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/valuetypenode.hpp" 71 #include "opto/vectornode.hpp" 72 #include "runtime/arguments.hpp" 73 #include "runtime/sharedRuntime.hpp" 74 #include "runtime/signature.hpp" 75 #include "runtime/stubRoutines.hpp" 76 #include "runtime/timer.hpp" 77 #include "utilities/align.hpp" 78 #include "utilities/copy.hpp" 79 #if INCLUDE_G1GC 80 #include "gc/g1/g1ThreadLocalData.hpp" 81 #endif // INCLUDE_G1GC 82 83 84 // -------------------- Compile::mach_constant_base_node ----------------------- 85 // Constant table base node singleton. 86 MachConstantBaseNode* Compile::mach_constant_base_node() { 87 if (_mach_constant_base_node == NULL) { 88 _mach_constant_base_node = new MachConstantBaseNode(); 89 _mach_constant_base_node->add_req(C->root()); 90 } 91 return _mach_constant_base_node; 92 } 93 94 95 /// Support for intrinsics. 96 97 // Return the index at which m must be inserted (or already exists). 98 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 99 class IntrinsicDescPair { 100 private: 101 ciMethod* _m; 102 bool _is_virtual; 103 public: 104 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 105 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 106 ciMethod* m= elt->method(); 107 ciMethod* key_m = key->_m; 108 if (key_m < m) return -1; 109 else if (key_m > m) return 1; 110 else { 111 bool is_virtual = elt->is_virtual(); 112 bool key_virtual = key->_is_virtual; 113 if (key_virtual < is_virtual) return -1; 114 else if (key_virtual > is_virtual) return 1; 115 else return 0; 116 } 117 } 118 }; 119 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 120 #ifdef ASSERT 121 for (int i = 1; i < _intrinsics->length(); i++) { 122 CallGenerator* cg1 = _intrinsics->at(i-1); 123 CallGenerator* cg2 = _intrinsics->at(i); 124 assert(cg1->method() != cg2->method() 125 ? cg1->method() < cg2->method() 126 : cg1->is_virtual() < cg2->is_virtual(), 127 "compiler intrinsics list must stay sorted"); 128 } 129 #endif 130 IntrinsicDescPair pair(m, is_virtual); 131 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 132 } 133 134 void Compile::register_intrinsic(CallGenerator* cg) { 135 if (_intrinsics == NULL) { 136 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 137 } 138 int len = _intrinsics->length(); 139 bool found = false; 140 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 141 assert(!found, "registering twice"); 142 _intrinsics->insert_before(index, cg); 143 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 144 } 145 146 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 147 assert(m->is_loaded(), "don't try this on unloaded methods"); 148 if (_intrinsics != NULL) { 149 bool found = false; 150 int index = intrinsic_insertion_index(m, is_virtual, found); 151 if (found) { 152 return _intrinsics->at(index); 153 } 154 } 155 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 156 if (m->intrinsic_id() != vmIntrinsics::_none && 157 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 158 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 159 if (cg != NULL) { 160 // Save it for next time: 161 register_intrinsic(cg); 162 return cg; 163 } else { 164 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 165 } 166 } 167 return NULL; 168 } 169 170 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 171 // in library_call.cpp. 172 173 174 #ifndef PRODUCT 175 // statistics gathering... 176 177 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 178 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 179 180 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 181 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 182 int oflags = _intrinsic_hist_flags[id]; 183 assert(flags != 0, "what happened?"); 184 if (is_virtual) { 185 flags |= _intrinsic_virtual; 186 } 187 bool changed = (flags != oflags); 188 if ((flags & _intrinsic_worked) != 0) { 189 juint count = (_intrinsic_hist_count[id] += 1); 190 if (count == 1) { 191 changed = true; // first time 192 } 193 // increment the overall count also: 194 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 195 } 196 if (changed) { 197 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 198 // Something changed about the intrinsic's virtuality. 199 if ((flags & _intrinsic_virtual) != 0) { 200 // This is the first use of this intrinsic as a virtual call. 201 if (oflags != 0) { 202 // We already saw it as a non-virtual, so note both cases. 203 flags |= _intrinsic_both; 204 } 205 } else if ((oflags & _intrinsic_both) == 0) { 206 // This is the first use of this intrinsic as a non-virtual 207 flags |= _intrinsic_both; 208 } 209 } 210 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 211 } 212 // update the overall flags also: 213 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 214 return changed; 215 } 216 217 static char* format_flags(int flags, char* buf) { 218 buf[0] = 0; 219 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 220 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 221 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 222 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 223 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 224 if (buf[0] == 0) strcat(buf, ","); 225 assert(buf[0] == ',', "must be"); 226 return &buf[1]; 227 } 228 229 void Compile::print_intrinsic_statistics() { 230 char flagsbuf[100]; 231 ttyLocker ttyl; 232 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 233 tty->print_cr("Compiler intrinsic usage:"); 234 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 235 if (total == 0) total = 1; // avoid div0 in case of no successes 236 #define PRINT_STAT_LINE(name, c, f) \ 237 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 238 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 239 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 240 int flags = _intrinsic_hist_flags[id]; 241 juint count = _intrinsic_hist_count[id]; 242 if ((flags | count) != 0) { 243 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 244 } 245 } 246 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 247 if (xtty != NULL) xtty->tail("statistics"); 248 } 249 250 void Compile::print_statistics() { 251 { ttyLocker ttyl; 252 if (xtty != NULL) xtty->head("statistics type='opto'"); 253 Parse::print_statistics(); 254 PhaseCCP::print_statistics(); 255 PhaseRegAlloc::print_statistics(); 256 Scheduling::print_statistics(); 257 PhasePeephole::print_statistics(); 258 PhaseIdealLoop::print_statistics(); 259 if (xtty != NULL) xtty->tail("statistics"); 260 } 261 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 262 // put this under its own <statistics> element. 263 print_intrinsic_statistics(); 264 } 265 } 266 #endif //PRODUCT 267 268 // Support for bundling info 269 Bundle* Compile::node_bundling(const Node *n) { 270 assert(valid_bundle_info(n), "oob"); 271 return &_node_bundling_base[n->_idx]; 272 } 273 274 bool Compile::valid_bundle_info(const Node *n) { 275 return (_node_bundling_limit > n->_idx); 276 } 277 278 279 void Compile::gvn_replace_by(Node* n, Node* nn) { 280 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 281 Node* use = n->last_out(i); 282 bool is_in_table = initial_gvn()->hash_delete(use); 283 uint uses_found = 0; 284 for (uint j = 0; j < use->len(); j++) { 285 if (use->in(j) == n) { 286 if (j < use->req()) 287 use->set_req(j, nn); 288 else 289 use->set_prec(j, nn); 290 uses_found++; 291 } 292 } 293 if (is_in_table) { 294 // reinsert into table 295 initial_gvn()->hash_find_insert(use); 296 } 297 record_for_igvn(use); 298 i -= uses_found; // we deleted 1 or more copies of this edge 299 } 300 } 301 302 303 static inline bool not_a_node(const Node* n) { 304 if (n == NULL) return true; 305 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 306 if (*(address*)n == badAddress) return true; // kill by Node::destruct 307 return false; 308 } 309 310 // Identify all nodes that are reachable from below, useful. 311 // Use breadth-first pass that records state in a Unique_Node_List, 312 // recursive traversal is slower. 313 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 314 int estimated_worklist_size = live_nodes(); 315 useful.map( estimated_worklist_size, NULL ); // preallocate space 316 317 // Initialize worklist 318 if (root() != NULL) { useful.push(root()); } 319 // If 'top' is cached, declare it useful to preserve cached node 320 if( cached_top_node() ) { useful.push(cached_top_node()); } 321 322 // Push all useful nodes onto the list, breadthfirst 323 for( uint next = 0; next < useful.size(); ++next ) { 324 assert( next < unique(), "Unique useful nodes < total nodes"); 325 Node *n = useful.at(next); 326 uint max = n->len(); 327 for( uint i = 0; i < max; ++i ) { 328 Node *m = n->in(i); 329 if (not_a_node(m)) continue; 330 useful.push(m); 331 } 332 } 333 } 334 335 // Update dead_node_list with any missing dead nodes using useful 336 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 337 void Compile::update_dead_node_list(Unique_Node_List &useful) { 338 uint max_idx = unique(); 339 VectorSet& useful_node_set = useful.member_set(); 340 341 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 342 // If node with index node_idx is not in useful set, 343 // mark it as dead in dead node list. 344 if (! useful_node_set.test(node_idx) ) { 345 record_dead_node(node_idx); 346 } 347 } 348 } 349 350 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 351 int shift = 0; 352 for (int i = 0; i < inlines->length(); i++) { 353 CallGenerator* cg = inlines->at(i); 354 CallNode* call = cg->call_node(); 355 if (shift > 0) { 356 inlines->at_put(i-shift, cg); 357 } 358 if (!useful.member(call)) { 359 shift++; 360 } 361 } 362 inlines->trunc_to(inlines->length()-shift); 363 } 364 365 // Disconnect all useless nodes by disconnecting those at the boundary. 366 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 367 uint next = 0; 368 while (next < useful.size()) { 369 Node *n = useful.at(next++); 370 if (n->is_SafePoint()) { 371 // We're done with a parsing phase. Replaced nodes are not valid 372 // beyond that point. 373 n->as_SafePoint()->delete_replaced_nodes(); 374 } 375 // Use raw traversal of out edges since this code removes out edges 376 int max = n->outcnt(); 377 for (int j = 0; j < max; ++j) { 378 Node* child = n->raw_out(j); 379 if (! useful.member(child)) { 380 assert(!child->is_top() || child != top(), 381 "If top is cached in Compile object it is in useful list"); 382 // Only need to remove this out-edge to the useless node 383 n->raw_del_out(j); 384 --j; 385 --max; 386 } 387 } 388 if (n->outcnt() == 1 && n->has_special_unique_user()) { 389 record_for_igvn(n->unique_out()); 390 } 391 } 392 // Remove useless macro and predicate opaq nodes 393 for (int i = C->macro_count()-1; i >= 0; i--) { 394 Node* n = C->macro_node(i); 395 if (!useful.member(n)) { 396 remove_macro_node(n); 397 } 398 } 399 // Remove useless CastII nodes with range check dependency 400 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 401 Node* cast = range_check_cast_node(i); 402 if (!useful.member(cast)) { 403 remove_range_check_cast(cast); 404 } 405 } 406 // Remove useless expensive nodes 407 for (int i = C->expensive_count()-1; i >= 0; i--) { 408 Node* n = C->expensive_node(i); 409 if (!useful.member(n)) { 410 remove_expensive_node(n); 411 } 412 } 413 // Remove useless Opaque4 nodes 414 for (int i = opaque4_count() - 1; i >= 0; i--) { 415 Node* opaq = opaque4_node(i); 416 if (!useful.member(opaq)) { 417 remove_opaque4_node(opaq); 418 } 419 } 420 // Remove useless value type nodes 421 if (_value_type_nodes != NULL) { 422 _value_type_nodes->remove_useless_nodes(useful.member_set()); 423 } 424 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 425 bs->eliminate_useless_gc_barriers(useful); 426 // clean up the late inline lists 427 remove_useless_late_inlines(&_string_late_inlines, useful); 428 remove_useless_late_inlines(&_boxing_late_inlines, useful); 429 remove_useless_late_inlines(&_late_inlines, useful); 430 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 431 } 432 433 //------------------------------frame_size_in_words----------------------------- 434 // frame_slots in units of words 435 int Compile::frame_size_in_words() const { 436 // shift is 0 in LP32 and 1 in LP64 437 const int shift = (LogBytesPerWord - LogBytesPerInt); 438 int words = _frame_slots >> shift; 439 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 440 return words; 441 } 442 443 // To bang the stack of this compiled method we use the stack size 444 // that the interpreter would need in case of a deoptimization. This 445 // removes the need to bang the stack in the deoptimization blob which 446 // in turn simplifies stack overflow handling. 447 int Compile::bang_size_in_bytes() const { 448 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 449 } 450 451 // ============================================================================ 452 //------------------------------CompileWrapper--------------------------------- 453 class CompileWrapper : public StackObj { 454 Compile *const _compile; 455 public: 456 CompileWrapper(Compile* compile); 457 458 ~CompileWrapper(); 459 }; 460 461 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 462 // the Compile* pointer is stored in the current ciEnv: 463 ciEnv* env = compile->env(); 464 assert(env == ciEnv::current(), "must already be a ciEnv active"); 465 assert(env->compiler_data() == NULL, "compile already active?"); 466 env->set_compiler_data(compile); 467 assert(compile == Compile::current(), "sanity"); 468 469 compile->set_type_dict(NULL); 470 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 471 compile->clone_map().set_clone_idx(0); 472 compile->set_type_hwm(NULL); 473 compile->set_type_last_size(0); 474 compile->set_last_tf(NULL, NULL); 475 compile->set_indexSet_arena(NULL); 476 compile->set_indexSet_free_block_list(NULL); 477 compile->init_type_arena(); 478 Type::Initialize(compile); 479 _compile->set_scratch_buffer_blob(NULL); 480 _compile->begin_method(); 481 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 482 } 483 CompileWrapper::~CompileWrapper() { 484 _compile->end_method(); 485 if (_compile->scratch_buffer_blob() != NULL) 486 BufferBlob::free(_compile->scratch_buffer_blob()); 487 _compile->env()->set_compiler_data(NULL); 488 } 489 490 491 //----------------------------print_compile_messages--------------------------- 492 void Compile::print_compile_messages() { 493 #ifndef PRODUCT 494 // Check if recompiling 495 if (_subsume_loads == false && PrintOpto) { 496 // Recompiling without allowing machine instructions to subsume loads 497 tty->print_cr("*********************************************************"); 498 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 499 tty->print_cr("*********************************************************"); 500 } 501 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 502 // Recompiling without escape analysis 503 tty->print_cr("*********************************************************"); 504 tty->print_cr("** Bailout: Recompile without escape analysis **"); 505 tty->print_cr("*********************************************************"); 506 } 507 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 508 // Recompiling without boxing elimination 509 tty->print_cr("*********************************************************"); 510 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 511 tty->print_cr("*********************************************************"); 512 } 513 if (C->directive()->BreakAtCompileOption) { 514 // Open the debugger when compiling this method. 515 tty->print("### Breaking when compiling: "); 516 method()->print_short_name(); 517 tty->cr(); 518 BREAKPOINT; 519 } 520 521 if( PrintOpto ) { 522 if (is_osr_compilation()) { 523 tty->print("[OSR]%3d", _compile_id); 524 } else { 525 tty->print("%3d", _compile_id); 526 } 527 } 528 #endif 529 } 530 531 532 //-----------------------init_scratch_buffer_blob------------------------------ 533 // Construct a temporary BufferBlob and cache it for this compile. 534 void Compile::init_scratch_buffer_blob(int const_size) { 535 // If there is already a scratch buffer blob allocated and the 536 // constant section is big enough, use it. Otherwise free the 537 // current and allocate a new one. 538 BufferBlob* blob = scratch_buffer_blob(); 539 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 540 // Use the current blob. 541 } else { 542 if (blob != NULL) { 543 BufferBlob::free(blob); 544 } 545 546 ResourceMark rm; 547 _scratch_const_size = const_size; 548 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 549 blob = BufferBlob::create("Compile::scratch_buffer", size); 550 // Record the buffer blob for next time. 551 set_scratch_buffer_blob(blob); 552 // Have we run out of code space? 553 if (scratch_buffer_blob() == NULL) { 554 // Let CompilerBroker disable further compilations. 555 record_failure("Not enough space for scratch buffer in CodeCache"); 556 return; 557 } 558 } 559 560 // Initialize the relocation buffers 561 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 562 set_scratch_locs_memory(locs_buf); 563 } 564 565 566 //-----------------------scratch_emit_size------------------------------------- 567 // Helper function that computes size by emitting code 568 uint Compile::scratch_emit_size(const Node* n) { 569 // Start scratch_emit_size section. 570 set_in_scratch_emit_size(true); 571 572 // Emit into a trash buffer and count bytes emitted. 573 // This is a pretty expensive way to compute a size, 574 // but it works well enough if seldom used. 575 // All common fixed-size instructions are given a size 576 // method by the AD file. 577 // Note that the scratch buffer blob and locs memory are 578 // allocated at the beginning of the compile task, and 579 // may be shared by several calls to scratch_emit_size. 580 // The allocation of the scratch buffer blob is particularly 581 // expensive, since it has to grab the code cache lock. 582 BufferBlob* blob = this->scratch_buffer_blob(); 583 assert(blob != NULL, "Initialize BufferBlob at start"); 584 assert(blob->size() > MAX_inst_size, "sanity"); 585 relocInfo* locs_buf = scratch_locs_memory(); 586 address blob_begin = blob->content_begin(); 587 address blob_end = (address)locs_buf; 588 assert(blob->contains(blob_end), "sanity"); 589 CodeBuffer buf(blob_begin, blob_end - blob_begin); 590 buf.initialize_consts_size(_scratch_const_size); 591 buf.initialize_stubs_size(MAX_stubs_size); 592 assert(locs_buf != NULL, "sanity"); 593 int lsize = MAX_locs_size / 3; 594 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 595 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 596 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 597 // Mark as scratch buffer. 598 buf.consts()->set_scratch_emit(); 599 buf.insts()->set_scratch_emit(); 600 buf.stubs()->set_scratch_emit(); 601 602 // Do the emission. 603 604 Label fakeL; // Fake label for branch instructions. 605 Label* saveL = NULL; 606 uint save_bnum = 0; 607 bool is_branch = n->is_MachBranch(); 608 if (is_branch) { 609 MacroAssembler masm(&buf); 610 masm.bind(fakeL); 611 n->as_MachBranch()->save_label(&saveL, &save_bnum); 612 n->as_MachBranch()->label_set(&fakeL, 0); 613 } 614 n->emit(buf, this->regalloc()); 615 616 // Emitting into the scratch buffer should not fail 617 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 618 619 if (is_branch) // Restore label. 620 n->as_MachBranch()->label_set(saveL, save_bnum); 621 622 // End scratch_emit_size section. 623 set_in_scratch_emit_size(false); 624 625 return buf.insts_size(); 626 } 627 628 629 // ============================================================================ 630 //------------------------------Compile standard------------------------------- 631 debug_only( int Compile::_debug_idx = 100000; ) 632 633 // Compile a method. entry_bci is -1 for normal compilations and indicates 634 // the continuation bci for on stack replacement. 