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/compilerDefinitions.hpp" 37 #include "compiler/compilerOracle.hpp" 38 #include "compiler/compiler_globals.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 if (for_preload()) { 573 tty->print("[AP]%3d", _compile_id); 574 } else { 575 tty->print("%3d", _compile_id); 576 } 577 } 578 #endif 579 } 580 581 #ifndef PRODUCT 582 void Compile::print_phase(const char* phase_name) { 583 tty->print_cr("%u.\t%s", ++_phase_counter, phase_name); 584 } 585 586 void Compile::print_ideal_ir(const char* phase_name) { 587 // keep the following output all in one block 588 // This output goes directly to the tty, not the compiler log. 589 // To enable tools to match it up with the compilation activity, 590 // be sure to tag this tty output with the compile ID. 591 592 // Node dumping can cause a safepoint, which can break the tty lock. 593 // Buffer all node dumps, so that all safepoints happen before we lock. 594 ResourceMark rm; 595 stringStream ss; 596 597 if (_output == nullptr) { 598 ss.print_cr("AFTER: %s", phase_name); 599 // Print out all nodes in ascending order of index. 600 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss); 601 } else { 602 // Dump the node blockwise if we have a scheduling 603 _output->print_scheduling(&ss); 604 } 605 606 // Check that the lock is not broken by a safepoint. 607 NoSafepointVerifier nsv; 608 ttyLocker ttyl; 609 if (xtty != nullptr) { 610 xtty->head("ideal compile_id='%d'%s compile_phase='%s'", 611 compile_id(), 612 is_osr_compilation() ? " compile_kind='osr'" : 613 (for_preload() ? " compile_kind='AP'" : ""), 614 phase_name); 615 } 616 617 tty->print("%s", ss.as_string()); 618 619 if (xtty != nullptr) { 620 xtty->tail("ideal"); 621 } 622 } 623 #endif 624 625 // ============================================================================ 626 //------------------------------Compile standard------------------------------- 627 628 // Compile a method. entry_bci is -1 for normal compilations and indicates 629 // the continuation bci for on stack replacement. 630 631 632 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci, 633 Options options, DirectiveSet* directive) 634 : Phase(Compiler), 635 _compile_id(ci_env->compile_id()), 636 _options(options), 637 _method(target), 638 _entry_bci(osr_bci), 639 _ilt(nullptr), 640 _stub_function(nullptr), 641 _stub_name(nullptr), 642 _stub_id(-1), 643 _stub_entry_point(nullptr), 644 _max_node_limit(MaxNodeLimit), 645 _post_loop_opts_phase(false), 646 _merge_stores_phase(false), 647 _allow_macro_nodes(true), 648 _inlining_progress(false), 649 _inlining_incrementally(false), 650 _do_cleanup(false), 651 _has_reserved_stack_access(target->has_reserved_stack_access()), 652 #ifndef PRODUCT 653 _igv_idx(0), 654 _trace_opto_output(directive->TraceOptoOutputOption), 655 #endif 656 _has_method_handle_invokes(false), 657 _clinit_barrier_on_entry(false), 658 _has_clinit_barriers(false), 659 _stress_seed(0), 660 _comp_arena(mtCompiler, Arena::Tag::tag_comp), 661 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 662 _env(ci_env), 663 _directive(directive), 664 _log(ci_env->log()), 665 _first_failure_details(nullptr), 666 _intrinsics(comp_arena(), 0, 0, nullptr), 667 _macro_nodes(comp_arena(), 8, 0, nullptr), 668 _parse_predicates(comp_arena(), 8, 0, nullptr), 669 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr), 670 _expensive_nodes(comp_arena(), 8, 0, nullptr), 671 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr), 672 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr), 673 _unstable_if_traps(comp_arena(), 8, 0, nullptr), 674 _coarsened_locks(comp_arena(), 8, 0, nullptr), 675 _congraph(nullptr), 676 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 677 _unique(0), 678 _dead_node_count(0), 679 _dead_node_list(comp_arena()), 680 _node_arena_one(mtCompiler, Arena::Tag::tag_node), 681 _node_arena_two(mtCompiler, Arena::Tag::tag_node), 682 _node_arena(&_node_arena_one), 683 _mach_constant_base_node(nullptr), 684 _Compile_types(mtCompiler, Arena::Tag::tag_type), 685 _initial_gvn(nullptr), 686 _igvn_worklist(nullptr), 687 _types(nullptr), 688 _node_hash(nullptr), 689 _late_inlines(comp_arena(), 2, 0, nullptr), 690 _string_late_inlines(comp_arena(), 2, 0, nullptr), 691 _boxing_late_inlines(comp_arena(), 2, 0, nullptr), 692 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr), 693 _late_inlines_pos(0), 694 _number_of_mh_late_inlines(0), 695 _oom(false), 696 _replay_inline_data(nullptr), 697 _inline_printer(this), 698 _java_calls(0), 699 _inner_loops(0), 700 _interpreter_frame_size(0), 701 _output(nullptr) 702 #ifndef PRODUCT 703 , 704 _in_dump_cnt(0) 705 #endif 706 { 707 C = this; 708 CompileWrapper cw(this); 709 710 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 711 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false); 712 713 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 714 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 715 // We can always print a disassembly, either abstract (hex dump) or 716 // with the help of a suitable hsdis library. Thus, we should not 717 // couple print_assembly and print_opto_assembly controls. 718 // But: always print opto and regular assembly on compile command 'print'. 719 bool print_assembly = directive->PrintAssemblyOption; 720 set_print_assembly(print_opto_assembly || print_assembly); 721 #else 722 set_print_assembly(false); // must initialize. 723 #endif 724 725 #ifndef PRODUCT 726 set_parsed_irreducible_loop(false); 727 #endif 728 729 if (directive->ReplayInlineOption) { 730 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 731 } 732 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 733 set_print_intrinsics(directive->PrintIntrinsicsOption); 734 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 735 736 if (ProfileTraps) { 737 // Make sure the method being compiled gets its own MDO, 738 // so we can at least track the decompile_count(). 739 method()->ensure_method_data(); 740 } 741 742 if (StressLCM || StressGCM || StressIGVN || StressCCP || 743 StressIncrementalInlining || StressMacroExpansion || StressUnstableIfTraps || StressBailout || 744 StressLoopPeeling) { 745 initialize_stress_seed(directive); 746 } 747 748 Init(/*do_aliasing=*/ true); 749 750 print_compile_messages(); 751 752 _ilt = InlineTree::build_inline_tree_root(); 753 754 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 755 assert(num_alias_types() >= AliasIdxRaw, ""); 756 757 #define MINIMUM_NODE_HASH 1023 758 759 // GVN that will be run immediately on new nodes 760 uint estimated_size = method()->code_size()*4+64; 761 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 762 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 763 _types = new (comp_arena()) Type_Array(comp_arena()); 764 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size); 765 PhaseGVN gvn; 766 set_initial_gvn(&gvn); 767 768 { // Scope for timing the parser 769 TracePhase tp(_t_parser); 770 771 // Put top into the hash table ASAP. 772 initial_gvn()->transform(top()); 773 774 // Set up tf(), start(), and find a CallGenerator. 775 CallGenerator* cg = nullptr; 776 if (is_osr_compilation()) { 777 const TypeTuple *domain = StartOSRNode::osr_domain(); 778 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 779 init_tf(TypeFunc::make(domain, range)); 780 StartNode* s = new StartOSRNode(root(), domain); 781 initial_gvn()->set_type_bottom(s); 782 verify_start(s); 783 cg = CallGenerator::for_osr(method(), entry_bci()); 784 } else { 785 // Normal case. 786 init_tf(TypeFunc::make(method())); 787 StartNode* s = new StartNode(root(), tf()->domain()); 788 initial_gvn()->set_type_bottom(s); 789 verify_start(s); 790 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 791 // With java.lang.ref.reference.get() we must go through the 792 // intrinsic - even when get() is the root 793 // method of the compile - so that, if necessary, the value in 794 // the referent field of the reference object gets recorded by 795 // the pre-barrier code. 796 cg = find_intrinsic(method(), false); 797 } 798 if (cg == nullptr) { 799 float past_uses = method()->interpreter_invocation_count(); 800 float expected_uses = past_uses; 801 cg = CallGenerator::for_inline(method(), expected_uses); 802 } 803 } 804 if (failing()) return; 805 if (cg == nullptr) { 806 const char* reason = InlineTree::check_can_parse(method()); 807 assert(reason != nullptr, "expect reason for parse failure"); 808 stringStream ss; 809 ss.print("cannot parse method: %s", reason); 810 record_method_not_compilable(ss.as_string()); 811 return; 812 } 813 814 gvn.set_type(root(), root()->bottom_type()); 815 816 JVMState* jvms = build_start_state(start(), tf()); 817 if ((jvms = cg->generate(jvms)) == nullptr) { 818 assert(failure_reason() != nullptr, "expect reason for parse failure"); 819 stringStream ss; 820 ss.print("method parse failed: %s", failure_reason()); 821 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true)); 822 return; 823 } 824 GraphKit kit(jvms); 825 826 if (!kit.stopped()) { 827 // Accept return values, and transfer control we know not where. 828 // This is done by a special, unique ReturnNode bound to root. 829 return_values(kit.jvms()); 830 } 831 832 if (kit.has_exceptions()) { 833 // Any exceptions that escape from this call must be rethrown 834 // to whatever caller is dynamically above us on the stack. 835 // This is done by a special, unique RethrowNode bound to root. 836 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 837 } 838 839 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 840 841 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 842 inline_string_calls(true); 843 } 844 845 if (failing()) return; 846 847 // Remove clutter produced by parsing. 848 if (!failing()) { 849 ResourceMark rm; 850 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 851 } 852 } 853 854 // Note: Large methods are capped off in do_one_bytecode(). 855 if (failing()) return; 856 857 // After parsing, node notes are no longer automagic. 858 // They must be propagated by register_new_node_with_optimizer(), 859 // clone(), or the like. 860 set_default_node_notes(nullptr); 861 862 #ifndef PRODUCT 863 if (should_print_igv(1)) { 864 _igv_printer->print_inlining(); 865 } 866 #endif 867 868 if (failing()) return; 869 NOT_PRODUCT( verify_graph_edges(); ) 870 871 // Now optimize 872 Optimize(); 873 if (failing()) return; 874 NOT_PRODUCT( verify_graph_edges(); ) 875 876 #ifndef PRODUCT 877 if (should_print_ideal()) { 878 print_ideal_ir("print_ideal"); 879 } 880 #endif 881 882 #ifdef ASSERT 883 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 884 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 885 #endif 886 887 // Dump compilation data to replay it. 888 if (directive->DumpReplayOption) { 889 env()->dump_replay_data(_compile_id); 890 } 891 if (directive->DumpInlineOption && (ilt() != nullptr)) { 892 env()->dump_inline_data(_compile_id); 893 } 894 895 // Now that we know the size of all the monitors we can add a fixed slot 896 // for the original deopt pc. 897 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size); 898 set_fixed_slots(next_slot); 899 900 // Compute when to use implicit null checks. Used by matching trap based 901 // nodes and NullCheck optimization. 902 set_allowed_deopt_reasons(); 903 904 // Now generate code 905 Code_Gen(); 906 } 907 908 //------------------------------Compile---------------------------------------- 909 // Compile a runtime stub 910 Compile::Compile(ciEnv* ci_env, 911 TypeFunc_generator generator, 912 address stub_function, 913 const char* stub_name, 914 int stub_id, 915 int is_fancy_jump, 916 bool pass_tls, 917 bool return_pc, 918 DirectiveSet* directive) 919 : Phase(Compiler), 920 _compile_id(0), 921 _options(Options::for_runtime_stub()), 922 _method(nullptr), 923 _entry_bci(InvocationEntryBci), 924 _stub_function(stub_function), 925 _stub_name(stub_name), 926 _stub_id(stub_id), 927 _stub_entry_point(nullptr), 928 _max_node_limit(MaxNodeLimit), 929 _post_loop_opts_phase(false), 930 _merge_stores_phase(false), 931 _allow_macro_nodes(true), 932 _inlining_progress(false), 933 _inlining_incrementally(false), 934 _has_reserved_stack_access(false), 935 #ifndef PRODUCT 936 _igv_idx(0), 937 _trace_opto_output(directive->TraceOptoOutputOption), 938 #endif 939 _has_method_handle_invokes(false), 940 _clinit_barrier_on_entry(false), 941 _has_clinit_barriers(false), 942 _stress_seed(0), 943 _comp_arena(mtCompiler, Arena::Tag::tag_comp), 944 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 945 _env(ci_env), 946 _directive(directive), 947 _log(ci_env->log()), 948 _first_failure_details(nullptr), 949 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr), 950 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr), 951 _congraph(nullptr), 952 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 953 _unique(0), 954 _dead_node_count(0), 955 _dead_node_list(comp_arena()), 956 _node_arena_one(mtCompiler, Arena::Tag::tag_node), 957 _node_arena_two(mtCompiler, Arena::Tag::tag_node), 958 _node_arena(&_node_arena_one), 959 _mach_constant_base_node(nullptr), 960 _Compile_types(mtCompiler, Arena::Tag::tag_type), 961 _initial_gvn(nullptr), 962 _igvn_worklist(nullptr), 963 _types(nullptr), 964 _node_hash(nullptr), 965 _number_of_mh_late_inlines(0), 966 _oom(false), 967 _replay_inline_data(nullptr), 968 _inline_printer(this), 969 _java_calls(0), 970 _inner_loops(0), 971 _interpreter_frame_size(0), 972 _output(nullptr), 973 #ifndef PRODUCT 974 _in_dump_cnt(0), 975 #endif 976 _allowed_reasons(0) { 977 C = this; 978 979 // try to reuse an existing stub 980 { 981 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, _stub_id, stub_name); 982 if (blob != nullptr) { 983 RuntimeStub* rs = blob->as_runtime_stub(); 984 _stub_entry_point = rs->entry_point(); 985 return; 986 } 987 } 988 989 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false); 990 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false); 991 992 #ifndef PRODUCT 993 set_print_assembly(PrintFrameConverterAssembly); 994 set_parsed_irreducible_loop(false); 995 #else 996 set_print_assembly(false); // Must initialize. 997 #endif 998 set_has_irreducible_loop(false); // no loops 999 1000 CompileWrapper cw(this); 1001 Init(/*do_aliasing=*/ false); 1002 init_tf((*generator)()); 1003 1004 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 1005 _types = new (comp_arena()) Type_Array(comp_arena()); 1006 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255); 1007 1008 if (StressLCM || StressGCM || StressBailout) { 1009 initialize_stress_seed(directive); 1010 } 1011 1012 { 1013 PhaseGVN gvn; 1014 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1015 gvn.transform(top()); 1016 1017 GraphKit kit; 1018 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1019 } 1020 1021 NOT_PRODUCT( verify_graph_edges(); ) 1022 1023 Code_Gen(); 1024 } 1025 1026 Compile::~Compile() { 1027 delete _first_failure_details; 1028 }; 1029 1030 //------------------------------Init------------------------------------------- 1031 // Prepare for a single compilation 1032 void Compile::Init(bool aliasing) { 1033 _do_aliasing = aliasing; 1034 _unique = 0; 1035 _regalloc = nullptr; 1036 1037 _tf = nullptr; // filled in later 1038 _top = nullptr; // cached later 1039 _matcher = nullptr; // filled in later 1040 _cfg = nullptr; // filled in later 1041 1042 _node_note_array = nullptr; 1043 _default_node_notes = nullptr; 1044 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize() 1045 1046 _immutable_memory = nullptr; // filled in at first inquiry 1047 1048 #ifdef ASSERT 1049 _phase_optimize_finished = false; 1050 _phase_verify_ideal_loop = false; 1051 _exception_backedge = false; 1052 _type_verify = nullptr; 1053 #endif 1054 1055 // Globally visible Nodes 1056 // First set TOP to null to give safe behavior during creation of RootNode 1057 set_cached_top_node(nullptr); 1058 set_root(new RootNode()); 1059 // Now that you have a Root to point to, create the real TOP 1060 set_cached_top_node( new ConNode(Type::TOP) ); 1061 set_recent_alloc(nullptr, nullptr); 1062 1063 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1064 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1065 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1066 env()->set_dependencies(new Dependencies(env())); 1067 1068 _fixed_slots = 0; 1069 set_has_split_ifs(false); 1070 set_has_loops(false); // first approximation 1071 set_has_stringbuilder(false); 1072 set_has_boxed_value(false); 1073 _trap_can_recompile = false; // no traps emitted yet 1074 _major_progress = true; // start out assuming good things will happen 1075 set_has_unsafe_access(false); 1076 set_max_vector_size(0); 1077 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1078 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1079 set_decompile_count(0); 1080 1081 #ifndef PRODUCT 1082 _phase_counter = 0; 1083 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter)); 1084 #endif 1085 1086 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1087 _loop_opts_cnt = LoopOptsCount; 1088 set_do_inlining(Inline); 1089 set_max_inline_size(MaxInlineSize); 1090 set_freq_inline_size(FreqInlineSize); 1091 set_do_scheduling(OptoScheduling); 1092 1093 set_do_vector_loop(false); 1094 set_has_monitors(false); 1095 set_has_scoped_access(false); 1096 1097 if (AllowVectorizeOnDemand) { 1098 if (has_method() && _directive->VectorizeOption) { 1099 set_do_vector_loop(true); 1100 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());}) 1101 } else if (has_method() && method()->name() != nullptr && 1102 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1103 set_do_vector_loop(true); 1104 } 1105 } 1106 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1107 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());}) 1108 1109 _max_node_limit = _directive->MaxNodeLimitOption; 1110 1111 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && 1112 (method()->needs_clinit_barrier() || (do_clinit_barriers() && method()->is_static()))) { 1113 set_clinit_barrier_on_entry(true); 1114 if (do_clinit_barriers()) { 1115 set_has_clinit_barriers(true); // Entry clinit barrier is in prolog code. 1116 } 1117 } 1118 if (debug_info()->recording_non_safepoints()) { 1119 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1120 (comp_arena(), 8, 0, nullptr)); 1121 set_default_node_notes(Node_Notes::make(this)); 1122 } 1123 1124 const int grow_ats = 16; 1125 _max_alias_types = grow_ats; 1126 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1127 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1128 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1129 { 1130 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1131 } 1132 // Initialize the first few types. 1133 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr); 1134 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1135 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1136 _num_alias_types = AliasIdxRaw+1; 1137 // Zero out the alias type cache. 1138 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1139 // A null adr_type hits in the cache right away. Preload the right answer. 1140 probe_alias_cache(nullptr)->_index = AliasIdxTop; 1141 } 1142 1143 #ifdef ASSERT 1144 // Verify that the current StartNode is valid. 1145 void Compile::verify_start(StartNode* s) const { 1146 assert(failing_internal() || s == start(), "should be StartNode"); 1147 } 1148 #endif 1149 1150 /** 1151 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1152 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1153 * the ideal graph. 1154 */ 1155 StartNode* Compile::start() const { 1156 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason()); 1157 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1158 Node* start = root()->fast_out(i); 1159 if (start->is_Start()) { 1160 return start->as_Start(); 1161 } 1162 } 1163 fatal("Did not find Start node!"); 1164 return nullptr; 1165 } 1166 1167 //-------------------------------immutable_memory------------------------------------- 1168 // Access immutable memory 1169 Node* Compile::immutable_memory() { 1170 if (_immutable_memory != nullptr) { 1171 return _immutable_memory; 1172 } 1173 StartNode* s = start(); 1174 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1175 Node *p = s->fast_out(i); 1176 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1177 _immutable_memory = p; 1178 return _immutable_memory; 1179 } 1180 } 1181 ShouldNotReachHere(); 1182 return nullptr; 1183 } 1184 1185 //----------------------set_cached_top_node------------------------------------ 1186 // Install the cached top node, and make sure Node::is_top works correctly. 