1 /* 2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "gc/shared/barrierSet.hpp" 37 #include "gc/shared/c2/barrierSetC2.hpp" 38 #include "memory/resourceArea.hpp" 39 #include "opto/addnode.hpp" 40 #include "opto/block.hpp" 41 #include "opto/c2compiler.hpp" 42 #include "opto/callGenerator.hpp" 43 #include "opto/callnode.hpp" 44 #include "opto/castnode.hpp" 45 #include "opto/cfgnode.hpp" 46 #include "opto/chaitin.hpp" 47 #include "opto/compile.hpp" 48 #include "opto/connode.hpp" 49 #include "opto/convertnode.hpp" 50 #include "opto/divnode.hpp" 51 #include "opto/escape.hpp" 52 #include "opto/idealGraphPrinter.hpp" 53 #include "opto/loopnode.hpp" 54 #include "opto/machnode.hpp" 55 #include "opto/macro.hpp" 56 #include "opto/matcher.hpp" 57 #include "opto/mathexactnode.hpp" 58 #include "opto/memnode.hpp" 59 #include "opto/mulnode.hpp" 60 #include "opto/narrowptrnode.hpp" 61 #include "opto/node.hpp" 62 #include "opto/opcodes.hpp" 63 #include "opto/output.hpp" 64 #include "opto/parse.hpp" 65 #include "opto/phaseX.hpp" 66 #include "opto/rootnode.hpp" 67 #include "opto/runtime.hpp" 68 #include "opto/stringopts.hpp" 69 #include "opto/type.hpp" 70 #include "opto/valuetypenode.hpp" 71 #include "opto/vectornode.hpp" 72 #include "runtime/arguments.hpp" 73 #include "runtime/sharedRuntime.hpp" 74 #include "runtime/signature.hpp" 75 #include "runtime/stubRoutines.hpp" 76 #include "runtime/timer.hpp" 77 #include "utilities/align.hpp" 78 #include "utilities/copy.hpp" 79 #include "utilities/macros.hpp" 80 #if INCLUDE_ZGC 81 #include "gc/z/c2/zBarrierSetC2.hpp" 82 #endif 83 84 85 // -------------------- Compile::mach_constant_base_node ----------------------- 86 // Constant table base node singleton. 87 MachConstantBaseNode* Compile::mach_constant_base_node() { 88 if (_mach_constant_base_node == NULL) { 89 _mach_constant_base_node = new MachConstantBaseNode(); 90 _mach_constant_base_node->add_req(C->root()); 91 } 92 return _mach_constant_base_node; 93 } 94 95 96 /// Support for intrinsics. 97 98 // Return the index at which m must be inserted (or already exists). 99 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 100 class IntrinsicDescPair { 101 private: 102 ciMethod* _m; 103 bool _is_virtual; 104 public: 105 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 106 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 107 ciMethod* m= elt->method(); 108 ciMethod* key_m = key->_m; 109 if (key_m < m) return -1; 110 else if (key_m > m) return 1; 111 else { 112 bool is_virtual = elt->is_virtual(); 113 bool key_virtual = key->_is_virtual; 114 if (key_virtual < is_virtual) return -1; 115 else if (key_virtual > is_virtual) return 1; 116 else return 0; 117 } 118 } 119 }; 120 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 121 #ifdef ASSERT 122 for (int i = 1; i < _intrinsics->length(); i++) { 123 CallGenerator* cg1 = _intrinsics->at(i-1); 124 CallGenerator* cg2 = _intrinsics->at(i); 125 assert(cg1->method() != cg2->method() 126 ? cg1->method() < cg2->method() 127 : cg1->is_virtual() < cg2->is_virtual(), 128 "compiler intrinsics list must stay sorted"); 129 } 130 #endif 131 IntrinsicDescPair pair(m, is_virtual); 132 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 133 } 134 135 void Compile::register_intrinsic(CallGenerator* cg) { 136 if (_intrinsics == NULL) { 137 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 138 } 139 int len = _intrinsics->length(); 140 bool found = false; 141 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 142 assert(!found, "registering twice"); 143 _intrinsics->insert_before(index, cg); 144 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 145 } 146 147 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 148 assert(m->is_loaded(), "don't try this on unloaded methods"); 149 if (_intrinsics != NULL) { 150 bool found = false; 151 int index = intrinsic_insertion_index(m, is_virtual, found); 152 if (found) { 153 return _intrinsics->at(index); 154 } 155 } 156 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 157 if (m->intrinsic_id() != vmIntrinsics::_none && 158 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 159 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 160 if (cg != NULL) { 161 // Save it for next time: 162 register_intrinsic(cg); 163 return cg; 164 } else { 165 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 166 } 167 } 168 return NULL; 169 } 170 171 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 172 // in library_call.cpp. 173 174 175 #ifndef PRODUCT 176 // statistics gathering... 177 178 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 179 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 180 181 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 182 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 183 int oflags = _intrinsic_hist_flags[id]; 184 assert(flags != 0, "what happened?"); 185 if (is_virtual) { 186 flags |= _intrinsic_virtual; 187 } 188 bool changed = (flags != oflags); 189 if ((flags & _intrinsic_worked) != 0) { 190 juint count = (_intrinsic_hist_count[id] += 1); 191 if (count == 1) { 192 changed = true; // first time 193 } 194 // increment the overall count also: 195 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 196 } 197 if (changed) { 198 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 199 // Something changed about the intrinsic's virtuality. 200 if ((flags & _intrinsic_virtual) != 0) { 201 // This is the first use of this intrinsic as a virtual call. 202 if (oflags != 0) { 203 // We already saw it as a non-virtual, so note both cases. 204 flags |= _intrinsic_both; 205 } 206 } else if ((oflags & _intrinsic_both) == 0) { 207 // This is the first use of this intrinsic as a non-virtual 208 flags |= _intrinsic_both; 209 } 210 } 211 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 212 } 213 // update the overall flags also: 214 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 215 return changed; 216 } 217 218 static char* format_flags(int flags, char* buf) { 219 buf[0] = 0; 220 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 221 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 222 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 223 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 224 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 225 if (buf[0] == 0) strcat(buf, ","); 226 assert(buf[0] == ',', "must be"); 227 return &buf[1]; 228 } 229 230 void Compile::print_intrinsic_statistics() { 231 char flagsbuf[100]; 232 ttyLocker ttyl; 233 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 234 tty->print_cr("Compiler intrinsic usage:"); 235 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 236 if (total == 0) total = 1; // avoid div0 in case of no successes 237 #define PRINT_STAT_LINE(name, c, f) \ 238 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 239 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 240 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 241 int flags = _intrinsic_hist_flags[id]; 242 juint count = _intrinsic_hist_count[id]; 243 if ((flags | count) != 0) { 244 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 245 } 246 } 247 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 248 if (xtty != NULL) xtty->tail("statistics"); 249 } 250 251 void Compile::print_statistics() { 252 { ttyLocker ttyl; 253 if (xtty != NULL) xtty->head("statistics type='opto'"); 254 Parse::print_statistics(); 255 PhaseCCP::print_statistics(); 256 PhaseRegAlloc::print_statistics(); 257 Scheduling::print_statistics(); 258 PhasePeephole::print_statistics(); 259 PhaseIdealLoop::print_statistics(); 260 if (xtty != NULL) xtty->tail("statistics"); 261 } 262 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 263 // put this under its own <statistics> element. 264 print_intrinsic_statistics(); 265 } 266 } 267 #endif //PRODUCT 268 269 // Support for bundling info 270 Bundle* Compile::node_bundling(const Node *n) { 271 assert(valid_bundle_info(n), "oob"); 272 return &_node_bundling_base[n->_idx]; 273 } 274 275 bool Compile::valid_bundle_info(const Node *n) { 276 return (_node_bundling_limit > n->_idx); 277 } 278 279 280 void Compile::gvn_replace_by(Node* n, Node* nn) { 281 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 282 Node* use = n->last_out(i); 283 bool is_in_table = initial_gvn()->hash_delete(use); 284 uint uses_found = 0; 285 for (uint j = 0; j < use->len(); j++) { 286 if (use->in(j) == n) { 287 if (j < use->req()) 288 use->set_req(j, nn); 289 else 290 use->set_prec(j, nn); 291 uses_found++; 292 } 293 } 294 if (is_in_table) { 295 // reinsert into table 296 initial_gvn()->hash_find_insert(use); 297 } 298 record_for_igvn(use); 299 i -= uses_found; // we deleted 1 or more copies of this edge 300 } 301 } 302 303 304 static inline bool not_a_node(const Node* n) { 305 if (n == NULL) return true; 306 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 307 if (*(address*)n == badAddress) return true; // kill by Node::destruct 308 return false; 309 } 310 311 // Identify all nodes that are reachable from below, useful. 312 // Use breadth-first pass that records state in a Unique_Node_List, 313 // recursive traversal is slower. 314 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 315 int estimated_worklist_size = live_nodes(); 316 useful.map( estimated_worklist_size, NULL ); // preallocate space 317 318 // Initialize worklist 319 if (root() != NULL) { useful.push(root()); } 320 // If 'top' is cached, declare it useful to preserve cached node 321 if( cached_top_node() ) { useful.push(cached_top_node()); } 322 323 // Push all useful nodes onto the list, breadthfirst 324 for( uint next = 0; next < useful.size(); ++next ) { 325 assert( next < unique(), "Unique useful nodes < total nodes"); 326 Node *n = useful.at(next); 327 uint max = n->len(); 328 for( uint i = 0; i < max; ++i ) { 329 Node *m = n->in(i); 330 if (not_a_node(m)) continue; 331 useful.push(m); 332 } 333 } 334 } 335 336 // Update dead_node_list with any missing dead nodes using useful 337 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 338 void Compile::update_dead_node_list(Unique_Node_List &useful) { 339 uint max_idx = unique(); 340 VectorSet& useful_node_set = useful.member_set(); 341 342 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 343 // If node with index node_idx is not in useful set, 344 // mark it as dead in dead node list. 345 if (! useful_node_set.test(node_idx) ) { 346 record_dead_node(node_idx); 347 } 348 } 349 } 350 351 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 352 int shift = 0; 353 for (int i = 0; i < inlines->length(); i++) { 354 CallGenerator* cg = inlines->at(i); 355 CallNode* call = cg->call_node(); 356 if (shift > 0) { 357 inlines->at_put(i-shift, cg); 358 } 359 if (!useful.member(call)) { 360 shift++; 361 } 362 } 363 inlines->trunc_to(inlines->length()-shift); 364 } 365 366 // Disconnect all useless nodes by disconnecting those at the boundary. 367 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 368 uint next = 0; 369 while (next < useful.size()) { 370 Node *n = useful.at(next++); 371 if (n->is_SafePoint()) { 372 // We're done with a parsing phase. Replaced nodes are not valid 373 // beyond that point. 374 n->as_SafePoint()->delete_replaced_nodes(); 375 } 376 // Use raw traversal of out edges since this code removes out edges 377 int max = n->outcnt(); 378 for (int j = 0; j < max; ++j) { 379 Node* child = n->raw_out(j); 380 if (! useful.member(child)) { 381 assert(!child->is_top() || child != top(), 382 "If top is cached in Compile object it is in useful list"); 383 // Only need to remove this out-edge to the useless node 384 n->raw_del_out(j); 385 --j; 386 --max; 387 } 388 } 389 if (n->outcnt() == 1 && n->has_special_unique_user()) { 390 record_for_igvn(n->unique_out()); 391 } 392 } 393 // Remove useless macro and predicate opaq nodes 394 for (int i = C->macro_count()-1; i >= 0; i--) { 395 Node* n = C->macro_node(i); 396 if (!useful.member(n)) { 397 remove_macro_node(n); 398 } 399 } 400 // Remove useless CastII nodes with range check dependency 401 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 402 Node* cast = range_check_cast_node(i); 403 if (!useful.member(cast)) { 404 remove_range_check_cast(cast); 405 } 406 } 407 // Remove useless expensive nodes 408 for (int i = C->expensive_count()-1; i >= 0; i--) { 409 Node* n = C->expensive_node(i); 410 if (!useful.member(n)) { 411 remove_expensive_node(n); 412 } 413 } 414 // Remove useless Opaque4 nodes 415 for (int i = opaque4_count() - 1; i >= 0; i--) { 416 Node* opaq = opaque4_node(i); 417 if (!useful.member(opaq)) { 418 remove_opaque4_node(opaq); 419 } 420 } 421 // Remove useless value type nodes 422 if (_value_type_nodes != NULL) { 423 _value_type_nodes->remove_useless_nodes(useful.member_set()); 424 } 425 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 426 bs->eliminate_useless_gc_barriers(useful, this); 427 // clean up the late inline lists 428 remove_useless_late_inlines(&_string_late_inlines, useful); 429 remove_useless_late_inlines(&_boxing_late_inlines, useful); 430 remove_useless_late_inlines(&_late_inlines, useful); 431 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 432 } 433 434 //------------------------------frame_size_in_words----------------------------- 435 // frame_slots in units of words 436 int Compile::frame_size_in_words() const { 437 // shift is 0 in LP32 and 1 in LP64 438 const int shift = (LogBytesPerWord - LogBytesPerInt); 439 int words = _frame_slots >> shift; 440 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 441 return words; 442 } 443 444 // To bang the stack of this compiled method we use the stack size 445 // that the interpreter would need in case of a deoptimization. This 446 // removes the need to bang the stack in the deoptimization blob which 447 // in turn simplifies stack overflow handling. 448 int Compile::bang_size_in_bytes() const { 449 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 450 } 451 452 // ============================================================================ 453 //------------------------------CompileWrapper--------------------------------- 454 class CompileWrapper : public StackObj { 455 Compile *const _compile; 456 public: 457 CompileWrapper(Compile* compile); 458 459 ~CompileWrapper(); 460 }; 461 462 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 463 // the Compile* pointer is stored in the current ciEnv: 464 ciEnv* env = compile->env(); 465 assert(env == ciEnv::current(), "must already be a ciEnv active"); 466 assert(env->compiler_data() == NULL, "compile already active?"); 467 env->set_compiler_data(compile); 468 assert(compile == Compile::current(), "sanity"); 469 470 compile->set_type_dict(NULL); 471 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 472 compile->clone_map().set_clone_idx(0); 473 compile->set_type_hwm(NULL); 474 compile->set_type_last_size(0); 475 compile->set_last_tf(NULL, NULL); 476 compile->set_indexSet_arena(NULL); 477 compile->set_indexSet_free_block_list(NULL); 478 compile->init_type_arena(); 479 Type::Initialize(compile); 480 _compile->set_scratch_buffer_blob(NULL); 481 _compile->begin_method(); 482 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 483 } 484 CompileWrapper::~CompileWrapper() { 485 _compile->end_method(); 486 if (_compile->scratch_buffer_blob() != NULL) 487 BufferBlob::free(_compile->scratch_buffer_blob()); 488 _compile->env()->set_compiler_data(NULL); 489 } 490 491 492 //----------------------------print_compile_messages--------------------------- 493 void Compile::print_compile_messages() { 494 #ifndef PRODUCT 495 // Check if recompiling 496 if (_subsume_loads == false && PrintOpto) { 497 // Recompiling without allowing machine instructions to subsume loads 498 tty->print_cr("*********************************************************"); 499 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 500 tty->print_cr("*********************************************************"); 501 } 502 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 503 // Recompiling without escape analysis 504 tty->print_cr("*********************************************************"); 505 tty->print_cr("** Bailout: Recompile without escape analysis **"); 506 tty->print_cr("*********************************************************"); 507 } 508 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 509 // Recompiling without boxing elimination 510 tty->print_cr("*********************************************************"); 511 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 512 tty->print_cr("*********************************************************"); 513 } 514 if (C->directive()->BreakAtCompileOption) { 515 // Open the debugger when compiling this method. 516 tty->print("### Breaking when compiling: "); 517 method()->print_short_name(); 518 tty->cr(); 519 BREAKPOINT; 520 } 521 522 if( PrintOpto ) { 523 if (is_osr_compilation()) { 524 tty->print("[OSR]%3d", _compile_id); 525 } else { 526 tty->print("%3d", _compile_id); 527 } 528 } 529 #endif 530 } 531 532 533 //-----------------------init_scratch_buffer_blob------------------------------ 534 // Construct a temporary BufferBlob and cache it for this compile. 535 void Compile::init_scratch_buffer_blob(int const_size) { 536 // If there is already a scratch buffer blob allocated and the 537 // constant section is big enough, use it. Otherwise free the 538 // current and allocate a new one. 539 BufferBlob* blob = scratch_buffer_blob(); 540 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 541 // Use the current blob. 542 } else { 543 if (blob != NULL) { 544 BufferBlob::free(blob); 545 } 546 547 ResourceMark rm; 548 _scratch_const_size = const_size; 549 int size = C2Compiler::initial_code_buffer_size(const_size); 550 blob = BufferBlob::create("Compile::scratch_buffer", size); 551 // Record the buffer blob for next time. 552 set_scratch_buffer_blob(blob); 553 // Have we run out of code space? 554 if (scratch_buffer_blob() == NULL) { 555 // Let CompilerBroker disable further compilations. 556 record_failure("Not enough space for scratch buffer in CodeCache"); 557 return; 558 } 559 } 560 561 // Initialize the relocation buffers 562 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 563 set_scratch_locs_memory(locs_buf); 564 } 565 566 567 //-----------------------scratch_emit_size------------------------------------- 568 // Helper function that computes size by emitting code 569 uint Compile::scratch_emit_size(const Node* n) { 570 // Start scratch_emit_size section. 571 set_in_scratch_emit_size(true); 572 573 // Emit into a trash buffer and count bytes emitted. 574 // This is a pretty expensive way to compute a size, 575 // but it works well enough if seldom used. 576 // All common fixed-size instructions are given a size 577 // method by the AD file. 578 // Note that the scratch buffer blob and locs memory are 579 // allocated at the beginning of the compile task, and 580 // may be shared by several calls to scratch_emit_size. 581 // The allocation of the scratch buffer blob is particularly 582 // expensive, since it has to grab the code cache lock. 583 BufferBlob* blob = this->scratch_buffer_blob(); 584 assert(blob != NULL, "Initialize BufferBlob at start"); 585 assert(blob->size() > MAX_inst_size, "sanity"); 586 relocInfo* locs_buf = scratch_locs_memory(); 587 address blob_begin = blob->content_begin(); 588 address blob_end = (address)locs_buf; 589 assert(blob->contains(blob_end), "sanity"); 590 CodeBuffer buf(blob_begin, blob_end - blob_begin); 591 buf.initialize_consts_size(_scratch_const_size); 592 buf.initialize_stubs_size(MAX_stubs_size); 593 assert(locs_buf != NULL, "sanity"); 594 int lsize = MAX_locs_size / 3; 595 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 596 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 597 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 598 // Mark as scratch buffer. 599 buf.consts()->set_scratch_emit(); 600 buf.insts()->set_scratch_emit(); 601 buf.stubs()->set_scratch_emit(); 602 603 // Do the emission. 604 605 Label fakeL; // Fake label for branch instructions. 