635 636 637 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 638 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 639 : Phase(Compiler), 640 _env(ci_env), 641 _directive(directive), 642 _log(ci_env->log()), 643 _compile_id(ci_env->compile_id()), 644 _save_argument_registers(false), 645 _stub_name(NULL), 646 _stub_function(NULL), 647 _stub_entry_point(NULL), 648 _method(target), 649 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 650 _entry_bci(osr_bci), 651 _initial_gvn(NULL), 652 _for_igvn(NULL), 653 _warm_calls(NULL), 654 _subsume_loads(subsume_loads), 655 _do_escape_analysis(do_escape_analysis), 656 _eliminate_boxing(eliminate_boxing), 657 _failure_reason(NULL), 658 _code_buffer("Compile::Fill_buffer"), 659 _orig_pc_slot(0), 660 _orig_pc_slot_offset_in_bytes(0), 661 _has_method_handle_invokes(false), 662 _mach_constant_base_node(NULL), 663 _node_bundling_limit(0), 664 _node_bundling_base(NULL), 665 _java_calls(0), 666 _inner_loops(0), 667 _scratch_const_size(-1), 668 _in_scratch_emit_size(false), 669 _dead_node_list(comp_arena()), 670 _dead_node_count(0), 671 #ifndef PRODUCT 672 _trace_opto_output(directive->TraceOptoOutputOption), 673 _in_dump_cnt(0), 674 _printer(IdealGraphPrinter::printer()), 675 #endif 676 _congraph(NULL), 677 _comp_arena(mtCompiler), 678 _node_arena(mtCompiler), 679 _old_arena(mtCompiler), 680 _Compile_types(mtCompiler), 681 _replay_inline_data(NULL), 682 _late_inlines(comp_arena(), 2, 0, NULL), 683 _string_late_inlines(comp_arena(), 2, 0, NULL), 684 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 685 _late_inlines_pos(0), 686 _number_of_mh_late_inlines(0), 687 _inlining_progress(false), 688 _inlining_incrementally(false), 689 _print_inlining_list(NULL), 690 _print_inlining_stream(NULL), 691 _print_inlining_idx(0), 692 _print_inlining_output(NULL), 693 _interpreter_frame_size(0), 694 _max_node_limit(MaxNodeLimit), 695 _has_reserved_stack_access(target->has_reserved_stack_access()) { 696 C = this; 697 #ifndef PRODUCT 698 if (_printer != NULL) { 699 _printer->set_compile(this); 700 } 701 #endif 702 CompileWrapper cw(this); 703 704 if (CITimeVerbose) { 705 tty->print(" "); 706 target->holder()->name()->print(); 707 tty->print("."); 708 target->print_short_name(); 709 tty->print(" "); 710 } 711 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 712 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 713 714 #ifndef PRODUCT 715 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 716 if (!print_opto_assembly) { 717 bool print_assembly = directive->PrintAssemblyOption; 718 if (print_assembly && !Disassembler::can_decode()) { 719 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 720 print_opto_assembly = true; 721 } 722 } 723 set_print_assembly(print_opto_assembly); 724 set_parsed_irreducible_loop(false); 725 726 if (directive->ReplayInlineOption) { 727 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 728 } 729 #endif 730 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 731 set_print_intrinsics(directive->PrintIntrinsicsOption); 732 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 733 734 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 735 // Make sure the method being compiled gets its own MDO, 736 // so we can at least track the decompile_count(). 737 // Need MDO to record RTM code generation state. 738 method()->ensure_method_data(); 739 } 740 741 Init(::AliasLevel); 742 743 744 print_compile_messages(); 745 746 _ilt = InlineTree::build_inline_tree_root(); 747 748 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 749 assert(num_alias_types() >= AliasIdxRaw, ""); 750 751 #define MINIMUM_NODE_HASH 1023 752 // Node list that Iterative GVN will start with 753 Unique_Node_List for_igvn(comp_arena()); 754 set_for_igvn(&for_igvn); 755 756 // GVN that will be run immediately on new nodes 757 uint estimated_size = method()->code_size()*4+64; 758 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 759 PhaseGVN gvn(node_arena(), estimated_size); 760 set_initial_gvn(&gvn); 761 762 print_inlining_init(); 763 { // Scope for timing the parser 764 TracePhase tp("parse", &timers[_t_parser]); 765 766 // Put top into the hash table ASAP. 767 initial_gvn()->transform_no_reclaim(top()); 768 769 // Set up tf(), start(), and find a CallGenerator. 770 CallGenerator* cg = NULL; 771 if (is_osr_compilation()) { 772 const TypeTuple *domain = StartOSRNode::osr_domain(); 773 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 774 init_tf(TypeFunc::make(domain, range)); 775 StartNode* s = new StartOSRNode(root(), domain); 776 initial_gvn()->set_type_bottom(s); 777 init_start(s); 778 cg = CallGenerator::for_osr(method(), entry_bci()); 779 } else { 780 // Normal case. 781 init_tf(TypeFunc::make(method())); 782 StartNode* s = new StartNode(root(), tf()->domain_cc()); 783 initial_gvn()->set_type_bottom(s); 784 init_start(s); 785 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 786 // With java.lang.ref.reference.get() we must go through the 787 // intrinsic - even when get() is the root 788 // method of the compile - so that, if necessary, the value in 789 // the referent field of the reference object gets recorded by 790 // the pre-barrier code. 791 cg = find_intrinsic(method(), false); 792 } 793 if (cg == NULL) { 794 float past_uses = method()->interpreter_invocation_count(); 795 float expected_uses = past_uses; 796 cg = CallGenerator::for_inline(method(), expected_uses); 797 } 798 } 799 if (failing()) return; 800 if (cg == NULL) { 801 record_method_not_compilable("cannot parse method"); 802 return; 803 } 804 JVMState* jvms = build_start_state(start(), tf()); 805 if ((jvms = cg->generate(jvms)) == NULL) { 806 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 807 record_method_not_compilable("method parse failed"); 808 } 809 return; 810 } 811 GraphKit kit(jvms); 812 813 if (!kit.stopped()) { 814 // Accept return values, and transfer control we know not where. 815 // This is done by a special, unique ReturnNode bound to root. 816 return_values(kit.jvms()); 817 } 818 819 if (kit.has_exceptions()) { 820 // Any exceptions that escape from this call must be rethrown 821 // to whatever caller is dynamically above us on the stack. 822 // This is done by a special, unique RethrowNode bound to root. 823 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 824 } 825 826 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 827 828 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 829 inline_string_calls(true); 830 } 831 832 if (failing()) return; 833 834 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 835 836 // Remove clutter produced by parsing. 837 if (!failing()) { 838 ResourceMark rm; 839 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 840 } 841 } 842 843 // Note: Large methods are capped off in do_one_bytecode(). 844 if (failing()) return; 845 846 // After parsing, node notes are no longer automagic. 847 // They must be propagated by register_new_node_with_optimizer(), 848 // clone(), or the like. 849 set_default_node_notes(NULL); 850 851 for (;;) { 852 int successes = Inline_Warm(); 853 if (failing()) return; 854 if (successes == 0) break; 855 } 856 857 // Drain the list. 858 Finish_Warm(); 859 #ifndef PRODUCT 860 if (_printer && _printer->should_print(1)) { 861 _printer->print_inlining(); 862 } 863 #endif 864 865 if (failing()) return; 866 NOT_PRODUCT( verify_graph_edges(); ) 867 868 // Now optimize 869 Optimize(); 870 if (failing()) return; 871 NOT_PRODUCT( verify_graph_edges(); ) 872 873 #ifndef PRODUCT 874 if (PrintIdeal) { 875 ttyLocker ttyl; // keep the following output all in one block 876 // This output goes directly to the tty, not the compiler log. 877 // To enable tools to match it up with the compilation activity, 878 // be sure to tag this tty output with the compile ID. 879 if (xtty != NULL) { 880 xtty->head("ideal compile_id='%d'%s", compile_id(), 881 is_osr_compilation() ? " compile_kind='osr'" : 882 ""); 883 } 884 root()->dump(9999); 885 if (xtty != NULL) { 886 xtty->tail("ideal"); 887 } 888 } 889 #endif 890 891 NOT_PRODUCT( verify_barriers(); ) 892 893 // Dump compilation data to replay it. 894 if (directive->DumpReplayOption) { 895 env()->dump_replay_data(_compile_id); 896 } 897 if (directive->DumpInlineOption && (ilt() != NULL)) { 898 env()->dump_inline_data(_compile_id); 899 } 900 901 // Now that we know the size of all the monitors we can add a fixed slot 902 // for the original deopt pc. 903 904 _orig_pc_slot = fixed_slots(); 905 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 906 set_fixed_slots(next_slot); 907 908 // Compute when to use implicit null checks. Used by matching trap based 909 // nodes and NullCheck optimization. 910 set_allowed_deopt_reasons(); 911 912 // Now generate code 913 Code_Gen(); 914 if (failing()) return; 915 916 // Check if we want to skip execution of all compiled code. 917 { 918 #ifndef PRODUCT 919 if (OptoNoExecute) { 920 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 921 return; 922 } 923 #endif 924 TracePhase tp("install_code", &timers[_t_registerMethod]); 925 926 if (is_osr_compilation()) { 927 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 928 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 929 } else { 930 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 931 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 932 } 933 934 env()->register_method(_method, _entry_bci, 935 &_code_offsets, 936 _orig_pc_slot_offset_in_bytes, 937 code_buffer(), 938 frame_size_in_words(), _oop_map_set, 939 &_handler_table, &_inc_table, 940 compiler, 941 has_unsafe_access(), 942 SharedRuntime::is_wide_vector(max_vector_size()), 943 rtm_state() 944 ); 945 946 if (log() != NULL) // Print code cache state into compiler log 947 log()->code_cache_state(); 948 } 949 } 950 951 //------------------------------Compile---------------------------------------- 952 // Compile a runtime stub 953 Compile::Compile( ciEnv* ci_env, 954 TypeFunc_generator generator, 955 address stub_function, 956 const char *stub_name, 957 int is_fancy_jump, 958 bool pass_tls, 959 bool save_arg_registers, 960 bool return_pc, 961 DirectiveSet* directive) 962 : Phase(Compiler), 963 _env(ci_env), 964 _directive(directive), 965 _log(ci_env->log()), 966 _compile_id(0), 967 _save_argument_registers(save_arg_registers), 968 _method(NULL), 969 _stub_name(stub_name), 970 _stub_function(stub_function), 971 _stub_entry_point(NULL), 972 _entry_bci(InvocationEntryBci), 973 _initial_gvn(NULL), 974 _for_igvn(NULL), 975 _warm_calls(NULL), 976 _orig_pc_slot(0), 977 _orig_pc_slot_offset_in_bytes(0), 978 _subsume_loads(true), 979 _do_escape_analysis(false), 980 _eliminate_boxing(false), 981 _failure_reason(NULL), 982 _code_buffer("Compile::Fill_buffer"), 983 _has_method_handle_invokes(false), 984 _mach_constant_base_node(NULL), 985 _node_bundling_limit(0), 986 _node_bundling_base(NULL), 987 _java_calls(0), 988 _inner_loops(0), 989 #ifndef PRODUCT 990 _trace_opto_output(directive->TraceOptoOutputOption), 991 _in_dump_cnt(0), 992 _printer(NULL), 993 #endif 994 _comp_arena(mtCompiler), 995 _node_arena(mtCompiler), 996 _old_arena(mtCompiler), 997 _Compile_types(mtCompiler), 998 _dead_node_list(comp_arena()), 999 _dead_node_count(0), 1000 _congraph(NULL), 1001 _replay_inline_data(NULL), 1002 _number_of_mh_late_inlines(0), 1003 _inlining_progress(false), 1004 _inlining_incrementally(false), 1005 _print_inlining_list(NULL), 1006 _print_inlining_stream(NULL), 1007 _print_inlining_idx(0), 1008 _print_inlining_output(NULL), 1009 _allowed_reasons(0), 1010 _interpreter_frame_size(0), 1011 _max_node_limit(MaxNodeLimit), 1012 _has_reserved_stack_access(false) { 1013 C = this; 1014 1015 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1016 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1017 1018 #ifndef PRODUCT 1019 set_print_assembly(PrintFrameConverterAssembly); 1020 set_parsed_irreducible_loop(false); 1021 #endif 1022 set_has_irreducible_loop(false); // no loops 1023 1024 CompileWrapper cw(this); 1025 Init(/*AliasLevel=*/ 0); 1026 init_tf((*generator)()); 1027 1028 { 1029 // The following is a dummy for the sake of GraphKit::gen_stub 1030 Unique_Node_List for_igvn(comp_arena()); 1031 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1032 PhaseGVN gvn(Thread::current()->resource_area(),255); 1033 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1034 gvn.transform_no_reclaim(top()); 1035 1036 GraphKit kit; 1037 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1038 } 1039 1040 NOT_PRODUCT( verify_graph_edges(); ) 1041 Code_Gen(); 1042 if (failing()) return; 1043 1044 1045 // Entry point will be accessed using compile->stub_entry_point(); 1046 if (code_buffer() == NULL) { 1047 Matcher::soft_match_failure(); 1048 } else { 1049 if (PrintAssembly && (WizardMode || Verbose)) 1050 tty->print_cr("### Stub::%s", stub_name); 1051 1052 if (!failing()) { 1053 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1054 1055 // Make the NMethod 1056 // For now we mark the frame as never safe for profile stackwalking 1057 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1058 code_buffer(), 1059 CodeOffsets::frame_never_safe, 1060 // _code_offsets.value(CodeOffsets::Frame_Complete), 1061 frame_size_in_words(), 1062 _oop_map_set, 1063 save_arg_registers); 1064 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1065 1066 _stub_entry_point = rs->entry_point(); 1067 } 1068 } 1069 } 1070 1071 //------------------------------Init------------------------------------------- 1072 // Prepare for a single compilation 1073 void Compile::Init(int aliaslevel) { 1074 _unique = 0; 1075 _regalloc = NULL; 1076 1077 _tf = NULL; // filled in later 1078 _top = NULL; // cached later 1079 _matcher = NULL; // filled in later 1080 _cfg = NULL; // filled in later 1081 1082 set_24_bit_selection_and_mode(Use24BitFP, false); 1083 1084 _node_note_array = NULL; 1085 _default_node_notes = NULL; 1086 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1087 1088 _immutable_memory = NULL; // filled in at first inquiry 1089 1090 // Globally visible Nodes 1091 // First set TOP to NULL to give safe behavior during creation of RootNode 1092 set_cached_top_node(NULL); 1093 set_root(new RootNode()); 1094 // Now that you have a Root to point to, create the real TOP 1095 set_cached_top_node( new ConNode(Type::TOP) ); 1096 set_recent_alloc(NULL, NULL); 1097 1098 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1099 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1100 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1101 env()->set_dependencies(new Dependencies(env())); 1102 1103 _fixed_slots = 0; 1104 set_has_split_ifs(false); 1105 set_has_loops(has_method() && method()->has_loops()); // first approximation 1106 set_has_stringbuilder(false); 1107 set_has_boxed_value(false); 1108 _trap_can_recompile = false; // no traps emitted yet 1109 _major_progress = true; // start out assuming good things will happen 1110 set_has_unsafe_access(false); 1111 set_max_vector_size(0); 1112 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1113 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1114 set_decompile_count(0); 1115 1116 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1117 set_num_loop_opts(LoopOptsCount); 1118 set_do_inlining(Inline); 1119 set_max_inline_size(MaxInlineSize); 1120 set_freq_inline_size(FreqInlineSize); 1121 set_do_scheduling(OptoScheduling); 1122 set_do_count_invocations(false); 1123 set_do_method_data_update(false); 1124 1125 set_do_vector_loop(false); 1126 1127 if (AllowVectorizeOnDemand) { 1128 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1129 set_do_vector_loop(true); 1130 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());}) 1131 } else if (has_method() && method()->name() != 0 && 1132 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1133 set_do_vector_loop(true); 1134 } 1135 } 1136 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1137 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());}) 1138 1139 set_age_code(has_method() && method()->profile_aging()); 1140 set_rtm_state(NoRTM); // No RTM lock eliding by default 1141 _max_node_limit = _directive->MaxNodeLimitOption; 1142 1143 #if INCLUDE_RTM_OPT 1144 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1145 int rtm_state = method()->method_data()->rtm_state(); 1146 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1147 // Don't generate RTM lock eliding code. 1148 set_rtm_state(NoRTM); 1149 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1150 // Generate RTM lock eliding code without abort ratio calculation code. 1151 set_rtm_state(UseRTM); 1152 } else if (UseRTMDeopt) { 1153 // Generate RTM lock eliding code and include abort ratio calculation 1154 // code if UseRTMDeopt is on. 1155 set_rtm_state(ProfileRTM); 1156 } 1157 } 1158 #endif 1159 if (debug_info()->recording_non_safepoints()) { 1160 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1161 (comp_arena(), 8, 0, NULL)); 1162 set_default_node_notes(Node_Notes::make(this)); 1163 } 1164 1165 // // -- Initialize types before each compile -- 1166 // // Update cached type information 1167 // if( _method && _method->constants() ) 1168 // Type::update_loaded_types(_method, _method->constants()); 1169 1170 // Init alias_type map. 1171 if (!_do_escape_analysis && aliaslevel == 3) 1172 aliaslevel = 2; // No unique types without escape analysis 1173 _AliasLevel = aliaslevel; 1174 const int grow_ats = 16; 1175 _max_alias_types = grow_ats; 1176 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1177 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1178 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1179 { 1180 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1181 } 1182 // Initialize the first few types. 1183 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1184 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1185 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1186 _num_alias_types = AliasIdxRaw+1; 1187 // Zero out the alias type cache. 1188 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1189 // A NULL adr_type hits in the cache right away. Preload the right answer. 1190 probe_alias_cache(NULL)->_index = AliasIdxTop; 1191 1192 _intrinsics = NULL; 1193 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1194 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1195 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1196 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1197 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1198 _value_type_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 1199 register_library_intrinsics(); 1200 } 1201 1202 //---------------------------init_start---------------------------------------- 1203 // Install the StartNode on this compile object. 1204 void Compile::init_start(StartNode* s) { 1205 if (failing()) 1206 return; // already failing 1207 assert(s == start(), ""); 1208 } 1209 1210 /** 1211 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1212 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1213 * the ideal graph. 1214 */ 1215 StartNode* Compile::start() const { 1216 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1217 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1218 Node* start = root()->fast_out(i); 1219 if (start->is_Start()) { 1220 return start->as_Start(); 1221 } 1222 } 1223 fatal("Did not find Start node!"); 1224 return NULL; 1225 } 1226 1227 //-------------------------------immutable_memory------------------------------------- 1228 // Access immutable memory 1229 Node* Compile::immutable_memory() { 1230 if (_immutable_memory != NULL) { 1231 return _immutable_memory; 1232 } 1233 StartNode* s = start(); 1234 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1235 Node *p = s->fast_out(i); 1236 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1237 _immutable_memory = p; 1238 return _immutable_memory; 1239 } 1240 } 1241 ShouldNotReachHere(); 1242 return NULL; 1243 } 1244 1245 //----------------------set_cached_top_node------------------------------------ 1246 // Install the cached top node, and make sure Node::is_top works correctly. 1247 void Compile::set_cached_top_node(Node* tn) { 1248 if (tn != NULL) verify_top(tn); 1249 Node* old_top = _top; 1250 _top = tn; 1251 // Calling Node::setup_is_top allows the nodes the chance to adjust 1252 // their _out arrays. 