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