606 Label* saveL = NULL; 607 uint save_bnum = 0; 608 bool is_branch = n->is_MachBranch(); 609 if (is_branch) { 610 MacroAssembler masm(&buf); 611 masm.bind(fakeL); 612 n->as_MachBranch()->save_label(&saveL, &save_bnum); 613 n->as_MachBranch()->label_set(&fakeL, 0); 614 } 615 n->emit(buf, this->regalloc()); 616 617 // Emitting into the scratch buffer should not fail 618 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 619 620 // Restore label. 621 if (is_branch) { 622 n->as_MachBranch()->label_set(saveL, save_bnum); 623 } 624 625 // End scratch_emit_size section. 626 set_in_scratch_emit_size(false); 627 628 return buf.insts_size(); 629 } 630 631 632 // ============================================================================ 633 //------------------------------Compile standard------------------------------- 634 debug_only( int Compile::_debug_idx = 100000; ) 635 636 // Compile a method. entry_bci is -1 for normal compilations and indicates 637 // the continuation bci for on stack replacement. 638 639 640 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 641 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 642 : Phase(Compiler), 643 _compile_id(ci_env->compile_id()), 644 _save_argument_registers(false), 645 _subsume_loads(subsume_loads), 646 _do_escape_analysis(do_escape_analysis), 647 _eliminate_boxing(eliminate_boxing), 648 _method(target), 649 _entry_bci(osr_bci), 650 _stub_function(NULL), 651 _stub_name(NULL), 652 _stub_entry_point(NULL), 653 _max_node_limit(MaxNodeLimit), 654 _orig_pc_slot(0), 655 _orig_pc_slot_offset_in_bytes(0), 656 _inlining_progress(false), 657 _inlining_incrementally(false), 658 _has_reserved_stack_access(target->has_reserved_stack_access()), 659 #ifndef PRODUCT 660 _trace_opto_output(directive->TraceOptoOutputOption), 661 #endif 662 _has_method_handle_invokes(false), 663 _comp_arena(mtCompiler), 664 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 665 _env(ci_env), 666 _directive(directive), 667 _log(ci_env->log()), 668 _failure_reason(NULL), 669 _congraph(NULL), 670 #ifndef PRODUCT 671 _printer(IdealGraphPrinter::printer()), 672 #endif 673 _dead_node_list(comp_arena()), 674 _dead_node_count(0), 675 _node_arena(mtCompiler), 676 _old_arena(mtCompiler), 677 _mach_constant_base_node(NULL), 678 _Compile_types(mtCompiler), 679 _initial_gvn(NULL), 680 _for_igvn(NULL), 681 _warm_calls(NULL), 682 _late_inlines(comp_arena(), 2, 0, NULL), 683 _string_late_inlines(comp_arena(), 2, 0, NULL), 684 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 685 _late_inlines_pos(0), 686 _number_of_mh_late_inlines(0), 687 _print_inlining_stream(NULL), 688 _print_inlining_list(NULL), 689 _print_inlining_idx(0), 690 _print_inlining_output(NULL), 691 _replay_inline_data(NULL), 692 _java_calls(0), 693 _inner_loops(0), 694 _interpreter_frame_size(0), 695 _node_bundling_limit(0), 696 _node_bundling_base(NULL), 697 _code_buffer("Compile::Fill_buffer"), 698 _scratch_const_size(-1), 699 _in_scratch_emit_size(false) 700 #ifndef PRODUCT 701 , _in_dump_cnt(0) 702 #endif 703 { 704 C = this; 705 #ifndef PRODUCT 706 if (_printer != NULL) { 707 _printer->set_compile(this); 708 } 709 #endif 710 CompileWrapper cw(this); 711 712 if (CITimeVerbose) { 713 tty->print(" "); 714 target->holder()->name()->print(); 715 tty->print("."); 716 target->print_short_name(); 717 tty->print(" "); 718 } 719 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 720 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 721 722 #ifndef PRODUCT 723 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 724 if (!print_opto_assembly) { 725 bool print_assembly = directive->PrintAssemblyOption; 726 if (print_assembly && !Disassembler::can_decode()) { 727 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 728 print_opto_assembly = true; 729 } 730 } 731 set_print_assembly(print_opto_assembly); 732 set_parsed_irreducible_loop(false); 733 734 if (directive->ReplayInlineOption) { 735 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 736 } 737 #endif 738 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 739 set_print_intrinsics(directive->PrintIntrinsicsOption); 740 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 741 742 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 743 // Make sure the method being compiled gets its own MDO, 744 // so we can at least track the decompile_count(). 745 // Need MDO to record RTM code generation state. 746 method()->ensure_method_data(); 747 } 748 749 Init(::AliasLevel); 750 751 752 print_compile_messages(); 753 754 _ilt = InlineTree::build_inline_tree_root(); 755 756 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 757 assert(num_alias_types() >= AliasIdxRaw, ""); 758 759 #define MINIMUM_NODE_HASH 1023 760 // Node list that Iterative GVN will start with 761 Unique_Node_List for_igvn(comp_arena()); 762 set_for_igvn(&for_igvn); 763 764 // GVN that will be run immediately on new nodes 765 uint estimated_size = method()->code_size()*4+64; 766 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 767 PhaseGVN gvn(node_arena(), estimated_size); 768 set_initial_gvn(&gvn); 769 770 print_inlining_init(); 771 { // Scope for timing the parser 772 TracePhase tp("parse", &timers[_t_parser]); 773 774 // Put top into the hash table ASAP. 775 initial_gvn()->transform_no_reclaim(top()); 776 777 // Set up tf(), start(), and find a CallGenerator. 778 CallGenerator* cg = NULL; 779 if (is_osr_compilation()) { 780 const TypeTuple *domain = StartOSRNode::osr_domain(); 781 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 782 init_tf(TypeFunc::make(domain, range)); 783 StartNode* s = new StartOSRNode(root(), domain); 784 initial_gvn()->set_type_bottom(s); 785 init_start(s); 786 cg = CallGenerator::for_osr(method(), entry_bci()); 787 } else { 788 // Normal case. 789 init_tf(TypeFunc::make(method())); 790 StartNode* s = new StartNode(root(), tf()->domain_cc()); 791 initial_gvn()->set_type_bottom(s); 792 init_start(s); 793 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 794 // With java.lang.ref.reference.get() we must go through the 795 // intrinsic - even when get() is the root 796 // method of the compile - so that, if necessary, the value in 797 // the referent field of the reference object gets recorded by 798 // the pre-barrier code. 799 cg = find_intrinsic(method(), false); 800 } 801 if (cg == NULL) { 802 float past_uses = method()->interpreter_invocation_count(); 803 float expected_uses = past_uses; 804 cg = CallGenerator::for_inline(method(), expected_uses); 805 } 806 } 807 if (failing()) return; 808 if (cg == NULL) { 809 record_method_not_compilable("cannot parse method"); 810 return; 811 } 812 JVMState* jvms = build_start_state(start(), tf()); 813 if ((jvms = cg->generate(jvms)) == NULL) { 814 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 815 record_method_not_compilable("method parse failed"); 816 } 817 return; 818 } 819 GraphKit kit(jvms); 820 821 if (!kit.stopped()) { 822 // Accept return values, and transfer control we know not where. 823 // This is done by a special, unique ReturnNode bound to root. 824 return_values(kit.jvms()); 825 } 826 827 if (kit.has_exceptions()) { 828 // Any exceptions that escape from this call must be rethrown 829 // to whatever caller is dynamically above us on the stack. 830 // This is done by a special, unique RethrowNode bound to root. 831 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 832 } 833 834 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 835 836 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 837 inline_string_calls(true); 838 } 839 840 if (failing()) return; 841 842 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 843 844 // Remove clutter produced by parsing. 845 if (!failing()) { 846 ResourceMark rm; 847 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 848 } 849 } 850 851 // Note: Large methods are capped off in do_one_bytecode(). 852 if (failing()) return; 853 854 // After parsing, node notes are no longer automagic. 855 // They must be propagated by register_new_node_with_optimizer(), 856 // clone(), or the like. 857 set_default_node_notes(NULL); 858 859 for (;;) { 860 int successes = Inline_Warm(); 861 if (failing()) return; 862 if (successes == 0) break; 863 } 864 865 // Drain the list. 866 Finish_Warm(); 867 #ifndef PRODUCT 868 if (_printer && _printer->should_print(1)) { 869 _printer->print_inlining(); 870 } 871 #endif 872 873 if (failing()) return; 874 NOT_PRODUCT( verify_graph_edges(); ) 875 876 // Now optimize 877 Optimize(); 878 if (failing()) return; 879 NOT_PRODUCT( verify_graph_edges(); ) 880 881 #ifndef PRODUCT 882 if (PrintIdeal) { 883 ttyLocker ttyl; // keep the following output all in one block 884 // This output goes directly to the tty, not the compiler log. 885 // To enable tools to match it up with the compilation activity, 886 // be sure to tag this tty output with the compile ID. 887 if (xtty != NULL) { 888 xtty->head("ideal compile_id='%d'%s", compile_id(), 889 is_osr_compilation() ? " compile_kind='osr'" : 890 ""); 891 } 892 root()->dump(9999); 893 if (xtty != NULL) { 894 xtty->tail("ideal"); 895 } 896 } 897 #endif 898 899 #ifdef ASSERT 900 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 901 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 902 #endif 903 904 // Dump compilation data to replay it. 905 if (directive->DumpReplayOption) { 906 env()->dump_replay_data(_compile_id); 907 } 908 if (directive->DumpInlineOption && (ilt() != NULL)) { 909 env()->dump_inline_data(_compile_id); 910 } 911 912 // Now that we know the size of all the monitors we can add a fixed slot 913 // for the original deopt pc. 914 915 _orig_pc_slot = fixed_slots(); 916 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 917 set_fixed_slots(next_slot); 918 919 // Compute when to use implicit null checks. Used by matching trap based 920 // nodes and NullCheck optimization. 921 set_allowed_deopt_reasons(); 922 923 // Now generate code 924 Code_Gen(); 925 if (failing()) return; 926 927 // Check if we want to skip execution of all compiled code. 928 { 929 #ifndef PRODUCT 930 if (OptoNoExecute) { 931 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 932 return; 933 } 934 #endif 935 TracePhase tp("install_code", &timers[_t_registerMethod]); 936 937 if (is_osr_compilation()) { 938 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 939 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 940 } else { 941 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 942 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 943 } 944 945 env()->register_method(_method, _entry_bci, 946 &_code_offsets, 947 _orig_pc_slot_offset_in_bytes, 948 code_buffer(), 949 frame_size_in_words(), _oop_map_set, 950 &_handler_table, &_inc_table, 951 compiler, 952 has_unsafe_access(), 953 SharedRuntime::is_wide_vector(max_vector_size()), 954 rtm_state() 955 ); 956 957 if (log() != NULL) // Print code cache state into compiler log 958 log()->code_cache_state(); 959 } 960 } 961 962 //------------------------------Compile---------------------------------------- 963 // Compile a runtime stub 964 Compile::Compile( ciEnv* ci_env, 965 TypeFunc_generator generator, 966 address stub_function, 967 const char *stub_name, 968 int is_fancy_jump, 969 bool pass_tls, 970 bool save_arg_registers, 971 bool return_pc, 972 DirectiveSet* directive) 973 : Phase(Compiler), 974 _compile_id(0), 975 _save_argument_registers(save_arg_registers), 976 _subsume_loads(true), 977 _do_escape_analysis(false), 978 _eliminate_boxing(false), 979 _method(NULL), 980 _entry_bci(InvocationEntryBci), 981 _stub_function(stub_function), 982 _stub_name(stub_name), 983 _stub_entry_point(NULL), 984 _max_node_limit(MaxNodeLimit), 985 _orig_pc_slot(0), 986 _orig_pc_slot_offset_in_bytes(0), 987 _inlining_progress(false), 988 _inlining_incrementally(false), 989 _has_reserved_stack_access(false), 990 #ifndef PRODUCT 991 _trace_opto_output(directive->TraceOptoOutputOption), 992 #endif 993 _has_method_handle_invokes(false), 994 _comp_arena(mtCompiler), 995 _env(ci_env), 996 _directive(directive), 997 _log(ci_env->log()), 998 _failure_reason(NULL), 999 _congraph(NULL), 1000 #ifndef PRODUCT 1001 _printer(NULL), 1002 #endif 1003 _dead_node_list(comp_arena()), 1004 _dead_node_count(0), 1005 _node_arena(mtCompiler), 1006 _old_arena(mtCompiler), 1007 _mach_constant_base_node(NULL), 1008 _Compile_types(mtCompiler), 1009 _initial_gvn(NULL), 1010 _for_igvn(NULL), 1011 _warm_calls(NULL), 1012 _number_of_mh_late_inlines(0), 1013 _print_inlining_stream(NULL), 1014 _print_inlining_list(NULL), 1015 _print_inlining_idx(0), 1016 _print_inlining_output(NULL), 1017 _replay_inline_data(NULL), 1018 _java_calls(0), 1019 _inner_loops(0), 1020 _interpreter_frame_size(0), 1021 _node_bundling_limit(0), 1022 _node_bundling_base(NULL), 1023 _code_buffer("Compile::Fill_buffer"), 1024 #ifndef PRODUCT 1025 _in_dump_cnt(0), 1026 #endif 1027 _allowed_reasons(0) { 1028 C = this; 1029 1030 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1031 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1032 1033 #ifndef PRODUCT 1034 set_print_assembly(PrintFrameConverterAssembly); 1035 set_parsed_irreducible_loop(false); 1036 #endif 1037 set_has_irreducible_loop(false); // no loops 1038 1039 CompileWrapper cw(this); 1040 Init(/*AliasLevel=*/ 0); 1041 init_tf((*generator)()); 1042 1043 { 1044 // The following is a dummy for the sake of GraphKit::gen_stub 1045 Unique_Node_List for_igvn(comp_arena()); 1046 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1047 PhaseGVN gvn(Thread::current()->resource_area(),255); 1048 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1049 gvn.transform_no_reclaim(top()); 1050 1051 GraphKit kit; 1052 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1053 } 1054 1055 NOT_PRODUCT( verify_graph_edges(); ) 1056 Code_Gen(); 1057 if (failing()) return; 1058 1059 1060 // Entry point will be accessed using compile->stub_entry_point(); 1061 if (code_buffer() == NULL) { 1062 Matcher::soft_match_failure(); 1063 } else { 1064 if (PrintAssembly && (WizardMode || Verbose)) 1065 tty->print_cr("### Stub::%s", stub_name); 1066 1067 if (!failing()) { 1068 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1069 1070 // Make the NMethod 1071 // For now we mark the frame as never safe for profile stackwalking 1072 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1073 code_buffer(), 1074 CodeOffsets::frame_never_safe, 1075 // _code_offsets.value(CodeOffsets::Frame_Complete), 1076 frame_size_in_words(), 1077 _oop_map_set, 1078 save_arg_registers); 1079 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1080 1081 _stub_entry_point = rs->entry_point(); 1082 } 1083 } 1084 } 1085 1086 //------------------------------Init------------------------------------------- 1087 // Prepare for a single compilation 1088 void Compile::Init(int aliaslevel) { 1089 _unique = 0; 1090 _regalloc = NULL; 1091 1092 _tf = NULL; // filled in later 1093 _top = NULL; // cached later 1094 _matcher = NULL; // filled in later 1095 _cfg = NULL; // filled in later 1096 1097 set_24_bit_selection_and_mode(Use24BitFP, false); 1098 1099 _node_note_array = NULL; 1100 _default_node_notes = NULL; 1101 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1102 1103 _immutable_memory = NULL; // filled in at first inquiry 1104 1105 // Globally visible Nodes 1106 // First set TOP to NULL to give safe behavior during creation of RootNode 1107 set_cached_top_node(NULL); 1108 set_root(new RootNode()); 1109 // Now that you have a Root to point to, create the real TOP 1110 set_cached_top_node( new ConNode(Type::TOP) ); 1111 set_recent_alloc(NULL, NULL); 1112 1113 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1114 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1115 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1116 env()->set_dependencies(new Dependencies(env())); 1117 1118 _fixed_slots = 0; 1119 set_has_split_ifs(false); 1120 set_has_loops(has_method() && method()->has_loops()); // first approximation 1121 set_has_stringbuilder(false); 1122 set_has_boxed_value(false); 1123 _trap_can_recompile = false; // no traps emitted yet 1124 _major_progress = true; // start out assuming good things will happen 1125 set_has_unsafe_access(false); 1126 set_max_vector_size(0); 1127 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1128 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1129 set_decompile_count(0); 1130 1131 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1132 set_num_loop_opts(LoopOptsCount); 1133 set_do_inlining(Inline); 1134 set_max_inline_size(MaxInlineSize); 1135 set_freq_inline_size(FreqInlineSize); 1136 set_do_scheduling(OptoScheduling); 1137 set_do_count_invocations(false); 1138 set_do_method_data_update(false); 1139 1140 set_do_vector_loop(false); 1141 1142 if (AllowVectorizeOnDemand) { 1143 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1144 set_do_vector_loop(true); 1145 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());}) 1146 } else if (has_method() && method()->name() != 0 && 1147 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1148 set_do_vector_loop(true); 1149 } 1150 } 1151 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1152 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());}) 1153 1154 set_age_code(has_method() && method()->profile_aging()); 1155 set_rtm_state(NoRTM); // No RTM lock eliding by default 1156 _max_node_limit = _directive->MaxNodeLimitOption; 1157 1158 #if INCLUDE_RTM_OPT 1159 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1160 int rtm_state = method()->method_data()->rtm_state(); 1161 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1162 // Don't generate RTM lock eliding code. 1163 set_rtm_state(NoRTM); 1164 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1165 // Generate RTM lock eliding code without abort ratio calculation code. 1166 set_rtm_state(UseRTM); 1167 } else if (UseRTMDeopt) { 1168 // Generate RTM lock eliding code and include abort ratio calculation 1169 // code if UseRTMDeopt is on. 1170 set_rtm_state(ProfileRTM); 1171 } 1172 } 1173 #endif 1174 if (debug_info()->recording_non_safepoints()) { 1175 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1176 (comp_arena(), 8, 0, NULL)); 1177 set_default_node_notes(Node_Notes::make(this)); 1178 } 1179 1180 // // -- Initialize types before each compile -- 1181 // // Update cached type information 1182 // if( _method && _method->constants() ) 1183 // Type::update_loaded_types(_method, _method->constants()); 1184 1185 // Init alias_type map. 1186 if (!_do_escape_analysis && aliaslevel == 3) 1187 aliaslevel = 2; // No unique types without escape analysis 1188 _AliasLevel = aliaslevel; 1189 const int grow_ats = 16; 1190 _max_alias_types = grow_ats; 1191 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1192 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1193 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1194 { 1195 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1196 } 1197 // Initialize the first few types. 1198 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1199 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1200 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1201 _num_alias_types = AliasIdxRaw+1; 1202 // Zero out the alias type cache. 1203 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1204 // A NULL adr_type hits in the cache right away. Preload the right answer. 1205 probe_alias_cache(NULL)->_index = AliasIdxTop; 1206 1207 _intrinsics = NULL; 1208 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1209 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1210 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1211 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1212 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1213 _value_type_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 1214 register_library_intrinsics(); 1215 } 1216 1217 //---------------------------init_start---------------------------------------- 1218 // Install the StartNode on this compile object. 1219 void Compile::init_start(StartNode* s) { 1220 if (failing()) 1221 return; // already failing 1222 assert(s == start(), ""); 1223 } 1224 1225 /** 1226 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1227 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1228 * the ideal graph. 