1253 if (_top != NULL) _top->setup_is_top(); 1254 if (old_top != NULL) old_top->setup_is_top(); 1255 assert(_top == NULL || top()->is_top(), ""); 1256 } 1257 1258 #ifdef ASSERT 1259 uint Compile::count_live_nodes_by_graph_walk() { 1260 Unique_Node_List useful(comp_arena()); 1261 // Get useful node list by walking the graph. 1262 identify_useful_nodes(useful); 1263 return useful.size(); 1264 } 1265 1266 void Compile::print_missing_nodes() { 1267 1268 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1269 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1270 return; 1271 } 1272 1273 // This is an expensive function. It is executed only when the user 1274 // specifies VerifyIdealNodeCount option or otherwise knows the 1275 // additional work that needs to be done to identify reachable nodes 1276 // by walking the flow graph and find the missing ones using 1277 // _dead_node_list. 1278 1279 Unique_Node_List useful(comp_arena()); 1280 // Get useful node list by walking the graph. 1281 identify_useful_nodes(useful); 1282 1283 uint l_nodes = C->live_nodes(); 1284 uint l_nodes_by_walk = useful.size(); 1285 1286 if (l_nodes != l_nodes_by_walk) { 1287 if (_log != NULL) { 1288 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1289 _log->stamp(); 1290 _log->end_head(); 1291 } 1292 VectorSet& useful_member_set = useful.member_set(); 1293 int last_idx = l_nodes_by_walk; 1294 for (int i = 0; i < last_idx; i++) { 1295 if (useful_member_set.test(i)) { 1296 if (_dead_node_list.test(i)) { 1297 if (_log != NULL) { 1298 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1299 } 1300 if (PrintIdealNodeCount) { 1301 // Print the log message to tty 1302 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1303 useful.at(i)->dump(); 1304 } 1305 } 1306 } 1307 else if (! _dead_node_list.test(i)) { 1308 if (_log != NULL) { 1309 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1310 } 1311 if (PrintIdealNodeCount) { 1312 // Print the log message to tty 1313 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1314 } 1315 } 1316 } 1317 if (_log != NULL) { 1318 _log->tail("mismatched_nodes"); 1319 } 1320 } 1321 } 1322 void Compile::record_modified_node(Node* n) { 1323 if (_modified_nodes != NULL && !_inlining_incrementally && 1324 n->outcnt() != 0 && !n->is_Con()) { 1325 _modified_nodes->push(n); 1326 } 1327 } 1328 1329 void Compile::remove_modified_node(Node* n) { 1330 if (_modified_nodes != NULL) { 1331 _modified_nodes->remove(n); 1332 } 1333 } 1334 #endif 1335 1336 #ifndef PRODUCT 1337 void Compile::verify_top(Node* tn) const { 1338 if (tn != NULL) { 1339 assert(tn->is_Con(), "top node must be a constant"); 1340 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1341 assert(tn->in(0) != NULL, "must have live top node"); 1342 } 1343 } 1344 #endif 1345 1346 1347 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1348 1349 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1350 guarantee(arr != NULL, ""); 1351 int num_blocks = arr->length(); 1352 if (grow_by < num_blocks) grow_by = num_blocks; 1353 int num_notes = grow_by * _node_notes_block_size; 1354 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1355 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1356 while (num_notes > 0) { 1357 arr->append(notes); 1358 notes += _node_notes_block_size; 1359 num_notes -= _node_notes_block_size; 1360 } 1361 assert(num_notes == 0, "exact multiple, please"); 1362 } 1363 1364 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1365 if (source == NULL || dest == NULL) return false; 1366 1367 if (dest->is_Con()) 1368 return false; // Do not push debug info onto constants. 1369 1370 #ifdef ASSERT 1371 // Leave a bread crumb trail pointing to the original node: 1372 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1373 dest->set_debug_orig(source); 1374 } 1375 #endif 1376 1377 if (node_note_array() == NULL) 1378 return false; // Not collecting any notes now. 1379 1380 // This is a copy onto a pre-existing node, which may already have notes. 1381 // If both nodes have notes, do not overwrite any pre-existing notes. 1382 Node_Notes* source_notes = node_notes_at(source->_idx); 1383 if (source_notes == NULL || source_notes->is_clear()) return false; 1384 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1385 if (dest_notes == NULL || dest_notes->is_clear()) { 1386 return set_node_notes_at(dest->_idx, source_notes); 1387 } 1388 1389 Node_Notes merged_notes = (*source_notes); 1390 // The order of operations here ensures that dest notes will win... 1391 merged_notes.update_from(dest_notes); 1392 return set_node_notes_at(dest->_idx, &merged_notes); 1393 } 1394 1395 1396 //--------------------------allow_range_check_smearing------------------------- 1397 // Gating condition for coalescing similar range checks. 1398 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1399 // single covering check that is at least as strong as any of them. 1400 // If the optimization succeeds, the simplified (strengthened) range check 1401 // will always succeed. If it fails, we will deopt, and then give up 1402 // on the optimization. 1403 bool Compile::allow_range_check_smearing() const { 1404 // If this method has already thrown a range-check, 1405 // assume it was because we already tried range smearing 1406 // and it failed. 1407 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1408 return !already_trapped; 1409 } 1410 1411 1412 //------------------------------flatten_alias_type----------------------------- 1413 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1414 int offset = tj->offset(); 1415 TypePtr::PTR ptr = tj->ptr(); 1416 1417 // Known instance (scalarizable allocation) alias only with itself. 1418 bool is_known_inst = tj->isa_oopptr() != NULL && 1419 tj->is_oopptr()->is_known_instance(); 1420 1421 // Process weird unsafe references. 1422 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1423 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1424 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1425 tj = TypeOopPtr::BOTTOM; 1426 ptr = tj->ptr(); 1427 offset = tj->offset(); 1428 } 1429 1430 // Array pointers need some flattening 1431 const TypeAryPtr *ta = tj->isa_aryptr(); 1432 if (ta && ta->is_stable()) { 1433 // Erase stability property for alias analysis. 1434 tj = ta = ta->cast_to_stable(false); 1435 } 1436 if( ta && is_known_inst ) { 1437 if ( offset != Type::OffsetBot && 1438 offset > arrayOopDesc::length_offset_in_bytes() ) { 1439 offset = Type::OffsetBot; // Flatten constant access into array body only 1440 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id()); 1441 } 1442 } else if( ta && _AliasLevel >= 2 ) { 1443 // For arrays indexed by constant indices, we flatten the alias 1444 // space to include all of the array body. Only the header, klass 1445 // and array length can be accessed un-aliased. 1446 // For flattened value type array, each field has its own slice so 1447 // we must include the field offset. 1448 if( offset != Type::OffsetBot ) { 1449 if( ta->const_oop() ) { // MethodData* or Method* 1450 offset = Type::OffsetBot; // Flatten constant access into array body 1451 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1452 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1453 // range is OK as-is. 1454 tj = ta = TypeAryPtr::RANGE; 1455 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1456 tj = TypeInstPtr::KLASS; // all klass loads look alike 1457 ta = TypeAryPtr::RANGE; // generic ignored junk 1458 ptr = TypePtr::BotPTR; 1459 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1460 tj = TypeInstPtr::MARK; 1461 ta = TypeAryPtr::RANGE; // generic ignored junk 1462 ptr = TypePtr::BotPTR; 1463 } else { // Random constant offset into array body 1464 offset = Type::OffsetBot; // Flatten constant access into array body 1465 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1466 } 1467 } 1468 // Arrays of fixed size alias with arrays of unknown size. 1469 if (ta->size() != TypeInt::POS) { 1470 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1471 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1472 } 1473 // Arrays of known objects become arrays of unknown objects. 1474 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1475 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1476 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1477 } 1478 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1479 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1480 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1481 } 1482 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1483 // cannot be distinguished by bytecode alone. 1484 if (ta->elem() == TypeInt::BOOL) { 1485 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1486 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1487 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset()); 1488 } 1489 // During the 2nd round of IterGVN, NotNull castings are removed. 1490 // Make sure the Bottom and NotNull variants alias the same. 1491 // Also, make sure exact and non-exact variants alias the same. 1492 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1493 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1494 } 1495 } 1496 1497 // Oop pointers need some flattening 1498 const TypeInstPtr *to = tj->isa_instptr(); 1499 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1500 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1501 if( ptr == TypePtr::Constant ) { 1502 if (to->klass() != ciEnv::current()->Class_klass() || 1503 offset < k->size_helper() * wordSize) { 1504 // No constant oop pointers (such as Strings); they alias with 1505 // unknown strings. 1506 assert(!is_known_inst, "not scalarizable allocation"); 1507 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1508 } 1509 } else if( is_known_inst ) { 1510 tj = to; // Keep NotNull and klass_is_exact for instance type 1511 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1512 // During the 2nd round of IterGVN, NotNull castings are removed. 1513 // Make sure the Bottom and NotNull variants alias the same. 1514 // Also, make sure exact and non-exact variants alias the same. 1515 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1516 } 1517 if (to->speculative() != NULL) { 1518 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->instance_id()); 1519 } 1520 // Canonicalize the holder of this field 1521 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1522 // First handle header references such as a LoadKlassNode, even if the 1523 // object's klass is unloaded at compile time (4965979). 1524 if (!is_known_inst) { // Do it only for non-instance types 1525 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset)); 1526 } 1527 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1528 // Static fields are in the space above the normal instance 1529 // fields in the java.lang.Class instance. 1530 if (to->klass() != ciEnv::current()->Class_klass()) { 1531 to = NULL; 1532 tj = TypeOopPtr::BOTTOM; 1533 offset = tj->offset(); 1534 } 1535 } else { 1536 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1537 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1538 if( is_known_inst ) { 1539 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), to->instance_id()); 1540 } else { 1541 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset)); 1542 } 1543 } 1544 } 1545 } 1546 1547 // Klass pointers to object array klasses need some flattening 1548 const TypeKlassPtr *tk = tj->isa_klassptr(); 1549 if( tk ) { 1550 // If we are referencing a field within a Klass, we need 1551 // to assume the worst case of an Object. Both exact and 1552 // inexact types must flatten to the same alias class so 1553 // use NotNull as the PTR. 1554 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1555 1556 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1557 TypeKlassPtr::OBJECT->klass(), 1558 Type::Offset(offset)); 1559 } 1560 1561 ciKlass* klass = tk->klass(); 1562 if (klass != NULL && klass->is_obj_array_klass()) { 1563 ciKlass* k = TypeAryPtr::OOPS->klass(); 1564 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1565 k = TypeInstPtr::BOTTOM->klass(); 1566 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset)); 1567 } 1568 1569 // Check for precise loads from the primary supertype array and force them 1570 // to the supertype cache alias index. Check for generic array loads from 1571 // the primary supertype array and also force them to the supertype cache 1572 // alias index. Since the same load can reach both, we need to merge 1573 // these 2 disparate memories into the same alias class. Since the 1574 // primary supertype array is read-only, there's no chance of confusion 1575 // where we bypass an array load and an array store. 1576 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1577 if (offset == Type::OffsetBot || 1578 (offset >= primary_supers_offset && 1579 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1580 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1581 offset = in_bytes(Klass::secondary_super_cache_offset()); 1582 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset)); 1583 } 1584 } 1585 1586 // Flatten all Raw pointers together. 1587 if (tj->base() == Type::RawPtr) 1588 tj = TypeRawPtr::BOTTOM; 1589 1590 if (tj->base() == Type::AnyPtr) 1591 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1592 1593 // Flatten all to bottom for now 1594 switch( _AliasLevel ) { 1595 case 0: 1596 tj = TypePtr::BOTTOM; 1597 break; 1598 case 1: // Flatten to: oop, static, field or array 1599 switch (tj->base()) { 1600 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1601 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1602 case Type::AryPtr: // do not distinguish arrays at all 1603 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1604 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1605 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1606 default: ShouldNotReachHere(); 1607 } 1608 break; 1609 case 2: // No collapsing at level 2; keep all splits 1610 case 3: // No collapsing at level 3; keep all splits 1611 break; 1612 default: 1613 Unimplemented(); 1614 } 1615 1616 offset = tj->offset(); 1617 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1618 1619 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1620 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1621 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1622 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1623 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1624 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1625 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1626 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1627 assert( tj->ptr() != TypePtr::TopPTR && 1628 tj->ptr() != TypePtr::AnyNull && 1629 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1630 // assert( tj->ptr() != TypePtr::Constant || 1631 // tj->base() == Type::RawPtr || 1632 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1633 1634 return tj; 1635 } 1636 1637 void Compile::AliasType::Init(int i, const TypePtr* at) { 1638 _index = i; 1639 _adr_type = at; 1640 _field = NULL; 1641 _element = NULL; 1642 _is_rewritable = true; // default 1643 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1644 if (atoop != NULL && atoop->is_known_instance()) { 1645 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1646 _general_index = Compile::current()->get_alias_index(gt); 1647 } else { 1648 _general_index = 0; 1649 } 1650 } 1651 1652 BasicType Compile::AliasType::basic_type() const { 1653 if (element() != NULL) { 1654 const Type* element = adr_type()->is_aryptr()->elem(); 1655 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1656 } if (field() != NULL) { 1657 return field()->layout_type(); 1658 } else { 1659 return T_ILLEGAL; // unknown 1660 } 1661 } 1662 1663 //---------------------------------print_on------------------------------------ 1664 #ifndef PRODUCT 1665 void Compile::AliasType::print_on(outputStream* st) { 1666 if (index() < 10) 1667 st->print("@ <%d> ", index()); 1668 else st->print("@ <%d>", index()); 1669 st->print(is_rewritable() ? " " : " RO"); 1670 int offset = adr_type()->offset(); 1671 if (offset == Type::OffsetBot) 1672 st->print(" +any"); 1673 else st->print(" +%-3d", offset); 1674 st->print(" in "); 1675 adr_type()->dump_on(st); 1676 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1677 if (field() != NULL && tjp) { 1678 if (tjp->klass() != field()->holder() || 1679 tjp->offset() != field()->offset_in_bytes()) { 1680 st->print(" != "); 1681 field()->print(); 1682 st->print(" ***"); 1683 } 1684 } 1685 } 1686 1687 void print_alias_types() { 1688 Compile* C = Compile::current(); 1689 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1690 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1691 C->alias_type(idx)->print_on(tty); 1692 tty->cr(); 1693 } 1694 } 1695 #endif 1696 1697 1698 //----------------------------probe_alias_cache-------------------------------- 1699 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1700 intptr_t key = (intptr_t) adr_type; 1701 key ^= key >> logAliasCacheSize; 1702 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1703 } 1704 1705 1706 //-----------------------------grow_alias_types-------------------------------- 1707 void Compile::grow_alias_types() { 1708 const int old_ats = _max_alias_types; // how many before? 1709 const int new_ats = old_ats; // how many more? 1710 const int grow_ats = old_ats+new_ats; // how many now? 1711 _max_alias_types = grow_ats; 1712 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1713 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1714 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1715 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1716 } 1717 1718 1719 //--------------------------------find_alias_type------------------------------ 1720 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1721 if (_AliasLevel == 0) 1722 return alias_type(AliasIdxBot); 1723 1724 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1725 if (ace->_adr_type == adr_type) { 1726 return alias_type(ace->_index); 1727 } 1728 1729 // Handle special cases. 1730 if (adr_type == NULL) return alias_type(AliasIdxTop); 1731 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1732 1733 // Do it the slow way. 1734 const TypePtr* flat = flatten_alias_type(adr_type); 1735 1736 #ifdef ASSERT 1737 { 1738 ResourceMark rm; 1739 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1740 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1741 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1742 Type::str(adr_type)); 1743 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1744 const TypeOopPtr* foop = flat->is_oopptr(); 1745 // Scalarizable allocations have exact klass always. 1746 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1747 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1748 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1749 Type::str(foop), Type::str(xoop)); 1750 } 1751 } 1752 #endif 1753 1754 int idx = AliasIdxTop; 1755 for (int i = 0; i < num_alias_types(); i++) { 1756 if (alias_type(i)->adr_type() == flat) { 1757 idx = i; 1758 break; 1759 } 1760 } 1761 1762 if (idx == AliasIdxTop) { 1763 if (no_create) return NULL; 1764 // Grow the array if necessary. 1765 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1766 // Add a new alias type. 1767 idx = _num_alias_types++; 1768 _alias_types[idx]->Init(idx, flat); 1769 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1770 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1771 if (flat->isa_instptr()) { 1772 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1773 && flat->is_instptr()->klass() == env()->Class_klass()) 1774 alias_type(idx)->set_rewritable(false); 1775 } 1776 ciField* field = NULL; 1777 if (flat->isa_aryptr()) { 1778 #ifdef ASSERT 1779 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1780 // (T_BYTE has the weakest alignment and size restrictions...) 