { // disabled by the user 5219 return false; 5220 } 5221 5222 bool need = directive()->IGVPrintLevelOption >= level; 5223 if (need) { 5224 Compile::init_igv(); 5225 } 5226 return need; 5227 } 5228 5229 IdealGraphPrinter* Compile::_debug_file_printer = nullptr; 5230 IdealGraphPrinter* Compile::_debug_network_printer = nullptr; 5231 5232 // Called from debugger. Prints method to the default file with the default phase name. 5233 // This works regardless of any Ideal Graph Visualizer flags set or not. 5234 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc). 5235 void igv_print(void* sp, void* fp, void* pc) { 5236 frame fr(sp, fp, pc); 5237 Compile::current()->igv_print_method_to_file(nullptr, false, &fr); 5238 } 5239 5240 // Same as igv_print() above but with a specified phase name. 5241 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) { 5242 frame fr(sp, fp, pc); 5243 Compile::current()->igv_print_method_to_file(phase_name, false, &fr); 5244 } 5245 5246 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 5247 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 5248 // This works regardless of any Ideal Graph Visualizer flags set or not. 5249 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc). 5250 void igv_print(bool network, void* sp, void* fp, void* pc) { 5251 frame fr(sp, fp, pc); 5252 if (network) { 5253 Compile::current()->igv_print_method_to_network(nullptr, &fr); 5254 } else { 5255 Compile::current()->igv_print_method_to_file(nullptr, false, &fr); 5256 } 5257 } 5258 5259 // Same as igv_print(bool network, ...) above but with a specified phase name. 5260 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc). 5261 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) { 5262 frame fr(sp, fp, pc); 5263 if (network) { 5264 Compile::current()->igv_print_method_to_network(phase_name, &fr); 5265 } else { 5266 Compile::current()->igv_print_method_to_file(phase_name, false, &fr); 5267 } 5268 } 5269 5270 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 5271 void igv_print_default() { 5272 Compile::current()->print_method(PHASE_DEBUG, 0); 5273 } 5274 5275 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 5276 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 5277 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 5278 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc). 5279 void igv_append(void* sp, void* fp, void* pc) { 5280 frame fr(sp, fp, pc); 5281 Compile::current()->igv_print_method_to_file(nullptr, true, &fr); 5282 } 5283 5284 // Same as igv_append(...) above but with a specified phase name. 5285 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc). 5286 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) { 5287 frame fr(sp, fp, pc); 5288 Compile::current()->igv_print_method_to_file(phase_name, true, &fr); 5289 } 5290 5291 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) { 5292 const char* file_name = "custom_debug.xml"; 5293 if (_debug_file_printer == nullptr) { 5294 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 5295 } else { 5296 _debug_file_printer->update_compiled_method(C->method()); 5297 } 5298 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 5299 _debug_file_printer->print_graph(phase_name, fr); 5300 } 5301 5302 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) { 5303 ResourceMark rm; 5304 GrowableArray<const Node*> empty_list; 5305 igv_print_graph_to_network(phase_name, empty_list, fr); 5306 } 5307 5308 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) { 5309 if (_debug_network_printer == nullptr) { 5310 _debug_network_printer = new IdealGraphPrinter(C); 5311 } else { 5312 _debug_network_printer->update_compiled_method(C->method()); 5313 } 5314 tty->print_cr("Method printed over network stream to IGV"); 5315 _debug_network_printer->print(name, C->root(), visible_nodes, fr); 5316 } 5317 #endif // !PRODUCT 5318 5319 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 5320 if (type != nullptr && phase->type(value)->higher_equal(type)) { 5321 return value; 5322 } 5323 Node* result = nullptr; 5324 if (bt == T_BYTE) { 5325 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 5326 result = new RShiftINode(result, phase->intcon(24)); 5327 } else if (bt == T_BOOLEAN) { 5328 result = new AndINode(value, phase->intcon(0xFF)); 5329 } else if (bt == T_CHAR) { 5330 result = new AndINode(value,phase->intcon(0xFFFF)); 5331 } else { 5332 assert(bt == T_SHORT, "unexpected narrow type"); 5333 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 5334 result = new RShiftINode(result, phase->intcon(16)); 5335 } 5336 if (transform_res) { 5337 result = phase->transform(result); 5338 } 5339 return result; 5340 } 5341 5342 void Compile::record_method_not_compilable_oom() { 5343 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit()); 5344 }