1 /* 2 * Copyright (c) 1997, 2023, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/javaClasses.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "gc/shared/barrierSet.hpp" 37 #include "gc/shared/c2/barrierSetC2.hpp" 38 #include "jfr/jfrEvents.hpp" 39 #include "jvm_io.h" 40 #include "memory/allocation.hpp" 41 #include "memory/resourceArea.hpp" 42 #include "opto/addnode.hpp" 43 #include "opto/block.hpp" 44 #include "opto/c2compiler.hpp" 45 #include "opto/callGenerator.hpp" 46 #include "opto/callnode.hpp" 47 #include "opto/castnode.hpp" 48 #include "opto/cfgnode.hpp" 49 #include "opto/chaitin.hpp" 50 #include "opto/compile.hpp" 51 #include "opto/connode.hpp" 52 #include "opto/convertnode.hpp" 53 #include "opto/divnode.hpp" 54 #include "opto/escape.hpp" 55 #include "opto/idealGraphPrinter.hpp" 56 #include "opto/loopnode.hpp" 57 #include "opto/machnode.hpp" 58 #include "opto/macro.hpp" 59 #include "opto/matcher.hpp" 60 #include "opto/mathexactnode.hpp" 61 #include "opto/memnode.hpp" 62 #include "opto/mulnode.hpp" 63 #include "opto/narrowptrnode.hpp" 64 #include "opto/node.hpp" 65 #include "opto/opcodes.hpp" 66 #include "opto/output.hpp" 67 #include "opto/parse.hpp" 68 #include "opto/phaseX.hpp" 69 #include "opto/rootnode.hpp" 70 #include "opto/runtime.hpp" 71 #include "opto/stringopts.hpp" 72 #include "opto/type.hpp" 73 #include "opto/vector.hpp" 74 #include "opto/vectornode.hpp" 75 #include "runtime/globals_extension.hpp" 76 #include "runtime/sharedRuntime.hpp" 77 #include "runtime/signature.hpp" 78 #include "runtime/stubRoutines.hpp" 79 #include "runtime/timer.hpp" 80 #include "utilities/align.hpp" 81 #include "utilities/copy.hpp" 82 #include "utilities/macros.hpp" 83 #include "utilities/resourceHash.hpp" 84 85 // -------------------- Compile::mach_constant_base_node ----------------------- 86 // Constant table base node singleton. 87 MachConstantBaseNode* Compile::mach_constant_base_node() { 88 if (_mach_constant_base_node == nullptr) { 89 _mach_constant_base_node = new MachConstantBaseNode(); 90 _mach_constant_base_node->add_req(C->root()); 91 } 92 return _mach_constant_base_node; 93 } 94 95 96 /// Support for intrinsics. 97 98 // Return the index at which m must be inserted (or already exists). 99 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 100 class IntrinsicDescPair { 101 private: 102 ciMethod* _m; 103 bool _is_virtual; 104 public: 105 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 106 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 107 ciMethod* m= elt->method(); 108 ciMethod* key_m = key->_m; 109 if (key_m < m) return -1; 110 else if (key_m > m) return 1; 111 else { 112 bool is_virtual = elt->is_virtual(); 113 bool key_virtual = key->_is_virtual; 114 if (key_virtual < is_virtual) return -1; 115 else if (key_virtual > is_virtual) return 1; 116 else return 0; 117 } 118 } 119 }; 120 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 121 #ifdef ASSERT 122 for (int i = 1; i < _intrinsics.length(); i++) { 123 CallGenerator* cg1 = _intrinsics.at(i-1); 124 CallGenerator* cg2 = _intrinsics.at(i); 125 assert(cg1->method() != cg2->method() 126 ? cg1->method() < cg2->method() 127 : cg1->is_virtual() < cg2->is_virtual(), 128 "compiler intrinsics list must stay sorted"); 129 } 130 #endif 131 IntrinsicDescPair pair(m, is_virtual); 132 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 133 } 134 135 void Compile::register_intrinsic(CallGenerator* cg) { 136 bool found = false; 137 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 138 assert(!found, "registering twice"); 139 _intrinsics.insert_before(index, cg); 140 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 141 } 142 143 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 144 assert(m->is_loaded(), "don't try this on unloaded methods"); 145 if (_intrinsics.length() > 0) { 146 bool found = false; 147 int index = intrinsic_insertion_index(m, is_virtual, found); 148 if (found) { 149 return _intrinsics.at(index); 150 } 151 } 152 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 153 if (m->intrinsic_id() != vmIntrinsics::_none && 154 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 155 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 156 if (cg != nullptr) { 157 // Save it for next time: 158 register_intrinsic(cg); 159 return cg; 160 } else { 161 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 162 } 163 } 164 return nullptr; 165 } 166 167 // Compile::make_vm_intrinsic is defined in library_call.cpp. 168 169 #ifndef PRODUCT 170 // statistics gathering... 171 172 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0}; 173 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0}; 174 175 inline int as_int(vmIntrinsics::ID id) { 176 return vmIntrinsics::as_int(id); 177 } 178 179 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 180 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 181 int oflags = _intrinsic_hist_flags[as_int(id)]; 182 assert(flags != 0, "what happened?"); 183 if (is_virtual) { 184 flags |= _intrinsic_virtual; 185 } 186 bool changed = (flags != oflags); 187 if ((flags & _intrinsic_worked) != 0) { 188 juint count = (_intrinsic_hist_count[as_int(id)] += 1); 189 if (count == 1) { 190 changed = true; // first time 191 } 192 // increment the overall count also: 193 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1; 194 } 195 if (changed) { 196 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 197 // Something changed about the intrinsic's virtuality. 198 if ((flags & _intrinsic_virtual) != 0) { 199 // This is the first use of this intrinsic as a virtual call. 200 if (oflags != 0) { 201 // We already saw it as a non-virtual, so note both cases. 202 flags |= _intrinsic_both; 203 } 204 } else if ((oflags & _intrinsic_both) == 0) { 205 // This is the first use of this intrinsic as a non-virtual 206 flags |= _intrinsic_both; 207 } 208 } 209 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags); 210 } 211 // update the overall flags also: 212 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags; 213 return changed; 214 } 215 216 static char* format_flags(int flags, char* buf) { 217 buf[0] = 0; 218 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 219 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 220 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 221 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 222 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 223 if (buf[0] == 0) strcat(buf, ","); 224 assert(buf[0] == ',', "must be"); 225 return &buf[1]; 226 } 227 228 void Compile::print_intrinsic_statistics() { 229 char flagsbuf[100]; 230 ttyLocker ttyl; 231 if (xtty != nullptr) xtty->head("statistics type='intrinsic'"); 232 tty->print_cr("Compiler intrinsic usage:"); 233 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)]; 234 if (total == 0) total = 1; // avoid div0 in case of no successes 235 #define PRINT_STAT_LINE(name, c, f) \ 236 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 237 for (auto id : EnumRange<vmIntrinsicID>{}) { 238 int flags = _intrinsic_hist_flags[as_int(id)]; 239 juint count = _intrinsic_hist_count[as_int(id)]; 240 if ((flags | count) != 0) { 241 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 242 } 243 } 244 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf)); 245 if (xtty != nullptr) xtty->tail("statistics"); 246 } 247 248 void Compile::print_statistics() { 249 { ttyLocker ttyl; 250 if (xtty != nullptr) xtty->head("statistics type='opto'"); 251 Parse::print_statistics(); 252 PhaseStringOpts::print_statistics(); 253 PhaseCCP::print_statistics(); 254 PhaseRegAlloc::print_statistics(); 255 PhaseOutput::print_statistics(); 256 PhasePeephole::print_statistics(); 257 PhaseIdealLoop::print_statistics(); 258 ConnectionGraph::print_statistics(); 259 PhaseMacroExpand::print_statistics(); 260 if (xtty != nullptr) xtty->tail("statistics"); 261 } 262 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) { 263 // put this under its own <statistics> element. 264 print_intrinsic_statistics(); 265 } 266 } 267 #endif //PRODUCT 268 269 void Compile::gvn_replace_by(Node* n, Node* nn) { 270 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 271 Node* use = n->last_out(i); 272 bool is_in_table = initial_gvn()->hash_delete(use); 273 uint uses_found = 0; 274 for (uint j = 0; j < use->len(); j++) { 275 if (use->in(j) == n) { 276 if (j < use->req()) 277 use->set_req(j, nn); 278 else 279 use->set_prec(j, nn); 280 uses_found++; 281 } 282 } 283 if (is_in_table) { 284 // reinsert into table 285 initial_gvn()->hash_find_insert(use); 286 } 287 record_for_igvn(use); 288 i -= uses_found; // we deleted 1 or more copies of this edge 289 } 290 } 291 292 293 // Identify all nodes that are reachable from below, useful. 294 // Use breadth-first pass that records state in a Unique_Node_List, 295 // recursive traversal is slower. 296 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 297 int estimated_worklist_size = live_nodes(); 298 useful.map( estimated_worklist_size, nullptr ); // preallocate space 299 300 // Initialize worklist 301 if (root() != nullptr) { useful.push(root()); } 302 // If 'top' is cached, declare it useful to preserve cached node 303 if (cached_top_node()) { useful.push(cached_top_node()); } 304 305 // Push all useful nodes onto the list, breadthfirst 306 for( uint next = 0; next < useful.size(); ++next ) { 307 assert( next < unique(), "Unique useful nodes < total nodes"); 308 Node *n = useful.at(next); 309 uint max = n->len(); 310 for( uint i = 0; i < max; ++i ) { 311 Node *m = n->in(i); 312 if (not_a_node(m)) continue; 313 useful.push(m); 314 } 315 } 316 } 317 318 // Update dead_node_list with any missing dead nodes using useful 319 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 320 void Compile::update_dead_node_list(Unique_Node_List &useful) { 321 uint max_idx = unique(); 322 VectorSet& useful_node_set = useful.member_set(); 323 324 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 325 // If node with index node_idx is not in useful set, 326 // mark it as dead in dead node list. 327 if (!useful_node_set.test(node_idx)) { 328 record_dead_node(node_idx); 329 } 330 } 331 } 332 333 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 334 int shift = 0; 335 for (int i = 0; i < inlines->length(); i++) { 336 CallGenerator* cg = inlines->at(i); 337 if (useful.member(cg->call_node())) { 338 if (shift > 0) { 339 inlines->at_put(i - shift, cg); 340 } 341 } else { 342 shift++; // skip over the dead element 343 } 344 } 345 if (shift > 0) { 346 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array 347 } 348 } 349 350 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) { 351 assert(dead != nullptr && dead->is_Call(), "sanity"); 352 int found = 0; 353 for (int i = 0; i < inlines->length(); i++) { 354 if (inlines->at(i)->call_node() == dead) { 355 inlines->remove_at(i); 356 found++; 357 NOT_DEBUG( break; ) // elements are unique, so exit early 358 } 359 } 360 assert(found <= 1, "not unique"); 361 } 362 363 void Compile::remove_useless_nodes(GrowableArray<Node*>& node_list, Unique_Node_List& useful) { 364 for (int i = node_list.length() - 1; i >= 0; i--) { 365 Node* n = node_list.at(i); 366 if (!useful.member(n)) { 367 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 368 } 369 } 370 } 371 372 void Compile::remove_useless_node(Node* dead) { 373 remove_modified_node(dead); 374 375 // Constant node that has no out-edges and has only one in-edge from 376 // root is usually dead. However, sometimes reshaping walk makes 377 // it reachable by adding use edges. So, we will NOT count Con nodes 378 // as dead to be conservative about the dead node count at any 379 // given time. 380 if (!dead->is_Con()) { 381 record_dead_node(dead->_idx); 382 } 383 if (dead->is_macro()) { 384 remove_macro_node(dead); 385 } 386 if (dead->is_expensive()) { 387 remove_expensive_node(dead); 388 } 389 if (dead->Opcode() == Op_Opaque4) { 390 remove_template_assertion_predicate_opaq(dead); 391 } 392 if (dead->for_post_loop_opts_igvn()) { 393 remove_from_post_loop_opts_igvn(dead); 394 } 395 if (dead->is_Call()) { 396 remove_useless_late_inlines( &_late_inlines, dead); 397 remove_useless_late_inlines( &_string_late_inlines, dead); 398 remove_useless_late_inlines( &_boxing_late_inlines, dead); 399 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead); 400 401 if (dead->is_CallStaticJava()) { 402 remove_unstable_if_trap(dead->as_CallStaticJava(), false); 403 } 404 } 405 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 406 bs->unregister_potential_barrier_node(dead); 407 } 408 409 // Disconnect all useless nodes by disconnecting those at the boundary. 410 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) { 411 uint next = 0; 412 while (next < useful.size()) { 413 Node *n = useful.at(next++); 414 if (n->is_SafePoint()) { 415 // We're done with a parsing phase. Replaced nodes are not valid 416 // beyond that point. 417 n->as_SafePoint()->delete_replaced_nodes(); 418 } 419 // Use raw traversal of out edges since this code removes out edges 420 int max = n->outcnt(); 421 for (int j = 0; j < max; ++j) { 422 Node* child = n->raw_out(j); 423 if (!useful.member(child)) { 424 assert(!child->is_top() || child != top(), 425 "If top is cached in Compile object it is in useful list"); 426 // Only need to remove this out-edge to the useless node 427 n->raw_del_out(j); 428 --j; 429 --max; 430 } 431 } 432 if (n->outcnt() == 1 && n->has_special_unique_user()) { 433 assert(useful.member(n->unique_out()), "do not push a useless node"); 434 worklist.push(n->unique_out()); 435 } 436 } 437 438 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes 439 remove_useless_nodes(_parse_predicate_opaqs, useful); // remove useless Parse Predicate opaque nodes 440 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes 441 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes 442 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass 443 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps 444 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes 445 #ifdef ASSERT 446 if (_modified_nodes != nullptr) { 447 _modified_nodes->remove_useless_nodes(useful.member_set()); 448 } 449 #endif 450 451 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 452 bs->eliminate_useless_gc_barriers(useful, this); 453 // clean up the late inline lists 454 remove_useless_late_inlines( &_late_inlines, useful); 455 remove_useless_late_inlines( &_string_late_inlines, useful); 456 remove_useless_late_inlines( &_boxing_late_inlines, useful); 457 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful); 458 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 459 } 460 461 // ============================================================================ 462 //------------------------------CompileWrapper--------------------------------- 463 class CompileWrapper : public StackObj { 464 Compile *const _compile; 465 public: 466 CompileWrapper(Compile* compile); 467 468 ~CompileWrapper(); 469 }; 470 471 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 472 // the Compile* pointer is stored in the current ciEnv: 473 ciEnv* env = compile->env(); 474 assert(env == ciEnv::current(), "must already be a ciEnv active"); 475 assert(env->compiler_data() == nullptr, "compile already active?"); 476 env->set_compiler_data(compile); 477 assert(compile == Compile::current(), "sanity"); 478 479 compile->set_type_dict(nullptr); 480 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 481 compile->clone_map().set_clone_idx(0); 482 compile->set_type_last_size(0); 483 compile->set_last_tf(nullptr, nullptr); 484 compile->set_indexSet_arena(nullptr); 485 compile->set_indexSet_free_block_list(nullptr); 486 compile->init_type_arena(); 487 Type::Initialize(compile); 488 _compile->begin_method(); 489 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 490 } 491 CompileWrapper::~CompileWrapper() { 492 // simulate crash during compilation 493 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned"); 494 495 _compile->end_method(); 496 _compile->env()->set_compiler_data(nullptr); 497 } 498 499 500 //----------------------------print_compile_messages--------------------------- 501 void Compile::print_compile_messages() { 502 #ifndef PRODUCT 503 // Check if recompiling 504 if (!subsume_loads() && PrintOpto) { 505 // Recompiling without allowing machine instructions to subsume loads 506 tty->print_cr("*********************************************************"); 507 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 508 tty->print_cr("*********************************************************"); 509 } 510 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) { 511 // Recompiling without escape analysis 512 tty->print_cr("*********************************************************"); 513 tty->print_cr("** Bailout: Recompile without escape analysis **"); 514 tty->print_cr("*********************************************************"); 515 } 516 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) { 517 // Recompiling without iterative escape analysis 518 tty->print_cr("*********************************************************"); 519 tty->print_cr("** Bailout: Recompile without iterative escape analysis**"); 520 tty->print_cr("*********************************************************"); 521 } 522 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) { 523 // Recompiling without boxing elimination 524 tty->print_cr("*********************************************************"); 525 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 526 tty->print_cr("*********************************************************"); 527 } 528 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) { 529 // Recompiling without locks coarsening 530 tty->print_cr("*********************************************************"); 531 tty->print_cr("** Bailout: Recompile without locks coarsening **"); 532 tty->print_cr("*********************************************************"); 533 } 534 if (env()->break_at_compile()) { 535 // Open the debugger when compiling this method. 536 tty->print("### Breaking when compiling: "); 537 method()->print_short_name(); 538 tty->cr(); 539 BREAKPOINT; 540 } 541 542 if( PrintOpto ) { 543 if (is_osr_compilation()) { 544 tty->print("[OSR]%3d", _compile_id); 545 } else { 546 tty->print("%3d", _compile_id); 547 } 548 } 549 #endif 550 } 551 552 #ifndef PRODUCT 553 void Compile::print_ideal_ir(const char* phase_name) { 554 ttyLocker ttyl; 555 // keep the following output all in one block 556 // This output goes directly to the tty, not the compiler log. 557 // To enable tools to match it up with the compilation activity, 558 // be sure to tag this tty output with the compile ID. 559 if (xtty != nullptr) { 560 xtty->head("ideal compile_id='%d'%s compile_phase='%s'", 561 compile_id(), 562 is_osr_compilation() ? " compile_kind='osr'" : "", 563 phase_name); 564 } 565 if (_output == nullptr) { 566 tty->print_cr("AFTER: %s", phase_name); 567 // Print out all nodes in ascending order of index. 568 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$"); 569 } else { 570 // Dump the node blockwise if we have a scheduling 571 _output->print_scheduling(); 572 } 573 574 if (xtty != nullptr) { 575 xtty->tail("ideal"); 576 } 577 } 578 #endif 579 580 // ============================================================================ 581 //------------------------------Compile standard------------------------------- 582 583 // Compile a method. entry_bci is -1 for normal compilations and indicates 584 // the continuation bci for on stack replacement. 585 586 587 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci, 588 Options options, DirectiveSet* directive) 589 : Phase(Compiler), 590 _compile_id(ci_env->compile_id()), 591 _options(options), 592 _method(target), 593 _entry_bci(osr_bci), 594 _ilt(nullptr), 595 _stub_function(nullptr), 596 _stub_name(nullptr), 597 _stub_entry_point(nullptr), 598 _max_node_limit(MaxNodeLimit), 599 _post_loop_opts_phase(false), 600 _inlining_progress(false), 601 _inlining_incrementally(false), 602 _do_cleanup(false), 603 _has_reserved_stack_access(target->has_reserved_stack_access()), 604 #ifndef PRODUCT 605 _igv_idx(0), 606 _trace_opto_output(directive->TraceOptoOutputOption), 607 #endif 608 _has_method_handle_invokes(false), 609 _clinit_barrier_on_entry(false), 610 _stress_seed(0), 611 _comp_arena(mtCompiler), 612 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 613 _env(ci_env), 614 _directive(directive), 615 _log(ci_env->log()), 616 _failure_reason(nullptr), 617 _intrinsics (comp_arena(), 0, 0, nullptr), 618 _macro_nodes (comp_arena(), 8, 0, nullptr), 619 _parse_predicate_opaqs (comp_arena(), 8, 0, nullptr), 620 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr), 621 _expensive_nodes (comp_arena(), 8, 0, nullptr), 622 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr), 623 _unstable_if_traps (comp_arena(), 8, 0, nullptr), 624 _coarsened_locks (comp_arena(), 8, 0, nullptr), 625 _congraph(nullptr), 626 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 627 _dead_node_list(comp_arena()), 628 _dead_node_count(0), 629 _node_arena(mtCompiler), 630 _old_arena(mtCompiler), 631 _mach_constant_base_node(nullptr), 632 _Compile_types(mtCompiler), 633 _initial_gvn(nullptr), 634 _igvn_worklist(nullptr), 635 _types(nullptr), 636 _node_hash(nullptr), 637 _late_inlines(comp_arena(), 2, 0, nullptr), 638 _string_late_inlines(comp_arena(), 2, 0, nullptr), 639 _boxing_late_inlines(comp_arena(), 2, 0, nullptr), 640 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr), 641 _late_inlines_pos(0), 642 _number_of_mh_late_inlines(0), 643 _print_inlining_stream(new (mtCompiler) stringStream()), 644 _print_inlining_list(nullptr), 645 _print_inlining_idx(0), 646 _print_inlining_output(nullptr), 647 _replay_inline_data(nullptr), 648 _java_calls(0), 649 _inner_loops(0), 650 _interpreter_frame_size(0), 651 _output(nullptr) 652 #ifndef PRODUCT 653 , _in_dump_cnt(0) 654 #endif 655 { 656 C = this; 657 CompileWrapper cw(this); 658 659 if (CITimeVerbose) { 660 tty->print(" "); 661 target->holder()->name()->print(); 662 tty->print("."); 663 target->print_short_name(); 664 tty->print(" "); 665 } 666 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 667 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false); 668 669 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 670 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 671 // We can always print a disassembly, either abstract (hex dump) or 672 // with the help of a suitable hsdis library. Thus, we should not 673 // couple print_assembly and print_opto_assembly controls. 