1 /* 2 * Copyright (c) 1997, 2018, 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/systemDictionary.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 "memory/resourceArea.hpp" 39 #include "opto/addnode.hpp" 40 #include "opto/block.hpp" 41 #include "opto/c2compiler.hpp" 42 #include "opto/callGenerator.hpp" 43 #include "opto/callnode.hpp" 44 #include "opto/castnode.hpp" 45 #include "opto/cfgnode.hpp" 46 #include "opto/chaitin.hpp" 47 #include "opto/compile.hpp" 48 #include "opto/connode.hpp" 49 #include "opto/convertnode.hpp" 50 #include "opto/divnode.hpp" 51 #include "opto/escape.hpp" 52 #include "opto/idealGraphPrinter.hpp" 53 #include "opto/loopnode.hpp" 54 #include "opto/machnode.hpp" 55 #include "opto/macro.hpp" 56 #include "opto/matcher.hpp" 57 #include "opto/mathexactnode.hpp" 58 #include "opto/memnode.hpp" 59 #include "opto/mulnode.hpp" 60 #include "opto/narrowptrnode.hpp" 61 #include "opto/node.hpp" 62 #include "opto/opcodes.hpp" 63 #include "opto/output.hpp" 64 #include "opto/parse.hpp" 65 #include "opto/phaseX.hpp" 66 #include "opto/rootnode.hpp" 67 #include "opto/runtime.hpp" 68 #include "opto/stringopts.hpp" 69 #include "opto/type.hpp" 70 #include "opto/vectornode.hpp" 71 #include "runtime/arguments.hpp" 72 #include "runtime/sharedRuntime.hpp" 73 #include "runtime/signature.hpp" 74 #include "runtime/stubRoutines.hpp" 75 #include "runtime/timer.hpp" 76 #include "utilities/align.hpp" 77 #include "utilities/copy.hpp" 78 #include "utilities/macros.hpp" 79 #if INCLUDE_G1GC 80 #include "gc/g1/g1ThreadLocalData.hpp" 81 #endif // INCLUDE_G1GC 82 #if INCLUDE_ZGC 83 #include "gc/z/c2/zBarrierSetC2.hpp" 84 #endif 85 86 87 // -------------------- Compile::mach_constant_base_node ----------------------- 88 // Constant table base node singleton. 89 MachConstantBaseNode* Compile::mach_constant_base_node() { 90 if (_mach_constant_base_node == NULL) { 91 _mach_constant_base_node = new MachConstantBaseNode(); 92 _mach_constant_base_node->add_req(C->root()); 93 } 94 return _mach_constant_base_node; 95 } 96 97 98 /// Support for intrinsics. 99 100 // Return the index at which m must be inserted (or already exists). 101 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 102 class IntrinsicDescPair { 103 private: 104 ciMethod* _m; 105 bool _is_virtual; 106 public: 107 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 108 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 109 ciMethod* m= elt->method(); 110 ciMethod* key_m = key->_m; 111 if (key_m < m) return -1; 112 else if (key_m > m) return 1; 113 else { 114 bool is_virtual = elt->is_virtual(); 115 bool key_virtual = key->_is_virtual; 116 if (key_virtual < is_virtual) return -1; 117 else if (key_virtual > is_virtual) return 1; 118 else return 0; 119 } 120 } 121 }; 122 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 123 #ifdef ASSERT 124 for (int i = 1; i < _intrinsics->length(); i++) { 125 CallGenerator* cg1 = _intrinsics->at(i-1); 126 CallGenerator* cg2 = _intrinsics->at(i); 127 assert(cg1->method() != cg2->method() 128 ? cg1->method() < cg2->method() 129 : cg1->is_virtual() < cg2->is_virtual(), 130 "compiler intrinsics list must stay sorted"); 131 } 132 #endif 133 IntrinsicDescPair pair(m, is_virtual); 134 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 135 } 136 137 void Compile::register_intrinsic(CallGenerator* cg) { 138 if (_intrinsics == NULL) { 139 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 140 } 141 int len = _intrinsics->length(); 142 bool found = false; 143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 144 assert(!found, "registering twice"); 145 _intrinsics->insert_before(index, cg); 146 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 147 } 148 149 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 150 assert(m->is_loaded(), "don't try this on unloaded methods"); 151 if (_intrinsics != NULL) { 152 bool found = false; 153 int index = intrinsic_insertion_index(m, is_virtual, found); 154 if (found) { 155 return _intrinsics->at(index); 156 } 157 } 158 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 159 if (m->intrinsic_id() != vmIntrinsics::_none && 160 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 161 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 162 if (cg != NULL) { 163 // Save it for next time: 164 register_intrinsic(cg); 165 return cg; 166 } else { 167 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 168 } 169 } 170 return NULL; 171 } 172 173 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 174 // in library_call.cpp. 175 176 177 #ifndef PRODUCT 178 // statistics gathering... 179 180 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 181 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 182 183 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 184 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 185 int oflags = _intrinsic_hist_flags[id]; 186 assert(flags != 0, "what happened?"); 187 if (is_virtual) { 188 flags |= _intrinsic_virtual; 189 } 190 bool changed = (flags != oflags); 191 if ((flags & _intrinsic_worked) != 0) { 192 juint count = (_intrinsic_hist_count[id] += 1); 193 if (count == 1) { 194 changed = true; // first time 195 } 196 // increment the overall count also: 197 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 198 } 199 if (changed) { 200 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 201 // Something changed about the intrinsic's virtuality. 202 if ((flags & _intrinsic_virtual) != 0) { 203 // This is the first use of this intrinsic as a virtual call. 204 if (oflags != 0) { 205 // We already saw it as a non-virtual, so note both cases. 206 flags |= _intrinsic_both; 207 } 208 } else if ((oflags & _intrinsic_both) == 0) { 209 // This is the first use of this intrinsic as a non-virtual 210 flags |= _intrinsic_both; 211 } 212 } 213 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 214 } 215 // update the overall flags also: 216 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 217 return changed; 218 } 219 220 static char* format_flags(int flags, char* buf) { 221 buf[0] = 0; 222 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 223 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 224 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 225 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 226 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 227 if (buf[0] == 0) strcat(buf, ","); 228 assert(buf[0] == ',', "must be"); 229 return &buf[1]; 230 } 231 232 void Compile::print_intrinsic_statistics() { 233 char flagsbuf[100]; 234 ttyLocker ttyl; 235 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 236 tty->print_cr("Compiler intrinsic usage:"); 237 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 238 if (total == 0) total = 1; // avoid div0 in case of no successes 239 #define PRINT_STAT_LINE(name, c, f) \ 240 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 241 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 242 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 243 int flags = _intrinsic_hist_flags[id]; 244 juint count = _intrinsic_hist_count[id]; 245 if ((flags | count) != 0) { 246 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 247 } 248 } 249 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 250 if (xtty != NULL) xtty->tail("statistics"); 251 } 252 253 void Compile::print_statistics() { 254 { ttyLocker ttyl; 255 if (xtty != NULL) xtty->head("statistics type='opto'"); 256 Parse::print_statistics(); 257 PhaseCCP::print_statistics(); 258 PhaseRegAlloc::print_statistics(); 259 Scheduling::print_statistics(); 260 PhasePeephole::print_statistics(); 261 PhaseIdealLoop::print_statistics(); 262 if (xtty != NULL) xtty->tail("statistics"); 263 } 264 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 265 // put this under its own <statistics> element. 266 print_intrinsic_statistics(); 267 } 268 } 269 #endif //PRODUCT 270 271 // Support for bundling info 272 Bundle* Compile::node_bundling(const Node *n) { 273 assert(valid_bundle_info(n), "oob"); 274 return &_node_bundling_base[n->_idx]; 275 } 276 277 bool Compile::valid_bundle_info(const Node *n) { 278 return (_node_bundling_limit > n->_idx); 279 } 280 281 282 void Compile::gvn_replace_by(Node* n, Node* nn) { 283 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 284 Node* use = n->last_out(i); 285 bool is_in_table = initial_gvn()->hash_delete(use); 286 uint uses_found = 0; 287 for (uint j = 0; j < use->len(); j++) { 288 if (use->in(j) == n) { 289 if (j < use->req()) 290 use->set_req(j, nn); 291 else 292 use->set_prec(j, nn); 293 uses_found++; 294 } 295 } 296 if (is_in_table) { 297 // reinsert into table 298 initial_gvn()->hash_find_insert(use); 299 } 300 record_for_igvn(use); 301 i -= uses_found; // we deleted 1 or more copies of this edge 302 } 303 } 304 305 306 static inline bool not_a_node(const Node* n) { 307 if (n == NULL) return true; 308 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 309 if (*(address*)n == badAddress) return true; // kill by Node::destruct 310 return false; 311 } 312 313 // Identify all nodes that are reachable from below, useful. 314 // Use breadth-first pass that records state in a Unique_Node_List, 315 // recursive traversal is slower. 316 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 317 int estimated_worklist_size = live_nodes(); 318 useful.map( estimated_worklist_size, NULL ); // preallocate space 319 320 // Initialize worklist 321 if (root() != NULL) { useful.push(root()); } 322 // If 'top' is cached, declare it useful to preserve cached node 323 if( cached_top_node() ) { useful.push(cached_top_node()); } 324 325 // Push all useful nodes onto the list, breadthfirst 326 for( uint next = 0; next < useful.size(); ++next ) { 327 assert( next < unique(), "Unique useful nodes < total nodes"); 328 Node *n = useful.at(next); 329 uint max = n->len(); 330 for( uint i = 0; i < max; ++i ) { 331 Node *m = n->in(i); 332 if (not_a_node(m)) continue; 333 useful.push(m); 334 } 335 } 336 } 337 338 // Update dead_node_list with any missing dead nodes using useful 339 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 340 void Compile::update_dead_node_list(Unique_Node_List &useful) { 341 uint max_idx = unique(); 342 VectorSet& useful_node_set = useful.member_set(); 343 344 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 345 // If node with index node_idx is not in useful set, 346 // mark it as dead in dead node list. 347 if (! useful_node_set.test(node_idx) ) { 348 record_dead_node(node_idx); 349 } 350 } 351 } 352 353 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 354 int shift = 0; 355 for (int i = 0; i < inlines->length(); i++) { 356 CallGenerator* cg = inlines->at(i); 357 CallNode* call = cg->call_node(); 358 if (shift > 0) { 359 inlines->at_put(i-shift, cg); 360 } 361 if (!useful.member(call)) { 362 shift++; 363 } 364 } 365 inlines->trunc_to(inlines->length()-shift); 366 } 367 368 // Disconnect all useless nodes by disconnecting those at the boundary. 369 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 370 uint next = 0; 371 while (next < useful.size()) { 372 Node *n = useful.at(next++); 373 if (n->is_SafePoint()) { 374 // We're done with a parsing phase. Replaced nodes are not valid 375 // beyond that point. 376 n->as_SafePoint()->delete_replaced_nodes(); 377 } 378 // Use raw traversal of out edges since this code removes out edges 379 int max = n->outcnt(); 380 for (int j = 0; j < max; ++j) { 381 Node* child = n->raw_out(j); 382 if (! useful.member(child)) { 383 assert(!child->is_top() || child != top(), 384 "If top is cached in Compile object it is in useful list"); 385 // Only need to remove this out-edge to the useless node 386 n->raw_del_out(j); 387 --j; 388 --max; 389 } 390 } 391 if (n->outcnt() == 1 && n->has_special_unique_user()) { 392 record_for_igvn(n->unique_out()); 393 } 394 } 395 // Remove useless macro and predicate opaq nodes 396 for (int i = C->macro_count()-1; i >= 0; i--) { 397 Node* n = C->macro_node(i); 398 if (!useful.member(n)) { 399 remove_macro_node(n); 400 } 401 } 402 // Remove useless CastII nodes with range check dependency 403 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 404 Node* cast = range_check_cast_node(i); 405 if (!useful.member(cast)) { 406 remove_range_check_cast(cast); 407 } 408 } 409 // Remove useless expensive nodes 410 for (int i = C->expensive_count()-1; i >= 0; i--) { 411 Node* n = C->expensive_node(i); 412 if (!useful.member(n)) { 413 remove_expensive_node(n); 414 } 415 } 416 // Remove useless Opaque4 nodes 417 for (int i = opaque4_count() - 1; i >= 0; i--) { 418 Node* opaq = opaque4_node(i); 419 if (!useful.member(opaq)) { 420 remove_opaque4_node(opaq); 421 } 422 } 423 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 424 bs->eliminate_useless_gc_barriers(useful); 425 // clean up the late inline lists 426 remove_useless_late_inlines(&_string_late_inlines, useful); 427 remove_useless_late_inlines(&_boxing_late_inlines, useful); 428 remove_useless_late_inlines(&_late_inlines, useful); 429 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 430 } 431 432 //------------------------------frame_size_in_words----------------------------- 433 // frame_slots in units of words 434 int Compile::frame_size_in_words() const { 435 // shift is 0 in LP32 and 1 in LP64 436 const int shift = (LogBytesPerWord - LogBytesPerInt); 437 int words = _frame_slots >> shift; 438 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 439 return words; 440 } 441 442 // To bang the stack of this compiled method we use the stack size 443 // that the interpreter would need in case of a deoptimization. This 444 // removes the need to bang the stack in the deoptimization blob which 445 // in turn simplifies stack overflow handling. 446 int Compile::bang_size_in_bytes() const { 447 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 448 } 449 450 // ============================================================================ 451 //------------------------------CompileWrapper--------------------------------- 452 class CompileWrapper : public StackObj { 453 Compile *const _compile; 454 public: 455 CompileWrapper(Compile* compile); 456 457 ~CompileWrapper(); 458 }; 459 460 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 461 // the Compile* pointer is stored in the current ciEnv: 462 ciEnv* env = compile->env(); 463 assert(env == ciEnv::current(), "must already be a ciEnv active"); 464 assert(env->compiler_data() == NULL, "compile already active?"); 465 env->set_compiler_data(compile); 466 assert(compile == Compile::current(), "sanity"); 467 468 compile->set_type_dict(NULL); 469 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 470 compile->clone_map().set_clone_idx(0); 471 compile->set_type_hwm(NULL); 472 compile->set_type_last_size(0); 473 compile->set_last_tf(NULL, NULL); 474 compile->set_indexSet_arena(NULL); 475 compile->set_indexSet_free_block_list(NULL); 476 compile->init_type_arena(); 477 Type::Initialize(compile); 478 _compile->set_scratch_buffer_blob(NULL); 479 _compile->begin_method(); 480 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 481 } 482 CompileWrapper::~CompileWrapper() { 483 _compile->end_method(); 484 if (_compile->scratch_buffer_blob() != NULL) 485 BufferBlob::free(_compile->scratch_buffer_blob()); 486 _compile->env()->set_compiler_data(NULL); 487 } 488 489 490 //----------------------------print_compile_messages--------------------------- 491 void Compile::print_compile_messages() { 492 #ifndef PRODUCT 493 // Check if recompiling 494 if (_subsume_loads == false && PrintOpto) { 495 // Recompiling without allowing machine instructions to subsume loads 496 tty->print_cr("*********************************************************"); 497 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 498 tty->print_cr("*********************************************************"); 499 } 500 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 501 // Recompiling without escape analysis 502 tty->print_cr("*********************************************************"); 503 tty->print_cr("** Bailout: Recompile without escape analysis **"); 504 tty->print_cr("*********************************************************"); 505 } 506 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 507 // Recompiling without boxing elimination 508 tty->print_cr("*********************************************************"); 509 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 510 tty->print_cr("*********************************************************"); 511 } 512 if (C->directive()->BreakAtCompileOption) { 513 // Open the debugger when compiling this method. 514 tty->print("### Breaking when compiling: "); 515 method()->print_short_name(); 516 tty->cr(); 517 BREAKPOINT; 518 } 519 520 if( PrintOpto ) { 521 if (is_osr_compilation()) { 522 tty->print("[OSR]%3d", _compile_id); 523 } else { 524 tty->print("%3d", _compile_id); 525 } 526 } 527 #endif 528 } 529 530 531 //-----------------------init_scratch_buffer_blob------------------------------ 532 // Construct a temporary BufferBlob and cache it for this compile. 533 void Compile::init_scratch_buffer_blob(int const_size) { 534 // If there is already a scratch buffer blob allocated and the 535 // constant section is big enough, use it. Otherwise free the 536 // current and allocate a new one. 537 BufferBlob* blob = scratch_buffer_blob(); 538 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 539 // Use the current blob. 540 } else { 541 if (blob != NULL) { 542 BufferBlob::free(blob); 543 } 544 545 ResourceMark rm; 546 _scratch_const_size = const_size; 547 int size = C2Compiler::initial_code_buffer_size(const_size); 548 blob = BufferBlob::create("Compile::scratch_buffer", size); 549 // Record the buffer blob for next time. 550 set_scratch_buffer_blob(blob); 551 // Have we run out of code space? 552 if (scratch_buffer_blob() == NULL) { 553 // Let CompilerBroker disable further compilations. 554 record_failure("Not enough space for scratch buffer in CodeCache"); 555 return; 556 } 557 } 558 559 // Initialize the relocation buffers 560 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 561 set_scratch_locs_memory(locs_buf); 562 } 563 564 565 //-----------------------scratch_emit_size------------------------------------- 566 // Helper function that computes size by emitting code 567 uint Compile::scratch_emit_size(const Node* n) { 568 // Start scratch_emit_size section. 569 set_in_scratch_emit_size(true); 570 571 // Emit into a trash buffer and count bytes emitted. 572 // This is a pretty expensive way to compute a size, 573 // but it works well enough if seldom used. 574 // All common fixed-size instructions are given a size 575 // method by the AD file. 576 // Note that the scratch buffer blob and locs memory are 577 // allocated at the beginning of the compile task, and 578 // may be shared by several calls to scratch_emit_size. 579 // The allocation of the scratch buffer blob is particularly 580 // expensive, since it has to grab the code cache lock. 581 BufferBlob* blob = this->scratch_buffer_blob(); 582 assert(blob != NULL, "Initialize BufferBlob at start"); 583 assert(blob->size() > MAX_inst_size, "sanity"); 584 relocInfo* locs_buf = scratch_locs_memory(); 585 address blob_begin = blob->content_begin(); 586 address blob_end = (address)locs_buf; 587 assert(blob->contains(blob_end), "sanity"); 588 CodeBuffer buf(blob_begin, blob_end - blob_begin); 589 buf.initialize_consts_size(_scratch_const_size); 590 buf.initialize_stubs_size(MAX_stubs_size); 591 assert(locs_buf != NULL, "sanity"); 592 int lsize = MAX_locs_size / 3; 593 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 594 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 595 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 596 // Mark as scratch buffer. 597 buf.consts()->set_scratch_emit(); 598 buf.insts()->set_scratch_emit(); 599 buf.stubs()->set_scratch_emit(); 600 601 // Do the emission. 602 603 Label fakeL; // Fake label for branch instructions. 