1187 void Compile::set_cached_top_node(Node* tn) { 1188 if (tn != nullptr) verify_top(tn); 1189 Node* old_top = _top; 1190 _top = tn; 1191 // Calling Node::setup_is_top allows the nodes the chance to adjust 1192 // their _out arrays. 1193 if (_top != nullptr) _top->setup_is_top(); 1194 if (old_top != nullptr) old_top->setup_is_top(); 1195 assert(_top == nullptr || top()->is_top(), ""); 1196 } 1197 1198 #ifdef ASSERT 1199 uint Compile::count_live_nodes_by_graph_walk() { 1200 Unique_Node_List useful(comp_arena()); 1201 // Get useful node list by walking the graph. 1202 identify_useful_nodes(useful); 1203 return useful.size(); 1204 } 1205 1206 void Compile::print_missing_nodes() { 1207 1208 // Return if CompileLog is null and PrintIdealNodeCount is false. 1209 if ((_log == nullptr) && (! PrintIdealNodeCount)) { 1210 return; 1211 } 1212 1213 // This is an expensive function. It is executed only when the user 1214 // specifies VerifyIdealNodeCount option or otherwise knows the 1215 // additional work that needs to be done to identify reachable nodes 1216 // by walking the flow graph and find the missing ones using 1217 // _dead_node_list. 1218 1219 Unique_Node_List useful(comp_arena()); 1220 // Get useful node list by walking the graph. 1221 identify_useful_nodes(useful); 1222 1223 uint l_nodes = C->live_nodes(); 1224 uint l_nodes_by_walk = useful.size(); 1225 1226 if (l_nodes != l_nodes_by_walk) { 1227 if (_log != nullptr) { 1228 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1229 _log->stamp(); 1230 _log->end_head(); 1231 } 1232 VectorSet& useful_member_set = useful.member_set(); 1233 int last_idx = l_nodes_by_walk; 1234 for (int i = 0; i < last_idx; i++) { 1235 if (useful_member_set.test(i)) { 1236 if (_dead_node_list.test(i)) { 1237 if (_log != nullptr) { 1238 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1239 } 1240 if (PrintIdealNodeCount) { 1241 // Print the log message to tty 1242 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1243 useful.at(i)->dump(); 1244 } 1245 } 1246 } 1247 else if (! _dead_node_list.test(i)) { 1248 if (_log != nullptr) { 1249 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1250 } 1251 if (PrintIdealNodeCount) { 1252 // Print the log message to tty 1253 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1254 } 1255 } 1256 } 1257 if (_log != nullptr) { 1258 _log->tail("mismatched_nodes"); 1259 } 1260 } 1261 } 1262 void Compile::record_modified_node(Node* n) { 1263 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) { 1264 _modified_nodes->push(n); 1265 } 1266 } 1267 1268 void Compile::remove_modified_node(Node* n) { 1269 if (_modified_nodes != nullptr) { 1270 _modified_nodes->remove(n); 1271 } 1272 } 1273 #endif 1274 1275 #ifndef PRODUCT 1276 void Compile::verify_top(Node* tn) const { 1277 if (tn != nullptr) { 1278 assert(tn->is_Con(), "top node must be a constant"); 1279 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1280 assert(tn->in(0) != nullptr, "must have live top node"); 1281 } 1282 } 1283 #endif 1284 1285 1286 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1287 1288 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1289 guarantee(arr != nullptr, ""); 1290 int num_blocks = arr->length(); 1291 if (grow_by < num_blocks) grow_by = num_blocks; 1292 int num_notes = grow_by * _node_notes_block_size; 1293 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1294 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1295 while (num_notes > 0) { 1296 arr->append(notes); 1297 notes += _node_notes_block_size; 1298 num_notes -= _node_notes_block_size; 1299 } 1300 assert(num_notes == 0, "exact multiple, please"); 1301 } 1302 1303 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1304 if (source == nullptr || dest == nullptr) return false; 1305 1306 if (dest->is_Con()) 1307 return false; // Do not push debug info onto constants. 1308 1309 #ifdef ASSERT 1310 // Leave a bread crumb trail pointing to the original node: 1311 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) { 1312 dest->set_debug_orig(source); 1313 } 1314 #endif 1315 1316 if (node_note_array() == nullptr) 1317 return false; // Not collecting any notes now. 1318 1319 // This is a copy onto a pre-existing node, which may already have notes. 1320 // If both nodes have notes, do not overwrite any pre-existing notes. 1321 Node_Notes* source_notes = node_notes_at(source->_idx); 1322 if (source_notes == nullptr || source_notes->is_clear()) return false; 1323 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1324 if (dest_notes == nullptr || dest_notes->is_clear()) { 1325 return set_node_notes_at(dest->_idx, source_notes); 1326 } 1327 1328 Node_Notes merged_notes = (*source_notes); 1329 // The order of operations here ensures that dest notes will win... 1330 merged_notes.update_from(dest_notes); 1331 return set_node_notes_at(dest->_idx, &merged_notes); 1332 } 1333 1334 1335 //--------------------------allow_range_check_smearing------------------------- 1336 // Gating condition for coalescing similar range checks. 1337 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1338 // single covering check that is at least as strong as any of them. 1339 // If the optimization succeeds, the simplified (strengthened) range check 1340 // will always succeed. If it fails, we will deopt, and then give up 1341 // on the optimization. 1342 bool Compile::allow_range_check_smearing() const { 1343 // If this method has already thrown a range-check, 1344 // assume it was because we already tried range smearing 1345 // and it failed. 1346 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1347 return !already_trapped; 1348 } 1349 1350 1351 //------------------------------flatten_alias_type----------------------------- 1352 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1353 assert(do_aliasing(), "Aliasing should be enabled"); 1354 int offset = tj->offset(); 1355 TypePtr::PTR ptr = tj->ptr(); 1356 1357 // Known instance (scalarizable allocation) alias only with itself. 1358 bool is_known_inst = tj->isa_oopptr() != nullptr && 1359 tj->is_oopptr()->is_known_instance(); 1360 1361 // Process weird unsafe references. 1362 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1363 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops"); 1364 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1365 tj = TypeOopPtr::BOTTOM; 1366 ptr = tj->ptr(); 1367 offset = tj->offset(); 1368 } 1369 1370 // Array pointers need some flattening 1371 const TypeAryPtr* ta = tj->isa_aryptr(); 1372 if (ta && ta->is_stable()) { 1373 // Erase stability property for alias analysis. 1374 tj = ta = ta->cast_to_stable(false); 1375 } 1376 if( ta && is_known_inst ) { 1377 if ( offset != Type::OffsetBot && 1378 offset > arrayOopDesc::length_offset_in_bytes() ) { 1379 offset = Type::OffsetBot; // Flatten constant access into array body only 1380 tj = ta = ta-> 1381 remove_speculative()-> 1382 cast_to_ptr_type(ptr)-> 1383 with_offset(offset); 1384 } 1385 } else if (ta) { 1386 // For arrays indexed by constant indices, we flatten the alias 1387 // space to include all of the array body. Only the header, klass 1388 // and array length can be accessed un-aliased. 1389 if( offset != Type::OffsetBot ) { 1390 if( ta->const_oop() ) { // MethodData* or Method* 1391 offset = Type::OffsetBot; // Flatten constant access into array body 1392 tj = ta = ta-> 1393 remove_speculative()-> 1394 cast_to_ptr_type(ptr)-> 1395 cast_to_exactness(false)-> 1396 with_offset(offset); 1397 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1398 // range is OK as-is. 1399 tj = ta = TypeAryPtr::RANGE; 1400 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1401 tj = TypeInstPtr::KLASS; // all klass loads look alike 1402 ta = TypeAryPtr::RANGE; // generic ignored junk 1403 ptr = TypePtr::BotPTR; 1404 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1405 tj = TypeInstPtr::MARK; 1406 ta = TypeAryPtr::RANGE; // generic ignored junk 1407 ptr = TypePtr::BotPTR; 1408 } else { // Random constant offset into array body 1409 offset = Type::OffsetBot; // Flatten constant access into array body 1410 tj = ta = ta-> 1411 remove_speculative()-> 1412 cast_to_ptr_type(ptr)-> 1413 cast_to_exactness(false)-> 1414 with_offset(offset); 1415 } 1416 } 1417 // Arrays of fixed size alias with arrays of unknown size. 1418 if (ta->size() != TypeInt::POS) { 1419 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1420 tj = ta = ta-> 1421 remove_speculative()-> 1422 cast_to_ptr_type(ptr)-> 1423 with_ary(tary)-> 1424 cast_to_exactness(false); 1425 } 1426 // Arrays of known objects become arrays of unknown objects. 1427 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1428 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1429 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1430 } 1431 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1432 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1433 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1434 } 1435 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1436 // cannot be distinguished by bytecode alone. 1437 if (ta->elem() == TypeInt::BOOL) { 1438 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1439 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1440 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1441 } 1442 // During the 2nd round of IterGVN, NotNull castings are removed. 1443 // Make sure the Bottom and NotNull variants alias the same. 1444 // Also, make sure exact and non-exact variants alias the same. 1445 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) { 1446 tj = ta = ta-> 1447 remove_speculative()-> 1448 cast_to_ptr_type(TypePtr::BotPTR)-> 1449 cast_to_exactness(false)-> 1450 with_offset(offset); 1451 } 1452 } 1453 1454 // Oop pointers need some flattening 1455 const TypeInstPtr *to = tj->isa_instptr(); 1456 if (to && to != TypeOopPtr::BOTTOM) { 1457 ciInstanceKlass* ik = to->instance_klass(); 1458 if( ptr == TypePtr::Constant ) { 1459 if (ik != ciEnv::current()->Class_klass() || 1460 offset < ik->layout_helper_size_in_bytes()) { 1461 // No constant oop pointers (such as Strings); they alias with 1462 // unknown strings. 1463 assert(!is_known_inst, "not scalarizable allocation"); 1464 tj = to = to-> 1465 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1466 remove_speculative()-> 1467 cast_to_ptr_type(TypePtr::BotPTR)-> 1468 cast_to_exactness(false); 1469 } 1470 } else if( is_known_inst ) { 1471 tj = to; // Keep NotNull and klass_is_exact for instance type 1472 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1473 // During the 2nd round of IterGVN, NotNull castings are removed. 1474 // Make sure the Bottom and NotNull variants alias the same. 1475 // Also, make sure exact and non-exact variants alias the same. 1476 tj = to = to-> 1477 remove_speculative()-> 1478 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1479 cast_to_ptr_type(TypePtr::BotPTR)-> 1480 cast_to_exactness(false); 1481 } 1482 if (to->speculative() != nullptr) { 1483 tj = to = to->remove_speculative(); 1484 } 1485 // Canonicalize the holder of this field 1486 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1487 // First handle header references such as a LoadKlassNode, even if the 1488 // object's klass is unloaded at compile time (4965979). 1489 if (!is_known_inst) { // Do it only for non-instance types 1490 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset); 1491 } 1492 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) { 1493 // Static fields are in the space above the normal instance 1494 // fields in the java.lang.Class instance. 1495 if (ik != ciEnv::current()->Class_klass()) { 1496 to = nullptr; 1497 tj = TypeOopPtr::BOTTOM; 1498 offset = tj->offset(); 1499 } 1500 } else { 1501 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset); 1502 assert(offset < canonical_holder->layout_helper_size_in_bytes(), ""); 1503 assert(tj->offset() == offset, "no change to offset expected"); 1504 bool xk = to->klass_is_exact(); 1505 int instance_id = to->instance_id(); 1506 1507 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder 1508 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only 1509 // its interfaces are included. 1510 if (xk && ik->equals(canonical_holder)) { 1511 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type"); 1512 } else { 1513 assert(xk || !is_known_inst, "Known instance should be exact type"); 1514 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id); 1515 } 1516 } 1517 } 1518 1519 // Klass pointers to object array klasses need some flattening 1520 const TypeKlassPtr *tk = tj->isa_klassptr(); 1521 if( tk ) { 1522 // If we are referencing a field within a Klass, we need 1523 // to assume the worst case of an Object. Both exact and 1524 // inexact types must flatten to the same alias class so 1525 // use NotNull as the PTR. 1526 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1527 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, 1528 env()->Object_klass(), 1529 offset); 1530 } 1531 1532 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) { 1533 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass()); 1534 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs 1535 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset); 1536 } else { 1537 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset); 1538 } 1539 } 1540 1541 // Check for precise loads from the primary supertype array and force them 1542 // to the supertype cache alias index. Check for generic array loads from 1543 // the primary supertype array and also force them to the supertype cache 1544 // alias index. Since the same load can reach both, we need to merge 1545 // these 2 disparate memories into the same alias class. Since the 1546 // primary supertype array is read-only, there's no chance of confusion 1547 // where we bypass an array load and an array store. 1548 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1549 if (offset == Type::OffsetBot || 1550 (offset >= primary_supers_offset && 1551 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1552 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1553 offset = in_bytes(Klass::secondary_super_cache_offset()); 1554 tj = tk = tk->with_offset(offset); 1555 } 1556 } 1557 1558 // Flatten all Raw pointers together. 1559 if (tj->base() == Type::RawPtr) 1560 tj = TypeRawPtr::BOTTOM; 1561 1562 if (tj->base() == Type::AnyPtr) 1563 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1564 1565 offset = tj->offset(); 1566 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1567 1568 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1569 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1570 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1571 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1572 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1573 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1574 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1575 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1576 assert( tj->ptr() != TypePtr::TopPTR && 1577 tj->ptr() != TypePtr::AnyNull && 1578 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1579 // assert( tj->ptr() != TypePtr::Constant || 1580 // tj->base() == Type::RawPtr || 1581 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1582 1583 return tj; 1584 } 1585 1586 void Compile::AliasType::Init(int i, const TypePtr* at) { 1587 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1588 _index = i; 1589 _adr_type = at; 1590 _field = nullptr; 1591 _element = nullptr; 1592 _is_rewritable = true; // default 1593 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr; 1594 if (atoop != nullptr && atoop->is_known_instance()) { 1595 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1596 _general_index = Compile::current()->get_alias_index(gt); 1597 } else { 1598 _general_index = 0; 1599 } 1600 } 1601 1602 BasicType Compile::AliasType::basic_type() const { 1603 if (element() != nullptr) { 1604 const Type* element = adr_type()->is_aryptr()->elem(); 1605 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1606 } if (field() != nullptr) { 1607 return field()->layout_type(); 1608 } else { 1609 return T_ILLEGAL; // unknown 1610 } 1611 } 1612 1613 //---------------------------------print_on------------------------------------ 1614 #ifndef PRODUCT 1615 void Compile::AliasType::print_on(outputStream* st) { 1616 if (index() < 10) 1617 st->print("@ <%d> ", index()); 1618 else st->print("@ <%d>", index()); 1619 st->print(is_rewritable() ? " " : " RO"); 1620 int offset = adr_type()->offset(); 1621 if (offset == Type::OffsetBot) 1622 st->print(" +any"); 1623 else st->print(" +%-3d", offset); 1624 st->print(" in "); 1625 adr_type()->dump_on(st); 1626 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1627 if (field() != nullptr && tjp) { 1628 if (tjp->is_instptr()->instance_klass() != field()->holder() || 1629 tjp->offset() != field()->offset_in_bytes()) { 1630 st->print(" != "); 1631 field()->print(); 1632 st->print(" ***"); 1633 } 1634 } 1635 } 1636 1637 void print_alias_types() { 1638 Compile* C = Compile::current(); 1639 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1640 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1641 C->alias_type(idx)->print_on(tty); 1642 tty->cr(); 1643 } 1644 } 1645 #endif 1646 1647 1648 //----------------------------probe_alias_cache-------------------------------- 1649 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1650 intptr_t key = (intptr_t) adr_type; 1651 key ^= key >> logAliasCacheSize; 1652 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1653 } 1654 1655 1656 //-----------------------------grow_alias_types-------------------------------- 1657 void Compile::grow_alias_types() { 1658 const int old_ats = _max_alias_types; // how many before? 1659 const int new_ats = old_ats; // how many more? 1660 const int grow_ats = old_ats+new_ats; // how many now? 1661 _max_alias_types = grow_ats; 1662 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1663 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1664 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1665 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1666 } 1667 1668 1669 //--------------------------------find_alias_type------------------------------ 1670 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1671 if (!do_aliasing()) { 1672 return alias_type(AliasIdxBot); 1673 } 1674 1675 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1676 if (ace->_adr_type == adr_type) { 1677 return alias_type(ace->_index); 1678 } 1679 1680 // Handle special cases. 1681 if (adr_type == nullptr) return alias_type(AliasIdxTop); 1682 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1683 1684 // Do it the slow way. 1685 const TypePtr* flat = flatten_alias_type(adr_type); 1686 1687 #ifdef ASSERT 1688 { 1689 ResourceMark rm; 1690 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1691 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1692 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1693 Type::str(adr_type)); 1694 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1695 const TypeOopPtr* foop = flat->is_oopptr(); 1696 // Scalarizable allocations have exact klass always. 1697 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1698 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1699 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1700 Type::str(foop), Type::str(xoop)); 1701 } 1702 } 1703 #endif 1704 1705 int idx = AliasIdxTop; 1706 for (int i = 0; i < num_alias_types(); i++) { 1707 if (alias_type(i)->adr_type() == flat) { 1708 idx = i; 1709 break; 1710 } 1711 } 1712 1713 if (idx == AliasIdxTop) { 1714 if (no_create) return nullptr; 1715 // Grow the array if necessary. 1716 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1717 // Add a new alias type. 1718 idx = _num_alias_types++; 1719 _alias_types[idx]->Init(idx, flat); 1720 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1721 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1722 if (flat->isa_instptr()) { 1723 if (flat->offset() == java_lang_Class::klass_offset() 1724 && flat->is_instptr()->instance_klass() == env()->Class_klass()) 1725 alias_type(idx)->set_rewritable(false); 1726 } 1727 if (flat->isa_aryptr()) { 1728 #ifdef ASSERT 1729 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1730 // (T_BYTE has the weakest alignment and size restrictions...) 