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 // Remove edges from "root" to each SafePoint at a backward branch. 2177 // They were inserted during parsing (see add_safepoint()) to make 2178 // infinite loops without calls or exceptions visible to root, i.e., 2179 // useful. 2180 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2181 Node *r = root(); 2182 if (r != NULL) { 2183 for (uint i = r->req(); i < r->len(); ++i) { 2184 Node *n = r->in(i); 2185 if (n != NULL && n->is_SafePoint()) { 2186 r->rm_prec(i); 2187 if (n->outcnt() == 0) { 2188 igvn.remove_dead_node(n); 2189 } 2190 --i; 2191 } 2192 } 2193 // Parsing may have added top inputs to the root node (Path 2194 // leading to the Halt node proven dead). Make sure we get a 2195 // chance to clean them up. 2196 igvn._worklist.push(r); 2197 igvn.optimize(); 2198 } 2199 } 2200 2201 //------------------------------Optimize--------------------------------------- 2202 // Given a graph, optimize it. 2203 void Compile::Optimize() { 2204 TracePhase tp("optimizer", &timers[_t_optimizer]); 2205 2206 #ifndef PRODUCT 2207 if (_directive->BreakAtCompileOption) { 2208 BREAKPOINT; 2209 } 2210 2211 #endif 2212 2213 #ifdef ASSERT 2214 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2215 bs->verify_gc_barriers(true); 2216 #endif 2217 2218 ResourceMark rm; 2219 int loop_opts_cnt; 2220 2221 print_inlining_reinit(); 2222 2223 NOT_PRODUCT( verify_graph_edges(); ) 2224 2225 print_method(PHASE_AFTER_PARSING); 2226 2227 { 2228 // Iterative Global Value Numbering, including ideal transforms 2229 // Initialize IterGVN with types and values from parse-time GVN 2230 PhaseIterGVN igvn(initial_gvn()); 2231 #ifdef ASSERT 2232 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2233 #endif 2234 { 2235 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2236 igvn.optimize(); 2237 } 2238 2239 if (failing()) return; 2240 2241 print_method(PHASE_ITER_GVN1, 2); 2242 2243 inline_incrementally(igvn); 2244 2245 print_method(PHASE_INCREMENTAL_INLINE, 2); 2246 2247 if (failing()) return; 2248 2249 if (eliminate_boxing()) { 2250 // Inline valueOf() methods now. 2251 inline_boxing_calls(igvn); 2252 2253 if (AlwaysIncrementalInline) { 2254 inline_incrementally(igvn); 2255 } 2256 2257 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2258 2259 if (failing()) return; 2260 } 2261 2262 // Now that all inlining is over, cut edge from root to loop 2263 // safepoints 2264 remove_root_to_sfpts_edges(igvn); 2265 2266 // Remove the speculative part of types and clean up the graph from 2267 // the extra CastPP nodes whose only purpose is to carry them. Do 2268 // that early so that optimizations are not disrupted by the extra 2269 // CastPP nodes. 2270 remove_speculative_types(igvn); 2271 2272 // No more new expensive nodes will be added to the list from here 2273 // so keep only the actual candidates for optimizations. 2274 cleanup_expensive_nodes(igvn); 2275 2276 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2277 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2278 initial_gvn()->replace_with(&igvn); 2279 for_igvn()->clear(); 2280 Unique_Node_List new_worklist(C->comp_arena()); 2281 { 2282 ResourceMark rm; 2283 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2284 } 2285 set_for_igvn(&new_worklist); 2286 igvn = PhaseIterGVN(initial_gvn()); 2287 igvn.optimize(); 2288 } 2289 2290 // Perform escape analysis 2291 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2292 if (has_loops()) { 2293 // Cleanup graph (remove dead nodes). 2294 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2295 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone); 2296 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2297 if (failing()) return; 2298 } 2299 ConnectionGraph::do_analysis(this, &igvn); 2300 2301 if (failing()) return; 2302 2303 // Optimize out fields loads from scalar replaceable allocations. 2304 igvn.optimize(); 2305 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2306 2307 if (failing()) return; 2308 2309 if (congraph() != NULL && macro_count() > 0) { 2310 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2311 PhaseMacroExpand mexp(igvn); 2312 mexp.eliminate_macro_nodes(); 2313 igvn.set_delay_transform(false); 2314 2315 igvn.optimize(); 2316 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2317 2318 if (failing()) return; 2319 } 2320 } 2321 2322 // Loop transforms on the ideal graph. Range Check Elimination, 2323 // peeling, unrolling, etc. 2324 2325 // Set loop opts counter 2326 loop_opts_cnt = num_loop_opts(); 2327 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2328 { 2329 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2330 PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault); 2331 loop_opts_cnt--; 2332 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2333 if (failing()) return; 2334 } 2335 // Loop opts pass if partial peeling occurred in previous pass 2336 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2337 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2338 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2339 loop_opts_cnt--; 2340 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2341 if (failing()) return; 2342 } 2343 // Loop opts pass for loop-unrolling before CCP 2344 if(major_progress() && (loop_opts_cnt > 0)) { 2345 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2346 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2347 loop_opts_cnt--; 2348 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2349 } 2350 if (!failing()) { 2351 // Verify that last round of loop opts produced a valid graph 2352 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2353 PhaseIdealLoop::verify(igvn); 2354 } 2355 } 2356 if (failing()) return; 2357 2358 // Conditional Constant Propagation; 2359 PhaseCCP ccp( &igvn ); 2360 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2361 { 2362 TracePhase tp("ccp", &timers[_t_ccp]); 2363 ccp.do_transform(); 2364 } 2365 print_method(PHASE_CPP1, 2); 2366 2367 assert( true, "Break here to ccp.dump_old2new_map()"); 2368 2369 // Iterative Global Value Numbering, including ideal transforms 2370 { 2371 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2372 igvn = ccp; 2373 igvn.optimize(); 2374 } 2375 2376 print_method(PHASE_ITER_GVN2, 2); 2377 2378 if (failing()) return; 2379 2380 // Loop transforms on the ideal graph. Range Check Elimination, 2381 // peeling, unrolling, etc. 2382 if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) { 2383 return; 2384 } 2385 2386 #if INCLUDE_ZGC 2387 if (UseZGC) { 2388 ZBarrierSetC2::find_dominating_barriers(igvn); 2389 } 2390 #endif 2391 2392 if (failing()) return; 2393 2394 // Ensure that major progress is now clear 2395 C->clear_major_progress(); 2396 2397 { 2398 // Verify that all previous optimizations produced a valid graph 2399 // at least to this point, even if no loop optimizations were done. 2400 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2401 PhaseIdealLoop::verify(igvn); 2402 } 2403 2404 if (range_check_cast_count() > 0) { 2405 // No more loop optimizations. Remove all range check dependent CastIINodes. 2406 C->remove_range_check_casts(igvn); 2407 igvn.optimize(); 2408 } 2409 2410 #ifdef ASSERT 2411 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2412 bs->verify_gc_barriers(false); 2413 #endif 2414 2415 { 2416 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2417 PhaseMacroExpand mex(igvn); 2418 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2); 2419 if (mex.expand_macro_nodes()) { 2420 assert(failing(), "must bail out w/ explicit message"); 2421 return; 2422 } 2423 } 2424 2425 print_method(PHASE_BEFORE_BARRIER_EXPAND, 2); 2426 2427 #if INCLUDE_SHENANDOAHGC 2428 if (UseShenandoahGC && ((ShenandoahBarrierSetC2*)BarrierSet::barrier_set()->barrier_set_c2())->expand_barriers(this, igvn)) { 2429 assert(failing(), "must bail out w/ explicit message"); 2430 return; 2431 } 2432 #endif 2433 2434 if (opaque4_count() > 0) { 2435 C->remove_opaque4_nodes(igvn); 2436 igvn.optimize(); 2437 } 2438 2439 DEBUG_ONLY( _modified_nodes = NULL; ) 2440 } // (End scope of igvn; run destructor if necessary for asserts.) 2441 2442 process_print_inlining(); 2443 // A method with only infinite loops has no edges entering loops from root 2444 { 2445 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2446 if (final_graph_reshaping()) { 2447 assert(failing(), "must bail out w/ explicit message"); 2448 return; 2449 } 2450 } 2451 2452 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2453 } 2454 2455 2456 //------------------------------Code_Gen--------------------------------------- 2457 // Given a graph, generate code for it 2458 void Compile::Code_Gen() { 2459 if (failing()) { 2460 return; 2461 } 2462 2463 // Perform instruction selection. You might think we could reclaim Matcher 2464 // memory PDQ, but actually the Matcher is used in generating spill code. 2465 // Internals of the Matcher (including some VectorSets) must remain live 2466 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2467 // set a bit in reclaimed memory. 2468 2469 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2470 // nodes. Mapping is only valid at the root of each matched subtree. 2471 NOT_PRODUCT( verify_graph_edges(); ) 2472 2473 Matcher matcher; 2474 _matcher = &matcher; 2475 { 2476 TracePhase tp("matcher", &timers[_t_matcher]); 2477 matcher.match(); 2478 } 2479 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2480 // nodes. Mapping is only valid at the root of each matched subtree. 2481 NOT_PRODUCT( verify_graph_edges(); ) 2482 2483 // If you have too many nodes, or if matching has failed, bail out 2484 check_node_count(0, "out of nodes matching instructions"); 2485 if (failing()) { 2486 return; 2487 } 2488 2489 print_method(PHASE_MATCHING, 2); 2490 2491 // Build a proper-looking CFG 2492 PhaseCFG cfg(node_arena(), root(), matcher); 2493 _cfg = &cfg; 2494 { 2495 TracePhase tp("scheduler", &timers[_t_scheduler]); 2496 bool success = cfg.do_global_code_motion(); 2497 if (!success) { 2498 return; 2499 } 2500 2501 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2502 NOT_PRODUCT( verify_graph_edges(); ) 2503 debug_only( cfg.verify(); ) 2504 } 2505 2506 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2507 _regalloc = ®alloc; 2508 { 2509 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2510 // Perform register allocation. After Chaitin, use-def chains are 2511 // no longer accurate (at spill code) and so must be ignored. 2512 // Node->LRG->reg mappings are still accurate. 