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(StubId::NO_STUBID), 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 StubId 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 BlobId blob_id = StubInfo::blob(_stub_id); 968 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, blob_id); 969 if (blob != nullptr) { 970 RuntimeStub* rs = blob->as_runtime_stub(); 971 _stub_entry_point = rs->entry_point(); 972 return; 973 } 974 } 975 976 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false); 977 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false); 978 979 #ifndef PRODUCT 980 set_print_assembly(PrintFrameConverterAssembly); 981 set_parsed_irreducible_loop(false); 982 #else 983 set_print_assembly(false); // Must initialize. 984 #endif 985 set_has_irreducible_loop(false); // no loops 986 987 CompileWrapper cw(this); 988 Init(/*do_aliasing=*/ false); 989 init_tf((*generator)()); 990 991 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 992 _types = new (comp_arena()) Type_Array(comp_arena()); 993 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255); 994 995 if (StressLCM || StressGCM || StressBailout) { 996 initialize_stress_seed(directive); 997 } 998 999 { 1000 PhaseGVN gvn; 1001 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1002 gvn.transform(top()); 1003 1004 GraphKit kit; 1005 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1006 } 1007 1008 NOT_PRODUCT( verify_graph_edges(); ) 1009 1010 Code_Gen(); 1011 } 1012 1013 Compile::~Compile() { 1014 delete _first_failure_details; 1015 }; 1016 1017 //------------------------------Init------------------------------------------- 1018 // Prepare for a single compilation 1019 void Compile::Init(bool aliasing) { 1020 _do_aliasing = aliasing; 1021 _unique = 0; 1022 _regalloc = nullptr; 1023 1024 _tf = nullptr; // filled in later 1025 _top = nullptr; // cached later 1026 _matcher = nullptr; // filled in later 1027 _cfg = nullptr; // filled in later 1028 1029 _node_note_array = nullptr; 1030 _default_node_notes = nullptr; 1031 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize() 1032 1033 _immutable_memory = nullptr; // filled in at first inquiry 1034 1035 #ifdef ASSERT 1036 _phase_optimize_finished = false; 1037 _phase_verify_ideal_loop = false; 1038 _exception_backedge = false; 1039 _type_verify = nullptr; 1040 #endif 1041 1042 // Globally visible Nodes 1043 // First set TOP to null to give safe behavior during creation of RootNode 1044 set_cached_top_node(nullptr); 1045 set_root(new RootNode()); 1046 // Now that you have a Root to point to, create the real TOP 1047 set_cached_top_node( new ConNode(Type::TOP) ); 1048 set_recent_alloc(nullptr, nullptr); 1049 1050 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1051 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1052 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1053 env()->set_dependencies(new Dependencies(env())); 1054 1055 _fixed_slots = 0; 1056 set_has_split_ifs(false); 1057 set_has_loops(false); // first approximation 1058 set_has_stringbuilder(false); 1059 set_has_boxed_value(false); 1060 _trap_can_recompile = false; // no traps emitted yet 1061 _major_progress = true; // start out assuming good things will happen 1062 set_has_unsafe_access(false); 1063 set_max_vector_size(0); 1064 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1065 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1066 set_decompile_count(0); 1067 1068 #ifndef PRODUCT 1069 _phase_counter = 0; 1070 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter)); 1071 #endif 1072 1073 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1074 _loop_opts_cnt = LoopOptsCount; 1075 set_do_inlining(Inline); 1076 set_max_inline_size(MaxInlineSize); 1077 set_freq_inline_size(FreqInlineSize); 1078 set_do_scheduling(OptoScheduling); 1079 1080 set_do_vector_loop(false); 1081 set_has_monitors(false); 1082 set_has_scoped_access(false); 1083 1084 if (AllowVectorizeOnDemand) { 1085 if (has_method() && _directive->VectorizeOption) { 1086 set_do_vector_loop(true); 1087 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());}) 1088 } else if (has_method() && method()->name() != nullptr && 1089 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1090 set_do_vector_loop(true); 1091 } 1092 } 1093 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1094 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());}) 1095 1096 _max_node_limit = _directive->MaxNodeLimitOption; 1097 1098 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1099 set_clinit_barrier_on_entry(true); 1100 } 1101 if (debug_info()->recording_non_safepoints()) { 1102 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1103 (comp_arena(), 8, 0, nullptr)); 1104 set_default_node_notes(Node_Notes::make(this)); 1105 } 1106 1107 const int grow_ats = 16; 1108 _max_alias_types = grow_ats; 1109 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1110 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1111 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1112 { 1113 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1114 } 1115 // Initialize the first few types. 1116 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr); 1117 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1118 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1119 _num_alias_types = AliasIdxRaw+1; 1120 // Zero out the alias type cache. 1121 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1122 // A null adr_type hits in the cache right away. Preload the right answer. 1123 probe_alias_cache(nullptr)->_index = AliasIdxTop; 1124 } 1125 1126 #ifdef ASSERT 1127 // Verify that the current StartNode is valid. 1128 void Compile::verify_start(StartNode* s) const { 1129 assert(failing_internal() || s == start(), "should be StartNode"); 1130 } 1131 #endif 1132 1133 /** 1134 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1135 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1136 * the ideal graph. 1137 */ 1138 StartNode* Compile::start() const { 1139 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason()); 1140 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1141 Node* start = root()->fast_out(i); 1142 if (start->is_Start()) { 1143 return start->as_Start(); 1144 } 1145 } 1146 fatal("Did not find Start node!"); 1147 return nullptr; 1148 } 1149 1150 //-------------------------------immutable_memory------------------------------------- 1151 // Access immutable memory 1152 Node* Compile::immutable_memory() { 1153 if (_immutable_memory != nullptr) { 1154 return _immutable_memory; 1155 } 1156 StartNode* s = start(); 1157 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1158 Node *p = s->fast_out(i); 1159 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1160 _immutable_memory = p; 1161 return _immutable_memory; 1162 } 1163 } 1164 ShouldNotReachHere(); 1165 return nullptr; 1166 } 1167 1168 //----------------------set_cached_top_node------------------------------------ 1169 // Install the cached top node, and make sure Node::is_top works correctly. 1170 void Compile::set_cached_top_node(Node* tn) { 1171 if (tn != nullptr) verify_top(tn); 1172 Node* old_top = _top; 1173 _top = tn; 1174 // Calling Node::setup_is_top allows the nodes the chance to adjust 1175 // their _out arrays. 1176 if (_top != nullptr) _top->setup_is_top(); 1177 if (old_top != nullptr) old_top->setup_is_top(); 1178 assert(_top == nullptr || top()->is_top(), ""); 1179 } 1180 1181 #ifdef ASSERT 1182 uint Compile::count_live_nodes_by_graph_walk() { 1183 Unique_Node_List useful(comp_arena()); 1184 // Get useful node list by walking the graph. 1185 identify_useful_nodes(useful); 1186 return useful.size(); 1187 } 1188 1189 void Compile::print_missing_nodes() { 1190 1191 // Return if CompileLog is null and PrintIdealNodeCount is false. 1192 if ((_log == nullptr) && (! PrintIdealNodeCount)) { 1193 return; 1194 } 1195 1196 // This is an expensive function. It is executed only when the user 1197 // specifies VerifyIdealNodeCount option or otherwise knows the 1198 // additional work that needs to be done to identify reachable nodes 1199 // by walking the flow graph and find the missing ones using 1200 // _dead_node_list. 1201 1202 Unique_Node_List useful(comp_arena()); 1203 // Get useful node list by walking the graph. 1204 identify_useful_nodes(useful); 1205 1206 uint l_nodes = C->live_nodes(); 1207 uint l_nodes_by_walk = useful.size(); 1208 1209 if (l_nodes != l_nodes_by_walk) { 1210 if (_log != nullptr) { 1211 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1212 _log->stamp(); 1213 _log->end_head(); 1214 } 1215 VectorSet& useful_member_set = useful.member_set(); 1216 int last_idx = l_nodes_by_walk; 1217 for (int i = 0; i < last_idx; i++) { 1218 if (useful_member_set.test(i)) { 1219 if (_dead_node_list.test(i)) { 1220 if (_log != nullptr) { 1221 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1222 } 1223 if (PrintIdealNodeCount) { 1224 // Print the log message to tty 1225 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1226 useful.at(i)->dump(); 1227 } 1228 } 1229 } 1230 else if (! _dead_node_list.test(i)) { 1231 if (_log != nullptr) { 1232 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1233 } 1234 if (PrintIdealNodeCount) { 1235 // Print the log message to tty 1236 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1237 } 1238 } 1239 } 1240 if (_log != nullptr) { 1241 _log->tail("mismatched_nodes"); 1242 } 1243 } 1244 } 1245 void Compile::record_modified_node(Node* n) { 1246 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) { 1247 _modified_nodes->push(n); 1248 } 1249 } 1250 1251 void Compile::remove_modified_node(Node* n) { 1252 if (_modified_nodes != nullptr) { 1253 _modified_nodes->remove(n); 1254 } 1255 } 1256 #endif 1257 1258 #ifndef PRODUCT 1259 void Compile::verify_top(Node* tn) const { 1260 if (tn != nullptr) { 1261 assert(tn->is_Con(), "top node must be a constant"); 1262 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1263 assert(tn->in(0) != nullptr, "must have live top node"); 1264 } 1265 } 1266 #endif 1267 1268 1269 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1270 1271 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1272 guarantee(arr != nullptr, ""); 1273 int num_blocks = arr->length(); 1274 if (grow_by < num_blocks) grow_by = num_blocks; 1275 int num_notes = grow_by * _node_notes_block_size; 1276 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1277 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1278 while (num_notes > 0) { 1279 arr->append(notes); 1280 notes += _node_notes_block_size; 1281 num_notes -= _node_notes_block_size; 1282 } 1283 assert(num_notes == 0, "exact multiple, please"); 1284 } 1285 1286 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1287 if (source == nullptr || dest == nullptr) return false; 1288 1289 if (dest->is_Con()) 1290 return false; // Do not push debug info onto constants. 1291 1292 #ifdef ASSERT 1293 // Leave a bread crumb trail pointing to the original node: 1294 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) { 1295 dest->set_debug_orig(source); 1296 } 1297 #endif 1298 1299 if (node_note_array() == nullptr) 1300 return false; // Not collecting any notes now. 1301 1302 // This is a copy onto a pre-existing node, which may already have notes. 1303 // If both nodes have notes, do not overwrite any pre-existing notes. 1304 Node_Notes* source_notes = node_notes_at(source->_idx); 1305 if (source_notes == nullptr || source_notes->is_clear()) return false; 1306 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1307 if (dest_notes == nullptr || dest_notes->is_clear()) { 1308 return set_node_notes_at(dest->_idx, source_notes); 1309 } 1310 1311 Node_Notes merged_notes = (*source_notes); 1312 // The order of operations here ensures that dest notes will win... 1313 merged_notes.update_from(dest_notes); 1314 return set_node_notes_at(dest->_idx, &merged_notes); 1315 } 1316 1317 1318 //--------------------------allow_range_check_smearing------------------------- 1319 // Gating condition for coalescing similar range checks. 1320 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1321 // single covering check that is at least as strong as any of them. 1322 // If the optimization succeeds, the simplified (strengthened) range check 1323 // will always succeed. If it fails, we will deopt, and then give up 1324 // on the optimization. 1325 bool Compile::allow_range_check_smearing() const { 1326 // If this method has already thrown a range-check, 1327 // assume it was because we already tried range smearing 1328 // and it failed. 1329 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1330 return !already_trapped; 1331 } 1332 1333 1334 //------------------------------flatten_alias_type----------------------------- 1335 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1336 assert(do_aliasing(), "Aliasing should be enabled"); 1337 int offset = tj->offset(); 1338 TypePtr::PTR ptr = tj->ptr(); 1339 1340 // Known instance (scalarizable allocation) alias only with itself. 1341 bool is_known_inst = tj->isa_oopptr() != nullptr && 1342 tj->is_oopptr()->is_known_instance(); 1343 1344 // Process weird unsafe references. 1345 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1346 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops"); 1347 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1348 tj = TypeOopPtr::BOTTOM; 1349 ptr = tj->ptr(); 1350 offset = tj->offset(); 1351 } 1352 1353 // Array pointers need some flattening 1354 const TypeAryPtr* ta = tj->isa_aryptr(); 1355 if (ta && ta->is_stable()) { 1356 // Erase stability property for alias analysis. 1357 tj = ta = ta->cast_to_stable(false); 1358 } 1359 if( ta && is_known_inst ) { 1360 if ( offset != Type::OffsetBot && 1361 offset > arrayOopDesc::length_offset_in_bytes() ) { 1362 offset = Type::OffsetBot; // Flatten constant access into array body only 1363 tj = ta = ta-> 1364 remove_speculative()-> 1365 cast_to_ptr_type(ptr)-> 1366 with_offset(offset); 1367 } 1368 } else if (ta) { 1369 // For arrays indexed by constant indices, we flatten the alias 1370 // space to include all of the array body. Only the header, klass 1371 // and array length can be accessed un-aliased. 1372 if( offset != Type::OffsetBot ) { 1373 if( ta->const_oop() ) { // MethodData* or Method* 1374 offset = Type::OffsetBot; // Flatten constant access into array body 1375 tj = ta = ta-> 1376 remove_speculative()-> 1377 cast_to_ptr_type(ptr)-> 1378 cast_to_exactness(false)-> 1379 with_offset(offset); 1380 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1381 // range is OK as-is. 1382 tj = ta = TypeAryPtr::RANGE; 1383 } else if( offset == Type::klass_offset() ) { 1384 tj = TypeInstPtr::KLASS; // all klass loads look alike 1385 ta = TypeAryPtr::RANGE; // generic ignored junk 1386 ptr = TypePtr::BotPTR; 1387 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1388 tj = TypeInstPtr::MARK; 1389 ta = TypeAryPtr::RANGE; // generic ignored junk 1390 ptr = TypePtr::BotPTR; 1391 } else { // Random constant offset into array body 1392 offset = Type::OffsetBot; // Flatten constant access into array body 1393 tj = ta = ta-> 1394 remove_speculative()-> 1395 cast_to_ptr_type(ptr)-> 1396 cast_to_exactness(false)-> 1397 with_offset(offset); 1398 } 1399 } 1400 // Arrays of fixed size alias with arrays of unknown size. 1401 if (ta->size() != TypeInt::POS) { 1402 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1403 tj = ta = ta-> 1404 remove_speculative()-> 1405 cast_to_ptr_type(ptr)-> 1406 with_ary(tary)-> 1407 cast_to_exactness(false); 1408 } 1409 // Arrays of known objects become arrays of unknown objects. 1410 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1411 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1412 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1413 } 1414 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1415 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1416 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1417 } 1418 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1419 // cannot be distinguished by bytecode alone. 1420 if (ta->elem() == TypeInt::BOOL) { 1421 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1422 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1423 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1424 } 1425 // During the 2nd round of IterGVN, NotNull castings are removed. 1426 // Make sure the Bottom and NotNull variants alias the same. 1427 // Also, make sure exact and non-exact variants alias the same. 1428 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) { 1429 tj = ta = ta-> 1430 remove_speculative()-> 1431 cast_to_ptr_type(TypePtr::BotPTR)-> 1432 cast_to_exactness(false)-> 1433 with_offset(offset); 1434 } 1435 } 1436 1437 // Oop pointers need some flattening 1438 const TypeInstPtr *to = tj->isa_instptr(); 1439 if (to && to != TypeOopPtr::BOTTOM) { 1440 ciInstanceKlass* ik = to->instance_klass(); 1441 if( ptr == TypePtr::Constant ) { 1442 if (ik != ciEnv::current()->Class_klass() || 1443 offset < ik->layout_helper_size_in_bytes()) { 1444 // No constant oop pointers (such as Strings); they alias with 1445 // unknown strings. 1446 assert(!is_known_inst, "not scalarizable allocation"); 1447 tj = to = to-> 1448 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1449 remove_speculative()-> 1450 cast_to_ptr_type(TypePtr::BotPTR)-> 1451 cast_to_exactness(false); 1452 } 1453 } else if( is_known_inst ) { 1454 tj = to; // Keep NotNull and klass_is_exact for instance type 1455 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1456 // During the 2nd round of IterGVN, NotNull castings are removed. 1457 // Make sure the Bottom and NotNull variants alias the same. 1458 // Also, make sure exact and non-exact variants alias the same. 1459 tj = to = to-> 1460 remove_speculative()-> 1461 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1462 cast_to_ptr_type(TypePtr::BotPTR)-> 1463 cast_to_exactness(false); 1464 } 1465 if (to->speculative() != nullptr) { 1466 tj = to = to->remove_speculative(); 1467 } 1468 // Canonicalize the holder of this field 1469 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1470 // First handle header references such as a LoadKlassNode, even if the 1471 // object's klass is unloaded at compile time (4965979). 1472 if (!is_known_inst) { // Do it only for non-instance types 1473 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset); 1474 } 1475 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) { 1476 // Static fields are in the space above the normal instance 1477 // fields in the java.lang.Class instance. 