604 Label* saveL = NULL; 605 uint save_bnum = 0; 606 bool is_branch = n->is_MachBranch(); 607 if (is_branch) { 608 MacroAssembler masm(&buf); 609 masm.bind(fakeL); 610 n->as_MachBranch()->save_label(&saveL, &save_bnum); 611 n->as_MachBranch()->label_set(&fakeL, 0); 612 } 613 n->emit(buf, this->regalloc()); 614 615 // Emitting into the scratch buffer should not fail 616 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 617 618 if (is_branch) // Restore label. 619 n->as_MachBranch()->label_set(saveL, save_bnum); 620 621 // End scratch_emit_size section. 622 set_in_scratch_emit_size(false); 623 624 return buf.insts_size(); 625 } 626 627 628 // ============================================================================ 629 //------------------------------Compile standard------------------------------- 630 debug_only( int Compile::_debug_idx = 100000; ) 631 632 // Compile a method. entry_bci is -1 for normal compilations and indicates 633 // the continuation bci for on stack replacement. 634 635 636 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 637 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 638 : Phase(Compiler), 639 _env(ci_env), 640 _directive(directive), 641 _log(ci_env->log()), 642 _compile_id(ci_env->compile_id()), 643 _save_argument_registers(false), 644 _stub_name(NULL), 645 _stub_function(NULL), 646 _stub_entry_point(NULL), 647 _method(target), 648 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 649 _entry_bci(osr_bci), 650 _initial_gvn(NULL), 651 _for_igvn(NULL), 652 _warm_calls(NULL), 653 _subsume_loads(subsume_loads), 654 _do_escape_analysis(do_escape_analysis), 655 _eliminate_boxing(eliminate_boxing), 656 _failure_reason(NULL), 657 _code_buffer("Compile::Fill_buffer"), 658 _orig_pc_slot(0), 659 _orig_pc_slot_offset_in_bytes(0), 660 _has_method_handle_invokes(false), 661 _mach_constant_base_node(NULL), 662 _node_bundling_limit(0), 663 _node_bundling_base(NULL), 664 _java_calls(0), 665 _inner_loops(0), 666 _scratch_const_size(-1), 667 _in_scratch_emit_size(false), 668 _dead_node_list(comp_arena()), 669 _dead_node_count(0), 670 #ifndef PRODUCT 671 _trace_opto_output(directive->TraceOptoOutputOption), 672 _in_dump_cnt(0), 673 _printer(IdealGraphPrinter::printer()), 674 #endif 675 _congraph(NULL), 676 _comp_arena(mtCompiler), 677 _node_arena(mtCompiler), 678 _old_arena(mtCompiler), 679 _Compile_types(mtCompiler), 680 _replay_inline_data(NULL), 681 _late_inlines(comp_arena(), 2, 0, NULL), 682 _string_late_inlines(comp_arena(), 2, 0, NULL), 683 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 684 _late_inlines_pos(0), 685 _number_of_mh_late_inlines(0), 686 _inlining_progress(false), 687 _inlining_incrementally(false), 688 _print_inlining_list(NULL), 689 _print_inlining_stream(NULL), 690 _print_inlining_idx(0), 691 _print_inlining_output(NULL), 692 _interpreter_frame_size(0), 693 _max_node_limit(MaxNodeLimit), 694 _has_reserved_stack_access(target->has_reserved_stack_access()) { 695 C = this; 696 #ifndef PRODUCT 697 if (_printer != NULL) { 698 _printer->set_compile(this); 699 } 700 #endif 701 CompileWrapper cw(this); 702 703 if (CITimeVerbose) { 704 tty->print(" "); 705 target->holder()->name()->print(); 706 tty->print("."); 707 target->print_short_name(); 708 tty->print(" "); 709 } 710 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 711 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 712 713 #ifndef PRODUCT 714 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 715 if (!print_opto_assembly) { 716 bool print_assembly = directive->PrintAssemblyOption; 717 if (print_assembly && !Disassembler::can_decode()) { 718 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 719 print_opto_assembly = true; 720 } 721 } 722 set_print_assembly(print_opto_assembly); 723 set_parsed_irreducible_loop(false); 724 725 if (directive->ReplayInlineOption) { 726 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 727 } 728 #endif 729 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 730 set_print_intrinsics(directive->PrintIntrinsicsOption); 731 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 732 733 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 734 // Make sure the method being compiled gets its own MDO, 735 // so we can at least track the decompile_count(). 736 // Need MDO to record RTM code generation state. 737 method()->ensure_method_data(); 738 } 739 740 Init(::AliasLevel); 741 742 743 print_compile_messages(); 744 745 _ilt = InlineTree::build_inline_tree_root(); 746 747 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 748 assert(num_alias_types() >= AliasIdxRaw, ""); 749 750 #define MINIMUM_NODE_HASH 1023 751 // Node list that Iterative GVN will start with 752 Unique_Node_List for_igvn(comp_arena()); 753 set_for_igvn(&for_igvn); 754 755 // GVN that will be run immediately on new nodes 756 uint estimated_size = method()->code_size()*4+64; 757 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 758 PhaseGVN gvn(node_arena(), estimated_size); 759 set_initial_gvn(&gvn); 760 761 print_inlining_init(); 762 { // Scope for timing the parser 763 TracePhase tp("parse", &timers[_t_parser]); 764 765 // Put top into the hash table ASAP. 766 initial_gvn()->transform_no_reclaim(top()); 767 768 // Set up tf(), start(), and find a CallGenerator. 769 CallGenerator* cg = NULL; 770 if (is_osr_compilation()) { 771 const TypeTuple *domain = StartOSRNode::osr_domain(); 772 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 773 init_tf(TypeFunc::make(domain, range)); 774 StartNode* s = new StartOSRNode(root(), domain); 775 initial_gvn()->set_type_bottom(s); 776 init_start(s); 777 cg = CallGenerator::for_osr(method(), entry_bci()); 778 } else { 779 // Normal case. 780 init_tf(TypeFunc::make(method())); 781 StartNode* s = new StartNode(root(), tf()->domain()); 782 initial_gvn()->set_type_bottom(s); 783 init_start(s); 784 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 785 // With java.lang.ref.reference.get() we must go through the 786 // intrinsic - even when get() is the root 787 // method of the compile - so that, if necessary, the value in 788 // the referent field of the reference object gets recorded by 789 // the pre-barrier code. 790 cg = find_intrinsic(method(), false); 791 } 792 if (cg == NULL) { 793 float past_uses = method()->interpreter_invocation_count(); 794 float expected_uses = past_uses; 795 cg = CallGenerator::for_inline(method(), expected_uses); 796 } 797 } 798 if (failing()) return; 799 if (cg == NULL) { 800 record_method_not_compilable("cannot parse method"); 801 return; 802 } 803 JVMState* jvms = build_start_state(start(), tf()); 804 if ((jvms = cg->generate(jvms)) == NULL) { 805 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 806 record_method_not_compilable("method parse failed"); 807 } 808 return; 809 } 810 GraphKit kit(jvms); 811 812 if (!kit.stopped()) { 813 // Accept return values, and transfer control we know not where. 814 // This is done by a special, unique ReturnNode bound to root. 815 return_values(kit.jvms()); 816 } 817 818 if (kit.has_exceptions()) { 819 // Any exceptions that escape from this call must be rethrown 820 // to whatever caller is dynamically above us on the stack. 821 // This is done by a special, unique RethrowNode bound to root. 822 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 823 } 824 825 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 826 827 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 828 inline_string_calls(true); 829 } 830 831 if (failing()) return; 832 833 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 834 835 // Remove clutter produced by parsing. 836 if (!failing()) { 837 ResourceMark rm; 838 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 839 } 840 } 841 842 // Note: Large methods are capped off in do_one_bytecode(). 843 if (failing()) return; 844 845 // After parsing, node notes are no longer automagic. 846 // They must be propagated by register_new_node_with_optimizer(), 847 // clone(), or the like. 848 set_default_node_notes(NULL); 849 850 for (;;) { 851 int successes = Inline_Warm(); 852 if (failing()) return; 853 if (successes == 0) break; 854 } 855 856 // Drain the list. 857 Finish_Warm(); 858 #ifndef PRODUCT 859 if (_printer && _printer->should_print(1)) { 860 _printer->print_inlining(); 861 } 862 #endif 863 864 if (failing()) return; 865 NOT_PRODUCT( verify_graph_edges(); ) 866 867 // Now optimize 868 Optimize(); 869 if (failing()) return; 870 NOT_PRODUCT( verify_graph_edges(); ) 871 872 #ifndef PRODUCT 873 if (PrintIdeal) { 874 ttyLocker ttyl; // keep the following output all in one block 875 // This output goes directly to the tty, not the compiler log. 876 // To enable tools to match it up with the compilation activity, 877 // be sure to tag this tty output with the compile ID. 878 if (xtty != NULL) { 879 xtty->head("ideal compile_id='%d'%s", compile_id(), 880 is_osr_compilation() ? " compile_kind='osr'" : 881 ""); 882 } 883 root()->dump(9999); 884 if (xtty != NULL) { 885 xtty->tail("ideal"); 886 } 887 } 888 #endif 889 890 NOT_PRODUCT( verify_barriers(); ) 891 892 // Dump compilation data to replay it. 893 if (directive->DumpReplayOption) { 894 env()->dump_replay_data(_compile_id); 895 } 896 if (directive->DumpInlineOption && (ilt() != NULL)) { 897 env()->dump_inline_data(_compile_id); 898 } 899 900 // Now that we know the size of all the monitors we can add a fixed slot 901 // for the original deopt pc. 902 903 _orig_pc_slot = fixed_slots(); 904 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 905 set_fixed_slots(next_slot); 906 907 // Compute when to use implicit null checks. Used by matching trap based 908 // nodes and NullCheck optimization. 909 set_allowed_deopt_reasons(); 910 911 // Now generate code 912 Code_Gen(); 913 if (failing()) return; 914 915 // Check if we want to skip execution of all compiled code. 916 { 917 #ifndef PRODUCT 918 if (OptoNoExecute) { 919 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 920 return; 921 } 922 #endif 923 TracePhase tp("install_code", &timers[_t_registerMethod]); 924 925 if (is_osr_compilation()) { 926 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 927 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 928 } else { 929 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 930 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 931 } 932 933 env()->register_method(_method, _entry_bci, 934 &_code_offsets, 935 _orig_pc_slot_offset_in_bytes, 936 code_buffer(), 937 frame_size_in_words(), _oop_map_set, 938 &_handler_table, &_inc_table, 939 compiler, 940 has_unsafe_access(), 941 SharedRuntime::is_wide_vector(max_vector_size()), 942 rtm_state() 943 ); 944 945 if (log() != NULL) // Print code cache state into compiler log 946 log()->code_cache_state(); 947 } 948 } 949 950 //------------------------------Compile---------------------------------------- 951 // Compile a runtime stub 952 Compile::Compile( ciEnv* ci_env, 953 TypeFunc_generator generator, 954 address stub_function, 955 const char *stub_name, 956 int is_fancy_jump, 957 bool pass_tls, 958 bool save_arg_registers, 959 bool return_pc, 960 DirectiveSet* directive) 961 : Phase(Compiler), 962 _env(ci_env), 963 _directive(directive), 964 _log(ci_env->log()), 965 _compile_id(0), 966 _save_argument_registers(save_arg_registers), 967 _method(NULL), 968 _stub_name(stub_name), 969 _stub_function(stub_function), 970 _stub_entry_point(NULL), 971 _entry_bci(InvocationEntryBci), 972 _initial_gvn(NULL), 973 _for_igvn(NULL), 974 _warm_calls(NULL), 975 _orig_pc_slot(0), 976 _orig_pc_slot_offset_in_bytes(0), 977 _subsume_loads(true), 978 _do_escape_analysis(false), 979 _eliminate_boxing(false), 980 _failure_reason(NULL), 981 _code_buffer("Compile::Fill_buffer"), 982 _has_method_handle_invokes(false), 983 _mach_constant_base_node(NULL), 984 _node_bundling_limit(0), 985 _node_bundling_base(NULL), 986 _java_calls(0), 987 _inner_loops(0), 988 #ifndef PRODUCT 989 _trace_opto_output(directive->TraceOptoOutputOption), 990 _in_dump_cnt(0), 991 _printer(NULL), 992 #endif 993 _comp_arena(mtCompiler), 994 _node_arena(mtCompiler), 995 _old_arena(mtCompiler), 996 _Compile_types(mtCompiler), 997 _dead_node_list(comp_arena()), 998 _dead_node_count(0), 999 _congraph(NULL), 1000 _replay_inline_data(NULL), 1001 _number_of_mh_late_inlines(0), 1002 _inlining_progress(false), 1003 _inlining_incrementally(false), 1004 _print_inlining_list(NULL), 1005 _print_inlining_stream(NULL), 1006 _print_inlining_idx(0), 1007 _print_inlining_output(NULL), 1008 _allowed_reasons(0), 1009 _interpreter_frame_size(0), 1010 _max_node_limit(MaxNodeLimit), 1011 _has_reserved_stack_access(false) { 1012 C = this; 1013 1014 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1015 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1016 1017 #ifndef PRODUCT 1018 set_print_assembly(PrintFrameConverterAssembly); 1019 set_parsed_irreducible_loop(false); 1020 #endif 1021 set_has_irreducible_loop(false); // no loops 1022 1023 CompileWrapper cw(this); 1024 Init(/*AliasLevel=*/ 0); 1025 init_tf((*generator)()); 1026 1027 { 1028 // The following is a dummy for the sake of GraphKit::gen_stub 1029 Unique_Node_List for_igvn(comp_arena()); 1030 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1031 PhaseGVN gvn(Thread::current()->resource_area(),255); 1032 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1033 gvn.transform_no_reclaim(top()); 1034 1035 GraphKit kit; 1036 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1037 } 1038 1039 NOT_PRODUCT( verify_graph_edges(); ) 1040 Code_Gen(); 1041 if (failing()) return; 1042 1043 1044 // Entry point will be accessed using compile->stub_entry_point(); 1045 if (code_buffer() == NULL) { 1046 Matcher::soft_match_failure(); 1047 } else { 1048 if (PrintAssembly && (WizardMode || Verbose)) 1049 tty->print_cr("### Stub::%s", stub_name); 1050 1051 if (!failing()) { 1052 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1053 1054 // Make the NMethod 1055 // For now we mark the frame as never safe for profile stackwalking 1056 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1057 code_buffer(), 1058 CodeOffsets::frame_never_safe, 1059 // _code_offsets.value(CodeOffsets::Frame_Complete), 1060 frame_size_in_words(), 1061 _oop_map_set, 1062 save_arg_registers); 1063 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1064 1065 _stub_entry_point = rs->entry_point(); 1066 } 1067 } 1068 } 1069 1070 //------------------------------Init------------------------------------------- 1071 // Prepare for a single compilation 1072 void Compile::Init(int aliaslevel) { 1073 _unique = 0; 1074 _regalloc = NULL; 1075 1076 _tf = NULL; // filled in later 1077 _top = NULL; // cached later 1078 _matcher = NULL; // filled in later 1079 _cfg = NULL; // filled in later 1080 1081 set_24_bit_selection_and_mode(Use24BitFP, false); 1082 1083 _node_note_array = NULL; 1084 _default_node_notes = NULL; 1085 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1086 1087 _immutable_memory = NULL; // filled in at first inquiry 1088 1089 // Globally visible Nodes 1090 // First set TOP to NULL to give safe behavior during creation of RootNode 1091 set_cached_top_node(NULL); 1092 set_root(new RootNode()); 1093 // Now that you have a Root to point to, create the real TOP 1094 set_cached_top_node( new ConNode(Type::TOP) ); 1095 set_recent_alloc(NULL, NULL); 1096 1097 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1098 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1099 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1100 env()->set_dependencies(new Dependencies(env())); 1101 1102 _fixed_slots = 0; 1103 set_has_split_ifs(false); 1104 set_has_loops(has_method() && method()->has_loops()); // first approximation 1105 set_has_stringbuilder(false); 1106 set_has_boxed_value(false); 1107 _trap_can_recompile = false; // no traps emitted yet 1108 _major_progress = true; // start out assuming good things will happen 1109 set_has_unsafe_access(false); 1110 set_max_vector_size(0); 1111 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1112 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1113 set_decompile_count(0); 1114 1115 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1116 set_num_loop_opts(LoopOptsCount); 1117 set_do_inlining(Inline); 1118 set_max_inline_size(MaxInlineSize); 1119 set_freq_inline_size(FreqInlineSize); 1120 set_do_scheduling(OptoScheduling); 1121 set_do_count_invocations(false); 1122 set_do_method_data_update(false); 1123 1124 set_do_vector_loop(false); 1125 1126 if (AllowVectorizeOnDemand) { 1127 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1128 set_do_vector_loop(true); 1129 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());}) 1130 } else if (has_method() && method()->name() != 0 && 1131 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1132 set_do_vector_loop(true); 1133 } 1134 } 1135 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1136 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());}) 1137 1138 set_age_code(has_method() && method()->profile_aging()); 1139 set_rtm_state(NoRTM); // No RTM lock eliding by default 1140 _max_node_limit = _directive->MaxNodeLimitOption; 1141 1142 #if INCLUDE_RTM_OPT 1143 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1144 int rtm_state = method()->method_data()->rtm_state(); 1145 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1146 // Don't generate RTM lock eliding code. 1147 set_rtm_state(NoRTM); 1148 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1149 // Generate RTM lock eliding code without abort ratio calculation code. 1150 set_rtm_state(UseRTM); 1151 } else if (UseRTMDeopt) { 1152 // Generate RTM lock eliding code and include abort ratio calculation 1153 // code if UseRTMDeopt is on. 1154 set_rtm_state(ProfileRTM); 1155 } 1156 } 1157 #endif 1158 if (debug_info()->recording_non_safepoints()) { 1159 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1160 (comp_arena(), 8, 0, NULL)); 1161 set_default_node_notes(Node_Notes::make(this)); 1162 } 1163 1164 // // -- Initialize types before each compile -- 1165 // // Update cached type information 1166 // if( _method && _method->constants() ) 1167 // Type::update_loaded_types(_method, _method->constants()); 1168 1169 // Init alias_type map. 1170 if (!_do_escape_analysis && aliaslevel == 3) 1171 aliaslevel = 2; // No unique types without escape analysis 1172 _AliasLevel = aliaslevel; 1173 const int grow_ats = 16; 1174 _max_alias_types = grow_ats; 1175 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1176 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1177 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1178 { 1179 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1180 } 1181 // Initialize the first few types. 1182 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1183 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1184 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1185 _num_alias_types = AliasIdxRaw+1; 1186 // Zero out the alias type cache. 1187 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1188 // A NULL adr_type hits in the cache right away. Preload the right answer. 1189 probe_alias_cache(NULL)->_index = AliasIdxTop; 1190 1191 _intrinsics = NULL; 1192 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1193 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1194 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1195 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1196 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1197 register_library_intrinsics(); 1198 } 1199 1200 //---------------------------init_start---------------------------------------- 1201 // Install the StartNode on this compile object. 1202 void Compile::init_start(StartNode* s) { 1203 if (failing()) 1204 return; // already failing 1205 assert(s == start(), ""); 1206 } 1207 1208 /** 1209 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1210 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1211 * the ideal graph. 