2513 _regalloc->Register_Allocate(); 2514 2515 // Bail out if the allocator builds too many nodes 2516 if (failing()) { 2517 return; 2518 } 2519 } 2520 2521 // Prior to register allocation we kept empty basic blocks in case the 2522 // the allocator needed a place to spill. After register allocation we 2523 // are not adding any new instructions. If any basic block is empty, we 2524 // can now safely remove it. 2525 { 2526 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2527 cfg.remove_empty_blocks(); 2528 if (do_freq_based_layout()) { 2529 PhaseBlockLayout layout(cfg); 2530 } else { 2531 cfg.set_loop_alignment(); 2532 } 2533 cfg.fixup_flow(); 2534 } 2535 2536 // Apply peephole optimizations 2537 if( OptoPeephole ) { 2538 TracePhase tp("peephole", &timers[_t_peephole]); 2539 PhasePeephole peep( _regalloc, cfg); 2540 peep.do_transform(); 2541 } 2542 2543 // Do late expand if CPU requires this. 2544 if (Matcher::require_postalloc_expand) { 2545 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2546 cfg.postalloc_expand(_regalloc); 2547 } 2548 2549 // Convert Nodes to instruction bits in a buffer 2550 { 2551 TraceTime tp("output", &timers[_t_output], CITime); 2552 Output(); 2553 } 2554 2555 print_method(PHASE_FINAL_CODE); 2556 2557 // He's dead, Jim. 2558 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2559 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2560 } 2561 2562 2563 //------------------------------dump_asm--------------------------------------- 2564 // Dump formatted assembly 2565 #ifndef PRODUCT 2566 void Compile::dump_asm(int *pcs, uint pc_limit) { 2567 bool cut_short = false; 2568 tty->print_cr("#"); 2569 tty->print("# "); _tf->dump(); tty->cr(); 2570 tty->print_cr("#"); 2571 2572 // For all blocks 2573 int pc = 0x0; // Program counter 2574 char starts_bundle = ' '; 2575 _regalloc->dump_frame(); 2576 2577 Node *n = NULL; 2578 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2579 if (VMThread::should_terminate()) { 2580 cut_short = true; 2581 break; 2582 } 2583 Block* block = _cfg->get_block(i); 2584 if (block->is_connector() && !Verbose) { 2585 continue; 2586 } 2587 n = block->head(); 2588 if (pcs && n->_idx < pc_limit) { 2589 tty->print("%3.3x ", pcs[n->_idx]); 2590 } else { 2591 tty->print(" "); 2592 } 2593 block->dump_head(_cfg); 2594 if (block->is_connector()) { 2595 tty->print_cr(" # Empty connector block"); 2596 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2597 tty->print_cr(" # Block is sole successor of call"); 2598 } 2599 2600 // For all instructions 2601 Node *delay = NULL; 2602 for (uint j = 0; j < block->number_of_nodes(); j++) { 2603 if (VMThread::should_terminate()) { 2604 cut_short = true; 2605 break; 2606 } 2607 n = block->get_node(j); 2608 if (valid_bundle_info(n)) { 2609 Bundle* bundle = node_bundling(n); 2610 if (bundle->used_in_unconditional_delay()) { 2611 delay = n; 2612 continue; 2613 } 2614 if (bundle->starts_bundle()) { 2615 starts_bundle = '+'; 2616 } 2617 } 2618 2619 if (WizardMode) { 2620 n->dump(); 2621 } 2622 2623 if( !n->is_Region() && // Dont print in the Assembly 2624 !n->is_Phi() && // a few noisely useless nodes 2625 !n->is_Proj() && 2626 !n->is_MachTemp() && 2627 !n->is_SafePointScalarObject() && 2628 !n->is_Catch() && // Would be nice to print exception table targets 2629 !n->is_MergeMem() && // Not very interesting 2630 !n->is_top() && // Debug info table constants 2631 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2632 ) { 2633 if (pcs && n->_idx < pc_limit) 2634 tty->print("%3.3x", pcs[n->_idx]); 2635 else 2636 tty->print(" "); 2637 tty->print(" %c ", starts_bundle); 2638 starts_bundle = ' '; 2639 tty->print("\t"); 2640 n->format(_regalloc, tty); 2641 tty->cr(); 2642 } 2643 2644 // If we have an instruction with a delay slot, and have seen a delay, 2645 // then back up and print it 2646 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2647 assert(delay != NULL, "no unconditional delay instruction"); 2648 if (WizardMode) delay->dump(); 2649 2650 if (node_bundling(delay)->starts_bundle()) 2651 starts_bundle = '+'; 2652 if (pcs && n->_idx < pc_limit) 2653 tty->print("%3.3x", pcs[n->_idx]); 2654 else 2655 tty->print(" "); 2656 tty->print(" %c ", starts_bundle); 2657 starts_bundle = ' '; 2658 tty->print("\t"); 2659 delay->format(_regalloc, tty); 2660 tty->cr(); 2661 delay = NULL; 2662 } 2663 2664 // Dump the exception table as well 2665 if( n->is_Catch() && (Verbose || WizardMode) ) { 2666 // Print the exception table for this offset 2667 _handler_table.print_subtable_for(pc); 2668 } 2669 } 2670 2671 if (pcs && n->_idx < pc_limit) 2672 tty->print_cr("%3.3x", pcs[n->_idx]); 2673 else 2674 tty->cr(); 2675 2676 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2677 2678 } // End of per-block dump 2679 tty->cr(); 2680 2681 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2682 } 2683 #endif 2684 2685 //------------------------------Final_Reshape_Counts--------------------------- 2686 // This class defines counters to help identify when a method 2687 // may/must be executed using hardware with only 24-bit precision. 2688 struct Final_Reshape_Counts : public StackObj { 2689 int _call_count; // count non-inlined 'common' calls 2690 int _float_count; // count float ops requiring 24-bit precision 2691 int _double_count; // count double ops requiring more precision 2692 int _java_call_count; // count non-inlined 'java' calls 2693 int _inner_loop_count; // count loops which need alignment 2694 VectorSet _visited; // Visitation flags 2695 Node_List _tests; // Set of IfNodes & PCTableNodes 2696 2697 Final_Reshape_Counts() : 2698 _call_count(0), _float_count(0), _double_count(0), 2699 _java_call_count(0), _inner_loop_count(0), 2700 _visited( Thread::current()->resource_area() ) { } 2701 2702 void inc_call_count () { _call_count ++; } 2703 void inc_float_count () { _float_count ++; } 2704 void inc_double_count() { _double_count++; } 2705 void inc_java_call_count() { _java_call_count++; } 2706 void inc_inner_loop_count() { _inner_loop_count++; } 2707 2708 int get_call_count () const { return _call_count ; } 2709 int get_float_count () const { return _float_count ; } 2710 int get_double_count() const { return _double_count; } 2711 int get_java_call_count() const { return _java_call_count; } 2712 int get_inner_loop_count() const { return _inner_loop_count; } 2713 }; 2714 2715 #ifdef ASSERT 2716 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2717 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2718 // Make sure the offset goes inside the instance layout. 2719 return k->contains_field_offset(tp->offset()); 2720 // Note that OffsetBot and OffsetTop are very negative. 2721 } 2722 #endif 2723 2724 // Eliminate trivially redundant StoreCMs and accumulate their 2725 // precedence edges. 2726 void Compile::eliminate_redundant_card_marks(Node* n) { 2727 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2728 if (n->in(MemNode::Address)->outcnt() > 1) { 2729 // There are multiple users of the same address so it might be 2730 // possible to eliminate some of the StoreCMs 2731 Node* mem = n->in(MemNode::Memory); 2732 Node* adr = n->in(MemNode::Address); 2733 Node* val = n->in(MemNode::ValueIn); 2734 Node* prev = n; 2735 bool done = false; 2736 // Walk the chain of StoreCMs eliminating ones that match. As 2737 // long as it's a chain of single users then the optimization is 2738 // safe. Eliminating partially redundant StoreCMs would require 2739 // cloning copies down the other paths. 2740 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2741 if (adr == mem->in(MemNode::Address) && 2742 val == mem->in(MemNode::ValueIn)) { 2743 // redundant StoreCM 2744 if (mem->req() > MemNode::OopStore) { 2745 // Hasn't been processed by this code yet. 2746 n->add_prec(mem->in(MemNode::OopStore)); 2747 } else { 2748 // Already converted to precedence edge 2749 for (uint i = mem->req(); i < mem->len(); i++) { 2750 // Accumulate any precedence edges 2751 if (mem->in(i) != NULL) { 2752 n->add_prec(mem->in(i)); 2753 } 2754 } 2755 // Everything above this point has been processed. 2756 done = true; 2757 } 2758 // Eliminate the previous StoreCM 2759 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2760 assert(mem->outcnt() == 0, "should be dead"); 2761 mem->disconnect_inputs(NULL, this); 2762 } else { 2763 prev = mem; 2764 } 2765 mem = prev->in(MemNode::Memory); 2766 } 2767 } 2768 } 2769 2770 //------------------------------final_graph_reshaping_impl---------------------- 2771 // Implement items 1-5 from final_graph_reshaping below. 2772 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2773 2774 if ( n->outcnt() == 0 ) return; // dead node 2775 uint nop = n->Opcode(); 2776 2777 // Check for 2-input instruction with "last use" on right input. 2778 // Swap to left input. Implements item (2). 2779 if( n->req() == 3 && // two-input instruction 2780 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2781 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2782 n->in(2)->outcnt() == 1 &&// right use IS a last use 2783 !n->in(2)->is_Con() ) { // right use is not a constant 2784 // Check for commutative opcode 2785 switch( nop ) { 2786 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2787 case Op_MaxI: case Op_MinI: 2788 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2789 case Op_AndL: case Op_XorL: case Op_OrL: 2790 case Op_AndI: case Op_XorI: case Op_OrI: { 2791 // Move "last use" input to left by swapping inputs 2792 n->swap_edges(1, 2); 2793 break; 2794 } 2795 default: 2796 break; 2797 } 2798 } 2799 2800 #ifdef ASSERT 2801 if( n->is_Mem() ) { 2802 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2803 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2804 // oop will be recorded in oop map if load crosses safepoint 2805 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2806 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2807 "raw memory operations should have control edge"); 2808 } 2809 if (n->is_MemBar()) { 2810 MemBarNode* mb = n->as_MemBar(); 2811 if (mb->trailing_store() || mb->trailing_load_store()) { 2812 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2813 Node* mem = mb->in(MemBarNode::Precedent); 2814 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2815 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2816 } else if (mb->leading()) { 2817 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2818 } 2819 } 2820 #endif 2821 // Count FPU ops and common calls, implements item (3) 2822 switch( nop ) { 2823 // Count all float operations that may use FPU 2824 case Op_AddF: 2825 case Op_SubF: 2826 case Op_MulF: 2827 case Op_DivF: 2828 case Op_NegF: 2829 case Op_ModF: 2830 case Op_ConvI2F: 2831 case Op_ConF: 2832 case Op_CmpF: 2833 case Op_CmpF3: 2834 // case Op_ConvL2F: // longs are split into 32-bit halves 2835 frc.inc_float_count(); 2836 break; 2837 2838 case Op_ConvF2D: 2839 case Op_ConvD2F: 2840 frc.inc_float_count(); 2841 frc.inc_double_count(); 2842 break; 2843 2844 // Count all double operations that may use FPU 2845 case Op_AddD: 2846 case Op_SubD: 2847 case Op_MulD: 2848 case Op_DivD: 2849 case Op_NegD: 2850 case Op_ModD: 2851 case Op_ConvI2D: 2852 case Op_ConvD2I: 2853 // case Op_ConvL2D: // handled by leaf call 2854 // case Op_ConvD2L: // handled by leaf call 2855 case Op_ConD: 2856 case Op_CmpD: 2857 case Op_CmpD3: 2858 frc.inc_double_count(); 2859 break; 2860 case Op_Opaque1: // Remove Opaque Nodes before matching 2861 case Op_Opaque2: // Remove Opaque Nodes before matching 2862 case Op_Opaque3: 2863 n->subsume_by(n->in(1), this); 2864 break; 2865 case Op_CallStaticJava: 2866 case Op_CallJava: 2867 case Op_CallDynamicJava: 2868 frc.inc_java_call_count(); // Count java call site; 2869 case Op_CallRuntime: 2870 case Op_CallLeaf: 2871 case Op_CallLeafNoFP: { 2872 assert (n->is_Call(), ""); 2873 CallNode *call = n->as_Call(); 2874 #if INCLUDE_SHENANDOAHGC 2875 if (UseShenandoahGC && ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) { 2876 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt(); 2877 if (call->req() > cnt) { 2878 assert(call->req() == cnt+1, "only one extra input"); 2879 Node* addp = call->in(cnt); 2880 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?"); 2881 call->del_req(cnt); 2882 } 2883 } 2884 #endif 2885 // Count call sites where the FP mode bit would have to be flipped. 