1781 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1782 #endif 1783 const Type* elemtype = flat->is_aryptr()->elem(); 1784 if (flat->offset() == TypePtr::OffsetBot) { 1785 alias_type(idx)->set_element(elemtype); 1786 } 1787 int field_offset = flat->is_aryptr()->field_offset().get(); 1788 if (elemtype->isa_valuetype() && field_offset != Type::OffsetBot) { 1789 ciValueKlass* vk = elemtype->is_valuetype()->value_klass(); 1790 field_offset += vk->first_field_offset(); 1791 field = vk->get_field_by_offset(field_offset, false); 1792 } 1793 } 1794 if (flat->isa_klassptr()) { 1795 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1796 alias_type(idx)->set_rewritable(false); 1797 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1798 alias_type(idx)->set_rewritable(false); 1799 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1800 alias_type(idx)->set_rewritable(false); 1801 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1802 alias_type(idx)->set_rewritable(false); 1803 } 1804 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1805 // but the base pointer type is not distinctive enough to identify 1806 // references into JavaThread.) 1807 1808 // Check for final fields. 1809 const TypeInstPtr* tinst = flat->isa_instptr(); 1810 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1811 if (tinst->const_oop() != NULL && 1812 tinst->klass() == ciEnv::current()->Class_klass() && 1813 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1814 // static field 1815 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1816 field = k->get_field_by_offset(tinst->offset(), true); 1817 } else if (tinst->klass()->is_valuetype()) { 1818 // Value type field 1819 ciValueKlass* vk = tinst->value_klass(); 1820 field = vk->get_field_by_offset(tinst->offset(), false); 1821 } else { 1822 ciInstanceKlass* k = tinst->klass()->as_instance_klass(); 1823 field = k->get_field_by_offset(tinst->offset(), false); 1824 } 1825 } 1826 assert(field == NULL || 1827 original_field == NULL || 1828 (field->holder() == original_field->holder() && 1829 field->offset() == original_field->offset() && 1830 field->is_static() == original_field->is_static()), "wrong field?"); 1831 // Set field() and is_rewritable() attributes. 1832 if (field != NULL) alias_type(idx)->set_field(field); 1833 } 1834 1835 // Fill the cache for next time. 1836 ace->_adr_type = adr_type; 1837 ace->_index = idx; 1838 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1839 1840 // Might as well try to fill the cache for the flattened version, too. 1841 AliasCacheEntry* face = probe_alias_cache(flat); 1842 if (face->_adr_type == NULL) { 1843 face->_adr_type = flat; 1844 face->_index = idx; 1845 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1846 } 1847 1848 return alias_type(idx); 1849 } 1850 1851 1852 Compile::AliasType* Compile::alias_type(ciField* field) { 1853 const TypeOopPtr* t; 1854 if (field->is_static()) 1855 t = TypeInstPtr::make(field->holder()->java_mirror()); 1856 else 1857 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1858 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1859 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1860 return atp; 1861 } 1862 1863 1864 //------------------------------have_alias_type-------------------------------- 1865 bool Compile::have_alias_type(const TypePtr* adr_type) { 1866 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1867 if (ace->_adr_type == adr_type) { 1868 return true; 1869 } 1870 1871 // Handle special cases. 1872 if (adr_type == NULL) return true; 1873 if (adr_type == TypePtr::BOTTOM) return true; 1874 1875 return find_alias_type(adr_type, true, NULL) != NULL; 1876 } 1877 1878 //-----------------------------must_alias-------------------------------------- 1879 // True if all values of the given address type are in the given alias category. 1880 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1881 if (alias_idx == AliasIdxBot) return true; // the universal category 1882 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1883 if (alias_idx == AliasIdxTop) return false; // the empty category 1884 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1885 1886 // the only remaining possible overlap is identity 1887 int adr_idx = get_alias_index(adr_type); 1888 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1889 assert(adr_idx == alias_idx || 1890 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1891 && adr_type != TypeOopPtr::BOTTOM), 1892 "should not be testing for overlap with an unsafe pointer"); 1893 return adr_idx == alias_idx; 1894 } 1895 1896 //------------------------------can_alias-------------------------------------- 1897 // True if any values of the given address type are in the given alias category. 1898 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1899 if (alias_idx == AliasIdxTop) return false; // the empty category 1900 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1901 if (alias_idx == AliasIdxBot) return true; // the universal category 1902 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1903 1904 // the only remaining possible overlap is identity 1905 int adr_idx = get_alias_index(adr_type); 1906 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1907 return adr_idx == alias_idx; 1908 } 1909 1910 1911 1912 //---------------------------pop_warm_call------------------------------------- 1913 WarmCallInfo* Compile::pop_warm_call() { 1914 WarmCallInfo* wci = _warm_calls; 1915 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1916 return wci; 1917 } 1918 1919 //----------------------------Inline_Warm-------------------------------------- 1920 int Compile::Inline_Warm() { 1921 // If there is room, try to inline some more warm call sites. 1922 // %%% Do a graph index compaction pass when we think we're out of space? 1923 if (!InlineWarmCalls) return 0; 1924 1925 int calls_made_hot = 0; 1926 int room_to_grow = NodeCountInliningCutoff - unique(); 1927 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1928 int amount_grown = 0; 1929 WarmCallInfo* call; 1930 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1931 int est_size = (int)call->size(); 1932 if (est_size > (room_to_grow - amount_grown)) { 1933 // This one won't fit anyway. Get rid of it. 1934 call->make_cold(); 1935 continue; 1936 } 1937 call->make_hot(); 1938 calls_made_hot++; 1939 amount_grown += est_size; 1940 amount_to_grow -= est_size; 1941 } 1942 1943 if (calls_made_hot > 0) set_major_progress(); 1944 return calls_made_hot; 1945 } 1946 1947 1948 //----------------------------Finish_Warm-------------------------------------- 1949 void Compile::Finish_Warm() { 1950 if (!InlineWarmCalls) return; 1951 if (failing()) return; 1952 if (warm_calls() == NULL) return; 1953 1954 // Clean up loose ends, if we are out of space for inlining. 1955 WarmCallInfo* call; 1956 while ((call = pop_warm_call()) != NULL) { 1957 call->make_cold(); 1958 } 1959 } 1960 1961 //---------------------cleanup_loop_predicates----------------------- 1962 // Remove the opaque nodes that protect the predicates so that all unused 1963 // checks and uncommon_traps will be eliminated from the ideal graph 1964 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1965 if (predicate_count()==0) return; 1966 for (int i = predicate_count(); i > 0; i--) { 1967 Node * n = predicate_opaque1_node(i-1); 1968 assert(n->Opcode() == Op_Opaque1, "must be"); 1969 igvn.replace_node(n, n->in(1)); 1970 } 1971 assert(predicate_count()==0, "should be clean!"); 1972 } 1973 1974 void Compile::add_range_check_cast(Node* n) { 1975 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1976 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1977 _range_check_casts->append(n); 1978 } 1979 1980 // Remove all range check dependent CastIINodes. 1981 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1982 for (int i = range_check_cast_count(); i > 0; i--) { 1983 Node* cast = range_check_cast_node(i-1); 1984 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1985 igvn.replace_node(cast, cast->in(1)); 1986 } 1987 assert(range_check_cast_count() == 0, "should be empty"); 1988 } 1989 1990 void Compile::add_opaque4_node(Node* n) { 1991 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 1992 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 1993 _opaque4_nodes->append(n); 1994 } 1995 1996 // Remove all Opaque4 nodes. 1997 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 1998 for (int i = opaque4_count(); i > 0; i--) { 1999 Node* opaq = opaque4_node(i-1); 2000 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 2001 igvn.replace_node(opaq, opaq->in(2)); 2002 } 2003 assert(opaque4_count() == 0, "should be empty"); 2004 } 2005 2006 void Compile::add_value_type(Node* n) { 2007 assert(n->is_ValueTypeBase(), "unexpected node"); 2008 if (_value_type_nodes != NULL) { 2009 _value_type_nodes->push(n); 2010 } 2011 } 2012 2013 void Compile::remove_value_type(Node* n) { 2014 assert(n->is_ValueTypeBase(), "unexpected node"); 2015 if (_value_type_nodes != NULL) { 2016 _value_type_nodes->remove(n); 2017 } 2018 } 2019 2020 void Compile::process_value_types(PhaseIterGVN &igvn) { 2021 // Make value types scalar in safepoints 2022 while (_value_type_nodes->size() != 0) { 2023 ValueTypeBaseNode* vt = _value_type_nodes->pop()->as_ValueTypeBase(); 2024 vt->make_scalar_in_safepoints(igvn.C->root(), &igvn); 2025 if (vt->is_ValueTypePtr()) { 2026 igvn.replace_node(vt, vt->get_oop()); 2027 } 2028 } 2029 _value_type_nodes = NULL; 2030 igvn.optimize(); 2031 } 2032 2033 // StringOpts and late inlining of string methods 2034 void Compile::inline_string_calls(bool parse_time) { 2035 { 2036 // remove useless nodes to make the usage analysis simpler 2037 ResourceMark rm; 2038 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2039 } 2040 2041 { 2042 ResourceMark rm; 2043 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2044 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2045 print_method(PHASE_AFTER_STRINGOPTS, 3); 2046 } 2047 2048 // now inline anything that we skipped the first time around 2049 if (!parse_time) { 2050 _late_inlines_pos = _late_inlines.length(); 2051 } 2052 2053 while (_string_late_inlines.length() > 0) { 2054 CallGenerator* cg = _string_late_inlines.pop(); 2055 cg->do_late_inline(); 2056 if (failing()) return; 2057 } 2058 _string_late_inlines.trunc_to(0); 2059 } 2060 2061 // Late inlining of boxing methods 2062 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2063 if (_boxing_late_inlines.length() > 0) { 2064 assert(has_boxed_value(), "inconsistent"); 2065 2066 PhaseGVN* gvn = initial_gvn(); 2067 set_inlining_incrementally(true); 2068 2069 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2070 for_igvn()->clear(); 2071 gvn->replace_with(&igvn); 2072 2073 _late_inlines_pos = _late_inlines.length(); 2074 2075 while (_boxing_late_inlines.length() > 0) { 2076 CallGenerator* cg = _boxing_late_inlines.pop(); 2077 cg->do_late_inline(); 2078 if (failing()) return; 2079 } 2080 _boxing_late_inlines.trunc_to(0); 2081 2082 { 2083 ResourceMark rm; 2084 PhaseRemoveUseless pru(gvn, for_igvn()); 2085 } 2086 2087 igvn = PhaseIterGVN(gvn); 2088 igvn.optimize(); 2089 2090 set_inlining_progress(false); 2091 set_inlining_incrementally(false); 2092 } 2093 } 2094 2095 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 2096 assert(IncrementalInline, "incremental inlining should be on"); 2097 PhaseGVN* gvn = initial_gvn(); 2098 2099 set_inlining_progress(false); 2100 for_igvn()->clear(); 2101 gvn->replace_with(&igvn); 2102 2103 { 2104 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2105 int i = 0; 2106 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2107 CallGenerator* cg = _late_inlines.at(i); 2108 _late_inlines_pos = i+1; 2109 cg->do_late_inline(); 2110 if (failing()) return; 2111 } 2112 int j = 0; 2113 for (; i < _late_inlines.length(); i++, j++) { 2114 _late_inlines.at_put(j, _late_inlines.at(i)); 2115 } 2116 _late_inlines.trunc_to(j); 2117 } 2118 2119 { 2120 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2121 ResourceMark rm; 2122 PhaseRemoveUseless pru(gvn, for_igvn()); 2123 } 2124 2125 { 2126 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2127 igvn = PhaseIterGVN(gvn); 2128 } 2129 } 2130 2131 // Perform incremental inlining until bound on number of live nodes is reached 2132 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2133 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2134 2135 PhaseGVN* gvn = initial_gvn(); 2136 2137 set_inlining_incrementally(true); 2138 set_inlining_progress(true); 2139 uint low_live_nodes = 0; 2140 2141 while(inlining_progress() && _late_inlines.length() > 0) { 2142 2143 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2144 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2145 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2146 // PhaseIdealLoop is expensive so we only try it once we are 2147 // out of live nodes and we only try it again if the previous 2148 // helped got the number of nodes down significantly 2149 PhaseIdealLoop ideal_loop( igvn, false, true ); 2150 if (failing()) return; 2151 low_live_nodes = live_nodes(); 2152 _major_progress = true; 2153 } 2154 2155 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2156 break; 2157 } 2158 } 2159 2160 inline_incrementally_one(igvn); 2161 2162 if (failing()) return; 2163 2164 { 2165 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2166 igvn.optimize(); 2167 } 2168 2169 if (failing()) return; 2170 } 2171 2172 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2173 2174 if (_string_late_inlines.length() > 0) { 2175 assert(has_stringbuilder(), "inconsistent"); 2176 for_igvn()->clear(); 2177 initial_gvn()->replace_with(&igvn); 2178 2179 inline_string_calls(false); 2180 2181 if (failing()) return; 2182 2183 { 2184 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2185 ResourceMark rm; 2186 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2187 } 2188 2189 { 2190 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2191 igvn = PhaseIterGVN(gvn); 2192 igvn.optimize(); 2193 } 2194 } 2195 2196 set_inlining_incrementally(false); 2197 } 2198 2199 2200 //------------------------------Optimize--------------------------------------- 2201 // Given a graph, optimize it. 2202 void Compile::Optimize() { 2203 TracePhase tp("optimizer", &timers[_t_optimizer]); 2204 2205 #ifndef PRODUCT 2206 if (_directive->BreakAtCompileOption) { 2207 BREAKPOINT; 2208 } 2209 2210 #endif 2211 2212 ResourceMark rm; 2213 int loop_opts_cnt; 2214 2215 print_inlining_reinit(); 2216 2217 NOT_PRODUCT( verify_graph_edges(); ) 2218 2219 print_method(PHASE_AFTER_PARSING); 2220 2221 { 2222 // Iterative Global Value Numbering, including ideal transforms 2223 // Initialize IterGVN with types and values from parse-time GVN 2224 PhaseIterGVN igvn(initial_gvn()); 2225 #ifdef ASSERT 2226 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2227 #endif 2228 { 2229 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2230 igvn.optimize(); 2231 } 2232 2233 print_method(PHASE_ITER_GVN1, 2); 2234 2235 if (failing()) return; 2236 2237 inline_incrementally(igvn); 2238 2239 print_method(PHASE_INCREMENTAL_INLINE, 2); 2240 2241 if (failing()) return; 2242 2243 if (eliminate_boxing()) { 2244 // Inline valueOf() methods now. 2245 inline_boxing_calls(igvn); 2246 2247 if (AlwaysIncrementalInline) { 2248 inline_incrementally(igvn); 2249 } 2250 2251 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2252 2253 if (failing()) return; 2254 } 2255 2256 // Remove the speculative part of types and clean up the graph from 2257 // the extra CastPP nodes whose only purpose is to carry them. Do 2258 // that early so that optimizations are not disrupted by the extra 2259 // CastPP nodes. 2260 remove_speculative_types(igvn); 2261 2262 // No more new expensive nodes will be added to the list from here 2263 // so keep only the actual candidates for optimizations. 2264 cleanup_expensive_nodes(igvn); 2265 2266 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2267 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2268 initial_gvn()->replace_with(&igvn); 2269 for_igvn()->clear(); 2270 Unique_Node_List new_worklist(C->comp_arena()); 2271 { 2272 ResourceMark rm; 2273 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2274 } 2275 set_for_igvn(&new_worklist); 2276 igvn = PhaseIterGVN(initial_gvn()); 2277 igvn.optimize(); 2278 } 2279 2280 if (_value_type_nodes->size() > 0) { 2281 // Do this once all inlining is over to avoid getting inconsistent debug info 2282 process_value_types(igvn); 2283 } 2284 2285 // Perform escape analysis 2286 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2287 if (has_loops()) { 2288 // Cleanup graph (remove dead nodes). 2289 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2290 PhaseIdealLoop ideal_loop( igvn, false, true ); 2291 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2292 if (failing()) return; 2293 } 2294 ConnectionGraph::do_analysis(this, &igvn); 2295 2296 if (failing()) return; 2297 2298 // Optimize out fields loads from scalar replaceable allocations. 2299 igvn.optimize(); 2300 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2301 2302 if (failing()) return; 2303 2304 if (congraph() != NULL && macro_count() > 0) { 2305 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2306 PhaseMacroExpand mexp(igvn); 2307 mexp.eliminate_macro_nodes(); 2308 igvn.set_delay_transform(false); 2309 2310 igvn.optimize(); 2311 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2312 2313 if (failing()) return; 2314 } 2315 } 2316 2317 // Loop transforms on the ideal graph. Range Check Elimination, 2318 // peeling, unrolling, etc. 2319 2320 // Set loop opts counter 2321 loop_opts_cnt = num_loop_opts(); 2322 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2323 { 2324 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2325 PhaseIdealLoop ideal_loop( igvn, true ); 2326 loop_opts_cnt--; 2327 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2328 if (failing()) return; 2329 } 2330 // Loop opts pass if partial peeling occurred in previous pass 2331 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2332 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2333 PhaseIdealLoop ideal_loop( igvn, false ); 2334 loop_opts_cnt--; 2335 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2336 if (failing()) return; 2337 } 2338 // Loop opts pass for loop-unrolling before CCP 2339 if(major_progress() && (loop_opts_cnt > 0)) { 2340 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2341 PhaseIdealLoop ideal_loop( igvn, false ); 2342 loop_opts_cnt--; 2343 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2344 } 2345 if (!failing()) { 2346 // Verify that last round of loop opts produced a valid graph 2347 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2348 PhaseIdealLoop::verify(igvn); 2349 } 2350 } 2351 if (failing()) return; 2352 2353 // Conditional Constant Propagation; 2354 PhaseCCP ccp( &igvn ); 2355 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2356 { 2357 TracePhase tp("ccp", &timers[_t_ccp]); 2358 ccp.do_transform(); 2359 } 2360 print_method(PHASE_CPP1, 2); 2361 2362 assert( true, "Break here to ccp.dump_old2new_map()"); 2363 2364 // Iterative Global Value Numbering, including ideal transforms 2365 { 2366 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2367 igvn = ccp; 2368 igvn.optimize(); 2369 } 2370 2371 print_method(PHASE_ITER_GVN2, 2); 2372 2373 if (failing()) return; 2374 2375 // Loop transforms on the ideal graph. Range Check Elimination, 2376 // peeling, unrolling, etc. 2377 if(loop_opts_cnt > 0) { 2378 debug_only( int cnt = 0; ); 2379 while(major_progress() && (loop_opts_cnt > 0)) { 2380 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2381 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2382 PhaseIdealLoop ideal_loop( igvn, true); 2383 loop_opts_cnt--; 2384 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2385 if (failing()) return; 2386 } 2387 } 2388 2389 if (failing()) return; 2390 2391 // Ensure that major progress is now clear 2392 C->clear_major_progress(); 2393 2394 { 2395 // Verify that all previous optimizations produced a valid graph 2396 // at least to this point, even if no loop optimizations were done. 2397 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2398 PhaseIdealLoop::verify(igvn); 2399 } 2400 2401 if (range_check_cast_count() > 0) { 2402 // No more loop optimizations. Remove all range check dependent CastIINodes. 