1212 */ 1213 StartNode* Compile::start() const { 1214 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1215 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1216 Node* start = root()->fast_out(i); 1217 if (start->is_Start()) { 1218 return start->as_Start(); 1219 } 1220 } 1221 fatal("Did not find Start node!"); 1222 return NULL; 1223 } 1224 1225 //-------------------------------immutable_memory------------------------------------- 1226 // Access immutable memory 1227 Node* Compile::immutable_memory() { 1228 if (_immutable_memory != NULL) { 1229 return _immutable_memory; 1230 } 1231 StartNode* s = start(); 1232 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1233 Node *p = s->fast_out(i); 1234 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1235 _immutable_memory = p; 1236 return _immutable_memory; 1237 } 1238 } 1239 ShouldNotReachHere(); 1240 return NULL; 1241 } 1242 1243 //----------------------set_cached_top_node------------------------------------ 1244 // Install the cached top node, and make sure Node::is_top works correctly. 1245 void Compile::set_cached_top_node(Node* tn) { 1246 if (tn != NULL) verify_top(tn); 1247 Node* old_top = _top; 1248 _top = tn; 1249 // Calling Node::setup_is_top allows the nodes the chance to adjust 1250 // their _out arrays. 1251 if (_top != NULL) _top->setup_is_top(); 1252 if (old_top != NULL) old_top->setup_is_top(); 1253 assert(_top == NULL || top()->is_top(), ""); 1254 } 1255 1256 #ifdef ASSERT 1257 uint Compile::count_live_nodes_by_graph_walk() { 1258 Unique_Node_List useful(comp_arena()); 1259 // Get useful node list by walking the graph. 1260 identify_useful_nodes(useful); 1261 return useful.size(); 1262 } 1263 1264 void Compile::print_missing_nodes() { 1265 1266 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1267 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1268 return; 1269 } 1270 1271 // This is an expensive function. It is executed only when the user 1272 // specifies VerifyIdealNodeCount option or otherwise knows the 1273 // additional work that needs to be done to identify reachable nodes 1274 // by walking the flow graph and find the missing ones using 1275 // _dead_node_list. 1276 1277 Unique_Node_List useful(comp_arena()); 1278 // Get useful node list by walking the graph. 1279 identify_useful_nodes(useful); 1280 1281 uint l_nodes = C->live_nodes(); 1282 uint l_nodes_by_walk = useful.size(); 1283 1284 if (l_nodes != l_nodes_by_walk) { 1285 if (_log != NULL) { 1286 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1287 _log->stamp(); 1288 _log->end_head(); 1289 } 1290 VectorSet& useful_member_set = useful.member_set(); 1291 int last_idx = l_nodes_by_walk; 1292 for (int i = 0; i < last_idx; i++) { 1293 if (useful_member_set.test(i)) { 1294 if (_dead_node_list.test(i)) { 1295 if (_log != NULL) { 1296 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1297 } 1298 if (PrintIdealNodeCount) { 1299 // Print the log message to tty 1300 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1301 useful.at(i)->dump(); 1302 } 1303 } 1304 } 1305 else if (! _dead_node_list.test(i)) { 1306 if (_log != NULL) { 1307 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1308 } 1309 if (PrintIdealNodeCount) { 1310 // Print the log message to tty 1311 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1312 } 1313 } 1314 } 1315 if (_log != NULL) { 1316 _log->tail("mismatched_nodes"); 1317 } 1318 } 1319 } 1320 void Compile::record_modified_node(Node* n) { 1321 if (_modified_nodes != NULL && !_inlining_incrementally && 1322 n->outcnt() != 0 && !n->is_Con()) { 1323 _modified_nodes->push(n); 1324 } 1325 } 1326 1327 void Compile::remove_modified_node(Node* n) { 1328 if (_modified_nodes != NULL) { 1329 _modified_nodes->remove(n); 1330 } 1331 } 1332 #endif 1333 1334 #ifndef PRODUCT 1335 void Compile::verify_top(Node* tn) const { 1336 if (tn != NULL) { 1337 assert(tn->is_Con(), "top node must be a constant"); 1338 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1339 assert(tn->in(0) != NULL, "must have live top node"); 1340 } 1341 } 1342 #endif 1343 1344 1345 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1346 1347 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1348 guarantee(arr != NULL, ""); 1349 int num_blocks = arr->length(); 1350 if (grow_by < num_blocks) grow_by = num_blocks; 1351 int num_notes = grow_by * _node_notes_block_size; 1352 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1353 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1354 while (num_notes > 0) { 1355 arr->append(notes); 1356 notes += _node_notes_block_size; 1357 num_notes -= _node_notes_block_size; 1358 } 1359 assert(num_notes == 0, "exact multiple, please"); 1360 } 1361 1362 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1363 if (source == NULL || dest == NULL) return false; 1364 1365 if (dest->is_Con()) 1366 return false; // Do not push debug info onto constants. 1367 1368 #ifdef ASSERT 1369 // Leave a bread crumb trail pointing to the original node: 1370 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1371 dest->set_debug_orig(source); 1372 } 1373 #endif 1374 1375 if (node_note_array() == NULL) 1376 return false; // Not collecting any notes now. 1377 1378 // This is a copy onto a pre-existing node, which may already have notes. 1379 // If both nodes have notes, do not overwrite any pre-existing notes. 1380 Node_Notes* source_notes = node_notes_at(source->_idx); 1381 if (source_notes == NULL || source_notes->is_clear()) return false; 1382 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1383 if (dest_notes == NULL || dest_notes->is_clear()) { 1384 return set_node_notes_at(dest->_idx, source_notes); 1385 } 1386 1387 Node_Notes merged_notes = (*source_notes); 1388 // The order of operations here ensures that dest notes will win... 1389 merged_notes.update_from(dest_notes); 1390 return set_node_notes_at(dest->_idx, &merged_notes); 1391 } 1392 1393 1394 //--------------------------allow_range_check_smearing------------------------- 1395 // Gating condition for coalescing similar range checks. 1396 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1397 // single covering check that is at least as strong as any of them. 1398 // If the optimization succeeds, the simplified (strengthened) range check 1399 // will always succeed. If it fails, we will deopt, and then give up 1400 // on the optimization. 1401 bool Compile::allow_range_check_smearing() const { 1402 // If this method has already thrown a range-check, 1403 // assume it was because we already tried range smearing 1404 // and it failed. 1405 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1406 return !already_trapped; 1407 } 1408 1409 1410 //------------------------------flatten_alias_type----------------------------- 1411 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1412 int offset = tj->offset(); 1413 TypePtr::PTR ptr = tj->ptr(); 1414 1415 // Known instance (scalarizable allocation) alias only with itself. 1416 bool is_known_inst = tj->isa_oopptr() != NULL && 1417 tj->is_oopptr()->is_known_instance(); 1418 1419 // Process weird unsafe references. 1420 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1421 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1422 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1423 tj = TypeOopPtr::BOTTOM; 1424 ptr = tj->ptr(); 1425 offset = tj->offset(); 1426 } 1427 1428 // Array pointers need some flattening 1429 const TypeAryPtr *ta = tj->isa_aryptr(); 1430 if (ta && ta->is_stable()) { 1431 // Erase stability property for alias analysis. 1432 tj = ta = ta->cast_to_stable(false); 1433 } 1434 if( ta && is_known_inst ) { 1435 if ( offset != Type::OffsetBot && 1436 offset > arrayOopDesc::length_offset_in_bytes() ) { 1437 offset = Type::OffsetBot; // Flatten constant access into array body only 1438 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1439 } 1440 } else if( ta && _AliasLevel >= 2 ) { 1441 // For arrays indexed by constant indices, we flatten the alias 1442 // space to include all of the array body. Only the header, klass 1443 // and array length can be accessed un-aliased. 1444 if( offset != Type::OffsetBot ) { 1445 if( ta->const_oop() ) { // MethodData* or Method* 1446 offset = Type::OffsetBot; // Flatten constant access into array body 1447 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1448 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1449 // range is OK as-is. 1450 tj = ta = TypeAryPtr::RANGE; 1451 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1452 tj = TypeInstPtr::KLASS; // all klass loads look alike 1453 ta = TypeAryPtr::RANGE; // generic ignored junk 1454 ptr = TypePtr::BotPTR; 1455 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1456 tj = TypeInstPtr::MARK; 1457 ta = TypeAryPtr::RANGE; // generic ignored junk 1458 ptr = TypePtr::BotPTR; 1459 } else { // Random constant offset into array body 1460 offset = Type::OffsetBot; // Flatten constant access into array body 1461 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1462 } 1463 } 1464 // Arrays of fixed size alias with arrays of unknown size. 1465 if (ta->size() != TypeInt::POS) { 1466 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1467 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1468 } 1469 // Arrays of known objects become arrays of unknown objects. 1470 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1471 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1472 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1473 } 1474 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1475 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1476 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1477 } 1478 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1479 // cannot be distinguished by bytecode alone. 1480 if (ta->elem() == TypeInt::BOOL) { 1481 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1482 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1483 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1484 } 1485 // During the 2nd round of IterGVN, NotNull castings are removed. 1486 // Make sure the Bottom and NotNull variants alias the same. 1487 // Also, make sure exact and non-exact variants alias the same. 1488 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1489 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1490 } 1491 } 1492 1493 // Oop pointers need some flattening 1494 const TypeInstPtr *to = tj->isa_instptr(); 1495 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1496 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1497 if( ptr == TypePtr::Constant ) { 1498 if (to->klass() != ciEnv::current()->Class_klass() || 1499 offset < k->size_helper() * wordSize) { 1500 // No constant oop pointers (such as Strings); they alias with 1501 // unknown strings. 1502 assert(!is_known_inst, "not scalarizable allocation"); 1503 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1504 } 1505 } else if( is_known_inst ) { 1506 tj = to; // Keep NotNull and klass_is_exact for instance type 1507 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1508 // During the 2nd round of IterGVN, NotNull castings are removed. 1509 // Make sure the Bottom and NotNull variants alias the same. 1510 // Also, make sure exact and non-exact variants alias the same. 1511 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1512 } 1513 if (to->speculative() != NULL) { 1514 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1515 } 1516 // Canonicalize the holder of this field 1517 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1518 // First handle header references such as a LoadKlassNode, even if the 1519 // object's klass is unloaded at compile time (4965979). 1520 if (!is_known_inst) { // Do it only for non-instance types 1521 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1522 } 1523 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1524 // Static fields are in the space above the normal instance 1525 // fields in the java.lang.Class instance. 1526 if (to->klass() != ciEnv::current()->Class_klass()) { 1527 to = NULL; 1528 tj = TypeOopPtr::BOTTOM; 1529 offset = tj->offset(); 1530 } 1531 } else { 1532 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1533 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1534 if( is_known_inst ) { 1535 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1536 } else { 1537 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1538 } 1539 } 1540 } 1541 } 1542 1543 // Klass pointers to object array klasses need some flattening 1544 const TypeKlassPtr *tk = tj->isa_klassptr(); 1545 if( tk ) { 1546 // If we are referencing a field within a Klass, we need 1547 // to assume the worst case of an Object. Both exact and 1548 // inexact types must flatten to the same alias class so 1549 // use NotNull as the PTR. 1550 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1551 1552 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1553 TypeKlassPtr::OBJECT->klass(), 1554 offset); 1555 } 1556 1557 ciKlass* klass = tk->klass(); 1558 if( klass->is_obj_array_klass() ) { 1559 ciKlass* k = TypeAryPtr::OOPS->klass(); 1560 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1561 k = TypeInstPtr::BOTTOM->klass(); 1562 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1563 } 1564 1565 // Check for precise loads from the primary supertype array and force them 1566 // to the supertype cache alias index. Check for generic array loads from 1567 // the primary supertype array and also force them to the supertype cache 1568 // alias index. Since the same load can reach both, we need to merge 1569 // these 2 disparate memories into the same alias class. Since the 1570 // primary supertype array is read-only, there's no chance of confusion 1571 // where we bypass an array load and an array store. 1572 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1573 if (offset == Type::OffsetBot || 1574 (offset >= primary_supers_offset && 1575 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1576 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1577 offset = in_bytes(Klass::secondary_super_cache_offset()); 1578 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1579 } 1580 } 1581 1582 // Flatten all Raw pointers together. 1583 if (tj->base() == Type::RawPtr) 1584 tj = TypeRawPtr::BOTTOM; 1585 1586 if (tj->base() == Type::AnyPtr) 1587 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1588 1589 // Flatten all to bottom for now 1590 switch( _AliasLevel ) { 1591 case 0: 1592 tj = TypePtr::BOTTOM; 1593 break; 1594 case 1: // Flatten to: oop, static, field or array 1595 switch (tj->base()) { 1596 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1597 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1598 case Type::AryPtr: // do not distinguish arrays at all 1599 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1600 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1601 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1602 default: ShouldNotReachHere(); 1603 } 1604 break; 1605 case 2: // No collapsing at level 2; keep all splits 1606 case 3: // No collapsing at level 3; keep all splits 1607 break; 1608 default: 1609 Unimplemented(); 1610 } 1611 1612 offset = tj->offset(); 1613 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1614 1615 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1616 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1617 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1618 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1619 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1620 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1621 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1622 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1623 assert( tj->ptr() != TypePtr::TopPTR && 1624 tj->ptr() != TypePtr::AnyNull && 1625 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1626 // assert( tj->ptr() != TypePtr::Constant || 1627 // tj->base() == Type::RawPtr || 1628 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1629 1630 return tj; 1631 } 1632 1633 void Compile::AliasType::Init(int i, const TypePtr* at) { 1634 _index = i; 1635 _adr_type = at; 1636 _field = NULL; 1637 _element = NULL; 1638 _is_rewritable = true; // default 1639 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1640 if (atoop != NULL && atoop->is_known_instance()) { 1641 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1642 _general_index = Compile::current()->get_alias_index(gt); 1643 } else { 1644 _general_index = 0; 1645 } 1646 } 1647 1648 BasicType Compile::AliasType::basic_type() const { 1649 if (element() != NULL) { 1650 const Type* element = adr_type()->is_aryptr()->elem(); 1651 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1652 } if (field() != NULL) { 1653 return field()->layout_type(); 1654 } else { 1655 return T_ILLEGAL; // unknown 1656 } 1657 } 1658 1659 //---------------------------------print_on------------------------------------ 1660 #ifndef PRODUCT 1661 void Compile::AliasType::print_on(outputStream* st) { 1662 if (index() < 10) 1663 st->print("@ <%d> ", index()); 1664 else st->print("@ <%d>", index()); 1665 st->print(is_rewritable() ? " " : " RO"); 1666 int offset = adr_type()->offset(); 1667 if (offset == Type::OffsetBot) 1668 st->print(" +any"); 1669 else st->print(" +%-3d", offset); 1670 st->print(" in "); 1671 adr_type()->dump_on(st); 1672 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1673 if (field() != NULL && tjp) { 1674 if (tjp->klass() != field()->holder() || 1675 tjp->offset() != field()->offset_in_bytes()) { 1676 st->print(" != "); 1677 field()->print(); 1678 st->print(" ***"); 1679 } 1680 } 1681 } 1682 1683 void print_alias_types() { 1684 Compile* C = Compile::current(); 1685 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1686 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1687 C->alias_type(idx)->print_on(tty); 1688 tty->cr(); 1689 } 1690 } 1691 #endif 1692 1693 1694 //----------------------------probe_alias_cache-------------------------------- 1695 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1696 intptr_t key = (intptr_t) adr_type; 1697 key ^= key >> logAliasCacheSize; 1698 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1699 } 1700 1701 1702 //-----------------------------grow_alias_types-------------------------------- 1703 void Compile::grow_alias_types() { 1704 const int old_ats = _max_alias_types; // how many before? 1705 const int new_ats = old_ats; // how many more? 1706 const int grow_ats = old_ats+new_ats; // how many now? 1707 _max_alias_types = grow_ats; 1708 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1709 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1710 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1711 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1712 } 1713 1714 1715 //--------------------------------find_alias_type------------------------------ 1716 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1717 if (_AliasLevel == 0) 1718 return alias_type(AliasIdxBot); 1719 1720 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1721 if (ace->_adr_type == adr_type) { 1722 return alias_type(ace->_index); 1723 } 1724 1725 // Handle special cases. 1726 if (adr_type == NULL) return alias_type(AliasIdxTop); 1727 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1728 1729 // Do it the slow way. 1730 const TypePtr* flat = flatten_alias_type(adr_type); 1731 1732 #ifdef ASSERT 1733 { 1734 ResourceMark rm; 1735 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1736 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1737 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1738 Type::str(adr_type)); 1739 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1740 const TypeOopPtr* foop = flat->is_oopptr(); 1741 // Scalarizable allocations have exact klass always. 1742 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1743 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1744 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1745 Type::str(foop), Type::str(xoop)); 1746 } 1747 } 1748 #endif 1749 1750 int idx = AliasIdxTop; 1751 for (int i = 0; i < num_alias_types(); i++) { 1752 if (alias_type(i)->adr_type() == flat) { 1753 idx = i; 1754 break; 1755 } 1756 } 1757 1758 if (idx == AliasIdxTop) { 1759 if (no_create) return NULL; 1760 // Grow the array if necessary. 