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("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 { 1712 ResourceMark rm; 1713 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1714 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1715 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1716 Type::str(adr_type)); 1717 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1718 const TypeOopPtr* foop = flat->is_oopptr(); 1719 // Scalarizable allocations have exact klass always. 1720 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1721 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1722 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1723 Type::str(foop), Type::str(xoop)); 1724 } 1725 } 1726 #endif 1727 1728 int idx = AliasIdxTop; 1729 for (int i = 0; i < num_alias_types(); i++) { 1730 if (alias_type(i)->adr_type() == flat) { 1731 idx = i; 1732 break; 1733 } 1734 } 1735 1736 if (idx == AliasIdxTop) { 1737 if (no_create) return NULL; 1738 // Grow the array if necessary. 1739 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1740 // Add a new alias type. 1741 idx = _num_alias_types++; 1742 _alias_types[idx]->Init(idx, flat); 1743 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1744 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1745 if (flat->isa_instptr()) { 1746 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1747 && flat->is_instptr()->klass() == env()->Class_klass()) 1748 alias_type(idx)->set_rewritable(false); 1749 } 1750 if (flat->isa_aryptr()) { 1751 #ifdef ASSERT 1752 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1753 // (T_BYTE has the weakest alignment and size restrictions...) 1754 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1755 #endif 1756 if (flat->offset() == TypePtr::OffsetBot) { 1757 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1758 } 1759 } 1760 if (flat->isa_klassptr()) { 1761 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1762 alias_type(idx)->set_rewritable(false); 1763 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1764 alias_type(idx)->set_rewritable(false); 1765 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1766 alias_type(idx)->set_rewritable(false); 1767 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1768 alias_type(idx)->set_rewritable(false); 1769 } 1770 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1771 // but the base pointer type is not distinctive enough to identify 1772 // references into JavaThread.) 1773 1774 // Check for final fields. 1775 const TypeInstPtr* tinst = flat->isa_instptr(); 1776 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1777 ciField* field; 1778 if (tinst->const_oop() != NULL && 1779 tinst->klass() == ciEnv::current()->Class_klass() && 1780 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1781 // static field 1782 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1783 field = k->get_field_by_offset(tinst->offset(), true); 1784 } else { 1785 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1786 field = k->get_field_by_offset(tinst->offset(), false); 1787 } 1788 assert(field == NULL || 1789 original_field == NULL || 1790 (field->holder() == original_field->holder() && 1791 field->offset() == original_field->offset() && 1792 field->is_static() == original_field->is_static()), "wrong field?"); 1793 // Set field() and is_rewritable() attributes. 1794 if (field != NULL) alias_type(idx)->set_field(field); 1795 } 1796 } 1797 1798 // Fill the cache for next time. 1799 ace->_adr_type = adr_type; 1800 ace->_index = idx; 1801 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1802 1803 // Might as well try to fill the cache for the flattened version, too. 1804 AliasCacheEntry* face = probe_alias_cache(flat); 1805 if (face->_adr_type == NULL) { 1806 face->_adr_type = flat; 1807 face->_index = idx; 1808 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1809 } 1810 1811 return alias_type(idx); 1812 } 1813 1814 1815 Compile::AliasType* Compile::alias_type(ciField* field) { 1816 const TypeOopPtr* t; 1817 if (field->is_static()) 1818 t = TypeInstPtr::make(field->holder()->java_mirror()); 1819 else 1820 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1821 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1822 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1823 return atp; 1824 } 1825 1826 1827 //------------------------------have_alias_type-------------------------------- 1828 bool Compile::have_alias_type(const TypePtr* adr_type) { 1829 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1830 if (ace->_adr_type == adr_type) { 1831 return true; 1832 } 1833 1834 // Handle special cases. 1835 if (adr_type == NULL) return true; 1836 if (adr_type == TypePtr::BOTTOM) return true; 1837 1838 return find_alias_type(adr_type, true, NULL) != NULL; 1839 } 1840 1841 //-----------------------------must_alias-------------------------------------- 1842 // True if all values of the given address type are in the given alias category. 1843 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1844 if (alias_idx == AliasIdxBot) return true; // the universal category 1845 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1846 if (alias_idx == AliasIdxTop) return false; // the empty category 1847 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1848 1849 // the only remaining possible overlap is identity 1850 int adr_idx = get_alias_index(adr_type); 1851 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1852 assert(adr_idx == alias_idx || 1853 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1854 && adr_type != TypeOopPtr::BOTTOM), 1855 "should not be testing for overlap with an unsafe pointer"); 1856 return adr_idx == alias_idx; 1857 } 1858 1859 //------------------------------can_alias-------------------------------------- 1860 // True if any values of the given address type are in the given alias category. 1861 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1862 if (alias_idx == AliasIdxTop) return false; // the empty category 1863 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1864 if (alias_idx == AliasIdxBot) return true; // the universal category 1865 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1866 1867 // the only remaining possible overlap is identity 1868 int adr_idx = get_alias_index(adr_type); 1869 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1870 return adr_idx == alias_idx; 1871 } 1872 1873 1874 1875 //---------------------------pop_warm_call------------------------------------- 1876 WarmCallInfo* Compile::pop_warm_call() { 1877 WarmCallInfo* wci = _warm_calls; 1878 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1879 return wci; 1880 } 1881 1882 //----------------------------Inline_Warm-------------------------------------- 1883 int Compile::Inline_Warm() { 1884 // If there is room, try to inline some more warm call sites. 1885 // %%% Do a graph index compaction pass when we think we're out of space? 1886 if (!InlineWarmCalls) return 0; 1887 1888 int calls_made_hot = 0; 1889 int room_to_grow = NodeCountInliningCutoff - unique(); 1890 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1891 int amount_grown = 0; 1892 WarmCallInfo* call; 1893 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1894 int est_size = (int)call->size(); 1895 if (est_size > (room_to_grow - amount_grown)) { 1896 // This one won't fit anyway. Get rid of it. 1897 call->make_cold(); 1898 continue; 1899 } 1900 call->make_hot(); 1901 calls_made_hot++; 1902 amount_grown += est_size; 1903 amount_to_grow -= est_size; 1904 } 1905 1906 if (calls_made_hot > 0) set_major_progress(); 1907 return calls_made_hot; 1908 } 1909 1910 1911 //----------------------------Finish_Warm-------------------------------------- 1912 void Compile::Finish_Warm() { 1913 if (!InlineWarmCalls) return; 1914 if (failing()) return; 1915 if (warm_calls() == NULL) return; 1916 1917 // Clean up loose ends, if we are out of space for inlining. 1918 WarmCallInfo* call; 1919 while ((call = pop_warm_call()) != NULL) { 1920 call->make_cold(); 1921 } 1922 } 1923 1924 //---------------------cleanup_loop_predicates----------------------- 1925 // Remove the opaque nodes that protect the predicates so that all unused 1926 // checks and uncommon_traps will be eliminated from the ideal graph 1927 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1928 if (predicate_count()==0) return; 1929 for (int i = predicate_count(); i > 0; i--) { 1930 Node * n = predicate_opaque1_node(i-1); 1931 assert(n->Opcode() == Op_Opaque1, "must be"); 1932 igvn.replace_node(n, n->in(1)); 1933 } 1934 assert(predicate_count()==0, "should be clean!"); 1935 } 1936 1937 void Compile::add_range_check_cast(Node* n) { 1938 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1939 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1940 _range_check_casts->append(n); 1941 } 1942 1943 // Remove all range check dependent CastIINodes. 1944 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1945 for (int i = range_check_cast_count(); i > 0; i--) { 1946 Node* cast = range_check_cast_node(i-1); 1947 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1948 igvn.replace_node(cast, cast->in(1)); 1949 } 1950 assert(range_check_cast_count() == 0, "should be empty"); 1951 } 1952 1953 // StringOpts and late inlining of string methods 1954 void Compile::inline_string_calls(bool parse_time) { 1955 { 1956 // remove useless nodes to make the usage analysis simpler 1957 ResourceMark rm; 1958 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1959 } 1960 1961 { 1962 ResourceMark rm; 1963 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1964 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1965 print_method(PHASE_AFTER_STRINGOPTS, 3); 1966 } 1967 1968 // now inline anything that we skipped the first time around 1969 if (!parse_time) { 1970 _late_inlines_pos = _late_inlines.length(); 1971 } 1972 1973 while (_string_late_inlines.length() > 0) { 1974 CallGenerator* cg = _string_late_inlines.pop(); 1975 cg->do_late_inline(); 1976 if (failing()) return; 1977 } 1978 _string_late_inlines.trunc_to(0); 1979 } 1980 1981 // Late inlining of boxing methods 1982 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1983 if (_boxing_late_inlines.length() > 0) { 1984 assert(has_boxed_value(), "inconsistent"); 1985 1986 PhaseGVN* gvn = initial_gvn(); 1987 set_inlining_incrementally(true); 1988 1989 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1990 for_igvn()->clear(); 1991 gvn->replace_with(&igvn); 1992 1993 _late_inlines_pos = _late_inlines.length(); 1994 1995 while (_boxing_late_inlines.length() > 0) { 1996 CallGenerator* cg = _boxing_late_inlines.pop(); 1997 cg->do_late_inline(); 1998 if (failing()) return; 1999 } 2000 _boxing_late_inlines.trunc_to(0); 2001 2002 { 2003 ResourceMark rm; 2004 PhaseRemoveUseless pru(gvn, for_igvn()); 2005 } 2006 2007 igvn = PhaseIterGVN(gvn); 2008 igvn.optimize(); 2009 2010 set_inlining_progress(false); 2011 set_inlining_incrementally(false); 2012 } 2013 } 2014 2015 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 2016 assert(IncrementalInline, "incremental inlining should be on"); 2017 PhaseGVN* gvn = initial_gvn(); 2018 2019 set_inlining_progress(false); 2020 for_igvn()->clear(); 2021 gvn->replace_with(&igvn); 2022 2023 { 2024 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2025 int i = 0; 2026 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2027 CallGenerator* cg = _late_inlines.at(i); 2028 _late_inlines_pos = i+1; 2029 cg->do_late_inline(); 2030 if (failing()) return; 2031 } 2032 int j = 0; 2033 for (; i < _late_inlines.length(); i++, j++) { 2034 _late_inlines.at_put(j, _late_inlines.at(i)); 2035 } 2036 _late_inlines.trunc_to(j); 2037 } 2038 2039 { 2040 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2041 ResourceMark rm; 2042 PhaseRemoveUseless pru(gvn, for_igvn()); 2043 } 2044 2045 { 2046 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2047 igvn = PhaseIterGVN(gvn); 2048 } 2049 } 2050 2051 // Perform incremental inlining until bound on number of live nodes is reached 2052 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2053 