2403 C->remove_range_check_casts(igvn); 2404 igvn.optimize(); 2405 } 2406 2407 #ifdef ASSERT 2408 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2409 bs->verify_gc_barriers(false); 2410 #endif 2411 2412 { 2413 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2414 PhaseMacroExpand mex(igvn); 2415 if (mex.expand_macro_nodes()) { 2416 assert(failing(), "must bail out w/ explicit message"); 2417 return; 2418 } 2419 } 2420 2421 if (opaque4_count() > 0) { 2422 C->remove_opaque4_nodes(igvn); 2423 igvn.optimize(); 2424 } 2425 2426 DEBUG_ONLY( _modified_nodes = NULL; ) 2427 } // (End scope of igvn; run destructor if necessary for asserts.) 2428 2429 process_print_inlining(); 2430 // A method with only infinite loops has no edges entering loops from root 2431 { 2432 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2433 if (final_graph_reshaping()) { 2434 assert(failing(), "must bail out w/ explicit message"); 2435 return; 2436 } 2437 } 2438 2439 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2440 } 2441 2442 //------------------------------Code_Gen--------------------------------------- 2443 // Given a graph, generate code for it 2444 void Compile::Code_Gen() { 2445 if (failing()) { 2446 return; 2447 } 2448 2449 // Perform instruction selection. You might think we could reclaim Matcher 2450 // memory PDQ, but actually the Matcher is used in generating spill code. 2451 // Internals of the Matcher (including some VectorSets) must remain live 2452 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2453 // set a bit in reclaimed memory. 2454 2455 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2456 // nodes. Mapping is only valid at the root of each matched subtree. 2457 NOT_PRODUCT( verify_graph_edges(); ) 2458 2459 Matcher matcher; 2460 _matcher = &matcher; 2461 { 2462 TracePhase tp("matcher", &timers[_t_matcher]); 2463 matcher.match(); 2464 } 2465 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2466 // nodes. Mapping is only valid at the root of each matched subtree. 2467 NOT_PRODUCT( verify_graph_edges(); ) 2468 2469 // If you have too many nodes, or if matching has failed, bail out 2470 check_node_count(0, "out of nodes matching instructions"); 2471 if (failing()) { 2472 return; 2473 } 2474 2475 print_method(PHASE_MATCHING, 2); 2476 2477 // Build a proper-looking CFG 2478 PhaseCFG cfg(node_arena(), root(), matcher); 2479 _cfg = &cfg; 2480 { 2481 TracePhase tp("scheduler", &timers[_t_scheduler]); 2482 bool success = cfg.do_global_code_motion(); 2483 if (!success) { 2484 return; 2485 } 2486 2487 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2488 NOT_PRODUCT( verify_graph_edges(); ) 2489 debug_only( cfg.verify(); ) 2490 } 2491 2492 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2493 _regalloc = ®alloc; 2494 { 2495 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2496 // Perform register allocation. After Chaitin, use-def chains are 2497 // no longer accurate (at spill code) and so must be ignored. 2498 // Node->LRG->reg mappings are still accurate. 2499 _regalloc->Register_Allocate(); 2500 2501 // Bail out if the allocator builds too many nodes 2502 if (failing()) { 2503 return; 2504 } 2505 } 2506 2507 // Prior to register allocation we kept empty basic blocks in case the 2508 // the allocator needed a place to spill. After register allocation we 2509 // are not adding any new instructions. If any basic block is empty, we 2510 // can now safely remove it. 2511 { 2512 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2513 cfg.remove_empty_blocks(); 2514 if (do_freq_based_layout()) { 2515 PhaseBlockLayout layout(cfg); 2516 } else { 2517 cfg.set_loop_alignment(); 2518 } 2519 cfg.fixup_flow(); 2520 } 2521 2522 // Apply peephole optimizations 2523 if( OptoPeephole ) { 2524 TracePhase tp("peephole", &timers[_t_peephole]); 2525 PhasePeephole peep( _regalloc, cfg); 2526 peep.do_transform(); 2527 } 2528 2529 // Do late expand if CPU requires this. 2530 if (Matcher::require_postalloc_expand) { 2531 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2532 cfg.postalloc_expand(_regalloc); 2533 } 2534 2535 // Convert Nodes to instruction bits in a buffer 2536 { 2537 TraceTime tp("output", &timers[_t_output], CITime); 2538 Output(); 2539 } 2540 2541 print_method(PHASE_FINAL_CODE); 2542 2543 // He's dead, Jim. 2544 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2545 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2546 } 2547 2548 2549 //------------------------------dump_asm--------------------------------------- 2550 // Dump formatted assembly 2551 #ifndef PRODUCT 2552 void Compile::dump_asm(int *pcs, uint pc_limit) { 2553 bool cut_short = false; 2554 tty->print_cr("#"); 2555 tty->print("# "); _tf->dump(); tty->cr(); 2556 tty->print_cr("#"); 2557 2558 // For all blocks 2559 int pc = 0x0; // Program counter 2560 char starts_bundle = ' '; 2561 _regalloc->dump_frame(); 2562 2563 Node *n = NULL; 2564 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2565 if (VMThread::should_terminate()) { 2566 cut_short = true; 2567 break; 2568 } 2569 Block* block = _cfg->get_block(i); 2570 if (block->is_connector() && !Verbose) { 2571 continue; 2572 } 2573 n = block->head(); 2574 if (pcs && n->_idx < pc_limit) { 2575 tty->print("%3.3x ", pcs[n->_idx]); 2576 } else { 2577 tty->print(" "); 2578 } 2579 block->dump_head(_cfg); 2580 if (block->is_connector()) { 2581 tty->print_cr(" # Empty connector block"); 2582 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2583 tty->print_cr(" # Block is sole successor of call"); 2584 } 2585 2586 // For all instructions 2587 Node *delay = NULL; 2588 for (uint j = 0; j < block->number_of_nodes(); j++) { 2589 if (VMThread::should_terminate()) { 2590 cut_short = true; 2591 break; 2592 } 2593 n = block->get_node(j); 2594 if (valid_bundle_info(n)) { 2595 Bundle* bundle = node_bundling(n); 2596 if (bundle->used_in_unconditional_delay()) { 2597 delay = n; 2598 continue; 2599 } 2600 if (bundle->starts_bundle()) { 2601 starts_bundle = '+'; 2602 } 2603 } 2604 2605 if (WizardMode) { 2606 n->dump(); 2607 } 2608 2609 if( !n->is_Region() && // Dont print in the Assembly 2610 !n->is_Phi() && // a few noisely useless nodes 2611 !n->is_Proj() && 2612 !n->is_MachTemp() && 2613 !n->is_SafePointScalarObject() && 2614 !n->is_Catch() && // Would be nice to print exception table targets 2615 !n->is_MergeMem() && // Not very interesting 2616 !n->is_top() && // Debug info table constants 2617 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2618 ) { 2619 if (pcs && n->_idx < pc_limit) 2620 tty->print("%3.3x", pcs[n->_idx]); 2621 else 2622 tty->print(" "); 2623 tty->print(" %c ", starts_bundle); 2624 starts_bundle = ' '; 2625 tty->print("\t"); 2626 n->format(_regalloc, tty); 2627 tty->cr(); 2628 } 2629 2630 // If we have an instruction with a delay slot, and have seen a delay, 2631 // then back up and print it 2632 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2633 assert(delay != NULL, "no unconditional delay instruction"); 2634 if (WizardMode) delay->dump(); 2635 2636 if (node_bundling(delay)->starts_bundle()) 2637 starts_bundle = '+'; 2638 if (pcs && n->_idx < pc_limit) 2639 tty->print("%3.3x", pcs[n->_idx]); 2640 else 2641 tty->print(" "); 2642 tty->print(" %c ", starts_bundle); 2643 starts_bundle = ' '; 2644 tty->print("\t"); 2645 delay->format(_regalloc, tty); 2646 tty->cr(); 2647 delay = NULL; 2648 } 2649 2650 // Dump the exception table as well 2651 if( n->is_Catch() && (Verbose || WizardMode) ) { 2652 // Print the exception table for this offset 2653 _handler_table.print_subtable_for(pc); 2654 } 2655 } 2656 2657 if (pcs && n->_idx < pc_limit) 2658 tty->print_cr("%3.3x", pcs[n->_idx]); 2659 else 2660 tty->cr(); 2661 2662 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2663 2664 } // End of per-block dump 2665 tty->cr(); 2666 2667 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2668 } 2669 #endif 2670 2671 //------------------------------Final_Reshape_Counts--------------------------- 2672 // This class defines counters to help identify when a method 2673 // may/must be executed using hardware with only 24-bit precision. 2674 struct Final_Reshape_Counts : public StackObj { 2675 int _call_count; // count non-inlined 'common' calls 2676 int _float_count; // count float ops requiring 24-bit precision 2677 int _double_count; // count double ops requiring more precision 2678 int _java_call_count; // count non-inlined 'java' calls 2679 int _inner_loop_count; // count loops which need alignment 2680 VectorSet _visited; // Visitation flags 2681 Node_List _tests; // Set of IfNodes & PCTableNodes 2682 2683 Final_Reshape_Counts() : 2684 _call_count(0), _float_count(0), _double_count(0), 2685 _java_call_count(0), _inner_loop_count(0), 2686 _visited( Thread::current()->resource_area() ) { } 2687 2688 void inc_call_count () { _call_count ++; } 2689 void inc_float_count () { _float_count ++; } 2690 void inc_double_count() { _double_count++; } 2691 void inc_java_call_count() { _java_call_count++; } 2692 void inc_inner_loop_count() { _inner_loop_count++; } 2693 2694 int get_call_count () const { return _call_count ; } 2695 int get_float_count () const { return _float_count ; } 2696 int get_double_count() const { return _double_count; } 2697 int get_java_call_count() const { return _java_call_count; } 2698 int get_inner_loop_count() const { return _inner_loop_count; } 2699 }; 2700 2701 #ifdef ASSERT 2702 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2703 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2704 // Make sure the offset goes inside the instance layout. 2705 return k->contains_field_offset(tp->offset()); 2706 // Note that OffsetBot and OffsetTop are very negative. 2707 } 2708 #endif 2709 2710 // Eliminate trivially redundant StoreCMs and accumulate their 2711 // precedence edges. 2712 void Compile::eliminate_redundant_card_marks(Node* n) { 2713 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2714 if (n->in(MemNode::Address)->outcnt() > 1) { 2715 // There are multiple users of the same address so it might be 2716 // possible to eliminate some of the StoreCMs 2717 Node* mem = n->in(MemNode::Memory); 2718 Node* adr = n->in(MemNode::Address); 2719 Node* val = n->in(MemNode::ValueIn); 2720 Node* prev = n; 2721 bool done = false; 2722 // Walk the chain of StoreCMs eliminating ones that match. As 2723 // long as it's a chain of single users then the optimization is 2724 // safe. Eliminating partially redundant StoreCMs would require 2725 // cloning copies down the other paths. 2726 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2727 if (adr == mem->in(MemNode::Address) && 2728 val == mem->in(MemNode::ValueIn)) { 2729 // redundant StoreCM 2730 if (mem->req() > MemNode::OopStore) { 2731 // Hasn't been processed by this code yet. 2732 n->add_prec(mem->in(MemNode::OopStore)); 2733 } else { 2734 // Already converted to precedence edge 2735 for (uint i = mem->req(); i < mem->len(); i++) { 2736 // Accumulate any precedence edges 2737 if (mem->in(i) != NULL) { 2738 n->add_prec(mem->in(i)); 2739 } 2740 } 2741 // Everything above this point has been processed. 2742 done = true; 2743 } 2744 // Eliminate the previous StoreCM 2745 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2746 assert(mem->outcnt() == 0, "should be dead"); 2747 mem->disconnect_inputs(NULL, this); 2748 } else { 2749 prev = mem; 2750 } 2751 mem = prev->in(MemNode::Memory); 2752 } 2753 } 2754 } 2755 2756 2757 //------------------------------final_graph_reshaping_impl---------------------- 2758 // Implement items 1-5 from final_graph_reshaping below. 2759 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2760 2761 if ( n->outcnt() == 0 ) return; // dead node 2762 uint nop = n->Opcode(); 2763 2764 // Check for 2-input instruction with "last use" on right input. 2765 // Swap to left input. Implements item (2). 2766 if( n->req() == 3 && // two-input instruction 2767 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2768 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2769 n->in(2)->outcnt() == 1 &&// right use IS a last use 2770 !n->in(2)->is_Con() ) { // right use is not a constant 2771 // Check for commutative opcode 2772 switch( nop ) { 2773 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2774 case Op_MaxI: case Op_MinI: 2775 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2776 case Op_AndL: case Op_XorL: case Op_OrL: 2777 case Op_AndI: case Op_XorI: case Op_OrI: { 2778 // Move "last use" input to left by swapping inputs 2779 n->swap_edges(1, 2); 2780 break; 2781 } 2782 default: 2783 break; 2784 } 2785 } 2786 2787 #ifdef ASSERT 2788 if( n->is_Mem() ) { 2789 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2790 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2791 // oop will be recorded in oop map if load crosses safepoint 2792 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2793 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2794 "raw memory operations should have control edge"); 2795 } 2796 #endif 2797 // Count FPU ops and common calls, implements item (3) 2798 switch( nop ) { 2799 // Count all float operations that may use FPU 2800 case Op_AddF: 2801 case Op_SubF: 2802 case Op_MulF: 2803 case Op_DivF: 2804 case Op_NegF: 2805 case Op_ModF: 2806 case Op_ConvI2F: 2807 case Op_ConF: 2808 case Op_CmpF: 2809 case Op_CmpF3: 2810 // case Op_ConvL2F: // longs are split into 32-bit halves 2811 frc.inc_float_count(); 2812 break; 2813 2814 case Op_ConvF2D: 2815 case Op_ConvD2F: 2816 frc.inc_float_count(); 2817 frc.inc_double_count(); 2818 break; 2819 2820 // Count all double operations that may use FPU 2821 case Op_AddD: 2822 case Op_SubD: 2823 case Op_MulD: 2824 case Op_DivD: 2825 case Op_NegD: 2826 case Op_ModD: 2827 case Op_ConvI2D: 2828 case Op_ConvD2I: 2829 // case Op_ConvL2D: // handled by leaf call 2830 // case Op_ConvD2L: // handled by leaf call 2831 case Op_ConD: 2832 case Op_CmpD: 2833 case Op_CmpD3: 2834 frc.inc_double_count(); 2835 break; 2836 case Op_Opaque1: // Remove Opaque Nodes before matching 2837 case Op_Opaque2: // Remove Opaque Nodes before matching 2838 case Op_Opaque3: 2839 n->subsume_by(n->in(1), this); 2840 break; 2841 case Op_CallStaticJava: 2842 case Op_CallJava: 2843 case Op_CallDynamicJava: 2844 frc.inc_java_call_count(); // Count java call site; 2845 case Op_CallRuntime: 2846 case Op_CallLeaf: 2847 case Op_CallLeafNoFP: { 2848 assert (n->is_Call(), ""); 2849 CallNode *call = n->as_Call(); 2850 // Count call sites where the FP mode bit would have to be flipped. 2851 // Do not count uncommon runtime calls: 2852 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2853 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2854 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 2855 frc.inc_call_count(); // Count the call site 2856 } else { // See if uncommon argument is shared 2857 Node *n = call->in(TypeFunc::Parms); 2858 int nop = n->Opcode(); 2859 // Clone shared simple arguments to uncommon calls, item (1). 2860 if (n->outcnt() > 1 && 2861 !n->is_Proj() && 2862 nop != Op_CreateEx && 2863 nop != Op_CheckCastPP && 2864 nop != Op_DecodeN && 2865 nop != Op_DecodeNKlass && 2866 !n->is_Mem() && 2867 !n->is_Phi()) { 2868 Node *x = n->clone(); 2869 call->set_req(TypeFunc::Parms, x); 2870 } 2871 } 2872 break; 2873 } 2874 2875 case Op_StoreD: 2876 case Op_LoadD: 2877 case Op_LoadD_unaligned: 2878 frc.inc_double_count(); 2879 goto handle_mem; 2880 case Op_StoreF: 2881 case Op_LoadF: 2882 frc.inc_float_count(); 2883 goto handle_mem; 2884 2885 case Op_StoreCM: 2886 { 2887 // Convert OopStore dependence into precedence edge 2888 Node* prec = n->in(MemNode::OopStore); 2889 n->del_req(MemNode::OopStore); 2890 n->add_prec(prec); 2891 eliminate_redundant_card_marks(n); 2892 } 2893 2894 // fall through 2895 2896 case Op_StoreB: 2897 case Op_StoreC: 2898 case Op_StorePConditional: 2899 case Op_StoreI: 2900 case Op_StoreL: 2901 case Op_StoreIConditional: 2902 case Op_StoreLConditional: 2903 case Op_CompareAndSwapB: 2904 case Op_CompareAndSwapS: 2905 case Op_CompareAndSwapI: 2906 case Op_CompareAndSwapL: 2907 case Op_CompareAndSwapP: 2908 case Op_CompareAndSwapN: 2909 case Op_WeakCompareAndSwapB: 2910 case Op_WeakCompareAndSwapS: 2911 case Op_WeakCompareAndSwapI: 2912 case Op_WeakCompareAndSwapL: 2913 case Op_WeakCompareAndSwapP: 2914 case Op_WeakCompareAndSwapN: 2915 case Op_CompareAndExchangeB: 2916 case Op_CompareAndExchangeS: 2917 case Op_CompareAndExchangeI: 2918 case Op_CompareAndExchangeL: 2919 case Op_CompareAndExchangeP: 2920 case Op_CompareAndExchangeN: 2921 case Op_GetAndAddS: 2922 case Op_GetAndAddB: 2923 case Op_GetAndAddI: 2924 case Op_GetAndAddL: 2925 case Op_GetAndSetS: 2926 case Op_GetAndSetB: 2927 case Op_GetAndSetI: 2928 case Op_GetAndSetL: 2929 case Op_GetAndSetP: 2930 case Op_GetAndSetN: 2931 case Op_StoreP: 2932 case Op_StoreN: 2933 case Op_StoreNKlass: 2934 case Op_LoadB: 2935 case Op_LoadUB: 2936 case Op_LoadUS: 2937 case Op_LoadI: 2938 case Op_LoadKlass: 2939 case Op_LoadNKlass: 2940 case Op_LoadL: 2941 case Op_LoadL_unaligned: 2942 case Op_LoadPLocked: 2943 case Op_LoadP: 2944 case Op_LoadN: 2945 case Op_LoadRange: 2946 case Op_LoadS: { 2947 handle_mem: 2948 #ifdef ASSERT 2949 if( VerifyOptoOopOffsets ) { 2950 assert( n->is_Mem(), "" ); 2951 MemNode *mem = (MemNode*)n; 2952 // Check to see if address types have grounded out somehow. 2953 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2954 assert( !tp || oop_offset_is_sane(tp), "" ); 2955 } 2956 #endif 2957 break; 2958 } 2959 2960 case Op_AddP: { // Assert sane base pointers 2961 Node *addp = n->in(AddPNode::Address); 2962 assert( !addp->is_AddP() || 2963 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2964 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2965 "Base pointers must match (addp %u)", addp->_idx ); 2966 #ifdef _LP64 2967 if ((UseCompressedOops || UseCompressedClassPointers) && 2968 addp->Opcode() == Op_ConP && 2969 addp == n->in(AddPNode::Base) && 2970 n->in(AddPNode::Offset)->is_Con()) { 2971 // If the transformation of ConP to ConN+DecodeN is beneficial depends 2972 // on the platform and on the compressed oops mode. 2973 // Use addressing with narrow klass to load with offset on x86. 2974 // Some platforms can use the constant pool to load ConP. 2975 // Do this transformation here since IGVN will convert ConN back to ConP. 2976 const Type* t = addp->bottom_type(); 2977 bool is_oop = t->isa_oopptr() != NULL; 2978 bool is_klass = t->isa_klassptr() != NULL; 2979 2980 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 2981 (is_klass && Matcher::const_klass_prefer_decode())) { 2982 Node* nn = NULL; 2983 2984 int op = is_oop ? Op_ConN : Op_ConNKlass; 2985 2986 // Look for existing ConN node of the same exact type. 2987 Node* r = root(); 2988 uint cnt = r->outcnt(); 2989 for (uint i = 0; i < cnt; i++) { 2990 Node* m = r->raw_out(i); 2991 if (m!= NULL && m->Opcode() == op && 2992 m->bottom_type()->make_ptr() == t) { 2993 nn = m; 2994 break; 2995 } 2996 } 2997 if (nn != NULL) { 2998 // Decode a narrow oop to match address 2999 // [R12 + narrow_oop_reg<<3 + offset] 3000 if (is_oop) { 3001 nn = new DecodeNNode(nn, t); 3002 } else { 3003 nn = new DecodeNKlassNode(nn, t); 3004 } 3005 // Check for succeeding AddP which uses the same Base. 3006 // Otherwise we will run into the assertion above when visiting that guy. 3007 for (uint i = 0; i < n->outcnt(); ++i) { 3008 Node *out_i = n->raw_out(i); 3009 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3010 out_i->set_req(AddPNode::Base, nn); 3011 #ifdef ASSERT 3012 for (uint j = 0; j < out_i->outcnt(); ++j) { 3013 Node *out_j = out_i->raw_out(j); 3014 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3015 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3016 } 3017 #endif 3018 } 3019 } 3020 n->set_req(AddPNode::Base, nn); 3021 n->set_req(AddPNode::Address, nn); 3022 if (addp->outcnt() == 0) { 3023 addp->disconnect_inputs(NULL, this); 3024 } 3025 } 3026 } 3027 } 3028 #endif 3029 // platform dependent reshaping of the address expression 3030 reshape_address(n->as_AddP()); 3031 break; 3032 } 3033 3034 case Op_CastPP: { 3035 // Remove CastPP nodes to gain more freedom during scheduling but 3036 // keep the dependency they encode as control or precedence edges 3037 // (if control is set already) on memory operations. Some CastPP 3038 // nodes don't have a control (don't carry a dependency): skip 3039 // those. 