1229 */ 1230 StartNode* Compile::start() const { 1231 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1232 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1233 Node* start = root()->fast_out(i); 1234 if (start->is_Start()) { 1235 return start->as_Start(); 1236 } 1237 } 1238 fatal("Did not find Start node!"); 1239 return NULL; 1240 } 1241 1242 //-------------------------------immutable_memory------------------------------------- 1243 // Access immutable memory 1244 Node* Compile::immutable_memory() { 1245 if (_immutable_memory != NULL) { 1246 return _immutable_memory; 1247 } 1248 StartNode* s = start(); 1249 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1250 Node *p = s->fast_out(i); 1251 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1252 _immutable_memory = p; 1253 return _immutable_memory; 1254 } 1255 } 1256 ShouldNotReachHere(); 1257 return NULL; 1258 } 1259 1260 //----------------------set_cached_top_node------------------------------------ 1261 // Install the cached top node, and make sure Node::is_top works correctly. 1262 void Compile::set_cached_top_node(Node* tn) { 1263 if (tn != NULL) verify_top(tn); 1264 Node* old_top = _top; 1265 _top = tn; 1266 // Calling Node::setup_is_top allows the nodes the chance to adjust 1267 // their _out arrays. 1268 if (_top != NULL) _top->setup_is_top(); 1269 if (old_top != NULL) old_top->setup_is_top(); 1270 assert(_top == NULL || top()->is_top(), ""); 1271 } 1272 1273 #ifdef ASSERT 1274 uint Compile::count_live_nodes_by_graph_walk() { 1275 Unique_Node_List useful(comp_arena()); 1276 // Get useful node list by walking the graph. 1277 identify_useful_nodes(useful); 1278 return useful.size(); 1279 } 1280 1281 void Compile::print_missing_nodes() { 1282 1283 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1284 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1285 return; 1286 } 1287 1288 // This is an expensive function. It is executed only when the user 1289 // specifies VerifyIdealNodeCount option or otherwise knows the 1290 // additional work that needs to be done to identify reachable nodes 1291 // by walking the flow graph and find the missing ones using 1292 // _dead_node_list. 1293 1294 Unique_Node_List useful(comp_arena()); 1295 // Get useful node list by walking the graph. 1296 identify_useful_nodes(useful); 1297 1298 uint l_nodes = C->live_nodes(); 1299 uint l_nodes_by_walk = useful.size(); 1300 1301 if (l_nodes != l_nodes_by_walk) { 1302 if (_log != NULL) { 1303 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1304 _log->stamp(); 1305 _log->end_head(); 1306 } 1307 VectorSet& useful_member_set = useful.member_set(); 1308 int last_idx = l_nodes_by_walk; 1309 for (int i = 0; i < last_idx; i++) { 1310 if (useful_member_set.test(i)) { 1311 if (_dead_node_list.test(i)) { 1312 if (_log != NULL) { 1313 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1314 } 1315 if (PrintIdealNodeCount) { 1316 // Print the log message to tty 1317 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1318 useful.at(i)->dump(); 1319 } 1320 } 1321 } 1322 else if (! _dead_node_list.test(i)) { 1323 if (_log != NULL) { 1324 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1325 } 1326 if (PrintIdealNodeCount) { 1327 // Print the log message to tty 1328 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1329 } 1330 } 1331 } 1332 if (_log != NULL) { 1333 _log->tail("mismatched_nodes"); 1334 } 1335 } 1336 } 1337 void Compile::record_modified_node(Node* n) { 1338 if (_modified_nodes != NULL && !_inlining_incrementally && 1339 n->outcnt() != 0 && !n->is_Con()) { 1340 _modified_nodes->push(n); 1341 } 1342 } 1343 1344 void Compile::remove_modified_node(Node* n) { 1345 if (_modified_nodes != NULL) { 1346 _modified_nodes->remove(n); 1347 } 1348 } 1349 #endif 1350 1351 #ifndef PRODUCT 1352 void Compile::verify_top(Node* tn) const { 1353 if (tn != NULL) { 1354 assert(tn->is_Con(), "top node must be a constant"); 1355 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1356 assert(tn->in(0) != NULL, "must have live top node"); 1357 } 1358 } 1359 #endif 1360 1361 1362 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1363 1364 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1365 guarantee(arr != NULL, ""); 1366 int num_blocks = arr->length(); 1367 if (grow_by < num_blocks) grow_by = num_blocks; 1368 int num_notes = grow_by * _node_notes_block_size; 1369 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1370 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1371 while (num_notes > 0) { 1372 arr->append(notes); 1373 notes += _node_notes_block_size; 1374 num_notes -= _node_notes_block_size; 1375 } 1376 assert(num_notes == 0, "exact multiple, please"); 1377 } 1378 1379 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1380 if (source == NULL || dest == NULL) return false; 1381 1382 if (dest->is_Con()) 1383 return false; // Do not push debug info onto constants. 1384 1385 #ifdef ASSERT 1386 // Leave a bread crumb trail pointing to the original node: 1387 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1388 dest->set_debug_orig(source); 1389 } 1390 #endif 1391 1392 if (node_note_array() == NULL) 1393 return false; // Not collecting any notes now. 1394 1395 // This is a copy onto a pre-existing node, which may already have notes. 1396 // If both nodes have notes, do not overwrite any pre-existing notes. 1397 Node_Notes* source_notes = node_notes_at(source->_idx); 1398 if (source_notes == NULL || source_notes->is_clear()) return false; 1399 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1400 if (dest_notes == NULL || dest_notes->is_clear()) { 1401 return set_node_notes_at(dest->_idx, source_notes); 1402 } 1403 1404 Node_Notes merged_notes = (*source_notes); 1405 // The order of operations here ensures that dest notes will win... 1406 merged_notes.update_from(dest_notes); 1407 return set_node_notes_at(dest->_idx, &merged_notes); 1408 } 1409 1410 1411 //--------------------------allow_range_check_smearing------------------------- 1412 // Gating condition for coalescing similar range checks. 1413 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1414 // single covering check that is at least as strong as any of them. 1415 // If the optimization succeeds, the simplified (strengthened) range check 1416 // will always succeed. If it fails, we will deopt, and then give up 1417 // on the optimization. 1418 bool Compile::allow_range_check_smearing() const { 1419 // If this method has already thrown a range-check, 1420 // assume it was because we already tried range smearing 1421 // and it failed. 1422 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1423 return !already_trapped; 1424 } 1425 1426 1427 //------------------------------flatten_alias_type----------------------------- 1428 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1429 int offset = tj->offset(); 1430 TypePtr::PTR ptr = tj->ptr(); 1431 1432 // Known instance (scalarizable allocation) alias only with itself. 1433 bool is_known_inst = tj->isa_oopptr() != NULL && 1434 tj->is_oopptr()->is_known_instance(); 1435 1436 // Process weird unsafe references. 1437 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1438 bool default_value_load = EnableValhalla && tj->is_instptr()->klass() == ciEnv::current()->Class_klass(); 1439 assert(InlineUnsafeOps || default_value_load, "indeterminate pointers come only from unsafe ops"); 1440 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1441 tj = TypeOopPtr::BOTTOM; 1442 ptr = tj->ptr(); 1443 offset = tj->offset(); 1444 } 1445 1446 // Array pointers need some flattening 1447 const TypeAryPtr *ta = tj->isa_aryptr(); 1448 if (ta && ta->is_stable()) { 1449 // Erase stability property for alias analysis. 1450 tj = ta = ta->cast_to_stable(false); 1451 } 1452 if( ta && is_known_inst ) { 1453 if ( offset != Type::OffsetBot && 1454 offset > arrayOopDesc::length_offset_in_bytes() ) { 1455 offset = Type::OffsetBot; // Flatten constant access into array body only 1456 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id()); 1457 } 1458 } else if( ta && _AliasLevel >= 2 ) { 1459 // For arrays indexed by constant indices, we flatten the alias 1460 // space to include all of the array body. Only the header, klass 1461 // and array length can be accessed un-aliased. 1462 // For flattened value type array, each field has its own slice so 1463 // we must include the field offset. 1464 if( offset != Type::OffsetBot ) { 1465 if( ta->const_oop() ) { // MethodData* or Method* 1466 offset = Type::OffsetBot; // Flatten constant access into array body 1467 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1468 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1469 // range is OK as-is. 1470 tj = ta = TypeAryPtr::RANGE; 1471 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1472 tj = TypeInstPtr::KLASS; // all klass loads look alike 1473 ta = TypeAryPtr::RANGE; // generic ignored junk 1474 ptr = TypePtr::BotPTR; 1475 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1476 tj = TypeInstPtr::MARK; 1477 ta = TypeAryPtr::RANGE; // generic ignored junk 1478 ptr = TypePtr::BotPTR; 1479 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { 1480 ta = tj->isa_aryptr(); 1481 } else { // Random constant offset into array body 1482 offset = Type::OffsetBot; // Flatten constant access into array body 1483 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1484 } 1485 } 1486 // Arrays of fixed size alias with arrays of unknown size. 1487 if (ta->size() != TypeInt::POS) { 1488 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1489 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1490 } 1491 // Arrays of known objects become arrays of unknown objects. 1492 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1493 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1494 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1495 } 1496 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1497 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1498 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1499 } 1500 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1501 // cannot be distinguished by bytecode alone. 1502 if (ta->elem() == TypeInt::BOOL) { 1503 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1504 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1505 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset()); 1506 } 1507 // During the 2nd round of IterGVN, NotNull castings are removed. 1508 // Make sure the Bottom and NotNull variants alias the same. 1509 // Also, make sure exact and non-exact variants alias the same. 1510 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1511 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1512 } 1513 } 1514 1515 // Oop pointers need some flattening 1516 const TypeInstPtr *to = tj->isa_instptr(); 1517 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1518 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1519 if( ptr == TypePtr::Constant ) { 1520 if (to->klass() != ciEnv::current()->Class_klass() || 1521 offset < k->size_helper() * wordSize) { 1522 // No constant oop pointers (such as Strings); they alias with 1523 // unknown strings. 1524 assert(!is_known_inst, "not scalarizable allocation"); 1525 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1526 } 1527 } else if( is_known_inst ) { 1528 tj = to; // Keep NotNull and klass_is_exact for instance type 1529 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1530 // During the 2nd round of IterGVN, NotNull castings are removed. 1531 // Make sure the Bottom and NotNull variants alias the same. 1532 // Also, make sure exact and non-exact variants alias the same. 1533 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1534 } 1535 if (to->speculative() != NULL) { 1536 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->instance_id()); 1537 } 1538 // Canonicalize the holder of this field 1539 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1540 // First handle header references such as a LoadKlassNode, even if the 1541 // object's klass is unloaded at compile time (4965979). 1542 if (!is_known_inst) { // Do it only for non-instance types 1543 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset)); 1544 } 1545 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { 1546 to = tj->is_instptr(); 1547 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1548 // Static fields are in the space above the normal instance 1549 // fields in the java.lang.Class instance. 1550 if (to->klass() != ciEnv::current()->Class_klass()) { 1551 to = NULL; 1552 tj = TypeOopPtr::BOTTOM; 1553 offset = tj->offset(); 1554 } 1555 } else { 1556 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1557 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1558 if( is_known_inst ) { 1559 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), to->instance_id()); 1560 } else { 1561 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset)); 1562 } 1563 } 1564 } 1565 } 1566 1567 // Klass pointers to object array klasses need some flattening 1568 const TypeKlassPtr *tk = tj->isa_klassptr(); 1569 if( tk ) { 1570 // If we are referencing a field within a Klass, we need 1571 // to assume the worst case of an Object. Both exact and 1572 // inexact types must flatten to the same alias class so 1573 // use NotNull as the PTR. 1574 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1575 1576 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1577 TypeKlassPtr::OBJECT->klass(), 1578 Type::Offset(offset)); 1579 } 1580 1581 ciKlass* klass = tk->klass(); 1582 if (klass != NULL && klass->is_obj_array_klass()) { 1583 ciKlass* k = TypeAryPtr::OOPS->klass(); 1584 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1585 k = TypeInstPtr::BOTTOM->klass(); 1586 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset)); 1587 } 1588 1589 // Check for precise loads from the primary supertype array and force them 1590 // to the supertype cache alias index. Check for generic array loads from 1591 // the primary supertype array and also force them to the supertype cache 1592 // alias index. Since the same load can reach both, we need to merge 1593 // these 2 disparate memories into the same alias class. Since the 1594 // primary supertype array is read-only, there's no chance of confusion 1595 // where we bypass an array load and an array store. 1596 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1597 if (offset == Type::OffsetBot || 1598 (offset >= primary_supers_offset && 1599 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1600 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1601 offset = in_bytes(Klass::secondary_super_cache_offset()); 1602 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset)); 1603 } 1604 } 1605 1606 // Flatten all Raw pointers together. 1607 if (tj->base() == Type::RawPtr) 1608 tj = TypeRawPtr::BOTTOM; 1609 1610 if (tj->base() == Type::AnyPtr) 1611 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1612 1613 // Flatten all to bottom for now 1614 switch( _AliasLevel ) { 1615 case 0: 1616 tj = TypePtr::BOTTOM; 1617 break; 1618 case 1: // Flatten to: oop, static, field or array 1619 switch (tj->base()) { 1620 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1621 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1622 case Type::AryPtr: // do not distinguish arrays at all 1623 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1624 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1625 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1626 default: ShouldNotReachHere(); 1627 } 1628 break; 1629 case 2: // No collapsing at level 2; keep all splits 1630 case 3: // No collapsing at level 3; keep all splits 1631 break; 1632 default: 1633 Unimplemented(); 1634 } 1635 1636 offset = tj->offset(); 1637 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1638 1639 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1640 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1641 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1642 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1643 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1644 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1645 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) || 1646 (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)), 1647 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1648 assert( tj->ptr() != TypePtr::TopPTR && 1649 tj->ptr() != TypePtr::AnyNull && 1650 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1651 // assert( tj->ptr() != TypePtr::Constant || 1652 // tj->base() == Type::RawPtr || 1653 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1654 1655 return tj; 1656 } 1657 1658 void Compile::AliasType::Init(int i, const TypePtr* at) { 1659 _index = i; 1660 _adr_type = at; 1661 _field = NULL; 1662 _element = NULL; 1663 _is_rewritable = true; // default 1664 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1665 if (atoop != NULL && atoop->is_known_instance()) { 1666 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1667 _general_index = Compile::current()->get_alias_index(gt); 1668 } else { 1669 _general_index = 0; 1670 } 1671 } 1672 1673 BasicType Compile::AliasType::basic_type() const { 1674 if (element() != NULL) { 1675 const Type* element = adr_type()->is_aryptr()->elem(); 1676 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1677 } if (field() != NULL) { 1678 return field()->layout_type(); 1679 } else { 1680 return T_ILLEGAL; // unknown 1681 } 1682 } 1683 1684 //---------------------------------print_on------------------------------------ 1685 #ifndef PRODUCT 1686 void Compile::AliasType::print_on(outputStream* st) { 1687 if (index() < 10) 1688 st->print("@ <%d> ", index()); 1689 else st->print("@ <%d>", index()); 1690 st->print(is_rewritable() ? " " : " RO"); 1691 int offset = adr_type()->offset(); 1692 if (offset == Type::OffsetBot) 1693 st->print(" +any"); 1694 else st->print(" +%-3d", offset); 1695 st->print(" in "); 1696 adr_type()->dump_on(st); 1697 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1698 if (field() != NULL && tjp) { 1699 if (tjp->klass() != field()->holder() || 1700 tjp->offset() != field()->offset_in_bytes()) { 1701 st->print(" != "); 1702 field()->print(); 1703 st->print(" ***"); 1704 } 1705 } 1706 } 1707 1708 void print_alias_types() { 1709 Compile* C = Compile::current(); 1710 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1711 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1712 C->alias_type(idx)->print_on(tty); 1713 tty->cr(); 1714 } 1715 } 1716 #endif 1717 1718 1719 //----------------------------probe_alias_cache-------------------------------- 1720 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1721 intptr_t key = (intptr_t) adr_type; 1722 key ^= key >> logAliasCacheSize; 1723 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1724 } 1725 1726 1727 //-----------------------------grow_alias_types-------------------------------- 1728 void Compile::grow_alias_types() { 1729 const int old_ats = _max_alias_types; // how many before? 1730 const int new_ats = old_ats; // how many more? 1731 const int grow_ats = old_ats+new_ats; // how many now? 1732 _max_alias_types = grow_ats; 1733 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1734 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1735 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1736 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1737 } 1738 1739 1740 //--------------------------------find_alias_type------------------------------ 1741 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1742 if (_AliasLevel == 0) 1743 return alias_type(AliasIdxBot); 1744 1745 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1746 if (ace->_adr_type == adr_type) { 1747 return alias_type(ace->_index); 1748 } 1749 1750 // Handle special cases. 