1 /* 2 * Copyright (c) 1997, 2017, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "memory/resourceArea.hpp" 37 #include "opto/addnode.hpp" 38 #include "opto/block.hpp" 39 #include "opto/c2compiler.hpp" 40 #include "opto/callGenerator.hpp" 41 #include "opto/callnode.hpp" 42 #include "opto/castnode.hpp" 43 #include "opto/cfgnode.hpp" 44 #include "opto/chaitin.hpp" 45 #include "opto/compile.hpp" 46 #include "opto/connode.hpp" 47 #include "opto/convertnode.hpp" 48 #include "opto/divnode.hpp" 49 #include "opto/escape.hpp" 50 #include "opto/idealGraphPrinter.hpp" 51 #include "opto/loopnode.hpp" 52 #include "opto/machnode.hpp" 53 #include "opto/macro.hpp" 54 #include "opto/matcher.hpp" 55 #include "opto/mathexactnode.hpp" 56 #include "opto/memnode.hpp" 57 #include "opto/mulnode.hpp" 58 #include "opto/narrowptrnode.hpp" 59 #include "opto/node.hpp" 60 #include "opto/opcodes.hpp" 61 #include "opto/output.hpp" 62 #include "opto/parse.hpp" 63 #include "opto/phaseX.hpp" 64 #include "opto/rootnode.hpp" 65 #include "opto/runtime.hpp" 66 #include "opto/stringopts.hpp" 67 #include "opto/type.hpp" 68 #include "opto/valuetypenode.hpp" 69 #include "opto/vectornode.hpp" 70 #include "runtime/arguments.hpp" 71 #include "runtime/sharedRuntime.hpp" 72 #include "runtime/signature.hpp" 73 #include "runtime/stubRoutines.hpp" 74 #include "runtime/timer.hpp" 75 #include "utilities/align.hpp" 76 #include "utilities/copy.hpp" 77 78 79 // -------------------- Compile::mach_constant_base_node ----------------------- 80 // Constant table base node singleton. 81 MachConstantBaseNode* Compile::mach_constant_base_node() { 82 if (_mach_constant_base_node == NULL) { 83 _mach_constant_base_node = new MachConstantBaseNode(); 84 _mach_constant_base_node->add_req(C->root()); 85 } 86 return _mach_constant_base_node; 87 } 88 89 90 /// Support for intrinsics. 91 92 // Return the index at which m must be inserted (or already exists). 93 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 94 class IntrinsicDescPair { 95 private: 96 ciMethod* _m; 97 bool _is_virtual; 98 public: 99 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 100 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 101 ciMethod* m= elt->method(); 102 ciMethod* key_m = key->_m; 103 if (key_m < m) return -1; 104 else if (key_m > m) return 1; 105 else { 106 bool is_virtual = elt->is_virtual(); 107 bool key_virtual = key->_is_virtual; 108 if (key_virtual < is_virtual) return -1; 109 else if (key_virtual > is_virtual) return 1; 110 else return 0; 111 } 112 } 113 }; 114 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 115 #ifdef ASSERT 116 for (int i = 1; i < _intrinsics->length(); i++) { 117 CallGenerator* cg1 = _intrinsics->at(i-1); 118 CallGenerator* cg2 = _intrinsics->at(i); 119 assert(cg1->method() != cg2->method() 120 ? cg1->method() < cg2->method() 121 : cg1->is_virtual() < cg2->is_virtual(), 122 "compiler intrinsics list must stay sorted"); 123 } 124 #endif 125 IntrinsicDescPair pair(m, is_virtual); 126 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 127 } 128 129 void Compile::register_intrinsic(CallGenerator* cg) { 130 if (_intrinsics == NULL) { 131 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 132 } 133 int len = _intrinsics->length(); 134 bool found = false; 135 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 136 assert(!found, "registering twice"); 137 _intrinsics->insert_before(index, cg); 138 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 139 } 140 141 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 142 assert(m->is_loaded(), "don't try this on unloaded methods"); 143 if (_intrinsics != NULL) { 144 bool found = false; 145 int index = intrinsic_insertion_index(m, is_virtual, found); 146 if (found) { 147 return _intrinsics->at(index); 148 } 149 } 150 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 151 if (m->intrinsic_id() != vmIntrinsics::_none && 152 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 153 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 154 if (cg != NULL) { 155 // Save it for next time: 156 register_intrinsic(cg); 157 return cg; 158 } else { 159 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 160 } 161 } 162 return NULL; 163 } 164 165 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 166 // in library_call.cpp. 167 168 169 #ifndef PRODUCT 170 // statistics gathering... 171 172 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 173 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 174 175 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 176 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 177 int oflags = _intrinsic_hist_flags[id]; 178 assert(flags != 0, "what happened?"); 179 if (is_virtual) { 180 flags |= _intrinsic_virtual; 181 } 182 bool changed = (flags != oflags); 183 if ((flags & _intrinsic_worked) != 0) { 184 juint count = (_intrinsic_hist_count[id] += 1); 185 if (count == 1) { 186 changed = true; // first time 187 } 188 // increment the overall count also: 189 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 190 } 191 if (changed) { 192 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 193 // Something changed about the intrinsic's virtuality. 194 if ((flags & _intrinsic_virtual) != 0) { 195 // This is the first use of this intrinsic as a virtual call. 196 if (oflags != 0) { 197 // We already saw it as a non-virtual, so note both cases. 198 flags |= _intrinsic_both; 199 } 200 } else if ((oflags & _intrinsic_both) == 0) { 201 // This is the first use of this intrinsic as a non-virtual 202 flags |= _intrinsic_both; 203 } 204 } 205 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 206 } 207 // update the overall flags also: 208 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 209 return changed; 210 } 211 212 static char* format_flags(int flags, char* buf) { 213 buf[0] = 0; 214 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 215 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 216 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 217 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 218 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 219 if (buf[0] == 0) strcat(buf, ","); 220 assert(buf[0] == ',', "must be"); 221 return &buf[1]; 222 } 223 224 void Compile::print_intrinsic_statistics() { 225 char flagsbuf[100]; 226 ttyLocker ttyl; 227 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 228 tty->print_cr("Compiler intrinsic usage:"); 229 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 230 if (total == 0) total = 1; // avoid div0 in case of no successes 231 #define PRINT_STAT_LINE(name, c, f) \ 232 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 233 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 234 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 235 int flags = _intrinsic_hist_flags[id]; 236 juint count = _intrinsic_hist_count[id]; 237 if ((flags | count) != 0) { 238 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 239 } 240 } 241 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 242 if (xtty != NULL) xtty->tail("statistics"); 243 } 244 245 void Compile::print_statistics() { 246 { ttyLocker ttyl; 247 if (xtty != NULL) xtty->head("statistics type='opto'"); 248 Parse::print_statistics(); 249 PhaseCCP::print_statistics(); 250 PhaseRegAlloc::print_statistics(); 251 Scheduling::print_statistics(); 252 PhasePeephole::print_statistics(); 253 PhaseIdealLoop::print_statistics(); 254 if (xtty != NULL) xtty->tail("statistics"); 255 } 256 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 257 // put this under its own <statistics> element. 258 print_intrinsic_statistics(); 259 } 260 } 261 #endif //PRODUCT 262 263 // Support for bundling info 264 Bundle* Compile::node_bundling(const Node *n) { 265 assert(valid_bundle_info(n), "oob"); 266 return &_node_bundling_base[n->_idx]; 267 } 268 269 bool Compile::valid_bundle_info(const Node *n) { 270 return (_node_bundling_limit > n->_idx); 271 } 272 273 274 void Compile::gvn_replace_by(Node* n, Node* nn) { 275 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 276 Node* use = n->last_out(i); 277 bool is_in_table = initial_gvn()->hash_delete(use); 278 uint uses_found = 0; 279 for (uint j = 0; j < use->len(); j++) { 280 if (use->in(j) == n) { 281 if (j < use->req()) 282 use->set_req(j, nn); 283 else 284 use->set_prec(j, nn); 285 uses_found++; 286 } 287 } 288 if (is_in_table) { 289 // reinsert into table 290 initial_gvn()->hash_find_insert(use); 291 } 292 record_for_igvn(use); 293 i -= uses_found; // we deleted 1 or more copies of this edge 294 } 295 } 296 297 298 static inline bool not_a_node(const Node* n) { 299 if (n == NULL) return true; 300 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 301 if (*(address*)n == badAddress) return true; // kill by Node::destruct 302 return false; 303 } 304 305 // Identify all nodes that are reachable from below, useful. 306 // Use breadth-first pass that records state in a Unique_Node_List, 307 // recursive traversal is slower. 308 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 309 int estimated_worklist_size = live_nodes(); 310 useful.map( estimated_worklist_size, NULL ); // preallocate space 311 312 // Initialize worklist 313 if (root() != NULL) { useful.push(root()); } 314 // If 'top' is cached, declare it useful to preserve cached node 315 if( cached_top_node() ) { useful.push(cached_top_node()); } 316 317 // Push all useful nodes onto the list, breadthfirst 318 for( uint next = 0; next < useful.size(); ++next ) { 319 assert( next < unique(), "Unique useful nodes < total nodes"); 320 Node *n = useful.at(next); 321 uint max = n->len(); 322 for( uint i = 0; i < max; ++i ) { 323 Node *m = n->in(i); 324 if (not_a_node(m)) continue; 325 useful.push(m); 326 } 327 } 328 } 329 330 // Update dead_node_list with any missing dead nodes using useful 331 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 332 void Compile::update_dead_node_list(Unique_Node_List &useful) { 333 uint max_idx = unique(); 334 VectorSet& useful_node_set = useful.member_set(); 335 336 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 337 // If node with index node_idx is not in useful set, 338 // mark it as dead in dead node list. 339 if (! useful_node_set.test(node_idx) ) { 340 record_dead_node(node_idx); 341 } 342 } 343 } 344 345 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 346 int shift = 0; 347 for (int i = 0; i < inlines->length(); i++) { 348 CallGenerator* cg = inlines->at(i); 349 CallNode* call = cg->call_node(); 350 if (shift > 0) { 351 inlines->at_put(i-shift, cg); 352 } 353 if (!useful.member(call)) { 354 shift++; 355 } 356 } 357 inlines->trunc_to(inlines->length()-shift); 358 } 359 360 // Disconnect all useless nodes by disconnecting those at the boundary. 361 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 362 uint next = 0; 363 while (next < useful.size()) { 364 Node *n = useful.at(next++); 365 if (n->is_SafePoint()) { 366 // We're done with a parsing phase. Replaced nodes are not valid 367 // beyond that point. 