1 /* 2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "gc/shared/barrierSet.hpp" 37 #include "gc/shared/c2/barrierSetC2.hpp" 38 #include "memory/resourceArea.hpp" 39 #include "opto/addnode.hpp" 40 #include "opto/block.hpp" 41 #include "opto/c2compiler.hpp" 42 #include "opto/callGenerator.hpp" 43 #include "opto/callnode.hpp" 44 #include "opto/castnode.hpp" 45 #include "opto/cfgnode.hpp" 46 #include "opto/chaitin.hpp" 47 #include "opto/compile.hpp" 48 #include "opto/connode.hpp" 49 #include "opto/convertnode.hpp" 50 #include "opto/divnode.hpp" 51 #include "opto/escape.hpp" 52 #include "opto/idealGraphPrinter.hpp" 53 #include "opto/loopnode.hpp" 54 #include "opto/machnode.hpp" 55 #include "opto/macro.hpp" 56 #include "opto/matcher.hpp" 57 #include "opto/mathexactnode.hpp" 58 #include "opto/memnode.hpp" 59 #include "opto/mulnode.hpp" 60 #include "opto/narrowptrnode.hpp" 61 #include "opto/node.hpp" 62 #include "opto/opcodes.hpp" 63 #include "opto/output.hpp" 64 #include "opto/parse.hpp" 65 #include "opto/phaseX.hpp" 66 #include "opto/rootnode.hpp" 67 #include "opto/runtime.hpp" 68 #include "opto/stringopts.hpp" 69 #include "opto/type.hpp" 70 #include "opto/vectornode.hpp" 71 #include "runtime/arguments.hpp" 72 #include "runtime/sharedRuntime.hpp" 73 #include "runtime/signature.hpp" 74 #include "runtime/stubRoutines.hpp" 75 #include "runtime/timer.hpp" 76 #include "utilities/align.hpp" 77 #include "utilities/copy.hpp" 78 #include "utilities/macros.hpp" 79 #if INCLUDE_G1GC 80 #include "gc/g1/g1ThreadLocalData.hpp" 81 #endif // INCLUDE_G1GC 82 #if INCLUDE_ZGC 83 #include "gc/z/c2/zBarrierSetC2.hpp" 84 #endif 85 #if INCLUDE_SHENANDOAHGC 86 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp" 87 #endif 88 89 90 // -------------------- Compile::mach_constant_base_node ----------------------- 91 // Constant table base node singleton. 92 MachConstantBaseNode* Compile::mach_constant_base_node() { 93 if (_mach_constant_base_node == NULL) { 94 _mach_constant_base_node = new MachConstantBaseNode(); 95 _mach_constant_base_node->add_req(C->root()); 96 } 97 return _mach_constant_base_node; 98 } 99 100 101 /// Support for intrinsics. 102 103 // Return the index at which m must be inserted (or already exists). 104 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 105 class IntrinsicDescPair { 106 private: 107 ciMethod* _m; 108 bool _is_virtual; 109 public: 110 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 111 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 112 ciMethod* m= elt->method(); 113 ciMethod* key_m = key->_m; 114 if (key_m < m) return -1; 115 else if (key_m > m) return 1; 116 else { 117 bool is_virtual = elt->is_virtual(); 118 bool key_virtual = key->_is_virtual; 119 if (key_virtual < is_virtual) return -1; 120 else if (key_virtual > is_virtual) return 1; 121 else return 0; 122 } 123 } 124 }; 125 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 126 #ifdef ASSERT 127 for (int i = 1; i < _intrinsics->length(); i++) { 128 CallGenerator* cg1 = _intrinsics->at(i-1); 129 CallGenerator* cg2 = _intrinsics->at(i); 130 assert(cg1->method() != cg2->method() 131 ? cg1->method() < cg2->method() 132 : cg1->is_virtual() < cg2->is_virtual(), 133 "compiler intrinsics list must stay sorted"); 134 } 135 #endif 136 IntrinsicDescPair pair(m, is_virtual); 137 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 138 } 139 140 void Compile::register_intrinsic(CallGenerator* cg) { 141 if (_intrinsics == NULL) { 142 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 143 } 144 int len = _intrinsics->length(); 145 bool found = false; 146 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 147 assert(!found, "registering twice"); 148 _intrinsics->insert_before(index, cg); 149 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 150 } 151 152 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 153 assert(m->is_loaded(), "don't try this on unloaded methods"); 154 if (_intrinsics != NULL) { 155 bool found = false; 156 int index = intrinsic_insertion_index(m, is_virtual, found); 157 if (found) { 158 return _intrinsics->at(index); 159 } 160 } 161 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 162 if (m->intrinsic_id() != vmIntrinsics::_none && 163 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 164 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 165 if (cg != NULL) { 166 // Save it for next time: 167 register_intrinsic(cg); 168 return cg; 169 } else { 170 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 171 } 172 } 173 return NULL; 174 } 175 176 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 177 // in library_call.cpp. 178 179 180 #ifndef PRODUCT 181 // statistics gathering... 182 183 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 184 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 185 186 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 187 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 188 int oflags = _intrinsic_hist_flags[id]; 189 assert(flags != 0, "what happened?"); 190 if (is_virtual) { 191 flags |= _intrinsic_virtual; 192 } 193 bool changed = (flags != oflags); 194 if ((flags & _intrinsic_worked) != 0) { 195 juint count = (_intrinsic_hist_count[id] += 1); 196 if (count == 1) { 197 changed = true; // first time 198 } 199 // increment the overall count also: 200 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 201 } 202 if (changed) { 203 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 204 // Something changed about the intrinsic's virtuality. 205 if ((flags & _intrinsic_virtual) != 0) { 206 // This is the first use of this intrinsic as a virtual call. 207 if (oflags != 0) { 208 // We already saw it as a non-virtual, so note both cases. 209 flags |= _intrinsic_both; 210 } 211 } else if ((oflags & _intrinsic_both) == 0) { 212 // This is the first use of this intrinsic as a non-virtual 213 flags |= _intrinsic_both; 214 } 215 } 216 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 217 } 218 // update the overall flags also: 219 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 220 return changed; 221 } 222 223 static char* format_flags(int flags, char* buf) { 224 buf[0] = 0; 225 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 226 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 227 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 228 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 229 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 230 if (buf[0] == 0) strcat(buf, ","); 231 assert(buf[0] == ',', "must be"); 232 return &buf[1]; 233 } 234 235 void Compile::print_intrinsic_statistics() { 236 char flagsbuf[100]; 237 ttyLocker ttyl; 238 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 239 tty->print_cr("Compiler intrinsic usage:"); 240 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 241 if (total == 0) total = 1; // avoid div0 in case of no successes 242 #define PRINT_STAT_LINE(name, c, f) \ 243 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 244 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 245 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 246 int flags = _intrinsic_hist_flags[id]; 247 juint count = _intrinsic_hist_count[id]; 248 if ((flags | count) != 0) { 249 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 250 } 251 } 252 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 253 if (xtty != NULL) xtty->tail("statistics"); 254 } 255 256 void Compile::print_statistics() { 257 { ttyLocker ttyl; 258 if (xtty != NULL) xtty->head("statistics type='opto'"); 259 Parse::print_statistics(); 260 PhaseCCP::print_statistics(); 261 PhaseRegAlloc::print_statistics(); 262 Scheduling::print_statistics(); 263 PhasePeephole::print_statistics(); 264 PhaseIdealLoop::print_statistics(); 265 if (xtty != NULL) xtty->tail("statistics"); 266 } 267 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 268 // put this under its own <statistics> element. 269 print_intrinsic_statistics(); 270 } 271 } 272 #endif //PRODUCT 273 274 // Support for bundling info 275 Bundle* Compile::node_bundling(const Node *n) { 276 assert(valid_bundle_info(n), "oob"); 277 return &_node_bundling_base[n->_idx]; 278 } 279 280 bool Compile::valid_bundle_info(const Node *n) { 281 return (_node_bundling_limit > n->_idx); 282 } 283 284 285 void Compile::gvn_replace_by(Node* n, Node* nn) { 286 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 287 Node* use = n->last_out(i); 288 bool is_in_table = initial_gvn()->hash_delete(use); 289 uint uses_found = 0; 290 for (uint j = 0; j < use->len(); j++) { 291 if (use->in(j) == n) { 292 if (j < use->req()) 293 use->set_req(j, nn); 294 else 295 use->set_prec(j, nn); 296 uses_found++; 297 } 298 } 299 if (is_in_table) { 300 // reinsert into table 301 initial_gvn()->hash_find_insert(use); 302 } 303 record_for_igvn(use); 304 i -= uses_found; // we deleted 1 or more copies of this edge 305 } 306 } 307 308 309 static inline bool not_a_node(const Node* n) { 310 if (n == NULL) return true; 311 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 312 if (*(address*)n == badAddress) return true; // kill by Node::destruct 313 return false; 314 } 315 316 // Identify all nodes that are reachable from below, useful. 317 // Use breadth-first pass that records state in a Unique_Node_List, 318 // recursive traversal is slower. 319 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 320 int estimated_worklist_size = live_nodes(); 321 useful.map( estimated_worklist_size, NULL ); // preallocate space 322 323 // Initialize worklist 324 if (root() != NULL) { useful.push(root()); } 325 // If 'top' is cached, declare it useful to preserve cached node 326 if( cached_top_node() ) { useful.push(cached_top_node()); } 327 328 // Push all useful nodes onto the list, breadthfirst 329 for( uint next = 0; next < useful.size(); ++next ) { 330 assert( next < unique(), "Unique useful nodes < total nodes"); 331 Node *n = useful.at(next); 332 uint max = n->len(); 333 for( uint i = 0; i < max; ++i ) { 334 Node *m = n->in(i); 335 if (not_a_node(m)) continue; 336 useful.push(m); 337 } 338 } 339 } 340 341 // Update dead_node_list with any missing dead nodes using useful 342 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 343 void Compile::update_dead_node_list(Unique_Node_List &useful) { 344 uint max_idx = unique(); 345 VectorSet& useful_node_set = useful.member_set(); 346 347 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 348 // If node with index node_idx is not in useful set, 349 // mark it as dead in dead node list. 350 if (! useful_node_set.test(node_idx) ) { 351 record_dead_node(node_idx); 352 } 353 } 354 } 355 356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 357 int shift = 0; 358 for (int i = 0; i < inlines->length(); i++) { 359 CallGenerator* cg = inlines->at(i); 360 CallNode* call = cg->call_node(); 361 if (shift > 0) { 362 inlines->at_put(i-shift, cg); 363 } 364 if (!useful.member(call)) { 365 shift++; 366 } 367 } 368 inlines->trunc_to(inlines->length()-shift); 369 } 370 371 // Disconnect all useless nodes by disconnecting those at the boundary. 372 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 373 uint next = 0; 374 while (next < useful.size()) { 375 Node *n = useful.at(next++); 376 if (n->is_SafePoint()) { 377 // We're done with a parsing phase. Replaced nodes are not valid 378 // beyond that point. 379 n->as_SafePoint()->delete_replaced_nodes(); 380 } 381 // Use raw traversal of out edges since this code removes out edges 382 int max = n->outcnt(); 383 for (int j = 0; j < max; ++j) { 384 Node* child = n->raw_out(j); 385 if (! useful.member(child)) { 386 assert(!child->is_top() || child != top(), 387 "If top is cached in Compile object it is in useful list"); 388 // Only need to remove this out-edge to the useless node 389 n->raw_del_out(j); 390 --j; 391 --max; 392 } 393 } 394 if (n->outcnt() == 1 && n->has_special_unique_user()) { 395 record_for_igvn(n->unique_out()); 396 } 397 } 398 // Remove useless macro and predicate opaq nodes 399 for (int i = C->macro_count()-1; i >= 0; i--) { 400 Node* n = C->macro_node(i); 401 if (!useful.member(n)) { 402 remove_macro_node(n); 403 } 404 } 405 // Remove useless CastII nodes with range check dependency 406 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 407 Node* cast = range_check_cast_node(i); 408 if (!useful.member(cast)) { 409 remove_range_check_cast(cast); 410 } 411 } 412 // Remove useless expensive nodes 413 for (int i = C->expensive_count()-1; i >= 0; i--) { 414 Node* n = C->expensive_node(i); 415 if (!useful.member(n)) { 416 remove_expensive_node(n); 417 } 418 } 419 // Remove useless Opaque4 nodes 420 for (int i = opaque4_count() - 1; i >= 0; i--) { 421 Node* opaq = opaque4_node(i); 422 if (!useful.member(opaq)) { 423 remove_opaque4_node(opaq); 424 } 425 } 426 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 427 bs->eliminate_useless_gc_barriers(useful); 428 // clean up the late inline lists 429 remove_useless_late_inlines(&_string_late_inlines, useful); 430 remove_useless_late_inlines(&_boxing_late_inlines, useful); 431 remove_useless_late_inlines(&_late_inlines, useful); 432 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 433 } 434 435 //------------------------------frame_size_in_words----------------------------- 436 // frame_slots in units of words 437 int Compile::frame_size_in_words() const { 438 // shift is 0 in LP32 and 1 in LP64 439 const int shift = (LogBytesPerWord - LogBytesPerInt); 440 int words = _frame_slots >> shift; 441 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 442 return words; 443 } 444 445 // To bang the stack of this compiled method we use the stack size 446 // that the interpreter would need in case of a deoptimization. This 447 // removes the need to bang the stack in the deoptimization blob which 448 // in turn simplifies stack overflow handling. 449 int Compile::bang_size_in_bytes() const { 450 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 451 } 452 453 // ============================================================================ 454 //------------------------------CompileWrapper--------------------------------- 455 class CompileWrapper : public StackObj { 456 Compile *const _compile; 457 public: 458 CompileWrapper(Compile* compile); 459 460 ~CompileWrapper(); 461 }; 462 463 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 464 // the Compile* pointer is stored in the current ciEnv: 465 ciEnv* env = compile->env(); 466 assert(env == ciEnv::current(), "must already be a ciEnv active"); 467 assert(env->compiler_data() == NULL, "compile already active?"); 468 env->set_compiler_data(compile); 469 assert(compile == Compile::current(), "sanity"); 470 471 compile->set_type_dict(NULL); 472 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 473 compile->clone_map().set_clone_idx(0); 474 compile->set_type_hwm(NULL); 475 compile->set_type_last_size(0); 476 compile->set_last_tf(NULL, NULL); 477 compile->set_indexSet_arena(NULL); 478 compile->set_indexSet_free_block_list(NULL); 479 compile->init_type_arena(); 480 Type::Initialize(compile); 481 _compile->set_scratch_buffer_blob(NULL); 482 _compile->begin_method(); 483 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 484 } 485 CompileWrapper::~CompileWrapper() { 486 _compile->end_method(); 487 if (_compile->scratch_buffer_blob() != NULL) 488 BufferBlob::free(_compile->scratch_buffer_blob()); 489 _compile->env()->set_compiler_data(NULL); 490 } 491 492 493 //----------------------------print_compile_messages--------------------------- 494 void Compile::print_compile_messages() { 495 #ifndef PRODUCT 496 // Check if recompiling 497 if (_subsume_loads == false && PrintOpto) { 498 // Recompiling without allowing machine instructions to subsume loads 499 tty->print_cr("*********************************************************"); 500 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 501 tty->print_cr("*********************************************************"); 502 } 503 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 504 // Recompiling without escape analysis 505 tty->print_cr("*********************************************************"); 506 tty->print_cr("** Bailout: Recompile without escape analysis **"); 507 tty->print_cr("*********************************************************"); 508 } 509 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 510 // Recompiling without boxing elimination 511 tty->print_cr("*********************************************************"); 512 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 513 tty->print_cr("*********************************************************"); 514 } 515 if (C->directive()->BreakAtCompileOption) { 516 // Open the debugger when compiling this method. 517 tty->print("### Breaking when compiling: "); 518 method()->print_short_name(); 519 tty->cr(); 520 BREAKPOINT; 521 } 522 523 if( PrintOpto ) { 524 if (is_osr_compilation()) { 525 tty->print("[OSR]%3d", _compile_id); 526 } else { 527 tty->print("%3d", _compile_id); 528 } 529 } 530 #endif 531 } 532 533 534 //-----------------------init_scratch_buffer_blob------------------------------ 535 // Construct a temporary BufferBlob and cache it for this compile. 536 void Compile::init_scratch_buffer_blob(int const_size) { 537 // If there is already a scratch buffer blob allocated and the 538 // constant section is big enough, use it. Otherwise free the 539 // current and allocate a new one. 540 BufferBlob* blob = scratch_buffer_blob(); 541 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 542 // Use the current blob. 543 } else { 544 if (blob != NULL) { 545 BufferBlob::free(blob); 546 } 547 548 ResourceMark rm; 549 _scratch_const_size = const_size; 550 int size = C2Compiler::initial_code_buffer_size(const_size); 551 blob = BufferBlob::create("Compile::scratch_buffer", size); 552 // Record the buffer blob for next time. 553 set_scratch_buffer_blob(blob); 554 // Have we run out of code space? 555 if (scratch_buffer_blob() == NULL) { 556 // Let CompilerBroker disable further compilations. 557 record_failure("Not enough space for scratch buffer in CodeCache"); 558 return; 559 } 560 } 561 562 // Initialize the relocation buffers 563 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 564 set_scratch_locs_memory(locs_buf); 565 } 566 567 568 //-----------------------scratch_emit_size------------------------------------- 569 // Helper function that computes size by emitting code 570 uint Compile::scratch_emit_size(const Node* n) { 571 // Start scratch_emit_size section. 572 set_in_scratch_emit_size(true); 573 574 // Emit into a trash buffer and count bytes emitted. 575 // This is a pretty expensive way to compute a size, 576 // but it works well enough if seldom used. 577 // All common fixed-size instructions are given a size 578 // method by the AD file. 579 // Note that the scratch buffer blob and locs memory are 580 // allocated at the beginning of the compile task, and 581 // may be shared by several calls to scratch_emit_size. 582 // The allocation of the scratch buffer blob is particularly 583 // expensive, since it has to grab the code cache lock. 584 BufferBlob* blob = this->scratch_buffer_blob(); 585 assert(blob != NULL, "Initialize BufferBlob at start"); 586 assert(blob->size() > MAX_inst_size, "sanity"); 587 relocInfo* locs_buf = scratch_locs_memory(); 588 address blob_begin = blob->content_begin(); 589 address blob_end = (address)locs_buf; 590 assert(blob->contains(blob_end), "sanity"); 591 CodeBuffer buf(blob_begin, blob_end - blob_begin); 592 buf.initialize_consts_size(_scratch_const_size); 593 buf.initialize_stubs_size(MAX_stubs_size); 594 assert(locs_buf != NULL, "sanity"); 595 int lsize = MAX_locs_size / 3; 596 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 597 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 598 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 599 // Mark as scratch buffer. 600 buf.consts()->set_scratch_emit(); 601 buf.insts()->set_scratch_emit(); 602 buf.stubs()->set_scratch_emit(); 603 604 // Do the emission. 605 606 Label fakeL; // Fake label for branch instructions. 