674 // But: always print opto and regular assembly on compile command 'print'. 675 bool print_assembly = directive->PrintAssemblyOption; 676 set_print_assembly(print_opto_assembly || print_assembly); 677 #else 678 set_print_assembly(false); // must initialize. 679 #endif 680 681 #ifndef PRODUCT 682 set_parsed_irreducible_loop(false); 683 #endif 684 685 if (directive->ReplayInlineOption) { 686 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 687 } 688 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 689 set_print_intrinsics(directive->PrintIntrinsicsOption); 690 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 691 692 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 693 // Make sure the method being compiled gets its own MDO, 694 // so we can at least track the decompile_count(). 695 // Need MDO to record RTM code generation state. 696 method()->ensure_method_data(); 697 } 698 699 Init(/*do_aliasing=*/ true); 700 701 print_compile_messages(); 702 703 _ilt = InlineTree::build_inline_tree_root(); 704 705 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 706 assert(num_alias_types() >= AliasIdxRaw, ""); 707 708 #define MINIMUM_NODE_HASH 1023 709 710 // GVN that will be run immediately on new nodes 711 uint estimated_size = method()->code_size()*4+64; 712 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 713 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 714 _types = new (comp_arena()) Type_Array(comp_arena()); 715 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size); 716 PhaseGVN gvn; 717 set_initial_gvn(&gvn); 718 719 print_inlining_init(); 720 { // Scope for timing the parser 721 TracePhase tp("parse", &timers[_t_parser]); 722 723 // Put top into the hash table ASAP. 724 initial_gvn()->transform_no_reclaim(top()); 725 726 // Set up tf(), start(), and find a CallGenerator. 727 CallGenerator* cg = nullptr; 728 if (is_osr_compilation()) { 729 const TypeTuple *domain = StartOSRNode::osr_domain(); 730 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 731 init_tf(TypeFunc::make(domain, range)); 732 StartNode* s = new StartOSRNode(root(), domain); 733 initial_gvn()->set_type_bottom(s); 734 init_start(s); 735 cg = CallGenerator::for_osr(method(), entry_bci()); 736 } else { 737 // Normal case. 738 init_tf(TypeFunc::make(method())); 739 StartNode* s = new StartNode(root(), tf()->domain()); 740 initial_gvn()->set_type_bottom(s); 741 init_start(s); 742 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 743 // With java.lang.ref.reference.get() we must go through the 744 // intrinsic - even when get() is the root 745 // method of the compile - so that, if necessary, the value in 746 // the referent field of the reference object gets recorded by 747 // the pre-barrier code. 748 cg = find_intrinsic(method(), false); 749 } 750 if (cg == nullptr) { 751 float past_uses = method()->interpreter_invocation_count(); 752 float expected_uses = past_uses; 753 cg = CallGenerator::for_inline(method(), expected_uses); 754 } 755 } 756 if (failing()) return; 757 if (cg == nullptr) { 758 const char* reason = InlineTree::check_can_parse(method()); 759 assert(reason != nullptr, "expect reason for parse failure"); 760 stringStream ss; 761 ss.print("cannot parse method: %s", reason); 762 record_method_not_compilable(ss.as_string()); 763 return; 764 } 765 766 gvn.set_type(root(), root()->bottom_type()); 767 768 JVMState* jvms = build_start_state(start(), tf()); 769 if ((jvms = cg->generate(jvms)) == nullptr) { 770 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 771 assert(failure_reason() != nullptr, "expect reason for parse failure"); 772 stringStream ss; 773 ss.print("method parse failed: %s", failure_reason()); 774 record_method_not_compilable(ss.as_string()); 775 } 776 return; 777 } 778 GraphKit kit(jvms); 779 780 if (!kit.stopped()) { 781 // Accept return values, and transfer control we know not where. 782 // This is done by a special, unique ReturnNode bound to root. 783 return_values(kit.jvms()); 784 } 785 786 if (kit.has_exceptions()) { 787 // Any exceptions that escape from this call must be rethrown 788 // to whatever caller is dynamically above us on the stack. 789 // This is done by a special, unique RethrowNode bound to root. 790 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 791 } 792 793 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 794 795 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 796 inline_string_calls(true); 797 } 798 799 if (failing()) return; 800 801 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 802 803 // Remove clutter produced by parsing. 804 if (!failing()) { 805 ResourceMark rm; 806 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 807 } 808 } 809 810 // Note: Large methods are capped off in do_one_bytecode(). 811 if (failing()) return; 812 813 // After parsing, node notes are no longer automagic. 814 // They must be propagated by register_new_node_with_optimizer(), 815 // clone(), or the like. 816 set_default_node_notes(nullptr); 817 818 #ifndef PRODUCT 819 if (should_print_igv(1)) { 820 _igv_printer->print_inlining(); 821 } 822 #endif 823 824 if (failing()) return; 825 NOT_PRODUCT( verify_graph_edges(); ) 826 827 // If any phase is randomized for stress testing, seed random number 828 // generation and log the seed for repeatability. 829 if (StressLCM || StressGCM || StressIGVN || StressCCP) { 830 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && RepeatCompilation)) { 831 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds()); 832 FLAG_SET_ERGO(StressSeed, _stress_seed); 833 } else { 834 _stress_seed = StressSeed; 835 } 836 if (_log != nullptr) { 837 _log->elem("stress_test seed='%u'", _stress_seed); 838 } 839 } 840 841 // Now optimize 842 Optimize(); 843 if (failing()) return; 844 NOT_PRODUCT( verify_graph_edges(); ) 845 846 #ifndef PRODUCT 847 if (should_print_ideal()) { 848 print_ideal_ir("print_ideal"); 849 } 850 #endif 851 852 #ifdef ASSERT 853 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 854 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 855 #endif 856 857 // Dump compilation data to replay it. 858 if (directive->DumpReplayOption) { 859 env()->dump_replay_data(_compile_id); 860 } 861 if (directive->DumpInlineOption && (ilt() != nullptr)) { 862 env()->dump_inline_data(_compile_id); 863 } 864 865 // Now that we know the size of all the monitors we can add a fixed slot 866 // for the original deopt pc. 867 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size); 868 set_fixed_slots(next_slot); 869 870 // Compute when to use implicit null checks. Used by matching trap based 871 // nodes and NullCheck optimization. 872 set_allowed_deopt_reasons(); 873 874 // Now generate code 875 Code_Gen(); 876 } 877 878 //------------------------------Compile---------------------------------------- 879 // Compile a runtime stub 880 Compile::Compile( ciEnv* ci_env, 881 TypeFunc_generator generator, 882 address stub_function, 883 const char *stub_name, 884 int is_fancy_jump, 885 bool pass_tls, 886 bool return_pc, 887 DirectiveSet* directive) 888 : Phase(Compiler), 889 _compile_id(0), 890 _options(Options::for_runtime_stub()), 891 _method(nullptr), 892 _entry_bci(InvocationEntryBci), 893 _stub_function(stub_function), 894 _stub_name(stub_name), 895 _stub_entry_point(nullptr), 896 _max_node_limit(MaxNodeLimit), 897 _post_loop_opts_phase(false), 898 _inlining_progress(false), 899 _inlining_incrementally(false), 900 _has_reserved_stack_access(false), 901 #ifndef PRODUCT 902 _igv_idx(0), 903 _trace_opto_output(directive->TraceOptoOutputOption), 904 #endif 905 _has_method_handle_invokes(false), 906 _clinit_barrier_on_entry(false), 907 _stress_seed(0), 908 _comp_arena(mtCompiler), 909 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 910 _env(ci_env), 911 _directive(directive), 912 _log(ci_env->log()), 913 _failure_reason(nullptr), 914 _congraph(nullptr), 915 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 916 _dead_node_list(comp_arena()), 917 _dead_node_count(0), 918 _node_arena(mtCompiler), 919 _old_arena(mtCompiler), 920 _mach_constant_base_node(nullptr), 921 _Compile_types(mtCompiler), 922 _initial_gvn(nullptr), 923 _igvn_worklist(nullptr), 924 _types(nullptr), 925 _node_hash(nullptr), 926 _number_of_mh_late_inlines(0), 927 _print_inlining_stream(new (mtCompiler) stringStream()), 928 _print_inlining_list(nullptr), 929 _print_inlining_idx(0), 930 _print_inlining_output(nullptr), 931 _replay_inline_data(nullptr), 932 _java_calls(0), 933 _inner_loops(0), 934 _interpreter_frame_size(0), 935 _output(nullptr), 936 #ifndef PRODUCT 937 _in_dump_cnt(0), 938 #endif 939 _allowed_reasons(0) { 940 C = this; 941 942 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false); 943 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false); 944 945 #ifndef PRODUCT 946 set_print_assembly(PrintFrameConverterAssembly); 947 set_parsed_irreducible_loop(false); 948 #else 949 set_print_assembly(false); // Must initialize. 950 #endif 951 set_has_irreducible_loop(false); // no loops 952 953 CompileWrapper cw(this); 954 Init(/*do_aliasing=*/ false); 955 init_tf((*generator)()); 956 957 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 958 _types = new (comp_arena()) Type_Array(comp_arena()); 959 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255); 960 { 961 PhaseGVN gvn; 962 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 963 gvn.transform_no_reclaim(top()); 964 965 GraphKit kit; 966 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 967 } 968 969 NOT_PRODUCT( verify_graph_edges(); ) 970 971 Code_Gen(); 972 } 973 974 //------------------------------Init------------------------------------------- 975 // Prepare for a single compilation 976 void Compile::Init(bool aliasing) { 977 _do_aliasing = aliasing; 978 _unique = 0; 979 _regalloc = nullptr; 980 981 _tf = nullptr; // filled in later 982 _top = nullptr; // cached later 983 _matcher = nullptr; // filled in later 984 _cfg = nullptr; // filled in later 985 986 IA32_ONLY( set_24_bit_selection_and_mode(true, false); ) 987 988 _node_note_array = nullptr; 989 _default_node_notes = nullptr; 990 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize() 991 992 _immutable_memory = nullptr; // filled in at first inquiry 993 994 #ifdef ASSERT 995 _phase_optimize_finished = false; 996 _exception_backedge = false; 997 _type_verify = nullptr; 998 #endif 999 1000 // Globally visible Nodes 1001 // First set TOP to null to give safe behavior during creation of RootNode 1002 set_cached_top_node(nullptr); 1003 set_root(new RootNode()); 1004 // Now that you have a Root to point to, create the real TOP 1005 set_cached_top_node( new ConNode(Type::TOP) ); 1006 set_recent_alloc(nullptr, nullptr); 1007 1008 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1009 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1010 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1011 env()->set_dependencies(new Dependencies(env())); 1012 1013 _fixed_slots = 0; 1014 set_has_split_ifs(false); 1015 set_has_loops(false); // first approximation 1016 set_has_stringbuilder(false); 1017 set_has_boxed_value(false); 1018 _trap_can_recompile = false; // no traps emitted yet 1019 _major_progress = true; // start out assuming good things will happen 1020 set_has_unsafe_access(false); 1021 set_max_vector_size(0); 1022 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1023 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1024 set_decompile_count(0); 1025 1026 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1027 _loop_opts_cnt = LoopOptsCount; 1028 set_do_inlining(Inline); 1029 set_max_inline_size(MaxInlineSize); 1030 set_freq_inline_size(FreqInlineSize); 1031 set_do_scheduling(OptoScheduling); 1032 1033 set_do_vector_loop(false); 1034 set_has_monitors(false); 1035 1036 if (AllowVectorizeOnDemand) { 1037 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1038 set_do_vector_loop(true); 1039 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());}) 1040 } else if (has_method() && method()->name() != 0 && 1041 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1042 set_do_vector_loop(true); 1043 } 1044 } 1045 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1046 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());}) 1047 1048 set_rtm_state(NoRTM); // No RTM lock eliding by default 1049 _max_node_limit = _directive->MaxNodeLimitOption; 1050 1051 #if INCLUDE_RTM_OPT 1052 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != nullptr)) { 1053 int rtm_state = method()->method_data()->rtm_state(); 1054 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) { 1055 // Don't generate RTM lock eliding code. 1056 set_rtm_state(NoRTM); 1057 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1058 // Generate RTM lock eliding code without abort ratio calculation code. 1059 set_rtm_state(UseRTM); 1060 } else if (UseRTMDeopt) { 1061 // Generate RTM lock eliding code and include abort ratio calculation 1062 // code if UseRTMDeopt is on. 1063 set_rtm_state(ProfileRTM); 1064 } 1065 } 1066 #endif 1067 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1068 set_clinit_barrier_on_entry(true); 1069 } 1070 if (debug_info()->recording_non_safepoints()) { 1071 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1072 (comp_arena(), 8, 0, nullptr)); 1073 set_default_node_notes(Node_Notes::make(this)); 1074 } 1075 1076 const int grow_ats = 16; 1077 _max_alias_types = grow_ats; 1078 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1079 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1080 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1081 { 1082 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1083 } 1084 // Initialize the first few types. 1085 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr); 1086 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1087 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1088 _num_alias_types = AliasIdxRaw+1; 1089 // Zero out the alias type cache. 1090 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1091 // A null adr_type hits in the cache right away. Preload the right answer. 1092 probe_alias_cache(nullptr)->_index = AliasIdxTop; 1093 } 1094 1095 //---------------------------init_start---------------------------------------- 1096 // Install the StartNode on this compile object. 1097 void Compile::init_start(StartNode* s) { 1098 if (failing()) 1099 return; // already failing 1100 assert(s == start(), ""); 1101 } 1102 1103 /** 1104 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1105 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1106 * the ideal graph. 1107 */ 1108 StartNode* Compile::start() const { 1109 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1110 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1111 Node* start = root()->fast_out(i); 1112 if (start->is_Start()) { 1113 return start->as_Start(); 1114 } 1115 } 1116 fatal("Did not find Start node!"); 1117 return nullptr; 1118 } 1119 1120 //-------------------------------immutable_memory------------------------------------- 1121 // Access immutable memory 1122 Node* Compile::immutable_memory() { 1123 if (_immutable_memory != nullptr) { 1124 return _immutable_memory; 1125 } 1126 StartNode* s = start(); 1127 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1128 Node *p = s->fast_out(i); 1129 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1130 _immutable_memory = p; 1131 return _immutable_memory; 1132 } 1133 } 1134 ShouldNotReachHere(); 1135 return nullptr; 1136 } 1137 1138 //----------------------set_cached_top_node------------------------------------ 1139 // Install the cached top node, and make sure Node::is_top works correctly. 1140 void Compile::set_cached_top_node(Node* tn) { 1141 if (tn != nullptr) verify_top(tn); 1142 Node* old_top = _top; 1143 _top = tn; 1144 // Calling Node::setup_is_top allows the nodes the chance to adjust 1145 // their _out arrays. 1146 if (_top != nullptr) _top->setup_is_top(); 1147 if (old_top != nullptr) old_top->setup_is_top(); 1148 assert(_top == nullptr || top()->is_top(), ""); 1149 } 1150 1151 #ifdef ASSERT 1152 uint Compile::count_live_nodes_by_graph_walk() { 1153 Unique_Node_List useful(comp_arena()); 1154 // Get useful node list by walking the graph. 1155 identify_useful_nodes(useful); 1156 return useful.size(); 1157 } 1158 1159 void Compile::print_missing_nodes() { 1160 1161 // Return if CompileLog is null and PrintIdealNodeCount is false. 1162 if ((_log == nullptr) && (! PrintIdealNodeCount)) { 1163 return; 1164 } 1165 1166 // This is an expensive function. It is executed only when the user 1167 // specifies VerifyIdealNodeCount option or otherwise knows the 1168 // additional work that needs to be done to identify reachable nodes 1169 // by walking the flow graph and find the missing ones using 1170 // _dead_node_list. 1171 1172 Unique_Node_List useful(comp_arena()); 1173 // Get useful node list by walking the graph. 1174 identify_useful_nodes(useful); 1175 1176 uint l_nodes = C->live_nodes(); 1177 uint l_nodes_by_walk = useful.size(); 1178 1179 if (l_nodes != l_nodes_by_walk) { 1180 if (_log != nullptr) { 1181 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1182 _log->stamp(); 1183 _log->end_head(); 1184 } 1185 VectorSet& useful_member_set = useful.member_set(); 1186 int last_idx = l_nodes_by_walk; 1187 for (int i = 0; i < last_idx; i++) { 1188 if (useful_member_set.test(i)) { 1189 if (_dead_node_list.test(i)) { 1190 if (_log != nullptr) { 1191 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1192 } 1193 if (PrintIdealNodeCount) { 1194 // Print the log message to tty 1195 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1196 useful.at(i)->dump(); 1197 } 1198 } 1199 } 1200 else if (! _dead_node_list.test(i)) { 1201 if (_log != nullptr) { 1202 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1203 } 1204 if (PrintIdealNodeCount) { 1205 // Print the log message to tty 1206 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1207 } 1208 } 1209 } 1210 if (_log != nullptr) { 1211 _log->tail("mismatched_nodes"); 1212 } 1213 } 1214 } 1215 void Compile::record_modified_node(Node* n) { 1216 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) { 1217 _modified_nodes->push(n); 1218 } 1219 } 1220 1221 void Compile::remove_modified_node(Node* n) { 1222 if (_modified_nodes != nullptr) { 1223 _modified_nodes->remove(n); 1224 } 1225 } 1226 #endif 1227 1228 #ifndef PRODUCT 1229 void Compile::verify_top(Node* tn) const { 1230 if (tn != nullptr) { 1231 assert(tn->is_Con(), "top node must be a constant"); 1232 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1233 assert(tn->in(0) != nullptr, "must have live top node"); 1234 } 1235 } 1236 #endif 1237 1238 1239 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1240 1241 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1242 guarantee(arr != nullptr, ""); 1243 int num_blocks = arr->length(); 1244 if (grow_by < num_blocks) grow_by = num_blocks; 1245 int num_notes = grow_by * _node_notes_block_size; 1246 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1247 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1248 while (num_notes > 0) { 1249 arr->append(notes); 1250 notes += _node_notes_block_size; 1251 num_notes -= _node_notes_block_size; 1252 } 1253 assert(num_notes == 0, "exact multiple, please"); 1254 } 1255 1256 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1257 if (source == nullptr || dest == nullptr) return false; 1258 1259 if (dest->is_Con()) 1260 return false; // Do not push debug info onto constants. 1261 1262 #ifdef ASSERT 1263 // Leave a bread crumb trail pointing to the original node: 1264 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) { 1265 dest->set_debug_orig(source); 1266 } 1267 #endif 1268 1269 if (node_note_array() == nullptr) 1270 return false; // Not collecting any notes now. 1271 1272 // This is a copy onto a pre-existing node, which may already have notes. 1273 // If both nodes have notes, do not overwrite any pre-existing notes. 1274 Node_Notes* source_notes = node_notes_at(source->_idx); 1275 if (source_notes == nullptr || source_notes->is_clear()) return false; 1276 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1277 if (dest_notes == nullptr || dest_notes->is_clear()) { 1278 return set_node_notes_at(dest->_idx, source_notes); 1279 } 1280 1281 Node_Notes merged_notes = (*source_notes); 1282 // The order of operations here ensures that dest notes will win... 1283 merged_notes.update_from(dest_notes); 1284 return set_node_notes_at(dest->_idx, &merged_notes); 1285 } 1286 1287 1288 //--------------------------allow_range_check_smearing------------------------- 1289 // Gating condition for coalescing similar range checks. 1290 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1291 // single covering check that is at least as strong as any of them. 1292 // If the optimization succeeds, the simplified (strengthened) range check 1293 // will always succeed. If it fails, we will deopt, and then give up 1294 // on the optimization. 1295 bool Compile::allow_range_check_smearing() const { 1296 // If this method has already thrown a range-check, 1297 // assume it was because we already tried range smearing 1298 // and it failed. 1299 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1300 return !already_trapped; 1301 } 1302 1303 1304 //------------------------------flatten_alias_type----------------------------- 1305 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1306 assert(do_aliasing(), "Aliasing should be enabled"); 1307 int offset = tj->offset(); 1308 TypePtr::PTR ptr = tj->ptr(); 1309 1310 // Known instance (scalarizable allocation) alias only with itself. 