368 n->as_SafePoint()->delete_replaced_nodes(); 369 } 370 // Use raw traversal of out edges since this code removes out edges 371 int max = n->outcnt(); 372 for (int j = 0; j < max; ++j) { 373 Node* child = n->raw_out(j); 374 if (! useful.member(child)) { 375 assert(!child->is_top() || child != top(), 376 "If top is cached in Compile object it is in useful list"); 377 // Only need to remove this out-edge to the useless node 378 n->raw_del_out(j); 379 --j; 380 --max; 381 } 382 } 383 if (n->outcnt() == 1 && n->has_special_unique_user()) { 384 record_for_igvn(n->unique_out()); 385 } 386 } 387 // Remove useless macro and predicate opaq nodes 388 for (int i = C->macro_count()-1; i >= 0; i--) { 389 Node* n = C->macro_node(i); 390 if (!useful.member(n)) { 391 remove_macro_node(n); 392 } 393 } 394 // Remove useless CastII nodes with range check dependency 395 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 396 Node* cast = range_check_cast_node(i); 397 if (!useful.member(cast)) { 398 remove_range_check_cast(cast); 399 } 400 } 401 // Remove useless expensive node 402 for (int i = C->expensive_count()-1; i >= 0; i--) { 403 Node* n = C->expensive_node(i); 404 if (!useful.member(n)) { 405 remove_expensive_node(n); 406 } 407 } 408 // Remove useless value type nodes 409 if (_value_type_nodes != NULL) { 410 _value_type_nodes->remove_useless_nodes(useful.member_set()); 411 } 412 // clean up the late inline lists 413 remove_useless_late_inlines(&_string_late_inlines, useful); 414 remove_useless_late_inlines(&_boxing_late_inlines, useful); 415 remove_useless_late_inlines(&_late_inlines, useful); 416 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 417 } 418 419 //------------------------------frame_size_in_words----------------------------- 420 // frame_slots in units of words 421 int Compile::frame_size_in_words() const { 422 // shift is 0 in LP32 and 1 in LP64 423 const int shift = (LogBytesPerWord - LogBytesPerInt); 424 int words = _frame_slots >> shift; 425 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 426 return words; 427 } 428 429 // To bang the stack of this compiled method we use the stack size 430 // that the interpreter would need in case of a deoptimization. This 431 // removes the need to bang the stack in the deoptimization blob which 432 // in turn simplifies stack overflow handling. 433 int Compile::bang_size_in_bytes() const { 434 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 435 } 436 437 // ============================================================================ 438 //------------------------------CompileWrapper--------------------------------- 439 class CompileWrapper : public StackObj { 440 Compile *const _compile; 441 public: 442 CompileWrapper(Compile* compile); 443 444 ~CompileWrapper(); 445 }; 446 447 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 448 // the Compile* pointer is stored in the current ciEnv: 449 ciEnv* env = compile->env(); 450 assert(env == ciEnv::current(), "must already be a ciEnv active"); 451 assert(env->compiler_data() == NULL, "compile already active?"); 452 env->set_compiler_data(compile); 453 assert(compile == Compile::current(), "sanity"); 454 455 compile->set_type_dict(NULL); 456 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 457 compile->clone_map().set_clone_idx(0); 458 compile->set_type_hwm(NULL); 459 compile->set_type_last_size(0); 460 compile->set_last_tf(NULL, NULL); 461 compile->set_indexSet_arena(NULL); 462 compile->set_indexSet_free_block_list(NULL); 463 compile->init_type_arena(); 464 Type::Initialize(compile); 465 _compile->set_scratch_buffer_blob(NULL); 466 _compile->begin_method(); 467 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 468 } 469 CompileWrapper::~CompileWrapper() { 470 _compile->end_method(); 471 if (_compile->scratch_buffer_blob() != NULL) 472 BufferBlob::free(_compile->scratch_buffer_blob()); 473 _compile->env()->set_compiler_data(NULL); 474 } 475 476 477 //----------------------------print_compile_messages--------------------------- 478 void Compile::print_compile_messages() { 479 #ifndef PRODUCT 480 // Check if recompiling 481 if (_subsume_loads == false && PrintOpto) { 482 // Recompiling without allowing machine instructions to subsume loads 483 tty->print_cr("*********************************************************"); 484 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 485 tty->print_cr("*********************************************************"); 486 } 487 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 488 // Recompiling without escape analysis 489 tty->print_cr("*********************************************************"); 490 tty->print_cr("** Bailout: Recompile without escape analysis **"); 491 tty->print_cr("*********************************************************"); 492 } 493 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 494 // Recompiling without boxing elimination 495 tty->print_cr("*********************************************************"); 496 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 497 tty->print_cr("*********************************************************"); 498 } 499 if (C->directive()->BreakAtCompileOption) { 500 // Open the debugger when compiling this method. 501 tty->print("### Breaking when compiling: "); 502 method()->print_short_name(); 503 tty->cr(); 504 BREAKPOINT; 505 } 506 507 if( PrintOpto ) { 508 if (is_osr_compilation()) { 509 tty->print("[OSR]%3d", _compile_id); 510 } else { 511 tty->print("%3d", _compile_id); 512 } 513 } 514 #endif 515 } 516 517 518 //-----------------------init_scratch_buffer_blob------------------------------ 519 // Construct a temporary BufferBlob and cache it for this compile. 520 void Compile::init_scratch_buffer_blob(int const_size) { 521 // If there is already a scratch buffer blob allocated and the 522 // constant section is big enough, use it. Otherwise free the 523 // current and allocate a new one. 524 BufferBlob* blob = scratch_buffer_blob(); 525 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 526 // Use the current blob. 527 } else { 528 if (blob != NULL) { 529 BufferBlob::free(blob); 530 } 531 532 ResourceMark rm; 533 _scratch_const_size = const_size; 534 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 535 blob = BufferBlob::create("Compile::scratch_buffer", size); 536 // Record the buffer blob for next time. 537 set_scratch_buffer_blob(blob); 538 // Have we run out of code space? 539 if (scratch_buffer_blob() == NULL) { 540 // Let CompilerBroker disable further compilations. 541 record_failure("Not enough space for scratch buffer in CodeCache"); 542 return; 543 } 544 } 545 546 // Initialize the relocation buffers 547 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 548 set_scratch_locs_memory(locs_buf); 549 } 550 551 552 //-----------------------scratch_emit_size------------------------------------- 553 // Helper function that computes size by emitting code 554 uint Compile::scratch_emit_size(const Node* n) { 555 // Start scratch_emit_size section. 556 set_in_scratch_emit_size(true); 557 558 // Emit into a trash buffer and count bytes emitted. 559 // This is a pretty expensive way to compute a size, 560 // but it works well enough if seldom used. 561 // All common fixed-size instructions are given a size 562 // method by the AD file. 563 // Note that the scratch buffer blob and locs memory are 564 // allocated at the beginning of the compile task, and 565 // may be shared by several calls to scratch_emit_size. 566 // The allocation of the scratch buffer blob is particularly 567 // expensive, since it has to grab the code cache lock. 568 BufferBlob* blob = this->scratch_buffer_blob(); 569 assert(blob != NULL, "Initialize BufferBlob at start"); 570 assert(blob->size() > MAX_inst_size, "sanity"); 571 relocInfo* locs_buf = scratch_locs_memory(); 572 address blob_begin = blob->content_begin(); 573 address blob_end = (address)locs_buf; 574 assert(blob->contains(blob_end), "sanity"); 575 CodeBuffer buf(blob_begin, blob_end - blob_begin); 576 buf.initialize_consts_size(_scratch_const_size); 577 buf.initialize_stubs_size(MAX_stubs_size); 578 assert(locs_buf != NULL, "sanity"); 579 int lsize = MAX_locs_size / 3; 580 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 581 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 582 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 583 // Mark as scratch buffer. 584 buf.consts()->set_scratch_emit(); 585 buf.insts()->set_scratch_emit(); 586 buf.stubs()->set_scratch_emit(); 587 588 // Do the emission. 589 590 Label fakeL; // Fake label for branch instructions. 591 Label* saveL = NULL; 592 uint save_bnum = 0; 593 bool is_branch = n->is_MachBranch(); 594 if (is_branch) { 595 MacroAssembler masm(&buf); 596 masm.bind(fakeL); 597 n->as_MachBranch()->save_label(&saveL, &save_bnum); 598 n->as_MachBranch()->label_set(&fakeL, 0); 599 } 600 n->emit(buf, this->regalloc()); 601 602 // Emitting into the scratch buffer should not fail 603 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 604 605 if (is_branch) // Restore label. 606 n->as_MachBranch()->label_set(saveL, save_bnum); 607 608 // End scratch_emit_size section. 609 set_in_scratch_emit_size(false); 610 611 return buf.insts_size(); 612 } 613 614 615 // ============================================================================ 616 //------------------------------Compile standard------------------------------- 617 debug_only( int Compile::_debug_idx = 100000; ) 618 619 // Compile a method. entry_bci is -1 for normal compilations and indicates 620 // the continuation bci for on stack replacement. 