607 Label* saveL = NULL; 608 uint save_bnum = 0; 609 bool is_branch = n->is_MachBranch(); 610 if (is_branch) { 611 MacroAssembler masm(&buf); 612 masm.bind(fakeL); 613 n->as_MachBranch()->save_label(&saveL, &save_bnum); 614 n->as_MachBranch()->label_set(&fakeL, 0); 615 } 616 n->emit(buf, this->regalloc()); 617 618 // Emitting into the scratch buffer should not fail 619 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 620 621 if (is_branch) // Restore label. 622 n->as_MachBranch()->label_set(saveL, save_bnum); 623 624 // End scratch_emit_size section. 625 set_in_scratch_emit_size(false); 626 627 return buf.insts_size(); 628 } 629 630 631 // ============================================================================ 632 //------------------------------Compile standard------------------------------- 633 debug_only( int Compile::_debug_idx = 100000; ) 634 635 // Compile a method. entry_bci is -1 for normal compilations and indicates 636 // the continuation bci for on stack replacement. 637 638 639 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 640 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 641 : Phase(Compiler), 642 _env(ci_env), 643 _directive(directive), 644 _log(ci_env->log()), 645 _compile_id(ci_env->compile_id()), 646 _save_argument_registers(false), 647 _stub_name(NULL), 648 _stub_function(NULL), 649 _stub_entry_point(NULL), 650 _method(target), 651 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 652 _entry_bci(osr_bci), 653 _initial_gvn(NULL), 654 _for_igvn(NULL), 655 _warm_calls(NULL), 656 _subsume_loads(subsume_loads), 657 _do_escape_analysis(do_escape_analysis), 658 _eliminate_boxing(eliminate_boxing), 659 _failure_reason(NULL), 660 _code_buffer("Compile::Fill_buffer"), 661 _orig_pc_slot(0), 662 _orig_pc_slot_offset_in_bytes(0), 663 _has_method_handle_invokes(false), 664 _mach_constant_base_node(NULL), 665 _node_bundling_limit(0), 666 _node_bundling_base(NULL), 667 _java_calls(0), 668 _inner_loops(0), 669 _scratch_const_size(-1), 670 _in_scratch_emit_size(false), 671 _dead_node_list(comp_arena()), 672 _dead_node_count(0), 673 #ifndef PRODUCT 674 _trace_opto_output(directive->TraceOptoOutputOption), 675 _in_dump_cnt(0), 676 _printer(IdealGraphPrinter::printer()), 677 #endif 678 _congraph(NULL), 679 _comp_arena(mtCompiler), 680 _node_arena(mtCompiler), 681 _old_arena(mtCompiler), 682 _Compile_types(mtCompiler), 683 _replay_inline_data(NULL), 684 _late_inlines(comp_arena(), 2, 0, NULL), 685 _string_late_inlines(comp_arena(), 2, 0, NULL), 686 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 687 _late_inlines_pos(0), 688 _number_of_mh_late_inlines(0), 689 _inlining_progress(false), 690 _inlining_incrementally(false), 691 _print_inlining_list(NULL), 692 _print_inlining_stream(NULL), 693 _print_inlining_idx(0), 694 _print_inlining_output(NULL), 695 _interpreter_frame_size(0), 696 _max_node_limit(MaxNodeLimit), 697 _has_reserved_stack_access(target->has_reserved_stack_access()) { 698 C = this; 699 #ifndef PRODUCT 700 if (_printer != NULL) { 701 _printer->set_compile(this); 702 } 703 #endif 704 CompileWrapper cw(this); 705 706 if (CITimeVerbose) { 707 tty->print(" "); 708 target->holder()->name()->print(); 709 tty->print("."); 710 target->print_short_name(); 711 tty->print(" "); 712 } 713 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 714 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 715 716 #ifndef PRODUCT 717 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 718 if (!print_opto_assembly) { 719 bool print_assembly = directive->PrintAssemblyOption; 720 if (print_assembly && !Disassembler::can_decode()) { 721 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 722 print_opto_assembly = true; 723 } 724 } 725 set_print_assembly(print_opto_assembly); 726 set_parsed_irreducible_loop(false); 727 728 if (directive->ReplayInlineOption) { 729 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 730 } 731 #endif 732 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 733 set_print_intrinsics(directive->PrintIntrinsicsOption); 734 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 735 736 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 737 // Make sure the method being compiled gets its own MDO, 738 // so we can at least track the decompile_count(). 739 // Need MDO to record RTM code generation state. 740 method()->ensure_method_data(); 741 } 742 743 Init(::AliasLevel); 744 745 746 print_compile_messages(); 747 748 _ilt = InlineTree::build_inline_tree_root(); 749 750 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 751 assert(num_alias_types() >= AliasIdxRaw, ""); 752 753 #define MINIMUM_NODE_HASH 1023 754 // Node list that Iterative GVN will start with 755 Unique_Node_List for_igvn(comp_arena()); 756 set_for_igvn(&for_igvn); 757 758 // GVN that will be run immediately on new nodes 759 uint estimated_size = method()->code_size()*4+64; 760 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 761 PhaseGVN gvn(node_arena(), estimated_size); 762 set_initial_gvn(&gvn); 763 764 print_inlining_init(); 765 { // Scope for timing the parser 766 TracePhase tp("parse", &timers[_t_parser]); 767 768 // Put top into the hash table ASAP. 769 initial_gvn()->transform_no_reclaim(top()); 770 771 // Set up tf(), start(), and find a CallGenerator. 772 CallGenerator* cg = NULL; 773 if (is_osr_compilation()) { 774 const TypeTuple *domain = StartOSRNode::osr_domain(); 775 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 776 init_tf(TypeFunc::make(domain, range)); 777 StartNode* s = new StartOSRNode(root(), domain); 778 initial_gvn()->set_type_bottom(s); 779 init_start(s); 780 cg = CallGenerator::for_osr(method(), entry_bci()); 781 } else { 782 // Normal case. 783 init_tf(TypeFunc::make(method())); 784 StartNode* s = new StartNode(root(), tf()->domain()); 785 initial_gvn()->set_type_bottom(s); 786 init_start(s); 787 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 788 // With java.lang.ref.reference.get() we must go through the 789 // intrinsic - even when get() is the root 790 // method of the compile - so that, if necessary, the value in 791 // the referent field of the reference object gets recorded by 792 // the pre-barrier code. 793 cg = find_intrinsic(method(), false); 794 } 795 if (cg == NULL) { 796 float past_uses = method()->interpreter_invocation_count(); 797 float expected_uses = past_uses; 798 cg = CallGenerator::for_inline(method(), expected_uses); 799 } 800 } 801 if (failing()) return; 802 if (cg == NULL) { 803 record_method_not_compilable("cannot parse method"); 804 return; 805 } 806 JVMState* jvms = build_start_state(start(), tf()); 807 if ((jvms = cg->generate(jvms)) == NULL) { 808 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 809 record_method_not_compilable("method parse failed"); 810 } 811 return; 812 } 813 GraphKit kit(jvms); 814 815 if (!kit.stopped()) { 816 // Accept return values, and transfer control we know not where. 817 // This is done by a special, unique ReturnNode bound to root. 818 return_values(kit.jvms()); 819 } 820 821 if (kit.has_exceptions()) { 822 // Any exceptions that escape from this call must be rethrown 823 // to whatever caller is dynamically above us on the stack. 824 // This is done by a special, unique RethrowNode bound to root. 825 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 826 } 827 828 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 829 830 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 831 inline_string_calls(true); 832 } 833 834 if (failing()) return; 835 836 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 837 838 // Remove clutter produced by parsing. 839 if (!failing()) { 840 ResourceMark rm; 841 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 842 } 843 } 844 845 // Note: Large methods are capped off in do_one_bytecode(). 846 if (failing()) return; 847 848 // After parsing, node notes are no longer automagic. 849 // They must be propagated by register_new_node_with_optimizer(), 850 // clone(), or the like. 851 set_default_node_notes(NULL); 852 853 for (;;) { 854 int successes = Inline_Warm(); 855 if (failing()) return; 856 if (successes == 0) break; 857 } 858 859 // Drain the list. 860 Finish_Warm(); 861 #ifndef PRODUCT 862 if (_printer && _printer->should_print(1)) { 863 _printer->print_inlining(); 864 } 865 #endif 866 867 if (failing()) return; 868 NOT_PRODUCT( verify_graph_edges(); ) 869 870 // Now optimize 871 Optimize(); 872 if (failing()) return; 873 NOT_PRODUCT( verify_graph_edges(); ) 874 875 #ifndef PRODUCT 876 if (PrintIdeal) { 877 ttyLocker ttyl; // keep the following output all in one block 878 // This output goes directly to the tty, not the compiler log. 879 // To enable tools to match it up with the compilation activity, 880 // be sure to tag this tty output with the compile ID. 881 if (xtty != NULL) { 882 xtty->head("ideal compile_id='%d'%s", compile_id(), 883 is_osr_compilation() ? " compile_kind='osr'" : 884 ""); 885 } 886 root()->dump(9999); 887 if (xtty != NULL) { 888 xtty->tail("ideal"); 889 } 890 } 891 #endif 892 893 NOT_PRODUCT( verify_barriers(); ) 894 895 // Dump compilation data to replay it. 896 if (directive->DumpReplayOption) { 897 env()->dump_replay_data(_compile_id); 898 } 899 if (directive->DumpInlineOption && (ilt() != NULL)) { 900 env()->dump_inline_data(_compile_id); 901 } 902 903 // Now that we know the size of all the monitors we can add a fixed slot 904 // for the original deopt pc. 905 906 _orig_pc_slot = fixed_slots(); 907 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 908 set_fixed_slots(next_slot); 909 910 // Compute when to use implicit null checks. Used by matching trap based 911 // nodes and NullCheck optimization. 912 set_allowed_deopt_reasons(); 913 914 // Now generate code 915 Code_Gen(); 916 if (failing()) return; 917 918 // Check if we want to skip execution of all compiled code. 919 { 920 #ifndef PRODUCT 921 if (OptoNoExecute) { 922 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 923 return; 924 } 925 #endif 926 TracePhase tp("install_code", &timers[_t_registerMethod]); 927 928 if (is_osr_compilation()) { 929 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 930 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 931 } else { 932 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 933 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 934 } 935 936 env()->register_method(_method, _entry_bci, 937 &_code_offsets, 938 _orig_pc_slot_offset_in_bytes, 939 code_buffer(), 940 frame_size_in_words(), _oop_map_set, 941 &_handler_table, &_inc_table, 942 compiler, 943 has_unsafe_access(), 944 SharedRuntime::is_wide_vector(max_vector_size()), 945 rtm_state() 946 ); 947 948 if (log() != NULL) // Print code cache state into compiler log 949 log()->code_cache_state(); 950 } 951 } 952 953 //------------------------------Compile---------------------------------------- 954 // Compile a runtime stub 955 Compile::Compile( ciEnv* ci_env, 956 TypeFunc_generator generator, 957 address stub_function, 958 const char *stub_name, 959 int is_fancy_jump, 960 bool pass_tls, 961 bool save_arg_registers, 962 bool return_pc, 963 DirectiveSet* directive) 964 : Phase(Compiler), 965 _env(ci_env), 966 _directive(directive), 967 _log(ci_env->log()), 968 _compile_id(0), 969 _save_argument_registers(save_arg_registers), 970 _method(NULL), 971 _stub_name(stub_name), 972 _stub_function(stub_function), 973 _stub_entry_point(NULL), 974 _entry_bci(InvocationEntryBci), 975 _initial_gvn(NULL), 976 _for_igvn(NULL), 977 _warm_calls(NULL), 978 _orig_pc_slot(0), 979 _orig_pc_slot_offset_in_bytes(0), 980 _subsume_loads(true), 981 _do_escape_analysis(false), 982 _eliminate_boxing(false), 983 _failure_reason(NULL), 984 _code_buffer("Compile::Fill_buffer"), 985 _has_method_handle_invokes(false), 986 _mach_constant_base_node(NULL), 987 _node_bundling_limit(0), 988 _node_bundling_base(NULL), 989 _java_calls(0), 990 _inner_loops(0), 991 #ifndef PRODUCT 992 _trace_opto_output(directive->TraceOptoOutputOption), 993 _in_dump_cnt(0), 994 _printer(NULL), 995 #endif 996 _comp_arena(mtCompiler), 997 _node_arena(mtCompiler), 998 _old_arena(mtCompiler), 999 _Compile_types(mtCompiler), 1000 _dead_node_list(comp_arena()), 1001 _dead_node_count(0), 1002 _congraph(NULL), 1003 _replay_inline_data(NULL), 1004 _number_of_mh_late_inlines(0), 1005 _inlining_progress(false), 1006 _inlining_incrementally(false), 1007 _print_inlining_list(NULL), 1008 _print_inlining_stream(NULL), 1009 _print_inlining_idx(0), 1010 _print_inlining_output(NULL), 1011 _allowed_reasons(0), 1012 _interpreter_frame_size(0), 1013 _max_node_limit(MaxNodeLimit), 1014 _has_reserved_stack_access(false) { 1015 C = this; 1016 1017 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1018 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1019 1020 #ifndef PRODUCT 1021 set_print_assembly(PrintFrameConverterAssembly); 1022 set_parsed_irreducible_loop(false); 1023 #endif 1024 set_has_irreducible_loop(false); // no loops 1025 1026 CompileWrapper cw(this); 1027 Init(/*AliasLevel=*/ 0); 1028 init_tf((*generator)()); 1029 1030 { 1031 // The following is a dummy for the sake of GraphKit::gen_stub 1032 Unique_Node_List for_igvn(comp_arena()); 1033 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1034 PhaseGVN gvn(Thread::current()->resource_area(),255); 1035 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1036 gvn.transform_no_reclaim(top()); 1037 1038 GraphKit kit; 1039 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1040 } 1041 1042 NOT_PRODUCT( verify_graph_edges(); ) 1043 Code_Gen(); 1044 if (failing()) return; 1045 1046 1047 // Entry point will be accessed using compile->stub_entry_point(); 1048 if (code_buffer() == NULL) { 1049 Matcher::soft_match_failure(); 1050 } else { 1051 if (PrintAssembly && (WizardMode || Verbose)) 1052 tty->print_cr("### Stub::%s", stub_name); 1053 1054 if (!failing()) { 1055 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1056 1057 // Make the NMethod 1058 // For now we mark the frame as never safe for profile stackwalking 1059 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1060 code_buffer(), 1061 CodeOffsets::frame_never_safe, 1062 // _code_offsets.value(CodeOffsets::Frame_Complete), 1063 frame_size_in_words(), 1064 _oop_map_set, 1065 save_arg_registers); 1066 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1067 1068 _stub_entry_point = rs->entry_point(); 1069 } 1070 } 1071 } 1072 1073 //------------------------------Init------------------------------------------- 1074 // Prepare for a single compilation 1075 void Compile::Init(int aliaslevel) { 1076 _unique = 0; 1077 _regalloc = NULL; 1078 1079 _tf = NULL; // filled in later 1080 _top = NULL; // cached later 1081 _matcher = NULL; // filled in later 1082 _cfg = NULL; // filled in later 1083 1084 set_24_bit_selection_and_mode(Use24BitFP, false); 1085 1086 _node_note_array = NULL; 1087 _default_node_notes = NULL; 1088 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1089 1090 _immutable_memory = NULL; // filled in at first inquiry 1091 1092 // Globally visible Nodes 1093 // First set TOP to NULL to give safe behavior during creation of RootNode 1094 set_cached_top_node(NULL); 1095 set_root(new RootNode()); 1096 // Now that you have a Root to point to, create the real TOP 1097 set_cached_top_node( new ConNode(Type::TOP) ); 1098 set_recent_alloc(NULL, NULL); 1099 1100 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1101 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1102 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1103 env()->set_dependencies(new Dependencies(env())); 1104 1105 _fixed_slots = 0; 1106 set_has_split_ifs(false); 1107 set_has_loops(has_method() && method()->has_loops()); // first approximation 1108 set_has_stringbuilder(false); 1109 set_has_boxed_value(false); 1110 _trap_can_recompile = false; // no traps emitted yet 1111 _major_progress = true; // start out assuming good things will happen 1112 set_has_unsafe_access(false); 1113 set_max_vector_size(0); 1114 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1115 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1116 set_decompile_count(0); 1117 1118 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1119 set_num_loop_opts(LoopOptsCount); 1120 set_do_inlining(Inline); 1121 set_max_inline_size(MaxInlineSize); 1122 set_freq_inline_size(FreqInlineSize); 1123 set_do_scheduling(OptoScheduling); 1124 set_do_count_invocations(false); 1125 set_do_method_data_update(false); 1126 1127 set_do_vector_loop(false); 1128 1129 if (AllowVectorizeOnDemand) { 1130 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1131 set_do_vector_loop(true); 1132 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());}) 1133 } else if (has_method() && method()->name() != 0 && 1134 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1135 set_do_vector_loop(true); 1136 } 1137 } 1138 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1139 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());}) 1140 1141 set_age_code(has_method() && method()->profile_aging()); 1142 set_rtm_state(NoRTM); // No RTM lock eliding by default 1143 _max_node_limit = _directive->MaxNodeLimitOption; 1144 1145 #if INCLUDE_RTM_OPT 1146 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1147 int rtm_state = method()->method_data()->rtm_state(); 1148 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1149 // Don't generate RTM lock eliding code. 1150 set_rtm_state(NoRTM); 1151 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1152 // Generate RTM lock eliding code without abort ratio calculation code. 1153 set_rtm_state(UseRTM); 1154 } else if (UseRTMDeopt) { 1155 // Generate RTM lock eliding code and include abort ratio calculation 1156 // code if UseRTMDeopt is on. 1157 set_rtm_state(ProfileRTM); 1158 } 1159 } 1160 #endif 1161 if (debug_info()->recording_non_safepoints()) { 1162 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1163 (comp_arena(), 8, 0, NULL)); 1164 set_default_node_notes(Node_Notes::make(this)); 1165 } 1166 1167 // // -- Initialize types before each compile -- 1168 // // Update cached type information 1169 // if( _method && _method->constants() ) 1170 // Type::update_loaded_types(_method, _method->constants()); 1171 1172 // Init alias_type map. 1173 if (!_do_escape_analysis && aliaslevel == 3) 1174 aliaslevel = 2; // No unique types without escape analysis 1175 _AliasLevel = aliaslevel; 1176 const int grow_ats = 16; 1177 _max_alias_types = grow_ats; 1178 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1179 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1180 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1181 { 1182 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1183 } 1184 // Initialize the first few types. 1185 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1186 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1187 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1188 _num_alias_types = AliasIdxRaw+1; 1189 // Zero out the alias type cache. 1190 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1191 // A NULL adr_type hits in the cache right away. Preload the right answer. 1192 probe_alias_cache(NULL)->_index = AliasIdxTop; 1193 1194 _intrinsics = NULL; 1195 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1196 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1197 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1198 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1199 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1200 register_library_intrinsics(); 1201 } 1202 1203 //---------------------------init_start---------------------------------------- 1204 // Install the StartNode on this compile object. 1205 void Compile::init_start(StartNode* s) { 1206 if (failing()) 1207 return; // already failing 1208 assert(s == start(), ""); 1209 } 1210 1211 /** 1212 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1213 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1214 * the ideal graph. 