2886 // Do not count uncommon runtime calls: 2887 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2888 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2889 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 2890 frc.inc_call_count(); // Count the call site 2891 } else { // See if uncommon argument is shared 2892 Node *n = call->in(TypeFunc::Parms); 2893 int nop = n->Opcode(); 2894 // Clone shared simple arguments to uncommon calls, item (1). 2895 if (n->outcnt() > 1 && 2896 !n->is_Proj() && 2897 nop != Op_CreateEx && 2898 nop != Op_CheckCastPP && 2899 nop != Op_DecodeN && 2900 nop != Op_DecodeNKlass && 2901 !n->is_Mem() && 2902 !n->is_Phi()) { 2903 Node *x = n->clone(); 2904 call->set_req(TypeFunc::Parms, x); 2905 } 2906 } 2907 break; 2908 } 2909 2910 case Op_StoreD: 2911 case Op_LoadD: 2912 case Op_LoadD_unaligned: 2913 frc.inc_double_count(); 2914 goto handle_mem; 2915 case Op_StoreF: 2916 case Op_LoadF: 2917 frc.inc_float_count(); 2918 goto handle_mem; 2919 2920 case Op_StoreCM: 2921 { 2922 // Convert OopStore dependence into precedence edge 2923 Node* prec = n->in(MemNode::OopStore); 2924 n->del_req(MemNode::OopStore); 2925 n->add_prec(prec); 2926 eliminate_redundant_card_marks(n); 2927 } 2928 2929 // fall through 2930 2931 case Op_StoreB: 2932 case Op_StoreC: 2933 case Op_StorePConditional: 2934 case Op_StoreI: 2935 case Op_StoreL: 2936 case Op_StoreIConditional: 2937 case Op_StoreLConditional: 2938 case Op_CompareAndSwapB: 2939 case Op_CompareAndSwapS: 2940 case Op_CompareAndSwapI: 2941 case Op_CompareAndSwapL: 2942 case Op_CompareAndSwapP: 2943 case Op_CompareAndSwapN: 2944 case Op_WeakCompareAndSwapB: 2945 case Op_WeakCompareAndSwapS: 2946 case Op_WeakCompareAndSwapI: 2947 case Op_WeakCompareAndSwapL: 2948 case Op_WeakCompareAndSwapP: 2949 case Op_WeakCompareAndSwapN: 2950 case Op_CompareAndExchangeB: 2951 case Op_CompareAndExchangeS: 2952 case Op_CompareAndExchangeI: 2953 case Op_CompareAndExchangeL: 2954 case Op_CompareAndExchangeP: 2955 case Op_CompareAndExchangeN: 2956 case Op_GetAndAddS: 2957 case Op_GetAndAddB: 2958 case Op_GetAndAddI: 2959 case Op_GetAndAddL: 2960 case Op_GetAndSetS: 2961 case Op_GetAndSetB: 2962 case Op_GetAndSetI: 2963 case Op_GetAndSetL: 2964 case Op_GetAndSetP: 2965 case Op_GetAndSetN: 2966 case Op_StoreP: 2967 case Op_StoreN: 2968 case Op_StoreNKlass: 2969 case Op_LoadB: 2970 case Op_LoadUB: 2971 case Op_LoadUS: 2972 case Op_LoadI: 2973 case Op_LoadKlass: 2974 case Op_LoadNKlass: 2975 case Op_LoadL: 2976 case Op_LoadL_unaligned: 2977 case Op_LoadPLocked: 2978 case Op_LoadP: 2979 #if INCLUDE_ZGC 2980 case Op_LoadBarrierSlowReg: 2981 case Op_LoadBarrierWeakSlowReg: 2982 #endif 2983 case Op_LoadN: 2984 case Op_LoadRange: 2985 case Op_LoadS: { 2986 handle_mem: 2987 #ifdef ASSERT 2988 if( VerifyOptoOopOffsets ) { 2989 assert( n->is_Mem(), "" ); 2990 MemNode *mem = (MemNode*)n; 2991 // Check to see if address types have grounded out somehow. 2992 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2993 assert( !tp || oop_offset_is_sane(tp), "" ); 2994 } 2995 #endif 2996 break; 2997 } 2998 2999 case Op_AddP: { // Assert sane base pointers 3000 Node *addp = n->in(AddPNode::Address); 3001 assert( !addp->is_AddP() || 3002 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3003 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3004 "Base pointers must match (addp %u)", addp->_idx ); 3005 #ifdef _LP64 3006 if ((UseCompressedOops || UseCompressedClassPointers) && 3007 addp->Opcode() == Op_ConP && 3008 addp == n->in(AddPNode::Base) && 3009 n->in(AddPNode::Offset)->is_Con()) { 3010 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3011 // on the platform and on the compressed oops mode. 3012 // Use addressing with narrow klass to load with offset on x86. 3013 // Some platforms can use the constant pool to load ConP. 3014 // Do this transformation here since IGVN will convert ConN back to ConP. 3015 const Type* t = addp->bottom_type(); 3016 bool is_oop = t->isa_oopptr() != NULL; 3017 bool is_klass = t->isa_klassptr() != NULL; 3018 3019 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3020 (is_klass && Matcher::const_klass_prefer_decode())) { 3021 Node* nn = NULL; 3022 3023 int op = is_oop ? Op_ConN : Op_ConNKlass; 3024 3025 // Look for existing ConN node of the same exact type. 3026 Node* r = root(); 3027 uint cnt = r->outcnt(); 3028 for (uint i = 0; i < cnt; i++) { 3029 Node* m = r->raw_out(i); 3030 if (m!= NULL && m->Opcode() == op && 3031 m->bottom_type()->make_ptr() == t) { 3032 nn = m; 3033 break; 3034 } 3035 } 3036 if (nn != NULL) { 3037 // Decode a narrow oop to match address 3038 // [R12 + narrow_oop_reg<<3 + offset] 3039 if (is_oop) { 3040 nn = new DecodeNNode(nn, t); 3041 } else { 3042 nn = new DecodeNKlassNode(nn, t); 3043 } 3044 // Check for succeeding AddP which uses the same Base. 3045 // Otherwise we will run into the assertion above when visiting that guy. 3046 for (uint i = 0; i < n->outcnt(); ++i) { 3047 Node *out_i = n->raw_out(i); 3048 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3049 out_i->set_req(AddPNode::Base, nn); 3050 #ifdef ASSERT 3051 for (uint j = 0; j < out_i->outcnt(); ++j) { 3052 Node *out_j = out_i->raw_out(j); 3053 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3054 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3055 } 3056 #endif 3057 } 3058 } 3059 n->set_req(AddPNode::Base, nn); 3060 n->set_req(AddPNode::Address, nn); 3061 if (addp->outcnt() == 0) { 3062 addp->disconnect_inputs(NULL, this); 3063 } 3064 } 3065 } 3066 } 3067 #endif 3068 // platform dependent reshaping of the address expression 3069 reshape_address(n->as_AddP()); 3070 break; 3071 } 3072 3073 case Op_CastPP: { 3074 // Remove CastPP nodes to gain more freedom during scheduling but 3075 // keep the dependency they encode as control or precedence edges 3076 // (if control is set already) on memory operations. Some CastPP 3077 // nodes don't have a control (don't carry a dependency): skip 3078 // those. 3079 if (n->in(0) != NULL) { 3080 ResourceMark rm; 3081 Unique_Node_List wq; 3082 wq.push(n); 3083 for (uint next = 0; next < wq.size(); ++next) { 3084 Node *m = wq.at(next); 3085 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3086 Node* use = m->fast_out(i); 3087 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3088 use->ensure_control_or_add_prec(n->in(0)); 3089 } else { 3090 switch(use->Opcode()) { 3091 case Op_AddP: 3092 case Op_DecodeN: 3093 case Op_DecodeNKlass: 3094 case Op_CheckCastPP: 3095 case Op_CastPP: 3096 wq.push(use); 3097 break; 3098 } 3099 } 3100 } 3101 } 3102 } 3103 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3104 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3105 Node* in1 = n->in(1); 3106 const Type* t = n->bottom_type(); 3107 Node* new_in1 = in1->clone(); 3108 new_in1->as_DecodeN()->set_type(t); 3109 3110 if (!Matcher::narrow_oop_use_complex_address()) { 3111 // 3112 // x86, ARM and friends can handle 2 adds in addressing mode 3113 // and Matcher can fold a DecodeN node into address by using 3114 // a narrow oop directly and do implicit NULL check in address: 3115 // 3116 // [R12 + narrow_oop_reg<<3 + offset] 3117 // NullCheck narrow_oop_reg 3118 // 3119 // On other platforms (Sparc) we have to keep new DecodeN node and 3120 // use it to do implicit NULL check in address: 3121 // 3122 // decode_not_null narrow_oop_reg, base_reg 3123 // [base_reg + offset] 3124 // NullCheck base_reg 3125 // 3126 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3127 // to keep the information to which NULL check the new DecodeN node 3128 // corresponds to use it as value in implicit_null_check(). 3129 // 3130 new_in1->set_req(0, n->in(0)); 3131 } 3132 3133 n->subsume_by(new_in1, this); 3134 if (in1->outcnt() == 0) { 3135 in1->disconnect_inputs(NULL, this); 3136 } 3137 } else { 3138 n->subsume_by(n->in(1), this); 3139 if (n->outcnt() == 0) { 3140 n->disconnect_inputs(NULL, this); 3141 } 3142 } 3143 break; 3144 } 3145 #ifdef _LP64 3146 case Op_CmpP: 3147 // Do this transformation here to preserve CmpPNode::sub() and 3148 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3149 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3150 Node* in1 = n->in(1); 3151 Node* in2 = n->in(2); 3152 if (!in1->is_DecodeNarrowPtr()) { 3153 in2 = in1; 3154 in1 = n->in(2); 3155 } 3156 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3157 3158 Node* new_in2 = NULL; 3159 if (in2->is_DecodeNarrowPtr()) { 3160 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3161 new_in2 = in2->in(1); 3162 } else if (in2->Opcode() == Op_ConP) { 3163 const Type* t = in2->bottom_type(); 3164 if (t == TypePtr::NULL_PTR) { 3165 assert(in1->is_DecodeN(), "compare klass to null?"); 3166 // Don't convert CmpP null check into CmpN if compressed 3167 // oops implicit null check is not generated. 3168 // This will allow to generate normal oop implicit null check. 3169 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3170 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3171 // 3172 // This transformation together with CastPP transformation above 3173 // will generated code for implicit NULL checks for compressed oops. 3174 // 3175 // The original code after Optimize() 3176 // 3177 // LoadN memory, narrow_oop_reg 3178 // decode narrow_oop_reg, base_reg 3179 // CmpP base_reg, NULL 3180 // CastPP base_reg // NotNull 3181 // Load [base_reg + offset], val_reg 3182 // 3183 // after these transformations will be 3184 // 3185 // LoadN memory, narrow_oop_reg 3186 // CmpN narrow_oop_reg, NULL 3187 // decode_not_null narrow_oop_reg, base_reg 3188 // Load [base_reg + offset], val_reg 3189 // 3190 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3191 // since narrow oops can be used in debug info now (see the code in 3192 // final_graph_reshaping_walk()). 3193 // 3194 // At the end the code will be matched to 3195 // on x86: 3196 // 3197 // Load_narrow_oop memory, narrow_oop_reg 3198 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3199 // NullCheck narrow_oop_reg 3200 // 3201 // and on sparc: 3202 // 3203 // Load_narrow_oop memory, narrow_oop_reg 3204 // decode_not_null narrow_oop_reg, base_reg 3205 // Load [base_reg + offset], val_reg 3206 // NullCheck base_reg 3207 // 3208 } else if (t->isa_oopptr()) { 3209 new_in2 = ConNode::make(t->make_narrowoop()); 3210 } else if (t->isa_klassptr()) { 3211 new_in2 = ConNode::make(t->make_narrowklass()); 3212 } 3213 } 3214 if (new_in2 != NULL) { 3215 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3216 n->subsume_by(cmpN, this); 3217 if (in1->outcnt() == 0) { 3218 in1->disconnect_inputs(NULL, this); 3219 } 3220 if (in2->outcnt() == 0) { 3221 in2->disconnect_inputs(NULL, this); 3222 } 3223 } 3224 } 3225 break; 3226 3227 case Op_DecodeN: 3228 case Op_DecodeNKlass: 3229 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3230 // DecodeN could be pinned when it can't be fold into 3231 // an address expression, see the code for Op_CastPP above. 