1311 bool is_known_inst = tj->isa_oopptr() != nullptr && 1312 tj->is_oopptr()->is_known_instance(); 1313 1314 // Process weird unsafe references. 1315 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1316 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops"); 1317 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1318 tj = TypeOopPtr::BOTTOM; 1319 ptr = tj->ptr(); 1320 offset = tj->offset(); 1321 } 1322 1323 // Array pointers need some flattening 1324 const TypeAryPtr* ta = tj->isa_aryptr(); 1325 if (ta && ta->is_stable()) { 1326 // Erase stability property for alias analysis. 1327 tj = ta = ta->cast_to_stable(false); 1328 } 1329 if( ta && is_known_inst ) { 1330 if ( offset != Type::OffsetBot && 1331 offset > arrayOopDesc::length_offset_in_bytes() ) { 1332 offset = Type::OffsetBot; // Flatten constant access into array body only 1333 tj = ta = ta-> 1334 remove_speculative()-> 1335 cast_to_ptr_type(ptr)-> 1336 with_offset(offset); 1337 } 1338 } else if (ta) { 1339 // For arrays indexed by constant indices, we flatten the alias 1340 // space to include all of the array body. Only the header, klass 1341 // and array length can be accessed un-aliased. 1342 if( offset != Type::OffsetBot ) { 1343 if( ta->const_oop() ) { // MethodData* or Method* 1344 offset = Type::OffsetBot; // Flatten constant access into array body 1345 tj = ta = ta-> 1346 remove_speculative()-> 1347 cast_to_ptr_type(ptr)-> 1348 cast_to_exactness(false)-> 1349 with_offset(offset); 1350 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1351 // range is OK as-is. 1352 tj = ta = TypeAryPtr::RANGE; 1353 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1354 tj = TypeInstPtr::KLASS; // all klass loads look alike 1355 ta = TypeAryPtr::RANGE; // generic ignored junk 1356 ptr = TypePtr::BotPTR; 1357 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1358 tj = TypeInstPtr::MARK; 1359 ta = TypeAryPtr::RANGE; // generic ignored junk 1360 ptr = TypePtr::BotPTR; 1361 } else { // Random constant offset into array body 1362 offset = Type::OffsetBot; // Flatten constant access into array body 1363 tj = ta = ta-> 1364 remove_speculative()-> 1365 cast_to_ptr_type(ptr)-> 1366 cast_to_exactness(false)-> 1367 with_offset(offset); 1368 } 1369 } 1370 // Arrays of fixed size alias with arrays of unknown size. 1371 if (ta->size() != TypeInt::POS) { 1372 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1373 tj = ta = ta-> 1374 remove_speculative()-> 1375 cast_to_ptr_type(ptr)-> 1376 with_ary(tary)-> 1377 cast_to_exactness(false); 1378 } 1379 // Arrays of known objects become arrays of unknown objects. 1380 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1381 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1382 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1383 } 1384 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1385 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1386 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1387 } 1388 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1389 // cannot be distinguished by bytecode alone. 1390 if (ta->elem() == TypeInt::BOOL) { 1391 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1392 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1393 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1394 } 1395 // During the 2nd round of IterGVN, NotNull castings are removed. 1396 // Make sure the Bottom and NotNull variants alias the same. 1397 // Also, make sure exact and non-exact variants alias the same. 1398 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) { 1399 tj = ta = ta-> 1400 remove_speculative()-> 1401 cast_to_ptr_type(TypePtr::BotPTR)-> 1402 cast_to_exactness(false)-> 1403 with_offset(offset); 1404 } 1405 } 1406 1407 // Oop pointers need some flattening 1408 const TypeInstPtr *to = tj->isa_instptr(); 1409 if (to && to != TypeOopPtr::BOTTOM) { 1410 ciInstanceKlass* ik = to->instance_klass(); 1411 if( ptr == TypePtr::Constant ) { 1412 if (ik != ciEnv::current()->Class_klass() || 1413 offset < ik->layout_helper_size_in_bytes()) { 1414 // No constant oop pointers (such as Strings); they alias with 1415 // unknown strings. 1416 assert(!is_known_inst, "not scalarizable allocation"); 1417 tj = to = to-> 1418 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1419 remove_speculative()-> 1420 cast_to_ptr_type(TypePtr::BotPTR)-> 1421 cast_to_exactness(false); 1422 } 1423 } else if( is_known_inst ) { 1424 tj = to; // Keep NotNull and klass_is_exact for instance type 1425 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1426 // During the 2nd round of IterGVN, NotNull castings are removed. 1427 // Make sure the Bottom and NotNull variants alias the same. 1428 // Also, make sure exact and non-exact variants alias the same. 1429 tj = to = to-> 1430 remove_speculative()-> 1431 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1432 cast_to_ptr_type(TypePtr::BotPTR)-> 1433 cast_to_exactness(false); 1434 } 1435 if (to->speculative() != nullptr) { 1436 tj = to = to->remove_speculative(); 1437 } 1438 // Canonicalize the holder of this field 1439 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1440 // First handle header references such as a LoadKlassNode, even if the 1441 // object's klass is unloaded at compile time (4965979). 1442 if (!is_known_inst) { // Do it only for non-instance types 1443 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset); 1444 } 1445 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) { 1446 // Static fields are in the space above the normal instance 1447 // fields in the java.lang.Class instance. 1448 if (ik != ciEnv::current()->Class_klass()) { 1449 to = nullptr; 1450 tj = TypeOopPtr::BOTTOM; 1451 offset = tj->offset(); 1452 } 1453 } else { 1454 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset); 1455 assert(offset < canonical_holder->layout_helper_size_in_bytes(), ""); 1456 if (!ik->equals(canonical_holder) || tj->offset() != offset) { 1457 if( is_known_inst ) { 1458 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id()); 1459 } else { 1460 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset); 1461 } 1462 } 1463 } 1464 } 1465 1466 // Klass pointers to object array klasses need some flattening 1467 const TypeKlassPtr *tk = tj->isa_klassptr(); 1468 if( tk ) { 1469 // If we are referencing a field within a Klass, we need 1470 // to assume the worst case of an Object. Both exact and 1471 // inexact types must flatten to the same alias class so 1472 // use NotNull as the PTR. 1473 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1474 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, 1475 env()->Object_klass(), 1476 offset); 1477 } 1478 1479 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) { 1480 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass()); 1481 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs 1482 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset); 1483 } else { 1484 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset); 1485 } 1486 } 1487 1488 // Check for precise loads from the primary supertype array and force them 1489 // to the supertype cache alias index. Check for generic array loads from 1490 // the primary supertype array and also force them to the supertype cache 1491 // alias index. Since the same load can reach both, we need to merge 1492 // these 2 disparate memories into the same alias class. Since the 1493 // primary supertype array is read-only, there's no chance of confusion 1494 // where we bypass an array load and an array store. 1495 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1496 if (offset == Type::OffsetBot || 1497 (offset >= primary_supers_offset && 1498 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1499 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1500 offset = in_bytes(Klass::secondary_super_cache_offset()); 1501 tj = tk = tk->with_offset(offset); 1502 } 1503 } 1504 1505 // Flatten all Raw pointers together. 1506 if (tj->base() == Type::RawPtr) 1507 tj = TypeRawPtr::BOTTOM; 1508 1509 if (tj->base() == Type::AnyPtr) 1510 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1511 1512 offset = tj->offset(); 1513 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1514 1515 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1516 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1517 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1518 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1519 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1520 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1521 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1522 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1523 assert( tj->ptr() != TypePtr::TopPTR && 1524 tj->ptr() != TypePtr::AnyNull && 1525 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1526 // assert( tj->ptr() != TypePtr::Constant || 1527 // tj->base() == Type::RawPtr || 1528 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1529 1530 return tj; 1531 } 1532 1533 void Compile::AliasType::Init(int i, const TypePtr* at) { 1534 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1535 _index = i; 1536 _adr_type = at; 1537 _field = nullptr; 1538 _element = nullptr; 1539 _is_rewritable = true; // default 1540 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr; 1541 if (atoop != nullptr && atoop->is_known_instance()) { 1542 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1543 _general_index = Compile::current()->get_alias_index(gt); 1544 } else { 1545 _general_index = 0; 1546 } 1547 } 1548 1549 BasicType Compile::AliasType::basic_type() const { 1550 if (element() != nullptr) { 1551 const Type* element = adr_type()->is_aryptr()->elem(); 1552 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1553 } if (field() != nullptr) { 1554 return field()->layout_type(); 1555 } else { 1556 return T_ILLEGAL; // unknown 1557 } 1558 } 1559 1560 //---------------------------------print_on------------------------------------ 1561 #ifndef PRODUCT 1562 void Compile::AliasType::print_on(outputStream* st) { 1563 if (index() < 10) 1564 st->print("@ <%d> ", index()); 1565 else st->print("@ <%d>", index()); 1566 st->print(is_rewritable() ? " " : " RO"); 1567 int offset = adr_type()->offset(); 1568 if (offset == Type::OffsetBot) 1569 st->print(" +any"); 1570 else st->print(" +%-3d", offset); 1571 st->print(" in "); 1572 adr_type()->dump_on(st); 1573 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1574 if (field() != nullptr && tjp) { 1575 if (tjp->is_instptr()->instance_klass() != field()->holder() || 1576 tjp->offset() != field()->offset_in_bytes()) { 1577 st->print(" != "); 1578 field()->print(); 1579 st->print(" ***"); 1580 } 1581 } 1582 } 1583 1584 void print_alias_types() { 1585 Compile* C = Compile::current(); 1586 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1587 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1588 C->alias_type(idx)->print_on(tty); 1589 tty->cr(); 1590 } 1591 } 1592 #endif 1593 1594 1595 //----------------------------probe_alias_cache-------------------------------- 1596 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1597 intptr_t key = (intptr_t) adr_type; 1598 key ^= key >> logAliasCacheSize; 1599 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1600 } 1601 1602 1603 //-----------------------------grow_alias_types-------------------------------- 1604 void Compile::grow_alias_types() { 1605 const int old_ats = _max_alias_types; // how many before? 1606 const int new_ats = old_ats; // how many more? 1607 const int grow_ats = old_ats+new_ats; // how many now? 1608 _max_alias_types = grow_ats; 1609 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1610 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1611 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1612 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1613 } 1614 1615 1616 //--------------------------------find_alias_type------------------------------ 1617 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1618 if (!do_aliasing()) { 1619 return alias_type(AliasIdxBot); 1620 } 1621 1622 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1623 if (ace->_adr_type == adr_type) { 1624 return alias_type(ace->_index); 1625 } 1626 1627 // Handle special cases. 1628 if (adr_type == nullptr) return alias_type(AliasIdxTop); 1629 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1630 1631 // Do it the slow way. 1632 const TypePtr* flat = flatten_alias_type(adr_type); 1633 1634 #ifdef ASSERT 1635 { 1636 ResourceMark rm; 1637 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1638 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1639 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1640 Type::str(adr_type)); 1641 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1642 const TypeOopPtr* foop = flat->is_oopptr(); 1643 // Scalarizable allocations have exact klass always. 1644 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1645 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1646 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1647 Type::str(foop), Type::str(xoop)); 1648 } 1649 } 1650 #endif 1651 1652 int idx = AliasIdxTop; 1653 for (int i = 0; i < num_alias_types(); i++) { 1654 if (alias_type(i)->adr_type() == flat) { 1655 idx = i; 1656 break; 1657 } 1658 } 1659 1660 if (idx == AliasIdxTop) { 1661 if (no_create) return nullptr; 1662 // Grow the array if necessary. 1663 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1664 // Add a new alias type. 1665 idx = _num_alias_types++; 1666 _alias_types[idx]->Init(idx, flat); 1667 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1668 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1669 if (flat->isa_instptr()) { 1670 if (flat->offset() == java_lang_Class::klass_offset() 1671 && flat->is_instptr()->instance_klass() == env()->Class_klass()) 1672 alias_type(idx)->set_rewritable(false); 1673 } 1674 if (flat->isa_aryptr()) { 1675 #ifdef ASSERT 1676 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1677 // (T_BYTE has the weakest alignment and size restrictions...) 1678 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1679 #endif 1680 if (flat->offset() == TypePtr::OffsetBot) { 1681 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1682 } 1683 } 1684 if (flat->isa_klassptr()) { 1685 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1686 alias_type(idx)->set_rewritable(false); 1687 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1688 alias_type(idx)->set_rewritable(false); 1689 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1690 alias_type(idx)->set_rewritable(false); 1691 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1692 alias_type(idx)->set_rewritable(false); 1693 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1694 alias_type(idx)->set_rewritable(false); 1695 } 1696 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1697 // but the base pointer type is not distinctive enough to identify 1698 // references into JavaThread.) 1699 1700 // Check for final fields. 1701 const TypeInstPtr* tinst = flat->isa_instptr(); 1702 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1703 ciField* field; 1704 if (tinst->const_oop() != nullptr && 1705 tinst->instance_klass() == ciEnv::current()->Class_klass() && 1706 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) { 1707 // static field 1708 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1709 field = k->get_field_by_offset(tinst->offset(), true); 1710 } else { 1711 ciInstanceKlass *k = tinst->instance_klass(); 1712 field = k->get_field_by_offset(tinst->offset(), false); 1713 } 1714 assert(field == nullptr || 1715 original_field == nullptr || 1716 (field->holder() == original_field->holder() && 1717 field->offset_in_bytes() == original_field->offset_in_bytes() && 1718 field->is_static() == original_field->is_static()), "wrong field?"); 1719 // Set field() and is_rewritable() attributes. 1720 if (field != nullptr) alias_type(idx)->set_field(field); 1721 } 1722 } 1723 1724 // Fill the cache for next time. 1725 ace->_adr_type = adr_type; 1726 ace->_index = idx; 1727 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1728 1729 // Might as well try to fill the cache for the flattened version, too. 1730 AliasCacheEntry* face = probe_alias_cache(flat); 1731 if (face->_adr_type == nullptr) { 1732 face->_adr_type = flat; 1733 face->_index = idx; 1734 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1735 } 1736 1737 return alias_type(idx); 1738 } 1739 1740 1741 Compile::AliasType* Compile::alias_type(ciField* field) { 1742 const TypeOopPtr* t; 1743 if (field->is_static()) 1744 t = TypeInstPtr::make(field->holder()->java_mirror()); 1745 else 1746 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1747 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1748 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1749 return atp; 1750 } 1751 1752 1753 //------------------------------have_alias_type-------------------------------- 1754 bool Compile::have_alias_type(const TypePtr* adr_type) { 1755 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1756 if (ace->_adr_type == adr_type) { 1757 return true; 1758 } 1759 1760 // Handle special cases. 1761 if (adr_type == nullptr) return true; 1762 if (adr_type == TypePtr::BOTTOM) return true; 1763 1764 return find_alias_type(adr_type, true, nullptr) != nullptr; 1765 } 1766 1767 //-----------------------------must_alias-------------------------------------- 1768 // True if all values of the given address type are in the given alias category. 1769 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1770 if (alias_idx == AliasIdxBot) return true; // the universal category 1771 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP 1772 if (alias_idx == AliasIdxTop) return false; // the empty category 1773 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1774 1775 // the only remaining possible overlap is identity 1776 int adr_idx = get_alias_index(adr_type); 1777 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1778 assert(adr_idx == alias_idx || 1779 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1780 && adr_type != TypeOopPtr::BOTTOM), 1781 "should not be testing for overlap with an unsafe pointer"); 1782 return adr_idx == alias_idx; 1783 } 1784 1785 //------------------------------can_alias-------------------------------------- 1786 // True if any values of the given address type are in the given alias category. 1787 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1788 if (alias_idx == AliasIdxTop) return false; // the empty category 1789 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP 1790 // Known instance doesn't alias with bottom memory 1791 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1792 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1793 1794 // the only remaining possible overlap is identity 1795 int adr_idx = get_alias_index(adr_type); 1796 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1797 return adr_idx == alias_idx; 1798 } 1799 1800 // Remove the opaque nodes that protect the Parse Predicates so that all unused 1801 // checks and uncommon_traps will be eliminated from the ideal graph. 1802 void Compile::cleanup_parse_predicates(PhaseIterGVN& igvn) const { 1803 if (parse_predicate_count() == 0) { 1804 return; 1805 } 1806 for (int i = parse_predicate_count(); i > 0; i--) { 1807 Node* n = parse_predicate_opaque1_node(i - 1); 1808 assert(n->Opcode() == Op_Opaque1, "must be"); 1809 igvn.replace_node(n, n->in(1)); 1810 } 1811 assert(parse_predicate_count() == 0, "should be clean!"); 1812 } 1813 1814 void Compile::record_for_post_loop_opts_igvn(Node* n) { 1815 if (!n->for_post_loop_opts_igvn()) { 1816 assert(!_for_post_loop_igvn.contains(n), "duplicate"); 1817 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1818 _for_post_loop_igvn.append(n); 1819 } 1820 } 1821 1822 void Compile::remove_from_post_loop_opts_igvn(Node* n) { 1823 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1824 _for_post_loop_igvn.remove(n); 1825 } 1826 1827 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) { 1828 // Verify that all previous optimizations produced a valid graph 1829 // at least to this point, even if no loop optimizations were done. 1830 PhaseIdealLoop::verify(igvn); 1831 1832 C->set_post_loop_opts_phase(); // no more loop opts allowed 1833 1834 assert(!C->major_progress(), "not cleared"); 1835 1836 if (_for_post_loop_igvn.length() > 0) { 1837 while (_for_post_loop_igvn.length() > 0) { 1838 Node* n = _for_post_loop_igvn.pop(); 1839 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1840 igvn._worklist.push(n); 1841 } 1842 igvn.optimize(); 1843 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed"); 1844 1845 // Sometimes IGVN sets major progress (e.g., when processing loop nodes). 1846 if (C->major_progress()) { 1847 C->clear_major_progress(); // ensure that major progress is now clear 1848 } 1849 } 1850 } 1851 1852 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) { 1853 if (OptimizeUnstableIf) { 1854 _unstable_if_traps.append(trap); 1855 } 1856 } 1857 1858 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) { 1859 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) { 1860 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1861 Node* n = trap->uncommon_trap(); 1862 if (!useful.member(n)) { 1863 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 1864 } 1865 } 1866 } 1867 1868 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead 1869 // or fold-compares case. Return true if succeed or not found. 1870 // 1871 // In rare cases, the found trap has been processed. It is too late to delete it. Return 1872 // false and ask fold-compares to yield. 1873 // 1874 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused 1875 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path 1876 // when deoptimization does happen. 1877 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) { 1878 for (int i = 0; i < _unstable_if_traps.length(); ++i) { 1879 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1880 if (trap->uncommon_trap() == unc) { 1881 if (yield && trap->modified()) { 1882 return false; 1883 } 1884 _unstable_if_traps.delete_at(i); 1885 break; 1886 } 1887 } 1888 return true; 1889 } 1890 1891 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path. 1892 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering. 