1215 */ 1216 StartNode* Compile::start() const { 1217 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1218 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1219 Node* start = root()->fast_out(i); 1220 if (start->is_Start()) { 1221 return start->as_Start(); 1222 } 1223 } 1224 fatal("Did not find Start node!"); 1225 return NULL; 1226 } 1227 1228 //-------------------------------immutable_memory------------------------------------- 1229 // Access immutable memory 1230 Node* Compile::immutable_memory() { 1231 if (_immutable_memory != NULL) { 1232 return _immutable_memory; 1233 } 1234 StartNode* s = start(); 1235 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1236 Node *p = s->fast_out(i); 1237 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1238 _immutable_memory = p; 1239 return _immutable_memory; 1240 } 1241 } 1242 ShouldNotReachHere(); 1243 return NULL; 1244 } 1245 1246 //----------------------set_cached_top_node------------------------------------ 1247 // Install the cached top node, and make sure Node::is_top works correctly. 1248 void Compile::set_cached_top_node(Node* tn) { 1249 if (tn != NULL) verify_top(tn); 1250 Node* old_top = _top; 1251 _top = tn; 1252 // Calling Node::setup_is_top allows the nodes the chance to adjust 1253 // their _out arrays. 1254 if (_top != NULL) _top->setup_is_top(); 1255 if (old_top != NULL) old_top->setup_is_top(); 1256 assert(_top == NULL || top()->is_top(), ""); 1257 } 1258 1259 #ifdef ASSERT 1260 uint Compile::count_live_nodes_by_graph_walk() { 1261 Unique_Node_List useful(comp_arena()); 1262 // Get useful node list by walking the graph. 1263 identify_useful_nodes(useful); 1264 return useful.size(); 1265 } 1266 1267 void Compile::print_missing_nodes() { 1268 1269 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1270 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1271 return; 1272 } 1273 1274 // This is an expensive function. It is executed only when the user 1275 // specifies VerifyIdealNodeCount option or otherwise knows the 1276 // additional work that needs to be done to identify reachable nodes 1277 // by walking the flow graph and find the missing ones using 1278 // _dead_node_list. 1279 1280 Unique_Node_List useful(comp_arena()); 1281 // Get useful node list by walking the graph. 1282 identify_useful_nodes(useful); 1283 1284 uint l_nodes = C->live_nodes(); 1285 uint l_nodes_by_walk = useful.size(); 1286 1287 if (l_nodes != l_nodes_by_walk) { 1288 if (_log != NULL) { 1289 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1290 _log->stamp(); 1291 _log->end_head(); 1292 } 1293 VectorSet& useful_member_set = useful.member_set(); 1294 int last_idx = l_nodes_by_walk; 1295 for (int i = 0; i < last_idx; i++) { 1296 if (useful_member_set.test(i)) { 1297 if (_dead_node_list.test(i)) { 1298 if (_log != NULL) { 1299 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1300 } 1301 if (PrintIdealNodeCount) { 1302 // Print the log message to tty 1303 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1304 useful.at(i)->dump(); 1305 } 1306 } 1307 } 1308 else if (! _dead_node_list.test(i)) { 1309 if (_log != NULL) { 1310 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1311 } 1312 if (PrintIdealNodeCount) { 1313 // Print the log message to tty 1314 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1315 } 1316 } 1317 } 1318 if (_log != NULL) { 1319 _log->tail("mismatched_nodes"); 1320 } 1321 } 1322 } 1323 void Compile::record_modified_node(Node* n) { 1324 if (_modified_nodes != NULL && !_inlining_incrementally && 1325 n->outcnt() != 0 && !n->is_Con()) { 1326 _modified_nodes->push(n); 1327 } 1328 } 1329 1330 void Compile::remove_modified_node(Node* n) { 1331 if (_modified_nodes != NULL) { 1332 _modified_nodes->remove(n); 1333 } 1334 } 1335 #endif 1336 1337 #ifndef PRODUCT 1338 void Compile::verify_top(Node* tn) const { 1339 if (tn != NULL) { 1340 assert(tn->is_Con(), "top node must be a constant"); 1341 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1342 assert(tn->in(0) != NULL, "must have live top node"); 1343 } 1344 } 1345 #endif 1346 1347 1348 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1349 1350 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1351 guarantee(arr != NULL, ""); 1352 int num_blocks = arr->length(); 1353 if (grow_by < num_blocks) grow_by = num_blocks; 1354 int num_notes = grow_by * _node_notes_block_size; 1355 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1356 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1357 while (num_notes > 0) { 1358 arr->append(notes); 1359 notes += _node_notes_block_size; 1360 num_notes -= _node_notes_block_size; 1361 } 1362 assert(num_notes == 0, "exact multiple, please"); 1363 } 1364 1365 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1366 if (source == NULL || dest == NULL) return false; 1367 1368 if (dest->is_Con()) 1369 return false; // Do not push debug info onto constants. 1370 1371 #ifdef ASSERT 1372 // Leave a bread crumb trail pointing to the original node: 1373 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1374 dest->set_debug_orig(source); 1375 } 1376 #endif 1377 1378 if (node_note_array() == NULL) 1379 return false; // Not collecting any notes now. 1380 1381 // This is a copy onto a pre-existing node, which may already have notes. 1382 // If both nodes have notes, do not overwrite any pre-existing notes. 1383 Node_Notes* source_notes = node_notes_at(source->_idx); 1384 if (source_notes == NULL || source_notes->is_clear()) return false; 1385 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1386 if (dest_notes == NULL || dest_notes->is_clear()) { 1387 return set_node_notes_at(dest->_idx, source_notes); 1388 } 1389 1390 Node_Notes merged_notes = (*source_notes); 1391 // The order of operations here ensures that dest notes will win... 1392 merged_notes.update_from(dest_notes); 1393 return set_node_notes_at(dest->_idx, &merged_notes); 1394 } 1395 1396 1397 //--------------------------allow_range_check_smearing------------------------- 1398 // Gating condition for coalescing similar range checks. 1399 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1400 // single covering check that is at least as strong as any of them. 1401 // If the optimization succeeds, the simplified (strengthened) range check 1402 // will always succeed. If it fails, we will deopt, and then give up 1403 // on the optimization. 1404 bool Compile::allow_range_check_smearing() const { 1405 // If this method has already thrown a range-check, 1406 // assume it was because we already tried range smearing 1407 // and it failed. 1408 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1409 return !already_trapped; 1410 } 1411 1412 1413 //------------------------------flatten_alias_type----------------------------- 1414 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1415 int offset = tj->offset(); 1416 TypePtr::PTR ptr = tj->ptr(); 1417 1418 // Known instance (scalarizable allocation) alias only with itself. 1419 bool is_known_inst = tj->isa_oopptr() != NULL && 1420 tj->is_oopptr()->is_known_instance(); 1421 1422 // Process weird unsafe references. 1423 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1424 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1425 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1426 tj = TypeOopPtr::BOTTOM; 1427 ptr = tj->ptr(); 1428 offset = tj->offset(); 1429 } 1430 1431 // Array pointers need some flattening 1432 const TypeAryPtr *ta = tj->isa_aryptr(); 1433 if (ta && ta->is_stable()) { 1434 // Erase stability property for alias analysis. 1435 tj = ta = ta->cast_to_stable(false); 1436 } 1437 if( ta && is_known_inst ) { 1438 if ( offset != Type::OffsetBot && 1439 offset > arrayOopDesc::length_offset_in_bytes() ) { 1440 offset = Type::OffsetBot; // Flatten constant access into array body only 1441 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1442 } 1443 } else if( ta && _AliasLevel >= 2 ) { 1444 // For arrays indexed by constant indices, we flatten the alias 1445 // space to include all of the array body. Only the header, klass 1446 // and array length can be accessed un-aliased. 1447 if( offset != Type::OffsetBot ) { 1448 if( ta->const_oop() ) { // MethodData* or Method* 1449 offset = Type::OffsetBot; // Flatten constant access into array body 1450 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1451 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1452 // range is OK as-is. 1453 tj = ta = TypeAryPtr::RANGE; 1454 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1455 tj = TypeInstPtr::KLASS; // all klass loads look alike 1456 ta = TypeAryPtr::RANGE; // generic ignored junk 1457 ptr = TypePtr::BotPTR; 1458 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1459 tj = TypeInstPtr::MARK; 1460 ta = TypeAryPtr::RANGE; // generic ignored junk 1461 ptr = TypePtr::BotPTR; 1462 } else { // Random constant offset into array body 1463 offset = Type::OffsetBot; // Flatten constant access into array body 1464 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1465 } 1466 } 1467 // Arrays of fixed size alias with arrays of unknown size. 1468 if (ta->size() != TypeInt::POS) { 1469 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1470 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1471 } 1472 // Arrays of known objects become arrays of unknown objects. 1473 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1474 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1475 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1476 } 1477 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1478 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1479 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1480 } 1481 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1482 // cannot be distinguished by bytecode alone. 1483 if (ta->elem() == TypeInt::BOOL) { 1484 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1485 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1486 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1487 } 1488 // During the 2nd round of IterGVN, NotNull castings are removed. 1489 // Make sure the Bottom and NotNull variants alias the same. 1490 // Also, make sure exact and non-exact variants alias the same. 1491 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1492 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1493 } 1494 } 1495 1496 // Oop pointers need some flattening 1497 const TypeInstPtr *to = tj->isa_instptr(); 1498 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1499 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1500 if( ptr == TypePtr::Constant ) { 1501 if (to->klass() != ciEnv::current()->Class_klass() || 1502 offset < k->size_helper() * wordSize) { 1503 // No constant oop pointers (such as Strings); they alias with 1504 // unknown strings. 1505 assert(!is_known_inst, "not scalarizable allocation"); 1506 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1507 } 1508 } else if( is_known_inst ) { 1509 tj = to; // Keep NotNull and klass_is_exact for instance type 1510 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1511 // During the 2nd round of IterGVN, NotNull castings are removed. 1512 // Make sure the Bottom and NotNull variants alias the same. 1513 // Also, make sure exact and non-exact variants alias the same. 1514 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1515 } 1516 if (to->speculative() != NULL) { 1517 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1518 } 1519 // Canonicalize the holder of this field 1520 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1521 // First handle header references such as a LoadKlassNode, even if the 1522 // object's klass is unloaded at compile time (4965979). 1523 if (!is_known_inst) { // Do it only for non-instance types 1524 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1525 } 1526 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1527 // Static fields are in the space above the normal instance 1528 // fields in the java.lang.Class instance. 1529 if (to->klass() != ciEnv::current()->Class_klass()) { 1530 to = NULL; 1531 tj = TypeOopPtr::BOTTOM; 1532 offset = tj->offset(); 1533 } 1534 } else { 1535 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1536 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1537 if( is_known_inst ) { 1538 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1539 } else { 1540 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1541 } 1542 } 1543 } 1544 } 1545 1546 // Klass pointers to object array klasses need some flattening 1547 const TypeKlassPtr *tk = tj->isa_klassptr(); 1548 if( tk ) { 1549 // If we are referencing a field within a Klass, we need 1550 // to assume the worst case of an Object. Both exact and 1551 // inexact types must flatten to the same alias class so 1552 // use NotNull as the PTR. 1553 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1554 1555 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1556 TypeKlassPtr::OBJECT->klass(), 1557 offset); 1558 } 1559 1560 ciKlass* klass = tk->klass(); 1561 if( klass->is_obj_array_klass() ) { 1562 ciKlass* k = TypeAryPtr::OOPS->klass(); 1563 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1564 k = TypeInstPtr::BOTTOM->klass(); 1565 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1566 } 1567 1568 // Check for precise loads from the primary supertype array and force them 1569 // to the supertype cache alias index. Check for generic array loads from 1570 // the primary supertype array and also force them to the supertype cache 1571 // alias index. Since the same load can reach both, we need to merge 1572 // these 2 disparate memories into the same alias class. Since the 1573 // primary supertype array is read-only, there's no chance of confusion 1574 // where we bypass an array load and an array store. 1575 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1576 if (offset == Type::OffsetBot || 1577 (offset >= primary_supers_offset && 1578 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1579 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1580 offset = in_bytes(Klass::secondary_super_cache_offset()); 1581 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1582 } 1583 } 1584 1585 // Flatten all Raw pointers together. 1586 if (tj->base() == Type::RawPtr) 1587 tj = TypeRawPtr::BOTTOM; 1588 1589 if (tj->base() == Type::AnyPtr) 1590 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1591 1592 // Flatten all to bottom for now 1593 switch( _AliasLevel ) { 1594 case 0: 1595 tj = TypePtr::BOTTOM; 1596 break; 1597 case 1: // Flatten to: oop, static, field or array 1598 switch (tj->base()) { 1599 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1600 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1601 case Type::AryPtr: // do not distinguish arrays at all 1602 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1603 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1604 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1605 default: ShouldNotReachHere(); 1606 } 1607 break; 1608 case 2: // No collapsing at level 2; keep all splits 1609 case 3: // No collapsing at level 3; keep all splits 1610 break; 1611 default: 1612 Unimplemented(); 1613 } 1614 1615 offset = tj->offset(); 1616 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1617 1618 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1619 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1620 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1621 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1622 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1623 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1624 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1625 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1626 assert( tj->ptr() != TypePtr::TopPTR && 1627 tj->ptr() != TypePtr::AnyNull && 1628 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1629 // assert( tj->ptr() != TypePtr::Constant || 1630 // tj->base() == Type::RawPtr || 1631 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1632 1633 return tj; 1634 } 1635 1636 void Compile::AliasType::Init(int i, const TypePtr* at) { 1637 _index = i; 1638 _adr_type = at; 1639 _field = NULL; 1640 _element = NULL; 1641 _is_rewritable = true; // default 1642 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1643 if (atoop != NULL && atoop->is_known_instance()) { 1644 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1645 _general_index = Compile::current()->get_alias_index(gt); 1646 } else { 1647 _general_index = 0; 1648 } 1649 } 1650 1651 BasicType Compile::AliasType::basic_type() const { 1652 if (element() != NULL) { 1653 const Type* element = adr_type()->is_aryptr()->elem(); 1654 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1655 } if (field() != NULL) { 1656 return field()->layout_type(); 1657 } else { 1658 return T_ILLEGAL; // unknown 1659 } 1660 } 1661 1662 //---------------------------------print_on------------------------------------ 1663 #ifndef PRODUCT 1664 void Compile::AliasType::print_on(outputStream* st) { 1665 if (index() < 10) 1666 st->print("@ <%d> ", index()); 1667 else st->print("@ <%d>", index()); 1668 st->print(is_rewritable() ? " " : " RO"); 1669 int offset = adr_type()->offset(); 1670 if (offset == Type::OffsetBot) 1671 st->print(" +any"); 1672 else st->print(" +%-3d", offset); 1673 st->print(" in "); 1674 adr_type()->dump_on(st); 1675 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1676 if (field() != NULL && tjp) { 1677 if (tjp->klass() != field()->holder() || 1678 tjp->offset() != field()->offset_in_bytes()) { 1679 st->print(" != "); 1680 field()->print(); 1681 st->print(" ***"); 1682 } 1683 } 1684 } 1685 1686 void print_alias_types() { 1687 Compile* C = Compile::current(); 1688 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1689 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1690 C->alias_type(idx)->print_on(tty); 1691 tty->cr(); 1692 } 1693 } 1694 #endif 1695 1696 1697 //----------------------------probe_alias_cache-------------------------------- 1698 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1699 intptr_t key = (intptr_t) adr_type; 1700 key ^= key >> logAliasCacheSize; 1701 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1702 } 1703 1704 1705 //-----------------------------grow_alias_types-------------------------------- 1706 void Compile::grow_alias_types() { 1707 const int old_ats = _max_alias_types; // how many before? 1708 const int new_ats = old_ats; // how many more? 1709 const int grow_ats = old_ats+new_ats; // how many now? 1710 _max_alias_types = grow_ats; 1711 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1712 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1713 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1714 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1715 } 1716 1717 1718 //--------------------------------find_alias_type------------------------------ 1719 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1720 if (_AliasLevel == 0) 1721 return alias_type(AliasIdxBot); 1722 1723 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1724 if (ace->_adr_type == adr_type) { 1725 return alias_type(ace->_index); 1726 } 1727 1728 // Handle special cases. 1729 if (adr_type == NULL) return alias_type(AliasIdxTop); 1730 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1731 1732 // Do it the slow way. 1733 const TypePtr* flat = flatten_alias_type(adr_type); 1734 1735 #ifdef ASSERT 1736 { 1737 ResourceMark rm; 1738 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1739 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1740 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1741 Type::str(adr_type)); 1742 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1743 const TypeOopPtr* foop = flat->is_oopptr(); 1744 // Scalarizable allocations have exact klass always. 1745 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1746 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1747 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1748 Type::str(foop), Type::str(xoop)); 1749 } 1750 } 1751 #endif 1752 1753 int idx = AliasIdxTop; 1754 for (int i = 0; i < num_alias_types(); i++) { 1755 if (alias_type(i)->adr_type() == flat) { 1756 idx = i; 1757 break; 1758 } 1759 } 1760 1761 if (idx == AliasIdxTop) { 1762 if (no_create) return NULL; 1763 // Grow the array if necessary. 