1731 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1732 #endif 1733 if (flat->offset() == TypePtr::OffsetBot) { 1734 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1735 } 1736 } 1737 if (flat->isa_klassptr()) { 1738 if (UseCompactObjectHeaders) { 1739 if (flat->offset() == in_bytes(Klass::prototype_header_offset())) 1740 alias_type(idx)->set_rewritable(false); 1741 } 1742 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1743 alias_type(idx)->set_rewritable(false); 1744 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1745 alias_type(idx)->set_rewritable(false); 1746 if (flat->offset() == in_bytes(Klass::misc_flags_offset())) 1747 alias_type(idx)->set_rewritable(false); 1748 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1749 alias_type(idx)->set_rewritable(false); 1750 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1751 alias_type(idx)->set_rewritable(false); 1752 } 1753 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1754 // but the base pointer type is not distinctive enough to identify 1755 // references into JavaThread.) 1756 1757 // Check for final fields. 1758 const TypeInstPtr* tinst = flat->isa_instptr(); 1759 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1760 ciField* field; 1761 if (tinst->const_oop() != nullptr && 1762 tinst->instance_klass() == ciEnv::current()->Class_klass() && 1763 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) { 1764 // static field 1765 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1766 field = k->get_field_by_offset(tinst->offset(), true); 1767 } else { 1768 ciInstanceKlass *k = tinst->instance_klass(); 1769 field = k->get_field_by_offset(tinst->offset(), false); 1770 } 1771 assert(field == nullptr || 1772 original_field == nullptr || 1773 (field->holder() == original_field->holder() && 1774 field->offset_in_bytes() == original_field->offset_in_bytes() && 1775 field->is_static() == original_field->is_static()), "wrong field?"); 1776 // Set field() and is_rewritable() attributes. 1777 if (field != nullptr) alias_type(idx)->set_field(field); 1778 } 1779 } 1780 1781 // Fill the cache for next time. 1782 ace->_adr_type = adr_type; 1783 ace->_index = idx; 1784 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1785 1786 // Might as well try to fill the cache for the flattened version, too. 1787 AliasCacheEntry* face = probe_alias_cache(flat); 1788 if (face->_adr_type == nullptr) { 1789 face->_adr_type = flat; 1790 face->_index = idx; 1791 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1792 } 1793 1794 return alias_type(idx); 1795 } 1796 1797 1798 Compile::AliasType* Compile::alias_type(ciField* field) { 1799 const TypeOopPtr* t; 1800 if (field->is_static()) 1801 t = TypeInstPtr::make(field->holder()->java_mirror()); 1802 else 1803 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1804 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1805 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1806 return atp; 1807 } 1808 1809 1810 //------------------------------have_alias_type-------------------------------- 1811 bool Compile::have_alias_type(const TypePtr* adr_type) { 1812 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1813 if (ace->_adr_type == adr_type) { 1814 return true; 1815 } 1816 1817 // Handle special cases. 1818 if (adr_type == nullptr) return true; 1819 if (adr_type == TypePtr::BOTTOM) return true; 1820 1821 return find_alias_type(adr_type, true, nullptr) != nullptr; 1822 } 1823 1824 //-----------------------------must_alias-------------------------------------- 1825 // True if all values of the given address type are in the given alias category. 1826 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1827 if (alias_idx == AliasIdxBot) return true; // the universal category 1828 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP 1829 if (alias_idx == AliasIdxTop) return false; // the empty category 1830 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1831 1832 // the only remaining possible overlap is identity 1833 int adr_idx = get_alias_index(adr_type); 1834 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1835 assert(adr_idx == alias_idx || 1836 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1837 && adr_type != TypeOopPtr::BOTTOM), 1838 "should not be testing for overlap with an unsafe pointer"); 1839 return adr_idx == alias_idx; 1840 } 1841 1842 //------------------------------can_alias-------------------------------------- 1843 // True if any values of the given address type are in the given alias category. 1844 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1845 if (alias_idx == AliasIdxTop) return false; // the empty category 1846 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP 1847 // Known instance doesn't alias with bottom memory 1848 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1849 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1850 1851 // the only remaining possible overlap is identity 1852 int adr_idx = get_alias_index(adr_type); 1853 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1854 return adr_idx == alias_idx; 1855 } 1856 1857 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their 1858 // uncommon traps if no Runtime Predicates were created from the Parse Predicates. 1859 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) { 1860 if (parse_predicate_count() == 0) { 1861 return; 1862 } 1863 for (int i = 0; i < parse_predicate_count(); i++) { 1864 ParsePredicateNode* parse_predicate = _parse_predicates.at(i); 1865 parse_predicate->mark_useless(igvn); 1866 } 1867 _parse_predicates.clear(); 1868 } 1869 1870 void Compile::record_for_post_loop_opts_igvn(Node* n) { 1871 if (!n->for_post_loop_opts_igvn()) { 1872 assert(!_for_post_loop_igvn.contains(n), "duplicate"); 1873 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1874 _for_post_loop_igvn.append(n); 1875 } 1876 } 1877 1878 void Compile::remove_from_post_loop_opts_igvn(Node* n) { 1879 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1880 _for_post_loop_igvn.remove(n); 1881 } 1882 1883 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) { 1884 // Verify that all previous optimizations produced a valid graph 1885 // at least to this point, even if no loop optimizations were done. 1886 PhaseIdealLoop::verify(igvn); 1887 1888 if (has_loops() || _loop_opts_cnt > 0) { 1889 print_method(PHASE_AFTER_LOOP_OPTS, 2); 1890 } 1891 C->set_post_loop_opts_phase(); // no more loop opts allowed 1892 1893 assert(!C->major_progress(), "not cleared"); 1894 1895 if (_for_post_loop_igvn.length() > 0) { 1896 while (_for_post_loop_igvn.length() > 0) { 1897 Node* n = _for_post_loop_igvn.pop(); 1898 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1899 igvn._worklist.push(n); 1900 } 1901 igvn.optimize(); 1902 if (failing()) return; 1903 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed"); 1904 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now"); 1905 1906 // Sometimes IGVN sets major progress (e.g., when processing loop nodes). 1907 if (C->major_progress()) { 1908 C->clear_major_progress(); // ensure that major progress is now clear 1909 } 1910 } 1911 } 1912 1913 void Compile::record_for_merge_stores_igvn(Node* n) { 1914 if (!n->for_merge_stores_igvn()) { 1915 assert(!_for_merge_stores_igvn.contains(n), "duplicate"); 1916 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn); 1917 _for_merge_stores_igvn.append(n); 1918 } 1919 } 1920 1921 void Compile::remove_from_merge_stores_igvn(Node* n) { 1922 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn); 1923 _for_merge_stores_igvn.remove(n); 1924 } 1925 1926 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during 1927 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between 1928 // the stores, and we merge the wrong sequence of stores. 1929 // Example: 1930 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI 1931 // Apply MergeStores: 1932 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI 1933 // Remove more RangeChecks: 1934 // StoreI [ StoreL ] StoreI 1935 // But now it would have been better to do this instead: 1936 // [ StoreL ] [ StoreL ] 1937 // 1938 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round, 1939 // since we never unset _merge_stores_phase. 1940 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) { 1941 C->set_merge_stores_phase(); 1942 1943 if (_for_merge_stores_igvn.length() > 0) { 1944 while (_for_merge_stores_igvn.length() > 0) { 1945 Node* n = _for_merge_stores_igvn.pop(); 1946 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn); 1947 igvn._worklist.push(n); 1948 } 1949 igvn.optimize(); 1950 if (failing()) return; 1951 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed"); 1952 print_method(PHASE_AFTER_MERGE_STORES, 3); 1953 } 1954 } 1955 1956 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) { 1957 if (OptimizeUnstableIf) { 1958 _unstable_if_traps.append(trap); 1959 } 1960 } 1961 1962 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) { 1963 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) { 1964 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1965 Node* n = trap->uncommon_trap(); 1966 if (!useful.member(n)) { 1967 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 1968 } 1969 } 1970 } 1971 1972 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead 1973 // or fold-compares case. Return true if succeed or not found. 1974 // 1975 // In rare cases, the found trap has been processed. It is too late to delete it. Return 1976 // false and ask fold-compares to yield. 1977 // 1978 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused 1979 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path 1980 // when deoptimization does happen. 1981 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) { 1982 for (int i = 0; i < _unstable_if_traps.length(); ++i) { 1983 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1984 if (trap->uncommon_trap() == unc) { 1985 if (yield && trap->modified()) { 1986 return false; 1987 } 1988 _unstable_if_traps.delete_at(i); 1989 break; 1990 } 1991 } 1992 return true; 1993 } 1994 1995 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path. 1996 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering. 1997 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) { 1998 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) { 1999 UnstableIfTrap* trap = _unstable_if_traps.at(i); 2000 CallStaticJavaNode* unc = trap->uncommon_trap(); 2001 int next_bci = trap->next_bci(); 2002 bool modified = trap->modified(); 2003 2004 if (next_bci != -1 && !modified) { 2005 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!"); 2006 JVMState* jvms = unc->jvms(); 2007 ciMethod* method = jvms->method(); 2008 ciBytecodeStream iter(method); 2009 2010 iter.force_bci(jvms->bci()); 2011 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if"); 2012 Bytecodes::Code c = iter.cur_bc(); 2013 Node* lhs = nullptr; 2014 Node* rhs = nullptr; 2015 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) { 2016 lhs = unc->peek_operand(0); 2017 rhs = unc->peek_operand(1); 2018 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) { 2019 lhs = unc->peek_operand(0); 2020 } 2021 2022 ResourceMark rm; 2023 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci); 2024 assert(live_locals.is_valid(), "broken liveness info"); 2025 int len = (int)live_locals.size(); 2026 2027 for (int i = 0; i < len; i++) { 2028 Node* local = unc->local(jvms, i); 2029 // kill local using the liveness of next_bci. 2030 // give up when the local looks like an operand to secure reexecution. 2031 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) { 2032 uint idx = jvms->locoff() + i; 2033 #ifdef ASSERT 2034 if (PrintOpto && Verbose) { 2035 tty->print("[unstable_if] kill local#%d: ", idx); 2036 local->dump(); 2037 tty->cr(); 2038 } 2039 #endif 2040 igvn.replace_input_of(unc, idx, top()); 2041 modified = true; 2042 } 2043 } 2044 } 2045 2046 // keep the mondified trap for late query 2047 if (modified) { 2048 trap->set_modified(); 2049 } else { 2050 _unstable_if_traps.delete_at(i); 2051 } 2052 } 2053 igvn.optimize(); 2054 } 2055 2056 // StringOpts and late inlining of string methods 2057 void Compile::inline_string_calls(bool parse_time) { 2058 { 2059 // remove useless nodes to make the usage analysis simpler 2060 ResourceMark rm; 2061 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2062 } 2063 2064 { 2065 ResourceMark rm; 2066 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2067 PhaseStringOpts pso(initial_gvn()); 2068 print_method(PHASE_AFTER_STRINGOPTS, 3); 2069 } 2070 2071 // now inline anything that we skipped the first time around 2072 if (!parse_time) { 2073 _late_inlines_pos = _late_inlines.length(); 2074 } 2075 2076 while (_string_late_inlines.length() > 0) { 2077 CallGenerator* cg = _string_late_inlines.pop(); 2078 cg->do_late_inline(); 2079 if (failing()) return; 2080 } 2081 _string_late_inlines.trunc_to(0); 2082 } 2083 2084 // Late inlining of boxing methods 2085 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2086 if (_boxing_late_inlines.length() > 0) { 2087 assert(has_boxed_value(), "inconsistent"); 2088 2089 set_inlining_incrementally(true); 2090 2091 igvn_worklist()->ensure_empty(); // should be done with igvn 2092 2093 _late_inlines_pos = _late_inlines.length(); 2094 2095 while (_boxing_late_inlines.length() > 0) { 2096 CallGenerator* cg = _boxing_late_inlines.pop(); 2097 cg->do_late_inline(); 2098 if (failing()) return; 2099 } 2100 _boxing_late_inlines.trunc_to(0); 2101 2102 inline_incrementally_cleanup(igvn); 2103 2104 set_inlining_incrementally(false); 2105 } 2106 } 2107 2108 bool Compile::inline_incrementally_one() { 2109 assert(IncrementalInline, "incremental inlining should be on"); 2110 2111 TracePhase tp(_t_incrInline_inline); 2112 2113 set_inlining_progress(false); 2114 set_do_cleanup(false); 2115 2116 for (int i = 0; i < _late_inlines.length(); i++) { 2117 _late_inlines_pos = i+1; 2118 CallGenerator* cg = _late_inlines.at(i); 2119 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node()); 2120 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline(); 2121 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call 2122 cg->do_late_inline(); 2123 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed"); 2124 if (failing()) { 2125 return false; 2126 } else if (inlining_progress()) { 2127 _late_inlines_pos = i+1; // restore the position in case new elements were inserted 2128 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node()); 2129 break; // process one call site at a time 2130 } else { 2131 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node()); 2132 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) { 2133 // Avoid potential infinite loop if node already in the IGVN list 2134 assert(false, "scheduled for IGVN during inlining attempt"); 2135 } else { 2136 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt 2137 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass"); 2138 cg->call_node()->set_generator(cg); 2139 } 2140 } 2141 } else { 2142 // Ignore late inline direct calls when inlining is not allowed. 2143 // They are left in the late inline list when node budget is exhausted until the list is fully drained. 2144 } 2145 } 2146 // Remove processed elements. 2147 _late_inlines.remove_till(_late_inlines_pos); 2148 _late_inlines_pos = 0; 2149 2150 assert(inlining_progress() || _late_inlines.length() == 0, "no progress"); 2151 2152 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2153 2154 set_inlining_progress(false); 2155 set_do_cleanup(false); 2156 2157 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption; 2158 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup; 2159 } 2160 2161 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2162 { 2163 TracePhase tp(_t_incrInline_pru); 2164 ResourceMark rm; 2165 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2166 } 2167 { 2168 TracePhase tp(_t_incrInline_igvn); 2169 igvn.reset_from_gvn(initial_gvn()); 2170 igvn.optimize(); 2171 if (failing()) return; 2172 } 2173 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3); 2174 } 2175 2176 // Perform incremental inlining until bound on number of live nodes is reached 2177 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2178 TracePhase tp(_t_incrInline); 2179 2180 set_inlining_incrementally(true); 2181 uint low_live_nodes = 0; 2182 2183 while (_late_inlines.length() > 0) { 2184 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2185 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2186 TracePhase tp(_t_incrInline_ideal); 2187 // PhaseIdealLoop is expensive so we only try it once we are 2188 // out of live nodes and we only try it again if the previous 2189 // helped got the number of nodes down significantly 2190 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2191 if (failing()) return; 2192 low_live_nodes = live_nodes(); 2193 _major_progress = true; 2194 } 2195 2196 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2197 bool do_print_inlining = print_inlining() || print_intrinsics(); 2198 if (do_print_inlining || log() != nullptr) { 2199 // Print inlining message for candidates that we couldn't inline for lack of space. 2200 for (int i = 0; i < _late_inlines.length(); i++) { 2201 CallGenerator* cg = _late_inlines.at(i); 2202 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 2203 if (do_print_inlining) { 2204 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg); 2205 } 2206 log_late_inline_failure(cg, msg); 2207 } 2208 } 2209 break; // finish 2210 } 2211 } 2212 2213 igvn_worklist()->ensure_empty(); // should be done with igvn 2214 2215 while (inline_incrementally_one()) { 2216 assert(!failing_internal() || failure_is_artificial(), "inconsistent"); 2217 } 2218 if (failing()) return; 2219 2220 inline_incrementally_cleanup(igvn); 2221 2222 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 2223 2224 if (failing()) return; 2225 2226 if (_late_inlines.length() == 0) { 2227 break; // no more progress 2228 } 2229 } 2230 2231 igvn_worklist()->ensure_empty(); // should be done with igvn 2232 2233 if (_string_late_inlines.length() > 0) { 2234 assert(has_stringbuilder(), "inconsistent"); 2235 2236 inline_string_calls(false); 2237 2238 if (failing()) return; 2239 2240 inline_incrementally_cleanup(igvn); 2241 } 2242 2243 set_inlining_incrementally(false); 2244 } 2245 2246 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) { 2247 // "inlining_incrementally() == false" is used to signal that no inlining is allowed 2248 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details). 2249 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr" 2250 // as if "inlining_incrementally() == true" were set. 2251 assert(inlining_incrementally() == false, "not allowed"); 2252 assert(_modified_nodes == nullptr, "not allowed"); 2253 assert(_late_inlines.length() > 0, "sanity"); 2254 2255 while (_late_inlines.length() > 0) { 2256 igvn_worklist()->ensure_empty(); // should be done with igvn 2257 2258 while (inline_incrementally_one()) { 2259 assert(!failing_internal() || failure_is_artificial(), "inconsistent"); 2260 } 2261 if (failing()) return; 2262 2263 inline_incrementally_cleanup(igvn); 2264 } 2265 } 2266 2267 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2268 if (_loop_opts_cnt > 0) { 2269 while (major_progress() && (_loop_opts_cnt > 0)) { 2270 TracePhase tp(_t_idealLoop); 2271 PhaseIdealLoop::optimize(igvn, mode); 2272 _loop_opts_cnt--; 2273 if (failing()) return false; 2274 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2275 } 2276 } 2277 return true; 2278 } 2279 2280 // Remove edges from "root" to each SafePoint at a backward branch. 2281 // They were inserted during parsing (see add_safepoint()) to make 2282 // infinite loops without calls or exceptions visible to root, i.e., 2283 // useful. 