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 #include "utilities/macros.hpp" 80 #if INCLUDE_G1GC 81 #include "gc/g1/g1ThreadLocalData.hpp" 82 #endif // INCLUDE_G1GC 83 #if INCLUDE_ZGC 84 #include "gc/z/c2/zBarrierSetC2.hpp" 85 #endif 86 87 88 // -------------------- Compile::mach_constant_base_node ----------------------- 89 // Constant table base node singleton. 90 MachConstantBaseNode* Compile::mach_constant_base_node() { 91 if (_mach_constant_base_node == NULL) { 92 _mach_constant_base_node = new MachConstantBaseNode(); 93 _mach_constant_base_node->add_req(C->root()); 94 } 95 return _mach_constant_base_node; 96 } 97 98 99 /// Support for intrinsics. 100 101 // Return the index at which m must be inserted (or already exists). 102 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 103 class IntrinsicDescPair { 104 private: 105 ciMethod* _m; 106 bool _is_virtual; 107 public: 108 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 109 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 110 ciMethod* m= elt->method(); 111 ciMethod* key_m = key->_m; 112 if (key_m < m) return -1; 113 else if (key_m > m) return 1; 114 else { 115 bool is_virtual = elt->is_virtual(); 116 bool key_virtual = key->_is_virtual; 117 if (key_virtual < is_virtual) return -1; 118 else if (key_virtual > is_virtual) return 1; 119 else return 0; 120 } 121 } 122 }; 123 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 124 #ifdef ASSERT 125 for (int i = 1; i < _intrinsics->length(); i++) { 126 CallGenerator* cg1 = _intrinsics->at(i-1); 127 CallGenerator* cg2 = _intrinsics->at(i); 128 assert(cg1->method() != cg2->method() 129 ? cg1->method() < cg2->method() 130 : cg1->is_virtual() < cg2->is_virtual(), 131 "compiler intrinsics list must stay sorted"); 132 } 133 #endif 134 IntrinsicDescPair pair(m, is_virtual); 135 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 136 } 137 138 void Compile::register_intrinsic(CallGenerator* cg) { 139 if (_intrinsics == NULL) { 140 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 141 } 142 int len = _intrinsics->length(); 143 bool found = false; 144 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 145 assert(!found, "registering twice"); 146 _intrinsics->insert_before(index, cg); 147 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 148 } 149 150 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 151 assert(m->is_loaded(), "don't try this on unloaded methods"); 152 if (_intrinsics != NULL) { 153 bool found = false; 154 int index = intrinsic_insertion_index(m, is_virtual, found); 155 if (found) { 156 return _intrinsics->at(index); 157 } 158 } 159 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 160 if (m->intrinsic_id() != vmIntrinsics::_none && 161 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 162 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 163 if (cg != NULL) { 164 // Save it for next time: 165 register_intrinsic(cg); 166 return cg; 167 } else { 168 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 169 } 170 } 171 return NULL; 172 } 173 174 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 175 // in library_call.cpp. 176 177 178 #ifndef PRODUCT 179 // statistics gathering... 180 181 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 182 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 183 184 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 185 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 186 int oflags = _intrinsic_hist_flags[id]; 187 assert(flags != 0, "what happened?"); 188 if (is_virtual) { 189 flags |= _intrinsic_virtual; 190 } 191 bool changed = (flags != oflags); 192 if ((flags & _intrinsic_worked) != 0) { 193 juint count = (_intrinsic_hist_count[id] += 1); 194 if (count == 1) { 195 changed = true; // first time 196 } 197 // increment the overall count also: 198 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 199 } 200 if (changed) { 201 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 202 // Something changed about the intrinsic's virtuality. 203 if ((flags & _intrinsic_virtual) != 0) { 204 // This is the first use of this intrinsic as a virtual call. 205 if (oflags != 0) { 206 // We already saw it as a non-virtual, so note both cases. 207 flags |= _intrinsic_both; 208 } 209 } else if ((oflags & _intrinsic_both) == 0) { 210 // This is the first use of this intrinsic as a non-virtual 211 flags |= _intrinsic_both; 212 } 213 } 214 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 215 } 216 // update the overall flags also: 217 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 218 return changed; 219 } 220 221 static char* format_flags(int flags, char* buf) { 222 buf[0] = 0; 223 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 224 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 225 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 226 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 227 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 228 if (buf[0] == 0) strcat(buf, ","); 229 assert(buf[0] == ',', "must be"); 230 return &buf[1]; 231 } 232 233 void Compile::print_intrinsic_statistics() { 234 char flagsbuf[100]; 235 ttyLocker ttyl; 236 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 237 tty->print_cr("Compiler intrinsic usage:"); 238 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 239 if (total == 0) total = 1; // avoid div0 in case of no successes 240 #define PRINT_STAT_LINE(name, c, f) \ 241 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 242 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 243 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 244 int flags = _intrinsic_hist_flags[id]; 245 juint count = _intrinsic_hist_count[id]; 246 if ((flags | count) != 0) { 247 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 248 } 249 } 250 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 251 if (xtty != NULL) xtty->tail("statistics"); 252 } 253 254 void Compile::print_statistics() { 255 { ttyLocker ttyl; 256 if (xtty != NULL) xtty->head("statistics type='opto'"); 257 Parse::print_statistics(); 258 PhaseCCP::print_statistics(); 259 PhaseRegAlloc::print_statistics(); 260 Scheduling::print_statistics(); 261 PhasePeephole::print_statistics(); 262 PhaseIdealLoop::print_statistics(); 263 if (xtty != NULL) xtty->tail("statistics"); 264 } 265 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 266 // put this under its own <statistics> element. 267 print_intrinsic_statistics(); 268 } 269 } 270 #endif //PRODUCT 271 272 // Support for bundling info 273 Bundle* Compile::node_bundling(const Node *n) { 274 assert(valid_bundle_info(n), "oob"); 275 return &_node_bundling_base[n->_idx]; 276 } 277 278 bool Compile::valid_bundle_info(const Node *n) { 279 return (_node_bundling_limit > n->_idx); 280 } 281 282 283 void Compile::gvn_replace_by(Node* n, Node* nn) { 284 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 285 Node* use = n->last_out(i); 286 bool is_in_table = initial_gvn()->hash_delete(use); 287 uint uses_found = 0; 288 for (uint j = 0; j < use->len(); j++) { 289 if (use->in(j) == n) { 290 if (j < use->req()) 291 use->set_req(j, nn); 292 else 293 use->set_prec(j, nn); 294 uses_found++; 295 } 296 } 297 if (is_in_table) { 298 // reinsert into table 299 initial_gvn()->hash_find_insert(use); 300 } 301 record_for_igvn(use); 302 i -= uses_found; // we deleted 1 or more copies of this edge 303 } 304 } 305 306 307 static inline bool not_a_node(const Node* n) { 308 if (n == NULL) return true; 309 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 310 if (*(address*)n == badAddress) return true; // kill by Node::destruct 311 return false; 312 } 313 314 // Identify all nodes that are reachable from below, useful. 315 // Use breadth-first pass that records state in a Unique_Node_List, 316 // recursive traversal is slower. 317 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 318 int estimated_worklist_size = live_nodes(); 319 useful.map( estimated_worklist_size, NULL ); // preallocate space 320 321 // Initialize worklist 322 if (root() != NULL) { useful.push(root()); } 323 // If 'top' is cached, declare it useful to preserve cached node 324 if( cached_top_node() ) { useful.push(cached_top_node()); } 325 326 // Push all useful nodes onto the list, breadthfirst 327 for( uint next = 0; next < useful.size(); ++next ) { 328 assert( next < unique(), "Unique useful nodes < total nodes"); 329 Node *n = useful.at(next); 330 uint max = n->len(); 331 for( uint i = 0; i < max; ++i ) { 332 Node *m = n->in(i); 333 if (not_a_node(m)) continue; 334 useful.push(m); 335 } 336 } 337 } 338 339 // Update dead_node_list with any missing dead nodes using useful 340 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 341 void Compile::update_dead_node_list(Unique_Node_List &useful) { 342 uint max_idx = unique(); 343 VectorSet& useful_node_set = useful.member_set(); 344 345 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 346 // If node with index node_idx is not in useful set, 347 // mark it as dead in dead node list. 348 if (! useful_node_set.test(node_idx) ) { 349 record_dead_node(node_idx); 350 } 351 } 352 } 353 354 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 355 int shift = 0; 356 for (int i = 0; i < inlines->length(); i++) { 357 CallGenerator* cg = inlines->at(i); 358 CallNode* call = cg->call_node(); 359 if (shift > 0) { 360 inlines->at_put(i-shift, cg); 361 } 362 if (!useful.member(call)) { 363 shift++; 364 } 365 } 366 inlines->trunc_to(inlines->length()-shift); 367 } 368 369 // Disconnect all useless nodes by disconnecting those at the boundary. 370 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 371 uint next = 0; 372 while (next < useful.size()) { 373 Node *n = useful.at(next++); 374 if (n->is_SafePoint()) { 375 // We're done with a parsing phase. Replaced nodes are not valid 376 // beyond that point. 377 n->as_SafePoint()->delete_replaced_nodes(); 378 } 379 // Use raw traversal of out edges since this code removes out edges 380 int max = n->outcnt(); 381 for (int j = 0; j < max; ++j) { 382 Node* child = n->raw_out(j); 383 if (! useful.member(child)) { 384 assert(!child->is_top() || child != top(), 385 "If top is cached in Compile object it is in useful list"); 386 // Only need to remove this out-edge to the useless node 387 n->raw_del_out(j); 388 --j; 389 --max; 390 } 391 } 392 if (n->outcnt() == 1 && n->has_special_unique_user()) { 393 record_for_igvn(n->unique_out()); 394 } 395 } 396 // Remove useless macro and predicate opaq nodes 397 for (int i = C->macro_count()-1; i >= 0; i--) { 398 Node* n = C->macro_node(i); 399 if (!useful.member(n)) { 400 remove_macro_node(n); 401 } 402 } 403 // Remove useless CastII nodes with range check dependency 404 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 405 Node* cast = range_check_cast_node(i); 406 if (!useful.member(cast)) { 407 remove_range_check_cast(cast); 408 } 409 } 410 // Remove useless expensive nodes 411 for (int i = C->expensive_count()-1; i >= 0; i--) { 412 Node* n = C->expensive_node(i); 413 if (!useful.member(n)) { 414 remove_expensive_node(n); 415 } 416 } 417 // Remove useless Opaque4 nodes 418 for (int i = opaque4_count() - 1; i >= 0; i--) { 419 Node* opaq = opaque4_node(i); 420 if (!useful.member(opaq)) { 421 remove_opaque4_node(opaq); 422 } 423 } 424 // Remove useless value type nodes 425 if (_value_type_nodes != NULL) { 426 _value_type_nodes->remove_useless_nodes(useful.member_set()); 427 } 428 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 429 bs->eliminate_useless_gc_barriers(useful); 430 // clean up the late inline lists 431 remove_useless_late_inlines(&_string_late_inlines, useful); 432 remove_useless_late_inlines(&_boxing_late_inlines, useful); 433 remove_useless_late_inlines(&_late_inlines, useful); 434 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 435 } 436 437 //------------------------------frame_size_in_words----------------------------- 438 // frame_slots in units of words 439 int Compile::frame_size_in_words() const { 440 // shift is 0 in LP32 and 1 in LP64 441 const int shift = (LogBytesPerWord - LogBytesPerInt); 442 int words = _frame_slots >> shift; 443 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 444 return words; 445 } 446 447 // To bang the stack of this compiled method we use the stack size 448 // that the interpreter would need in case of a deoptimization. This 449 // removes the need to bang the stack in the deoptimization blob which 450 // in turn simplifies stack overflow handling. 451 int Compile::bang_size_in_bytes() const { 452 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 453 } 454 455 // ============================================================================ 456 //------------------------------CompileWrapper--------------------------------- 457 class CompileWrapper : public StackObj { 458 Compile *const _compile; 459 public: 460 CompileWrapper(Compile* compile); 461 462 ~CompileWrapper(); 463 }; 464 465 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 466 // the Compile* pointer is stored in the current ciEnv: 467 ciEnv* env = compile->env(); 468 assert(env == ciEnv::current(), "must already be a ciEnv active"); 469 assert(env->compiler_data() == NULL, "compile already active?"); 470 env->set_compiler_data(compile); 471 assert(compile == Compile::current(), "sanity"); 472 473 compile->set_type_dict(NULL); 474 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 475 compile->clone_map().set_clone_idx(0); 476 compile->set_type_hwm(NULL); 477 compile->set_type_last_size(0); 478 compile->set_last_tf(NULL, NULL); 479 compile->set_indexSet_arena(NULL); 480 compile->set_indexSet_free_block_list(NULL); 481 compile->init_type_arena(); 482 Type::Initialize(compile); 483 _compile->set_scratch_buffer_blob(NULL); 484 _compile->begin_method(); 485 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 486 } 487 CompileWrapper::~CompileWrapper() { 488 _compile->end_method(); 489 if (_compile->scratch_buffer_blob() != NULL) 490 BufferBlob::free(_compile->scratch_buffer_blob()); 491 _compile->env()->set_compiler_data(NULL); 492 } 493 494 495 //----------------------------print_compile_messages--------------------------- 496 void Compile::print_compile_messages() { 497 #ifndef PRODUCT 498 // Check if recompiling 499 if (_subsume_loads == false && PrintOpto) { 500 // Recompiling without allowing machine instructions to subsume loads 501 tty->print_cr("*********************************************************"); 502 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 503 tty->print_cr("*********************************************************"); 504 } 505 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 506 // Recompiling without escape analysis 507 tty->print_cr("*********************************************************"); 508 tty->print_cr("** Bailout: Recompile without escape analysis **"); 509 tty->print_cr("*********************************************************"); 510 } 511 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 512 // Recompiling without boxing elimination 513 tty->print_cr("*********************************************************"); 514 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 515 tty->print_cr("*********************************************************"); 516 } 517 if (C->directive()->BreakAtCompileOption) { 518 // Open the debugger when compiling this method. 519 tty->print("### Breaking when compiling: "); 520 method()->print_short_name(); 521 tty->cr(); 522 BREAKPOINT; 523 } 524 525 if( PrintOpto ) { 526 if (is_osr_compilation()) { 527 tty->print("[OSR]%3d", _compile_id); 528 } else { 529 tty->print("%3d", _compile_id); 530 } 531 } 532 #endif 533 } 534 535 536 //-----------------------init_scratch_buffer_blob------------------------------ 537 // Construct a temporary BufferBlob and cache it for this compile. 538 void Compile::init_scratch_buffer_blob(int const_size) { 539 // If there is already a scratch buffer blob allocated and the 540 // constant section is big enough, use it. Otherwise free the 541 // current and allocate a new one. 542 BufferBlob* blob = scratch_buffer_blob(); 543 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 544 // Use the current blob. 545 } else { 546 if (blob != NULL) { 547 BufferBlob::free(blob); 548 } 549 550 ResourceMark rm; 551 _scratch_const_size = const_size; 552 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 553 blob = BufferBlob::create("Compile::scratch_buffer", size); 554 // Record the buffer blob for next time. 555 set_scratch_buffer_blob(blob); 556 // Have we run out of code space? 557 if (scratch_buffer_blob() == NULL) { 558 // Let CompilerBroker disable further compilations. 559 record_failure("Not enough space for scratch buffer in CodeCache"); 560 return; 561 } 562 } 563 564 // Initialize the relocation buffers 565 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 566 set_scratch_locs_memory(locs_buf); 567 } 568 569 570 //-----------------------scratch_emit_size------------------------------------- 571 // Helper function that computes size by emitting code 572 uint Compile::scratch_emit_size(const Node* n) { 573 // Start scratch_emit_size section. 574 set_in_scratch_emit_size(true); 575 576 // Emit into a trash buffer and count bytes emitted. 577 // This is a pretty expensive way to compute a size, 578 // but it works well enough if seldom used. 579 // All common fixed-size instructions are given a size 580 // method by the AD file. 581 // Note that the scratch buffer blob and locs memory are 582 // allocated at the beginning of the compile task, and 583 // may be shared by several calls to scratch_emit_size. 584 // The allocation of the scratch buffer blob is particularly 585 // expensive, since it has to grab the code cache lock. 586 BufferBlob* blob = this->scratch_buffer_blob(); 587 assert(blob != NULL, "Initialize BufferBlob at start"); 588 assert(blob->size() > MAX_inst_size, "sanity"); 589 relocInfo* locs_buf = scratch_locs_memory(); 590 address blob_begin = blob->content_begin(); 591 address blob_end = (address)locs_buf; 592 assert(blob->contains(blob_end), "sanity"); 593 CodeBuffer buf(blob_begin, blob_end - blob_begin); 594 buf.initialize_consts_size(_scratch_const_size); 595 buf.initialize_stubs_size(MAX_stubs_size); 596 assert(locs_buf != NULL, "sanity"); 597 int lsize = MAX_locs_size / 3; 598 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 599 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 600 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 601 // Mark as scratch buffer. 602 buf.consts()->set_scratch_emit(); 603 buf.insts()->set_scratch_emit(); 604 buf.stubs()->set_scratch_emit(); 605 606 // Do the emission. 607 608 Label fakeL; // Fake label for branch instructions. 609 Label* saveL = NULL; 610 uint save_bnum = 0; 611 bool is_branch = n->is_MachBranch(); 612 if (is_branch) { 613 MacroAssembler masm(&buf); 614 masm.bind(fakeL); 615 n->as_MachBranch()->save_label(&saveL, &save_bnum); 616 n->as_MachBranch()->label_set(&fakeL, 0); 617 } else if (n->is_MachProlog()) { 618 MacroAssembler masm(&buf); 619 masm.bind(fakeL); 620 saveL = ((MachPrologNode*)n)->_verified_entry; 621 ((MachPrologNode*)n)->_verified_entry = &fakeL; 622 } 623 n->emit(buf, this->regalloc()); 624 625 // Emitting into the scratch buffer should not fail 626 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 627 628 // Restore label. 629 if (is_branch) { 630 n->as_MachBranch()->label_set(saveL, save_bnum); 631 } else if (n->is_MachProlog()) { 632 ((MachPrologNode*)n)->_verified_entry = saveL; 633 } 634 635 // End scratch_emit_size section. 636 set_in_scratch_emit_size(false); 637 638 return buf.insts_size(); 639 } 640 641 642 // ============================================================================ 643 //------------------------------Compile standard------------------------------- 644 debug_only( int Compile::_debug_idx = 100000; ) 645 646 // Compile a method. entry_bci is -1 for normal compilations and indicates 647 // the continuation bci for on stack replacement. 