621 622 623 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 624 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 625 : Phase(Compiler), 626 _env(ci_env), 627 _directive(directive), 628 _log(ci_env->log()), 629 _compile_id(ci_env->compile_id()), 630 _save_argument_registers(false), 631 _stub_name(NULL), 632 _stub_function(NULL), 633 _stub_entry_point(NULL), 634 _method(target), 635 _entry_bci(osr_bci), 636 _initial_gvn(NULL), 637 _for_igvn(NULL), 638 _warm_calls(NULL), 639 _subsume_loads(subsume_loads), 640 _do_escape_analysis(do_escape_analysis), 641 _eliminate_boxing(eliminate_boxing), 642 _failure_reason(NULL), 643 _code_buffer("Compile::Fill_buffer"), 644 _orig_pc_slot(0), 645 _orig_pc_slot_offset_in_bytes(0), 646 _has_method_handle_invokes(false), 647 _mach_constant_base_node(NULL), 648 _node_bundling_limit(0), 649 _node_bundling_base(NULL), 650 _java_calls(0), 651 _inner_loops(0), 652 _scratch_const_size(-1), 653 _in_scratch_emit_size(false), 654 _dead_node_list(comp_arena()), 655 _dead_node_count(0), 656 #ifndef PRODUCT 657 _trace_opto_output(directive->TraceOptoOutputOption), 658 _in_dump_cnt(0), 659 _printer(IdealGraphPrinter::printer()), 660 #endif 661 _congraph(NULL), 662 _comp_arena(mtCompiler), 663 _node_arena(mtCompiler), 664 _old_arena(mtCompiler), 665 _Compile_types(mtCompiler), 666 _replay_inline_data(NULL), 667 _late_inlines(comp_arena(), 2, 0, NULL), 668 _string_late_inlines(comp_arena(), 2, 0, NULL), 669 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 670 _late_inlines_pos(0), 671 _number_of_mh_late_inlines(0), 672 _inlining_progress(false), 673 _inlining_incrementally(false), 674 _print_inlining_list(NULL), 675 _print_inlining_stream(NULL), 676 _print_inlining_idx(0), 677 _print_inlining_output(NULL), 678 _interpreter_frame_size(0), 679 _max_node_limit(MaxNodeLimit), 680 _has_reserved_stack_access(target->has_reserved_stack_access()) { 681 C = this; 682 #ifndef PRODUCT 683 if (_printer != NULL) { 684 _printer->set_compile(this); 685 } 686 #endif 687 CompileWrapper cw(this); 688 689 if (CITimeVerbose) { 690 tty->print(" "); 691 target->holder()->name()->print(); 692 tty->print("."); 693 target->print_short_name(); 694 tty->print(" "); 695 } 696 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 697 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 698 699 #ifndef PRODUCT 700 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 701 if (!print_opto_assembly) { 702 bool print_assembly = directive->PrintAssemblyOption; 703 if (print_assembly && !Disassembler::can_decode()) { 704 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 705 print_opto_assembly = true; 706 } 707 } 708 set_print_assembly(print_opto_assembly); 709 set_parsed_irreducible_loop(false); 710 711 if (directive->ReplayInlineOption) { 712 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 713 } 714 #endif 715 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 716 set_print_intrinsics(directive->PrintIntrinsicsOption); 717 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 718 719 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 720 // Make sure the method being compiled gets its own MDO, 721 // so we can at least track the decompile_count(). 722 // Need MDO to record RTM code generation state. 723 method()->ensure_method_data(); 724 } 725 726 Init(::AliasLevel); 727 728 729 print_compile_messages(); 730 731 _ilt = InlineTree::build_inline_tree_root(); 732 733 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 734 assert(num_alias_types() >= AliasIdxRaw, ""); 735 736 #define MINIMUM_NODE_HASH 1023 737 // Node list that Iterative GVN will start with 738 Unique_Node_List for_igvn(comp_arena()); 739 set_for_igvn(&for_igvn); 740 741 // GVN that will be run immediately on new nodes 742 uint estimated_size = method()->code_size()*4+64; 743 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 744 PhaseGVN gvn(node_arena(), estimated_size); 745 set_initial_gvn(&gvn); 746 747 print_inlining_init(); 748 { // Scope for timing the parser 749 TracePhase tp("parse", &timers[_t_parser]); 750 751 // Put top into the hash table ASAP. 752 initial_gvn()->transform_no_reclaim(top()); 753 754 // Set up tf(), start(), and find a CallGenerator. 755 CallGenerator* cg = NULL; 756 if (is_osr_compilation()) { 757 const TypeTuple *domain = StartOSRNode::osr_domain(); 758 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 759 init_tf(TypeFunc::make(domain, range)); 760 StartNode* s = new StartOSRNode(root(), domain); 761 initial_gvn()->set_type_bottom(s); 762 init_start(s); 763 cg = CallGenerator::for_osr(method(), entry_bci()); 764 } else { 765 // Normal case. 766 init_tf(TypeFunc::make(method())); 767 StartNode* s = new StartNode(root(), tf()->domain_cc()); 768 initial_gvn()->set_type_bottom(s); 769 init_start(s); 770 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 771 // With java.lang.ref.reference.get() we must go through the 772 // intrinsic when G1 is enabled - even when get() is the root 773 // method of the compile - so that, if necessary, the value in 774 // the referent field of the reference object gets recorded by 775 // the pre-barrier code. 776 // Specifically, if G1 is enabled, the value in the referent 777 // field is recorded by the G1 SATB pre barrier. This will 778 // result in the referent being marked live and the reference 779 // object removed from the list of discovered references during 780 // reference processing. 781 cg = find_intrinsic(method(), false); 782 } 783 if (cg == NULL) { 784 float past_uses = method()->interpreter_invocation_count(); 785 float expected_uses = past_uses; 786 cg = CallGenerator::for_inline(method(), expected_uses); 787 } 788 } 789 if (failing()) return; 790 if (cg == NULL) { 791 record_method_not_compilable("cannot parse method"); 792 return; 793 } 794 JVMState* jvms = build_start_state(start(), tf()); 795 if ((jvms = cg->generate(jvms)) == NULL) { 796 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 797 record_method_not_compilable("method parse failed"); 798 } 799 return; 800 } 801 GraphKit kit(jvms); 802 803 if (!kit.stopped()) { 804 // Accept return values, and transfer control we know not where. 805 // This is done by a special, unique ReturnNode bound to root. 806 return_values(kit.jvms()); 807 } 808 809 if (kit.has_exceptions()) { 810 // Any exceptions that escape from this call must be rethrown 811 // to whatever caller is dynamically above us on the stack. 812 // This is done by a special, unique RethrowNode bound to root. 813 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 814 } 815 816 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 817 818 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 819 inline_string_calls(true); 820 } 821 822 if (failing()) return; 823 824 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 825 826 // Remove clutter produced by parsing. 827 if (!failing()) { 828 ResourceMark rm; 829 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 830 } 831 } 832 833 // Note: Large methods are capped off in do_one_bytecode(). 834 if (failing()) return; 835 836 // After parsing, node notes are no longer automagic. 837 // They must be propagated by register_new_node_with_optimizer(), 838 // clone(), or the like. 839 set_default_node_notes(NULL); 840 841 for (;;) { 842 int successes = Inline_Warm(); 843 if (failing()) return; 844 if (successes == 0) break; 845 } 846 847 // Drain the list. 848 Finish_Warm(); 849 #ifndef PRODUCT 850 if (_printer && _printer->should_print(1)) { 851 _printer->print_inlining(); 852 } 853 #endif 854 855 if (failing()) return; 856 NOT_PRODUCT( verify_graph_edges(); ) 857 858 // Now optimize 859 Optimize(); 860 if (failing()) return; 861 NOT_PRODUCT( verify_graph_edges(); ) 862 863 #ifndef PRODUCT 864 if (PrintIdeal) { 865 ttyLocker ttyl; // keep the following output all in one block 866 // This output goes directly to the tty, not the compiler log. 867 // To enable tools to match it up with the compilation activity, 868 // be sure to tag this tty output with the compile ID. 869 if (xtty != NULL) { 870 xtty->head("ideal compile_id='%d'%s", compile_id(), 871 is_osr_compilation() ? " compile_kind='osr'" : 872 ""); 873 } 874 root()->dump(9999); 875 if (xtty != NULL) { 876 xtty->tail("ideal"); 877 } 878 } 879 #endif 880 881 NOT_PRODUCT( verify_barriers(); ) 882 883 // Dump compilation data to replay it. 884 if (directive->DumpReplayOption) { 885 env()->dump_replay_data(_compile_id); 886 } 887 if (directive->DumpInlineOption && (ilt() != NULL)) { 888 env()->dump_inline_data(_compile_id); 889 } 890 891 // Now that we know the size of all the monitors we can add a fixed slot 892 // for the original deopt pc. 893 894 _orig_pc_slot = fixed_slots(); 895 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 896 set_fixed_slots(next_slot); 897 898 // Compute when to use implicit null checks. Used by matching trap based 899 // nodes and NullCheck optimization. 900 set_allowed_deopt_reasons(); 901 902 // Now generate code 903 Code_Gen(); 904 if (failing()) return; 905 906 // Check if we want to skip execution of all compiled code. 907 { 908 #ifndef PRODUCT 909 if (OptoNoExecute) { 910 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 911 return; 912 } 913 #endif 914 TracePhase tp("install_code", &timers[_t_registerMethod]); 915 916 if (is_osr_compilation()) { 917 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 918 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 919 } else { 920 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 921 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 922 } 923 924 env()->register_method(_method, _entry_bci, 925 &_code_offsets, 926 _orig_pc_slot_offset_in_bytes, 927 code_buffer(), 928 frame_size_in_words(), _oop_map_set, 929 &_handler_table, &_inc_table, 930 compiler, 931 has_unsafe_access(), 932 SharedRuntime::is_wide_vector(max_vector_size()), 933 rtm_state() 934 ); 935 936 if (log() != NULL) // Print code cache state into compiler log 937 log()->code_cache_state(); 938 } 939 } 940 941 //------------------------------Compile---------------------------------------- 942 // Compile a runtime stub 943 Compile::Compile( ciEnv* ci_env, 944 TypeFunc_generator generator, 945 address stub_function, 946 const char *stub_name, 947 int is_fancy_jump, 948 bool pass_tls, 949 bool save_arg_registers, 950 bool return_pc, 951 DirectiveSet* directive) 952 : Phase(Compiler), 953 _env(ci_env), 954 _directive(directive), 955 _log(ci_env->log()), 956 _compile_id(0), 957 _save_argument_registers(save_arg_registers), 958 _method(NULL), 959 _stub_name(stub_name), 960 _stub_function(stub_function), 961 _stub_entry_point(NULL), 962 _entry_bci(InvocationEntryBci), 963 _initial_gvn(NULL), 964 _for_igvn(NULL), 965 _warm_calls(NULL), 966 _orig_pc_slot(0), 967 _orig_pc_slot_offset_in_bytes(0), 968 _subsume_loads(true), 969 _do_escape_analysis(false), 970 _eliminate_boxing(false), 971 _failure_reason(NULL), 972 _code_buffer("Compile::Fill_buffer"), 973 _has_method_handle_invokes(false), 974 _mach_constant_base_node(NULL), 975 _node_bundling_limit(0), 976 _node_bundling_base(NULL), 977 _java_calls(0), 978 _inner_loops(0), 979 #ifndef PRODUCT 980 _trace_opto_output(directive->TraceOptoOutputOption), 981 _in_dump_cnt(0), 982 _printer(NULL), 983 #endif 984 _comp_arena(mtCompiler), 985 _node_arena(mtCompiler), 986 _old_arena(mtCompiler), 987 _Compile_types(mtCompiler), 988 _dead_node_list(comp_arena()), 989 _dead_node_count(0), 990 _congraph(NULL), 991 _replay_inline_data(NULL), 992 _number_of_mh_late_inlines(0), 993 _inlining_progress(false), 994 _inlining_incrementally(false), 995 _print_inlining_list(NULL), 996 _print_inlining_stream(NULL), 997 _print_inlining_idx(0), 998 _print_inlining_output(NULL), 999 _allowed_reasons(0), 1000 _interpreter_frame_size(0), 1001 _max_node_limit(MaxNodeLimit), 1002 _has_reserved_stack_access(false) { 1003 C = this; 1004 1005 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1006 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1007 1008 #ifndef PRODUCT 1009 set_print_assembly(PrintFrameConverterAssembly); 1010 set_parsed_irreducible_loop(false); 1011 #endif 1012 set_has_irreducible_loop(false); // no loops 1013 1014 CompileWrapper cw(this); 1015 Init(/*AliasLevel=*/ 0); 1016 init_tf((*generator)()); 1017 1018 { 1019 // The following is a dummy for the sake of GraphKit::gen_stub 1020 Unique_Node_List for_igvn(comp_arena()); 1021 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1022 PhaseGVN gvn(Thread::current()->resource_area(),255); 1023 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1024 gvn.transform_no_reclaim(top()); 1025 1026 GraphKit kit; 