1751 if (adr_type == NULL) return alias_type(AliasIdxTop); 1752 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1753 1754 // Do it the slow way. 1755 const TypePtr* flat = flatten_alias_type(adr_type); 1756 1757 #ifdef ASSERT 1758 { 1759 ResourceMark rm; 1760 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1761 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1762 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1763 Type::str(adr_type)); 1764 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1765 const TypeOopPtr* foop = flat->is_oopptr(); 1766 // Scalarizable allocations have exact klass always. 1767 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1768 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1769 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1770 Type::str(foop), Type::str(xoop)); 1771 } 1772 } 1773 #endif 1774 1775 int idx = AliasIdxTop; 1776 for (int i = 0; i < num_alias_types(); i++) { 1777 if (alias_type(i)->adr_type() == flat) { 1778 idx = i; 1779 break; 1780 } 1781 } 1782 1783 if (idx == AliasIdxTop) { 1784 if (no_create) return NULL; 1785 // Grow the array if necessary. 1786 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1787 // Add a new alias type. 1788 idx = _num_alias_types++; 1789 _alias_types[idx]->Init(idx, flat); 1790 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1791 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1792 if (flat->isa_instptr()) { 1793 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1794 && flat->is_instptr()->klass() == env()->Class_klass()) 1795 alias_type(idx)->set_rewritable(false); 1796 } 1797 ciField* field = NULL; 1798 if (flat->isa_aryptr()) { 1799 #ifdef ASSERT 1800 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1801 // (T_BYTE has the weakest alignment and size restrictions...) 1802 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1803 #endif 1804 const Type* elemtype = flat->is_aryptr()->elem(); 1805 if (flat->offset() == TypePtr::OffsetBot) { 1806 alias_type(idx)->set_element(elemtype); 1807 } 1808 int field_offset = flat->is_aryptr()->field_offset().get(); 1809 if (elemtype->isa_valuetype() && field_offset != Type::OffsetBot) { 1810 ciValueKlass* vk = elemtype->is_valuetype()->value_klass(); 1811 field_offset += vk->first_field_offset(); 1812 field = vk->get_field_by_offset(field_offset, false); 1813 } 1814 } 1815 if (flat->isa_klassptr()) { 1816 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1817 alias_type(idx)->set_rewritable(false); 1818 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1819 alias_type(idx)->set_rewritable(false); 1820 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1821 alias_type(idx)->set_rewritable(false); 1822 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1823 alias_type(idx)->set_rewritable(false); 1824 } 1825 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1826 // but the base pointer type is not distinctive enough to identify 1827 // references into JavaThread.) 1828 1829 // Check for final fields. 1830 const TypeInstPtr* tinst = flat->isa_instptr(); 1831 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1832 if (tinst->const_oop() != NULL && 1833 tinst->klass() == ciEnv::current()->Class_klass() && 1834 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1835 // static field 1836 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1837 field = k->get_field_by_offset(tinst->offset(), true); 1838 } else if (tinst->klass()->is_valuetype()) { 1839 // Value type field 1840 ciValueKlass* vk = tinst->value_klass(); 1841 field = vk->get_field_by_offset(tinst->offset(), false); 1842 } else { 1843 ciInstanceKlass* k = tinst->klass()->as_instance_klass(); 1844 field = k->get_field_by_offset(tinst->offset(), false); 1845 } 1846 } 1847 assert(field == NULL || 1848 original_field == NULL || 1849 (field->holder() == original_field->holder() && 1850 field->offset() == original_field->offset() && 1851 field->is_static() == original_field->is_static()), "wrong field?"); 1852 // Set field() and is_rewritable() attributes. 1853 if (field != NULL) alias_type(idx)->set_field(field); 1854 } 1855 1856 // Fill the cache for next time. 1857 ace->_adr_type = adr_type; 1858 ace->_index = idx; 1859 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1860 1861 // Might as well try to fill the cache for the flattened version, too. 1862 AliasCacheEntry* face = probe_alias_cache(flat); 1863 if (face->_adr_type == NULL) { 1864 face->_adr_type = flat; 1865 face->_index = idx; 1866 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1867 } 1868 1869 return alias_type(idx); 1870 } 1871 1872 1873 Compile::AliasType* Compile::alias_type(ciField* field) { 1874 const TypeOopPtr* t; 1875 if (field->is_static()) 1876 t = TypeInstPtr::make(field->holder()->java_mirror()); 1877 else 1878 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1879 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1880 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1881 return atp; 1882 } 1883 1884 1885 //------------------------------have_alias_type-------------------------------- 1886 bool Compile::have_alias_type(const TypePtr* adr_type) { 1887 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1888 if (ace->_adr_type == adr_type) { 1889 return true; 1890 } 1891 1892 // Handle special cases. 1893 if (adr_type == NULL) return true; 1894 if (adr_type == TypePtr::BOTTOM) return true; 1895 1896 return find_alias_type(adr_type, true, NULL) != NULL; 1897 } 1898 1899 //-----------------------------must_alias-------------------------------------- 1900 // True if all values of the given address type are in the given alias category. 1901 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1902 if (alias_idx == AliasIdxBot) return true; // the universal category 1903 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1904 if (alias_idx == AliasIdxTop) return false; // the empty category 1905 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1906 1907 // the only remaining possible overlap is identity 1908 int adr_idx = get_alias_index(adr_type); 1909 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1910 assert(adr_idx == alias_idx || 1911 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1912 && adr_type != TypeOopPtr::BOTTOM), 1913 "should not be testing for overlap with an unsafe pointer"); 1914 return adr_idx == alias_idx; 1915 } 1916 1917 //------------------------------can_alias-------------------------------------- 1918 // True if any values of the given address type are in the given alias category. 1919 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1920 if (alias_idx == AliasIdxTop) return false; // the empty category 1921 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1922 if (alias_idx == AliasIdxBot) return true; // the universal category 1923 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1924 1925 // the only remaining possible overlap is identity 1926 int adr_idx = get_alias_index(adr_type); 1927 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1928 return adr_idx == alias_idx; 1929 } 1930 1931 1932 1933 //---------------------------pop_warm_call------------------------------------- 1934 WarmCallInfo* Compile::pop_warm_call() { 1935 WarmCallInfo* wci = _warm_calls; 1936 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1937 return wci; 1938 } 1939 1940 //----------------------------Inline_Warm-------------------------------------- 1941 int Compile::Inline_Warm() { 1942 // If there is room, try to inline some more warm call sites. 1943 // %%% Do a graph index compaction pass when we think we're out of space? 1944 if (!InlineWarmCalls) return 0; 1945 1946 int calls_made_hot = 0; 1947 int room_to_grow = NodeCountInliningCutoff - unique(); 1948 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1949 int amount_grown = 0; 1950 WarmCallInfo* call; 1951 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1952 int est_size = (int)call->size(); 1953 if (est_size > (room_to_grow - amount_grown)) { 1954 // This one won't fit anyway. Get rid of it. 1955 call->make_cold(); 1956 continue; 1957 } 1958 call->make_hot(); 1959 calls_made_hot++; 1960 amount_grown += est_size; 1961 amount_to_grow -= est_size; 1962 } 1963 1964 if (calls_made_hot > 0) set_major_progress(); 1965 return calls_made_hot; 1966 } 1967 1968 1969 //----------------------------Finish_Warm-------------------------------------- 1970 void Compile::Finish_Warm() { 1971 if (!InlineWarmCalls) return; 1972 if (failing()) return; 1973 if (warm_calls() == NULL) return; 1974 1975 // Clean up loose ends, if we are out of space for inlining. 1976 WarmCallInfo* call; 1977 while ((call = pop_warm_call()) != NULL) { 1978 call->make_cold(); 1979 } 1980 } 1981 1982 //---------------------cleanup_loop_predicates----------------------- 1983 // Remove the opaque nodes that protect the predicates so that all unused 1984 // checks and uncommon_traps will be eliminated from the ideal graph 1985 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1986 if (predicate_count()==0) return; 1987 for (int i = predicate_count(); i > 0; i--) { 1988 Node * n = predicate_opaque1_node(i-1); 1989 assert(n->Opcode() == Op_Opaque1, "must be"); 1990 igvn.replace_node(n, n->in(1)); 1991 } 1992 assert(predicate_count()==0, "should be clean!"); 1993 } 1994 1995 void Compile::add_range_check_cast(Node* n) { 1996 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1997 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1998 _range_check_casts->append(n); 1999 } 2000 2001 // Remove all range check dependent CastIINodes. 2002 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 2003 for (int i = range_check_cast_count(); i > 0; i--) { 2004 Node* cast = range_check_cast_node(i-1); 2005 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 2006 igvn.replace_node(cast, cast->in(1)); 2007 } 2008 assert(range_check_cast_count() == 0, "should be empty"); 2009 } 2010 2011 void Compile::add_opaque4_node(Node* n) { 2012 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 2013 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 2014 _opaque4_nodes->append(n); 2015 } 2016 2017 // Remove all Opaque4 nodes. 2018 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 2019 for (int i = opaque4_count(); i > 0; i--) { 2020 Node* opaq = opaque4_node(i-1); 2021 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 2022 igvn.replace_node(opaq, opaq->in(2)); 2023 } 2024 assert(opaque4_count() == 0, "should be empty"); 2025 } 2026 2027 void Compile::add_value_type(Node* n) { 2028 assert(n->is_ValueTypeBase(), "unexpected node"); 2029 if (_value_type_nodes != NULL) { 2030 _value_type_nodes->push(n); 2031 } 2032 } 2033 2034 void Compile::remove_value_type(Node* n) { 2035 assert(n->is_ValueTypeBase(), "unexpected node"); 2036 if (_value_type_nodes != NULL) { 2037 _value_type_nodes->remove(n); 2038 } 2039 } 2040 2041 void Compile::process_value_types(PhaseIterGVN &igvn) { 2042 // Make value types scalar in safepoints 2043 while (_value_type_nodes->size() != 0) { 2044 ValueTypeBaseNode* vt = _value_type_nodes->pop()->as_ValueTypeBase(); 2045 vt->make_scalar_in_safepoints(igvn.C->root(), &igvn); 2046 if (vt->is_ValueTypePtr()) { 2047 igvn.replace_node(vt, vt->get_oop()); 2048 } 2049 } 2050 _value_type_nodes = NULL; 2051 igvn.optimize(); 2052 } 2053 2054 // StringOpts and late inlining of string methods 2055 void Compile::inline_string_calls(bool parse_time) { 2056 { 2057 // remove useless nodes to make the usage analysis simpler 2058 ResourceMark rm; 2059 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2060 } 2061 2062 { 2063 ResourceMark rm; 2064 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2065 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2066 print_method(PHASE_AFTER_STRINGOPTS, 3); 2067 } 2068 2069 // now inline anything that we skipped the first time around 2070 if (!parse_time) { 2071 _late_inlines_pos = _late_inlines.length(); 2072 } 2073 2074 while (_string_late_inlines.length() > 0) { 2075 CallGenerator* cg = _string_late_inlines.pop(); 2076 cg->do_late_inline(); 2077 if (failing()) return; 2078 } 2079 _string_late_inlines.trunc_to(0); 2080 } 2081 2082 // Late inlining of boxing methods 2083 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2084 if (_boxing_late_inlines.length() > 0) { 2085 assert(has_boxed_value(), "inconsistent"); 2086 2087 PhaseGVN* gvn = initial_gvn(); 2088 set_inlining_incrementally(true); 2089 2090 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2091 for_igvn()->clear(); 2092 gvn->replace_with(&igvn); 2093 2094 _late_inlines_pos = _late_inlines.length(); 2095 2096 while (_boxing_late_inlines.length() > 0) { 2097 CallGenerator* cg = _boxing_late_inlines.pop(); 2098 cg->do_late_inline(); 2099 if (failing()) return; 2100 } 2101 _boxing_late_inlines.trunc_to(0); 2102 2103 { 2104 ResourceMark rm; 2105 PhaseRemoveUseless pru(gvn, for_igvn()); 2106 } 2107 2108 igvn = PhaseIterGVN(gvn); 2109 igvn.optimize(); 2110 2111 set_inlining_progress(false); 2112 set_inlining_incrementally(false); 2113 } 2114 } 2115 2116 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 2117 assert(IncrementalInline, "incremental inlining should be on"); 2118 PhaseGVN* gvn = initial_gvn(); 2119 2120 set_inlining_progress(false); 2121 for_igvn()->clear(); 2122 gvn->replace_with(&igvn); 2123 2124 { 2125 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2126 int i = 0; 2127 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2128 CallGenerator* cg = _late_inlines.at(i); 2129 _late_inlines_pos = i+1; 2130 cg->do_late_inline(); 2131 if (failing()) return; 2132 } 2133 int j = 0; 2134 for (; i < _late_inlines.length(); i++, j++) { 2135 _late_inlines.at_put(j, _late_inlines.at(i)); 2136 } 2137 _late_inlines.trunc_to(j); 2138 } 2139 2140 { 2141 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2142 ResourceMark rm; 2143 PhaseRemoveUseless pru(gvn, for_igvn()); 2144 } 2145 2146 { 2147 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2148 igvn = PhaseIterGVN(gvn); 2149 } 2150 } 2151 2152 // Perform incremental inlining until bound on number of live nodes is reached 2153 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2154 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2155 2156 PhaseGVN* gvn = initial_gvn(); 2157 2158 set_inlining_incrementally(true); 2159 set_inlining_progress(true); 2160 uint low_live_nodes = 0; 2161 2162 while(inlining_progress() && _late_inlines.length() > 0) { 2163 2164 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2165 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2166 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2167 // PhaseIdealLoop is expensive so we only try it once we are 2168 // out of live nodes and we only try it again if the previous 2169 // helped got the number of nodes down significantly 2170 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone); 2171 if (failing()) return; 2172 low_live_nodes = live_nodes(); 2173 _major_progress = true; 2174 } 2175 2176 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2177 break; 2178 } 2179 } 2180 2181 inline_incrementally_one(igvn); 2182 2183 if (failing()) return; 2184 2185 { 2186 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2187 igvn.optimize(); 2188 } 2189 2190 if (failing()) return; 2191 } 2192 2193 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2194 2195 if (_string_late_inlines.length() > 0) { 2196 assert(has_stringbuilder(), "inconsistent"); 2197 for_igvn()->clear(); 2198 initial_gvn()->replace_with(&igvn); 2199 2200 inline_string_calls(false); 2201 2202 if (failing()) return; 2203 2204 { 2205 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2206 ResourceMark rm; 2207 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2208 } 2209 2210 { 2211 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2212 igvn = PhaseIterGVN(gvn); 2213 igvn.optimize(); 2214 } 2215 } 2216 2217 set_inlining_incrementally(false); 2218 } 2219 2220 2221 bool Compile::optimize_loops(int& loop_opts_cnt, PhaseIterGVN& igvn, LoopOptsMode mode) { 2222 if(loop_opts_cnt > 0) { 2223 debug_only( int cnt = 0; ); 2224 while(major_progress() && (loop_opts_cnt > 0)) { 2225 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2226 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2227 PhaseIdealLoop ideal_loop(igvn, mode); 2228 loop_opts_cnt--; 2229 if (failing()) return false; 2230 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2231 } 2232 } 2233 return true; 2234 } 2235 2236 //------------------------------Optimize--------------------------------------- 2237 // Given a graph, optimize it. 2238 void Compile::Optimize() { 2239 TracePhase tp("optimizer", &timers[_t_optimizer]); 2240 2241 #ifndef PRODUCT 2242 if (_directive->BreakAtCompileOption) { 2243 BREAKPOINT; 2244 } 2245 2246 #endif 2247 2248 #ifdef ASSERT 2249 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2250 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2251 #endif 2252 2253 ResourceMark rm; 2254 int loop_opts_cnt; 2255 2256 print_inlining_reinit(); 2257 2258 NOT_PRODUCT( verify_graph_edges(); ) 2259 2260 print_method(PHASE_AFTER_PARSING); 2261 2262 { 2263 // Iterative Global Value Numbering, including ideal transforms 2264 // Initialize IterGVN with types and values from parse-time GVN 2265 PhaseIterGVN igvn(initial_gvn()); 2266 #ifdef ASSERT 2267 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2268 #endif 2269 { 2270 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2271 igvn.optimize(); 2272 } 2273 2274 if (failing()) return; 2275 2276 print_method(PHASE_ITER_GVN1, 2); 2277 2278 inline_incrementally(igvn); 2279 2280 print_method(PHASE_INCREMENTAL_INLINE, 2); 2281 2282 if (failing()) return; 2283 2284 if (eliminate_boxing()) { 2285 // Inline valueOf() methods now. 2286 inline_boxing_calls(igvn); 2287 2288 if (AlwaysIncrementalInline) { 2289 inline_incrementally(igvn); 2290 } 2291 2292 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2293 2294 if (failing()) return; 2295 } 2296 2297 // Remove the speculative part of types and clean up the graph from 2298 // the extra CastPP nodes whose only purpose is to carry them. Do 2299 // that early so that optimizations are not disrupted by the extra 2300 // CastPP nodes. 2301 remove_speculative_types(igvn); 2302 2303 // No more new expensive nodes will be added to the list from here 2304 // so keep only the actual candidates for optimizations. 2305 cleanup_expensive_nodes(igvn); 2306 2307 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2308 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2309 initial_gvn()->replace_with(&igvn); 2310 for_igvn()->clear(); 2311 Unique_Node_List new_worklist(C->comp_arena()); 2312 { 2313 ResourceMark rm; 2314 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2315 } 2316 set_for_igvn(&new_worklist); 2317 igvn = PhaseIterGVN(initial_gvn()); 2318 igvn.optimize(); 2319 } 2320 2321 if (_value_type_nodes->size() > 0) { 2322 // Do this once all inlining is over to avoid getting inconsistent debug info 2323 process_value_types(igvn); 2324 } 2325 2326 // Perform escape analysis 2327 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2328 if (has_loops()) { 2329 // Cleanup graph (remove dead nodes). 2330 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2331 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone); 2332 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2333 if (failing()) return; 2334 } 2335 ConnectionGraph::do_analysis(this, &igvn); 2336 2337 if (failing()) return; 2338 2339 // Optimize out fields loads from scalar replaceable allocations. 