648 649 650 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 651 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 652 : Phase(Compiler), 653 _compile_id(ci_env->compile_id()), 654 _save_argument_registers(false), 655 _subsume_loads(subsume_loads), 656 _do_escape_analysis(do_escape_analysis), 657 _eliminate_boxing(eliminate_boxing), 658 _method(target), 659 _entry_bci(osr_bci), 660 _stub_function(NULL), 661 _stub_name(NULL), 662 _stub_entry_point(NULL), 663 _max_node_limit(MaxNodeLimit), 664 _orig_pc_slot(0), 665 _orig_pc_slot_offset_in_bytes(0), 666 _inlining_progress(false), 667 _inlining_incrementally(false), 668 _has_reserved_stack_access(target->has_reserved_stack_access()), 669 #ifndef PRODUCT 670 _trace_opto_output(directive->TraceOptoOutputOption), 671 #endif 672 _has_method_handle_invokes(false), 673 _comp_arena(mtCompiler), 674 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 675 _env(ci_env), 676 _directive(directive), 677 _log(ci_env->log()), 678 _failure_reason(NULL), 679 _congraph(NULL), 680 #ifndef PRODUCT 681 _printer(IdealGraphPrinter::printer()), 682 #endif 683 _dead_node_list(comp_arena()), 684 _dead_node_count(0), 685 _node_arena(mtCompiler), 686 _old_arena(mtCompiler), 687 _mach_constant_base_node(NULL), 688 _Compile_types(mtCompiler), 689 _initial_gvn(NULL), 690 _for_igvn(NULL), 691 _warm_calls(NULL), 692 _late_inlines(comp_arena(), 2, 0, NULL), 693 _string_late_inlines(comp_arena(), 2, 0, NULL), 694 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 695 _late_inlines_pos(0), 696 _number_of_mh_late_inlines(0), 697 _print_inlining_stream(NULL), 698 _print_inlining_list(NULL), 699 _print_inlining_idx(0), 700 _print_inlining_output(NULL), 701 _replay_inline_data(NULL), 702 _java_calls(0), 703 _inner_loops(0), 704 _interpreter_frame_size(0), 705 _node_bundling_limit(0), 706 _node_bundling_base(NULL), 707 _code_buffer("Compile::Fill_buffer"), 708 _scratch_const_size(-1), 709 _in_scratch_emit_size(false) 710 #ifndef PRODUCT 711 , _in_dump_cnt(0) 712 #endif 713 { 714 C = this; 715 #ifndef PRODUCT 716 if (_printer != NULL) { 717 _printer->set_compile(this); 718 } 719 #endif 720 CompileWrapper cw(this); 721 722 if (CITimeVerbose) { 723 tty->print(" "); 724 target->holder()->name()->print(); 725 tty->print("."); 726 target->print_short_name(); 727 tty->print(" "); 728 } 729 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 730 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 731 732 #ifndef PRODUCT 733 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 734 if (!print_opto_assembly) { 735 bool print_assembly = directive->PrintAssemblyOption; 736 if (print_assembly && !Disassembler::can_decode()) { 737 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 738 print_opto_assembly = true; 739 } 740 } 741 set_print_assembly(print_opto_assembly); 742 set_parsed_irreducible_loop(false); 743 744 if (directive->ReplayInlineOption) { 745 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 746 } 747 #endif 748 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 749 set_print_intrinsics(directive->PrintIntrinsicsOption); 750 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 751 752 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 753 // Make sure the method being compiled gets its own MDO, 754 // so we can at least track the decompile_count(). 755 // Need MDO to record RTM code generation state. 756 method()->ensure_method_data(); 757 } 758 759 Init(::AliasLevel); 760 761 762 print_compile_messages(); 763 764 _ilt = InlineTree::build_inline_tree_root(); 765 766 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 767 assert(num_alias_types() >= AliasIdxRaw, ""); 768 769 #define MINIMUM_NODE_HASH 1023 770 // Node list that Iterative GVN will start with 771 Unique_Node_List for_igvn(comp_arena()); 772 set_for_igvn(&for_igvn); 773 774 // GVN that will be run immediately on new nodes 775 uint estimated_size = method()->code_size()*4+64; 776 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 777 PhaseGVN gvn(node_arena(), estimated_size); 778 set_initial_gvn(&gvn); 779 780 print_inlining_init(); 781 { // Scope for timing the parser 782 TracePhase tp("parse", &timers[_t_parser]); 783 784 // Put top into the hash table ASAP. 785 initial_gvn()->transform_no_reclaim(top()); 786 787 // Set up tf(), start(), and find a CallGenerator. 788 CallGenerator* cg = NULL; 789 if (is_osr_compilation()) { 790 const TypeTuple *domain = StartOSRNode::osr_domain(); 791 const TypeTuple *range = TypeTuple::make_range(method()); 792 init_tf(TypeFunc::make(domain, range)); 793 StartNode* s = new StartOSRNode(root(), domain); 794 initial_gvn()->set_type_bottom(s); 795 init_start(s); 796 cg = CallGenerator::for_osr(method(), entry_bci()); 797 } else { 798 // Normal case. 799 init_tf(TypeFunc::make(method())); 800 StartNode* s = new StartNode(root(), tf()->domain_cc()); 801 initial_gvn()->set_type_bottom(s); 802 init_start(s); 803 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 804 // With java.lang.ref.reference.get() we must go through the 805 // intrinsic - even when get() is the root 806 // method of the compile - so that, if necessary, the value in 807 // the referent field of the reference object gets recorded by 808 // the pre-barrier code. 809 cg = find_intrinsic(method(), false); 810 } 811 if (cg == NULL) { 812 float past_uses = method()->interpreter_invocation_count(); 813 float expected_uses = past_uses; 814 cg = CallGenerator::for_inline(method(), expected_uses); 815 } 816 } 817 if (failing()) return; 818 if (cg == NULL) { 819 record_method_not_compilable("cannot parse method"); 820 return; 821 } 822 JVMState* jvms = build_start_state(start(), tf()); 823 if ((jvms = cg->generate(jvms)) == NULL) { 824 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 825 record_method_not_compilable("method parse failed"); 826 } 827 return; 828 } 829 GraphKit kit(jvms); 830 831 if (!kit.stopped()) { 832 // Accept return values, and transfer control we know not where. 833 // This is done by a special, unique ReturnNode bound to root. 834 return_values(kit.jvms()); 835 } 836 837 if (kit.has_exceptions()) { 838 // Any exceptions that escape from this call must be rethrown 839 // to whatever caller is dynamically above us on the stack. 840 // This is done by a special, unique RethrowNode bound to root. 841 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 842 } 843 844 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 845 846 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 847 inline_string_calls(true); 848 } 849 850 if (failing()) return; 851 852 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 853 854 // Remove clutter produced by parsing. 855 if (!failing()) { 856 ResourceMark rm; 857 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 858 } 859 } 860 861 // Note: Large methods are capped off in do_one_bytecode(). 862 if (failing()) return; 863 864 // After parsing, node notes are no longer automagic. 865 // They must be propagated by register_new_node_with_optimizer(), 866 // clone(), or the like. 867 set_default_node_notes(NULL); 868 869 for (;;) { 870 int successes = Inline_Warm(); 871 if (failing()) return; 872 if (successes == 0) break; 873 } 874 875 // Drain the list. 876 Finish_Warm(); 877 #ifndef PRODUCT 878 if (_printer && _printer->should_print(1)) { 879 _printer->print_inlining(); 880 } 881 #endif 882 883 if (failing()) return; 884 NOT_PRODUCT( verify_graph_edges(); ) 885 886 // Now optimize 887 Optimize(); 888 if (failing()) return; 889 NOT_PRODUCT( verify_graph_edges(); ) 890 891 #ifndef PRODUCT 892 if (PrintIdeal) { 893 ttyLocker ttyl; // keep the following output all in one block 894 // This output goes directly to the tty, not the compiler log. 895 // To enable tools to match it up with the compilation activity, 896 // be sure to tag this tty output with the compile ID. 897 if (xtty != NULL) { 898 xtty->head("ideal compile_id='%d'%s", compile_id(), 899 is_osr_compilation() ? " compile_kind='osr'" : 900 ""); 901 } 902 root()->dump(9999); 903 if (xtty != NULL) { 904 xtty->tail("ideal"); 905 } 906 } 907 #endif 908 909 NOT_PRODUCT( verify_barriers(); ) 910 911 // Dump compilation data to replay it. 912 if (directive->DumpReplayOption) { 913 env()->dump_replay_data(_compile_id); 914 } 915 if (directive->DumpInlineOption && (ilt() != NULL)) { 916 env()->dump_inline_data(_compile_id); 917 } 918 919 // Now that we know the size of all the monitors we can add a fixed slot 920 // for the original deopt pc. 921 922 _orig_pc_slot = fixed_slots(); 923 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 924 set_fixed_slots(next_slot); 925 926 // Compute when to use implicit null checks. Used by matching trap based 927 // nodes and NullCheck optimization. 928 set_allowed_deopt_reasons(); 929 930 // Now generate code 931 Code_Gen(); 932 if (failing()) return; 933 934 // Check if we want to skip execution of all compiled code. 935 { 936 #ifndef PRODUCT 937 if (OptoNoExecute) { 938 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 939 return; 940 } 941 #endif 942 TracePhase tp("install_code", &timers[_t_registerMethod]); 943 944 if (is_osr_compilation()) { 945 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 946 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 947 } else { 948 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 949 if (_code_offsets.value(CodeOffsets::Entry) == -1) { 950 // We emitted a value type entry point, adjust normal entry 951 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size); 952 } 953 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 954 } 955 956 env()->register_method(_method, _entry_bci, 957 &_code_offsets, 958 _orig_pc_slot_offset_in_bytes, 959 code_buffer(), 960 frame_size_in_words(), _oop_map_set, 961 &_handler_table, &_inc_table, 962 compiler, 963 has_unsafe_access(), 964 SharedRuntime::is_wide_vector(max_vector_size()), 965 rtm_state() 966 ); 967 968 if (log() != NULL) // Print code cache state into compiler log 969 log()->code_cache_state(); 970 } 971 } 972 973 //------------------------------Compile---------------------------------------- 974 // Compile a runtime stub 975 Compile::Compile( ciEnv* ci_env, 976 TypeFunc_generator generator, 977 address stub_function, 978 const char *stub_name, 979 int is_fancy_jump, 980 bool pass_tls, 981 bool save_arg_registers, 982 bool return_pc, 983 DirectiveSet* directive) 984 : Phase(Compiler), 985 _compile_id(0), 986 _save_argument_registers(save_arg_registers), 987 _subsume_loads(true), 988 _do_escape_analysis(false), 989 _eliminate_boxing(false), 990 _method(NULL), 991 _entry_bci(InvocationEntryBci), 992 _stub_function(stub_function), 993 _stub_name(stub_name), 994 _stub_entry_point(NULL), 995 _max_node_limit(MaxNodeLimit), 996 _orig_pc_slot(0), 997 _orig_pc_slot_offset_in_bytes(0), 998 _inlining_progress(false), 999 _inlining_incrementally(false), 1000 _has_reserved_stack_access(false), 1001 #ifndef PRODUCT 1002 _trace_opto_output(directive->TraceOptoOutputOption), 1003 #endif 1004 _has_method_handle_invokes(false), 1005 _comp_arena(mtCompiler), 1006 _env(ci_env), 1007 _directive(directive), 1008 _log(ci_env->log()), 1009 _failure_reason(NULL), 1010 _congraph(NULL), 1011 #ifndef PRODUCT 1012 _printer(NULL), 1013 #endif 1014 _dead_node_list(comp_arena()), 1015 _dead_node_count(0), 1016 _node_arena(mtCompiler), 1017 _old_arena(mtCompiler), 1018 _mach_constant_base_node(NULL), 1019 _Compile_types(mtCompiler), 1020 _initial_gvn(NULL), 1021 _for_igvn(NULL), 1022 _warm_calls(NULL), 1023 _number_of_mh_late_inlines(0), 1024 _print_inlining_stream(NULL), 1025 _print_inlining_list(NULL), 1026 _print_inlining_idx(0), 1027 _print_inlining_output(NULL), 1028 _replay_inline_data(NULL), 1029 _java_calls(0), 1030 _inner_loops(0), 1031 _interpreter_frame_size(0), 1032 _node_bundling_limit(0), 1033 _node_bundling_base(NULL), 1034 _code_buffer("Compile::Fill_buffer"), 1035 #ifndef PRODUCT 1036 _in_dump_cnt(0), 1037 #endif 1038 _allowed_reasons(0) { 1039 C = this; 1040 1041 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1042 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1043 1044 #ifndef PRODUCT 1045 set_print_assembly(PrintFrameConverterAssembly); 1046 set_parsed_irreducible_loop(false); 1047 #endif 1048 set_has_irreducible_loop(false); // no loops 1049 1050 CompileWrapper cw(this); 1051 Init(/*AliasLevel=*/ 0); 1052 init_tf((*generator)()); 1053 1054 { 1055 // The following is a dummy for the sake of GraphKit::gen_stub 1056 Unique_Node_List for_igvn(comp_arena()); 1057 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1058 PhaseGVN gvn(Thread::current()->resource_area(),255); 1059 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1060 gvn.transform_no_reclaim(top()); 1061 1062 GraphKit kit; 1063 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1064 } 1065 1066 NOT_PRODUCT( verify_graph_edges(); ) 1067 Code_Gen(); 1068 if (failing()) return; 1069 1070 1071 // Entry point will be accessed using compile->stub_entry_point(); 1072 if (code_buffer() == NULL) { 1073 Matcher::soft_match_failure(); 1074 } else { 1075 if (PrintAssembly && (WizardMode || Verbose)) 1076 tty->print_cr("### Stub::%s", stub_name); 1077 1078 if (!failing()) { 1079 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1080 1081 // Make the NMethod 1082 // For now we mark the frame as never safe for profile stackwalking 1083 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1084 code_buffer(), 1085 CodeOffsets::frame_never_safe, 1086 // _code_offsets.value(CodeOffsets::Frame_Complete), 1087 frame_size_in_words(), 1088 _oop_map_set, 1089 save_arg_registers); 1090 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1091 1092 _stub_entry_point = rs->entry_point(); 1093 } 1094 } 1095 } 1096 1097 //------------------------------Init------------------------------------------- 1098 // Prepare for a single compilation 1099 void Compile::Init(int aliaslevel) { 1100 _unique = 0; 1101 _regalloc = NULL; 1102 1103 _tf = NULL; // filled in later 1104 _top = NULL; // cached later 1105 _matcher = NULL; // filled in later 1106 _cfg = NULL; // filled in later 1107 1108 set_24_bit_selection_and_mode(Use24BitFP, false); 1109 1110 _node_note_array = NULL; 1111 _default_node_notes = NULL; 1112 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1113 1114 _immutable_memory = NULL; // filled in at first inquiry 1115 1116 // Globally visible Nodes 1117 // First set TOP to NULL to give safe behavior during creation of RootNode 1118 set_cached_top_node(NULL); 1119 set_root(new RootNode()); 1120 // Now that you have a Root to point to, create the real TOP 1121 set_cached_top_node( new ConNode(Type::TOP) ); 1122 set_recent_alloc(NULL, NULL); 1123 1124 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1125 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1126 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1127 env()->set_dependencies(new Dependencies(env())); 1128 1129 _fixed_slots = 0; 1130 set_has_split_ifs(false); 1131 set_has_loops(has_method() && method()->has_loops()); // first approximation 1132 set_has_stringbuilder(false); 1133 set_has_boxed_value(false); 1134 _trap_can_recompile = false; // no traps emitted yet 1135 _major_progress = true; // start out assuming good things will happen 1136 set_has_unsafe_access(false); 1137 set_max_vector_size(0); 1138 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1139 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1140 set_decompile_count(0); 1141 1142 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1143 set_num_loop_opts(LoopOptsCount); 1144 set_do_inlining(Inline); 1145 set_max_inline_size(MaxInlineSize); 1146 set_freq_inline_size(FreqInlineSize); 1147 set_do_scheduling(OptoScheduling); 1148 set_do_count_invocations(false); 1149 set_do_method_data_update(false); 1150 1151 set_do_vector_loop(false); 1152 1153 if (AllowVectorizeOnDemand) { 1154 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1155 set_do_vector_loop(true); 1156 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());}) 1157 } else if (has_method() && method()->name() != 0 && 1158 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1159 set_do_vector_loop(true); 1160 } 1161 } 1162 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1163 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());}) 1164 1165 set_age_code(has_method() && method()->profile_aging()); 1166 set_rtm_state(NoRTM); // No RTM lock eliding by default 1167 _max_node_limit = _directive->MaxNodeLimitOption; 1168 1169 #if INCLUDE_RTM_OPT 1170 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1171 int rtm_state = method()->method_data()->rtm_state(); 1172 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1173 // Don't generate RTM lock eliding code. 1174 set_rtm_state(NoRTM); 1175 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1176 // Generate RTM lock eliding code without abort ratio calculation code. 1177 set_rtm_state(UseRTM); 1178 } else if (UseRTMDeopt) { 1179 // Generate RTM lock eliding code and include abort ratio calculation 1180 // code if UseRTMDeopt is on. 1181 set_rtm_state(ProfileRTM); 1182 } 1183 } 1184 #endif 1185 if (debug_info()->recording_non_safepoints()) { 1186 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1187 (comp_arena(), 8, 0, NULL)); 1188 set_default_node_notes(Node_Notes::make(this)); 1189 } 1190 1191 // // -- Initialize types before each compile -- 1192 // // Update cached type information 1193 // if( _method && _method->constants() ) 1194 // Type::update_loaded_types(_method, _method->constants()); 1195 1196 // Init alias_type map. 1197 if (!_do_escape_analysis && aliaslevel == 3) 1198 aliaslevel = 2; // No unique types without escape analysis 1199 _AliasLevel = aliaslevel; 1200 const int grow_ats = 16; 1201 _max_alias_types = grow_ats; 1202 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1203 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1204 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1205 { 1206 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1207 } 1208 // Initialize the first few types. 1209 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1210 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1211 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1212 _num_alias_types = AliasIdxRaw+1; 1213 // Zero out the alias type cache. 1214 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1215 // A NULL adr_type hits in the cache right away. Preload the right answer. 1216 probe_alias_cache(NULL)->_index = AliasIdxTop; 1217 1218 _intrinsics = NULL; 1219 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1220 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1221 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1222 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1223 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1224 _value_type_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 1225 register_library_intrinsics(); 1226 } 1227 1228 //---------------------------init_start---------------------------------------- 1229 // Install the StartNode on this compile object. 1230 void Compile::init_start(StartNode* s) { 1231 if (failing()) 1232 return; // already failing 1233 assert(s == start(), ""); 1234 } 1235 1236 /** 1237 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1238 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1239 * the ideal graph. 1240 */ 1241 StartNode* Compile::start() const { 1242 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1243 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1244 Node* start = root()->fast_out(i); 1245 if (start->is_Start()) { 1246 return start->as_Start(); 1247 } 1248 } 1249 fatal("Did not find Start node!"); 1250 return NULL; 1251 } 1252 1253 //-------------------------------immutable_memory------------------------------------- 1254 // Access immutable memory 1255 Node* Compile::immutable_memory() { 1256 if (_immutable_memory != NULL) { 1257 return _immutable_memory; 1258 } 1259 StartNode* s = start(); 1260 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1261 Node *p = s->fast_out(i); 1262 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1263 _immutable_memory = p; 1264 return _immutable_memory; 1265 } 1266 } 1267 ShouldNotReachHere(); 1268 return NULL; 1269 } 1270 1271 //----------------------set_cached_top_node------------------------------------ 1272 // Install the cached top node, and make sure Node::is_top works correctly. 