2284 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2285 Node *r = root(); 2286 if (r != nullptr) { 2287 for (uint i = r->req(); i < r->len(); ++i) { 2288 Node *n = r->in(i); 2289 if (n != nullptr && n->is_SafePoint()) { 2290 r->rm_prec(i); 2291 if (n->outcnt() == 0) { 2292 igvn.remove_dead_node(n); 2293 } 2294 --i; 2295 } 2296 } 2297 // Parsing may have added top inputs to the root node (Path 2298 // leading to the Halt node proven dead). Make sure we get a 2299 // chance to clean them up. 2300 igvn._worklist.push(r); 2301 igvn.optimize(); 2302 } 2303 } 2304 2305 //------------------------------Optimize--------------------------------------- 2306 // Given a graph, optimize it. 2307 void Compile::Optimize() { 2308 TracePhase tp(_t_optimizer); 2309 2310 #ifndef PRODUCT 2311 if (env()->break_at_compile()) { 2312 BREAKPOINT; 2313 } 2314 2315 #endif 2316 2317 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2318 #ifdef ASSERT 2319 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2320 #endif 2321 2322 ResourceMark rm; 2323 2324 NOT_PRODUCT( verify_graph_edges(); ) 2325 2326 print_method(PHASE_AFTER_PARSING, 1); 2327 2328 { 2329 // Iterative Global Value Numbering, including ideal transforms 2330 // Initialize IterGVN with types and values from parse-time GVN 2331 PhaseIterGVN igvn(initial_gvn()); 2332 #ifdef ASSERT 2333 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2334 #endif 2335 { 2336 TracePhase tp(_t_iterGVN); 2337 igvn.optimize(); 2338 } 2339 2340 if (failing()) return; 2341 2342 print_method(PHASE_ITER_GVN1, 2); 2343 2344 process_for_unstable_if_traps(igvn); 2345 2346 if (failing()) return; 2347 2348 inline_incrementally(igvn); 2349 2350 print_method(PHASE_INCREMENTAL_INLINE, 2); 2351 2352 if (failing()) return; 2353 2354 if (eliminate_boxing()) { 2355 // Inline valueOf() methods now. 2356 inline_boxing_calls(igvn); 2357 2358 if (failing()) return; 2359 2360 if (AlwaysIncrementalInline || StressIncrementalInlining) { 2361 inline_incrementally(igvn); 2362 } 2363 2364 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2365 2366 if (failing()) return; 2367 } 2368 2369 // Remove the speculative part of types and clean up the graph from 2370 // the extra CastPP nodes whose only purpose is to carry them. Do 2371 // that early so that optimizations are not disrupted by the extra 2372 // CastPP nodes. 2373 remove_speculative_types(igvn); 2374 2375 if (failing()) return; 2376 2377 // No more new expensive nodes will be added to the list from here 2378 // so keep only the actual candidates for optimizations. 2379 cleanup_expensive_nodes(igvn); 2380 2381 if (failing()) return; 2382 2383 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2384 if (EnableVectorSupport && has_vbox_nodes()) { 2385 TracePhase tp(_t_vector); 2386 PhaseVector pv(igvn); 2387 pv.optimize_vector_boxes(); 2388 if (failing()) return; 2389 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2390 } 2391 assert(!has_vbox_nodes(), "sanity"); 2392 2393 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2394 Compile::TracePhase tp(_t_renumberLive); 2395 igvn_worklist()->ensure_empty(); // should be done with igvn 2396 { 2397 ResourceMark rm; 2398 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist()); 2399 } 2400 igvn.reset_from_gvn(initial_gvn()); 2401 igvn.optimize(); 2402 if (failing()) return; 2403 } 2404 2405 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop 2406 // safepoints 2407 remove_root_to_sfpts_edges(igvn); 2408 2409 if (failing()) return; 2410 2411 if (has_loops()) { 2412 print_method(PHASE_BEFORE_LOOP_OPTS, 2); 2413 } 2414 2415 // Perform escape analysis 2416 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) { 2417 if (has_loops()) { 2418 // Cleanup graph (remove dead nodes). 2419 TracePhase tp(_t_idealLoop); 2420 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2421 if (failing()) return; 2422 } 2423 bool progress; 2424 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2425 do { 2426 ConnectionGraph::do_analysis(this, &igvn); 2427 2428 if (failing()) return; 2429 2430 int mcount = macro_count(); // Record number of allocations and locks before IGVN 2431 2432 // Optimize out fields loads from scalar replaceable allocations. 2433 igvn.optimize(); 2434 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2435 2436 if (failing()) return; 2437 2438 if (congraph() != nullptr && macro_count() > 0) { 2439 TracePhase tp(_t_macroEliminate); 2440 PhaseMacroExpand mexp(igvn); 2441 mexp.eliminate_macro_nodes(); 2442 if (failing()) return; 2443 2444 igvn.set_delay_transform(false); 2445 igvn.optimize(); 2446 if (failing()) return; 2447 2448 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2449 } 2450 2451 ConnectionGraph::verify_ram_nodes(this, root()); 2452 if (failing()) return; 2453 2454 progress = do_iterative_escape_analysis() && 2455 (macro_count() < mcount) && 2456 ConnectionGraph::has_candidates(this); 2457 // Try again if candidates exist and made progress 2458 // by removing some allocations and/or locks. 2459 } while (progress); 2460 } 2461 2462 // Loop transforms on the ideal graph. Range Check Elimination, 2463 // peeling, unrolling, etc. 2464 2465 // Set loop opts counter 2466 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2467 { 2468 TracePhase tp(_t_idealLoop); 2469 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2470 _loop_opts_cnt--; 2471 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2472 if (failing()) return; 2473 } 2474 // Loop opts pass if partial peeling occurred in previous pass 2475 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2476 TracePhase tp(_t_idealLoop); 2477 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2478 _loop_opts_cnt--; 2479 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2480 if (failing()) return; 2481 } 2482 // Loop opts pass for loop-unrolling before CCP 2483 if(major_progress() && (_loop_opts_cnt > 0)) { 2484 TracePhase tp(_t_idealLoop); 2485 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2486 _loop_opts_cnt--; 2487 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2488 } 2489 if (!failing()) { 2490 // Verify that last round of loop opts produced a valid graph 2491 PhaseIdealLoop::verify(igvn); 2492 } 2493 } 2494 if (failing()) return; 2495 2496 // Conditional Constant Propagation; 2497 print_method(PHASE_BEFORE_CCP1, 2); 2498 PhaseCCP ccp( &igvn ); 2499 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2500 { 2501 TracePhase tp(_t_ccp); 2502 ccp.do_transform(); 2503 } 2504 print_method(PHASE_CCP1, 2); 2505 2506 assert( true, "Break here to ccp.dump_old2new_map()"); 2507 2508 // Iterative Global Value Numbering, including ideal transforms 2509 { 2510 TracePhase tp(_t_iterGVN2); 2511 igvn.reset_from_igvn(&ccp); 2512 igvn.optimize(); 2513 } 2514 print_method(PHASE_ITER_GVN2, 2); 2515 2516 if (failing()) return; 2517 2518 // Loop transforms on the ideal graph. Range Check Elimination, 2519 // peeling, unrolling, etc. 2520 if (!optimize_loops(igvn, LoopOptsDefault)) { 2521 return; 2522 } 2523 2524 if (failing()) return; 2525 2526 C->clear_major_progress(); // ensure that major progress is now clear 2527 2528 process_for_post_loop_opts_igvn(igvn); 2529 2530 process_for_merge_stores_igvn(igvn); 2531 2532 if (failing()) return; 2533 2534 #ifdef ASSERT 2535 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2536 #endif 2537 2538 { 2539 TracePhase tp(_t_macroExpand); 2540 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3); 2541 PhaseMacroExpand mex(igvn); 2542 if (mex.expand_macro_nodes()) { 2543 assert(failing(), "must bail out w/ explicit message"); 2544 return; 2545 } 2546 print_method(PHASE_AFTER_MACRO_EXPANSION, 2); 2547 } 2548 2549 { 2550 TracePhase tp(_t_barrierExpand); 2551 if (bs->expand_barriers(this, igvn)) { 2552 assert(failing(), "must bail out w/ explicit message"); 2553 return; 2554 } 2555 print_method(PHASE_BARRIER_EXPANSION, 2); 2556 } 2557 2558 if (C->max_vector_size() > 0) { 2559 C->optimize_logic_cones(igvn); 2560 igvn.optimize(); 2561 if (failing()) return; 2562 } 2563 2564 DEBUG_ONLY( _modified_nodes = nullptr; ) 2565 2566 assert(igvn._worklist.size() == 0, "not empty"); 2567 2568 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty"); 2569 2570 if (_late_inlines.length() > 0) { 2571 // More opportunities to optimize virtual and MH calls. 2572 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option. 2573 process_late_inline_calls_no_inline(igvn); 2574 if (failing()) return; 2575 } 2576 } // (End scope of igvn; run destructor if necessary for asserts.) 2577 2578 check_no_dead_use(); 2579 2580 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have 2581 // to remove hashes to unlock nodes for modifications. 2582 C->node_hash()->clear(); 2583 2584 // A method with only infinite loops has no edges entering loops from root 2585 { 2586 TracePhase tp(_t_graphReshaping); 2587 if (final_graph_reshaping()) { 2588 assert(failing(), "must bail out w/ explicit message"); 2589 return; 2590 } 2591 } 2592 2593 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2594 DEBUG_ONLY(set_phase_optimize_finished();) 2595 } 2596 2597 #ifdef ASSERT 2598 void Compile::check_no_dead_use() const { 2599 ResourceMark rm; 2600 Unique_Node_List wq; 2601 wq.push(root()); 2602 for (uint i = 0; i < wq.size(); ++i) { 2603 Node* n = wq.at(i); 2604 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2605 Node* u = n->fast_out(j); 2606 if (u->outcnt() == 0 && !u->is_Con()) { 2607 u->dump(); 2608 fatal("no reachable node should have no use"); 2609 } 2610 wq.push(u); 2611 } 2612 } 2613 } 2614 #endif 2615 2616 void Compile::inline_vector_reboxing_calls() { 2617 if (C->_vector_reboxing_late_inlines.length() > 0) { 2618 _late_inlines_pos = C->_late_inlines.length(); 2619 while (_vector_reboxing_late_inlines.length() > 0) { 2620 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 2621 cg->do_late_inline(); 2622 if (failing()) return; 2623 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node()); 2624 } 2625 _vector_reboxing_late_inlines.trunc_to(0); 2626 } 2627 } 2628 2629 bool Compile::has_vbox_nodes() { 2630 if (C->_vector_reboxing_late_inlines.length() > 0) { 2631 return true; 2632 } 2633 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 2634 Node * n = C->macro_node(macro_idx); 2635 assert(n->is_macro(), "only macro nodes expected here"); 2636 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 2637 return true; 2638 } 2639 } 2640 return false; 2641 } 2642 2643 //---------------------------- Bitwise operation packing optimization --------------------------- 2644 2645 static bool is_vector_unary_bitwise_op(Node* n) { 2646 return n->Opcode() == Op_XorV && 2647 VectorNode::is_vector_bitwise_not_pattern(n); 2648 } 2649 2650 static bool is_vector_binary_bitwise_op(Node* n) { 2651 switch (n->Opcode()) { 2652 case Op_AndV: 2653 case Op_OrV: 2654 return true; 2655 2656 case Op_XorV: 2657 return !is_vector_unary_bitwise_op(n); 2658 2659 default: 2660 return false; 2661 } 2662 } 2663 2664 static bool is_vector_ternary_bitwise_op(Node* n) { 2665 return n->Opcode() == Op_MacroLogicV; 2666 } 2667 2668 static bool is_vector_bitwise_op(Node* n) { 2669 return is_vector_unary_bitwise_op(n) || 2670 is_vector_binary_bitwise_op(n) || 2671 is_vector_ternary_bitwise_op(n); 2672 } 2673 2674 static bool is_vector_bitwise_cone_root(Node* n) { 2675 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) { 2676 return false; 2677 } 2678 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2679 if (is_vector_bitwise_op(n->fast_out(i))) { 2680 return false; 2681 } 2682 } 2683 return true; 2684 } 2685 2686 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) { 2687 uint cnt = 0; 2688 if (is_vector_bitwise_op(n)) { 2689 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req(); 2690 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2691 for (uint i = 1; i < inp_cnt; i++) { 2692 Node* in = n->in(i); 2693 bool skip = VectorNode::is_all_ones_vector(in); 2694 if (!skip && !inputs.member(in)) { 2695 inputs.push(in); 2696 cnt++; 2697 } 2698 } 2699 assert(cnt <= 1, "not unary"); 2700 } else { 2701 uint last_req = inp_cnt; 2702 if (is_vector_ternary_bitwise_op(n)) { 2703 last_req = inp_cnt - 1; // skip last input 2704 } 2705 for (uint i = 1; i < last_req; i++) { 2706 Node* def = n->in(i); 2707 if (!inputs.member(def)) { 2708 inputs.push(def); 2709 cnt++; 2710 } 2711 } 2712 } 2713 } else { // not a bitwise operations 2714 if (!inputs.member(n)) { 2715 inputs.push(n); 2716 cnt++; 2717 } 2718 } 2719 return cnt; 2720 } 2721 2722 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 2723 Unique_Node_List useful_nodes; 2724 C->identify_useful_nodes(useful_nodes); 2725 2726 for (uint i = 0; i < useful_nodes.size(); i++) { 2727 Node* n = useful_nodes.at(i); 2728 if (is_vector_bitwise_cone_root(n)) { 2729 list.push(n); 2730 } 2731 } 2732 } 2733 2734 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 2735 const TypeVect* vt, 2736 Unique_Node_List& partition, 2737 Unique_Node_List& inputs) { 2738 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 2739 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 2740 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 2741 2742 Node* in1 = inputs.at(0); 2743 Node* in2 = inputs.at(1); 2744 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 2745 2746 uint func = compute_truth_table(partition, inputs); 2747 2748 Node* pn = partition.at(partition.size() - 1); 2749 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2750 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt)); 2751 } 2752 2753 static uint extract_bit(uint func, uint pos) { 2754 return (func & (1 << pos)) >> pos; 2755 } 2756 2757 // 2758 // A macro logic node represents a truth table. It has 4 inputs, 2759 // First three inputs corresponds to 3 columns of a truth table 2760 // and fourth input captures the logic function. 2761 // 2762 // eg. fn = (in1 AND in2) OR in3; 2763 // 2764 // MacroNode(in1,in2,in3,fn) 2765 // 2766 // ----------------- 2767 // in1 in2 in3 fn 2768 // ----------------- 2769 // 0 0 0 0 2770 // 0 0 1 1 2771 // 0 1 0 0 2772 // 0 1 1 1 2773 // 1 0 0 0 2774 // 1 0 1 1 2775 // 1 1 0 1 2776 // 1 1 1 1 2777 // 2778 2779 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 2780 int res = 0; 2781 for (int i = 0; i < 8; i++) { 2782 int bit1 = extract_bit(in1, i); 2783 int bit2 = extract_bit(in2, i); 2784 int bit3 = extract_bit(in3, i); 2785 2786 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 2787 int func_bit = extract_bit(func, func_bit_pos); 2788 2789 res |= func_bit << i; 2790 } 2791 return res; 2792 } 2793 2794 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 2795 assert(n != nullptr, ""); 2796 assert(eval_map.contains(n), "absent"); 2797 return *(eval_map.get(n)); 2798 } 2799 2800 static void eval_operands(Node* n, 2801 uint& func1, uint& func2, uint& func3, 2802 ResourceHashtable<Node*,uint>& eval_map) { 2803 assert(is_vector_bitwise_op(n), ""); 2804 2805 if (is_vector_unary_bitwise_op(n)) { 2806 Node* opnd = n->in(1); 2807 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 2808 opnd = n->in(2); 2809 } 2810 func1 = eval_operand(opnd, eval_map); 2811 } else if (is_vector_binary_bitwise_op(n)) { 2812 func1 = eval_operand(n->in(1), eval_map); 2813 func2 = eval_operand(n->in(2), eval_map); 2814 } else { 2815 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 2816 func1 = eval_operand(n->in(1), eval_map); 2817 func2 = eval_operand(n->in(2), eval_map); 2818 func3 = eval_operand(n->in(3), eval_map); 2819 } 2820 } 2821 2822 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 2823 assert(inputs.size() <= 3, "sanity"); 2824 ResourceMark rm; 2825 uint res = 0; 2826 ResourceHashtable<Node*,uint> eval_map; 2827 2828 // Populate precomputed functions for inputs. 2829 // Each input corresponds to one column of 3 input truth-table. 2830 uint input_funcs[] = { 0xAA, // (_, _, c) -> c 2831 0xCC, // (_, b, _) -> b 2832 0xF0 }; // (a, _, _) -> a 2833 for (uint i = 0; i < inputs.size(); i++) { 2834 eval_map.put(inputs.at(i), input_funcs[2-i]); 2835 } 2836 2837 for (uint i = 0; i < partition.size(); i++) { 2838 Node* n = partition.at(i); 2839 2840 uint func1 = 0, func2 = 0, func3 = 0; 2841 eval_operands(n, func1, func2, func3, eval_map); 2842 2843 switch (n->Opcode()) { 2844 case Op_OrV: 2845 assert(func3 == 0, "not binary"); 2846 res = func1 | func2; 2847 break; 2848 case Op_AndV: 2849 assert(func3 == 0, "not binary"); 2850 res = func1 & func2; 2851 break; 2852 case Op_XorV: 2853 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2854 assert(func2 == 0 && func3 == 0, "not unary"); 2855 res = (~func1) & 0xFF; 2856 } else { 2857 assert(func3 == 0, "not binary"); 2858 res = func1 ^ func2; 2859 } 2860 break; 2861 case Op_MacroLogicV: 2862 // Ordering of inputs may change during evaluation of sub-tree 2863 // containing MacroLogic node as a child node, thus a re-evaluation 2864 // makes sure that function is evaluated in context of current 2865 // inputs. 2866 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 2867 break; 2868 2869 default: assert(false, "not supported: %s", n->Name()); 2870 } 2871 assert(res <= 0xFF, "invalid"); 2872 eval_map.put(n, res); 2873 } 2874 return res; 2875 } 2876 2877 // Criteria under which nodes gets packed into a macro logic node:- 2878 // 1) Parent and both child nodes are all unmasked or masked with 2879 // same predicates. 2880 // 2) Masked parent can be packed with left child if it is predicated 2881 // and both have same predicates. 2882 // 3) Masked parent can be packed with right child if its un-predicated 2883 // or has matching predication condition. 2884 // 4) An unmasked parent can be packed with an unmasked child. 2885 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2886 assert(partition.size() == 0, "not empty"); 2887 assert(inputs.size() == 0, "not empty"); 2888 if (is_vector_ternary_bitwise_op(n)) { 2889 return false; 2890 } 2891 2892 bool is_unary_op = is_vector_unary_bitwise_op(n); 2893 if (is_unary_op) { 2894 assert(collect_unique_inputs(n, inputs) == 1, "not unary"); 2895 return false; // too few inputs 2896 } 2897 2898 bool pack_left_child = true; 2899 bool pack_right_child = true; 2900 2901 bool left_child_LOP = is_vector_bitwise_op(n->in(1)); 2902 bool right_child_LOP = is_vector_bitwise_op(n->in(2)); 2903 2904 int left_child_input_cnt = 0; 2905 int right_child_input_cnt = 0; 2906 2907 bool parent_is_predicated = n->is_predicated_vector(); 2908 bool left_child_predicated = n->in(1)->is_predicated_vector(); 2909 bool right_child_predicated = n->in(2)->is_predicated_vector(); 2910 2911 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr; 2912 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2913 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2914 2915 do { 2916 if (pack_left_child && left_child_LOP && 2917 ((!parent_is_predicated && !left_child_predicated) || 2918 ((parent_is_predicated && left_child_predicated && 2919 parent_pred == left_child_pred)))) { 2920 partition.push(n->in(1)); 2921 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs); 2922 } else { 2923 inputs.push(n->in(1)); 2924 left_child_input_cnt = 1; 2925 } 2926 2927 if (pack_right_child && right_child_LOP && 2928 (!right_child_predicated || 2929 (right_child_predicated && parent_is_predicated && 2930 parent_pred == right_child_pred))) { 2931 partition.push(n->in(2)); 2932 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs); 2933 } else { 2934 inputs.push(n->in(2)); 2935 right_child_input_cnt = 1; 2936 } 2937 2938 if (inputs.size() > 3) { 2939 assert(partition.size() > 0, ""); 2940 inputs.clear(); 2941 partition.clear(); 2942 if (left_child_input_cnt > right_child_input_cnt) { 2943 pack_left_child = false; 2944 } else { 2945 pack_right_child = false; 2946 } 2947 } else { 2948 break; 2949 } 2950 } while(true); 2951 2952 if(partition.size()) { 2953 partition.push(n); 2954 } 2955 2956 return (partition.size() == 2 || partition.size() == 3) && 2957 (inputs.size() == 2 || inputs.size() == 3); 2958 } 2959 2960 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 2961 assert(is_vector_bitwise_op(n), "not a root"); 2962 2963 visited.set(n->_idx); 2964 2965 // 1) Do a DFS walk over the logic cone. 2966 for (uint i = 1; i < n->req(); i++) { 2967 Node* in = n->in(i); 2968 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 2969 process_logic_cone_root(igvn, in, visited); 2970 } 2971 } 2972 2973 // 2) Bottom up traversal: Merge node[s] with 2974 // the parent to form macro logic node. 2975 Unique_Node_List partition; 2976 Unique_Node_List inputs; 2977 if (compute_logic_cone(n, partition, inputs)) { 2978 const TypeVect* vt = n->bottom_type()->is_vect(); 2979 Node* pn = partition.at(partition.size() - 1); 2980 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2981 if (mask == nullptr || 2982 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) { 2983 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 2984 VectorNode::trace_new_vector(macro_logic, "MacroLogic"); 2985 igvn.replace_node(n, macro_logic); 2986 } 2987 } 2988 } 2989 2990 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 2991 ResourceMark rm; 2992 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 2993 Unique_Node_List list; 2994 collect_logic_cone_roots(list); 2995 2996 while (list.size() > 0) { 2997 Node* n = list.pop(); 2998 const TypeVect* vt = n->bottom_type()->is_vect(); 2999 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 3000 if (supported) { 3001 VectorSet visited(comp_arena()); 3002 process_logic_cone_root(igvn, n, visited); 3003 } 3004 } 3005 } 3006 } 3007 3008 //------------------------------Code_Gen--------------------------------------- 3009 // Given a graph, generate code for it 3010 void Compile::Code_Gen() { 3011 if (failing()) { 3012 return; 3013 } 3014 3015 // Perform instruction selection. You might think we could reclaim Matcher 3016 // memory PDQ, but actually the Matcher is used in generating spill code. 3017 // Internals of the Matcher (including some VectorSets) must remain live 3018 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 3019 // set a bit in reclaimed memory. 