1761 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1762 // Add a new alias type. 1763 idx = _num_alias_types++; 1764 _alias_types[idx]->Init(idx, flat); 1765 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1766 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1767 if (flat->isa_instptr()) { 1768 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1769 && flat->is_instptr()->klass() == env()->Class_klass()) 1770 alias_type(idx)->set_rewritable(false); 1771 } 1772 if (flat->isa_aryptr()) { 1773 #ifdef ASSERT 1774 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1775 // (T_BYTE has the weakest alignment and size restrictions...) 1776 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1777 #endif 1778 if (flat->offset() == TypePtr::OffsetBot) { 1779 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1780 } 1781 } 1782 if (flat->isa_klassptr()) { 1783 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1784 alias_type(idx)->set_rewritable(false); 1785 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1786 alias_type(idx)->set_rewritable(false); 1787 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1788 alias_type(idx)->set_rewritable(false); 1789 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1790 alias_type(idx)->set_rewritable(false); 1791 } 1792 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1793 // but the base pointer type is not distinctive enough to identify 1794 // references into JavaThread.) 1795 1796 // Check for final fields. 1797 const TypeInstPtr* tinst = flat->isa_instptr(); 1798 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1799 ciField* field; 1800 if (tinst->const_oop() != NULL && 1801 tinst->klass() == ciEnv::current()->Class_klass() && 1802 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1803 // static field 1804 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1805 field = k->get_field_by_offset(tinst->offset(), true); 1806 } else { 1807 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1808 field = k->get_field_by_offset(tinst->offset(), false); 1809 } 1810 assert(field == NULL || 1811 original_field == NULL || 1812 (field->holder() == original_field->holder() && 1813 field->offset() == original_field->offset() && 1814 field->is_static() == original_field->is_static()), "wrong field?"); 1815 // Set field() and is_rewritable() attributes. 1816 if (field != NULL) alias_type(idx)->set_field(field); 1817 } 1818 } 1819 1820 // Fill the cache for next time. 1821 ace->_adr_type = adr_type; 1822 ace->_index = idx; 1823 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1824 1825 // Might as well try to fill the cache for the flattened version, too. 1826 AliasCacheEntry* face = probe_alias_cache(flat); 1827 if (face->_adr_type == NULL) { 1828 face->_adr_type = flat; 1829 face->_index = idx; 1830 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1831 } 1832 1833 return alias_type(idx); 1834 } 1835 1836 1837 Compile::AliasType* Compile::alias_type(ciField* field) { 1838 const TypeOopPtr* t; 1839 if (field->is_static()) 1840 t = TypeInstPtr::make(field->holder()->java_mirror()); 1841 else 1842 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1843 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1844 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1845 return atp; 1846 } 1847 1848 1849 //------------------------------have_alias_type-------------------------------- 1850 bool Compile::have_alias_type(const TypePtr* adr_type) { 1851 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1852 if (ace->_adr_type == adr_type) { 1853 return true; 1854 } 1855 1856 // Handle special cases. 1857 if (adr_type == NULL) return true; 1858 if (adr_type == TypePtr::BOTTOM) return true; 1859 1860 return find_alias_type(adr_type, true, NULL) != NULL; 1861 } 1862 1863 //-----------------------------must_alias-------------------------------------- 1864 // True if all values of the given address type are in the given alias category. 1865 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1866 if (alias_idx == AliasIdxBot) return true; // the universal category 1867 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1868 if (alias_idx == AliasIdxTop) return false; // the empty category 1869 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1870 1871 // the only remaining possible overlap is identity 1872 int adr_idx = get_alias_index(adr_type); 1873 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1874 assert(adr_idx == alias_idx || 1875 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1876 && adr_type != TypeOopPtr::BOTTOM), 1877 "should not be testing for overlap with an unsafe pointer"); 1878 return adr_idx == alias_idx; 1879 } 1880 1881 //------------------------------can_alias-------------------------------------- 1882 // True if any values of the given address type are in the given alias category. 1883 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1884 if (alias_idx == AliasIdxTop) return false; // the empty category 1885 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1886 if (alias_idx == AliasIdxBot) return true; // the universal category 1887 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1888 1889 // the only remaining possible overlap is identity 1890 int adr_idx = get_alias_index(adr_type); 1891 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1892 return adr_idx == alias_idx; 1893 } 1894 1895 1896 1897 //---------------------------pop_warm_call------------------------------------- 1898 WarmCallInfo* Compile::pop_warm_call() { 1899 WarmCallInfo* wci = _warm_calls; 1900 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1901 return wci; 1902 } 1903 1904 //----------------------------Inline_Warm-------------------------------------- 1905 int Compile::Inline_Warm() { 1906 // If there is room, try to inline some more warm call sites. 1907 // %%% Do a graph index compaction pass when we think we're out of space? 1908 if (!InlineWarmCalls) return 0; 1909 1910 int calls_made_hot = 0; 1911 int room_to_grow = NodeCountInliningCutoff - unique(); 1912 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1913 int amount_grown = 0; 1914 WarmCallInfo* call; 1915 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1916 int est_size = (int)call->size(); 1917 if (est_size > (room_to_grow - amount_grown)) { 1918 // This one won't fit anyway. Get rid of it. 1919 call->make_cold(); 1920 continue; 1921 } 1922 call->make_hot(); 1923 calls_made_hot++; 1924 amount_grown += est_size; 1925 amount_to_grow -= est_size; 1926 } 1927 1928 if (calls_made_hot > 0) set_major_progress(); 1929 return calls_made_hot; 1930 } 1931 1932 1933 //----------------------------Finish_Warm-------------------------------------- 1934 void Compile::Finish_Warm() { 1935 if (!InlineWarmCalls) return; 1936 if (failing()) return; 1937 if (warm_calls() == NULL) return; 1938 1939 // Clean up loose ends, if we are out of space for inlining. 1940 WarmCallInfo* call; 1941 while ((call = pop_warm_call()) != NULL) { 1942 call->make_cold(); 1943 } 1944 } 1945 1946 //---------------------cleanup_loop_predicates----------------------- 1947 // Remove the opaque nodes that protect the predicates so that all unused 1948 // checks and uncommon_traps will be eliminated from the ideal graph 1949 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1950 if (predicate_count()==0) return; 1951 for (int i = predicate_count(); i > 0; i--) { 1952 Node * n = predicate_opaque1_node(i-1); 1953 assert(n->Opcode() == Op_Opaque1, "must be"); 1954 igvn.replace_node(n, n->in(1)); 1955 } 1956 assert(predicate_count()==0, "should be clean!"); 1957 } 1958 1959 void Compile::add_range_check_cast(Node* n) { 1960 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1961 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1962 _range_check_casts->append(n); 1963 } 1964 1965 // Remove all range check dependent CastIINodes. 1966 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1967 for (int i = range_check_cast_count(); i > 0; i--) { 1968 Node* cast = range_check_cast_node(i-1); 1969 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1970 igvn.replace_node(cast, cast->in(1)); 1971 } 1972 assert(range_check_cast_count() == 0, "should be empty"); 1973 } 1974 1975 void Compile::add_opaque4_node(Node* n) { 1976 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 1977 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 1978 _opaque4_nodes->append(n); 1979 } 1980 1981 // Remove all Opaque4 nodes. 1982 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 1983 for (int i = opaque4_count(); i > 0; i--) { 1984 Node* opaq = opaque4_node(i-1); 1985 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 1986 igvn.replace_node(opaq, opaq->in(2)); 1987 } 1988 assert(opaque4_count() == 0, "should be empty"); 1989 } 1990 1991 // StringOpts and late inlining of string methods 1992 void Compile::inline_string_calls(bool parse_time) { 1993 { 1994 // remove useless nodes to make the usage analysis simpler 1995 ResourceMark rm; 1996 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1997 } 1998 1999 { 2000 ResourceMark rm; 2001 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2002 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2003 print_method(PHASE_AFTER_STRINGOPTS, 3); 2004 } 2005 2006 // now inline anything that we skipped the first time around 2007 if (!parse_time) { 2008 _late_inlines_pos = _late_inlines.length(); 2009 } 2010 2011 while (_string_late_inlines.length() > 0) { 2012 CallGenerator* cg = _string_late_inlines.pop(); 2013 cg->do_late_inline(); 2014 if (failing()) return; 2015 } 2016 _string_late_inlines.trunc_to(0); 2017 } 2018 2019 // Late inlining of boxing methods 2020 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2021 if (_boxing_late_inlines.length() > 0) { 2022 assert(has_boxed_value(), "inconsistent"); 2023 2024 PhaseGVN* gvn = initial_gvn(); 2025 set_inlining_incrementally(true); 2026 2027 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2028 for_igvn()->clear(); 2029 gvn->replace_with(&igvn); 2030 2031 _late_inlines_pos = _late_inlines.length(); 2032 2033 while (_boxing_late_inlines.length() > 0) { 2034 CallGenerator* cg = _boxing_late_inlines.pop(); 2035 cg->do_late_inline(); 2036 if (failing()) return; 2037 } 2038 _boxing_late_inlines.trunc_to(0); 2039 2040 { 2041 ResourceMark rm; 2042 PhaseRemoveUseless pru(gvn, for_igvn()); 2043 } 2044 2045 igvn = PhaseIterGVN(gvn); 2046 igvn.optimize(); 2047 2048 set_inlining_progress(false); 2049 set_inlining_incrementally(false); 2050 } 2051 } 2052 2053 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 2054 assert(IncrementalInline, "incremental inlining should be on"); 2055 PhaseGVN* gvn = initial_gvn(); 2056 2057 set_inlining_progress(false); 2058 for_igvn()->clear(); 2059 gvn->replace_with(&igvn); 2060 2061 { 2062 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2063 int i = 0; 2064 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2065 CallGenerator* cg = _late_inlines.at(i); 2066 _late_inlines_pos = i+1; 2067 cg->do_late_inline(); 2068 if (failing()) return; 2069 } 2070 int j = 0; 2071 for (; i < _late_inlines.length(); i++, j++) { 2072 _late_inlines.at_put(j, _late_inlines.at(i)); 2073 } 2074 _late_inlines.trunc_to(j); 2075 } 2076 2077 { 2078 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2079 ResourceMark rm; 2080 PhaseRemoveUseless pru(gvn, for_igvn()); 2081 } 2082 2083 { 2084 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2085 igvn = PhaseIterGVN(gvn); 2086 } 2087 } 2088 2089 // Perform incremental inlining until bound on number of live nodes is reached 2090 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2091 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2092 2093 PhaseGVN* gvn = initial_gvn(); 2094 2095 set_inlining_incrementally(true); 2096 set_inlining_progress(true); 2097 uint low_live_nodes = 0; 2098 2099 while(inlining_progress() && _late_inlines.length() > 0) { 2100 2101 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2102 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2103 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2104 // PhaseIdealLoop is expensive so we only try it once we are 2105 // out of live nodes and we only try it again if the previous 2106 // helped got the number of nodes down significantly 2107 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone); 2108 if (failing()) return; 2109 low_live_nodes = live_nodes(); 2110 _major_progress = true; 2111 } 2112 2113 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2114 break; 2115 } 2116 } 2117 2118 inline_incrementally_one(igvn); 2119 2120 if (failing()) return; 2121 2122 { 2123 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2124 igvn.optimize(); 2125 } 2126 2127 if (failing()) return; 2128 } 2129 2130 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2131 2132 if (_string_late_inlines.length() > 0) { 2133 assert(has_stringbuilder(), "inconsistent"); 2134 for_igvn()->clear(); 2135 initial_gvn()->replace_with(&igvn); 2136 2137 inline_string_calls(false); 2138 2139 if (failing()) return; 2140 2141 { 2142 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2143 ResourceMark rm; 2144 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2145 } 2146 2147 { 2148 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2149 igvn = PhaseIterGVN(gvn); 2150 igvn.optimize(); 2151 } 2152 } 2153 2154 set_inlining_incrementally(false); 2155 } 2156 2157 2158 bool Compile::optimize_loops(int& loop_opts_cnt, PhaseIterGVN& igvn, LoopOptsMode mode) { 2159 if(loop_opts_cnt > 0) { 2160 debug_only( int cnt = 0; ); 2161 while(major_progress() && (loop_opts_cnt > 0)) { 2162 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2163 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2164 PhaseIdealLoop ideal_loop(igvn, mode); 2165 loop_opts_cnt--; 2166 if (failing()) return false; 2167 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2168 } 2169 } 2170 return true; 2171 } 2172 2173 //------------------------------Optimize--------------------------------------- 2174 // Given a graph, optimize it. 2175 void Compile::Optimize() { 2176 TracePhase tp("optimizer", &timers[_t_optimizer]); 2177 2178 #ifndef PRODUCT 2179 if (_directive->BreakAtCompileOption) { 2180 BREAKPOINT; 2181 } 2182 2183 #endif 2184 2185 #ifdef ASSERT 2186 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2187 bs->verify_gc_barriers(true); 2188 #endif 2189 2190 ResourceMark rm; 2191 int loop_opts_cnt; 2192 2193 print_inlining_reinit(); 2194 2195 NOT_PRODUCT( verify_graph_edges(); ) 2196 2197 print_method(PHASE_AFTER_PARSING); 2198 2199 { 2200 // Iterative Global Value Numbering, including ideal transforms 2201 // Initialize IterGVN with types and values from parse-time GVN 2202 PhaseIterGVN igvn(initial_gvn()); 2203 #ifdef ASSERT 2204 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2205 #endif 2206 { 2207 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2208 igvn.optimize(); 2209 } 2210 2211 if (failing()) return; 2212 2213 print_method(PHASE_ITER_GVN1, 2); 2214 2215 inline_incrementally(igvn); 2216 2217 print_method(PHASE_INCREMENTAL_INLINE, 2); 2218 2219 if (failing()) return; 2220 2221 if (eliminate_boxing()) { 2222 // Inline valueOf() methods now. 2223 inline_boxing_calls(igvn); 2224 2225 if (AlwaysIncrementalInline) { 2226 inline_incrementally(igvn); 2227 } 2228 2229 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2230 2231 if (failing()) return; 2232 } 2233 2234 // Remove the speculative part of types and clean up the graph from 2235 // the extra CastPP nodes whose only purpose is to carry them. Do 2236 // that early so that optimizations are not disrupted by the extra 2237 // CastPP nodes. 2238 remove_speculative_types(igvn); 2239 2240 // No more new expensive nodes will be added to the list from here 2241 // so keep only the actual candidates for optimizations. 2242 cleanup_expensive_nodes(igvn); 2243 2244 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2245 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2246 initial_gvn()->replace_with(&igvn); 2247 for_igvn()->clear(); 2248 Unique_Node_List new_worklist(C->comp_arena()); 2249 { 2250 ResourceMark rm; 2251 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2252 } 2253 set_for_igvn(&new_worklist); 2254 igvn = PhaseIterGVN(initial_gvn()); 2255 igvn.optimize(); 2256 } 2257 2258 // Perform escape analysis 2259 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2260 if (has_loops()) { 2261 // Cleanup graph (remove dead nodes). 2262 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2263 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone); 2264 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2265 if (failing()) return; 2266 } 2267 ConnectionGraph::do_analysis(this, &igvn); 2268 2269 if (failing()) return; 2270 2271 // Optimize out fields loads from scalar replaceable allocations. 2272 igvn.optimize(); 2273 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2274 2275 if (failing()) return; 2276 2277 if (congraph() != NULL && macro_count() > 0) { 2278 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2279 PhaseMacroExpand mexp(igvn); 2280 mexp.eliminate_macro_nodes(); 2281 igvn.set_delay_transform(false); 2282 2283 igvn.optimize(); 2284 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2285 2286 if (failing()) return; 2287 } 2288 } 2289 2290 // Loop transforms on the ideal graph. Range Check Elimination, 2291 // peeling, unrolling, etc. 2292 2293 // Set loop opts counter 2294 loop_opts_cnt = num_loop_opts(); 2295 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2296 { 2297 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2298 PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault); 2299 loop_opts_cnt--; 2300 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2301 if (failing()) return; 2302 } 2303 // Loop opts pass if partial peeling occurred in previous pass 2304 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2305 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2306 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2307 loop_opts_cnt--; 2308 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2309 if (failing()) return; 2310 } 2311 // Loop opts pass for loop-unrolling before CCP 2312 if(major_progress() && (loop_opts_cnt > 0)) { 2313 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2314 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2315 loop_opts_cnt--; 2316 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2317 } 2318 if (!failing()) { 2319 // Verify that last round of loop opts produced a valid graph 2320 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2321 PhaseIdealLoop::verify(igvn); 2322 } 2323 } 2324 if (failing()) return; 2325 2326 // Conditional Constant Propagation; 2327 PhaseCCP ccp( &igvn ); 2328 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2329 { 2330 TracePhase tp("ccp", &timers[_t_ccp]); 2331 ccp.do_transform(); 2332 } 2333 print_method(PHASE_CPP1, 2); 2334 2335 assert( true, "Break here to ccp.dump_old2new_map()"); 2336 2337 // Iterative Global Value Numbering, including ideal transforms 2338 { 2339 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2340 igvn = ccp; 2341 igvn.optimize(); 2342 } 2343 2344 print_method(PHASE_ITER_GVN2, 2); 2345 2346 if (failing()) return; 2347 2348 // Loop transforms on the ideal graph. Range Check Elimination, 2349 // peeling, unrolling, etc. 2350 if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) { 2351 return; 2352 } 2353 2354 #if INCLUDE_ZGC 2355 if (UseZGC) { 2356 ZBarrierSetC2::find_dominating_barriers(igvn); 2357 } 2358 #endif 2359 2360 if (failing()) return; 2361 2362 // Ensure that major progress is now clear 2363 C->clear_major_progress(); 2364 2365 { 2366 // Verify that all previous optimizations produced a valid graph 2367 // at least to this point, even if no loop optimizations were done. 2368 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2369 PhaseIdealLoop::verify(igvn); 2370 } 2371 2372 if (range_check_cast_count() > 0) { 2373 // No more loop optimizations. Remove all range check dependent CastIINodes. 