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