1478 if (ik != ciEnv::current()->Class_klass()) { 1479 to = nullptr; 1480 tj = TypeOopPtr::BOTTOM; 1481 offset = tj->offset(); 1482 } 1483 } else { 1484 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset); 1485 assert(offset < canonical_holder->layout_helper_size_in_bytes(), ""); 1486 assert(tj->offset() == offset, "no change to offset expected"); 1487 bool xk = to->klass_is_exact(); 1488 int instance_id = to->instance_id(); 1489 1490 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder 1491 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only 1492 // its interfaces are included. 1493 if (xk && ik->equals(canonical_holder)) { 1494 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type"); 1495 } else { 1496 assert(xk || !is_known_inst, "Known instance should be exact type"); 1497 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id); 1498 } 1499 } 1500 } 1501 1502 // Klass pointers to object array klasses need some flattening 1503 const TypeKlassPtr *tk = tj->isa_klassptr(); 1504 if( tk ) { 1505 // If we are referencing a field within a Klass, we need 1506 // to assume the worst case of an Object. Both exact and 1507 // inexact types must flatten to the same alias class so 1508 // use NotNull as the PTR. 1509 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1510 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, 1511 env()->Object_klass(), 1512 offset); 1513 } 1514 1515 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) { 1516 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass()); 1517 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs 1518 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset); 1519 } else { 1520 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset); 1521 } 1522 } 1523 1524 // Check for precise loads from the primary supertype array and force them 1525 // to the supertype cache alias index. Check for generic array loads from 1526 // the primary supertype array and also force them to the supertype cache 1527 // alias index. Since the same load can reach both, we need to merge 1528 // these 2 disparate memories into the same alias class. Since the 1529 // primary supertype array is read-only, there's no chance of confusion 1530 // where we bypass an array load and an array store. 1531 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1532 if (offset == Type::OffsetBot || 1533 (offset >= primary_supers_offset && 1534 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1535 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1536 offset = in_bytes(Klass::secondary_super_cache_offset()); 1537 tj = tk = tk->with_offset(offset); 1538 } 1539 } 1540 1541 // Flatten all Raw pointers together. 1542 if (tj->base() == Type::RawPtr) 1543 tj = TypeRawPtr::BOTTOM; 1544 1545 if (tj->base() == Type::AnyPtr) 1546 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1547 1548 offset = tj->offset(); 1549 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1550 1551 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1552 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1553 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1554 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1555 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1556 (offset == Type::klass_offset() && tj->base() == Type::AryPtr) || 1557 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1558 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1559 assert( tj->ptr() != TypePtr::TopPTR && 1560 tj->ptr() != TypePtr::AnyNull && 1561 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1562 // assert( tj->ptr() != TypePtr::Constant || 1563 // tj->base() == Type::RawPtr || 1564 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1565 1566 return tj; 1567 } 1568 1569 void Compile::AliasType::Init(int i, const TypePtr* at) { 1570 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1571 _index = i; 1572 _adr_type = at; 1573 _field = nullptr; 1574 _element = nullptr; 1575 _is_rewritable = true; // default 1576 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr; 1577 if (atoop != nullptr && atoop->is_known_instance()) { 1578 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1579 _general_index = Compile::current()->get_alias_index(gt); 1580 } else { 1581 _general_index = 0; 1582 } 1583 } 1584 1585 BasicType Compile::AliasType::basic_type() const { 1586 if (element() != nullptr) { 1587 const Type* element = adr_type()->is_aryptr()->elem(); 1588 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1589 } if (field() != nullptr) { 1590 return field()->layout_type(); 1591 } else { 1592 return T_ILLEGAL; // unknown 1593 } 1594 } 1595 1596 //---------------------------------print_on------------------------------------ 1597 #ifndef PRODUCT 1598 void Compile::AliasType::print_on(outputStream* st) { 1599 if (index() < 10) 1600 st->print("@ <%d> ", index()); 1601 else st->print("@ <%d>", index()); 1602 st->print(is_rewritable() ? " " : " RO"); 1603 int offset = adr_type()->offset(); 1604 if (offset == Type::OffsetBot) 1605 st->print(" +any"); 1606 else st->print(" +%-3d", offset); 1607 st->print(" in "); 1608 adr_type()->dump_on(st); 1609 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1610 if (field() != nullptr && tjp) { 1611 if (tjp->is_instptr()->instance_klass() != field()->holder() || 1612 tjp->offset() != field()->offset_in_bytes()) { 1613 st->print(" != "); 1614 field()->print(); 1615 st->print(" ***"); 1616 } 1617 } 1618 } 1619 1620 void print_alias_types() { 1621 Compile* C = Compile::current(); 1622 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1623 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1624 C->alias_type(idx)->print_on(tty); 1625 tty->cr(); 1626 } 1627 } 1628 #endif 1629 1630 1631 //----------------------------probe_alias_cache-------------------------------- 1632 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1633 intptr_t key = (intptr_t) adr_type; 1634 key ^= key >> logAliasCacheSize; 1635 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1636 } 1637 1638 1639 //-----------------------------grow_alias_types-------------------------------- 1640 void Compile::grow_alias_types() { 1641 const int old_ats = _max_alias_types; // how many before? 1642 const int new_ats = old_ats; // how many more? 1643 const int grow_ats = old_ats+new_ats; // how many now? 1644 _max_alias_types = grow_ats; 1645 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1646 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1647 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1648 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1649 } 1650 1651 1652 //--------------------------------find_alias_type------------------------------ 1653 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1654 if (!do_aliasing()) { 1655 return alias_type(AliasIdxBot); 1656 } 1657 1658 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1659 if (ace->_adr_type == adr_type) { 1660 return alias_type(ace->_index); 1661 } 1662 1663 // Handle special cases. 1664 if (adr_type == nullptr) return alias_type(AliasIdxTop); 1665 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1666 1667 // Do it the slow way. 1668 const TypePtr* flat = flatten_alias_type(adr_type); 1669 1670 #ifdef ASSERT 1671 { 1672 ResourceMark rm; 1673 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1674 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1675 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1676 Type::str(adr_type)); 1677 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1678 const TypeOopPtr* foop = flat->is_oopptr(); 1679 // Scalarizable allocations have exact klass always. 1680 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1681 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1682 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1683 Type::str(foop), Type::str(xoop)); 1684 } 1685 } 1686 #endif 1687 1688 int idx = AliasIdxTop; 1689 for (int i = 0; i < num_alias_types(); i++) { 1690 if (alias_type(i)->adr_type() == flat) { 1691 idx = i; 1692 break; 1693 } 1694 } 1695 1696 if (idx == AliasIdxTop) { 1697 if (no_create) return nullptr; 1698 // Grow the array if necessary. 1699 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1700 // Add a new alias type. 1701 idx = _num_alias_types++; 1702 _alias_types[idx]->Init(idx, flat); 1703 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1704 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1705 if (flat->isa_instptr()) { 1706 if (flat->offset() == java_lang_Class::klass_offset() 1707 && flat->is_instptr()->instance_klass() == env()->Class_klass()) 1708 alias_type(idx)->set_rewritable(false); 1709 } 1710 if (flat->isa_aryptr()) { 1711 #ifdef ASSERT 1712 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1713 // (T_BYTE has the weakest alignment and size restrictions...) 1714 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1715 #endif 1716 if (flat->offset() == TypePtr::OffsetBot) { 1717 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1718 } 1719 } 1720 if (flat->isa_klassptr()) { 1721 if (UseCompactObjectHeaders) { 1722 if (flat->offset() == in_bytes(Klass::prototype_header_offset())) 1723 alias_type(idx)->set_rewritable(false); 1724 } 1725 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1726 alias_type(idx)->set_rewritable(false); 1727 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1728 alias_type(idx)->set_rewritable(false); 1729 if (flat->offset() == in_bytes(Klass::misc_flags_offset())) 1730 alias_type(idx)->set_rewritable(false); 1731 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1732 alias_type(idx)->set_rewritable(false); 1733 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1734 alias_type(idx)->set_rewritable(false); 1735 } 1736 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1737 // but the base pointer type is not distinctive enough to identify 1738 // references into JavaThread.) 1739 1740 // Check for final fields. 1741 const TypeInstPtr* tinst = flat->isa_instptr(); 1742 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1743 ciField* field; 1744 if (tinst->const_oop() != nullptr && 1745 tinst->instance_klass() == ciEnv::current()->Class_klass() && 1746 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) { 1747 // static field 1748 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1749 field = k->get_field_by_offset(tinst->offset(), true); 1750 } else { 1751 ciInstanceKlass *k = tinst->instance_klass(); 1752 field = k->get_field_by_offset(tinst->offset(), false); 1753 } 1754 assert(field == nullptr || 1755 original_field == nullptr || 1756 (field->holder() == original_field->holder() && 1757 field->offset_in_bytes() == original_field->offset_in_bytes() && 1758 field->is_static() == original_field->is_static()), "wrong field?"); 1759 // Set field() and is_rewritable() attributes. 1760 if (field != nullptr) alias_type(idx)->set_field(field); 1761 } 1762 } 1763 1764 // Fill the cache for next time. 1765 ace->_adr_type = adr_type; 1766 ace->_index = idx; 1767 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1768 1769 // Might as well try to fill the cache for the flattened version, too. 1770 AliasCacheEntry* face = probe_alias_cache(flat); 1771 if (face->_adr_type == nullptr) { 1772 face->_adr_type = flat; 1773 face->_index = idx; 1774 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1775 } 1776 1777 return alias_type(idx); 1778 } 1779 1780 1781 Compile::AliasType* Compile::alias_type(ciField* field) { 1782 const TypeOopPtr* t; 1783 if (field->is_static()) 1784 t = TypeInstPtr::make(field->holder()->java_mirror()); 1785 else 1786 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1787 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1788 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1789 return atp; 1790 } 1791 1792 1793 //------------------------------have_alias_type-------------------------------- 1794 bool Compile::have_alias_type(const TypePtr* adr_type) { 1795 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1796 if (ace->_adr_type == adr_type) { 1797 return true; 1798 } 1799 1800 // Handle special cases. 1801 if (adr_type == nullptr) return true; 1802 if (adr_type == TypePtr::BOTTOM) return true; 1803 1804 return find_alias_type(adr_type, true, nullptr) != nullptr; 1805 } 1806 1807 //-----------------------------must_alias-------------------------------------- 1808 // True if all values of the given address type are in the given alias category. 1809 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1810 if (alias_idx == AliasIdxBot) return true; // the universal category 1811 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP 1812 if (alias_idx == AliasIdxTop) return false; // the empty category 1813 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1814 1815 // the only remaining possible overlap is identity 1816 int adr_idx = get_alias_index(adr_type); 1817 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1818 assert(adr_idx == alias_idx || 1819 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1820 && adr_type != TypeOopPtr::BOTTOM), 1821 "should not be testing for overlap with an unsafe pointer"); 1822 return adr_idx == alias_idx; 1823 } 1824 1825 //------------------------------can_alias-------------------------------------- 1826 // True if any values of the given address type are in the given alias category. 1827 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1828 if (alias_idx == AliasIdxTop) return false; // the empty category 1829 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP 1830 // Known instance doesn't alias with bottom memory 1831 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1832 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1833 1834 // the only remaining possible overlap is identity 1835 int adr_idx = get_alias_index(adr_type); 1836 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1837 return adr_idx == alias_idx; 1838 } 1839 1840 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their 1841 // uncommon traps if no Runtime Predicates were created from the Parse Predicates. 1842 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) { 1843 if (parse_predicate_count() == 0) { 1844 return; 1845 } 1846 for (int i = 0; i < parse_predicate_count(); i++) { 1847 ParsePredicateNode* parse_predicate = _parse_predicates.at(i); 1848 parse_predicate->mark_useless(igvn); 1849 } 1850 _parse_predicates.clear(); 1851 } 1852 1853 void Compile::record_for_post_loop_opts_igvn(Node* n) { 1854 if (!n->for_post_loop_opts_igvn()) { 1855 assert(!_for_post_loop_igvn.contains(n), "duplicate"); 1856 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1857 _for_post_loop_igvn.append(n); 1858 } 1859 } 1860 1861 void Compile::remove_from_post_loop_opts_igvn(Node* n) { 1862 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1863 _for_post_loop_igvn.remove(n); 1864 } 1865 1866 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) { 1867 // Verify that all previous optimizations produced a valid graph 1868 // at least to this point, even if no loop optimizations were done. 1869 PhaseIdealLoop::verify(igvn); 1870 1871 if (has_loops() || _loop_opts_cnt > 0) { 1872 print_method(PHASE_AFTER_LOOP_OPTS, 2); 1873 } 1874 C->set_post_loop_opts_phase(); // no more loop opts allowed 1875 1876 assert(!C->major_progress(), "not cleared"); 1877 1878 if (_for_post_loop_igvn.length() > 0) { 1879 while (_for_post_loop_igvn.length() > 0) { 1880 Node* n = _for_post_loop_igvn.pop(); 1881 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1882 igvn._worklist.push(n); 1883 } 1884 igvn.optimize(); 1885 if (failing()) return; 1886 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed"); 1887 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now"); 1888 1889 // Sometimes IGVN sets major progress (e.g., when processing loop nodes). 1890 if (C->major_progress()) { 1891 C->clear_major_progress(); // ensure that major progress is now clear 1892 } 1893 } 1894 } 1895 1896 void Compile::record_for_merge_stores_igvn(Node* n) { 1897 if (!n->for_merge_stores_igvn()) { 1898 assert(!_for_merge_stores_igvn.contains(n), "duplicate"); 1899 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn); 1900 _for_merge_stores_igvn.append(n); 1901 } 1902 } 1903 1904 void Compile::remove_from_merge_stores_igvn(Node* n) { 1905 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn); 1906 _for_merge_stores_igvn.remove(n); 1907 } 1908 1909 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during 1910 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between 1911 // the stores, and we merge the wrong sequence of stores. 1912 // Example: 1913 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI 1914 // Apply MergeStores: 1915 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI 1916 // Remove more RangeChecks: 1917 // StoreI [ StoreL ] StoreI 1918 // But now it would have been better to do this instead: 1919 // [ StoreL ] [ StoreL ] 1920 // 1921 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round, 1922 // since we never unset _merge_stores_phase. 1923 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) { 1924 C->set_merge_stores_phase(); 1925 1926 if (_for_merge_stores_igvn.length() > 0) { 1927 while (_for_merge_stores_igvn.length() > 0) { 1928 Node* n = _for_merge_stores_igvn.pop(); 1929 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn); 1930 igvn._worklist.push(n); 1931 } 1932 igvn.optimize(); 1933 if (failing()) return; 1934 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed"); 1935 print_method(PHASE_AFTER_MERGE_STORES, 3); 1936 } 1937 } 1938 1939 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) { 1940 if (OptimizeUnstableIf) { 1941 _unstable_if_traps.append(trap); 1942 } 1943 } 1944 1945 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) { 1946 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) { 1947 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1948 Node* n = trap->uncommon_trap(); 1949 if (!useful.member(n)) { 1950 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 1951 } 1952 } 1953 } 1954 1955 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead 1956 // or fold-compares case. Return true if succeed or not found. 1957 // 1958 // In rare cases, the found trap has been processed. It is too late to delete it. Return 1959 // false and ask fold-compares to yield. 1960 // 1961 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused 1962 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path 1963 // when deoptimization does happen. 1964 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) { 1965 for (int i = 0; i < _unstable_if_traps.length(); ++i) { 1966 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1967 if (trap->uncommon_trap() == unc) { 1968 if (yield && trap->modified()) { 1969 return false; 1970 } 1971 _unstable_if_traps.delete_at(i); 1972 break; 1973 } 1974 } 1975 return true; 1976 } 1977 1978 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path. 1979 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering. 