2374 C->remove_range_check_casts(igvn); 2375 igvn.optimize(); 2376 } 2377 2378 #ifdef ASSERT 2379 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2380 bs->verify_gc_barriers(false); 2381 #endif 2382 2383 { 2384 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2385 PhaseMacroExpand mex(igvn); 2386 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2); 2387 if (mex.expand_macro_nodes()) { 2388 assert(failing(), "must bail out w/ explicit message"); 2389 return; 2390 } 2391 } 2392 2393 if (opaque4_count() > 0) { 2394 C->remove_opaque4_nodes(igvn); 2395 igvn.optimize(); 2396 } 2397 2398 DEBUG_ONLY( _modified_nodes = NULL; ) 2399 } // (End scope of igvn; run destructor if necessary for asserts.) 2400 2401 process_print_inlining(); 2402 // A method with only infinite loops has no edges entering loops from root 2403 { 2404 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2405 if (final_graph_reshaping()) { 2406 assert(failing(), "must bail out w/ explicit message"); 2407 return; 2408 } 2409 } 2410 2411 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2412 } 2413 2414 2415 //------------------------------Code_Gen--------------------------------------- 2416 // Given a graph, generate code for it 2417 void Compile::Code_Gen() { 2418 if (failing()) { 2419 return; 2420 } 2421 2422 // Perform instruction selection. You might think we could reclaim Matcher 2423 // memory PDQ, but actually the Matcher is used in generating spill code. 2424 // Internals of the Matcher (including some VectorSets) must remain live 2425 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2426 // set a bit in reclaimed memory. 2427 2428 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2429 // nodes. Mapping is only valid at the root of each matched subtree. 2430 NOT_PRODUCT( verify_graph_edges(); ) 2431 2432 Matcher matcher; 2433 _matcher = &matcher; 2434 { 2435 TracePhase tp("matcher", &timers[_t_matcher]); 2436 matcher.match(); 2437 } 2438 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2439 // nodes. Mapping is only valid at the root of each matched subtree. 2440 NOT_PRODUCT( verify_graph_edges(); ) 2441 2442 // If you have too many nodes, or if matching has failed, bail out 2443 check_node_count(0, "out of nodes matching instructions"); 2444 if (failing()) { 2445 return; 2446 } 2447 2448 print_method(PHASE_MATCHING, 2); 2449 2450 // Build a proper-looking CFG 2451 PhaseCFG cfg(node_arena(), root(), matcher); 2452 _cfg = &cfg; 2453 { 2454 TracePhase tp("scheduler", &timers[_t_scheduler]); 2455 bool success = cfg.do_global_code_motion(); 2456 if (!success) { 2457 return; 2458 } 2459 2460 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2461 NOT_PRODUCT( verify_graph_edges(); ) 2462 debug_only( cfg.verify(); ) 2463 } 2464 2465 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2466 _regalloc = ®alloc; 2467 { 2468 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2469 // Perform register allocation. After Chaitin, use-def chains are 2470 // no longer accurate (at spill code) and so must be ignored. 2471 // Node->LRG->reg mappings are still accurate. 2472 _regalloc->Register_Allocate(); 2473 2474 // Bail out if the allocator builds too many nodes 2475 if (failing()) { 2476 return; 2477 } 2478 } 2479 2480 // Prior to register allocation we kept empty basic blocks in case the 2481 // the allocator needed a place to spill. After register allocation we 2482 // are not adding any new instructions. If any basic block is empty, we 2483 // can now safely remove it. 2484 { 2485 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2486 cfg.remove_empty_blocks(); 2487 if (do_freq_based_layout()) { 2488 PhaseBlockLayout layout(cfg); 2489 } else { 2490 cfg.set_loop_alignment(); 2491 } 2492 cfg.fixup_flow(); 2493 } 2494 2495 // Apply peephole optimizations 2496 if( OptoPeephole ) { 2497 TracePhase tp("peephole", &timers[_t_peephole]); 2498 PhasePeephole peep( _regalloc, cfg); 2499 peep.do_transform(); 2500 } 2501 2502 // Do late expand if CPU requires this. 2503 if (Matcher::require_postalloc_expand) { 2504 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2505 cfg.postalloc_expand(_regalloc); 2506 } 2507 2508 // Convert Nodes to instruction bits in a buffer 2509 { 2510 TraceTime tp("output", &timers[_t_output], CITime); 2511 Output(); 2512 } 2513 2514 print_method(PHASE_FINAL_CODE); 2515 2516 // He's dead, Jim. 2517 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2518 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2519 } 2520 2521 2522 //------------------------------dump_asm--------------------------------------- 2523 // Dump formatted assembly 2524 #ifndef PRODUCT 2525 void Compile::dump_asm(int *pcs, uint pc_limit) { 2526 bool cut_short = false; 2527 tty->print_cr("#"); 2528 tty->print("# "); _tf->dump(); tty->cr(); 2529 tty->print_cr("#"); 2530 2531 // For all blocks 2532 int pc = 0x0; // Program counter 2533 char starts_bundle = ' '; 2534 _regalloc->dump_frame(); 2535 2536 Node *n = NULL; 2537 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2538 if (VMThread::should_terminate()) { 2539 cut_short = true; 2540 break; 2541 } 2542 Block* block = _cfg->get_block(i); 2543 if (block->is_connector() && !Verbose) { 2544 continue; 2545 } 2546 n = block->head(); 2547 if (pcs && n->_idx < pc_limit) { 2548 tty->print("%3.3x ", pcs[n->_idx]); 2549 } else { 2550 tty->print(" "); 2551 } 2552 block->dump_head(_cfg); 2553 if (block->is_connector()) { 2554 tty->print_cr(" # Empty connector block"); 2555 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2556 tty->print_cr(" # Block is sole successor of call"); 2557 } 2558 2559 // For all instructions 2560 Node *delay = NULL; 2561 for (uint j = 0; j < block->number_of_nodes(); j++) { 2562 if (VMThread::should_terminate()) { 2563 cut_short = true; 2564 break; 2565 } 2566 n = block->get_node(j); 2567 if (valid_bundle_info(n)) { 2568 Bundle* bundle = node_bundling(n); 2569 if (bundle->used_in_unconditional_delay()) { 2570 delay = n; 2571 continue; 2572 } 2573 if (bundle->starts_bundle()) { 2574 starts_bundle = '+'; 2575 } 2576 } 2577 2578 if (WizardMode) { 2579 n->dump(); 2580 } 2581 2582 if( !n->is_Region() && // Dont print in the Assembly 2583 !n->is_Phi() && // a few noisely useless nodes 2584 !n->is_Proj() && 2585 !n->is_MachTemp() && 2586 !n->is_SafePointScalarObject() && 2587 !n->is_Catch() && // Would be nice to print exception table targets 2588 !n->is_MergeMem() && // Not very interesting 2589 !n->is_top() && // Debug info table constants 2590 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2591 ) { 2592 if (pcs && n->_idx < pc_limit) 2593 tty->print("%3.3x", pcs[n->_idx]); 2594 else 2595 tty->print(" "); 2596 tty->print(" %c ", starts_bundle); 2597 starts_bundle = ' '; 2598 tty->print("\t"); 2599 n->format(_regalloc, tty); 2600 tty->cr(); 2601 } 2602 2603 // If we have an instruction with a delay slot, and have seen a delay, 2604 // then back up and print it 2605 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2606 assert(delay != NULL, "no unconditional delay instruction"); 2607 if (WizardMode) delay->dump(); 2608 2609 if (node_bundling(delay)->starts_bundle()) 2610 starts_bundle = '+'; 2611 if (pcs && n->_idx < pc_limit) 2612 tty->print("%3.3x", pcs[n->_idx]); 2613 else 2614 tty->print(" "); 2615 tty->print(" %c ", starts_bundle); 2616 starts_bundle = ' '; 2617 tty->print("\t"); 2618 delay->format(_regalloc, tty); 2619 tty->cr(); 2620 delay = NULL; 2621 } 2622 2623 // Dump the exception table as well 2624 if( n->is_Catch() && (Verbose || WizardMode) ) { 2625 // Print the exception table for this offset 2626 _handler_table.print_subtable_for(pc); 2627 } 2628 } 2629 2630 if (pcs && n->_idx < pc_limit) 2631 tty->print_cr("%3.3x", pcs[n->_idx]); 2632 else 2633 tty->cr(); 2634 2635 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2636 2637 } // End of per-block dump 2638 tty->cr(); 2639 2640 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2641 } 2642 #endif 2643 2644 //------------------------------Final_Reshape_Counts--------------------------- 2645 // This class defines counters to help identify when a method 2646 // may/must be executed using hardware with only 24-bit precision. 2647 struct Final_Reshape_Counts : public StackObj { 2648 int _call_count; // count non-inlined 'common' calls 2649 int _float_count; // count float ops requiring 24-bit precision 2650 int _double_count; // count double ops requiring more precision 2651 int _java_call_count; // count non-inlined 'java' calls 2652 int _inner_loop_count; // count loops which need alignment 2653 VectorSet _visited; // Visitation flags 2654 Node_List _tests; // Set of IfNodes & PCTableNodes 2655 2656 Final_Reshape_Counts() : 2657 _call_count(0), _float_count(0), _double_count(0), 2658 _java_call_count(0), _inner_loop_count(0), 2659 _visited( Thread::current()->resource_area() ) { } 2660 2661 void inc_call_count () { _call_count ++; } 2662 void inc_float_count () { _float_count ++; } 2663 void inc_double_count() { _double_count++; } 2664 void inc_java_call_count() { _java_call_count++; } 2665 void inc_inner_loop_count() { _inner_loop_count++; } 2666 2667 int get_call_count () const { return _call_count ; } 2668 int get_float_count () const { return _float_count ; } 2669 int get_double_count() const { return _double_count; } 2670 int get_java_call_count() const { return _java_call_count; } 2671 int get_inner_loop_count() const { return _inner_loop_count; } 2672 }; 2673 2674 #ifdef ASSERT 2675 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2676 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2677 // Make sure the offset goes inside the instance layout. 2678 return k->contains_field_offset(tp->offset()); 2679 // Note that OffsetBot and OffsetTop are very negative. 2680 } 2681 #endif 2682 2683 // Eliminate trivially redundant StoreCMs and accumulate their 2684 // precedence edges. 2685 void Compile::eliminate_redundant_card_marks(Node* n) { 2686 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2687 if (n->in(MemNode::Address)->outcnt() > 1) { 2688 // There are multiple users of the same address so it might be 2689 // possible to eliminate some of the StoreCMs 2690 Node* mem = n->in(MemNode::Memory); 2691 Node* adr = n->in(MemNode::Address); 2692 Node* val = n->in(MemNode::ValueIn); 2693 Node* prev = n; 2694 bool done = false; 2695 // Walk the chain of StoreCMs eliminating ones that match. As 2696 // long as it's a chain of single users then the optimization is 2697 // safe. Eliminating partially redundant StoreCMs would require 2698 // cloning copies down the other paths. 2699 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2700 if (adr == mem->in(MemNode::Address) && 2701 val == mem->in(MemNode::ValueIn)) { 2702 // redundant StoreCM 2703 if (mem->req() > MemNode::OopStore) { 2704 // Hasn't been processed by this code yet. 2705 n->add_prec(mem->in(MemNode::OopStore)); 2706 } else { 2707 // Already converted to precedence edge 2708 for (uint i = mem->req(); i < mem->len(); i++) { 2709 // Accumulate any precedence edges 2710 if (mem->in(i) != NULL) { 2711 n->add_prec(mem->in(i)); 2712 } 2713 } 2714 // Everything above this point has been processed. 2715 done = true; 2716 } 2717 // Eliminate the previous StoreCM 2718 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2719 assert(mem->outcnt() == 0, "should be dead"); 2720 mem->disconnect_inputs(NULL, this); 2721 } else { 2722 prev = mem; 2723 } 2724 mem = prev->in(MemNode::Memory); 2725 } 2726 } 2727 } 2728 2729 //------------------------------final_graph_reshaping_impl---------------------- 2730 // Implement items 1-5 from final_graph_reshaping below. 2731 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2732 2733 if ( n->outcnt() == 0 ) return; // dead node 2734 uint nop = n->Opcode(); 2735 2736 // Check for 2-input instruction with "last use" on right input. 2737 // Swap to left input. Implements item (2). 2738 if( n->req() == 3 && // two-input instruction 2739 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2740 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2741 n->in(2)->outcnt() == 1 &&// right use IS a last use 2742 !n->in(2)->is_Con() ) { // right use is not a constant 2743 // Check for commutative opcode 2744 switch( nop ) { 2745 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2746 case Op_MaxI: case Op_MinI: 2747 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2748 case Op_AndL: case Op_XorL: case Op_OrL: 2749 case Op_AndI: case Op_XorI: case Op_OrI: { 2750 // Move "last use" input to left by swapping inputs 2751 n->swap_edges(1, 2); 2752 break; 2753 } 2754 default: 2755 break; 2756 } 2757 } 2758 2759 #ifdef ASSERT 2760 if( n->is_Mem() ) { 2761 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2762 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2763 // oop will be recorded in oop map if load crosses safepoint 2764 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2765 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2766 "raw memory operations should have control edge"); 2767 } 2768 if (n->is_MemBar()) { 2769 MemBarNode* mb = n->as_MemBar(); 2770 if (mb->trailing_store() || mb->trailing_load_store()) { 2771 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2772 Node* mem = mb->in(MemBarNode::Precedent); 2773 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2774 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2775 } else if (mb->leading()) { 2776 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2777 } 2778 } 2779 #endif 2780 // Count FPU ops and common calls, implements item (3) 2781 switch( nop ) { 2782 // Count all float operations that may use FPU 2783 case Op_AddF: 2784 case Op_SubF: 2785 case Op_MulF: 2786 case Op_DivF: 2787 case Op_NegF: 2788 case Op_ModF: 2789 case Op_ConvI2F: 2790 case Op_ConF: 2791 case Op_CmpF: 2792 case Op_CmpF3: 2793 // case Op_ConvL2F: // longs are split into 32-bit halves 2794 frc.inc_float_count(); 2795 break; 2796 2797 case Op_ConvF2D: 2798 case Op_ConvD2F: 2799 frc.inc_float_count(); 2800 frc.inc_double_count(); 2801 break; 2802 2803 // Count all double operations that may use FPU 2804 case Op_AddD: 2805 case Op_SubD: 2806 case Op_MulD: 2807 case Op_DivD: 2808 case Op_NegD: 2809 case Op_ModD: 2810 case Op_ConvI2D: 2811 case Op_ConvD2I: 2812 // case Op_ConvL2D: // handled by leaf call 2813 // case Op_ConvD2L: // handled by leaf call 2814 case Op_ConD: 2815 case Op_CmpD: 2816 case Op_CmpD3: 2817 frc.inc_double_count(); 2818 break; 2819 case Op_Opaque1: // Remove Opaque Nodes before matching 2820 case Op_Opaque2: // Remove Opaque Nodes before matching 2821 case Op_Opaque3: 2822 n->subsume_by(n->in(1), this); 2823 break; 2824 case Op_CallStaticJava: 2825 case Op_CallJava: 2826 case Op_CallDynamicJava: 2827 frc.inc_java_call_count(); // Count java call site; 2828 case Op_CallRuntime: 2829 case Op_CallLeaf: 2830 case Op_CallLeafNoFP: { 2831 assert (n->is_Call(), ""); 2832 CallNode *call = n->as_Call(); 2833 // Count call sites where the FP mode bit would have to be flipped. 2834 // Do not count uncommon runtime calls: 2835 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2836 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2837 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 2838 frc.inc_call_count(); // Count the call site 2839 } else { // See if uncommon argument is shared 2840 Node *n = call->in(TypeFunc::Parms); 2841 int nop = n->Opcode(); 2842 // Clone shared simple arguments to uncommon calls, item (1). 2843 if (n->outcnt() > 1 && 2844 !n->is_Proj() && 2845 nop != Op_CreateEx && 2846 nop != Op_CheckCastPP && 2847 nop != Op_DecodeN && 2848 nop != Op_DecodeNKlass && 2849 !n->is_Mem() && 2850 !n->is_Phi()) { 2851 Node *x = n->clone(); 2852 call->set_req(TypeFunc::Parms, x); 2853 } 2854 } 2855 break; 2856 } 2857 2858 case Op_StoreD: 2859 case Op_LoadD: 2860 case Op_LoadD_unaligned: 2861 frc.inc_double_count(); 2862 goto handle_mem; 2863 case Op_StoreF: 2864 case Op_LoadF: 2865 frc.inc_float_count(); 2866 goto handle_mem; 2867 2868 case Op_StoreCM: 2869 { 2870 // Convert OopStore dependence into precedence edge 2871 Node* prec = n->in(MemNode::OopStore); 2872 n->del_req(MemNode::OopStore); 2873 n->add_prec(prec); 2874 eliminate_redundant_card_marks(n); 2875 } 2876 2877 // fall through 2878 2879 case Op_StoreB: 2880 case Op_StoreC: 2881 case Op_StorePConditional: 2882 case Op_StoreI: 2883 case Op_StoreL: 2884 case Op_StoreIConditional: 2885 case Op_StoreLConditional: 2886 case Op_CompareAndSwapB: 2887 case Op_CompareAndSwapS: 2888 case Op_CompareAndSwapI: 2889 case Op_CompareAndSwapL: 2890 case Op_CompareAndSwapP: 2891 case Op_CompareAndSwapN: 2892 case Op_WeakCompareAndSwapB: 2893 case Op_WeakCompareAndSwapS: 2894 case Op_WeakCompareAndSwapI: 2895 case Op_WeakCompareAndSwapL: 2896 case Op_WeakCompareAndSwapP: 2897 case Op_WeakCompareAndSwapN: 2898 case Op_CompareAndExchangeB: 2899 case Op_CompareAndExchangeS: 2900 case Op_CompareAndExchangeI: 2901 case Op_CompareAndExchangeL: 2902 case Op_CompareAndExchangeP: 2903 case Op_CompareAndExchangeN: 2904 case Op_GetAndAddS: 2905 case Op_GetAndAddB: 2906 case Op_GetAndAddI: 2907 case Op_GetAndAddL: 2908 case Op_GetAndSetS: 2909 case Op_GetAndSetB: 2910 case Op_GetAndSetI: 2911 case Op_GetAndSetL: 2912 case Op_GetAndSetP: 2913 case Op_GetAndSetN: 2914 case Op_StoreP: 2915 case Op_StoreN: 2916 case Op_StoreNKlass: 2917 case Op_LoadB: 2918 case Op_LoadUB: 2919 case Op_LoadUS: 2920 case Op_LoadI: 2921 case Op_LoadKlass: 2922 case Op_LoadNKlass: 2923 case Op_LoadL: 2924 case Op_LoadL_unaligned: 2925 case Op_LoadPLocked: 2926 case Op_LoadP: 2927 #if INCLUDE_ZGC 2928 case Op_LoadBarrierSlowReg: 2929 case Op_LoadBarrierWeakSlowReg: 2930 #endif 2931 case Op_LoadN: 2932 case Op_LoadRange: 2933 case Op_LoadS: { 2934 handle_mem: 2935 #ifdef ASSERT 2936 if( VerifyOptoOopOffsets ) { 2937 assert( n->is_Mem(), "" ); 2938 MemNode *mem = (MemNode*)n; 2939 // Check to see if address types have grounded out somehow. 2940 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2941 assert( !tp || oop_offset_is_sane(tp), "" ); 2942 } 2943 #endif 2944 break; 2945 } 2946 2947 case Op_AddP: { // Assert sane base pointers 2948 Node *addp = n->in(AddPNode::Address); 2949 assert( !addp->is_AddP() || 2950 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2951 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2952 "Base pointers must match (addp %u)", addp->_idx ); 2953 #ifdef _LP64 2954 if ((UseCompressedOops || UseCompressedClassPointers) && 2955 addp->Opcode() == Op_ConP && 2956 addp == n->in(AddPNode::Base) && 2957 n->in(AddPNode::Offset)->is_Con()) { 2958 // If the transformation of ConP to ConN+DecodeN is beneficial depends 2959 // on the platform and on the compressed oops mode. 2960 // Use addressing with narrow klass to load with offset on x86. 2961 // Some platforms can use the constant pool to load ConP. 2962 // Do this transformation here since IGVN will convert ConN back to ConP. 2963 const Type* t = addp->bottom_type(); 2964 bool is_oop = t->isa_oopptr() != NULL; 2965 bool is_klass = t->isa_klassptr() != NULL; 2966 2967 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 2968 (is_klass && Matcher::const_klass_prefer_decode())) { 2969 Node* nn = NULL; 2970 2971 int op = is_oop ? Op_ConN : Op_ConNKlass; 2972 2973 // Look for existing ConN node of the same exact type. 2974 Node* r = root(); 2975 uint cnt = r->outcnt(); 2976 for (uint i = 0; i < cnt; i++) { 2977 Node* m = r->raw_out(i); 2978 if (m!