1273 void Compile::set_cached_top_node(Node* tn) { 1274 if (tn != NULL) verify_top(tn); 1275 Node* old_top = _top; 1276 _top = tn; 1277 // Calling Node::setup_is_top allows the nodes the chance to adjust 1278 // their _out arrays. 1279 if (_top != NULL) _top->setup_is_top(); 1280 if (old_top != NULL) old_top->setup_is_top(); 1281 assert(_top == NULL || top()->is_top(), ""); 1282 } 1283 1284 #ifdef ASSERT 1285 uint Compile::count_live_nodes_by_graph_walk() { 1286 Unique_Node_List useful(comp_arena()); 1287 // Get useful node list by walking the graph. 1288 identify_useful_nodes(useful); 1289 return useful.size(); 1290 } 1291 1292 void Compile::print_missing_nodes() { 1293 1294 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1295 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1296 return; 1297 } 1298 1299 // This is an expensive function. It is executed only when the user 1300 // specifies VerifyIdealNodeCount option or otherwise knows the 1301 // additional work that needs to be done to identify reachable nodes 1302 // by walking the flow graph and find the missing ones using 1303 // _dead_node_list. 1304 1305 Unique_Node_List useful(comp_arena()); 1306 // Get useful node list by walking the graph. 1307 identify_useful_nodes(useful); 1308 1309 uint l_nodes = C->live_nodes(); 1310 uint l_nodes_by_walk = useful.size(); 1311 1312 if (l_nodes != l_nodes_by_walk) { 1313 if (_log != NULL) { 1314 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1315 _log->stamp(); 1316 _log->end_head(); 1317 } 1318 VectorSet& useful_member_set = useful.member_set(); 1319 int last_idx = l_nodes_by_walk; 1320 for (int i = 0; i < last_idx; i++) { 1321 if (useful_member_set.test(i)) { 1322 if (_dead_node_list.test(i)) { 1323 if (_log != NULL) { 1324 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1325 } 1326 if (PrintIdealNodeCount) { 1327 // Print the log message to tty 1328 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1329 useful.at(i)->dump(); 1330 } 1331 } 1332 } 1333 else if (! _dead_node_list.test(i)) { 1334 if (_log != NULL) { 1335 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1336 } 1337 if (PrintIdealNodeCount) { 1338 // Print the log message to tty 1339 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1340 } 1341 } 1342 } 1343 if (_log != NULL) { 1344 _log->tail("mismatched_nodes"); 1345 } 1346 } 1347 } 1348 void Compile::record_modified_node(Node* n) { 1349 if (_modified_nodes != NULL && !_inlining_incrementally && 1350 n->outcnt() != 0 && !n->is_Con()) { 1351 _modified_nodes->push(n); 1352 } 1353 } 1354 1355 void Compile::remove_modified_node(Node* n) { 1356 if (_modified_nodes != NULL) { 1357 _modified_nodes->remove(n); 1358 } 1359 } 1360 #endif 1361 1362 #ifndef PRODUCT 1363 void Compile::verify_top(Node* tn) const { 1364 if (tn != NULL) { 1365 assert(tn->is_Con(), "top node must be a constant"); 1366 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1367 assert(tn->in(0) != NULL, "must have live top node"); 1368 } 1369 } 1370 #endif 1371 1372 1373 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1374 1375 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1376 guarantee(arr != NULL, ""); 1377 int num_blocks = arr->length(); 1378 if (grow_by < num_blocks) grow_by = num_blocks; 1379 int num_notes = grow_by * _node_notes_block_size; 1380 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1381 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1382 while (num_notes > 0) { 1383 arr->append(notes); 1384 notes += _node_notes_block_size; 1385 num_notes -= _node_notes_block_size; 1386 } 1387 assert(num_notes == 0, "exact multiple, please"); 1388 } 1389 1390 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1391 if (source == NULL || dest == NULL) return false; 1392 1393 if (dest->is_Con()) 1394 return false; // Do not push debug info onto constants. 1395 1396 #ifdef ASSERT 1397 // Leave a bread crumb trail pointing to the original node: 1398 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1399 dest->set_debug_orig(source); 1400 } 1401 #endif 1402 1403 if (node_note_array() == NULL) 1404 return false; // Not collecting any notes now. 1405 1406 // This is a copy onto a pre-existing node, which may already have notes. 1407 // If both nodes have notes, do not overwrite any pre-existing notes. 1408 Node_Notes* source_notes = node_notes_at(source->_idx); 1409 if (source_notes == NULL || source_notes->is_clear()) return false; 1410 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1411 if (dest_notes == NULL || dest_notes->is_clear()) { 1412 return set_node_notes_at(dest->_idx, source_notes); 1413 } 1414 1415 Node_Notes merged_notes = (*source_notes); 1416 // The order of operations here ensures that dest notes will win... 1417 merged_notes.update_from(dest_notes); 1418 return set_node_notes_at(dest->_idx, &merged_notes); 1419 } 1420 1421 1422 //--------------------------allow_range_check_smearing------------------------- 1423 // Gating condition for coalescing similar range checks. 1424 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1425 // single covering check that is at least as strong as any of them. 1426 // If the optimization succeeds, the simplified (strengthened) range check 1427 // will always succeed. If it fails, we will deopt, and then give up 1428 // on the optimization. 1429 bool Compile::allow_range_check_smearing() const { 1430 // If this method has already thrown a range-check, 1431 // assume it was because we already tried range smearing 1432 // and it failed. 1433 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1434 return !already_trapped; 1435 } 1436 1437 1438 //------------------------------flatten_alias_type----------------------------- 1439 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1440 int offset = tj->offset(); 1441 TypePtr::PTR ptr = tj->ptr(); 1442 1443 // Known instance (scalarizable allocation) alias only with itself. 1444 bool is_known_inst = tj->isa_oopptr() != NULL && 1445 tj->is_oopptr()->is_known_instance(); 1446 1447 // Process weird unsafe references. 1448 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1449 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1450 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1451 tj = TypeOopPtr::BOTTOM; 1452 ptr = tj->ptr(); 1453 offset = tj->offset(); 1454 } 1455 1456 // Array pointers need some flattening 1457 const TypeAryPtr *ta = tj->isa_aryptr(); 1458 if (ta && ta->is_stable()) { 1459 // Erase stability property for alias analysis. 1460 tj = ta = ta->cast_to_stable(false); 1461 } 1462 if( ta && is_known_inst ) { 1463 if ( offset != Type::OffsetBot && 1464 offset > arrayOopDesc::length_offset_in_bytes() ) { 1465 offset = Type::OffsetBot; // Flatten constant access into array body only 1466 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id()); 1467 } 1468 } else if( ta && _AliasLevel >= 2 ) { 1469 // For arrays indexed by constant indices, we flatten the alias 1470 // space to include all of the array body. Only the header, klass 1471 // and array length can be accessed un-aliased. 1472 // For flattened value type array, each field has its own slice so 1473 // we must include the field offset. 1474 if( offset != Type::OffsetBot ) { 1475 if( ta->const_oop() ) { // MethodData* or Method* 1476 offset = Type::OffsetBot; // Flatten constant access into array body 1477 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1478 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1479 // range is OK as-is. 1480 tj = ta = TypeAryPtr::RANGE; 1481 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1482 tj = TypeInstPtr::KLASS; // all klass loads look alike 1483 ta = TypeAryPtr::RANGE; // generic ignored junk 1484 ptr = TypePtr::BotPTR; 1485 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1486 tj = TypeInstPtr::MARK; 1487 ta = TypeAryPtr::RANGE; // generic ignored junk 1488 ptr = TypePtr::BotPTR; 1489 } else { // Random constant offset into array body 1490 offset = Type::OffsetBot; // Flatten constant access into array body 1491 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1492 } 1493 } 1494 // Arrays of fixed size alias with arrays of unknown size. 1495 if (ta->size() != TypeInt::POS) { 1496 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1497 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1498 } 1499 // Arrays of known objects become arrays of unknown objects. 1500 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1501 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1502 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1503 } 1504 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1505 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1506 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1507 } 1508 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1509 // cannot be distinguished by bytecode alone. 1510 if (ta->elem() == TypeInt::BOOL) { 1511 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1512 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1513 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset()); 1514 } 1515 // During the 2nd round of IterGVN, NotNull castings are removed. 1516 // Make sure the Bottom and NotNull variants alias the same. 1517 // Also, make sure exact and non-exact variants alias the same. 1518 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1519 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1520 } 1521 } 1522 1523 // Oop pointers need some flattening 1524 const TypeInstPtr *to = tj->isa_instptr(); 1525 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1526 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1527 if( ptr == TypePtr::Constant ) { 1528 if (to->klass() != ciEnv::current()->Class_klass() || 1529 offset < k->size_helper() * wordSize) { 1530 // No constant oop pointers (such as Strings); they alias with 1531 // unknown strings. 1532 assert(!is_known_inst, "not scalarizable allocation"); 1533 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1534 } 1535 } else if( is_known_inst ) { 1536 tj = to; // Keep NotNull and klass_is_exact for instance type 1537 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1538 // During the 2nd round of IterGVN, NotNull castings are removed. 1539 // Make sure the Bottom and NotNull variants alias the same. 1540 // Also, make sure exact and non-exact variants alias the same. 1541 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1542 } 1543 if (to->speculative() != NULL) { 1544 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->instance_id()); 1545 } 1546 // Canonicalize the holder of this field 1547 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1548 // First handle header references such as a LoadKlassNode, even if the 1549 // object's klass is unloaded at compile time (4965979). 1550 if (!is_known_inst) { // Do it only for non-instance types 1551 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset)); 1552 } 1553 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1554 // Static fields are in the space above the normal instance 1555 // fields in the java.lang.Class instance. 1556 if (to->klass() != ciEnv::current()->Class_klass()) { 1557 to = NULL; 1558 tj = TypeOopPtr::BOTTOM; 1559 offset = tj->offset(); 1560 } 1561 } else { 1562 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1563 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1564 if( is_known_inst ) { 1565 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), to->instance_id()); 1566 } else { 1567 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset)); 1568 } 1569 } 1570 } 1571 } 1572 1573 // Klass pointers to object array klasses need some flattening 1574 const TypeKlassPtr *tk = tj->isa_klassptr(); 1575 if( tk ) { 1576 // If we are referencing a field within a Klass, we need 1577 // to assume the worst case of an Object. Both exact and 1578 // inexact types must flatten to the same alias class so 1579 // use NotNull as the PTR. 1580 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1581 1582 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1583 TypeKlassPtr::OBJECT->klass(), 1584 Type::Offset(offset)); 1585 } 1586 1587 ciKlass* klass = tk->klass(); 1588 if (klass != NULL && klass->is_obj_array_klass()) { 1589 ciKlass* k = TypeAryPtr::OOPS->klass(); 1590 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1591 k = TypeInstPtr::BOTTOM->klass(); 1592 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset)); 1593 } 1594 1595 // Check for precise loads from the primary supertype array and force them 1596 // to the supertype cache alias index. Check for generic array loads from 1597 // the primary supertype array and also force them to the supertype cache 1598 // alias index. Since the same load can reach both, we need to merge 1599 // these 2 disparate memories into the same alias class. Since the 1600 // primary supertype array is read-only, there's no chance of confusion 1601 // where we bypass an array load and an array store. 1602 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1603 if (offset == Type::OffsetBot || 1604 (offset >= primary_supers_offset && 1605 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1606 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1607 offset = in_bytes(Klass::secondary_super_cache_offset()); 1608 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset)); 1609 } 1610 } 1611 1612 // Flatten all Raw pointers together. 1613 if (tj->base() == Type::RawPtr) 1614 tj = TypeRawPtr::BOTTOM; 1615 1616 if (tj->base() == Type::AnyPtr) 1617 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1618 1619 // Flatten all to bottom for now 1620 switch( _AliasLevel ) { 1621 case 0: 1622 tj = TypePtr::BOTTOM; 1623 break; 1624 case 1: // Flatten to: oop, static, field or array 1625 switch (tj->base()) { 1626 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1627 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1628 case Type::AryPtr: // do not distinguish arrays at all 1629 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1630 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1631 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1632 default: ShouldNotReachHere(); 1633 } 1634 break; 1635 case 2: // No collapsing at level 2; keep all splits 1636 case 3: // No collapsing at level 3; keep all splits 1637 break; 1638 default: 1639 Unimplemented(); 1640 } 1641 1642 offset = tj->offset(); 1643 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1644 1645 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1646 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1647 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1648 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1649 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1650 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1651 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1652 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1653 assert( tj->ptr() != TypePtr::TopPTR && 1654 tj->ptr() != TypePtr::AnyNull && 1655 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1656 // assert( tj->ptr() != TypePtr::Constant || 1657 // tj->base() == Type::RawPtr || 1658 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1659 1660 return tj; 1661 } 1662 1663 void Compile::AliasType::Init(int i, const TypePtr* at) { 1664 _index = i; 1665 _adr_type = at; 1666 _field = NULL; 1667 _element = NULL; 1668 _is_rewritable = true; // default 1669 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1670 if (atoop != NULL && atoop->is_known_instance()) { 1671 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1672 _general_index = Compile::current()->get_alias_index(gt); 1673 } else { 1674 _general_index = 0; 1675 } 1676 } 1677 1678 BasicType Compile::AliasType::basic_type() const { 1679 if (element() != NULL) { 1680 const Type* element = adr_type()->is_aryptr()->elem(); 1681 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1682 } if (field() != NULL) { 1683 return field()->layout_type(); 1684 } else { 1685 return T_ILLEGAL; // unknown 1686 } 1687 } 1688 1689 //---------------------------------print_on------------------------------------ 1690 #ifndef PRODUCT 1691 void Compile::AliasType::print_on(outputStream* st) { 1692 if (index() < 10) 1693 st->print("@ <%d> ", index()); 1694 else st->print("@ <%d>", index()); 1695 st->print(is_rewritable() ? " " : " RO"); 1696 int offset = adr_type()->offset(); 1697 if (offset == Type::OffsetBot) 1698 st->print(" +any"); 1699 else st->print(" +%-3d", offset); 1700 st->print(" in "); 1701 adr_type()->dump_on(st); 1702 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1703 if (field() != NULL && tjp) { 1704 if (tjp->klass() != field()->holder() || 1705 tjp->offset() != field()->offset_in_bytes()) { 1706 st->print(" != "); 1707 field()->print(); 1708 st->print(" ***"); 1709 } 1710 } 1711 } 1712 1713 void print_alias_types() { 1714 Compile* C = Compile::current(); 1715 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1716 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1717 C->alias_type(idx)->print_on(tty); 1718 tty->cr(); 1719 } 1720 } 1721 #endif 1722 1723 1724 //----------------------------probe_alias_cache-------------------------------- 1725 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1726 intptr_t key = (intptr_t) adr_type; 1727 key ^= key >> logAliasCacheSize; 1728 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1729 } 1730 1731 1732 //-----------------------------grow_alias_types-------------------------------- 1733 void Compile::grow_alias_types() { 1734 const int old_ats = _max_alias_types; // how many before? 1735 const int new_ats = old_ats; // how many more? 1736 const int grow_ats = old_ats+new_ats; // how many now? 1737 _max_alias_types = grow_ats; 1738 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1739 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1740 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1741 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1742 } 1743 1744 1745 //--------------------------------find_alias_type------------------------------ 1746 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1747 if (_AliasLevel == 0) 1748 return alias_type(AliasIdxBot); 1749 1750 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1751 if (ace->_adr_type == adr_type) { 1752 return alias_type(ace->_index); 1753 } 1754 1755 // Handle special cases. 1756 if (adr_type == NULL) return alias_type(AliasIdxTop); 1757 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1758 1759 // Do it the slow way. 1760 const TypePtr* flat = flatten_alias_type(adr_type); 1761 1762 #ifdef ASSERT 1763 { 1764 ResourceMark rm; 1765 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1766 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1767 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1768 Type::str(adr_type)); 1769 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1770 const TypeOopPtr* foop = flat->is_oopptr(); 1771 // Scalarizable allocations have exact klass always. 1772 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1773 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1774 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1775 Type::str(foop), Type::str(xoop)); 1776 } 1777 } 1778 #endif 1779 1780 int idx = AliasIdxTop; 1781 for (int i = 0; i < num_alias_types(); i++) { 1782 if (alias_type(i)->adr_type() == flat) { 1783 idx = i; 1784 break; 1785 } 1786 } 1787 1788 if (idx == AliasIdxTop) { 1789 if (no_create) return NULL; 1790 // Grow the array if necessary. 1791 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1792 // Add a new alias type. 1793 idx = _num_alias_types++; 1794 _alias_types[idx]->Init(idx, flat); 1795 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1796 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1797 if (flat->isa_instptr()) { 1798 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1799 && flat->is_instptr()->klass() == env()->Class_klass()) 1800 alias_type(idx)->set_rewritable(false); 1801 } 1802 ciField* field = NULL; 1803 if (flat->isa_aryptr()) { 1804 #ifdef ASSERT 1805 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1806 // (T_BYTE has the weakest alignment and size restrictions...) 