1893 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) { 1894 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) { 1895 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1896 CallStaticJavaNode* unc = trap->uncommon_trap(); 1897 int next_bci = trap->next_bci(); 1898 bool modified = trap->modified(); 1899 1900 if (next_bci != -1 && !modified) { 1901 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!"); 1902 JVMState* jvms = unc->jvms(); 1903 ciMethod* method = jvms->method(); 1904 ciBytecodeStream iter(method); 1905 1906 iter.force_bci(jvms->bci()); 1907 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if"); 1908 Bytecodes::Code c = iter.cur_bc(); 1909 Node* lhs = nullptr; 1910 Node* rhs = nullptr; 1911 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) { 1912 lhs = unc->peek_operand(0); 1913 rhs = unc->peek_operand(1); 1914 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) { 1915 lhs = unc->peek_operand(0); 1916 } 1917 1918 ResourceMark rm; 1919 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci); 1920 assert(live_locals.is_valid(), "broken liveness info"); 1921 int len = (int)live_locals.size(); 1922 1923 for (int i = 0; i < len; i++) { 1924 Node* local = unc->local(jvms, i); 1925 // kill local using the liveness of next_bci. 1926 // give up when the local looks like an operand to secure reexecution. 1927 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) { 1928 uint idx = jvms->locoff() + i; 1929 #ifdef ASSERT 1930 if (Verbose) { 1931 tty->print("[unstable_if] kill local#%d: ", idx); 1932 local->dump(); 1933 tty->cr(); 1934 } 1935 #endif 1936 igvn.replace_input_of(unc, idx, top()); 1937 modified = true; 1938 } 1939 } 1940 } 1941 1942 // keep the mondified trap for late query 1943 if (modified) { 1944 trap->set_modified(); 1945 } else { 1946 _unstable_if_traps.delete_at(i); 1947 } 1948 } 1949 igvn.optimize(); 1950 } 1951 1952 // StringOpts and late inlining of string methods 1953 void Compile::inline_string_calls(bool parse_time) { 1954 { 1955 // remove useless nodes to make the usage analysis simpler 1956 ResourceMark rm; 1957 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 1958 } 1959 1960 { 1961 ResourceMark rm; 1962 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1963 PhaseStringOpts pso(initial_gvn()); 1964 print_method(PHASE_AFTER_STRINGOPTS, 3); 1965 } 1966 1967 // now inline anything that we skipped the first time around 1968 if (!parse_time) { 1969 _late_inlines_pos = _late_inlines.length(); 1970 } 1971 1972 while (_string_late_inlines.length() > 0) { 1973 CallGenerator* cg = _string_late_inlines.pop(); 1974 cg->do_late_inline(); 1975 if (failing()) return; 1976 } 1977 _string_late_inlines.trunc_to(0); 1978 } 1979 1980 // Late inlining of boxing methods 1981 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1982 if (_boxing_late_inlines.length() > 0) { 1983 assert(has_boxed_value(), "inconsistent"); 1984 1985 PhaseGVN* gvn = initial_gvn(); 1986 set_inlining_incrementally(true); 1987 1988 igvn_worklist()->ensure_empty(); // should be done with igvn 1989 1990 _late_inlines_pos = _late_inlines.length(); 1991 1992 while (_boxing_late_inlines.length() > 0) { 1993 CallGenerator* cg = _boxing_late_inlines.pop(); 1994 cg->do_late_inline(); 1995 if (failing()) return; 1996 } 1997 _boxing_late_inlines.trunc_to(0); 1998 1999 inline_incrementally_cleanup(igvn); 2000 2001 set_inlining_incrementally(false); 2002 } 2003 } 2004 2005 bool Compile::inline_incrementally_one() { 2006 assert(IncrementalInline, "incremental inlining should be on"); 2007 2008 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2009 2010 set_inlining_progress(false); 2011 set_do_cleanup(false); 2012 2013 for (int i = 0; i < _late_inlines.length(); i++) { 2014 _late_inlines_pos = i+1; 2015 CallGenerator* cg = _late_inlines.at(i); 2016 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline(); 2017 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call 2018 cg->do_late_inline(); 2019 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed"); 2020 if (failing()) { 2021 return false; 2022 } else if (inlining_progress()) { 2023 _late_inlines_pos = i+1; // restore the position in case new elements were inserted 2024 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node()); 2025 break; // process one call site at a time 2026 } 2027 } else { 2028 // Ignore late inline direct calls when inlining is not allowed. 2029 // They are left in the late inline list when node budget is exhausted until the list is fully drained. 2030 } 2031 } 2032 // Remove processed elements. 2033 _late_inlines.remove_till(_late_inlines_pos); 2034 _late_inlines_pos = 0; 2035 2036 assert(inlining_progress() || _late_inlines.length() == 0, "no progress"); 2037 2038 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2039 2040 set_inlining_progress(false); 2041 set_do_cleanup(false); 2042 2043 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption; 2044 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup; 2045 } 2046 2047 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2048 { 2049 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2050 ResourceMark rm; 2051 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2052 } 2053 { 2054 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2055 igvn.reset_from_gvn(initial_gvn()); 2056 igvn.optimize(); 2057 } 2058 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3); 2059 } 2060 2061 // Perform incremental inlining until bound on number of live nodes is reached 2062 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2063 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2064 2065 set_inlining_incrementally(true); 2066 uint low_live_nodes = 0; 2067 2068 while (_late_inlines.length() > 0) { 2069 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2070 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2071 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2072 // PhaseIdealLoop is expensive so we only try it once we are 2073 // out of live nodes and we only try it again if the previous 2074 // helped got the number of nodes down significantly 2075 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2076 if (failing()) return; 2077 low_live_nodes = live_nodes(); 2078 _major_progress = true; 2079 } 2080 2081 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2082 bool do_print_inlining = print_inlining() || print_intrinsics(); 2083 if (do_print_inlining || log() != nullptr) { 2084 // Print inlining message for candidates that we couldn't inline for lack of space. 2085 for (int i = 0; i < _late_inlines.length(); i++) { 2086 CallGenerator* cg = _late_inlines.at(i); 2087 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 2088 if (do_print_inlining) { 2089 cg->print_inlining_late(msg); 2090 } 2091 log_late_inline_failure(cg, msg); 2092 } 2093 } 2094 break; // finish 2095 } 2096 } 2097 2098 igvn_worklist()->ensure_empty(); // should be done with igvn 2099 2100 while (inline_incrementally_one()) { 2101 assert(!failing(), "inconsistent"); 2102 } 2103 if (failing()) return; 2104 2105 inline_incrementally_cleanup(igvn); 2106 2107 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 2108 2109 if (failing()) return; 2110 2111 if (_late_inlines.length() == 0) { 2112 break; // no more progress 2113 } 2114 } 2115 2116 igvn_worklist()->ensure_empty(); // should be done with igvn 2117 2118 if (_string_late_inlines.length() > 0) { 2119 assert(has_stringbuilder(), "inconsistent"); 2120 2121 inline_string_calls(false); 2122 2123 if (failing()) return; 2124 2125 inline_incrementally_cleanup(igvn); 2126 } 2127 2128 set_inlining_incrementally(false); 2129 } 2130 2131 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) { 2132 // "inlining_incrementally() == false" is used to signal that no inlining is allowed 2133 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details). 2134 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr" 2135 // as if "inlining_incrementally() == true" were set. 2136 assert(inlining_incrementally() == false, "not allowed"); 2137 assert(_modified_nodes == nullptr, "not allowed"); 2138 assert(_late_inlines.length() > 0, "sanity"); 2139 2140 while (_late_inlines.length() > 0) { 2141 igvn_worklist()->ensure_empty(); // should be done with igvn 2142 2143 while (inline_incrementally_one()) { 2144 assert(!failing(), "inconsistent"); 2145 } 2146 if (failing()) return; 2147 2148 inline_incrementally_cleanup(igvn); 2149 } 2150 } 2151 2152 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2153 if (_loop_opts_cnt > 0) { 2154 while (major_progress() && (_loop_opts_cnt > 0)) { 2155 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2156 PhaseIdealLoop::optimize(igvn, mode); 2157 _loop_opts_cnt--; 2158 if (failing()) return false; 2159 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2160 } 2161 } 2162 return true; 2163 } 2164 2165 // Remove edges from "root" to each SafePoint at a backward branch. 2166 // They were inserted during parsing (see add_safepoint()) to make 2167 // infinite loops without calls or exceptions visible to root, i.e., 2168 // useful. 2169 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2170 Node *r = root(); 2171 if (r != nullptr) { 2172 for (uint i = r->req(); i < r->len(); ++i) { 2173 Node *n = r->in(i); 2174 if (n != nullptr && n->is_SafePoint()) { 2175 r->rm_prec(i); 2176 if (n->outcnt() == 0) { 2177 igvn.remove_dead_node(n); 2178 } 2179 --i; 2180 } 2181 } 2182 // Parsing may have added top inputs to the root node (Path 2183 // leading to the Halt node proven dead). Make sure we get a 2184 // chance to clean them up. 2185 igvn._worklist.push(r); 2186 igvn.optimize(); 2187 } 2188 } 2189 2190 //------------------------------Optimize--------------------------------------- 2191 // Given a graph, optimize it. 2192 void Compile::Optimize() { 2193 TracePhase tp("optimizer", &timers[_t_optimizer]); 2194 2195 #ifndef PRODUCT 2196 if (env()->break_at_compile()) { 2197 BREAKPOINT; 2198 } 2199 2200 #endif 2201 2202 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2203 #ifdef ASSERT 2204 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2205 #endif 2206 2207 ResourceMark rm; 2208 2209 print_inlining_reinit(); 2210 2211 NOT_PRODUCT( verify_graph_edges(); ) 2212 2213 print_method(PHASE_AFTER_PARSING, 1); 2214 2215 { 2216 // Iterative Global Value Numbering, including ideal transforms 2217 // Initialize IterGVN with types and values from parse-time GVN 2218 PhaseIterGVN igvn(initial_gvn()); 2219 #ifdef ASSERT 2220 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2221 #endif 2222 { 2223 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2224 igvn.optimize(); 2225 } 2226 2227 if (failing()) return; 2228 2229 print_method(PHASE_ITER_GVN1, 2); 2230 2231 process_for_unstable_if_traps(igvn); 2232 2233 inline_incrementally(igvn); 2234 2235 print_method(PHASE_INCREMENTAL_INLINE, 2); 2236 2237 if (failing()) return; 2238 2239 if (eliminate_boxing()) { 2240 // Inline valueOf() methods now. 2241 inline_boxing_calls(igvn); 2242 2243 if (AlwaysIncrementalInline) { 2244 inline_incrementally(igvn); 2245 } 2246 2247 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2248 2249 if (failing()) return; 2250 } 2251 2252 // Remove the speculative part of types and clean up the graph from 2253 // the extra CastPP nodes whose only purpose is to carry them. Do 2254 // that early so that optimizations are not disrupted by the extra 2255 // CastPP nodes. 2256 remove_speculative_types(igvn); 2257 2258 // No more new expensive nodes will be added to the list from here 2259 // so keep only the actual candidates for optimizations. 2260 cleanup_expensive_nodes(igvn); 2261 2262 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2263 if (EnableVectorSupport && has_vbox_nodes()) { 2264 TracePhase tp("", &timers[_t_vector]); 2265 PhaseVector pv(igvn); 2266 pv.optimize_vector_boxes(); 2267 2268 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2269 } 2270 assert(!has_vbox_nodes(), "sanity"); 2271 2272 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2273 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2274 igvn_worklist()->ensure_empty(); // should be done with igvn 2275 { 2276 ResourceMark rm; 2277 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist()); 2278 } 2279 igvn.reset_from_gvn(initial_gvn()); 2280 igvn.optimize(); 2281 } 2282 2283 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop 2284 // safepoints 2285 remove_root_to_sfpts_edges(igvn); 2286 2287 // Perform escape analysis 2288 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) { 2289 if (has_loops()) { 2290 // Cleanup graph (remove dead nodes). 2291 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2292 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2293 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2294 if (failing()) return; 2295 } 2296 bool progress; 2297 do { 2298 ConnectionGraph::do_analysis(this, &igvn); 2299 2300 if (failing()) return; 2301 2302 int mcount = macro_count(); // Record number of allocations and locks before IGVN 2303 2304 // Optimize out fields loads from scalar replaceable allocations. 2305 igvn.optimize(); 2306 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2307 2308 if (failing()) return; 2309 2310 if (congraph() != nullptr && macro_count() > 0) { 2311 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2312 PhaseMacroExpand mexp(igvn); 2313 mexp.eliminate_macro_nodes(); 2314 igvn.set_delay_transform(false); 2315 2316 igvn.optimize(); 2317 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2318 2319 if (failing()) return; 2320 } 2321 progress = do_iterative_escape_analysis() && 2322 (macro_count() < mcount) && 2323 ConnectionGraph::has_candidates(this); 2324 // Try again if candidates exist and made progress 2325 // by removing some allocations and/or locks. 2326 } while (progress); 2327 } 2328 2329 // Loop transforms on the ideal graph. Range Check Elimination, 2330 // peeling, unrolling, etc. 2331 2332 // Set loop opts counter 2333 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2334 { 2335 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2336 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2337 _loop_opts_cnt--; 2338 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2339 if (failing()) return; 2340 } 2341 // Loop opts pass if partial peeling occurred in previous pass 2342 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2343 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2344 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2345 _loop_opts_cnt--; 2346 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2347 if (failing()) return; 2348 } 2349 // Loop opts pass for loop-unrolling before CCP 2350 if(major_progress() && (_loop_opts_cnt > 0)) { 2351 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2352 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2353 _loop_opts_cnt--; 2354 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2355 } 2356 if (!failing()) { 2357 // Verify that last round of loop opts produced a valid graph 2358 PhaseIdealLoop::verify(igvn); 2359 } 2360 } 2361 if (failing()) return; 2362 2363 // Conditional Constant Propagation; 2364 PhaseCCP ccp( &igvn ); 2365 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2366 { 2367 TracePhase tp("ccp", &timers[_t_ccp]); 2368 ccp.do_transform(); 2369 } 2370 print_method(PHASE_CCP1, 2); 2371 2372 assert( true, "Break here to ccp.dump_old2new_map()"); 2373 2374 // Iterative Global Value Numbering, including ideal transforms 2375 { 2376 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2377 igvn.reset_from_igvn(&ccp); 2378 igvn.optimize(); 2379 } 2380 print_method(PHASE_ITER_GVN2, 2); 2381 2382 if (failing()) return; 2383 2384 // Loop transforms on the ideal graph. Range Check Elimination, 2385 // peeling, unrolling, etc. 2386 if (!optimize_loops(igvn, LoopOptsDefault)) { 2387 return; 2388 } 2389 2390 if (failing()) return; 2391 2392 C->clear_major_progress(); // ensure that major progress is now clear 2393 2394 process_for_post_loop_opts_igvn(igvn); 2395 2396 #ifdef ASSERT 2397 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2398 #endif 2399 2400 { 2401 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2402 PhaseMacroExpand mex(igvn); 2403 if (mex.expand_macro_nodes()) { 2404 assert(failing(), "must bail out w/ explicit message"); 2405 return; 2406 } 2407 print_method(PHASE_MACRO_EXPANSION, 2); 2408 } 2409 2410 { 2411 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2412 if (bs->expand_barriers(this, igvn)) { 2413 assert(failing(), "must bail out w/ explicit message"); 2414 return; 2415 } 2416 print_method(PHASE_BARRIER_EXPANSION, 2); 2417 } 2418 2419 if (C->max_vector_size() > 0) { 2420 C->optimize_logic_cones(igvn); 2421 igvn.optimize(); 2422 } 2423 2424 DEBUG_ONLY( _modified_nodes = nullptr; ) 2425 2426 assert(igvn._worklist.size() == 0, "not empty"); 2427 2428 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty"); 2429 2430 if (_late_inlines.length() > 0) { 2431 // More opportunities to optimize virtual and MH calls. 2432 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option. 2433 process_late_inline_calls_no_inline(igvn); 2434 } 2435 } // (End scope of igvn; run destructor if necessary for asserts.) 2436 2437 check_no_dead_use(); 2438 2439 process_print_inlining(); 2440 2441 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have 2442 // to remove hashes to unlock nodes for modifications. 2443 C->node_hash()->clear(); 2444 2445 // A method with only infinite loops has no edges entering loops from root 2446 { 2447 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2448 if (final_graph_reshaping()) { 2449 assert(failing(), "must bail out w/ explicit message"); 2450 return; 2451 } 2452 } 2453 2454 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2455 DEBUG_ONLY(set_phase_optimize_finished();) 2456 } 2457 2458 #ifdef ASSERT 2459 void Compile::check_no_dead_use() const { 2460 ResourceMark rm; 2461 Unique_Node_List wq; 2462 wq.push(root()); 2463 for (uint i = 0; i < wq.size(); ++i) { 2464 Node* n = wq.at(i); 2465 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2466 Node* u = n->fast_out(j); 2467 if (u->outcnt() == 0 && !u->is_Con()) { 2468 u->dump(); 2469 fatal("no reachable node should have no use"); 2470 } 2471 wq.push(u); 2472 } 2473 } 2474 } 2475 #endif 2476 2477 void Compile::inline_vector_reboxing_calls() { 2478 if (C->_vector_reboxing_late_inlines.length() > 0) { 2479 _late_inlines_pos = C->_late_inlines.length(); 2480 while (_vector_reboxing_late_inlines.length() > 0) { 2481 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 2482 cg->do_late_inline(); 2483 if (failing()) return; 2484 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node()); 2485 } 2486 _vector_reboxing_late_inlines.trunc_to(0); 2487 } 2488 } 2489 2490 bool Compile::has_vbox_nodes() { 2491 if (C->_vector_reboxing_late_inlines.length() > 0) { 2492 return true; 2493 } 2494 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 2495 Node * n = C->macro_node(macro_idx); 2496 assert(n->is_macro(), "only macro nodes expected here"); 2497 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 2498 return true; 2499 } 2500 } 2501 return false; 2502 } 2503 2504 //---------------------------- Bitwise operation packing optimization --------------------------- 2505 2506 static bool is_vector_unary_bitwise_op(Node* n) { 2507 return n->Opcode() == Op_XorV && 2508 VectorNode::is_vector_bitwise_not_pattern(n); 2509 } 2510 2511 static bool is_vector_binary_bitwise_op(Node* n) { 2512 switch (n->Opcode()) { 2513 case Op_AndV: 2514 case Op_OrV: 2515 return true; 2516 2517 case Op_XorV: 2518 return !is_vector_unary_bitwise_op(n); 2519 2520 default: 2521 return false; 2522 } 2523 } 2524 2525 static bool is_vector_ternary_bitwise_op(Node* n) { 2526 return n->Opcode() == Op_MacroLogicV; 2527 } 2528 2529 static bool is_vector_bitwise_op(Node* n) { 2530 return is_vector_unary_bitwise_op(n) || 2531 is_vector_binary_bitwise_op(n) || 2532 is_vector_ternary_bitwise_op(n); 2533 } 2534 2535 static bool is_vector_bitwise_cone_root(Node* n) { 2536 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) { 2537 return false; 2538 } 2539 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2540 if (is_vector_bitwise_op(n->fast_out(i))) { 2541 return false; 2542 } 2543 } 2544 return true; 2545 } 2546 2547 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) { 2548 uint cnt = 0; 2549 if (is_vector_bitwise_op(n)) { 2550 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req(); 2551 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2552 for (uint i = 1; i < inp_cnt; i++) { 2553 Node* in = n->in(i); 2554 bool skip = VectorNode::is_all_ones_vector(in); 2555 if (!skip && !inputs.member(in)) { 2556 inputs.push(in); 2557 cnt++; 2558 } 2559 } 2560 assert(cnt <= 1, "not unary"); 2561 } else { 2562 uint last_req = inp_cnt; 2563 if (is_vector_ternary_bitwise_op(n)) { 2564 last_req = inp_cnt - 1; // skip last input 2565 } 2566 for (uint i = 1; i < last_req; i++) { 2567 Node* def = n->in(i); 2568 if (!inputs.member(def)) { 2569 inputs.push(def); 2570 cnt++; 2571 } 2572 } 2573 } 2574 } else { // not a bitwise operations 2575 if (!inputs.member(n)) { 2576 inputs.push(n); 2577 cnt++; 2578 } 2579 } 2580 return cnt; 2581 } 2582 2583 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 2584 Unique_Node_List useful_nodes; 2585 C->identify_useful_nodes(useful_nodes); 2586 2587 for (uint i = 0; i < useful_nodes.size(); i++) { 2588 Node* n = useful_nodes.at(i); 2589 if (is_vector_bitwise_cone_root(n)) { 2590 list.push(n); 2591 } 2592 } 2593 } 2594 2595 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 2596 const TypeVect* vt, 2597 Unique_Node_List& partition, 2598 Unique_Node_List& inputs) { 2599 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 2600 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 2601 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 2602 2603 Node* in1 = inputs.at(0); 2604 Node* in2 = inputs.at(1); 2605 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 2606 2607 uint func = compute_truth_table(partition, inputs); 2608 2609 Node* pn = partition.at(partition.size() - 1); 2610 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2611 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt)); 2612 } 2613 2614 static uint extract_bit(uint func, uint pos) { 2615 return (func & (1 << pos)) >> pos; 2616 } 2617 2618 // 2619 // A macro logic node represents a truth table. It has 4 inputs, 2620 // First three inputs corresponds to 3 columns of a truth table 2621 // and fourth input captures the logic function. 