3232 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3233 break; 3234 3235 case Op_EncodeP: 3236 case Op_EncodePKlass: { 3237 Node* in1 = n->in(1); 3238 if (in1->is_DecodeNarrowPtr()) { 3239 n->subsume_by(in1->in(1), this); 3240 } else if (in1->Opcode() == Op_ConP) { 3241 const Type* t = in1->bottom_type(); 3242 if (t == TypePtr::NULL_PTR) { 3243 assert(t->isa_oopptr(), "null klass?"); 3244 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3245 } else if (t->isa_oopptr()) { 3246 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3247 } else if (t->isa_klassptr()) { 3248 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3249 } 3250 } 3251 if (in1->outcnt() == 0) { 3252 in1->disconnect_inputs(NULL, this); 3253 } 3254 break; 3255 } 3256 3257 case Op_Proj: { 3258 if (OptimizeStringConcat) { 3259 ProjNode* p = n->as_Proj(); 3260 if (p->_is_io_use) { 3261 // Separate projections were used for the exception path which 3262 // are normally removed by a late inline. If it wasn't inlined 3263 // then they will hang around and should just be replaced with 3264 // the original one. 3265 Node* proj = NULL; 3266 // Replace with just one 3267 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3268 Node *use = i.get(); 3269 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3270 proj = use; 3271 break; 3272 } 3273 } 3274 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop"); 3275 if (proj != NULL) { 3276 p->subsume_by(proj, this); 3277 } 3278 } 3279 } 3280 break; 3281 } 3282 3283 case Op_Phi: 3284 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3285 // The EncodeP optimization may create Phi with the same edges 3286 // for all paths. It is not handled well by Register Allocator. 3287 Node* unique_in = n->in(1); 3288 assert(unique_in != NULL, ""); 3289 uint cnt = n->req(); 3290 for (uint i = 2; i < cnt; i++) { 3291 Node* m = n->in(i); 3292 assert(m != NULL, ""); 3293 if (unique_in != m) 3294 unique_in = NULL; 3295 } 3296 if (unique_in != NULL) { 3297 n->subsume_by(unique_in, this); 3298 } 3299 } 3300 break; 3301 3302 #endif 3303 3304 #ifdef ASSERT 3305 case Op_CastII: 3306 // Verify that all range check dependent CastII nodes were removed. 3307 if (n->isa_CastII()->has_range_check()) { 3308 n->dump(3); 3309 assert(false, "Range check dependent CastII node was not removed"); 3310 } 3311 break; 3312 #endif 3313 3314 case Op_ModI: 3315 if (UseDivMod) { 3316 // Check if a%b and a/b both exist 3317 Node* d = n->find_similar(Op_DivI); 3318 if (d) { 3319 // Replace them with a fused divmod if supported 3320 if (Matcher::has_match_rule(Op_DivModI)) { 3321 DivModINode* divmod = DivModINode::make(n); 3322 d->subsume_by(divmod->div_proj(), this); 3323 n->subsume_by(divmod->mod_proj(), this); 3324 } else { 3325 // replace a%b with a-((a/b)*b) 3326 Node* mult = new MulINode(d, d->in(2)); 3327 Node* sub = new SubINode(d->in(1), mult); 3328 n->subsume_by(sub, this); 3329 } 3330 } 3331 } 3332 break; 3333 3334 case Op_ModL: 3335 if (UseDivMod) { 3336 // Check if a%b and a/b both exist 3337 Node* d = n->find_similar(Op_DivL); 3338 if (d) { 3339 // Replace them with a fused divmod if supported 3340 if (Matcher::has_match_rule(Op_DivModL)) { 3341 DivModLNode* divmod = DivModLNode::make(n); 3342 d->subsume_by(divmod->div_proj(), this); 3343 n->subsume_by(divmod->mod_proj(), this); 3344 } else { 3345 // replace a%b with a-((a/b)*b) 3346 Node* mult = new MulLNode(d, d->in(2)); 3347 Node* sub = new SubLNode(d->in(1), mult); 3348 n->subsume_by(sub, this); 3349 } 3350 } 3351 } 3352 break; 3353 3354 case Op_LoadVector: 3355 case Op_StoreVector: 3356 break; 3357 3358 case Op_AddReductionVI: 3359 case Op_AddReductionVL: 3360 case Op_AddReductionVF: 3361 case Op_AddReductionVD: 3362 case Op_MulReductionVI: 3363 case Op_MulReductionVL: 3364 case Op_MulReductionVF: 3365 case Op_MulReductionVD: 3366 break; 3367 3368 case Op_PackB: 3369 case Op_PackS: 3370 case Op_PackI: 3371 case Op_PackF: 3372 case Op_PackL: 3373 case Op_PackD: 3374 if (n->req()-1 > 2) { 3375 // Replace many operand PackNodes with a binary tree for matching 3376 PackNode* p = (PackNode*) n; 3377 Node* btp = p->binary_tree_pack(1, n->req()); 3378 n->subsume_by(btp, this); 3379 } 3380 break; 3381 case Op_Loop: 3382 case Op_CountedLoop: 3383 case Op_OuterStripMinedLoop: 3384 if (n->as_Loop()->is_inner_loop()) { 3385 frc.inc_inner_loop_count(); 3386 } 3387 n->as_Loop()->verify_strip_mined(0); 3388 break; 3389 case Op_LShiftI: 3390 case Op_RShiftI: 3391 case Op_URShiftI: 3392 case Op_LShiftL: 3393 case Op_RShiftL: 3394 case Op_URShiftL: 3395 if (Matcher::need_masked_shift_count) { 3396 // The cpu's shift instructions don't restrict the count to the 3397 // lower 5/6 bits. We need to do the masking ourselves. 3398 Node* in2 = n->in(2); 3399 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3400 const TypeInt* t = in2->find_int_type(); 3401 if (t != NULL && t->is_con()) { 3402 juint shift = t->get_con(); 3403 if (shift > mask) { // Unsigned cmp 3404 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3405 } 3406 } else { 3407 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3408 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3409 n->set_req(2, shift); 3410 } 3411 } 3412 if (in2->outcnt() == 0) { // Remove dead node 3413 in2->disconnect_inputs(NULL, this); 3414 } 3415 } 3416 break; 3417 case Op_MemBarStoreStore: 3418 case Op_MemBarRelease: 3419 // Break the link with AllocateNode: it is no longer useful and 3420 // confuses register allocation. 3421 if (n->req() > MemBarNode::Precedent) { 3422 n->set_req(MemBarNode::Precedent, top()); 3423 } 3424 break; 3425 case Op_MemBarAcquire: { 3426 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3427 // At parse time, the trailing MemBarAcquire for a volatile load 3428 // is created with an edge to the load. After optimizations, 3429 // that input may be a chain of Phis. If those phis have no 3430 // other use, then the MemBarAcquire keeps them alive and 3431 // register allocation can be confused. 3432 ResourceMark rm; 3433 Unique_Node_List wq; 3434 wq.push(n->in(MemBarNode::Precedent)); 3435 n->set_req(MemBarNode::Precedent, top()); 3436 while (wq.size() > 0) { 3437 Node* m = wq.pop(); 3438 if (m->outcnt() == 0) { 3439 for (uint j = 0; j < m->req(); j++) { 3440 Node* in = m->in(j); 3441 if (in != NULL) { 3442 wq.push(in); 3443 } 3444 } 3445 m->disconnect_inputs(NULL, this); 3446 } 3447 } 3448 } 3449 break; 3450 } 3451 #if INCLUDE_SHENANDOAHGC 3452 case Op_ShenandoahCompareAndSwapP: 3453 case Op_ShenandoahCompareAndSwapN: 3454 case Op_ShenandoahWeakCompareAndSwapN: 3455 case Op_ShenandoahWeakCompareAndSwapP: 3456 case Op_ShenandoahCompareAndExchangeP: 3457 case Op_ShenandoahCompareAndExchangeN: 3458 #ifdef ASSERT 3459 if( VerifyOptoOopOffsets ) { 3460 MemNode* mem = n->as_Mem(); 3461 // Check to see if address types have grounded out somehow. 3462 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 3463 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 3464 bool oop_offset_is_sane = k->contains_field_offset(tp->offset()); 3465 assert( !tp || oop_offset_is_sane, "" ); 3466 } 3467 #endif 3468 break; 3469 case Op_ShenandoahLoadReferenceBarrier: 3470 assert(false, "should have been expanded already"); 3471 break; 3472 #endif 3473 case Op_RangeCheck: { 3474 RangeCheckNode* rc = n->as_RangeCheck(); 3475 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3476 n->subsume_by(iff, this); 3477 frc._tests.push(iff); 3478 break; 3479 } 3480 case Op_ConvI2L: { 3481 if (!Matcher::convi2l_type_required) { 3482 // Code generation on some platforms doesn't need accurate 3483 // ConvI2L types. Widening the type can help remove redundant 3484 // address computations. 3485 n->as_Type()->set_type(TypeLong::INT); 3486 ResourceMark rm; 3487 Unique_Node_List wq; 3488 wq.push(n); 3489 for (uint next = 0; next < wq.size(); next++) { 3490 Node *m = wq.at(next); 3491 3492 for(;;) { 3493 // Loop over all nodes with identical inputs edges as m 3494 Node* k = m->find_similar(m->Opcode()); 3495 if (k == NULL) { 3496 break; 3497 } 3498 // Push their uses so we get a chance to remove node made 3499 // redundant 3500 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3501 Node* u = k->fast_out(i); 3502 if (u->Opcode() == Op_LShiftL || 3503 u->Opcode() == Op_AddL || 3504 u->Opcode() == Op_SubL || 3505 u->Opcode() == Op_AddP) { 3506 wq.push(u); 3507 } 3508 } 3509 // Replace all nodes with identical edges as m with m 3510 k->subsume_by(m, this); 3511 } 3512 } 3513 } 3514 break; 3515 } 3516 case Op_CmpUL: { 3517 if (!Matcher::has_match_rule(Op_CmpUL)) { 3518 // No support for unsigned long comparisons 3519 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3520 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3521 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3522 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3523 Node* andl = new AndLNode(orl, remove_sign_mask); 3524 Node* cmp = new CmpLNode(andl, n->in(2)); 3525 n->subsume_by(cmp, this); 3526 } 3527 break; 3528 } 3529 default: 3530 assert( !n->is_Call(), "" ); 3531 assert( !n->is_Mem(), "" ); 3532 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3533 break; 3534 } 3535 3536 // Collect CFG split points 3537 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3538 frc._tests.push(n); 3539 } 3540 } 3541 3542 //------------------------------final_graph_reshaping_walk--------------------- 3543 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3544 // requires that the walk visits a node's inputs before visiting the node. 3545 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3546 ResourceArea *area = Thread::current()->resource_area(); 3547 Unique_Node_List sfpt(area); 3548 3549 frc._visited.set(root->_idx); // first, mark node as visited 3550 uint cnt = root->req(); 3551 Node *n = root; 3552 uint i = 0; 3553 while (true) { 3554 if (i < cnt) { 3555 // Place all non-visited non-null inputs onto stack 3556 Node* m = n->in(i); 3557 ++i; 3558 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3559 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3560 // compute worst case interpreter size in case of a deoptimization 3561 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3562 3563 sfpt.push(m); 3564 } 3565 cnt = m->req(); 3566 nstack.push(n, i); // put on stack parent and next input's index 3567 n = m; 3568 i = 0; 3569 } 3570 } else { 3571 // Now do post-visit work 3572 final_graph_reshaping_impl( n, frc ); 3573 if (nstack.is_empty()) 3574 break; // finished 3575 n = nstack.node(); // Get node from stack 3576 cnt = n->req(); 3577 i = nstack.index(); 3578 nstack.pop(); // Shift to the next node on stack 3579 } 3580 } 3581 3582 // Skip next transformation if compressed oops are not used. 3583 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3584 (!UseCompressedOops && !UseCompressedClassPointers)) 3585 return; 3586 3587 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3588 // It could be done for an uncommon traps or any safepoints/calls 3589 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3590 while (sfpt.size() > 0) { 3591 n = sfpt.pop(); 3592 JVMState *jvms = n->as_SafePoint()->jvms(); 3593 assert(jvms != NULL, "sanity"); 3594 int start = jvms->debug_start(); 3595 int end = n->req(); 3596 bool is_uncommon = (n->is_CallStaticJava() && 3597 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3598 for (int j = start; j < end; j++) { 3599 Node* in = n->in(j); 3600 if (in->is_DecodeNarrowPtr()) { 3601 bool safe_to_skip = true; 3602 if (!is_uncommon ) { 3603 // Is it safe to skip? 3604 for (uint i = 0; i < in->outcnt(); i++) { 3605 Node* u = in->raw_out(i); 3606 if (!u->is_SafePoint() || 3607 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3608 safe_to_skip = false; 3609 } 3610 } 3611 } 3612 if (safe_to_skip) { 3613 n->set_req(j, in->in(1)); 3614 } 3615 if (in->outcnt() == 0) { 3616 in->disconnect_inputs(NULL, this); 3617 } 3618 } 3619 } 3620 } 3621 } 3622 3623 //------------------------------final_graph_reshaping-------------------------- 3624 // Final Graph Reshaping. 