2340 igvn.optimize(); 2341 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2342 2343 if (failing()) return; 2344 2345 if (congraph() != NULL && macro_count() > 0) { 2346 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2347 PhaseMacroExpand mexp(igvn); 2348 mexp.eliminate_macro_nodes(); 2349 igvn.set_delay_transform(false); 2350 2351 igvn.optimize(); 2352 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2353 2354 if (failing()) return; 2355 } 2356 } 2357 2358 // Loop transforms on the ideal graph. Range Check Elimination, 2359 // peeling, unrolling, etc. 2360 2361 // Set loop opts counter 2362 loop_opts_cnt = num_loop_opts(); 2363 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2364 { 2365 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2366 PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault); 2367 loop_opts_cnt--; 2368 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2369 if (failing()) return; 2370 } 2371 // Loop opts pass if partial peeling occurred in previous pass 2372 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2373 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2374 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2375 loop_opts_cnt--; 2376 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2377 if (failing()) return; 2378 } 2379 // Loop opts pass for loop-unrolling before CCP 2380 if(major_progress() && (loop_opts_cnt > 0)) { 2381 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2382 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2383 loop_opts_cnt--; 2384 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2385 } 2386 if (!failing()) { 2387 // Verify that last round of loop opts produced a valid graph 2388 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2389 PhaseIdealLoop::verify(igvn); 2390 } 2391 } 2392 if (failing()) return; 2393 2394 // Conditional Constant Propagation; 2395 PhaseCCP ccp( &igvn ); 2396 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2397 { 2398 TracePhase tp("ccp", &timers[_t_ccp]); 2399 ccp.do_transform(); 2400 } 2401 print_method(PHASE_CPP1, 2); 2402 2403 assert( true, "Break here to ccp.dump_old2new_map()"); 2404 2405 // Iterative Global Value Numbering, including ideal transforms 2406 { 2407 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2408 igvn = ccp; 2409 igvn.optimize(); 2410 } 2411 2412 print_method(PHASE_ITER_GVN2, 2); 2413 2414 if (failing()) return; 2415 2416 // Loop transforms on the ideal graph. Range Check Elimination, 2417 // peeling, unrolling, etc. 2418 if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) { 2419 return; 2420 } 2421 2422 #if INCLUDE_ZGC 2423 if (UseZGC) { 2424 ZBarrierSetC2::find_dominating_barriers(igvn); 2425 } 2426 #endif 2427 2428 if (failing()) return; 2429 2430 // Ensure that major progress is now clear 2431 C->clear_major_progress(); 2432 2433 { 2434 // Verify that all previous optimizations produced a valid graph 2435 // at least to this point, even if no loop optimizations were done. 2436 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2437 PhaseIdealLoop::verify(igvn); 2438 } 2439 2440 if (range_check_cast_count() > 0) { 2441 // No more loop optimizations. Remove all range check dependent CastIINodes. 2442 C->remove_range_check_casts(igvn); 2443 igvn.optimize(); 2444 } 2445 2446 #ifdef ASSERT 2447 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2448 bs->verify_gc_barriers(this, BarrierSetC2::BeforeExpand); 2449 #endif 2450 2451 { 2452 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2453 PhaseMacroExpand mex(igvn); 2454 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2); 2455 if (mex.expand_macro_nodes()) { 2456 assert(failing(), "must bail out w/ explicit message"); 2457 return; 2458 } 2459 } 2460 2461 if (opaque4_count() > 0) { 2462 C->remove_opaque4_nodes(igvn); 2463 igvn.optimize(); 2464 } 2465 2466 DEBUG_ONLY( _modified_nodes = NULL; ) 2467 } // (End scope of igvn; run destructor if necessary for asserts.) 2468 2469 process_print_inlining(); 2470 // A method with only infinite loops has no edges entering loops from root 2471 { 2472 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2473 if (final_graph_reshaping()) { 2474 assert(failing(), "must bail out w/ explicit message"); 2475 return; 2476 } 2477 } 2478 2479 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2480 } 2481 2482 //------------------------------Code_Gen--------------------------------------- 2483 // Given a graph, generate code for it 2484 void Compile::Code_Gen() { 2485 if (failing()) { 2486 return; 2487 } 2488 2489 // Perform instruction selection. You might think we could reclaim Matcher 2490 // memory PDQ, but actually the Matcher is used in generating spill code. 2491 // Internals of the Matcher (including some VectorSets) must remain live 2492 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2493 // set a bit in reclaimed memory. 2494 2495 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2496 // nodes. Mapping is only valid at the root of each matched subtree. 2497 NOT_PRODUCT( verify_graph_edges(); ) 2498 2499 Matcher matcher; 2500 _matcher = &matcher; 2501 { 2502 TracePhase tp("matcher", &timers[_t_matcher]); 2503 matcher.match(); 2504 } 2505 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2506 // nodes. Mapping is only valid at the root of each matched subtree. 2507 NOT_PRODUCT( verify_graph_edges(); ) 2508 2509 // If you have too many nodes, or if matching has failed, bail out 2510 check_node_count(0, "out of nodes matching instructions"); 2511 if (failing()) { 2512 return; 2513 } 2514 2515 print_method(PHASE_MATCHING, 2); 2516 2517 // Build a proper-looking CFG 2518 PhaseCFG cfg(node_arena(), root(), matcher); 2519 _cfg = &cfg; 2520 { 2521 TracePhase tp("scheduler", &timers[_t_scheduler]); 2522 bool success = cfg.do_global_code_motion(); 2523 if (!success) { 2524 return; 2525 } 2526 2527 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2528 NOT_PRODUCT( verify_graph_edges(); ) 2529 debug_only( cfg.verify(); ) 2530 } 2531 2532 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2533 _regalloc = ®alloc; 2534 { 2535 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2536 // Perform register allocation. After Chaitin, use-def chains are 2537 // no longer accurate (at spill code) and so must be ignored. 2538 // Node->LRG->reg mappings are still accurate. 2539 _regalloc->Register_Allocate(); 2540 2541 // Bail out if the allocator builds too many nodes 2542 if (failing()) { 2543 return; 2544 } 2545 } 2546 2547 // Prior to register allocation we kept empty basic blocks in case the 2548 // the allocator needed a place to spill. After register allocation we 2549 // are not adding any new instructions. If any basic block is empty, we 2550 // can now safely remove it. 2551 { 2552 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2553 cfg.remove_empty_blocks(); 2554 if (do_freq_based_layout()) { 2555 PhaseBlockLayout layout(cfg); 2556 } else { 2557 cfg.set_loop_alignment(); 2558 } 2559 cfg.fixup_flow(); 2560 } 2561 2562 // Apply peephole optimizations 2563 if( OptoPeephole ) { 2564 TracePhase tp("peephole", &timers[_t_peephole]); 2565 PhasePeephole peep( _regalloc, cfg); 2566 peep.do_transform(); 2567 } 2568 2569 // Do late expand if CPU requires this. 2570 if (Matcher::require_postalloc_expand) { 2571 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2572 cfg.postalloc_expand(_regalloc); 2573 } 2574 2575 // Convert Nodes to instruction bits in a buffer 2576 { 2577 TraceTime tp("output", &timers[_t_output], CITime); 2578 Output(); 2579 } 2580 2581 print_method(PHASE_FINAL_CODE); 2582 2583 // He's dead, Jim. 2584 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2585 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2586 } 2587 2588 2589 //------------------------------dump_asm--------------------------------------- 2590 // Dump formatted assembly 2591 #ifndef PRODUCT 2592 void Compile::dump_asm(int *pcs, uint pc_limit) { 2593 bool cut_short = false; 2594 tty->print_cr("#"); 2595 tty->print("# "); _tf->dump(); tty->cr(); 2596 tty->print_cr("#"); 2597 2598 // For all blocks 2599 int pc = 0x0; // Program counter 2600 char starts_bundle = ' '; 2601 _regalloc->dump_frame(); 2602 2603 Node *n = NULL; 2604 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2605 if (VMThread::should_terminate()) { 2606 cut_short = true; 2607 break; 2608 } 2609 Block* block = _cfg->get_block(i); 2610 if (block->is_connector() && !Verbose) { 2611 continue; 2612 } 2613 n = block->head(); 2614 if (pcs && n->_idx < pc_limit) { 2615 tty->print("%3.3x ", pcs[n->_idx]); 2616 } else { 2617 tty->print(" "); 2618 } 2619 block->dump_head(_cfg); 2620 if (block->is_connector()) { 2621 tty->print_cr(" # Empty connector block"); 2622 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2623 tty->print_cr(" # Block is sole successor of call"); 2624 } 2625 2626 // For all instructions 2627 Node *delay = NULL; 2628 for (uint j = 0; j < block->number_of_nodes(); j++) { 2629 if (VMThread::should_terminate()) { 2630 cut_short = true; 2631 break; 2632 } 2633 n = block->get_node(j); 2634 if (valid_bundle_info(n)) { 2635 Bundle* bundle = node_bundling(n); 2636 if (bundle->used_in_unconditional_delay()) { 2637 delay = n; 2638 continue; 2639 } 2640 if (bundle->starts_bundle()) { 2641 starts_bundle = '+'; 2642 } 2643 } 2644 2645 if (WizardMode) { 2646 n->dump(); 2647 } 2648 2649 if( !n->is_Region() && // Dont print in the Assembly 2650 !n->is_Phi() && // a few noisely useless nodes 2651 !n->is_Proj() && 2652 !n->is_MachTemp() && 2653 !n->is_SafePointScalarObject() && 2654 !n->is_Catch() && // Would be nice to print exception table targets 2655 !n->is_MergeMem() && // Not very interesting 2656 !n->is_top() && // Debug info table constants 2657 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2658 ) { 2659 if (pcs && n->_idx < pc_limit) 2660 tty->print("%3.3x", pcs[n->_idx]); 2661 else 2662 tty->print(" "); 2663 tty->print(" %c ", starts_bundle); 2664 starts_bundle = ' '; 2665 tty->print("\t"); 2666 n->format(_regalloc, tty); 2667 tty->cr(); 2668 } 2669 2670 // If we have an instruction with a delay slot, and have seen a delay, 2671 // then back up and print it 2672 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2673 assert(delay != NULL, "no unconditional delay instruction"); 2674 if (WizardMode) delay->dump(); 2675 2676 if (node_bundling(delay)->starts_bundle()) 2677 starts_bundle = '+'; 2678 if (pcs && n->_idx < pc_limit) 2679 tty->print("%3.3x", pcs[n->_idx]); 2680 else 2681 tty->print(" "); 2682 tty->print(" %c ", starts_bundle); 2683 starts_bundle = ' '; 2684 tty->print("\t"); 2685 delay->format(_regalloc, tty); 2686 tty->cr(); 2687 delay = NULL; 2688 } 2689 2690 // Dump the exception table as well 2691 if( n->is_Catch() && (Verbose || WizardMode) ) { 2692 // Print the exception table for this offset 2693 _handler_table.print_subtable_for(pc); 2694 } 2695 } 2696 2697 if (pcs && n->_idx < pc_limit) 2698 tty->print_cr("%3.3x", pcs[n->_idx]); 2699 else 2700 tty->cr(); 2701 2702 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2703 2704 } // End of per-block dump 2705 tty->cr(); 2706 2707 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2708 } 2709 #endif 2710 2711 //------------------------------Final_Reshape_Counts--------------------------- 2712 // This class defines counters to help identify when a method 2713 // may/must be executed using hardware with only 24-bit precision. 2714 struct Final_Reshape_Counts : public StackObj { 2715 int _call_count; // count non-inlined 'common' calls 2716 int _float_count; // count float ops requiring 24-bit precision 2717 int _double_count; // count double ops requiring more precision 2718 int _java_call_count; // count non-inlined 'java' calls 2719 int _inner_loop_count; // count loops which need alignment 2720 VectorSet _visited; // Visitation flags 2721 Node_List _tests; // Set of IfNodes & PCTableNodes 2722 2723 Final_Reshape_Counts() : 2724 _call_count(0), _float_count(0), _double_count(0), 2725 _java_call_count(0), _inner_loop_count(0), 2726 _visited( Thread::current()->resource_area() ) { } 2727 2728 void inc_call_count () { _call_count ++; } 2729 void inc_float_count () { _float_count ++; } 2730 void inc_double_count() { _double_count++; } 2731 void inc_java_call_count() { _java_call_count++; } 2732 void inc_inner_loop_count() { _inner_loop_count++; } 2733 2734 int get_call_count () const { return _call_count ; } 2735 int get_float_count () const { return _float_count ; } 2736 int get_double_count() const { return _double_count; } 2737 int get_java_call_count() const { return _java_call_count; } 2738 int get_inner_loop_count() const { return _inner_loop_count; } 2739 }; 2740 2741 #ifdef ASSERT 2742 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2743 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2744 // Make sure the offset goes inside the instance layout. 2745 return k->contains_field_offset(tp->offset()); 2746 // Note that OffsetBot and OffsetTop are very negative. 2747 } 2748 #endif 2749 2750 // Eliminate trivially redundant StoreCMs and accumulate their 2751 // precedence edges. 2752 void Compile::eliminate_redundant_card_marks(Node* n) { 2753 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2754 if (n->in(MemNode::Address)->outcnt() > 1) { 2755 // There are multiple users of the same address so it might be 2756 // possible to eliminate some of the StoreCMs 2757 Node* mem = n->in(MemNode::Memory); 2758 Node* adr = n->in(MemNode::Address); 2759 Node* val = n->in(MemNode::ValueIn); 2760 Node* prev = n; 2761 bool done = false; 2762 // Walk the chain of StoreCMs eliminating ones that match. As 2763 // long as it's a chain of single users then the optimization is 2764 // safe. Eliminating partially redundant StoreCMs would require 2765 // cloning copies down the other paths. 2766 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2767 if (adr == mem->in(MemNode::Address) && 2768 val == mem->in(MemNode::ValueIn)) { 2769 // redundant StoreCM 2770 if (mem->req() > MemNode::OopStore) { 2771 // Hasn't been processed by this code yet. 2772 n->add_prec(mem->in(MemNode::OopStore)); 2773 } else { 2774 // Already converted to precedence edge 2775 for (uint i = mem->req(); i < mem->len(); i++) { 2776 // Accumulate any precedence edges 2777 if (mem->in(i) != NULL) { 2778 n->add_prec(mem->in(i)); 2779 } 2780 } 2781 // Everything above this point has been processed. 2782 done = true; 2783 } 2784 // Eliminate the previous StoreCM 2785 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2786 assert(mem->outcnt() == 0, "should be dead"); 2787 mem->disconnect_inputs(NULL, this); 2788 } else { 2789 prev = mem; 2790 } 2791 mem = prev->in(MemNode::Memory); 2792 } 2793 } 2794 } 2795 2796 2797 //------------------------------final_graph_reshaping_impl---------------------- 2798 // Implement items 1-5 from final_graph_reshaping below. 2799 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2800 2801 if ( n->outcnt() == 0 ) return; // dead node 2802 uint nop = n->Opcode(); 2803 2804 // Check for 2-input instruction with "last use" on right input. 2805 // Swap to left input. Implements item (2). 2806 if( n->req() == 3 && // two-input instruction 2807 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2808 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2809 n->in(2)->outcnt() == 1 &&// right use IS a last use 2810 !n->in(2)->is_Con() ) { // right use is not a constant 2811 // Check for commutative opcode 2812 switch( nop ) { 2813 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2814 case Op_MaxI: case Op_MinI: 2815 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2816 case Op_AndL: case Op_XorL: case Op_OrL: 2817 case Op_AndI: case Op_XorI: case Op_OrI: { 2818 // Move "last use" input to left by swapping inputs 2819 n->swap_edges(1, 2); 2820 break; 2821 } 2822 default: 2823 break; 2824 } 2825 } 2826 2827 #ifdef ASSERT 2828 if( n->is_Mem() ) { 2829 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2830 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2831 // oop will be recorded in oop map if load crosses safepoint 2832 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2833 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2834 "raw memory operations should have control edge"); 2835 } 2836 if (n->is_MemBar()) { 2837 MemBarNode* mb = n->as_MemBar(); 2838 if (mb->trailing_store() || mb->trailing_load_store()) { 2839 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2840 Node* mem = mb->in(MemBarNode::Precedent); 2841 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2842 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2843 } else if (mb->leading()) { 2844 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2845 } 2846 } 2847 #endif 2848 // Count FPU ops and common calls, implements item (3) 2849 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop); 2850 if (!gc_handled) { 2851 final_graph_reshaping_main_switch(n, frc, nop); 2852 } 2853 2854 // Collect CFG split points 2855 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 2856 frc._tests.push(n); 2857 } 2858 } 2859 2860 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) { 2861 switch( nop ) { 2862 // Count all float operations that may use FPU 2863 case Op_AddF: 2864 case Op_SubF: 2865 case Op_MulF: 2866 case Op_DivF: 2867 case Op_NegF: 2868 case Op_ModF: 2869 case Op_ConvI2F: 2870 case Op_ConF: 2871 case Op_CmpF: 2872 case Op_CmpF3: 2873 // case Op_ConvL2F: // longs are split into 32-bit halves 2874 frc.inc_float_count(); 2875 break; 2876 2877 case Op_ConvF2D: 2878 case Op_ConvD2F: 2879 frc.inc_float_count(); 2880 frc.inc_double_count(); 2881 break; 2882 2883 // Count all double operations that may use FPU 2884 case Op_AddD: 2885 case Op_SubD: 2886 case Op_MulD: 2887 case Op_DivD: 2888 case Op_NegD: 2889 case Op_ModD: 2890 case Op_ConvI2D: 2891 case Op_ConvD2I: 2892 // case Op_ConvL2D: // handled by leaf call 2893 // case Op_ConvD2L: // handled by leaf call 2894 case Op_ConD: 2895 case Op_CmpD: 2896 case Op_CmpD3: 2897 frc.inc_double_count(); 2898 break; 2899 case Op_Opaque1: // Remove Opaque Nodes before matching 2900 case Op_Opaque2: // Remove Opaque Nodes before matching 2901 case Op_Opaque3: 2902 n->subsume_by(n->in(1), this); 2903 break; 2904 case Op_CallStaticJava: 2905 case Op_CallJava: 2906 case Op_CallDynamicJava: 2907 frc.inc_java_call_count(); // Count java call site; 2908 case Op_CallRuntime: 2909 case Op_CallLeaf: 2910 case Op_CallLeafNoFP: { 2911 assert (n->is_Call(), ""); 2912 CallNode *call = n->as_Call(); 2913 // Count call sites where the FP mode bit would have to be flipped. 2914 // Do not count uncommon runtime calls: 2915 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2916 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2917 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 2918 frc.inc_call_count(); // Count the call site 2919 } else { // See if uncommon argument is shared 2920 Node *n = call->in(TypeFunc::Parms); 2921 int nop = n->Opcode(); 2922 // Clone shared simple arguments to uncommon calls, item (1). 2923 if (n->outcnt() > 1 && 2924 !n->is_Proj() && 2925 nop != Op_CreateEx && 2926 nop != Op_CheckCastPP && 2927 nop != Op_DecodeN && 2928 nop != Op_DecodeNKlass && 2929 !n->is_Mem() && 2930 !n->is_Phi()) { 2931 Node *x = n->clone(); 2932 call->set_req(TypeFunc::Parms, x); 2933 } 2934 } 2935 break; 2936 } 2937 2938 case Op_StoreD: 2939 case Op_LoadD: 2940 case Op_LoadD_unaligned: 2941 frc.inc_double_count(); 2942 goto handle_mem; 2943 case Op_StoreF: 2944 case Op_LoadF: 2945 frc.inc_float_count(); 2946 goto handle_mem; 2947 2948 case Op_StoreCM: 2949 { 2950 // Convert OopStore dependence into precedence edge 2951 Node* prec = n->in(MemNode::OopStore); 2952 n->del_req(MemNode::OopStore); 2953 n->add_prec(prec); 2954 eliminate_redundant_card_marks(n); 2955 } 2956 2957 // fall through 2958 2959 case Op_StoreB: 2960 case Op_StoreC: 2961 case Op_StorePConditional: 2962 case Op_StoreI: 2963 case Op_StoreL: 2964 case Op_StoreIConditional: 2965 case Op_StoreLConditional: 2966 case Op_CompareAndSwapB: 2967 case Op_CompareAndSwapS: 2968 case Op_CompareAndSwapI: 2969 case Op_CompareAndSwapL: 2970 case Op_CompareAndSwapP: 2971 case Op_CompareAndSwapN: 2972 case Op_WeakCompareAndSwapB: 2973 case Op_WeakCompareAndSwapS: 2974 case Op_WeakCompareAndSwapI: 2975 case Op_WeakCompareAndSwapL: 2976 case Op_WeakCompareAndSwapP: 2977 case Op_WeakCompareAndSwapN: 2978 case Op_CompareAndExchangeB: 2979 case Op_CompareAndExchangeS: 2980 case Op_CompareAndExchangeI: 2981 case Op_CompareAndExchangeL: 2982 case Op_CompareAndExchangeP: 2983 case Op_CompareAndExchangeN: 2984 case Op_GetAndAddS: 2985 case Op_GetAndAddB: 2986 case Op_GetAndAddI: 2987 case Op_GetAndAddL: 2988 case Op_GetAndSetS: 2989 case Op_GetAndSetB: 2990 case Op_GetAndSetI: 2991 case Op_GetAndSetL: 2992 case Op_GetAndSetP: 2993 case Op_GetAndSetN: 2994 case Op_StoreP: 2995 case Op_StoreN: 2996 case Op_StoreNKlass: 2997 case Op_LoadB: 2998 case Op_LoadUB: 2999 case Op_LoadUS: 3000 case Op_LoadI: 3001 case Op_LoadKlass: 3002 case Op_LoadNKlass: 3003 case Op_LoadL: 3004 case Op_LoadL_unaligned: 3005 case Op_LoadPLocked: 3006 case Op_LoadP: 3007 case Op_LoadN: 3008 case Op_LoadRange: 3009 case Op_LoadS: { 3010 handle_mem: 3011 #ifdef ASSERT 3012 if( VerifyOptoOopOffsets ) { 3013 MemNode* mem = n->as_Mem(); 3014 // Check to see if address types have grounded out somehow. 