1980 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) { 1981 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) { 1982 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1983 CallStaticJavaNode* unc = trap->uncommon_trap(); 1984 int next_bci = trap->next_bci(); 1985 bool modified = trap->modified(); 1986 1987 if (next_bci != -1 && !modified) { 1988 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!"); 1989 JVMState* jvms = unc->jvms(); 1990 ciMethod* method = jvms->method(); 1991 ciBytecodeStream iter(method); 1992 1993 iter.force_bci(jvms->bci()); 1994 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if"); 1995 Bytecodes::Code c = iter.cur_bc(); 1996 Node* lhs = nullptr; 1997 Node* rhs = nullptr; 1998 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) { 1999 lhs = unc->peek_operand(0); 2000 rhs = unc->peek_operand(1); 2001 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) { 2002 lhs = unc->peek_operand(0); 2003 } 2004 2005 ResourceMark rm; 2006 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci); 2007 assert(live_locals.is_valid(), "broken liveness info"); 2008 int len = (int)live_locals.size(); 2009 2010 for (int i = 0; i < len; i++) { 2011 Node* local = unc->local(jvms, i); 2012 // kill local using the liveness of next_bci. 2013 // give up when the local looks like an operand to secure reexecution. 2014 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) { 2015 uint idx = jvms->locoff() + i; 2016 #ifdef ASSERT 2017 if (PrintOpto && Verbose) { 2018 tty->print("[unstable_if] kill local#%d: ", idx); 2019 local->dump(); 2020 tty->cr(); 2021 } 2022 #endif 2023 igvn.replace_input_of(unc, idx, top()); 2024 modified = true; 2025 } 2026 } 2027 } 2028 2029 // keep the mondified trap for late query 2030 if (modified) { 2031 trap->set_modified(); 2032 } else { 2033 _unstable_if_traps.delete_at(i); 2034 } 2035 } 2036 igvn.optimize(); 2037 } 2038 2039 // StringOpts and late inlining of string methods 2040 void Compile::inline_string_calls(bool parse_time) { 2041 { 2042 // remove useless nodes to make the usage analysis simpler 2043 ResourceMark rm; 2044 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2045 } 2046 2047 { 2048 ResourceMark rm; 2049 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2050 PhaseStringOpts pso(initial_gvn()); 2051 print_method(PHASE_AFTER_STRINGOPTS, 3); 2052 } 2053 2054 // now inline anything that we skipped the first time around 2055 if (!parse_time) { 2056 _late_inlines_pos = _late_inlines.length(); 2057 } 2058 2059 while (_string_late_inlines.length() > 0) { 2060 CallGenerator* cg = _string_late_inlines.pop(); 2061 cg->do_late_inline(); 2062 if (failing()) return; 2063 } 2064 _string_late_inlines.trunc_to(0); 2065 } 2066 2067 // Late inlining of boxing methods 2068 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2069 if (_boxing_late_inlines.length() > 0) { 2070 assert(has_boxed_value(), "inconsistent"); 2071 2072 set_inlining_incrementally(true); 2073 2074 igvn_worklist()->ensure_empty(); // should be done with igvn 2075 2076 _late_inlines_pos = _late_inlines.length(); 2077 2078 while (_boxing_late_inlines.length() > 0) { 2079 CallGenerator* cg = _boxing_late_inlines.pop(); 2080 cg->do_late_inline(); 2081 if (failing()) return; 2082 } 2083 _boxing_late_inlines.trunc_to(0); 2084 2085 inline_incrementally_cleanup(igvn); 2086 2087 set_inlining_incrementally(false); 2088 } 2089 } 2090 2091 bool Compile::inline_incrementally_one() { 2092 assert(IncrementalInline, "incremental inlining should be on"); 2093 2094 TracePhase tp(_t_incrInline_inline); 2095 2096 set_inlining_progress(false); 2097 set_do_cleanup(false); 2098 2099 for (int i = 0; i < _late_inlines.length(); i++) { 2100 _late_inlines_pos = i+1; 2101 CallGenerator* cg = _late_inlines.at(i); 2102 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node()); 2103 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline(); 2104 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call 2105 cg->do_late_inline(); 2106 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed"); 2107 if (failing()) { 2108 return false; 2109 } else if (inlining_progress()) { 2110 _late_inlines_pos = i+1; // restore the position in case new elements were inserted 2111 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node()); 2112 break; // process one call site at a time 2113 } else { 2114 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node()); 2115 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) { 2116 // Avoid potential infinite loop if node already in the IGVN list 2117 assert(false, "scheduled for IGVN during inlining attempt"); 2118 } else { 2119 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt 2120 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass"); 2121 cg->call_node()->set_generator(cg); 2122 } 2123 } 2124 } else { 2125 // Ignore late inline direct calls when inlining is not allowed. 2126 // They are left in the late inline list when node budget is exhausted until the list is fully drained. 2127 } 2128 } 2129 // Remove processed elements. 2130 _late_inlines.remove_till(_late_inlines_pos); 2131 _late_inlines_pos = 0; 2132 2133 assert(inlining_progress() || _late_inlines.length() == 0, "no progress"); 2134 2135 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2136 2137 set_inlining_progress(false); 2138 set_do_cleanup(false); 2139 2140 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption; 2141 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup; 2142 } 2143 2144 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2145 { 2146 TracePhase tp(_t_incrInline_pru); 2147 ResourceMark rm; 2148 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2149 } 2150 { 2151 TracePhase tp(_t_incrInline_igvn); 2152 igvn.reset_from_gvn(initial_gvn()); 2153 igvn.optimize(); 2154 if (failing()) return; 2155 } 2156 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3); 2157 } 2158 2159 // Perform incremental inlining until bound on number of live nodes is reached 2160 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2161 TracePhase tp(_t_incrInline); 2162 2163 set_inlining_incrementally(true); 2164 uint low_live_nodes = 0; 2165 2166 while (_late_inlines.length() > 0) { 2167 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2168 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2169 TracePhase tp(_t_incrInline_ideal); 2170 // PhaseIdealLoop is expensive so we only try it once we are 2171 // out of live nodes and we only try it again if the previous 2172 // helped got the number of nodes down significantly 2173 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2174 if (failing()) return; 2175 low_live_nodes = live_nodes(); 2176 _major_progress = true; 2177 } 2178 2179 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2180 bool do_print_inlining = print_inlining() || print_intrinsics(); 2181 if (do_print_inlining || log() != nullptr) { 2182 // Print inlining message for candidates that we couldn't inline for lack of space. 2183 for (int i = 0; i < _late_inlines.length(); i++) { 2184 CallGenerator* cg = _late_inlines.at(i); 2185 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 2186 if (do_print_inlining) { 2187 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg); 2188 } 2189 log_late_inline_failure(cg, msg); 2190 } 2191 } 2192 break; // finish 2193 } 2194 } 2195 2196 igvn_worklist()->ensure_empty(); // should be done with igvn 2197 2198 while (inline_incrementally_one()) { 2199 assert(!failing_internal() || failure_is_artificial(), "inconsistent"); 2200 } 2201 if (failing()) return; 2202 2203 inline_incrementally_cleanup(igvn); 2204 2205 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 2206 2207 if (failing()) return; 2208 2209 if (_late_inlines.length() == 0) { 2210 break; // no more progress 2211 } 2212 } 2213 2214 igvn_worklist()->ensure_empty(); // should be done with igvn 2215 2216 if (_string_late_inlines.length() > 0) { 2217 assert(has_stringbuilder(), "inconsistent"); 2218 2219 inline_string_calls(false); 2220 2221 if (failing()) return; 2222 2223 inline_incrementally_cleanup(igvn); 2224 } 2225 2226 set_inlining_incrementally(false); 2227 } 2228 2229 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) { 2230 // "inlining_incrementally() == false" is used to signal that no inlining is allowed 2231 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details). 2232 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr" 2233 // as if "inlining_incrementally() == true" were set. 2234 assert(inlining_incrementally() == false, "not allowed"); 2235 assert(_modified_nodes == nullptr, "not allowed"); 2236 assert(_late_inlines.length() > 0, "sanity"); 2237 2238 while (_late_inlines.length() > 0) { 2239 igvn_worklist()->ensure_empty(); // should be done with igvn 2240 2241 while (inline_incrementally_one()) { 2242 assert(!failing_internal() || failure_is_artificial(), "inconsistent"); 2243 } 2244 if (failing()) return; 2245 2246 inline_incrementally_cleanup(igvn); 2247 } 2248 } 2249 2250 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2251 if (_loop_opts_cnt > 0) { 2252 while (major_progress() && (_loop_opts_cnt > 0)) { 2253 TracePhase tp(_t_idealLoop); 2254 PhaseIdealLoop::optimize(igvn, mode); 2255 _loop_opts_cnt--; 2256 if (failing()) return false; 2257 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2258 } 2259 } 2260 return true; 2261 } 2262 2263 // Remove edges from "root" to each SafePoint at a backward branch. 2264 // They were inserted during parsing (see add_safepoint()) to make 2265 // infinite loops without calls or exceptions visible to root, i.e., 2266 // useful. 2267 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2268 Node *r = root(); 2269 if (r != nullptr) { 2270 for (uint i = r->req(); i < r->len(); ++i) { 2271 Node *n = r->in(i); 2272 if (n != nullptr && n->is_SafePoint()) { 2273 r->rm_prec(i); 2274 if (n->outcnt() == 0) { 2275 igvn.remove_dead_node(n); 2276 } 2277 --i; 2278 } 2279 } 2280 // Parsing may have added top inputs to the root node (Path 2281 // leading to the Halt node proven dead). Make sure we get a 2282 // chance to clean them up. 2283 igvn._worklist.push(r); 2284 igvn.optimize(); 2285 } 2286 } 2287 2288 //------------------------------Optimize--------------------------------------- 2289 // Given a graph, optimize it. 2290 void Compile::Optimize() { 2291 TracePhase tp(_t_optimizer); 2292 2293 #ifndef PRODUCT 2294 if (env()->break_at_compile()) { 2295 BREAKPOINT; 2296 } 2297 2298 #endif 2299 2300 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2301 #ifdef ASSERT 2302 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2303 #endif 2304 2305 ResourceMark rm; 2306 2307 NOT_PRODUCT( verify_graph_edges(); ) 2308 2309 print_method(PHASE_AFTER_PARSING, 1); 2310 2311 { 2312 // Iterative Global Value Numbering, including ideal transforms 2313 // Initialize IterGVN with types and values from parse-time GVN 2314 PhaseIterGVN igvn(initial_gvn()); 2315 #ifdef ASSERT 2316 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2317 #endif 2318 { 2319 TracePhase tp(_t_iterGVN); 2320 igvn.optimize(); 2321 } 2322 2323 if (failing()) return; 2324 2325 print_method(PHASE_ITER_GVN1, 2); 2326 2327 process_for_unstable_if_traps(igvn); 2328 2329 if (failing()) return; 2330 2331 inline_incrementally(igvn); 2332 2333 print_method(PHASE_INCREMENTAL_INLINE, 2); 2334 2335 if (failing()) return; 2336 2337 if (eliminate_boxing()) { 2338 // Inline valueOf() methods now. 2339 inline_boxing_calls(igvn); 2340 2341 if (failing()) return; 2342 2343 if (AlwaysIncrementalInline || StressIncrementalInlining) { 2344 inline_incrementally(igvn); 2345 } 2346 2347 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2348 2349 if (failing()) return; 2350 } 2351 2352 // Remove the speculative part of types and clean up the graph from 2353 // the extra CastPP nodes whose only purpose is to carry them. Do 2354 // that early so that optimizations are not disrupted by the extra 2355 // CastPP nodes. 2356 remove_speculative_types(igvn); 2357 2358 if (failing()) return; 2359 2360 // No more new expensive nodes will be added to the list from here 2361 // so keep only the actual candidates for optimizations. 2362 cleanup_expensive_nodes(igvn); 2363 2364 if (failing()) return; 2365 2366 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2367 if (EnableVectorSupport && has_vbox_nodes()) { 2368 TracePhase tp(_t_vector); 2369 PhaseVector pv(igvn); 2370 pv.optimize_vector_boxes(); 2371 if (failing()) return; 2372 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2373 } 2374 assert(!has_vbox_nodes(), "sanity"); 2375 2376 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2377 Compile::TracePhase tp(_t_renumberLive); 2378 igvn_worklist()->ensure_empty(); // should be done with igvn 2379 { 2380 ResourceMark rm; 2381 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist()); 2382 } 2383 igvn.reset_from_gvn(initial_gvn()); 2384 igvn.optimize(); 2385 if (failing()) return; 2386 } 2387 2388 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop 2389 // safepoints 2390 remove_root_to_sfpts_edges(igvn); 2391 2392 if (failing()) return; 2393 2394 if (has_loops()) { 2395 print_method(PHASE_BEFORE_LOOP_OPTS, 2); 2396 } 2397 2398 // Perform escape analysis 2399 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) { 2400 if (has_loops()) { 2401 // Cleanup graph (remove dead nodes). 2402 TracePhase tp(_t_idealLoop); 2403 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2404 if (failing()) return; 2405 } 2406 bool progress; 2407 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2408 do { 2409 ConnectionGraph::do_analysis(this, &igvn); 2410 2411 if (failing()) return; 2412 2413 int mcount = macro_count(); // Record number of allocations and locks before IGVN 2414 2415 // Optimize out fields loads from scalar replaceable allocations. 2416 igvn.optimize(); 2417 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2418 2419 if (failing()) return; 2420 2421 if (congraph() != nullptr && macro_count() > 0) { 2422 TracePhase tp(_t_macroEliminate); 2423 PhaseMacroExpand mexp(igvn); 2424 mexp.eliminate_macro_nodes(); 2425 if (failing()) return; 2426 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2); 2427 2428 igvn.set_delay_transform(false); 2429 igvn.optimize(); 2430 if (failing()) return; 2431 2432 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2433 } 2434 2435 ConnectionGraph::verify_ram_nodes(this, root()); 2436 if (failing()) return; 2437 2438 progress = do_iterative_escape_analysis() && 2439 (macro_count() < mcount) && 2440 ConnectionGraph::has_candidates(this); 2441 // Try again if candidates exist and made progress 2442 // by removing some allocations and/or locks. 2443 } while (progress); 2444 } 2445 2446 // Loop transforms on the ideal graph. Range Check Elimination, 2447 // peeling, unrolling, etc. 2448 2449 // Set loop opts counter 2450 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2451 { 2452 TracePhase tp(_t_idealLoop); 2453 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2454 _loop_opts_cnt--; 2455 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2456 if (failing()) return; 2457 } 2458 // Loop opts pass if partial peeling occurred in previous pass 2459 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2460 TracePhase tp(_t_idealLoop); 2461 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2462 _loop_opts_cnt--; 2463 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2464 if (failing()) return; 2465 } 2466 // Loop opts pass for loop-unrolling before CCP 2467 if(major_progress() && (_loop_opts_cnt > 0)) { 2468 TracePhase tp(_t_idealLoop); 2469 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2470 _loop_opts_cnt--; 2471 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2472 } 2473 if (!failing()) { 2474 // Verify that last round of loop opts produced a valid graph 2475 PhaseIdealLoop::verify(igvn); 2476 } 2477 } 2478 if (failing()) return; 2479 2480 // Conditional Constant Propagation; 2481 print_method(PHASE_BEFORE_CCP1, 2); 2482 PhaseCCP ccp( &igvn ); 2483 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2484 { 2485 TracePhase tp(_t_ccp); 2486 ccp.do_transform(); 2487 } 2488 print_method(PHASE_CCP1, 2); 2489 2490 assert( true, "Break here to ccp.dump_old2new_map()"); 2491 2492 // Iterative Global Value Numbering, including ideal transforms 2493 { 2494 TracePhase tp(_t_iterGVN2); 2495 igvn.reset_from_igvn(&ccp); 2496 igvn.optimize(); 2497 } 2498 print_method(PHASE_ITER_GVN2, 2); 2499 2500 if (failing()) return; 2501 2502 // Loop transforms on the ideal graph. Range Check Elimination, 2503 // peeling, unrolling, etc. 2504 if (!optimize_loops(igvn, LoopOptsDefault)) { 2505 return; 2506 } 2507 2508 if (failing()) return; 2509 2510 C->clear_major_progress(); // ensure that major progress is now clear 2511 2512 process_for_post_loop_opts_igvn(igvn); 2513 2514 process_for_merge_stores_igvn(igvn); 2515 2516 if (failing()) return; 2517 2518 #ifdef ASSERT 2519 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2520 #endif 2521 2522 { 2523 TracePhase tp(_t_macroExpand); 2524 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3); 2525 PhaseMacroExpand mex(igvn); 2526 // Do not allow new macro nodes once we start to eliminate and expand 2527 C->reset_allow_macro_nodes(); 2528 // Last attempt to eliminate macro nodes before expand 2529 mex.eliminate_macro_nodes(); 2530 if (failing()) { 2531 return; 2532 } 2533 mex.eliminate_opaque_looplimit_macro_nodes(); 2534 if (failing()) { 2535 return; 2536 } 2537 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2); 2538 if (mex.expand_macro_nodes()) { 2539 assert(failing(), "must bail out w/ explicit message"); 2540 return; 2541 } 2542 print_method(PHASE_AFTER_MACRO_EXPANSION, 2); 2543 } 2544 2545 { 2546 TracePhase tp(_t_barrierExpand); 2547 if (bs->expand_barriers(this, igvn)) { 2548 assert(failing(), "must bail out w/ explicit message"); 2549 return; 2550 } 2551 print_method(PHASE_BARRIER_EXPANSION, 2); 2552 } 2553 2554 if (C->max_vector_size() > 0) { 2555 C->optimize_logic_cones(igvn); 2556 igvn.optimize(); 2557 if (failing()) return; 2558 } 2559 2560 DEBUG_ONLY( _modified_nodes = nullptr; ) 2561 2562 assert(igvn._worklist.size() == 0, "not empty"); 2563 2564 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty"); 2565 2566 if (_late_inlines.length() > 0) { 2567 // More opportunities to optimize virtual and MH calls. 2568 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option. 