3625 // 3626 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3627 // and not commoned up and forced early. Must come after regular 3628 // optimizations to avoid GVN undoing the cloning. Clone constant 3629 // inputs to Loop Phis; these will be split by the allocator anyways. 3630 // Remove Opaque nodes. 3631 // (2) Move last-uses by commutative operations to the left input to encourage 3632 // Intel update-in-place two-address operations and better register usage 3633 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3634 // calls canonicalizing them back. 3635 // (3) Count the number of double-precision FP ops, single-precision FP ops 3636 // and call sites. On Intel, we can get correct rounding either by 3637 // forcing singles to memory (requires extra stores and loads after each 3638 // FP bytecode) or we can set a rounding mode bit (requires setting and 3639 // clearing the mode bit around call sites). The mode bit is only used 3640 // if the relative frequency of single FP ops to calls is low enough. 3641 // This is a key transform for SPEC mpeg_audio. 3642 // (4) Detect infinite loops; blobs of code reachable from above but not 3643 // below. Several of the Code_Gen algorithms fail on such code shapes, 3644 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3645 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3646 // Detection is by looking for IfNodes where only 1 projection is 3647 // reachable from below or CatchNodes missing some targets. 3648 // (5) Assert for insane oop offsets in debug mode. 3649 3650 bool Compile::final_graph_reshaping() { 3651 // an infinite loop may have been eliminated by the optimizer, 3652 // in which case the graph will be empty. 3653 if (root()->req() == 1) { 3654 record_method_not_compilable("trivial infinite loop"); 3655 return true; 3656 } 3657 3658 // Expensive nodes have their control input set to prevent the GVN 3659 // from freely commoning them. There's no GVN beyond this point so 3660 // no need to keep the control input. We want the expensive nodes to 3661 // be freely moved to the least frequent code path by gcm. 3662 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3663 for (int i = 0; i < expensive_count(); i++) { 3664 _expensive_nodes->at(i)->set_req(0, NULL); 3665 } 3666 3667 Final_Reshape_Counts frc; 3668 3669 // Visit everybody reachable! 3670 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3671 Node_Stack nstack(live_nodes() >> 1); 3672 final_graph_reshaping_walk(nstack, root(), frc); 3673 3674 // Check for unreachable (from below) code (i.e., infinite loops). 3675 for( uint i = 0; i < frc._tests.size(); i++ ) { 3676 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3677 // Get number of CFG targets. 3678 // Note that PCTables include exception targets after calls. 3679 uint required_outcnt = n->required_outcnt(); 3680 if (n->outcnt() != required_outcnt) { 3681 // Check for a few special cases. Rethrow Nodes never take the 3682 // 'fall-thru' path, so expected kids is 1 less. 3683 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3684 if (n->in(0)->in(0)->is_Call()) { 3685 CallNode *call = n->in(0)->in(0)->as_Call(); 3686 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3687 required_outcnt--; // Rethrow always has 1 less kid 3688 } else if (call->req() > TypeFunc::Parms && 3689 call->is_CallDynamicJava()) { 3690 // Check for null receiver. In such case, the optimizer has 3691 // detected that the virtual call will always result in a null 3692 // pointer exception. The fall-through projection of this CatchNode 3693 // will not be populated. 3694 Node *arg0 = call->in(TypeFunc::Parms); 3695 if (arg0->is_Type() && 3696 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3697 required_outcnt--; 3698 } 3699 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3700 call->req() > TypeFunc::Parms+1 && 3701 call->is_CallStaticJava()) { 3702 // Check for negative array length. In such case, the optimizer has 3703 // detected that the allocation attempt will always result in an 3704 // exception. There is no fall-through projection of this CatchNode . 3705 Node *arg1 = call->in(TypeFunc::Parms+1); 3706 if (arg1->is_Type() && 3707 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3708 required_outcnt--; 3709 } 3710 } 3711 } 3712 } 3713 // Recheck with a better notion of 'required_outcnt' 3714 if (n->outcnt() != required_outcnt) { 3715 record_method_not_compilable("malformed control flow"); 3716 return true; // Not all targets reachable! 3717 } 3718 } 3719 // Check that I actually visited all kids. Unreached kids 3720 // must be infinite loops. 3721 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3722 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3723 record_method_not_compilable("infinite loop"); 3724 return true; // Found unvisited kid; must be unreach 3725 } 3726 3727 // Here so verification code in final_graph_reshaping_walk() 3728 // always see an OuterStripMinedLoopEnd 3729 if (n->is_OuterStripMinedLoopEnd()) { 3730 IfNode* init_iff = n->as_If(); 3731 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3732 n->subsume_by(iff, this); 3733 } 3734 } 3735 3736 // If original bytecodes contained a mixture of floats and doubles 3737 // check if the optimizer has made it homogenous, item (3). 3738 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3739 frc.get_float_count() > 32 && 3740 frc.get_double_count() == 0 && 3741 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3742 set_24_bit_selection_and_mode( false, true ); 3743 } 3744 3745 set_java_calls(frc.get_java_call_count()); 3746 set_inner_loops(frc.get_inner_loop_count()); 3747 3748 // No infinite loops, no reason to bail out. 3749 return false; 3750 } 3751 3752 //-----------------------------too_many_traps---------------------------------- 3753 // Report if there are too many traps at the current method and bci. 3754 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3755 bool Compile::too_many_traps(ciMethod* method, 3756 int bci, 3757 Deoptimization::DeoptReason reason) { 3758 ciMethodData* md = method->method_data(); 3759 if (md->is_empty()) { 3760 // Assume the trap has not occurred, or that it occurred only 3761 // because of a transient condition during start-up in the interpreter. 3762 return false; 3763 } 3764 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3765 if (md->has_trap_at(bci, m, reason) != 0) { 3766 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3767 // Also, if there are multiple reasons, or if there is no per-BCI record, 3768 // assume the worst. 3769 if (log()) 3770 log()->elem("observe trap='%s' count='%d'", 3771 Deoptimization::trap_reason_name(reason), 3772 md->trap_count(reason)); 3773 return true; 3774 } else { 3775 // Ignore method/bci and see if there have been too many globally. 3776 return too_many_traps(reason, md); 3777 } 3778 } 3779 3780 // Less-accurate variant which does not require a method and bci. 3781 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3782 ciMethodData* logmd) { 3783 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3784 // Too many traps globally. 3785 // Note that we use cumulative trap_count, not just md->trap_count. 3786 if (log()) { 3787 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3788 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3789 Deoptimization::trap_reason_name(reason), 3790 mcount, trap_count(reason)); 3791 } 3792 return true; 3793 } else { 3794 // The coast is clear. 3795 return false; 3796 } 3797 } 3798 3799 //--------------------------too_many_recompiles-------------------------------- 3800 // Report if there are too many recompiles at the current method and bci. 3801 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3802 // Is not eager to return true, since this will cause the compiler to use 3803 // Action_none for a trap point, to avoid too many recompilations. 3804 bool Compile::too_many_recompiles(ciMethod* method, 3805 int bci, 3806 Deoptimization::DeoptReason reason) { 3807 ciMethodData* md = method->method_data(); 3808 if (md->is_empty()) { 3809 // Assume the trap has not occurred, or that it occurred only 3810 // because of a transient condition during start-up in the interpreter. 3811 return false; 3812 } 3813 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3814 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3815 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3816 Deoptimization::DeoptReason per_bc_reason 3817 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3818 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3819 if ((per_bc_reason == Deoptimization::Reason_none 3820 || md->has_trap_at(bci, m, reason) != 0) 3821 // The trap frequency measure we care about is the recompile count: 3822 && md->trap_recompiled_at(bci, m) 3823 && md->overflow_recompile_count() >= bc_cutoff) { 3824 // Do not emit a trap here if it has already caused recompilations. 3825 // Also, if there are multiple reasons, or if there is no per-BCI record, 3826 // assume the worst. 3827 if (log()) 3828 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3829 Deoptimization::trap_reason_name(reason), 3830 md->trap_count(reason), 3831 md->overflow_recompile_count()); 3832 return true; 3833 } else if (trap_count(reason) != 0 3834 && decompile_count() >= m_cutoff) { 3835 // Too many recompiles globally, and we have seen this sort of trap. 3836 // Use cumulative decompile_count, not just md->decompile_count. 3837 if (log()) 3838 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3839 Deoptimization::trap_reason_name(reason), 3840 md->trap_count(reason), trap_count(reason), 3841 md->decompile_count(), decompile_count()); 3842 return true; 3843 } else { 3844 // The coast is clear. 3845 return false; 3846 } 3847 } 3848 3849 // Compute when not to trap. Used by matching trap based nodes and 3850 // NullCheck optimization. 