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