1027 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1028 } 1029 1030 NOT_PRODUCT( verify_graph_edges(); ) 1031 Code_Gen(); 1032 if (failing()) return; 1033 1034 1035 // Entry point will be accessed using compile->stub_entry_point(); 1036 if (code_buffer() == NULL) { 1037 Matcher::soft_match_failure(); 1038 } else { 1039 if (PrintAssembly && (WizardMode || Verbose)) 1040 tty->print_cr("### Stub::%s", stub_name); 1041 1042 if (!failing()) { 1043 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1044 1045 // Make the NMethod 1046 // For now we mark the frame as never safe for profile stackwalking 1047 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1048 code_buffer(), 1049 CodeOffsets::frame_never_safe, 1050 // _code_offsets.value(CodeOffsets::Frame_Complete), 1051 frame_size_in_words(), 1052 _oop_map_set, 1053 save_arg_registers); 1054 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1055 1056 _stub_entry_point = rs->entry_point(); 1057 } 1058 } 1059 } 1060 1061 //------------------------------Init------------------------------------------- 1062 // Prepare for a single compilation 1063 void Compile::Init(int aliaslevel) { 1064 _unique = 0; 1065 _regalloc = NULL; 1066 1067 _tf = NULL; // filled in later 1068 _top = NULL; // cached later 1069 _matcher = NULL; // filled in later 1070 _cfg = NULL; // filled in later 1071 1072 set_24_bit_selection_and_mode(Use24BitFP, false); 1073 1074 _node_note_array = NULL; 1075 _default_node_notes = NULL; 1076 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1077 1078 _immutable_memory = NULL; // filled in at first inquiry 1079 1080 // Globally visible Nodes 1081 // First set TOP to NULL to give safe behavior during creation of RootNode 1082 set_cached_top_node(NULL); 1083 set_root(new RootNode()); 1084 // Now that you have a Root to point to, create the real TOP 1085 set_cached_top_node( new ConNode(Type::TOP) ); 1086 set_recent_alloc(NULL, NULL); 1087 1088 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1089 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1090 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1091 env()->set_dependencies(new Dependencies(env())); 1092 1093 _fixed_slots = 0; 1094 set_has_split_ifs(false); 1095 set_has_loops(has_method() && method()->has_loops()); // first approximation 1096 set_has_stringbuilder(false); 1097 set_has_boxed_value(false); 1098 _trap_can_recompile = false; // no traps emitted yet 1099 _major_progress = true; // start out assuming good things will happen 1100 set_has_unsafe_access(false); 1101 set_max_vector_size(0); 1102 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1103 set_decompile_count(0); 1104 1105 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1106 set_num_loop_opts(LoopOptsCount); 1107 set_do_inlining(Inline); 1108 set_max_inline_size(MaxInlineSize); 1109 set_freq_inline_size(FreqInlineSize); 1110 set_do_scheduling(OptoScheduling); 1111 set_do_count_invocations(false); 1112 set_do_method_data_update(false); 1113 1114 set_do_vector_loop(false); 1115 1116 if (AllowVectorizeOnDemand) { 1117 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1118 set_do_vector_loop(true); 1119 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());}) 1120 } else if (has_method() && method()->name() != 0 && 1121 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1122 set_do_vector_loop(true); 1123 } 1124 } 1125 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1126 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());}) 1127 1128 set_age_code(has_method() && method()->profile_aging()); 1129 set_rtm_state(NoRTM); // No RTM lock eliding by default 1130 _max_node_limit = _directive->MaxNodeLimitOption; 1131 1132 #if INCLUDE_RTM_OPT 1133 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1134 int rtm_state = method()->method_data()->rtm_state(); 1135 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1136 // Don't generate RTM lock eliding code. 1137 set_rtm_state(NoRTM); 1138 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1139 // Generate RTM lock eliding code without abort ratio calculation code. 1140 set_rtm_state(UseRTM); 1141 } else if (UseRTMDeopt) { 1142 // Generate RTM lock eliding code and include abort ratio calculation 1143 // code if UseRTMDeopt is on. 1144 set_rtm_state(ProfileRTM); 1145 } 1146 } 1147 #endif 1148 if (debug_info()->recording_non_safepoints()) { 1149 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1150 (comp_arena(), 8, 0, NULL)); 1151 set_default_node_notes(Node_Notes::make(this)); 1152 } 1153 1154 // // -- Initialize types before each compile -- 1155 // // Update cached type information 1156 // if( _method && _method->constants() ) 1157 // Type::update_loaded_types(_method, _method->constants()); 1158 1159 // Init alias_type map. 1160 if (!_do_escape_analysis && aliaslevel == 3) 1161 aliaslevel = 2; // No unique types without escape analysis 1162 _AliasLevel = aliaslevel; 1163 const int grow_ats = 16; 1164 _max_alias_types = grow_ats; 1165 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1166 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1167 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1168 { 1169 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1170 } 1171 // Initialize the first few types. 1172 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1173 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1174 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1175 _num_alias_types = AliasIdxRaw+1; 1176 // Zero out the alias type cache. 1177 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1178 // A NULL adr_type hits in the cache right away. Preload the right answer. 1179 probe_alias_cache(NULL)->_index = AliasIdxTop; 1180 1181 _intrinsics = NULL; 1182 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1183 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1184 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1185 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1186 _value_type_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 1187 register_library_intrinsics(); 1188 } 1189 1190 //---------------------------init_start---------------------------------------- 1191 // Install the StartNode on this compile object. 1192 void Compile::init_start(StartNode* s) { 1193 if (failing()) 1194 return; // already failing 1195 assert(s == start(), ""); 1196 } 1197 1198 /** 1199 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1200 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1201 * the ideal graph. 1202 */ 1203 StartNode* Compile::start() const { 1204 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1205 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1206 Node* start = root()->fast_out(i); 1207 if (start->is_Start()) { 1208 return start->as_Start(); 1209 } 1210 } 1211 fatal("Did not find Start node!"); 1212 return NULL; 1213 } 1214 1215 //-------------------------------immutable_memory------------------------------------- 1216 // Access immutable memory 1217 Node* Compile::immutable_memory() { 1218 if (_immutable_memory != NULL) { 1219 return _immutable_memory; 1220 } 1221 StartNode* s = start(); 1222 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1223 Node *p = s->fast_out(i); 1224 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1225 _immutable_memory = p; 1226 return _immutable_memory; 1227 } 1228 } 1229 ShouldNotReachHere(); 1230 return NULL; 1231 } 1232 1233 //----------------------set_cached_top_node------------------------------------ 1234 // Install the cached top node, and make sure Node::is_top works correctly. 1235 void Compile::set_cached_top_node(Node* tn) { 1236 if (tn != NULL) verify_top(tn); 1237 Node* old_top = _top; 1238 _top = tn; 1239 // Calling Node::setup_is_top allows the nodes the chance to adjust 1240 // their _out arrays. 1241 if (_top != NULL) _top->setup_is_top(); 1242 if (old_top != NULL) old_top->setup_is_top(); 1243 assert(_top == NULL || top()->is_top(), ""); 1244 } 1245 1246 #ifdef ASSERT 1247 uint Compile::count_live_nodes_by_graph_walk() { 1248 Unique_Node_List useful(comp_arena()); 1249 // Get useful node list by walking the graph. 1250 identify_useful_nodes(useful); 1251 return useful.size(); 1252 } 1253 1254 void Compile::print_missing_nodes() { 1255 1256 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1257 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1258 return; 1259 } 1260 1261 // This is an expensive function. It is executed only when the user 1262 // specifies VerifyIdealNodeCount option or otherwise knows the 1263 // additional work that needs to be done to identify reachable nodes 1264 // by walking the flow graph and find the missing ones using 1265 // _dead_node_list. 1266 1267 Unique_Node_List useful(comp_arena()); 1268 // Get useful node list by walking the graph. 1269 identify_useful_nodes(useful); 1270 1271 uint l_nodes = C->live_nodes(); 1272 uint l_nodes_by_walk = useful.size(); 1273 1274 if (l_nodes != l_nodes_by_walk) { 1275 if (_log != NULL) { 1276 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1277 _log->stamp(); 1278 _log->end_head(); 1279 } 1280 VectorSet& useful_member_set = useful.member_set(); 1281 int last_idx = l_nodes_by_walk; 1282 for (int i = 0; i < last_idx; i++) { 1283 if (useful_member_set.test(i)) { 1284 if (_dead_node_list.test(i)) { 1285 if (_log != NULL) { 1286 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1287 } 1288 if (PrintIdealNodeCount) { 1289 // Print the log message to tty 1290 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1291 useful.at(i)->dump(); 1292 } 1293 } 1294 } 1295 else if (! _dead_node_list.test(i)) { 1296 if (_log != NULL) { 1297 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1298 } 1299 if (PrintIdealNodeCount) { 1300 // Print the log message to tty 1301 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1302 } 1303 } 1304 } 1305 if (_log != NULL) { 1306 _log->tail("mismatched_nodes"); 1307 } 1308 } 1309 } 1310 void Compile::record_modified_node(Node* n) { 1311 if (_modified_nodes != NULL && !_inlining_incrementally && 1312 n->outcnt() != 0 && !n->is_Con()) { 1313 _modified_nodes->push(n); 1314 } 1315 } 1316 1317 void Compile::remove_modified_node(Node* n) { 1318 if (_modified_nodes != NULL) { 1319 _modified_nodes->remove(n); 1320 } 1321 } 1322 #endif 1323 1324 #ifndef PRODUCT 1325 void Compile::verify_top(Node* tn) const { 1326 if (tn != NULL) { 1327 assert(tn->is_Con(), "top node must be a constant"); 1328 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1329 assert(tn->in(0) != NULL, "must have live top node"); 1330 } 1331 } 1332 #endif 1333 1334 1335 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1336 1337 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1338 guarantee(arr != NULL, ""); 1339 int num_blocks = arr->length(); 1340 if (grow_by < num_blocks) grow_by = num_blocks; 1341 int num_notes = grow_by * _node_notes_block_size; 1342 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1343 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1344 while (num_notes > 0) { 1345 arr->append(notes); 1346 notes += _node_notes_block_size; 1347 num_notes -= _node_notes_block_size; 1348 } 1349 assert(num_notes == 0, "exact multiple, please"); 1350 } 1351 1352 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1353 if (source == NULL || dest == NULL) return false; 1354 1355 if (dest->is_Con()) 1356 return false; // Do not push debug info onto constants. 