3851 void Compile::set_allowed_deopt_reasons() { 3852 _allowed_reasons = 0; 3853 if (is_method_compilation()) { 3854 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3855 assert(rs < BitsPerInt, "recode bit map"); 3856 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3857 _allowed_reasons |= nth_bit(rs); 3858 } 3859 } 3860 } 3861 } 3862 3863 bool Compile::is_compiling_clinit_for(ciKlass* k) { 3864 ciMethod* root = method(); // the root method of compilation 3865 return root->is_static_initializer() && root->holder() == k; // access in the context of clinit 3866 } 3867 3868 #ifndef PRODUCT 3869 //------------------------------verify_graph_edges--------------------------- 3870 // Walk the Graph and verify that there is a one-to-one correspondence 3871 // between Use-Def edges and Def-Use edges in the graph. 3872 void Compile::verify_graph_edges(bool no_dead_code) { 3873 if (VerifyGraphEdges) { 3874 ResourceArea *area = Thread::current()->resource_area(); 3875 Unique_Node_List visited(area); 3876 // Call recursive graph walk to check edges 3877 _root->verify_edges(visited); 3878 if (no_dead_code) { 3879 // Now make sure that no visited node is used by an unvisited node. 3880 bool dead_nodes = false; 3881 Unique_Node_List checked(area); 3882 while (visited.size() > 0) { 3883 Node* n = visited.pop(); 3884 checked.push(n); 3885 for (uint i = 0; i < n->outcnt(); i++) { 3886 Node* use = n->raw_out(i); 3887 if (checked.member(use)) continue; // already checked 3888 if (visited.member(use)) continue; // already in the graph 3889 if (use->is_Con()) continue; // a dead ConNode is OK 3890 // At this point, we have found a dead node which is DU-reachable. 3891 if (!dead_nodes) { 3892 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3893 dead_nodes = true; 3894 } 3895 use->dump(2); 3896 tty->print_cr("---"); 3897 checked.push(use); // No repeats; pretend it is now checked. 3898 } 3899 } 3900 assert(!dead_nodes, "using nodes must be reachable from root"); 3901 } 3902 } 3903 } 3904 3905 // Verify GC barriers consistency 3906 // Currently supported: 3907 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3908 void Compile::verify_barriers() { 3909 #if INCLUDE_G1GC || INCLUDE_SHENANDOAHGC 3910 if (UseG1GC || UseShenandoahGC) { 3911 // Verify G1 pre-barriers 3912 3913 #if INCLUDE_G1GC && INCLUDE_SHENANDOAHGC 3914 const int marking_offset = in_bytes(UseG1GC ? G1ThreadLocalData::satb_mark_queue_active_offset() 3915 : ShenandoahThreadLocalData::satb_mark_queue_active_offset()); 3916 #elif INCLUDE_G1GC 3917 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()); 3918 #else 3919 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()); 3920 #endif 3921 3922 ResourceArea *area = Thread::current()->resource_area(); 3923 Unique_Node_List visited(area); 3924 Node_List worklist(area); 3925 // We're going to walk control flow backwards starting from the Root 3926 worklist.push(_root); 3927 while (worklist.size() > 0) { 3928 Node* x = worklist.pop(); 3929 if (x == NULL || x == top()) continue; 3930 if (visited.member(x)) { 3931 continue; 3932 } else { 3933 visited.push(x); 3934 } 3935 3936 if (x->is_Region()) { 3937 for (uint i = 1; i < x->req(); i++) { 3938 worklist.push(x->in(i)); 3939 } 3940 } else { 3941 worklist.push(x->in(0)); 3942 // We are looking for the pattern: 3943 // /->ThreadLocal 3944 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3945 // \->ConI(0) 3946 // We want to verify that the If and the LoadB have the same control 3947 // See GraphKit::g1_write_barrier_pre() 3948 if (x->is_If()) { 3949 IfNode *iff = x->as_If(); 3950 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3951 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3952 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3953 && cmp->in(1)->is_Load()) { 3954 LoadNode* load = cmp->in(1)->as_Load(); 3955 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3956 && load->in(2)->in(3)->is_Con() 3957 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3958 3959 Node* if_ctrl = iff->in(0); 3960 Node* load_ctrl = load->in(0); 3961 3962 if (if_ctrl != load_ctrl) { 3963 // Skip possible CProj->NeverBranch in infinite loops 3964 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3965 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3966 if_ctrl = if_ctrl->in(0)->in(0); 3967 } 3968 } 3969 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3970 } 3971 } 3972 } 3973 } 3974 } 3975 } 3976 } 3977 #endif 3978 } 3979 3980 #endif 3981 3982 // The Compile object keeps track of failure reasons separately from the ciEnv. 3983 // This is required because there is not quite a 1-1 relation between the 3984 // ciEnv and its compilation task and the Compile object. Note that one 3985 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3986 // to backtrack and retry without subsuming loads. Other than this backtracking 3987 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3988 // by the logic in C2Compiler. 3989 void Compile::record_failure(const char* reason) { 3990 if (log() != NULL) { 3991 log()->elem("failure reason='%s' phase='compile'", reason); 3992 } 3993 if (_failure_reason == NULL) { 3994 // Record the first failure reason. 3995 _failure_reason = reason; 3996 } 3997 3998 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3999 C->print_method(PHASE_FAILURE); 4000 } 4001 _root = NULL; // flush the graph, too 4002 } 4003 4004 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 4005 : TraceTime(name, accumulator, CITime, CITimeVerbose), 4006 _phase_name(name), _dolog(CITimeVerbose) 4007 { 4008 if (_dolog) { 4009 C = Compile::current(); 4010 _log = C->log(); 4011 } else { 4012 C = NULL; 4013 _log = NULL; 4014 } 4015 if (_log != NULL) { 4016 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4017 _log->stamp(); 4018 _log->end_head(); 4019 } 4020 } 4021 4022 Compile::TracePhase::~TracePhase() { 4023 4024 C = Compile::current(); 4025 if (_dolog) { 4026 _log = C->log(); 4027 } else { 4028 _log = NULL; 4029 } 4030 4031 #ifdef ASSERT 4032 if (PrintIdealNodeCount) { 4033 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4034 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 4035 } 4036 4037 if (VerifyIdealNodeCount) { 4038 Compile::current()->print_missing_nodes(); 4039 } 4040 #endif 4041 4042 if (_log != NULL) { 4043 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4044 } 4045 } 4046 4047 //============================================================================= 4048 // Two Constant's are equal when the type and the value are equal. 4049 bool Compile::Constant::operator==(const Constant& other) { 4050 if (type() != other.type() ) return false; 4051 if (can_be_reused() != other.can_be_reused()) return false; 4052 // For floating point values we compare the bit pattern. 4053 switch (type()) { 4054 case T_INT: 4055 case T_FLOAT: return (_v._value.i == other._v._value.i); 4056 case T_LONG: 4057 case T_DOUBLE: return (_v._value.j == other._v._value.j); 4058 case T_OBJECT: 4059 case T_ADDRESS: return (_v._value.l == other._v._value.l); 4060 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 4061 case T_METADATA: return (_v._metadata == other._v._metadata); 4062 default: ShouldNotReachHere(); return false; 4063 } 4064 } 4065 4066 static int type_to_size_in_bytes(BasicType t) { 4067 switch (t) { 4068 case T_INT: return sizeof(jint ); 4069 case T_LONG: return sizeof(jlong ); 4070 case T_FLOAT: return sizeof(jfloat ); 4071 case T_DOUBLE: return sizeof(jdouble); 4072 case T_METADATA: return sizeof(Metadata*); 4073 // We use T_VOID as marker for jump-table entries (labels) which 4074 // need an internal word relocation. 4075 case T_VOID: 4076 case T_ADDRESS: 4077 case T_OBJECT: return sizeof(jobject); 4078 default: 4079 ShouldNotReachHere(); 4080 return -1; 4081 } 4082 } 4083 4084 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 4085 // sort descending 4086 if (a->freq() > b->freq()) return -1; 4087 if (a->freq() < b->freq()) return 1; 4088 return 0; 4089 } 4090 4091 void Compile::ConstantTable::calculate_offsets_and_size() { 4092 // First, sort the array by frequencies. 4093 _constants.sort(qsort_comparator); 4094 4095 #ifdef ASSERT 4096 // Make sure all jump-table entries were sorted to the end of the 4097 // array (they have a negative frequency). 4098 bool found_void = false; 4099 for (int i = 0; i < _constants.length(); i++) { 4100 Constant con = _constants.at(i); 4101 if (con.type() == T_VOID) 4102 found_void = true; // jump-tables 4103 else 4104 assert(!found_void, "wrong sorting"); 4105 } 4106 #endif 4107 4108 int offset = 0; 4109 for (int i = 0; i < _constants.length(); i++) { 4110 Constant* con = _constants.adr_at(i); 4111 4112 // Align offset for type. 4113 int typesize = type_to_size_in_bytes(con->type()); 4114 offset = align_up(offset, typesize); 4115 con->set_offset(offset); // set constant's offset 4116 4117 if (con->type() == T_VOID) { 4118 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4119 offset = offset + typesize * n->outcnt(); // expand jump-table 4120 } else { 4121 offset = offset + typesize; 4122 } 4123 } 4124 4125 // Align size up to the next section start (which is insts; see 4126 // CodeBuffer::align_at_start). 4127 assert(_size == -1, "already set?"); 4128 _size = align_up(offset, (int)CodeEntryAlignment); 4129 } 4130 4131 void Compile::ConstantTable::emit(CodeBuffer& cb) { 4132 MacroAssembler _masm(&cb); 4133 for (int i = 0; i < _constants.length(); i++) { 4134 Constant con = _constants.at(i); 4135 address constant_addr = NULL; 4136 switch (con.type()) { 4137 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4138 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4139 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4140 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4141 case T_OBJECT: { 4142 jobject obj = con.get_jobject(); 4143 int oop_index = _masm.oop_recorder()->find_index(obj); 4144 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4145 break; 4146 } 4147 case T_ADDRESS: { 4148 address addr = (address) con.get_jobject(); 4149 constant_addr = _masm.address_constant(addr); 4150 break; 4151 } 4152 // We use T_VOID as marker for jump-table entries (labels) which 4153 // need an internal word relocation. 4154 case T_VOID: { 4155 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4156 // Fill the jump-table with a dummy word. The real value is 4157 // filled in later in fill_jump_table. 