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2054 2055 PhaseGVN* gvn = initial_gvn(); 2056 2057 set_inlining_incrementally(true); 2058 set_inlining_progress(true); 2059 uint low_live_nodes = 0; 2060 2061 while(inlining_progress() && _late_inlines.length() > 0) { 2062 2063 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2064 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2065 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2066 // PhaseIdealLoop is expensive so we only try it once we are 2067 // out of live nodes and we only try it again if the previous 2068 // helped got the number of nodes down significantly 2069 PhaseIdealLoop ideal_loop( igvn, false, true ); 2070 if (failing()) return; 2071 low_live_nodes = live_nodes(); 2072 _major_progress = true; 2073 } 2074 2075 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2076 break; 2077 } 2078 } 2079 2080 inline_incrementally_one(igvn); 2081 2082 if (failing()) return; 2083 2084 { 2085 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2086 igvn.optimize(); 2087 } 2088 2089 if (failing()) return; 2090 } 2091 2092 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2093 2094 if (_string_late_inlines.length() > 0) { 2095 assert(has_stringbuilder(), "inconsistent"); 2096 for_igvn()->clear(); 2097 initial_gvn()->replace_with(&igvn); 2098 2099 inline_string_calls(false); 2100 2101 if (failing()) return; 2102 2103 { 2104 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2105 ResourceMark rm; 2106 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2107 } 2108 2109 { 2110 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2111 igvn = PhaseIterGVN(gvn); 2112 igvn.optimize(); 2113 } 2114 } 2115 2116 set_inlining_incrementally(false); 2117 } 2118 2119 2120 //------------------------------Optimize--------------------------------------- 2121 // Given a graph, optimize it. 2122 void Compile::Optimize() { 2123 TracePhase tp("optimizer", &timers[_t_optimizer]); 2124 2125 #ifndef PRODUCT 2126 if (_directive->BreakAtCompileOption) { 2127 BREAKPOINT; 2128 } 2129 2130 #endif 2131 2132 ResourceMark rm; 2133 int loop_opts_cnt; 2134 2135 print_inlining_reinit(); 2136 2137 NOT_PRODUCT( verify_graph_edges(); ) 2138 2139 print_method(PHASE_AFTER_PARSING); 2140 2141 { 2142 // Iterative Global Value Numbering, including ideal transforms 2143 // Initialize IterGVN with types and values from parse-time GVN 2144 PhaseIterGVN igvn(initial_gvn()); 2145 #ifdef ASSERT 2146 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2147 #endif 2148 { 2149 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2150 igvn.optimize(); 2151 } 2152 2153 print_method(PHASE_ITER_GVN1, 2); 2154 2155 if (failing()) return; 2156 2157 inline_incrementally(igvn); 2158 2159 print_method(PHASE_INCREMENTAL_INLINE, 2); 2160 2161 if (failing()) return; 2162 2163 if (eliminate_boxing()) { 2164 // Inline valueOf() methods now. 2165 inline_boxing_calls(igvn); 2166 2167 if (AlwaysIncrementalInline) { 2168 inline_incrementally(igvn); 2169 } 2170 2171 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2172 2173 if (failing()) return; 2174 } 2175 2176 // Remove the speculative part of types and clean up the graph from 2177 // the extra CastPP nodes whose only purpose is to carry them. Do 2178 // that early so that optimizations are not disrupted by the extra 2179 // CastPP nodes. 2180 remove_speculative_types(igvn); 2181 2182 // No more new expensive nodes will be added to the list from here 2183 // so keep only the actual candidates for optimizations. 2184 cleanup_expensive_nodes(igvn); 2185 2186 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2187 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2188 initial_gvn()->replace_with(&igvn); 2189 for_igvn()->clear(); 2190 Unique_Node_List new_worklist(C->comp_arena()); 2191 { 2192 ResourceMark rm; 2193 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2194 } 2195 set_for_igvn(&new_worklist); 2196 igvn = PhaseIterGVN(initial_gvn()); 2197 igvn.optimize(); 2198 } 2199 2200 // Perform escape analysis 2201 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2202 if (has_loops()) { 2203 // Cleanup graph (remove dead nodes). 2204 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2205 PhaseIdealLoop ideal_loop( igvn, false, true ); 2206 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2207 if (failing()) return; 2208 } 2209 ConnectionGraph::do_analysis(this, &igvn); 2210 2211 if (failing()) return; 2212 2213 // Optimize out fields loads from scalar replaceable allocations. 2214 igvn.optimize(); 2215 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2216 2217 if (failing()) return; 2218 2219 if (congraph() != NULL && macro_count() > 0) { 2220 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2221 PhaseMacroExpand mexp(igvn); 2222 mexp.eliminate_macro_nodes(); 2223 igvn.set_delay_transform(false); 2224 2225 igvn.optimize(); 2226 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2227 2228 if (failing()) return; 2229 } 2230 } 2231 2232 // Loop transforms on the ideal graph. Range Check Elimination, 2233 // peeling, unrolling, etc. 2234 2235 // Set loop opts counter 2236 loop_opts_cnt = num_loop_opts(); 2237 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2238 { 2239 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2240 PhaseIdealLoop ideal_loop( igvn, true ); 2241 loop_opts_cnt--; 2242 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2243 if (failing()) return; 2244 } 2245 // Loop opts pass if partial peeling occurred in previous pass 2246 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2247 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2248 PhaseIdealLoop ideal_loop( igvn, false ); 2249 loop_opts_cnt--; 2250 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2251 if (failing()) return; 2252 } 2253 // Loop opts pass for loop-unrolling before CCP 2254 if(major_progress() && (loop_opts_cnt > 0)) { 2255 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2256 PhaseIdealLoop ideal_loop( igvn, false ); 2257 loop_opts_cnt--; 2258 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2259 } 2260 if (!failing()) { 2261 // Verify that last round of loop opts produced a valid graph 2262 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2263 PhaseIdealLoop::verify(igvn); 2264 } 2265 } 2266 if (failing()) return; 2267 2268 // Conditional Constant Propagation; 2269 PhaseCCP ccp( &igvn ); 2270 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2271 { 2272 TracePhase tp("ccp", &timers[_t_ccp]); 2273 ccp.do_transform(); 2274 } 2275 print_method(PHASE_CPP1, 2); 2276 2277 assert( true, "Break here to ccp.dump_old2new_map()"); 2278 2279 // Iterative Global Value Numbering, including ideal transforms 2280 { 2281 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2282 igvn = ccp; 2283 igvn.optimize(); 2284 } 2285 2286 print_method(PHASE_ITER_GVN2, 2); 2287 2288 if (failing()) return; 2289 2290 // Loop transforms on the ideal graph. Range Check Elimination, 2291 // peeling, unrolling, etc. 2292 if(loop_opts_cnt > 0) { 2293 debug_only( int cnt = 0; ); 2294 while(major_progress() && (loop_opts_cnt > 0)) { 2295 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2296 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2297 PhaseIdealLoop ideal_loop( igvn, true); 2298 loop_opts_cnt--; 2299 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2300 if (failing()) return; 2301 } 2302 } 2303 // Ensure that major progress is now clear 2304 C->clear_major_progress(); 2305 2306 { 2307 // Verify that all previous optimizations produced a valid graph 2308 // at least to this point, even if no loop optimizations were done. 2309 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2310 PhaseIdealLoop::verify(igvn); 2311 } 2312 2313 if (range_check_cast_count() > 0) { 2314 // No more loop optimizations. Remove all range check dependent CastIINodes. 2315 C->remove_range_check_casts(igvn); 2316 igvn.optimize(); 2317 } 2318 2319 { 2320 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2321 PhaseMacroExpand mex(igvn); 2322 if (mex.expand_macro_nodes()) { 2323 assert(failing(), "must bail out w/ explicit message"); 2324 return; 2325 } 2326 } 2327 2328 DEBUG_ONLY( _modified_nodes = NULL; ) 2329 } // (End scope of igvn; run destructor if necessary for asserts.) 2330 2331 process_print_inlining(); 2332 // A method with only infinite loops has no edges entering loops from root 2333 { 2334 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2335 if (final_graph_reshaping()) { 2336 assert(failing(), "must bail out w/ explicit message"); 2337 return; 2338 } 2339 } 2340 2341 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2342 } 2343 2344 2345 //------------------------------Code_Gen--------------------------------------- 2346 // Given a graph, generate code for it 2347 void Compile::Code_Gen() { 2348 if (failing()) { 2349 return; 2350 } 2351 2352 // Perform instruction selection. You might think we could reclaim Matcher 2353 // memory PDQ, but actually the Matcher is used in generating spill code. 2354 // Internals of the Matcher (including some VectorSets) must remain live 2355 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2356 // set a bit in reclaimed memory. 2357 2358 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2359 // nodes. Mapping is only valid at the root of each matched subtree. 2360 NOT_PRODUCT( verify_graph_edges(); ) 2361 2362 Matcher matcher; 2363 _matcher = &matcher; 2364 { 2365 TracePhase tp("matcher", &timers[_t_matcher]); 2366 matcher.match(); 2367 } 2368 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2369 // nodes. Mapping is only valid at the root of each matched subtree. 2370 NOT_PRODUCT( verify_graph_edges(); ) 2371 2372 // If you have too many nodes, or if matching has failed, bail out 2373 check_node_count(0, "out of nodes matching instructions"); 2374 if (failing()) { 2375 return; 2376 } 2377 2378 // Build a proper-looking CFG 2379 PhaseCFG cfg(node_arena(), root(), matcher); 2380 _cfg = &cfg; 2381 { 2382 TracePhase tp("scheduler", &timers[_t_scheduler]); 2383 bool success = cfg.do_global_code_motion(); 2384 if (!success) { 2385 return; 2386 } 2387 2388 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2389 NOT_PRODUCT( verify_graph_edges(); ) 2390 debug_only( cfg.verify(); ) 2391 } 2392 2393 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2394 _regalloc = ®alloc; 2395 { 2396 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2397 // Perform register allocation. After Chaitin, use-def chains are 2398 // no longer accurate (at spill code) and so must be ignored. 2399 // Node->LRG->reg mappings are still accurate. 