1807 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1808 #endif 1809 const Type* elemtype = flat->is_aryptr()->elem(); 1810 if (flat->offset() == TypePtr::OffsetBot) { 1811 alias_type(idx)->set_element(elemtype); 1812 } 1813 int field_offset = flat->is_aryptr()->field_offset().get(); 1814 if (elemtype->isa_valuetype() && field_offset != Type::OffsetBot) { 1815 ciValueKlass* vk = elemtype->is_valuetype()->value_klass(); 1816 field_offset += vk->first_field_offset(); 1817 field = vk->get_field_by_offset(field_offset, false); 1818 } 1819 } 1820 if (flat->isa_klassptr()) { 1821 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1822 alias_type(idx)->set_rewritable(false); 1823 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1824 alias_type(idx)->set_rewritable(false); 1825 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1826 alias_type(idx)->set_rewritable(false); 1827 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1828 alias_type(idx)->set_rewritable(false); 1829 } 1830 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1831 // but the base pointer type is not distinctive enough to identify 1832 // references into JavaThread.) 1833 1834 // Check for final fields. 1835 const TypeInstPtr* tinst = flat->isa_instptr(); 1836 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1837 if (tinst->const_oop() != NULL && 1838 tinst->klass() == ciEnv::current()->Class_klass() && 1839 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1840 // static field 1841 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1842 field = k->get_field_by_offset(tinst->offset(), true); 1843 } else if (tinst->klass()->is_valuetype()) { 1844 // Value type field 1845 ciValueKlass* vk = tinst->value_klass(); 1846 field = vk->get_field_by_offset(tinst->offset(), false); 1847 } else { 1848 ciInstanceKlass* k = tinst->klass()->as_instance_klass(); 1849 field = k->get_field_by_offset(tinst->offset(), false); 1850 } 1851 } 1852 assert(field == NULL || 1853 original_field == NULL || 1854 (field->holder() == original_field->holder() && 1855 field->offset() == original_field->offset() && 1856 field->is_static() == original_field->is_static()), "wrong field?"); 1857 // Set field() and is_rewritable() attributes. 1858 if (field != NULL) alias_type(idx)->set_field(field); 1859 } 1860 1861 // Fill the cache for next time. 1862 ace->_adr_type = adr_type; 1863 ace->_index = idx; 1864 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1865 1866 // Might as well try to fill the cache for the flattened version, too. 1867 AliasCacheEntry* face = probe_alias_cache(flat); 1868 if (face->_adr_type == NULL) { 1869 face->_adr_type = flat; 1870 face->_index = idx; 1871 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1872 } 1873 1874 return alias_type(idx); 1875 } 1876 1877 1878 Compile::AliasType* Compile::alias_type(ciField* field) { 1879 const TypeOopPtr* t; 1880 if (field->is_static()) 1881 t = TypeInstPtr::make(field->holder()->java_mirror()); 1882 else 1883 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1884 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1885 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1886 return atp; 1887 } 1888 1889 1890 //------------------------------have_alias_type-------------------------------- 1891 bool Compile::have_alias_type(const TypePtr* adr_type) { 1892 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1893 if (ace->_adr_type == adr_type) { 1894 return true; 1895 } 1896 1897 // Handle special cases. 1898 if (adr_type == NULL) return true; 1899 if (adr_type == TypePtr::BOTTOM) return true; 1900 1901 return find_alias_type(adr_type, true, NULL) != NULL; 1902 } 1903 1904 //-----------------------------must_alias-------------------------------------- 1905 // True if all values of the given address type are in the given alias category. 1906 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1907 if (alias_idx == AliasIdxBot) return true; // the universal category 1908 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1909 if (alias_idx == AliasIdxTop) return false; // the empty category 1910 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1911 1912 // the only remaining possible overlap is identity 1913 int adr_idx = get_alias_index(adr_type); 1914 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1915 assert(adr_idx == alias_idx || 1916 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1917 && adr_type != TypeOopPtr::BOTTOM), 1918 "should not be testing for overlap with an unsafe pointer"); 1919 return adr_idx == alias_idx; 1920 } 1921 1922 //------------------------------can_alias-------------------------------------- 1923 // True if any values of the given address type are in the given alias category. 1924 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1925 if (alias_idx == AliasIdxTop) return false; // the empty category 1926 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1927 if (alias_idx == AliasIdxBot) return true; // the universal category 1928 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1929 1930 // the only remaining possible overlap is identity 1931 int adr_idx = get_alias_index(adr_type); 1932 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1933 return adr_idx == alias_idx; 1934 } 1935 1936 1937 1938 //---------------------------pop_warm_call------------------------------------- 1939 WarmCallInfo* Compile::pop_warm_call() { 1940 WarmCallInfo* wci = _warm_calls; 1941 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1942 return wci; 1943 } 1944 1945 //----------------------------Inline_Warm-------------------------------------- 1946 int Compile::Inline_Warm() { 1947 // If there is room, try to inline some more warm call sites. 1948 // %%% Do a graph index compaction pass when we think we're out of space? 1949 if (!InlineWarmCalls) return 0; 1950 1951 int calls_made_hot = 0; 1952 int room_to_grow = NodeCountInliningCutoff - unique(); 1953 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1954 int amount_grown = 0; 1955 WarmCallInfo* call; 1956 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1957 int est_size = (int)call->size(); 1958 if (est_size > (room_to_grow - amount_grown)) { 1959 // This one won't fit anyway. Get rid of it. 1960 call->make_cold(); 1961 continue; 1962 } 1963 call->make_hot(); 1964 calls_made_hot++; 1965 amount_grown += est_size; 1966 amount_to_grow -= est_size; 1967 } 1968 1969 if (calls_made_hot > 0) set_major_progress(); 1970 return calls_made_hot; 1971 } 1972 1973 1974 //----------------------------Finish_Warm-------------------------------------- 1975 void Compile::Finish_Warm() { 1976 if (!InlineWarmCalls) return; 1977 if (failing()) return; 1978 if (warm_calls() == NULL) return; 1979 1980 // Clean up loose ends, if we are out of space for inlining. 1981 WarmCallInfo* call; 1982 while ((call = pop_warm_call()) != NULL) { 1983 call->make_cold(); 1984 } 1985 } 1986 1987 //---------------------cleanup_loop_predicates----------------------- 1988 // Remove the opaque nodes that protect the predicates so that all unused 1989 // checks and uncommon_traps will be eliminated from the ideal graph 1990 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1991 if (predicate_count()==0) return; 1992 for (int i = predicate_count(); i > 0; i--) { 1993 Node * n = predicate_opaque1_node(i-1); 1994 assert(n->Opcode() == Op_Opaque1, "must be"); 1995 igvn.replace_node(n, n->in(1)); 1996 } 1997 assert(predicate_count()==0, "should be clean!"); 1998 } 1999 2000 void Compile::add_range_check_cast(Node* n) { 2001 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 2002 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 2003 _range_check_casts->append(n); 2004 } 2005 2006 // Remove all range check dependent CastIINodes. 2007 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 2008 for (int i = range_check_cast_count(); i > 0; i--) { 2009 Node* cast = range_check_cast_node(i-1); 2010 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 2011 igvn.replace_node(cast, cast->in(1)); 2012 } 2013 assert(range_check_cast_count() == 0, "should be empty"); 2014 } 2015 2016 void Compile::add_opaque4_node(Node* n) { 2017 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 2018 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 2019 _opaque4_nodes->append(n); 2020 } 2021 2022 // Remove all Opaque4 nodes. 2023 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 2024 for (int i = opaque4_count(); i > 0; i--) { 2025 Node* opaq = opaque4_node(i-1); 2026 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 2027 igvn.replace_node(opaq, opaq->in(2)); 2028 } 2029 assert(opaque4_count() == 0, "should be empty"); 2030 } 2031 2032 void Compile::add_value_type(Node* n) { 2033 assert(n->is_ValueTypeBase(), "unexpected node"); 2034 if (_value_type_nodes != NULL) { 2035 _value_type_nodes->push(n); 2036 } 2037 } 2038 2039 void Compile::remove_value_type(Node* n) { 2040 assert(n->is_ValueTypeBase(), "unexpected node"); 2041 if (_value_type_nodes != NULL) { 2042 _value_type_nodes->remove(n); 2043 } 2044 } 2045 2046 void Compile::process_value_types(PhaseIterGVN &igvn) { 2047 // Make value types scalar in safepoints 2048 while (_value_type_nodes->size() != 0) { 2049 ValueTypeBaseNode* vt = _value_type_nodes->pop()->as_ValueTypeBase(); 2050 vt->make_scalar_in_safepoints(igvn.C->root(), &igvn); 2051 if (vt->is_ValueTypePtr()) { 2052 igvn.replace_node(vt, vt->get_oop()); 2053 } 2054 } 2055 _value_type_nodes = NULL; 2056 igvn.optimize(); 2057 } 2058 2059 // StringOpts and late inlining of string methods 2060 void Compile::inline_string_calls(bool parse_time) { 2061 { 2062 // remove useless nodes to make the usage analysis simpler 2063 ResourceMark rm; 2064 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2065 } 2066 2067 { 2068 ResourceMark rm; 2069 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2070 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2071 print_method(PHASE_AFTER_STRINGOPTS, 3); 2072 } 2073 2074 // now inline anything that we skipped the first time around 2075 if (!parse_time) { 2076 _late_inlines_pos = _late_inlines.length(); 2077 } 2078 2079 while (_string_late_inlines.length() > 0) { 2080 CallGenerator* cg = _string_late_inlines.pop(); 2081 cg->do_late_inline(); 2082 if (failing()) return; 2083 } 2084 _string_late_inlines.trunc_to(0); 2085 } 2086 2087 // Late inlining of boxing methods 2088 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2089 if (_boxing_late_inlines.length() > 0) { 2090 assert(has_boxed_value(), "inconsistent"); 2091 2092 PhaseGVN* gvn = initial_gvn(); 2093 set_inlining_incrementally(true); 2094 2095 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2096 for_igvn()->clear(); 2097 gvn->replace_with(&igvn); 2098 2099 _late_inlines_pos = _late_inlines.length(); 2100 2101 while (_boxing_late_inlines.length() > 0) { 2102 CallGenerator* cg = _boxing_late_inlines.pop(); 2103 cg->do_late_inline(); 2104 if (failing()) return; 2105 } 2106 _boxing_late_inlines.trunc_to(0); 2107 2108 { 2109 ResourceMark rm; 2110 PhaseRemoveUseless pru(gvn, for_igvn()); 2111 } 2112 2113 igvn = PhaseIterGVN(gvn); 2114 igvn.optimize(); 2115 2116 set_inlining_progress(false); 2117 set_inlining_incrementally(false); 2118 } 2119 } 2120 2121 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 2122 assert(IncrementalInline, "incremental inlining should be on"); 2123 PhaseGVN* gvn = initial_gvn(); 2124 2125 set_inlining_progress(false); 2126 for_igvn()->clear(); 2127 gvn->replace_with(&igvn); 2128 2129 { 2130 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2131 int i = 0; 2132 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2133 CallGenerator* cg = _late_inlines.at(i); 2134 _late_inlines_pos = i+1; 2135 cg->do_late_inline(); 2136 if (failing()) return; 2137 } 2138 int j = 0; 2139 for (; i < _late_inlines.length(); i++, j++) { 2140 _late_inlines.at_put(j, _late_inlines.at(i)); 2141 } 2142 _late_inlines.trunc_to(j); 2143 } 2144 2145 { 2146 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2147 ResourceMark rm; 2148 PhaseRemoveUseless pru(gvn, for_igvn()); 2149 } 2150 2151 { 2152 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2153 igvn = PhaseIterGVN(gvn); 2154 } 2155 } 2156 2157 // Perform incremental inlining until bound on number of live nodes is reached 2158 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2159 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2160 2161 PhaseGVN* gvn = initial_gvn(); 2162 2163 set_inlining_incrementally(true); 2164 set_inlining_progress(true); 2165 uint low_live_nodes = 0; 2166 2167 while(inlining_progress() && _late_inlines.length() > 0) { 2168 2169 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2170 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2171 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2172 // PhaseIdealLoop is expensive so we only try it once we are 2173 // out of live nodes and we only try it again if the previous 2174 // helped got the number of nodes down significantly 2175 PhaseIdealLoop ideal_loop( igvn, false, true ); 2176 if (failing()) return; 2177 low_live_nodes = live_nodes(); 2178 _major_progress = true; 2179 } 2180 2181 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2182 break; 2183 } 2184 } 2185 2186 inline_incrementally_one(igvn); 2187 2188 if (failing()) return; 2189 2190 { 2191 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2192 igvn.optimize(); 2193 } 2194 2195 if (failing()) return; 2196 } 2197 2198 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2199 2200 if (_string_late_inlines.length() > 0) { 2201 assert(has_stringbuilder(), "inconsistent"); 2202 for_igvn()->clear(); 2203 initial_gvn()->replace_with(&igvn); 2204 2205 inline_string_calls(false); 2206 2207 if (failing()) return; 2208 2209 { 2210 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2211 ResourceMark rm; 2212 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2213 } 2214 2215 { 2216 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2217 igvn = PhaseIterGVN(gvn); 2218 igvn.optimize(); 2219 } 2220 } 2221 2222 set_inlining_incrementally(false); 2223 } 2224 2225 2226 //------------------------------Optimize--------------------------------------- 2227 // Given a graph, optimize it. 2228 void Compile::Optimize() { 2229 TracePhase tp("optimizer", &timers[_t_optimizer]); 2230 2231 #ifndef PRODUCT 2232 if (_directive->BreakAtCompileOption) { 2233 BREAKPOINT; 2234 } 2235 2236 #endif 2237 2238 #ifdef ASSERT 2239 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2240 bs->verify_gc_barriers(true); 2241 #endif 2242 2243 ResourceMark rm; 2244 int loop_opts_cnt; 2245 2246 print_inlining_reinit(); 2247 2248 NOT_PRODUCT( verify_graph_edges(); ) 2249 2250 print_method(PHASE_AFTER_PARSING); 2251 2252 { 2253 // Iterative Global Value Numbering, including ideal transforms 2254 // Initialize IterGVN with types and values from parse-time GVN 2255 PhaseIterGVN igvn(initial_gvn()); 2256 #ifdef ASSERT 2257 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2258 #endif 2259 { 2260 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2261 igvn.optimize(); 2262 } 2263 2264 print_method(PHASE_ITER_GVN1, 2); 2265 2266 if (failing()) return; 2267 2268 inline_incrementally(igvn); 2269 2270 print_method(PHASE_INCREMENTAL_INLINE, 2); 2271 2272 if (failing()) return; 2273 2274 if (eliminate_boxing()) { 2275 // Inline valueOf() methods now. 2276 inline_boxing_calls(igvn); 2277 2278 if (AlwaysIncrementalInline) { 2279 inline_incrementally(igvn); 2280 } 2281 2282 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2283 2284 if (failing()) return; 2285 } 2286 2287 // Remove the speculative part of types and clean up the graph from 2288 // the extra CastPP nodes whose only purpose is to carry them. Do 2289 // that early so that optimizations are not disrupted by the extra 2290 // CastPP nodes. 2291 remove_speculative_types(igvn); 2292 2293 // No more new expensive nodes will be added to the list from here 2294 // so keep only the actual candidates for optimizations. 2295 cleanup_expensive_nodes(igvn); 2296 2297 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2298 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2299 initial_gvn()->replace_with(&igvn); 2300 for_igvn()->clear(); 2301 Unique_Node_List new_worklist(C->comp_arena()); 2302 { 2303 ResourceMark rm; 2304 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2305 } 2306 set_for_igvn(&new_worklist); 2307 igvn = PhaseIterGVN(initial_gvn()); 2308 igvn.optimize(); 2309 } 2310 2311 if (_value_type_nodes->size() > 0) { 2312 // Do this once all inlining is over to avoid getting inconsistent debug info 2313 process_value_types(igvn); 2314 } 2315 2316 // Perform escape analysis 2317 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2318 if (has_loops()) { 2319 // Cleanup graph (remove dead nodes). 2320 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2321 PhaseIdealLoop ideal_loop( igvn, false, true ); 2322 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2323 if (failing()) return; 2324 } 2325 ConnectionGraph::do_analysis(this, &igvn); 2326 2327 if (failing()) return; 2328 2329 // Optimize out fields loads from scalar replaceable allocations. 2330 igvn.optimize(); 2331 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2332 2333 if (failing()) return; 2334 2335 if (congraph() != NULL && macro_count() > 0) { 2336 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2337 PhaseMacroExpand mexp(igvn); 2338 mexp.eliminate_macro_nodes(); 2339 igvn.set_delay_transform(false); 2340 2341 igvn.optimize(); 2342 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2343 2344 if (failing()) return; 2345 } 2346 } 2347 2348 // Loop transforms on the ideal graph. Range Check Elimination, 2349 // peeling, unrolling, etc. 2350 2351 // Set loop opts counter 2352 loop_opts_cnt = num_loop_opts(); 2353 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2354 { 2355 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2356 PhaseIdealLoop ideal_loop( igvn, true ); 2357 loop_opts_cnt--; 2358 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2359 if (failing()) return; 2360 } 2361 // Loop opts pass if partial peeling occurred in previous pass 2362 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2363 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2364 PhaseIdealLoop ideal_loop( igvn, false ); 2365 loop_opts_cnt--; 2366 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2367 if (failing()) return; 2368 } 2369 // Loop opts pass for loop-unrolling before CCP 2370 if(major_progress() && (loop_opts_cnt > 0)) { 2371 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2372 PhaseIdealLoop ideal_loop( igvn, false ); 2373 loop_opts_cnt--; 2374 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2375 } 2376 if (!failing()) { 2377 // Verify that last round of loop opts produced a valid graph 2378 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2379 PhaseIdealLoop::verify(igvn); 2380 } 2381 } 2382 if (failing()) return; 2383 2384 // Conditional Constant Propagation; 2385 PhaseCCP ccp( &igvn ); 2386 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2387 { 2388 TracePhase tp("ccp", &timers[_t_ccp]); 2389 ccp.do_transform(); 2390 } 2391 print_method(PHASE_CPP1, 2); 2392 2393 assert( true, "Break here to ccp.dump_old2new_map()"); 2394 2395 // Iterative Global Value Numbering, including ideal transforms 2396 { 2397 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2398 igvn = ccp; 2399 igvn.optimize(); 2400 } 2401 2402 print_method(PHASE_ITER_GVN2, 2); 2403 2404 if (failing()) return; 2405 2406 // Loop transforms on the ideal graph. Range Check Elimination, 2407 // peeling, unrolling, etc. 2408 if(loop_opts_cnt > 0) { 2409 debug_only( int cnt = 0; ); 2410 while(major_progress() && (loop_opts_cnt > 0)) { 2411 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2412 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2413 PhaseIdealLoop ideal_loop( igvn, true); 2414 loop_opts_cnt--; 2415 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2416 if (failing()) return; 2417 } 2418 } 2419 2420 #if INCLUDE_ZGC 2421 if (UseZGC) { 2422 ZBarrierSetC2::find_dominating_barriers(igvn); 2423 } 2424 #endif 2425 2426 if (failing()) return; 2427 2428 // Ensure that major progress is now clear 2429 C->clear_major_progress(); 2430 2431 { 2432 // Verify that all previous optimizations produced a valid graph 2433 // at least to this point, even if no loop optimizations were done. 2434 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2435 PhaseIdealLoop::verify(igvn); 2436 } 2437 2438 if (range_check_cast_count() > 0) { 2439 // No more loop optimizations. Remove all range check dependent CastIINodes. 