1764 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1765 // Add a new alias type. 1766 idx = _num_alias_types++; 1767 _alias_types[idx]->Init(idx, flat); 1768 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1769 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1770 if (flat->isa_instptr()) { 1771 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1772 && flat->is_instptr()->klass() == env()->Class_klass()) 1773 alias_type(idx)->set_rewritable(false); 1774 } 1775 if (flat->isa_aryptr()) { 1776 #ifdef ASSERT 1777 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1778 // (T_BYTE has the weakest alignment and size restrictions...) 1779 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1780 #endif 1781 if (flat->offset() == TypePtr::OffsetBot) { 1782 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1783 } 1784 } 1785 if (flat->isa_klassptr()) { 1786 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1787 alias_type(idx)->set_rewritable(false); 1788 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1789 alias_type(idx)->set_rewritable(false); 1790 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1791 alias_type(idx)->set_rewritable(false); 1792 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1793 alias_type(idx)->set_rewritable(false); 1794 } 1795 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1796 // but the base pointer type is not distinctive enough to identify 1797 // references into JavaThread.) 1798 1799 // Check for final fields. 1800 const TypeInstPtr* tinst = flat->isa_instptr(); 1801 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1802 ciField* field; 1803 if (tinst->const_oop() != NULL && 1804 tinst->klass() == ciEnv::current()->Class_klass() && 1805 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1806 // static field 1807 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1808 field = k->get_field_by_offset(tinst->offset(), true); 1809 } else { 1810 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1811 field = k->get_field_by_offset(tinst->offset(), false); 1812 } 1813 assert(field == NULL || 1814 original_field == NULL || 1815 (field->holder() == original_field->holder() && 1816 field->offset() == original_field->offset() && 1817 field->is_static() == original_field->is_static()), "wrong field?"); 1818 // Set field() and is_rewritable() attributes. 1819 if (field != NULL) alias_type(idx)->set_field(field); 1820 } 1821 } 1822 1823 // Fill the cache for next time. 1824 ace->_adr_type = adr_type; 1825 ace->_index = idx; 1826 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1827 1828 // Might as well try to fill the cache for the flattened version, too. 1829 AliasCacheEntry* face = probe_alias_cache(flat); 1830 if (face->_adr_type == NULL) { 1831 face->_adr_type = flat; 1832 face->_index = idx; 1833 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1834 } 1835 1836 return alias_type(idx); 1837 } 1838 1839 1840 Compile::AliasType* Compile::alias_type(ciField* field) { 1841 const TypeOopPtr* t; 1842 if (field->is_static()) 1843 t = TypeInstPtr::make(field->holder()->java_mirror()); 1844 else 1845 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1846 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1847 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1848 return atp; 1849 } 1850 1851 1852 //------------------------------have_alias_type-------------------------------- 1853 bool Compile::have_alias_type(const TypePtr* adr_type) { 1854 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1855 if (ace->_adr_type == adr_type) { 1856 return true; 1857 } 1858 1859 // Handle special cases. 1860 if (adr_type == NULL) return true; 1861 if (adr_type == TypePtr::BOTTOM) return true; 1862 1863 return find_alias_type(adr_type, true, NULL) != NULL; 1864 } 1865 1866 //-----------------------------must_alias-------------------------------------- 1867 // True if all values of the given address type are in the given alias category. 1868 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1869 if (alias_idx == AliasIdxBot) return true; // the universal category 1870 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1871 if (alias_idx == AliasIdxTop) return false; // the empty category 1872 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1873 1874 // the only remaining possible overlap is identity 1875 int adr_idx = get_alias_index(adr_type); 1876 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1877 assert(adr_idx == alias_idx || 1878 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1879 && adr_type != TypeOopPtr::BOTTOM), 1880 "should not be testing for overlap with an unsafe pointer"); 1881 return adr_idx == alias_idx; 1882 } 1883 1884 //------------------------------can_alias-------------------------------------- 1885 // True if any values of the given address type are in the given alias category. 1886 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1887 if (alias_idx == AliasIdxTop) return false; // the empty category 1888 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1889 if (alias_idx == AliasIdxBot) return true; // the universal category 1890 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1891 1892 // the only remaining possible overlap is identity 1893 int adr_idx = get_alias_index(adr_type); 1894 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1895 return adr_idx == alias_idx; 1896 } 1897 1898 1899 1900 //---------------------------pop_warm_call------------------------------------- 1901 WarmCallInfo* Compile::pop_warm_call() { 1902 WarmCallInfo* wci = _warm_calls; 1903 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1904 return wci; 1905 } 1906 1907 //----------------------------Inline_Warm-------------------------------------- 1908 int Compile::Inline_Warm() { 1909 // If there is room, try to inline some more warm call sites. 1910 // %%% Do a graph index compaction pass when we think we're out of space? 1911 if (!InlineWarmCalls) return 0; 1912 1913 int calls_made_hot = 0; 1914 int room_to_grow = NodeCountInliningCutoff - unique(); 1915 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1916 int amount_grown = 0; 1917 WarmCallInfo* call; 1918 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1919 int est_size = (int)call->size(); 1920 if (est_size > (room_to_grow - amount_grown)) { 1921 // This one won't fit anyway. Get rid of it. 1922 call->make_cold(); 1923 continue; 1924 } 1925 call->make_hot(); 1926 calls_made_hot++; 1927 amount_grown += est_size; 1928 amount_to_grow -= est_size; 1929 } 1930 1931 if (calls_made_hot > 0) set_major_progress(); 1932 return calls_made_hot; 1933 } 1934 1935 1936 //----------------------------Finish_Warm-------------------------------------- 1937 void Compile::Finish_Warm() { 1938 if (!InlineWarmCalls) return; 1939 if (failing()) return; 1940 if (warm_calls() == NULL) return; 1941 1942 // Clean up loose ends, if we are out of space for inlining. 1943 WarmCallInfo* call; 1944 while ((call = pop_warm_call()) != NULL) { 1945 call->make_cold(); 1946 } 1947 } 1948 1949 //---------------------cleanup_loop_predicates----------------------- 1950 // Remove the opaque nodes that protect the predicates so that all unused 1951 // checks and uncommon_traps will be eliminated from the ideal graph 1952 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1953 if (predicate_count()==0) return; 1954 for (int i = predicate_count(); i > 0; i--) { 1955 Node * n = predicate_opaque1_node(i-1); 1956 assert(n->Opcode() == Op_Opaque1, "must be"); 1957 igvn.replace_node(n, n->in(1)); 1958 } 1959 assert(predicate_count()==0, "should be clean!"); 1960 } 1961 1962 void Compile::add_range_check_cast(Node* n) { 1963 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1964 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1965 _range_check_casts->append(n); 1966 } 1967 1968 // Remove all range check dependent CastIINodes. 1969 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1970 for (int i = range_check_cast_count(); i > 0; i--) { 1971 Node* cast = range_check_cast_node(i-1); 1972 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1973 igvn.replace_node(cast, cast->in(1)); 1974 } 1975 assert(range_check_cast_count() == 0, "should be empty"); 1976 } 1977 1978 void Compile::add_opaque4_node(Node* n) { 1979 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 1980 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 1981 _opaque4_nodes->append(n); 1982 } 1983 1984 // Remove all Opaque4 nodes. 1985 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 1986 for (int i = opaque4_count(); i > 0; i--) { 1987 Node* opaq = opaque4_node(i-1); 1988 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 1989 igvn.replace_node(opaq, opaq->in(2)); 1990 } 1991 assert(opaque4_count() == 0, "should be empty"); 1992 } 1993 1994 // StringOpts and late inlining of string methods 1995 void Compile::inline_string_calls(bool parse_time) { 1996 { 1997 // remove useless nodes to make the usage analysis simpler 1998 ResourceMark rm; 1999 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2000 } 2001 2002 { 2003 ResourceMark rm; 2004 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2005 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2006 print_method(PHASE_AFTER_STRINGOPTS, 3); 2007 } 2008 2009 // now inline anything that we skipped the first time around 2010 if (!parse_time) { 2011 _late_inlines_pos = _late_inlines.length(); 2012 } 2013 2014 while (_string_late_inlines.length() > 0) { 2015 CallGenerator* cg = _string_late_inlines.pop(); 2016 cg->do_late_inline(); 2017 if (failing()) return; 2018 } 2019 _string_late_inlines.trunc_to(0); 2020 } 2021 2022 // Late inlining of boxing methods 2023 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2024 if (_boxing_late_inlines.length() > 0) { 2025 assert(has_boxed_value(), "inconsistent"); 2026 2027 PhaseGVN* gvn = initial_gvn(); 2028 set_inlining_incrementally(true); 2029 2030 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2031 for_igvn()->clear(); 2032 gvn->replace_with(&igvn); 2033 2034 _late_inlines_pos = _late_inlines.length(); 2035 2036 while (_boxing_late_inlines.length() > 0) { 2037 CallGenerator* cg = _boxing_late_inlines.pop(); 2038 cg->do_late_inline(); 2039 if (failing()) return; 2040 } 2041 _boxing_late_inlines.trunc_to(0); 2042 2043 { 2044 ResourceMark rm; 2045 PhaseRemoveUseless pru(gvn, for_igvn()); 2046 } 2047 2048 igvn = PhaseIterGVN(gvn); 2049 igvn.optimize(); 2050 2051 set_inlining_progress(false); 2052 set_inlining_incrementally(false); 2053 } 2054 } 2055 2056 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 2057 assert(IncrementalInline, "incremental inlining should be on"); 2058 PhaseGVN* gvn = initial_gvn(); 2059 2060 set_inlining_progress(false); 2061 for_igvn()->clear(); 2062 gvn->replace_with(&igvn); 2063 2064 { 2065 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2066 int i = 0; 2067 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2068 CallGenerator* cg = _late_inlines.at(i); 2069 _late_inlines_pos = i+1; 2070 cg->do_late_inline(); 2071 if (failing()) return; 2072 } 2073 int j = 0; 2074 for (; i < _late_inlines.length(); i++, j++) { 2075 _late_inlines.at_put(j, _late_inlines.at(i)); 2076 } 2077 _late_inlines.trunc_to(j); 2078 } 2079 2080 { 2081 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2082 ResourceMark rm; 2083 PhaseRemoveUseless pru(gvn, for_igvn()); 2084 } 2085 2086 { 2087 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2088 igvn = PhaseIterGVN(gvn); 2089 } 2090 } 2091 2092 // Perform incremental inlining until bound on number of live nodes is reached 2093 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2094 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2095 2096 PhaseGVN* gvn = initial_gvn(); 2097 2098 set_inlining_incrementally(true); 2099 set_inlining_progress(true); 2100 uint low_live_nodes = 0; 2101 2102 while(inlining_progress() && _late_inlines.length() > 0) { 2103 2104 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2105 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2106 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2107 // PhaseIdealLoop is expensive so we only try it once we are 2108 // out of live nodes and we only try it again if the previous 2109 // helped got the number of nodes down significantly 2110 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone); 2111 if (failing()) return; 2112 low_live_nodes = live_nodes(); 2113 _major_progress = true; 2114 } 2115 2116 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2117 break; 2118 } 2119 } 2120 2121 inline_incrementally_one(igvn); 2122 2123 if (failing()) return; 2124 2125 { 2126 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2127 igvn.optimize(); 2128 } 2129 2130 if (failing()) return; 2131 } 2132 2133 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2134 2135 if (_string_late_inlines.length() > 0) { 2136 assert(has_stringbuilder(), "inconsistent"); 2137 for_igvn()->clear(); 2138 initial_gvn()->replace_with(&igvn); 2139 2140 inline_string_calls(false); 2141 2142 if (failing()) return; 2143 2144 { 2145 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2146 ResourceMark rm; 2147 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2148 } 2149 2150 { 2151 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2152 igvn = PhaseIterGVN(gvn); 2153 igvn.optimize(); 2154 } 2155 } 2156 2157 set_inlining_incrementally(false); 2158 } 2159 2160 2161 bool Compile::optimize_loops(int& loop_opts_cnt, PhaseIterGVN& igvn, LoopOptsMode mode) { 2162 if(loop_opts_cnt > 0) { 2163 debug_only( int cnt = 0; ); 2164 while(major_progress() && (loop_opts_cnt > 0)) { 2165 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2166 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2167 PhaseIdealLoop ideal_loop(igvn, mode); 2168 loop_opts_cnt--; 2169 if (failing()) return false; 2170 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2171 } 2172 } 2173 return true; 2174 } 2175 2176 //------------------------------Optimize--------------------------------------- 2177 // Given a graph, optimize it. 2178 void Compile::Optimize() { 2179 TracePhase tp("optimizer", &timers[_t_optimizer]); 2180 2181 #ifndef PRODUCT 2182 if (_directive->BreakAtCompileOption) { 2183 BREAKPOINT; 2184 } 2185 2186 #endif 2187 2188 #ifdef ASSERT 2189 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2190 bs->verify_gc_barriers(true); 2191 #endif 2192 2193 ResourceMark rm; 2194 int loop_opts_cnt; 2195 2196 print_inlining_reinit(); 2197 2198 NOT_PRODUCT( verify_graph_edges(); ) 2199 2200 print_method(PHASE_AFTER_PARSING); 2201 2202 { 2203 // Iterative Global Value Numbering, including ideal transforms 2204 // Initialize IterGVN with types and values from parse-time GVN 2205 PhaseIterGVN igvn(initial_gvn()); 2206 #ifdef ASSERT 2207 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2208 #endif 2209 { 2210 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2211 igvn.optimize(); 2212 } 2213 2214 if (failing()) return; 2215 2216 print_method(PHASE_ITER_GVN1, 2); 2217 2218 inline_incrementally(igvn); 2219 2220 print_method(PHASE_INCREMENTAL_INLINE, 2); 2221 2222 if (failing()) return; 2223 2224 if (eliminate_boxing()) { 2225 // Inline valueOf() methods now. 2226 inline_boxing_calls(igvn); 2227 2228 if (AlwaysIncrementalInline) { 2229 inline_incrementally(igvn); 2230 } 2231 2232 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2233 2234 if (failing()) return; 2235 } 2236 2237 // Remove the speculative part of types and clean up the graph from 2238 // the extra CastPP nodes whose only purpose is to carry them. Do 2239 // that early so that optimizations are not disrupted by the extra 2240 // CastPP nodes. 2241 remove_speculative_types(igvn); 2242 2243 // No more new expensive nodes will be added to the list from here 2244 // so keep only the actual candidates for optimizations. 2245 cleanup_expensive_nodes(igvn); 2246 2247 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2248 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2249 initial_gvn()->replace_with(&igvn); 2250 for_igvn()->clear(); 2251 Unique_Node_List new_worklist(C->comp_arena()); 2252 { 2253 ResourceMark rm; 2254 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2255 } 2256 set_for_igvn(&new_worklist); 2257 igvn = PhaseIterGVN(initial_gvn()); 2258 igvn.optimize(); 2259 } 2260 2261 // Perform escape analysis 2262 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2263 if (has_loops()) { 2264 // Cleanup graph (remove dead nodes). 2265 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2266 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone); 2267 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2268 if (failing()) return; 2269 } 2270 ConnectionGraph::do_analysis(this, &igvn); 2271 2272 if (failing()) return; 2273 2274 // Optimize out fields loads from scalar replaceable allocations. 2275 igvn.optimize(); 2276 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2277 2278 if (failing()) return; 2279 2280 if (congraph() != NULL && macro_count() > 0) { 2281 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2282 PhaseMacroExpand mexp(igvn); 2283 mexp.eliminate_macro_nodes(); 2284 igvn.set_delay_transform(false); 2285 2286 igvn.optimize(); 2287 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2288 2289 if (failing()) return; 2290 } 2291 } 2292 2293 // Loop transforms on the ideal graph. Range Check Elimination, 2294 // peeling, unrolling, etc. 2295 2296 // Set loop opts counter 2297 loop_opts_cnt = num_loop_opts(); 2298 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2299 { 2300 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2301 PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault); 2302 loop_opts_cnt--; 2303 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2304 if (failing()) return; 2305 } 2306 // Loop opts pass if partial peeling occurred in previous pass 2307 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2308 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2309 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2310 loop_opts_cnt--; 2311 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2312 if (failing()) return; 2313 } 2314 // Loop opts pass for loop-unrolling before CCP 2315 if(major_progress() && (loop_opts_cnt > 0)) { 2316 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2317 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2318 loop_opts_cnt--; 2319 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2320 } 2321 if (!failing()) { 2322 // Verify that last round of loop opts produced a valid graph 2323 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2324 PhaseIdealLoop::verify(igvn); 2325 } 2326 } 2327 if (failing()) return; 2328 2329 // Conditional Constant Propagation; 2330 PhaseCCP ccp( &igvn ); 2331 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2332 { 2333 TracePhase tp("ccp", &timers[_t_ccp]); 2334 ccp.do_transform(); 2335 } 2336 print_method(PHASE_CPP1, 2); 2337 2338 assert( true, "Break here to ccp.dump_old2new_map()"); 2339 2340 // Iterative Global Value Numbering, including ideal transforms 2341 { 2342 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2343 igvn = ccp; 2344 igvn.optimize(); 2345 } 2346 2347 print_method(PHASE_ITER_GVN2, 2); 2348 2349 if (failing()) return; 2350 2351 // Loop transforms on the ideal graph. Range Check Elimination, 2352 // peeling, unrolling, etc. 2353 if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) { 2354 return; 2355 } 2356 2357 #if INCLUDE_ZGC 2358 if (UseZGC) { 2359 ZBarrierSetC2::find_dominating_barriers(igvn); 2360 } 2361 #endif 2362 2363 if (failing()) return; 2364 2365 // Ensure that major progress is now clear 2366 C->clear_major_progress(); 2367 2368 { 2369 // Verify that all previous optimizations produced a valid graph 2370 // at least to this point, even if no loop optimizations were done. 2371 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2372 PhaseIdealLoop::verify(igvn); 2373 } 2374 2375 if (range_check_cast_count() > 0) { 2376 // No more loop optimizations. Remove all range check dependent CastIINodes. 