2400 _regalloc->Register_Allocate(); 2401 2402 // Bail out if the allocator builds too many nodes 2403 if (failing()) { 2404 return; 2405 } 2406 } 2407 2408 // Prior to register allocation we kept empty basic blocks in case the 2409 // the allocator needed a place to spill. After register allocation we 2410 // are not adding any new instructions. If any basic block is empty, we 2411 // can now safely remove it. 2412 { 2413 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2414 cfg.remove_empty_blocks(); 2415 if (do_freq_based_layout()) { 2416 PhaseBlockLayout layout(cfg); 2417 } else { 2418 cfg.set_loop_alignment(); 2419 } 2420 cfg.fixup_flow(); 2421 } 2422 2423 // Apply peephole optimizations 2424 if( OptoPeephole ) { 2425 TracePhase tp("peephole", &timers[_t_peephole]); 2426 PhasePeephole peep( _regalloc, cfg); 2427 peep.do_transform(); 2428 } 2429 2430 // Do late expand if CPU requires this. 2431 if (Matcher::require_postalloc_expand) { 2432 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2433 cfg.postalloc_expand(_regalloc); 2434 } 2435 2436 // Convert Nodes to instruction bits in a buffer 2437 { 2438 TraceTime tp("output", &timers[_t_output], CITime); 2439 Output(); 2440 } 2441 2442 print_method(PHASE_FINAL_CODE); 2443 2444 // He's dead, Jim. 2445 _cfg = (PhaseCFG*)0xdeadbeef; 2446 _regalloc = (PhaseChaitin*)0xdeadbeef; 2447 } 2448 2449 2450 //------------------------------dump_asm--------------------------------------- 2451 // Dump formatted assembly 2452 #ifndef PRODUCT 2453 void Compile::dump_asm(int *pcs, uint pc_limit) { 2454 bool cut_short = false; 2455 tty->print_cr("#"); 2456 tty->print("# "); _tf->dump(); tty->cr(); 2457 tty->print_cr("#"); 2458 2459 // For all blocks 2460 int pc = 0x0; // Program counter 2461 char starts_bundle = ' '; 2462 _regalloc->dump_frame(); 2463 2464 Node *n = NULL; 2465 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2466 if (VMThread::should_terminate()) { 2467 cut_short = true; 2468 break; 2469 } 2470 Block* block = _cfg->get_block(i); 2471 if (block->is_connector() && !Verbose) { 2472 continue; 2473 } 2474 n = block->head(); 2475 if (pcs && n->_idx < pc_limit) { 2476 tty->print("%3.3x ", pcs[n->_idx]); 2477 } else { 2478 tty->print(" "); 2479 } 2480 block->dump_head(_cfg); 2481 if (block->is_connector()) { 2482 tty->print_cr(" # Empty connector block"); 2483 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2484 tty->print_cr(" # Block is sole successor of call"); 2485 } 2486 2487 // For all instructions 2488 Node *delay = NULL; 2489 for (uint j = 0; j < block->number_of_nodes(); j++) { 2490 if (VMThread::should_terminate()) { 2491 cut_short = true; 2492 break; 2493 } 2494 n = block->get_node(j); 2495 if (valid_bundle_info(n)) { 2496 Bundle* bundle = node_bundling(n); 2497 if (bundle->used_in_unconditional_delay()) { 2498 delay = n; 2499 continue; 2500 } 2501 if (bundle->starts_bundle()) { 2502 starts_bundle = '+'; 2503 } 2504 } 2505 2506 if (WizardMode) { 2507 n->dump(); 2508 } 2509 2510 if( !n->is_Region() && // Dont print in the Assembly 2511 !n->is_Phi() && // a few noisely useless nodes 2512 !n->is_Proj() && 2513 !n->is_MachTemp() && 2514 !n->is_SafePointScalarObject() && 2515 !n->is_Catch() && // Would be nice to print exception table targets 2516 !n->is_MergeMem() && // Not very interesting 2517 !n->is_top() && // Debug info table constants 2518 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2519 ) { 2520 if (pcs && n->_idx < pc_limit) 2521 tty->print("%3.3x", pcs[n->_idx]); 2522 else 2523 tty->print(" "); 2524 tty->print(" %c ", starts_bundle); 2525 starts_bundle = ' '; 2526 tty->print("\t"); 2527 n->format(_regalloc, tty); 2528 tty->cr(); 2529 } 2530 2531 // If we have an instruction with a delay slot, and have seen a delay, 2532 // then back up and print it 2533 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2534 assert(delay != NULL, "no unconditional delay instruction"); 2535 if (WizardMode) delay->dump(); 2536 2537 if (node_bundling(delay)->starts_bundle()) 2538 starts_bundle = '+'; 2539 if (pcs && n->_idx < pc_limit) 2540 tty->print("%3.3x", pcs[n->_idx]); 2541 else 2542 tty->print(" "); 2543 tty->print(" %c ", starts_bundle); 2544 starts_bundle = ' '; 2545 tty->print("\t"); 2546 delay->format(_regalloc, tty); 2547 tty->cr(); 2548 delay = NULL; 2549 } 2550 2551 // Dump the exception table as well 2552 if( n->is_Catch() && (Verbose || WizardMode) ) { 2553 // Print the exception table for this offset 2554 _handler_table.print_subtable_for(pc); 2555 } 2556 } 2557 2558 if (pcs && n->_idx < pc_limit) 2559 tty->print_cr("%3.3x", pcs[n->_idx]); 2560 else 2561 tty->cr(); 2562 2563 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2564 2565 } // End of per-block dump 2566 tty->cr(); 2567 2568 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2569 } 2570 #endif 2571 2572 //------------------------------Final_Reshape_Counts--------------------------- 2573 // This class defines counters to help identify when a method 2574 // may/must be executed using hardware with only 24-bit precision. 2575 struct Final_Reshape_Counts : public StackObj { 2576 int _call_count; // count non-inlined 'common' calls 2577 int _float_count; // count float ops requiring 24-bit precision 2578 int _double_count; // count double ops requiring more precision 2579 int _java_call_count; // count non-inlined 'java' calls 2580 int _inner_loop_count; // count loops which need alignment 2581 VectorSet _visited; // Visitation flags 2582 Node_List _tests; // Set of IfNodes & PCTableNodes 2583 2584 Final_Reshape_Counts() : 2585 _call_count(0), _float_count(0), _double_count(0), 2586 _java_call_count(0), _inner_loop_count(0), 2587 _visited( Thread::current()->resource_area() ) { } 2588 2589 void inc_call_count () { _call_count ++; } 2590 void inc_float_count () { _float_count ++; } 2591 void inc_double_count() { _double_count++; } 2592 void inc_java_call_count() { _java_call_count++; } 2593 void inc_inner_loop_count() { _inner_loop_count++; } 2594 2595 int get_call_count () const { return _call_count ; } 2596 int get_float_count () const { return _float_count ; } 2597 int get_double_count() const { return _double_count; } 2598 int get_java_call_count() const { return _java_call_count; } 2599 int get_inner_loop_count() const { return _inner_loop_count; } 2600 }; 2601 2602 #ifdef ASSERT 2603 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2604 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2605 // Make sure the offset goes inside the instance layout. 2606 return k->contains_field_offset(tp->offset()); 2607 // Note that OffsetBot and OffsetTop are very negative. 2608 } 2609 #endif 2610 2611 // Eliminate trivially redundant StoreCMs and accumulate their 2612 // precedence edges. 2613 void Compile::eliminate_redundant_card_marks(Node* n) { 2614 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2615 if (n->in(MemNode::Address)->outcnt() > 1) { 2616 // There are multiple users of the same address so it might be 2617 // possible to eliminate some of the StoreCMs 2618 Node* mem = n->in(MemNode::Memory); 2619 Node* adr = n->in(MemNode::Address); 2620 Node* val = n->in(MemNode::ValueIn); 2621 Node* prev = n; 2622 bool done = false; 2623 // Walk the chain of StoreCMs eliminating ones that match. As 2624 // long as it's a chain of single users then the optimization is 2625 // safe. Eliminating partially redundant StoreCMs would require 2626 // cloning copies down the other paths. 2627 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2628 if (adr == mem->in(MemNode::Address) && 2629 val == mem->in(MemNode::ValueIn)) { 2630 // redundant StoreCM 2631 if (mem->req() > MemNode::OopStore) { 2632 // Hasn't been processed by this code yet. 2633 n->add_prec(mem->in(MemNode::OopStore)); 2634 } else { 2635 // Already converted to precedence edge 2636 for (uint i = mem->req(); i < mem->len(); i++) { 2637 // Accumulate any precedence edges 2638 if (mem->in(i) != NULL) { 2639 n->add_prec(mem->in(i)); 2640 } 2641 } 2642 // Everything above this point has been processed. 2643 done = true; 2644 } 2645 // Eliminate the previous StoreCM 2646 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2647 assert(mem->outcnt() == 0, "should be dead"); 2648 mem->disconnect_inputs(NULL, this); 2649 } else { 2650 prev = mem; 2651 } 2652 mem = prev->in(MemNode::Memory); 2653 } 2654 } 2655 } 2656 2657 //------------------------------final_graph_reshaping_impl---------------------- 2658 // Implement items 1-5 from final_graph_reshaping below. 2659 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2660 2661 if ( n->outcnt() == 0 ) return; // dead node 2662 uint nop = n->Opcode(); 2663 2664 // Check for 2-input instruction with "last use" on right input. 2665 // Swap to left input. Implements item (2). 2666 if( n->req() == 3 && // two-input instruction 2667 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2668 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2669 n->in(2)->outcnt() == 1 &&// right use IS a last use 2670 !n->in(2)->is_Con() ) { // right use is not a constant 2671 // Check for commutative opcode 2672 switch( nop ) { 2673 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2674 case Op_MaxI: case Op_MinI: 2675 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2676 case Op_AndL: case Op_XorL: case Op_OrL: 2677 case Op_AndI: case Op_XorI: case Op_OrI: { 2678 // Move "last use" input to left by swapping inputs 2679 n->swap_edges(1, 2); 2680 break; 2681 } 2682 default: 2683 break; 2684 } 2685 } 2686 2687 #ifdef ASSERT 2688 if( n->is_Mem() ) { 2689 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2690 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2691 // oop will be recorded in oop map if load crosses safepoint 2692 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2693 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2694 "raw memory operations should have control edge"); 2695 } 2696 #endif 2697 // Count FPU ops and common calls, implements item (3) 2698 switch( nop ) { 2699 // Count all float operations that may use FPU 2700 case Op_AddF: 2701 case Op_SubF: 2702 case Op_MulF: 2703 case Op_DivF: 2704 case Op_NegF: 2705 case Op_ModF: 2706 case Op_ConvI2F: 2707 case Op_ConF: 2708 case Op_CmpF: 2709 case Op_CmpF3: 2710 // case Op_ConvL2F: // longs are split into 32-bit halves 2711 frc.inc_float_count(); 2712 break; 2713 2714 case Op_ConvF2D: 2715 case Op_ConvD2F: 2716 frc.inc_float_count(); 2717 frc.inc_double_count(); 2718 break; 2719 2720 // Count all double operations that may use FPU 2721 case Op_AddD: 2722 case Op_SubD: 2723 case Op_MulD: 2724 case Op_DivD: 2725 case Op_NegD: 2726 case Op_ModD: 2727 case Op_ConvI2D: 2728 case Op_ConvD2I: 2729 // case Op_ConvL2D: // handled by leaf call 2730 // case Op_ConvD2L: // handled by leaf call 2731 case Op_ConD: 2732 case Op_CmpD: 2733 case Op_CmpD3: 2734 frc.inc_double_count(); 2735 break; 2736 case Op_Opaque1: // Remove Opaque Nodes before matching 2737 case Op_Opaque2: // Remove Opaque Nodes before matching 2738 case Op_Opaque3: 2739 n->subsume_by(n->in(1), this); 2740 break; 2741 case Op_CallStaticJava: 2742 case Op_CallJava: 2743 case Op_CallDynamicJava: 2744 frc.inc_java_call_count(); // Count java call site; 2745 case Op_CallRuntime: 2746 case Op_CallLeaf: 2747 case Op_CallLeafNoFP: { 2748 assert( n->is_Call(), "" ); 2749 CallNode *call = n->as_Call(); 2750 // Count call sites where the FP mode bit would have to be flipped. 2751 // Do not count uncommon runtime calls: 2752 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2753 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2754 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2755 frc.inc_call_count(); // Count the call site 2756 } else { // See if uncommon argument is shared 2757 Node *n = call->in(TypeFunc::Parms); 2758 int nop = n->Opcode(); 2759 // Clone shared simple arguments to uncommon calls, item (1). 2760 if( n->outcnt() > 1 && 2761 !n->is_Proj() && 2762 nop != Op_CreateEx && 2763 nop != Op_CheckCastPP && 2764 nop != Op_DecodeN && 2765 nop != Op_DecodeNKlass && 2766 !n->is_Mem() ) { 2767 Node *x = n->clone(); 2768 call->set_req( TypeFunc::Parms, x ); 2769 } 2770 } 2771 break; 2772 } 2773 2774 case Op_StoreD: 2775 case Op_LoadD: 2776 case Op_LoadD_unaligned: 2777 frc.inc_double_count(); 2778 goto handle_mem; 2779 case Op_StoreF: 2780 case Op_LoadF: 2781 frc.inc_float_count(); 2782 goto handle_mem; 2783 2784 case Op_StoreCM: 2785 { 2786 // Convert OopStore dependence into precedence edge 2787 Node* prec = n->in(MemNode::OopStore); 2788 n->del_req(MemNode::OopStore); 2789 n->add_prec(prec); 2790 eliminate_redundant_card_marks(n); 2791 } 2792 2793 // fall through 2794 2795 case Op_StoreB: 2796 case Op_StoreC: 2797 case Op_StorePConditional: 2798 case Op_StoreI: 2799 case Op_StoreL: 2800 case Op_StoreIConditional: 2801 case Op_StoreLConditional: 2802 case Op_CompareAndSwapB: 2803 case Op_CompareAndSwapS: 2804 case Op_CompareAndSwapI: 2805 case Op_CompareAndSwapL: 2806 case Op_CompareAndSwapP: 2807 case Op_CompareAndSwapN: 2808 case Op_WeakCompareAndSwapB: 2809 case Op_WeakCompareAndSwapS: 2810 case Op_WeakCompareAndSwapI: 2811 case Op_WeakCompareAndSwapL: 2812 case Op_WeakCompareAndSwapP: 2813 case Op_WeakCompareAndSwapN: 2814 case Op_CompareAndExchangeB: 2815 case Op_CompareAndExchangeS: 2816 case Op_CompareAndExchangeI: 2817 case Op_CompareAndExchangeL: 2818 case Op_CompareAndExchangeP: 2819 case Op_CompareAndExchangeN: 2820 case Op_GetAndAddS: 2821 case Op_GetAndAddB: 2822 case Op_GetAndAddI: 2823 case Op_GetAndAddL: 2824 case Op_GetAndSetS: 2825 case Op_GetAndSetB: 2826 case Op_GetAndSetI: 2827 case Op_GetAndSetL: 2828 case Op_GetAndSetP: 2829 case Op_GetAndSetN: 2830 case Op_StoreP: 2831 case Op_StoreN: 2832 case Op_StoreNKlass: 2833 case Op_LoadB: 2834 case Op_LoadUB: 2835 case Op_LoadUS: 2836 case Op_LoadI: 2837 case Op_LoadKlass: 2838 case Op_LoadNKlass: 2839 case Op_LoadL: 2840 case Op_LoadL_unaligned: 2841 case Op_LoadPLocked: 2842 case Op_LoadP: 2843 case Op_LoadN: 2844 case Op_LoadRange: 2845 case Op_LoadS: { 2846 handle_mem: 2847 #ifdef ASSERT 2848 if( VerifyOptoOopOffsets ) { 2849 assert( n->is_Mem(), "" ); 2850 MemNode *mem = (MemNode*)n; 2851 // Check to see if address types have grounded out somehow. 