3020 3021 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 3022 // nodes. Mapping is only valid at the root of each matched subtree. 3023 NOT_PRODUCT( verify_graph_edges(); ) 3024 3025 Matcher matcher; 3026 _matcher = &matcher; 3027 { 3028 TracePhase tp(_t_matcher); 3029 matcher.match(); 3030 if (failing()) { 3031 return; 3032 } 3033 } 3034 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 3035 // nodes. Mapping is only valid at the root of each matched subtree. 3036 NOT_PRODUCT( verify_graph_edges(); ) 3037 3038 // If you have too many nodes, or if matching has failed, bail out 3039 check_node_count(0, "out of nodes matching instructions"); 3040 if (failing()) { 3041 return; 3042 } 3043 3044 print_method(PHASE_MATCHING, 2); 3045 3046 // Build a proper-looking CFG 3047 PhaseCFG cfg(node_arena(), root(), matcher); 3048 if (failing()) { 3049 return; 3050 } 3051 _cfg = &cfg; 3052 { 3053 TracePhase tp(_t_scheduler); 3054 bool success = cfg.do_global_code_motion(); 3055 if (!success) { 3056 return; 3057 } 3058 3059 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 3060 NOT_PRODUCT( verify_graph_edges(); ) 3061 cfg.verify(); 3062 if (failing()) { 3063 return; 3064 } 3065 } 3066 3067 PhaseChaitin regalloc(unique(), cfg, matcher, false); 3068 _regalloc = ®alloc; 3069 { 3070 TracePhase tp(_t_registerAllocation); 3071 // Perform register allocation. After Chaitin, use-def chains are 3072 // no longer accurate (at spill code) and so must be ignored. 3073 // Node->LRG->reg mappings are still accurate. 3074 _regalloc->Register_Allocate(); 3075 3076 // Bail out if the allocator builds too many nodes 3077 if (failing()) { 3078 return; 3079 } 3080 3081 print_method(PHASE_REGISTER_ALLOCATION, 2); 3082 } 3083 3084 // Prior to register allocation we kept empty basic blocks in case the 3085 // the allocator needed a place to spill. After register allocation we 3086 // are not adding any new instructions. If any basic block is empty, we 3087 // can now safely remove it. 3088 { 3089 TracePhase tp(_t_blockOrdering); 3090 cfg.remove_empty_blocks(); 3091 if (do_freq_based_layout()) { 3092 PhaseBlockLayout layout(cfg); 3093 } else { 3094 cfg.set_loop_alignment(); 3095 } 3096 cfg.fixup_flow(); 3097 cfg.remove_unreachable_blocks(); 3098 cfg.verify_dominator_tree(); 3099 print_method(PHASE_BLOCK_ORDERING, 3); 3100 } 3101 3102 // Apply peephole optimizations 3103 if( OptoPeephole ) { 3104 TracePhase tp(_t_peephole); 3105 PhasePeephole peep( _regalloc, cfg); 3106 peep.do_transform(); 3107 print_method(PHASE_PEEPHOLE, 3); 3108 } 3109 3110 // Do late expand if CPU requires this. 3111 if (Matcher::require_postalloc_expand) { 3112 TracePhase tp(_t_postalloc_expand); 3113 cfg.postalloc_expand(_regalloc); 3114 print_method(PHASE_POSTALLOC_EXPAND, 3); 3115 } 3116 3117 #ifdef ASSERT 3118 { 3119 CompilationMemoryStatistic::do_test_allocations(); 3120 if (failing()) return; 3121 } 3122 #endif 3123 3124 // Convert Nodes to instruction bits in a buffer 3125 { 3126 TracePhase tp(_t_output); 3127 PhaseOutput output; 3128 output.Output(); 3129 if (failing()) return; 3130 output.install(); 3131 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here 3132 } 3133 3134 // He's dead, Jim. 3135 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 3136 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 3137 } 3138 3139 //------------------------------Final_Reshape_Counts--------------------------- 3140 // This class defines counters to help identify when a method 3141 // may/must be executed using hardware with only 24-bit precision. 3142 struct Final_Reshape_Counts : public StackObj { 3143 int _call_count; // count non-inlined 'common' calls 3144 int _float_count; // count float ops requiring 24-bit precision 3145 int _double_count; // count double ops requiring more precision 3146 int _java_call_count; // count non-inlined 'java' calls 3147 int _inner_loop_count; // count loops which need alignment 3148 VectorSet _visited; // Visitation flags 3149 Node_List _tests; // Set of IfNodes & PCTableNodes 3150 3151 Final_Reshape_Counts() : 3152 _call_count(0), _float_count(0), _double_count(0), 3153 _java_call_count(0), _inner_loop_count(0) { } 3154 3155 void inc_call_count () { _call_count ++; } 3156 void inc_float_count () { _float_count ++; } 3157 void inc_double_count() { _double_count++; } 3158 void inc_java_call_count() { _java_call_count++; } 3159 void inc_inner_loop_count() { _inner_loop_count++; } 3160 3161 int get_call_count () const { return _call_count ; } 3162 int get_float_count () const { return _float_count ; } 3163 int get_double_count() const { return _double_count; } 3164 int get_java_call_count() const { return _java_call_count; } 3165 int get_inner_loop_count() const { return _inner_loop_count; } 3166 }; 3167 3168 //------------------------------final_graph_reshaping_impl---------------------- 3169 // Implement items 1-5 from final_graph_reshaping below. 3170 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3171 3172 if ( n->outcnt() == 0 ) return; // dead node 3173 uint nop = n->Opcode(); 3174 3175 // Check for 2-input instruction with "last use" on right input. 3176 // Swap to left input. Implements item (2). 3177 if( n->req() == 3 && // two-input instruction 3178 n->in(1)->outcnt() > 1 && // left use is NOT a last use 3179 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 3180 n->in(2)->outcnt() == 1 &&// right use IS a last use 3181 !n->in(2)->is_Con() ) { // right use is not a constant 3182 // Check for commutative opcode 3183 switch( nop ) { 3184 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 3185 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 3186 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 3187 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 3188 case Op_AndL: case Op_XorL: case Op_OrL: 3189 case Op_AndI: case Op_XorI: case Op_OrI: { 3190 // Move "last use" input to left by swapping inputs 3191 n->swap_edges(1, 2); 3192 break; 3193 } 3194 default: 3195 break; 3196 } 3197 } 3198 3199 #ifdef ASSERT 3200 if( n->is_Mem() ) { 3201 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 3202 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw || 3203 // oop will be recorded in oop map if load crosses safepoint 3204 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 3205 LoadNode::is_immutable_value(n->in(MemNode::Address)))), 3206 "raw memory operations should have control edge"); 3207 } 3208 if (n->is_MemBar()) { 3209 MemBarNode* mb = n->as_MemBar(); 3210 if (mb->trailing_store() || mb->trailing_load_store()) { 3211 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 3212 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 3213 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 3214 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 3215 } else if (mb->leading()) { 3216 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 3217 } 3218 } 3219 #endif 3220 // Count FPU ops and common calls, implements item (3) 3221 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes); 3222 if (!gc_handled) { 3223 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes); 3224 } 3225 3226 // Collect CFG split points 3227 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3228 frc._tests.push(n); 3229 } 3230 } 3231 3232 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) { 3233 if (!UseDivMod) { 3234 return; 3235 } 3236 3237 // Check if "a % b" and "a / b" both exist 3238 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned)); 3239 if (d == nullptr) { 3240 return; 3241 } 3242 3243 // Replace them with a fused divmod if supported 3244 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) { 3245 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned); 3246 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero. 3247 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the 3248 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency 3249 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the 3250 // DivMod node so the dependency is not lost. 3251 divmod->add_prec_from(n); 3252 divmod->add_prec_from(d); 3253 d->subsume_by(divmod->div_proj(), this); 3254 n->subsume_by(divmod->mod_proj(), this); 3255 } else { 3256 // Replace "a % b" with "a - ((a / b) * b)" 3257 Node* mult = MulNode::make(d, d->in(2), bt); 3258 Node* sub = SubNode::make(d->in(1), mult, bt); 3259 n->subsume_by(sub, this); 3260 } 3261 } 3262 3263 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) { 3264 switch( nop ) { 3265 // Count all float operations that may use FPU 3266 case Op_AddF: 3267 case Op_SubF: 3268 case Op_MulF: 3269 case Op_DivF: 3270 case Op_NegF: 3271 case Op_ModF: 3272 case Op_ConvI2F: 3273 case Op_ConF: 3274 case Op_CmpF: 3275 case Op_CmpF3: 3276 case Op_StoreF: 3277 case Op_LoadF: 3278 // case Op_ConvL2F: // longs are split into 32-bit halves 3279 frc.inc_float_count(); 3280 break; 3281 3282 case Op_ConvF2D: 3283 case Op_ConvD2F: 3284 frc.inc_float_count(); 3285 frc.inc_double_count(); 3286 break; 3287 3288 // Count all double operations that may use FPU 3289 case Op_AddD: 3290 case Op_SubD: 3291 case Op_MulD: 3292 case Op_DivD: 3293 case Op_NegD: 3294 case Op_ModD: 3295 case Op_ConvI2D: 3296 case Op_ConvD2I: 3297 // case Op_ConvL2D: // handled by leaf call 3298 // case Op_ConvD2L: // handled by leaf call 3299 case Op_ConD: 3300 case Op_CmpD: 3301 case Op_CmpD3: 3302 case Op_StoreD: 3303 case Op_LoadD: 3304 case Op_LoadD_unaligned: 3305 frc.inc_double_count(); 3306 break; 3307 case Op_Opaque1: // Remove Opaque Nodes before matching 3308 n->subsume_by(n->in(1), this); 3309 break; 3310 case Op_CallStaticJava: 3311 case Op_CallJava: 3312 case Op_CallDynamicJava: 3313 frc.inc_java_call_count(); // Count java call site; 3314 case Op_CallRuntime: 3315 case Op_CallLeaf: 3316 case Op_CallLeafVector: 3317 case Op_CallLeafNoFP: { 3318 assert (n->is_Call(), ""); 3319 CallNode *call = n->as_Call(); 3320 // Count call sites where the FP mode bit would have to be flipped. 3321 // Do not count uncommon runtime calls: 3322 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3323 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3324 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3325 frc.inc_call_count(); // Count the call site 3326 } else { // See if uncommon argument is shared 3327 Node *n = call->in(TypeFunc::Parms); 3328 int nop = n->Opcode(); 3329 // Clone shared simple arguments to uncommon calls, item (1). 3330 if (n->outcnt() > 1 && 3331 !n->is_Proj() && 3332 nop != Op_CreateEx && 3333 nop != Op_CheckCastPP && 3334 nop != Op_DecodeN && 3335 nop != Op_DecodeNKlass && 3336 !n->is_Mem() && 3337 !n->is_Phi()) { 3338 Node *x = n->clone(); 3339 call->set_req(TypeFunc::Parms, x); 3340 } 3341 } 3342 break; 3343 } 3344 case Op_StoreB: 3345 case Op_StoreC: 3346 case Op_StoreI: 3347 case Op_StoreL: 3348 case Op_CompareAndSwapB: 3349 case Op_CompareAndSwapS: 3350 case Op_CompareAndSwapI: 3351 case Op_CompareAndSwapL: 3352 case Op_CompareAndSwapP: 3353 case Op_CompareAndSwapN: 3354 case Op_WeakCompareAndSwapB: 3355 case Op_WeakCompareAndSwapS: 3356 case Op_WeakCompareAndSwapI: 3357 case Op_WeakCompareAndSwapL: 3358 case Op_WeakCompareAndSwapP: 3359 case Op_WeakCompareAndSwapN: 3360 case Op_CompareAndExchangeB: 3361 case Op_CompareAndExchangeS: 3362 case Op_CompareAndExchangeI: 3363 case Op_CompareAndExchangeL: 3364 case Op_CompareAndExchangeP: 3365 case Op_CompareAndExchangeN: 3366 case Op_GetAndAddS: 3367 case Op_GetAndAddB: 3368 case Op_GetAndAddI: 3369 case Op_GetAndAddL: 3370 case Op_GetAndSetS: 3371 case Op_GetAndSetB: 3372 case Op_GetAndSetI: 3373 case Op_GetAndSetL: 3374 case Op_GetAndSetP: 3375 case Op_GetAndSetN: 3376 case Op_StoreP: 3377 case Op_StoreN: 3378 case Op_StoreNKlass: 3379 case Op_LoadB: 3380 case Op_LoadUB: 3381 case Op_LoadUS: 3382 case Op_LoadI: 3383 case Op_LoadKlass: 3384 case Op_LoadNKlass: 3385 case Op_LoadL: 3386 case Op_LoadL_unaligned: 3387 case Op_LoadP: 3388 case Op_LoadN: 3389 case Op_LoadRange: 3390 case Op_LoadS: 3391 break; 3392 3393 case Op_AddP: { // Assert sane base pointers 3394 Node *addp = n->in(AddPNode::Address); 3395 assert( !addp->is_AddP() || 3396 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3397 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3398 "Base pointers must match (addp %u)", addp->_idx ); 3399 #ifdef _LP64 3400 if ((UseCompressedOops || UseCompressedClassPointers) && 3401 addp->Opcode() == Op_ConP && 3402 addp == n->in(AddPNode::Base) && 3403 n->in(AddPNode::Offset)->is_Con()) { 3404 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3405 // on the platform and on the compressed oops mode. 3406 // Use addressing with narrow klass to load with offset on x86. 3407 // Some platforms can use the constant pool to load ConP. 3408 // Do this transformation here since IGVN will convert ConN back to ConP. 3409 const Type* t = addp->bottom_type(); 3410 bool is_oop = t->isa_oopptr() != nullptr; 3411 bool is_klass = t->isa_klassptr() != nullptr; 3412 3413 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) || 3414 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() && 3415 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) { 3416 Node* nn = nullptr; 3417 3418 int op = is_oop ? Op_ConN : Op_ConNKlass; 3419 3420 // Look for existing ConN node of the same exact type. 3421 Node* r = root(); 3422 uint cnt = r->outcnt(); 3423 for (uint i = 0; i < cnt; i++) { 3424 Node* m = r->raw_out(i); 3425 if (m!= nullptr && m->Opcode() == op && 3426 m->bottom_type()->make_ptr() == t) { 3427 nn = m; 3428 break; 3429 } 3430 } 3431 if (nn != nullptr) { 3432 // Decode a narrow oop to match address 3433 // [R12 + narrow_oop_reg<<3 + offset] 3434 if (is_oop) { 3435 nn = new DecodeNNode(nn, t); 3436 } else { 3437 nn = new DecodeNKlassNode(nn, t); 3438 } 3439 // Check for succeeding AddP which uses the same Base. 3440 // Otherwise we will run into the assertion above when visiting that guy. 3441 for (uint i = 0; i < n->outcnt(); ++i) { 3442 Node *out_i = n->raw_out(i); 3443 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3444 out_i->set_req(AddPNode::Base, nn); 3445 #ifdef ASSERT 3446 for (uint j = 0; j < out_i->outcnt(); ++j) { 3447 Node *out_j = out_i->raw_out(j); 3448 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3449 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3450 } 3451 #endif 3452 } 3453 } 3454 n->set_req(AddPNode::Base, nn); 3455 n->set_req(AddPNode::Address, nn); 3456 if (addp->outcnt() == 0) { 3457 addp->disconnect_inputs(this); 3458 } 3459 } 3460 } 3461 } 3462 #endif 3463 break; 3464 } 3465 3466 case Op_CastPP: { 3467 // Remove CastPP nodes to gain more freedom during scheduling but 3468 // keep the dependency they encode as control or precedence edges 3469 // (if control is set already) on memory operations. Some CastPP 3470 // nodes don't have a control (don't carry a dependency): skip 3471 // those. 3472 if (n->in(0) != nullptr) { 3473 ResourceMark rm; 3474 Unique_Node_List wq; 3475 wq.push(n); 3476 for (uint next = 0; next < wq.size(); ++next) { 3477 Node *m = wq.at(next); 3478 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3479 Node* use = m->fast_out(i); 3480 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3481 use->ensure_control_or_add_prec(n->in(0)); 3482 } else { 3483 switch(use->Opcode()) { 3484 case Op_AddP: 3485 case Op_DecodeN: 3486 case Op_DecodeNKlass: 3487 case Op_CheckCastPP: 3488 case Op_CastPP: 3489 wq.push(use); 3490 break; 3491 } 3492 } 3493 } 3494 } 3495 } 3496 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3497 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3498 Node* in1 = n->in(1); 3499 const Type* t = n->bottom_type(); 3500 Node* new_in1 = in1->clone(); 3501 new_in1->as_DecodeN()->set_type(t); 3502 3503 if (!Matcher::narrow_oop_use_complex_address()) { 3504 // 3505 // x86, ARM and friends can handle 2 adds in addressing mode 3506 // and Matcher can fold a DecodeN node into address by using 3507 // a narrow oop directly and do implicit null check in address: 3508 // 3509 // [R12 + narrow_oop_reg<<3 + offset] 3510 // NullCheck narrow_oop_reg 3511 // 3512 // On other platforms (Sparc) we have to keep new DecodeN node and 3513 // use it to do implicit null check in address: 3514 // 3515 // decode_not_null narrow_oop_reg, base_reg 3516 // [base_reg + offset] 3517 // NullCheck base_reg 3518 // 3519 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3520 // to keep the information to which null check the new DecodeN node 3521 // corresponds to use it as value in implicit_null_check(). 3522 // 3523 new_in1->set_req(0, n->in(0)); 3524 } 3525 3526 n->subsume_by(new_in1, this); 3527 if (in1->outcnt() == 0) { 3528 in1->disconnect_inputs(this); 3529 } 3530 } else { 3531 n->subsume_by(n->in(1), this); 3532 if (n->outcnt() == 0) { 3533 n->disconnect_inputs(this); 3534 } 3535 } 3536 break; 3537 } 3538 case Op_CastII: { 3539 n->as_CastII()->remove_range_check_cast(this); 3540 break; 3541 } 3542 #ifdef _LP64 3543 case Op_CmpP: 3544 // Do this transformation here to preserve CmpPNode::sub() and 3545 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3546 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3547 Node* in1 = n->in(1); 3548 Node* in2 = n->in(2); 3549 if (!in1->is_DecodeNarrowPtr()) { 3550 in2 = in1; 3551 in1 = n->in(2); 3552 } 3553 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3554 3555 Node* new_in2 = nullptr; 3556 if (in2->is_DecodeNarrowPtr()) { 3557 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3558 new_in2 = in2->in(1); 3559 } else if (in2->Opcode() == Op_ConP) { 3560 const Type* t = in2->bottom_type(); 3561 if (t == TypePtr::NULL_PTR) { 3562 assert(in1->is_DecodeN(), "compare klass to null?"); 3563 // Don't convert CmpP null check into CmpN if compressed 3564 // oops implicit null check is not generated. 3565 // This will allow to generate normal oop implicit null check. 3566 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3567 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3568 // 3569 // This transformation together with CastPP transformation above 3570 // will generated code for implicit null checks for compressed oops. 3571 // 3572 // The original code after Optimize() 3573 // 3574 // LoadN memory, narrow_oop_reg 3575 // decode narrow_oop_reg, base_reg 3576 // CmpP base_reg, nullptr 3577 // CastPP base_reg // NotNull 3578 // Load [base_reg + offset], val_reg 3579 // 3580 // after these transformations will be 3581 // 3582 // LoadN memory, narrow_oop_reg 3583 // CmpN narrow_oop_reg, nullptr 3584 // decode_not_null narrow_oop_reg, base_reg 3585 // Load [base_reg + offset], val_reg 3586 // 3587 // and the uncommon path (== nullptr) will use narrow_oop_reg directly 3588 // since narrow oops can be used in debug info now (see the code in 3589 // final_graph_reshaping_walk()). 