2622 // 2623 // eg. fn = (in1 AND in2) OR in3; 2624 // 2625 // MacroNode(in1,in2,in3,fn) 2626 // 2627 // ----------------- 2628 // in1 in2 in3 fn 2629 // ----------------- 2630 // 0 0 0 0 2631 // 0 0 1 1 2632 // 0 1 0 0 2633 // 0 1 1 1 2634 // 1 0 0 0 2635 // 1 0 1 1 2636 // 1 1 0 1 2637 // 1 1 1 1 2638 // 2639 2640 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 2641 int res = 0; 2642 for (int i = 0; i < 8; i++) { 2643 int bit1 = extract_bit(in1, i); 2644 int bit2 = extract_bit(in2, i); 2645 int bit3 = extract_bit(in3, i); 2646 2647 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 2648 int func_bit = extract_bit(func, func_bit_pos); 2649 2650 res |= func_bit << i; 2651 } 2652 return res; 2653 } 2654 2655 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 2656 assert(n != nullptr, ""); 2657 assert(eval_map.contains(n), "absent"); 2658 return *(eval_map.get(n)); 2659 } 2660 2661 static void eval_operands(Node* n, 2662 uint& func1, uint& func2, uint& func3, 2663 ResourceHashtable<Node*,uint>& eval_map) { 2664 assert(is_vector_bitwise_op(n), ""); 2665 2666 if (is_vector_unary_bitwise_op(n)) { 2667 Node* opnd = n->in(1); 2668 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 2669 opnd = n->in(2); 2670 } 2671 func1 = eval_operand(opnd, eval_map); 2672 } else if (is_vector_binary_bitwise_op(n)) { 2673 func1 = eval_operand(n->in(1), eval_map); 2674 func2 = eval_operand(n->in(2), eval_map); 2675 } else { 2676 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 2677 func1 = eval_operand(n->in(1), eval_map); 2678 func2 = eval_operand(n->in(2), eval_map); 2679 func3 = eval_operand(n->in(3), eval_map); 2680 } 2681 } 2682 2683 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 2684 assert(inputs.size() <= 3, "sanity"); 2685 ResourceMark rm; 2686 uint res = 0; 2687 ResourceHashtable<Node*,uint> eval_map; 2688 2689 // Populate precomputed functions for inputs. 2690 // Each input corresponds to one column of 3 input truth-table. 2691 uint input_funcs[] = { 0xAA, // (_, _, c) -> c 2692 0xCC, // (_, b, _) -> b 2693 0xF0 }; // (a, _, _) -> a 2694 for (uint i = 0; i < inputs.size(); i++) { 2695 eval_map.put(inputs.at(i), input_funcs[2-i]); 2696 } 2697 2698 for (uint i = 0; i < partition.size(); i++) { 2699 Node* n = partition.at(i); 2700 2701 uint func1 = 0, func2 = 0, func3 = 0; 2702 eval_operands(n, func1, func2, func3, eval_map); 2703 2704 switch (n->Opcode()) { 2705 case Op_OrV: 2706 assert(func3 == 0, "not binary"); 2707 res = func1 | func2; 2708 break; 2709 case Op_AndV: 2710 assert(func3 == 0, "not binary"); 2711 res = func1 & func2; 2712 break; 2713 case Op_XorV: 2714 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2715 assert(func2 == 0 && func3 == 0, "not unary"); 2716 res = (~func1) & 0xFF; 2717 } else { 2718 assert(func3 == 0, "not binary"); 2719 res = func1 ^ func2; 2720 } 2721 break; 2722 case Op_MacroLogicV: 2723 // Ordering of inputs may change during evaluation of sub-tree 2724 // containing MacroLogic node as a child node, thus a re-evaluation 2725 // makes sure that function is evaluated in context of current 2726 // inputs. 2727 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 2728 break; 2729 2730 default: assert(false, "not supported: %s", n->Name()); 2731 } 2732 assert(res <= 0xFF, "invalid"); 2733 eval_map.put(n, res); 2734 } 2735 return res; 2736 } 2737 2738 // Criteria under which nodes gets packed into a macro logic node:- 2739 // 1) Parent and both child nodes are all unmasked or masked with 2740 // same predicates. 2741 // 2) Masked parent can be packed with left child if it is predicated 2742 // and both have same predicates. 2743 // 3) Masked parent can be packed with right child if its un-predicated 2744 // or has matching predication condition. 2745 // 4) An unmasked parent can be packed with an unmasked child. 2746 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2747 assert(partition.size() == 0, "not empty"); 2748 assert(inputs.size() == 0, "not empty"); 2749 if (is_vector_ternary_bitwise_op(n)) { 2750 return false; 2751 } 2752 2753 bool is_unary_op = is_vector_unary_bitwise_op(n); 2754 if (is_unary_op) { 2755 assert(collect_unique_inputs(n, inputs) == 1, "not unary"); 2756 return false; // too few inputs 2757 } 2758 2759 bool pack_left_child = true; 2760 bool pack_right_child = true; 2761 2762 bool left_child_LOP = is_vector_bitwise_op(n->in(1)); 2763 bool right_child_LOP = is_vector_bitwise_op(n->in(2)); 2764 2765 int left_child_input_cnt = 0; 2766 int right_child_input_cnt = 0; 2767 2768 bool parent_is_predicated = n->is_predicated_vector(); 2769 bool left_child_predicated = n->in(1)->is_predicated_vector(); 2770 bool right_child_predicated = n->in(2)->is_predicated_vector(); 2771 2772 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr; 2773 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2774 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2775 2776 do { 2777 if (pack_left_child && left_child_LOP && 2778 ((!parent_is_predicated && !left_child_predicated) || 2779 ((parent_is_predicated && left_child_predicated && 2780 parent_pred == left_child_pred)))) { 2781 partition.push(n->in(1)); 2782 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs); 2783 } else { 2784 inputs.push(n->in(1)); 2785 left_child_input_cnt = 1; 2786 } 2787 2788 if (pack_right_child && right_child_LOP && 2789 (!right_child_predicated || 2790 (right_child_predicated && parent_is_predicated && 2791 parent_pred == right_child_pred))) { 2792 partition.push(n->in(2)); 2793 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs); 2794 } else { 2795 inputs.push(n->in(2)); 2796 right_child_input_cnt = 1; 2797 } 2798 2799 if (inputs.size() > 3) { 2800 assert(partition.size() > 0, ""); 2801 inputs.clear(); 2802 partition.clear(); 2803 if (left_child_input_cnt > right_child_input_cnt) { 2804 pack_left_child = false; 2805 } else { 2806 pack_right_child = false; 2807 } 2808 } else { 2809 break; 2810 } 2811 } while(true); 2812 2813 if(partition.size()) { 2814 partition.push(n); 2815 } 2816 2817 return (partition.size() == 2 || partition.size() == 3) && 2818 (inputs.size() == 2 || inputs.size() == 3); 2819 } 2820 2821 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 2822 assert(is_vector_bitwise_op(n), "not a root"); 2823 2824 visited.set(n->_idx); 2825 2826 // 1) Do a DFS walk over the logic cone. 2827 for (uint i = 1; i < n->req(); i++) { 2828 Node* in = n->in(i); 2829 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 2830 process_logic_cone_root(igvn, in, visited); 2831 } 2832 } 2833 2834 // 2) Bottom up traversal: Merge node[s] with 2835 // the parent to form macro logic node. 2836 Unique_Node_List partition; 2837 Unique_Node_List inputs; 2838 if (compute_logic_cone(n, partition, inputs)) { 2839 const TypeVect* vt = n->bottom_type()->is_vect(); 2840 Node* pn = partition.at(partition.size() - 1); 2841 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2842 if (mask == nullptr || 2843 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) { 2844 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 2845 VectorNode::trace_new_vector(macro_logic, "MacroLogic"); 2846 igvn.replace_node(n, macro_logic); 2847 } 2848 } 2849 } 2850 2851 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 2852 ResourceMark rm; 2853 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 2854 Unique_Node_List list; 2855 collect_logic_cone_roots(list); 2856 2857 while (list.size() > 0) { 2858 Node* n = list.pop(); 2859 const TypeVect* vt = n->bottom_type()->is_vect(); 2860 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 2861 if (supported) { 2862 VectorSet visited(comp_arena()); 2863 process_logic_cone_root(igvn, n, visited); 2864 } 2865 } 2866 } 2867 } 2868 2869 //------------------------------Code_Gen--------------------------------------- 2870 // Given a graph, generate code for it 2871 void Compile::Code_Gen() { 2872 if (failing()) { 2873 return; 2874 } 2875 2876 // Perform instruction selection. You might think we could reclaim Matcher 2877 // memory PDQ, but actually the Matcher is used in generating spill code. 2878 // Internals of the Matcher (including some VectorSets) must remain live 2879 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2880 // set a bit in reclaimed memory. 2881 2882 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2883 // nodes. Mapping is only valid at the root of each matched subtree. 2884 NOT_PRODUCT( verify_graph_edges(); ) 2885 2886 Matcher matcher; 2887 _matcher = &matcher; 2888 { 2889 TracePhase tp("matcher", &timers[_t_matcher]); 2890 matcher.match(); 2891 if (failing()) { 2892 return; 2893 } 2894 } 2895 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2896 // nodes. Mapping is only valid at the root of each matched subtree. 2897 NOT_PRODUCT( verify_graph_edges(); ) 2898 2899 // If you have too many nodes, or if matching has failed, bail out 2900 check_node_count(0, "out of nodes matching instructions"); 2901 if (failing()) { 2902 return; 2903 } 2904 2905 print_method(PHASE_MATCHING, 2); 2906 2907 // Build a proper-looking CFG 2908 PhaseCFG cfg(node_arena(), root(), matcher); 2909 _cfg = &cfg; 2910 { 2911 TracePhase tp("scheduler", &timers[_t_scheduler]); 2912 bool success = cfg.do_global_code_motion(); 2913 if (!success) { 2914 return; 2915 } 2916 2917 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2918 NOT_PRODUCT( verify_graph_edges(); ) 2919 cfg.verify(); 2920 } 2921 2922 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2923 _regalloc = ®alloc; 2924 { 2925 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2926 // Perform register allocation. After Chaitin, use-def chains are 2927 // no longer accurate (at spill code) and so must be ignored. 2928 // Node->LRG->reg mappings are still accurate. 2929 _regalloc->Register_Allocate(); 2930 2931 // Bail out if the allocator builds too many nodes 2932 if (failing()) { 2933 return; 2934 } 2935 } 2936 2937 // Prior to register allocation we kept empty basic blocks in case the 2938 // the allocator needed a place to spill. After register allocation we 2939 // are not adding any new instructions. If any basic block is empty, we 2940 // can now safely remove it. 2941 { 2942 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2943 cfg.remove_empty_blocks(); 2944 if (do_freq_based_layout()) { 2945 PhaseBlockLayout layout(cfg); 2946 } else { 2947 cfg.set_loop_alignment(); 2948 } 2949 cfg.fixup_flow(); 2950 cfg.remove_unreachable_blocks(); 2951 cfg.verify_dominator_tree(); 2952 } 2953 2954 // Apply peephole optimizations 2955 if( OptoPeephole ) { 2956 TracePhase tp("peephole", &timers[_t_peephole]); 2957 PhasePeephole peep( _regalloc, cfg); 2958 peep.do_transform(); 2959 } 2960 2961 // Do late expand if CPU requires this. 2962 if (Matcher::require_postalloc_expand) { 2963 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2964 cfg.postalloc_expand(_regalloc); 2965 } 2966 2967 // Convert Nodes to instruction bits in a buffer 2968 { 2969 TracePhase tp("output", &timers[_t_output]); 2970 PhaseOutput output; 2971 output.Output(); 2972 if (failing()) return; 2973 output.install(); 2974 } 2975 2976 print_method(PHASE_FINAL_CODE, 1); 2977 2978 // He's dead, Jim. 2979 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2980 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2981 } 2982 2983 //------------------------------Final_Reshape_Counts--------------------------- 2984 // This class defines counters to help identify when a method 2985 // may/must be executed using hardware with only 24-bit precision. 2986 struct Final_Reshape_Counts : public StackObj { 2987 int _call_count; // count non-inlined 'common' calls 2988 int _float_count; // count float ops requiring 24-bit precision 2989 int _double_count; // count double ops requiring more precision 2990 int _java_call_count; // count non-inlined 'java' calls 2991 int _inner_loop_count; // count loops which need alignment 2992 VectorSet _visited; // Visitation flags 2993 Node_List _tests; // Set of IfNodes & PCTableNodes 2994 2995 Final_Reshape_Counts() : 2996 _call_count(0), _float_count(0), _double_count(0), 2997 _java_call_count(0), _inner_loop_count(0) { } 2998 2999 void inc_call_count () { _call_count ++; } 3000 void inc_float_count () { _float_count ++; } 3001 void inc_double_count() { _double_count++; } 3002 void inc_java_call_count() { _java_call_count++; } 3003 void inc_inner_loop_count() { _inner_loop_count++; } 3004 3005 int get_call_count () const { return _call_count ; } 3006 int get_float_count () const { return _float_count ; } 3007 int get_double_count() const { return _double_count; } 3008 int get_java_call_count() const { return _java_call_count; } 3009 int get_inner_loop_count() const { return _inner_loop_count; } 3010 }; 3011 3012 // Eliminate trivially redundant StoreCMs and accumulate their 3013 // precedence edges. 3014 void Compile::eliminate_redundant_card_marks(Node* n) { 3015 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 3016 if (n->in(MemNode::Address)->outcnt() > 1) { 3017 // There are multiple users of the same address so it might be 3018 // possible to eliminate some of the StoreCMs 3019 Node* mem = n->in(MemNode::Memory); 3020 Node* adr = n->in(MemNode::Address); 3021 Node* val = n->in(MemNode::ValueIn); 3022 Node* prev = n; 3023 bool done = false; 3024 // Walk the chain of StoreCMs eliminating ones that match. As 3025 // long as it's a chain of single users then the optimization is 3026 // safe. Eliminating partially redundant StoreCMs would require 3027 // cloning copies down the other paths. 3028 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 3029 if (adr == mem->in(MemNode::Address) && 3030 val == mem->in(MemNode::ValueIn)) { 3031 // redundant StoreCM 3032 if (mem->req() > MemNode::OopStore) { 3033 // Hasn't been processed by this code yet. 3034 n->add_prec(mem->in(MemNode::OopStore)); 3035 } else { 3036 // Already converted to precedence edge 3037 for (uint i = mem->req(); i < mem->len(); i++) { 3038 // Accumulate any precedence edges 3039 if (mem->in(i) != nullptr) { 3040 n->add_prec(mem->in(i)); 3041 } 3042 } 3043 // Everything above this point has been processed. 3044 done = true; 3045 } 3046 // Eliminate the previous StoreCM 3047 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 3048 assert(mem->outcnt() == 0, "should be dead"); 3049 mem->disconnect_inputs(this); 3050 } else { 3051 prev = mem; 3052 } 3053 mem = prev->in(MemNode::Memory); 3054 } 3055 } 3056 } 3057 3058 //------------------------------final_graph_reshaping_impl---------------------- 3059 // Implement items 1-5 from final_graph_reshaping below. 3060 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3061 3062 if ( n->outcnt() == 0 ) return; // dead node 3063 uint nop = n->Opcode(); 3064 3065 // Check for 2-input instruction with "last use" on right input. 3066 // Swap to left input. Implements item (2). 3067 if( n->req() == 3 && // two-input instruction 3068 n->in(1)->outcnt() > 1 && // left use is NOT a last use 3069 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 3070 n->in(2)->outcnt() == 1 &&// right use IS a last use 3071 !n->in(2)->is_Con() ) { // right use is not a constant 3072 // Check for commutative opcode 3073 switch( nop ) { 3074 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 3075 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 3076 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 3077 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 3078 case Op_AndL: case Op_XorL: case Op_OrL: 3079 case Op_AndI: case Op_XorI: case Op_OrI: { 3080 // Move "last use" input to left by swapping inputs 3081 n->swap_edges(1, 2); 3082 break; 3083 } 3084 default: 3085 break; 3086 } 3087 } 3088 3089 #ifdef ASSERT 3090 if( n->is_Mem() ) { 3091 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 3092 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw || 3093 // oop will be recorded in oop map if load crosses safepoint 3094 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 3095 LoadNode::is_immutable_value(n->in(MemNode::Address))), 3096 "raw memory operations should have control edge"); 3097 } 3098 if (n->is_MemBar()) { 3099 MemBarNode* mb = n->as_MemBar(); 3100 if (mb->trailing_store() || mb->trailing_load_store()) { 3101 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 3102 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 3103 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 3104 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 3105 } else if (mb->leading()) { 3106 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 3107 } 3108 } 3109 #endif 3110 // Count FPU ops and common calls, implements item (3) 3111 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes); 3112 if (!gc_handled) { 3113 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes); 3114 } 3115 3116 // Collect CFG split points 3117 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3118 frc._tests.push(n); 3119 } 3120 } 3121 3122 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) { 3123 switch( nop ) { 3124 // Count all float operations that may use FPU 3125 case Op_AddF: 3126 case Op_SubF: 3127 case Op_MulF: 3128 case Op_DivF: 3129 case Op_NegF: 3130 case Op_ModF: 3131 case Op_ConvI2F: 3132 case Op_ConF: 3133 case Op_CmpF: 3134 case Op_CmpF3: 3135 case Op_StoreF: 3136 case Op_LoadF: 3137 // case Op_ConvL2F: // longs are split into 32-bit halves 3138 frc.inc_float_count(); 3139 break; 3140 3141 case Op_ConvF2D: 3142 case Op_ConvD2F: 3143 frc.inc_float_count(); 3144 frc.inc_double_count(); 3145 break; 3146 3147 // Count all double operations that may use FPU 3148 case Op_AddD: 3149 case Op_SubD: 3150 case Op_MulD: 3151 case Op_DivD: 3152 case Op_NegD: 3153 case Op_ModD: 3154 case Op_ConvI2D: 3155 case Op_ConvD2I: 3156 // case Op_ConvL2D: // handled by leaf call 3157 // case Op_ConvD2L: // handled by leaf call 3158 case Op_ConD: 3159 case Op_CmpD: 3160 case Op_CmpD3: 3161 case Op_StoreD: 3162 case Op_LoadD: 3163 case Op_LoadD_unaligned: 3164 frc.inc_double_count(); 3165 break; 3166 case Op_Opaque1: // Remove Opaque Nodes before matching 3167 case Op_Opaque3: 3168 n->subsume_by(n->in(1), this); 3169 break; 3170 case Op_CallStaticJava: 3171 case Op_CallJava: 3172 case Op_CallDynamicJava: 3173 frc.inc_java_call_count(); // Count java call site; 3174 case Op_CallRuntime: 3175 case Op_CallLeaf: 3176 case Op_CallLeafVector: 3177 case Op_CallLeafNoFP: { 3178 assert (n->is_Call(), ""); 3179 CallNode *call = n->as_Call(); 3180 // Count call sites where the FP mode bit would have to be flipped. 3181 // Do not count uncommon runtime calls: 3182 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3183 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3184 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3185 frc.inc_call_count(); // Count the call site 3186 } else { // See if uncommon argument is shared 3187 Node *n = call->in(TypeFunc::Parms); 3188 int nop = n->Opcode(); 3189 // Clone shared simple arguments to uncommon calls, item (1). 3190 if (n->outcnt() > 1 && 3191 !n->is_Proj() && 3192 nop != Op_CreateEx && 3193 nop != Op_CheckCastPP && 3194 nop != Op_DecodeN && 3195 nop != Op_DecodeNKlass && 3196 !n->is_Mem() && 3197 !n->is_Phi()) { 3198 Node *x = n->clone(); 3199 call->set_req(TypeFunc::Parms, x); 3200 } 3201 } 3202 break; 3203 } 3204 3205 case Op_StoreCM: 3206 { 3207 // Convert OopStore dependence into precedence edge 3208 Node* prec = n->in(MemNode::OopStore); 3209 n->del_req(MemNode::OopStore); 3210 n->add_prec(prec); 3211 eliminate_redundant_card_marks(n); 3212 } 3213 3214 // fall through 3215 3216 case Op_StoreB: 3217 case Op_StoreC: 3218 case Op_StoreI: 3219 case Op_StoreL: 3220 case Op_CompareAndSwapB: 3221 case Op_CompareAndSwapS: 3222 case Op_CompareAndSwapI: 3223 case Op_CompareAndSwapL: 3224 case Op_CompareAndSwapP: 3225 case Op_CompareAndSwapN: 3226 case Op_WeakCompareAndSwapB: 3227 case Op_WeakCompareAndSwapS: 3228 case Op_WeakCompareAndSwapI: 3229 case Op_WeakCompareAndSwapL: 3230 case Op_WeakCompareAndSwapP: 3231 case Op_WeakCompareAndSwapN: 3232 case Op_CompareAndExchangeB: 3233 case Op_CompareAndExchangeS: 3234 case Op_CompareAndExchangeI: 3235 case Op_CompareAndExchangeL: 3236 case Op_CompareAndExchangeP: 3237 case Op_CompareAndExchangeN: 3238 case Op_GetAndAddS: 3239 case Op_GetAndAddB: 3240 case Op_GetAndAddI: 3241 case Op_GetAndAddL: 3242 case Op_GetAndSetS: 3243 case Op_GetAndSetB: 3244 case Op_GetAndSetI: 3245 case Op_GetAndSetL: 3246 case Op_GetAndSetP: 3247 case Op_GetAndSetN: 3248 case Op_StoreP: 3249 case Op_StoreN: 3250 case Op_StoreNKlass: 3251 case Op_LoadB: 3252 case Op_LoadUB: 3253 case Op_LoadUS: 3254 case Op_LoadI: 3255 case Op_LoadKlass: 3256 case Op_LoadNKlass: 3257 case Op_LoadL: 3258 case Op_LoadL_unaligned: 3259 case Op_LoadP: 3260 case Op_LoadN: 3261 case Op_LoadRange: 3262 case Op_LoadS: 3263 break; 3264 3265 case Op_AddP: { // Assert sane base pointers 3266 Node *addp = n->in(AddPNode::Address); 3267 assert( !addp->is_AddP() || 3268 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3269 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3270 "Base pointers must match (addp %u)", addp->_idx ); 3271 #ifdef _LP64 3272 if ((UseCompressedOops || UseCompressedClassPointers) && 3273 addp->Opcode() == Op_ConP && 3274 addp == n->in(AddPNode::Base) && 3275 n->in(AddPNode::Offset)->is_Con()) { 3276 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3277 // on the platform and on the compressed oops mode. 3278 // Use addressing with narrow klass to load with offset on x86. 