1357 1358 #ifdef ASSERT 1359 // Leave a bread crumb trail pointing to the original node: 1360 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1361 dest->set_debug_orig(source); 1362 } 1363 #endif 1364 1365 if (node_note_array() == NULL) 1366 return false; // Not collecting any notes now. 1367 1368 // This is a copy onto a pre-existing node, which may already have notes. 1369 // If both nodes have notes, do not overwrite any pre-existing notes. 1370 Node_Notes* source_notes = node_notes_at(source->_idx); 1371 if (source_notes == NULL || source_notes->is_clear()) return false; 1372 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1373 if (dest_notes == NULL || dest_notes->is_clear()) { 1374 return set_node_notes_at(dest->_idx, source_notes); 1375 } 1376 1377 Node_Notes merged_notes = (*source_notes); 1378 // The order of operations here ensures that dest notes will win... 1379 merged_notes.update_from(dest_notes); 1380 return set_node_notes_at(dest->_idx, &merged_notes); 1381 } 1382 1383 1384 //--------------------------allow_range_check_smearing------------------------- 1385 // Gating condition for coalescing similar range checks. 1386 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1387 // single covering check that is at least as strong as any of them. 1388 // If the optimization succeeds, the simplified (strengthened) range check 1389 // will always succeed. If it fails, we will deopt, and then give up 1390 // on the optimization. 1391 bool Compile::allow_range_check_smearing() const { 1392 // If this method has already thrown a range-check, 1393 // assume it was because we already tried range smearing 1394 // and it failed. 1395 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1396 return !already_trapped; 1397 } 1398 1399 1400 //------------------------------flatten_alias_type----------------------------- 1401 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1402 int offset = tj->offset(); 1403 TypePtr::PTR ptr = tj->ptr(); 1404 1405 // Known instance (scalarizable allocation) alias only with itself. 1406 bool is_known_inst = tj->isa_oopptr() != NULL && 1407 tj->is_oopptr()->is_known_instance(); 1408 1409 // Process weird unsafe references. 1410 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1411 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1412 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1413 tj = TypeOopPtr::BOTTOM; 1414 ptr = tj->ptr(); 1415 offset = tj->offset(); 1416 } 1417 1418 // Array pointers need some flattening 1419 const TypeAryPtr *ta = tj->isa_aryptr(); 1420 if (ta && ta->is_stable()) { 1421 // Erase stability property for alias analysis. 1422 tj = ta = ta->cast_to_stable(false); 1423 } 1424 if( ta && is_known_inst ) { 1425 if ( offset != Type::OffsetBot && 1426 offset > arrayOopDesc::length_offset_in_bytes() ) { 1427 offset = Type::OffsetBot; // Flatten constant access into array body only 1428 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id()); 1429 } 1430 } else if( ta && _AliasLevel >= 2 ) { 1431 // For arrays indexed by constant indices, we flatten the alias 1432 // space to include all of the array body. Only the header, klass 1433 // and array length can be accessed un-aliased. 1434 // For flattened value type array, each field has its own slice so 1435 // we must include the field offset. 1436 if( offset != Type::OffsetBot ) { 1437 if( ta->const_oop() ) { // MethodData* or Method* 1438 offset = Type::OffsetBot; // Flatten constant access into array body 1439 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1440 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1441 // range is OK as-is. 1442 tj = ta = TypeAryPtr::RANGE; 1443 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1444 tj = TypeInstPtr::KLASS; // all klass loads look alike 1445 ta = TypeAryPtr::RANGE; // generic ignored junk 1446 ptr = TypePtr::BotPTR; 1447 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1448 tj = TypeInstPtr::MARK; 1449 ta = TypeAryPtr::RANGE; // generic ignored junk 1450 ptr = TypePtr::BotPTR; 1451 } else { // Random constant offset into array body 1452 offset = Type::OffsetBot; // Flatten constant access into array body 1453 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1454 } 1455 } 1456 // Arrays of fixed size alias with arrays of unknown size. 1457 if (ta->size() != TypeInt::POS) { 1458 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1459 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1460 } 1461 // Arrays of known objects become arrays of unknown objects. 1462 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1463 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1464 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1465 } 1466 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1467 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1468 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1469 } 1470 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1471 // cannot be distinguished by bytecode alone. 1472 if (ta->elem() == TypeInt::BOOL) { 1473 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1474 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1475 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset()); 1476 } 1477 // During the 2nd round of IterGVN, NotNull castings are removed. 1478 // Make sure the Bottom and NotNull variants alias the same. 1479 // Also, make sure exact and non-exact variants alias the same. 1480 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1481 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1482 } 1483 } 1484 1485 // Oop pointers need some flattening 1486 const TypeInstPtr *to = tj->isa_instptr(); 1487 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1488 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1489 if( ptr == TypePtr::Constant ) { 1490 if (to->klass() != ciEnv::current()->Class_klass() || 1491 offset < k->size_helper() * wordSize) { 1492 // No constant oop pointers (such as Strings); they alias with 1493 // unknown strings. 1494 assert(!is_known_inst, "not scalarizable allocation"); 1495 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1496 } 1497 } else if( is_known_inst ) { 1498 tj = to; // Keep NotNull and klass_is_exact for instance type 1499 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1500 // During the 2nd round of IterGVN, NotNull castings are removed. 1501 // Make sure the Bottom and NotNull variants alias the same. 1502 // Also, make sure exact and non-exact variants alias the same. 1503 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1504 } 1505 if (to->speculative() != NULL) { 1506 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->instance_id()); 1507 } 1508 // Canonicalize the holder of this field 1509 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1510 // First handle header references such as a LoadKlassNode, even if the 1511 // object's klass is unloaded at compile time (4965979). 1512 if (!is_known_inst) { // Do it only for non-instance types 1513 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset)); 1514 } 1515 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1516 // Static fields are in the space above the normal instance 1517 // fields in the java.lang.Class instance. 1518 if (to->klass() != ciEnv::current()->Class_klass()) { 1519 to = NULL; 1520 tj = TypeOopPtr::BOTTOM; 1521 offset = tj->offset(); 1522 } 1523 } else { 1524 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1525 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1526 if( is_known_inst ) { 1527 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), to->instance_id()); 1528 } else { 1529 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset)); 1530 } 1531 } 1532 } 1533 } 1534 1535 // Klass pointers to object array klasses need some flattening 1536 const TypeKlassPtr *tk = tj->isa_klassptr(); 1537 if( tk ) { 1538 // If we are referencing a field within a Klass, we need 1539 // to assume the worst case of an Object. Both exact and 1540 // inexact types must flatten to the same alias class so 1541 // use NotNull as the PTR. 1542 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1543 1544 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1545 TypeKlassPtr::OBJECT->klass(), 1546 Type::Offset(offset)); 1547 } 1548 1549 ciKlass* klass = tk->klass(); 1550 if (klass != NULL && klass->is_obj_array_klass()) { 1551 ciKlass* k = TypeAryPtr::OOPS->klass(); 1552 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1553 k = TypeInstPtr::BOTTOM->klass(); 1554 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset)); 1555 } 1556 1557 // Check for precise loads from the primary supertype array and force them 1558 // to the supertype cache alias index. Check for generic array loads from 1559 // the primary supertype array and also force them to the supertype cache 1560 // alias index. Since the same load can reach both, we need to merge 1561 // these 2 disparate memories into the same alias class. Since the 1562 // primary supertype array is read-only, there's no chance of confusion 1563 // where we bypass an array load and an array store. 1564 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1565 if (offset == Type::OffsetBot || 1566 (offset >= primary_supers_offset && 1567 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1568 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1569 offset = in_bytes(Klass::secondary_super_cache_offset()); 1570 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset)); 1571 } 1572 } 1573 1574 // Flatten all Raw pointers together. 