2852 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2853 assert( !tp || oop_offset_is_sane(tp), "" ); 2854 } 2855 #endif 2856 break; 2857 } 2858 2859 case Op_AddP: { // Assert sane base pointers 2860 Node *addp = n->in(AddPNode::Address); 2861 assert( !addp->is_AddP() || 2862 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2863 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2864 "Base pointers must match (addp %u)", addp->_idx ); 2865 #ifdef _LP64 2866 if ((UseCompressedOops || UseCompressedClassPointers) && 2867 addp->Opcode() == Op_ConP && 2868 addp == n->in(AddPNode::Base) && 2869 n->in(AddPNode::Offset)->is_Con()) { 2870 // Use addressing with narrow klass to load with offset on x86. 2871 // On sparc loading 32-bits constant and decoding it have less 2872 // instructions (4) then load 64-bits constant (7). 2873 // Do this transformation here since IGVN will convert ConN back to ConP. 2874 const Type* t = addp->bottom_type(); 2875 if (t->isa_oopptr() || t->isa_klassptr()) { 2876 Node* nn = NULL; 2877 2878 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2879 2880 // Look for existing ConN node of the same exact type. 2881 Node* r = root(); 2882 uint cnt = r->outcnt(); 2883 for (uint i = 0; i < cnt; i++) { 2884 Node* m = r->raw_out(i); 2885 if (m!= NULL && m->Opcode() == op && 2886 m->bottom_type()->make_ptr() == t) { 2887 nn = m; 2888 break; 2889 } 2890 } 2891 if (nn != NULL) { 2892 // Decode a narrow oop to match address 2893 // [R12 + narrow_oop_reg<<3 + offset] 2894 if (t->isa_oopptr()) { 2895 nn = new DecodeNNode(nn, t); 2896 } else { 2897 nn = new DecodeNKlassNode(nn, t); 2898 } 2899 // Check for succeeding AddP which uses the same Base. 2900 // Otherwise we will run into the assertion above when visiting that guy. 2901 for (uint i = 0; i < n->outcnt(); ++i) { 2902 Node *out_i = n->raw_out(i); 2903 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 2904 out_i->set_req(AddPNode::Base, nn); 2905 #ifdef ASSERT 2906 for (uint j = 0; j < out_i->outcnt(); ++j) { 2907 Node *out_j = out_i->raw_out(j); 2908 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 2909 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 2910 } 2911 #endif 2912 } 2913 } 2914 n->set_req(AddPNode::Base, nn); 2915 n->set_req(AddPNode::Address, nn); 2916 if (addp->outcnt() == 0) { 2917 addp->disconnect_inputs(NULL, this); 2918 } 2919 } 2920 } 2921 } 2922 #endif 2923 // platform dependent reshaping of the address expression 2924 reshape_address(n->as_AddP()); 2925 break; 2926 } 2927 2928 case Op_CastPP: { 2929 // Remove CastPP nodes to gain more freedom during scheduling but 2930 // keep the dependency they encode as control or precedence edges 2931 // (if control is set already) on memory operations. Some CastPP 2932 // nodes don't have a control (don't carry a dependency): skip 2933 // those. 2934 if (n->in(0) != NULL) { 2935 ResourceMark rm; 2936 Unique_Node_List wq; 2937 wq.push(n); 2938 for (uint next = 0; next < wq.size(); ++next) { 2939 Node *m = wq.at(next); 2940 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 2941 Node* use = m->fast_out(i); 2942 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 2943 use->ensure_control_or_add_prec(n->in(0)); 2944 } else { 2945 switch(use->Opcode()) { 2946 case Op_AddP: 2947 case Op_DecodeN: 2948 case Op_DecodeNKlass: 2949 case Op_CheckCastPP: 2950 case Op_CastPP: 2951 wq.push(use); 2952 break; 2953 } 2954 } 2955 } 2956 } 2957 } 2958 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 2959 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2960 Node* in1 = n->in(1); 2961 const Type* t = n->bottom_type(); 2962 Node* new_in1 = in1->clone(); 2963 new_in1->as_DecodeN()->set_type(t); 2964 2965 if (!Matcher::narrow_oop_use_complex_address()) { 2966 // 2967 // x86, ARM and friends can handle 2 adds in addressing mode 2968 // and Matcher can fold a DecodeN node into address by using 2969 // a narrow oop directly and do implicit NULL check in address: 2970 // 2971 // [R12 + narrow_oop_reg<<3 + offset] 2972 // NullCheck narrow_oop_reg 2973 // 2974 // On other platforms (Sparc) we have to keep new DecodeN node and 2975 // use it to do implicit NULL check in address: 2976 // 2977 // decode_not_null narrow_oop_reg, base_reg 2978 // [base_reg + offset] 2979 // NullCheck base_reg 2980 // 2981 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2982 // to keep the information to which NULL check the new DecodeN node 2983 // corresponds to use it as value in implicit_null_check(). 2984 // 2985 new_in1->set_req(0, n->in(0)); 2986 } 2987 2988 n->subsume_by(new_in1, this); 2989 if (in1->outcnt() == 0) { 2990 in1->disconnect_inputs(NULL, this); 2991 } 2992 } else { 2993 n->subsume_by(n->in(1), this); 2994 if (n->outcnt() == 0) { 2995 n->disconnect_inputs(NULL, this); 2996 } 2997 } 2998 break; 2999 } 3000 #ifdef _LP64 3001 case Op_CmpP: 3002 // Do this transformation here to preserve CmpPNode::sub() and 3003 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3004 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3005 Node* in1 = n->in(1); 3006 Node* in2 = n->in(2); 3007 if (!in1->is_DecodeNarrowPtr()) { 3008 in2 = in1; 3009 in1 = n->in(2); 3010 } 3011 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3012 3013 Node* new_in2 = NULL; 3014 if (in2->is_DecodeNarrowPtr()) { 3015 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3016 new_in2 = in2->in(1); 3017 } else if (in2->Opcode() == Op_ConP) { 3018 const Type* t = in2->bottom_type(); 3019 if (t == TypePtr::NULL_PTR) { 3020 assert(in1->is_DecodeN(), "compare klass to null?"); 3021 // Don't convert CmpP null check into CmpN if compressed 3022 // oops implicit null check is not generated. 3023 // This will allow to generate normal oop implicit null check. 3024 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3025 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3026 // 3027 // This transformation together with CastPP transformation above 3028 // will generated code for implicit NULL checks for compressed oops. 3029 // 3030 // The original code after Optimize() 3031 // 3032 // LoadN memory, narrow_oop_reg 3033 // decode narrow_oop_reg, base_reg 3034 // CmpP base_reg, NULL 3035 // CastPP base_reg // NotNull 3036 // Load [base_reg + offset], val_reg 3037 // 3038 // after these transformations will be 3039 // 3040 // LoadN memory, narrow_oop_reg 3041 // CmpN narrow_oop_reg, NULL 3042 // decode_not_null narrow_oop_reg, base_reg 3043 // Load [base_reg + offset], val_reg 3044 // 3045 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3046 // since narrow oops can be used in debug info now (see the code in 3047 // final_graph_reshaping_walk()). 3048 // 3049 // At the end the code will be matched to 3050 // on x86: 3051 // 3052 // Load_narrow_oop memory, narrow_oop_reg 3053 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3054 // NullCheck narrow_oop_reg 3055 // 3056 // and on sparc: 3057 // 3058 // Load_narrow_oop memory, narrow_oop_reg 3059 // decode_not_null narrow_oop_reg, base_reg 3060 // Load [base_reg + offset], val_reg 3061 // NullCheck base_reg 3062 // 3063 } else if (t->isa_oopptr()) { 3064 new_in2 = ConNode::make(t->make_narrowoop()); 3065 } else if (t->isa_klassptr()) { 3066 new_in2 = ConNode::make(t->make_narrowklass()); 3067 } 3068 } 3069 if (new_in2 != NULL) { 3070 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3071 n->subsume_by(cmpN, this); 3072 if (in1->outcnt() == 0) { 3073 in1->disconnect_inputs(NULL, this); 3074 } 3075 if (in2->outcnt() == 0) { 3076 in2->disconnect_inputs(NULL, this); 3077 } 3078 } 3079 } 3080 break; 3081 3082 case Op_DecodeN: 3083 case Op_DecodeNKlass: 3084 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3085 // DecodeN could be pinned when it can't be fold into 3086 // an address expression, see the code for Op_CastPP above. 3087 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3088 break; 3089 3090 case Op_EncodeP: 3091 case Op_EncodePKlass: { 3092 Node* in1 = n->in(1); 3093 if (in1->is_DecodeNarrowPtr()) { 3094 n->subsume_by(in1->in(1), this); 3095 } else if (in1->Opcode() == Op_ConP) { 3096 const Type* t = in1->bottom_type(); 3097 if (t == TypePtr::NULL_PTR) { 3098 assert(t->isa_oopptr(), "null klass?"); 3099 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3100 } else if (t->isa_oopptr()) { 3101 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3102 } else if (t->isa_klassptr()) { 3103 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3104 } 3105 } 3106 if (in1->outcnt() == 0) { 3107 in1->disconnect_inputs(NULL, this); 3108 } 3109 break; 3110 } 3111 3112 case Op_Proj: { 3113 if (OptimizeStringConcat) { 3114 ProjNode* p = n->as_Proj(); 3115 if (p->_is_io_use) { 3116 // Separate projections were used for the exception path which 3117 // are normally removed by a late inline. If it wasn't inlined 3118 // then they will hang around and should just be replaced with 3119 // the original one. 3120 Node* proj = NULL; 3121 // Replace with just one 3122 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3123 Node *use = i.get(); 3124 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3125 proj = use; 3126 break; 3127 } 3128 } 3129 assert(proj != NULL, "must be found"); 3130 p->subsume_by(proj, this); 3131 } 3132 } 3133 break; 3134 } 3135 3136 case Op_Phi: 3137 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3138 // The EncodeP optimization may create Phi with the same edges 3139 // for all paths. It is not handled well by Register Allocator. 3140 Node* unique_in = n->in(1); 3141 assert(unique_in != NULL, ""); 3142 uint cnt = n->req(); 3143 for (uint i = 2; i < cnt; i++) { 3144 Node* m = n->in(i); 3145 assert(m != NULL, ""); 3146 if (unique_in != m) 3147 unique_in = NULL; 3148 } 3149 if (unique_in != NULL) { 3150 n->subsume_by(unique_in, this); 3151 } 3152 } 3153 break; 3154 3155 #endif 3156 3157 #ifdef ASSERT 3158 case Op_CastII: 3159 // Verify that all range check dependent CastII nodes were removed. 3160 if (n->isa_CastII()->has_range_check()) { 3161 n->dump(3); 3162 assert(false, "Range check dependent CastII node was not removed"); 3163 } 3164 break; 3165 #endif 3166 3167 case Op_ModI: { 3168 Node* di = NULL; 3169 if (UseDivMod) { 3170 // Check if a%b and a/b both exist 3171 di = n->find_similar(Op_DivI); 3172 if (di) { 3173 // Replace them with a fused divmod if supported 3174 if (Matcher::has_match_rule(Op_DivModI)) { 3175 DivModINode* divmod = DivModINode::make(n); 3176 di->subsume_by(divmod->div_proj(), this); 3177 n->subsume_by(divmod->mod_proj(), this); 3178 } else { 3179 // replace a%b with a-((a/b)*b) 3180 Node* mult = new MulINode(di, di->in(2)); 3181 Node* sub = new SubINode(di->in(1), mult); 3182 n->subsume_by(sub, this); 3183 } 3184 } 3185 } 3186 if (di == NULL) { 3187 // Remove useless control edge in case of not mod-zero. 3188 const Type *t = n->in(2)->bottom_type(); 3189 const TypeInt *ti = t->is_int(); 3190 if (n->in(0) && (ti->_hi < 0 || ti->_lo > 0)) { 3191 n->set_req(0, NULL); 3192 } 3193 } 3194 break; 3195 } 3196 3197 case Op_ModL: { 3198 Node* dl = NULL; 3199 if (UseDivMod) { 3200 // Check if a%b and a/b both exist 3201 dl = n->find_similar(Op_DivL); 3202 if (dl) { 3203 // Replace them with a fused divmod if supported 3204 if (Matcher::has_match_rule(Op_DivModL)) { 3205 DivModLNode* divmod = DivModLNode::make(n); 3206 dl->subsume_by(divmod->div_proj(), this); 3207 n->subsume_by(divmod->mod_proj(), this); 3208 } else { 3209 // replace a%b with a-((a/b)*b) 3210 Node* mult = new MulLNode(dl, dl->in(2)); 3211 Node* sub = new SubLNode(dl->in(1), mult); 3212 n->subsume_by(sub, this); 3213 } 3214 } 3215 } 3216 if (dl == NULL) { 3217 // Remove useless control edge in case of not mod-zero. 3218 const Type *t = n->in(2)->bottom_type(); 3219 const TypeLong *tl = t->is_long(); 3220 if (n->in(0) && (tl->_hi < 0 || tl->_lo > 0)) { 3221 n->set_req(0, NULL); 3222 } 3223 } 3224 break; 3225 } 3226 3227 case Op_LoadVector: 3228 case Op_StoreVector: 3229 break; 3230 3231 case Op_AddReductionVI: 3232 case Op_AddReductionVL: 3233 case Op_AddReductionVF: 3234 case Op_AddReductionVD: 3235 case Op_MulReductionVI: 3236 case Op_MulReductionVL: 3237 case Op_MulReductionVF: 3238 case Op_MulReductionVD: 3239 break; 3240 3241 case Op_PackB: 3242 case Op_PackS: 3243 case Op_PackI: 3244 case Op_PackF: 3245 case Op_PackL: 3246 case Op_PackD: 3247 if (n->req()-1 > 2) { 3248 // Replace many operand PackNodes with a binary tree for matching 3249 PackNode* p = (PackNode*) n; 3250 Node* btp = p->binary_tree_pack(1, n->req()); 3251 n->subsume_by(btp, this); 3252 } 3253 break; 3254 case Op_Loop: 3255 case Op_CountedLoop: 3256 if (n->as_Loop()->is_inner_loop()) { 3257 frc.inc_inner_loop_count(); 3258 } 3259 break; 3260 case Op_LShiftI: 3261 case Op_RShiftI: 3262 case Op_URShiftI: 3263 case Op_LShiftL: 3264 case Op_RShiftL: 3265 case Op_URShiftL: 3266 if (Matcher::need_masked_shift_count) { 3267 // The cpu's shift instructions don't restrict the count to the 3268 // lower 5/6 bits. We need to do the masking ourselves. 