3015 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 3016 assert( !tp || oop_offset_is_sane(tp), "" ); 3017 } 3018 #endif 3019 break; 3020 } 3021 3022 case Op_AddP: { // Assert sane base pointers 3023 Node *addp = n->in(AddPNode::Address); 3024 assert( !addp->is_AddP() || 3025 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3026 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3027 "Base pointers must match (addp %u)", addp->_idx ); 3028 #ifdef _LP64 3029 if ((UseCompressedOops || UseCompressedClassPointers) && 3030 addp->Opcode() == Op_ConP && 3031 addp == n->in(AddPNode::Base) && 3032 n->in(AddPNode::Offset)->is_Con()) { 3033 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3034 // on the platform and on the compressed oops mode. 3035 // Use addressing with narrow klass to load with offset on x86. 3036 // Some platforms can use the constant pool to load ConP. 3037 // Do this transformation here since IGVN will convert ConN back to ConP. 3038 const Type* t = addp->bottom_type(); 3039 bool is_oop = t->isa_oopptr() != NULL; 3040 bool is_klass = t->isa_klassptr() != NULL; 3041 3042 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3043 (is_klass && Matcher::const_klass_prefer_decode())) { 3044 Node* nn = NULL; 3045 3046 int op = is_oop ? Op_ConN : Op_ConNKlass; 3047 3048 // Look for existing ConN node of the same exact type. 3049 Node* r = root(); 3050 uint cnt = r->outcnt(); 3051 for (uint i = 0; i < cnt; i++) { 3052 Node* m = r->raw_out(i); 3053 if (m!= NULL && m->Opcode() == op && 3054 m->bottom_type()->make_ptr() == t) { 3055 nn = m; 3056 break; 3057 } 3058 } 3059 if (nn != NULL) { 3060 // Decode a narrow oop to match address 3061 // [R12 + narrow_oop_reg<<3 + offset] 3062 if (is_oop) { 3063 nn = new DecodeNNode(nn, t); 3064 } else { 3065 nn = new DecodeNKlassNode(nn, t); 3066 } 3067 // Check for succeeding AddP which uses the same Base. 3068 // Otherwise we will run into the assertion above when visiting that guy. 3069 for (uint i = 0; i < n->outcnt(); ++i) { 3070 Node *out_i = n->raw_out(i); 3071 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3072 out_i->set_req(AddPNode::Base, nn); 3073 #ifdef ASSERT 3074 for (uint j = 0; j < out_i->outcnt(); ++j) { 3075 Node *out_j = out_i->raw_out(j); 3076 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3077 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3078 } 3079 #endif 3080 } 3081 } 3082 n->set_req(AddPNode::Base, nn); 3083 n->set_req(AddPNode::Address, nn); 3084 if (addp->outcnt() == 0) { 3085 addp->disconnect_inputs(NULL, this); 3086 } 3087 } 3088 } 3089 } 3090 #endif 3091 // platform dependent reshaping of the address expression 3092 reshape_address(n->as_AddP()); 3093 break; 3094 } 3095 3096 case Op_CastPP: { 3097 // Remove CastPP nodes to gain more freedom during scheduling but 3098 // keep the dependency they encode as control or precedence edges 3099 // (if control is set already) on memory operations. Some CastPP 3100 // nodes don't have a control (don't carry a dependency): skip 3101 // those. 3102 if (n->in(0) != NULL) { 3103 ResourceMark rm; 3104 Unique_Node_List wq; 3105 wq.push(n); 3106 for (uint next = 0; next < wq.size(); ++next) { 3107 Node *m = wq.at(next); 3108 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3109 Node* use = m->fast_out(i); 3110 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3111 use->ensure_control_or_add_prec(n->in(0)); 3112 } else { 3113 switch(use->Opcode()) { 3114 case Op_AddP: 3115 case Op_DecodeN: 3116 case Op_DecodeNKlass: 3117 case Op_CheckCastPP: 3118 case Op_CastPP: 3119 wq.push(use); 3120 break; 3121 } 3122 } 3123 } 3124 } 3125 } 3126 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3127 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3128 Node* in1 = n->in(1); 3129 const Type* t = n->bottom_type(); 3130 Node* new_in1 = in1->clone(); 3131 new_in1->as_DecodeN()->set_type(t); 3132 3133 if (!Matcher::narrow_oop_use_complex_address()) { 3134 // 3135 // x86, ARM and friends can handle 2 adds in addressing mode 3136 // and Matcher can fold a DecodeN node into address by using 3137 // a narrow oop directly and do implicit NULL check in address: 3138 // 3139 // [R12 + narrow_oop_reg<<3 + offset] 3140 // NullCheck narrow_oop_reg 3141 // 3142 // On other platforms (Sparc) we have to keep new DecodeN node and 3143 // use it to do implicit NULL check in address: 3144 // 3145 // decode_not_null narrow_oop_reg, base_reg 3146 // [base_reg + offset] 3147 // NullCheck base_reg 3148 // 3149 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3150 // to keep the information to which NULL check the new DecodeN node 3151 // corresponds to use it as value in implicit_null_check(). 3152 // 3153 new_in1->set_req(0, n->in(0)); 3154 } 3155 3156 n->subsume_by(new_in1, this); 3157 if (in1->outcnt() == 0) { 3158 in1->disconnect_inputs(NULL, this); 3159 } 3160 } else { 3161 n->subsume_by(n->in(1), this); 3162 if (n->outcnt() == 0) { 3163 n->disconnect_inputs(NULL, this); 3164 } 3165 } 3166 break; 3167 } 3168 #ifdef _LP64 3169 case Op_CmpP: 3170 // Do this transformation here to preserve CmpPNode::sub() and 3171 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3172 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3173 Node* in1 = n->in(1); 3174 Node* in2 = n->in(2); 3175 if (!in1->is_DecodeNarrowPtr()) { 3176 in2 = in1; 3177 in1 = n->in(2); 3178 } 3179 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3180 3181 Node* new_in2 = NULL; 3182 if (in2->is_DecodeNarrowPtr()) { 3183 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3184 new_in2 = in2->in(1); 3185 } else if (in2->Opcode() == Op_ConP) { 3186 const Type* t = in2->bottom_type(); 3187 if (t == TypePtr::NULL_PTR) { 3188 assert(in1->is_DecodeN(), "compare klass to null?"); 3189 // Don't convert CmpP null check into CmpN if compressed 3190 // oops implicit null check is not generated. 3191 // This will allow to generate normal oop implicit null check. 3192 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3193 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3194 // 3195 // This transformation together with CastPP transformation above 3196 // will generated code for implicit NULL checks for compressed oops. 3197 // 3198 // The original code after Optimize() 3199 // 3200 // LoadN memory, narrow_oop_reg 3201 // decode narrow_oop_reg, base_reg 3202 // CmpP base_reg, NULL 3203 // CastPP base_reg // NotNull 3204 // Load [base_reg + offset], val_reg 3205 // 3206 // after these transformations will be 3207 // 3208 // LoadN memory, narrow_oop_reg 3209 // CmpN narrow_oop_reg, NULL 3210 // decode_not_null narrow_oop_reg, base_reg 3211 // Load [base_reg + offset], val_reg 3212 // 3213 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3214 // since narrow oops can be used in debug info now (see the code in 3215 // final_graph_reshaping_walk()). 3216 // 3217 // At the end the code will be matched to 3218 // on x86: 3219 // 3220 // Load_narrow_oop memory, narrow_oop_reg 3221 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3222 // NullCheck narrow_oop_reg 3223 // 3224 // and on sparc: 3225 // 3226 // Load_narrow_oop memory, narrow_oop_reg 3227 // decode_not_null narrow_oop_reg, base_reg 3228 // Load [base_reg + offset], val_reg 3229 // NullCheck base_reg 3230 // 3231 } else if (t->isa_oopptr()) { 3232 new_in2 = ConNode::make(t->make_narrowoop()); 3233 } else if (t->isa_klassptr()) { 3234 new_in2 = ConNode::make(t->make_narrowklass()); 3235 } 3236 } 3237 if (new_in2 != NULL) { 3238 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3239 n->subsume_by(cmpN, this); 3240 if (in1->outcnt() == 0) { 3241 in1->disconnect_inputs(NULL, this); 3242 } 3243 if (in2->outcnt() == 0) { 3244 in2->disconnect_inputs(NULL, this); 3245 } 3246 } 3247 } 3248 break; 3249 3250 case Op_DecodeN: 3251 case Op_DecodeNKlass: 3252 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3253 // DecodeN could be pinned when it can't be fold into 3254 // an address expression, see the code for Op_CastPP above. 3255 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3256 break; 3257 3258 case Op_EncodeP: 3259 case Op_EncodePKlass: { 3260 Node* in1 = n->in(1); 3261 if (in1->is_DecodeNarrowPtr()) { 3262 n->subsume_by(in1->in(1), this); 3263 } else if (in1->Opcode() == Op_ConP) { 3264 const Type* t = in1->bottom_type(); 3265 if (t == TypePtr::NULL_PTR) { 3266 assert(t->isa_oopptr(), "null klass?"); 3267 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3268 } else if (t->isa_oopptr()) { 3269 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3270 } else if (t->isa_klassptr()) { 3271 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3272 } 3273 } 3274 if (in1->outcnt() == 0) { 3275 in1->disconnect_inputs(NULL, this); 3276 } 3277 break; 3278 } 3279 3280 case Op_Proj: { 3281 if (OptimizeStringConcat) { 3282 ProjNode* p = n->as_Proj(); 3283 if (p->_is_io_use) { 3284 // Separate projections were used for the exception path which 3285 // are normally removed by a late inline. If it wasn't inlined 3286 // then they will hang around and should just be replaced with 3287 // the original one. 3288 Node* proj = NULL; 3289 // Replace with just one 3290 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3291 Node *use = i.get(); 3292 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3293 proj = use; 3294 break; 3295 } 3296 } 3297 assert(proj != NULL, "must be found"); 3298 p->subsume_by(proj, this); 3299 } 3300 } 3301 break; 3302 } 3303 3304 case Op_Phi: 3305 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3306 // The EncodeP optimization may create Phi with the same edges 3307 // for all paths. It is not handled well by Register Allocator. 3308 Node* unique_in = n->in(1); 3309 assert(unique_in != NULL, ""); 3310 uint cnt = n->req(); 3311 for (uint i = 2; i < cnt; i++) { 3312 Node* m = n->in(i); 3313 assert(m != NULL, ""); 3314 if (unique_in != m) 3315 unique_in = NULL; 3316 } 3317 if (unique_in != NULL) { 3318 n->subsume_by(unique_in, this); 3319 } 3320 } 3321 break; 3322 3323 #endif 3324 3325 #ifdef ASSERT 3326 case Op_CastII: 3327 // Verify that all range check dependent CastII nodes were removed. 3328 if (n->isa_CastII()->has_range_check()) { 3329 n->dump(3); 3330 assert(false, "Range check dependent CastII node was not removed"); 3331 } 3332 break; 3333 #endif 3334 3335 case Op_ModI: 3336 if (UseDivMod) { 3337 // Check if a%b and a/b both exist 3338 Node* d = n->find_similar(Op_DivI); 3339 if (d) { 3340 // Replace them with a fused divmod if supported 3341 if (Matcher::has_match_rule(Op_DivModI)) { 3342 DivModINode* divmod = DivModINode::make(n); 3343 d->subsume_by(divmod->div_proj(), this); 3344 n->subsume_by(divmod->mod_proj(), this); 3345 } else { 3346 // replace a%b with a-((a/b)*b) 3347 Node* mult = new MulINode(d, d->in(2)); 3348 Node* sub = new SubINode(d->in(1), mult); 3349 n->subsume_by(sub, this); 3350 } 3351 } 3352 } 3353 break; 3354 3355 case Op_ModL: 3356 if (UseDivMod) { 3357 // Check if a%b and a/b both exist 3358 Node* d = n->find_similar(Op_DivL); 3359 if (d) { 3360 // Replace them with a fused divmod if supported 3361 if (Matcher::has_match_rule(Op_DivModL)) { 3362 DivModLNode* divmod = DivModLNode::make(n); 3363 d->subsume_by(divmod->div_proj(), this); 3364 n->subsume_by(divmod->mod_proj(), this); 3365 } else { 3366 // replace a%b with a-((a/b)*b) 3367 Node* mult = new MulLNode(d, d->in(2)); 3368 Node* sub = new SubLNode(d->in(1), mult); 3369 n->subsume_by(sub, this); 3370 } 3371 } 3372 } 3373 break; 3374 3375 case Op_LoadVector: 3376 case Op_StoreVector: 3377 break; 3378 3379 case Op_AddReductionVI: 3380 case Op_AddReductionVL: 3381 case Op_AddReductionVF: 3382 case Op_AddReductionVD: 3383 case Op_MulReductionVI: 3384 case Op_MulReductionVL: 3385 case Op_MulReductionVF: 3386 case Op_MulReductionVD: 3387 break; 3388 3389 case Op_PackB: 3390 case Op_PackS: 3391 case Op_PackI: 3392 case Op_PackF: 3393 case Op_PackL: 3394 case Op_PackD: 3395 if (n->req()-1 > 2) { 3396 // Replace many operand PackNodes with a binary tree for matching 3397 PackNode* p = (PackNode*) n; 3398 Node* btp = p->binary_tree_pack(1, n->req()); 3399 n->subsume_by(btp, this); 3400 } 3401 break; 3402 case Op_Loop: 3403 case Op_CountedLoop: 3404 case Op_OuterStripMinedLoop: 3405 if (n->as_Loop()->is_inner_loop()) { 3406 frc.inc_inner_loop_count(); 3407 } 3408 n->as_Loop()->verify_strip_mined(0); 3409 break; 3410 case Op_LShiftI: 3411 case Op_RShiftI: 3412 case Op_URShiftI: 3413 case Op_LShiftL: 3414 case Op_RShiftL: 3415 case Op_URShiftL: 3416 if (Matcher::need_masked_shift_count) { 3417 // The cpu's shift instructions don't restrict the count to the 3418 // lower 5/6 bits. We need to do the masking ourselves. 3419 Node* in2 = n->in(2); 3420 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3421 const TypeInt* t = in2->find_int_type(); 3422 if (t != NULL && t->is_con()) { 3423 juint shift = t->get_con(); 3424 if (shift > mask) { // Unsigned cmp 3425 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3426 } 3427 } else { 3428 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3429 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3430 n->set_req(2, shift); 3431 } 3432 } 3433 if (in2->outcnt() == 0) { // Remove dead node 3434 in2->disconnect_inputs(NULL, this); 3435 } 3436 } 3437 break; 3438 case Op_MemBarStoreStore: 3439 case Op_MemBarRelease: 3440 // Break the link with AllocateNode: it is no longer useful and 3441 // confuses register allocation. 3442 if (n->req() > MemBarNode::Precedent) { 3443 n->set_req(MemBarNode::Precedent, top()); 3444 } 3445 break; 3446 case Op_MemBarAcquire: { 3447 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3448 // At parse time, the trailing MemBarAcquire for a volatile load 3449 // is created with an edge to the load. After optimizations, 3450 // that input may be a chain of Phis. If those phis have no 3451 // other use, then the MemBarAcquire keeps them alive and 3452 // register allocation can be confused. 3453 ResourceMark rm; 3454 Unique_Node_List wq; 3455 wq.push(n->in(MemBarNode::Precedent)); 3456 n->set_req(MemBarNode::Precedent, top()); 3457 while (wq.size() > 0) { 3458 Node* m = wq.pop(); 3459 if (m->outcnt() == 0) { 3460 for (uint j = 0; j < m->req(); j++) { 3461 Node* in = m->in(j); 3462 if (in != NULL) { 3463 wq.push(in); 3464 } 3465 } 3466 m->disconnect_inputs(NULL, this); 3467 } 3468 } 3469 } 3470 break; 3471 } 3472 case Op_RangeCheck: { 3473 RangeCheckNode* rc = n->as_RangeCheck(); 3474 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3475 n->subsume_by(iff, this); 3476 frc._tests.push(iff); 3477 break; 3478 } 3479 case Op_ConvI2L: { 3480 if (!Matcher::convi2l_type_required) { 3481 // Code generation on some platforms doesn't need accurate 3482 // ConvI2L types. Widening the type can help remove redundant 3483 // address computations. 3484 n->as_Type()->set_type(TypeLong::INT); 3485 ResourceMark rm; 3486 Node_List wq; 3487 wq.push(n); 3488 for (uint next = 0; next < wq.size(); next++) { 3489 Node *m = wq.at(next); 3490 3491 for(;;) { 3492 // Loop over all nodes with identical inputs edges as m 3493 Node* k = m->find_similar(m->Opcode()); 3494 if (k == NULL) { 3495 break; 3496 } 3497 // Push their uses so we get a chance to remove node made 3498 // redundant 3499 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3500 Node* u = k->fast_out(i); 3501 assert(!wq.contains(u), "shouldn't process one node several times"); 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 // We don't support unsigned long comparisons. Set 'max_idx_expr' 3519 // to max_julong if < 0 to make the signed comparison fail. 3520 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3521 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3522 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3523 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3524 Node* andl = new AndLNode(orl, remove_sign_mask); 3525 Node* cmp = new CmpLNode(andl, n->in(2)); 3526 n->subsume_by(cmp, this); 3527 } 3528 break; 3529 } 3530 #ifdef ASSERT 3531 case Op_ValueTypePtr: 3532 case Op_ValueType: { 3533 n->dump(-1); 3534 assert(false, "value type node was not removed"); 3535 break; 3536 } 3537 #endif 3538 default: 3539 assert(!n->is_Call(), ""); 3540 assert(!n->is_Mem(), ""); 3541 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3542 break; 3543 } 3544 } 3545 3546 //------------------------------final_graph_reshaping_walk--------------------- 3547 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3548 // requires that the walk visits a node's inputs before visiting the node. 3549 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3550 ResourceArea *area = Thread::current()->resource_area(); 3551 Unique_Node_List sfpt(area); 3552 3553 frc._visited.set(root->_idx); // first, mark node as visited 3554 uint cnt = root->req(); 3555 Node *n = root; 3556 uint i = 0; 3557 while (true) { 3558 if (i < cnt) { 3559 // Place all non-visited non-null inputs onto stack 3560 Node* m = n->in(i); 3561 ++i; 3562 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3563 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3564 // compute worst case interpreter size in case of a deoptimization 3565 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3566 3567 sfpt.push(m); 3568 } 3569 cnt = m->req(); 3570 nstack.push(n, i); // put on stack parent and next input's index 3571 n = m; 3572 i = 0; 3573 } 3574 } else { 3575 // Now do post-visit work 3576 final_graph_reshaping_impl( n, frc ); 3577 if (nstack.is_empty()) 3578 break; // finished 3579 n = nstack.node(); // Get node from stack 3580 cnt = n->req(); 3581 i = nstack.index(); 3582 nstack.pop(); // Shift to the next node on stack 3583 } 3584 } 3585 3586 // Skip next transformation if compressed oops are not used. 3587 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3588 (!UseCompressedOops && !UseCompressedClassPointers)) 3589 return; 3590 3591 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3592 // It could be done for an uncommon traps or any safepoints/calls 3593 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3594 while (sfpt.size() > 0) { 3595 n = sfpt.pop(); 3596 JVMState *jvms = n->as_SafePoint()->jvms(); 3597 assert(jvms != NULL, "sanity"); 3598 int start = jvms->debug_start(); 3599 int end = n->req(); 3600 bool is_uncommon = (n->is_CallStaticJava() && 3601 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3602 for (int j = start; j < end; j++) { 3603 Node* in = n->in(j); 3604 if (in->is_DecodeNarrowPtr()) { 3605 bool safe_to_skip = true; 3606 if (!is_uncommon ) { 3607 // Is it safe to skip? 3608 for (uint i = 0; i < in->outcnt(); i++) { 3609 Node* u = in->raw_out(i); 3610 if (!u->is_SafePoint() || 3611 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3612 safe_to_skip = false; 3613 } 3614 } 3615 } 3616 if (safe_to_skip) { 3617 n->set_req(j, in->in(1)); 3618 } 3619 if (in->outcnt() == 0) { 3620 in->disconnect_inputs(NULL, this); 3621 } 3622 } 3623 } 3624 } 3625 } 3626 3627 //------------------------------final_graph_reshaping-------------------------- 3628 // Final Graph Reshaping. 3629 // 3630 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3631 // and not commoned up and forced early. Must come after regular 3632 // optimizations to avoid GVN undoing the cloning. Clone constant 3633 // inputs to Loop Phis; these will be split by the allocator anyways. 3634 // Remove Opaque nodes. 