= NULL && m->Opcode() == op && 2979 m->bottom_type()->make_ptr() == t) { 2980 nn = m; 2981 break; 2982 } 2983 } 2984 if (nn != NULL) { 2985 // Decode a narrow oop to match address 2986 // [R12 + narrow_oop_reg<<3 + offset] 2987 if (is_oop) { 2988 nn = new DecodeNNode(nn, t); 2989 } else { 2990 nn = new DecodeNKlassNode(nn, t); 2991 } 2992 // Check for succeeding AddP which uses the same Base. 2993 // Otherwise we will run into the assertion above when visiting that guy. 2994 for (uint i = 0; i < n->outcnt(); ++i) { 2995 Node *out_i = n->raw_out(i); 2996 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 2997 out_i->set_req(AddPNode::Base, nn); 2998 #ifdef ASSERT 2999 for (uint j = 0; j < out_i->outcnt(); ++j) { 3000 Node *out_j = out_i->raw_out(j); 3001 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3002 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3003 } 3004 #endif 3005 } 3006 } 3007 n->set_req(AddPNode::Base, nn); 3008 n->set_req(AddPNode::Address, nn); 3009 if (addp->outcnt() == 0) { 3010 addp->disconnect_inputs(NULL, this); 3011 } 3012 } 3013 } 3014 } 3015 #endif 3016 // platform dependent reshaping of the address expression 3017 reshape_address(n->as_AddP()); 3018 break; 3019 } 3020 3021 case Op_CastPP: { 3022 // Remove CastPP nodes to gain more freedom during scheduling but 3023 // keep the dependency they encode as control or precedence edges 3024 // (if control is set already) on memory operations. Some CastPP 3025 // nodes don't have a control (don't carry a dependency): skip 3026 // those. 3027 if (n->in(0) != NULL) { 3028 ResourceMark rm; 3029 Unique_Node_List wq; 3030 wq.push(n); 3031 for (uint next = 0; next < wq.size(); ++next) { 3032 Node *m = wq.at(next); 3033 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3034 Node* use = m->fast_out(i); 3035 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3036 use->ensure_control_or_add_prec(n->in(0)); 3037 } else { 3038 switch(use->Opcode()) { 3039 case Op_AddP: 3040 case Op_DecodeN: 3041 case Op_DecodeNKlass: 3042 case Op_CheckCastPP: 3043 case Op_CastPP: 3044 wq.push(use); 3045 break; 3046 } 3047 } 3048 } 3049 } 3050 } 3051 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3052 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3053 Node* in1 = n->in(1); 3054 const Type* t = n->bottom_type(); 3055 Node* new_in1 = in1->clone(); 3056 new_in1->as_DecodeN()->set_type(t); 3057 3058 if (!Matcher::narrow_oop_use_complex_address()) { 3059 // 3060 // x86, ARM and friends can handle 2 adds in addressing mode 3061 // and Matcher can fold a DecodeN node into address by using 3062 // a narrow oop directly and do implicit NULL check in address: 3063 // 3064 // [R12 + narrow_oop_reg<<3 + offset] 3065 // NullCheck narrow_oop_reg 3066 // 3067 // On other platforms (Sparc) we have to keep new DecodeN node and 3068 // use it to do implicit NULL check in address: 3069 // 3070 // decode_not_null narrow_oop_reg, base_reg 3071 // [base_reg + offset] 3072 // NullCheck base_reg 3073 // 3074 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3075 // to keep the information to which NULL check the new DecodeN node 3076 // corresponds to use it as value in implicit_null_check(). 3077 // 3078 new_in1->set_req(0, n->in(0)); 3079 } 3080 3081 n->subsume_by(new_in1, this); 3082 if (in1->outcnt() == 0) { 3083 in1->disconnect_inputs(NULL, this); 3084 } 3085 } else { 3086 n->subsume_by(n->in(1), this); 3087 if (n->outcnt() == 0) { 3088 n->disconnect_inputs(NULL, this); 3089 } 3090 } 3091 break; 3092 } 3093 #ifdef _LP64 3094 case Op_CmpP: 3095 // Do this transformation here to preserve CmpPNode::sub() and 3096 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3097 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3098 Node* in1 = n->in(1); 3099 Node* in2 = n->in(2); 3100 if (!in1->is_DecodeNarrowPtr()) { 3101 in2 = in1; 3102 in1 = n->in(2); 3103 } 3104 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3105 3106 Node* new_in2 = NULL; 3107 if (in2->is_DecodeNarrowPtr()) { 3108 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3109 new_in2 = in2->in(1); 3110 } else if (in2->Opcode() == Op_ConP) { 3111 const Type* t = in2->bottom_type(); 3112 if (t == TypePtr::NULL_PTR) { 3113 assert(in1->is_DecodeN(), "compare klass to null?"); 3114 // Don't convert CmpP null check into CmpN if compressed 3115 // oops implicit null check is not generated. 3116 // This will allow to generate normal oop implicit null check. 3117 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3118 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3119 // 3120 // This transformation together with CastPP transformation above 3121 // will generated code for implicit NULL checks for compressed oops. 3122 // 3123 // The original code after Optimize() 3124 // 3125 // LoadN memory, narrow_oop_reg 3126 // decode narrow_oop_reg, base_reg 3127 // CmpP base_reg, NULL 3128 // CastPP base_reg // NotNull 3129 // Load [base_reg + offset], val_reg 3130 // 3131 // after these transformations will be 3132 // 3133 // LoadN memory, narrow_oop_reg 3134 // CmpN narrow_oop_reg, NULL 3135 // decode_not_null narrow_oop_reg, base_reg 3136 // Load [base_reg + offset], val_reg 3137 // 3138 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3139 // since narrow oops can be used in debug info now (see the code in 3140 // final_graph_reshaping_walk()). 3141 // 3142 // At the end the code will be matched to 3143 // on x86: 3144 // 3145 // Load_narrow_oop memory, narrow_oop_reg 3146 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3147 // NullCheck narrow_oop_reg 3148 // 3149 // and on sparc: 3150 // 3151 // Load_narrow_oop memory, narrow_oop_reg 3152 // decode_not_null narrow_oop_reg, base_reg 3153 // Load [base_reg + offset], val_reg 3154 // NullCheck base_reg 3155 // 3156 } else if (t->isa_oopptr()) { 3157 new_in2 = ConNode::make(t->make_narrowoop()); 3158 } else if (t->isa_klassptr()) { 3159 new_in2 = ConNode::make(t->make_narrowklass()); 3160 } 3161 } 3162 if (new_in2 != NULL) { 3163 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3164 n->subsume_by(cmpN, this); 3165 if (in1->outcnt() == 0) { 3166 in1->disconnect_inputs(NULL, this); 3167 } 3168 if (in2->outcnt() == 0) { 3169 in2->disconnect_inputs(NULL, this); 3170 } 3171 } 3172 } 3173 break; 3174 3175 case Op_DecodeN: 3176 case Op_DecodeNKlass: 3177 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3178 // DecodeN could be pinned when it can't be fold into 3179 // an address expression, see the code for Op_CastPP above. 3180 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3181 break; 3182 3183 case Op_EncodeP: 3184 case Op_EncodePKlass: { 3185 Node* in1 = n->in(1); 3186 if (in1->is_DecodeNarrowPtr()) { 3187 n->subsume_by(in1->in(1), this); 3188 } else if (in1->Opcode() == Op_ConP) { 3189 const Type* t = in1->bottom_type(); 3190 if (t == TypePtr::NULL_PTR) { 3191 assert(t->isa_oopptr(), "null klass?"); 3192 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3193 } else if (t->isa_oopptr()) { 3194 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3195 } else if (t->isa_klassptr()) { 3196 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3197 } 3198 } 3199 if (in1->outcnt() == 0) { 3200 in1->disconnect_inputs(NULL, this); 3201 } 3202 break; 3203 } 3204 3205 case Op_Proj: { 3206 if (OptimizeStringConcat) { 3207 ProjNode* p = n->as_Proj(); 3208 if (p->_is_io_use) { 3209 // Separate projections were used for the exception path which 3210 // are normally removed by a late inline. If it wasn't inlined 3211 // then they will hang around and should just be replaced with 3212 // the original one. 3213 Node* proj = NULL; 3214 // Replace with just one 3215 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3216 Node *use = i.get(); 3217 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3218 proj = use; 3219 break; 3220 } 3221 } 3222 assert(proj != NULL, "must be found"); 3223 p->subsume_by(proj, this); 3224 } 3225 } 3226 break; 3227 } 3228 3229 case Op_Phi: 3230 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3231 // The EncodeP optimization may create Phi with the same edges 3232 // for all paths. It is not handled well by Register Allocator. 3233 Node* unique_in = n->in(1); 3234 assert(unique_in != NULL, ""); 3235 uint cnt = n->req(); 3236 for (uint i = 2; i < cnt; i++) { 3237 Node* m = n->in(i); 3238 assert(m != NULL, ""); 3239 if (unique_in != m) 3240 unique_in = NULL; 3241 } 3242 if (unique_in != NULL) { 3243 n->subsume_by(unique_in, this); 3244 } 3245 } 3246 break; 3247 3248 #endif 3249 3250 #ifdef ASSERT 3251 case Op_CastII: 3252 // Verify that all range check dependent CastII nodes were removed. 3253 if (n->isa_CastII()->has_range_check()) { 3254 n->dump(3); 3255 assert(false, "Range check dependent CastII node was not removed"); 3256 } 3257 break; 3258 #endif 3259 3260 case Op_ModI: 3261 if (UseDivMod) { 3262 // Check if a%b and a/b both exist 3263 Node* d = n->find_similar(Op_DivI); 3264 if (d) { 3265 // Replace them with a fused divmod if supported 3266 if (Matcher::has_match_rule(Op_DivModI)) { 3267 DivModINode* divmod = DivModINode::make(n); 3268 d->subsume_by(divmod->div_proj(), this); 3269 n->subsume_by(divmod->mod_proj(), this); 3270 } else { 3271 // replace a%b with a-((a/b)*b) 3272 Node* mult = new MulINode(d, d->in(2)); 3273 Node* sub = new SubINode(d->in(1), mult); 3274 n->subsume_by(sub, this); 3275 } 3276 } 3277 } 3278 break; 3279 3280 case Op_ModL: 3281 if (UseDivMod) { 3282 // Check if a%b and a/b both exist 3283 Node* d = n->find_similar(Op_DivL); 3284 if (d) { 3285 // Replace them with a fused divmod if supported 3286 if (Matcher::has_match_rule(Op_DivModL)) { 3287 DivModLNode* divmod = DivModLNode::make(n); 3288 d->subsume_by(divmod->div_proj(), this); 3289 n->subsume_by(divmod->mod_proj(), this); 3290 } else { 3291 // replace a%b with a-((a/b)*b) 3292 Node* mult = new MulLNode(d, d->in(2)); 3293 Node* sub = new SubLNode(d->in(1), mult); 3294 n->subsume_by(sub, this); 3295 } 3296 } 3297 } 3298 break; 3299 3300 case Op_LoadVector: 3301 case Op_StoreVector: 3302 break; 3303 3304 case Op_AddReductionVI: 3305 case Op_AddReductionVL: 3306 case Op_AddReductionVF: 3307 case Op_AddReductionVD: 3308 case Op_MulReductionVI: 3309 case Op_MulReductionVL: 3310 case Op_MulReductionVF: 3311 case Op_MulReductionVD: 3312 break; 3313 3314 case Op_PackB: 3315 case Op_PackS: 3316 case Op_PackI: 3317 case Op_PackF: 3318 case Op_PackL: 3319 case Op_PackD: 3320 if (n->req()-1 > 2) { 3321 // Replace many operand PackNodes with a binary tree for matching 3322 PackNode* p = (PackNode*) n; 3323 Node* btp = p->binary_tree_pack(1, n->req()); 3324 n->subsume_by(btp, this); 3325 } 3326 break; 3327 case Op_Loop: 3328 case Op_CountedLoop: 3329 case Op_OuterStripMinedLoop: 3330 if (n->as_Loop()->is_inner_loop()) { 3331 frc.inc_inner_loop_count(); 3332 } 3333 n->as_Loop()->verify_strip_mined(0); 3334 break; 3335 case Op_LShiftI: 3336 case Op_RShiftI: 3337 case Op_URShiftI: 3338 case Op_LShiftL: 3339 case Op_RShiftL: 3340 case Op_URShiftL: 3341 if (Matcher::need_masked_shift_count) { 3342 // The cpu's shift instructions don't restrict the count to the 3343 // lower 5/6 bits. We need to do the masking ourselves. 3344 Node* in2 = n->in(2); 3345 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3346 const TypeInt* t = in2->find_int_type(); 3347 if (t != NULL && t->is_con()) { 3348 juint shift = t->get_con(); 3349 if (shift > mask) { // Unsigned cmp 3350 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3351 } 3352 } else { 3353 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3354 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3355 n->set_req(2, shift); 3356 } 3357 } 3358 if (in2->outcnt() == 0) { // Remove dead node 3359 in2->disconnect_inputs(NULL, this); 3360 } 3361 } 3362 break; 3363 case Op_MemBarStoreStore: 3364 case Op_MemBarRelease: 3365 // Break the link with AllocateNode: it is no longer useful and 3366 // confuses register allocation. 3367 if (n->req() > MemBarNode::Precedent) { 3368 n->set_req(MemBarNode::Precedent, top()); 3369 } 3370 break; 3371 case Op_MemBarAcquire: { 3372 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3373 // At parse time, the trailing MemBarAcquire for a volatile load 3374 // is created with an edge to the load. After optimizations, 3375 // that input may be a chain of Phis. If those phis have no 3376 // other use, then the MemBarAcquire keeps them alive and 3377 // register allocation can be confused. 3378 ResourceMark rm; 3379 Unique_Node_List wq; 3380 wq.push(n->in(MemBarNode::Precedent)); 3381 n->set_req(MemBarNode::Precedent, top()); 3382 while (wq.size() > 0) { 3383 Node* m = wq.pop(); 3384 if (m->outcnt() == 0) { 3385 for (uint j = 0; j < m->req(); j++) { 3386 Node* in = m->in(j); 3387 if (in != NULL) { 3388 wq.push(in); 3389 } 3390 } 3391 m->disconnect_inputs(NULL, this); 3392 } 3393 } 3394 } 3395 break; 3396 } 3397 case Op_RangeCheck: { 3398 RangeCheckNode* rc = n->as_RangeCheck(); 3399 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3400 n->subsume_by(iff, this); 3401 frc._tests.push(iff); 3402 break; 3403 } 3404 case Op_ConvI2L: { 3405 if (!Matcher::convi2l_type_required) { 3406 // Code generation on some platforms doesn't need accurate 3407 // ConvI2L types. Widening the type can help remove redundant 3408 // address computations. 3409 n->as_Type()->set_type(TypeLong::INT); 3410 ResourceMark rm; 3411 Unique_Node_List wq; 3412 wq.push(n); 3413 for (uint next = 0; next < wq.size(); next++) { 3414 Node *m = wq.at(next); 3415 3416 for(;;) { 3417 // Loop over all nodes with identical inputs edges as m 3418 Node* k = m->find_similar(m->Opcode()); 3419 if (k == NULL) { 3420 break; 3421 } 3422 // Push their uses so we get a chance to remove node made 3423 // redundant 3424 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3425 Node* u = k->fast_out(i); 3426 if (u->Opcode() == Op_LShiftL || 3427 u->Opcode() == Op_AddL || 3428 u->Opcode() == Op_SubL || 3429 u->Opcode() == Op_AddP) { 3430 wq.push(u); 3431 } 3432 } 3433 // Replace all nodes with identical edges as m with m 3434 k->subsume_by(m, this); 3435 } 3436 } 3437 } 3438 break; 3439 } 3440 case Op_CmpUL: { 3441 if (!Matcher::has_match_rule(Op_CmpUL)) { 3442 // No support for unsigned long comparisons 3443 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3444 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3445 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3446 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3447 Node* andl = new AndLNode(orl, remove_sign_mask); 3448 Node* cmp = new CmpLNode(andl, n->in(2)); 3449 n->subsume_by(cmp, this); 3450 } 3451 break; 3452 } 3453 default: 3454 assert( !n->is_Call(), "" ); 3455 assert( !n->is_Mem(), "" ); 3456 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3457 break; 3458 } 3459 3460 // Collect CFG split points 3461 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3462 frc._tests.push(n); 3463 } 3464 } 3465 3466 //------------------------------final_graph_reshaping_walk--------------------- 3467 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3468 // requires that the walk visits a node's inputs before visiting the node. 3469 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3470 ResourceArea *area = Thread::current()->resource_area(); 3471 Unique_Node_List sfpt(area); 3472 3473 frc._visited.set(root->_idx); // first, mark node as visited 3474 uint cnt = root->req(); 3475 Node *n = root; 3476 uint i = 0; 3477 while (true) { 3478 if (i < cnt) { 3479 // Place all non-visited non-null inputs onto stack 3480 Node* m = n->in(i); 3481 ++i; 3482 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3483 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3484 // compute worst case interpreter size in case of a deoptimization 3485 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3486 3487 sfpt.push(m); 3488 } 3489 cnt = m->req(); 3490 nstack.push(n, i); // put on stack parent and next input's index 3491 n = m; 3492 i = 0; 3493 } 3494 } else { 3495 // Now do post-visit work 3496 final_graph_reshaping_impl( n, frc ); 3497 if (nstack.is_empty()) 3498 break; // finished 3499 n = nstack.node(); // Get node from stack 3500 cnt = n->req(); 3501 i = nstack.index(); 3502 nstack.pop(); // Shift to the next node on stack 3503 } 3504 } 3505 3506 // Skip next transformation if compressed oops are not used. 3507 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3508 (!UseCompressedOops && !UseCompressedClassPointers)) 3509 return; 3510 3511 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3512 // It could be done for an uncommon traps or any safepoints/calls 3513 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3514 while (sfpt.size() > 0) { 3515 n = sfpt.pop(); 3516 JVMState *jvms = n->as_SafePoint()->jvms(); 3517 assert(jvms != NULL, "sanity"); 3518 int start = jvms->debug_start(); 3519 int end = n->req(); 3520 bool is_uncommon = (n->is_CallStaticJava() && 3521 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3522 for (int j = start; j < end; j++) { 3523 Node* in = n->in(j); 3524 if (in->is_DecodeNarrowPtr()) { 3525 bool safe_to_skip = true; 3526 if (!is_uncommon ) { 3527 // Is it safe to skip? 3528 for (uint i = 0; i < in->outcnt(); i++) { 3529 Node* u = in->raw_out(i); 3530 if (!u->is_SafePoint() || 3531 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3532 safe_to_skip = false; 3533 } 3534 } 3535 } 3536 if (safe_to_skip) { 3537 n->set_req(j, in->in(1)); 3538 } 3539 if (in->outcnt() == 0) { 3540 in->disconnect_inputs(NULL, this); 3541 } 3542 } 3543 } 3544 } 3545 } 3546 3547 //------------------------------final_graph_reshaping-------------------------- 3548 // Final Graph Reshaping. 3549 // 3550 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3551 // and not commoned up and forced early. Must come after regular 3552 // optimizations to avoid GVN undoing the cloning. Clone constant 3553 // inputs to Loop Phis; these will be split by the allocator anyways. 3554 // Remove Opaque nodes. 3555 // (2) Move last-uses by commutative operations to the left input to encourage 3556 // Intel update-in-place two-address operations and better register usage 3557 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3558 // calls canonicalizing them back. 3559 // (3) Count the number of double-precision FP ops, single-precision FP ops 3560 // and call sites. On Intel, we can get correct rounding either by 3561 // forcing singles to memory (requires extra stores and loads after each 3562 // FP bytecode) or we can set a rounding mode bit (requires setting and 3563 // clearing the mode bit around call sites). The mode bit is only used 3564 // if the relative frequency of single FP ops to calls is low enough. 3565 // This is a key transform for SPEC mpeg_audio. 3566 // (4) Detect infinite loops; blobs of code reachable from above but not 3567 // below. Several of the Code_Gen algorithms fail on such code shapes, 3568 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3569 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3570 // Detection is by looking for IfNodes where only 1 projection is 3571 // reachable from below or CatchNodes missing some targets. 3572 // (5) Assert for insane oop offsets in debug mode. 3573 3574 bool Compile::final_graph_reshaping() { 3575 // an infinite loop may have been eliminated by the optimizer, 3576 // in which case the graph will be empty. 3577 if (root()->req() == 1) { 3578 record_method_not_compilable("trivial infinite loop"); 3579 return true; 3580 } 3581 3582 // Expensive nodes have their control input set to prevent the GVN 3583 // from freely commoning them. There's no GVN beyond this point so 3584 // no need to keep the control input. We want the expensive nodes to 3585 // be freely moved to the least frequent code path by gcm. 3586 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3587 for (int i = 0; i < expensive_count(); i++) { 3588 _expensive_nodes->at(i)->set_req(0, NULL); 3589 } 3590 3591 Final_Reshape_Counts frc; 3592 3593 // Visit everybody reachable! 3594 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3595 Node_Stack nstack(live_nodes() >> 1); 3596 final_graph_reshaping_walk(nstack, root(), frc); 3597 3598 // Check for unreachable (from below) code (i.e., infinite loops). 3599 for( uint i = 0; i < frc._tests.size(); i++ ) { 3600 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3601 // Get number of CFG targets. 3602 // Note that PCTables include exception targets after calls. 