2569 process_late_inline_calls_no_inline(igvn); 2570 if (failing()) return; 2571 } 2572 } // (End scope of igvn; run destructor if necessary for asserts.) 2573 2574 check_no_dead_use(); 2575 2576 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have 2577 // to remove hashes to unlock nodes for modifications. 2578 C->node_hash()->clear(); 2579 2580 // A method with only infinite loops has no edges entering loops from root 2581 { 2582 TracePhase tp(_t_graphReshaping); 2583 if (final_graph_reshaping()) { 2584 assert(failing(), "must bail out w/ explicit message"); 2585 return; 2586 } 2587 } 2588 2589 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2590 DEBUG_ONLY(set_phase_optimize_finished();) 2591 } 2592 2593 #ifdef ASSERT 2594 void Compile::check_no_dead_use() const { 2595 ResourceMark rm; 2596 Unique_Node_List wq; 2597 wq.push(root()); 2598 for (uint i = 0; i < wq.size(); ++i) { 2599 Node* n = wq.at(i); 2600 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2601 Node* u = n->fast_out(j); 2602 if (u->outcnt() == 0 && !u->is_Con()) { 2603 u->dump(); 2604 fatal("no reachable node should have no use"); 2605 } 2606 wq.push(u); 2607 } 2608 } 2609 } 2610 #endif 2611 2612 void Compile::inline_vector_reboxing_calls() { 2613 if (C->_vector_reboxing_late_inlines.length() > 0) { 2614 _late_inlines_pos = C->_late_inlines.length(); 2615 while (_vector_reboxing_late_inlines.length() > 0) { 2616 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 2617 cg->do_late_inline(); 2618 if (failing()) return; 2619 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node()); 2620 } 2621 _vector_reboxing_late_inlines.trunc_to(0); 2622 } 2623 } 2624 2625 bool Compile::has_vbox_nodes() { 2626 if (C->_vector_reboxing_late_inlines.length() > 0) { 2627 return true; 2628 } 2629 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 2630 Node * n = C->macro_node(macro_idx); 2631 assert(n->is_macro(), "only macro nodes expected here"); 2632 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 2633 return true; 2634 } 2635 } 2636 return false; 2637 } 2638 2639 //---------------------------- Bitwise operation packing optimization --------------------------- 2640 2641 static bool is_vector_unary_bitwise_op(Node* n) { 2642 return n->Opcode() == Op_XorV && 2643 VectorNode::is_vector_bitwise_not_pattern(n); 2644 } 2645 2646 static bool is_vector_binary_bitwise_op(Node* n) { 2647 switch (n->Opcode()) { 2648 case Op_AndV: 2649 case Op_OrV: 2650 return true; 2651 2652 case Op_XorV: 2653 return !is_vector_unary_bitwise_op(n); 2654 2655 default: 2656 return false; 2657 } 2658 } 2659 2660 static bool is_vector_ternary_bitwise_op(Node* n) { 2661 return n->Opcode() == Op_MacroLogicV; 2662 } 2663 2664 static bool is_vector_bitwise_op(Node* n) { 2665 return is_vector_unary_bitwise_op(n) || 2666 is_vector_binary_bitwise_op(n) || 2667 is_vector_ternary_bitwise_op(n); 2668 } 2669 2670 static bool is_vector_bitwise_cone_root(Node* n) { 2671 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) { 2672 return false; 2673 } 2674 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2675 if (is_vector_bitwise_op(n->fast_out(i))) { 2676 return false; 2677 } 2678 } 2679 return true; 2680 } 2681 2682 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) { 2683 uint cnt = 0; 2684 if (is_vector_bitwise_op(n)) { 2685 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req(); 2686 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2687 for (uint i = 1; i < inp_cnt; i++) { 2688 Node* in = n->in(i); 2689 bool skip = VectorNode::is_all_ones_vector(in); 2690 if (!skip && !inputs.member(in)) { 2691 inputs.push(in); 2692 cnt++; 2693 } 2694 } 2695 assert(cnt <= 1, "not unary"); 2696 } else { 2697 uint last_req = inp_cnt; 2698 if (is_vector_ternary_bitwise_op(n)) { 2699 last_req = inp_cnt - 1; // skip last input 2700 } 2701 for (uint i = 1; i < last_req; i++) { 2702 Node* def = n->in(i); 2703 if (!inputs.member(def)) { 2704 inputs.push(def); 2705 cnt++; 2706 } 2707 } 2708 } 2709 } else { // not a bitwise operations 2710 if (!inputs.member(n)) { 2711 inputs.push(n); 2712 cnt++; 2713 } 2714 } 2715 return cnt; 2716 } 2717 2718 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 2719 Unique_Node_List useful_nodes; 2720 C->identify_useful_nodes(useful_nodes); 2721 2722 for (uint i = 0; i < useful_nodes.size(); i++) { 2723 Node* n = useful_nodes.at(i); 2724 if (is_vector_bitwise_cone_root(n)) { 2725 list.push(n); 2726 } 2727 } 2728 } 2729 2730 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 2731 const TypeVect* vt, 2732 Unique_Node_List& partition, 2733 Unique_Node_List& inputs) { 2734 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 2735 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 2736 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 2737 2738 Node* in1 = inputs.at(0); 2739 Node* in2 = inputs.at(1); 2740 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 2741 2742 uint func = compute_truth_table(partition, inputs); 2743 2744 Node* pn = partition.at(partition.size() - 1); 2745 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2746 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt)); 2747 } 2748 2749 static uint extract_bit(uint func, uint pos) { 2750 return (func & (1 << pos)) >> pos; 2751 } 2752 2753 // 2754 // A macro logic node represents a truth table. It has 4 inputs, 2755 // First three inputs corresponds to 3 columns of a truth table 2756 // and fourth input captures the logic function. 2757 // 2758 // eg. fn = (in1 AND in2) OR in3; 2759 // 2760 // MacroNode(in1,in2,in3,fn) 2761 // 2762 // ----------------- 2763 // in1 in2 in3 fn 2764 // ----------------- 2765 // 0 0 0 0 2766 // 0 0 1 1 2767 // 0 1 0 0 2768 // 0 1 1 1 2769 // 1 0 0 0 2770 // 1 0 1 1 2771 // 1 1 0 1 2772 // 1 1 1 1 2773 // 2774 2775 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 2776 int res = 0; 2777 for (int i = 0; i < 8; i++) { 2778 int bit1 = extract_bit(in1, i); 2779 int bit2 = extract_bit(in2, i); 2780 int bit3 = extract_bit(in3, i); 2781 2782 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 2783 int func_bit = extract_bit(func, func_bit_pos); 2784 2785 res |= func_bit << i; 2786 } 2787 return res; 2788 } 2789 2790 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 2791 assert(n != nullptr, ""); 2792 assert(eval_map.contains(n), "absent"); 2793 return *(eval_map.get(n)); 2794 } 2795 2796 static void eval_operands(Node* n, 2797 uint& func1, uint& func2, uint& func3, 2798 ResourceHashtable<Node*,uint>& eval_map) { 2799 assert(is_vector_bitwise_op(n), ""); 2800 2801 if (is_vector_unary_bitwise_op(n)) { 2802 Node* opnd = n->in(1); 2803 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 2804 opnd = n->in(2); 2805 } 2806 func1 = eval_operand(opnd, eval_map); 2807 } else if (is_vector_binary_bitwise_op(n)) { 2808 func1 = eval_operand(n->in(1), eval_map); 2809 func2 = eval_operand(n->in(2), eval_map); 2810 } else { 2811 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 2812 func1 = eval_operand(n->in(1), eval_map); 2813 func2 = eval_operand(n->in(2), eval_map); 2814 func3 = eval_operand(n->in(3), eval_map); 2815 } 2816 } 2817 2818 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 2819 assert(inputs.size() <= 3, "sanity"); 2820 ResourceMark rm; 2821 uint res = 0; 2822 ResourceHashtable<Node*,uint> eval_map; 2823 2824 // Populate precomputed functions for inputs. 2825 // Each input corresponds to one column of 3 input truth-table. 2826 uint input_funcs[] = { 0xAA, // (_, _, c) -> c 2827 0xCC, // (_, b, _) -> b 2828 0xF0 }; // (a, _, _) -> a 2829 for (uint i = 0; i < inputs.size(); i++) { 2830 eval_map.put(inputs.at(i), input_funcs[2-i]); 2831 } 2832 2833 for (uint i = 0; i < partition.size(); i++) { 2834 Node* n = partition.at(i); 2835 2836 uint func1 = 0, func2 = 0, func3 = 0; 2837 eval_operands(n, func1, func2, func3, eval_map); 2838 2839 switch (n->Opcode()) { 2840 case Op_OrV: 2841 assert(func3 == 0, "not binary"); 2842 res = func1 | func2; 2843 break; 2844 case Op_AndV: 2845 assert(func3 == 0, "not binary"); 2846 res = func1 & func2; 2847 break; 2848 case Op_XorV: 2849 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2850 assert(func2 == 0 && func3 == 0, "not unary"); 2851 res = (~func1) & 0xFF; 2852 } else { 2853 assert(func3 == 0, "not binary"); 2854 res = func1 ^ func2; 2855 } 2856 break; 2857 case Op_MacroLogicV: 2858 // Ordering of inputs may change during evaluation of sub-tree 2859 // containing MacroLogic node as a child node, thus a re-evaluation 2860 // makes sure that function is evaluated in context of current 2861 // inputs. 2862 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 2863 break; 2864 2865 default: assert(false, "not supported: %s", n->Name()); 2866 } 2867 assert(res <= 0xFF, "invalid"); 2868 eval_map.put(n, res); 2869 } 2870 return res; 2871 } 2872 2873 // Criteria under which nodes gets packed into a macro logic node:- 2874 // 1) Parent and both child nodes are all unmasked or masked with 2875 // same predicates. 2876 // 2) Masked parent can be packed with left child if it is predicated 2877 // and both have same predicates. 2878 // 3) Masked parent can be packed with right child if its un-predicated 2879 // or has matching predication condition. 2880 // 4) An unmasked parent can be packed with an unmasked child. 2881 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2882 assert(partition.size() == 0, "not empty"); 2883 assert(inputs.size() == 0, "not empty"); 2884 if (is_vector_ternary_bitwise_op(n)) { 2885 return false; 2886 } 2887 2888 bool is_unary_op = is_vector_unary_bitwise_op(n); 2889 if (is_unary_op) { 2890 assert(collect_unique_inputs(n, inputs) == 1, "not unary"); 2891 return false; // too few inputs 2892 } 2893 2894 bool pack_left_child = true; 2895 bool pack_right_child = true; 2896 2897 bool left_child_LOP = is_vector_bitwise_op(n->in(1)); 2898 bool right_child_LOP = is_vector_bitwise_op(n->in(2)); 2899 2900 int left_child_input_cnt = 0; 2901 int right_child_input_cnt = 0; 2902 2903 bool parent_is_predicated = n->is_predicated_vector(); 2904 bool left_child_predicated = n->in(1)->is_predicated_vector(); 2905 bool right_child_predicated = n->in(2)->is_predicated_vector(); 2906 2907 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr; 2908 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2909 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2910 2911 do { 2912 if (pack_left_child && left_child_LOP && 2913 ((!parent_is_predicated && !left_child_predicated) || 2914 ((parent_is_predicated && left_child_predicated && 2915 parent_pred == left_child_pred)))) { 2916 partition.push(n->in(1)); 2917 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs); 2918 } else { 2919 inputs.push(n->in(1)); 2920 left_child_input_cnt = 1; 2921 } 2922 2923 if (pack_right_child && right_child_LOP && 2924 (!right_child_predicated || 2925 (right_child_predicated && parent_is_predicated && 2926 parent_pred == right_child_pred))) { 2927 partition.push(n->in(2)); 2928 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs); 2929 } else { 2930 inputs.push(n->in(2)); 2931 right_child_input_cnt = 1; 2932 } 2933 2934 if (inputs.size() > 3) { 2935 assert(partition.size() > 0, ""); 2936 inputs.clear(); 2937 partition.clear(); 2938 if (left_child_input_cnt > right_child_input_cnt) { 2939 pack_left_child = false; 2940 } else { 2941 pack_right_child = false; 2942 } 2943 } else { 2944 break; 2945 } 2946 } while(true); 2947 2948 if(partition.size()) { 2949 partition.push(n); 2950 } 2951 2952 return (partition.size() == 2 || partition.size() == 3) && 2953 (inputs.size() == 2 || inputs.size() == 3); 2954 } 2955 2956 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 2957 assert(is_vector_bitwise_op(n), "not a root"); 2958 2959 visited.set(n->_idx); 2960 2961 // 1) Do a DFS walk over the logic cone. 2962 for (uint i = 1; i < n->req(); i++) { 2963 Node* in = n->in(i); 2964 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 2965 process_logic_cone_root(igvn, in, visited); 2966 } 2967 } 2968 2969 // 2) Bottom up traversal: Merge node[s] with 2970 // the parent to form macro logic node. 2971 Unique_Node_List partition; 2972 Unique_Node_List inputs; 2973 if (compute_logic_cone(n, partition, inputs)) { 2974 const TypeVect* vt = n->bottom_type()->is_vect(); 2975 Node* pn = partition.at(partition.size() - 1); 2976 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2977 if (mask == nullptr || 2978 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) { 2979 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 2980 VectorNode::trace_new_vector(macro_logic, "MacroLogic"); 2981 igvn.replace_node(n, macro_logic); 2982 } 2983 } 2984 } 2985 2986 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 2987 ResourceMark rm; 2988 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 2989 Unique_Node_List list; 2990 collect_logic_cone_roots(list); 2991 2992 while (list.size() > 0) { 2993 Node* n = list.pop(); 2994 const TypeVect* vt = n->bottom_type()->is_vect(); 2995 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 2996 if (supported) { 2997 VectorSet visited(comp_arena()); 2998 process_logic_cone_root(igvn, n, visited); 2999 } 3000 } 3001 } 3002 } 3003 3004 //------------------------------Code_Gen--------------------------------------- 3005 // Given a graph, generate code for it 3006 void Compile::Code_Gen() { 3007 if (failing()) { 3008 return; 3009 } 3010 3011 // Perform instruction selection. You might think we could reclaim Matcher 3012 // memory PDQ, but actually the Matcher is used in generating spill code. 3013 // Internals of the Matcher (including some VectorSets) must remain live 3014 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 3015 // set a bit in reclaimed memory. 3016 3017 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 3018 // nodes. Mapping is only valid at the root of each matched subtree. 3019 NOT_PRODUCT( verify_graph_edges(); ) 3020 3021 Matcher matcher; 3022 _matcher = &matcher; 3023 { 3024 TracePhase tp(_t_matcher); 3025 matcher.match(); 3026 if (failing()) { 3027 return; 3028 } 3029 } 3030 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 3031 // nodes. Mapping is only valid at the root of each matched subtree. 3032 NOT_PRODUCT( verify_graph_edges(); ) 3033 3034 // If you have too many nodes, or if matching has failed, bail out 3035 check_node_count(0, "out of nodes matching instructions"); 3036 if (failing()) { 3037 return; 3038 } 3039 3040 print_method(PHASE_MATCHING, 2); 3041 3042 // Build a proper-looking CFG 3043 PhaseCFG cfg(node_arena(), root(), matcher); 3044 if (failing()) { 3045 return; 3046 } 3047 _cfg = &cfg; 3048 { 3049 TracePhase tp(_t_scheduler); 3050 bool success = cfg.do_global_code_motion(); 3051 if (!success) { 3052 return; 3053 } 3054 3055 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 3056 NOT_PRODUCT( verify_graph_edges(); ) 3057 cfg.verify(); 3058 if (failing()) { 3059 return; 3060 } 3061 } 3062 3063 PhaseChaitin regalloc(unique(), cfg, matcher, false); 3064 _regalloc = ®alloc; 3065 { 3066 TracePhase tp(_t_registerAllocation); 3067 // Perform register allocation. After Chaitin, use-def chains are 3068 // no longer accurate (at spill code) and so must be ignored. 3069 // Node->LRG->reg mappings are still accurate. 3070 _regalloc->Register_Allocate(); 3071 3072 // Bail out if the allocator builds too many nodes 3073 if (failing()) { 3074 return; 3075 } 3076 3077 print_method(PHASE_REGISTER_ALLOCATION, 2); 3078 } 3079 3080 // Prior to register allocation we kept empty basic blocks in case the 3081 // the allocator needed a place to spill. After register allocation we 3082 // are not adding any new instructions. If any basic block is empty, we 3083 // can now safely remove it. 3084 { 3085 TracePhase tp(_t_blockOrdering); 3086 cfg.remove_empty_blocks(); 3087 if (do_freq_based_layout()) { 3088 PhaseBlockLayout layout(cfg); 3089 } else { 3090 cfg.set_loop_alignment(); 3091 } 3092 cfg.fixup_flow(); 3093 cfg.remove_unreachable_blocks(); 3094 cfg.verify_dominator_tree(); 3095 print_method(PHASE_BLOCK_ORDERING, 3); 3096 } 3097 3098 // Apply peephole optimizations 3099 if( OptoPeephole ) { 3100 TracePhase tp(_t_peephole); 3101 PhasePeephole peep( _regalloc, cfg); 3102 peep.do_transform(); 3103 print_method(PHASE_PEEPHOLE, 3); 3104 } 3105 3106 // Do late expand if CPU requires this. 3107 if (Matcher::require_postalloc_expand) { 3108 TracePhase tp(_t_postalloc_expand); 3109 cfg.postalloc_expand(_regalloc); 3110 print_method(PHASE_POSTALLOC_EXPAND, 3); 3111 } 3112 3113 #ifdef ASSERT 3114 { 3115 CompilationMemoryStatistic::do_test_allocations(); 3116 if (failing()) return; 3117 } 3118 #endif 3119 3120 // Convert Nodes to instruction bits in a buffer 3121 { 3122 TracePhase tp(_t_output); 3123 PhaseOutput output; 3124 output.Output(); 3125 if (failing()) return; 3126 output.install(); 3127 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here 3128 } 3129 3130 // He's dead, Jim. 3131 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 3132 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 3133 } 3134 3135 //------------------------------Final_Reshape_Counts--------------------------- 3136 // This class defines counters to help identify when a method 3137 // may/must be executed using hardware with only 24-bit precision. 3138 struct Final_Reshape_Counts : public StackObj { 3139 int _call_count; // count non-inlined 'common' calls 3140 int _float_count; // count float ops requiring 24-bit precision 3141 int _double_count; // count double ops requiring more precision 3142 int _java_call_count; // count non-inlined 'java' calls 3143 int _inner_loop_count; // count loops which need alignment 3144 VectorSet _visited; // Visitation flags 3145 Node_List _tests; // Set of IfNodes & PCTableNodes 3146 3147 Final_Reshape_Counts() : 3148 _call_count(0), _float_count(0), _double_count(0), 3149 _java_call_count(0), _inner_loop_count(0) { } 3150 3151 void inc_call_count () { _call_count ++; } 3152 void inc_float_count () { _float_count ++; } 3153 void inc_double_count() { _double_count++; } 3154 void inc_java_call_count() { _java_call_count++; } 3155 void inc_inner_loop_count() { _inner_loop_count++; } 3156 3157 int get_call_count () const { return _call_count ; } 3158 int get_float_count () const { return _float_count ; } 3159 int get_double_count() const { return _double_count; } 3160 int get_java_call_count() const { return _java_call_count; } 3161 int get_inner_loop_count() const { return _inner_loop_count; } 3162 }; 3163 3164 //------------------------------final_graph_reshaping_impl---------------------- 3165 // Implement items 1-5 from final_graph_reshaping below. 3166 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3167 3168 if ( n->outcnt() == 0 ) return; // dead node 3169 uint nop = n->Opcode(); 3170 3171 // Check for 2-input instruction with "last use" on right input. 