1575 if (tj->base() == Type::RawPtr) 1576 tj = TypeRawPtr::BOTTOM; 1577 1578 if (tj->base() == Type::AnyPtr) 1579 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1580 1581 // Flatten all to bottom for now 1582 switch( _AliasLevel ) { 1583 case 0: 1584 tj = TypePtr::BOTTOM; 1585 break; 1586 case 1: // Flatten to: oop, static, field or array 1587 switch (tj->base()) { 1588 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1589 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1590 case Type::AryPtr: // do not distinguish arrays at all 1591 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1592 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1593 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1594 default: ShouldNotReachHere(); 1595 } 1596 break; 1597 case 2: // No collapsing at level 2; keep all splits 1598 case 3: // No collapsing at level 3; keep all splits 1599 break; 1600 default: 1601 Unimplemented(); 1602 } 1603 1604 offset = tj->offset(); 1605 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1606 1607 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1608 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1609 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1610 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1611 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1612 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1613 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1614 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1615 assert( tj->ptr() != TypePtr::TopPTR && 1616 tj->ptr() != TypePtr::AnyNull && 1617 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1618 // assert( tj->ptr() != TypePtr::Constant || 1619 // tj->base() == Type::RawPtr || 1620 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1621 1622 return tj; 1623 } 1624 1625 void Compile::AliasType::Init(int i, const TypePtr* at) { 1626 _index = i; 1627 _adr_type = at; 1628 _field = NULL; 1629 _element = NULL; 1630 _is_rewritable = true; // default 1631 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1632 if (atoop != NULL && atoop->is_known_instance()) { 1633 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1634 _general_index = Compile::current()->get_alias_index(gt); 1635 } else { 1636 _general_index = 0; 1637 } 1638 } 1639 1640 BasicType Compile::AliasType::basic_type() const { 1641 if (element() != NULL) { 1642 const Type* element = adr_type()->is_aryptr()->elem(); 1643 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1644 } if (field() != NULL) { 1645 return field()->layout_type(); 1646 } else { 1647 return T_ILLEGAL; // unknown 1648 } 1649 } 1650 1651 //---------------------------------print_on------------------------------------ 1652 #ifndef PRODUCT 1653 void Compile::AliasType::print_on(outputStream* st) { 1654 if (index() < 10) 1655 st->print("@ <%d> ", index()); 1656 else st->print("@ <%d>", index()); 1657 st->print(is_rewritable() ? " " : " RO"); 1658 int offset = adr_type()->offset(); 1659 if (offset == Type::OffsetBot) 1660 st->print(" +any"); 1661 else st->print(" +%-3d", offset); 1662 st->print(" in "); 1663 adr_type()->dump_on(st); 1664 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1665 if (field() != NULL && tjp) { 1666 if (tjp->klass() != field()->holder() || 1667 tjp->offset() != field()->offset_in_bytes()) { 1668 st->print(" != "); 1669 field()->print(); 1670 st->print(" ***"); 1671 } 1672 } 1673 } 1674 1675 void print_alias_types() { 1676 Compile* C = Compile::current(); 1677 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1678 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1679 C->alias_type(idx)->print_on(tty); 1680 tty->cr(); 1681 } 1682 } 1683 #endif 1684 1685 1686 //----------------------------probe_alias_cache-------------------------------- 1687 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1688 intptr_t key = (intptr_t) adr_type; 1689 key ^= key >> logAliasCacheSize; 1690 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1691 } 1692 1693 1694 //-----------------------------grow_alias_types-------------------------------- 1695 void Compile::grow_alias_types() { 1696 const int old_ats = _max_alias_types; // how many before? 1697 const int new_ats = old_ats; // how many more? 1698 const int grow_ats = old_ats+new_ats; // how many now? 1699 _max_alias_types = grow_ats; 1700 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1701 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1702 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1703 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1704 } 1705 1706 1707 //--------------------------------find_alias_type------------------------------ 1708 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1709 if (_AliasLevel == 0) 1710 return alias_type(AliasIdxBot); 1711 1712 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1713 if (ace->_adr_type == adr_type) { 1714 return alias_type(ace->_index); 1715 } 1716 1717 // Handle special cases. 1718 if (adr_type == NULL) return alias_type(AliasIdxTop); 1719 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1720 1721 // Do it the slow way. 1722 const TypePtr* flat = flatten_alias_type(adr_type); 1723 1724 #ifdef ASSERT 1725 { 1726 ResourceMark rm; 1727 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1728 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1729 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1730 Type::str(adr_type)); 1731 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1732 const TypeOopPtr* foop = flat->is_oopptr(); 1733 // Scalarizable allocations have exact klass always. 1734 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1735 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1736 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1737 Type::str(foop), Type::str(xoop)); 1738 } 1739 } 1740 #endif 1741 1742 int idx = AliasIdxTop; 1743 for (int i = 0; i < num_alias_types(); i++) { 1744 if (alias_type(i)->adr_type() == flat) { 1745 idx = i; 1746 break; 1747 } 1748 } 1749 1750 if (idx == AliasIdxTop) { 1751 if (no_create) return NULL; 1752 // Grow the array if necessary. 1753 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1754 // Add a new alias type. 1755 idx = _num_alias_types++; 1756 _alias_types[idx]->Init(idx, flat); 1757 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1758 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1759 if (flat->isa_instptr()) { 1760 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1761 && flat->is_instptr()->klass() == env()->Class_klass()) 1762 alias_type(idx)->set_rewritable(false); 1763 } 1764 if (flat->isa_aryptr()) { 1765 #ifdef ASSERT 1766 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1767 // (T_BYTE has the weakest alignment and size restrictions...) 1768 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1769 #endif 1770 if (flat->offset() == TypePtr::OffsetBot) { 1771 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1772 } 1773 } 1774 if (flat->isa_klassptr()) { 1775 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1776 alias_type(idx)->set_rewritable(false); 1777 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1778 alias_type(idx)->set_rewritable(false); 1779 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1780 alias_type(idx)->set_rewritable(false); 1781 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1782 alias_type(idx)->set_rewritable(false); 1783 } 1784 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1785 // but the base pointer type is not distinctive enough to identify 1786 // references into JavaThread.) 1787 1788 // Check for final fields. 1789 const TypeInstPtr* tinst = flat->isa_instptr(); 1790 const TypeValueTypePtr* vtptr = flat->isa_valuetypeptr(); 1791 ciField* field = NULL; 1792 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1793 if (tinst->const_oop() != NULL && 1794 tinst->klass() == ciEnv::current()->Class_klass() && 1795 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1796 // static field 1797 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1798 field = k->get_field_by_offset(tinst->offset(), true); 1799 } else { 1800 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1801 field = k->get_field_by_offset(tinst->offset(), false); 1802 } 1803 } else if (vtptr) { 1804 // Value type field 1805 ciValueKlass* vk = vtptr->klass()->as_value_klass(); 1806 field = vk->get_field_by_offset(vtptr->offset(), false); 1807 } 1808 assert(field == NULL || 1809 original_field == NULL || 1810 (field->holder() == original_field->holder() && 1811 field->offset() == original_field->offset() && 1812 field->is_static() == original_field->is_static()), "wrong field?"); 1813 // Set field() and is_rewritable() attributes. 1814 if (field != NULL) alias_type(idx)->set_field(field); 1815 } 1816 1817 // Fill the cache for next time. 1818 ace->_adr_type = adr_type; 1819 ace->_index = idx; 1820 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1821 1822 // Might as well try to fill the cache for the flattened version, too. 1823 AliasCacheEntry* face = probe_alias_cache(flat); 1824 if (face->_adr_type == NULL) { 1825 face->_adr_type = flat; 1826 face->_index = idx; 1827 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1828 } 1829 1830 return alias_type(idx); 1831 } 1832 1833 1834 Compile::AliasType* Compile::alias_type(ciField* field) { 1835 const TypeOopPtr* t; 1836 if (field->is_static()) 1837 t = TypeInstPtr::make(field->holder()->java_mirror()); 1838 else 1839 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1840 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1841 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1842 return atp; 1843 } 1844 1845 1846 //------------------------------have_alias_type-------------------------------- 1847 bool Compile::have_alias_type(const TypePtr* adr_type) { 1848 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1849 if (ace->_adr_type == adr_type) { 1850 return true; 1851 } 1852 1853 // Handle special cases. 1854 if (adr_type == NULL) return true; 1855 if (adr_type == TypePtr::BOTTOM) return true; 1856 1857 return find_alias_type(adr_type, true, NULL) != NULL; 1858 } 1859 1860 //-----------------------------must_alias-------------------------------------- 1861 // True if all values of the given address type are in the given alias category. 1862 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1863 if (alias_idx == AliasIdxBot) return true; // the universal category 1864 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1865 if (alias_idx == AliasIdxTop) return false; // the empty category 1866 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1867 1868 // the only remaining possible overlap is identity 1869 int adr_idx = get_alias_index(adr_type); 1870 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1871 assert(adr_idx == alias_idx || 1872 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1873 && adr_type != TypeOopPtr::BOTTOM), 1874 "should not be testing for overlap with an unsafe pointer"); 1875 return adr_idx == alias_idx; 1876 } 1877 1878 //------------------------------can_alias-------------------------------------- 1879 // True if any values of the given address type are in the given alias category. 1880 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1881 if (alias_idx == AliasIdxTop) return false; // the empty category 1882 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1883 if (alias_idx == AliasIdxBot) return true; // the universal category 1884 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1885 1886 // the only remaining possible overlap is identity 1887 int adr_idx = get_alias_index(adr_type); 1888 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1889 return adr_idx == alias_idx; 1890 } 1891 1892 1893 1894 //---------------------------pop_warm_call------------------------------------- 1895 WarmCallInfo* Compile::pop_warm_call() { 1896 WarmCallInfo* wci = _warm_calls; 1897 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1898 return wci; 1899 } 1900 1901 //----------------------------Inline_Warm-------------------------------------- 1902 int Compile::Inline_Warm() { 1903 // If there is room, try to inline some more warm call sites. 1904 // %%% Do a graph index compaction pass when we think we're out of space? 