2440 C->remove_range_check_casts(igvn); 2441 igvn.optimize(); 2442 } 2443 2444 #ifdef ASSERT 2445 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2446 bs->verify_gc_barriers(false); 2447 #endif 2448 2449 { 2450 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2451 PhaseMacroExpand mex(igvn); 2452 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2); 2453 if (mex.expand_macro_nodes()) { 2454 assert(failing(), "must bail out w/ explicit message"); 2455 return; 2456 } 2457 } 2458 2459 if (opaque4_count() > 0) { 2460 C->remove_opaque4_nodes(igvn); 2461 igvn.optimize(); 2462 } 2463 2464 DEBUG_ONLY( _modified_nodes = NULL; ) 2465 } // (End scope of igvn; run destructor if necessary for asserts.) 2466 2467 process_print_inlining(); 2468 // A method with only infinite loops has no edges entering loops from root 2469 { 2470 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2471 if (final_graph_reshaping()) { 2472 assert(failing(), "must bail out w/ explicit message"); 2473 return; 2474 } 2475 } 2476 2477 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2478 } 2479 2480 //------------------------------Code_Gen--------------------------------------- 2481 // Given a graph, generate code for it 2482 void Compile::Code_Gen() { 2483 if (failing()) { 2484 return; 2485 } 2486 2487 // Perform instruction selection. You might think we could reclaim Matcher 2488 // memory PDQ, but actually the Matcher is used in generating spill code. 2489 // Internals of the Matcher (including some VectorSets) must remain live 2490 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2491 // set a bit in reclaimed memory. 2492 2493 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2494 // nodes. Mapping is only valid at the root of each matched subtree. 2495 NOT_PRODUCT( verify_graph_edges(); ) 2496 2497 Matcher matcher; 2498 _matcher = &matcher; 2499 { 2500 TracePhase tp("matcher", &timers[_t_matcher]); 2501 matcher.match(); 2502 } 2503 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2504 // nodes. Mapping is only valid at the root of each matched subtree. 2505 NOT_PRODUCT( verify_graph_edges(); ) 2506 2507 // If you have too many nodes, or if matching has failed, bail out 2508 check_node_count(0, "out of nodes matching instructions"); 2509 if (failing()) { 2510 return; 2511 } 2512 2513 print_method(PHASE_MATCHING, 2); 2514 2515 // Build a proper-looking CFG 2516 PhaseCFG cfg(node_arena(), root(), matcher); 2517 _cfg = &cfg; 2518 { 2519 TracePhase tp("scheduler", &timers[_t_scheduler]); 2520 bool success = cfg.do_global_code_motion(); 2521 if (!success) { 2522 return; 2523 } 2524 2525 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2526 NOT_PRODUCT( verify_graph_edges(); ) 2527 debug_only( cfg.verify(); ) 2528 } 2529 2530 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2531 _regalloc = ®alloc; 2532 { 2533 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2534 // Perform register allocation. After Chaitin, use-def chains are 2535 // no longer accurate (at spill code) and so must be ignored. 2536 // Node->LRG->reg mappings are still accurate. 2537 _regalloc->Register_Allocate(); 2538 2539 // Bail out if the allocator builds too many nodes 2540 if (failing()) { 2541 return; 2542 } 2543 } 2544 2545 // Prior to register allocation we kept empty basic blocks in case the 2546 // the allocator needed a place to spill. After register allocation we 2547 // are not adding any new instructions. If any basic block is empty, we 2548 // can now safely remove it. 2549 { 2550 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2551 cfg.remove_empty_blocks(); 2552 if (do_freq_based_layout()) { 2553 PhaseBlockLayout layout(cfg); 2554 } else { 2555 cfg.set_loop_alignment(); 2556 } 2557 cfg.fixup_flow(); 2558 } 2559 2560 // Apply peephole optimizations 2561 if( OptoPeephole ) { 2562 TracePhase tp("peephole", &timers[_t_peephole]); 2563 PhasePeephole peep( _regalloc, cfg); 2564 peep.do_transform(); 2565 } 2566 2567 // Do late expand if CPU requires this. 2568 if (Matcher::require_postalloc_expand) { 2569 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2570 cfg.postalloc_expand(_regalloc); 2571 } 2572 2573 // Convert Nodes to instruction bits in a buffer 2574 { 2575 TraceTime tp("output", &timers[_t_output], CITime); 2576 Output(); 2577 } 2578 2579 print_method(PHASE_FINAL_CODE); 2580 2581 // He's dead, Jim. 2582 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2583 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2584 } 2585 2586 2587 //------------------------------dump_asm--------------------------------------- 2588 // Dump formatted assembly 2589 #ifndef PRODUCT 2590 void Compile::dump_asm(int *pcs, uint pc_limit) { 2591 bool cut_short = false; 2592 tty->print_cr("#"); 2593 tty->print("# "); _tf->dump(); tty->cr(); 2594 tty->print_cr("#"); 2595 2596 // For all blocks 2597 int pc = 0x0; // Program counter 2598 char starts_bundle = ' '; 2599 _regalloc->dump_frame(); 2600 2601 Node *n = NULL; 2602 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2603 if (VMThread::should_terminate()) { 2604 cut_short = true; 2605 break; 2606 } 2607 Block* block = _cfg->get_block(i); 2608 if (block->is_connector() && !Verbose) { 2609 continue; 2610 } 2611 n = block->head(); 2612 if (pcs && n->_idx < pc_limit) { 2613 tty->print("%3.3x ", pcs[n->_idx]); 2614 } else { 2615 tty->print(" "); 2616 } 2617 block->dump_head(_cfg); 2618 if (block->is_connector()) { 2619 tty->print_cr(" # Empty connector block"); 2620 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2621 tty->print_cr(" # Block is sole successor of call"); 2622 } 2623 2624 // For all instructions 2625 Node *delay = NULL; 2626 for (uint j = 0; j < block->number_of_nodes(); j++) { 2627 if (VMThread::should_terminate()) { 2628 cut_short = true; 2629 break; 2630 } 2631 n = block->get_node(j); 2632 if (valid_bundle_info(n)) { 2633 Bundle* bundle = node_bundling(n); 2634 if (bundle->used_in_unconditional_delay()) { 2635 delay = n; 2636 continue; 2637 } 2638 if (bundle->starts_bundle()) { 2639 starts_bundle = '+'; 2640 } 2641 } 2642 2643 if (WizardMode) { 2644 n->dump(); 2645 } 2646 2647 if( !n->is_Region() && // Dont print in the Assembly 2648 !n->is_Phi() && // a few noisely useless nodes 2649 !n->is_Proj() && 2650 !n->is_MachTemp() && 2651 !n->is_SafePointScalarObject() && 2652 !n->is_Catch() && // Would be nice to print exception table targets 2653 !n->is_MergeMem() && // Not very interesting 2654 !n->is_top() && // Debug info table constants 2655 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2656 ) { 2657 if (pcs && n->_idx < pc_limit) 2658 tty->print("%3.3x", pcs[n->_idx]); 2659 else 2660 tty->print(" "); 2661 tty->print(" %c ", starts_bundle); 2662 starts_bundle = ' '; 2663 tty->print("\t"); 2664 n->format(_regalloc, tty); 2665 tty->cr(); 2666 } 2667 2668 // If we have an instruction with a delay slot, and have seen a delay, 2669 // then back up and print it 2670 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2671 assert(delay != NULL, "no unconditional delay instruction"); 2672 if (WizardMode) delay->dump(); 2673 2674 if (node_bundling(delay)->starts_bundle()) 2675 starts_bundle = '+'; 2676 if (pcs && n->_idx < pc_limit) 2677 tty->print("%3.3x", pcs[n->_idx]); 2678 else 2679 tty->print(" "); 2680 tty->print(" %c ", starts_bundle); 2681 starts_bundle = ' '; 2682 tty->print("\t"); 2683 delay->format(_regalloc, tty); 2684 tty->cr(); 2685 delay = NULL; 2686 } 2687 2688 // Dump the exception table as well 2689 if( n->is_Catch() && (Verbose || WizardMode) ) { 2690 // Print the exception table for this offset 2691 _handler_table.print_subtable_for(pc); 2692 } 2693 } 2694 2695 if (pcs && n->_idx < pc_limit) 2696 tty->print_cr("%3.3x", pcs[n->_idx]); 2697 else 2698 tty->cr(); 2699 2700 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2701 2702 } // End of per-block dump 2703 tty->cr(); 2704 2705 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2706 } 2707 #endif 2708 2709 //------------------------------Final_Reshape_Counts--------------------------- 2710 // This class defines counters to help identify when a method 2711 // may/must be executed using hardware with only 24-bit precision. 2712 struct Final_Reshape_Counts : public StackObj { 2713 int _call_count; // count non-inlined 'common' calls 2714 int _float_count; // count float ops requiring 24-bit precision 2715 int _double_count; // count double ops requiring more precision 2716 int _java_call_count; // count non-inlined 'java' calls 2717 int _inner_loop_count; // count loops which need alignment 2718 VectorSet _visited; // Visitation flags 2719 Node_List _tests; // Set of IfNodes & PCTableNodes 2720 2721 Final_Reshape_Counts() : 2722 _call_count(0), _float_count(0), _double_count(0), 2723 _java_call_count(0), _inner_loop_count(0), 2724 _visited( Thread::current()->resource_area() ) { } 2725 2726 void inc_call_count () { _call_count ++; } 2727 void inc_float_count () { _float_count ++; } 2728 void inc_double_count() { _double_count++; } 2729 void inc_java_call_count() { _java_call_count++; } 2730 void inc_inner_loop_count() { _inner_loop_count++; } 2731 2732 int get_call_count () const { return _call_count ; } 2733 int get_float_count () const { return _float_count ; } 2734 int get_double_count() const { return _double_count; } 2735 int get_java_call_count() const { return _java_call_count; } 2736 int get_inner_loop_count() const { return _inner_loop_count; } 2737 }; 2738 2739 #ifdef ASSERT 2740 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2741 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2742 // Make sure the offset goes inside the instance layout. 2743 return k->contains_field_offset(tp->offset()); 2744 // Note that OffsetBot and OffsetTop are very negative. 2745 } 2746 #endif 2747 2748 // Eliminate trivially redundant StoreCMs and accumulate their 2749 // precedence edges. 2750 void Compile::eliminate_redundant_card_marks(Node* n) { 2751 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2752 if (n->in(MemNode::Address)->outcnt() > 1) { 2753 // There are multiple users of the same address so it might be 2754 // possible to eliminate some of the StoreCMs 2755 Node* mem = n->in(MemNode::Memory); 2756 Node* adr = n->in(MemNode::Address); 2757 Node* val = n->in(MemNode::ValueIn); 2758 Node* prev = n; 2759 bool done = false; 2760 // Walk the chain of StoreCMs eliminating ones that match. As 2761 // long as it's a chain of single users then the optimization is 2762 // safe. Eliminating partially redundant StoreCMs would require 2763 // cloning copies down the other paths. 2764 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2765 if (adr == mem->in(MemNode::Address) && 2766 val == mem->in(MemNode::ValueIn)) { 2767 // redundant StoreCM 2768 if (mem->req() > MemNode::OopStore) { 2769 // Hasn't been processed by this code yet. 2770 n->add_prec(mem->in(MemNode::OopStore)); 2771 } else { 2772 // Already converted to precedence edge 2773 for (uint i = mem->req(); i < mem->len(); i++) { 2774 // Accumulate any precedence edges 2775 if (mem->in(i) != NULL) { 2776 n->add_prec(mem->in(i)); 2777 } 2778 } 2779 // Everything above this point has been processed. 2780 done = true; 2781 } 2782 // Eliminate the previous StoreCM 2783 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2784 assert(mem->outcnt() == 0, "should be dead"); 2785 mem->disconnect_inputs(NULL, this); 2786 } else { 2787 prev = mem; 2788 } 2789 mem = prev->in(MemNode::Memory); 2790 } 2791 } 2792 } 2793 2794 2795 //------------------------------final_graph_reshaping_impl---------------------- 2796 // Implement items 1-5 from final_graph_reshaping below. 2797 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2798 2799 if ( n->outcnt() == 0 ) return; // dead node 2800 uint nop = n->Opcode(); 2801 2802 // Check for 2-input instruction with "last use" on right input. 2803 // Swap to left input. Implements item (2). 2804 if( n->req() == 3 && // two-input instruction 2805 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2806 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2807 n->in(2)->outcnt() == 1 &&// right use IS a last use 2808 !n->in(2)->is_Con() ) { // right use is not a constant 2809 // Check for commutative opcode 2810 switch( nop ) { 2811 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2812 case Op_MaxI: case Op_MinI: 2813 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2814 case Op_AndL: case Op_XorL: case Op_OrL: 2815 case Op_AndI: case Op_XorI: case Op_OrI: { 2816 // Move "last use" input to left by swapping inputs 2817 n->swap_edges(1, 2); 2818 break; 2819 } 2820 default: 2821 break; 2822 } 2823 } 2824 2825 #ifdef ASSERT 2826 if( n->is_Mem() ) { 2827 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2828 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2829 // oop will be recorded in oop map if load crosses safepoint 2830 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2831 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2832 "raw memory operations should have control edge"); 2833 } 2834 #endif 2835 // Count FPU ops and common calls, implements item (3) 2836 switch( nop ) { 2837 // Count all float operations that may use FPU 2838 case Op_AddF: 2839 case Op_SubF: 2840 case Op_MulF: 2841 case Op_DivF: 2842 case Op_NegF: 2843 case Op_ModF: 2844 case Op_ConvI2F: 2845 case Op_ConF: 2846 case Op_CmpF: 2847 case Op_CmpF3: 2848 // case Op_ConvL2F: // longs are split into 32-bit halves 2849 frc.inc_float_count(); 2850 break; 2851 2852 case Op_ConvF2D: 2853 case Op_ConvD2F: 2854 frc.inc_float_count(); 2855 frc.inc_double_count(); 2856 break; 2857 2858 // Count all double operations that may use FPU 2859 case Op_AddD: 2860 case Op_SubD: 2861 case Op_MulD: 2862 case Op_DivD: 2863 case Op_NegD: 2864 case Op_ModD: 2865 case Op_ConvI2D: 2866 case Op_ConvD2I: 2867 // case Op_ConvL2D: // handled by leaf call 2868 // case Op_ConvD2L: // handled by leaf call 2869 case Op_ConD: 2870 case Op_CmpD: 2871 case Op_CmpD3: 2872 frc.inc_double_count(); 2873 break; 2874 case Op_Opaque1: // Remove Opaque Nodes before matching 2875 case Op_Opaque2: // Remove Opaque Nodes before matching 2876 case Op_Opaque3: 2877 n->subsume_by(n->in(1), this); 2878 break; 2879 case Op_CallStaticJava: 2880 case Op_CallJava: 2881 case Op_CallDynamicJava: 2882 frc.inc_java_call_count(); // Count java call site; 2883 case Op_CallRuntime: 2884 case Op_CallLeaf: 2885 case Op_CallLeafNoFP: { 2886 assert (n->is_Call(), ""); 2887 CallNode *call = n->as_Call(); 2888 // Count call sites where the FP mode bit would have to be flipped. 2889 // Do not count uncommon runtime calls: 2890 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2891 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2892 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 2893 frc.inc_call_count(); // Count the call site 2894 } else { // See if uncommon argument is shared 2895 Node *n = call->in(TypeFunc::Parms); 2896 int nop = n->Opcode(); 2897 // Clone shared simple arguments to uncommon calls, item (1). 2898 if (n->outcnt() > 1 && 2899 !n->is_Proj() && 2900 nop != Op_CreateEx && 2901 nop != Op_CheckCastPP && 2902 nop != Op_DecodeN && 2903 nop != Op_DecodeNKlass && 2904 !n->is_Mem() && 2905 !n->is_Phi()) { 2906 Node *x = n->clone(); 2907 call->set_req(TypeFunc::Parms, x); 2908 } 2909 } 2910 break; 2911 } 2912 2913 case Op_StoreD: 2914 case Op_LoadD: 2915 case Op_LoadD_unaligned: 2916 frc.inc_double_count(); 2917 goto handle_mem; 2918 case Op_StoreF: 2919 case Op_LoadF: 2920 frc.inc_float_count(); 2921 goto handle_mem; 2922 2923 case Op_StoreCM: 2924 { 2925 // Convert OopStore dependence into precedence edge 2926 Node* prec = n->in(MemNode::OopStore); 2927 n->del_req(MemNode::OopStore); 2928 n->add_prec(prec); 2929 eliminate_redundant_card_marks(n); 2930 } 2931 2932 // fall through 2933 2934 case Op_StoreB: 2935 case Op_StoreC: 2936 case Op_StorePConditional: 2937 case Op_StoreI: 2938 case Op_StoreL: 2939 case Op_StoreIConditional: 2940 case Op_StoreLConditional: 2941 case Op_CompareAndSwapB: 2942 case Op_CompareAndSwapS: 2943 case Op_CompareAndSwapI: 2944 case Op_CompareAndSwapL: 2945 case Op_CompareAndSwapP: 2946 case Op_CompareAndSwapN: 2947 case Op_WeakCompareAndSwapB: 2948 case Op_WeakCompareAndSwapS: 2949 case Op_WeakCompareAndSwapI: 2950 case Op_WeakCompareAndSwapL: 2951 case Op_WeakCompareAndSwapP: 2952 case Op_WeakCompareAndSwapN: 2953 case Op_CompareAndExchangeB: 2954 case Op_CompareAndExchangeS: 2955 case Op_CompareAndExchangeI: 2956 case Op_CompareAndExchangeL: 2957 case Op_CompareAndExchangeP: 2958 case Op_CompareAndExchangeN: 2959 case Op_GetAndAddS: 2960 case Op_GetAndAddB: 2961 case Op_GetAndAddI: 2962 case Op_GetAndAddL: 2963 case Op_GetAndSetS: 2964 case Op_GetAndSetB: 2965 case Op_GetAndSetI: 2966 case Op_GetAndSetL: 2967 case Op_GetAndSetP: 2968 case Op_GetAndSetN: 2969 case Op_StoreP: 2970 case Op_StoreN: 2971 case Op_StoreNKlass: 2972 case Op_LoadB: 2973 case Op_LoadUB: 2974 case Op_LoadUS: 2975 case Op_LoadI: 2976 case Op_LoadKlass: 2977 case Op_LoadNKlass: 2978 case Op_LoadL: 2979 case Op_LoadL_unaligned: 2980 case Op_LoadPLocked: 2981 case Op_LoadP: 2982 #if INCLUDE_ZGC 2983 case Op_LoadBarrierSlowReg: 2984 case Op_LoadBarrierWeakSlowReg: 2985 #endif 2986 case Op_LoadN: 2987 case Op_LoadRange: 2988 case Op_LoadS: { 2989 handle_mem: 2990 #ifdef ASSERT 2991 if( VerifyOptoOopOffsets ) { 2992 assert( n->is_Mem(), "" ); 2993 MemNode *mem = (MemNode*)n; 2994 // Check to see if address types have grounded out somehow. 2995 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2996 assert( !tp || oop_offset_is_sane(tp), "" ); 2997 } 2998 #endif 2999 break; 3000 } 3001 3002 case Op_AddP: { // Assert sane base pointers 3003 Node *addp = n->in(AddPNode::Address); 3004 assert( !addp->is_AddP() || 3005 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3006 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3007 "Base pointers must match (addp %u)", addp->_idx ); 3008 #ifdef _LP64 3009 if ((UseCompressedOops || UseCompressedClassPointers) && 3010 addp->Opcode() == Op_ConP && 3011 addp == n->in(AddPNode::Base) && 3012 n->in(AddPNode::Offset)->is_Con()) { 3013 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3014 // on the platform and on the compressed oops mode. 3015 // Use addressing with narrow klass to load with offset on x86. 3016 // Some platforms can use the constant pool to load ConP. 3017 // Do this transformation here since IGVN will convert ConN back to ConP. 3018 const Type* t = addp->bottom_type(); 3019 bool is_oop = t->isa_oopptr() != NULL; 3020 bool is_klass = t->isa_klassptr() != NULL; 3021 3022 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3023 (is_klass && Matcher::const_klass_prefer_decode())) { 3024 Node* nn = NULL; 3025 3026 int op = is_oop ? Op_ConN : Op_ConNKlass; 3027 3028 // Look for existing ConN node of the same exact type. 3029 Node* r = root(); 3030 uint cnt = r->outcnt(); 3031 for (uint i = 0; i < cnt; i++) { 3032 Node* m = r->raw_out(i); 3033 if (m!= NULL && m->Opcode() == op && 3034 m->bottom_type()->make_ptr() == t) { 3035 nn = m; 3036 break; 3037 } 3038 } 3039 if (nn != NULL) { 3040 // Decode a narrow oop to match address 3041 // [R12 + narrow_oop_reg<<3 + offset] 3042 if (is_oop) { 3043 nn = new DecodeNNode(nn, t); 3044 } else { 3045 nn = new DecodeNKlassNode(nn, t); 3046 } 3047 // Check for succeeding AddP which uses the same Base. 3048 // Otherwise we will run into the assertion above when visiting that guy. 3049 for (uint i = 0; i < n->outcnt(); ++i) { 3050 Node *out_i = n->raw_out(i); 3051 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3052 out_i->set_req(AddPNode::Base, nn); 3053 #ifdef ASSERT 3054 for (uint j = 0; j < out_i->outcnt(); ++j) { 3055 Node *out_j = out_i->raw_out(j); 3056 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3057 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3058 } 3059 #endif 3060 } 3061 } 3062 n->set_req(AddPNode::Base, nn); 3063 n->set_req(AddPNode::Address, nn); 3064 if (addp->outcnt() == 0) { 3065 addp->disconnect_inputs(NULL, this); 3066 } 3067 } 3068 } 3069 } 3070 #endif 3071 // platform dependent reshaping of the address expression 3072 reshape_address(n->as_AddP()); 3073 break; 3074 } 3075 3076 case Op_CastPP: { 3077 // Remove CastPP nodes to gain more freedom during scheduling but 3078 // keep the dependency they encode as control or precedence edges 3079 // (if control is set already) on memory operations. Some CastPP 3080 // nodes don't have a control (don't carry a dependency): skip 3081 // those. 