3269 Node* in2 = n->in(2); 3270 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3271 const TypeInt* t = in2->find_int_type(); 3272 if (t != NULL && t->is_con()) { 3273 juint shift = t->get_con(); 3274 if (shift > mask) { // Unsigned cmp 3275 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3276 } 3277 } else { 3278 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3279 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3280 n->set_req(2, shift); 3281 } 3282 } 3283 if (in2->outcnt() == 0) { // Remove dead node 3284 in2->disconnect_inputs(NULL, this); 3285 } 3286 } 3287 break; 3288 case Op_MemBarStoreStore: 3289 case Op_MemBarRelease: 3290 // Break the link with AllocateNode: it is no longer useful and 3291 // confuses register allocation. 3292 if (n->req() > MemBarNode::Precedent) { 3293 n->set_req(MemBarNode::Precedent, top()); 3294 } 3295 break; 3296 case Op_RangeCheck: { 3297 RangeCheckNode* rc = n->as_RangeCheck(); 3298 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3299 n->subsume_by(iff, this); 3300 frc._tests.push(iff); 3301 break; 3302 } 3303 case Op_ConvI2L: { 3304 if (!Matcher::convi2l_type_required) { 3305 // Code generation on some platforms doesn't need accurate 3306 // ConvI2L types. Widening the type can help remove redundant 3307 // address computations. 3308 n->as_Type()->set_type(TypeLong::INT); 3309 ResourceMark rm; 3310 Node_List wq; 3311 wq.push(n); 3312 for (uint next = 0; next < wq.size(); next++) { 3313 Node *m = wq.at(next); 3314 3315 for(;;) { 3316 // Loop over all nodes with identical inputs edges as m 3317 Node* k = m->find_similar(m->Opcode()); 3318 if (k == NULL) { 3319 break; 3320 } 3321 // Push their uses so we get a chance to remove node made 3322 // redundant 3323 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3324 Node* u = k->fast_out(i); 3325 assert(!wq.contains(u), "shouldn't process one node several times"); 3326 if (u->Opcode() == Op_LShiftL || 3327 u->Opcode() == Op_AddL || 3328 u->Opcode() == Op_SubL || 3329 u->Opcode() == Op_AddP) { 3330 wq.push(u); 3331 } 3332 } 3333 // Replace all nodes with identical edges as m with m 3334 k->subsume_by(m, this); 3335 } 3336 } 3337 } 3338 break; 3339 } 3340 default: 3341 assert( !n->is_Call(), "" ); 3342 assert( !n->is_Mem(), "" ); 3343 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3344 break; 3345 } 3346 3347 // Collect CFG split points 3348 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3349 frc._tests.push(n); 3350 } 3351 } 3352 3353 //------------------------------final_graph_reshaping_walk--------------------- 3354 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3355 // requires that the walk visits a node's inputs before visiting the node. 3356 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3357 ResourceArea *area = Thread::current()->resource_area(); 3358 Unique_Node_List sfpt(area); 3359 3360 frc._visited.set(root->_idx); // first, mark node as visited 3361 uint cnt = root->req(); 3362 Node *n = root; 3363 uint i = 0; 3364 while (true) { 3365 if (i < cnt) { 3366 // Place all non-visited non-null inputs onto stack 3367 Node* m = n->in(i); 3368 ++i; 3369 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3370 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3371 // compute worst case interpreter size in case of a deoptimization 3372 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3373 3374 sfpt.push(m); 3375 } 3376 cnt = m->req(); 3377 nstack.push(n, i); // put on stack parent and next input's index 3378 n = m; 3379 i = 0; 3380 } 3381 } else { 3382 // Now do post-visit work 3383 final_graph_reshaping_impl( n, frc ); 3384 if (nstack.is_empty()) 3385 break; // finished 3386 n = nstack.node(); // Get node from stack 3387 cnt = n->req(); 3388 i = nstack.index(); 3389 nstack.pop(); // Shift to the next node on stack 3390 } 3391 } 3392 3393 // Skip next transformation if compressed oops are not used. 3394 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3395 (!UseCompressedOops && !UseCompressedClassPointers)) 3396 return; 3397 3398 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3399 // It could be done for an uncommon traps or any safepoints/calls 3400 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3401 while (sfpt.size() > 0) { 3402 n = sfpt.pop(); 3403 JVMState *jvms = n->as_SafePoint()->jvms(); 3404 assert(jvms != NULL, "sanity"); 3405 int start = jvms->debug_start(); 3406 int end = n->req(); 3407 bool is_uncommon = (n->is_CallStaticJava() && 3408 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3409 for (int j = start; j < end; j++) { 3410 Node* in = n->in(j); 3411 if (in->is_DecodeNarrowPtr()) { 3412 bool safe_to_skip = true; 3413 if (!is_uncommon ) { 3414 // Is it safe to skip? 3415 for (uint i = 0; i < in->outcnt(); i++) { 3416 Node* u = in->raw_out(i); 3417 if (!u->is_SafePoint() || 3418 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3419 safe_to_skip = false; 3420 } 3421 } 3422 } 3423 if (safe_to_skip) { 3424 n->set_req(j, in->in(1)); 3425 } 3426 if (in->outcnt() == 0) { 3427 in->disconnect_inputs(NULL, this); 3428 } 3429 } 3430 } 3431 } 3432 } 3433 3434 //------------------------------final_graph_reshaping-------------------------- 3435 // Final Graph Reshaping. 3436 // 3437 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3438 // and not commoned up and forced early. Must come after regular 3439 // optimizations to avoid GVN undoing the cloning. Clone constant 3440 // inputs to Loop Phis; these will be split by the allocator anyways. 3441 // Remove Opaque nodes. 3442 // (2) Move last-uses by commutative operations to the left input to encourage 3443 // Intel update-in-place two-address operations and better register usage 3444 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3445 // calls canonicalizing them back. 3446 // (3) Count the number of double-precision FP ops, single-precision FP ops 3447 // and call sites. On Intel, we can get correct rounding either by 3448 // forcing singles to memory (requires extra stores and loads after each 3449 // FP bytecode) or we can set a rounding mode bit (requires setting and 3450 // clearing the mode bit around call sites). The mode bit is only used 3451 // if the relative frequency of single FP ops to calls is low enough. 3452 // This is a key transform for SPEC mpeg_audio. 3453 // (4) Detect infinite loops; blobs of code reachable from above but not 3454 // below. Several of the Code_Gen algorithms fail on such code shapes, 3455 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3456 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3457 // Detection is by looking for IfNodes where only 1 projection is 3458 // reachable from below or CatchNodes missing some targets. 3459 // (5) Assert for insane oop offsets in debug mode. 3460 3461 bool Compile::final_graph_reshaping() { 3462 // an infinite loop may have been eliminated by the optimizer, 3463 // in which case the graph will be empty. 3464 if (root()->req() == 1) { 3465 record_method_not_compilable("trivial infinite loop"); 3466 return true; 3467 } 3468 3469 // Expensive nodes have their control input set to prevent the GVN 3470 // from freely commoning them. There's no GVN beyond this point so 3471 // no need to keep the control input. We want the expensive nodes to 3472 // be freely moved to the least frequent code path by gcm. 3473 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3474 for (int i = 0; i < expensive_count(); i++) { 3475 _expensive_nodes->at(i)->set_req(0, NULL); 3476 } 3477 3478 Final_Reshape_Counts frc; 3479 3480 // Visit everybody reachable! 3481 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3482 Node_Stack nstack(live_nodes() >> 1); 3483 final_graph_reshaping_walk(nstack, root(), frc); 3484 3485 // Check for unreachable (from below) code (i.e., infinite loops). 3486 for( uint i = 0; i < frc._tests.size(); i++ ) { 3487 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3488 // Get number of CFG targets. 3489 // Note that PCTables include exception targets after calls. 3490 uint required_outcnt = n->required_outcnt(); 3491 if (n->outcnt() != required_outcnt) { 3492 // Check for a few special cases. Rethrow Nodes never take the 3493 // 'fall-thru' path, so expected kids is 1 less. 3494 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3495 if (n->in(0)->in(0)->is_Call()) { 3496 CallNode *call = n->in(0)->in(0)->as_Call(); 3497 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3498 required_outcnt--; // Rethrow always has 1 less kid 3499 } else if (call->req() > TypeFunc::Parms && 3500 call->is_CallDynamicJava()) { 3501 // Check for null receiver. In such case, the optimizer has 3502 // detected that the virtual call will always result in a null 3503 // pointer exception. The fall-through projection of this CatchNode 3504 // will not be populated. 3505 Node *arg0 = call->in(TypeFunc::Parms); 3506 if (arg0->is_Type() && 3507 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3508 required_outcnt--; 3509 } 3510 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3511 call->req() > TypeFunc::Parms+1 && 3512 call->is_CallStaticJava()) { 3513 // Check for negative array length. In such case, the optimizer has 3514 // detected that the allocation attempt will always result in an 3515 // exception. There is no fall-through projection of this CatchNode . 3516 Node *arg1 = call->in(TypeFunc::Parms+1); 3517 if (arg1->is_Type() && 3518 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3519 required_outcnt--; 3520 } 3521 } 3522 } 3523 } 3524 // Recheck with a better notion of 'required_outcnt' 3525 if (n->outcnt() != required_outcnt) { 3526 record_method_not_compilable("malformed control flow"); 3527 return true; // Not all targets reachable! 3528 } 3529 } 3530 // Check that I actually visited all kids. Unreached kids 3531 // must be infinite loops. 3532 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3533 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3534 record_method_not_compilable("infinite loop"); 3535 return true; // Found unvisited kid; must be unreach 3536 } 3537 } 3538 3539 // If original bytecodes contained a mixture of floats and doubles 3540 // check if the optimizer has made it homogenous, item (3). 