3172 // Swap to left input. Implements item (2). 3173 if( n->req() == 3 && // two-input instruction 3174 n->in(1)->outcnt() > 1 && // left use is NOT a last use 3175 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 3176 n->in(2)->outcnt() == 1 &&// right use IS a last use 3177 !n->in(2)->is_Con() ) { // right use is not a constant 3178 // Check for commutative opcode 3179 switch( nop ) { 3180 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 3181 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 3182 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 3183 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 3184 case Op_AndL: case Op_XorL: case Op_OrL: 3185 case Op_AndI: case Op_XorI: case Op_OrI: { 3186 // Move "last use" input to left by swapping inputs 3187 n->swap_edges(1, 2); 3188 break; 3189 } 3190 default: 3191 break; 3192 } 3193 } 3194 3195 #ifdef ASSERT 3196 if( n->is_Mem() ) { 3197 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 3198 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw || 3199 // oop will be recorded in oop map if load crosses safepoint 3200 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 3201 LoadNode::is_immutable_value(n->in(MemNode::Address)))), 3202 "raw memory operations should have control edge"); 3203 } 3204 if (n->is_MemBar()) { 3205 MemBarNode* mb = n->as_MemBar(); 3206 if (mb->trailing_store() || mb->trailing_load_store()) { 3207 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 3208 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 3209 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 3210 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 3211 } else if (mb->leading()) { 3212 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 3213 } 3214 } 3215 #endif 3216 // Count FPU ops and common calls, implements item (3) 3217 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes); 3218 if (!gc_handled) { 3219 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes); 3220 } 3221 3222 // Collect CFG split points 3223 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3224 frc._tests.push(n); 3225 } 3226 } 3227 3228 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) { 3229 if (!UseDivMod) { 3230 return; 3231 } 3232 3233 // Check if "a % b" and "a / b" both exist 3234 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned)); 3235 if (d == nullptr) { 3236 return; 3237 } 3238 3239 // Replace them with a fused divmod if supported 3240 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) { 3241 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned); 3242 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero. 3243 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the 3244 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency 3245 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the 3246 // DivMod node so the dependency is not lost. 3247 divmod->add_prec_from(n); 3248 divmod->add_prec_from(d); 3249 d->subsume_by(divmod->div_proj(), this); 3250 n->subsume_by(divmod->mod_proj(), this); 3251 } else { 3252 // Replace "a % b" with "a - ((a / b) * b)" 3253 Node* mult = MulNode::make(d, d->in(2), bt); 3254 Node* sub = SubNode::make(d->in(1), mult, bt); 3255 n->subsume_by(sub, this); 3256 } 3257 } 3258 3259 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) { 3260 switch( nop ) { 3261 // Count all float operations that may use FPU 3262 case Op_AddF: 3263 case Op_SubF: 3264 case Op_MulF: 3265 case Op_DivF: 3266 case Op_NegF: 3267 case Op_ModF: 3268 case Op_ConvI2F: 3269 case Op_ConF: 3270 case Op_CmpF: 3271 case Op_CmpF3: 3272 case Op_StoreF: 3273 case Op_LoadF: 3274 // case Op_ConvL2F: // longs are split into 32-bit halves 3275 frc.inc_float_count(); 3276 break; 3277 3278 case Op_ConvF2D: 3279 case Op_ConvD2F: 3280 frc.inc_float_count(); 3281 frc.inc_double_count(); 3282 break; 3283 3284 // Count all double operations that may use FPU 3285 case Op_AddD: 3286 case Op_SubD: 3287 case Op_MulD: 3288 case Op_DivD: 3289 case Op_NegD: 3290 case Op_ModD: 3291 case Op_ConvI2D: 3292 case Op_ConvD2I: 3293 // case Op_ConvL2D: // handled by leaf call 3294 // case Op_ConvD2L: // handled by leaf call 3295 case Op_ConD: 3296 case Op_CmpD: 3297 case Op_CmpD3: 3298 case Op_StoreD: 3299 case Op_LoadD: 3300 case Op_LoadD_unaligned: 3301 frc.inc_double_count(); 3302 break; 3303 case Op_Opaque1: // Remove Opaque Nodes before matching 3304 n->subsume_by(n->in(1), this); 3305 break; 3306 case Op_CallStaticJava: 3307 case Op_CallJava: 3308 case Op_CallDynamicJava: 3309 frc.inc_java_call_count(); // Count java call site; 3310 case Op_CallRuntime: 3311 case Op_CallLeaf: 3312 case Op_CallLeafVector: 3313 case Op_CallLeafNoFP: { 3314 assert (n->is_Call(), ""); 3315 CallNode *call = n->as_Call(); 3316 // Count call sites where the FP mode bit would have to be flipped. 3317 // Do not count uncommon runtime calls: 3318 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3319 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3320 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3321 frc.inc_call_count(); // Count the call site 3322 } else { // See if uncommon argument is shared 3323 Node *n = call->in(TypeFunc::Parms); 3324 int nop = n->Opcode(); 3325 // Clone shared simple arguments to uncommon calls, item (1). 3326 if (n->outcnt() > 1 && 3327 !n->is_Proj() && 3328 nop != Op_CreateEx && 3329 nop != Op_CheckCastPP && 3330 nop != Op_DecodeN && 3331 nop != Op_DecodeNKlass && 3332 !n->is_Mem() && 3333 !n->is_Phi()) { 3334 Node *x = n->clone(); 3335 call->set_req(TypeFunc::Parms, x); 3336 } 3337 } 3338 break; 3339 } 3340 case Op_StoreB: 3341 case Op_StoreC: 3342 case Op_StoreI: 3343 case Op_StoreL: 3344 case Op_CompareAndSwapB: 3345 case Op_CompareAndSwapS: 3346 case Op_CompareAndSwapI: 3347 case Op_CompareAndSwapL: 3348 case Op_CompareAndSwapP: 3349 case Op_CompareAndSwapN: 3350 case Op_WeakCompareAndSwapB: 3351 case Op_WeakCompareAndSwapS: 3352 case Op_WeakCompareAndSwapI: 3353 case Op_WeakCompareAndSwapL: 3354 case Op_WeakCompareAndSwapP: 3355 case Op_WeakCompareAndSwapN: 3356 case Op_CompareAndExchangeB: 3357 case Op_CompareAndExchangeS: 3358 case Op_CompareAndExchangeI: 3359 case Op_CompareAndExchangeL: 3360 case Op_CompareAndExchangeP: 3361 case Op_CompareAndExchangeN: 3362 case Op_GetAndAddS: 3363 case Op_GetAndAddB: 3364 case Op_GetAndAddI: 3365 case Op_GetAndAddL: 3366 case Op_GetAndSetS: 3367 case Op_GetAndSetB: 3368 case Op_GetAndSetI: 3369 case Op_GetAndSetL: 3370 case Op_GetAndSetP: 3371 case Op_GetAndSetN: 3372 case Op_StoreP: 3373 case Op_StoreN: 3374 case Op_StoreNKlass: 3375 case Op_LoadB: 3376 case Op_LoadUB: 3377 case Op_LoadUS: 3378 case Op_LoadI: 3379 case Op_LoadKlass: 3380 case Op_LoadNKlass: 3381 case Op_LoadL: 3382 case Op_LoadL_unaligned: 3383 case Op_LoadP: 3384 case Op_LoadN: 3385 case Op_LoadRange: 3386 case Op_LoadS: 3387 break; 3388 3389 case Op_AddP: { // Assert sane base pointers 3390 Node *addp = n->in(AddPNode::Address); 3391 assert( !addp->is_AddP() || 3392 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3393 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3394 "Base pointers must match (addp %u)", addp->_idx ); 3395 #ifdef _LP64 3396 if ((UseCompressedOops || UseCompressedClassPointers) && 3397 addp->Opcode() == Op_ConP && 3398 addp == n->in(AddPNode::Base) && 3399 n->in(AddPNode::Offset)->is_Con()) { 3400 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3401 // on the platform and on the compressed oops mode. 3402 // Use addressing with narrow klass to load with offset on x86. 3403 // Some platforms can use the constant pool to load ConP. 3404 // Do this transformation here since IGVN will convert ConN back to ConP. 3405 const Type* t = addp->bottom_type(); 3406 bool is_oop = t->isa_oopptr() != nullptr; 3407 bool is_klass = t->isa_klassptr() != nullptr; 3408 3409 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) || 3410 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() && 3411 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) { 3412 Node* nn = nullptr; 3413 3414 int op = is_oop ? Op_ConN : Op_ConNKlass; 3415 3416 // Look for existing ConN node of the same exact type. 3417 Node* r = root(); 3418 uint cnt = r->outcnt(); 3419 for (uint i = 0; i < cnt; i++) { 3420 Node* m = r->raw_out(i); 3421 if (m!= nullptr && m->Opcode() == op && 3422 m->bottom_type()->make_ptr() == t) { 3423 nn = m; 3424 break; 3425 } 3426 } 3427 if (nn != nullptr) { 3428 // Decode a narrow oop to match address 3429 // [R12 + narrow_oop_reg<<3 + offset] 3430 if (is_oop) { 3431 nn = new DecodeNNode(nn, t); 3432 } else { 3433 nn = new DecodeNKlassNode(nn, t); 3434 } 3435 // Check for succeeding AddP which uses the same Base. 3436 // Otherwise we will run into the assertion above when visiting that guy. 3437 for (uint i = 0; i < n->outcnt(); ++i) { 3438 Node *out_i = n->raw_out(i); 3439 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3440 out_i->set_req(AddPNode::Base, nn); 3441 #ifdef ASSERT 3442 for (uint j = 0; j < out_i->outcnt(); ++j) { 3443 Node *out_j = out_i->raw_out(j); 3444 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3445 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3446 } 3447 #endif 3448 } 3449 } 3450 n->set_req(AddPNode::Base, nn); 3451 n->set_req(AddPNode::Address, nn); 3452 if (addp->outcnt() == 0) { 3453 addp->disconnect_inputs(this); 3454 } 3455 } 3456 } 3457 } 3458 #endif 3459 break; 3460 } 3461 3462 case Op_CastPP: { 3463 // Remove CastPP nodes to gain more freedom during scheduling but 3464 // keep the dependency they encode as control or precedence edges 3465 // (if control is set already) on memory operations. Some CastPP 3466 // nodes don't have a control (don't carry a dependency): skip 3467 // those. 3468 if (n->in(0) != nullptr) { 3469 ResourceMark rm; 3470 Unique_Node_List wq; 3471 wq.push(n); 3472 for (uint next = 0; next < wq.size(); ++next) { 3473 Node *m = wq.at(next); 3474 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3475 Node* use = m->fast_out(i); 3476 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3477 use->ensure_control_or_add_prec(n->in(0)); 3478 } else { 3479 switch(use->Opcode()) { 3480 case Op_AddP: 3481 case Op_DecodeN: 3482 case Op_DecodeNKlass: 3483 case Op_CheckCastPP: 3484 case Op_CastPP: 3485 wq.push(use); 3486 break; 3487 } 3488 } 3489 } 3490 } 3491 } 3492 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3493 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3494 Node* in1 = n->in(1); 3495 const Type* t = n->bottom_type(); 3496 Node* new_in1 = in1->clone(); 3497 new_in1->as_DecodeN()->set_type(t); 3498 3499 if (!Matcher::narrow_oop_use_complex_address()) { 3500 // 3501 // x86, ARM and friends can handle 2 adds in addressing mode 3502 // and Matcher can fold a DecodeN node into address by using 3503 // a narrow oop directly and do implicit null check in address: 3504 // 3505 // [R12 + narrow_oop_reg<<3 + offset] 3506 // NullCheck narrow_oop_reg 3507 // 3508 // On other platforms (Sparc) we have to keep new DecodeN node and 3509 // use it to do implicit null check in address: 3510 // 3511 // decode_not_null narrow_oop_reg, base_reg 3512 // [base_reg + offset] 3513 // NullCheck base_reg 3514 // 3515 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3516 // to keep the information to which null check the new DecodeN node 3517 // corresponds to use it as value in implicit_null_check(). 3518 // 3519 new_in1->set_req(0, n->in(0)); 3520 } 3521 3522 n->subsume_by(new_in1, this); 3523 if (in1->outcnt() == 0) { 3524 in1->disconnect_inputs(this); 3525 } 3526 } else { 3527 n->subsume_by(n->in(1), this); 3528 if (n->outcnt() == 0) { 3529 n->disconnect_inputs(this); 3530 } 3531 } 3532 break; 3533 } 3534 case Op_CastII: { 3535 n->as_CastII()->remove_range_check_cast(this); 3536 break; 3537 } 3538 #ifdef _LP64 3539 case Op_CmpP: 3540 // Do this transformation here to preserve CmpPNode::sub() and 3541 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3542 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3543 Node* in1 = n->in(1); 3544 Node* in2 = n->in(2); 3545 if (!in1->is_DecodeNarrowPtr()) { 3546 in2 = in1; 3547 in1 = n->in(2); 3548 } 3549 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3550 3551 Node* new_in2 = nullptr; 3552 if (in2->is_DecodeNarrowPtr()) { 3553 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3554 new_in2 = in2->in(1); 3555 } else if (in2->Opcode() == Op_ConP) { 3556 const Type* t = in2->bottom_type(); 3557 if (t == TypePtr::NULL_PTR) { 3558 assert(in1->is_DecodeN(), "compare klass to null?"); 3559 // Don't convert CmpP null check into CmpN if compressed 3560 // oops implicit null check is not generated. 3561 // This will allow to generate normal oop implicit null check. 3562 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3563 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3564 // 3565 // This transformation together with CastPP transformation above 3566 // will generated code for implicit null checks for compressed oops. 3567 // 3568 // The original code after Optimize() 3569 // 3570 // LoadN memory, narrow_oop_reg 3571 // decode narrow_oop_reg, base_reg 3572 // CmpP base_reg, nullptr 3573 // CastPP base_reg // NotNull 3574 // Load [base_reg + offset], val_reg 3575 // 3576 // after these transformations will be 3577 // 3578 // LoadN memory, narrow_oop_reg 3579 // CmpN narrow_oop_reg, nullptr 3580 // decode_not_null narrow_oop_reg, base_reg 3581 // Load [base_reg + offset], val_reg 3582 // 3583 // and the uncommon path (== nullptr) will use narrow_oop_reg directly 3584 // since narrow oops can be used in debug info now (see the code in 3585 // final_graph_reshaping_walk()). 3586 // 3587 // At the end the code will be matched to 3588 // on x86: 3589 // 3590 // Load_narrow_oop memory, narrow_oop_reg 3591 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3592 // NullCheck narrow_oop_reg 3593 // 3594 // and on sparc: 3595 // 3596 // Load_narrow_oop memory, narrow_oop_reg 3597 // decode_not_null narrow_oop_reg, base_reg 3598 // Load [base_reg + offset], val_reg 3599 // NullCheck base_reg 3600 // 3601 } else if (t->isa_oopptr()) { 3602 new_in2 = ConNode::make(t->make_narrowoop()); 3603 } else if (t->isa_klassptr()) { 3604 new_in2 = ConNode::make(t->make_narrowklass()); 3605 } 3606 } 3607 if (new_in2 != nullptr) { 3608 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3609 n->subsume_by(cmpN, this); 3610 if (in1->outcnt() == 0) { 3611 in1->disconnect_inputs(this); 3612 } 3613 if (in2->outcnt() == 0) { 3614 in2->disconnect_inputs(this); 3615 } 3616 } 3617 } 3618 break; 3619 3620 case Op_DecodeN: 3621 case Op_DecodeNKlass: 3622 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3623 // DecodeN could be pinned when it can't be fold into 3624 // an address expression, see the code for Op_CastPP above. 3625 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3626 break; 3627 3628 case Op_EncodeP: 3629 case Op_EncodePKlass: { 3630 Node* in1 = n->in(1); 3631 if (in1->is_DecodeNarrowPtr()) { 3632 n->subsume_by(in1->in(1), this); 3633 } else if (in1->Opcode() == Op_ConP) { 3634 const Type* t = in1->bottom_type(); 3635 if (t == TypePtr::NULL_PTR) { 3636 assert(t->isa_oopptr(), "null klass?"); 3637 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3638 } else if (t->isa_oopptr()) { 3639 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3640 } else if (t->isa_klassptr()) { 3641 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3642 } 3643 } 3644 if (in1->outcnt() == 0) { 3645 in1->disconnect_inputs(this); 3646 } 3647 break; 3648 } 3649 3650 case Op_Proj: { 3651 if (OptimizeStringConcat || IncrementalInline) { 3652 ProjNode* proj = n->as_Proj(); 3653 if (proj->_is_io_use) { 3654 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, ""); 3655 // Separate projections were used for the exception path which 3656 // are normally removed by a late inline. If it wasn't inlined 3657 // then they will hang around and should just be replaced with 3658 // the original one. Merge them. 3659 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/); 3660 if (non_io_proj != nullptr) { 3661 proj->subsume_by(non_io_proj , this); 3662 } 3663 } 3664 } 3665 break; 3666 } 3667 3668 case Op_Phi: 3669 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3670 // The EncodeP optimization may create Phi with the same edges 3671 // for all paths. It is not handled well by Register Allocator. 3672 Node* unique_in = n->in(1); 3673 assert(unique_in != nullptr, ""); 3674 uint cnt = n->req(); 3675 for (uint i = 2; i < cnt; i++) { 3676 Node* m = n->in(i); 3677 assert(m != nullptr, ""); 3678 if (unique_in != m) 3679 unique_in = nullptr; 3680 } 3681 if (unique_in != nullptr) { 3682 n->subsume_by(unique_in, this); 3683 } 3684 } 3685 break; 3686 3687 #endif 3688 3689 case Op_ModI: 3690 handle_div_mod_op(n, T_INT, false); 3691 break; 3692 3693 case Op_ModL: 3694 handle_div_mod_op(n, T_LONG, false); 3695 break; 3696 3697 case Op_UModI: 3698 handle_div_mod_op(n, T_INT, true); 3699 break; 3700 3701 case Op_UModL: 3702 handle_div_mod_op(n, T_LONG, true); 3703 break; 3704 3705 case Op_LoadVector: 3706 case Op_StoreVector: 3707 #ifdef ASSERT 3708 // Add VerifyVectorAlignment node between adr and load / store. 3709 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) { 3710 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() : 3711 n->as_StoreVector()->must_verify_alignment(); 3712 if (must_verify_alignment) { 3713 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() : 3714 n->as_StoreVector()->memory_size(); 3715 // The memory access should be aligned to the vector width in bytes. 