3082 if (n->in(0) != NULL) { 3083 ResourceMark rm; 3084 Unique_Node_List wq; 3085 wq.push(n); 3086 for (uint next = 0; next < wq.size(); ++next) { 3087 Node *m = wq.at(next); 3088 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3089 Node* use = m->fast_out(i); 3090 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3091 use->ensure_control_or_add_prec(n->in(0)); 3092 } else { 3093 switch(use->Opcode()) { 3094 case Op_AddP: 3095 case Op_DecodeN: 3096 case Op_DecodeNKlass: 3097 case Op_CheckCastPP: 3098 case Op_CastPP: 3099 wq.push(use); 3100 break; 3101 } 3102 } 3103 } 3104 } 3105 } 3106 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3107 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3108 Node* in1 = n->in(1); 3109 const Type* t = n->bottom_type(); 3110 Node* new_in1 = in1->clone(); 3111 new_in1->as_DecodeN()->set_type(t); 3112 3113 if (!Matcher::narrow_oop_use_complex_address()) { 3114 // 3115 // x86, ARM and friends can handle 2 adds in addressing mode 3116 // and Matcher can fold a DecodeN node into address by using 3117 // a narrow oop directly and do implicit NULL check in address: 3118 // 3119 // [R12 + narrow_oop_reg<<3 + offset] 3120 // NullCheck narrow_oop_reg 3121 // 3122 // On other platforms (Sparc) we have to keep new DecodeN node and 3123 // use it to do implicit NULL check in address: 3124 // 3125 // decode_not_null narrow_oop_reg, base_reg 3126 // [base_reg + offset] 3127 // NullCheck base_reg 3128 // 3129 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3130 // to keep the information to which NULL check the new DecodeN node 3131 // corresponds to use it as value in implicit_null_check(). 3132 // 3133 new_in1->set_req(0, n->in(0)); 3134 } 3135 3136 n->subsume_by(new_in1, this); 3137 if (in1->outcnt() == 0) { 3138 in1->disconnect_inputs(NULL, this); 3139 } 3140 } else { 3141 n->subsume_by(n->in(1), this); 3142 if (n->outcnt() == 0) { 3143 n->disconnect_inputs(NULL, this); 3144 } 3145 } 3146 break; 3147 } 3148 #ifdef _LP64 3149 case Op_CmpP: 3150 // Do this transformation here to preserve CmpPNode::sub() and 3151 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3152 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3153 Node* in1 = n->in(1); 3154 Node* in2 = n->in(2); 3155 if (!in1->is_DecodeNarrowPtr()) { 3156 in2 = in1; 3157 in1 = n->in(2); 3158 } 3159 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3160 3161 Node* new_in2 = NULL; 3162 if (in2->is_DecodeNarrowPtr()) { 3163 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3164 new_in2 = in2->in(1); 3165 } else if (in2->Opcode() == Op_ConP) { 3166 const Type* t = in2->bottom_type(); 3167 if (t == TypePtr::NULL_PTR) { 3168 assert(in1->is_DecodeN(), "compare klass to null?"); 3169 // Don't convert CmpP null check into CmpN if compressed 3170 // oops implicit null check is not generated. 3171 // This will allow to generate normal oop implicit null check. 3172 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3173 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3174 // 3175 // This transformation together with CastPP transformation above 3176 // will generated code for implicit NULL checks for compressed oops. 3177 // 3178 // The original code after Optimize() 3179 // 3180 // LoadN memory, narrow_oop_reg 3181 // decode narrow_oop_reg, base_reg 3182 // CmpP base_reg, NULL 3183 // CastPP base_reg // NotNull 3184 // Load [base_reg + offset], val_reg 3185 // 3186 // after these transformations will be 3187 // 3188 // LoadN memory, narrow_oop_reg 3189 // CmpN narrow_oop_reg, NULL 3190 // decode_not_null narrow_oop_reg, base_reg 3191 // Load [base_reg + offset], val_reg 3192 // 3193 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3194 // since narrow oops can be used in debug info now (see the code in 3195 // final_graph_reshaping_walk()). 3196 // 3197 // At the end the code will be matched to 3198 // on x86: 3199 // 3200 // Load_narrow_oop memory, narrow_oop_reg 3201 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3202 // NullCheck narrow_oop_reg 3203 // 3204 // and on sparc: 3205 // 3206 // Load_narrow_oop memory, narrow_oop_reg 3207 // decode_not_null narrow_oop_reg, base_reg 3208 // Load [base_reg + offset], val_reg 3209 // NullCheck base_reg 3210 // 3211 } else if (t->isa_oopptr()) { 3212 new_in2 = ConNode::make(t->make_narrowoop()); 3213 } else if (t->isa_klassptr()) { 3214 new_in2 = ConNode::make(t->make_narrowklass()); 3215 } 3216 } 3217 if (new_in2 != NULL) { 3218 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3219 n->subsume_by(cmpN, this); 3220 if (in1->outcnt() == 0) { 3221 in1->disconnect_inputs(NULL, this); 3222 } 3223 if (in2->outcnt() == 0) { 3224 in2->disconnect_inputs(NULL, this); 3225 } 3226 } 3227 } 3228 break; 3229 3230 case Op_DecodeN: 3231 case Op_DecodeNKlass: 3232 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3233 // DecodeN could be pinned when it can't be fold into 3234 // an address expression, see the code for Op_CastPP above. 3235 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3236 break; 3237 3238 case Op_EncodeP: 3239 case Op_EncodePKlass: { 3240 Node* in1 = n->in(1); 3241 if (in1->is_DecodeNarrowPtr()) { 3242 n->subsume_by(in1->in(1), this); 3243 } else if (in1->Opcode() == Op_ConP) { 3244 const Type* t = in1->bottom_type(); 3245 if (t == TypePtr::NULL_PTR) { 3246 assert(t->isa_oopptr(), "null klass?"); 3247 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3248 } else if (t->isa_oopptr()) { 3249 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3250 } else if (t->isa_klassptr()) { 3251 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3252 } 3253 } 3254 if (in1->outcnt() == 0) { 3255 in1->disconnect_inputs(NULL, this); 3256 } 3257 break; 3258 } 3259 3260 case Op_Proj: { 3261 if (OptimizeStringConcat) { 3262 ProjNode* p = n->as_Proj(); 3263 if (p->_is_io_use) { 3264 // Separate projections were used for the exception path which 3265 // are normally removed by a late inline. If it wasn't inlined 3266 // then they will hang around and should just be replaced with 3267 // the original one. 3268 Node* proj = NULL; 3269 // Replace with just one 3270 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3271 Node *use = i.get(); 3272 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3273 proj = use; 3274 break; 3275 } 3276 } 3277 assert(proj != NULL, "must be found"); 3278 p->subsume_by(proj, this); 3279 } 3280 } 3281 break; 3282 } 3283 3284 case Op_Phi: 3285 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3286 // The EncodeP optimization may create Phi with the same edges 3287 // for all paths. It is not handled well by Register Allocator. 3288 Node* unique_in = n->in(1); 3289 assert(unique_in != NULL, ""); 3290 uint cnt = n->req(); 3291 for (uint i = 2; i < cnt; i++) { 3292 Node* m = n->in(i); 3293 assert(m != NULL, ""); 3294 if (unique_in != m) 3295 unique_in = NULL; 3296 } 3297 if (unique_in != NULL) { 3298 n->subsume_by(unique_in, this); 3299 } 3300 } 3301 break; 3302 3303 #endif 3304 3305 #ifdef ASSERT 3306 case Op_CastII: 3307 // Verify that all range check dependent CastII nodes were removed. 3308 if (n->isa_CastII()->has_range_check()) { 3309 n->dump(3); 3310 assert(false, "Range check dependent CastII node was not removed"); 3311 } 3312 break; 3313 #endif 3314 3315 case Op_ModI: 3316 if (UseDivMod) { 3317 // Check if a%b and a/b both exist 3318 Node* d = n->find_similar(Op_DivI); 3319 if (d) { 3320 // Replace them with a fused divmod if supported 3321 if (Matcher::has_match_rule(Op_DivModI)) { 3322 DivModINode* divmod = DivModINode::make(n); 3323 d->subsume_by(divmod->div_proj(), this); 3324 n->subsume_by(divmod->mod_proj(), this); 3325 } else { 3326 // replace a%b with a-((a/b)*b) 3327 Node* mult = new MulINode(d, d->in(2)); 3328 Node* sub = new SubINode(d->in(1), mult); 3329 n->subsume_by(sub, this); 3330 } 3331 } 3332 } 3333 break; 3334 3335 case Op_ModL: 3336 if (UseDivMod) { 3337 // Check if a%b and a/b both exist 3338 Node* d = n->find_similar(Op_DivL); 3339 if (d) { 3340 // Replace them with a fused divmod if supported 3341 if (Matcher::has_match_rule(Op_DivModL)) { 3342 DivModLNode* divmod = DivModLNode::make(n); 3343 d->subsume_by(divmod->div_proj(), this); 3344 n->subsume_by(divmod->mod_proj(), this); 3345 } else { 3346 // replace a%b with a-((a/b)*b) 3347 Node* mult = new MulLNode(d, d->in(2)); 3348 Node* sub = new SubLNode(d->in(1), mult); 3349 n->subsume_by(sub, this); 3350 } 3351 } 3352 } 3353 break; 3354 3355 case Op_LoadVector: 3356 case Op_StoreVector: 3357 break; 3358 3359 case Op_AddReductionVI: 3360 case Op_AddReductionVL: 3361 case Op_AddReductionVF: 3362 case Op_AddReductionVD: 3363 case Op_MulReductionVI: 3364 case Op_MulReductionVL: 3365 case Op_MulReductionVF: 3366 case Op_MulReductionVD: 3367 break; 3368 3369 case Op_PackB: 3370 case Op_PackS: 3371 case Op_PackI: 3372 case Op_PackF: 3373 case Op_PackL: 3374 case Op_PackD: 3375 if (n->req()-1 > 2) { 3376 // Replace many operand PackNodes with a binary tree for matching 3377 PackNode* p = (PackNode*) n; 3378 Node* btp = p->binary_tree_pack(1, n->req()); 3379 n->subsume_by(btp, this); 3380 } 3381 break; 3382 case Op_Loop: 3383 case Op_CountedLoop: 3384 case Op_OuterStripMinedLoop: 3385 if (n->as_Loop()->is_inner_loop()) { 3386 frc.inc_inner_loop_count(); 3387 } 3388 n->as_Loop()->verify_strip_mined(0); 3389 break; 3390 case Op_LShiftI: 3391 case Op_RShiftI: 3392 case Op_URShiftI: 3393 case Op_LShiftL: 3394 case Op_RShiftL: 3395 case Op_URShiftL: 3396 if (Matcher::need_masked_shift_count) { 3397 // The cpu's shift instructions don't restrict the count to the 3398 // lower 5/6 bits. We need to do the masking ourselves. 3399 Node* in2 = n->in(2); 3400 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3401 const TypeInt* t = in2->find_int_type(); 3402 if (t != NULL && t->is_con()) { 3403 juint shift = t->get_con(); 3404 if (shift > mask) { // Unsigned cmp 3405 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3406 } 3407 } else { 3408 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3409 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3410 n->set_req(2, shift); 3411 } 3412 } 3413 if (in2->outcnt() == 0) { // Remove dead node 3414 in2->disconnect_inputs(NULL, this); 3415 } 3416 } 3417 break; 3418 case Op_MemBarStoreStore: 3419 case Op_MemBarRelease: 3420 // Break the link with AllocateNode: it is no longer useful and 3421 // confuses register allocation. 3422 if (n->req() > MemBarNode::Precedent) { 3423 n->set_req(MemBarNode::Precedent, top()); 3424 } 3425 break; 3426 case Op_RangeCheck: { 3427 RangeCheckNode* rc = n->as_RangeCheck(); 3428 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3429 n->subsume_by(iff, this); 3430 frc._tests.push(iff); 3431 break; 3432 } 3433 case Op_ConvI2L: { 3434 if (!Matcher::convi2l_type_required) { 3435 // Code generation on some platforms doesn't need accurate 3436 // ConvI2L types. Widening the type can help remove redundant 3437 // address computations. 3438 n->as_Type()->set_type(TypeLong::INT); 3439 ResourceMark rm; 3440 Node_List wq; 3441 wq.push(n); 3442 for (uint next = 0; next < wq.size(); next++) { 3443 Node *m = wq.at(next); 3444 3445 for(;;) { 3446 // Loop over all nodes with identical inputs edges as m 3447 Node* k = m->find_similar(m->Opcode()); 3448 if (k == NULL) { 3449 break; 3450 } 3451 // Push their uses so we get a chance to remove node made 3452 // redundant 3453 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3454 Node* u = k->fast_out(i); 3455 assert(!wq.contains(u), "shouldn't process one node several times"); 3456 if (u->Opcode() == Op_LShiftL || 3457 u->Opcode() == Op_AddL || 3458 u->Opcode() == Op_SubL || 3459 u->Opcode() == Op_AddP) { 3460 wq.push(u); 3461 } 3462 } 3463 // Replace all nodes with identical edges as m with m 3464 k->subsume_by(m, this); 3465 } 3466 } 3467 } 3468 break; 3469 } 3470 case Op_CmpUL: { 3471 if (!Matcher::has_match_rule(Op_CmpUL)) { 3472 // We don't support unsigned long comparisons. Set 'max_idx_expr' 3473 // to max_julong if < 0 to make the signed comparison fail. 3474 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3475 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3476 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3477 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3478 Node* andl = new AndLNode(orl, remove_sign_mask); 3479 Node* cmp = new CmpLNode(andl, n->in(2)); 3480 n->subsume_by(cmp, this); 3481 } 3482 break; 3483 } 3484 #ifdef ASSERT 3485 case Op_ValueTypePtr: 3486 case Op_ValueType: { 3487 n->dump(-1); 3488 assert(false, "value type node was not removed"); 3489 break; 3490 } 3491 #endif 3492 default: 3493 assert( !n->is_Call(), "" ); 3494 assert( !n->is_Mem(), "" ); 3495 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3496 break; 3497 } 3498 3499 // Collect CFG split points 3500 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3501 frc._tests.push(n); 3502 } 3503 } 3504 3505 //------------------------------final_graph_reshaping_walk--------------------- 3506 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3507 // requires that the walk visits a node's inputs before visiting the node. 3508 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3509 ResourceArea *area = Thread::current()->resource_area(); 3510 Unique_Node_List sfpt(area); 3511 3512 frc._visited.set(root->_idx); // first, mark node as visited 3513 uint cnt = root->req(); 3514 Node *n = root; 3515 uint i = 0; 3516 while (true) { 3517 if (i < cnt) { 3518 // Place all non-visited non-null inputs onto stack 3519 Node* m = n->in(i); 3520 ++i; 3521 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3522 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3523 // compute worst case interpreter size in case of a deoptimization 3524 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3525 3526 sfpt.push(m); 3527 } 3528 cnt = m->req(); 3529 nstack.push(n, i); // put on stack parent and next input's index 3530 n = m; 3531 i = 0; 3532 } 3533 } else { 3534 // Now do post-visit work 3535 final_graph_reshaping_impl( n, frc ); 3536 if (nstack.is_empty()) 3537 break; // finished 3538 n = nstack.node(); // Get node from stack 3539 cnt = n->req(); 3540 i = nstack.index(); 3541 nstack.pop(); // Shift to the next node on stack 3542 } 3543 } 3544 3545 // Skip next transformation if compressed oops are not used. 3546 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3547 (!UseCompressedOops && !UseCompressedClassPointers)) 3548 return; 3549 3550 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3551 // It could be done for an uncommon traps or any safepoints/calls 3552 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3553 while (sfpt.size() > 0) { 3554 n = sfpt.pop(); 3555 JVMState *jvms = n->as_SafePoint()->jvms(); 3556 assert(jvms != NULL, "sanity"); 3557 int start = jvms->debug_start(); 3558 int end = n->req(); 3559 bool is_uncommon = (n->is_CallStaticJava() && 3560 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3561 for (int j = start; j < end; j++) { 3562 Node* in = n->in(j); 3563 if (in->is_DecodeNarrowPtr()) { 3564 bool safe_to_skip = true; 3565 if (!is_uncommon ) { 3566 // Is it safe to skip? 3567 for (uint i = 0; i < in->outcnt(); i++) { 3568 Node* u = in->raw_out(i); 3569 if (!u->is_SafePoint() || 3570 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3571 safe_to_skip = false; 3572 } 3573 } 3574 } 3575 if (safe_to_skip) { 3576 n->set_req(j, in->in(1)); 3577 } 3578 if (in->outcnt() == 0) { 3579 in->disconnect_inputs(NULL, this); 3580 } 3581 } 3582 } 3583 } 3584 } 3585 3586 //------------------------------final_graph_reshaping-------------------------- 3587 // Final Graph Reshaping. 3588 // 3589 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3590 // and not commoned up and forced early. Must come after regular 3591 // optimizations to avoid GVN undoing the cloning. Clone constant 3592 // inputs to Loop Phis; these will be split by the allocator anyways. 3593 // Remove Opaque nodes. 3594 // (2) Move last-uses by commutative operations to the left input to encourage 3595 // Intel update-in-place two-address operations and better register usage 3596 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3597 // calls canonicalizing them back. 3598 // (3) Count the number of double-precision FP ops, single-precision FP ops 3599 // and call sites. On Intel, we can get correct rounding either by 3600 // forcing singles to memory (requires extra stores and loads after each 3601 // FP bytecode) or we can set a rounding mode bit (requires setting and 3602 // clearing the mode bit around call sites). The mode bit is only used 3603 // if the relative frequency of single FP ops to calls is low enough. 3604 // This is a key transform for SPEC mpeg_audio. 3605 // (4) Detect infinite loops; blobs of code reachable from above but not 3606 // below. Several of the Code_Gen algorithms fail on such code shapes, 3607 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3608 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3609 // Detection is by looking for IfNodes where only 1 projection is 3610 // reachable from below or CatchNodes missing some targets. 3611 // (5) Assert for insane oop offsets in debug mode. 3612 3613 bool Compile::final_graph_reshaping() { 3614 // an infinite loop may have been eliminated by the optimizer, 3615 // in which case the graph will be empty. 3616 if (root()->req() == 1) { 3617 record_method_not_compilable("trivial infinite loop"); 3618 return true; 3619 } 3620 3621 // Expensive nodes have their control input set to prevent the GVN 3622 // from freely commoning them. There's no GVN beyond this point so 3623 // no need to keep the control input. We want the expensive nodes to 3624 // be freely moved to the least frequent code path by gcm. 3625 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3626 for (int i = 0; i < expensive_count(); i++) { 3627 _expensive_nodes->at(i)->set_req(0, NULL); 3628 } 3629 3630 Final_Reshape_Counts frc; 3631 3632 // Visit everybody reachable! 3633 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3634 Node_Stack nstack(live_nodes() >> 1); 3635 final_graph_reshaping_walk(nstack, root(), frc); 3636 3637 // Check for unreachable (from below) code (i.e., infinite loops). 3638 for( uint i = 0; i < frc._tests.size(); i++ ) { 3639 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3640 // Get number of CFG targets. 3641 // Note that PCTables include exception targets after calls. 3642 uint required_outcnt = n->required_outcnt(); 3643 if (n->outcnt() != required_outcnt) { 3644 // Check for a few special cases. Rethrow Nodes never take the 3645 // 'fall-thru' path, so expected kids is 1 less. 3646 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3647 if (n->in(0)->in(0)->is_Call()) { 3648 CallNode *call = n->in(0)->in(0)->as_Call(); 3649 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3650 required_outcnt--; // Rethrow always has 1 less kid 3651 } else if (call->req() > TypeFunc::Parms && 3652 call->is_CallDynamicJava()) { 3653 // Check for null receiver. In such case, the optimizer has 3654 // detected that the virtual call will always result in a null 3655 // pointer exception. The fall-through projection of this CatchNode 3656 // will not be populated. 3657 Node *arg0 = call->in(TypeFunc::Parms); 3658 if (arg0->is_Type() && 3659 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3660 required_outcnt--; 3661 } 3662 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3663 call->req() > TypeFunc::Parms+1 && 3664 call->is_CallStaticJava()) { 3665 // Check for negative array length. In such case, the optimizer has 3666 // detected that the allocation attempt will always result in an 3667 // exception. There is no fall-through projection of this CatchNode . 