3279 // Some platforms can use the constant pool to load ConP. 3280 // Do this transformation here since IGVN will convert ConN back to ConP. 3281 const Type* t = addp->bottom_type(); 3282 bool is_oop = t->isa_oopptr() != nullptr; 3283 bool is_klass = t->isa_klassptr() != nullptr; 3284 3285 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3286 (is_klass && Matcher::const_klass_prefer_decode())) { 3287 Node* nn = nullptr; 3288 3289 int op = is_oop ? Op_ConN : Op_ConNKlass; 3290 3291 // Look for existing ConN node of the same exact type. 3292 Node* r = root(); 3293 uint cnt = r->outcnt(); 3294 for (uint i = 0; i < cnt; i++) { 3295 Node* m = r->raw_out(i); 3296 if (m!= nullptr && m->Opcode() == op && 3297 m->bottom_type()->make_ptr() == t) { 3298 nn = m; 3299 break; 3300 } 3301 } 3302 if (nn != nullptr) { 3303 // Decode a narrow oop to match address 3304 // [R12 + narrow_oop_reg<<3 + offset] 3305 if (is_oop) { 3306 nn = new DecodeNNode(nn, t); 3307 } else { 3308 nn = new DecodeNKlassNode(nn, t); 3309 } 3310 // Check for succeeding AddP which uses the same Base. 3311 // Otherwise we will run into the assertion above when visiting that guy. 3312 for (uint i = 0; i < n->outcnt(); ++i) { 3313 Node *out_i = n->raw_out(i); 3314 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3315 out_i->set_req(AddPNode::Base, nn); 3316 #ifdef ASSERT 3317 for (uint j = 0; j < out_i->outcnt(); ++j) { 3318 Node *out_j = out_i->raw_out(j); 3319 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3320 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3321 } 3322 #endif 3323 } 3324 } 3325 n->set_req(AddPNode::Base, nn); 3326 n->set_req(AddPNode::Address, nn); 3327 if (addp->outcnt() == 0) { 3328 addp->disconnect_inputs(this); 3329 } 3330 } 3331 } 3332 } 3333 #endif 3334 break; 3335 } 3336 3337 case Op_CastPP: { 3338 // Remove CastPP nodes to gain more freedom during scheduling but 3339 // keep the dependency they encode as control or precedence edges 3340 // (if control is set already) on memory operations. Some CastPP 3341 // nodes don't have a control (don't carry a dependency): skip 3342 // those. 3343 if (n->in(0) != nullptr) { 3344 ResourceMark rm; 3345 Unique_Node_List wq; 3346 wq.push(n); 3347 for (uint next = 0; next < wq.size(); ++next) { 3348 Node *m = wq.at(next); 3349 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3350 Node* use = m->fast_out(i); 3351 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3352 use->ensure_control_or_add_prec(n->in(0)); 3353 } else { 3354 switch(use->Opcode()) { 3355 case Op_AddP: 3356 case Op_DecodeN: 3357 case Op_DecodeNKlass: 3358 case Op_CheckCastPP: 3359 case Op_CastPP: 3360 wq.push(use); 3361 break; 3362 } 3363 } 3364 } 3365 } 3366 } 3367 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3368 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3369 Node* in1 = n->in(1); 3370 const Type* t = n->bottom_type(); 3371 Node* new_in1 = in1->clone(); 3372 new_in1->as_DecodeN()->set_type(t); 3373 3374 if (!Matcher::narrow_oop_use_complex_address()) { 3375 // 3376 // x86, ARM and friends can handle 2 adds in addressing mode 3377 // and Matcher can fold a DecodeN node into address by using 3378 // a narrow oop directly and do implicit null check in address: 3379 // 3380 // [R12 + narrow_oop_reg<<3 + offset] 3381 // NullCheck narrow_oop_reg 3382 // 3383 // On other platforms (Sparc) we have to keep new DecodeN node and 3384 // use it to do implicit null check in address: 3385 // 3386 // decode_not_null narrow_oop_reg, base_reg 3387 // [base_reg + offset] 3388 // NullCheck base_reg 3389 // 3390 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3391 // to keep the information to which null check the new DecodeN node 3392 // corresponds to use it as value in implicit_null_check(). 3393 // 3394 new_in1->set_req(0, n->in(0)); 3395 } 3396 3397 n->subsume_by(new_in1, this); 3398 if (in1->outcnt() == 0) { 3399 in1->disconnect_inputs(this); 3400 } 3401 } else { 3402 n->subsume_by(n->in(1), this); 3403 if (n->outcnt() == 0) { 3404 n->disconnect_inputs(this); 3405 } 3406 } 3407 break; 3408 } 3409 #ifdef _LP64 3410 case Op_CmpP: 3411 // Do this transformation here to preserve CmpPNode::sub() and 3412 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3413 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3414 Node* in1 = n->in(1); 3415 Node* in2 = n->in(2); 3416 if (!in1->is_DecodeNarrowPtr()) { 3417 in2 = in1; 3418 in1 = n->in(2); 3419 } 3420 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3421 3422 Node* new_in2 = nullptr; 3423 if (in2->is_DecodeNarrowPtr()) { 3424 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3425 new_in2 = in2->in(1); 3426 } else if (in2->Opcode() == Op_ConP) { 3427 const Type* t = in2->bottom_type(); 3428 if (t == TypePtr::NULL_PTR) { 3429 assert(in1->is_DecodeN(), "compare klass to null?"); 3430 // Don't convert CmpP null check into CmpN if compressed 3431 // oops implicit null check is not generated. 3432 // This will allow to generate normal oop implicit null check. 3433 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3434 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3435 // 3436 // This transformation together with CastPP transformation above 3437 // will generated code for implicit null checks for compressed oops. 3438 // 3439 // The original code after Optimize() 3440 // 3441 // LoadN memory, narrow_oop_reg 3442 // decode narrow_oop_reg, base_reg 3443 // CmpP base_reg, nullptr 3444 // CastPP base_reg // NotNull 3445 // Load [base_reg + offset], val_reg 3446 // 3447 // after these transformations will be 3448 // 3449 // LoadN memory, narrow_oop_reg 3450 // CmpN narrow_oop_reg, nullptr 3451 // decode_not_null narrow_oop_reg, base_reg 3452 // Load [base_reg + offset], val_reg 3453 // 3454 // and the uncommon path (== nullptr) will use narrow_oop_reg directly 3455 // since narrow oops can be used in debug info now (see the code in 3456 // final_graph_reshaping_walk()). 3457 // 3458 // At the end the code will be matched to 3459 // on x86: 3460 // 3461 // Load_narrow_oop memory, narrow_oop_reg 3462 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3463 // NullCheck narrow_oop_reg 3464 // 3465 // and on sparc: 3466 // 3467 // Load_narrow_oop memory, narrow_oop_reg 3468 // decode_not_null narrow_oop_reg, base_reg 3469 // Load [base_reg + offset], val_reg 3470 // NullCheck base_reg 3471 // 3472 } else if (t->isa_oopptr()) { 3473 new_in2 = ConNode::make(t->make_narrowoop()); 3474 } else if (t->isa_klassptr()) { 3475 new_in2 = ConNode::make(t->make_narrowklass()); 3476 } 3477 } 3478 if (new_in2 != nullptr) { 3479 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3480 n->subsume_by(cmpN, this); 3481 if (in1->outcnt() == 0) { 3482 in1->disconnect_inputs(this); 3483 } 3484 if (in2->outcnt() == 0) { 3485 in2->disconnect_inputs(this); 3486 } 3487 } 3488 } 3489 break; 3490 3491 case Op_DecodeN: 3492 case Op_DecodeNKlass: 3493 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3494 // DecodeN could be pinned when it can't be fold into 3495 // an address expression, see the code for Op_CastPP above. 3496 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3497 break; 3498 3499 case Op_EncodeP: 3500 case Op_EncodePKlass: { 3501 Node* in1 = n->in(1); 3502 if (in1->is_DecodeNarrowPtr()) { 3503 n->subsume_by(in1->in(1), this); 3504 } else if (in1->Opcode() == Op_ConP) { 3505 const Type* t = in1->bottom_type(); 3506 if (t == TypePtr::NULL_PTR) { 3507 assert(t->isa_oopptr(), "null klass?"); 3508 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3509 } else if (t->isa_oopptr()) { 3510 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3511 } else if (t->isa_klassptr()) { 3512 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3513 } 3514 } 3515 if (in1->outcnt() == 0) { 3516 in1->disconnect_inputs(this); 3517 } 3518 break; 3519 } 3520 3521 case Op_Proj: { 3522 if (OptimizeStringConcat || IncrementalInline) { 3523 ProjNode* proj = n->as_Proj(); 3524 if (proj->_is_io_use) { 3525 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, ""); 3526 // Separate projections were used for the exception path which 3527 // are normally removed by a late inline. If it wasn't inlined 3528 // then they will hang around and should just be replaced with 3529 // the original one. Merge them. 3530 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/); 3531 if (non_io_proj != nullptr) { 3532 proj->subsume_by(non_io_proj , this); 3533 } 3534 } 3535 } 3536 break; 3537 } 3538 3539 case Op_Phi: 3540 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3541 // The EncodeP optimization may create Phi with the same edges 3542 // for all paths. It is not handled well by Register Allocator. 3543 Node* unique_in = n->in(1); 3544 assert(unique_in != nullptr, ""); 3545 uint cnt = n->req(); 3546 for (uint i = 2; i < cnt; i++) { 3547 Node* m = n->in(i); 3548 assert(m != nullptr, ""); 3549 if (unique_in != m) 3550 unique_in = nullptr; 3551 } 3552 if (unique_in != nullptr) { 3553 n->subsume_by(unique_in, this); 3554 } 3555 } 3556 break; 3557 3558 #endif 3559 3560 #ifdef ASSERT 3561 case Op_CastII: 3562 // Verify that all range check dependent CastII nodes were removed. 3563 if (n->isa_CastII()->has_range_check()) { 3564 n->dump(3); 3565 assert(false, "Range check dependent CastII node was not removed"); 3566 } 3567 break; 3568 #endif 3569 3570 case Op_ModI: 3571 if (UseDivMod) { 3572 // Check if a%b and a/b both exist 3573 Node* d = n->find_similar(Op_DivI); 3574 if (d) { 3575 // Replace them with a fused divmod if supported 3576 if (Matcher::has_match_rule(Op_DivModI)) { 3577 DivModINode* divmod = DivModINode::make(n); 3578 d->subsume_by(divmod->div_proj(), this); 3579 n->subsume_by(divmod->mod_proj(), this); 3580 } else { 3581 // replace a%b with a-((a/b)*b) 3582 Node* mult = new MulINode(d, d->in(2)); 3583 Node* sub = new SubINode(d->in(1), mult); 3584 n->subsume_by(sub, this); 3585 } 3586 } 3587 } 3588 break; 3589 3590 case Op_ModL: 3591 if (UseDivMod) { 3592 // Check if a%b and a/b both exist 3593 Node* d = n->find_similar(Op_DivL); 3594 if (d) { 3595 // Replace them with a fused divmod if supported 3596 if (Matcher::has_match_rule(Op_DivModL)) { 3597 DivModLNode* divmod = DivModLNode::make(n); 3598 d->subsume_by(divmod->div_proj(), this); 3599 n->subsume_by(divmod->mod_proj(), this); 3600 } else { 3601 // replace a%b with a-((a/b)*b) 3602 Node* mult = new MulLNode(d, d->in(2)); 3603 Node* sub = new SubLNode(d->in(1), mult); 3604 n->subsume_by(sub, this); 3605 } 3606 } 3607 } 3608 break; 3609 3610 case Op_UModI: 3611 if (UseDivMod) { 3612 // Check if a%b and a/b both exist 3613 Node* d = n->find_similar(Op_UDivI); 3614 if (d) { 3615 // Replace them with a fused unsigned divmod if supported 3616 if (Matcher::has_match_rule(Op_UDivModI)) { 3617 UDivModINode* divmod = UDivModINode::make(n); 3618 d->subsume_by(divmod->div_proj(), this); 3619 n->subsume_by(divmod->mod_proj(), this); 3620 } else { 3621 // replace a%b with a-((a/b)*b) 3622 Node* mult = new MulINode(d, d->in(2)); 3623 Node* sub = new SubINode(d->in(1), mult); 3624 n->subsume_by(sub, this); 3625 } 3626 } 3627 } 3628 break; 3629 3630 case Op_UModL: 3631 if (UseDivMod) { 3632 // Check if a%b and a/b both exist 3633 Node* d = n->find_similar(Op_UDivL); 3634 if (d) { 3635 // Replace them with a fused unsigned divmod if supported 3636 if (Matcher::has_match_rule(Op_UDivModL)) { 3637 UDivModLNode* divmod = UDivModLNode::make(n); 3638 d->subsume_by(divmod->div_proj(), this); 3639 n->subsume_by(divmod->mod_proj(), this); 3640 } else { 3641 // replace a%b with a-((a/b)*b) 3642 Node* mult = new MulLNode(d, d->in(2)); 3643 Node* sub = new SubLNode(d->in(1), mult); 3644 n->subsume_by(sub, this); 3645 } 3646 } 3647 } 3648 break; 3649 3650 case Op_LoadVector: 3651 case Op_StoreVector: 3652 case Op_LoadVectorGather: 3653 case Op_StoreVectorScatter: 3654 case Op_LoadVectorGatherMasked: 3655 case Op_StoreVectorScatterMasked: 3656 case Op_VectorCmpMasked: 3657 case Op_VectorMaskGen: 3658 case Op_LoadVectorMasked: 3659 case Op_StoreVectorMasked: 3660 break; 3661 3662 case Op_AddReductionVI: 3663 case Op_AddReductionVL: 3664 case Op_AddReductionVF: 3665 case Op_AddReductionVD: 3666 case Op_MulReductionVI: 3667 case Op_MulReductionVL: 3668 case Op_MulReductionVF: 3669 case Op_MulReductionVD: 3670 case Op_MinReductionV: 3671 case Op_MaxReductionV: 3672 case Op_AndReductionV: 3673 case Op_OrReductionV: 3674 case Op_XorReductionV: 3675 break; 3676 3677 case Op_PackB: 3678 case Op_PackS: 3679 case Op_PackI: 3680 case Op_PackF: 3681 case Op_PackL: 3682 case Op_PackD: 3683 if (n->req()-1 > 2) { 3684 // Replace many operand PackNodes with a binary tree for matching 3685 PackNode* p = (PackNode*) n; 3686 Node* btp = p->binary_tree_pack(1, n->req()); 3687 n->subsume_by(btp, this); 3688 } 3689 break; 3690 case Op_Loop: 3691 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop"); 3692 case Op_CountedLoop: 3693 case Op_LongCountedLoop: 3694 case Op_OuterStripMinedLoop: 3695 if (n->as_Loop()->is_inner_loop()) { 3696 frc.inc_inner_loop_count(); 3697 } 3698 n->as_Loop()->verify_strip_mined(0); 3699 break; 3700 case Op_LShiftI: 3701 case Op_RShiftI: 3702 case Op_URShiftI: 3703 case Op_LShiftL: 3704 case Op_RShiftL: 3705 case Op_URShiftL: 3706 if (Matcher::need_masked_shift_count) { 3707 // The cpu's shift instructions don't restrict the count to the 3708 // lower 5/6 bits. We need to do the masking ourselves. 3709 Node* in2 = n->in(2); 3710 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3711 const TypeInt* t = in2->find_int_type(); 3712 if (t != nullptr && t->is_con()) { 3713 juint shift = t->get_con(); 3714 if (shift > mask) { // Unsigned cmp 3715 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3716 } 3717 } else { 3718 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) { 3719 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3720 n->set_req(2, shift); 3721 } 3722 } 3723 if (in2->outcnt() == 0) { // Remove dead node 3724 in2->disconnect_inputs(this); 3725 } 3726 } 3727 break; 3728 case Op_MemBarStoreStore: 3729 case Op_MemBarRelease: 3730 // Break the link with AllocateNode: it is no longer useful and 3731 // confuses register allocation. 3732 if (n->req() > MemBarNode::Precedent) { 3733 n->set_req(MemBarNode::Precedent, top()); 3734 } 3735 break; 3736 case Op_MemBarAcquire: { 3737 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3738 // At parse time, the trailing MemBarAcquire for a volatile load 3739 // is created with an edge to the load. After optimizations, 3740 // that input may be a chain of Phis. If those phis have no 3741 // other use, then the MemBarAcquire keeps them alive and 3742 // register allocation can be confused. 3743 dead_nodes.push(n->in(MemBarNode::Precedent)); 3744 n->set_req(MemBarNode::Precedent, top()); 3745 } 3746 break; 3747 } 3748 case Op_Blackhole: 3749 break; 3750 case Op_RangeCheck: { 3751 RangeCheckNode* rc = n->as_RangeCheck(); 3752 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3753 n->subsume_by(iff, this); 3754 frc._tests.push(iff); 3755 break; 3756 } 3757 case Op_ConvI2L: { 3758 if (!Matcher::convi2l_type_required) { 3759 // Code generation on some platforms doesn't need accurate 3760 // ConvI2L types. Widening the type can help remove redundant 3761 // address computations. 3762 n->as_Type()->set_type(TypeLong::INT); 3763 ResourceMark rm; 3764 Unique_Node_List wq; 3765 wq.push(n); 3766 for (uint next = 0; next < wq.size(); next++) { 3767 Node *m = wq.at(next); 3768 3769 for(;;) { 3770 // Loop over all nodes with identical inputs edges as m 3771 Node* k = m->find_similar(m->Opcode()); 3772 if (k == nullptr) { 3773 break; 3774 } 3775 // Push their uses so we get a chance to remove node made 3776 // redundant 3777 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3778 Node* u = k->fast_out(i); 3779 if (u->Opcode() == Op_LShiftL || 3780 u->Opcode() == Op_AddL || 3781 u->Opcode() == Op_SubL || 3782 u->Opcode() == Op_AddP) { 3783 wq.push(u); 3784 } 3785 } 3786 // Replace all nodes with identical edges as m with m 3787 k->subsume_by(m, this); 3788 } 3789 } 3790 } 3791 break; 3792 } 3793 case Op_CmpUL: { 3794 if (!Matcher::has_match_rule(Op_CmpUL)) { 3795 // No support for unsigned long comparisons 3796 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3797 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3798 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3799 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3800 Node* andl = new AndLNode(orl, remove_sign_mask); 3801 Node* cmp = new CmpLNode(andl, n->in(2)); 3802 n->subsume_by(cmp, this); 3803 } 3804 break; 3805 } 3806 default: 3807 assert(!n->is_Call(), ""); 3808 assert(!n->is_Mem(), ""); 3809 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3810 break; 3811 } 3812 } 3813 3814 //------------------------------final_graph_reshaping_walk--------------------- 3815 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3816 // requires that the walk visits a node's inputs before visiting the node. 3817 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3818 Unique_Node_List sfpt; 3819 3820 frc._visited.set(root->_idx); // first, mark node as visited 3821 uint cnt = root->req(); 3822 Node *n = root; 3823 uint i = 0; 3824 while (true) { 3825 if (i < cnt) { 3826 // Place all non-visited non-null inputs onto stack 3827 Node* m = n->in(i); 3828 ++i; 3829 if (m != nullptr && !frc._visited.test_set(m->_idx)) { 3830 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) { 3831 // compute worst case interpreter size in case of a deoptimization 3832 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3833 3834 sfpt.push(m); 3835 } 3836 cnt = m->req(); 3837 nstack.push(n, i); // put on stack parent and next input's index 3838 n = m; 3839 i = 0; 3840 } 3841 } else { 3842 // Now do post-visit work 3843 final_graph_reshaping_impl(n, frc, dead_nodes); 3844 if (nstack.is_empty()) 3845 break; // finished 3846 n = nstack.node(); // Get node from stack 3847 cnt = n->req(); 3848 i = nstack.index(); 3849 nstack.pop(); // Shift to the next node on stack 3850 } 3851 } 3852 3853 // Skip next transformation if compressed oops are not used. 3854 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3855 (!UseCompressedOops && !UseCompressedClassPointers)) 3856 return; 3857 3858 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3859 // It could be done for an uncommon traps or any safepoints/calls 3860 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3861 while (sfpt.size() > 0) { 3862 n = sfpt.pop(); 3863 JVMState *jvms = n->as_SafePoint()->jvms(); 3864 assert(jvms != nullptr, "sanity"); 3865 int start = jvms->debug_start(); 3866 int end = n->req(); 3867 bool is_uncommon = (n->is_CallStaticJava() && 3868 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3869 for (int j = start; j < end; j++) { 3870 Node* in = n->in(j); 3871 if (in->is_DecodeNarrowPtr()) { 3872 bool safe_to_skip = true; 3873 if (!is_uncommon ) { 3874 // Is it safe to skip? 3875 for (uint i = 0; i < in->outcnt(); i++) { 3876 Node* u = in->raw_out(i); 3877 if (!u->is_SafePoint() || 3878 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3879 safe_to_skip = false; 3880 } 3881 } 3882 } 3883 if (safe_to_skip) { 3884 n->set_req(j, in->in(1)); 3885 } 3886 if (in->outcnt() == 0) { 3887 in->disconnect_inputs(this); 3888 } 3889 } 3890 } 3891 } 3892 } 3893 3894 //------------------------------final_graph_reshaping-------------------------- 3895 // Final Graph Reshaping. 3896 // 3897 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3898 // and not commoned up and forced early. Must come after regular 3899 // optimizations to avoid GVN undoing the cloning. Clone constant 3900 // inputs to Loop Phis; these will be split by the allocator anyways. 3901 // Remove Opaque nodes. 3902 // (2) Move last-uses by commutative operations to the left input to encourage 3903 // Intel update-in-place two-address operations and better register usage 3904 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3905 // calls canonicalizing them back. 3906 // (3) Count the number of double-precision FP ops, single-precision FP ops 3907 // and call sites. On Intel, we can get correct rounding either by 3908 // forcing singles to memory (requires extra stores and loads after each 3909 // FP bytecode) or we can set a rounding mode bit (requires setting and 3910 // clearing the mode bit around call sites). The mode bit is only used 3911 // if the relative frequency of single FP ops to calls is low enough. 3912 // This is a key transform for SPEC mpeg_audio. 3913 // (4) Detect infinite loops; blobs of code reachable from above but not 3914 // below. Several of the Code_Gen algorithms fail on such code shapes, 3915 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3916 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3917 // Detection is by looking for IfNodes where only 1 projection is 3918 // reachable from below or CatchNodes missing some targets. 3919 // (5) Assert for insane oop offsets in debug mode. 3920 3921 bool Compile::final_graph_reshaping() { 3922 // an infinite loop may have been eliminated by the optimizer, 3923 // in which case the graph will be empty. 3924 if (root()->req() == 1) { 3925 // Do not compile method that is only a trivial infinite loop, 3926 // since the content of the loop may have been eliminated. 3927 record_method_not_compilable("trivial infinite loop"); 3928 return true; 3929 } 3930 3931 // Expensive nodes have their control input set to prevent the GVN 3932 // from freely commoning them. There's no GVN beyond this point so 3933 // no need to keep the control input. We want the expensive nodes to 3934 // be freely moved to the least frequent code path by gcm. 3935 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3936 for (int i = 0; i < expensive_count(); i++) { 3937 _expensive_nodes.at(i)->set_req(0, nullptr); 3938 } 3939 3940 Final_Reshape_Counts frc; 3941 3942 // Visit everybody reachable! 