1905 if (!InlineWarmCalls) return 0; 1906 1907 int calls_made_hot = 0; 1908 int room_to_grow = NodeCountInliningCutoff - unique(); 1909 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1910 int amount_grown = 0; 1911 WarmCallInfo* call; 1912 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1913 int est_size = (int)call->size(); 1914 if (est_size > (room_to_grow - amount_grown)) { 1915 // This one won't fit anyway. Get rid of it. 1916 call->make_cold(); 1917 continue; 1918 } 1919 call->make_hot(); 1920 calls_made_hot++; 1921 amount_grown += est_size; 1922 amount_to_grow -= est_size; 1923 } 1924 1925 if (calls_made_hot > 0) set_major_progress(); 1926 return calls_made_hot; 1927 } 1928 1929 1930 //----------------------------Finish_Warm-------------------------------------- 1931 void Compile::Finish_Warm() { 1932 if (!InlineWarmCalls) return; 1933 if (failing()) return; 1934 if (warm_calls() == NULL) return; 1935 1936 // Clean up loose ends, if we are out of space for inlining. 1937 WarmCallInfo* call; 1938 while ((call = pop_warm_call()) != NULL) { 1939 call->make_cold(); 1940 } 1941 } 1942 1943 //---------------------cleanup_loop_predicates----------------------- 1944 // Remove the opaque nodes that protect the predicates so that all unused 1945 // checks and uncommon_traps will be eliminated from the ideal graph 1946 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1947 if (predicate_count()==0) return; 1948 for (int i = predicate_count(); i > 0; i--) { 1949 Node * n = predicate_opaque1_node(i-1); 1950 assert(n->Opcode() == Op_Opaque1, "must be"); 1951 igvn.replace_node(n, n->in(1)); 1952 } 1953 assert(predicate_count()==0, "should be clean!"); 1954 } 1955 1956 void Compile::add_range_check_cast(Node* n) { 1957 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1958 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1959 _range_check_casts->append(n); 1960 } 1961 1962 // Remove all range check dependent CastIINodes. 1963 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1964 for (int i = range_check_cast_count(); i > 0; i--) { 1965 Node* cast = range_check_cast_node(i-1); 1966 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1967 igvn.replace_node(cast, cast->in(1)); 1968 } 1969 assert(range_check_cast_count() == 0, "should be empty"); 1970 } 1971 1972 void Compile::add_value_type(Node* n) { 1973 assert(n->is_ValueTypeBase(), "unexpected node"); 1974 if (_value_type_nodes != NULL) { 1975 _value_type_nodes->push(n); 1976 } 1977 } 1978 1979 void Compile::remove_value_type(Node* n) { 1980 assert(n->is_ValueTypeBase(), "unexpected node"); 1981 if (_value_type_nodes != NULL) { 1982 _value_type_nodes->remove(n); 1983 } 1984 } 1985 1986 void Compile::process_value_types(PhaseIterGVN &igvn) { 1987 // Make value types scalar in safepoints 1988 while (_value_type_nodes->size() != 0) { 1989 ValueTypeBaseNode* vt = _value_type_nodes->pop()->as_ValueTypeBase(); 1990 vt->make_scalar_in_safepoints(igvn.C->root(), &igvn); 1991 if (vt->is_ValueTypePtr()) { 1992 igvn.replace_node(vt, vt->get_oop()); 1993 } 1994 } 1995 _value_type_nodes = NULL; 1996 igvn.optimize(); 1997 } 1998 1999 // StringOpts and late inlining of string methods 2000 void Compile::inline_string_calls(bool parse_time) { 2001 { 2002 // remove useless nodes to make the usage analysis simpler 2003 ResourceMark rm; 2004 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2005 } 2006 2007 { 2008 ResourceMark rm; 2009 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2010 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2011 print_method(PHASE_AFTER_STRINGOPTS, 3); 2012 } 2013 2014 // now inline anything that we skipped the first time around 2015 if (!parse_time) { 2016 _late_inlines_pos = _late_inlines.length(); 2017 } 2018 2019 while (_string_late_inlines.length() > 0) { 2020 CallGenerator* cg = _string_late_inlines.pop(); 2021 cg->do_late_inline(); 2022 if (failing()) return; 2023 } 2024 _string_late_inlines.trunc_to(0); 2025 } 2026 2027 // Late inlining of boxing methods 2028 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2029 if (_boxing_late_inlines.length() > 0) { 2030 assert(has_boxed_value(), "inconsistent"); 2031 2032 PhaseGVN* gvn = initial_gvn(); 2033 set_inlining_incrementally(true); 2034 2035 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2036 for_igvn()->clear(); 2037 gvn->replace_with(&igvn); 2038 2039 _late_inlines_pos = _late_inlines.length(); 2040 2041 while (_boxing_late_inlines.length() > 0) { 2042 CallGenerator* cg = _boxing_late_inlines.pop(); 2043 cg->do_late_inline(); 2044 if (failing()) return; 2045 } 2046 _boxing_late_inlines.trunc_to(0); 2047 2048 { 2049 ResourceMark rm; 2050 PhaseRemoveUseless pru(gvn, for_igvn()); 2051 } 2052 2053 igvn = PhaseIterGVN(gvn); 2054 igvn.optimize(); 2055 2056 set_inlining_progress(false); 2057 set_inlining_incrementally(false); 2058 } 2059 } 2060 2061 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 2062 assert(IncrementalInline, "incremental inlining should be on"); 2063 PhaseGVN* gvn = initial_gvn(); 2064 2065 set_inlining_progress(false); 2066 for_igvn()->clear(); 2067 gvn->replace_with(&igvn); 2068 2069 { 2070 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2071 int i = 0; 2072 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2073 CallGenerator* cg = _late_inlines.at(i); 2074 _late_inlines_pos = i+1; 2075 cg->do_late_inline(); 2076 if (failing()) return; 2077 } 2078 int j = 0; 2079 for (; i < _late_inlines.length(); i++, j++) { 2080 _late_inlines.at_put(j, _late_inlines.at(i)); 2081 } 2082 _late_inlines.trunc_to(j); 2083 } 2084 2085 { 2086 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2087 ResourceMark rm; 2088 PhaseRemoveUseless pru(gvn, for_igvn()); 2089 } 2090 2091 { 2092 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2093 igvn = PhaseIterGVN(gvn); 2094 } 2095 } 2096 2097 // Perform incremental inlining until bound on number of live nodes is reached 2098 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2099 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2100 2101 PhaseGVN* gvn = initial_gvn(); 2102 2103 set_inlining_incrementally(true); 2104 set_inlining_progress(true); 2105 uint low_live_nodes = 0; 2106 2107 while(inlining_progress() && _late_inlines.length() > 0) { 2108 2109 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2110 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2111 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2112 // PhaseIdealLoop is expensive so we only try it once we are 2113 // out of live nodes and we only try it again if the previous 2114 // helped got the number of nodes down significantly 2115 PhaseIdealLoop ideal_loop( igvn, false, true ); 2116 if (failing()) return; 2117 low_live_nodes = live_nodes(); 2118 _major_progress = true; 2119 } 2120 2121 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2122 break; 2123 } 2124 } 2125 2126 inline_incrementally_one(igvn); 2127 2128 if (failing()) return; 2129 2130 { 2131 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2132 igvn.optimize(); 2133 } 2134 2135 if (failing()) return; 2136 } 2137 2138 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2139 2140 if (_string_late_inlines.length() > 0) { 2141 assert(has_stringbuilder(), "inconsistent"); 2142 for_igvn()->clear(); 2143 initial_gvn()->replace_with(&igvn); 2144 2145 inline_string_calls(false); 2146 2147 if (failing()) return; 2148 2149 { 2150 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2151 ResourceMark rm; 2152 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2153 } 2154 2155 { 2156 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2157 igvn = PhaseIterGVN(gvn); 2158 igvn.optimize(); 2159 } 2160 } 2161 2162 set_inlining_incrementally(false); 2163 } 2164 2165 2166 //------------------------------Optimize--------------------------------------- 2167 // Given a graph, optimize it. 2168 void Compile::Optimize() { 2169 TracePhase tp("optimizer", &timers[_t_optimizer]); 2170 2171 #ifndef PRODUCT 2172 if (_directive->BreakAtCompileOption) { 2173 BREAKPOINT; 2174 } 2175 2176 #endif 2177 2178 ResourceMark rm; 2179 int loop_opts_cnt; 2180 2181 print_inlining_reinit(); 2182 2183 NOT_PRODUCT( verify_graph_edges(); ) 2184 2185 print_method(PHASE_AFTER_PARSING); 2186 2187 { 2188 // Iterative Global Value Numbering, including ideal transforms 2189 // Initialize IterGVN with types and values from parse-time GVN 2190 PhaseIterGVN igvn(initial_gvn()); 2191 #ifdef ASSERT 2192 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2193 #endif 2194 { 2195 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2196 igvn.optimize(); 2197 } 2198 2199 print_method(PHASE_ITER_GVN1, 2); 2200 2201 if (failing()) return; 2202 2203 inline_incrementally(igvn); 2204 2205 print_method(PHASE_INCREMENTAL_INLINE, 2); 2206 2207 if (failing()) return; 2208 2209 if (eliminate_boxing()) { 2210 // Inline valueOf() methods now. 2211 inline_boxing_calls(igvn); 2212 2213 if (AlwaysIncrementalInline) { 2214 inline_incrementally(igvn); 2215 } 2216 2217 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2218 2219 if (failing()) return; 2220 } 2221 2222 // Remove the speculative part of types and clean up the graph from 2223 // the extra CastPP nodes whose only purpose is to carry them. Do 2224 // that early so that optimizations are not disrupted by the extra 2225 // CastPP nodes. 2226 remove_speculative_types(igvn); 2227 2228 // No more new expensive nodes will be added to the list from here 2229 // so keep only the actual candidates for optimizations. 2230 cleanup_expensive_nodes(igvn); 2231 2232 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2233 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2234 initial_gvn()->replace_with(&igvn); 2235 for_igvn()->clear(); 2236 Unique_Node_List new_worklist(C->comp_arena()); 2237 { 2238 ResourceMark rm; 2239 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2240 } 2241 set_for_igvn(&new_worklist); 2242 igvn = PhaseIterGVN(initial_gvn()); 2243 igvn.optimize(); 2244 } 2245 2246 if (_value_type_nodes->size() > 0) { 2247 // Do this once all inlining is over to avoid getting inconsistent debug info 2248 process_value_types(igvn); 2249 } 2250 2251 // Perform escape analysis 2252 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2253 if (has_loops()) { 2254 // Cleanup graph (remove dead nodes). 2255 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2256 PhaseIdealLoop ideal_loop( igvn, false, true ); 2257 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2258 if (failing()) return; 2259 } 2260 ConnectionGraph::do_analysis(this, &igvn); 2261 2262 if (failing()) return; 2263 2264 // Optimize out fields loads from scalar replaceable allocations. 