3668 Node *arg1 = call->in(TypeFunc::Parms+1); 3669 if (arg1->is_Type() && 3670 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3671 required_outcnt--; 3672 } 3673 } 3674 } 3675 } 3676 // Recheck with a better notion of 'required_outcnt' 3677 if (n->outcnt() != required_outcnt) { 3678 record_method_not_compilable("malformed control flow"); 3679 return true; // Not all targets reachable! 3680 } 3681 } 3682 // Check that I actually visited all kids. Unreached kids 3683 // must be infinite loops. 3684 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3685 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3686 record_method_not_compilable("infinite loop"); 3687 return true; // Found unvisited kid; must be unreach 3688 } 3689 3690 // Here so verification code in final_graph_reshaping_walk() 3691 // always see an OuterStripMinedLoopEnd 3692 if (n->is_OuterStripMinedLoopEnd()) { 3693 IfNode* init_iff = n->as_If(); 3694 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3695 n->subsume_by(iff, this); 3696 } 3697 } 3698 3699 // If original bytecodes contained a mixture of floats and doubles 3700 // check if the optimizer has made it homogenous, item (3). 3701 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3702 frc.get_float_count() > 32 && 3703 frc.get_double_count() == 0 && 3704 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3705 set_24_bit_selection_and_mode( false, true ); 3706 } 3707 3708 set_java_calls(frc.get_java_call_count()); 3709 set_inner_loops(frc.get_inner_loop_count()); 3710 3711 // No infinite loops, no reason to bail out. 3712 return false; 3713 } 3714 3715 //-----------------------------too_many_traps---------------------------------- 3716 // Report if there are too many traps at the current method and bci. 3717 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3718 bool Compile::too_many_traps(ciMethod* method, 3719 int bci, 3720 Deoptimization::DeoptReason reason) { 3721 ciMethodData* md = method->method_data(); 3722 if (md->is_empty()) { 3723 // Assume the trap has not occurred, or that it occurred only 3724 // because of a transient condition during start-up in the interpreter. 3725 return false; 3726 } 3727 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3728 if (md->has_trap_at(bci, m, reason) != 0) { 3729 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3730 // Also, if there are multiple reasons, or if there is no per-BCI record, 3731 // assume the worst. 3732 if (log()) 3733 log()->elem("observe trap='%s' count='%d'", 3734 Deoptimization::trap_reason_name(reason), 3735 md->trap_count(reason)); 3736 return true; 3737 } else { 3738 // Ignore method/bci and see if there have been too many globally. 3739 return too_many_traps(reason, md); 3740 } 3741 } 3742 3743 // Less-accurate variant which does not require a method and bci. 3744 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3745 ciMethodData* logmd) { 3746 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3747 // Too many traps globally. 3748 // Note that we use cumulative trap_count, not just md->trap_count. 3749 if (log()) { 3750 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3751 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3752 Deoptimization::trap_reason_name(reason), 3753 mcount, trap_count(reason)); 3754 } 3755 return true; 3756 } else { 3757 // The coast is clear. 3758 return false; 3759 } 3760 } 3761 3762 //--------------------------too_many_recompiles-------------------------------- 3763 // Report if there are too many recompiles at the current method and bci. 3764 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3765 // Is not eager to return true, since this will cause the compiler to use 3766 // Action_none for a trap point, to avoid too many recompilations. 3767 bool Compile::too_many_recompiles(ciMethod* method, 3768 int bci, 3769 Deoptimization::DeoptReason reason) { 3770 ciMethodData* md = method->method_data(); 3771 if (md->is_empty()) { 3772 // Assume the trap has not occurred, or that it occurred only 3773 // because of a transient condition during start-up in the interpreter. 3774 return false; 3775 } 3776 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3777 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3778 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3779 Deoptimization::DeoptReason per_bc_reason 3780 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3781 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3782 if ((per_bc_reason == Deoptimization::Reason_none 3783 || md->has_trap_at(bci, m, reason) != 0) 3784 // The trap frequency measure we care about is the recompile count: 3785 && md->trap_recompiled_at(bci, m) 3786 && md->overflow_recompile_count() >= bc_cutoff) { 3787 // Do not emit a trap here if it has already caused recompilations. 3788 // Also, if there are multiple reasons, or if there is no per-BCI record, 3789 // assume the worst. 3790 if (log()) 3791 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3792 Deoptimization::trap_reason_name(reason), 3793 md->trap_count(reason), 3794 md->overflow_recompile_count()); 3795 return true; 3796 } else if (trap_count(reason) != 0 3797 && decompile_count() >= m_cutoff) { 3798 // Too many recompiles globally, and we have seen this sort of trap. 3799 // Use cumulative decompile_count, not just md->decompile_count. 3800 if (log()) 3801 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3802 Deoptimization::trap_reason_name(reason), 3803 md->trap_count(reason), trap_count(reason), 3804 md->decompile_count(), decompile_count()); 3805 return true; 3806 } else { 3807 // The coast is clear. 3808 return false; 3809 } 3810 } 3811 3812 // Compute when not to trap. Used by matching trap based nodes and 3813 // NullCheck optimization. 3814 void Compile::set_allowed_deopt_reasons() { 3815 _allowed_reasons = 0; 3816 if (is_method_compilation()) { 3817 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3818 assert(rs < BitsPerInt, "recode bit map"); 3819 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3820 _allowed_reasons |= nth_bit(rs); 3821 } 3822 } 3823 } 3824 } 3825 3826 #ifndef PRODUCT 3827 //------------------------------verify_graph_edges--------------------------- 3828 // Walk the Graph and verify that there is a one-to-one correspondence 3829 // between Use-Def edges and Def-Use edges in the graph. 3830 void Compile::verify_graph_edges(bool no_dead_code) { 3831 if (VerifyGraphEdges) { 3832 ResourceArea *area = Thread::current()->resource_area(); 3833 Unique_Node_List visited(area); 3834 // Call recursive graph walk to check edges 3835 _root->verify_edges(visited); 3836 if (no_dead_code) { 3837 // Now make sure that no visited node is used by an unvisited node. 3838 bool dead_nodes = false; 3839 Unique_Node_List checked(area); 3840 while (visited.size() > 0) { 3841 Node* n = visited.pop(); 3842 checked.push(n); 3843 for (uint i = 0; i < n->outcnt(); i++) { 3844 Node* use = n->raw_out(i); 3845 if (checked.member(use)) continue; // already checked 3846 if (visited.member(use)) continue; // already in the graph 3847 if (use->is_Con()) continue; // a dead ConNode is OK 3848 // At this point, we have found a dead node which is DU-reachable. 3849 if (!dead_nodes) { 3850 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3851 dead_nodes = true; 3852 } 3853 use->dump(2); 3854 tty->print_cr("---"); 3855 checked.push(use); // No repeats; pretend it is now checked. 3856 } 3857 } 3858 assert(!dead_nodes, "using nodes must be reachable from root"); 3859 } 3860 } 3861 } 3862 3863 // Verify GC barriers consistency 3864 // Currently supported: 3865 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3866 void Compile::verify_barriers() { 3867 #if INCLUDE_G1GC 3868 if (UseG1GC) { 3869 // Verify G1 pre-barriers 3870 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()); 3871 3872 ResourceArea *area = Thread::current()->resource_area(); 3873 Unique_Node_List visited(area); 3874 Node_List worklist(area); 3875 // We're going to walk control flow backwards starting from the Root 3876 worklist.push(_root); 3877 while (worklist.size() > 0) { 3878 Node* x = worklist.pop(); 3879 if (x == NULL || x == top()) continue; 3880 if (visited.member(x)) { 3881 continue; 3882 } else { 3883 visited.push(x); 3884 } 3885 3886 if (x->is_Region()) { 3887 for (uint i = 1; i < x->req(); i++) { 3888 worklist.push(x->in(i)); 3889 } 3890 } else { 3891 worklist.push(x->in(0)); 3892 // We are looking for the pattern: 3893 // /->ThreadLocal 3894 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3895 // \->ConI(0) 3896 // We want to verify that the If and the LoadB have the same control 3897 // See GraphKit::g1_write_barrier_pre() 3898 if (x->is_If()) { 3899 IfNode *iff = x->as_If(); 3900 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3901 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3902 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3903 && cmp->in(1)->is_Load()) { 3904 LoadNode* load = cmp->in(1)->as_Load(); 3905 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3906 && load->in(2)->in(3)->is_Con() 3907 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3908 3909 Node* if_ctrl = iff->in(0); 3910 Node* load_ctrl = load->in(0); 3911 3912 if (if_ctrl != load_ctrl) { 3913 // Skip possible CProj->NeverBranch in infinite loops 3914 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3915 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3916 if_ctrl = if_ctrl->in(0)->in(0); 3917 } 3918 } 3919 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3920 } 3921 } 3922 } 3923 } 3924 } 3925 } 3926 } 3927 #endif 3928 } 3929 3930 #endif 3931 3932 // The Compile object keeps track of failure reasons separately from the ciEnv. 3933 // This is required because there is not quite a 1-1 relation between the 3934 // ciEnv and its compilation task and the Compile object. Note that one 3935 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3936 // to backtrack and retry without subsuming loads. Other than this backtracking 3937 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3938 // by the logic in C2Compiler. 3939 void Compile::record_failure(const char* reason) { 3940 if (log() != NULL) { 3941 log()->elem("failure reason='%s' phase='compile'", reason); 3942 } 3943 if (_failure_reason == NULL) { 3944 // Record the first failure reason. 3945 _failure_reason = reason; 3946 } 3947 3948 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3949 C->print_method(PHASE_FAILURE); 3950 } 3951 _root = NULL; // flush the graph, too 3952 } 3953 3954 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3955 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3956 _phase_name(name), _dolog(CITimeVerbose) 3957 { 3958 if (_dolog) { 3959 C = Compile::current(); 3960 _log = C->log(); 3961 } else { 3962 C = NULL; 3963 _log = NULL; 3964 } 3965 if (_log != NULL) { 3966 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3967 _log->stamp(); 3968 _log->end_head(); 3969 } 3970 } 3971 3972 Compile::TracePhase::~TracePhase() { 3973 3974 C = Compile::current(); 3975 if (_dolog) { 3976 _log = C->log(); 3977 } else { 3978 _log = NULL; 3979 } 3980 3981 #ifdef ASSERT 3982 if (PrintIdealNodeCount) { 3983 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3984 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3985 } 3986 3987 if (VerifyIdealNodeCount) { 3988 Compile::current()->print_missing_nodes(); 3989 } 3990 #endif 3991 3992 if (_log != NULL) { 3993 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3994 } 3995 } 3996 3997 //============================================================================= 3998 // Two Constant's are equal when the type and the value are equal. 3999 bool Compile::Constant::operator==(const Constant& other) { 4000 if (type() != other.type() ) return false; 4001 if (can_be_reused() != other.can_be_reused()) return false; 4002 // For floating point values we compare the bit pattern. 4003 switch (type()) { 4004 case T_INT: 4005 case T_FLOAT: return (_v._value.i == other._v._value.i); 4006 case T_LONG: 4007 case T_DOUBLE: return (_v._value.j == other._v._value.j); 4008 case T_OBJECT: 4009 case T_ADDRESS: return (_v._value.l == other._v._value.l); 4010 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 4011 case T_METADATA: return (_v._metadata == other._v._metadata); 4012 default: ShouldNotReachHere(); return false; 4013 } 4014 } 4015 4016 static int type_to_size_in_bytes(BasicType t) { 4017 switch (t) { 4018 case T_INT: return sizeof(jint ); 4019 case T_LONG: return sizeof(jlong ); 4020 case T_FLOAT: return sizeof(jfloat ); 4021 case T_DOUBLE: return sizeof(jdouble); 4022 case T_METADATA: return sizeof(Metadata*); 4023 // We use T_VOID as marker for jump-table entries (labels) which 4024 // need an internal word relocation. 4025 case T_VOID: 4026 case T_ADDRESS: 4027 case T_OBJECT: return sizeof(jobject); 4028 default: 4029 ShouldNotReachHere(); 4030 return -1; 4031 } 4032 } 4033 4034 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 4035 // sort descending 4036 if (a->freq() > b->freq()) return -1; 4037 if (a->freq() < b->freq()) return 1; 4038 return 0; 4039 } 4040 4041 void Compile::ConstantTable::calculate_offsets_and_size() { 4042 // First, sort the array by frequencies. 4043 _constants.sort(qsort_comparator); 4044 4045 #ifdef ASSERT 4046 // Make sure all jump-table entries were sorted to the end of the 4047 // array (they have a negative frequency). 4048 bool found_void = false; 4049 for (int i = 0; i < _constants.length(); i++) { 4050 Constant con = _constants.at(i); 4051 if (con.type() == T_VOID) 4052 found_void = true; // jump-tables 4053 else 4054 assert(!found_void, "wrong sorting"); 4055 } 4056 #endif 4057 4058 int offset = 0; 4059 for (int i = 0; i < _constants.length(); i++) { 4060 Constant* con = _constants.adr_at(i); 4061 4062 // Align offset for type. 4063 int typesize = type_to_size_in_bytes(con->type()); 4064 offset = align_up(offset, typesize); 4065 con->set_offset(offset); // set constant's offset 4066 4067 if (con->type() == T_VOID) { 4068 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4069 offset = offset + typesize * n->outcnt(); // expand jump-table 4070 } else { 4071 offset = offset + typesize; 4072 } 4073 } 4074 4075 // Align size up to the next section start (which is insts; see 4076 // CodeBuffer::align_at_start). 4077 assert(_size == -1, "already set?"); 4078 _size = align_up(offset, (int)CodeEntryAlignment); 4079 } 4080 4081 void Compile::ConstantTable::emit(CodeBuffer& cb) { 4082 MacroAssembler _masm(&cb); 4083 for (int i = 0; i < _constants.length(); i++) { 4084 Constant con = _constants.at(i); 4085 address constant_addr = NULL; 4086 switch (con.type()) { 4087 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4088 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4089 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4090 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4091 case T_OBJECT: { 4092 jobject obj = con.get_jobject(); 4093 int oop_index = _masm.oop_recorder()->find_index(obj); 4094 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4095 break; 4096 } 4097 case T_ADDRESS: { 4098 address addr = (address) con.get_jobject(); 4099 constant_addr = _masm.address_constant(addr); 4100 break; 4101 } 4102 // We use T_VOID as marker for jump-table entries (labels) which 4103 // need an internal word relocation. 4104 case T_VOID: { 4105 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4106 // Fill the jump-table with a dummy word. The real value is 4107 // filled in later in fill_jump_table. 4108 address dummy = (address) n; 4109 constant_addr = _masm.address_constant(dummy); 4110 // Expand jump-table 4111 for (uint i = 1; i < n->outcnt(); i++) { 4112 address temp_addr = _masm.address_constant(dummy + i); 4113 assert(temp_addr, "consts section too small"); 4114 } 4115 break; 4116 } 4117 case T_METADATA: { 4118 Metadata* obj = con.get_metadata(); 4119 int metadata_index = _masm.oop_recorder()->find_index(obj); 4120 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4121 break; 4122 } 4123 default: ShouldNotReachHere(); 4124 } 4125 assert(constant_addr, "consts section too small"); 4126 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4127 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4128 } 4129 } 4130 4131 int Compile::ConstantTable::find_offset(Constant& con) const { 4132 int idx = _constants.find(con); 4133 guarantee(idx != -1, "constant must be in constant table"); 4134 int offset = _constants.at(idx).offset(); 4135 guarantee(offset != -1, "constant table not emitted yet?"); 4136 return offset; 4137 } 4138 4139 void Compile::ConstantTable::add(Constant& con) { 4140 if (con.can_be_reused()) { 4141 int idx = _constants.find(con); 4142 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4143 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4144 return; 4145 } 4146 } 4147 (void) _constants.append(con); 4148 } 4149 4150 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4151 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4152 Constant con(type, value, b->_freq); 4153 add(con); 4154 return con; 4155 } 4156 4157 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4158 Constant con(metadata); 4159 add(con); 4160 return con; 4161 } 4162 4163 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4164 jvalue value; 4165 BasicType type = oper->type()->basic_type(); 4166 switch (type) { 4167 case T_LONG: value.j = oper->constantL(); break; 4168 case T_FLOAT: value.f = oper->constantF(); break; 4169 case T_DOUBLE: value.d = oper->constantD(); break; 4170 case T_OBJECT: 4171 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4172 case T_METADATA: return add((Metadata*)oper->constant()); break; 4173 default: guarantee(false, "unhandled type: %s", type2name(type)); 4174 } 4175 return add(n, type, value); 4176 } 4177 4178 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4179 jvalue value; 4180 // We can use the node pointer here to identify the right jump-table 4181 // as this method is called from Compile::Fill_buffer right before 4182 // the MachNodes are emitted and the jump-table is filled (means the 4183 // MachNode pointers do not change anymore). 4184 value.l = (jobject) n; 4185 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4186 add(con); 4187 return con; 4188 } 4189 4190 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4191 // If called from Compile::scratch_emit_size do nothing. 4192 if (Compile::current()->in_scratch_emit_size()) return; 4193 4194 assert(labels.is_nonempty(), "must be"); 4195 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4196 4197 // Since MachConstantNode::constant_offset() also contains 4198 // table_base_offset() we need to subtract the table_base_offset() 4199 // to get the plain offset into the constant table. 4200 int offset = n->constant_offset() - table_base_offset(); 4201 4202 MacroAssembler _masm(&cb); 4203 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4204 4205 for (uint i = 0; i < n->outcnt(); i++) { 4206 address* constant_addr = &jump_table_base[i]; 4207 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)); 4208 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4209 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4210 } 4211 } 4212 4213 //----------------------------static_subtype_check----------------------------- 4214 // Shortcut important common cases when superklass is exact: 4215 // (0) superklass is java.lang.Object (can occur in reflective code) 4216 // (1) subklass is already limited to a subtype of superklass => always ok 4217 // (2) subklass does not overlap with superklass => always fail 4218 // (3) superklass has NO subtypes and we can check with a simple compare. 