3541 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3542 frc.get_float_count() > 32 && 3543 frc.get_double_count() == 0 && 3544 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3545 set_24_bit_selection_and_mode( false, true ); 3546 } 3547 3548 set_java_calls(frc.get_java_call_count()); 3549 set_inner_loops(frc.get_inner_loop_count()); 3550 3551 // No infinite loops, no reason to bail out. 3552 return false; 3553 } 3554 3555 //-----------------------------too_many_traps---------------------------------- 3556 // Report if there are too many traps at the current method and bci. 3557 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3558 bool Compile::too_many_traps(ciMethod* method, 3559 int bci, 3560 Deoptimization::DeoptReason reason) { 3561 ciMethodData* md = method->method_data(); 3562 if (md->is_empty()) { 3563 // Assume the trap has not occurred, or that it occurred only 3564 // because of a transient condition during start-up in the interpreter. 3565 return false; 3566 } 3567 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3568 if (md->has_trap_at(bci, m, reason) != 0) { 3569 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3570 // Also, if there are multiple reasons, or if there is no per-BCI record, 3571 // assume the worst. 3572 if (log()) 3573 log()->elem("observe trap='%s' count='%d'", 3574 Deoptimization::trap_reason_name(reason), 3575 md->trap_count(reason)); 3576 return true; 3577 } else { 3578 // Ignore method/bci and see if there have been too many globally. 3579 return too_many_traps(reason, md); 3580 } 3581 } 3582 3583 // Less-accurate variant which does not require a method and bci. 3584 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3585 ciMethodData* logmd) { 3586 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3587 // Too many traps globally. 3588 // Note that we use cumulative trap_count, not just md->trap_count. 3589 if (log()) { 3590 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3591 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3592 Deoptimization::trap_reason_name(reason), 3593 mcount, trap_count(reason)); 3594 } 3595 return true; 3596 } else { 3597 // The coast is clear. 3598 return false; 3599 } 3600 } 3601 3602 //--------------------------too_many_recompiles-------------------------------- 3603 // Report if there are too many recompiles at the current method and bci. 3604 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3605 // Is not eager to return true, since this will cause the compiler to use 3606 // Action_none for a trap point, to avoid too many recompilations. 3607 bool Compile::too_many_recompiles(ciMethod* method, 3608 int bci, 3609 Deoptimization::DeoptReason reason) { 3610 ciMethodData* md = method->method_data(); 3611 if (md->is_empty()) { 3612 // Assume the trap has not occurred, or that it occurred only 3613 // because of a transient condition during start-up in the interpreter. 3614 return false; 3615 } 3616 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3617 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3618 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3619 Deoptimization::DeoptReason per_bc_reason 3620 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3621 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3622 if ((per_bc_reason == Deoptimization::Reason_none 3623 || md->has_trap_at(bci, m, reason) != 0) 3624 // The trap frequency measure we care about is the recompile count: 3625 && md->trap_recompiled_at(bci, m) 3626 && md->overflow_recompile_count() >= bc_cutoff) { 3627 // Do not emit a trap here if it has already caused recompilations. 3628 // Also, if there are multiple reasons, or if there is no per-BCI record, 3629 // assume the worst. 3630 if (log()) 3631 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3632 Deoptimization::trap_reason_name(reason), 3633 md->trap_count(reason), 3634 md->overflow_recompile_count()); 3635 return true; 3636 } else if (trap_count(reason) != 0 3637 && decompile_count() >= m_cutoff) { 3638 // Too many recompiles globally, and we have seen this sort of trap. 3639 // Use cumulative decompile_count, not just md->decompile_count. 3640 if (log()) 3641 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3642 Deoptimization::trap_reason_name(reason), 3643 md->trap_count(reason), trap_count(reason), 3644 md->decompile_count(), decompile_count()); 3645 return true; 3646 } else { 3647 // The coast is clear. 3648 return false; 3649 } 3650 } 3651 3652 // Compute when not to trap. Used by matching trap based nodes and 3653 // NullCheck optimization. 3654 void Compile::set_allowed_deopt_reasons() { 3655 _allowed_reasons = 0; 3656 if (is_method_compilation()) { 3657 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3658 assert(rs < BitsPerInt, "recode bit map"); 3659 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3660 _allowed_reasons |= nth_bit(rs); 3661 } 3662 } 3663 } 3664 } 3665 3666 #ifndef PRODUCT 3667 //------------------------------verify_graph_edges--------------------------- 3668 // Walk the Graph and verify that there is a one-to-one correspondence 3669 // between Use-Def edges and Def-Use edges in the graph. 3670 void Compile::verify_graph_edges(bool no_dead_code) { 3671 if (VerifyGraphEdges) { 3672 ResourceArea *area = Thread::current()->resource_area(); 3673 Unique_Node_List visited(area); 3674 // Call recursive graph walk to check edges 3675 _root->verify_edges(visited); 3676 if (no_dead_code) { 3677 // Now make sure that no visited node is used by an unvisited node. 3678 bool dead_nodes = false; 3679 Unique_Node_List checked(area); 3680 while (visited.size() > 0) { 3681 Node* n = visited.pop(); 3682 checked.push(n); 3683 for (uint i = 0; i < n->outcnt(); i++) { 3684 Node* use = n->raw_out(i); 3685 if (checked.member(use)) continue; // already checked 3686 if (visited.member(use)) continue; // already in the graph 3687 if (use->is_Con()) continue; // a dead ConNode is OK 3688 // At this point, we have found a dead node which is DU-reachable. 3689 if (!dead_nodes) { 3690 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3691 dead_nodes = true; 3692 } 3693 use->dump(2); 3694 tty->print_cr("---"); 3695 checked.push(use); // No repeats; pretend it is now checked. 3696 } 3697 } 3698 assert(!dead_nodes, "using nodes must be reachable from root"); 3699 } 3700 } 3701 } 3702 3703 // Verify GC barriers consistency 3704 // Currently supported: 3705 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3706 void Compile::verify_barriers() { 3707 if (UseG1GC) { 3708 // Verify G1 pre-barriers 3709 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + SATBMarkQueue::byte_offset_of_active()); 3710 3711 ResourceArea *area = Thread::current()->resource_area(); 3712 Unique_Node_List visited(area); 3713 Node_List worklist(area); 3714 // We're going to walk control flow backwards starting from the Root 3715 worklist.push(_root); 3716 while (worklist.size() > 0) { 3717 Node* x = worklist.pop(); 3718 if (x == NULL || x == top()) continue; 3719 if (visited.member(x)) { 3720 continue; 3721 } else { 3722 visited.push(x); 3723 } 3724 3725 if (x->is_Region()) { 3726 for (uint i = 1; i < x->req(); i++) { 3727 worklist.push(x->in(i)); 3728 } 3729 } else { 3730 worklist.push(x->in(0)); 3731 // We are looking for the pattern: 3732 // /->ThreadLocal 3733 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3734 // \->ConI(0) 3735 // We want to verify that the If and the LoadB have the same control 3736 // See GraphKit::g1_write_barrier_pre() 3737 if (x->is_If()) { 3738 IfNode *iff = x->as_If(); 3739 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3740 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3741 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3742 && cmp->in(1)->is_Load()) { 3743 LoadNode* load = cmp->in(1)->as_Load(); 3744 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3745 && load->in(2)->in(3)->is_Con() 3746 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3747 3748 Node* if_ctrl = iff->in(0); 3749 Node* load_ctrl = load->in(0); 3750 3751 if (if_ctrl != load_ctrl) { 3752 // Skip possible CProj->NeverBranch in infinite loops 3753 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3754 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3755 if_ctrl = if_ctrl->in(0)->in(0); 3756 } 3757 } 3758 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3759 } 3760 } 3761 } 3762 } 3763 } 3764 } 3765 } 3766 } 3767 3768 #endif 3769 3770 // The Compile object keeps track of failure reasons separately from the ciEnv. 3771 // This is required because there is not quite a 1-1 relation between the 3772 // ciEnv and its compilation task and the Compile object. Note that one 3773 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3774 // to backtrack and retry without subsuming loads. Other than this backtracking 3775 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3776 // by the logic in C2Compiler. 3777 void Compile::record_failure(const char* reason) { 3778 if (log() != NULL) { 3779 log()->elem("failure reason='%s' phase='compile'", reason); 3780 } 3781 if (_failure_reason == NULL) { 3782 // Record the first failure reason. 3783 _failure_reason = reason; 3784 } 3785 3786 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3787 C->print_method(PHASE_FAILURE); 3788 } 3789 _root = NULL; // flush the graph, too 3790 } 3791 3792 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3793 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3794 _phase_name(name), _dolog(CITimeVerbose) 3795 { 3796 if (_dolog) { 3797 C = Compile::current(); 3798 _log = C->log(); 3799 } else { 3800 C = NULL; 3801 _log = NULL; 3802 } 3803 if (_log != NULL) { 3804 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3805 _log->stamp(); 3806 _log->end_head(); 3807 } 3808 } 3809 3810 Compile::TracePhase::~TracePhase() { 3811 3812 C = Compile::current(); 3813 if (_dolog) { 3814 _log = C->log(); 3815 } else { 3816 _log = NULL; 3817 } 3818 3819 #ifdef ASSERT 3820 if (PrintIdealNodeCount) { 3821 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3822 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3823 } 3824 3825 if (VerifyIdealNodeCount) { 3826 Compile::current()->print_missing_nodes(); 3827 } 3828 #endif 3829 3830 if (_log != NULL) { 3831 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3832 } 3833 } 3834 3835 //============================================================================= 3836 // Two Constant's are equal when the type and the value are equal. 3837 bool Compile::Constant::operator==(const Constant& other) { 3838 if (type() != other.type() ) return false; 3839 if (can_be_reused() != other.can_be_reused()) return false; 3840 // For floating point values we compare the bit pattern. 3841 switch (type()) { 3842 case T_INT: 3843 case T_FLOAT: return (_v._value.i == other._v._value.i); 3844 case T_LONG: 3845 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3846 case T_OBJECT: 3847 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3848 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3849 case T_METADATA: return (_v._metadata == other._v._metadata); 3850 default: ShouldNotReachHere(); 3851 } 3852 return false; 3853 } 3854 3855 static int type_to_size_in_bytes(BasicType t) { 3856 switch (t) { 3857 case T_INT: return sizeof(jint ); 3858 case T_LONG: return sizeof(jlong ); 3859 case T_FLOAT: return sizeof(jfloat ); 3860 case T_DOUBLE: return sizeof(jdouble); 3861 case T_METADATA: return sizeof(Metadata*); 3862 // We use T_VOID as marker for jump-table entries (labels) which 3863 // need an internal word relocation. 3864 case T_VOID: 3865 case T_ADDRESS: 3866 case T_OBJECT: return sizeof(jobject); 3867 } 3868 3869 ShouldNotReachHere(); 3870 return -1; 3871 } 3872 3873 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3874 // sort descending 3875 if (a->freq() > b->freq()) return -1; 3876 if (a->freq() < b->freq()) return 1; 3877 return 0; 3878 } 3879 3880 void Compile::ConstantTable::calculate_offsets_and_size() { 3881 // First, sort the array by frequencies. 3882 _constants.sort(qsort_comparator); 3883 3884 #ifdef ASSERT 3885 // Make sure all jump-table entries were sorted to the end of the 3886 // array (they have a negative frequency). 