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