3040 if (n->in(0) != NULL) { 3041 ResourceMark rm; 3042 Unique_Node_List wq; 3043 wq.push(n); 3044 for (uint next = 0; next < wq.size(); ++next) { 3045 Node *m = wq.at(next); 3046 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3047 Node* use = m->fast_out(i); 3048 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3049 use->ensure_control_or_add_prec(n->in(0)); 3050 } else { 3051 switch(use->Opcode()) { 3052 case Op_AddP: 3053 case Op_DecodeN: 3054 case Op_DecodeNKlass: 3055 case Op_CheckCastPP: 3056 case Op_CastPP: 3057 wq.push(use); 3058 break; 3059 } 3060 } 3061 } 3062 } 3063 } 3064 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3065 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3066 Node* in1 = n->in(1); 3067 const Type* t = n->bottom_type(); 3068 Node* new_in1 = in1->clone(); 3069 new_in1->as_DecodeN()->set_type(t); 3070 3071 if (!Matcher::narrow_oop_use_complex_address()) { 3072 // 3073 // x86, ARM and friends can handle 2 adds in addressing mode 3074 // and Matcher can fold a DecodeN node into address by using 3075 // a narrow oop directly and do implicit NULL check in address: 3076 // 3077 // [R12 + narrow_oop_reg<<3 + offset] 3078 // NullCheck narrow_oop_reg 3079 // 3080 // On other platforms (Sparc) we have to keep new DecodeN node and 3081 // use it to do implicit NULL check in address: 3082 // 3083 // decode_not_null narrow_oop_reg, base_reg 3084 // [base_reg + offset] 3085 // NullCheck base_reg 3086 // 3087 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3088 // to keep the information to which NULL check the new DecodeN node 3089 // corresponds to use it as value in implicit_null_check(). 3090 // 3091 new_in1->set_req(0, n->in(0)); 3092 } 3093 3094 n->subsume_by(new_in1, this); 3095 if (in1->outcnt() == 0) { 3096 in1->disconnect_inputs(NULL, this); 3097 } 3098 } else { 3099 n->subsume_by(n->in(1), this); 3100 if (n->outcnt() == 0) { 3101 n->disconnect_inputs(NULL, this); 3102 } 3103 } 3104 break; 3105 } 3106 #ifdef _LP64 3107 case Op_CmpP: 3108 // Do this transformation here to preserve CmpPNode::sub() and 3109 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3110 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3111 Node* in1 = n->in(1); 3112 Node* in2 = n->in(2); 3113 if (!in1->is_DecodeNarrowPtr()) { 3114 in2 = in1; 3115 in1 = n->in(2); 3116 } 3117 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3118 3119 Node* new_in2 = NULL; 3120 if (in2->is_DecodeNarrowPtr()) { 3121 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3122 new_in2 = in2->in(1); 3123 } else if (in2->Opcode() == Op_ConP) { 3124 const Type* t = in2->bottom_type(); 3125 if (t == TypePtr::NULL_PTR) { 3126 assert(in1->is_DecodeN(), "compare klass to null?"); 3127 // Don't convert CmpP null check into CmpN if compressed 3128 // oops implicit null check is not generated. 3129 // This will allow to generate normal oop implicit null check. 3130 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3131 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3132 // 3133 // This transformation together with CastPP transformation above 3134 // will generated code for implicit NULL checks for compressed oops. 3135 // 3136 // The original code after Optimize() 3137 // 3138 // LoadN memory, narrow_oop_reg 3139 // decode narrow_oop_reg, base_reg 3140 // CmpP base_reg, NULL 3141 // CastPP base_reg // NotNull 3142 // Load [base_reg + offset], val_reg 3143 // 3144 // after these transformations will be 3145 // 3146 // LoadN memory, narrow_oop_reg 3147 // CmpN narrow_oop_reg, NULL 3148 // decode_not_null narrow_oop_reg, base_reg 3149 // Load [base_reg + offset], val_reg 3150 // 3151 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3152 // since narrow oops can be used in debug info now (see the code in 3153 // final_graph_reshaping_walk()). 3154 // 3155 // At the end the code will be matched to 3156 // on x86: 3157 // 3158 // Load_narrow_oop memory, narrow_oop_reg 3159 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3160 // NullCheck narrow_oop_reg 3161 // 3162 // and on sparc: 3163 // 3164 // Load_narrow_oop memory, narrow_oop_reg 3165 // decode_not_null narrow_oop_reg, base_reg 3166 // Load [base_reg + offset], val_reg 3167 // NullCheck base_reg 3168 // 3169 } else if (t->isa_oopptr()) { 3170 new_in2 = ConNode::make(t->make_narrowoop()); 3171 } else if (t->isa_klassptr()) { 3172 new_in2 = ConNode::make(t->make_narrowklass()); 3173 } 3174 } 3175 if (new_in2 != NULL) { 3176 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3177 n->subsume_by(cmpN, this); 3178 if (in1->outcnt() == 0) { 3179 in1->disconnect_inputs(NULL, this); 3180 } 3181 if (in2->outcnt() == 0) { 3182 in2->disconnect_inputs(NULL, this); 3183 } 3184 } 3185 } 3186 break; 3187 3188 case Op_DecodeN: 3189 case Op_DecodeNKlass: 3190 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3191 // DecodeN could be pinned when it can't be fold into 3192 // an address expression, see the code for Op_CastPP above. 3193 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3194 break; 3195 3196 case Op_EncodeP: 3197 case Op_EncodePKlass: { 3198 Node* in1 = n->in(1); 3199 if (in1->is_DecodeNarrowPtr()) { 3200 n->subsume_by(in1->in(1), this); 3201 } else if (in1->Opcode() == Op_ConP) { 3202 const Type* t = in1->bottom_type(); 3203 if (t == TypePtr::NULL_PTR) { 3204 assert(t->isa_oopptr(), "null klass?"); 3205 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3206 } else if (t->isa_oopptr()) { 3207 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3208 } else if (t->isa_klassptr()) { 3209 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3210 } 3211 } 3212 if (in1->outcnt() == 0) { 3213 in1->disconnect_inputs(NULL, this); 3214 } 3215 break; 3216 } 3217 3218 case Op_Proj: { 3219 if (OptimizeStringConcat) { 3220 ProjNode* p = n->as_Proj(); 3221 if (p->_is_io_use) { 3222 // Separate projections were used for the exception path which 3223 // are normally removed by a late inline. If it wasn't inlined 3224 // then they will hang around and should just be replaced with 3225 // the original one. 3226 Node* proj = NULL; 3227 // Replace with just one 3228 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3229 Node *use = i.get(); 3230 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3231 proj = use; 3232 break; 3233 } 3234 } 3235 assert(proj != NULL, "must be found"); 3236 p->subsume_by(proj, this); 3237 } 3238 } 3239 break; 3240 } 3241 3242 case Op_Phi: 3243 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3244 // The EncodeP optimization may create Phi with the same edges 3245 // for all paths. It is not handled well by Register Allocator. 3246 Node* unique_in = n->in(1); 3247 assert(unique_in != NULL, ""); 3248 uint cnt = n->req(); 3249 for (uint i = 2; i < cnt; i++) { 3250 Node* m = n->in(i); 3251 assert(m != NULL, ""); 3252 if (unique_in != m) 3253 unique_in = NULL; 3254 } 3255 if (unique_in != NULL) { 3256 n->subsume_by(unique_in, this); 3257 } 3258 } 3259 break; 3260 3261 #endif 3262 3263 #ifdef ASSERT 3264 case Op_CastII: 3265 // Verify that all range check dependent CastII nodes were removed. 3266 if (n->isa_CastII()->has_range_check()) { 3267 n->dump(3); 3268 assert(false, "Range check dependent CastII node was not removed"); 3269 } 3270 break; 3271 #endif 3272 3273 case Op_ModI: 3274 if (UseDivMod) { 3275 // Check if a%b and a/b both exist 3276 Node* d = n->find_similar(Op_DivI); 3277 if (d) { 3278 // Replace them with a fused divmod if supported 3279 if (Matcher::has_match_rule(Op_DivModI)) { 3280 DivModINode* divmod = DivModINode::make(n); 3281 d->subsume_by(divmod->div_proj(), this); 3282 n->subsume_by(divmod->mod_proj(), this); 3283 } else { 3284 // replace a%b with a-((a/b)*b) 3285 Node* mult = new MulINode(d, d->in(2)); 3286 Node* sub = new SubINode(d->in(1), mult); 3287 n->subsume_by(sub, this); 3288 } 3289 } 3290 } 3291 break; 3292 3293 case Op_ModL: 3294 if (UseDivMod) { 3295 // Check if a%b and a/b both exist 3296 Node* d = n->find_similar(Op_DivL); 3297 if (d) { 3298 // Replace them with a fused divmod if supported 3299 if (Matcher::has_match_rule(Op_DivModL)) { 3300 DivModLNode* divmod = DivModLNode::make(n); 3301 d->subsume_by(divmod->div_proj(), this); 3302 n->subsume_by(divmod->mod_proj(), this); 3303 } else { 3304 // replace a%b with a-((a/b)*b) 3305 Node* mult = new MulLNode(d, d->in(2)); 3306 Node* sub = new SubLNode(d->in(1), mult); 3307 n->subsume_by(sub, this); 3308 } 3309 } 3310 } 3311 break; 3312 3313 case Op_LoadVector: 3314 case Op_StoreVector: 3315 break; 3316 3317 case Op_AddReductionVI: 3318 case Op_AddReductionVL: 3319 case Op_AddReductionVF: 3320 case Op_AddReductionVD: 3321 case Op_MulReductionVI: 3322 case Op_MulReductionVL: 3323 case Op_MulReductionVF: 3324 case Op_MulReductionVD: 3325 break; 3326 3327 case Op_PackB: 3328 case Op_PackS: 3329 case Op_PackI: 3330 case Op_PackF: 3331 case Op_PackL: 3332 case Op_PackD: 3333 if (n->req()-1 > 2) { 3334 // Replace many operand PackNodes with a binary tree for matching 3335 PackNode* p = (PackNode*) n; 3336 Node* btp = p->binary_tree_pack(1, n->req()); 3337 n->subsume_by(btp, this); 3338 } 3339 break; 3340 case Op_Loop: 3341 case Op_CountedLoop: 3342 case Op_OuterStripMinedLoop: 3343 if (n->as_Loop()->is_inner_loop()) { 3344 frc.inc_inner_loop_count(); 3345 } 3346 n->as_Loop()->verify_strip_mined(0); 3347 break; 3348 case Op_LShiftI: 3349 case Op_RShiftI: 3350 case Op_URShiftI: 3351 case Op_LShiftL: 3352 case Op_RShiftL: 3353 case Op_URShiftL: 3354 if (Matcher::need_masked_shift_count) { 3355 // The cpu's shift instructions don't restrict the count to the 3356 // lower 5/6 bits. We need to do the masking ourselves. 3357 Node* in2 = n->in(2); 3358 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3359 const TypeInt* t = in2->find_int_type(); 3360 if (t != NULL && t->is_con()) { 3361 juint shift = t->get_con(); 3362 if (shift > mask) { // Unsigned cmp 3363 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3364 } 3365 } else { 3366 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3367 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3368 n->set_req(2, shift); 3369 } 3370 } 3371 if (in2->outcnt() == 0) { // Remove dead node 3372 in2->disconnect_inputs(NULL, this); 3373 } 3374 } 3375 break; 3376 case Op_MemBarStoreStore: 3377 case Op_MemBarRelease: 3378 // Break the link with AllocateNode: it is no longer useful and 3379 // confuses register allocation. 3380 if (n->req() > MemBarNode::Precedent) { 3381 n->set_req(MemBarNode::Precedent, top()); 3382 } 3383 break; 3384 case Op_RangeCheck: { 3385 RangeCheckNode* rc = n->as_RangeCheck(); 3386 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3387 n->subsume_by(iff, this); 3388 frc._tests.push(iff); 3389 break; 3390 } 3391 case Op_ConvI2L: { 3392 if (!Matcher::convi2l_type_required) { 3393 // Code generation on some platforms doesn't need accurate 3394 // ConvI2L types. Widening the type can help remove redundant 3395 // address computations. 3396 n->as_Type()->set_type(TypeLong::INT); 3397 ResourceMark rm; 3398 Node_List wq; 3399 wq.push(n); 3400 for (uint next = 0; next < wq.size(); next++) { 3401 Node *m = wq.at(next); 3402 3403 for(;;) { 3404 // Loop over all nodes with identical inputs edges as m 3405 Node* k = m->find_similar(m->Opcode()); 3406 if (k == NULL) { 3407 break; 3408 } 3409 // Push their uses so we get a chance to remove node made 3410 // redundant 3411 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3412 Node* u = k->fast_out(i); 3413 assert(!wq.contains(u), "shouldn't process one node several times"); 3414 if (u->Opcode() == Op_LShiftL || 3415 u->Opcode() == Op_AddL || 3416 u->Opcode() == Op_SubL || 3417 u->Opcode() == Op_AddP) { 3418 wq.push(u); 3419 } 3420 } 3421 // Replace all nodes with identical edges as m with m 3422 k->subsume_by(m, this); 3423 } 3424 } 3425 } 3426 break; 3427 } 3428 case Op_CmpUL: { 3429 if (!Matcher::has_match_rule(Op_CmpUL)) { 3430 // We don't support unsigned long comparisons. Set 'max_idx_expr' 3431 // to max_julong if < 0 to make the signed comparison fail. 3432 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3433 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3434 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3435 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3436 Node* andl = new AndLNode(orl, remove_sign_mask); 3437 Node* cmp = new CmpLNode(andl, n->in(2)); 3438 n->subsume_by(cmp, this); 3439 } 3440 break; 3441 } 3442 #ifdef ASSERT 3443 case Op_ValueTypePtr: 3444 case Op_ValueType: { 3445 n->dump(-1); 3446 assert(false, "value type node was not removed"); 3447 break; 3448 } 3449 #endif 3450 default: 3451 assert( !n->is_Call(), "" ); 3452 assert( !n->is_Mem(), "" ); 3453 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3454 break; 3455 } 3456 3457 // Collect CFG split points 3458 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3459 frc._tests.push(n); 3460 } 3461 } 3462 3463 //------------------------------final_graph_reshaping_walk--------------------- 3464 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3465 // requires that the walk visits a node's inputs before visiting the node. 3466 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3467 ResourceArea *area = Thread::current()->resource_area(); 3468 Unique_Node_List sfpt(area); 3469 3470 frc._visited.set(root->_idx); // first, mark node as visited 3471 uint cnt = root->req(); 3472 Node *n = root; 3473 uint i = 0; 3474 while (true) { 3475 if (i < cnt) { 3476 // Place all non-visited non-null inputs onto stack 3477 Node* m = n->in(i); 3478 ++i; 3479 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3480 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3481 // compute worst case interpreter size in case of a deoptimization 3482 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3483 3484 sfpt.push(m); 3485 } 3486 cnt = m->req(); 3487 nstack.push(n, i); // put on stack parent and next input's index 3488 n = m; 3489 i = 0; 3490 } 3491 } else { 3492 // Now do post-visit work 3493 final_graph_reshaping_impl( n, frc ); 3494 if (nstack.is_empty()) 3495 break; // finished 3496 n = nstack.node(); // Get node from stack 3497 cnt = n->req(); 3498 i = nstack.index(); 3499 nstack.pop(); // Shift to the next node on stack 3500 } 3501 } 3502 3503 // Skip next transformation if compressed oops are not used. 3504 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3505 (!UseCompressedOops && !UseCompressedClassPointers)) 3506 return; 3507 3508 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3509 // It could be done for an uncommon traps or any safepoints/calls 3510 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3511 while (sfpt.size() > 0) { 3512 n = sfpt.pop(); 3513 JVMState *jvms = n->as_SafePoint()->jvms(); 3514 assert(jvms != NULL, "sanity"); 3515 int start = jvms->debug_start(); 3516 int end = n->req(); 3517 bool is_uncommon = (n->is_CallStaticJava() && 3518 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3519 for (int j = start; j < end; j++) { 3520 Node* in = n->in(j); 3521 if (in->is_DecodeNarrowPtr()) { 3522 bool safe_to_skip = true; 3523 if (!is_uncommon ) { 3524 // Is it safe to skip? 3525 for (uint i = 0; i < in->outcnt(); i++) { 3526 Node* u = in->raw_out(i); 3527 if (!u->is_SafePoint() || 3528 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3529 safe_to_skip = false; 3530 } 3531 } 3532 } 3533 if (safe_to_skip) { 3534 n->set_req(j, in->in(1)); 3535 } 3536 if (in->outcnt() == 0) { 3537 in->disconnect_inputs(NULL, this); 3538 } 3539 } 3540 } 3541 } 3542 } 3543 3544 //------------------------------final_graph_reshaping-------------------------- 3545 // Final Graph Reshaping. 3546 // 3547 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3548 // and not commoned up and forced early. Must come after regular 3549 // optimizations to avoid GVN undoing the cloning. Clone constant 3550 // inputs to Loop Phis; these will be split by the allocator anyways. 3551 // Remove Opaque nodes. 3552 // (2) Move last-uses by commutative operations to the left input to encourage 3553 // Intel update-in-place two-address operations and better register usage 3554 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3555 // calls canonicalizing them back. 3556 // (3) Count the number of double-precision FP ops, single-precision FP ops 3557 // and call sites. On Intel, we can get correct rounding either by 3558 // forcing singles to memory (requires extra stores and loads after each 3559 // FP bytecode) or we can set a rounding mode bit (requires setting and 3560 // clearing the mode bit around call sites). The mode bit is only used 3561 // if the relative frequency of single FP ops to calls is low enough. 3562 // This is a key transform for SPEC mpeg_audio. 3563 // (4) Detect infinite loops; blobs of code reachable from above but not 3564 // below. Several of the Code_Gen algorithms fail on such code shapes, 3565 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3566 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3567 // Detection is by looking for IfNodes where only 1 projection is 3568 // reachable from below or CatchNodes missing some targets. 3569 // (5) Assert for insane oop offsets in debug mode. 3570 3571 bool Compile::final_graph_reshaping() { 3572 // an infinite loop may have been eliminated by the optimizer, 3573 // in which case the graph will be empty. 3574 if (root()->req() == 1) { 3575 record_method_not_compilable("trivial infinite loop"); 3576 return true; 3577 } 3578 3579 // Expensive nodes have their control input set to prevent the GVN 3580 // from freely commoning them. There's no GVN beyond this point so 3581 // no need to keep the control input. We want the expensive nodes to 3582 // be freely moved to the least frequent code path by gcm. 3583 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3584 for (int i = 0; i < expensive_count(); i++) { 3585 _expensive_nodes->at(i)->set_req(0, NULL); 3586 } 3587 3588 Final_Reshape_Counts frc; 3589 3590 // Visit everybody reachable! 3591 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3592 Node_Stack nstack(live_nodes() >> 1); 3593 final_graph_reshaping_walk(nstack, root(), frc); 3594 3595 // Check for unreachable (from below) code (i.e., infinite loops). 3596 for( uint i = 0; i < frc._tests.size(); i++ ) { 3597 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3598 // Get number of CFG targets. 3599 // Note that PCTables include exception targets after calls. 3600 uint required_outcnt = n->required_outcnt(); 3601 if (n->outcnt() != required_outcnt) { 3602 // Check for a few special cases. Rethrow Nodes never take the 3603 // 'fall-thru' path, so expected kids is 1 less. 3604 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3605 if (n->in(0)->in(0)->is_Call()) { 3606 CallNode *call = n->in(0)->in(0)->as_Call(); 3607 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3608 required_outcnt--; // Rethrow always has 1 less kid 3609 } else if (call->req() > TypeFunc::Parms && 3610 call->is_CallDynamicJava()) { 3611 // Check for null receiver. In such case, the optimizer has 3612 // detected that the virtual call will always result in a null 3613 // pointer exception. The fall-through projection of this CatchNode 3614 // will not be populated. 3615 Node *arg0 = call->in(TypeFunc::Parms); 3616 if (arg0->is_Type() && 3617 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3618 required_outcnt--; 3619 } 3620 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3621 call->req() > TypeFunc::Parms+1 && 3622 call->is_CallStaticJava()) { 3623 // Check for negative array length. In such case, the optimizer has 3624 // detected that the allocation attempt will always result in an 3625 // exception. There is no fall-through projection of this CatchNode . 