2377 C->remove_range_check_casts(igvn); 2378 igvn.optimize(); 2379 } 2380 2381 #ifdef ASSERT 2382 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2383 bs->verify_gc_barriers(false); 2384 #endif 2385 2386 { 2387 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2388 PhaseMacroExpand mex(igvn); 2389 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2); 2390 if (mex.expand_macro_nodes()) { 2391 assert(failing(), "must bail out w/ explicit message"); 2392 return; 2393 } 2394 } 2395 2396 print_method(PHASE_BEFORE_BARRIER_EXPAND, 2); 2397 2398 #if INCLUDE_SHENANDOAHGC 2399 if (UseShenandoahGC && ((ShenandoahBarrierSetC2*)BarrierSet::barrier_set()->barrier_set_c2())->expand_barriers(this, igvn)) { 2400 assert(failing(), "must bail out w/ explicit message"); 2401 return; 2402 } 2403 #endif 2404 2405 if (opaque4_count() > 0) { 2406 C->remove_opaque4_nodes(igvn); 2407 igvn.optimize(); 2408 } 2409 2410 DEBUG_ONLY( _modified_nodes = NULL; ) 2411 } // (End scope of igvn; run destructor if necessary for asserts.) 2412 2413 process_print_inlining(); 2414 // A method with only infinite loops has no edges entering loops from root 2415 { 2416 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2417 if (final_graph_reshaping()) { 2418 assert(failing(), "must bail out w/ explicit message"); 2419 return; 2420 } 2421 } 2422 2423 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2424 } 2425 2426 2427 //------------------------------Code_Gen--------------------------------------- 2428 // Given a graph, generate code for it 2429 void Compile::Code_Gen() { 2430 if (failing()) { 2431 return; 2432 } 2433 2434 // Perform instruction selection. You might think we could reclaim Matcher 2435 // memory PDQ, but actually the Matcher is used in generating spill code. 2436 // Internals of the Matcher (including some VectorSets) must remain live 2437 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2438 // set a bit in reclaimed memory. 2439 2440 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2441 // nodes. Mapping is only valid at the root of each matched subtree. 2442 NOT_PRODUCT( verify_graph_edges(); ) 2443 2444 Matcher matcher; 2445 _matcher = &matcher; 2446 { 2447 TracePhase tp("matcher", &timers[_t_matcher]); 2448 matcher.match(); 2449 } 2450 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2451 // nodes. Mapping is only valid at the root of each matched subtree. 2452 NOT_PRODUCT( verify_graph_edges(); ) 2453 2454 // If you have too many nodes, or if matching has failed, bail out 2455 check_node_count(0, "out of nodes matching instructions"); 2456 if (failing()) { 2457 return; 2458 } 2459 2460 print_method(PHASE_MATCHING, 2); 2461 2462 // Build a proper-looking CFG 2463 PhaseCFG cfg(node_arena(), root(), matcher); 2464 _cfg = &cfg; 2465 { 2466 TracePhase tp("scheduler", &timers[_t_scheduler]); 2467 bool success = cfg.do_global_code_motion(); 2468 if (!success) { 2469 return; 2470 } 2471 2472 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2473 NOT_PRODUCT( verify_graph_edges(); ) 2474 debug_only( cfg.verify(); ) 2475 } 2476 2477 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2478 _regalloc = ®alloc; 2479 { 2480 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2481 // Perform register allocation. After Chaitin, use-def chains are 2482 // no longer accurate (at spill code) and so must be ignored. 2483 // Node->LRG->reg mappings are still accurate. 2484 _regalloc->Register_Allocate(); 2485 2486 // Bail out if the allocator builds too many nodes 2487 if (failing()) { 2488 return; 2489 } 2490 } 2491 2492 // Prior to register allocation we kept empty basic blocks in case the 2493 // the allocator needed a place to spill. After register allocation we 2494 // are not adding any new instructions. If any basic block is empty, we 2495 // can now safely remove it. 2496 { 2497 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2498 cfg.remove_empty_blocks(); 2499 if (do_freq_based_layout()) { 2500 PhaseBlockLayout layout(cfg); 2501 } else { 2502 cfg.set_loop_alignment(); 2503 } 2504 cfg.fixup_flow(); 2505 } 2506 2507 // Apply peephole optimizations 2508 if( OptoPeephole ) { 2509 TracePhase tp("peephole", &timers[_t_peephole]); 2510 PhasePeephole peep( _regalloc, cfg); 2511 peep.do_transform(); 2512 } 2513 2514 // Do late expand if CPU requires this. 2515 if (Matcher::require_postalloc_expand) { 2516 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2517 cfg.postalloc_expand(_regalloc); 2518 } 2519 2520 // Convert Nodes to instruction bits in a buffer 2521 { 2522 TraceTime tp("output", &timers[_t_output], CITime); 2523 Output(); 2524 } 2525 2526 print_method(PHASE_FINAL_CODE); 2527 2528 // He's dead, Jim. 2529 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2530 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2531 } 2532 2533 2534 //------------------------------dump_asm--------------------------------------- 2535 // Dump formatted assembly 2536 #ifndef PRODUCT 2537 void Compile::dump_asm(int *pcs, uint pc_limit) { 2538 bool cut_short = false; 2539 tty->print_cr("#"); 2540 tty->print("# "); _tf->dump(); tty->cr(); 2541 tty->print_cr("#"); 2542 2543 // For all blocks 2544 int pc = 0x0; // Program counter 2545 char starts_bundle = ' '; 2546 _regalloc->dump_frame(); 2547 2548 Node *n = NULL; 2549 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2550 if (VMThread::should_terminate()) { 2551 cut_short = true; 2552 break; 2553 } 2554 Block* block = _cfg->get_block(i); 2555 if (block->is_connector() && !Verbose) { 2556 continue; 2557 } 2558 n = block->head(); 2559 if (pcs && n->_idx < pc_limit) { 2560 tty->print("%3.3x ", pcs[n->_idx]); 2561 } else { 2562 tty->print(" "); 2563 } 2564 block->dump_head(_cfg); 2565 if (block->is_connector()) { 2566 tty->print_cr(" # Empty connector block"); 2567 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2568 tty->print_cr(" # Block is sole successor of call"); 2569 } 2570 2571 // For all instructions 2572 Node *delay = NULL; 2573 for (uint j = 0; j < block->number_of_nodes(); j++) { 2574 if (VMThread::should_terminate()) { 2575 cut_short = true; 2576 break; 2577 } 2578 n = block->get_node(j); 2579 if (valid_bundle_info(n)) { 2580 Bundle* bundle = node_bundling(n); 2581 if (bundle->used_in_unconditional_delay()) { 2582 delay = n; 2583 continue; 2584 } 2585 if (bundle->starts_bundle()) { 2586 starts_bundle = '+'; 2587 } 2588 } 2589 2590 if (WizardMode) { 2591 n->dump(); 2592 } 2593 2594 if( !n->is_Region() && // Dont print in the Assembly 2595 !n->is_Phi() && // a few noisely useless nodes 2596 !n->is_Proj() && 2597 !n->is_MachTemp() && 2598 !n->is_SafePointScalarObject() && 2599 !n->is_Catch() && // Would be nice to print exception table targets 2600 !n->is_MergeMem() && // Not very interesting 2601 !n->is_top() && // Debug info table constants 2602 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2603 ) { 2604 if (pcs && n->_idx < pc_limit) 2605 tty->print("%3.3x", pcs[n->_idx]); 2606 else 2607 tty->print(" "); 2608 tty->print(" %c ", starts_bundle); 2609 starts_bundle = ' '; 2610 tty->print("\t"); 2611 n->format(_regalloc, tty); 2612 tty->cr(); 2613 } 2614 2615 // If we have an instruction with a delay slot, and have seen a delay, 2616 // then back up and print it 2617 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2618 assert(delay != NULL, "no unconditional delay instruction"); 2619 if (WizardMode) delay->dump(); 2620 2621 if (node_bundling(delay)->starts_bundle()) 2622 starts_bundle = '+'; 2623 if (pcs && n->_idx < pc_limit) 2624 tty->print("%3.3x", pcs[n->_idx]); 2625 else 2626 tty->print(" "); 2627 tty->print(" %c ", starts_bundle); 2628 starts_bundle = ' '; 2629 tty->print("\t"); 2630 delay->format(_regalloc, tty); 2631 tty->cr(); 2632 delay = NULL; 2633 } 2634 2635 // Dump the exception table as well 2636 if( n->is_Catch() && (Verbose || WizardMode) ) { 2637 // Print the exception table for this offset 2638 _handler_table.print_subtable_for(pc); 2639 } 2640 } 2641 2642 if (pcs && n->_idx < pc_limit) 2643 tty->print_cr("%3.3x", pcs[n->_idx]); 2644 else 2645 tty->cr(); 2646 2647 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2648 2649 } // End of per-block dump 2650 tty->cr(); 2651 2652 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2653 } 2654 #endif 2655 2656 //------------------------------Final_Reshape_Counts--------------------------- 2657 // This class defines counters to help identify when a method 2658 // may/must be executed using hardware with only 24-bit precision. 2659 struct Final_Reshape_Counts : public StackObj { 2660 int _call_count; // count non-inlined 'common' calls 2661 int _float_count; // count float ops requiring 24-bit precision 2662 int _double_count; // count double ops requiring more precision 2663 int _java_call_count; // count non-inlined 'java' calls 2664 int _inner_loop_count; // count loops which need alignment 2665 VectorSet _visited; // Visitation flags 2666 Node_List _tests; // Set of IfNodes & PCTableNodes 2667 2668 Final_Reshape_Counts() : 2669 _call_count(0), _float_count(0), _double_count(0), 2670 _java_call_count(0), _inner_loop_count(0), 2671 _visited( Thread::current()->resource_area() ) { } 2672 2673 void inc_call_count () { _call_count ++; } 2674 void inc_float_count () { _float_count ++; } 2675 void inc_double_count() { _double_count++; } 2676 void inc_java_call_count() { _java_call_count++; } 2677 void inc_inner_loop_count() { _inner_loop_count++; } 2678 2679 int get_call_count () const { return _call_count ; } 2680 int get_float_count () const { return _float_count ; } 2681 int get_double_count() const { return _double_count; } 2682 int get_java_call_count() const { return _java_call_count; } 2683 int get_inner_loop_count() const { return _inner_loop_count; } 2684 }; 2685 2686 #ifdef ASSERT 2687 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2688 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2689 // Make sure the offset goes inside the instance layout. 2690 return k->contains_field_offset(tp->offset()); 2691 // Note that OffsetBot and OffsetTop are very negative. 2692 } 2693 #endif 2694 2695 // Eliminate trivially redundant StoreCMs and accumulate their 2696 // precedence edges. 2697 void Compile::eliminate_redundant_card_marks(Node* n) { 2698 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2699 if (n->in(MemNode::Address)->outcnt() > 1) { 2700 // There are multiple users of the same address so it might be 2701 // possible to eliminate some of the StoreCMs 2702 Node* mem = n->in(MemNode::Memory); 2703 Node* adr = n->in(MemNode::Address); 2704 Node* val = n->in(MemNode::ValueIn); 2705 Node* prev = n; 2706 bool done = false; 2707 // Walk the chain of StoreCMs eliminating ones that match. As 2708 // long as it's a chain of single users then the optimization is 2709 // safe. Eliminating partially redundant StoreCMs would require 2710 // cloning copies down the other paths. 2711 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2712 if (adr == mem->in(MemNode::Address) && 2713 val == mem->in(MemNode::ValueIn)) { 2714 // redundant StoreCM 2715 if (mem->req() > MemNode::OopStore) { 2716 // Hasn't been processed by this code yet. 2717 n->add_prec(mem->in(MemNode::OopStore)); 2718 } else { 2719 // Already converted to precedence edge 2720 for (uint i = mem->req(); i < mem->len(); i++) { 2721 // Accumulate any precedence edges 2722 if (mem->in(i) != NULL) { 2723 n->add_prec(mem->in(i)); 2724 } 2725 } 2726 // Everything above this point has been processed. 2727 done = true; 2728 } 2729 // Eliminate the previous StoreCM 2730 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2731 assert(mem->outcnt() == 0, "should be dead"); 2732 mem->disconnect_inputs(NULL, this); 2733 } else { 2734 prev = mem; 2735 } 2736 mem = prev->in(MemNode::Memory); 2737 } 2738 } 2739 } 2740 2741 //------------------------------final_graph_reshaping_impl---------------------- 2742 // Implement items 1-5 from final_graph_reshaping below. 2743 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2744 2745 if ( n->outcnt() == 0 ) return; // dead node 2746 uint nop = n->Opcode(); 2747 2748 // Check for 2-input instruction with "last use" on right input. 2749 // Swap to left input. Implements item (2). 2750 if( n->req() == 3 && // two-input instruction 2751 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2752 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2753 n->in(2)->outcnt() == 1 &&// right use IS a last use 2754 !n->in(2)->is_Con() ) { // right use is not a constant 2755 // Check for commutative opcode 2756 switch( nop ) { 2757 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2758 case Op_MaxI: case Op_MinI: 2759 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2760 case Op_AndL: case Op_XorL: case Op_OrL: 2761 case Op_AndI: case Op_XorI: case Op_OrI: { 2762 // Move "last use" input to left by swapping inputs 2763 n->swap_edges(1, 2); 2764 break; 2765 } 2766 default: 2767 break; 2768 } 2769 } 2770 2771 #ifdef ASSERT 2772 if( n->is_Mem() ) { 2773 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2774 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2775 // oop will be recorded in oop map if load crosses safepoint 2776 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2777 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2778 "raw memory operations should have control edge"); 2779 } 2780 if (n->is_MemBar()) { 2781 MemBarNode* mb = n->as_MemBar(); 2782 if (mb->trailing_store() || mb->trailing_load_store()) { 2783 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2784 Node* mem = mb->in(MemBarNode::Precedent); 2785 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2786 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2787 } else if (mb->leading()) { 2788 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2789 } 2790 } 2791 #endif 2792 // Count FPU ops and common calls, implements item (3) 2793 switch( nop ) { 2794 // Count all float operations that may use FPU 2795 case Op_AddF: 2796 case Op_SubF: 2797 case Op_MulF: 2798 case Op_DivF: 2799 case Op_NegF: 2800 case Op_ModF: 2801 case Op_ConvI2F: 2802 case Op_ConF: 2803 case Op_CmpF: 2804 case Op_CmpF3: 2805 // case Op_ConvL2F: // longs are split into 32-bit halves 2806 frc.inc_float_count(); 2807 break; 2808 2809 case Op_ConvF2D: 2810 case Op_ConvD2F: 2811 frc.inc_float_count(); 2812 frc.inc_double_count(); 2813 break; 2814 2815 // Count all double operations that may use FPU 2816 case Op_AddD: 2817 case Op_SubD: 2818 case Op_MulD: 2819 case Op_DivD: 2820 case Op_NegD: 2821 case Op_ModD: 2822 case Op_ConvI2D: 2823 case Op_ConvD2I: 2824 // case Op_ConvL2D: // handled by leaf call 2825 // case Op_ConvD2L: // handled by leaf call 2826 case Op_ConD: 2827 case Op_CmpD: 2828 case Op_CmpD3: 2829 frc.inc_double_count(); 2830 break; 2831 case Op_Opaque1: // Remove Opaque Nodes before matching 2832 case Op_Opaque2: // Remove Opaque Nodes before matching 2833 case Op_Opaque3: 2834 n->subsume_by(n->in(1), this); 2835 break; 2836 case Op_CallStaticJava: 2837 case Op_CallJava: 2838 case Op_CallDynamicJava: 2839 frc.inc_java_call_count(); // Count java call site; 2840 case Op_CallRuntime: 2841 case Op_CallLeaf: 2842 case Op_CallLeafNoFP: { 2843 assert (n->is_Call(), ""); 2844 CallNode *call = n->as_Call(); 2845 #if INCLUDE_SHENANDOAHGC 2846 if (UseShenandoahGC && ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) { 2847 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt(); 2848 if (call->req() > cnt) { 2849 assert(call->req() == cnt+1, "only one extra input"); 2850 Node* addp = call->in(cnt); 2851 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?"); 2852 call->del_req(cnt); 2853 } 2854 } 2855 #endif 2856 // Count call sites where the FP mode bit would have to be flipped. 2857 // Do not count uncommon runtime calls: 2858 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2859 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2860 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 2861 frc.inc_call_count(); // Count the call site 2862 } else { // See if uncommon argument is shared 2863 Node *n = call->in(TypeFunc::Parms); 2864 int nop = n->Opcode(); 2865 // Clone shared simple arguments to uncommon calls, item (1). 2866 if (n->outcnt() > 1 && 2867 !n->is_Proj() && 2868 nop != Op_CreateEx && 2869 nop != Op_CheckCastPP && 2870 nop != Op_DecodeN && 2871 nop != Op_DecodeNKlass && 2872 !n->is_Mem() && 2873 !n->is_Phi()) { 2874 Node *x = n->clone(); 2875 call->set_req(TypeFunc::Parms, x); 2876 } 2877 } 2878 break; 2879 } 2880 2881 case Op_StoreD: 2882 case Op_LoadD: 2883 case Op_LoadD_unaligned: 2884 frc.inc_double_count(); 2885 goto handle_mem; 2886 case Op_StoreF: 2887 case Op_LoadF: 2888 frc.inc_float_count(); 2889 goto handle_mem; 2890 2891 case Op_StoreCM: 2892 { 2893 // Convert OopStore dependence into precedence edge 2894 Node* prec = n->in(MemNode::OopStore); 2895 n->del_req(MemNode::OopStore); 2896 n->add_prec(prec); 2897 eliminate_redundant_card_marks(n); 2898 } 2899 2900 // fall through 2901 2902 case Op_StoreB: 2903 case Op_StoreC: 2904 case Op_StorePConditional: 2905 case Op_StoreI: 2906 case Op_StoreL: 2907 case Op_StoreIConditional: 2908 case Op_StoreLConditional: 2909 case Op_CompareAndSwapB: 2910 case Op_CompareAndSwapS: 2911 case Op_CompareAndSwapI: 2912 case Op_CompareAndSwapL: 2913 case Op_CompareAndSwapP: 2914 case Op_CompareAndSwapN: 2915 case Op_WeakCompareAndSwapB: 2916 case Op_WeakCompareAndSwapS: 2917 case Op_WeakCompareAndSwapI: 2918 case Op_WeakCompareAndSwapL: 2919 case Op_WeakCompareAndSwapP: 2920 case Op_WeakCompareAndSwapN: 2921 case Op_CompareAndExchangeB: 2922 case Op_CompareAndExchangeS: 2923 case Op_CompareAndExchangeI: 2924 case Op_CompareAndExchangeL: 2925 case Op_CompareAndExchangeP: 2926 case Op_CompareAndExchangeN: 2927 case Op_GetAndAddS: 2928 case Op_GetAndAddB: 2929 case Op_GetAndAddI: 2930 case Op_GetAndAddL: 2931 case Op_GetAndSetS: 2932 case Op_GetAndSetB: 2933 case Op_GetAndSetI: 2934 case Op_GetAndSetL: 2935 case Op_GetAndSetP: 2936 case Op_GetAndSetN: 2937 case Op_StoreP: 2938 case Op_StoreN: 2939 case Op_StoreNKlass: 2940 case Op_LoadB: 2941 case Op_LoadUB: 2942 case Op_LoadUS: 2943 case Op_LoadI: 2944 case Op_LoadKlass: 2945 case Op_LoadNKlass: 2946 case Op_LoadL: 2947 case Op_LoadL_unaligned: 2948 case Op_LoadPLocked: 2949 case Op_LoadP: 2950 #if INCLUDE_ZGC 2951 case Op_LoadBarrierSlowReg: 2952 case Op_LoadBarrierWeakSlowReg: 2953 #endif 2954 case Op_LoadN: 2955 case Op_LoadRange: 2956 case Op_LoadS: { 2957 handle_mem: 2958 #ifdef ASSERT 2959 if( VerifyOptoOopOffsets ) { 2960 assert( n->is_Mem(), "" ); 2961 MemNode *mem = (MemNode*)n; 2962 // Check to see if address types have grounded out somehow. 2963 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2964 assert( !tp || oop_offset_is_sane(tp), "" ); 2965 } 2966 #endif 2967 break; 2968 } 2969 2970 case Op_AddP: { // Assert sane base pointers 2971 Node *addp = n->in(AddPNode::Address); 2972 assert( !addp->is_AddP() || 2973 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2974 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2975 "Base pointers must match (addp %u)", addp->_idx ); 2976 #ifdef _LP64 2977 if ((UseCompressedOops || UseCompressedClassPointers) && 2978 addp->Opcode() == Op_ConP && 2979 addp == n->in(AddPNode::Base) && 2980 n->in(AddPNode::Offset)->is_Con()) { 2981 // If the transformation of ConP to ConN+DecodeN is beneficial depends 2982 // on the platform and on the compressed oops mode. 2983 // Use addressing with narrow klass to load with offset on x86. 2984 // Some platforms can use the constant pool to load ConP. 2985 // Do this transformation here since IGVN will convert ConN back to ConP. 2986 const Type* t = addp->bottom_type(); 2987 bool is_oop = t->isa_oopptr() != NULL; 2988 bool is_klass = t->isa_klassptr() != NULL; 2989 2990 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 2991 (is_klass && Matcher::const_klass_prefer_decode())) { 2992 Node* nn = NULL; 2993 2994 int op = is_oop ? Op_ConN : Op_ConNKlass; 2995 2996 // Look for existing ConN node of the same exact type. 2997 Node* r = root(); 2998 uint cnt = r->outcnt(); 2999 for (uint i = 0; i < cnt; i++) { 3000 Node* m = r->raw_out(i); 3001 if (m!