3590 // 3591 // At the end the code will be matched to 3592 // on x86: 3593 // 3594 // Load_narrow_oop memory, narrow_oop_reg 3595 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3596 // NullCheck narrow_oop_reg 3597 // 3598 // and on sparc: 3599 // 3600 // Load_narrow_oop memory, narrow_oop_reg 3601 // decode_not_null narrow_oop_reg, base_reg 3602 // Load [base_reg + offset], val_reg 3603 // NullCheck base_reg 3604 // 3605 } else if (t->isa_oopptr()) { 3606 new_in2 = ConNode::make(t->make_narrowoop()); 3607 } else if (t->isa_klassptr()) { 3608 new_in2 = ConNode::make(t->make_narrowklass()); 3609 } 3610 } 3611 if (new_in2 != nullptr) { 3612 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3613 n->subsume_by(cmpN, this); 3614 if (in1->outcnt() == 0) { 3615 in1->disconnect_inputs(this); 3616 } 3617 if (in2->outcnt() == 0) { 3618 in2->disconnect_inputs(this); 3619 } 3620 } 3621 } 3622 break; 3623 3624 case Op_DecodeN: 3625 case Op_DecodeNKlass: 3626 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3627 // DecodeN could be pinned when it can't be fold into 3628 // an address expression, see the code for Op_CastPP above. 3629 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3630 break; 3631 3632 case Op_EncodeP: 3633 case Op_EncodePKlass: { 3634 Node* in1 = n->in(1); 3635 if (in1->is_DecodeNarrowPtr()) { 3636 n->subsume_by(in1->in(1), this); 3637 } else if (in1->Opcode() == Op_ConP) { 3638 const Type* t = in1->bottom_type(); 3639 if (t == TypePtr::NULL_PTR) { 3640 assert(t->isa_oopptr(), "null klass?"); 3641 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3642 } else if (t->isa_oopptr()) { 3643 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3644 } else if (t->isa_klassptr()) { 3645 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3646 } 3647 } 3648 if (in1->outcnt() == 0) { 3649 in1->disconnect_inputs(this); 3650 } 3651 break; 3652 } 3653 3654 case Op_Proj: { 3655 if (OptimizeStringConcat || IncrementalInline) { 3656 ProjNode* proj = n->as_Proj(); 3657 if (proj->_is_io_use) { 3658 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, ""); 3659 // Separate projections were used for the exception path which 3660 // are normally removed by a late inline. If it wasn't inlined 3661 // then they will hang around and should just be replaced with 3662 // the original one. Merge them. 3663 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/); 3664 if (non_io_proj != nullptr) { 3665 proj->subsume_by(non_io_proj , this); 3666 } 3667 } 3668 } 3669 break; 3670 } 3671 3672 case Op_Phi: 3673 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3674 // The EncodeP optimization may create Phi with the same edges 3675 // for all paths. It is not handled well by Register Allocator. 3676 Node* unique_in = n->in(1); 3677 assert(unique_in != nullptr, ""); 3678 uint cnt = n->req(); 3679 for (uint i = 2; i < cnt; i++) { 3680 Node* m = n->in(i); 3681 assert(m != nullptr, ""); 3682 if (unique_in != m) 3683 unique_in = nullptr; 3684 } 3685 if (unique_in != nullptr) { 3686 n->subsume_by(unique_in, this); 3687 } 3688 } 3689 break; 3690 3691 #endif 3692 3693 case Op_ModI: 3694 handle_div_mod_op(n, T_INT, false); 3695 break; 3696 3697 case Op_ModL: 3698 handle_div_mod_op(n, T_LONG, false); 3699 break; 3700 3701 case Op_UModI: 3702 handle_div_mod_op(n, T_INT, true); 3703 break; 3704 3705 case Op_UModL: 3706 handle_div_mod_op(n, T_LONG, true); 3707 break; 3708 3709 case Op_LoadVector: 3710 case Op_StoreVector: 3711 #ifdef ASSERT 3712 // Add VerifyVectorAlignment node between adr and load / store. 3713 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) { 3714 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() : 3715 n->as_StoreVector()->must_verify_alignment(); 3716 if (must_verify_alignment) { 3717 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() : 3718 n->as_StoreVector()->memory_size(); 3719 // The memory access should be aligned to the vector width in bytes. 3720 // However, the underlying array is possibly less well aligned, but at least 3721 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in 3722 // a loop we can expect at least the following alignment: 3723 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes); 3724 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range"); 3725 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2"); 3726 // Create mask from alignment. e.g. 0b1000 -> 0b0111 3727 jlong mask = guaranteed_alignment - 1; 3728 Node* mask_con = ConLNode::make(mask); 3729 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con); 3730 n->set_req(MemNode::Address, va); 3731 } 3732 } 3733 #endif 3734 break; 3735 3736 case Op_LoadVectorGather: 3737 case Op_StoreVectorScatter: 3738 case Op_LoadVectorGatherMasked: 3739 case Op_StoreVectorScatterMasked: 3740 case Op_VectorCmpMasked: 3741 case Op_VectorMaskGen: 3742 case Op_LoadVectorMasked: 3743 case Op_StoreVectorMasked: 3744 break; 3745 3746 case Op_AddReductionVI: 3747 case Op_AddReductionVL: 3748 case Op_AddReductionVF: 3749 case Op_AddReductionVD: 3750 case Op_MulReductionVI: 3751 case Op_MulReductionVL: 3752 case Op_MulReductionVF: 3753 case Op_MulReductionVD: 3754 case Op_MinReductionV: 3755 case Op_MaxReductionV: 3756 case Op_AndReductionV: 3757 case Op_OrReductionV: 3758 case Op_XorReductionV: 3759 break; 3760 3761 case Op_PackB: 3762 case Op_PackS: 3763 case Op_PackI: 3764 case Op_PackF: 3765 case Op_PackL: 3766 case Op_PackD: 3767 if (n->req()-1 > 2) { 3768 // Replace many operand PackNodes with a binary tree for matching 3769 PackNode* p = (PackNode*) n; 3770 Node* btp = p->binary_tree_pack(1, n->req()); 3771 n->subsume_by(btp, this); 3772 } 3773 break; 3774 case Op_Loop: 3775 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop"); 3776 case Op_CountedLoop: 3777 case Op_LongCountedLoop: 3778 case Op_OuterStripMinedLoop: 3779 if (n->as_Loop()->is_inner_loop()) { 3780 frc.inc_inner_loop_count(); 3781 } 3782 n->as_Loop()->verify_strip_mined(0); 3783 break; 3784 case Op_LShiftI: 3785 case Op_RShiftI: 3786 case Op_URShiftI: 3787 case Op_LShiftL: 3788 case Op_RShiftL: 3789 case Op_URShiftL: 3790 if (Matcher::need_masked_shift_count) { 3791 // The cpu's shift instructions don't restrict the count to the 3792 // lower 5/6 bits. We need to do the masking ourselves. 3793 Node* in2 = n->in(2); 3794 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3795 const TypeInt* t = in2->find_int_type(); 3796 if (t != nullptr && t->is_con()) { 3797 juint shift = t->get_con(); 3798 if (shift > mask) { // Unsigned cmp 3799 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3800 } 3801 } else { 3802 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) { 3803 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3804 n->set_req(2, shift); 3805 } 3806 } 3807 if (in2->outcnt() == 0) { // Remove dead node 3808 in2->disconnect_inputs(this); 3809 } 3810 } 3811 break; 3812 case Op_MemBarStoreStore: 3813 case Op_MemBarRelease: 3814 // Break the link with AllocateNode: it is no longer useful and 3815 // confuses register allocation. 3816 if (n->req() > MemBarNode::Precedent) { 3817 n->set_req(MemBarNode::Precedent, top()); 3818 } 3819 break; 3820 case Op_MemBarAcquire: { 3821 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3822 // At parse time, the trailing MemBarAcquire for a volatile load 3823 // is created with an edge to the load. After optimizations, 3824 // that input may be a chain of Phis. If those phis have no 3825 // other use, then the MemBarAcquire keeps them alive and 3826 // register allocation can be confused. 3827 dead_nodes.push(n->in(MemBarNode::Precedent)); 3828 n->set_req(MemBarNode::Precedent, top()); 3829 } 3830 break; 3831 } 3832 case Op_Blackhole: 3833 break; 3834 case Op_RangeCheck: { 3835 RangeCheckNode* rc = n->as_RangeCheck(); 3836 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3837 n->subsume_by(iff, this); 3838 frc._tests.push(iff); 3839 break; 3840 } 3841 case Op_ConvI2L: { 3842 if (!Matcher::convi2l_type_required) { 3843 // Code generation on some platforms doesn't need accurate 3844 // ConvI2L types. Widening the type can help remove redundant 3845 // address computations. 3846 n->as_Type()->set_type(TypeLong::INT); 3847 ResourceMark rm; 3848 Unique_Node_List wq; 3849 wq.push(n); 3850 for (uint next = 0; next < wq.size(); next++) { 3851 Node *m = wq.at(next); 3852 3853 for(;;) { 3854 // Loop over all nodes with identical inputs edges as m 3855 Node* k = m->find_similar(m->Opcode()); 3856 if (k == nullptr) { 3857 break; 3858 } 3859 // Push their uses so we get a chance to remove node made 3860 // redundant 3861 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3862 Node* u = k->fast_out(i); 3863 if (u->Opcode() == Op_LShiftL || 3864 u->Opcode() == Op_AddL || 3865 u->Opcode() == Op_SubL || 3866 u->Opcode() == Op_AddP) { 3867 wq.push(u); 3868 } 3869 } 3870 // Replace all nodes with identical edges as m with m 3871 k->subsume_by(m, this); 3872 } 3873 } 3874 } 3875 break; 3876 } 3877 case Op_CmpUL: { 3878 if (!Matcher::has_match_rule(Op_CmpUL)) { 3879 // No support for unsigned long comparisons 3880 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3881 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3882 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3883 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3884 Node* andl = new AndLNode(orl, remove_sign_mask); 3885 Node* cmp = new CmpLNode(andl, n->in(2)); 3886 n->subsume_by(cmp, this); 3887 } 3888 break; 3889 } 3890 #ifdef ASSERT 3891 case Op_ConNKlass: { 3892 const TypePtr* tp = n->as_Type()->type()->make_ptr(); 3893 ciKlass* klass = tp->is_klassptr()->exact_klass(); 3894 assert(klass->is_in_encoding_range(), "klass cannot be compressed"); 3895 break; 3896 } 3897 #endif 3898 default: 3899 assert(!n->is_Call(), ""); 3900 assert(!n->is_Mem(), ""); 3901 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3902 break; 3903 } 3904 } 3905 3906 //------------------------------final_graph_reshaping_walk--------------------- 3907 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3908 // requires that the walk visits a node's inputs before visiting the node. 3909 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3910 Unique_Node_List sfpt; 3911 3912 frc._visited.set(root->_idx); // first, mark node as visited 3913 uint cnt = root->req(); 3914 Node *n = root; 3915 uint i = 0; 3916 while (true) { 3917 if (i < cnt) { 3918 // Place all non-visited non-null inputs onto stack 3919 Node* m = n->in(i); 3920 ++i; 3921 if (m != nullptr && !frc._visited.test_set(m->_idx)) { 3922 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) { 3923 // compute worst case interpreter size in case of a deoptimization 3924 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3925 3926 sfpt.push(m); 3927 } 3928 cnt = m->req(); 3929 nstack.push(n, i); // put on stack parent and next input's index 3930 n = m; 3931 i = 0; 3932 } 3933 } else { 3934 // Now do post-visit work 3935 final_graph_reshaping_impl(n, frc, dead_nodes); 3936 if (nstack.is_empty()) 3937 break; // finished 3938 n = nstack.node(); // Get node from stack 3939 cnt = n->req(); 3940 i = nstack.index(); 3941 nstack.pop(); // Shift to the next node on stack 3942 } 3943 } 3944 3945 // Skip next transformation if compressed oops are not used. 3946 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3947 (!UseCompressedOops && !UseCompressedClassPointers)) 3948 return; 3949 3950 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3951 // It could be done for an uncommon traps or any safepoints/calls 3952 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3953 while (sfpt.size() > 0) { 3954 n = sfpt.pop(); 3955 JVMState *jvms = n->as_SafePoint()->jvms(); 3956 assert(jvms != nullptr, "sanity"); 3957 int start = jvms->debug_start(); 3958 int end = n->req(); 3959 bool is_uncommon = (n->is_CallStaticJava() && 3960 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3961 for (int j = start; j < end; j++) { 3962 Node* in = n->in(j); 3963 if (in->is_DecodeNarrowPtr()) { 3964 bool safe_to_skip = true; 3965 if (!is_uncommon ) { 3966 // Is it safe to skip? 3967 for (uint i = 0; i < in->outcnt(); i++) { 3968 Node* u = in->raw_out(i); 3969 if (!u->is_SafePoint() || 3970 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3971 safe_to_skip = false; 3972 } 3973 } 3974 } 3975 if (safe_to_skip) { 3976 n->set_req(j, in->in(1)); 3977 } 3978 if (in->outcnt() == 0) { 3979 in->disconnect_inputs(this); 3980 } 3981 } 3982 } 3983 } 3984 } 3985 3986 //------------------------------final_graph_reshaping-------------------------- 3987 // Final Graph Reshaping. 3988 // 3989 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3990 // and not commoned up and forced early. Must come after regular 3991 // optimizations to avoid GVN undoing the cloning. Clone constant 3992 // inputs to Loop Phis; these will be split by the allocator anyways. 3993 // Remove Opaque nodes. 3994 // (2) Move last-uses by commutative operations to the left input to encourage 3995 // Intel update-in-place two-address operations and better register usage 3996 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3997 // calls canonicalizing them back. 3998 // (3) Count the number of double-precision FP ops, single-precision FP ops 3999 // and call sites. On Intel, we can get correct rounding either by 4000 // forcing singles to memory (requires extra stores and loads after each 4001 // FP bytecode) or we can set a rounding mode bit (requires setting and 4002 // clearing the mode bit around call sites). The mode bit is only used 4003 // if the relative frequency of single FP ops to calls is low enough. 4004 // This is a key transform for SPEC mpeg_audio. 4005 // (4) Detect infinite loops; blobs of code reachable from above but not 4006 // below. Several of the Code_Gen algorithms fail on such code shapes, 4007 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 4008 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 4009 // Detection is by looking for IfNodes where only 1 projection is 4010 // reachable from below or CatchNodes missing some targets. 4011 // (5) Assert for insane oop offsets in debug mode. 4012 4013 bool Compile::final_graph_reshaping() { 4014 // an infinite loop may have been eliminated by the optimizer, 4015 // in which case the graph will be empty. 4016 if (root()->req() == 1) { 4017 // Do not compile method that is only a trivial infinite loop, 4018 // since the content of the loop may have been eliminated. 4019 record_method_not_compilable("trivial infinite loop"); 4020 return true; 4021 } 4022 4023 // Expensive nodes have their control input set to prevent the GVN 4024 // from freely commoning them. There's no GVN beyond this point so 4025 // no need to keep the control input. We want the expensive nodes to 4026 // be freely moved to the least frequent code path by gcm. 4027 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 4028 for (int i = 0; i < expensive_count(); i++) { 4029 _expensive_nodes.at(i)->set_req(0, nullptr); 4030 } 4031 4032 Final_Reshape_Counts frc; 4033 4034 // Visit everybody reachable! 4035 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 4036 Node_Stack nstack(live_nodes() >> 1); 4037 Unique_Node_List dead_nodes; 4038 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes); 4039 4040 // Check for unreachable (from below) code (i.e., infinite loops). 4041 for( uint i = 0; i < frc._tests.size(); i++ ) { 4042 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 4043 // Get number of CFG targets. 4044 // Note that PCTables include exception targets after calls. 4045 uint required_outcnt = n->required_outcnt(); 4046 if (n->outcnt() != required_outcnt) { 4047 // Check for a few special cases. Rethrow Nodes never take the 4048 // 'fall-thru' path, so expected kids is 1 less. 4049 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 4050 if (n->in(0)->in(0)->is_Call()) { 4051 CallNode* call = n->in(0)->in(0)->as_Call(); 4052 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 4053 required_outcnt--; // Rethrow always has 1 less kid 4054 } else if (call->req() > TypeFunc::Parms && 4055 call->is_CallDynamicJava()) { 4056 // Check for null receiver. In such case, the optimizer has 4057 // detected that the virtual call will always result in a null 4058 // pointer exception. The fall-through projection of this CatchNode 4059 // will not be populated. 4060 Node* arg0 = call->in(TypeFunc::Parms); 4061 if (arg0->is_Type() && 4062 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 4063 required_outcnt--; 4064 } 4065 } else if (call->entry_point() == OptoRuntime::new_array_Java() || 4066 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4067 // Check for illegal array length. In such case, the optimizer has 4068 // detected that the allocation attempt will always result in an 4069 // exception. There is no fall-through projection of this CatchNode . 4070 assert(call->is_CallStaticJava(), "static call expected"); 4071 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4072 uint valid_length_test_input = call->req() - 1; 4073 Node* valid_length_test = call->in(valid_length_test_input); 4074 call->del_req(valid_length_test_input); 4075 if (valid_length_test->find_int_con(1) == 0) { 4076 required_outcnt--; 4077 } 4078 dead_nodes.push(valid_length_test); 4079 assert(n->outcnt() == required_outcnt, "malformed control flow"); 4080 continue; 4081 } 4082 } 4083 } 4084 4085 // Recheck with a better notion of 'required_outcnt' 4086 if (n->outcnt() != required_outcnt) { 4087 record_method_not_compilable("malformed control flow"); 4088 return true; // Not all targets reachable! 4089 } 4090 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) { 4091 CallNode* call = n->in(0)->in(0)->as_Call(); 4092 if (call->entry_point() == OptoRuntime::new_array_Java() || 4093 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4094 assert(call->is_CallStaticJava(), "static call expected"); 4095 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4096 uint valid_length_test_input = call->req() - 1; 4097 dead_nodes.push(call->in(valid_length_test_input)); 4098 call->del_req(valid_length_test_input); // valid length test useless now 4099 } 4100 } 4101 // Check that I actually visited all kids. Unreached kids 4102 // must be infinite loops. 4103 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 4104 if (!frc._visited.test(n->fast_out(j)->_idx)) { 4105 record_method_not_compilable("infinite loop"); 4106 return true; // Found unvisited kid; must be unreach 4107 } 4108 4109 // Here so verification code in final_graph_reshaping_walk() 4110 // always see an OuterStripMinedLoopEnd 4111 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) { 4112 IfNode* init_iff = n->as_If(); 4113 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 4114 n->subsume_by(iff, this); 4115 } 4116 } 4117 4118 while (dead_nodes.size() > 0) { 4119 Node* m = dead_nodes.pop(); 4120 if (m->outcnt() == 0 && m != top()) { 4121 for (uint j = 0; j < m->req(); j++) { 4122 Node* in = m->in(j); 4123 if (in != nullptr) { 4124 dead_nodes.push(in); 4125 } 4126 } 4127 m->disconnect_inputs(this); 4128 } 4129 } 4130 4131 set_java_calls(frc.get_java_call_count()); 4132 set_inner_loops(frc.get_inner_loop_count()); 4133 4134 // No infinite loops, no reason to bail out. 4135 return false; 4136 } 4137 4138 //-----------------------------too_many_traps---------------------------------- 4139 // Report if there are too many traps at the current method and bci. 4140 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 4141 bool Compile::too_many_traps(ciMethod* method, 4142 int bci, 4143 Deoptimization::DeoptReason reason) { 4144 if (method->has_trap_at(bci)) { 4145 return true; 4146 } 4147 if (PreloadReduceTraps && for_preload()) { 4148 // Preload code should not have traps, if possible. 4149 return true; 4150 } 4151 ciMethodData* md = method->method_data(); 4152 if (md->is_empty()) { 4153 // Assume the trap has not occurred, or that it occurred only 4154 // because of a transient condition during start-up in the interpreter. 4155 return false; 4156 } 4157 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4158 if (md->has_trap_at(bci, m, reason) != 0) { 4159 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 4160 // Also, if there are multiple reasons, or if there is no per-BCI record, 4161 // assume the worst. 4162 if (log()) 4163 log()->elem("observe trap='%s' count='%d'", 4164 Deoptimization::trap_reason_name(reason), 4165 md->trap_count(reason)); 4166 return true; 4167 } else { 4168 // Ignore method/bci and see if there have been too many globally. 