3716 // However, the underlying array is possibly less well aligned, but at least 3717 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in 3718 // a loop we can expect at least the following alignment: 3719 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes); 3720 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range"); 3721 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2"); 3722 // Create mask from alignment. e.g. 0b1000 -> 0b0111 3723 jlong mask = guaranteed_alignment - 1; 3724 Node* mask_con = ConLNode::make(mask); 3725 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con); 3726 n->set_req(MemNode::Address, va); 3727 } 3728 } 3729 #endif 3730 break; 3731 3732 case Op_LoadVectorGather: 3733 case Op_StoreVectorScatter: 3734 case Op_LoadVectorGatherMasked: 3735 case Op_StoreVectorScatterMasked: 3736 case Op_VectorCmpMasked: 3737 case Op_VectorMaskGen: 3738 case Op_LoadVectorMasked: 3739 case Op_StoreVectorMasked: 3740 break; 3741 3742 case Op_AddReductionVI: 3743 case Op_AddReductionVL: 3744 case Op_AddReductionVF: 3745 case Op_AddReductionVD: 3746 case Op_MulReductionVI: 3747 case Op_MulReductionVL: 3748 case Op_MulReductionVF: 3749 case Op_MulReductionVD: 3750 case Op_MinReductionV: 3751 case Op_MaxReductionV: 3752 case Op_AndReductionV: 3753 case Op_OrReductionV: 3754 case Op_XorReductionV: 3755 break; 3756 3757 case Op_PackB: 3758 case Op_PackS: 3759 case Op_PackI: 3760 case Op_PackF: 3761 case Op_PackL: 3762 case Op_PackD: 3763 if (n->req()-1 > 2) { 3764 // Replace many operand PackNodes with a binary tree for matching 3765 PackNode* p = (PackNode*) n; 3766 Node* btp = p->binary_tree_pack(1, n->req()); 3767 n->subsume_by(btp, this); 3768 } 3769 break; 3770 case Op_Loop: 3771 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop"); 3772 case Op_CountedLoop: 3773 case Op_LongCountedLoop: 3774 case Op_OuterStripMinedLoop: 3775 if (n->as_Loop()->is_inner_loop()) { 3776 frc.inc_inner_loop_count(); 3777 } 3778 n->as_Loop()->verify_strip_mined(0); 3779 break; 3780 case Op_LShiftI: 3781 case Op_RShiftI: 3782 case Op_URShiftI: 3783 case Op_LShiftL: 3784 case Op_RShiftL: 3785 case Op_URShiftL: 3786 if (Matcher::need_masked_shift_count) { 3787 // The cpu's shift instructions don't restrict the count to the 3788 // lower 5/6 bits. We need to do the masking ourselves. 3789 Node* in2 = n->in(2); 3790 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3791 const TypeInt* t = in2->find_int_type(); 3792 if (t != nullptr && t->is_con()) { 3793 juint shift = t->get_con(); 3794 if (shift > mask) { // Unsigned cmp 3795 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3796 } 3797 } else { 3798 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) { 3799 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3800 n->set_req(2, shift); 3801 } 3802 } 3803 if (in2->outcnt() == 0) { // Remove dead node 3804 in2->disconnect_inputs(this); 3805 } 3806 } 3807 break; 3808 case Op_MemBarStoreStore: 3809 case Op_MemBarRelease: 3810 // Break the link with AllocateNode: it is no longer useful and 3811 // confuses register allocation. 3812 if (n->req() > MemBarNode::Precedent) { 3813 n->set_req(MemBarNode::Precedent, top()); 3814 } 3815 break; 3816 case Op_MemBarAcquire: { 3817 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3818 // At parse time, the trailing MemBarAcquire for a volatile load 3819 // is created with an edge to the load. After optimizations, 3820 // that input may be a chain of Phis. If those phis have no 3821 // other use, then the MemBarAcquire keeps them alive and 3822 // register allocation can be confused. 3823 dead_nodes.push(n->in(MemBarNode::Precedent)); 3824 n->set_req(MemBarNode::Precedent, top()); 3825 } 3826 break; 3827 } 3828 case Op_Blackhole: 3829 break; 3830 case Op_RangeCheck: { 3831 RangeCheckNode* rc = n->as_RangeCheck(); 3832 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3833 n->subsume_by(iff, this); 3834 frc._tests.push(iff); 3835 break; 3836 } 3837 case Op_ConvI2L: { 3838 if (!Matcher::convi2l_type_required) { 3839 // Code generation on some platforms doesn't need accurate 3840 // ConvI2L types. Widening the type can help remove redundant 3841 // address computations. 3842 n->as_Type()->set_type(TypeLong::INT); 3843 ResourceMark rm; 3844 Unique_Node_List wq; 3845 wq.push(n); 3846 for (uint next = 0; next < wq.size(); next++) { 3847 Node *m = wq.at(next); 3848 3849 for(;;) { 3850 // Loop over all nodes with identical inputs edges as m 3851 Node* k = m->find_similar(m->Opcode()); 3852 if (k == nullptr) { 3853 break; 3854 } 3855 // Push their uses so we get a chance to remove node made 3856 // redundant 3857 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3858 Node* u = k->fast_out(i); 3859 if (u->Opcode() == Op_LShiftL || 3860 u->Opcode() == Op_AddL || 3861 u->Opcode() == Op_SubL || 3862 u->Opcode() == Op_AddP) { 3863 wq.push(u); 3864 } 3865 } 3866 // Replace all nodes with identical edges as m with m 3867 k->subsume_by(m, this); 3868 } 3869 } 3870 } 3871 break; 3872 } 3873 case Op_CmpUL: { 3874 if (!Matcher::has_match_rule(Op_CmpUL)) { 3875 // No support for unsigned long comparisons 3876 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3877 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3878 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3879 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3880 Node* andl = new AndLNode(orl, remove_sign_mask); 3881 Node* cmp = new CmpLNode(andl, n->in(2)); 3882 n->subsume_by(cmp, this); 3883 } 3884 break; 3885 } 3886 #ifdef ASSERT 3887 case Op_ConNKlass: { 3888 const TypePtr* tp = n->as_Type()->type()->make_ptr(); 3889 ciKlass* klass = tp->is_klassptr()->exact_klass(); 3890 assert(klass->is_in_encoding_range(), "klass cannot be compressed"); 3891 break; 3892 } 3893 #endif 3894 default: 3895 assert(!n->is_Call(), ""); 3896 assert(!n->is_Mem(), ""); 3897 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3898 break; 3899 } 3900 } 3901 3902 //------------------------------final_graph_reshaping_walk--------------------- 3903 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3904 // requires that the walk visits a node's inputs before visiting the node. 3905 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3906 Unique_Node_List sfpt; 3907 3908 frc._visited.set(root->_idx); // first, mark node as visited 3909 uint cnt = root->req(); 3910 Node *n = root; 3911 uint i = 0; 3912 while (true) { 3913 if (i < cnt) { 3914 // Place all non-visited non-null inputs onto stack 3915 Node* m = n->in(i); 3916 ++i; 3917 if (m != nullptr && !frc._visited.test_set(m->_idx)) { 3918 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) { 3919 // compute worst case interpreter size in case of a deoptimization 3920 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3921 3922 sfpt.push(m); 3923 } 3924 cnt = m->req(); 3925 nstack.push(n, i); // put on stack parent and next input's index 3926 n = m; 3927 i = 0; 3928 } 3929 } else { 3930 // Now do post-visit work 3931 final_graph_reshaping_impl(n, frc, dead_nodes); 3932 if (nstack.is_empty()) 3933 break; // finished 3934 n = nstack.node(); // Get node from stack 3935 cnt = n->req(); 3936 i = nstack.index(); 3937 nstack.pop(); // Shift to the next node on stack 3938 } 3939 } 3940 3941 // Skip next transformation if compressed oops are not used. 3942 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3943 (!UseCompressedOops && !UseCompressedClassPointers)) 3944 return; 3945 3946 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3947 // It could be done for an uncommon traps or any safepoints/calls 3948 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3949 while (sfpt.size() > 0) { 3950 n = sfpt.pop(); 3951 JVMState *jvms = n->as_SafePoint()->jvms(); 3952 assert(jvms != nullptr, "sanity"); 3953 int start = jvms->debug_start(); 3954 int end = n->req(); 3955 bool is_uncommon = (n->is_CallStaticJava() && 3956 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3957 for (int j = start; j < end; j++) { 3958 Node* in = n->in(j); 3959 if (in->is_DecodeNarrowPtr()) { 3960 bool safe_to_skip = true; 3961 if (!is_uncommon ) { 3962 // Is it safe to skip? 3963 for (uint i = 0; i < in->outcnt(); i++) { 3964 Node* u = in->raw_out(i); 3965 if (!u->is_SafePoint() || 3966 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3967 safe_to_skip = false; 3968 } 3969 } 3970 } 3971 if (safe_to_skip) { 3972 n->set_req(j, in->in(1)); 3973 } 3974 if (in->outcnt() == 0) { 3975 in->disconnect_inputs(this); 3976 } 3977 } 3978 } 3979 } 3980 } 3981 3982 //------------------------------final_graph_reshaping-------------------------- 3983 // Final Graph Reshaping. 3984 // 3985 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3986 // and not commoned up and forced early. Must come after regular 3987 // optimizations to avoid GVN undoing the cloning. Clone constant 3988 // inputs to Loop Phis; these will be split by the allocator anyways. 3989 // Remove Opaque nodes. 3990 // (2) Move last-uses by commutative operations to the left input to encourage 3991 // Intel update-in-place two-address operations and better register usage 3992 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3993 // calls canonicalizing them back. 3994 // (3) Count the number of double-precision FP ops, single-precision FP ops 3995 // and call sites. On Intel, we can get correct rounding either by 3996 // forcing singles to memory (requires extra stores and loads after each 3997 // FP bytecode) or we can set a rounding mode bit (requires setting and 3998 // clearing the mode bit around call sites). The mode bit is only used 3999 // if the relative frequency of single FP ops to calls is low enough. 4000 // This is a key transform for SPEC mpeg_audio. 4001 // (4) Detect infinite loops; blobs of code reachable from above but not 4002 // below. Several of the Code_Gen algorithms fail on such code shapes, 4003 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 4004 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 4005 // Detection is by looking for IfNodes where only 1 projection is 4006 // reachable from below or CatchNodes missing some targets. 4007 // (5) Assert for insane oop offsets in debug mode. 4008 4009 bool Compile::final_graph_reshaping() { 4010 // an infinite loop may have been eliminated by the optimizer, 4011 // in which case the graph will be empty. 4012 if (root()->req() == 1) { 4013 // Do not compile method that is only a trivial infinite loop, 4014 // since the content of the loop may have been eliminated. 4015 record_method_not_compilable("trivial infinite loop"); 4016 return true; 4017 } 4018 4019 // Expensive nodes have their control input set to prevent the GVN 4020 // from freely commoning them. There's no GVN beyond this point so 4021 // no need to keep the control input. We want the expensive nodes to 4022 // be freely moved to the least frequent code path by gcm. 4023 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 4024 for (int i = 0; i < expensive_count(); i++) { 4025 _expensive_nodes.at(i)->set_req(0, nullptr); 4026 } 4027 4028 Final_Reshape_Counts frc; 4029 4030 // Visit everybody reachable! 4031 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 4032 Node_Stack nstack(live_nodes() >> 1); 4033 Unique_Node_List dead_nodes; 4034 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes); 4035 4036 // Check for unreachable (from below) code (i.e., infinite loops). 4037 for( uint i = 0; i < frc._tests.size(); i++ ) { 4038 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 4039 // Get number of CFG targets. 4040 // Note that PCTables include exception targets after calls. 4041 uint required_outcnt = n->required_outcnt(); 4042 if (n->outcnt() != required_outcnt) { 4043 // Check for a few special cases. Rethrow Nodes never take the 4044 // 'fall-thru' path, so expected kids is 1 less. 4045 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 4046 if (n->in(0)->in(0)->is_Call()) { 4047 CallNode* call = n->in(0)->in(0)->as_Call(); 4048 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 4049 required_outcnt--; // Rethrow always has 1 less kid 4050 } else if (call->req() > TypeFunc::Parms && 4051 call->is_CallDynamicJava()) { 4052 // Check for null receiver. In such case, the optimizer has 4053 // detected that the virtual call will always result in a null 4054 // pointer exception. The fall-through projection of this CatchNode 4055 // will not be populated. 4056 Node* arg0 = call->in(TypeFunc::Parms); 4057 if (arg0->is_Type() && 4058 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 4059 required_outcnt--; 4060 } 4061 } else if (call->entry_point() == OptoRuntime::new_array_Java() || 4062 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4063 // Check for illegal array length. In such case, the optimizer has 4064 // detected that the allocation attempt will always result in an 4065 // exception. There is no fall-through projection of this CatchNode . 4066 assert(call->is_CallStaticJava(), "static call expected"); 4067 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4068 uint valid_length_test_input = call->req() - 1; 4069 Node* valid_length_test = call->in(valid_length_test_input); 4070 call->del_req(valid_length_test_input); 4071 if (valid_length_test->find_int_con(1) == 0) { 4072 required_outcnt--; 4073 } 4074 dead_nodes.push(valid_length_test); 4075 assert(n->outcnt() == required_outcnt, "malformed control flow"); 4076 continue; 4077 } 4078 } 4079 } 4080 4081 // Recheck with a better notion of 'required_outcnt' 4082 if (n->outcnt() != required_outcnt) { 4083 record_method_not_compilable("malformed control flow"); 4084 return true; // Not all targets reachable! 4085 } 4086 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) { 4087 CallNode* call = n->in(0)->in(0)->as_Call(); 4088 if (call->entry_point() == OptoRuntime::new_array_Java() || 4089 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4090 assert(call->is_CallStaticJava(), "static call expected"); 4091 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4092 uint valid_length_test_input = call->req() - 1; 4093 dead_nodes.push(call->in(valid_length_test_input)); 4094 call->del_req(valid_length_test_input); // valid length test useless now 4095 } 4096 } 4097 // Check that I actually visited all kids. Unreached kids 4098 // must be infinite loops. 4099 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 4100 if (!frc._visited.test(n->fast_out(j)->_idx)) { 4101 record_method_not_compilable("infinite loop"); 4102 return true; // Found unvisited kid; must be unreach 4103 } 4104 4105 // Here so verification code in final_graph_reshaping_walk() 4106 // always see an OuterStripMinedLoopEnd 4107 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) { 4108 IfNode* init_iff = n->as_If(); 4109 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 4110 n->subsume_by(iff, this); 4111 } 4112 } 4113 4114 while (dead_nodes.size() > 0) { 4115 Node* m = dead_nodes.pop(); 4116 if (m->outcnt() == 0 && m != top()) { 4117 for (uint j = 0; j < m->req(); j++) { 4118 Node* in = m->in(j); 4119 if (in != nullptr) { 4120 dead_nodes.push(in); 4121 } 4122 } 4123 m->disconnect_inputs(this); 4124 } 4125 } 4126 4127 set_java_calls(frc.get_java_call_count()); 4128 set_inner_loops(frc.get_inner_loop_count()); 4129 4130 // No infinite loops, no reason to bail out. 4131 return false; 4132 } 4133 4134 //-----------------------------too_many_traps---------------------------------- 4135 // Report if there are too many traps at the current method and bci. 4136 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 4137 bool Compile::too_many_traps(ciMethod* method, 4138 int bci, 4139 Deoptimization::DeoptReason reason) { 4140 ciMethodData* md = method->method_data(); 4141 if (md->is_empty()) { 4142 // Assume the trap has not occurred, or that it occurred only 4143 // because of a transient condition during start-up in the interpreter. 4144 return false; 4145 } 4146 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4147 if (md->has_trap_at(bci, m, reason) != 0) { 4148 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 4149 // Also, if there are multiple reasons, or if there is no per-BCI record, 4150 // assume the worst. 4151 if (log()) 4152 log()->elem("observe trap='%s' count='%d'", 4153 Deoptimization::trap_reason_name(reason), 4154 md->trap_count(reason)); 4155 return true; 4156 } else { 4157 // Ignore method/bci and see if there have been too many globally. 4158 return too_many_traps(reason, md); 4159 } 4160 } 4161 4162 // Less-accurate variant which does not require a method and bci. 4163 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 4164 ciMethodData* logmd) { 4165 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 4166 // Too many traps globally. 4167 // Note that we use cumulative trap_count, not just md->trap_count. 4168 if (log()) { 4169 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason); 4170 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 4171 Deoptimization::trap_reason_name(reason), 4172 mcount, trap_count(reason)); 4173 } 4174 return true; 4175 } else { 4176 // The coast is clear. 4177 return false; 4178 } 4179 } 4180 4181 //--------------------------too_many_recompiles-------------------------------- 4182 // Report if there are too many recompiles at the current method and bci. 4183 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 4184 // Is not eager to return true, since this will cause the compiler to use 4185 // Action_none for a trap point, to avoid too many recompilations. 4186 bool Compile::too_many_recompiles(ciMethod* method, 4187 int bci, 4188 Deoptimization::DeoptReason reason) { 4189 ciMethodData* md = method->method_data(); 4190 if (md->is_empty()) { 4191 // Assume the trap has not occurred, or that it occurred only 4192 // because of a transient condition during start-up in the interpreter. 4193 return false; 4194 } 4195 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 4196 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 4197 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 4198 Deoptimization::DeoptReason per_bc_reason 4199 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 4200 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4201 if ((per_bc_reason == Deoptimization::Reason_none 4202 || md->has_trap_at(bci, m, reason) != 0) 4203 // The trap frequency measure we care about is the recompile count: 4204 && md->trap_recompiled_at(bci, m) 4205 && md->overflow_recompile_count() >= bc_cutoff) { 4206 // Do not emit a trap here if it has already caused recompilations. 4207 // Also, if there are multiple reasons, or if there is no per-BCI record, 4208 // assume the worst. 4209 if (log()) 4210 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 4211 Deoptimization::trap_reason_name(reason), 4212 md->trap_count(reason), 4213 md->overflow_recompile_count()); 4214 return true; 4215 } else if (trap_count(reason) != 0 4216 && decompile_count() >= m_cutoff) { 4217 // Too many recompiles globally, and we have seen this sort of trap. 4218 // Use cumulative decompile_count, not just md->decompile_count. 4219 if (log()) 4220 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 4221 Deoptimization::trap_reason_name(reason), 4222 md->trap_count(reason), trap_count(reason), 4223 md->decompile_count(), decompile_count()); 4224 return true; 4225 } else { 4226 // The coast is clear. 4227 return false; 4228 } 4229 } 4230 4231 // Compute when not to trap. Used by matching trap based nodes and 4232 // NullCheck optimization. 