3626 Node *arg1 = call->in(TypeFunc::Parms+1); 3627 if (arg1->is_Type() && 3628 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3629 required_outcnt--; 3630 } 3631 } 3632 } 3633 } 3634 // Recheck with a better notion of 'required_outcnt' 3635 if (n->outcnt() != required_outcnt) { 3636 record_method_not_compilable("malformed control flow"); 3637 return true; // Not all targets reachable! 3638 } 3639 } 3640 // Check that I actually visited all kids. Unreached kids 3641 // must be infinite loops. 3642 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3643 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3644 record_method_not_compilable("infinite loop"); 3645 return true; // Found unvisited kid; must be unreach 3646 } 3647 3648 // Here so verification code in final_graph_reshaping_walk() 3649 // always see an OuterStripMinedLoopEnd 3650 if (n->is_OuterStripMinedLoopEnd()) { 3651 IfNode* init_iff = n->as_If(); 3652 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3653 n->subsume_by(iff, this); 3654 } 3655 } 3656 3657 // If original bytecodes contained a mixture of floats and doubles 3658 // check if the optimizer has made it homogenous, item (3). 3659 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3660 frc.get_float_count() > 32 && 3661 frc.get_double_count() == 0 && 3662 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3663 set_24_bit_selection_and_mode( false, true ); 3664 } 3665 3666 set_java_calls(frc.get_java_call_count()); 3667 set_inner_loops(frc.get_inner_loop_count()); 3668 3669 // No infinite loops, no reason to bail out. 3670 return false; 3671 } 3672 3673 //-----------------------------too_many_traps---------------------------------- 3674 // Report if there are too many traps at the current method and bci. 3675 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3676 bool Compile::too_many_traps(ciMethod* method, 3677 int bci, 3678 Deoptimization::DeoptReason reason) { 3679 ciMethodData* md = method->method_data(); 3680 if (md->is_empty()) { 3681 // Assume the trap has not occurred, or that it occurred only 3682 // because of a transient condition during start-up in the interpreter. 3683 return false; 3684 } 3685 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3686 if (md->has_trap_at(bci, m, reason) != 0) { 3687 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3688 // Also, if there are multiple reasons, or if there is no per-BCI record, 3689 // assume the worst. 3690 if (log()) 3691 log()->elem("observe trap='%s' count='%d'", 3692 Deoptimization::trap_reason_name(reason), 3693 md->trap_count(reason)); 3694 return true; 3695 } else { 3696 // Ignore method/bci and see if there have been too many globally. 3697 return too_many_traps(reason, md); 3698 } 3699 } 3700 3701 // Less-accurate variant which does not require a method and bci. 3702 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3703 ciMethodData* logmd) { 3704 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3705 // Too many traps globally. 3706 // Note that we use cumulative trap_count, not just md->trap_count. 3707 if (log()) { 3708 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3709 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3710 Deoptimization::trap_reason_name(reason), 3711 mcount, trap_count(reason)); 3712 } 3713 return true; 3714 } else { 3715 // The coast is clear. 3716 return false; 3717 } 3718 } 3719 3720 //--------------------------too_many_recompiles-------------------------------- 3721 // Report if there are too many recompiles at the current method and bci. 3722 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3723 // Is not eager to return true, since this will cause the compiler to use 3724 // Action_none for a trap point, to avoid too many recompilations. 3725 bool Compile::too_many_recompiles(ciMethod* method, 3726 int bci, 3727 Deoptimization::DeoptReason reason) { 3728 ciMethodData* md = method->method_data(); 3729 if (md->is_empty()) { 3730 // Assume the trap has not occurred, or that it occurred only 3731 // because of a transient condition during start-up in the interpreter. 3732 return false; 3733 } 3734 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3735 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3736 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3737 Deoptimization::DeoptReason per_bc_reason 3738 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3739 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3740 if ((per_bc_reason == Deoptimization::Reason_none 3741 || md->has_trap_at(bci, m, reason) != 0) 3742 // The trap frequency measure we care about is the recompile count: 3743 && md->trap_recompiled_at(bci, m) 3744 && md->overflow_recompile_count() >= bc_cutoff) { 3745 // Do not emit a trap here if it has already caused recompilations. 3746 // Also, if there are multiple reasons, or if there is no per-BCI record, 3747 // assume the worst. 3748 if (log()) 3749 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3750 Deoptimization::trap_reason_name(reason), 3751 md->trap_count(reason), 3752 md->overflow_recompile_count()); 3753 return true; 3754 } else if (trap_count(reason) != 0 3755 && decompile_count() >= m_cutoff) { 3756 // Too many recompiles globally, and we have seen this sort of trap. 3757 // Use cumulative decompile_count, not just md->decompile_count. 3758 if (log()) 3759 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3760 Deoptimization::trap_reason_name(reason), 3761 md->trap_count(reason), trap_count(reason), 3762 md->decompile_count(), decompile_count()); 3763 return true; 3764 } else { 3765 // The coast is clear. 3766 return false; 3767 } 3768 } 3769 3770 // Compute when not to trap. Used by matching trap based nodes and 3771 // NullCheck optimization. 3772 void Compile::set_allowed_deopt_reasons() { 3773 _allowed_reasons = 0; 3774 if (is_method_compilation()) { 3775 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3776 assert(rs < BitsPerInt, "recode bit map"); 3777 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3778 _allowed_reasons |= nth_bit(rs); 3779 } 3780 } 3781 } 3782 } 3783 3784 #ifndef PRODUCT 3785 //------------------------------verify_graph_edges--------------------------- 3786 // Walk the Graph and verify that there is a one-to-one correspondence 3787 // between Use-Def edges and Def-Use edges in the graph. 3788 void Compile::verify_graph_edges(bool no_dead_code) { 3789 if (VerifyGraphEdges) { 3790 ResourceArea *area = Thread::current()->resource_area(); 3791 Unique_Node_List visited(area); 3792 // Call recursive graph walk to check edges 3793 _root->verify_edges(visited); 3794 if (no_dead_code) { 3795 // Now make sure that no visited node is used by an unvisited node. 3796 bool dead_nodes = false; 3797 Unique_Node_List checked(area); 3798 while (visited.size() > 0) { 3799 Node* n = visited.pop(); 3800 checked.push(n); 3801 for (uint i = 0; i < n->outcnt(); i++) { 3802 Node* use = n->raw_out(i); 3803 if (checked.member(use)) continue; // already checked 3804 if (visited.member(use)) continue; // already in the graph 3805 if (use->is_Con()) continue; // a dead ConNode is OK 3806 // At this point, we have found a dead node which is DU-reachable. 3807 if (!dead_nodes) { 3808 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3809 dead_nodes = true; 3810 } 3811 use->dump(2); 3812 tty->print_cr("---"); 3813 checked.push(use); // No repeats; pretend it is now checked. 3814 } 3815 } 3816 assert(!dead_nodes, "using nodes must be reachable from root"); 3817 } 3818 } 3819 } 3820 3821 // Verify GC barriers consistency 3822 // Currently supported: 3823 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3824 void Compile::verify_barriers() { 3825 #if INCLUDE_G1GC 3826 if (UseG1GC) { 3827 // Verify G1 pre-barriers 3828 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()); 3829 3830 ResourceArea *area = Thread::current()->resource_area(); 3831 Unique_Node_List visited(area); 3832 Node_List worklist(area); 3833 // We're going to walk control flow backwards starting from the Root 3834 worklist.push(_root); 3835 while (worklist.size() > 0) { 3836 Node* x = worklist.pop(); 3837 if (x == NULL || x == top()) continue; 3838 if (visited.member(x)) { 3839 continue; 3840 } else { 3841 visited.push(x); 3842 } 3843 3844 if (x->is_Region()) { 3845 for (uint i = 1; i < x->req(); i++) { 3846 worklist.push(x->in(i)); 3847 } 3848 } else { 3849 worklist.push(x->in(0)); 3850 // We are looking for the pattern: 3851 // /->ThreadLocal 3852 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3853 // \->ConI(0) 3854 // We want to verify that the If and the LoadB have the same control 3855 // See GraphKit::g1_write_barrier_pre() 3856 if (x->is_If()) { 3857 IfNode *iff = x->as_If(); 3858 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3859 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3860 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3861 && cmp->in(1)->is_Load()) { 3862 LoadNode* load = cmp->in(1)->as_Load(); 3863 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3864 && load->in(2)->in(3)->is_Con() 3865 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3866 3867 Node* if_ctrl = iff->in(0); 3868 Node* load_ctrl = load->in(0); 3869 3870 if (if_ctrl != load_ctrl) { 3871 // Skip possible CProj->NeverBranch in infinite loops 3872 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3873 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3874 if_ctrl = if_ctrl->in(0)->in(0); 3875 } 3876 } 3877 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3878 } 3879 } 3880 } 3881 } 3882 } 3883 } 3884 } 3885 #endif 3886 } 3887 3888 #endif 3889 3890 // The Compile object keeps track of failure reasons separately from the ciEnv. 3891 // This is required because there is not quite a 1-1 relation between the 3892 // ciEnv and its compilation task and the Compile object. Note that one 3893 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3894 // to backtrack and retry without subsuming loads. Other than this backtracking 3895 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3896 // by the logic in C2Compiler. 3897 void Compile::record_failure(const char* reason) { 3898 if (log() != NULL) { 3899 log()->elem("failure reason='%s' phase='compile'", reason); 3900 } 3901 if (_failure_reason == NULL) { 3902 // Record the first failure reason. 3903 _failure_reason = reason; 3904 } 3905 3906 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3907 C->print_method(PHASE_FAILURE); 3908 } 3909 _root = NULL; // flush the graph, too 3910 } 3911 3912 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3913 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3914 _phase_name(name), _dolog(CITimeVerbose) 3915 { 3916 if (_dolog) { 3917 C = Compile::current(); 3918 _log = C->log(); 3919 } else { 3920 C = NULL; 3921 _log = NULL; 3922 } 3923 if (_log != NULL) { 3924 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3925 _log->stamp(); 3926 _log->end_head(); 3927 } 3928 } 3929 3930 Compile::TracePhase::~TracePhase() { 3931 3932 C = Compile::current(); 3933 if (_dolog) { 3934 _log = C->log(); 3935 } else { 3936 _log = NULL; 3937 } 3938 3939 #ifdef ASSERT 3940 if (PrintIdealNodeCount) { 3941 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3942 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3943 } 3944 3945 if (VerifyIdealNodeCount) { 3946 Compile::current()->print_missing_nodes(); 3947 } 3948 #endif 3949 3950 if (_log != NULL) { 3951 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3952 } 3953 } 3954 3955 //============================================================================= 3956 // Two Constant's are equal when the type and the value are equal. 3957 bool Compile::Constant::operator==(const Constant& other) { 3958 if (type() != other.type() ) return false; 3959 if (can_be_reused() != other.can_be_reused()) return false; 3960 // For floating point values we compare the bit pattern. 3961 switch (type()) { 3962 case T_INT: 3963 case T_FLOAT: return (_v._value.i == other._v._value.i); 3964 case T_LONG: 3965 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3966 case T_OBJECT: 3967 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3968 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3969 case T_METADATA: return (_v._metadata == other._v._metadata); 3970 default: ShouldNotReachHere(); return false; 3971 } 3972 } 3973 3974 static int type_to_size_in_bytes(BasicType t) { 3975 switch (t) { 3976 case T_INT: return sizeof(jint ); 3977 case T_LONG: return sizeof(jlong ); 3978 case T_FLOAT: return sizeof(jfloat ); 3979 case T_DOUBLE: return sizeof(jdouble); 3980 case T_METADATA: return sizeof(Metadata*); 3981 // We use T_VOID as marker for jump-table entries (labels) which 3982 // need an internal word relocation. 3983 case T_VOID: 3984 case T_ADDRESS: 3985 case T_OBJECT: return sizeof(jobject); 3986 default: 3987 ShouldNotReachHere(); 3988 return -1; 3989 } 3990 } 3991 3992 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3993 // sort descending 3994 if (a->freq() > b->freq()) return -1; 3995 if (a->freq() < b->freq()) return 1; 3996 return 0; 3997 } 3998 3999 void Compile::ConstantTable::calculate_offsets_and_size() { 4000 // First, sort the array by frequencies. 4001 _constants.sort(qsort_comparator); 4002 4003 #ifdef ASSERT 4004 // Make sure all jump-table entries were sorted to the end of the 4005 // array (they have a negative frequency). 4006 bool found_void = false; 4007 for (int i = 0; i < _constants.length(); i++) { 4008 Constant con = _constants.at(i); 4009 if (con.type() == T_VOID) 4010 found_void = true; // jump-tables 4011 else 4012 assert(!found_void, "wrong sorting"); 4013 } 4014 #endif 4015 4016 int offset = 0; 4017 for (int i = 0; i < _constants.length(); i++) { 4018 Constant* con = _constants.adr_at(i); 4019 4020 // Align offset for type. 4021 int typesize = type_to_size_in_bytes(con->type()); 4022 offset = align_up(offset, typesize); 4023 con->set_offset(offset); // set constant's offset 4024 4025 if (con->type() == T_VOID) { 4026 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4027 offset = offset + typesize * n->outcnt(); // expand jump-table 4028 } else { 4029 offset = offset + typesize; 4030 } 4031 } 4032 4033 // Align size up to the next section start (which is insts; see 4034 // CodeBuffer::align_at_start). 4035 assert(_size == -1, "already set?"); 4036 _size = align_up(offset, (int)CodeEntryAlignment); 4037 } 4038 4039 void Compile::ConstantTable::emit(CodeBuffer& cb) { 4040 MacroAssembler _masm(&cb); 4041 for (int i = 0; i < _constants.length(); i++) { 4042 Constant con = _constants.at(i); 4043 address constant_addr = NULL; 4044 switch (con.type()) { 4045 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4046 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4047 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4048 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4049 case T_OBJECT: { 4050 jobject obj = con.get_jobject(); 4051 int oop_index = _masm.oop_recorder()->find_index(obj); 4052 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4053 break; 4054 } 4055 case T_ADDRESS: { 4056 address addr = (address) con.get_jobject(); 4057 constant_addr = _masm.address_constant(addr); 4058 break; 4059 } 4060 // We use T_VOID as marker for jump-table entries (labels) which 4061 // need an internal word relocation. 4062 case T_VOID: { 4063 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4064 // Fill the jump-table with a dummy word. The real value is 4065 // filled in later in fill_jump_table. 4066 address dummy = (address) n; 4067 constant_addr = _masm.address_constant(dummy); 4068 // Expand jump-table 4069 for (uint i = 1; i < n->outcnt(); i++) { 4070 address temp_addr = _masm.address_constant(dummy + i); 4071 assert(temp_addr, "consts section too small"); 4072 } 4073 break; 4074 } 4075 case T_METADATA: { 4076 Metadata* obj = con.get_metadata(); 4077 int metadata_index = _masm.oop_recorder()->find_index(obj); 4078 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4079 break; 4080 } 4081 default: ShouldNotReachHere(); 4082 } 4083 assert(constant_addr, "consts section too small"); 4084 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4085 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4086 } 4087 } 4088 4089 int Compile::ConstantTable::find_offset(Constant& con) const { 4090 int idx = _constants.find(con); 4091 assert(idx != -1, "constant must be in constant table"); 4092 int offset = _constants.at(idx).offset(); 4093 assert(offset != -1, "constant table not emitted yet?"); 4094 return offset; 4095 } 4096 4097 void Compile::ConstantTable::add(Constant& con) { 4098 if (con.can_be_reused()) { 4099 int idx = _constants.find(con); 4100 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4101 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4102 return; 4103 } 4104 } 4105 (void) _constants.append(con); 4106 } 4107 4108 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4109 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4110 Constant con(type, value, b->_freq); 4111 add(con); 4112 return con; 4113 } 4114 4115 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4116 Constant con(metadata); 4117 add(con); 4118 return con; 4119 } 4120 4121 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4122 jvalue value; 4123 BasicType type = oper->type()->basic_type(); 4124 switch (type) { 4125 case T_LONG: value.j = oper->constantL(); break; 4126 case T_FLOAT: value.f = oper->constantF(); break; 4127 case T_DOUBLE: value.d = oper->constantD(); break; 4128 case T_OBJECT: 4129 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4130 case T_METADATA: return add((Metadata*)oper->constant()); break; 4131 default: guarantee(false, "unhandled type: %s", type2name(type)); 4132 } 4133 return add(n, type, value); 4134 } 4135 4136 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4137 jvalue value; 4138 // We can use the node pointer here to identify the right jump-table 4139 // as this method is called from Compile::Fill_buffer right before 4140 // the MachNodes are emitted and the jump-table is filled (means the 4141 // MachNode pointers do not change anymore). 4142 value.l = (jobject) n; 4143 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4144 add(con); 4145 return con; 4146 } 4147 4148 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4149 // If called from Compile::scratch_emit_size do nothing. 4150 if (Compile::current()->in_scratch_emit_size()) return; 4151 4152 assert(labels.is_nonempty(), "must be"); 4153 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4154 4155 // Since MachConstantNode::constant_offset() also contains 4156 // table_base_offset() we need to subtract the table_base_offset() 4157 // to get the plain offset into the constant table. 4158 int offset = n->constant_offset() - table_base_offset(); 4159 4160 MacroAssembler _masm(&cb); 4161 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4162 4163 for (uint i = 0; i < n->outcnt(); i++) { 4164 address* constant_addr = &jump_table_base[i]; 4165 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)); 4166 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4167 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4168 } 4169 } 4170 4171 //----------------------------static_subtype_check----------------------------- 4172 // Shortcut important common cases when superklass is exact: 4173 // (0) superklass is java.lang.Object (can occur in reflective code) 4174 // (1) subklass is already limited to a subtype of superklass => always ok 4175 // (2) subklass does not overlap with superklass => always fail 4176 // (3) superklass has NO subtypes and we can check with a simple compare. 