4169 return too_many_traps(reason, md); 4170 } 4171 } 4172 4173 // Less-accurate variant which does not require a method and bci. 4174 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 4175 ciMethodData* logmd) { 4176 if (PreloadReduceTraps && for_preload()) { 4177 // Preload code should not have traps, if possible. 4178 return true; 4179 } 4180 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 4181 // Too many traps globally. 4182 // Note that we use cumulative trap_count, not just md->trap_count. 4183 if (log()) { 4184 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason); 4185 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 4186 Deoptimization::trap_reason_name(reason), 4187 mcount, trap_count(reason)); 4188 } 4189 return true; 4190 } else { 4191 // The coast is clear. 4192 return false; 4193 } 4194 } 4195 4196 //--------------------------too_many_recompiles-------------------------------- 4197 // Report if there are too many recompiles at the current method and bci. 4198 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 4199 // Is not eager to return true, since this will cause the compiler to use 4200 // Action_none for a trap point, to avoid too many recompilations. 4201 bool Compile::too_many_recompiles(ciMethod* method, 4202 int bci, 4203 Deoptimization::DeoptReason reason) { 4204 ciMethodData* md = method->method_data(); 4205 if (md->is_empty()) { 4206 // Assume the trap has not occurred, or that it occurred only 4207 // because of a transient condition during start-up in the interpreter. 4208 return false; 4209 } 4210 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 4211 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 4212 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 4213 Deoptimization::DeoptReason per_bc_reason 4214 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 4215 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4216 if ((per_bc_reason == Deoptimization::Reason_none 4217 || md->has_trap_at(bci, m, reason) != 0) 4218 // The trap frequency measure we care about is the recompile count: 4219 && md->trap_recompiled_at(bci, m) 4220 && md->overflow_recompile_count() >= bc_cutoff) { 4221 // Do not emit a trap here if it has already caused recompilations. 4222 // Also, if there are multiple reasons, or if there is no per-BCI record, 4223 // assume the worst. 4224 if (log()) 4225 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 4226 Deoptimization::trap_reason_name(reason), 4227 md->trap_count(reason), 4228 md->overflow_recompile_count()); 4229 return true; 4230 } else if (trap_count(reason) != 0 4231 && decompile_count() >= m_cutoff) { 4232 // Too many recompiles globally, and we have seen this sort of trap. 4233 // Use cumulative decompile_count, not just md->decompile_count. 4234 if (log()) 4235 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 4236 Deoptimization::trap_reason_name(reason), 4237 md->trap_count(reason), trap_count(reason), 4238 md->decompile_count(), decompile_count()); 4239 return true; 4240 } else { 4241 // The coast is clear. 4242 return false; 4243 } 4244 } 4245 4246 // Compute when not to trap. Used by matching trap based nodes and 4247 // NullCheck optimization. 4248 void Compile::set_allowed_deopt_reasons() { 4249 _allowed_reasons = 0; 4250 if (is_method_compilation()) { 4251 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 4252 assert(rs < BitsPerInt, "recode bit map"); 4253 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 4254 _allowed_reasons |= nth_bit(rs); 4255 } 4256 } 4257 } 4258 } 4259 4260 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 4261 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 4262 } 4263 4264 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 4265 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 4266 } 4267 4268 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 4269 if (holder->is_initialized() && !do_clinit_barriers()) { 4270 return false; 4271 } 4272 if (holder->is_being_initialized() || do_clinit_barriers()) { 4273 if (accessing_method->holder() == holder) { 4274 // Access inside a class. The barrier can be elided when access happens in <clinit>, 4275 // <init>, or a static method. In all those cases, there was an initialization 4276 // barrier on the holder klass passed. 4277 if (accessing_method->is_static_initializer() || 4278 accessing_method->is_object_initializer() || 4279 accessing_method->is_static()) { 4280 return false; 4281 } 4282 } else if (accessing_method->holder()->is_subclass_of(holder)) { 4283 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 4284 // In case of <init> or a static method, the barrier is on the subclass is not enough: 4285 // child class can become fully initialized while its parent class is still being initialized. 4286 if (accessing_method->is_static_initializer()) { 4287 return false; 4288 } 4289 } 4290 ciMethod* root = method(); // the root method of compilation 4291 if (root != accessing_method) { 4292 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 4293 } 4294 } 4295 return true; 4296 } 4297 4298 #ifndef PRODUCT 4299 //------------------------------verify_bidirectional_edges--------------------- 4300 // For each input edge to a node (ie - for each Use-Def edge), verify that 4301 // there is a corresponding Def-Use edge. 4302 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const { 4303 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc 4304 uint stack_size = live_nodes() >> 4; 4305 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize)); 4306 if (root_and_safepoints != nullptr) { 4307 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints"); 4308 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) { 4309 Node* root_or_safepoint = root_and_safepoints->at(i); 4310 // If the node is a safepoint, let's check if it still has a control input 4311 // Lack of control input signifies that this node was killed by CCP or 4312 // recursively by remove_globally_dead_node and it shouldn't be a starting 4313 // point. 4314 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) { 4315 nstack.push(root_or_safepoint); 4316 } 4317 } 4318 } else { 4319 nstack.push(_root); 4320 } 4321 4322 while (nstack.size() > 0) { 4323 Node* n = nstack.pop(); 4324 if (visited.member(n)) { 4325 continue; 4326 } 4327 visited.push(n); 4328 4329 // Walk over all input edges, checking for correspondence 4330 uint length = n->len(); 4331 for (uint i = 0; i < length; i++) { 4332 Node* in = n->in(i); 4333 if (in != nullptr && !visited.member(in)) { 4334 nstack.push(in); // Put it on stack 4335 } 4336 if (in != nullptr && !in->is_top()) { 4337 // Count instances of `next` 4338 int cnt = 0; 4339 for (uint idx = 0; idx < in->_outcnt; idx++) { 4340 if (in->_out[idx] == n) { 4341 cnt++; 4342 } 4343 } 4344 assert(cnt > 0, "Failed to find Def-Use edge."); 4345 // Check for duplicate edges 4346 // walk the input array downcounting the input edges to n 4347 for (uint j = 0; j < length; j++) { 4348 if (n->in(j) == in) { 4349 cnt--; 4350 } 4351 } 4352 assert(cnt == 0, "Mismatched edge count."); 4353 } else if (in == nullptr) { 4354 assert(i == 0 || i >= n->req() || 4355 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() || 4356 (n->is_Unlock() && i == (n->req() - 1)) || 4357 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion 4358 "only region, phi, arraycopy, unlock or membar nodes have null data edges"); 4359 } else { 4360 assert(in->is_top(), "sanity"); 4361 // Nothing to check. 4362 } 4363 } 4364 } 4365 } 4366 4367 //------------------------------verify_graph_edges--------------------------- 4368 // Walk the Graph and verify that there is a one-to-one correspondence 4369 // between Use-Def edges and Def-Use edges in the graph. 4370 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const { 4371 if (VerifyGraphEdges) { 4372 Unique_Node_List visited; 4373 4374 // Call graph walk to check edges 4375 verify_bidirectional_edges(visited, root_and_safepoints); 4376 if (no_dead_code) { 4377 // Now make sure that no visited node is used by an unvisited node. 4378 bool dead_nodes = false; 4379 Unique_Node_List checked; 4380 while (visited.size() > 0) { 4381 Node* n = visited.pop(); 4382 checked.push(n); 4383 for (uint i = 0; i < n->outcnt(); i++) { 4384 Node* use = n->raw_out(i); 4385 if (checked.member(use)) continue; // already checked 4386 if (visited.member(use)) continue; // already in the graph 4387 if (use->is_Con()) continue; // a dead ConNode is OK 4388 // At this point, we have found a dead node which is DU-reachable. 4389 if (!dead_nodes) { 4390 tty->print_cr("*** Dead nodes reachable via DU edges:"); 4391 dead_nodes = true; 4392 } 4393 use->dump(2); 4394 tty->print_cr("---"); 4395 checked.push(use); // No repeats; pretend it is now checked. 4396 } 4397 } 4398 assert(!dead_nodes, "using nodes must be reachable from root"); 4399 } 4400 } 4401 } 4402 #endif 4403 4404 // The Compile object keeps track of failure reasons separately from the ciEnv. 4405 // This is required because there is not quite a 1-1 relation between the 4406 // ciEnv and its compilation task and the Compile object. Note that one 4407 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4408 // to backtrack and retry without subsuming loads. Other than this backtracking 4409 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4410 // by the logic in C2Compiler. 4411 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) { 4412 if (log() != nullptr) { 4413 log()->elem("failure reason='%s' phase='compile'", reason); 4414 } 4415 if (_failure_reason.get() == nullptr) { 4416 // Record the first failure reason. 4417 _failure_reason.set(reason); 4418 if (CaptureBailoutInformation) { 4419 _first_failure_details = new CompilationFailureInfo(reason); 4420 } 4421 } else { 4422 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure."); 4423 } 4424 4425 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4426 C->print_method(PHASE_FAILURE, 1); 4427 } 4428 _root = nullptr; // flush the graph, too 4429 } 4430 4431 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id) 4432 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose), 4433 _compile(Compile::current()), 4434 _log(nullptr), 4435 _dolog(CITimeVerbose) 4436 { 4437 assert(_compile != nullptr, "sanity check"); 4438 assert(id != PhaseTraceId::_t_none, "Don't use none"); 4439 if (_dolog) { 4440 _log = _compile->log(); 4441 } 4442 if (_log != nullptr) { 4443 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes()); 4444 _log->stamp(); 4445 _log->end_head(); 4446 } 4447 4448 // Inform memory statistic, if enabled 4449 if (CompilationMemoryStatistic::enabled()) { 4450 CompilationMemoryStatistic::on_phase_start((int)id, name); 4451 } 4452 } 4453 4454 Compile::TracePhase::TracePhase(PhaseTraceId id) 4455 : TracePhase(Phase::get_phase_trace_id_text(id), id) {} 4456 4457 Compile::TracePhase::~TracePhase() { 4458 4459 // Inform memory statistic, if enabled 4460 if (CompilationMemoryStatistic::enabled()) { 4461 CompilationMemoryStatistic::on_phase_end(); 4462 } 4463 4464 if (_compile->failing_internal()) { 4465 if (_log != nullptr) { 4466 _log->done("phase"); 4467 } 4468 return; // timing code, not stressing bailouts. 4469 } 4470 #ifdef ASSERT 4471 if (PrintIdealNodeCount) { 4472 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4473 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk()); 4474 } 4475 4476 if (VerifyIdealNodeCount) { 4477 _compile->print_missing_nodes(); 4478 } 4479 #endif 4480 4481 if (_log != nullptr) { 4482 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes()); 4483 } 4484 } 4485 4486 //----------------------------static_subtype_check----------------------------- 4487 // Shortcut important common cases when superklass is exact: 4488 // (0) superklass is java.lang.Object (can occur in reflective code) 4489 // (1) subklass is already limited to a subtype of superklass => always ok 4490 // (2) subklass does not overlap with superklass => always fail 4491 // (3) superklass has NO subtypes and we can check with a simple compare. 4492 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) { 4493 if (skip) { 4494 return SSC_full_test; // Let caller generate the general case. 4495 } 4496 4497 if (subk->is_java_subtype_of(superk)) { 4498 return SSC_always_true; // (0) and (1) this test cannot fail 4499 } 4500 4501 if (!subk->maybe_java_subtype_of(superk)) { 4502 return SSC_always_false; // (2) true path dead; no dynamic test needed 4503 } 4504 4505 const Type* superelem = superk; 4506 if (superk->isa_aryklassptr()) { 4507 int ignored; 4508 superelem = superk->is_aryklassptr()->base_element_type(ignored); 4509 } 4510 4511 if (superelem->isa_instklassptr()) { 4512 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass(); 4513 if (!ik->has_subklass()) { 4514 if (!ik->is_final()) { 4515 // Add a dependency if there is a chance of a later subclass. 4516 dependencies()->assert_leaf_type(ik); 4517 } 4518 if (!superk->maybe_java_subtype_of(subk)) { 4519 return SSC_always_false; 4520 } 4521 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4522 } 4523 } else { 4524 // A primitive array type has no subtypes. 4525 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4526 } 4527 4528 return SSC_full_test; 4529 } 4530 4531 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4532 #ifdef _LP64 4533 // The scaled index operand to AddP must be a clean 64-bit value. 4534 // Java allows a 32-bit int to be incremented to a negative 4535 // value, which appears in a 64-bit register as a large 4536 // positive number. Using that large positive number as an 4537 // operand in pointer arithmetic has bad consequences. 4538 // On the other hand, 32-bit overflow is rare, and the possibility 4539 // can often be excluded, if we annotate the ConvI2L node with 4540 // a type assertion that its value is known to be a small positive 4541 // number. (The prior range check has ensured this.) 4542 // This assertion is used by ConvI2LNode::Ideal. 4543 int index_max = max_jint - 1; // array size is max_jint, index is one less 4544 if (sizetype != nullptr) index_max = sizetype->_hi - 1; 4545 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4546 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4547 #endif 4548 return idx; 4549 } 4550 4551 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4552 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 4553 if (ctrl != nullptr) { 4554 // Express control dependency by a CastII node with a narrow type. 4555 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4556 // node from floating above the range check during loop optimizations. Otherwise, the 4557 // ConvI2L node may be eliminated independently of the range check, causing the data path 4558 // to become TOP while the control path is still there (although it's unreachable). 4559 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */); 4560 value = phase->transform(value); 4561 } 4562 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4563 return phase->transform(new ConvI2LNode(value, ltype)); 4564 } 4565 4566 void Compile::dump_print_inlining() { 4567 inline_printer()->print_on(tty); 4568 } 4569 4570 void Compile::log_late_inline(CallGenerator* cg) { 4571 if (log() != nullptr) { 4572 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4573 cg->unique_id()); 4574 JVMState* p = cg->call_node()->jvms(); 4575 while (p != nullptr) { 4576 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4577 p = p->caller(); 4578 } 4579 log()->tail("late_inline"); 4580 } 4581 } 4582 4583 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4584 log_late_inline(cg); 4585 if (log() != nullptr) { 4586 log()->inline_fail(msg); 4587 } 4588 } 4589 4590 void Compile::log_inline_id(CallGenerator* cg) { 4591 if (log() != nullptr) { 4592 // The LogCompilation tool needs a unique way to identify late 4593 // inline call sites. This id must be unique for this call site in 4594 // this compilation. Try to have it unique across compilations as 4595 // well because it can be convenient when grepping through the log 4596 // file. 4597 // Distinguish OSR compilations from others in case CICountOSR is 4598 // on. 4599 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4600 cg->set_unique_id(id); 4601 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4602 } 4603 } 4604 4605 void Compile::log_inline_failure(const char* msg) { 4606 if (C->log() != nullptr) { 4607 C->log()->inline_fail(msg); 4608 } 4609 } 4610 4611 4612 // Dump inlining replay data to the stream. 4613 // Don't change thread state and acquire any locks. 4614 void Compile::dump_inline_data(outputStream* out) { 4615 InlineTree* inl_tree = ilt(); 4616 if (inl_tree != nullptr) { 4617 out->print(" inline %d", inl_tree->count()); 4618 inl_tree->dump_replay_data(out); 4619 } 4620 } 4621 4622 void Compile::dump_inline_data_reduced(outputStream* out) { 4623 assert(ReplayReduce, ""); 4624 4625 InlineTree* inl_tree = ilt(); 4626 if (inl_tree == nullptr) { 4627 return; 4628 } 4629 // Enable iterative replay file reduction 4630 // Output "compile" lines for depth 1 subtrees, 4631 // simulating that those trees were compiled 4632 // instead of inlined. 4633 for (int i = 0; i < inl_tree->subtrees().length(); ++i) { 4634 InlineTree* sub = inl_tree->subtrees().at(i); 4635 if (sub->inline_level() != 1) { 4636 continue; 4637 } 4638 4639 ciMethod* method = sub->method(); 4640 int entry_bci = -1; 4641 int comp_level = env()->task()->comp_level(); 4642 out->print("compile "); 4643 method->dump_name_as_ascii(out); 4644 out->print(" %d %d", entry_bci, comp_level); 4645 out->print(" inline %d", sub->count()); 4646 sub->dump_replay_data(out, -1); 4647 out->cr(); 4648 } 4649 } 4650 4651 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4652 if (n1->Opcode() < n2->Opcode()) return -1; 4653 else if (n1->Opcode() > n2->Opcode()) return 1; 4654 4655 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4656 for (uint i = 1; i < n1->req(); i++) { 4657 if (n1->in(i) < n2->in(i)) return -1; 4658 else if (n1->in(i) > n2->in(i)) return 1; 4659 } 4660 4661 return 0; 4662 } 4663 4664 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4665 Node* n1 = *n1p; 4666 Node* n2 = *n2p; 4667 4668 return cmp_expensive_nodes(n1, n2); 4669 } 4670 4671 void Compile::sort_expensive_nodes() { 4672 if (!expensive_nodes_sorted()) { 4673 _expensive_nodes.sort(cmp_expensive_nodes); 4674 } 4675 } 4676 4677 bool Compile::expensive_nodes_sorted() const { 4678 for (int i = 1; i < _expensive_nodes.length(); i++) { 4679 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) { 4680 return false; 4681 } 4682 } 4683 return true; 4684 } 4685 4686 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4687 if (_expensive_nodes.length() == 0) { 4688 return false; 4689 } 4690 4691 assert(OptimizeExpensiveOps, "optimization off?"); 4692 4693 // Take this opportunity to remove dead nodes from the list 4694 int j = 0; 4695 for (int i = 0; i < _expensive_nodes.length(); i++) { 4696 Node* n = _expensive_nodes.at(i); 4697 if (!n->is_unreachable(igvn)) { 4698 assert(n->is_expensive(), "should be expensive"); 4699 _expensive_nodes.at_put(j, n); 4700 j++; 4701 } 4702 } 4703 _expensive_nodes.trunc_to(j); 4704 4705 // Then sort the list so that similar nodes are next to each other 4706 // and check for at least two nodes of identical kind with same data 4707 // inputs. 4708 sort_expensive_nodes(); 4709 4710 for (int i = 0; i < _expensive_nodes.length()-1; i++) { 4711 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) { 4712 return true; 4713 } 4714 } 4715 4716 return false; 4717 } 4718 4719 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4720 if (_expensive_nodes.length() == 0) { 4721 return; 4722 } 4723 4724 assert(OptimizeExpensiveOps, "optimization off?"); 4725 4726 // Sort to bring similar nodes next to each other and clear the 4727 // control input of nodes for which there's only a single copy. 4728 sort_expensive_nodes(); 4729 4730 int j = 0; 4731 int identical = 0; 4732 int i = 0; 4733 bool modified = false; 4734 for (; i < _expensive_nodes.length()-1; i++) { 4735 assert(j <= i, "can't write beyond current index"); 4736 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) { 4737 identical++; 4738 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4739 continue; 4740 } 4741 if (identical > 0) { 4742 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4743 identical = 0; 4744 } else { 4745 Node* n = _expensive_nodes.at(i); 4746 igvn.replace_input_of(n, 0, nullptr); 4747 igvn.hash_insert(n); 4748 modified = true; 4749 } 4750 } 4751 if (identical > 0) { 4752 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4753 } else if (_expensive_nodes.length() >= 1) { 4754 Node* n = _expensive_nodes.at(i); 4755 igvn.replace_input_of(n, 0, nullptr); 4756 igvn.hash_insert(n); 4757 modified = true; 4758 } 4759 _expensive_nodes.trunc_to(j); 4760 if (modified) { 4761 igvn.optimize(); 4762 } 4763 } 4764 4765 void Compile::add_expensive_node(Node * n) { 4766 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list"); 4767 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4768 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4769 if (OptimizeExpensiveOps) { 4770 _expensive_nodes.append(n); 4771 } else { 4772 // Clear control input and let IGVN optimize expensive nodes if 4773 // OptimizeExpensiveOps is off. 4774 n->set_req(0, nullptr); 4775 } 4776 } 4777 4778 /** 4779 * Track coarsened Lock and Unlock nodes. 