3603 uint required_outcnt = n->required_outcnt(); 3604 if (n->outcnt() != required_outcnt) { 3605 // Check for a few special cases. Rethrow Nodes never take the 3606 // 'fall-thru' path, so expected kids is 1 less. 3607 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3608 if (n->in(0)->in(0)->is_Call()) { 3609 CallNode *call = n->in(0)->in(0)->as_Call(); 3610 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3611 required_outcnt--; // Rethrow always has 1 less kid 3612 } else if (call->req() > TypeFunc::Parms && 3613 call->is_CallDynamicJava()) { 3614 // Check for null receiver. In such case, the optimizer has 3615 // detected that the virtual call will always result in a null 3616 // pointer exception. The fall-through projection of this CatchNode 3617 // will not be populated. 3618 Node *arg0 = call->in(TypeFunc::Parms); 3619 if (arg0->is_Type() && 3620 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3621 required_outcnt--; 3622 } 3623 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3624 call->req() > TypeFunc::Parms+1 && 3625 call->is_CallStaticJava()) { 3626 // Check for negative array length. In such case, the optimizer has 3627 // detected that the allocation attempt will always result in an 3628 // exception. There is no fall-through projection of this CatchNode . 3629 Node *arg1 = call->in(TypeFunc::Parms+1); 3630 if (arg1->is_Type() && 3631 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3632 required_outcnt--; 3633 } 3634 } 3635 } 3636 } 3637 // Recheck with a better notion of 'required_outcnt' 3638 if (n->outcnt() != required_outcnt) { 3639 record_method_not_compilable("malformed control flow"); 3640 return true; // Not all targets reachable! 3641 } 3642 } 3643 // Check that I actually visited all kids. Unreached kids 3644 // must be infinite loops. 3645 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3646 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3647 record_method_not_compilable("infinite loop"); 3648 return true; // Found unvisited kid; must be unreach 3649 } 3650 3651 // Here so verification code in final_graph_reshaping_walk() 3652 // always see an OuterStripMinedLoopEnd 3653 if (n->is_OuterStripMinedLoopEnd()) { 3654 IfNode* init_iff = n->as_If(); 3655 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3656 n->subsume_by(iff, this); 3657 } 3658 } 3659 3660 // If original bytecodes contained a mixture of floats and doubles 3661 // check if the optimizer has made it homogenous, item (3). 3662 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3663 frc.get_float_count() > 32 && 3664 frc.get_double_count() == 0 && 3665 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3666 set_24_bit_selection_and_mode( false, true ); 3667 } 3668 3669 set_java_calls(frc.get_java_call_count()); 3670 set_inner_loops(frc.get_inner_loop_count()); 3671 3672 // No infinite loops, no reason to bail out. 3673 return false; 3674 } 3675 3676 //-----------------------------too_many_traps---------------------------------- 3677 // Report if there are too many traps at the current method and bci. 3678 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3679 bool Compile::too_many_traps(ciMethod* method, 3680 int bci, 3681 Deoptimization::DeoptReason reason) { 3682 ciMethodData* md = method->method_data(); 3683 if (md->is_empty()) { 3684 // Assume the trap has not occurred, or that it occurred only 3685 // because of a transient condition during start-up in the interpreter. 3686 return false; 3687 } 3688 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3689 if (md->has_trap_at(bci, m, reason) != 0) { 3690 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3691 // Also, if there are multiple reasons, or if there is no per-BCI record, 3692 // assume the worst. 3693 if (log()) 3694 log()->elem("observe trap='%s' count='%d'", 3695 Deoptimization::trap_reason_name(reason), 3696 md->trap_count(reason)); 3697 return true; 3698 } else { 3699 // Ignore method/bci and see if there have been too many globally. 3700 return too_many_traps(reason, md); 3701 } 3702 } 3703 3704 // Less-accurate variant which does not require a method and bci. 3705 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3706 ciMethodData* logmd) { 3707 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3708 // Too many traps globally. 3709 // Note that we use cumulative trap_count, not just md->trap_count. 3710 if (log()) { 3711 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3712 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3713 Deoptimization::trap_reason_name(reason), 3714 mcount, trap_count(reason)); 3715 } 3716 return true; 3717 } else { 3718 // The coast is clear. 3719 return false; 3720 } 3721 } 3722 3723 //--------------------------too_many_recompiles-------------------------------- 3724 // Report if there are too many recompiles at the current method and bci. 3725 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3726 // Is not eager to return true, since this will cause the compiler to use 3727 // Action_none for a trap point, to avoid too many recompilations. 3728 bool Compile::too_many_recompiles(ciMethod* method, 3729 int bci, 3730 Deoptimization::DeoptReason reason) { 3731 ciMethodData* md = method->method_data(); 3732 if (md->is_empty()) { 3733 // Assume the trap has not occurred, or that it occurred only 3734 // because of a transient condition during start-up in the interpreter. 3735 return false; 3736 } 3737 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3738 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3739 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3740 Deoptimization::DeoptReason per_bc_reason 3741 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3742 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3743 if ((per_bc_reason == Deoptimization::Reason_none 3744 || md->has_trap_at(bci, m, reason) != 0) 3745 // The trap frequency measure we care about is the recompile count: 3746 && md->trap_recompiled_at(bci, m) 3747 && md->overflow_recompile_count() >= bc_cutoff) { 3748 // Do not emit a trap here if it has already caused recompilations. 3749 // Also, if there are multiple reasons, or if there is no per-BCI record, 3750 // assume the worst. 3751 if (log()) 3752 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3753 Deoptimization::trap_reason_name(reason), 3754 md->trap_count(reason), 3755 md->overflow_recompile_count()); 3756 return true; 3757 } else if (trap_count(reason) != 0 3758 && decompile_count() >= m_cutoff) { 3759 // Too many recompiles globally, and we have seen this sort of trap. 3760 // Use cumulative decompile_count, not just md->decompile_count. 3761 if (log()) 3762 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3763 Deoptimization::trap_reason_name(reason), 3764 md->trap_count(reason), trap_count(reason), 3765 md->decompile_count(), decompile_count()); 3766 return true; 3767 } else { 3768 // The coast is clear. 3769 return false; 3770 } 3771 } 3772 3773 // Compute when not to trap. Used by matching trap based nodes and 3774 // NullCheck optimization. 3775 void Compile::set_allowed_deopt_reasons() { 3776 _allowed_reasons = 0; 3777 if (is_method_compilation()) { 3778 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3779 assert(rs < BitsPerInt, "recode bit map"); 3780 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3781 _allowed_reasons |= nth_bit(rs); 3782 } 3783 } 3784 } 3785 } 3786 3787 bool Compile::is_compiling_clinit_for(ciKlass* k) { 3788 ciMethod* root = method(); // the root method of compilation 3789 return root->is_static_initializer() && root->holder() == k; // access in the context of clinit 3790 } 3791 3792 #ifndef PRODUCT 3793 //------------------------------verify_graph_edges--------------------------- 3794 // Walk the Graph and verify that there is a one-to-one correspondence 3795 // between Use-Def edges and Def-Use edges in the graph. 3796 void Compile::verify_graph_edges(bool no_dead_code) { 3797 if (VerifyGraphEdges) { 3798 ResourceArea *area = Thread::current()->resource_area(); 3799 Unique_Node_List visited(area); 3800 // Call recursive graph walk to check edges 3801 _root->verify_edges(visited); 3802 if (no_dead_code) { 3803 // Now make sure that no visited node is used by an unvisited node. 3804 bool dead_nodes = false; 3805 Unique_Node_List checked(area); 3806 while (visited.size() > 0) { 3807 Node* n = visited.pop(); 3808 checked.push(n); 3809 for (uint i = 0; i < n->outcnt(); i++) { 3810 Node* use = n->raw_out(i); 3811 if (checked.member(use)) continue; // already checked 3812 if (visited.member(use)) continue; // already in the graph 3813 if (use->is_Con()) continue; // a dead ConNode is OK 3814 // At this point, we have found a dead node which is DU-reachable. 3815 if (!dead_nodes) { 3816 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3817 dead_nodes = true; 3818 } 3819 use->dump(2); 3820 tty->print_cr("---"); 3821 checked.push(use); // No repeats; pretend it is now checked. 3822 } 3823 } 3824 assert(!dead_nodes, "using nodes must be reachable from root"); 3825 } 3826 } 3827 } 3828 3829 // Verify GC barriers consistency 3830 // Currently supported: 3831 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3832 void Compile::verify_barriers() { 3833 #if INCLUDE_G1GC 3834 if (UseG1GC) { 3835 // Verify G1 pre-barriers 3836 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()); 3837 3838 ResourceArea *area = Thread::current()->resource_area(); 3839 Unique_Node_List visited(area); 3840 Node_List worklist(area); 3841 // We're going to walk control flow backwards starting from the Root 3842 worklist.push(_root); 3843 while (worklist.size() > 0) { 3844 Node* x = worklist.pop(); 3845 if (x == NULL || x == top()) continue; 3846 if (visited.member(x)) { 3847 continue; 3848 } else { 3849 visited.push(x); 3850 } 3851 3852 if (x->is_Region()) { 3853 for (uint i = 1; i < x->req(); i++) { 3854 worklist.push(x->in(i)); 3855 } 3856 } else { 3857 worklist.push(x->in(0)); 3858 // We are looking for the pattern: 3859 // /->ThreadLocal 3860 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3861 // \->ConI(0) 3862 // We want to verify that the If and the LoadB have the same control 3863 // See GraphKit::g1_write_barrier_pre() 3864 if (x->is_If()) { 3865 IfNode *iff = x->as_If(); 3866 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3867 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3868 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3869 && cmp->in(1)->is_Load()) { 3870 LoadNode* load = cmp->in(1)->as_Load(); 3871 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3872 && load->in(2)->in(3)->is_Con() 3873 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3874 3875 Node* if_ctrl = iff->in(0); 3876 Node* load_ctrl = load->in(0); 3877 3878 if (if_ctrl != load_ctrl) { 3879 // Skip possible CProj->NeverBranch in infinite loops 3880 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3881 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3882 if_ctrl = if_ctrl->in(0)->in(0); 3883 } 3884 } 3885 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3886 } 3887 } 3888 } 3889 } 3890 } 3891 } 3892 } 3893 #endif 3894 } 3895 3896 #endif 3897 3898 // The Compile object keeps track of failure reasons separately from the ciEnv. 3899 // This is required because there is not quite a 1-1 relation between the 3900 // ciEnv and its compilation task and the Compile object. Note that one 3901 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3902 // to backtrack and retry without subsuming loads. Other than this backtracking 3903 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3904 // by the logic in C2Compiler. 3905 void Compile::record_failure(const char* reason) { 3906 if (log() != NULL) { 3907 log()->elem("failure reason='%s' phase='compile'", reason); 3908 } 3909 if (_failure_reason == NULL) { 3910 // Record the first failure reason. 3911 _failure_reason = reason; 3912 } 3913 3914 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3915 C->print_method(PHASE_FAILURE); 3916 } 3917 _root = NULL; // flush the graph, too 3918 } 3919 3920 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3921 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3922 _phase_name(name), _dolog(CITimeVerbose) 3923 { 3924 if (_dolog) { 3925 C = Compile::current(); 3926 _log = C->log(); 3927 } else { 3928 C = NULL; 3929 _log = NULL; 3930 } 3931 if (_log != NULL) { 3932 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3933 _log->stamp(); 3934 _log->end_head(); 3935 } 3936 } 3937 3938 Compile::TracePhase::~TracePhase() { 3939 3940 C = Compile::current(); 3941 if (_dolog) { 3942 _log = C->log(); 3943 } else { 3944 _log = NULL; 3945 } 3946 3947 #ifdef ASSERT 3948 if (PrintIdealNodeCount) { 3949 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3950 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3951 } 3952 3953 if (VerifyIdealNodeCount) { 3954 Compile::current()->print_missing_nodes(); 3955 } 3956 #endif 3957 3958 if (_log != NULL) { 3959 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3960 } 3961 } 3962 3963 //============================================================================= 3964 // Two Constant's are equal when the type and the value are equal. 3965 bool Compile::Constant::operator==(const Constant& other) { 3966 if (type() != other.type() ) return false; 3967 if (can_be_reused() != other.can_be_reused()) return false; 3968 // For floating point values we compare the bit pattern. 3969 switch (type()) { 3970 case T_INT: 3971 case T_FLOAT: return (_v._value.i == other._v._value.i); 3972 case T_LONG: 3973 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3974 case T_OBJECT: 3975 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3976 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3977 case T_METADATA: return (_v._metadata == other._v._metadata); 3978 default: ShouldNotReachHere(); return false; 3979 } 3980 } 3981 3982 static int type_to_size_in_bytes(BasicType t) { 3983 switch (t) { 3984 case T_INT: return sizeof(jint ); 3985 case T_LONG: return sizeof(jlong ); 3986 case T_FLOAT: return sizeof(jfloat ); 3987 case T_DOUBLE: return sizeof(jdouble); 3988 case T_METADATA: return sizeof(Metadata*); 3989 // We use T_VOID as marker for jump-table entries (labels) which 3990 // need an internal word relocation. 3991 case T_VOID: 3992 case T_ADDRESS: 3993 case T_OBJECT: return sizeof(jobject); 3994 default: 3995 ShouldNotReachHere(); 3996 return -1; 3997 } 3998 } 3999 4000 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 4001 // sort descending 4002 if (a->freq() > b->freq()) return -1; 4003 if (a->freq() < b->freq()) return 1; 4004 return 0; 4005 } 4006 4007 void Compile::ConstantTable::calculate_offsets_and_size() { 4008 // First, sort the array by frequencies. 4009 _constants.sort(qsort_comparator); 4010 4011 #ifdef ASSERT 4012 // Make sure all jump-table entries were sorted to the end of the 4013 // array (they have a negative frequency). 4014 bool found_void = false; 4015 for (int i = 0; i < _constants.length(); i++) { 4016 Constant con = _constants.at(i); 4017 if (con.type() == T_VOID) 4018 found_void = true; // jump-tables 4019 else 4020 assert(!found_void, "wrong sorting"); 4021 } 4022 #endif 4023 4024 int offset = 0; 4025 for (int i = 0; i < _constants.length(); i++) { 4026 Constant* con = _constants.adr_at(i); 4027 4028 // Align offset for type. 4029 int typesize = type_to_size_in_bytes(con->type()); 4030 offset = align_up(offset, typesize); 4031 con->set_offset(offset); // set constant's offset 4032 4033 if (con->type() == T_VOID) { 4034 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4035 offset = offset + typesize * n->outcnt(); // expand jump-table 4036 } else { 4037 offset = offset + typesize; 4038 } 4039 } 4040 4041 // Align size up to the next section start (which is insts; see 4042 // CodeBuffer::align_at_start). 4043 assert(_size == -1, "already set?"); 4044 _size = align_up(offset, (int)CodeEntryAlignment); 4045 } 4046 4047 void Compile::ConstantTable::emit(CodeBuffer& cb) { 4048 MacroAssembler _masm(&cb); 4049 for (int i = 0; i < _constants.length(); i++) { 4050 Constant con = _constants.at(i); 4051 address constant_addr = NULL; 4052 switch (con.type()) { 4053 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4054 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4055 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4056 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4057 case T_OBJECT: { 4058 jobject obj = con.get_jobject(); 4059 int oop_index = _masm.oop_recorder()->find_index(obj); 4060 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4061 break; 4062 } 4063 case T_ADDRESS: { 4064 address addr = (address) con.get_jobject(); 4065 constant_addr = _masm.address_constant(addr); 4066 break; 4067 } 4068 // We use T_VOID as marker for jump-table entries (labels) which 4069 // need an internal word relocation. 4070 case T_VOID: { 4071 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4072 // Fill the jump-table with a dummy word. The real value is 4073 // filled in later in fill_jump_table. 4074 address dummy = (address) n; 4075 constant_addr = _masm.address_constant(dummy); 4076 // Expand jump-table 4077 for (uint i = 1; i < n->outcnt(); i++) { 4078 address temp_addr = _masm.address_constant(dummy + i); 4079 assert(temp_addr, "consts section too small"); 4080 } 4081 break; 4082 } 4083 case T_METADATA: { 4084 Metadata* obj = con.get_metadata(); 4085 int metadata_index = _masm.oop_recorder()->find_index(obj); 4086 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4087 break; 4088 } 4089 default: ShouldNotReachHere(); 4090 } 4091 assert(constant_addr, "consts section too small"); 4092 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4093 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4094 } 4095 } 4096 4097 int Compile::ConstantTable::find_offset(Constant& con) const { 4098 int idx = _constants.find(con); 4099 guarantee(idx != -1, "constant must be in constant table"); 4100 int offset = _constants.at(idx).offset(); 4101 guarantee(offset != -1, "constant table not emitted yet?"); 4102 return offset; 4103 } 4104 4105 void Compile::ConstantTable::add(Constant& con) { 4106 if (con.can_be_reused()) { 4107 int idx = _constants.find(con); 4108 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4109 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4110 return; 4111 } 4112 } 4113 (void) _constants.append(con); 4114 } 4115 4116 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4117 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4118 Constant con(type, value, b->_freq); 4119 add(con); 4120 return con; 4121 } 4122 4123 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4124 Constant con(metadata); 4125 add(con); 4126 return con; 4127 } 4128 4129 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4130 jvalue value; 4131 BasicType type = oper->type()->basic_type(); 4132 switch (type) { 4133 case T_LONG: value.j = oper->constantL(); break; 4134 case T_FLOAT: value.f = oper->constantF(); break; 4135 case T_DOUBLE: value.d = oper->constantD(); break; 4136 case T_OBJECT: 4137 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4138 case T_METADATA: return add((Metadata*)oper->constant()); break; 4139 default: guarantee(false, "unhandled type: %s", type2name(type)); 4140 } 4141 return add(n, type, value); 4142 } 4143 4144 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4145 jvalue value; 4146 // We can use the node pointer here to identify the right jump-table 4147 // as this method is called from Compile::Fill_buffer right before 4148 // the MachNodes are emitted and the jump-table is filled (means the 4149 // MachNode pointers do not change anymore). 4150 value.l = (jobject) n; 4151 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4152 add(con); 4153 return con; 4154 } 4155 4156 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4157 // If called from Compile::scratch_emit_size do nothing. 4158 if (Compile::current()->in_scratch_emit_size()) return; 4159 4160 assert(labels.is_nonempty(), "must be"); 4161 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4162 4163 // Since MachConstantNode::constant_offset() also contains 4164 // table_base_offset() we need to subtract the table_base_offset() 4165 // to get the plain offset into the constant table. 