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