4780 */ 4781 4782 class Lock_List : public Node_List { 4783 uint _origin_cnt; 4784 public: 4785 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {} 4786 uint origin_cnt() const { return _origin_cnt; } 4787 }; 4788 4789 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) { 4790 int length = locks.length(); 4791 if (length > 0) { 4792 // Have to keep this list until locks elimination during Macro nodes elimination. 4793 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length); 4794 AbstractLockNode* alock = locks.at(0); 4795 BoxLockNode* box = alock->box_node()->as_BoxLock(); 4796 for (int i = 0; i < length; i++) { 4797 AbstractLockNode* lock = locks.at(i); 4798 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx); 4799 locks_list->push(lock); 4800 BoxLockNode* this_box = lock->box_node()->as_BoxLock(); 4801 if (this_box != box) { 4802 // Locking regions (BoxLock) could be Unbalanced here: 4803 // - its coarsened locks were eliminated in earlier 4804 // macro nodes elimination followed by loop unroll 4805 // - it is OSR locking region (no Lock node) 4806 // Preserve Unbalanced status in such cases. 4807 if (!this_box->is_unbalanced()) { 4808 this_box->set_coarsened(); 4809 } 4810 if (!box->is_unbalanced()) { 4811 box->set_coarsened(); 4812 } 4813 } 4814 } 4815 _coarsened_locks.append(locks_list); 4816 } 4817 } 4818 4819 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) { 4820 int count = coarsened_count(); 4821 for (int i = 0; i < count; i++) { 4822 Node_List* locks_list = _coarsened_locks.at(i); 4823 for (uint j = 0; j < locks_list->size(); j++) { 4824 Node* lock = locks_list->at(j); 4825 assert(lock->is_AbstractLock(), "sanity"); 4826 if (!useful.member(lock)) { 4827 locks_list->yank(lock); 4828 } 4829 } 4830 } 4831 } 4832 4833 void Compile::remove_coarsened_lock(Node* n) { 4834 if (n->is_AbstractLock()) { 4835 int count = coarsened_count(); 4836 for (int i = 0; i < count; i++) { 4837 Node_List* locks_list = _coarsened_locks.at(i); 4838 locks_list->yank(n); 4839 } 4840 } 4841 } 4842 4843 bool Compile::coarsened_locks_consistent() { 4844 int count = coarsened_count(); 4845 for (int i = 0; i < count; i++) { 4846 bool unbalanced = false; 4847 bool modified = false; // track locks kind modifications 4848 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i); 4849 uint size = locks_list->size(); 4850 if (size == 0) { 4851 unbalanced = false; // All locks were eliminated - good 4852 } else if (size != locks_list->origin_cnt()) { 4853 unbalanced = true; // Some locks were removed from list 4854 } else { 4855 for (uint j = 0; j < size; j++) { 4856 Node* lock = locks_list->at(j); 4857 // All nodes in group should have the same state (modified or not) 4858 if (!lock->as_AbstractLock()->is_coarsened()) { 4859 if (j == 0) { 4860 // first on list was modified, the rest should be too for consistency 4861 modified = true; 4862 } else if (!modified) { 4863 // this lock was modified but previous locks on the list were not 4864 unbalanced = true; 4865 break; 4866 } 4867 } else if (modified) { 4868 // previous locks on list were modified but not this lock 4869 unbalanced = true; 4870 break; 4871 } 4872 } 4873 } 4874 if (unbalanced) { 4875 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified 4876 #ifdef ASSERT 4877 if (PrintEliminateLocks) { 4878 tty->print_cr("=== unbalanced coarsened locks ==="); 4879 for (uint l = 0; l < size; l++) { 4880 locks_list->at(l)->dump(); 4881 } 4882 } 4883 #endif 4884 record_failure(C2Compiler::retry_no_locks_coarsening()); 4885 return false; 4886 } 4887 } 4888 return true; 4889 } 4890 4891 // Mark locking regions (identified by BoxLockNode) as unbalanced if 4892 // locks coarsening optimization removed Lock/Unlock nodes from them. 4893 // Such regions become unbalanced because coarsening only removes part 4894 // of Lock/Unlock nodes in region. As result we can't execute other 4895 // locks elimination optimizations which assume all code paths have 4896 // corresponding pair of Lock/Unlock nodes - they are balanced. 4897 void Compile::mark_unbalanced_boxes() const { 4898 int count = coarsened_count(); 4899 for (int i = 0; i < count; i++) { 4900 Node_List* locks_list = _coarsened_locks.at(i); 4901 uint size = locks_list->size(); 4902 if (size > 0) { 4903 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock(); 4904 BoxLockNode* box = alock->box_node()->as_BoxLock(); 4905 if (alock->is_coarsened()) { 4906 // coarsened_locks_consistent(), which is called before this method, verifies 4907 // that the rest of Lock/Unlock nodes on locks_list are also coarsened. 4908 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated"); 4909 for (uint j = 1; j < size; j++) { 4910 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here"); 4911 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock(); 4912 if (box != this_box) { 4913 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated"); 4914 box->set_unbalanced(); 4915 this_box->set_unbalanced(); 4916 } 4917 } 4918 } 4919 } 4920 } 4921 } 4922 4923 /** 4924 * Remove the speculative part of types and clean up the graph 4925 */ 4926 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4927 if (UseTypeSpeculation) { 4928 Unique_Node_List worklist; 4929 worklist.push(root()); 4930 int modified = 0; 4931 // Go over all type nodes that carry a speculative type, drop the 4932 // speculative part of the type and enqueue the node for an igvn 4933 // which may optimize it out. 4934 for (uint next = 0; next < worklist.size(); ++next) { 4935 Node *n = worklist.at(next); 4936 if (n->is_Type()) { 4937 TypeNode* tn = n->as_Type(); 4938 const Type* t = tn->type(); 4939 const Type* t_no_spec = t->remove_speculative(); 4940 if (t_no_spec != t) { 4941 bool in_hash = igvn.hash_delete(n); 4942 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table"); 4943 tn->set_type(t_no_spec); 4944 igvn.hash_insert(n); 4945 igvn._worklist.push(n); // give it a chance to go away 4946 modified++; 4947 } 4948 } 4949 // Iterate over outs - endless loops is unreachable from below 4950 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4951 Node *m = n->fast_out(i); 4952 if (not_a_node(m)) { 4953 continue; 4954 } 4955 worklist.push(m); 4956 } 4957 } 4958 // Drop the speculative part of all types in the igvn's type table 4959 igvn.remove_speculative_types(); 4960 if (modified > 0) { 4961 igvn.optimize(); 4962 if (failing()) return; 4963 } 4964 #ifdef ASSERT 4965 // Verify that after the IGVN is over no speculative type has resurfaced 4966 worklist.clear(); 4967 worklist.push(root()); 4968 for (uint next = 0; next < worklist.size(); ++next) { 4969 Node *n = worklist.at(next); 4970 const Type* t = igvn.type_or_null(n); 4971 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types"); 4972 if (n->is_Type()) { 4973 t = n->as_Type()->type(); 4974 assert(t == t->remove_speculative(), "no more speculative types"); 4975 } 4976 // Iterate over outs - endless loops is unreachable from below 4977 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4978 Node *m = n->fast_out(i); 4979 if (not_a_node(m)) { 4980 continue; 4981 } 4982 worklist.push(m); 4983 } 4984 } 4985 igvn.check_no_speculative_types(); 4986 #endif 4987 } 4988 } 4989 4990 // Auxiliary methods to support randomized stressing/fuzzing. 4991 4992 void Compile::initialize_stress_seed(const DirectiveSet* directive) { 4993 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) { 4994 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds()); 4995 FLAG_SET_ERGO(StressSeed, _stress_seed); 4996 } else { 4997 _stress_seed = StressSeed; 4998 } 4999 if (_log != nullptr) { 5000 _log->elem("stress_test seed='%u'", _stress_seed); 5001 } 5002 } 5003 5004 int Compile::random() { 5005 _stress_seed = os::next_random(_stress_seed); 5006 return static_cast<int>(_stress_seed); 5007 } 5008 5009 // This method can be called the arbitrary number of times, with current count 5010 // as the argument. The logic allows selecting a single candidate from the 5011 // running list of candidates as follows: 5012 // int count = 0; 5013 // Cand* selected = null; 5014 // while(cand = cand->next()) { 5015 // if (randomized_select(++count)) { 5016 // selected = cand; 5017 // } 5018 // } 5019 // 5020 // Including count equalizes the chances any candidate is "selected". 5021 // This is useful when we don't have the complete list of candidates to choose 5022 // from uniformly. In this case, we need to adjust the randomicity of the 5023 // selection, or else we will end up biasing the selection towards the latter 5024 // candidates. 5025 // 5026 // Quick back-envelope calculation shows that for the list of n candidates 5027 // the equal probability for the candidate to persist as "best" can be 5028 // achieved by replacing it with "next" k-th candidate with the probability 5029 // of 1/k. It can be easily shown that by the end of the run, the 5030 // probability for any candidate is converged to 1/n, thus giving the 5031 // uniform distribution among all the candidates. 5032 // 5033 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 5034 #define RANDOMIZED_DOMAIN_POW 29 5035 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 5036 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 5037 bool Compile::randomized_select(int count) { 5038 assert(count > 0, "only positive"); 5039 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 5040 } 5041 5042 #ifdef ASSERT 5043 // Failures are geometrically distributed with probability 1/StressBailoutMean. 5044 bool Compile::fail_randomly() { 5045 if ((random() % StressBailoutMean) != 0) { 5046 return false; 5047 } 5048 record_failure("StressBailout"); 5049 return true; 5050 } 5051 5052 bool Compile::failure_is_artificial() { 5053 return C->failure_reason_is("StressBailout"); 5054 } 5055 #endif 5056 5057 CloneMap& Compile::clone_map() { return _clone_map; } 5058 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 5059 5060 void NodeCloneInfo::dump_on(outputStream* st) const { 5061 st->print(" {%d:%d} ", idx(), gen()); 5062 } 5063 5064 void CloneMap::clone(Node* old, Node* nnn, int gen) { 5065 uint64_t val = value(old->_idx); 5066 NodeCloneInfo cio(val); 5067 assert(val != 0, "old node should be in the map"); 5068 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 5069 insert(nnn->_idx, cin.get()); 5070 #ifndef PRODUCT 5071 if (is_debug()) { 5072 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 5073 } 5074 #endif 5075 } 5076 5077 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 5078 NodeCloneInfo cio(value(old->_idx)); 5079 if (cio.get() == 0) { 5080 cio.set(old->_idx, 0); 5081 insert(old->_idx, cio.get()); 5082 #ifndef PRODUCT 5083 if (is_debug()) { 5084 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 5085 } 5086 #endif 5087 } 5088 clone(old, nnn, gen); 5089 } 5090 5091 int CloneMap::max_gen() const { 5092 int g = 0; 5093 DictI di(_dict); 5094 for(; di.test(); ++di) { 5095 int t = gen(di._key); 5096 if (g < t) { 5097 g = t; 5098 #ifndef PRODUCT 5099 if (is_debug()) { 5100 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 5101 } 5102 #endif 5103 } 5104 } 5105 return g; 5106 } 5107 5108 void CloneMap::dump(node_idx_t key, outputStream* st) const { 5109 uint64_t val = value(key); 5110 if (val != 0) { 5111 NodeCloneInfo ni(val); 5112 ni.dump_on(st); 5113 } 5114 } 5115 5116 void Compile::shuffle_macro_nodes() { 5117 if (_macro_nodes.length() < 2) { 5118 return; 5119 } 5120 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) { 5121 uint j = C->random() % (i + 1); 5122 swap(_macro_nodes.at(i), _macro_nodes.at(j)); 5123 } 5124 } 5125 5126 // Move Allocate nodes to the start of the list 5127 void Compile::sort_macro_nodes() { 5128 int count = macro_count(); 5129 int allocates = 0; 5130 for (int i = 0; i < count; i++) { 5131 Node* n = macro_node(i); 5132 if (n->is_Allocate()) { 5133 if (i != allocates) { 5134 Node* tmp = macro_node(allocates); 5135 _macro_nodes.at_put(allocates, n); 5136 _macro_nodes.at_put(i, tmp); 5137 } 5138 allocates++; 5139 } 5140 } 5141 } 5142 5143 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) { 5144 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code. 5145 EventCompilerPhase event(UNTIMED); 5146 if (event.should_commit()) { 5147 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 5148 } 5149 #ifndef PRODUCT 5150 ResourceMark rm; 5151 stringStream ss; 5152 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt)); 5153 int iter = ++_igv_phase_iter[cpt]; 5154 if (iter > 1) { 5155 ss.print(" %d", iter); 5156 } 5157 if (n != nullptr) { 5158 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]); 5159 if (n->is_Call()) { 5160 CallNode* call = n->as_Call(); 5161 if (call->_name != nullptr) { 5162 // E.g. uncommon traps etc. 5163 ss.print(" - %s", call->_name); 5164 } else if (call->is_CallJava()) { 5165 CallJavaNode* call_java = call->as_CallJava(); 5166 if (call_java->method() != nullptr) { 5167 ss.print(" -"); 5168 call_java->method()->print_short_name(&ss); 5169 } 5170 } 5171 } 5172 } 5173 5174 const char* name = ss.as_string(); 5175 if (should_print_igv(level)) { 5176 _igv_printer->print_graph(name); 5177 } 5178 if (should_print_phase(level)) { 5179 print_phase(name); 5180 } 5181 if (should_print_ideal_phase(cpt)) { 5182 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt)); 5183 } 5184 #endif 5185 C->_latest_stage_start_counter.stamp(); 5186 } 5187 5188 // Only used from CompileWrapper 5189 void Compile::begin_method() { 5190 #ifndef PRODUCT 5191 if (_method != nullptr && should_print_igv(1)) { 5192 _igv_printer->begin_method(); 5193 } 5194 #endif 5195 C->_latest_stage_start_counter.stamp(); 5196 } 5197 5198 // Only used from CompileWrapper 5199 void Compile::end_method() { 5200 EventCompilerPhase event(UNTIMED); 5201 if (event.should_commit()) { 5202 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1); 5203 } 5204 5205 #ifndef PRODUCT 5206 if (_method != nullptr && should_print_igv(1)) { 5207 _igv_printer->end_method(); 5208 } 5209 #endif 5210 } 5211 5212 #ifndef PRODUCT 5213 bool Compile::should_print_phase(const int level) const { 5214 return PrintPhaseLevel > 0 && directive()->PhasePrintLevelOption >= level && 5215 _method != nullptr; // Do not print phases for stubs. 5216 } 5217 5218 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const { 5219 return _directive->should_print_ideal_phase(cpt); 5220 } 5221 5222 void Compile::init_igv() { 5223 if (_igv_printer == nullptr) { 5224 _igv_printer = IdealGraphPrinter::printer(); 5225 _igv_printer->set_compile(this); 5226 } 5227 } 5228 5229 bool Compile::should_print_igv(const int level) { 5230 PRODUCT_RETURN_(return false;); 5231 5232 if (PrintIdealGraphLevel < 0) { // disabled by the user 5233 return false; 5234 } 5235 5236 bool need = directive()->IGVPrintLevelOption >= level; 5237 if (need) { 5238 Compile::init_igv(); 5239 } 5240 return need; 5241 } 5242 5243 IdealGraphPrinter* Compile::_debug_file_printer = nullptr; 5244 IdealGraphPrinter* Compile::_debug_network_printer = nullptr; 5245 5246 // Called from debugger. Prints method to the default file with the default phase name. 5247 // This works regardless of any Ideal Graph Visualizer flags set or not. 5248 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc). 5249 void igv_print(void* sp, void* fp, void* pc) { 5250 frame fr(sp, fp, pc); 5251 Compile::current()->igv_print_method_to_file(nullptr, false, &fr); 5252 } 5253 5254 // Same as igv_print() above but with a specified phase name. 5255 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) { 5256 frame fr(sp, fp, pc); 5257 Compile::current()->igv_print_method_to_file(phase_name, false, &fr); 5258 } 5259 5260 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 5261 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 5262 // This works regardless of any Ideal Graph Visualizer flags set or not. 5263 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc). 5264 void igv_print(bool network, void* sp, void* fp, void* pc) { 5265 frame fr(sp, fp, pc); 5266 if (network) { 5267 Compile::current()->igv_print_method_to_network(nullptr, &fr); 5268 } else { 5269 Compile::current()->igv_print_method_to_file(nullptr, false, &fr); 5270 } 5271 } 5272 5273 // Same as igv_print(bool network, ...) above but with a specified phase name. 5274 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc). 5275 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) { 5276 frame fr(sp, fp, pc); 5277 if (network) { 5278 Compile::current()->igv_print_method_to_network(phase_name, &fr); 5279 } else { 5280 Compile::current()->igv_print_method_to_file(phase_name, false, &fr); 5281 } 5282 } 5283 5284 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 5285 void igv_print_default() { 5286 Compile::current()->print_method(PHASE_DEBUG, 0); 5287 } 5288 5289 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 5290 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 5291 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 5292 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc). 5293 void igv_append(void* sp, void* fp, void* pc) { 5294 frame fr(sp, fp, pc); 5295 Compile::current()->igv_print_method_to_file(nullptr, true, &fr); 5296 } 5297 5298 // Same as igv_append(...) above but with a specified phase name. 5299 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc). 5300 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) { 5301 frame fr(sp, fp, pc); 5302 Compile::current()->igv_print_method_to_file(phase_name, true, &fr); 5303 } 5304 5305 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) { 5306 const char* file_name = "custom_debug.xml"; 5307 if (_debug_file_printer == nullptr) { 5308 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 5309 } else { 5310 _debug_file_printer->update_compiled_method(C->method()); 5311 } 5312 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 5313 _debug_file_printer->print_graph(phase_name, fr); 5314 } 5315 5316 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) { 5317 ResourceMark rm; 5318 GrowableArray<const Node*> empty_list; 5319 igv_print_graph_to_network(phase_name, empty_list, fr); 5320 } 5321 5322 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) { 5323 if (_debug_network_printer == nullptr) { 5324 _debug_network_printer = new IdealGraphPrinter(C); 5325 } else { 5326 _debug_network_printer->update_compiled_method(C->method()); 5327 } 5328 tty->print_cr("Method printed over network stream to IGV"); 5329 _debug_network_printer->print(name, C->root(), visible_nodes, fr); 5330 } 5331 #endif // !PRODUCT 5332 5333 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 5334 if (type != nullptr && phase->type(value)->higher_equal(type)) { 5335 return value; 5336 } 5337 Node* result = nullptr; 5338 if (bt == T_BYTE) { 5339 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 5340 result = new RShiftINode(result, phase->intcon(24)); 5341 } else if (bt == T_BOOLEAN) { 5342 result = new AndINode(value, phase->intcon(0xFF)); 5343 } else if (bt == T_CHAR) { 5344 result = new AndINode(value,phase->intcon(0xFFFF)); 5345 } else { 5346 assert(bt == T_SHORT, "unexpected narrow type"); 5347 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 5348 result = new RShiftINode(result, phase->intcon(16)); 5349 } 5350 if (transform_res) { 5351 result = phase->transform(result); 5352 } 5353 return result; 5354 } 5355 5356 void Compile::record_method_not_compilable_oom() { 5357 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit()); 5358 }