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