3943 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3944 Node_Stack nstack(live_nodes() >> 1); 3945 Unique_Node_List dead_nodes; 3946 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes); 3947 3948 // Check for unreachable (from below) code (i.e., infinite loops). 3949 for( uint i = 0; i < frc._tests.size(); i++ ) { 3950 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3951 // Get number of CFG targets. 3952 // Note that PCTables include exception targets after calls. 3953 uint required_outcnt = n->required_outcnt(); 3954 if (n->outcnt() != required_outcnt) { 3955 // Check for a few special cases. Rethrow Nodes never take the 3956 // 'fall-thru' path, so expected kids is 1 less. 3957 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3958 if (n->in(0)->in(0)->is_Call()) { 3959 CallNode* call = n->in(0)->in(0)->as_Call(); 3960 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3961 required_outcnt--; // Rethrow always has 1 less kid 3962 } else if (call->req() > TypeFunc::Parms && 3963 call->is_CallDynamicJava()) { 3964 // Check for null receiver. In such case, the optimizer has 3965 // detected that the virtual call will always result in a null 3966 // pointer exception. The fall-through projection of this CatchNode 3967 // will not be populated. 3968 Node* arg0 = call->in(TypeFunc::Parms); 3969 if (arg0->is_Type() && 3970 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3971 required_outcnt--; 3972 } 3973 } else if (call->entry_point() == OptoRuntime::new_array_Java() || 3974 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 3975 // Check for illegal array length. In such case, the optimizer has 3976 // detected that the allocation attempt will always result in an 3977 // exception. There is no fall-through projection of this CatchNode . 3978 assert(call->is_CallStaticJava(), "static call expected"); 3979 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 3980 uint valid_length_test_input = call->req() - 1; 3981 Node* valid_length_test = call->in(valid_length_test_input); 3982 call->del_req(valid_length_test_input); 3983 if (valid_length_test->find_int_con(1) == 0) { 3984 required_outcnt--; 3985 } 3986 dead_nodes.push(valid_length_test); 3987 assert(n->outcnt() == required_outcnt, "malformed control flow"); 3988 continue; 3989 } 3990 } 3991 } 3992 3993 // Recheck with a better notion of 'required_outcnt' 3994 if (n->outcnt() != required_outcnt) { 3995 record_method_not_compilable("malformed control flow"); 3996 return true; // Not all targets reachable! 3997 } 3998 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) { 3999 CallNode* call = n->in(0)->in(0)->as_Call(); 4000 if (call->entry_point() == OptoRuntime::new_array_Java() || 4001 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4002 assert(call->is_CallStaticJava(), "static call expected"); 4003 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4004 uint valid_length_test_input = call->req() - 1; 4005 dead_nodes.push(call->in(valid_length_test_input)); 4006 call->del_req(valid_length_test_input); // valid length test useless now 4007 } 4008 } 4009 // Check that I actually visited all kids. Unreached kids 4010 // must be infinite loops. 4011 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 4012 if (!frc._visited.test(n->fast_out(j)->_idx)) { 4013 record_method_not_compilable("infinite loop"); 4014 return true; // Found unvisited kid; must be unreach 4015 } 4016 4017 // Here so verification code in final_graph_reshaping_walk() 4018 // always see an OuterStripMinedLoopEnd 4019 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) { 4020 IfNode* init_iff = n->as_If(); 4021 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 4022 n->subsume_by(iff, this); 4023 } 4024 } 4025 4026 while (dead_nodes.size() > 0) { 4027 Node* m = dead_nodes.pop(); 4028 if (m->outcnt() == 0 && m != top()) { 4029 for (uint j = 0; j < m->req(); j++) { 4030 Node* in = m->in(j); 4031 if (in != nullptr) { 4032 dead_nodes.push(in); 4033 } 4034 } 4035 m->disconnect_inputs(this); 4036 } 4037 } 4038 4039 #ifdef IA32 4040 // If original bytecodes contained a mixture of floats and doubles 4041 // check if the optimizer has made it homogeneous, item (3). 4042 if (UseSSE == 0 && 4043 frc.get_float_count() > 32 && 4044 frc.get_double_count() == 0 && 4045 (10 * frc.get_call_count() < frc.get_float_count()) ) { 4046 set_24_bit_selection_and_mode(false, true); 4047 } 4048 #endif // IA32 4049 4050 set_java_calls(frc.get_java_call_count()); 4051 set_inner_loops(frc.get_inner_loop_count()); 4052 4053 // No infinite loops, no reason to bail out. 4054 return false; 4055 } 4056 4057 //-----------------------------too_many_traps---------------------------------- 4058 // Report if there are too many traps at the current method and bci. 4059 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 4060 bool Compile::too_many_traps(ciMethod* method, 4061 int bci, 4062 Deoptimization::DeoptReason reason) { 4063 ciMethodData* md = method->method_data(); 4064 if (md->is_empty()) { 4065 // Assume the trap has not occurred, or that it occurred only 4066 // because of a transient condition during start-up in the interpreter. 4067 return false; 4068 } 4069 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4070 if (md->has_trap_at(bci, m, reason) != 0) { 4071 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 4072 // Also, if there are multiple reasons, or if there is no per-BCI record, 4073 // assume the worst. 4074 if (log()) 4075 log()->elem("observe trap='%s' count='%d'", 4076 Deoptimization::trap_reason_name(reason), 4077 md->trap_count(reason)); 4078 return true; 4079 } else { 4080 // Ignore method/bci and see if there have been too many globally. 4081 return too_many_traps(reason, md); 4082 } 4083 } 4084 4085 // Less-accurate variant which does not require a method and bci. 4086 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 4087 ciMethodData* logmd) { 4088 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 4089 // Too many traps globally. 4090 // Note that we use cumulative trap_count, not just md->trap_count. 4091 if (log()) { 4092 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason); 4093 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 4094 Deoptimization::trap_reason_name(reason), 4095 mcount, trap_count(reason)); 4096 } 4097 return true; 4098 } else { 4099 // The coast is clear. 4100 return false; 4101 } 4102 } 4103 4104 //--------------------------too_many_recompiles-------------------------------- 4105 // Report if there are too many recompiles at the current method and bci. 4106 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 4107 // Is not eager to return true, since this will cause the compiler to use 4108 // Action_none for a trap point, to avoid too many recompilations. 4109 bool Compile::too_many_recompiles(ciMethod* method, 4110 int bci, 4111 Deoptimization::DeoptReason reason) { 4112 ciMethodData* md = method->method_data(); 4113 if (md->is_empty()) { 4114 // Assume the trap has not occurred, or that it occurred only 4115 // because of a transient condition during start-up in the interpreter. 4116 return false; 4117 } 4118 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 4119 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 4120 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 4121 Deoptimization::DeoptReason per_bc_reason 4122 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 4123 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4124 if ((per_bc_reason == Deoptimization::Reason_none 4125 || md->has_trap_at(bci, m, reason) != 0) 4126 // The trap frequency measure we care about is the recompile count: 4127 && md->trap_recompiled_at(bci, m) 4128 && md->overflow_recompile_count() >= bc_cutoff) { 4129 // Do not emit a trap here if it has already caused recompilations. 4130 // Also, if there are multiple reasons, or if there is no per-BCI record, 4131 // assume the worst. 4132 if (log()) 4133 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 4134 Deoptimization::trap_reason_name(reason), 4135 md->trap_count(reason), 4136 md->overflow_recompile_count()); 4137 return true; 4138 } else if (trap_count(reason) != 0 4139 && decompile_count() >= m_cutoff) { 4140 // Too many recompiles globally, and we have seen this sort of trap. 4141 // Use cumulative decompile_count, not just md->decompile_count. 4142 if (log()) 4143 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 4144 Deoptimization::trap_reason_name(reason), 4145 md->trap_count(reason), trap_count(reason), 4146 md->decompile_count(), decompile_count()); 4147 return true; 4148 } else { 4149 // The coast is clear. 4150 return false; 4151 } 4152 } 4153 4154 // Compute when not to trap. Used by matching trap based nodes and 4155 // NullCheck optimization. 4156 void Compile::set_allowed_deopt_reasons() { 4157 _allowed_reasons = 0; 4158 if (is_method_compilation()) { 4159 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 4160 assert(rs < BitsPerInt, "recode bit map"); 4161 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 4162 _allowed_reasons |= nth_bit(rs); 4163 } 4164 } 4165 } 4166 } 4167 4168 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 4169 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 4170 } 4171 4172 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 4173 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 4174 } 4175 4176 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 4177 if (holder->is_initialized()) { 4178 return false; 4179 } 4180 if (holder->is_being_initialized()) { 4181 if (accessing_method->holder() == holder) { 4182 // Access inside a class. The barrier can be elided when access happens in <clinit>, 4183 // <init>, or a static method. In all those cases, there was an initialization 4184 // barrier on the holder klass passed. 4185 if (accessing_method->is_static_initializer() || 4186 accessing_method->is_object_initializer() || 4187 accessing_method->is_static()) { 4188 return false; 4189 } 4190 } else if (accessing_method->holder()->is_subclass_of(holder)) { 4191 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 4192 // In case of <init> or a static method, the barrier is on the subclass is not enough: 4193 // child class can become fully initialized while its parent class is still being initialized. 4194 if (accessing_method->is_static_initializer()) { 4195 return false; 4196 } 4197 } 4198 ciMethod* root = method(); // the root method of compilation 4199 if (root != accessing_method) { 4200 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 4201 } 4202 } 4203 return true; 4204 } 4205 4206 #ifndef PRODUCT 4207 //------------------------------verify_bidirectional_edges--------------------- 4208 // For each input edge to a node (ie - for each Use-Def edge), verify that 4209 // there is a corresponding Def-Use edge. 4210 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) { 4211 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc 4212 uint stack_size = live_nodes() >> 4; 4213 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize)); 4214 nstack.push(_root); 4215 4216 while (nstack.size() > 0) { 4217 Node* n = nstack.pop(); 4218 if (visited.member(n)) { 4219 continue; 4220 } 4221 visited.push(n); 4222 4223 // Walk over all input edges, checking for correspondence 4224 uint length = n->len(); 4225 for (uint i = 0; i < length; i++) { 4226 Node* in = n->in(i); 4227 if (in != nullptr && !visited.member(in)) { 4228 nstack.push(in); // Put it on stack 4229 } 4230 if (in != nullptr && !in->is_top()) { 4231 // Count instances of `next` 4232 int cnt = 0; 4233 for (uint idx = 0; idx < in->_outcnt; idx++) { 4234 if (in->_out[idx] == n) { 4235 cnt++; 4236 } 4237 } 4238 assert(cnt > 0, "Failed to find Def-Use edge."); 4239 // Check for duplicate edges 4240 // walk the input array downcounting the input edges to n 4241 for (uint j = 0; j < length; j++) { 4242 if (n->in(j) == in) { 4243 cnt--; 4244 } 4245 } 4246 assert(cnt == 0, "Mismatched edge count."); 4247 } else if (in == nullptr) { 4248 assert(i == 0 || i >= n->req() || 4249 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() || 4250 (n->is_Unlock() && i == (n->req() - 1)) || 4251 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion 4252 "only region, phi, arraycopy, unlock or membar nodes have null data edges"); 4253 } else { 4254 assert(in->is_top(), "sanity"); 4255 // Nothing to check. 4256 } 4257 } 4258 } 4259 } 4260 4261 //------------------------------verify_graph_edges--------------------------- 4262 // Walk the Graph and verify that there is a one-to-one correspondence 4263 // between Use-Def edges and Def-Use edges in the graph. 4264 void Compile::verify_graph_edges(bool no_dead_code) { 4265 if (VerifyGraphEdges) { 4266 Unique_Node_List visited; 4267 4268 // Call graph walk to check edges 4269 verify_bidirectional_edges(visited); 4270 if (no_dead_code) { 4271 // Now make sure that no visited node is used by an unvisited node. 4272 bool dead_nodes = false; 4273 Unique_Node_List checked; 4274 while (visited.size() > 0) { 4275 Node* n = visited.pop(); 4276 checked.push(n); 4277 for (uint i = 0; i < n->outcnt(); i++) { 4278 Node* use = n->raw_out(i); 4279 if (checked.member(use)) continue; // already checked 4280 if (visited.member(use)) continue; // already in the graph 4281 if (use->is_Con()) continue; // a dead ConNode is OK 4282 // At this point, we have found a dead node which is DU-reachable. 4283 if (!dead_nodes) { 4284 tty->print_cr("*** Dead nodes reachable via DU edges:"); 4285 dead_nodes = true; 4286 } 4287 use->dump(2); 4288 tty->print_cr("---"); 4289 checked.push(use); // No repeats; pretend it is now checked. 4290 } 4291 } 4292 assert(!dead_nodes, "using nodes must be reachable from root"); 4293 } 4294 } 4295 } 4296 #endif 4297 4298 // The Compile object keeps track of failure reasons separately from the ciEnv. 4299 // This is required because there is not quite a 1-1 relation between the 4300 // ciEnv and its compilation task and the Compile object. Note that one 4301 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4302 // to backtrack and retry without subsuming loads. Other than this backtracking 4303 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4304 // by the logic in C2Compiler. 4305 void Compile::record_failure(const char* reason) { 4306 if (log() != nullptr) { 4307 log()->elem("failure reason='%s' phase='compile'", reason); 4308 } 4309 if (_failure_reason == nullptr) { 4310 // Record the first failure reason. 4311 _failure_reason = reason; 4312 } 4313 4314 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4315 C->print_method(PHASE_FAILURE, 1); 4316 } 4317 _root = nullptr; // flush the graph, too 4318 } 4319 4320 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 4321 : TraceTime(name, accumulator, CITime, CITimeVerbose), 4322 _phase_name(name), _dolog(CITimeVerbose) 4323 { 4324 if (_dolog) { 4325 C = Compile::current(); 4326 _log = C->log(); 4327 } else { 4328 C = nullptr; 4329 _log = nullptr; 4330 } 4331 if (_log != nullptr) { 4332 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4333 _log->stamp(); 4334 _log->end_head(); 4335 } 4336 } 4337 4338 Compile::TracePhase::~TracePhase() { 4339 4340 C = Compile::current(); 4341 if (_dolog) { 4342 _log = C->log(); 4343 } else { 4344 _log = nullptr; 4345 } 4346 4347 #ifdef ASSERT 4348 if (PrintIdealNodeCount) { 4349 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4350 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 4351 } 4352 4353 if (VerifyIdealNodeCount) { 4354 Compile::current()->print_missing_nodes(); 4355 } 4356 #endif 4357 4358 if (_log != nullptr) { 4359 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4360 } 4361 } 4362 4363 //----------------------------static_subtype_check----------------------------- 4364 // Shortcut important common cases when superklass is exact: 4365 // (0) superklass is java.lang.Object (can occur in reflective code) 4366 // (1) subklass is already limited to a subtype of superklass => always ok 4367 // (2) subklass does not overlap with superklass => always fail 4368 // (3) superklass has NO subtypes and we can check with a simple compare. 4369 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) { 4370 if (skip) { 4371 return SSC_full_test; // Let caller generate the general case. 4372 } 4373 4374 if (subk->is_java_subtype_of(superk)) { 4375 return SSC_always_true; // (0) and (1) this test cannot fail 4376 } 4377 4378 if (!subk->maybe_java_subtype_of(superk)) { 4379 return SSC_always_false; // (2) true path dead; no dynamic test needed 4380 } 4381 4382 const Type* superelem = superk; 4383 if (superk->isa_aryklassptr()) { 4384 int ignored; 4385 superelem = superk->is_aryklassptr()->base_element_type(ignored); 4386 } 4387 4388 if (superelem->isa_instklassptr()) { 4389 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass(); 4390 if (!ik->has_subklass()) { 4391 if (!ik->is_final()) { 4392 // Add a dependency if there is a chance of a later subclass. 4393 dependencies()->assert_leaf_type(ik); 4394 } 4395 if (!superk->maybe_java_subtype_of(subk)) { 4396 return SSC_always_false; 4397 } 4398 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4399 } 4400 } else { 4401 // A primitive array type has no subtypes. 4402 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4403 } 4404 4405 return SSC_full_test; 4406 } 4407 4408 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4409 #ifdef _LP64 4410 // The scaled index operand to AddP must be a clean 64-bit value. 4411 // Java allows a 32-bit int to be incremented to a negative 4412 // value, which appears in a 64-bit register as a large 4413 // positive number. Using that large positive number as an 4414 // operand in pointer arithmetic has bad consequences. 4415 // On the other hand, 32-bit overflow is rare, and the possibility 4416 // can often be excluded, if we annotate the ConvI2L node with 4417 // a type assertion that its value is known to be a small positive 4418 // number. (The prior range check has ensured this.) 4419 // This assertion is used by ConvI2LNode::Ideal. 4420 int index_max = max_jint - 1; // array size is max_jint, index is one less 4421 if (sizetype != nullptr) index_max = sizetype->_hi - 1; 4422 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4423 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4424 #endif 4425 return idx; 4426 } 4427 4428 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4429 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 4430 if (ctrl != nullptr) { 4431 // Express control dependency by a CastII node with a narrow type. 4432 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */); 4433 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4434 // node from floating above the range check during loop optimizations. Otherwise, the 4435 // ConvI2L node may be eliminated independently of the range check, causing the data path 4436 // to become TOP while the control path is still there (although it's unreachable). 4437 value->set_req(0, ctrl); 4438 value = phase->transform(value); 4439 } 4440 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4441 return phase->transform(new ConvI2LNode(value, ltype)); 4442 } 4443 4444 // The message about the current inlining is accumulated in 4445 // _print_inlining_stream and transferred into the _print_inlining_list 4446 // once we know whether inlining succeeds or not. For regular 4447 // inlining, messages are appended to the buffer pointed by 4448 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4449 // a new buffer is added after _print_inlining_idx in the list. This 4450 // way we can update the inlining message for late inlining call site 4451 // when the inlining is attempted again. 4452 void Compile::print_inlining_init() { 4453 if (print_inlining() || print_intrinsics()) { 4454 // print_inlining_init is actually called several times. 4455 print_inlining_reset(); 4456 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer()); 4457 } 4458 } 4459 4460 void Compile::print_inlining_reinit() { 4461 if (print_inlining() || print_intrinsics()) { 4462 print_inlining_reset(); 4463 } 4464 } 4465 4466 void Compile::print_inlining_reset() { 4467 _print_inlining_stream->reset(); 4468 } 4469 4470 void Compile::print_inlining_commit() { 4471 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4472 // Transfer the message from _print_inlining_stream to the current 4473 // _print_inlining_list buffer and clear _print_inlining_stream. 4474 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 4475 print_inlining_reset(); 4476 } 4477 4478 void Compile::print_inlining_push() { 4479 // Add new buffer to the _print_inlining_list at current position 4480 _print_inlining_idx++; 4481 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer()); 4482 } 4483 4484 Compile::PrintInliningBuffer* Compile::print_inlining_current() { 4485 return _print_inlining_list->at(_print_inlining_idx); 4486 } 4487 4488 void Compile::print_inlining_update(CallGenerator* cg) { 4489 if (print_inlining() || print_intrinsics()) { 4490 if (cg->is_late_inline()) { 4491 if (print_inlining_current()->cg() != cg && 4492 (print_inlining_current()->cg() != nullptr || 4493 print_inlining_current()->ss()->size() != 0)) { 4494 print_inlining_push(); 4495 } 4496 print_inlining_commit(); 4497 print_inlining_current()->set_cg(cg); 4498 } else { 4499 if (print_inlining_current()->cg() != nullptr) { 4500 print_inlining_push(); 4501 } 4502 print_inlining_commit(); 4503 } 4504 } 4505 } 4506 4507 void Compile::print_inlining_move_to(CallGenerator* cg) { 4508 // We resume inlining at a late inlining call site. Locate the 4509 // corresponding inlining buffer so that we can update it. 4510 if (print_inlining() || print_intrinsics()) { 4511 for (int i = 0; i < _print_inlining_list->length(); i++) { 4512 if (_print_inlining_list->at(i)->cg() == cg) { 4513 _print_inlining_idx = i; 4514 return; 4515 } 4516 } 4517 ShouldNotReachHere(); 4518 } 4519 } 4520 4521 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4522 if (print_inlining() || print_intrinsics()) { 4523 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4524 assert(print_inlining_current()->cg() == cg, "wrong entry"); 4525 // replace message with new message 4526 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer()); 4527 print_inlining_commit(); 4528 print_inlining_current()->set_cg(cg); 4529 } 4530 } 4531 4532 void Compile::print_inlining_assert_ready() { 4533 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data"); 4534 } 4535 4536 void Compile::process_print_inlining() { 4537 assert(_late_inlines.length() == 0, "not drained yet"); 4538 if (print_inlining() || print_intrinsics()) { 4539 ResourceMark rm; 4540 stringStream ss; 4541 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once."); 4542 for (int i = 0; i < _print_inlining_list->length(); i++) { 4543 PrintInliningBuffer* pib = _print_inlining_list->at(i); 4544 ss.print("%s", pib->ss()->freeze()); 4545 delete pib; 4546 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr)); 4547 } 4548 // Reset _print_inlining_list, it only contains destructed objects. 