3635 // (2) Move last-uses by commutative operations to the left input to encourage 3636 // Intel update-in-place two-address operations and better register usage 3637 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3638 // calls canonicalizing them back. 3639 // (3) Count the number of double-precision FP ops, single-precision FP ops 3640 // and call sites. On Intel, we can get correct rounding either by 3641 // forcing singles to memory (requires extra stores and loads after each 3642 // FP bytecode) or we can set a rounding mode bit (requires setting and 3643 // clearing the mode bit around call sites). The mode bit is only used 3644 // if the relative frequency of single FP ops to calls is low enough. 3645 // This is a key transform for SPEC mpeg_audio. 3646 // (4) Detect infinite loops; blobs of code reachable from above but not 3647 // below. Several of the Code_Gen algorithms fail on such code shapes, 3648 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3649 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3650 // Detection is by looking for IfNodes where only 1 projection is 3651 // reachable from below or CatchNodes missing some targets. 3652 // (5) Assert for insane oop offsets in debug mode. 3653 3654 bool Compile::final_graph_reshaping() { 3655 // an infinite loop may have been eliminated by the optimizer, 3656 // in which case the graph will be empty. 3657 if (root()->req() == 1) { 3658 record_method_not_compilable("trivial infinite loop"); 3659 return true; 3660 } 3661 3662 // Expensive nodes have their control input set to prevent the GVN 3663 // from freely commoning them. There's no GVN beyond this point so 3664 // no need to keep the control input. We want the expensive nodes to 3665 // be freely moved to the least frequent code path by gcm. 3666 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3667 for (int i = 0; i < expensive_count(); i++) { 3668 _expensive_nodes->at(i)->set_req(0, NULL); 3669 } 3670 3671 Final_Reshape_Counts frc; 3672 3673 // Visit everybody reachable! 3674 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3675 Node_Stack nstack(live_nodes() >> 1); 3676 final_graph_reshaping_walk(nstack, root(), frc); 3677 3678 // Check for unreachable (from below) code (i.e., infinite loops). 3679 for( uint i = 0; i < frc._tests.size(); i++ ) { 3680 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3681 // Get number of CFG targets. 3682 // Note that PCTables include exception targets after calls. 3683 uint required_outcnt = n->required_outcnt(); 3684 if (n->outcnt() != required_outcnt) { 3685 // Check for a few special cases. Rethrow Nodes never take the 3686 // 'fall-thru' path, so expected kids is 1 less. 3687 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3688 if (n->in(0)->in(0)->is_Call()) { 3689 CallNode *call = n->in(0)->in(0)->as_Call(); 3690 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3691 required_outcnt--; // Rethrow always has 1 less kid 3692 } else if (call->req() > TypeFunc::Parms && 3693 call->is_CallDynamicJava()) { 3694 // Check for null receiver. In such case, the optimizer has 3695 // detected that the virtual call will always result in a null 3696 // pointer exception. The fall-through projection of this CatchNode 3697 // will not be populated. 3698 Node *arg0 = call->in(TypeFunc::Parms); 3699 if (arg0->is_Type() && 3700 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3701 required_outcnt--; 3702 } 3703 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3704 call->req() > TypeFunc::Parms+1 && 3705 call->is_CallStaticJava()) { 3706 // Check for negative array length. In such case, the optimizer has 3707 // detected that the allocation attempt will always result in an 3708 // exception. There is no fall-through projection of this CatchNode . 3709 Node *arg1 = call->in(TypeFunc::Parms+1); 3710 if (arg1->is_Type() && 3711 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3712 required_outcnt--; 3713 } 3714 } 3715 } 3716 } 3717 // Recheck with a better notion of 'required_outcnt' 3718 if (n->outcnt() != required_outcnt) { 3719 record_method_not_compilable("malformed control flow"); 3720 return true; // Not all targets reachable! 3721 } 3722 } 3723 // Check that I actually visited all kids. Unreached kids 3724 // must be infinite loops. 3725 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3726 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3727 record_method_not_compilable("infinite loop"); 3728 return true; // Found unvisited kid; must be unreach 3729 } 3730 3731 // Here so verification code in final_graph_reshaping_walk() 3732 // always see an OuterStripMinedLoopEnd 3733 if (n->is_OuterStripMinedLoopEnd()) { 3734 IfNode* init_iff = n->as_If(); 3735 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3736 n->subsume_by(iff, this); 3737 } 3738 } 3739 3740 // If original bytecodes contained a mixture of floats and doubles 3741 // check if the optimizer has made it homogenous, item (3). 3742 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3743 frc.get_float_count() > 32 && 3744 frc.get_double_count() == 0 && 3745 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3746 set_24_bit_selection_and_mode( false, true ); 3747 } 3748 3749 set_java_calls(frc.get_java_call_count()); 3750 set_inner_loops(frc.get_inner_loop_count()); 3751 3752 // No infinite loops, no reason to bail out. 3753 return false; 3754 } 3755 3756 //-----------------------------too_many_traps---------------------------------- 3757 // Report if there are too many traps at the current method and bci. 3758 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3759 bool Compile::too_many_traps(ciMethod* method, 3760 int bci, 3761 Deoptimization::DeoptReason reason) { 3762 ciMethodData* md = method->method_data(); 3763 if (md->is_empty()) { 3764 // Assume the trap has not occurred, or that it occurred only 3765 // because of a transient condition during start-up in the interpreter. 3766 return false; 3767 } 3768 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3769 if (md->has_trap_at(bci, m, reason) != 0) { 3770 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3771 // Also, if there are multiple reasons, or if there is no per-BCI record, 3772 // assume the worst. 3773 if (log()) 3774 log()->elem("observe trap='%s' count='%d'", 3775 Deoptimization::trap_reason_name(reason), 3776 md->trap_count(reason)); 3777 return true; 3778 } else { 3779 // Ignore method/bci and see if there have been too many globally. 3780 return too_many_traps(reason, md); 3781 } 3782 } 3783 3784 // Less-accurate variant which does not require a method and bci. 3785 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3786 ciMethodData* logmd) { 3787 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3788 // Too many traps globally. 3789 // Note that we use cumulative trap_count, not just md->trap_count. 3790 if (log()) { 3791 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3792 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3793 Deoptimization::trap_reason_name(reason), 3794 mcount, trap_count(reason)); 3795 } 3796 return true; 3797 } else { 3798 // The coast is clear. 3799 return false; 3800 } 3801 } 3802 3803 //--------------------------too_many_recompiles-------------------------------- 3804 // Report if there are too many recompiles at the current method and bci. 3805 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3806 // Is not eager to return true, since this will cause the compiler to use 3807 // Action_none for a trap point, to avoid too many recompilations. 3808 bool Compile::too_many_recompiles(ciMethod* method, 3809 int bci, 3810 Deoptimization::DeoptReason reason) { 3811 ciMethodData* md = method->method_data(); 3812 if (md->is_empty()) { 3813 // Assume the trap has not occurred, or that it occurred only 3814 // because of a transient condition during start-up in the interpreter. 3815 return false; 3816 } 3817 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3818 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3819 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3820 Deoptimization::DeoptReason per_bc_reason 3821 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3822 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3823 if ((per_bc_reason == Deoptimization::Reason_none 3824 || md->has_trap_at(bci, m, reason) != 0) 3825 // The trap frequency measure we care about is the recompile count: 3826 && md->trap_recompiled_at(bci, m) 3827 && md->overflow_recompile_count() >= bc_cutoff) { 3828 // Do not emit a trap here if it has already caused recompilations. 3829 // Also, if there are multiple reasons, or if there is no per-BCI record, 3830 // assume the worst. 3831 if (log()) 3832 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3833 Deoptimization::trap_reason_name(reason), 3834 md->trap_count(reason), 3835 md->overflow_recompile_count()); 3836 return true; 3837 } else if (trap_count(reason) != 0 3838 && decompile_count() >= m_cutoff) { 3839 // Too many recompiles globally, and we have seen this sort of trap. 3840 // Use cumulative decompile_count, not just md->decompile_count. 3841 if (log()) 3842 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3843 Deoptimization::trap_reason_name(reason), 3844 md->trap_count(reason), trap_count(reason), 3845 md->decompile_count(), decompile_count()); 3846 return true; 3847 } else { 3848 // The coast is clear. 3849 return false; 3850 } 3851 } 3852 3853 // Compute when not to trap. Used by matching trap based nodes and 3854 // NullCheck optimization. 3855 void Compile::set_allowed_deopt_reasons() { 3856 _allowed_reasons = 0; 3857 if (is_method_compilation()) { 3858 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3859 assert(rs < BitsPerInt, "recode bit map"); 3860 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3861 _allowed_reasons |= nth_bit(rs); 3862 } 3863 } 3864 } 3865 } 3866 3867 #ifndef PRODUCT 3868 //------------------------------verify_graph_edges--------------------------- 3869 // Walk the Graph and verify that there is a one-to-one correspondence 3870 // between Use-Def edges and Def-Use edges in the graph. 3871 void Compile::verify_graph_edges(bool no_dead_code) { 3872 if (VerifyGraphEdges) { 3873 ResourceArea *area = Thread::current()->resource_area(); 3874 Unique_Node_List visited(area); 3875 // Call recursive graph walk to check edges 3876 _root->verify_edges(visited); 3877 if (no_dead_code) { 3878 // Now make sure that no visited node is used by an unvisited node. 3879 bool dead_nodes = false; 3880 Unique_Node_List checked(area); 3881 while (visited.size() > 0) { 3882 Node* n = visited.pop(); 3883 checked.push(n); 3884 for (uint i = 0; i < n->outcnt(); i++) { 3885 Node* use = n->raw_out(i); 3886 if (checked.member(use)) continue; // already checked 3887 if (visited.member(use)) continue; // already in the graph 3888 if (use->is_Con()) continue; // a dead ConNode is OK 3889 // At this point, we have found a dead node which is DU-reachable. 3890 if (!dead_nodes) { 3891 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3892 dead_nodes = true; 3893 } 3894 use->dump(2); 3895 tty->print_cr("---"); 3896 checked.push(use); // No repeats; pretend it is now checked. 3897 } 3898 } 3899 assert(!dead_nodes, "using nodes must be reachable from root"); 3900 } 3901 } 3902 } 3903 #endif 3904 3905 // The Compile object keeps track of failure reasons separately from the ciEnv. 3906 // This is required because there is not quite a 1-1 relation between the 3907 // ciEnv and its compilation task and the Compile object. Note that one 3908 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3909 // to backtrack and retry without subsuming loads. Other than this backtracking 3910 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3911 // by the logic in C2Compiler. 3912 void Compile::record_failure(const char* reason) { 3913 if (log() != NULL) { 3914 log()->elem("failure reason='%s' phase='compile'", reason); 3915 } 3916 if (_failure_reason == NULL) { 3917 // Record the first failure reason. 3918 _failure_reason = reason; 3919 } 3920 3921 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3922 C->print_method(PHASE_FAILURE); 3923 } 3924 _root = NULL; // flush the graph, too 3925 } 3926 3927 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3928 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3929 _phase_name(name), _dolog(CITimeVerbose) 3930 { 3931 if (_dolog) { 3932 C = Compile::current(); 3933 _log = C->log(); 3934 } else { 3935 C = NULL; 3936 _log = NULL; 3937 } 3938 if (_log != NULL) { 3939 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3940 _log->stamp(); 3941 _log->end_head(); 3942 } 3943 } 3944 3945 Compile::TracePhase::~TracePhase() { 3946 3947 C = Compile::current(); 3948 if (_dolog) { 3949 _log = C->log(); 3950 } else { 3951 _log = NULL; 3952 } 3953 3954 #ifdef ASSERT 3955 if (PrintIdealNodeCount) { 3956 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3957 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3958 } 3959 3960 if (VerifyIdealNodeCount) { 3961 Compile::current()->print_missing_nodes(); 3962 } 3963 #endif 3964 3965 if (_log != NULL) { 3966 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3967 } 3968 } 3969 3970 //============================================================================= 3971 // Two Constant's are equal when the type and the value are equal. 3972 bool Compile::Constant::operator==(const Constant& other) { 3973 if (type() != other.type() ) return false; 3974 if (can_be_reused() != other.can_be_reused()) return false; 3975 // For floating point values we compare the bit pattern. 3976 switch (type()) { 3977 case T_INT: 3978 case T_FLOAT: return (_v._value.i == other._v._value.i); 3979 case T_LONG: 3980 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3981 case T_OBJECT: 3982 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3983 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3984 case T_METADATA: return (_v._metadata == other._v._metadata); 3985 default: ShouldNotReachHere(); return false; 3986 } 3987 } 3988 3989 static int type_to_size_in_bytes(BasicType t) { 3990 switch (t) { 3991 case T_INT: return sizeof(jint ); 3992 case T_LONG: return sizeof(jlong ); 3993 case T_FLOAT: return sizeof(jfloat ); 3994 case T_DOUBLE: return sizeof(jdouble); 3995 case T_METADATA: return sizeof(Metadata*); 3996 // We use T_VOID as marker for jump-table entries (labels) which 3997 // need an internal word relocation. 3998 case T_VOID: 3999 case T_ADDRESS: 4000 case T_OBJECT: return sizeof(jobject); 4001 default: 4002 ShouldNotReachHere(); 4003 return -1; 4004 } 4005 } 4006 4007 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 4008 // sort descending 4009 if (a->freq() > b->freq()) return -1; 4010 if (a->freq() < b->freq()) return 1; 4011 return 0; 4012 } 4013 4014 void Compile::ConstantTable::calculate_offsets_and_size() { 4015 // First, sort the array by frequencies. 4016 _constants.sort(qsort_comparator); 4017 4018 #ifdef ASSERT 4019 // Make sure all jump-table entries were sorted to the end of the 4020 // array (they have a negative frequency). 4021 bool found_void = false; 4022 for (int i = 0; i < _constants.length(); i++) { 4023 Constant con = _constants.at(i); 4024 if (con.type() == T_VOID) 4025 found_void = true; // jump-tables 4026 else 4027 assert(!found_void, "wrong sorting"); 4028 } 4029 #endif 4030 4031 int offset = 0; 4032 for (int i = 0; i < _constants.length(); i++) { 4033 Constant* con = _constants.adr_at(i); 4034 4035 // Align offset for type. 4036 int typesize = type_to_size_in_bytes(con->type()); 4037 offset = align_up(offset, typesize); 4038 con->set_offset(offset); // set constant's offset 4039 4040 if (con->type() == T_VOID) { 4041 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4042 offset = offset + typesize * n->outcnt(); // expand jump-table 4043 } else { 4044 offset = offset + typesize; 4045 } 4046 } 4047 4048 // Align size up to the next section start (which is insts; see 4049 // CodeBuffer::align_at_start). 4050 assert(_size == -1, "already set?"); 4051 _size = align_up(offset, (int)CodeEntryAlignment); 4052 } 4053 4054 void Compile::ConstantTable::emit(CodeBuffer& cb) { 4055 MacroAssembler _masm(&cb); 4056 for (int i = 0; i < _constants.length(); i++) { 4057 Constant con = _constants.at(i); 4058 address constant_addr = NULL; 4059 switch (con.type()) { 4060 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4061 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4062 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4063 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4064 case T_OBJECT: { 4065 jobject obj = con.get_jobject(); 4066 int oop_index = _masm.oop_recorder()->find_index(obj); 4067 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4068 break; 4069 } 4070 case T_ADDRESS: { 4071 address addr = (address) con.get_jobject(); 4072 constant_addr = _masm.address_constant(addr); 4073 break; 4074 } 4075 // We use T_VOID as marker for jump-table entries (labels) which 4076 // need an internal word relocation. 4077 case T_VOID: { 4078 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4079 // Fill the jump-table with a dummy word. The real value is 4080 // filled in later in fill_jump_table. 4081 address dummy = (address) n; 4082 constant_addr = _masm.address_constant(dummy); 4083 // Expand jump-table 4084 for (uint i = 1; i < n->outcnt(); i++) { 4085 address temp_addr = _masm.address_constant(dummy + i); 4086 assert(temp_addr, "consts section too small"); 4087 } 4088 break; 4089 } 4090 case T_METADATA: { 4091 Metadata* obj = con.get_metadata(); 4092 int metadata_index = _masm.oop_recorder()->find_index(obj); 4093 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4094 break; 4095 } 4096 default: ShouldNotReachHere(); 4097 } 4098 assert(constant_addr, "consts section too small"); 4099 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4100 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4101 } 4102 } 4103 4104 int Compile::ConstantTable::find_offset(Constant& con) const { 4105 int idx = _constants.find(con); 4106 guarantee(idx != -1, "constant must be in constant table"); 4107 int offset = _constants.at(idx).offset(); 4108 guarantee(offset != -1, "constant table not emitted yet?"); 4109 return offset; 4110 } 4111 4112 void Compile::ConstantTable::add(Constant& con) { 4113 if (con.can_be_reused()) { 4114 int idx = _constants.find(con); 4115 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4116 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4117 return; 4118 } 4119 } 4120 (void) _constants.append(con); 4121 } 4122 4123 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4124 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4125 Constant con(type, value, b->_freq); 4126 add(con); 4127 return con; 4128 } 4129 4130 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4131 Constant con(metadata); 4132 add(con); 4133 return con; 4134 } 4135 4136 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4137 jvalue value; 4138 BasicType type = oper->type()->basic_type(); 4139 switch (type) { 4140 case T_LONG: value.j = oper->constantL(); break; 4141 case T_FLOAT: value.f = oper->constantF(); break; 4142 case T_DOUBLE: value.d = oper->constantD(); break; 4143 case T_OBJECT: 4144 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4145 case T_METADATA: return add((Metadata*)oper->constant()); break; 4146 default: guarantee(false, "unhandled type: %s", type2name(type)); 4147 } 4148 return add(n, type, value); 4149 } 4150 4151 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4152 jvalue value; 4153 // We can use the node pointer here to identify the right jump-table 4154 // as this method is called from Compile::Fill_buffer right before 4155 // the MachNodes are emitted and the jump-table is filled (means the 4156 // MachNode pointers do not change anymore). 4157 value.l = (jobject) n; 4158 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4159 add(con); 4160 return con; 4161 } 4162 4163 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4164 // If called from Compile::scratch_emit_size do nothing. 4165 if (Compile::current()->in_scratch_emit_size()) return; 4166 4167 assert(labels.is_nonempty(), "must be"); 4168 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4169 4170 // Since MachConstantNode::constant_offset() also contains 4171 // table_base_offset() we need to subtract the table_base_offset() 4172 // to get the plain offset into the constant table. 