3887 bool found_void = false; 3888 for (int i = 0; i < _constants.length(); i++) { 3889 Constant con = _constants.at(i); 3890 if (con.type() == T_VOID) 3891 found_void = true; // jump-tables 3892 else 3893 assert(!found_void, "wrong sorting"); 3894 } 3895 #endif 3896 3897 int offset = 0; 3898 for (int i = 0; i < _constants.length(); i++) { 3899 Constant* con = _constants.adr_at(i); 3900 3901 // Align offset for type. 3902 int typesize = type_to_size_in_bytes(con->type()); 3903 offset = align_size_up(offset, typesize); 3904 con->set_offset(offset); // set constant's offset 3905 3906 if (con->type() == T_VOID) { 3907 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3908 offset = offset + typesize * n->outcnt(); // expand jump-table 3909 } else { 3910 offset = offset + typesize; 3911 } 3912 } 3913 3914 // Align size up to the next section start (which is insts; see 3915 // CodeBuffer::align_at_start). 3916 assert(_size == -1, "already set?"); 3917 _size = align_size_up(offset, CodeEntryAlignment); 3918 } 3919 3920 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3921 MacroAssembler _masm(&cb); 3922 for (int i = 0; i < _constants.length(); i++) { 3923 Constant con = _constants.at(i); 3924 address constant_addr = NULL; 3925 switch (con.type()) { 3926 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 3927 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3928 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3929 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3930 case T_OBJECT: { 3931 jobject obj = con.get_jobject(); 3932 int oop_index = _masm.oop_recorder()->find_index(obj); 3933 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3934 break; 3935 } 3936 case T_ADDRESS: { 3937 address addr = (address) con.get_jobject(); 3938 constant_addr = _masm.address_constant(addr); 3939 break; 3940 } 3941 // We use T_VOID as marker for jump-table entries (labels) which 3942 // need an internal word relocation. 3943 case T_VOID: { 3944 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3945 // Fill the jump-table with a dummy word. The real value is 3946 // filled in later in fill_jump_table. 3947 address dummy = (address) n; 3948 constant_addr = _masm.address_constant(dummy); 3949 // Expand jump-table 3950 for (uint i = 1; i < n->outcnt(); i++) { 3951 address temp_addr = _masm.address_constant(dummy + i); 3952 assert(temp_addr, "consts section too small"); 3953 } 3954 break; 3955 } 3956 case T_METADATA: { 3957 Metadata* obj = con.get_metadata(); 3958 int metadata_index = _masm.oop_recorder()->find_index(obj); 3959 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3960 break; 3961 } 3962 default: ShouldNotReachHere(); 3963 } 3964 assert(constant_addr, "consts section too small"); 3965 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3966 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 3967 } 3968 } 3969 3970 int Compile::ConstantTable::find_offset(Constant& con) const { 3971 int idx = _constants.find(con); 3972 assert(idx != -1, "constant must be in constant table"); 3973 int offset = _constants.at(idx).offset(); 3974 assert(offset != -1, "constant table not emitted yet?"); 3975 return offset; 3976 } 3977 3978 void Compile::ConstantTable::add(Constant& con) { 3979 if (con.can_be_reused()) { 3980 int idx = _constants.find(con); 3981 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3982 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3983 return; 3984 } 3985 } 3986 (void) _constants.append(con); 3987 } 3988 3989 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3990 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3991 Constant con(type, value, b->_freq); 3992 add(con); 3993 return con; 3994 } 3995 3996 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3997 Constant con(metadata); 3998 add(con); 3999 return con; 4000 } 4001 4002 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4003 jvalue value; 4004 BasicType type = oper->type()->basic_type(); 4005 switch (type) { 4006 case T_LONG: value.j = oper->constantL(); break; 4007 case T_FLOAT: value.f = oper->constantF(); break; 4008 case T_DOUBLE: value.d = oper->constantD(); break; 4009 case T_OBJECT: 4010 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4011 case T_METADATA: return add((Metadata*)oper->constant()); break; 4012 default: guarantee(false, "unhandled type: %s", type2name(type)); 4013 } 4014 return add(n, type, value); 4015 } 4016 4017 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4018 jvalue value; 4019 // We can use the node pointer here to identify the right jump-table 4020 // as this method is called from Compile::Fill_buffer right before 4021 // the MachNodes are emitted and the jump-table is filled (means the 4022 // MachNode pointers do not change anymore). 4023 value.l = (jobject) n; 4024 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4025 add(con); 4026 return con; 4027 } 4028 4029 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4030 // If called from Compile::scratch_emit_size do nothing. 4031 if (Compile::current()->in_scratch_emit_size()) return; 4032 4033 assert(labels.is_nonempty(), "must be"); 4034 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4035 4036 // Since MachConstantNode::constant_offset() also contains 4037 // table_base_offset() we need to subtract the table_base_offset() 4038 // to get the plain offset into the constant table. 4039 int offset = n->constant_offset() - table_base_offset(); 4040 4041 MacroAssembler _masm(&cb); 4042 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4043 4044 for (uint i = 0; i < n->outcnt(); i++) { 4045 address* constant_addr = &jump_table_base[i]; 4046 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)); 4047 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4048 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4049 } 4050 } 4051 4052 //----------------------------static_subtype_check----------------------------- 4053 // Shortcut important common cases when superklass is exact: 4054 // (0) superklass is java.lang.Object (can occur in reflective code) 4055 // (1) subklass is already limited to a subtype of superklass => always ok 4056 // (2) subklass does not overlap with superklass => always fail 4057 // (3) superklass has NO subtypes and we can check with a simple compare. 4058 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4059 if (StressReflectiveCode) { 4060 return SSC_full_test; // Let caller generate the general case. 4061 } 4062 4063 if (superk == env()->Object_klass()) { 4064 return SSC_always_true; // (0) this test cannot fail 4065 } 4066 4067 ciType* superelem = superk; 4068 if (superelem->is_array_klass()) 4069 superelem = superelem->as_array_klass()->base_element_type(); 4070 4071 if (!subk->is_interface()) { // cannot trust static interface types yet 4072 if (subk->is_subtype_of(superk)) { 4073 return SSC_always_true; // (1) false path dead; no dynamic test needed 4074 } 4075 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4076 !superk->is_subtype_of(subk)) { 4077 return SSC_always_false; 4078 } 4079 } 4080 4081 // If casting to an instance klass, it must have no subtypes 4082 if (superk->is_interface()) { 4083 // Cannot trust interfaces yet. 4084 // %%% S.B. superk->nof_implementors() == 1 4085 } else if (superelem->is_instance_klass()) { 4086 ciInstanceKlass* ik = superelem->as_instance_klass(); 4087 if (!ik->has_subklass() && !ik->is_interface()) { 4088 if (!ik->is_final()) { 4089 // Add a dependency if there is a chance of a later subclass. 4090 dependencies()->assert_leaf_type(ik); 4091 } 4092 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4093 } 4094 } else { 4095 // A primitive array type has no subtypes. 4096 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4097 } 4098 4099 return SSC_full_test; 4100 } 4101 4102 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4103 #ifdef _LP64 4104 // The scaled index operand to AddP must be a clean 64-bit value. 4105 // Java allows a 32-bit int to be incremented to a negative 4106 // value, which appears in a 64-bit register as a large 4107 // positive number. Using that large positive number as an 4108 // operand in pointer arithmetic has bad consequences. 4109 // On the other hand, 32-bit overflow is rare, and the possibility 4110 // can often be excluded, if we annotate the ConvI2L node with 4111 // a type assertion that its value is known to be a small positive 4112 // number. (The prior range check has ensured this.) 4113 // This assertion is used by ConvI2LNode::Ideal. 4114 int index_max = max_jint - 1; // array size is max_jint, index is one less 4115 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4116 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4117 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4118 #endif 4119 return idx; 4120 } 4121 4122 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4123 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4124 if (ctrl != NULL) { 4125 // Express control dependency by a CastII node with a narrow type. 4126 value = new CastIINode(value, itype, false, true /* range check dependency */); 4127 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4128 // node from floating above the range check during loop optimizations. Otherwise, the 4129 // ConvI2L node may be eliminated independently of the range check, causing the data path 4130 // to become TOP while the control path is still there (although it's unreachable). 4131 value->set_req(0, ctrl); 4132 // Save CastII node to remove it after loop optimizations. 4133 phase->C->add_range_check_cast(value); 4134 value = phase->transform(value); 4135 } 4136 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4137 return phase->transform(new ConvI2LNode(value, ltype)); 4138 } 4139 4140 // The message about the current inlining is accumulated in 4141 // _print_inlining_stream and transfered into the _print_inlining_list 4142 // once we know whether inlining succeeds or not. For regular 4143 // inlining, messages are appended to the buffer pointed by 4144 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4145 // a new buffer is added after _print_inlining_idx in the list. This 4146 // way we can update the inlining message for late inlining call site 4147 // when the inlining is attempted again. 4148 void Compile::print_inlining_init() { 4149 if (print_inlining() || print_intrinsics()) { 4150 _print_inlining_stream = new stringStream(); 4151 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4152 } 4153 } 4154 4155 void Compile::print_inlining_reinit() { 4156 if (print_inlining() || print_intrinsics()) { 4157 // Re allocate buffer when we change ResourceMark 4158 _print_inlining_stream = new stringStream(); 4159 } 4160 } 4161 4162 void Compile::print_inlining_reset() { 4163 _print_inlining_stream->reset(); 4164 } 4165 4166 void Compile::print_inlining_commit() { 4167 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4168 // Transfer the message from _print_inlining_stream to the current 4169 // _print_inlining_list buffer and clear _print_inlining_stream. 