4549 // It is on the arena, so it will be freed when the arena is reset. 4550 _print_inlining_list = nullptr; 4551 // _print_inlining_stream won't be used anymore, either. 4552 print_inlining_reset(); 4553 size_t end = ss.size(); 4554 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4555 strncpy(_print_inlining_output, ss.freeze(), end+1); 4556 _print_inlining_output[end] = 0; 4557 } 4558 } 4559 4560 void Compile::dump_print_inlining() { 4561 if (_print_inlining_output != nullptr) { 4562 tty->print_raw(_print_inlining_output); 4563 } 4564 } 4565 4566 void Compile::log_late_inline(CallGenerator* cg) { 4567 if (log() != nullptr) { 4568 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4569 cg->unique_id()); 4570 JVMState* p = cg->call_node()->jvms(); 4571 while (p != nullptr) { 4572 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4573 p = p->caller(); 4574 } 4575 log()->tail("late_inline"); 4576 } 4577 } 4578 4579 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4580 log_late_inline(cg); 4581 if (log() != nullptr) { 4582 log()->inline_fail(msg); 4583 } 4584 } 4585 4586 void Compile::log_inline_id(CallGenerator* cg) { 4587 if (log() != nullptr) { 4588 // The LogCompilation tool needs a unique way to identify late 4589 // inline call sites. This id must be unique for this call site in 4590 // this compilation. Try to have it unique across compilations as 4591 // well because it can be convenient when grepping through the log 4592 // file. 4593 // Distinguish OSR compilations from others in case CICountOSR is 4594 // on. 4595 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4596 cg->set_unique_id(id); 4597 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4598 } 4599 } 4600 4601 void Compile::log_inline_failure(const char* msg) { 4602 if (C->log() != nullptr) { 4603 C->log()->inline_fail(msg); 4604 } 4605 } 4606 4607 4608 // Dump inlining replay data to the stream. 4609 // Don't change thread state and acquire any locks. 4610 void Compile::dump_inline_data(outputStream* out) { 4611 InlineTree* inl_tree = ilt(); 4612 if (inl_tree != nullptr) { 4613 out->print(" inline %d", inl_tree->count()); 4614 inl_tree->dump_replay_data(out); 4615 } 4616 } 4617 4618 void Compile::dump_inline_data_reduced(outputStream* out) { 4619 assert(ReplayReduce, ""); 4620 4621 InlineTree* inl_tree = ilt(); 4622 if (inl_tree == nullptr) { 4623 return; 4624 } 4625 // Enable iterative replay file reduction 4626 // Output "compile" lines for depth 1 subtrees, 4627 // simulating that those trees were compiled 4628 // instead of inlined. 4629 for (int i = 0; i < inl_tree->subtrees().length(); ++i) { 4630 InlineTree* sub = inl_tree->subtrees().at(i); 4631 if (sub->inline_level() != 1) { 4632 continue; 4633 } 4634 4635 ciMethod* method = sub->method(); 4636 int entry_bci = -1; 4637 int comp_level = env()->task()->comp_level(); 4638 out->print("compile "); 4639 method->dump_name_as_ascii(out); 4640 out->print(" %d %d", entry_bci, comp_level); 4641 out->print(" inline %d", sub->count()); 4642 sub->dump_replay_data(out, -1); 4643 out->cr(); 4644 } 4645 } 4646 4647 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4648 if (n1->Opcode() < n2->Opcode()) return -1; 4649 else if (n1->Opcode() > n2->Opcode()) return 1; 4650 4651 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4652 for (uint i = 1; i < n1->req(); i++) { 4653 if (n1->in(i) < n2->in(i)) return -1; 4654 else if (n1->in(i) > n2->in(i)) return 1; 4655 } 4656 4657 return 0; 4658 } 4659 4660 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4661 Node* n1 = *n1p; 4662 Node* n2 = *n2p; 4663 4664 return cmp_expensive_nodes(n1, n2); 4665 } 4666 4667 void Compile::sort_expensive_nodes() { 4668 if (!expensive_nodes_sorted()) { 4669 _expensive_nodes.sort(cmp_expensive_nodes); 4670 } 4671 } 4672 4673 bool Compile::expensive_nodes_sorted() const { 4674 for (int i = 1; i < _expensive_nodes.length(); i++) { 4675 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) { 4676 return false; 4677 } 4678 } 4679 return true; 4680 } 4681 4682 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4683 if (_expensive_nodes.length() == 0) { 4684 return false; 4685 } 4686 4687 assert(OptimizeExpensiveOps, "optimization off?"); 4688 4689 // Take this opportunity to remove dead nodes from the list 4690 int j = 0; 4691 for (int i = 0; i < _expensive_nodes.length(); i++) { 4692 Node* n = _expensive_nodes.at(i); 4693 if (!n->is_unreachable(igvn)) { 4694 assert(n->is_expensive(), "should be expensive"); 4695 _expensive_nodes.at_put(j, n); 4696 j++; 4697 } 4698 } 4699 _expensive_nodes.trunc_to(j); 4700 4701 // Then sort the list so that similar nodes are next to each other 4702 // and check for at least two nodes of identical kind with same data 4703 // inputs. 4704 sort_expensive_nodes(); 4705 4706 for (int i = 0; i < _expensive_nodes.length()-1; i++) { 4707 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) { 4708 return true; 4709 } 4710 } 4711 4712 return false; 4713 } 4714 4715 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4716 if (_expensive_nodes.length() == 0) { 4717 return; 4718 } 4719 4720 assert(OptimizeExpensiveOps, "optimization off?"); 4721 4722 // Sort to bring similar nodes next to each other and clear the 4723 // control input of nodes for which there's only a single copy. 4724 sort_expensive_nodes(); 4725 4726 int j = 0; 4727 int identical = 0; 4728 int i = 0; 4729 bool modified = false; 4730 for (; i < _expensive_nodes.length()-1; i++) { 4731 assert(j <= i, "can't write beyond current index"); 4732 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) { 4733 identical++; 4734 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4735 continue; 4736 } 4737 if (identical > 0) { 4738 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4739 identical = 0; 4740 } else { 4741 Node* n = _expensive_nodes.at(i); 4742 igvn.replace_input_of(n, 0, nullptr); 4743 igvn.hash_insert(n); 4744 modified = true; 4745 } 4746 } 4747 if (identical > 0) { 4748 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4749 } else if (_expensive_nodes.length() >= 1) { 4750 Node* n = _expensive_nodes.at(i); 4751 igvn.replace_input_of(n, 0, nullptr); 4752 igvn.hash_insert(n); 4753 modified = true; 4754 } 4755 _expensive_nodes.trunc_to(j); 4756 if (modified) { 4757 igvn.optimize(); 4758 } 4759 } 4760 4761 void Compile::add_expensive_node(Node * n) { 4762 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list"); 4763 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4764 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4765 if (OptimizeExpensiveOps) { 4766 _expensive_nodes.append(n); 4767 } else { 4768 // Clear control input and let IGVN optimize expensive nodes if 4769 // OptimizeExpensiveOps is off. 4770 n->set_req(0, nullptr); 4771 } 4772 } 4773 4774 /** 4775 * Track coarsened Lock and Unlock nodes. 4776 */ 4777 4778 class Lock_List : public Node_List { 4779 uint _origin_cnt; 4780 public: 4781 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {} 4782 uint origin_cnt() const { return _origin_cnt; } 4783 }; 4784 4785 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) { 4786 int length = locks.length(); 4787 if (length > 0) { 4788 // Have to keep this list until locks elimination during Macro nodes elimination. 4789 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length); 4790 for (int i = 0; i < length; i++) { 4791 AbstractLockNode* lock = locks.at(i); 4792 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx); 4793 locks_list->push(lock); 4794 } 4795 _coarsened_locks.append(locks_list); 4796 } 4797 } 4798 4799 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) { 4800 int count = coarsened_count(); 4801 for (int i = 0; i < count; i++) { 4802 Node_List* locks_list = _coarsened_locks.at(i); 4803 for (uint j = 0; j < locks_list->size(); j++) { 4804 Node* lock = locks_list->at(j); 4805 assert(lock->is_AbstractLock(), "sanity"); 4806 if (!useful.member(lock)) { 4807 locks_list->yank(lock); 4808 } 4809 } 4810 } 4811 } 4812 4813 void Compile::remove_coarsened_lock(Node* n) { 4814 if (n->is_AbstractLock()) { 4815 int count = coarsened_count(); 4816 for (int i = 0; i < count; i++) { 4817 Node_List* locks_list = _coarsened_locks.at(i); 4818 locks_list->yank(n); 4819 } 4820 } 4821 } 4822 4823 bool Compile::coarsened_locks_consistent() { 4824 int count = coarsened_count(); 4825 for (int i = 0; i < count; i++) { 4826 bool unbalanced = false; 4827 bool modified = false; // track locks kind modifications 4828 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i); 4829 uint size = locks_list->size(); 4830 if (size == 0) { 4831 unbalanced = false; // All locks were eliminated - good 4832 } else if (size != locks_list->origin_cnt()) { 4833 unbalanced = true; // Some locks were removed from list 4834 } else { 4835 for (uint j = 0; j < size; j++) { 4836 Node* lock = locks_list->at(j); 4837 // All nodes in group should have the same state (modified or not) 4838 if (!lock->as_AbstractLock()->is_coarsened()) { 4839 if (j == 0) { 4840 // first on list was modified, the rest should be too for consistency 4841 modified = true; 4842 } else if (!modified) { 4843 // this lock was modified but previous locks on the list were not 4844 unbalanced = true; 4845 break; 4846 } 4847 } else if (modified) { 4848 // previous locks on list were modified but not this lock 4849 unbalanced = true; 4850 break; 4851 } 4852 } 4853 } 4854 if (unbalanced) { 4855 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified 4856 #ifdef ASSERT 4857 if (PrintEliminateLocks) { 4858 tty->print_cr("=== unbalanced coarsened locks ==="); 4859 for (uint l = 0; l < size; l++) { 4860 locks_list->at(l)->dump(); 4861 } 4862 } 4863 #endif 4864 record_failure(C2Compiler::retry_no_locks_coarsening()); 4865 return false; 4866 } 4867 } 4868 return true; 4869 } 4870 4871 /** 4872 * Remove the speculative part of types and clean up the graph 4873 */ 4874 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4875 if (UseTypeSpeculation) { 4876 Unique_Node_List worklist; 4877 worklist.push(root()); 4878 int modified = 0; 4879 // Go over all type nodes that carry a speculative type, drop the 4880 // speculative part of the type and enqueue the node for an igvn 4881 // which may optimize it out. 4882 for (uint next = 0; next < worklist.size(); ++next) { 4883 Node *n = worklist.at(next); 4884 if (n->is_Type()) { 4885 TypeNode* tn = n->as_Type(); 4886 const Type* t = tn->type(); 4887 const Type* t_no_spec = t->remove_speculative(); 4888 if (t_no_spec != t) { 4889 bool in_hash = igvn.hash_delete(n); 4890 assert(in_hash, "node should be in igvn hash table"); 4891 tn->set_type(t_no_spec); 4892 igvn.hash_insert(n); 4893 igvn._worklist.push(n); // give it a chance to go away 4894 modified++; 4895 } 4896 } 4897 // Iterate over outs - endless loops is unreachable from below 4898 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4899 Node *m = n->fast_out(i); 4900 if (not_a_node(m)) { 4901 continue; 4902 } 4903 worklist.push(m); 4904 } 4905 } 4906 // Drop the speculative part of all types in the igvn's type table 4907 igvn.remove_speculative_types(); 4908 if (modified > 0) { 4909 igvn.optimize(); 4910 } 4911 #ifdef ASSERT 4912 // Verify that after the IGVN is over no speculative type has resurfaced 4913 worklist.clear(); 4914 worklist.push(root()); 4915 for (uint next = 0; next < worklist.size(); ++next) { 4916 Node *n = worklist.at(next); 4917 const Type* t = igvn.type_or_null(n); 4918 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types"); 4919 if (n->is_Type()) { 4920 t = n->as_Type()->type(); 4921 assert(t == t->remove_speculative(), "no more speculative types"); 4922 } 4923 // Iterate over outs - endless loops is unreachable from below 4924 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4925 Node *m = n->fast_out(i); 4926 if (not_a_node(m)) { 4927 continue; 4928 } 4929 worklist.push(m); 4930 } 4931 } 4932 igvn.check_no_speculative_types(); 4933 #endif 4934 } 4935 } 4936 4937 // Auxiliary methods to support randomized stressing/fuzzing. 4938 4939 int Compile::random() { 4940 _stress_seed = os::next_random(_stress_seed); 4941 return static_cast<int>(_stress_seed); 4942 } 4943 4944 // This method can be called the arbitrary number of times, with current count 4945 // as the argument. The logic allows selecting a single candidate from the 4946 // running list of candidates as follows: 4947 // int count = 0; 4948 // Cand* selected = null; 4949 // while(cand = cand->next()) { 4950 // if (randomized_select(++count)) { 4951 // selected = cand; 4952 // } 4953 // } 4954 // 4955 // Including count equalizes the chances any candidate is "selected". 4956 // This is useful when we don't have the complete list of candidates to choose 4957 // from uniformly. In this case, we need to adjust the randomicity of the 4958 // selection, or else we will end up biasing the selection towards the latter 4959 // candidates. 4960 // 4961 // Quick back-envelope calculation shows that for the list of n candidates 4962 // the equal probability for the candidate to persist as "best" can be 4963 // achieved by replacing it with "next" k-th candidate with the probability 4964 // of 1/k. It can be easily shown that by the end of the run, the 4965 // probability for any candidate is converged to 1/n, thus giving the 4966 // uniform distribution among all the candidates. 4967 // 4968 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4969 #define RANDOMIZED_DOMAIN_POW 29 4970 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4971 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4972 bool Compile::randomized_select(int count) { 4973 assert(count > 0, "only positive"); 4974 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4975 } 4976 4977 CloneMap& Compile::clone_map() { return _clone_map; } 4978 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4979 4980 void NodeCloneInfo::dump() const { 4981 tty->print(" {%d:%d} ", idx(), gen()); 4982 } 4983 4984 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4985 uint64_t val = value(old->_idx); 4986 NodeCloneInfo cio(val); 4987 assert(val != 0, "old node should be in the map"); 4988 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4989 insert(nnn->_idx, cin.get()); 4990 #ifndef PRODUCT 4991 if (is_debug()) { 4992 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4993 } 4994 #endif 4995 } 4996 4997 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4998 NodeCloneInfo cio(value(old->_idx)); 4999 if (cio.get() == 0) { 5000 cio.set(old->_idx, 0); 5001 insert(old->_idx, cio.get()); 5002 #ifndef PRODUCT 5003 if (is_debug()) { 5004 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 5005 } 5006 #endif 5007 } 5008 clone(old, nnn, gen); 5009 } 5010 5011 int CloneMap::max_gen() const { 5012 int g = 0; 5013 DictI di(_dict); 5014 for(; di.test(); ++di) { 5015 int t = gen(di._key); 5016 if (g < t) { 5017 g = t; 5018 #ifndef PRODUCT 5019 if (is_debug()) { 5020 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 5021 } 5022 #endif 5023 } 5024 } 5025 return g; 5026 } 5027 5028 void CloneMap::dump(node_idx_t key) const { 5029 uint64_t val = value(key); 5030 if (val != 0) { 5031 NodeCloneInfo ni(val); 5032 ni.dump(); 5033 } 5034 } 5035 5036 // Move Allocate nodes to the start of the list 5037 void Compile::sort_macro_nodes() { 5038 int count = macro_count(); 5039 int allocates = 0; 5040 for (int i = 0; i < count; i++) { 5041 Node* n = macro_node(i); 5042 if (n->is_Allocate()) { 5043 if (i != allocates) { 5044 Node* tmp = macro_node(allocates); 5045 _macro_nodes.at_put(allocates, n); 5046 _macro_nodes.at_put(i, tmp); 5047 } 5048 allocates++; 5049 } 5050 } 5051 } 5052 5053 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) { 5054 EventCompilerPhase event; 5055 if (event.should_commit()) { 5056 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 5057 } 5058 #ifndef PRODUCT 5059 ResourceMark rm; 5060 stringStream ss; 5061 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt)); 5062 if (n != nullptr) { 5063 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]); 5064 } 5065 5066 const char* name = ss.as_string(); 5067 if (should_print_igv(level)) { 5068 _igv_printer->print_method(name, level); 5069 } 5070 if (should_print_phase(cpt)) { 5071 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt)); 5072 } 5073 #endif 5074 C->_latest_stage_start_counter.stamp(); 5075 } 5076 5077 // Only used from CompileWrapper 5078 void Compile::begin_method() { 5079 #ifndef PRODUCT 5080 if (_method != nullptr && should_print_igv(1)) { 5081 _igv_printer->begin_method(); 5082 } 5083 #endif 5084 C->_latest_stage_start_counter.stamp(); 5085 } 5086 5087 // Only used from CompileWrapper 5088 void Compile::end_method() { 5089 EventCompilerPhase event; 5090 if (event.should_commit()) { 5091 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1); 5092 } 5093 5094 #ifndef PRODUCT 5095 if (_method != nullptr && should_print_igv(1)) { 5096 _igv_printer->end_method(); 5097 } 5098 #endif 5099 } 5100 5101 bool Compile::should_print_phase(CompilerPhaseType cpt) { 5102 #ifndef PRODUCT 5103 if ((_directive->ideal_phase_mask() & CompilerPhaseTypeHelper::to_bitmask(cpt)) != 0) { 5104 return true; 5105 } 5106 #endif 5107 return false; 5108 } 5109 5110 bool Compile::should_print_igv(int level) { 5111 #ifndef PRODUCT 5112 if (PrintIdealGraphLevel < 0) { // disabled by the user 5113 return false; 5114 } 5115 5116 bool need = directive()->IGVPrintLevelOption >= level; 5117 if (need && !_igv_printer) { 5118 _igv_printer = IdealGraphPrinter::printer(); 5119 _igv_printer->set_compile(this); 5120 } 5121 return need; 5122 #else 5123 return false; 5124 #endif 5125 } 5126 5127 #ifndef PRODUCT 5128 IdealGraphPrinter* Compile::_debug_file_printer = nullptr; 5129 IdealGraphPrinter* Compile::_debug_network_printer = nullptr; 5130 5131 // Called from debugger. Prints method to the default file with the default phase name. 5132 // This works regardless of any Ideal Graph Visualizer flags set or not. 5133 void igv_print() { 5134 Compile::current()->igv_print_method_to_file(); 5135 } 5136 5137 // Same as igv_print() above but with a specified phase name. 5138 void igv_print(const char* phase_name) { 5139 Compile::current()->igv_print_method_to_file(phase_name); 5140 } 5141 5142 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 5143 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 5144 // This works regardless of any Ideal Graph Visualizer flags set or not. 5145 void igv_print(bool network) { 5146 if (network) { 5147 Compile::current()->igv_print_method_to_network(); 5148 } else { 5149 Compile::current()->igv_print_method_to_file(); 5150 } 5151 } 5152 5153 // Same as igv_print(bool network) above but with a specified phase name. 5154 void igv_print(bool network, const char* phase_name) { 5155 if (network) { 5156 Compile::current()->igv_print_method_to_network(phase_name); 5157 } else { 5158 Compile::current()->igv_print_method_to_file(phase_name); 5159 } 5160 } 5161 5162 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 5163 void igv_print_default() { 5164 Compile::current()->print_method(PHASE_DEBUG, 0); 5165 } 5166 5167 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 5168 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 5169 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 5170 void igv_append() { 5171 Compile::current()->igv_print_method_to_file("Debug", true); 5172 } 5173 5174 // Same as igv_append() above but with a specified phase name. 5175 void igv_append(const char* phase_name) { 5176 Compile::current()->igv_print_method_to_file(phase_name, true); 5177 } 5178 5179 void Compile::igv_print_method_to_file(const char* phase_name, bool append) { 5180 const char* file_name = "custom_debug.xml"; 5181 if (_debug_file_printer == nullptr) { 5182 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 5183 } else { 5184 _debug_file_printer->update_compiled_method(C->method()); 5185 } 5186 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 5187 _debug_file_printer->print(phase_name, (Node*)C->root()); 5188 } 5189 5190 void Compile::igv_print_method_to_network(const char* phase_name) { 5191 if (_debug_network_printer == nullptr) { 5192 _debug_network_printer = new IdealGraphPrinter(C); 5193 } else { 5194 _debug_network_printer->update_compiled_method(C->method()); 5195 } 5196 tty->print_cr("Method printed over network stream to IGV"); 5197 _debug_network_printer->print(phase_name, (Node*)C->root()); 5198 } 5199 #endif 5200 5201 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 5202 if (type != nullptr && phase->type(value)->higher_equal(type)) { 5203 return value; 5204 } 5205 Node* result = nullptr; 5206 if (bt == T_BYTE) { 5207 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 5208 result = new RShiftINode(result, phase->intcon(24)); 5209 } else if (bt == T_BOOLEAN) { 5210 result = new AndINode(value, phase->intcon(0xFF)); 5211 } else if (bt == T_CHAR) { 5212 result = new AndINode(value,phase->intcon(0xFFFF)); 5213 } else { 5214 assert(bt == T_SHORT, "unexpected narrow type"); 5215 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 5216 result = new RShiftINode(result, phase->intcon(16)); 5217 } 5218 if (transform_res) { 5219 result = phase->transform(result); 5220 } 5221 return result; 5222 } 5223