4166 int offset = n->constant_offset() - table_base_offset(); 4167 4168 MacroAssembler _masm(&cb); 4169 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4170 4171 for (uint i = 0; i < n->outcnt(); i++) { 4172 address* constant_addr = &jump_table_base[i]; 4173 assert(*constant_addr == (((address) n) + i), "all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i)); 4174 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4175 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4176 } 4177 } 4178 4179 //----------------------------static_subtype_check----------------------------- 4180 // Shortcut important common cases when superklass is exact: 4181 // (0) superklass is java.lang.Object (can occur in reflective code) 4182 // (1) subklass is already limited to a subtype of superklass => always ok 4183 // (2) subklass does not overlap with superklass => always fail 4184 // (3) superklass has NO subtypes and we can check with a simple compare. 4185 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4186 if (StressReflectiveCode) { 4187 return SSC_full_test; // Let caller generate the general case. 4188 } 4189 4190 if (superk == env()->Object_klass()) { 4191 return SSC_always_true; // (0) this test cannot fail 4192 } 4193 4194 ciType* superelem = superk; 4195 if (superelem->is_array_klass()) 4196 superelem = superelem->as_array_klass()->base_element_type(); 4197 4198 if (!subk->is_interface()) { // cannot trust static interface types yet 4199 if (subk->is_subtype_of(superk)) { 4200 return SSC_always_true; // (1) false path dead; no dynamic test needed 4201 } 4202 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4203 !superk->is_subtype_of(subk)) { 4204 return SSC_always_false; 4205 } 4206 } 4207 4208 // If casting to an instance klass, it must have no subtypes 4209 if (superk->is_interface()) { 4210 // Cannot trust interfaces yet. 4211 // %%% S.B. superk->nof_implementors() == 1 4212 } else if (superelem->is_instance_klass()) { 4213 ciInstanceKlass* ik = superelem->as_instance_klass(); 4214 if (!ik->has_subklass() && !ik->is_interface()) { 4215 if (!ik->is_final()) { 4216 // Add a dependency if there is a chance of a later subclass. 4217 dependencies()->assert_leaf_type(ik); 4218 } 4219 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4220 } 4221 } else { 4222 // A primitive array type has no subtypes. 4223 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4224 } 4225 4226 return SSC_full_test; 4227 } 4228 4229 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4230 #ifdef _LP64 4231 // The scaled index operand to AddP must be a clean 64-bit value. 4232 // Java allows a 32-bit int to be incremented to a negative 4233 // value, which appears in a 64-bit register as a large 4234 // positive number. Using that large positive number as an 4235 // operand in pointer arithmetic has bad consequences. 4236 // On the other hand, 32-bit overflow is rare, and the possibility 4237 // can often be excluded, if we annotate the ConvI2L node with 4238 // a type assertion that its value is known to be a small positive 4239 // number. (The prior range check has ensured this.) 4240 // This assertion is used by ConvI2LNode::Ideal. 4241 int index_max = max_jint - 1; // array size is max_jint, index is one less 4242 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4243 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4244 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4245 #endif 4246 return idx; 4247 } 4248 4249 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4250 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4251 if (ctrl != NULL) { 4252 // Express control dependency by a CastII node with a narrow type. 4253 value = new CastIINode(value, itype, false, true /* range check dependency */); 4254 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4255 // node from floating above the range check during loop optimizations. Otherwise, the 4256 // ConvI2L node may be eliminated independently of the range check, causing the data path 4257 // to become TOP while the control path is still there (although it's unreachable). 4258 value->set_req(0, ctrl); 4259 // Save CastII node to remove it after loop optimizations. 4260 phase->C->add_range_check_cast(value); 4261 value = phase->transform(value); 4262 } 4263 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4264 return phase->transform(new ConvI2LNode(value, ltype)); 4265 } 4266 4267 void Compile::print_inlining_stream_free() { 4268 if (_print_inlining_stream != NULL) { 4269 _print_inlining_stream->~stringStream(); 4270 _print_inlining_stream = NULL; 4271 } 4272 } 4273 4274 // The message about the current inlining is accumulated in 4275 // _print_inlining_stream and transfered into the _print_inlining_list 4276 // once we know whether inlining succeeds or not. For regular 4277 // inlining, messages are appended to the buffer pointed by 4278 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4279 // a new buffer is added after _print_inlining_idx in the list. This 4280 // way we can update the inlining message for late inlining call site 4281 // when the inlining is attempted again. 4282 void Compile::print_inlining_init() { 4283 if (print_inlining() || print_intrinsics()) { 4284 // print_inlining_init is actually called several times. 4285 print_inlining_stream_free(); 4286 _print_inlining_stream = new stringStream(); 4287 // Watch out: The memory initialized by the constructor call PrintInliningBuffer() 4288 // will be copied into the only initial element. The default destructor of 4289 // PrintInliningBuffer will be called when leaving the scope here. If it 4290 // would destuct the enclosed stringStream _print_inlining_list[0]->_ss 4291 // would be destructed, too! 4292 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4293 } 4294 } 4295 4296 void Compile::print_inlining_reinit() { 4297 if (print_inlining() || print_intrinsics()) { 4298 print_inlining_stream_free(); 4299 // Re allocate buffer when we change ResourceMark 4300 _print_inlining_stream = new stringStream(); 4301 } 4302 } 4303 4304 void Compile::print_inlining_reset() { 4305 _print_inlining_stream->reset(); 4306 } 4307 4308 void Compile::print_inlining_commit() { 4309 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4310 // Transfer the message from _print_inlining_stream to the current 4311 // _print_inlining_list buffer and clear _print_inlining_stream. 4312 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 4313 print_inlining_reset(); 4314 } 4315 4316 void Compile::print_inlining_push() { 4317 // Add new buffer to the _print_inlining_list at current position 4318 _print_inlining_idx++; 4319 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4320 } 4321 4322 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4323 return _print_inlining_list->at(_print_inlining_idx); 4324 } 4325 4326 void Compile::print_inlining_update(CallGenerator* cg) { 4327 if (print_inlining() || print_intrinsics()) { 4328 if (!cg->is_late_inline()) { 4329 if (print_inlining_current().cg() != NULL) { 4330 print_inlining_push(); 4331 } 4332 print_inlining_commit(); 4333 } else { 4334 if (print_inlining_current().cg() != cg && 4335 (print_inlining_current().cg() != NULL || 4336 print_inlining_current().ss()->size() != 0)) { 4337 print_inlining_push(); 4338 } 4339 print_inlining_commit(); 4340 print_inlining_current().set_cg(cg); 4341 } 4342 } 4343 } 4344 4345 void Compile::print_inlining_move_to(CallGenerator* cg) { 4346 // We resume inlining at a late inlining call site. Locate the 4347 // corresponding inlining buffer so that we can update it. 4348 if (print_inlining()) { 4349 for (int i = 0; i < _print_inlining_list->length(); i++) { 4350 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4351 _print_inlining_idx = i; 4352 return; 4353 } 4354 } 4355 ShouldNotReachHere(); 4356 } 4357 } 4358 4359 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4360 if (print_inlining()) { 4361 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4362 assert(print_inlining_current().cg() == cg, "wrong entry"); 4363 // replace message with new message 4364 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4365 print_inlining_commit(); 4366 print_inlining_current().set_cg(cg); 4367 } 4368 } 4369 4370 void Compile::print_inlining_assert_ready() { 4371 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4372 } 4373 4374 void Compile::process_print_inlining() { 4375 bool do_print_inlining = print_inlining() || print_intrinsics(); 4376 if (do_print_inlining || log() != NULL) { 4377 // Print inlining message for candidates that we couldn't inline 4378 // for lack of space 4379 for (int i = 0; i < _late_inlines.length(); i++) { 4380 CallGenerator* cg = _late_inlines.at(i); 4381 if (!cg->is_mh_late_inline()) { 4382 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4383 if (do_print_inlining) { 4384 cg->print_inlining_late(msg); 4385 } 4386 log_late_inline_failure(cg, msg); 4387 } 4388 } 4389 } 4390 if (do_print_inlining) { 4391 ResourceMark rm; 4392 stringStream ss; 4393 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once."); 4394 for (int i = 0; i < _print_inlining_list->length(); i++) { 4395 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4396 _print_inlining_list->at(i).freeStream(); 4397 } 4398 // Reset _print_inlining_list, it only contains destructed objects. 4399 // It is on the arena, so it will be freed when the arena is reset. 4400 _print_inlining_list = NULL; 4401 // _print_inlining_stream won't be used anymore, either. 4402 print_inlining_stream_free(); 4403 size_t end = ss.size(); 4404 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4405 strncpy(_print_inlining_output, ss.base(), end+1); 4406 _print_inlining_output[end] = 0; 4407 } 4408 } 4409 4410 void Compile::dump_print_inlining() { 4411 if (_print_inlining_output != NULL) { 4412 tty->print_raw(_print_inlining_output); 4413 } 4414 } 4415 4416 void Compile::log_late_inline(CallGenerator* cg) { 4417 if (log() != NULL) { 4418 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4419 cg->unique_id()); 4420 JVMState* p = cg->call_node()->jvms(); 4421 while (p != NULL) { 4422 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4423 p = p->caller(); 4424 } 4425 log()->tail("late_inline"); 4426 } 4427 } 4428 4429 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4430 log_late_inline(cg); 4431 if (log() != NULL) { 4432 log()->inline_fail(msg); 4433 } 4434 } 4435 4436 void Compile::log_inline_id(CallGenerator* cg) { 4437 if (log() != NULL) { 4438 // The LogCompilation tool needs a unique way to identify late 4439 // inline call sites. This id must be unique for this call site in 4440 // this compilation. Try to have it unique across compilations as 4441 // well because it can be convenient when grepping through the log 4442 // file. 4443 // Distinguish OSR compilations from others in case CICountOSR is 4444 // on. 4445 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4446 cg->set_unique_id(id); 4447 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4448 } 4449 } 4450 4451 void Compile::log_inline_failure(const char* msg) { 4452 if (C->log() != NULL) { 4453 C->log()->inline_fail(msg); 4454 } 4455 } 4456 4457 4458 // Dump inlining replay data to the stream. 4459 // Don't change thread state and acquire any locks. 4460 void Compile::dump_inline_data(outputStream* out) { 4461 InlineTree* inl_tree = ilt(); 4462 if (inl_tree != NULL) { 4463 out->print(" inline %d", inl_tree->count()); 4464 inl_tree->dump_replay_data(out); 4465 } 4466 } 4467 4468 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4469 if (n1->Opcode() < n2->Opcode()) return -1; 4470 else if (n1->Opcode() > n2->Opcode()) return 1; 4471 4472 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4473 for (uint i = 1; i < n1->req(); i++) { 4474 if (n1->in(i) < n2->in(i)) return -1; 4475 else if (n1->in(i) > n2->in(i)) return 1; 4476 } 4477 4478 return 0; 4479 } 4480 4481 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4482 Node* n1 = *n1p; 4483 Node* n2 = *n2p; 4484 4485 return cmp_expensive_nodes(n1, n2); 4486 } 4487 4488 void Compile::sort_expensive_nodes() { 4489 if (!expensive_nodes_sorted()) { 4490 _expensive_nodes->sort(cmp_expensive_nodes); 4491 } 4492 } 4493 4494 bool Compile::expensive_nodes_sorted() const { 4495 for (int i = 1; i < _expensive_nodes->length(); i++) { 4496 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4497 return false; 4498 } 4499 } 4500 return true; 4501 } 4502 4503 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4504 if (_expensive_nodes->length() == 0) { 4505 return false; 4506 } 4507 4508 assert(OptimizeExpensiveOps, "optimization off?"); 4509 4510 // Take this opportunity to remove dead nodes from the list 4511 int j = 0; 4512 for (int i = 0; i < _expensive_nodes->length(); i++) { 4513 Node* n = _expensive_nodes->at(i); 4514 if (!n->is_unreachable(igvn)) { 4515 assert(n->is_expensive(), "should be expensive"); 4516 _expensive_nodes->at_put(j, n); 4517 j++; 4518 } 4519 } 4520 _expensive_nodes->trunc_to(j); 4521 4522 // Then sort the list so that similar nodes are next to each other 4523 // and check for at least two nodes of identical kind with same data 4524 // inputs. 4525 sort_expensive_nodes(); 4526 4527 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4528 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4529 return true; 4530 } 4531 } 4532 4533 return false; 4534 } 4535 4536 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4537 if (_expensive_nodes->length() == 0) { 4538 return; 4539 } 4540 4541 assert(OptimizeExpensiveOps, "optimization off?"); 4542 4543 // Sort to bring similar nodes next to each other and clear the 4544 // control input of nodes for which there's only a single copy. 4545 sort_expensive_nodes(); 4546 4547 int j = 0; 4548 int identical = 0; 4549 int i = 0; 4550 bool modified = false; 4551 for (; i < _expensive_nodes->length()-1; i++) { 4552 assert(j <= i, "can't write beyond current index"); 4553 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4554 identical++; 4555 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4556 continue; 4557 } 4558 if (identical > 0) { 4559 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4560 identical = 0; 4561 } else { 4562 Node* n = _expensive_nodes->at(i); 4563 igvn.replace_input_of(n, 0, NULL); 4564 igvn.hash_insert(n); 4565 modified = true; 4566 } 4567 } 4568 if (identical > 0) { 4569 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4570 } else if (_expensive_nodes->length() >= 1) { 4571 Node* n = _expensive_nodes->at(i); 4572 igvn.replace_input_of(n, 0, NULL); 4573 igvn.hash_insert(n); 4574 modified = true; 4575 } 4576 _expensive_nodes->trunc_to(j); 4577 if (modified) { 4578 igvn.optimize(); 4579 } 4580 } 4581 4582 void Compile::add_expensive_node(Node * n) { 4583 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4584 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4585 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4586 if (OptimizeExpensiveOps) { 4587 _expensive_nodes->append(n); 4588 } else { 4589 // Clear control input and let IGVN optimize expensive nodes if 4590 // OptimizeExpensiveOps is off. 4591 n->set_req(0, NULL); 4592 } 4593 } 4594 4595 /** 4596 * Remove the speculative part of types and clean up the graph 4597 */ 4598 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4599 if (UseTypeSpeculation) { 4600 Unique_Node_List worklist; 4601 worklist.push(root()); 4602 int modified = 0; 4603 // Go over all type nodes that carry a speculative type, drop the 4604 // speculative part of the type and enqueue the node for an igvn 4605 // which may optimize it out. 4606 for (uint next = 0; next < worklist.size(); ++next) { 4607 Node *n = worklist.at(next); 4608 if (n->is_Type()) { 4609 TypeNode* tn = n->as_Type(); 4610 const Type* t = tn->type(); 4611 const Type* t_no_spec = t->remove_speculative(); 4612 if (t_no_spec != t) { 4613 bool in_hash = igvn.hash_delete(n); 4614 assert(in_hash, "node should be in igvn hash table"); 4615 tn->set_type(t_no_spec); 4616 igvn.hash_insert(n); 4617 igvn._worklist.push(n); // give it a chance to go away 4618 modified++; 4619 } 4620 } 4621 uint max = n->len(); 4622 for( uint i = 0; i < max; ++i ) { 4623 Node *m = n->in(i); 4624 if (not_a_node(m)) continue; 4625 worklist.push(m); 4626 } 4627 } 4628 // Drop the speculative part of all types in the igvn's type table 4629 igvn.remove_speculative_types(); 4630 if (modified > 0) { 4631 igvn.optimize(); 4632 } 4633 #ifdef ASSERT 4634 // Verify that after the IGVN is over no speculative type has resurfaced 4635 worklist.clear(); 4636 worklist.push(root()); 4637 for (uint next = 0; next < worklist.size(); ++next) { 4638 Node *n = worklist.at(next); 4639 const Type* t = igvn.type_or_null(n); 4640 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4641 if (n->is_Type()) { 4642 t = n->as_Type()->type(); 4643 assert(t == t->remove_speculative(), "no more speculative types"); 4644 } 4645 uint max = n->len(); 4646 for( uint i = 0; i < max; ++i ) { 4647 Node *m = n->in(i); 4648 if (not_a_node(m)) continue; 4649 worklist.push(m); 4650 } 4651 } 4652 igvn.check_no_speculative_types(); 4653 #endif 4654 } 4655 } 4656 4657 // Auxiliary method to support randomized stressing/fuzzing. 4658 // 4659 // This method can be called the arbitrary number of times, with current count 4660 // as the argument. The logic allows selecting a single candidate from the 4661 // running list of candidates as follows: 4662 // int count = 0; 4663 // Cand* selected = null; 4664 // while(cand = cand->next()) { 4665 // if (randomized_select(++count)) { 4666 // selected = cand; 4667 // } 4668 // } 4669 // 4670 // Including count equalizes the chances any candidate is "selected". 4671 // This is useful when we don't have the complete list of candidates to choose 4672 // from uniformly. In this case, we need to adjust the randomicity of the 4673 // selection, or else we will end up biasing the selection towards the latter 4674 // candidates. 4675 // 4676 // Quick back-envelope calculation shows that for the list of n candidates 4677 // the equal probability for the candidate to persist as "best" can be 4678 // achieved by replacing it with "next" k-th candidate with the probability 4679 // of 1/k. It can be easily shown that by the end of the run, the 4680 // probability for any candidate is converged to 1/n, thus giving the 4681 // uniform distribution among all the candidates. 4682 // 4683 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4684 #define RANDOMIZED_DOMAIN_POW 29 4685 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4686 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4687 bool Compile::randomized_select(int count) { 4688 assert(count > 0, "only positive"); 4689 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4690 } 4691 4692 CloneMap& Compile::clone_map() { return _clone_map; } 4693 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4694 4695 void NodeCloneInfo::dump() const { 4696 tty->print(" {%d:%d} ", idx(), gen()); 4697 } 4698 4699 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4700 uint64_t val = value(old->_idx); 4701 NodeCloneInfo cio(val); 4702 assert(val != 0, "old node should be in the map"); 4703 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4704 insert(nnn->_idx, cin.get()); 4705 #ifndef PRODUCT 4706 if (is_debug()) { 4707 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4708 } 4709 #endif 4710 } 4711 4712 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4713 NodeCloneInfo cio(value(old->_idx)); 4714 if (cio.get() == 0) { 4715 cio.set(old->_idx, 0); 4716 insert(old->_idx, cio.get()); 4717 #ifndef PRODUCT 4718 if (is_debug()) { 4719 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4720 } 4721 #endif 4722 } 4723 clone(old, nnn, gen); 4724 } 4725 4726 int CloneMap::max_gen() const { 4727 int g = 0; 4728 DictI di(_dict); 4729 for(; di.test(); ++di) { 4730 int t = gen(di._key); 4731 if (g < t) { 4732 g = t; 4733 #ifndef PRODUCT 4734 if (is_debug()) { 4735 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4736 } 4737 #endif 4738 } 4739 } 4740 return g; 4741 } 4742 4743 void CloneMap::dump(node_idx_t key) const { 4744 uint64_t val = value(key); 4745 if (val != 0) { 4746 NodeCloneInfo ni(val); 4747 ni.dump(); 4748 } 4749 }