= NULL && m->Opcode() == op && 3002 m->bottom_type()->make_ptr() == t) { 3003 nn = m; 3004 break; 3005 } 3006 } 3007 if (nn != NULL) { 3008 // Decode a narrow oop to match address 3009 // [R12 + narrow_oop_reg<<3 + offset] 3010 if (is_oop) { 3011 nn = new DecodeNNode(nn, t); 3012 } else { 3013 nn = new DecodeNKlassNode(nn, t); 3014 } 3015 // Check for succeeding AddP which uses the same Base. 3016 // Otherwise we will run into the assertion above when visiting that guy. 3017 for (uint i = 0; i < n->outcnt(); ++i) { 3018 Node *out_i = n->raw_out(i); 3019 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3020 out_i->set_req(AddPNode::Base, nn); 3021 #ifdef ASSERT 3022 for (uint j = 0; j < out_i->outcnt(); ++j) { 3023 Node *out_j = out_i->raw_out(j); 3024 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3025 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3026 } 3027 #endif 3028 } 3029 } 3030 n->set_req(AddPNode::Base, nn); 3031 n->set_req(AddPNode::Address, nn); 3032 if (addp->outcnt() == 0) { 3033 addp->disconnect_inputs(NULL, this); 3034 } 3035 } 3036 } 3037 } 3038 #endif 3039 // platform dependent reshaping of the address expression 3040 reshape_address(n->as_AddP()); 3041 break; 3042 } 3043 3044 case Op_CastPP: { 3045 // Remove CastPP nodes to gain more freedom during scheduling but 3046 // keep the dependency they encode as control or precedence edges 3047 // (if control is set already) on memory operations. Some CastPP 3048 // nodes don't have a control (don't carry a dependency): skip 3049 // those. 3050 if (n->in(0) != NULL) { 3051 ResourceMark rm; 3052 Unique_Node_List wq; 3053 wq.push(n); 3054 for (uint next = 0; next < wq.size(); ++next) { 3055 Node *m = wq.at(next); 3056 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3057 Node* use = m->fast_out(i); 3058 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3059 use->ensure_control_or_add_prec(n->in(0)); 3060 } else { 3061 switch(use->Opcode()) { 3062 case Op_AddP: 3063 case Op_DecodeN: 3064 case Op_DecodeNKlass: 3065 case Op_CheckCastPP: 3066 case Op_CastPP: 3067 wq.push(use); 3068 break; 3069 } 3070 } 3071 } 3072 } 3073 } 3074 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3075 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3076 Node* in1 = n->in(1); 3077 const Type* t = n->bottom_type(); 3078 Node* new_in1 = in1->clone(); 3079 new_in1->as_DecodeN()->set_type(t); 3080 3081 if (!Matcher::narrow_oop_use_complex_address()) { 3082 // 3083 // x86, ARM and friends can handle 2 adds in addressing mode 3084 // and Matcher can fold a DecodeN node into address by using 3085 // a narrow oop directly and do implicit NULL check in address: 3086 // 3087 // [R12 + narrow_oop_reg<<3 + offset] 3088 // NullCheck narrow_oop_reg 3089 // 3090 // On other platforms (Sparc) we have to keep new DecodeN node and 3091 // use it to do implicit NULL check in address: 3092 // 3093 // decode_not_null narrow_oop_reg, base_reg 3094 // [base_reg + offset] 3095 // NullCheck base_reg 3096 // 3097 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3098 // to keep the information to which NULL check the new DecodeN node 3099 // corresponds to use it as value in implicit_null_check(). 3100 // 3101 new_in1->set_req(0, n->in(0)); 3102 } 3103 3104 n->subsume_by(new_in1, this); 3105 if (in1->outcnt() == 0) { 3106 in1->disconnect_inputs(NULL, this); 3107 } 3108 } else { 3109 n->subsume_by(n->in(1), this); 3110 if (n->outcnt() == 0) { 3111 n->disconnect_inputs(NULL, this); 3112 } 3113 } 3114 break; 3115 } 3116 #ifdef _LP64 3117 case Op_CmpP: 3118 // Do this transformation here to preserve CmpPNode::sub() and 3119 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3120 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3121 Node* in1 = n->in(1); 3122 Node* in2 = n->in(2); 3123 if (!in1->is_DecodeNarrowPtr()) { 3124 in2 = in1; 3125 in1 = n->in(2); 3126 } 3127 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3128 3129 Node* new_in2 = NULL; 3130 if (in2->is_DecodeNarrowPtr()) { 3131 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3132 new_in2 = in2->in(1); 3133 } else if (in2->Opcode() == Op_ConP) { 3134 const Type* t = in2->bottom_type(); 3135 if (t == TypePtr::NULL_PTR) { 3136 assert(in1->is_DecodeN(), "compare klass to null?"); 3137 // Don't convert CmpP null check into CmpN if compressed 3138 // oops implicit null check is not generated. 3139 // This will allow to generate normal oop implicit null check. 3140 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3141 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3142 // 3143 // This transformation together with CastPP transformation above 3144 // will generated code for implicit NULL checks for compressed oops. 3145 // 3146 // The original code after Optimize() 3147 // 3148 // LoadN memory, narrow_oop_reg 3149 // decode narrow_oop_reg, base_reg 3150 // CmpP base_reg, NULL 3151 // CastPP base_reg // NotNull 3152 // Load [base_reg + offset], val_reg 3153 // 3154 // after these transformations will be 3155 // 3156 // LoadN memory, narrow_oop_reg 3157 // CmpN narrow_oop_reg, NULL 3158 // decode_not_null narrow_oop_reg, base_reg 3159 // Load [base_reg + offset], val_reg 3160 // 3161 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3162 // since narrow oops can be used in debug info now (see the code in 3163 // final_graph_reshaping_walk()). 3164 // 3165 // At the end the code will be matched to 3166 // on x86: 3167 // 3168 // Load_narrow_oop memory, narrow_oop_reg 3169 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3170 // NullCheck narrow_oop_reg 3171 // 3172 // and on sparc: 3173 // 3174 // Load_narrow_oop memory, narrow_oop_reg 3175 // decode_not_null narrow_oop_reg, base_reg 3176 // Load [base_reg + offset], val_reg 3177 // NullCheck base_reg 3178 // 3179 } else if (t->isa_oopptr()) { 3180 new_in2 = ConNode::make(t->make_narrowoop()); 3181 } else if (t->isa_klassptr()) { 3182 new_in2 = ConNode::make(t->make_narrowklass()); 3183 } 3184 } 3185 if (new_in2 != NULL) { 3186 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3187 n->subsume_by(cmpN, this); 3188 if (in1->outcnt() == 0) { 3189 in1->disconnect_inputs(NULL, this); 3190 } 3191 if (in2->outcnt() == 0) { 3192 in2->disconnect_inputs(NULL, this); 3193 } 3194 } 3195 } 3196 break; 3197 3198 case Op_DecodeN: 3199 case Op_DecodeNKlass: 3200 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3201 // DecodeN could be pinned when it can't be fold into 3202 // an address expression, see the code for Op_CastPP above. 3203 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3204 break; 3205 3206 case Op_EncodeP: 3207 case Op_EncodePKlass: { 3208 Node* in1 = n->in(1); 3209 if (in1->is_DecodeNarrowPtr()) { 3210 n->subsume_by(in1->in(1), this); 3211 } else if (in1->Opcode() == Op_ConP) { 3212 const Type* t = in1->bottom_type(); 3213 if (t == TypePtr::NULL_PTR) { 3214 assert(t->isa_oopptr(), "null klass?"); 3215 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3216 } else if (t->isa_oopptr()) { 3217 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3218 } else if (t->isa_klassptr()) { 3219 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3220 } 3221 } 3222 if (in1->outcnt() == 0) { 3223 in1->disconnect_inputs(NULL, this); 3224 } 3225 break; 3226 } 3227 3228 case Op_Proj: { 3229 if (OptimizeStringConcat) { 3230 ProjNode* p = n->as_Proj(); 3231 if (p->_is_io_use) { 3232 // Separate projections were used for the exception path which 3233 // are normally removed by a late inline. If it wasn't inlined 3234 // then they will hang around and should just be replaced with 3235 // the original one. 3236 Node* proj = NULL; 3237 // Replace with just one 3238 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3239 Node *use = i.get(); 3240 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3241 proj = use; 3242 break; 3243 } 3244 } 3245 assert(proj != NULL, "must be found"); 3246 p->subsume_by(proj, this); 3247 } 3248 } 3249 break; 3250 } 3251 3252 case Op_Phi: 3253 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3254 // The EncodeP optimization may create Phi with the same edges 3255 // for all paths. It is not handled well by Register Allocator. 3256 Node* unique_in = n->in(1); 3257 assert(unique_in != NULL, ""); 3258 uint cnt = n->req(); 3259 for (uint i = 2; i < cnt; i++) { 3260 Node* m = n->in(i); 3261 assert(m != NULL, ""); 3262 if (unique_in != m) 3263 unique_in = NULL; 3264 } 3265 if (unique_in != NULL) { 3266 n->subsume_by(unique_in, this); 3267 } 3268 } 3269 break; 3270 3271 #endif 3272 3273 #ifdef ASSERT 3274 case Op_CastII: 3275 // Verify that all range check dependent CastII nodes were removed. 3276 if (n->isa_CastII()->has_range_check()) { 3277 n->dump(3); 3278 assert(false, "Range check dependent CastII node was not removed"); 3279 } 3280 break; 3281 #endif 3282 3283 case Op_ModI: 3284 if (UseDivMod) { 3285 // Check if a%b and a/b both exist 3286 Node* d = n->find_similar(Op_DivI); 3287 if (d) { 3288 // Replace them with a fused divmod if supported 3289 if (Matcher::has_match_rule(Op_DivModI)) { 3290 DivModINode* divmod = DivModINode::make(n); 3291 d->subsume_by(divmod->div_proj(), this); 3292 n->subsume_by(divmod->mod_proj(), this); 3293 } else { 3294 // replace a%b with a-((a/b)*b) 3295 Node* mult = new MulINode(d, d->in(2)); 3296 Node* sub = new SubINode(d->in(1), mult); 3297 n->subsume_by(sub, this); 3298 } 3299 } 3300 } 3301 break; 3302 3303 case Op_ModL: 3304 if (UseDivMod) { 3305 // Check if a%b and a/b both exist 3306 Node* d = n->find_similar(Op_DivL); 3307 if (d) { 3308 // Replace them with a fused divmod if supported 3309 if (Matcher::has_match_rule(Op_DivModL)) { 3310 DivModLNode* divmod = DivModLNode::make(n); 3311 d->subsume_by(divmod->div_proj(), this); 3312 n->subsume_by(divmod->mod_proj(), this); 3313 } else { 3314 // replace a%b with a-((a/b)*b) 3315 Node* mult = new MulLNode(d, d->in(2)); 3316 Node* sub = new SubLNode(d->in(1), mult); 3317 n->subsume_by(sub, this); 3318 } 3319 } 3320 } 3321 break; 3322 3323 case Op_LoadVector: 3324 case Op_StoreVector: 3325 break; 3326 3327 case Op_AddReductionVI: 3328 case Op_AddReductionVL: 3329 case Op_AddReductionVF: 3330 case Op_AddReductionVD: 3331 case Op_MulReductionVI: 3332 case Op_MulReductionVL: 3333 case Op_MulReductionVF: 3334 case Op_MulReductionVD: 3335 break; 3336 3337 case Op_PackB: 3338 case Op_PackS: 3339 case Op_PackI: 3340 case Op_PackF: 3341 case Op_PackL: 3342 case Op_PackD: 3343 if (n->req()-1 > 2) { 3344 // Replace many operand PackNodes with a binary tree for matching 3345 PackNode* p = (PackNode*) n; 3346 Node* btp = p->binary_tree_pack(1, n->req()); 3347 n->subsume_by(btp, this); 3348 } 3349 break; 3350 case Op_Loop: 3351 case Op_CountedLoop: 3352 case Op_OuterStripMinedLoop: 3353 if (n->as_Loop()->is_inner_loop()) { 3354 frc.inc_inner_loop_count(); 3355 } 3356 n->as_Loop()->verify_strip_mined(0); 3357 break; 3358 case Op_LShiftI: 3359 case Op_RShiftI: 3360 case Op_URShiftI: 3361 case Op_LShiftL: 3362 case Op_RShiftL: 3363 case Op_URShiftL: 3364 if (Matcher::need_masked_shift_count) { 3365 // The cpu's shift instructions don't restrict the count to the 3366 // lower 5/6 bits. We need to do the masking ourselves. 3367 Node* in2 = n->in(2); 3368 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3369 const TypeInt* t = in2->find_int_type(); 3370 if (t != NULL && t->is_con()) { 3371 juint shift = t->get_con(); 3372 if (shift > mask) { // Unsigned cmp 3373 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3374 } 3375 } else { 3376 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3377 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3378 n->set_req(2, shift); 3379 } 3380 } 3381 if (in2->outcnt() == 0) { // Remove dead node 3382 in2->disconnect_inputs(NULL, this); 3383 } 3384 } 3385 break; 3386 case Op_MemBarStoreStore: 3387 case Op_MemBarRelease: 3388 // Break the link with AllocateNode: it is no longer useful and 3389 // confuses register allocation. 3390 if (n->req() > MemBarNode::Precedent) { 3391 n->set_req(MemBarNode::Precedent, top()); 3392 } 3393 break; 3394 case Op_MemBarAcquire: { 3395 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3396 // At parse time, the trailing MemBarAcquire for a volatile load 3397 // is created with an edge to the load. After optimizations, 3398 // that input may be a chain of Phis. If those phis have no 3399 // other use, then the MemBarAcquire keeps them alive and 3400 // register allocation can be confused. 3401 ResourceMark rm; 3402 Unique_Node_List wq; 3403 wq.push(n->in(MemBarNode::Precedent)); 3404 n->set_req(MemBarNode::Precedent, top()); 3405 while (wq.size() > 0) { 3406 Node* m = wq.pop(); 3407 if (m->outcnt() == 0) { 3408 for (uint j = 0; j < m->req(); j++) { 3409 Node* in = m->in(j); 3410 if (in != NULL) { 3411 wq.push(in); 3412 } 3413 } 3414 m->disconnect_inputs(NULL, this); 3415 } 3416 } 3417 } 3418 break; 3419 } 3420 #if INCLUDE_SHENANDOAHGC 3421 case Op_ShenandoahCompareAndSwapP: 3422 case Op_ShenandoahCompareAndSwapN: 3423 case Op_ShenandoahWeakCompareAndSwapN: 3424 case Op_ShenandoahWeakCompareAndSwapP: 3425 case Op_ShenandoahCompareAndExchangeP: 3426 case Op_ShenandoahCompareAndExchangeN: 3427 #ifdef ASSERT 3428 if( VerifyOptoOopOffsets ) { 3429 MemNode* mem = n->as_Mem(); 3430 // Check to see if address types have grounded out somehow. 3431 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 3432 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 3433 bool oop_offset_is_sane = k->contains_field_offset(tp->offset()); 3434 assert( !tp || oop_offset_is_sane, "" ); 3435 } 3436 #endif 3437 break; 3438 case Op_ShenandoahLoadReferenceBarrier: 3439 assert(false, "should have been expanded already"); 3440 break; 3441 #endif 3442 case Op_RangeCheck: { 3443 RangeCheckNode* rc = n->as_RangeCheck(); 3444 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3445 n->subsume_by(iff, this); 3446 frc._tests.push(iff); 3447 break; 3448 } 3449 case Op_ConvI2L: { 3450 if (!Matcher::convi2l_type_required) { 3451 // Code generation on some platforms doesn't need accurate 3452 // ConvI2L types. Widening the type can help remove redundant 3453 // address computations. 3454 n->as_Type()->set_type(TypeLong::INT); 3455 ResourceMark rm; 3456 Unique_Node_List wq; 3457 wq.push(n); 3458 for (uint next = 0; next < wq.size(); next++) { 3459 Node *m = wq.at(next); 3460 3461 for(;;) { 3462 // Loop over all nodes with identical inputs edges as m 3463 Node* k = m->find_similar(m->Opcode()); 3464 if (k == NULL) { 3465 break; 3466 } 3467 // Push their uses so we get a chance to remove node made 3468 // redundant 3469 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3470 Node* u = k->fast_out(i); 3471 if (u->Opcode() == Op_LShiftL || 3472 u->Opcode() == Op_AddL || 3473 u->Opcode() == Op_SubL || 3474 u->Opcode() == Op_AddP) { 3475 wq.push(u); 3476 } 3477 } 3478 // Replace all nodes with identical edges as m with m 3479 k->subsume_by(m, this); 3480 } 3481 } 3482 } 3483 break; 3484 } 3485 case Op_CmpUL: { 3486 if (!Matcher::has_match_rule(Op_CmpUL)) { 3487 // No support for unsigned long comparisons 3488 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3489 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3490 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3491 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3492 Node* andl = new AndLNode(orl, remove_sign_mask); 3493 Node* cmp = new CmpLNode(andl, n->in(2)); 3494 n->subsume_by(cmp, this); 3495 } 3496 break; 3497 } 3498 default: 3499 assert( !n->is_Call(), "" ); 3500 assert( !n->is_Mem(), "" ); 3501 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3502 break; 3503 } 3504 3505 // Collect CFG split points 3506 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3507 frc._tests.push(n); 3508 } 3509 } 3510 3511 //------------------------------final_graph_reshaping_walk--------------------- 3512 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3513 // requires that the walk visits a node's inputs before visiting the node. 3514 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3515 ResourceArea *area = Thread::current()->resource_area(); 3516 Unique_Node_List sfpt(area); 3517 3518 frc._visited.set(root->_idx); // first, mark node as visited 3519 uint cnt = root->req(); 3520 Node *n = root; 3521 uint i = 0; 3522 while (true) { 3523 if (i < cnt) { 3524 // Place all non-visited non-null inputs onto stack 3525 Node* m = n->in(i); 3526 ++i; 3527 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3528 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3529 // compute worst case interpreter size in case of a deoptimization 3530 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3531 3532 sfpt.push(m); 3533 } 3534 cnt = m->req(); 3535 nstack.push(n, i); // put on stack parent and next input's index 3536 n = m; 3537 i = 0; 3538 } 3539 } else { 3540 // Now do post-visit work 3541 final_graph_reshaping_impl( n, frc ); 3542 if (nstack.is_empty()) 3543 break; // finished 3544 n = nstack.node(); // Get node from stack 3545 cnt = n->req(); 3546 i = nstack.index(); 3547 nstack.pop(); // Shift to the next node on stack 3548 } 3549 } 3550 3551 // Skip next transformation if compressed oops are not used. 3552 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3553 (!UseCompressedOops && !UseCompressedClassPointers)) 3554 return; 3555 3556 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3557 // It could be done for an uncommon traps or any safepoints/calls 3558 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3559 while (sfpt.size() > 0) { 3560 n = sfpt.pop(); 3561 JVMState *jvms = n->as_SafePoint()->jvms(); 3562 assert(jvms != NULL, "sanity"); 3563 int start = jvms->debug_start(); 3564 int end = n->req(); 3565 bool is_uncommon = (n->is_CallStaticJava() && 3566 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3567 for (int j = start; j < end; j++) { 3568 Node* in = n->in(j); 3569 if (in->is_DecodeNarrowPtr()) { 3570 bool safe_to_skip = true; 3571 if (!is_uncommon ) { 3572 // Is it safe to skip? 3573 for (uint i = 0; i < in->outcnt(); i++) { 3574 Node* u = in->raw_out(i); 3575 if (!u->is_SafePoint() || 3576 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3577 safe_to_skip = false; 3578 } 3579 } 3580 } 3581 if (safe_to_skip) { 3582 n->set_req(j, in->in(1)); 3583 } 3584 if (in->outcnt() == 0) { 3585 in->disconnect_inputs(NULL, this); 3586 } 3587 } 3588 } 3589 } 3590 } 3591 3592 //------------------------------final_graph_reshaping-------------------------- 3593 // Final Graph Reshaping. 3594 // 3595 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3596 // and not commoned up and forced early. Must come after regular 3597 // optimizations to avoid GVN undoing the cloning. Clone constant 3598 // inputs to Loop Phis; these will be split by the allocator anyways. 3599 // Remove Opaque nodes. 3600 // (2) Move last-uses by commutative operations to the left input to encourage 3601 // Intel update-in-place two-address operations and better register usage 3602 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3603 // calls canonicalizing them back. 3604 // (3) Count the number of double-precision FP ops, single-precision FP ops 3605 // and call sites. On Intel, we can get correct rounding either by 3606 // forcing singles to memory (requires extra stores and loads after each 3607 // FP bytecode) or we can set a rounding mode bit (requires setting and 3608 // clearing the mode bit around call sites). The mode bit is only used 3609 // if the relative frequency of single FP ops to calls is low enough. 3610 // This is a key transform for SPEC mpeg_audio. 3611 // (4) Detect infinite loops; blobs of code reachable from above but not 3612 // below. Several of the Code_Gen algorithms fail on such code shapes, 3613 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3614 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3615 // Detection is by looking for IfNodes where only 1 projection is 3616 // reachable from below or CatchNodes missing some targets. 3617 // (5) Assert for insane oop offsets in debug mode. 3618 3619 bool Compile::final_graph_reshaping() { 3620 // an infinite loop may have been eliminated by the optimizer, 3621 // in which case the graph will be empty. 3622 if (root()->req() == 1) { 3623 record_method_not_compilable("trivial infinite loop"); 3624 return true; 3625 } 3626 3627 // Expensive nodes have their control input set to prevent the GVN 3628 // from freely commoning them. There's no GVN beyond this point so 3629 // no need to keep the control input. We want the expensive nodes to 3630 // be freely moved to the least frequent code path by gcm. 3631 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3632 for (int i = 0; i < expensive_count(); i++) { 3633 _expensive_nodes->at(i)->set_req(0, NULL); 3634 } 3635 3636 Final_Reshape_Counts frc; 3637 3638 // Visit everybody reachable! 