4158 address dummy = (address) n; 4159 constant_addr = _masm.address_constant(dummy); 4160 // Expand jump-table 4161 for (uint i = 1; i < n->outcnt(); i++) { 4162 address temp_addr = _masm.address_constant(dummy + i); 4163 assert(temp_addr, "consts section too small"); 4164 } 4165 break; 4166 } 4167 case T_METADATA: { 4168 Metadata* obj = con.get_metadata(); 4169 int metadata_index = _masm.oop_recorder()->find_index(obj); 4170 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4171 break; 4172 } 4173 default: ShouldNotReachHere(); 4174 } 4175 assert(constant_addr, "consts section too small"); 4176 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4177 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4178 } 4179 } 4180 4181 int Compile::ConstantTable::find_offset(Constant& con) const { 4182 int idx = _constants.find(con); 4183 guarantee(idx != -1, "constant must be in constant table"); 4184 int offset = _constants.at(idx).offset(); 4185 guarantee(offset != -1, "constant table not emitted yet?"); 4186 return offset; 4187 } 4188 4189 void Compile::ConstantTable::add(Constant& con) { 4190 if (con.can_be_reused()) { 4191 int idx = _constants.find(con); 4192 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4193 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4194 return; 4195 } 4196 } 4197 (void) _constants.append(con); 4198 } 4199 4200 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4201 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4202 Constant con(type, value, b->_freq); 4203 add(con); 4204 return con; 4205 } 4206 4207 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4208 Constant con(metadata); 4209 add(con); 4210 return con; 4211 } 4212 4213 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4214 jvalue value; 4215 BasicType type = oper->type()->basic_type(); 4216 switch (type) { 4217 case T_LONG: value.j = oper->constantL(); break; 4218 case T_FLOAT: value.f = oper->constantF(); break; 4219 case T_DOUBLE: value.d = oper->constantD(); break; 4220 case T_OBJECT: 4221 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4222 case T_METADATA: return add((Metadata*)oper->constant()); break; 4223 default: guarantee(false, "unhandled type: %s", type2name(type)); 4224 } 4225 return add(n, type, value); 4226 } 4227 4228 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4229 jvalue value; 4230 // We can use the node pointer here to identify the right jump-table 4231 // as this method is called from Compile::Fill_buffer right before 4232 // the MachNodes are emitted and the jump-table is filled (means the 4233 // MachNode pointers do not change anymore). 4234 value.l = (jobject) n; 4235 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4236 add(con); 4237 return con; 4238 } 4239 4240 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4241 // If called from Compile::scratch_emit_size do nothing. 4242 if (Compile::current()->in_scratch_emit_size()) return; 4243 4244 assert(labels.is_nonempty(), "must be"); 4245 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4246 4247 // Since MachConstantNode::constant_offset() also contains 4248 // table_base_offset() we need to subtract the table_base_offset() 4249 // to get the plain offset into the constant table. 4250 int offset = n->constant_offset() - table_base_offset(); 4251 4252 MacroAssembler _masm(&cb); 4253 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4254 4255 for (uint i = 0; i < n->outcnt(); i++) { 4256 address* constant_addr = &jump_table_base[i]; 4257 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)); 4258 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4259 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4260 } 4261 } 4262 4263 //----------------------------static_subtype_check----------------------------- 4264 // Shortcut important common cases when superklass is exact: 4265 // (0) superklass is java.lang.Object (can occur in reflective code) 4266 // (1) subklass is already limited to a subtype of superklass => always ok 4267 // (2) subklass does not overlap with superklass => always fail 4268 // (3) superklass has NO subtypes and we can check with a simple compare. 4269 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4270 if (StressReflectiveCode) { 4271 return SSC_full_test; // Let caller generate the general case. 4272 } 4273 4274 if (superk == env()->Object_klass()) { 4275 return SSC_always_true; // (0) this test cannot fail 4276 } 4277 4278 ciType* superelem = superk; 4279 if (superelem->is_array_klass()) 4280 superelem = superelem->as_array_klass()->base_element_type(); 4281 4282 if (!subk->is_interface()) { // cannot trust static interface types yet 4283 if (subk->is_subtype_of(superk)) { 4284 return SSC_always_true; // (1) false path dead; no dynamic test needed 4285 } 4286 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4287 !superk->is_subtype_of(subk)) { 4288 return SSC_always_false; 4289 } 4290 } 4291 4292 // If casting to an instance klass, it must have no subtypes 4293 if (superk->is_interface()) { 4294 // Cannot trust interfaces yet. 4295 // %%% S.B. superk->nof_implementors() == 1 4296 } else if (superelem->is_instance_klass()) { 4297 ciInstanceKlass* ik = superelem->as_instance_klass(); 4298 if (!ik->has_subklass() && !ik->is_interface()) { 4299 if (!ik->is_final()) { 4300 // Add a dependency if there is a chance of a later subclass. 4301 dependencies()->assert_leaf_type(ik); 4302 } 4303 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4304 } 4305 } else { 4306 // A primitive array type has no subtypes. 4307 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4308 } 4309 4310 return SSC_full_test; 4311 } 4312 4313 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4314 #ifdef _LP64 4315 // The scaled index operand to AddP must be a clean 64-bit value. 4316 // Java allows a 32-bit int to be incremented to a negative 4317 // value, which appears in a 64-bit register as a large 4318 // positive number. Using that large positive number as an 4319 // operand in pointer arithmetic has bad consequences. 4320 // On the other hand, 32-bit overflow is rare, and the possibility 4321 // can often be excluded, if we annotate the ConvI2L node with 4322 // a type assertion that its value is known to be a small positive 4323 // number. (The prior range check has ensured this.) 4324 // This assertion is used by ConvI2LNode::Ideal. 4325 int index_max = max_jint - 1; // array size is max_jint, index is one less 4326 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4327 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4328 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4329 #endif 4330 return idx; 4331 } 4332 4333 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4334 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4335 if (ctrl != NULL) { 4336 // Express control dependency by a CastII node with a narrow type. 4337 value = new CastIINode(value, itype, false, true /* range check dependency */); 4338 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4339 // node from floating above the range check during loop optimizations. Otherwise, the 4340 // ConvI2L node may be eliminated independently of the range check, causing the data path 4341 // to become TOP while the control path is still there (although it's unreachable). 4342 value->set_req(0, ctrl); 4343 // Save CastII node to remove it after loop optimizations. 4344 phase->C->add_range_check_cast(value); 4345 value = phase->transform(value); 4346 } 4347 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4348 return phase->transform(new ConvI2LNode(value, ltype)); 4349 } 4350 4351 void Compile::print_inlining_stream_free() { 4352 if (_print_inlining_stream != NULL) { 4353 _print_inlining_stream->~stringStream(); 4354 _print_inlining_stream = NULL; 4355 } 4356 } 4357 4358 // The message about the current inlining is accumulated in 4359 // _print_inlining_stream and transfered into the _print_inlining_list 4360 // once we know whether inlining succeeds or not. For regular 4361 // inlining, messages are appended to the buffer pointed by 4362 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4363 // a new buffer is added after _print_inlining_idx in the list. This 4364 // way we can update the inlining message for late inlining call site 4365 // when the inlining is attempted again. 4366 void Compile::print_inlining_init() { 4367 if (print_inlining() || print_intrinsics()) { 4368 // print_inlining_init is actually called several times. 4369 print_inlining_stream_free(); 4370 _print_inlining_stream = new stringStream(); 4371 // Watch out: The memory initialized by the constructor call PrintInliningBuffer() 4372 // will be copied into the only initial element. The default destructor of 4373 // PrintInliningBuffer will be called when leaving the scope here. If it 4374 // would destuct the enclosed stringStream _print_inlining_list[0]->_ss 4375 // would be destructed, too! 4376 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4377 } 4378 } 4379 4380 void Compile::print_inlining_reinit() { 4381 if (print_inlining() || print_intrinsics()) { 4382 print_inlining_stream_free(); 4383 // Re allocate buffer when we change ResourceMark 4384 _print_inlining_stream = new stringStream(); 4385 } 4386 } 4387 4388 void Compile::print_inlining_reset() { 4389 _print_inlining_stream->reset(); 4390 } 4391 4392 void Compile::print_inlining_commit() { 4393 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4394 // Transfer the message from _print_inlining_stream to the current 4395 // _print_inlining_list buffer and clear _print_inlining_stream. 4396 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 4397 print_inlining_reset(); 4398 } 4399 4400 void Compile::print_inlining_push() { 4401 // Add new buffer to the _print_inlining_list at current position 4402 _print_inlining_idx++; 4403 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4404 } 4405 4406 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4407 return _print_inlining_list->at(_print_inlining_idx); 4408 } 4409 4410 void Compile::print_inlining_update(CallGenerator* cg) { 4411 if (print_inlining() || print_intrinsics()) { 4412 if (!cg->is_late_inline()) { 4413 if (print_inlining_current().cg() != NULL) { 4414 print_inlining_push(); 4415 } 4416 print_inlining_commit(); 4417 } else { 4418 if (print_inlining_current().cg() != cg && 4419 (print_inlining_current().cg() != NULL || 4420 print_inlining_current().ss()->size() != 0)) { 4421 print_inlining_push(); 4422 } 4423 print_inlining_commit(); 4424 print_inlining_current().set_cg(cg); 4425 } 4426 } 4427 } 4428 4429 void Compile::print_inlining_move_to(CallGenerator* cg) { 4430 // We resume inlining at a late inlining call site. Locate the 4431 // corresponding inlining buffer so that we can update it. 4432 if (print_inlining()) { 4433 for (int i = 0; i < _print_inlining_list->length(); i++) { 4434 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4435 _print_inlining_idx = i; 4436 return; 4437 } 4438 } 4439 ShouldNotReachHere(); 4440 } 4441 } 4442 4443 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4444 if (print_inlining()) { 4445 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4446 assert(print_inlining_current().cg() == cg, "wrong entry"); 4447 // replace message with new message 4448 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4449 print_inlining_commit(); 4450 print_inlining_current().set_cg(cg); 4451 } 4452 } 4453 4454 void Compile::print_inlining_assert_ready() { 4455 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4456 } 4457 4458 void Compile::process_print_inlining() { 4459 bool do_print_inlining = print_inlining() || print_intrinsics(); 4460 if (do_print_inlining || log() != NULL) { 4461 // Print inlining message for candidates that we couldn't inline 4462 // for lack of space 4463 for (int i = 0; i < _late_inlines.length(); i++) { 4464 CallGenerator* cg = _late_inlines.at(i); 4465 if (!cg->is_mh_late_inline()) { 4466 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4467 if (do_print_inlining) { 4468 cg->print_inlining_late(msg); 4469 } 4470 log_late_inline_failure(cg, msg); 4471 } 4472 } 4473 } 4474 if (do_print_inlining) { 4475 ResourceMark rm; 4476 stringStream ss; 4477 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once."); 4478 for (int i = 0; i < _print_inlining_list->length(); i++) { 4479 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4480 _print_inlining_list->at(i).freeStream(); 4481 } 4482 // Reset _print_inlining_list, it only contains destructed objects. 