4233 void Compile::set_allowed_deopt_reasons() { 4234 _allowed_reasons = 0; 4235 if (is_method_compilation()) { 4236 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 4237 assert(rs < BitsPerInt, "recode bit map"); 4238 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 4239 _allowed_reasons |= nth_bit(rs); 4240 } 4241 } 4242 } 4243 } 4244 4245 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 4246 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 4247 } 4248 4249 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 4250 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 4251 } 4252 4253 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 4254 if (holder->is_initialized()) { 4255 return false; 4256 } 4257 if (holder->is_being_initialized()) { 4258 if (accessing_method->holder() == holder) { 4259 // Access inside a class. The barrier can be elided when access happens in <clinit>, 4260 // <init>, or a static method. In all those cases, there was an initialization 4261 // barrier on the holder klass passed. 4262 if (accessing_method->is_static_initializer() || 4263 accessing_method->is_object_initializer() || 4264 accessing_method->is_static()) { 4265 return false; 4266 } 4267 } else if (accessing_method->holder()->is_subclass_of(holder)) { 4268 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 4269 // In case of <init> or a static method, the barrier is on the subclass is not enough: 4270 // child class can become fully initialized while its parent class is still being initialized. 4271 if (accessing_method->is_static_initializer()) { 4272 return false; 4273 } 4274 } 4275 ciMethod* root = method(); // the root method of compilation 4276 if (root != accessing_method) { 4277 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 4278 } 4279 } 4280 return true; 4281 } 4282 4283 #ifndef PRODUCT 4284 //------------------------------verify_bidirectional_edges--------------------- 4285 // For each input edge to a node (ie - for each Use-Def edge), verify that 4286 // there is a corresponding Def-Use edge. 4287 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const { 4288 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc 4289 uint stack_size = live_nodes() >> 4; 4290 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize)); 4291 if (root_and_safepoints != nullptr) { 4292 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints"); 4293 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) { 4294 Node* root_or_safepoint = root_and_safepoints->at(i); 4295 // If the node is a safepoint, let's check if it still has a control input 4296 // Lack of control input signifies that this node was killed by CCP or 4297 // recursively by remove_globally_dead_node and it shouldn't be a starting 4298 // point. 4299 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) { 4300 nstack.push(root_or_safepoint); 4301 } 4302 } 4303 } else { 4304 nstack.push(_root); 4305 } 4306 4307 while (nstack.size() > 0) { 4308 Node* n = nstack.pop(); 4309 if (visited.member(n)) { 4310 continue; 4311 } 4312 visited.push(n); 4313 4314 // Walk over all input edges, checking for correspondence 4315 uint length = n->len(); 4316 for (uint i = 0; i < length; i++) { 4317 Node* in = n->in(i); 4318 if (in != nullptr && !visited.member(in)) { 4319 nstack.push(in); // Put it on stack 4320 } 4321 if (in != nullptr && !in->is_top()) { 4322 // Count instances of `next` 4323 int cnt = 0; 4324 for (uint idx = 0; idx < in->_outcnt; idx++) { 4325 if (in->_out[idx] == n) { 4326 cnt++; 4327 } 4328 } 4329 assert(cnt > 0, "Failed to find Def-Use edge."); 4330 // Check for duplicate edges 4331 // walk the input array downcounting the input edges to n 4332 for (uint j = 0; j < length; j++) { 4333 if (n->in(j) == in) { 4334 cnt--; 4335 } 4336 } 4337 assert(cnt == 0, "Mismatched edge count."); 4338 } else if (in == nullptr) { 4339 assert(i == 0 || i >= n->req() || 4340 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() || 4341 (n->is_Unlock() && i == (n->req() - 1)) || 4342 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion 4343 "only region, phi, arraycopy, unlock or membar nodes have null data edges"); 4344 } else { 4345 assert(in->is_top(), "sanity"); 4346 // Nothing to check. 4347 } 4348 } 4349 } 4350 } 4351 4352 //------------------------------verify_graph_edges--------------------------- 4353 // Walk the Graph and verify that there is a one-to-one correspondence 4354 // between Use-Def edges and Def-Use edges in the graph. 4355 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const { 4356 if (VerifyGraphEdges) { 4357 Unique_Node_List visited; 4358 4359 // Call graph walk to check edges 4360 verify_bidirectional_edges(visited, root_and_safepoints); 4361 if (no_dead_code) { 4362 // Now make sure that no visited node is used by an unvisited node. 4363 bool dead_nodes = false; 4364 Unique_Node_List checked; 4365 while (visited.size() > 0) { 4366 Node* n = visited.pop(); 4367 checked.push(n); 4368 for (uint i = 0; i < n->outcnt(); i++) { 4369 Node* use = n->raw_out(i); 4370 if (checked.member(use)) continue; // already checked 4371 if (visited.member(use)) continue; // already in the graph 4372 if (use->is_Con()) continue; // a dead ConNode is OK 4373 // At this point, we have found a dead node which is DU-reachable. 4374 if (!dead_nodes) { 4375 tty->print_cr("*** Dead nodes reachable via DU edges:"); 4376 dead_nodes = true; 4377 } 4378 use->dump(2); 4379 tty->print_cr("---"); 4380 checked.push(use); // No repeats; pretend it is now checked. 4381 } 4382 } 4383 assert(!dead_nodes, "using nodes must be reachable from root"); 4384 } 4385 } 4386 } 4387 #endif 4388 4389 // The Compile object keeps track of failure reasons separately from the ciEnv. 4390 // This is required because there is not quite a 1-1 relation between the 4391 // ciEnv and its compilation task and the Compile object. Note that one 4392 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4393 // to backtrack and retry without subsuming loads. Other than this backtracking 4394 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4395 // by the logic in C2Compiler. 4396 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) { 4397 if (log() != nullptr) { 4398 log()->elem("failure reason='%s' phase='compile'", reason); 4399 } 4400 if (_failure_reason.get() == nullptr) { 4401 // Record the first failure reason. 4402 _failure_reason.set(reason); 4403 if (CaptureBailoutInformation) { 4404 _first_failure_details = new CompilationFailureInfo(reason); 4405 } 4406 } else { 4407 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure."); 4408 } 4409 4410 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4411 C->print_method(PHASE_FAILURE, 1); 4412 } 4413 _root = nullptr; // flush the graph, too 4414 } 4415 4416 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id) 4417 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose), 4418 _compile(Compile::current()), 4419 _log(nullptr), 4420 _dolog(CITimeVerbose) 4421 { 4422 assert(_compile != nullptr, "sanity check"); 4423 assert(id != PhaseTraceId::_t_none, "Don't use none"); 4424 if (_dolog) { 4425 _log = _compile->log(); 4426 } 4427 if (_log != nullptr) { 4428 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes()); 4429 _log->stamp(); 4430 _log->end_head(); 4431 } 4432 4433 // Inform memory statistic, if enabled 4434 if (CompilationMemoryStatistic::enabled()) { 4435 CompilationMemoryStatistic::on_phase_start((int)id, name); 4436 } 4437 } 4438 4439 Compile::TracePhase::TracePhase(PhaseTraceId id) 4440 : TracePhase(Phase::get_phase_trace_id_text(id), id) {} 4441 4442 Compile::TracePhase::~TracePhase() { 4443 4444 // Inform memory statistic, if enabled 4445 if (CompilationMemoryStatistic::enabled()) { 4446 CompilationMemoryStatistic::on_phase_end(); 4447 } 4448 4449 if (_compile->failing_internal()) { 4450 if (_log != nullptr) { 4451 _log->done("phase"); 4452 } 4453 return; // timing code, not stressing bailouts. 4454 } 4455 #ifdef ASSERT 4456 if (PrintIdealNodeCount) { 4457 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4458 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk()); 4459 } 4460 4461 if (VerifyIdealNodeCount) { 4462 _compile->print_missing_nodes(); 4463 } 4464 #endif 4465 4466 if (_log != nullptr) { 4467 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes()); 4468 } 4469 } 4470 4471 //----------------------------static_subtype_check----------------------------- 4472 // Shortcut important common cases when superklass is exact: 4473 // (0) superklass is java.lang.Object (can occur in reflective code) 4474 // (1) subklass is already limited to a subtype of superklass => always ok 4475 // (2) subklass does not overlap with superklass => always fail 4476 // (3) superklass has NO subtypes and we can check with a simple compare. 4477 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) { 4478 if (skip) { 4479 return SSC_full_test; // Let caller generate the general case. 4480 } 4481 4482 if (subk->is_java_subtype_of(superk)) { 4483 return SSC_always_true; // (0) and (1) this test cannot fail 4484 } 4485 4486 if (!subk->maybe_java_subtype_of(superk)) { 4487 return SSC_always_false; // (2) true path dead; no dynamic test needed 4488 } 4489 4490 const Type* superelem = superk; 4491 if (superk->isa_aryklassptr()) { 4492 int ignored; 4493 superelem = superk->is_aryklassptr()->base_element_type(ignored); 4494 } 4495 4496 if (superelem->isa_instklassptr()) { 4497 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass(); 4498 if (!ik->has_subklass()) { 4499 if (!ik->is_final()) { 4500 // Add a dependency if there is a chance of a later subclass. 4501 dependencies()->assert_leaf_type(ik); 4502 } 4503 if (!superk->maybe_java_subtype_of(subk)) { 4504 return SSC_always_false; 4505 } 4506 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4507 } 4508 } else { 4509 // A primitive array type has no subtypes. 4510 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4511 } 4512 4513 return SSC_full_test; 4514 } 4515 4516 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4517 #ifdef _LP64 4518 // The scaled index operand to AddP must be a clean 64-bit value. 4519 // Java allows a 32-bit int to be incremented to a negative 4520 // value, which appears in a 64-bit register as a large 4521 // positive number. Using that large positive number as an 4522 // operand in pointer arithmetic has bad consequences. 4523 // On the other hand, 32-bit overflow is rare, and the possibility 4524 // can often be excluded, if we annotate the ConvI2L node with 4525 // a type assertion that its value is known to be a small positive 4526 // number. (The prior range check has ensured this.) 4527 // This assertion is used by ConvI2LNode::Ideal. 4528 int index_max = max_jint - 1; // array size is max_jint, index is one less 4529 if (sizetype != nullptr && sizetype->_hi > 0) { 4530 index_max = sizetype->_hi - 1; 4531 } 4532 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4533 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4534 #endif 4535 return idx; 4536 } 4537 4538 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4539 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 4540 if (ctrl != nullptr) { 4541 // Express control dependency by a CastII node with a narrow type. 4542 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4543 // node from floating above the range check during loop optimizations. Otherwise, the 4544 // ConvI2L node may be eliminated independently of the range check, causing the data path 4545 // to become TOP while the control path is still there (although it's unreachable). 4546 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */); 4547 value = phase->transform(value); 4548 } 4549 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4550 return phase->transform(new ConvI2LNode(value, ltype)); 4551 } 4552 4553 void Compile::dump_print_inlining() { 4554 inline_printer()->print_on(tty); 4555 } 4556 4557 void Compile::log_late_inline(CallGenerator* cg) { 4558 if (log() != nullptr) { 4559 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4560 cg->unique_id()); 4561 JVMState* p = cg->call_node()->jvms(); 4562 while (p != nullptr) { 4563 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4564 p = p->caller(); 4565 } 4566 log()->tail("late_inline"); 4567 } 4568 } 4569 4570 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4571 log_late_inline(cg); 4572 if (log() != nullptr) { 4573 log()->inline_fail(msg); 4574 } 4575 } 4576 4577 void Compile::log_inline_id(CallGenerator* cg) { 4578 if (log() != nullptr) { 4579 // The LogCompilation tool needs a unique way to identify late 4580 // inline call sites. This id must be unique for this call site in 4581 // this compilation. Try to have it unique across compilations as 4582 // well because it can be convenient when grepping through the log 4583 // file. 4584 // Distinguish OSR compilations from others in case CICountOSR is 4585 // on. 4586 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4587 cg->set_unique_id(id); 4588 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4589 } 4590 } 4591 4592 void Compile::log_inline_failure(const char* msg) { 4593 if (C->log() != nullptr) { 4594 C->log()->inline_fail(msg); 4595 } 4596 } 4597 4598 4599 // Dump inlining replay data to the stream. 4600 // Don't change thread state and acquire any locks. 4601 void Compile::dump_inline_data(outputStream* out) { 4602 InlineTree* inl_tree = ilt(); 4603 if (inl_tree != nullptr) { 4604 out->print(" inline %d", inl_tree->count()); 4605 inl_tree->dump_replay_data(out); 4606 } 4607 } 4608 4609 void Compile::dump_inline_data_reduced(outputStream* out) { 4610 assert(ReplayReduce, ""); 4611 4612 InlineTree* inl_tree = ilt(); 4613 if (inl_tree == nullptr) { 4614 return; 4615 } 4616 // Enable iterative replay file reduction 4617 // Output "compile" lines for depth 1 subtrees, 4618 // simulating that those trees were compiled 4619 // instead of inlined. 4620 for (int i = 0; i < inl_tree->subtrees().length(); ++i) { 4621 InlineTree* sub = inl_tree->subtrees().at(i); 4622 if (sub->inline_level() != 1) { 4623 continue; 4624 } 4625 4626 ciMethod* method = sub->method(); 4627 int entry_bci = -1; 4628 int comp_level = env()->task()->comp_level(); 4629 out->print("compile "); 4630 method->dump_name_as_ascii(out); 4631 out->print(" %d %d", entry_bci, comp_level); 4632 out->print(" inline %d", sub->count()); 4633 sub->dump_replay_data(out, -1); 4634 out->cr(); 4635 } 4636 } 4637 4638 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4639 if (n1->Opcode() < n2->Opcode()) return -1; 4640 else if (n1->Opcode() > n2->Opcode()) return 1; 4641 4642 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4643 for (uint i = 1; i < n1->req(); i++) { 4644 if (n1->in(i) < n2->in(i)) return -1; 4645 else if (n1->in(i) > n2->in(i)) return 1; 4646 } 4647 4648 return 0; 4649 } 4650 4651 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4652 Node* n1 = *n1p; 4653 Node* n2 = *n2p; 4654 4655 return cmp_expensive_nodes(n1, n2); 4656 } 4657 4658 void Compile::sort_expensive_nodes() { 4659 if (!expensive_nodes_sorted()) { 4660 _expensive_nodes.sort(cmp_expensive_nodes); 4661 } 4662 } 4663 4664 bool Compile::expensive_nodes_sorted() const { 4665 for (int i = 1; i < _expensive_nodes.length(); i++) { 4666 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) { 4667 return false; 4668 } 4669 } 4670 return true; 4671 } 4672 4673 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4674 if (_expensive_nodes.length() == 0) { 4675 return false; 4676 } 4677 4678 assert(OptimizeExpensiveOps, "optimization off?"); 4679 4680 // Take this opportunity to remove dead nodes from the list 4681 int j = 0; 4682 for (int i = 0; i < _expensive_nodes.length(); i++) { 4683 Node* n = _expensive_nodes.at(i); 4684 if (!n->is_unreachable(igvn)) { 4685 assert(n->is_expensive(), "should be expensive"); 4686 _expensive_nodes.at_put(j, n); 4687 j++; 4688 } 4689 } 4690 _expensive_nodes.trunc_to(j); 4691 4692 // Then sort the list so that similar nodes are next to each other 4693 // and check for at least two nodes of identical kind with same data 4694 // inputs. 4695 sort_expensive_nodes(); 4696 4697 for (int i = 0; i < _expensive_nodes.length()-1; i++) { 4698 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) { 4699 return true; 4700 } 4701 } 4702 4703 return false; 4704 } 4705 4706 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4707 if (_expensive_nodes.length() == 0) { 4708 return; 4709 } 4710 4711 assert(OptimizeExpensiveOps, "optimization off?"); 4712 4713 // Sort to bring similar nodes next to each other and clear the 4714 // control input of nodes for which there's only a single copy. 4715 sort_expensive_nodes(); 4716 4717 int j = 0; 4718 int identical = 0; 4719 int i = 0; 4720 bool modified = false; 4721 for (; i < _expensive_nodes.length()-1; i++) { 4722 assert(j <= i, "can't write beyond current index"); 4723 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) { 4724 identical++; 4725 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4726 continue; 4727 } 4728 if (identical > 0) { 4729 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4730 identical = 0; 4731 } else { 4732 Node* n = _expensive_nodes.at(i); 4733 igvn.replace_input_of(n, 0, nullptr); 4734 igvn.hash_insert(n); 4735 modified = true; 4736 } 4737 } 4738 if (identical > 0) { 4739 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4740 } else if (_expensive_nodes.length() >= 1) { 4741 Node* n = _expensive_nodes.at(i); 4742 igvn.replace_input_of(n, 0, nullptr); 4743 igvn.hash_insert(n); 4744 modified = true; 4745 } 4746 _expensive_nodes.trunc_to(j); 4747 if (modified) { 4748 igvn.optimize(); 4749 } 4750 } 4751 4752 void Compile::add_expensive_node(Node * n) { 4753 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list"); 4754 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4755 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4756 if (OptimizeExpensiveOps) { 4757 _expensive_nodes.append(n); 4758 } else { 4759 // Clear control input and let IGVN optimize expensive nodes if 4760 // OptimizeExpensiveOps is off. 4761 n->set_req(0, nullptr); 4762 } 4763 } 4764 4765 /** 4766 * Track coarsened Lock and Unlock nodes. 4767 */ 4768 4769 class Lock_List : public Node_List { 4770 uint _origin_cnt; 4771 public: 4772 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {} 4773 uint origin_cnt() const { return _origin_cnt; } 4774 }; 4775 4776 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) { 4777 int length = locks.length(); 4778 if (length > 0) { 4779 // Have to keep this list until locks elimination during Macro nodes elimination. 