3639 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3640 Node_Stack nstack(live_nodes() >> 1); 3641 final_graph_reshaping_walk(nstack, root(), frc); 3642 3643 // Check for unreachable (from below) code (i.e., infinite loops). 3644 for( uint i = 0; i < frc._tests.size(); i++ ) { 3645 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3646 // Get number of CFG targets. 3647 // Note that PCTables include exception targets after calls. 3648 uint required_outcnt = n->required_outcnt(); 3649 if (n->outcnt() != required_outcnt) { 3650 // Check for a few special cases. Rethrow Nodes never take the 3651 // 'fall-thru' path, so expected kids is 1 less. 3652 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3653 if (n->in(0)->in(0)->is_Call()) { 3654 CallNode *call = n->in(0)->in(0)->as_Call(); 3655 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3656 required_outcnt--; // Rethrow always has 1 less kid 3657 } else if (call->req() > TypeFunc::Parms && 3658 call->is_CallDynamicJava()) { 3659 // Check for null receiver. In such case, the optimizer has 3660 // detected that the virtual call will always result in a null 3661 // pointer exception. The fall-through projection of this CatchNode 3662 // will not be populated. 3663 Node *arg0 = call->in(TypeFunc::Parms); 3664 if (arg0->is_Type() && 3665 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3666 required_outcnt--; 3667 } 3668 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3669 call->req() > TypeFunc::Parms+1 && 3670 call->is_CallStaticJava()) { 3671 // Check for negative array length. In such case, the optimizer has 3672 // detected that the allocation attempt will always result in an 3673 // exception. There is no fall-through projection of this CatchNode . 3674 Node *arg1 = call->in(TypeFunc::Parms+1); 3675 if (arg1->is_Type() && 3676 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3677 required_outcnt--; 3678 } 3679 } 3680 } 3681 } 3682 // Recheck with a better notion of 'required_outcnt' 3683 if (n->outcnt() != required_outcnt) { 3684 record_method_not_compilable("malformed control flow"); 3685 return true; // Not all targets reachable! 3686 } 3687 } 3688 // Check that I actually visited all kids. Unreached kids 3689 // must be infinite loops. 3690 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3691 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3692 record_method_not_compilable("infinite loop"); 3693 return true; // Found unvisited kid; must be unreach 3694 } 3695 3696 // Here so verification code in final_graph_reshaping_walk() 3697 // always see an OuterStripMinedLoopEnd 3698 if (n->is_OuterStripMinedLoopEnd()) { 3699 IfNode* init_iff = n->as_If(); 3700 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3701 n->subsume_by(iff, this); 3702 } 3703 } 3704 3705 // If original bytecodes contained a mixture of floats and doubles 3706 // check if the optimizer has made it homogenous, item (3). 3707 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3708 frc.get_float_count() > 32 && 3709 frc.get_double_count() == 0 && 3710 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3711 set_24_bit_selection_and_mode( false, true ); 3712 } 3713 3714 set_java_calls(frc.get_java_call_count()); 3715 set_inner_loops(frc.get_inner_loop_count()); 3716 3717 // No infinite loops, no reason to bail out. 3718 return false; 3719 } 3720 3721 //-----------------------------too_many_traps---------------------------------- 3722 // Report if there are too many traps at the current method and bci. 3723 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3724 bool Compile::too_many_traps(ciMethod* method, 3725 int bci, 3726 Deoptimization::DeoptReason reason) { 3727 ciMethodData* md = method->method_data(); 3728 if (md->is_empty()) { 3729 // Assume the trap has not occurred, or that it occurred only 3730 // because of a transient condition during start-up in the interpreter. 3731 return false; 3732 } 3733 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3734 if (md->has_trap_at(bci, m, reason) != 0) { 3735 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3736 // Also, if there are multiple reasons, or if there is no per-BCI record, 3737 // assume the worst. 3738 if (log()) 3739 log()->elem("observe trap='%s' count='%d'", 3740 Deoptimization::trap_reason_name(reason), 3741 md->trap_count(reason)); 3742 return true; 3743 } else { 3744 // Ignore method/bci and see if there have been too many globally. 3745 return too_many_traps(reason, md); 3746 } 3747 } 3748 3749 // Less-accurate variant which does not require a method and bci. 3750 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3751 ciMethodData* logmd) { 3752 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3753 // Too many traps globally. 3754 // Note that we use cumulative trap_count, not just md->trap_count. 3755 if (log()) { 3756 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3757 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3758 Deoptimization::trap_reason_name(reason), 3759 mcount, trap_count(reason)); 3760 } 3761 return true; 3762 } else { 3763 // The coast is clear. 3764 return false; 3765 } 3766 } 3767 3768 //--------------------------too_many_recompiles-------------------------------- 3769 // Report if there are too many recompiles at the current method and bci. 3770 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3771 // Is not eager to return true, since this will cause the compiler to use 3772 // Action_none for a trap point, to avoid too many recompilations. 3773 bool Compile::too_many_recompiles(ciMethod* method, 3774 int bci, 3775 Deoptimization::DeoptReason reason) { 3776 ciMethodData* md = method->method_data(); 3777 if (md->is_empty()) { 3778 // Assume the trap has not occurred, or that it occurred only 3779 // because of a transient condition during start-up in the interpreter. 3780 return false; 3781 } 3782 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3783 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3784 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3785 Deoptimization::DeoptReason per_bc_reason 3786 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3787 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3788 if ((per_bc_reason == Deoptimization::Reason_none 3789 || md->has_trap_at(bci, m, reason) != 0) 3790 // The trap frequency measure we care about is the recompile count: 3791 && md->trap_recompiled_at(bci, m) 3792 && md->overflow_recompile_count() >= bc_cutoff) { 3793 // Do not emit a trap here if it has already caused recompilations. 3794 // Also, if there are multiple reasons, or if there is no per-BCI record, 3795 // assume the worst. 3796 if (log()) 3797 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3798 Deoptimization::trap_reason_name(reason), 3799 md->trap_count(reason), 3800 md->overflow_recompile_count()); 3801 return true; 3802 } else if (trap_count(reason) != 0 3803 && decompile_count() >= m_cutoff) { 3804 // Too many recompiles globally, and we have seen this sort of trap. 3805 // Use cumulative decompile_count, not just md->decompile_count. 3806 if (log()) 3807 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3808 Deoptimization::trap_reason_name(reason), 3809 md->trap_count(reason), trap_count(reason), 3810 md->decompile_count(), decompile_count()); 3811 return true; 3812 } else { 3813 // The coast is clear. 3814 return false; 3815 } 3816 } 3817 3818 // Compute when not to trap. Used by matching trap based nodes and 3819 // NullCheck optimization. 3820 void Compile::set_allowed_deopt_reasons() { 3821 _allowed_reasons = 0; 3822 if (is_method_compilation()) { 3823 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3824 assert(rs < BitsPerInt, "recode bit map"); 3825 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3826 _allowed_reasons |= nth_bit(rs); 3827 } 3828 } 3829 } 3830 } 3831 3832 bool Compile::is_compiling_clinit_for(ciKlass* k) { 3833 ciMethod* root = method(); // the root method of compilation 3834 return root->is_static_initializer() && root->holder() == k; // access in the context of clinit 3835 } 3836 3837 #ifndef PRODUCT 3838 //------------------------------verify_graph_edges--------------------------- 3839 // Walk the Graph and verify that there is a one-to-one correspondence 3840 // between Use-Def edges and Def-Use edges in the graph. 3841 void Compile::verify_graph_edges(bool no_dead_code) { 3842 if (VerifyGraphEdges) { 3843 ResourceArea *area = Thread::current()->resource_area(); 3844 Unique_Node_List visited(area); 3845 // Call recursive graph walk to check edges 3846 _root->verify_edges(visited); 3847 if (no_dead_code) { 3848 // Now make sure that no visited node is used by an unvisited node. 3849 bool dead_nodes = false; 3850 Unique_Node_List checked(area); 3851 while (visited.size() > 0) { 3852 Node* n = visited.pop(); 3853 checked.push(n); 3854 for (uint i = 0; i < n->outcnt(); i++) { 3855 Node* use = n->raw_out(i); 3856 if (checked.member(use)) continue; // already checked 3857 if (visited.member(use)) continue; // already in the graph 3858 if (use->is_Con()) continue; // a dead ConNode is OK 3859 // At this point, we have found a dead node which is DU-reachable. 3860 if (!dead_nodes) { 3861 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3862 dead_nodes = true; 3863 } 3864 use->dump(2); 3865 tty->print_cr("---"); 3866 checked.push(use); // No repeats; pretend it is now checked. 3867 } 3868 } 3869 assert(!dead_nodes, "using nodes must be reachable from root"); 3870 } 3871 } 3872 } 3873 3874 // Verify GC barriers consistency 3875 // Currently supported: 3876 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3877 void Compile::verify_barriers() { 3878 #if INCLUDE_G1GC || INCLUDE_SHENANDOAHGC 3879 if (UseG1GC || UseShenandoahGC) { 3880 // Verify G1 pre-barriers 3881 3882 #if INCLUDE_G1GC && INCLUDE_SHENANDOAHGC 3883 const int marking_offset = in_bytes(UseG1GC ? G1ThreadLocalData::satb_mark_queue_active_offset() 3884 : ShenandoahThreadLocalData::satb_mark_queue_active_offset()); 3885 #elif INCLUDE_G1GC 3886 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()); 3887 #else 3888 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()); 3889 #endif 3890 3891 ResourceArea *area = Thread::current()->resource_area(); 3892 Unique_Node_List visited(area); 3893 Node_List worklist(area); 3894 // We're going to walk control flow backwards starting from the Root 3895 worklist.push(_root); 3896 while (worklist.size() > 0) { 3897 Node* x = worklist.pop(); 3898 if (x == NULL || x == top()) continue; 3899 if (visited.member(x)) { 3900 continue; 3901 } else { 3902 visited.push(x); 3903 } 3904 3905 if (x->is_Region()) { 3906 for (uint i = 1; i < x->req(); i++) { 3907 worklist.push(x->in(i)); 3908 } 3909 } else { 3910 worklist.push(x->in(0)); 3911 // We are looking for the pattern: 3912 // /->ThreadLocal 3913 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3914 // \->ConI(0) 3915 // We want to verify that the If and the LoadB have the same control 3916 // See GraphKit::g1_write_barrier_pre() 3917 if (x->is_If()) { 3918 IfNode *iff = x->as_If(); 3919 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3920 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3921 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3922 && cmp->in(1)->is_Load()) { 3923 LoadNode* load = cmp->in(1)->as_Load(); 3924 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3925 && load->in(2)->in(3)->is_Con() 3926 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3927 3928 Node* if_ctrl = iff->in(0); 3929 Node* load_ctrl = load->in(0); 3930 3931 if (if_ctrl != load_ctrl) { 3932 // Skip possible CProj->NeverBranch in infinite loops 3933 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3934 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3935 if_ctrl = if_ctrl->in(0)->in(0); 3936 } 3937 } 3938 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3939 } 3940 } 3941 } 3942 } 3943 } 3944 } 3945 } 3946 #endif 3947 } 3948 3949 #endif 3950 3951 // The Compile object keeps track of failure reasons separately from the ciEnv. 3952 // This is required because there is not quite a 1-1 relation between the 3953 // ciEnv and its compilation task and the Compile object. Note that one 3954 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3955 // to backtrack and retry without subsuming loads. Other than this backtracking 3956 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3957 // by the logic in C2Compiler. 3958 void Compile::record_failure(const char* reason) { 3959 if (log() != NULL) { 3960 log()->elem("failure reason='%s' phase='compile'", reason); 3961 } 3962 if (_failure_reason == NULL) { 3963 // Record the first failure reason. 3964 _failure_reason = reason; 3965 } 3966 3967 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3968 C->print_method(PHASE_FAILURE); 3969 } 3970 _root = NULL; // flush the graph, too 3971 } 3972 3973 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3974 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3975 _phase_name(name), _dolog(CITimeVerbose) 3976 { 3977 if (_dolog) { 3978 C = Compile::current(); 3979 _log = C->log(); 3980 } else { 3981 C = NULL; 3982 _log = NULL; 3983 } 3984 if (_log != NULL) { 3985 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3986 _log->stamp(); 3987 _log->end_head(); 3988 } 3989 } 3990 3991 Compile::TracePhase::~TracePhase() { 3992 3993 C = Compile::current(); 3994 if (_dolog) { 3995 _log = C->log(); 3996 } else { 3997 _log = NULL; 3998 } 3999 4000 #ifdef ASSERT 4001 if (PrintIdealNodeCount) { 4002 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4003 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 4004 } 4005 4006 if (VerifyIdealNodeCount) { 4007 Compile::current()->print_missing_nodes(); 4008 } 4009 #endif 4010 4011 if (_log != NULL) { 4012 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4013 } 4014 } 4015 4016 //============================================================================= 4017 // Two Constant's are equal when the type and the value are equal. 4018 bool Compile::Constant::operator==(const Constant& other) { 4019 if (type() != other.type() ) return false; 4020 if (can_be_reused() != other.can_be_reused()) return false; 4021 // For floating point values we compare the bit pattern. 4022 switch (type()) { 4023 case T_INT: 4024 case T_FLOAT: return (_v._value.i == other._v._value.i); 4025 case T_LONG: 4026 case T_DOUBLE: return (_v._value.j == other._v._value.j); 4027 case T_OBJECT: 4028 case T_ADDRESS: return (_v._value.l == other._v._value.l); 4029 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 4030 case T_METADATA: return (_v._metadata == other._v._metadata); 4031 default: ShouldNotReachHere(); return false; 4032 } 4033 } 4034 4035 static int type_to_size_in_bytes(BasicType t) { 4036 switch (t) { 4037 case T_INT: return sizeof(jint ); 4038 case T_LONG: return sizeof(jlong ); 4039 case T_FLOAT: return sizeof(jfloat ); 4040 case T_DOUBLE: return sizeof(jdouble); 4041 case T_METADATA: return sizeof(Metadata*); 4042 // We use T_VOID as marker for jump-table entries (labels) which 4043 // need an internal word relocation. 4044 case T_VOID: 4045 case T_ADDRESS: 4046 case T_OBJECT: return sizeof(jobject); 4047 default: 4048 ShouldNotReachHere(); 4049 return -1; 4050 } 4051 } 4052 4053 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 4054 // sort descending 4055 if (a->freq() > b->freq()) return -1; 4056 if (a->freq() < b->freq()) return 1; 4057 return 0; 4058 } 4059 4060 void Compile::ConstantTable::calculate_offsets_and_size() { 4061 // First, sort the array by frequencies. 4062 _constants.sort(qsort_comparator); 4063 4064 #ifdef ASSERT 4065 // Make sure all jump-table entries were sorted to the end of the 4066 // array (they have a negative frequency). 4067 bool found_void = false; 4068 for (int i = 0; i < _constants.length(); i++) { 4069 Constant con = _constants.at(i); 4070 if (con.type() == T_VOID) 4071 found_void = true; // jump-tables 4072 else 4073 assert(!found_void, "wrong sorting"); 4074 } 4075 #endif 4076 4077 int offset = 0; 4078 for (int i = 0; i < _constants.length(); i++) { 4079 Constant* con = _constants.adr_at(i); 4080 4081 // Align offset for type. 4082 int typesize = type_to_size_in_bytes(con->type()); 4083 offset = align_up(offset, typesize); 4084 con->set_offset(offset); // set constant's offset 4085 4086 if (con->type() == T_VOID) { 4087 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4088 offset = offset + typesize * n->outcnt(); // expand jump-table 4089 } else { 4090 offset = offset + typesize; 4091 } 4092 } 4093 4094 // Align size up to the next section start (which is insts; see 4095 // CodeBuffer::align_at_start). 4096 assert(_size == -1, "already set?"); 4097 _size = align_up(offset, (int)CodeEntryAlignment); 4098 } 4099 4100 void Compile::ConstantTable::emit(CodeBuffer& cb) { 4101 MacroAssembler _masm(&cb); 4102 for (int i = 0; i < _constants.length(); i++) { 4103 Constant con = _constants.at(i); 4104 address constant_addr = NULL; 4105 switch (con.type()) { 4106 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4107 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4108 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4109 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4110 case T_OBJECT: { 4111 jobject obj = con.get_jobject(); 4112 int oop_index = _masm.oop_recorder()->find_index(obj); 4113 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4114 break; 4115 } 4116 case T_ADDRESS: { 4117 address addr = (address) con.get_jobject(); 4118 constant_addr = _masm.address_constant(addr); 4119 break; 4120 } 4121 // We use T_VOID as marker for jump-table entries (labels) which 4122 // need an internal word relocation. 4123 case T_VOID: { 4124 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4125 // Fill the jump-table with a dummy word. The real value is 4126 // filled in later in fill_jump_table. 4127 address dummy = (address) n; 4128 constant_addr = _masm.address_constant(dummy); 4129 // Expand jump-table 4130 for (uint i = 1; i < n->outcnt(); i++) { 4131 address temp_addr = _masm.address_constant(dummy + i); 4132 assert(temp_addr, "consts section too small"); 4133 } 4134 break; 4135 } 4136 case T_METADATA: { 4137 Metadata* obj = con.get_metadata(); 4138 int metadata_index = _masm.oop_recorder()->find_index(obj); 4139 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4140 break; 4141 } 4142 default: ShouldNotReachHere(); 4143 } 4144 assert(constant_addr, "consts section too small"); 4145 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4146 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4147 } 4148 } 4149 4150 int Compile::ConstantTable::find_offset(Constant& con) const { 4151 int idx = _constants.find(con); 4152 guarantee(idx != -1, "constant must be in constant table"); 4153 int offset = _constants.at(idx).offset(); 4154 guarantee(offset != -1, "constant table not emitted yet?"); 4155 return offset; 4156 } 4157 4158 void Compile::ConstantTable::add(Constant& con) { 4159 if (con.can_be_reused()) { 4160 int idx = _constants.find(con); 4161 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4162 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4163 return; 4164 } 4165 } 4166 (void) _constants.append(con); 4167 } 4168 4169 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4170 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4171 Constant con(type, value, b->_freq); 4172 add(con); 4173 return con; 4174 } 4175 4176 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4177 Constant con(metadata); 4178 add(con); 4179 return con; 4180 } 4181 4182 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4183 jvalue value; 4184 BasicType type = oper->type()->basic_type(); 4185 switch (type) { 4186 case T_LONG: value.j = oper->constantL(); break; 4187 case T_FLOAT: value.f = oper->constantF(); break; 4188 case T_DOUBLE: value.d = oper->constantD(); break; 4189 case T_OBJECT: 4190 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4191 case T_METADATA: return add((Metadata*)oper->constant()); break; 4192 default: guarantee(false, "unhandled type: %s", type2name(type)); 4193 } 4194 return add(n, type, value); 4195 } 4196 4197 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4198 jvalue value; 4199 // We can use the node pointer here to identify the right jump-table 4200 // as this method is called from Compile::Fill_buffer right before 4201 // the MachNodes are emitted and the jump-table is filled (means the 4202 // MachNode pointers do not change anymore). 