4177 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4178 if (StressReflectiveCode || superk == NULL || subk == NULL) { 4179 return SSC_full_test; // Let caller generate the general case. 4180 } 4181 4182 if (!EnableValhalla && superk == env()->Object_klass()) { 4183 return SSC_always_true; // (0) this test cannot fail 4184 } 4185 4186 ciType* superelem = superk; 4187 if (superelem->is_array_klass()) 4188 superelem = superelem->as_array_klass()->base_element_type(); 4189 4190 if (!subk->is_interface()) { // cannot trust static interface types yet 4191 if (subk->is_subtype_of(superk)) { 4192 return SSC_always_true; // (1) false path dead; no dynamic test needed 4193 } 4194 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4195 !superk->is_subtype_of(subk)) { 4196 return SSC_always_false; 4197 } 4198 } 4199 4200 // If casting to an instance klass, it must have no subtypes 4201 if (superk->is_interface()) { 4202 // Cannot trust interfaces yet. 4203 // %%% S.B. superk->nof_implementors() == 1 4204 } else if (superelem->is_instance_klass()) { 4205 ciInstanceKlass* ik = superelem->as_instance_klass(); 4206 if (!ik->has_subklass() && !ik->is_interface()) { 4207 if (!ik->is_final()) { 4208 // Add a dependency if there is a chance of a later subclass. 4209 dependencies()->assert_leaf_type(ik); 4210 } 4211 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4212 } 4213 } else { 4214 // A primitive array type has no subtypes. 4215 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4216 } 4217 4218 return SSC_full_test; 4219 } 4220 4221 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4222 #ifdef _LP64 4223 // The scaled index operand to AddP must be a clean 64-bit value. 4224 // Java allows a 32-bit int to be incremented to a negative 4225 // value, which appears in a 64-bit register as a large 4226 // positive number. Using that large positive number as an 4227 // operand in pointer arithmetic has bad consequences. 4228 // On the other hand, 32-bit overflow is rare, and the possibility 4229 // can often be excluded, if we annotate the ConvI2L node with 4230 // a type assertion that its value is known to be a small positive 4231 // number. (The prior range check has ensured this.) 4232 // This assertion is used by ConvI2LNode::Ideal. 4233 int index_max = max_jint - 1; // array size is max_jint, index is one less 4234 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4235 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4236 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4237 #endif 4238 return idx; 4239 } 4240 4241 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4242 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4243 if (ctrl != NULL) { 4244 // Express control dependency by a CastII node with a narrow type. 4245 value = new CastIINode(value, itype, false, true /* range check dependency */); 4246 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4247 // node from floating above the range check during loop optimizations. Otherwise, the 4248 // ConvI2L node may be eliminated independently of the range check, causing the data path 4249 // to become TOP while the control path is still there (although it's unreachable). 4250 value->set_req(0, ctrl); 4251 // Save CastII node to remove it after loop optimizations. 4252 phase->C->add_range_check_cast(value); 4253 value = phase->transform(value); 4254 } 4255 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4256 return phase->transform(new ConvI2LNode(value, ltype)); 4257 } 4258 4259 // The message about the current inlining is accumulated in 4260 // _print_inlining_stream and transfered into the _print_inlining_list 4261 // once we know whether inlining succeeds or not. For regular 4262 // inlining, messages are appended to the buffer pointed by 4263 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4264 // a new buffer is added after _print_inlining_idx in the list. This 4265 // way we can update the inlining message for late inlining call site 4266 // when the inlining is attempted again. 4267 void Compile::print_inlining_init() { 4268 if (print_inlining() || print_intrinsics()) { 4269 _print_inlining_stream = new stringStream(); 4270 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4271 } 4272 } 4273 4274 void Compile::print_inlining_reinit() { 4275 if (print_inlining() || print_intrinsics()) { 4276 // Re allocate buffer when we change ResourceMark 4277 _print_inlining_stream = new stringStream(); 4278 } 4279 } 4280 4281 void Compile::print_inlining_reset() { 4282 _print_inlining_stream->reset(); 4283 } 4284 4285 void Compile::print_inlining_commit() { 4286 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4287 // Transfer the message from _print_inlining_stream to the current 4288 // _print_inlining_list buffer and clear _print_inlining_stream. 4289 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4290 print_inlining_reset(); 4291 } 4292 4293 void Compile::print_inlining_push() { 4294 // Add new buffer to the _print_inlining_list at current position 4295 _print_inlining_idx++; 4296 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4297 } 4298 4299 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4300 return _print_inlining_list->at(_print_inlining_idx); 4301 } 4302 4303 void Compile::print_inlining_update(CallGenerator* cg) { 4304 if (print_inlining() || print_intrinsics()) { 4305 if (!cg->is_late_inline()) { 4306 if (print_inlining_current().cg() != NULL) { 4307 print_inlining_push(); 4308 } 4309 print_inlining_commit(); 4310 } else { 4311 if (print_inlining_current().cg() != cg && 4312 (print_inlining_current().cg() != NULL || 4313 print_inlining_current().ss()->size() != 0)) { 4314 print_inlining_push(); 4315 } 4316 print_inlining_commit(); 4317 print_inlining_current().set_cg(cg); 4318 } 4319 } 4320 } 4321 4322 void Compile::print_inlining_move_to(CallGenerator* cg) { 4323 // We resume inlining at a late inlining call site. Locate the 4324 // corresponding inlining buffer so that we can update it. 4325 if (print_inlining()) { 4326 for (int i = 0; i < _print_inlining_list->length(); i++) { 4327 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4328 _print_inlining_idx = i; 4329 return; 4330 } 4331 } 4332 ShouldNotReachHere(); 4333 } 4334 } 4335 4336 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4337 if (print_inlining()) { 4338 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4339 assert(print_inlining_current().cg() == cg, "wrong entry"); 4340 // replace message with new message 4341 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4342 print_inlining_commit(); 4343 print_inlining_current().set_cg(cg); 4344 } 4345 } 4346 4347 void Compile::print_inlining_assert_ready() { 4348 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4349 } 4350 4351 void Compile::process_print_inlining() { 4352 bool do_print_inlining = print_inlining() || print_intrinsics(); 4353 if (do_print_inlining || log() != NULL) { 4354 // Print inlining message for candidates that we couldn't inline 4355 // for lack of space 4356 for (int i = 0; i < _late_inlines.length(); i++) { 4357 CallGenerator* cg = _late_inlines.at(i); 4358 if (!cg->is_mh_late_inline()) { 4359 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4360 if (do_print_inlining) { 4361 cg->print_inlining_late(msg); 4362 } 4363 log_late_inline_failure(cg, msg); 4364 } 4365 } 4366 } 4367 if (do_print_inlining) { 4368 ResourceMark rm; 4369 stringStream ss; 4370 for (int i = 0; i < _print_inlining_list->length(); i++) { 4371 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4372 } 4373 size_t end = ss.size(); 4374 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4375 strncpy(_print_inlining_output, ss.base(), end+1); 4376 _print_inlining_output[end] = 0; 4377 } 4378 } 4379 4380 void Compile::dump_print_inlining() { 4381 if (_print_inlining_output != NULL) { 4382 tty->print_raw(_print_inlining_output); 4383 } 4384 } 4385 4386 void Compile::log_late_inline(CallGenerator* cg) { 4387 if (log() != NULL) { 4388 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4389 cg->unique_id()); 4390 JVMState* p = cg->call_node()->jvms(); 4391 while (p != NULL) { 4392 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4393 p = p->caller(); 4394 } 4395 log()->tail("late_inline"); 4396 } 4397 } 4398 4399 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4400 log_late_inline(cg); 4401 if (log() != NULL) { 4402 log()->inline_fail(msg); 4403 } 4404 } 4405 4406 void Compile::log_inline_id(CallGenerator* cg) { 4407 if (log() != NULL) { 4408 // The LogCompilation tool needs a unique way to identify late 4409 // inline call sites. This id must be unique for this call site in 4410 // this compilation. Try to have it unique across compilations as 4411 // well because it can be convenient when grepping through the log 4412 // file. 4413 // Distinguish OSR compilations from others in case CICountOSR is 4414 // on. 4415 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4416 cg->set_unique_id(id); 4417 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4418 } 4419 } 4420 4421 void Compile::log_inline_failure(const char* msg) { 4422 if (C->log() != NULL) { 4423 C->log()->inline_fail(msg); 4424 } 4425 } 4426 4427 4428 // Dump inlining replay data to the stream. 4429 // Don't change thread state and acquire any locks. 4430 void Compile::dump_inline_data(outputStream* out) { 4431 InlineTree* inl_tree = ilt(); 4432 if (inl_tree != NULL) { 4433 out->print(" inline %d", inl_tree->count()); 4434 inl_tree->dump_replay_data(out); 4435 } 4436 } 4437 4438 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4439 if (n1->Opcode() < n2->Opcode()) return -1; 4440 else if (n1->Opcode() > n2->Opcode()) return 1; 4441 4442 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4443 for (uint i = 1; i < n1->req(); i++) { 4444 if (n1->in(i) < n2->in(i)) return -1; 4445 else if (n1->in(i) > n2->in(i)) return 1; 4446 } 4447 4448 return 0; 4449 } 4450 4451 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4452 Node* n1 = *n1p; 4453 Node* n2 = *n2p; 4454 4455 return cmp_expensive_nodes(n1, n2); 4456 } 4457 4458 void Compile::sort_expensive_nodes() { 4459 if (!expensive_nodes_sorted()) { 4460 _expensive_nodes->sort(cmp_expensive_nodes); 4461 } 4462 } 4463 4464 bool Compile::expensive_nodes_sorted() const { 4465 for (int i = 1; i < _expensive_nodes->length(); i++) { 4466 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4467 return false; 4468 } 4469 } 4470 return true; 4471 } 4472 4473 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4474 if (_expensive_nodes->length() == 0) { 4475 return false; 4476 } 4477 4478 assert(OptimizeExpensiveOps, "optimization off?"); 4479 4480 // Take this opportunity to remove dead nodes from the list 4481 int j = 0; 4482 for (int i = 0; i < _expensive_nodes->length(); i++) { 4483 Node* n = _expensive_nodes->at(i); 4484 if (!n->is_unreachable(igvn)) { 4485 assert(n->is_expensive(), "should be expensive"); 4486 _expensive_nodes->at_put(j, n); 4487 j++; 4488 } 4489 } 4490 _expensive_nodes->trunc_to(j); 4491 4492 // Then sort the list so that similar nodes are next to each other 4493 // and check for at least two nodes of identical kind with same data 4494 // inputs. 4495 sort_expensive_nodes(); 4496 4497 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4498 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4499 return true; 4500 } 4501 } 4502 4503 return false; 4504 } 4505 4506 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4507 if (_expensive_nodes->length() == 0) { 4508 return; 4509 } 4510 4511 assert(OptimizeExpensiveOps, "optimization off?"); 4512 4513 // Sort to bring similar nodes next to each other and clear the 4514 // control input of nodes for which there's only a single copy. 4515 sort_expensive_nodes(); 4516 4517 int j = 0; 4518 int identical = 0; 4519 int i = 0; 4520 bool modified = false; 4521 for (; i < _expensive_nodes->length()-1; i++) { 4522 assert(j <= i, "can't write beyond current index"); 4523 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4524 identical++; 4525 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4526 continue; 4527 } 4528 if (identical > 0) { 4529 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4530 identical = 0; 4531 } else { 4532 Node* n = _expensive_nodes->at(i); 4533 igvn.replace_input_of(n, 0, NULL); 4534 igvn.hash_insert(n); 4535 modified = true; 4536 } 4537 } 4538 if (identical > 0) { 4539 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4540 } else if (_expensive_nodes->length() >= 1) { 4541 Node* n = _expensive_nodes->at(i); 4542 igvn.replace_input_of(n, 0, NULL); 4543 igvn.hash_insert(n); 4544 modified = true; 4545 } 4546 _expensive_nodes->trunc_to(j); 4547 if (modified) { 4548 igvn.optimize(); 4549 } 4550 } 4551 4552 void Compile::add_expensive_node(Node * n) { 4553 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4554 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4555 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4556 if (OptimizeExpensiveOps) { 4557 _expensive_nodes->append(n); 4558 } else { 4559 // Clear control input and let IGVN optimize expensive nodes if 4560 // OptimizeExpensiveOps is off. 4561 n->set_req(0, NULL); 4562 } 4563 } 4564 4565 /** 4566 * Remove the speculative part of types and clean up the graph 4567 */ 4568 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4569 if (UseTypeSpeculation) { 4570 Unique_Node_List worklist; 4571 worklist.push(root()); 4572 int modified = 0; 4573 // Go over all type nodes that carry a speculative type, drop the 4574 // speculative part of the type and enqueue the node for an igvn 4575 // which may optimize it out. 4576 for (uint next = 0; next < worklist.size(); ++next) { 4577 Node *n = worklist.at(next); 4578 if (n->is_Type()) { 4579 TypeNode* tn = n->as_Type(); 4580 const Type* t = tn->type(); 4581 const Type* t_no_spec = t->remove_speculative(); 4582 if (t_no_spec != t) { 4583 bool in_hash = igvn.hash_delete(n); 4584 assert(in_hash, "node should be in igvn hash table"); 4585 tn->set_type(t_no_spec); 4586 igvn.hash_insert(n); 4587 igvn._worklist.push(n); // give it a chance to go away 4588 modified++; 4589 } 4590 } 4591 uint max = n->len(); 4592 for( uint i = 0; i < max; ++i ) { 4593 Node *m = n->in(i); 4594 if (not_a_node(m)) continue; 4595 worklist.push(m); 4596 } 4597 } 4598 // Drop the speculative part of all types in the igvn's type table 4599 igvn.remove_speculative_types(); 4600 if (modified > 0) { 4601 igvn.optimize(); 4602 } 4603 #ifdef ASSERT 4604 // Verify that after the IGVN is over no speculative type has resurfaced 4605 worklist.clear(); 4606 worklist.push(root()); 4607 for (uint next = 0; next < worklist.size(); ++next) { 4608 Node *n = worklist.at(next); 4609 const Type* t = igvn.type_or_null(n); 4610 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4611 if (n->is_Type()) { 4612 t = n->as_Type()->type(); 4613 assert(t == t->remove_speculative(), "no more speculative types"); 4614 } 4615 uint max = n->len(); 4616 for( uint i = 0; i < max; ++i ) { 4617 Node *m = n->in(i); 4618 if (not_a_node(m)) continue; 4619 worklist.push(m); 4620 } 4621 } 4622 igvn.check_no_speculative_types(); 4623 #endif 4624 } 4625 } 4626 4627 Node* Compile::load_is_value_bit(PhaseGVN* phase, Node* oop) { 4628 // Load the klass pointer and check if it's odd, i.e., if it defines a value type 4629 // is_value = (klass & oop_metadata_valuetype_mask) >> LogKlassAlignmentInBytes 4630 Node* k_adr = phase->transform(new AddPNode(oop, oop, phase->MakeConX(oopDesc::klass_offset_in_bytes()))); 4631 Node* klass = NULL; 4632 if (UseCompressedClassPointers) { 4633 klass = phase->transform(new LoadNKlassNode(NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT->make_narrowklass(), MemNode::unordered)); 4634 } else { 4635 klass = phase->transform(new LoadKlassNode(NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT, MemNode::unordered)); 4636 } 4637 const int mask = Universe::oop_metadata_valuetype_mask(); 4638 Node* is_value = phase->transform(new CastP2XNode(NULL, klass)); 4639 is_value = phase->transform(new AndXNode(is_value, phase->MakeConX(mask))); 4640 // Check if a shift is required for perturbation to affect aligned bits of oop 4641 if (mask == KlassPtrValueTypeMask && ObjectAlignmentInBytes <= KlassAlignmentInBytes) { 4642 assert((mask >> LogKlassAlignmentInBytes) == 1, "invalid shift"); 4643 is_value = phase->transform(new URShiftXNode(is_value, phase->intcon(LogKlassAlignmentInBytes))); 4644 } else { 4645 assert(mask < ObjectAlignmentInBytes, "invalid mask"); 4646 } 4647 return is_value; 4648 } 4649 4650 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) { 4651 const TypeInstPtr* ta = phase->type(a)->isa_instptr(); 4652 const TypeInstPtr* tb = phase->type(b)->isa_instptr(); 4653 if (!EnableValhalla || ta == NULL || tb == NULL || 4654 ta->is_zero_type() || tb->is_zero_type() || 4655 !ta->can_be_value_type() || !tb->can_be_value_type()) { 4656 // Use old acmp if one operand is null or not a value type 4657 return new CmpPNode(a, b); 4658 } else if (ta->is_valuetypeptr() || tb->is_valuetypeptr()) { 4659 // We know that one operand is a value type. Therefore, 4660 // new acmp will only return true if both operands are NULL. 4661 // Check if both operands are null by or'ing the oops. 4662 a = phase->transform(new CastP2XNode(NULL, a)); 4663 b = phase->transform(new CastP2XNode(NULL, b)); 4664 a = phase->transform(new OrXNode(a, b)); 4665 return new CmpXNode(a, phase->MakeConX(0)); 4666 } 4667 // Use new acmp 4668 return NULL; 4669 } 4670 4671 // Auxiliary method to support randomized stressing/fuzzing. 4672 // 4673 // This method can be called the arbitrary number of times, with current count 4674 // as the argument. The logic allows selecting a single candidate from the 4675 // running list of candidates as follows: 4676 // int count = 0; 4677 // Cand* selected = null; 4678 // while(cand = cand->next()) { 4679 // if (randomized_select(++count)) { 4680 // selected = cand; 4681 // } 4682 // } 4683 // 4684 // Including count equalizes the chances any candidate is "selected". 4685 // This is useful when we don't have the complete list of candidates to choose 4686 // from uniformly. In this case, we need to adjust the randomicity of the 4687 // selection, or else we will end up biasing the selection towards the latter 4688 // candidates. 4689 // 4690 // Quick back-envelope calculation shows that for the list of n candidates 4691 // the equal probability for the candidate to persist as "best" can be 4692 // achieved by replacing it with "next" k-th candidate with the probability 4693 // of 1/k. It can be easily shown that by the end of the run, the 4694 // probability for any candidate is converged to 1/n, thus giving the 4695 // uniform distribution among all the candidates. 4696 // 4697 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4698 #define RANDOMIZED_DOMAIN_POW 29 4699 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4700 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4701 bool Compile::randomized_select(int count) { 4702 assert(count > 0, "only positive"); 4703 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4704 } 4705 4706 CloneMap& Compile::clone_map() { return _clone_map; } 4707 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4708 4709 void NodeCloneInfo::dump() const { 4710 tty->print(" {%d:%d} ", idx(), gen()); 4711 } 4712 4713 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4714 uint64_t val = value(old->_idx); 4715 NodeCloneInfo cio(val); 4716 assert(val != 0, "old node should be in the map"); 4717 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4718 insert(nnn->_idx, cin.get()); 4719 #ifndef PRODUCT 4720 if (is_debug()) { 4721 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4722 } 4723 #endif 4724 } 4725 4726 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4727 NodeCloneInfo cio(value(old->_idx)); 4728 if (cio.get() == 0) { 4729 cio.set(old->_idx, 0); 4730 insert(old->_idx, cio.get()); 4731 #ifndef PRODUCT 4732 if (is_debug()) { 4733 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4734 } 4735 #endif 4736 } 4737 clone(old, nnn, gen); 4738 } 4739 4740 int CloneMap::max_gen() const { 4741 int g = 0; 4742 DictI di(_dict); 4743 for(; di.test(); ++di) { 4744 int t = gen(di._key); 4745 if (g < t) { 4746 g = t; 4747 #ifndef PRODUCT 4748 if (is_debug()) { 4749 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4750 } 4751 #endif 4752 } 4753 } 4754 return g; 4755 } 4756 4757 void CloneMap::dump(node_idx_t key) const { 4758 uint64_t val = value(key); 4759 if (val != 0) { 4760 NodeCloneInfo ni(val); 4761 ni.dump(); 4762 } 4763 }