4780 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length); 4781 AbstractLockNode* alock = locks.at(0); 4782 BoxLockNode* box = alock->box_node()->as_BoxLock(); 4783 for (int i = 0; i < length; i++) { 4784 AbstractLockNode* lock = locks.at(i); 4785 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx); 4786 locks_list->push(lock); 4787 BoxLockNode* this_box = lock->box_node()->as_BoxLock(); 4788 if (this_box != box) { 4789 // Locking regions (BoxLock) could be Unbalanced here: 4790 // - its coarsened locks were eliminated in earlier 4791 // macro nodes elimination followed by loop unroll 4792 // - it is OSR locking region (no Lock node) 4793 // Preserve Unbalanced status in such cases. 4794 if (!this_box->is_unbalanced()) { 4795 this_box->set_coarsened(); 4796 } 4797 if (!box->is_unbalanced()) { 4798 box->set_coarsened(); 4799 } 4800 } 4801 } 4802 _coarsened_locks.append(locks_list); 4803 } 4804 } 4805 4806 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) { 4807 int count = coarsened_count(); 4808 for (int i = 0; i < count; i++) { 4809 Node_List* locks_list = _coarsened_locks.at(i); 4810 for (uint j = 0; j < locks_list->size(); j++) { 4811 Node* lock = locks_list->at(j); 4812 assert(lock->is_AbstractLock(), "sanity"); 4813 if (!useful.member(lock)) { 4814 locks_list->yank(lock); 4815 } 4816 } 4817 } 4818 } 4819 4820 void Compile::remove_coarsened_lock(Node* n) { 4821 if (n->is_AbstractLock()) { 4822 int count = coarsened_count(); 4823 for (int i = 0; i < count; i++) { 4824 Node_List* locks_list = _coarsened_locks.at(i); 4825 locks_list->yank(n); 4826 } 4827 } 4828 } 4829 4830 bool Compile::coarsened_locks_consistent() { 4831 int count = coarsened_count(); 4832 for (int i = 0; i < count; i++) { 4833 bool unbalanced = false; 4834 bool modified = false; // track locks kind modifications 4835 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i); 4836 uint size = locks_list->size(); 4837 if (size == 0) { 4838 unbalanced = false; // All locks were eliminated - good 4839 } else if (size != locks_list->origin_cnt()) { 4840 unbalanced = true; // Some locks were removed from list 4841 } else { 4842 for (uint j = 0; j < size; j++) { 4843 Node* lock = locks_list->at(j); 4844 // All nodes in group should have the same state (modified or not) 4845 if (!lock->as_AbstractLock()->is_coarsened()) { 4846 if (j == 0) { 4847 // first on list was modified, the rest should be too for consistency 4848 modified = true; 4849 } else if (!modified) { 4850 // this lock was modified but previous locks on the list were not 4851 unbalanced = true; 4852 break; 4853 } 4854 } else if (modified) { 4855 // previous locks on list were modified but not this lock 4856 unbalanced = true; 4857 break; 4858 } 4859 } 4860 } 4861 if (unbalanced) { 4862 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified 4863 #ifdef ASSERT 4864 if (PrintEliminateLocks) { 4865 tty->print_cr("=== unbalanced coarsened locks ==="); 4866 for (uint l = 0; l < size; l++) { 4867 locks_list->at(l)->dump(); 4868 } 4869 } 4870 #endif 4871 record_failure(C2Compiler::retry_no_locks_coarsening()); 4872 return false; 4873 } 4874 } 4875 return true; 4876 } 4877 4878 // Mark locking regions (identified by BoxLockNode) as unbalanced if 4879 // locks coarsening optimization removed Lock/Unlock nodes from them. 4880 // Such regions become unbalanced because coarsening only removes part 4881 // of Lock/Unlock nodes in region. As result we can't execute other 4882 // locks elimination optimizations which assume all code paths have 4883 // corresponding pair of Lock/Unlock nodes - they are balanced. 4884 void Compile::mark_unbalanced_boxes() const { 4885 int count = coarsened_count(); 4886 for (int i = 0; i < count; i++) { 4887 Node_List* locks_list = _coarsened_locks.at(i); 4888 uint size = locks_list->size(); 4889 if (size > 0) { 4890 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock(); 4891 BoxLockNode* box = alock->box_node()->as_BoxLock(); 4892 if (alock->is_coarsened()) { 4893 // coarsened_locks_consistent(), which is called before this method, verifies 4894 // that the rest of Lock/Unlock nodes on locks_list are also coarsened. 4895 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated"); 4896 for (uint j = 1; j < size; j++) { 4897 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here"); 4898 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock(); 4899 if (box != this_box) { 4900 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated"); 4901 box->set_unbalanced(); 4902 this_box->set_unbalanced(); 4903 } 4904 } 4905 } 4906 } 4907 } 4908 } 4909 4910 /** 4911 * Remove the speculative part of types and clean up the graph 4912 */ 4913 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4914 if (UseTypeSpeculation) { 4915 Unique_Node_List worklist; 4916 worklist.push(root()); 4917 int modified = 0; 4918 // Go over all type nodes that carry a speculative type, drop the 4919 // speculative part of the type and enqueue the node for an igvn 4920 // which may optimize it out. 4921 for (uint next = 0; next < worklist.size(); ++next) { 4922 Node *n = worklist.at(next); 4923 if (n->is_Type()) { 4924 TypeNode* tn = n->as_Type(); 4925 const Type* t = tn->type(); 4926 const Type* t_no_spec = t->remove_speculative(); 4927 if (t_no_spec != t) { 4928 bool in_hash = igvn.hash_delete(n); 4929 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table"); 4930 tn->set_type(t_no_spec); 4931 igvn.hash_insert(n); 4932 igvn._worklist.push(n); // give it a chance to go away 4933 modified++; 4934 } 4935 } 4936 // Iterate over outs - endless loops is unreachable from below 4937 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4938 Node *m = n->fast_out(i); 4939 if (not_a_node(m)) { 4940 continue; 4941 } 4942 worklist.push(m); 4943 } 4944 } 4945 // Drop the speculative part of all types in the igvn's type table 4946 igvn.remove_speculative_types(); 4947 if (modified > 0) { 4948 igvn.optimize(); 4949 if (failing()) return; 4950 } 4951 #ifdef ASSERT 4952 // Verify that after the IGVN is over no speculative type has resurfaced 4953 worklist.clear(); 4954 worklist.push(root()); 4955 for (uint next = 0; next < worklist.size(); ++next) { 4956 Node *n = worklist.at(next); 4957 const Type* t = igvn.type_or_null(n); 4958 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types"); 4959 if (n->is_Type()) { 4960 t = n->as_Type()->type(); 4961 assert(t == t->remove_speculative(), "no more speculative types"); 4962 } 4963 // Iterate over outs - endless loops is unreachable from below 4964 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4965 Node *m = n->fast_out(i); 4966 if (not_a_node(m)) { 4967 continue; 4968 } 4969 worklist.push(m); 4970 } 4971 } 4972 igvn.check_no_speculative_types(); 4973 #endif 4974 } 4975 } 4976 4977 // Auxiliary methods to support randomized stressing/fuzzing. 4978 4979 void Compile::initialize_stress_seed(const DirectiveSet* directive) { 4980 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) { 4981 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds()); 4982 FLAG_SET_ERGO(StressSeed, _stress_seed); 4983 } else { 4984 _stress_seed = StressSeed; 4985 } 4986 if (_log != nullptr) { 4987 _log->elem("stress_test seed='%u'", _stress_seed); 4988 } 4989 } 4990 4991 int Compile::random() { 4992 _stress_seed = os::next_random(_stress_seed); 4993 return static_cast<int>(_stress_seed); 4994 } 4995 4996 // This method can be called the arbitrary number of times, with current count 4997 // as the argument. The logic allows selecting a single candidate from the 4998 // running list of candidates as follows: 4999 // int count = 0; 5000 // Cand* selected = null; 5001 // while(cand = cand->next()) { 5002 // if (randomized_select(++count)) { 5003 // selected = cand; 5004 // } 5005 // } 5006 // 5007 // Including count equalizes the chances any candidate is "selected". 5008 // This is useful when we don't have the complete list of candidates to choose 5009 // from uniformly. In this case, we need to adjust the randomicity of the 5010 // selection, or else we will end up biasing the selection towards the latter 5011 // candidates. 5012 // 5013 // Quick back-envelope calculation shows that for the list of n candidates 5014 // the equal probability for the candidate to persist as "best" can be 5015 // achieved by replacing it with "next" k-th candidate with the probability 5016 // of 1/k. It can be easily shown that by the end of the run, the 5017 // probability for any candidate is converged to 1/n, thus giving the 5018 // uniform distribution among all the candidates. 5019 // 5020 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 5021 #define RANDOMIZED_DOMAIN_POW 29 5022 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 5023 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 5024 bool Compile::randomized_select(int count) { 5025 assert(count > 0, "only positive"); 5026 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 5027 } 5028 5029 #ifdef ASSERT 5030 // Failures are geometrically distributed with probability 1/StressBailoutMean. 5031 bool Compile::fail_randomly() { 5032 if ((random() % StressBailoutMean) != 0) { 5033 return false; 5034 } 5035 record_failure("StressBailout"); 5036 return true; 5037 } 5038 5039 bool Compile::failure_is_artificial() { 5040 return C->failure_reason_is("StressBailout"); 5041 } 5042 #endif 5043 5044 CloneMap& Compile::clone_map() { return _clone_map; } 5045 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 5046 5047 void NodeCloneInfo::dump_on(outputStream* st) const { 5048 st->print(" {%d:%d} ", idx(), gen()); 5049 } 5050 5051 void CloneMap::clone(Node* old, Node* nnn, int gen) { 5052 uint64_t val = value(old->_idx); 5053 NodeCloneInfo cio(val); 5054 assert(val != 0, "old node should be in the map"); 5055 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 5056 insert(nnn->_idx, cin.get()); 5057 #ifndef PRODUCT 5058 if (is_debug()) { 5059 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 5060 } 5061 #endif 5062 } 5063 5064 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 5065 NodeCloneInfo cio(value(old->_idx)); 5066 if (cio.get() == 0) { 5067 cio.set(old->_idx, 0); 5068 insert(old->_idx, cio.get()); 5069 #ifndef PRODUCT 5070 if (is_debug()) { 5071 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 5072 } 5073 #endif 5074 } 5075 clone(old, nnn, gen); 5076 } 5077 5078 int CloneMap::max_gen() const { 5079 int g = 0; 5080 DictI di(_dict); 5081 for(; di.test(); ++di) { 5082 int t = gen(di._key); 5083 if (g < t) { 5084 g = t; 5085 #ifndef PRODUCT 5086 if (is_debug()) { 5087 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 5088 } 5089 #endif 5090 } 5091 } 5092 return g; 5093 } 5094 5095 void CloneMap::dump(node_idx_t key, outputStream* st) const { 5096 uint64_t val = value(key); 5097 if (val != 0) { 5098 NodeCloneInfo ni(val); 5099 ni.dump_on(st); 5100 } 5101 } 5102 5103 void Compile::shuffle_macro_nodes() { 5104 if (_macro_nodes.length() < 2) { 5105 return; 5106 } 5107 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) { 5108 uint j = C->random() % (i + 1); 5109 swap(_macro_nodes.at(i), _macro_nodes.at(j)); 5110 } 5111 } 5112 5113 // Move Allocate nodes to the start of the list 5114 void Compile::sort_macro_nodes() { 5115 int count = macro_count(); 5116 int allocates = 0; 5117 for (int i = 0; i < count; i++) { 5118 Node* n = macro_node(i); 5119 if (n->is_Allocate()) { 5120 if (i != allocates) { 5121 Node* tmp = macro_node(allocates); 5122 _macro_nodes.at_put(allocates, n); 5123 _macro_nodes.at_put(i, tmp); 5124 } 5125 allocates++; 5126 } 5127 } 5128 } 5129 5130 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) { 5131 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code. 5132 EventCompilerPhase event(UNTIMED); 5133 if (event.should_commit()) { 5134 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 5135 } 5136 #ifndef PRODUCT 5137 ResourceMark rm; 5138 stringStream ss; 5139 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt)); 5140 int iter = ++_igv_phase_iter[cpt]; 5141 if (iter > 1) { 5142 ss.print(" %d", iter); 5143 } 5144 if (n != nullptr) { 5145 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]); 5146 if (n->is_Call()) { 5147 CallNode* call = n->as_Call(); 5148 if (call->_name != nullptr) { 5149 // E.g. uncommon traps etc. 5150 ss.print(" - %s", call->_name); 5151 } else if (call->is_CallJava()) { 5152 CallJavaNode* call_java = call->as_CallJava(); 5153 if (call_java->method() != nullptr) { 5154 ss.print(" -"); 5155 call_java->method()->print_short_name(&ss); 5156 } 5157 } 5158 } 5159 } 5160 5161 const char* name = ss.as_string(); 5162 if (should_print_igv(level)) { 5163 _igv_printer->print_graph(name); 5164 } 5165 if (should_print_phase(level)) { 5166 print_phase(name); 5167 } 5168 if (should_print_ideal_phase(cpt)) { 5169 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt)); 5170 } 5171 #endif 5172 C->_latest_stage_start_counter.stamp(); 5173 } 5174 5175 // Only used from CompileWrapper 5176 void Compile::begin_method() { 5177 #ifndef PRODUCT 5178 if (_method != nullptr && should_print_igv(1)) { 5179 _igv_printer->begin_method(); 5180 } 5181 #endif 5182 C->_latest_stage_start_counter.stamp(); 5183 } 5184 5185 // Only used from CompileWrapper 5186 void Compile::end_method() { 5187 EventCompilerPhase event(UNTIMED); 5188 if (event.should_commit()) { 5189 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1); 5190 } 5191 5192 #ifndef PRODUCT 5193 if (_method != nullptr && should_print_igv(1)) { 5194 _igv_printer->end_method(); 5195 } 5196 #endif 5197 } 5198 5199 #ifndef PRODUCT 5200 bool Compile::should_print_phase(const int level) const { 5201 return PrintPhaseLevel > 0 && directive()->PhasePrintLevelOption >= level && 5202 _method != nullptr; // Do not print phases for stubs. 5203 } 5204 5205 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const { 5206 return _directive->should_print_ideal_phase(cpt); 5207 } 5208 5209 void Compile::init_igv() { 5210 if (_igv_printer == nullptr) { 5211 _igv_printer = IdealGraphPrinter::printer(); 5212 _igv_printer->set_compile(this); 5213 } 5214 } 5215 5216 bool Compile::should_print_igv(const int level) { 5217 PRODUCT_RETURN_(return false;); 5218 5219 if (PrintIdealGraphLevel < 0) { // disabled by the user 5220 return false; 5221 } 5222 5223 bool need = directive()->IGVPrintLevelOption >= level; 5224 if (need) { 5225 Compile::init_igv(); 5226 } 5227 return need; 5228 } 5229 5230 IdealGraphPrinter* Compile::_debug_file_printer = nullptr; 5231 IdealGraphPrinter* Compile::_debug_network_printer = nullptr; 5232 5233 // Called from debugger. Prints method to the default file with the default phase name. 5234 // This works regardless of any Ideal Graph Visualizer flags set or not. 5235 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc). 5236 void igv_print(void* sp, void* fp, void* pc) { 5237 frame fr(sp, fp, pc); 5238 Compile::current()->igv_print_method_to_file(nullptr, false, &fr); 5239 } 5240 5241 // Same as igv_print() above but with a specified phase name. 5242 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) { 5243 frame fr(sp, fp, pc); 5244 Compile::current()->igv_print_method_to_file(phase_name, false, &fr); 5245 } 5246 5247 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 5248 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 5249 // This works regardless of any Ideal Graph Visualizer flags set or not. 5250 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc). 5251 void igv_print(bool network, void* sp, void* fp, void* pc) { 5252 frame fr(sp, fp, pc); 5253 if (network) { 5254 Compile::current()->igv_print_method_to_network(nullptr, &fr); 5255 } else { 5256 Compile::current()->igv_print_method_to_file(nullptr, false, &fr); 5257 } 5258 } 5259 5260 // Same as igv_print(bool network, ...) above but with a specified phase name. 5261 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc). 5262 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) { 5263 frame fr(sp, fp, pc); 5264 if (network) { 5265 Compile::current()->igv_print_method_to_network(phase_name, &fr); 5266 } else { 5267 Compile::current()->igv_print_method_to_file(phase_name, false, &fr); 5268 } 5269 } 5270 5271 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 5272 void igv_print_default() { 5273 Compile::current()->print_method(PHASE_DEBUG, 0); 5274 } 5275 5276 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 5277 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 5278 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 5279 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc). 5280 void igv_append(void* sp, void* fp, void* pc) { 5281 frame fr(sp, fp, pc); 5282 Compile::current()->igv_print_method_to_file(nullptr, true, &fr); 5283 } 5284 5285 // Same as igv_append(...) above but with a specified phase name. 5286 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc). 5287 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) { 5288 frame fr(sp, fp, pc); 5289 Compile::current()->igv_print_method_to_file(phase_name, true, &fr); 5290 } 5291 5292 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) { 5293 const char* file_name = "custom_debug.xml"; 5294 if (_debug_file_printer == nullptr) { 5295 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 5296 } else { 5297 _debug_file_printer->update_compiled_method(C->method()); 5298 } 5299 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 5300 _debug_file_printer->print_graph(phase_name, fr); 5301 } 5302 5303 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) { 5304 ResourceMark rm; 5305 GrowableArray<const Node*> empty_list; 5306 igv_print_graph_to_network(phase_name, empty_list, fr); 5307 } 5308 5309 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) { 5310 if (_debug_network_printer == nullptr) { 5311 _debug_network_printer = new IdealGraphPrinter(C); 5312 } else { 5313 _debug_network_printer->update_compiled_method(C->method()); 5314 } 5315 tty->print_cr("Method printed over network stream to IGV"); 5316 _debug_network_printer->print(name, C->root(), visible_nodes, fr); 5317 } 5318 #endif // !PRODUCT 5319 5320 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 5321 if (type != nullptr && phase->type(value)->higher_equal(type)) { 5322 return value; 5323 } 5324 Node* result = nullptr; 5325 if (bt == T_BYTE) { 5326 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 5327 result = new RShiftINode(result, phase->intcon(24)); 5328 } else if (bt == T_BOOLEAN) { 5329 result = new AndINode(value, phase->intcon(0xFF)); 5330 } else if (bt == T_CHAR) { 5331 result = new AndINode(value,phase->intcon(0xFFFF)); 5332 } else { 5333 assert(bt == T_SHORT, "unexpected narrow type"); 5334 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 5335 result = new RShiftINode(result, phase->intcon(16)); 5336 } 5337 if (transform_res) { 5338 result = phase->transform(result); 5339 } 5340 return result; 5341 } 5342 5343 void Compile::record_method_not_compilable_oom() { 5344 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit()); 5345 }