4483 // It is on the arena, so it will be freed when the arena is reset. 4484 _print_inlining_list = NULL; 4485 // _print_inlining_stream won't be used anymore, either. 4486 print_inlining_stream_free(); 4487 size_t end = ss.size(); 4488 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4489 strncpy(_print_inlining_output, ss.base(), end+1); 4490 _print_inlining_output[end] = 0; 4491 } 4492 } 4493 4494 void Compile::dump_print_inlining() { 4495 if (_print_inlining_output != NULL) { 4496 tty->print_raw(_print_inlining_output); 4497 } 4498 } 4499 4500 void Compile::log_late_inline(CallGenerator* cg) { 4501 if (log() != NULL) { 4502 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4503 cg->unique_id()); 4504 JVMState* p = cg->call_node()->jvms(); 4505 while (p != NULL) { 4506 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4507 p = p->caller(); 4508 } 4509 log()->tail("late_inline"); 4510 } 4511 } 4512 4513 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4514 log_late_inline(cg); 4515 if (log() != NULL) { 4516 log()->inline_fail(msg); 4517 } 4518 } 4519 4520 void Compile::log_inline_id(CallGenerator* cg) { 4521 if (log() != NULL) { 4522 // The LogCompilation tool needs a unique way to identify late 4523 // inline call sites. This id must be unique for this call site in 4524 // this compilation. Try to have it unique across compilations as 4525 // well because it can be convenient when grepping through the log 4526 // file. 4527 // Distinguish OSR compilations from others in case CICountOSR is 4528 // on. 4529 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4530 cg->set_unique_id(id); 4531 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4532 } 4533 } 4534 4535 void Compile::log_inline_failure(const char* msg) { 4536 if (C->log() != NULL) { 4537 C->log()->inline_fail(msg); 4538 } 4539 } 4540 4541 4542 // Dump inlining replay data to the stream. 4543 // Don't change thread state and acquire any locks. 4544 void Compile::dump_inline_data(outputStream* out) { 4545 InlineTree* inl_tree = ilt(); 4546 if (inl_tree != NULL) { 4547 out->print(" inline %d", inl_tree->count()); 4548 inl_tree->dump_replay_data(out); 4549 } 4550 } 4551 4552 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4553 if (n1->Opcode() < n2->Opcode()) return -1; 4554 else if (n1->Opcode() > n2->Opcode()) return 1; 4555 4556 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4557 for (uint i = 1; i < n1->req(); i++) { 4558 if (n1->in(i) < n2->in(i)) return -1; 4559 else if (n1->in(i) > n2->in(i)) return 1; 4560 } 4561 4562 return 0; 4563 } 4564 4565 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4566 Node* n1 = *n1p; 4567 Node* n2 = *n2p; 4568 4569 return cmp_expensive_nodes(n1, n2); 4570 } 4571 4572 void Compile::sort_expensive_nodes() { 4573 if (!expensive_nodes_sorted()) { 4574 _expensive_nodes->sort(cmp_expensive_nodes); 4575 } 4576 } 4577 4578 bool Compile::expensive_nodes_sorted() const { 4579 for (int i = 1; i < _expensive_nodes->length(); i++) { 4580 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4581 return false; 4582 } 4583 } 4584 return true; 4585 } 4586 4587 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4588 if (_expensive_nodes->length() == 0) { 4589 return false; 4590 } 4591 4592 assert(OptimizeExpensiveOps, "optimization off?"); 4593 4594 // Take this opportunity to remove dead nodes from the list 4595 int j = 0; 4596 for (int i = 0; i < _expensive_nodes->length(); i++) { 4597 Node* n = _expensive_nodes->at(i); 4598 if (!n->is_unreachable(igvn)) { 4599 assert(n->is_expensive(), "should be expensive"); 4600 _expensive_nodes->at_put(j, n); 4601 j++; 4602 } 4603 } 4604 _expensive_nodes->trunc_to(j); 4605 4606 // Then sort the list so that similar nodes are next to each other 4607 // and check for at least two nodes of identical kind with same data 4608 // inputs. 4609 sort_expensive_nodes(); 4610 4611 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4612 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4613 return true; 4614 } 4615 } 4616 4617 return false; 4618 } 4619 4620 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4621 if (_expensive_nodes->length() == 0) { 4622 return; 4623 } 4624 4625 assert(OptimizeExpensiveOps, "optimization off?"); 4626 4627 // Sort to bring similar nodes next to each other and clear the 4628 // control input of nodes for which there's only a single copy. 4629 sort_expensive_nodes(); 4630 4631 int j = 0; 4632 int identical = 0; 4633 int i = 0; 4634 bool modified = false; 4635 for (; i < _expensive_nodes->length()-1; i++) { 4636 assert(j <= i, "can't write beyond current index"); 4637 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4638 identical++; 4639 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4640 continue; 4641 } 4642 if (identical > 0) { 4643 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4644 identical = 0; 4645 } else { 4646 Node* n = _expensive_nodes->at(i); 4647 igvn.replace_input_of(n, 0, NULL); 4648 igvn.hash_insert(n); 4649 modified = true; 4650 } 4651 } 4652 if (identical > 0) { 4653 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4654 } else if (_expensive_nodes->length() >= 1) { 4655 Node* n = _expensive_nodes->at(i); 4656 igvn.replace_input_of(n, 0, NULL); 4657 igvn.hash_insert(n); 4658 modified = true; 4659 } 4660 _expensive_nodes->trunc_to(j); 4661 if (modified) { 4662 igvn.optimize(); 4663 } 4664 } 4665 4666 void Compile::add_expensive_node(Node * n) { 4667 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4668 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4669 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4670 if (OptimizeExpensiveOps) { 4671 _expensive_nodes->append(n); 4672 } else { 4673 // Clear control input and let IGVN optimize expensive nodes if 4674 // OptimizeExpensiveOps is off. 4675 n->set_req(0, NULL); 4676 } 4677 } 4678 4679 /** 4680 * Remove the speculative part of types and clean up the graph 4681 */ 4682 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4683 if (UseTypeSpeculation) { 4684 Unique_Node_List worklist; 4685 worklist.push(root()); 4686 int modified = 0; 4687 // Go over all type nodes that carry a speculative type, drop the 4688 // speculative part of the type and enqueue the node for an igvn 4689 // which may optimize it out. 4690 for (uint next = 0; next < worklist.size(); ++next) { 4691 Node *n = worklist.at(next); 4692 if (n->is_Type()) { 4693 TypeNode* tn = n->as_Type(); 4694 const Type* t = tn->type(); 4695 const Type* t_no_spec = t->remove_speculative(); 4696 if (t_no_spec != t) { 4697 bool in_hash = igvn.hash_delete(n); 4698 assert(in_hash, "node should be in igvn hash table"); 4699 tn->set_type(t_no_spec); 4700 igvn.hash_insert(n); 4701 igvn._worklist.push(n); // give it a chance to go away 4702 modified++; 4703 } 4704 } 4705 uint max = n->len(); 4706 for( uint i = 0; i < max; ++i ) { 4707 Node *m = n->in(i); 4708 if (not_a_node(m)) continue; 4709 worklist.push(m); 4710 } 4711 } 4712 // Drop the speculative part of all types in the igvn's type table 4713 igvn.remove_speculative_types(); 4714 if (modified > 0) { 4715 igvn.optimize(); 4716 } 4717 #ifdef ASSERT 4718 // Verify that after the IGVN is over no speculative type has resurfaced 4719 worklist.clear(); 4720 worklist.push(root()); 4721 for (uint next = 0; next < worklist.size(); ++next) { 4722 Node *n = worklist.at(next); 4723 const Type* t = igvn.type_or_null(n); 4724 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4725 if (n->is_Type()) { 4726 t = n->as_Type()->type(); 4727 assert(t == t->remove_speculative(), "no more speculative types"); 4728 } 4729 uint max = n->len(); 4730 for( uint i = 0; i < max; ++i ) { 4731 Node *m = n->in(i); 4732 if (not_a_node(m)) continue; 4733 worklist.push(m); 4734 } 4735 } 4736 igvn.check_no_speculative_types(); 4737 #endif 4738 } 4739 } 4740 4741 // Auxiliary method to support randomized stressing/fuzzing. 4742 // 4743 // This method can be called the arbitrary number of times, with current count 4744 // as the argument. The logic allows selecting a single candidate from the 4745 // running list of candidates as follows: 4746 // int count = 0; 4747 // Cand* selected = null; 4748 // while(cand = cand->next()) { 4749 // if (randomized_select(++count)) { 4750 // selected = cand; 4751 // } 4752 // } 4753 // 4754 // Including count equalizes the chances any candidate is "selected". 4755 // This is useful when we don't have the complete list of candidates to choose 4756 // from uniformly. In this case, we need to adjust the randomicity of the 4757 // selection, or else we will end up biasing the selection towards the latter 4758 // candidates. 4759 // 4760 // Quick back-envelope calculation shows that for the list of n candidates 4761 // the equal probability for the candidate to persist as "best" can be 4762 // achieved by replacing it with "next" k-th candidate with the probability 4763 // of 1/k. It can be easily shown that by the end of the run, the 4764 // probability for any candidate is converged to 1/n, thus giving the 4765 // uniform distribution among all the candidates. 4766 // 4767 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4768 #define RANDOMIZED_DOMAIN_POW 29 4769 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4770 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4771 bool Compile::randomized_select(int count) { 4772 assert(count > 0, "only positive"); 4773 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4774 } 4775 4776 CloneMap& Compile::clone_map() { return _clone_map; } 4777 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4778 4779 void NodeCloneInfo::dump() const { 4780 tty->print(" {%d:%d} ", idx(), gen()); 4781 } 4782 4783 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4784 uint64_t val = value(old->_idx); 4785 NodeCloneInfo cio(val); 4786 assert(val != 0, "old node should be in the map"); 4787 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4788 insert(nnn->_idx, cin.get()); 4789 #ifndef PRODUCT 4790 if (is_debug()) { 4791 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4792 } 4793 #endif 4794 } 4795 4796 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4797 NodeCloneInfo cio(value(old->_idx)); 4798 if (cio.get() == 0) { 4799 cio.set(old->_idx, 0); 4800 insert(old->_idx, cio.get()); 4801 #ifndef PRODUCT 4802 if (is_debug()) { 4803 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4804 } 4805 #endif 4806 } 4807 clone(old, nnn, gen); 4808 } 4809 4810 int CloneMap::max_gen() const { 4811 int g = 0; 4812 DictI di(_dict); 4813 for(; di.test(); ++di) { 4814 int t = gen(di._key); 4815 if (g < t) { 4816 g = t; 4817 #ifndef PRODUCT 4818 if (is_debug()) { 4819 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4820 } 4821 #endif 4822 } 4823 } 4824 return g; 4825 } 4826 4827 void CloneMap::dump(node_idx_t key) const { 4828 uint64_t val = value(key); 4829 if (val != 0) { 4830 NodeCloneInfo ni(val); 4831 ni.dump(); 4832 } 4833 }