4219 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4220 if (StressReflectiveCode || superk == NULL || subk == NULL) { 4221 return SSC_full_test; // Let caller generate the general case. 4222 } 4223 4224 if (!EnableValhalla && superk == env()->Object_klass()) { 4225 return SSC_always_true; // (0) this test cannot fail 4226 } 4227 4228 ciType* superelem = superk; 4229 if (superelem->is_array_klass()) 4230 superelem = superelem->as_array_klass()->base_element_type(); 4231 4232 if (!subk->is_interface()) { // cannot trust static interface types yet 4233 if (subk->is_subtype_of(superk)) { 4234 return SSC_always_true; // (1) false path dead; no dynamic test needed 4235 } 4236 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4237 !superk->is_subtype_of(subk)) { 4238 return SSC_always_false; 4239 } 4240 } 4241 4242 // If casting to an instance klass, it must have no subtypes 4243 if (superk->is_interface()) { 4244 // Cannot trust interfaces yet. 4245 // %%% S.B. superk->nof_implementors() == 1 4246 } else if (superelem->is_instance_klass()) { 4247 ciInstanceKlass* ik = superelem->as_instance_klass(); 4248 if (!ik->has_subklass() && !ik->is_interface()) { 4249 if (!ik->is_final()) { 4250 // Add a dependency if there is a chance of a later subclass. 4251 dependencies()->assert_leaf_type(ik); 4252 } 4253 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4254 } 4255 } else { 4256 // A primitive array type has no subtypes. 4257 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4258 } 4259 4260 return SSC_full_test; 4261 } 4262 4263 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4264 #ifdef _LP64 4265 // The scaled index operand to AddP must be a clean 64-bit value. 4266 // Java allows a 32-bit int to be incremented to a negative 4267 // value, which appears in a 64-bit register as a large 4268 // positive number. Using that large positive number as an 4269 // operand in pointer arithmetic has bad consequences. 4270 // On the other hand, 32-bit overflow is rare, and the possibility 4271 // can often be excluded, if we annotate the ConvI2L node with 4272 // a type assertion that its value is known to be a small positive 4273 // number. (The prior range check has ensured this.) 4274 // This assertion is used by ConvI2LNode::Ideal. 4275 int index_max = max_jint - 1; // array size is max_jint, index is one less 4276 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4277 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4278 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4279 #endif 4280 return idx; 4281 } 4282 4283 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4284 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4285 if (ctrl != NULL) { 4286 // Express control dependency by a CastII node with a narrow type. 4287 value = new CastIINode(value, itype, false, true /* range check dependency */); 4288 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4289 // node from floating above the range check during loop optimizations. Otherwise, the 4290 // ConvI2L node may be eliminated independently of the range check, causing the data path 4291 // to become TOP while the control path is still there (although it's unreachable). 4292 value->set_req(0, ctrl); 4293 // Save CastII node to remove it after loop optimizations. 4294 phase->C->add_range_check_cast(value); 4295 value = phase->transform(value); 4296 } 4297 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4298 return phase->transform(new ConvI2LNode(value, ltype)); 4299 } 4300 4301 // The message about the current inlining is accumulated in 4302 // _print_inlining_stream and transfered into the _print_inlining_list 4303 // once we know whether inlining succeeds or not. For regular 4304 // inlining, messages are appended to the buffer pointed by 4305 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4306 // a new buffer is added after _print_inlining_idx in the list. This 4307 // way we can update the inlining message for late inlining call site 4308 // when the inlining is attempted again. 4309 void Compile::print_inlining_init() { 4310 if (print_inlining() || print_intrinsics()) { 4311 _print_inlining_stream = new stringStream(); 4312 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4313 } 4314 } 4315 4316 void Compile::print_inlining_reinit() { 4317 if (print_inlining() || print_intrinsics()) { 4318 // Re allocate buffer when we change ResourceMark 4319 _print_inlining_stream = new stringStream(); 4320 } 4321 } 4322 4323 void Compile::print_inlining_reset() { 4324 _print_inlining_stream->reset(); 4325 } 4326 4327 void Compile::print_inlining_commit() { 4328 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4329 // Transfer the message from _print_inlining_stream to the current 4330 // _print_inlining_list buffer and clear _print_inlining_stream. 4331 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4332 print_inlining_reset(); 4333 } 4334 4335 void Compile::print_inlining_push() { 4336 // Add new buffer to the _print_inlining_list at current position 4337 _print_inlining_idx++; 4338 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4339 } 4340 4341 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4342 return _print_inlining_list->at(_print_inlining_idx); 4343 } 4344 4345 void Compile::print_inlining_update(CallGenerator* cg) { 4346 if (print_inlining() || print_intrinsics()) { 4347 if (!cg->is_late_inline()) { 4348 if (print_inlining_current().cg() != NULL) { 4349 print_inlining_push(); 4350 } 4351 print_inlining_commit(); 4352 } else { 4353 if (print_inlining_current().cg() != cg && 4354 (print_inlining_current().cg() != NULL || 4355 print_inlining_current().ss()->size() != 0)) { 4356 print_inlining_push(); 4357 } 4358 print_inlining_commit(); 4359 print_inlining_current().set_cg(cg); 4360 } 4361 } 4362 } 4363 4364 void Compile::print_inlining_move_to(CallGenerator* cg) { 4365 // We resume inlining at a late inlining call site. Locate the 4366 // corresponding inlining buffer so that we can update it. 4367 if (print_inlining()) { 4368 for (int i = 0; i < _print_inlining_list->length(); i++) { 4369 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4370 _print_inlining_idx = i; 4371 return; 4372 } 4373 } 4374 ShouldNotReachHere(); 4375 } 4376 } 4377 4378 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4379 if (print_inlining()) { 4380 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4381 assert(print_inlining_current().cg() == cg, "wrong entry"); 4382 // replace message with new message 4383 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4384 print_inlining_commit(); 4385 print_inlining_current().set_cg(cg); 4386 } 4387 } 4388 4389 void Compile::print_inlining_assert_ready() { 4390 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4391 } 4392 4393 void Compile::process_print_inlining() { 4394 bool do_print_inlining = print_inlining() || print_intrinsics(); 4395 if (do_print_inlining || log() != NULL) { 4396 // Print inlining message for candidates that we couldn't inline 4397 // for lack of space 4398 for (int i = 0; i < _late_inlines.length(); i++) { 4399 CallGenerator* cg = _late_inlines.at(i); 4400 if (!cg->is_mh_late_inline()) { 4401 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4402 if (do_print_inlining) { 4403 cg->print_inlining_late(msg); 4404 } 4405 log_late_inline_failure(cg, msg); 4406 } 4407 } 4408 } 4409 if (do_print_inlining) { 4410 ResourceMark rm; 4411 stringStream ss; 4412 for (int i = 0; i < _print_inlining_list->length(); i++) { 4413 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4414 } 4415 size_t end = ss.size(); 4416 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4417 strncpy(_print_inlining_output, ss.base(), end+1); 4418 _print_inlining_output[end] = 0; 4419 } 4420 } 4421 4422 void Compile::dump_print_inlining() { 4423 if (_print_inlining_output != NULL) { 4424 tty->print_raw(_print_inlining_output); 4425 } 4426 } 4427 4428 void Compile::log_late_inline(CallGenerator* cg) { 4429 if (log() != NULL) { 4430 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4431 cg->unique_id()); 4432 JVMState* p = cg->call_node()->jvms(); 4433 while (p != NULL) { 4434 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4435 p = p->caller(); 4436 } 4437 log()->tail("late_inline"); 4438 } 4439 } 4440 4441 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4442 log_late_inline(cg); 4443 if (log() != NULL) { 4444 log()->inline_fail(msg); 4445 } 4446 } 4447 4448 void Compile::log_inline_id(CallGenerator* cg) { 4449 if (log() != NULL) { 4450 // The LogCompilation tool needs a unique way to identify late 4451 // inline call sites. This id must be unique for this call site in 4452 // this compilation. Try to have it unique across compilations as 4453 // well because it can be convenient when grepping through the log 4454 // file. 4455 // Distinguish OSR compilations from others in case CICountOSR is 4456 // on. 4457 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4458 cg->set_unique_id(id); 4459 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4460 } 4461 } 4462 4463 void Compile::log_inline_failure(const char* msg) { 4464 if (C->log() != NULL) { 4465 C->log()->inline_fail(msg); 4466 } 4467 } 4468 4469 4470 // Dump inlining replay data to the stream. 4471 // Don't change thread state and acquire any locks. 4472 void Compile::dump_inline_data(outputStream* out) { 4473 InlineTree* inl_tree = ilt(); 4474 if (inl_tree != NULL) { 4475 out->print(" inline %d", inl_tree->count()); 4476 inl_tree->dump_replay_data(out); 4477 } 4478 } 4479 4480 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4481 if (n1->Opcode() < n2->Opcode()) return -1; 4482 else if (n1->Opcode() > n2->Opcode()) return 1; 4483 4484 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4485 for (uint i = 1; i < n1->req(); i++) { 4486 if (n1->in(i) < n2->in(i)) return -1; 4487 else if (n1->in(i) > n2->in(i)) return 1; 4488 } 4489 4490 return 0; 4491 } 4492 4493 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4494 Node* n1 = *n1p; 4495 Node* n2 = *n2p; 4496 4497 return cmp_expensive_nodes(n1, n2); 4498 } 4499 4500 void Compile::sort_expensive_nodes() { 4501 if (!expensive_nodes_sorted()) { 4502 _expensive_nodes->sort(cmp_expensive_nodes); 4503 } 4504 } 4505 4506 bool Compile::expensive_nodes_sorted() const { 4507 for (int i = 1; i < _expensive_nodes->length(); i++) { 4508 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4509 return false; 4510 } 4511 } 4512 return true; 4513 } 4514 4515 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4516 if (_expensive_nodes->length() == 0) { 4517 return false; 4518 } 4519 4520 assert(OptimizeExpensiveOps, "optimization off?"); 4521 4522 // Take this opportunity to remove dead nodes from the list 4523 int j = 0; 4524 for (int i = 0; i < _expensive_nodes->length(); i++) { 4525 Node* n = _expensive_nodes->at(i); 4526 if (!n->is_unreachable(igvn)) { 4527 assert(n->is_expensive(), "should be expensive"); 4528 _expensive_nodes->at_put(j, n); 4529 j++; 4530 } 4531 } 4532 _expensive_nodes->trunc_to(j); 4533 4534 // Then sort the list so that similar nodes are next to each other 4535 // and check for at least two nodes of identical kind with same data 4536 // inputs. 4537 sort_expensive_nodes(); 4538 4539 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4540 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4541 return true; 4542 } 4543 } 4544 4545 return false; 4546 } 4547 4548 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4549 if (_expensive_nodes->length() == 0) { 4550 return; 4551 } 4552 4553 assert(OptimizeExpensiveOps, "optimization off?"); 4554 4555 // Sort to bring similar nodes next to each other and clear the 4556 // control input of nodes for which there's only a single copy. 4557 sort_expensive_nodes(); 4558 4559 int j = 0; 4560 int identical = 0; 4561 int i = 0; 4562 bool modified = false; 4563 for (; i < _expensive_nodes->length()-1; i++) { 4564 assert(j <= i, "can't write beyond current index"); 4565 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4566 identical++; 4567 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4568 continue; 4569 } 4570 if (identical > 0) { 4571 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4572 identical = 0; 4573 } else { 4574 Node* n = _expensive_nodes->at(i); 4575 igvn.replace_input_of(n, 0, NULL); 4576 igvn.hash_insert(n); 4577 modified = true; 4578 } 4579 } 4580 if (identical > 0) { 4581 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4582 } else if (_expensive_nodes->length() >= 1) { 4583 Node* n = _expensive_nodes->at(i); 4584 igvn.replace_input_of(n, 0, NULL); 4585 igvn.hash_insert(n); 4586 modified = true; 4587 } 4588 _expensive_nodes->trunc_to(j); 4589 if (modified) { 4590 igvn.optimize(); 4591 } 4592 } 4593 4594 void Compile::add_expensive_node(Node * n) { 4595 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4596 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4597 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4598 if (OptimizeExpensiveOps) { 4599 _expensive_nodes->append(n); 4600 } else { 4601 // Clear control input and let IGVN optimize expensive nodes if 4602 // OptimizeExpensiveOps is off. 4603 n->set_req(0, NULL); 4604 } 4605 } 4606 4607 /** 4608 * Remove the speculative part of types and clean up the graph 4609 */ 4610 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4611 if (UseTypeSpeculation) { 4612 Unique_Node_List worklist; 4613 worklist.push(root()); 4614 int modified = 0; 4615 // Go over all type nodes that carry a speculative type, drop the 4616 // speculative part of the type and enqueue the node for an igvn 4617 // which may optimize it out. 4618 for (uint next = 0; next < worklist.size(); ++next) { 4619 Node *n = worklist.at(next); 4620 if (n->is_Type()) { 4621 TypeNode* tn = n->as_Type(); 4622 const Type* t = tn->type(); 4623 const Type* t_no_spec = t->remove_speculative(); 4624 if (t_no_spec != t) { 4625 bool in_hash = igvn.hash_delete(n); 4626 assert(in_hash, "node should be in igvn hash table"); 4627 tn->set_type(t_no_spec); 4628 igvn.hash_insert(n); 4629 igvn._worklist.push(n); // give it a chance to go away 4630 modified++; 4631 } 4632 } 4633 uint max = n->len(); 4634 for( uint i = 0; i < max; ++i ) { 4635 Node *m = n->in(i); 4636 if (not_a_node(m)) continue; 4637 worklist.push(m); 4638 } 4639 } 4640 // Drop the speculative part of all types in the igvn's type table 4641 igvn.remove_speculative_types(); 4642 if (modified > 0) { 4643 igvn.optimize(); 4644 } 4645 #ifdef ASSERT 4646 // Verify that after the IGVN is over no speculative type has resurfaced 4647 worklist.clear(); 4648 worklist.push(root()); 4649 for (uint next = 0; next < worklist.size(); ++next) { 4650 Node *n = worklist.at(next); 4651 const Type* t = igvn.type_or_null(n); 4652 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4653 if (n->is_Type()) { 4654 t = n->as_Type()->type(); 4655 assert(t == t->remove_speculative(), "no more speculative types"); 4656 } 4657 uint max = n->len(); 4658 for( uint i = 0; i < max; ++i ) { 4659 Node *m = n->in(i); 4660 if (not_a_node(m)) continue; 4661 worklist.push(m); 4662 } 4663 } 4664 igvn.check_no_speculative_types(); 4665 #endif 4666 } 4667 } 4668 4669 Node* Compile::load_is_value_bit(PhaseGVN* phase, Node* oop) { 4670 // Load the klass pointer and check if it's odd, i.e., if it defines a value type 4671 // is_value = (klass & oop_metadata_valuetype_mask) >> LogKlassAlignmentInBytes 4672 Node* k_adr = phase->transform(new AddPNode(oop, oop, phase->MakeConX(oopDesc::klass_offset_in_bytes()))); 4673 Node* klass = NULL; 4674 if (UseCompressedClassPointers) { 4675 klass = phase->transform(new LoadNKlassNode(NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT->make_narrowklass(), MemNode::unordered)); 4676 } else { 4677 klass = phase->transform(new LoadKlassNode(NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT, MemNode::unordered)); 4678 } 4679 const int mask = Universe::oop_metadata_valuetype_mask(); 4680 Node* is_value = phase->transform(new CastP2XNode(NULL, klass)); 4681 is_value = phase->transform(new AndXNode(is_value, phase->MakeConX(mask))); 4682 // Check if a shift is required for perturbation to affect aligned bits of oop 4683 if (mask == KlassPtrValueTypeMask && ObjectAlignmentInBytes <= KlassAlignmentInBytes) { 4684 assert((mask >> LogKlassAlignmentInBytes) == 1, "invalid shift"); 4685 is_value = phase->transform(new URShiftXNode(is_value, phase->intcon(LogKlassAlignmentInBytes))); 4686 } else { 4687 assert(mask < ObjectAlignmentInBytes, "invalid mask"); 4688 } 4689 return is_value; 4690 } 4691 4692 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) { 4693 const TypeInstPtr* ta = phase->type(a)->isa_instptr(); 4694 const TypeInstPtr* tb = phase->type(b)->isa_instptr(); 4695 if (!EnableValhalla || ta == NULL || tb == NULL || 4696 ta->is_zero_type() || tb->is_zero_type() || 4697 !ta->can_be_value_type() || !tb->can_be_value_type()) { 4698 // Use old acmp if one operand is null or not a value type 4699 return new CmpPNode(a, b); 4700 } else if (ta->is_valuetypeptr() || tb->is_valuetypeptr()) { 4701 // We know that one operand is a value type. Therefore, 4702 // new acmp will only return true if both operands are NULL. 4703 // Check if both operands are null by or'ing the oops. 4704 a = phase->transform(new CastP2XNode(NULL, a)); 4705 b = phase->transform(new CastP2XNode(NULL, b)); 4706 a = phase->transform(new OrXNode(a, b)); 4707 return new CmpXNode(a, phase->MakeConX(0)); 4708 } 4709 // Use new acmp 4710 return NULL; 4711 } 4712 4713 // Auxiliary method to support randomized stressing/fuzzing. 4714 // 4715 // This method can be called the arbitrary number of times, with current count 4716 // as the argument. The logic allows selecting a single candidate from the 4717 // running list of candidates as follows: 4718 // int count = 0; 4719 // Cand* selected = null; 4720 // while(cand = cand->next()) { 4721 // if (randomized_select(++count)) { 4722 // selected = cand; 4723 // } 4724 // } 4725 // 4726 // Including count equalizes the chances any candidate is "selected". 4727 // This is useful when we don't have the complete list of candidates to choose 4728 // from uniformly. In this case, we need to adjust the randomicity of the 4729 // selection, or else we will end up biasing the selection towards the latter 4730 // candidates. 4731 // 4732 // Quick back-envelope calculation shows that for the list of n candidates 4733 // the equal probability for the candidate to persist as "best" can be 4734 // achieved by replacing it with "next" k-th candidate with the probability 4735 // of 1/k. It can be easily shown that by the end of the run, the 4736 // probability for any candidate is converged to 1/n, thus giving the 4737 // uniform distribution among all the candidates. 4738 // 4739 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4740 #define RANDOMIZED_DOMAIN_POW 29 4741 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4742 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4743 bool Compile::randomized_select(int count) { 4744 assert(count > 0, "only positive"); 4745 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4746 } 4747 4748 CloneMap& Compile::clone_map() { return _clone_map; } 4749 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4750 4751 void NodeCloneInfo::dump() const { 4752 tty->print(" {%d:%d} ", idx(), gen()); 4753 } 4754 4755 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4756 uint64_t val = value(old->_idx); 4757 NodeCloneInfo cio(val); 4758 assert(val != 0, "old node should be in the map"); 4759 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4760 insert(nnn->_idx, cin.get()); 4761 #ifndef PRODUCT 4762 if (is_debug()) { 4763 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4764 } 4765 #endif 4766 } 4767 4768 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4769 NodeCloneInfo cio(value(old->_idx)); 4770 if (cio.get() == 0) { 4771 cio.set(old->_idx, 0); 4772 insert(old->_idx, cio.get()); 4773 #ifndef PRODUCT 4774 if (is_debug()) { 4775 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4776 } 4777 #endif 4778 } 4779 clone(old, nnn, gen); 4780 } 4781 4782 int CloneMap::max_gen() const { 4783 int g = 0; 4784 DictI di(_dict); 4785 for(; di.test(); ++di) { 4786 int t = gen(di._key); 4787 if (g < t) { 4788 g = t; 4789 #ifndef PRODUCT 4790 if (is_debug()) { 4791 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4792 } 4793 #endif 4794 } 4795 } 4796 return g; 4797 } 4798 4799 void CloneMap::dump(node_idx_t key) const { 4800 uint64_t val = value(key); 4801 if (val != 0) { 4802 NodeCloneInfo ni(val); 4803 ni.dump(); 4804 } 4805 }