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