4170 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4171 print_inlining_reset(); 4172 } 4173 4174 void Compile::print_inlining_push() { 4175 // Add new buffer to the _print_inlining_list at current position 4176 _print_inlining_idx++; 4177 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4178 } 4179 4180 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4181 return _print_inlining_list->at(_print_inlining_idx); 4182 } 4183 4184 void Compile::print_inlining_update(CallGenerator* cg) { 4185 if (print_inlining() || print_intrinsics()) { 4186 if (!cg->is_late_inline()) { 4187 if (print_inlining_current().cg() != NULL) { 4188 print_inlining_push(); 4189 } 4190 print_inlining_commit(); 4191 } else { 4192 if (print_inlining_current().cg() != cg && 4193 (print_inlining_current().cg() != NULL || 4194 print_inlining_current().ss()->size() != 0)) { 4195 print_inlining_push(); 4196 } 4197 print_inlining_commit(); 4198 print_inlining_current().set_cg(cg); 4199 } 4200 } 4201 } 4202 4203 void Compile::print_inlining_move_to(CallGenerator* cg) { 4204 // We resume inlining at a late inlining call site. Locate the 4205 // corresponding inlining buffer so that we can update it. 4206 if (print_inlining()) { 4207 for (int i = 0; i < _print_inlining_list->length(); i++) { 4208 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4209 _print_inlining_idx = i; 4210 return; 4211 } 4212 } 4213 ShouldNotReachHere(); 4214 } 4215 } 4216 4217 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4218 if (print_inlining()) { 4219 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4220 assert(print_inlining_current().cg() == cg, "wrong entry"); 4221 // replace message with new message 4222 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4223 print_inlining_commit(); 4224 print_inlining_current().set_cg(cg); 4225 } 4226 } 4227 4228 void Compile::print_inlining_assert_ready() { 4229 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4230 } 4231 4232 void Compile::process_print_inlining() { 4233 bool do_print_inlining = print_inlining() || print_intrinsics(); 4234 if (do_print_inlining || log() != NULL) { 4235 // Print inlining message for candidates that we couldn't inline 4236 // for lack of space 4237 for (int i = 0; i < _late_inlines.length(); i++) { 4238 CallGenerator* cg = _late_inlines.at(i); 4239 if (!cg->is_mh_late_inline()) { 4240 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4241 if (do_print_inlining) { 4242 cg->print_inlining_late(msg); 4243 } 4244 log_late_inline_failure(cg, msg); 4245 } 4246 } 4247 } 4248 if (do_print_inlining) { 4249 ResourceMark rm; 4250 stringStream ss; 4251 for (int i = 0; i < _print_inlining_list->length(); i++) { 4252 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4253 } 4254 size_t end = ss.size(); 4255 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4256 strncpy(_print_inlining_output, ss.base(), end+1); 4257 _print_inlining_output[end] = 0; 4258 } 4259 } 4260 4261 void Compile::dump_print_inlining() { 4262 if (_print_inlining_output != NULL) { 4263 tty->print_raw(_print_inlining_output); 4264 } 4265 } 4266 4267 void Compile::log_late_inline(CallGenerator* cg) { 4268 if (log() != NULL) { 4269 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4270 cg->unique_id()); 4271 JVMState* p = cg->call_node()->jvms(); 4272 while (p != NULL) { 4273 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4274 p = p->caller(); 4275 } 4276 log()->tail("late_inline"); 4277 } 4278 } 4279 4280 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4281 log_late_inline(cg); 4282 if (log() != NULL) { 4283 log()->inline_fail(msg); 4284 } 4285 } 4286 4287 void Compile::log_inline_id(CallGenerator* cg) { 4288 if (log() != NULL) { 4289 // The LogCompilation tool needs a unique way to identify late 4290 // inline call sites. This id must be unique for this call site in 4291 // this compilation. Try to have it unique across compilations as 4292 // well because it can be convenient when grepping through the log 4293 // file. 4294 // Distinguish OSR compilations from others in case CICountOSR is 4295 // on. 4296 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4297 cg->set_unique_id(id); 4298 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4299 } 4300 } 4301 4302 void Compile::log_inline_failure(const char* msg) { 4303 if (C->log() != NULL) { 4304 C->log()->inline_fail(msg); 4305 } 4306 } 4307 4308 4309 // Dump inlining replay data to the stream. 4310 // Don't change thread state and acquire any locks. 4311 void Compile::dump_inline_data(outputStream* out) { 4312 InlineTree* inl_tree = ilt(); 4313 if (inl_tree != NULL) { 4314 out->print(" inline %d", inl_tree->count()); 4315 inl_tree->dump_replay_data(out); 4316 } 4317 } 4318 4319 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4320 if (n1->Opcode() < n2->Opcode()) return -1; 4321 else if (n1->Opcode() > n2->Opcode()) return 1; 4322 4323 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4324 for (uint i = 1; i < n1->req(); i++) { 4325 if (n1->in(i) < n2->in(i)) return -1; 4326 else if (n1->in(i) > n2->in(i)) return 1; 4327 } 4328 4329 return 0; 4330 } 4331 4332 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4333 Node* n1 = *n1p; 4334 Node* n2 = *n2p; 4335 4336 return cmp_expensive_nodes(n1, n2); 4337 } 4338 4339 void Compile::sort_expensive_nodes() { 4340 if (!expensive_nodes_sorted()) { 4341 _expensive_nodes->sort(cmp_expensive_nodes); 4342 } 4343 } 4344 4345 bool Compile::expensive_nodes_sorted() const { 4346 for (int i = 1; i < _expensive_nodes->length(); i++) { 4347 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4348 return false; 4349 } 4350 } 4351 return true; 4352 } 4353 4354 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4355 if (_expensive_nodes->length() == 0) { 4356 return false; 4357 } 4358 4359 assert(OptimizeExpensiveOps, "optimization off?"); 4360 4361 // Take this opportunity to remove dead nodes from the list 4362 int j = 0; 4363 for (int i = 0; i < _expensive_nodes->length(); i++) { 4364 Node* n = _expensive_nodes->at(i); 4365 if (!n->is_unreachable(igvn)) { 4366 assert(n->is_expensive(), "should be expensive"); 4367 _expensive_nodes->at_put(j, n); 4368 j++; 4369 } 4370 } 4371 _expensive_nodes->trunc_to(j); 4372 4373 // Then sort the list so that similar nodes are next to each other 4374 // and check for at least two nodes of identical kind with same data 4375 // inputs. 4376 sort_expensive_nodes(); 4377 4378 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4379 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4380 return true; 4381 } 4382 } 4383 4384 return false; 4385 } 4386 4387 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4388 if (_expensive_nodes->length() == 0) { 4389 return; 4390 } 4391 4392 assert(OptimizeExpensiveOps, "optimization off?"); 4393 4394 // Sort to bring similar nodes next to each other and clear the 4395 // control input of nodes for which there's only a single copy. 4396 sort_expensive_nodes(); 4397 4398 int j = 0; 4399 int identical = 0; 4400 int i = 0; 4401 bool modified = false; 4402 for (; i < _expensive_nodes->length()-1; i++) { 4403 assert(j <= i, "can't write beyond current index"); 4404 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4405 identical++; 4406 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4407 continue; 4408 } 4409 if (identical > 0) { 4410 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4411 identical = 0; 4412 } else { 4413 Node* n = _expensive_nodes->at(i); 4414 igvn.replace_input_of(n, 0, NULL); 4415 igvn.hash_insert(n); 4416 modified = true; 4417 } 4418 } 4419 if (identical > 0) { 4420 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4421 } else if (_expensive_nodes->length() >= 1) { 4422 Node* n = _expensive_nodes->at(i); 4423 igvn.replace_input_of(n, 0, NULL); 4424 igvn.hash_insert(n); 4425 modified = true; 4426 } 4427 _expensive_nodes->trunc_to(j); 4428 if (modified) { 4429 igvn.optimize(); 4430 } 4431 } 4432 4433 void Compile::add_expensive_node(Node * n) { 4434 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4435 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4436 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4437 if (OptimizeExpensiveOps) { 4438 _expensive_nodes->append(n); 4439 } else { 4440 // Clear control input and let IGVN optimize expensive nodes if 4441 // OptimizeExpensiveOps is off. 4442 n->set_req(0, NULL); 4443 } 4444 } 4445 4446 /** 4447 * Remove the speculative part of types and clean up the graph 4448 */ 4449 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4450 if (UseTypeSpeculation) { 4451 Unique_Node_List worklist; 4452 worklist.push(root()); 4453 int modified = 0; 4454 // Go over all type nodes that carry a speculative type, drop the 4455 // speculative part of the type and enqueue the node for an igvn 4456 // which may optimize it out. 4457 for (uint next = 0; next < worklist.size(); ++next) { 4458 Node *n = worklist.at(next); 4459 if (n->is_Type()) { 4460 TypeNode* tn = n->as_Type(); 4461 const Type* t = tn->type(); 4462 const Type* t_no_spec = t->remove_speculative(); 4463 if (t_no_spec != t) { 4464 bool in_hash = igvn.hash_delete(n); 4465 assert(in_hash, "node should be in igvn hash table"); 4466 tn->set_type(t_no_spec); 4467 igvn.hash_insert(n); 4468 igvn._worklist.push(n); // give it a chance to go away 4469 modified++; 4470 } 4471 } 4472 uint max = n->len(); 4473 for( uint i = 0; i < max; ++i ) { 4474 Node *m = n->in(i); 4475 if (not_a_node(m)) continue; 4476 worklist.push(m); 4477 } 4478 } 4479 // Drop the speculative part of all types in the igvn's type table 4480 igvn.remove_speculative_types(); 4481 if (modified > 0) { 4482 igvn.optimize(); 4483 } 4484 #ifdef ASSERT 4485 // Verify that after the IGVN is over no speculative type has resurfaced 4486 worklist.clear(); 4487 worklist.push(root()); 4488 for (uint next = 0; next < worklist.size(); ++next) { 4489 Node *n = worklist.at(next); 4490 const Type* t = igvn.type_or_null(n); 4491 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4492 if (n->is_Type()) { 4493 t = n->as_Type()->type(); 4494 assert(t == t->remove_speculative(), "no more speculative types"); 4495 } 4496 uint max = n->len(); 4497 for( uint i = 0; i < max; ++i ) { 4498 Node *m = n->in(i); 4499 if (not_a_node(m)) continue; 4500 worklist.push(m); 4501 } 4502 } 4503 igvn.check_no_speculative_types(); 4504 #endif 4505 } 4506 } 4507 4508 // Auxiliary method to support randomized stressing/fuzzing. 4509 // 4510 // This method can be called the arbitrary number of times, with current count 4511 // as the argument. The logic allows selecting a single candidate from the 4512 // running list of candidates as follows: 4513 // int count = 0; 4514 // Cand* selected = null; 4515 // while(cand = cand->next()) { 4516 // if (randomized_select(++count)) { 4517 // selected = cand; 4518 // } 4519 // } 4520 // 4521 // Including count equalizes the chances any candidate is "selected". 4522 // This is useful when we don't have the complete list of candidates to choose 4523 // from uniformly. In this case, we need to adjust the randomicity of the 4524 // selection, or else we will end up biasing the selection towards the latter 4525 // candidates. 4526 // 4527 // Quick back-envelope calculation shows that for the list of n candidates 4528 // the equal probability for the candidate to persist as "best" can be 4529 // achieved by replacing it with "next" k-th candidate with the probability 4530 // of 1/k. It can be easily shown that by the end of the run, the 4531 // probability for any candidate is converged to 1/n, thus giving the 4532 // uniform distribution among all the candidates. 4533 // 4534 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4535 #define RANDOMIZED_DOMAIN_POW 29 4536 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4537 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4538 bool Compile::randomized_select(int count) { 4539 assert(count > 0, "only positive"); 4540 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4541 } 4542 4543 CloneMap& Compile::clone_map() { return _clone_map; } 4544 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4545 4546 void NodeCloneInfo::dump() const { 4547 tty->print(" {%d:%d} ", idx(), gen()); 4548 } 4549 4550 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4551 uint64_t val = value(old->_idx); 4552 NodeCloneInfo cio(val); 4553 assert(val != 0, "old node should be in the map"); 4554 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4555 insert(nnn->_idx, cin.get()); 4556 #ifndef PRODUCT 4557 if (is_debug()) { 4558 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4559 } 4560 #endif 4561 } 4562 4563 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4564 NodeCloneInfo cio(value(old->_idx)); 4565 if (cio.get() == 0) { 4566 cio.set(old->_idx, 0); 4567 insert(old->_idx, cio.get()); 4568 #ifndef PRODUCT 4569 if (is_debug()) { 4570 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4571 } 4572 #endif 4573 } 4574 clone(old, nnn, gen); 4575 } 4576 4577 int CloneMap::max_gen() const { 4578 int g = 0; 4579 DictI di(_dict); 4580 for(; di.test(); ++di) { 4581 int t = gen(di._key); 4582 if (g < t) { 4583 g = t; 4584 #ifndef PRODUCT 4585 if (is_debug()) { 4586 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4587 } 4588 #endif 4589 } 4590 } 4591 return g; 4592 } 4593 4594 void CloneMap::dump(node_idx_t key) const { 4595 uint64_t val = value(key); 4596 if (val != 0) { 4597 NodeCloneInfo ni(val); 4598 ni.dump(); 4599 } 4600 }