4203 value.l = (jobject) n; 4204 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4205 add(con); 4206 return con; 4207 } 4208 4209 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4210 // If called from Compile::scratch_emit_size do nothing. 4211 if (Compile::current()->in_scratch_emit_size()) return; 4212 4213 assert(labels.is_nonempty(), "must be"); 4214 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4215 4216 // Since MachConstantNode::constant_offset() also contains 4217 // table_base_offset() we need to subtract the table_base_offset() 4218 // to get the plain offset into the constant table. 4219 int offset = n->constant_offset() - table_base_offset(); 4220 4221 MacroAssembler _masm(&cb); 4222 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4223 4224 for (uint i = 0; i < n->outcnt(); i++) { 4225 address* constant_addr = &jump_table_base[i]; 4226 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)); 4227 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4228 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4229 } 4230 } 4231 4232 //----------------------------static_subtype_check----------------------------- 4233 // Shortcut important common cases when superklass is exact: 4234 // (0) superklass is java.lang.Object (can occur in reflective code) 4235 // (1) subklass is already limited to a subtype of superklass => always ok 4236 // (2) subklass does not overlap with superklass => always fail 4237 // (3) superklass has NO subtypes and we can check with a simple compare. 4238 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4239 if (StressReflectiveCode) { 4240 return SSC_full_test; // Let caller generate the general case. 4241 } 4242 4243 if (superk == env()->Object_klass()) { 4244 return SSC_always_true; // (0) this test cannot fail 4245 } 4246 4247 ciType* superelem = superk; 4248 if (superelem->is_array_klass()) 4249 superelem = superelem->as_array_klass()->base_element_type(); 4250 4251 if (!subk->is_interface()) { // cannot trust static interface types yet 4252 if (subk->is_subtype_of(superk)) { 4253 return SSC_always_true; // (1) false path dead; no dynamic test needed 4254 } 4255 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4256 !superk->is_subtype_of(subk)) { 4257 return SSC_always_false; 4258 } 4259 } 4260 4261 // If casting to an instance klass, it must have no subtypes 4262 if (superk->is_interface()) { 4263 // Cannot trust interfaces yet. 4264 // %%% S.B. superk->nof_implementors() == 1 4265 } else if (superelem->is_instance_klass()) { 4266 ciInstanceKlass* ik = superelem->as_instance_klass(); 4267 if (!ik->has_subklass() && !ik->is_interface()) { 4268 if (!ik->is_final()) { 4269 // Add a dependency if there is a chance of a later subclass. 4270 dependencies()->assert_leaf_type(ik); 4271 } 4272 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4273 } 4274 } else { 4275 // A primitive array type has no subtypes. 4276 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4277 } 4278 4279 return SSC_full_test; 4280 } 4281 4282 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4283 #ifdef _LP64 4284 // The scaled index operand to AddP must be a clean 64-bit value. 4285 // Java allows a 32-bit int to be incremented to a negative 4286 // value, which appears in a 64-bit register as a large 4287 // positive number. Using that large positive number as an 4288 // operand in pointer arithmetic has bad consequences. 4289 // On the other hand, 32-bit overflow is rare, and the possibility 4290 // can often be excluded, if we annotate the ConvI2L node with 4291 // a type assertion that its value is known to be a small positive 4292 // number. (The prior range check has ensured this.) 4293 // This assertion is used by ConvI2LNode::Ideal. 4294 int index_max = max_jint - 1; // array size is max_jint, index is one less 4295 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4296 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4297 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4298 #endif 4299 return idx; 4300 } 4301 4302 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4303 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4304 if (ctrl != NULL) { 4305 // Express control dependency by a CastII node with a narrow type. 4306 value = new CastIINode(value, itype, false, true /* range check dependency */); 4307 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4308 // node from floating above the range check during loop optimizations. Otherwise, the 4309 // ConvI2L node may be eliminated independently of the range check, causing the data path 4310 // to become TOP while the control path is still there (although it's unreachable). 4311 value->set_req(0, ctrl); 4312 // Save CastII node to remove it after loop optimizations. 4313 phase->C->add_range_check_cast(value); 4314 value = phase->transform(value); 4315 } 4316 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4317 return phase->transform(new ConvI2LNode(value, ltype)); 4318 } 4319 4320 // The message about the current inlining is accumulated in 4321 // _print_inlining_stream and transfered into the _print_inlining_list 4322 // once we know whether inlining succeeds or not. For regular 4323 // inlining, messages are appended to the buffer pointed by 4324 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4325 // a new buffer is added after _print_inlining_idx in the list. This 4326 // way we can update the inlining message for late inlining call site 4327 // when the inlining is attempted again. 4328 void Compile::print_inlining_init() { 4329 if (print_inlining() || print_intrinsics()) { 4330 _print_inlining_stream = new stringStream(); 4331 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4332 } 4333 } 4334 4335 void Compile::print_inlining_reinit() { 4336 if (print_inlining() || print_intrinsics()) { 4337 // Re allocate buffer when we change ResourceMark 4338 _print_inlining_stream = new stringStream(); 4339 } 4340 } 4341 4342 void Compile::print_inlining_reset() { 4343 _print_inlining_stream->reset(); 4344 } 4345 4346 void Compile::print_inlining_commit() { 4347 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4348 // Transfer the message from _print_inlining_stream to the current 4349 // _print_inlining_list buffer and clear _print_inlining_stream. 4350 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4351 print_inlining_reset(); 4352 } 4353 4354 void Compile::print_inlining_push() { 4355 // Add new buffer to the _print_inlining_list at current position 4356 _print_inlining_idx++; 4357 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4358 } 4359 4360 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4361 return _print_inlining_list->at(_print_inlining_idx); 4362 } 4363 4364 void Compile::print_inlining_update(CallGenerator* cg) { 4365 if (print_inlining() || print_intrinsics()) { 4366 if (!cg->is_late_inline()) { 4367 if (print_inlining_current().cg() != NULL) { 4368 print_inlining_push(); 4369 } 4370 print_inlining_commit(); 4371 } else { 4372 if (print_inlining_current().cg() != cg && 4373 (print_inlining_current().cg() != NULL || 4374 print_inlining_current().ss()->size() != 0)) { 4375 print_inlining_push(); 4376 } 4377 print_inlining_commit(); 4378 print_inlining_current().set_cg(cg); 4379 } 4380 } 4381 } 4382 4383 void Compile::print_inlining_move_to(CallGenerator* cg) { 4384 // We resume inlining at a late inlining call site. Locate the 4385 // corresponding inlining buffer so that we can update it. 4386 if (print_inlining()) { 4387 for (int i = 0; i < _print_inlining_list->length(); i++) { 4388 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4389 _print_inlining_idx = i; 4390 return; 4391 } 4392 } 4393 ShouldNotReachHere(); 4394 } 4395 } 4396 4397 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4398 if (print_inlining()) { 4399 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4400 assert(print_inlining_current().cg() == cg, "wrong entry"); 4401 // replace message with new message 4402 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4403 print_inlining_commit(); 4404 print_inlining_current().set_cg(cg); 4405 } 4406 } 4407 4408 void Compile::print_inlining_assert_ready() { 4409 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4410 } 4411 4412 void Compile::process_print_inlining() { 4413 bool do_print_inlining = print_inlining() || print_intrinsics(); 4414 if (do_print_inlining || log() != NULL) { 4415 // Print inlining message for candidates that we couldn't inline 4416 // for lack of space 4417 for (int i = 0; i < _late_inlines.length(); i++) { 4418 CallGenerator* cg = _late_inlines.at(i); 4419 if (!cg->is_mh_late_inline()) { 4420 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4421 if (do_print_inlining) { 4422 cg->print_inlining_late(msg); 4423 } 4424 log_late_inline_failure(cg, msg); 4425 } 4426 } 4427 } 4428 if (do_print_inlining) { 4429 ResourceMark rm; 4430 stringStream ss; 4431 for (int i = 0; i < _print_inlining_list->length(); i++) { 4432 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4433 } 4434 size_t end = ss.size(); 4435 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4436 strncpy(_print_inlining_output, ss.base(), end+1); 4437 _print_inlining_output[end] = 0; 4438 } 4439 } 4440 4441 void Compile::dump_print_inlining() { 4442 if (_print_inlining_output != NULL) { 4443 tty->print_raw(_print_inlining_output); 4444 } 4445 } 4446 4447 void Compile::log_late_inline(CallGenerator* cg) { 4448 if (log() != NULL) { 4449 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4450 cg->unique_id()); 4451 JVMState* p = cg->call_node()->jvms(); 4452 while (p != NULL) { 4453 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4454 p = p->caller(); 4455 } 4456 log()->tail("late_inline"); 4457 } 4458 } 4459 4460 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4461 log_late_inline(cg); 4462 if (log() != NULL) { 4463 log()->inline_fail(msg); 4464 } 4465 } 4466 4467 void Compile::log_inline_id(CallGenerator* cg) { 4468 if (log() != NULL) { 4469 // The LogCompilation tool needs a unique way to identify late 4470 // inline call sites. This id must be unique for this call site in 4471 // this compilation. Try to have it unique across compilations as 4472 // well because it can be convenient when grepping through the log 4473 // file. 4474 // Distinguish OSR compilations from others in case CICountOSR is 4475 // on. 4476 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4477 cg->set_unique_id(id); 4478 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4479 } 4480 } 4481 4482 void Compile::log_inline_failure(const char* msg) { 4483 if (C->log() != NULL) { 4484 C->log()->inline_fail(msg); 4485 } 4486 } 4487 4488 4489 // Dump inlining replay data to the stream. 4490 // Don't change thread state and acquire any locks. 4491 void Compile::dump_inline_data(outputStream* out) { 4492 InlineTree* inl_tree = ilt(); 4493 if (inl_tree != NULL) { 4494 out->print(" inline %d", inl_tree->count()); 4495 inl_tree->dump_replay_data(out); 4496 } 4497 } 4498 4499 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4500 if (n1->Opcode() < n2->Opcode()) return -1; 4501 else if (n1->Opcode() > n2->Opcode()) return 1; 4502 4503 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4504 for (uint i = 1; i < n1->req(); i++) { 4505 if (n1->in(i) < n2->in(i)) return -1; 4506 else if (n1->in(i) > n2->in(i)) return 1; 4507 } 4508 4509 return 0; 4510 } 4511 4512 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4513 Node* n1 = *n1p; 4514 Node* n2 = *n2p; 4515 4516 return cmp_expensive_nodes(n1, n2); 4517 } 4518 4519 void Compile::sort_expensive_nodes() { 4520 if (!expensive_nodes_sorted()) { 4521 _expensive_nodes->sort(cmp_expensive_nodes); 4522 } 4523 } 4524 4525 bool Compile::expensive_nodes_sorted() const { 4526 for (int i = 1; i < _expensive_nodes->length(); i++) { 4527 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4528 return false; 4529 } 4530 } 4531 return true; 4532 } 4533 4534 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4535 if (_expensive_nodes->length() == 0) { 4536 return false; 4537 } 4538 4539 assert(OptimizeExpensiveOps, "optimization off?"); 4540 4541 // Take this opportunity to remove dead nodes from the list 4542 int j = 0; 4543 for (int i = 0; i < _expensive_nodes->length(); i++) { 4544 Node* n = _expensive_nodes->at(i); 4545 if (!n->is_unreachable(igvn)) { 4546 assert(n->is_expensive(), "should be expensive"); 4547 _expensive_nodes->at_put(j, n); 4548 j++; 4549 } 4550 } 4551 _expensive_nodes->trunc_to(j); 4552 4553 // Then sort the list so that similar nodes are next to each other 4554 // and check for at least two nodes of identical kind with same data 4555 // inputs. 4556 sort_expensive_nodes(); 4557 4558 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4559 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4560 return true; 4561 } 4562 } 4563 4564 return false; 4565 } 4566 4567 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4568 if (_expensive_nodes->length() == 0) { 4569 return; 4570 } 4571 4572 assert(OptimizeExpensiveOps, "optimization off?"); 4573 4574 // Sort to bring similar nodes next to each other and clear the 4575 // control input of nodes for which there's only a single copy. 4576 sort_expensive_nodes(); 4577 4578 int j = 0; 4579 int identical = 0; 4580 int i = 0; 4581 bool modified = false; 4582 for (; i < _expensive_nodes->length()-1; i++) { 4583 assert(j <= i, "can't write beyond current index"); 4584 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4585 identical++; 4586 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4587 continue; 4588 } 4589 if (identical > 0) { 4590 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4591 identical = 0; 4592 } else { 4593 Node* n = _expensive_nodes->at(i); 4594 igvn.replace_input_of(n, 0, NULL); 4595 igvn.hash_insert(n); 4596 modified = true; 4597 } 4598 } 4599 if (identical > 0) { 4600 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4601 } else if (_expensive_nodes->length() >= 1) { 4602 Node* n = _expensive_nodes->at(i); 4603 igvn.replace_input_of(n, 0, NULL); 4604 igvn.hash_insert(n); 4605 modified = true; 4606 } 4607 _expensive_nodes->trunc_to(j); 4608 if (modified) { 4609 igvn.optimize(); 4610 } 4611 } 4612 4613 void Compile::add_expensive_node(Node * n) { 4614 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4615 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4616 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4617 if (OptimizeExpensiveOps) { 4618 _expensive_nodes->append(n); 4619 } else { 4620 // Clear control input and let IGVN optimize expensive nodes if 4621 // OptimizeExpensiveOps is off. 4622 n->set_req(0, NULL); 4623 } 4624 } 4625 4626 /** 4627 * Remove the speculative part of types and clean up the graph 4628 */ 4629 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4630 if (UseTypeSpeculation) { 4631 Unique_Node_List worklist; 4632 worklist.push(root()); 4633 int modified = 0; 4634 // Go over all type nodes that carry a speculative type, drop the 4635 // speculative part of the type and enqueue the node for an igvn 4636 // which may optimize it out. 4637 for (uint next = 0; next < worklist.size(); ++next) { 4638 Node *n = worklist.at(next); 4639 if (n->is_Type()) { 4640 TypeNode* tn = n->as_Type(); 4641 const Type* t = tn->type(); 4642 const Type* t_no_spec = t->remove_speculative(); 4643 if (t_no_spec != t) { 4644 bool in_hash = igvn.hash_delete(n); 4645 assert(in_hash, "node should be in igvn hash table"); 4646 tn->set_type(t_no_spec); 4647 igvn.hash_insert(n); 4648 igvn._worklist.push(n); // give it a chance to go away 4649 modified++; 4650 } 4651 } 4652 uint max = n->len(); 4653 for( uint i = 0; i < max; ++i ) { 4654 Node *m = n->in(i); 4655 if (not_a_node(m)) continue; 4656 worklist.push(m); 4657 } 4658 } 4659 // Drop the speculative part of all types in the igvn's type table 4660 igvn.remove_speculative_types(); 4661 if (modified > 0) { 4662 igvn.optimize(); 4663 } 4664 #ifdef ASSERT 4665 // Verify that after the IGVN is over no speculative type has resurfaced 4666 worklist.clear(); 4667 worklist.push(root()); 4668 for (uint next = 0; next < worklist.size(); ++next) { 4669 Node *n = worklist.at(next); 4670 const Type* t = igvn.type_or_null(n); 4671 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4672 if (n->is_Type()) { 4673 t = n->as_Type()->type(); 4674 assert(t == t->remove_speculative(), "no more speculative types"); 4675 } 4676 uint max = n->len(); 4677 for( uint i = 0; i < max; ++i ) { 4678 Node *m = n->in(i); 4679 if (not_a_node(m)) continue; 4680 worklist.push(m); 4681 } 4682 } 4683 igvn.check_no_speculative_types(); 4684 #endif 4685 } 4686 } 4687 4688 // Auxiliary method to support randomized stressing/fuzzing. 4689 // 4690 // This method can be called the arbitrary number of times, with current count 4691 // as the argument. The logic allows selecting a single candidate from the 4692 // running list of candidates as follows: 4693 // int count = 0; 4694 // Cand* selected = null; 4695 // while(cand = cand->next()) { 4696 // if (randomized_select(++count)) { 4697 // selected = cand; 4698 // } 4699 // } 4700 // 4701 // Including count equalizes the chances any candidate is "selected". 4702 // This is useful when we don't have the complete list of candidates to choose 4703 // from uniformly. In this case, we need to adjust the randomicity of the 4704 // selection, or else we will end up biasing the selection towards the latter 4705 // candidates. 4706 // 4707 // Quick back-envelope calculation shows that for the list of n candidates 4708 // the equal probability for the candidate to persist as "best" can be 4709 // achieved by replacing it with "next" k-th candidate with the probability 4710 // of 1/k. It can be easily shown that by the end of the run, the 4711 // probability for any candidate is converged to 1/n, thus giving the 4712 // uniform distribution among all the candidates. 4713 // 4714 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4715 #define RANDOMIZED_DOMAIN_POW 29 4716 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4717 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4718 bool Compile::randomized_select(int count) { 4719 assert(count > 0, "only positive"); 4720 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4721 } 4722 4723 CloneMap& Compile::clone_map() { return _clone_map; } 4724 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4725 4726 void NodeCloneInfo::dump() const { 4727 tty->print(" {%d:%d} ", idx(), gen()); 4728 } 4729 4730 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4731 uint64_t val = value(old->_idx); 4732 NodeCloneInfo cio(val); 4733 assert(val != 0, "old node should be in the map"); 4734 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4735 insert(nnn->_idx, cin.get()); 4736 #ifndef PRODUCT 4737 if (is_debug()) { 4738 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4739 } 4740 #endif 4741 } 4742 4743 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4744 NodeCloneInfo cio(value(old->_idx)); 4745 if (cio.get() == 0) { 4746 cio.set(old->_idx, 0); 4747 insert(old->_idx, cio.get()); 4748 #ifndef PRODUCT 4749 if (is_debug()) { 4750 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4751 } 4752 #endif 4753 } 4754 clone(old, nnn, gen); 4755 } 4756 4757 int CloneMap::max_gen() const { 4758 int g = 0; 4759 DictI di(_dict); 4760 for(; di.test(); ++di) { 4761 int t = gen(di._key); 4762 if (g < t) { 4763 g = t; 4764 #ifndef PRODUCT 4765 if (is_debug()) { 4766 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4767 } 4768 #endif 4769 } 4770 } 4771 return g; 4772 } 4773 4774 void CloneMap::dump(node_idx_t key) const { 4775 uint64_t val = value(key); 4776 if (val != 0) { 4777 NodeCloneInfo ni(val); 4778 ni.dump(); 4779 } 4780 }