4173 int offset = n->constant_offset() - table_base_offset(); 4174 4175 MacroAssembler _masm(&cb); 4176 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4177 4178 for (uint i = 0; i < n->outcnt(); i++) { 4179 address* constant_addr = &jump_table_base[i]; 4180 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)); 4181 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4182 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4183 } 4184 } 4185 4186 //----------------------------static_subtype_check----------------------------- 4187 // Shortcut important common cases when superklass is exact: 4188 // (0) superklass is java.lang.Object (can occur in reflective code) 4189 // (1) subklass is already limited to a subtype of superklass => always ok 4190 // (2) subklass does not overlap with superklass => always fail 4191 // (3) superklass has NO subtypes and we can check with a simple compare. 4192 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4193 if (StressReflectiveCode || superk == NULL || subk == NULL) { 4194 return SSC_full_test; // Let caller generate the general case. 4195 } 4196 4197 if (superk == env()->Object_klass()) { 4198 return SSC_always_true; // (0) this test cannot fail 4199 } 4200 4201 ciType* superelem = superk; 4202 if (superelem->is_array_klass()) 4203 superelem = superelem->as_array_klass()->base_element_type(); 4204 4205 if (!subk->is_interface()) { // cannot trust static interface types yet 4206 if (subk->is_subtype_of(superk)) { 4207 return SSC_always_true; // (1) false path dead; no dynamic test needed 4208 } 4209 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4210 !superk->is_subtype_of(subk)) { 4211 return SSC_always_false; 4212 } 4213 } 4214 4215 // If casting to an instance klass, it must have no subtypes 4216 if (superk->is_interface()) { 4217 // Cannot trust interfaces yet. 4218 // %%% S.B. superk->nof_implementors() == 1 4219 } else if (superelem->is_instance_klass()) { 4220 ciInstanceKlass* ik = superelem->as_instance_klass(); 4221 if (!ik->has_subklass() && !ik->is_interface()) { 4222 if (!ik->is_final()) { 4223 // Add a dependency if there is a chance of a later subclass. 4224 dependencies()->assert_leaf_type(ik); 4225 } 4226 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4227 } 4228 } else { 4229 // A primitive array type has no subtypes. 4230 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4231 } 4232 4233 return SSC_full_test; 4234 } 4235 4236 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4237 #ifdef _LP64 4238 // The scaled index operand to AddP must be a clean 64-bit value. 4239 // Java allows a 32-bit int to be incremented to a negative 4240 // value, which appears in a 64-bit register as a large 4241 // positive number. Using that large positive number as an 4242 // operand in pointer arithmetic has bad consequences. 4243 // On the other hand, 32-bit overflow is rare, and the possibility 4244 // can often be excluded, if we annotate the ConvI2L node with 4245 // a type assertion that its value is known to be a small positive 4246 // number. (The prior range check has ensured this.) 4247 // This assertion is used by ConvI2LNode::Ideal. 4248 int index_max = max_jint - 1; // array size is max_jint, index is one less 4249 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4250 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4251 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4252 #endif 4253 return idx; 4254 } 4255 4256 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4257 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4258 if (ctrl != NULL) { 4259 // Express control dependency by a CastII node with a narrow type. 4260 value = new CastIINode(value, itype, false, true /* range check dependency */); 4261 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4262 // node from floating above the range check during loop optimizations. Otherwise, the 4263 // ConvI2L node may be eliminated independently of the range check, causing the data path 4264 // to become TOP while the control path is still there (although it's unreachable). 4265 value->set_req(0, ctrl); 4266 // Save CastII node to remove it after loop optimizations. 4267 phase->C->add_range_check_cast(value); 4268 value = phase->transform(value); 4269 } 4270 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4271 return phase->transform(new ConvI2LNode(value, ltype)); 4272 } 4273 4274 // The message about the current inlining is accumulated in 4275 // _print_inlining_stream and transfered into the _print_inlining_list 4276 // once we know whether inlining succeeds or not. For regular 4277 // inlining, messages are appended to the buffer pointed by 4278 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4279 // a new buffer is added after _print_inlining_idx in the list. This 4280 // way we can update the inlining message for late inlining call site 4281 // when the inlining is attempted again. 4282 void Compile::print_inlining_init() { 4283 if (print_inlining() || print_intrinsics()) { 4284 _print_inlining_stream = new stringStream(); 4285 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4286 } 4287 } 4288 4289 void Compile::print_inlining_reinit() { 4290 if (print_inlining() || print_intrinsics()) { 4291 // Re allocate buffer when we change ResourceMark 4292 _print_inlining_stream = new stringStream(); 4293 } 4294 } 4295 4296 void Compile::print_inlining_reset() { 4297 _print_inlining_stream->reset(); 4298 } 4299 4300 void Compile::print_inlining_commit() { 4301 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4302 // Transfer the message from _print_inlining_stream to the current 4303 // _print_inlining_list buffer and clear _print_inlining_stream. 4304 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4305 print_inlining_reset(); 4306 } 4307 4308 void Compile::print_inlining_push() { 4309 // Add new buffer to the _print_inlining_list at current position 4310 _print_inlining_idx++; 4311 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4312 } 4313 4314 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4315 return _print_inlining_list->at(_print_inlining_idx); 4316 } 4317 4318 void Compile::print_inlining_update(CallGenerator* cg) { 4319 if (print_inlining() || print_intrinsics()) { 4320 if (!cg->is_late_inline()) { 4321 if (print_inlining_current().cg() != NULL) { 4322 print_inlining_push(); 4323 } 4324 print_inlining_commit(); 4325 } else { 4326 if (print_inlining_current().cg() != cg && 4327 (print_inlining_current().cg() != NULL || 4328 print_inlining_current().ss()->size() != 0)) { 4329 print_inlining_push(); 4330 } 4331 print_inlining_commit(); 4332 print_inlining_current().set_cg(cg); 4333 } 4334 } 4335 } 4336 4337 void Compile::print_inlining_move_to(CallGenerator* cg) { 4338 // We resume inlining at a late inlining call site. Locate the 4339 // corresponding inlining buffer so that we can update it. 4340 if (print_inlining()) { 4341 for (int i = 0; i < _print_inlining_list->length(); i++) { 4342 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4343 _print_inlining_idx = i; 4344 return; 4345 } 4346 } 4347 ShouldNotReachHere(); 4348 } 4349 } 4350 4351 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4352 if (print_inlining()) { 4353 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4354 assert(print_inlining_current().cg() == cg, "wrong entry"); 4355 // replace message with new message 4356 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4357 print_inlining_commit(); 4358 print_inlining_current().set_cg(cg); 4359 } 4360 } 4361 4362 void Compile::print_inlining_assert_ready() { 4363 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4364 } 4365 4366 void Compile::process_print_inlining() { 4367 bool do_print_inlining = print_inlining() || print_intrinsics(); 4368 if (do_print_inlining || log() != NULL) { 4369 // Print inlining message for candidates that we couldn't inline 4370 // for lack of space 4371 for (int i = 0; i < _late_inlines.length(); i++) { 4372 CallGenerator* cg = _late_inlines.at(i); 4373 if (!cg->is_mh_late_inline()) { 4374 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4375 if (do_print_inlining) { 4376 cg->print_inlining_late(msg); 4377 } 4378 log_late_inline_failure(cg, msg); 4379 } 4380 } 4381 } 4382 if (do_print_inlining) { 4383 ResourceMark rm; 4384 stringStream ss; 4385 for (int i = 0; i < _print_inlining_list->length(); i++) { 4386 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4387 } 4388 size_t end = ss.size(); 4389 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4390 strncpy(_print_inlining_output, ss.base(), end+1); 4391 _print_inlining_output[end] = 0; 4392 } 4393 } 4394 4395 void Compile::dump_print_inlining() { 4396 if (_print_inlining_output != NULL) { 4397 tty->print_raw(_print_inlining_output); 4398 } 4399 } 4400 4401 void Compile::log_late_inline(CallGenerator* cg) { 4402 if (log() != NULL) { 4403 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4404 cg->unique_id()); 4405 JVMState* p = cg->call_node()->jvms(); 4406 while (p != NULL) { 4407 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4408 p = p->caller(); 4409 } 4410 log()->tail("late_inline"); 4411 } 4412 } 4413 4414 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4415 log_late_inline(cg); 4416 if (log() != NULL) { 4417 log()->inline_fail(msg); 4418 } 4419 } 4420 4421 void Compile::log_inline_id(CallGenerator* cg) { 4422 if (log() != NULL) { 4423 // The LogCompilation tool needs a unique way to identify late 4424 // inline call sites. This id must be unique for this call site in 4425 // this compilation. Try to have it unique across compilations as 4426 // well because it can be convenient when grepping through the log 4427 // file. 4428 // Distinguish OSR compilations from others in case CICountOSR is 4429 // on. 4430 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4431 cg->set_unique_id(id); 4432 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4433 } 4434 } 4435 4436 void Compile::log_inline_failure(const char* msg) { 4437 if (C->log() != NULL) { 4438 C->log()->inline_fail(msg); 4439 } 4440 } 4441 4442 4443 // Dump inlining replay data to the stream. 4444 // Don't change thread state and acquire any locks. 4445 void Compile::dump_inline_data(outputStream* out) { 4446 InlineTree* inl_tree = ilt(); 4447 if (inl_tree != NULL) { 4448 out->print(" inline %d", inl_tree->count()); 4449 inl_tree->dump_replay_data(out); 4450 } 4451 } 4452 4453 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4454 if (n1->Opcode() < n2->Opcode()) return -1; 4455 else if (n1->Opcode() > n2->Opcode()) return 1; 4456 4457 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4458 for (uint i = 1; i < n1->req(); i++) { 4459 if (n1->in(i) < n2->in(i)) return -1; 4460 else if (n1->in(i) > n2->in(i)) return 1; 4461 } 4462 4463 return 0; 4464 } 4465 4466 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4467 Node* n1 = *n1p; 4468 Node* n2 = *n2p; 4469 4470 return cmp_expensive_nodes(n1, n2); 4471 } 4472 4473 void Compile::sort_expensive_nodes() { 4474 if (!expensive_nodes_sorted()) { 4475 _expensive_nodes->sort(cmp_expensive_nodes); 4476 } 4477 } 4478 4479 bool Compile::expensive_nodes_sorted() const { 4480 for (int i = 1; i < _expensive_nodes->length(); i++) { 4481 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4482 return false; 4483 } 4484 } 4485 return true; 4486 } 4487 4488 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4489 if (_expensive_nodes->length() == 0) { 4490 return false; 4491 } 4492 4493 assert(OptimizeExpensiveOps, "optimization off?"); 4494 4495 // Take this opportunity to remove dead nodes from the list 4496 int j = 0; 4497 for (int i = 0; i < _expensive_nodes->length(); i++) { 4498 Node* n = _expensive_nodes->at(i); 4499 if (!n->is_unreachable(igvn)) { 4500 assert(n->is_expensive(), "should be expensive"); 4501 _expensive_nodes->at_put(j, n); 4502 j++; 4503 } 4504 } 4505 _expensive_nodes->trunc_to(j); 4506 4507 // Then sort the list so that similar nodes are next to each other 4508 // and check for at least two nodes of identical kind with same data 4509 // inputs. 4510 sort_expensive_nodes(); 4511 4512 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4513 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4514 return true; 4515 } 4516 } 4517 4518 return false; 4519 } 4520 4521 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4522 if (_expensive_nodes->length() == 0) { 4523 return; 4524 } 4525 4526 assert(OptimizeExpensiveOps, "optimization off?"); 4527 4528 // Sort to bring similar nodes next to each other and clear the 4529 // control input of nodes for which there's only a single copy. 4530 sort_expensive_nodes(); 4531 4532 int j = 0; 4533 int identical = 0; 4534 int i = 0; 4535 bool modified = false; 4536 for (; i < _expensive_nodes->length()-1; i++) { 4537 assert(j <= i, "can't write beyond current index"); 4538 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4539 identical++; 4540 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4541 continue; 4542 } 4543 if (identical > 0) { 4544 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4545 identical = 0; 4546 } else { 4547 Node* n = _expensive_nodes->at(i); 4548 igvn.replace_input_of(n, 0, NULL); 4549 igvn.hash_insert(n); 4550 modified = true; 4551 } 4552 } 4553 if (identical > 0) { 4554 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4555 } else if (_expensive_nodes->length() >= 1) { 4556 Node* n = _expensive_nodes->at(i); 4557 igvn.replace_input_of(n, 0, NULL); 4558 igvn.hash_insert(n); 4559 modified = true; 4560 } 4561 _expensive_nodes->trunc_to(j); 4562 if (modified) { 4563 igvn.optimize(); 4564 } 4565 } 4566 4567 void Compile::add_expensive_node(Node * n) { 4568 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4569 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4570 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4571 if (OptimizeExpensiveOps) { 4572 _expensive_nodes->append(n); 4573 } else { 4574 // Clear control input and let IGVN optimize expensive nodes if 4575 // OptimizeExpensiveOps is off. 4576 n->set_req(0, NULL); 4577 } 4578 } 4579 4580 /** 4581 * Remove the speculative part of types and clean up the graph 4582 */ 4583 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4584 if (UseTypeSpeculation) { 4585 Unique_Node_List worklist; 4586 worklist.push(root()); 4587 int modified = 0; 4588 // Go over all type nodes that carry a speculative type, drop the 4589 // speculative part of the type and enqueue the node for an igvn 4590 // which may optimize it out. 4591 for (uint next = 0; next < worklist.size(); ++next) { 4592 Node *n = worklist.at(next); 4593 if (n->is_Type()) { 4594 TypeNode* tn = n->as_Type(); 4595 const Type* t = tn->type(); 4596 const Type* t_no_spec = t->remove_speculative(); 4597 if (t_no_spec != t) { 4598 bool in_hash = igvn.hash_delete(n); 4599 assert(in_hash, "node should be in igvn hash table"); 4600 tn->set_type(t_no_spec); 4601 igvn.hash_insert(n); 4602 igvn._worklist.push(n); // give it a chance to go away 4603 modified++; 4604 } 4605 } 4606 uint max = n->len(); 4607 for( uint i = 0; i < max; ++i ) { 4608 Node *m = n->in(i); 4609 if (not_a_node(m)) continue; 4610 worklist.push(m); 4611 } 4612 } 4613 // Drop the speculative part of all types in the igvn's type table 4614 igvn.remove_speculative_types(); 4615 if (modified > 0) { 4616 igvn.optimize(); 4617 } 4618 #ifdef ASSERT 4619 // Verify that after the IGVN is over no speculative type has resurfaced 4620 worklist.clear(); 4621 worklist.push(root()); 4622 for (uint next = 0; next < worklist.size(); ++next) { 4623 Node *n = worklist.at(next); 4624 const Type* t = igvn.type_or_null(n); 4625 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4626 if (n->is_Type()) { 4627 t = n->as_Type()->type(); 4628 assert(t == t->remove_speculative(), "no more speculative types"); 4629 } 4630 uint max = n->len(); 4631 for( uint i = 0; i < max; ++i ) { 4632 Node *m = n->in(i); 4633 if (not_a_node(m)) continue; 4634 worklist.push(m); 4635 } 4636 } 4637 igvn.check_no_speculative_types(); 4638 #endif 4639 } 4640 } 4641 4642 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) { 4643 const TypeInstPtr* ta = phase->type(a)->isa_instptr(); 4644 const TypeInstPtr* tb = phase->type(b)->isa_instptr(); 4645 if (!EnableValhalla || ta == NULL || tb == NULL || 4646 ta->is_zero_type() || tb->is_zero_type() || 4647 !ta->can_be_value_type() || !tb->can_be_value_type()) { 4648 // Use old acmp if one operand is null or not a value type 4649 return new CmpPNode(a, b); 4650 } else if (ta->is_valuetypeptr() || tb->is_valuetypeptr()) { 4651 // We know that one operand is a value type. Therefore, 4652 // new acmp will only return true if both operands are NULL. 4653 // Check if both operands are null by or'ing the oops. 4654 a = phase->transform(new CastP2XNode(NULL, a)); 4655 b = phase->transform(new CastP2XNode(NULL, b)); 4656 a = phase->transform(new OrXNode(a, b)); 4657 return new CmpXNode(a, phase->MakeConX(0)); 4658 } 4659 // Use new acmp 4660 return NULL; 4661 } 4662 4663 // Auxiliary method to support randomized stressing/fuzzing. 4664 // 4665 // This method can be called the arbitrary number of times, with current count 4666 // as the argument. The logic allows selecting a single candidate from the 4667 // running list of candidates as follows: 4668 // int count = 0; 4669 // Cand* selected = null; 4670 // while(cand = cand->next()) { 4671 // if (randomized_select(++count)) { 4672 // selected = cand; 4673 // } 4674 // } 4675 // 4676 // Including count equalizes the chances any candidate is "selected". 4677 // This is useful when we don't have the complete list of candidates to choose 4678 // from uniformly. In this case, we need to adjust the randomicity of the 4679 // selection, or else we will end up biasing the selection towards the latter 4680 // candidates. 4681 // 4682 // Quick back-envelope calculation shows that for the list of n candidates 4683 // the equal probability for the candidate to persist as "best" can be 4684 // achieved by replacing it with "next" k-th candidate with the probability 4685 // of 1/k. It can be easily shown that by the end of the run, the 4686 // probability for any candidate is converged to 1/n, thus giving the 4687 // uniform distribution among all the candidates. 4688 // 4689 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4690 #define RANDOMIZED_DOMAIN_POW 29 4691 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4692 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4693 bool Compile::randomized_select(int count) { 4694 assert(count > 0, "only positive"); 4695 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4696 } 4697 4698 CloneMap& Compile::clone_map() { return _clone_map; } 4699 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4700 4701 void NodeCloneInfo::dump() const { 4702 tty->print(" {%d:%d} ", idx(), gen()); 4703 } 4704 4705 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4706 uint64_t val = value(old->_idx); 4707 NodeCloneInfo cio(val); 4708 assert(val != 0, "old node should be in the map"); 4709 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4710 insert(nnn->_idx, cin.get()); 4711 #ifndef PRODUCT 4712 if (is_debug()) { 4713 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4714 } 4715 #endif 4716 } 4717 4718 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4719 NodeCloneInfo cio(value(old->_idx)); 4720 if (cio.get() == 0) { 4721 cio.set(old->_idx, 0); 4722 insert(old->_idx, cio.get()); 4723 #ifndef PRODUCT 4724 if (is_debug()) { 4725 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4726 } 4727 #endif 4728 } 4729 clone(old, nnn, gen); 4730 } 4731 4732 int CloneMap::max_gen() const { 4733 int g = 0; 4734 DictI di(_dict); 4735 for(; di.test(); ++di) { 4736 int t = gen(di._key); 4737 if (g < t) { 4738 g = t; 4739 #ifndef PRODUCT 4740 if (is_debug()) { 4741 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4742 } 4743 #endif 4744 } 4745 } 4746 return g; 4747 } 4748 4749 void CloneMap::dump(node_idx_t key) const { 4750 uint64_t val = value(key); 4751 if (val != 0) { 4752 NodeCloneInfo ni(val); 4753 ni.dump(); 4754 } 4755 }