2265 igvn.optimize(); 2266 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2267 2268 if (failing()) return; 2269 2270 if (congraph() != NULL && macro_count() > 0) { 2271 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2272 PhaseMacroExpand mexp(igvn); 2273 mexp.eliminate_macro_nodes(); 2274 igvn.set_delay_transform(false); 2275 2276 igvn.optimize(); 2277 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2278 2279 if (failing()) return; 2280 } 2281 } 2282 2283 // Loop transforms on the ideal graph. Range Check Elimination, 2284 // peeling, unrolling, etc. 2285 2286 // Set loop opts counter 2287 loop_opts_cnt = num_loop_opts(); 2288 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2289 { 2290 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2291 PhaseIdealLoop ideal_loop( igvn, true ); 2292 loop_opts_cnt--; 2293 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2294 if (failing()) return; 2295 } 2296 // Loop opts pass if partial peeling occurred in previous pass 2297 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2298 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2299 PhaseIdealLoop ideal_loop( igvn, false ); 2300 loop_opts_cnt--; 2301 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2302 if (failing()) return; 2303 } 2304 // Loop opts pass for loop-unrolling before CCP 2305 if(major_progress() && (loop_opts_cnt > 0)) { 2306 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2307 PhaseIdealLoop ideal_loop( igvn, false ); 2308 loop_opts_cnt--; 2309 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2310 } 2311 if (!failing()) { 2312 // Verify that last round of loop opts produced a valid graph 2313 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2314 PhaseIdealLoop::verify(igvn); 2315 } 2316 } 2317 if (failing()) return; 2318 2319 // Conditional Constant Propagation; 2320 PhaseCCP ccp( &igvn ); 2321 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2322 { 2323 TracePhase tp("ccp", &timers[_t_ccp]); 2324 ccp.do_transform(); 2325 } 2326 print_method(PHASE_CPP1, 2); 2327 2328 assert( true, "Break here to ccp.dump_old2new_map()"); 2329 2330 // Iterative Global Value Numbering, including ideal transforms 2331 { 2332 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2333 igvn = ccp; 2334 igvn.optimize(); 2335 } 2336 2337 print_method(PHASE_ITER_GVN2, 2); 2338 2339 if (failing()) return; 2340 2341 // Loop transforms on the ideal graph. Range Check Elimination, 2342 // peeling, unrolling, etc. 2343 if(loop_opts_cnt > 0) { 2344 debug_only( int cnt = 0; ); 2345 while(major_progress() && (loop_opts_cnt > 0)) { 2346 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2347 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2348 PhaseIdealLoop ideal_loop( igvn, true); 2349 loop_opts_cnt--; 2350 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2351 if (failing()) return; 2352 } 2353 } 2354 // Ensure that major progress is now clear 2355 C->clear_major_progress(); 2356 2357 { 2358 // Verify that all previous optimizations produced a valid graph 2359 // at least to this point, even if no loop optimizations were done. 2360 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2361 PhaseIdealLoop::verify(igvn); 2362 } 2363 2364 if (range_check_cast_count() > 0) { 2365 // No more loop optimizations. Remove all range check dependent CastIINodes. 2366 C->remove_range_check_casts(igvn); 2367 igvn.optimize(); 2368 } 2369 2370 { 2371 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2372 PhaseMacroExpand mex(igvn); 2373 if (mex.expand_macro_nodes()) { 2374 assert(failing(), "must bail out w/ explicit message"); 2375 return; 2376 } 2377 } 2378 2379 DEBUG_ONLY( _modified_nodes = NULL; ) 2380 } // (End scope of igvn; run destructor if necessary for asserts.) 2381 2382 process_print_inlining(); 2383 // A method with only infinite loops has no edges entering loops from root 2384 { 2385 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2386 if (final_graph_reshaping()) { 2387 assert(failing(), "must bail out w/ explicit message"); 2388 return; 2389 } 2390 } 2391 2392 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2393 } 2394 2395 //------------------------------Code_Gen--------------------------------------- 2396 // Given a graph, generate code for it 2397 void Compile::Code_Gen() { 2398 if (failing()) { 2399 return; 2400 } 2401 2402 // Perform instruction selection. You might think we could reclaim Matcher 2403 // memory PDQ, but actually the Matcher is used in generating spill code. 2404 // Internals of the Matcher (including some VectorSets) must remain live 2405 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2406 // set a bit in reclaimed memory. 2407 2408 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2409 // nodes. Mapping is only valid at the root of each matched subtree. 2410 NOT_PRODUCT( verify_graph_edges(); ) 2411 2412 Matcher matcher; 2413 _matcher = &matcher; 2414 { 2415 TracePhase tp("matcher", &timers[_t_matcher]); 2416 matcher.match(); 2417 } 2418 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2419 // nodes. Mapping is only valid at the root of each matched subtree. 2420 NOT_PRODUCT( verify_graph_edges(); ) 2421 2422 // If you have too many nodes, or if matching has failed, bail out 2423 check_node_count(0, "out of nodes matching instructions"); 2424 if (failing()) { 2425 return; 2426 } 2427 2428 // Build a proper-looking CFG 2429 PhaseCFG cfg(node_arena(), root(), matcher); 2430 _cfg = &cfg; 2431 { 2432 TracePhase tp("scheduler", &timers[_t_scheduler]); 2433 bool success = cfg.do_global_code_motion(); 2434 if (!success) { 2435 return; 2436 } 2437 2438 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2439 NOT_PRODUCT( verify_graph_edges(); ) 2440 debug_only( cfg.verify(); ) 2441 } 2442 2443 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2444 _regalloc = ®alloc; 2445 { 2446 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2447 // Perform register allocation. After Chaitin, use-def chains are 2448 // no longer accurate (at spill code) and so must be ignored. 2449 // Node->LRG->reg mappings are still accurate. 2450 _regalloc->Register_Allocate(); 2451 2452 // Bail out if the allocator builds too many nodes 2453 if (failing()) { 2454 return; 2455 } 2456 } 2457 2458 // Prior to register allocation we kept empty basic blocks in case the 2459 // the allocator needed a place to spill. After register allocation we 2460 // are not adding any new instructions. If any basic block is empty, we 2461 // can now safely remove it. 2462 { 2463 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2464 cfg.remove_empty_blocks(); 2465 if (do_freq_based_layout()) { 2466 PhaseBlockLayout layout(cfg); 2467 } else { 2468 cfg.set_loop_alignment(); 2469 } 2470 cfg.fixup_flow(); 2471 } 2472 2473 // Apply peephole optimizations 2474 if( OptoPeephole ) { 2475 TracePhase tp("peephole", &timers[_t_peephole]); 2476 PhasePeephole peep( _regalloc, cfg); 2477 peep.do_transform(); 2478 } 2479 2480 // Do late expand if CPU requires this. 2481 if (Matcher::require_postalloc_expand) { 2482 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2483 cfg.postalloc_expand(_regalloc); 2484 } 2485 2486 // Convert Nodes to instruction bits in a buffer 2487 { 2488 TraceTime tp("output", &timers[_t_output], CITime); 2489 Output(); 2490 } 2491 2492 print_method(PHASE_FINAL_CODE); 2493 2494 // He's dead, Jim. 2495 _cfg = (PhaseCFG*)0xdeadbeef; 2496 _regalloc = (PhaseChaitin*)0xdeadbeef; 2497 } 2498 2499 2500 //------------------------------dump_asm--------------------------------------- 2501 // Dump formatted assembly 2502 #ifndef PRODUCT 2503 void Compile::dump_asm(int *pcs, uint pc_limit) { 2504 bool cut_short = false; 2505 tty->print_cr("#"); 2506 tty->print("# "); _tf->dump(); tty->cr(); 2507 tty->print_cr("#"); 2508 2509 // For all blocks 2510 int pc = 0x0; // Program counter 2511 char starts_bundle = ' '; 2512 _regalloc->dump_frame(); 2513 2514 Node *n = NULL; 2515 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2516 if (VMThread::should_terminate()) { 2517 cut_short = true; 2518 break; 2519 } 2520 Block* block = _cfg->get_block(i); 2521 if (block->is_connector() && !Verbose) { 2522 continue; 2523 } 2524 n = block->head(); 2525 if (pcs && n->_idx < pc_limit) { 2526 tty->print("%3.3x ", pcs[n->_idx]); 2527 } else { 2528 tty->print(" "); 2529 } 2530 block->dump_head(_cfg); 2531 if (block->is_connector()) { 2532 tty->print_cr(" # Empty connector block"); 2533 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2534 tty->print_cr(" # Block is sole successor of call"); 2535 } 2536 2537 // For all instructions 2538 Node *delay = NULL; 2539 for (uint j = 0; j < block->number_of_nodes(); j++) { 2540 if (VMThread::should_terminate()) { 2541 cut_short = true; 2542 break; 2543 } 2544 n = block->get_node(j); 2545 if (valid_bundle_info(n)) { 2546 Bundle* bundle = node_bundling(n); 2547 if (bundle->used_in_unconditional_delay()) { 2548 delay = n; 2549 continue; 2550 } 2551 if (bundle->starts_bundle()) { 2552 starts_bundle = '+'; 2553 } 2554 } 2555 2556 if (WizardMode) { 2557 n->dump(); 2558 } 2559 2560 if( !n->is_Region() && // Dont print in the Assembly 2561 !n->is_Phi() && // a few noisely useless nodes 2562 !n->is_Proj() && 2563 !n->is_MachTemp() && 2564 !n->is_SafePointScalarObject() && 2565 !n->is_Catch() && // Would be nice to print exception table targets 2566 !n->is_MergeMem() && // Not very interesting 2567 !n->is_top() && // Debug info table constants 2568 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2569 ) { 2570 if (pcs && n->_idx < pc_limit) 2571 tty->print("%3.3x", pcs[n->_idx]); 2572 else 2573 tty->print(" "); 2574 tty->print(" %c ", starts_bundle); 2575 starts_bundle = ' '; 2576 tty->print("\t"); 2577 n->format(_regalloc, tty); 2578 tty->cr(); 2579 } 2580 2581 // If we have an instruction with a delay slot, and have seen a delay, 2582 // then back up and print it 2583 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2584 assert(delay != NULL, "no unconditional delay instruction"); 2585 if (WizardMode) delay->dump(); 2586 2587 if (node_bundling(delay)->starts_bundle()) 2588 starts_bundle = '+'; 2589 if (pcs && n->_idx < pc_limit) 2590 tty->print("%3.3x", pcs[n->_idx]); 2591 else 2592 tty->print(" "); 2593 tty->print(" %c ", starts_bundle); 2594 starts_bundle = ' '; 2595 tty->print("\t"); 2596 delay->format(_regalloc, tty); 2597 tty->cr(); 2598 delay = NULL; 2599 } 2600 2601 // Dump the exception table as well 2602 if( n->is_Catch() && (Verbose || WizardMode) ) { 2603 // Print the exception table for this offset 2604 _handler_table.print_subtable_for(pc); 2605 } 2606 } 2607 2608 if (pcs && n->_idx < pc_limit) 2609 tty->print_cr("%3.3x", pcs[n->_idx]); 2610 else 2611 tty->cr(); 2612 2613 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2614 2615 } // End of per-block dump 2616 tty->cr(); 2617 2618 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2619 } 2620 #endif 2621 2622 //------------------------------Final_Reshape_Counts--------------------------- 2623 // This class defines counters to help identify when a method 2624 // may/must be executed using hardware with only 24-bit precision. 2625 struct Final_Reshape_Counts : public StackObj { 2626 int _call_count; // count non-inlined 'common' calls 2627 int _float_count; // count float ops requiring 24-bit precision 2628 int _double_count; // count double ops requiring more precision 2629 int _java_call_count; // count non-inlined 'java' calls 2630 int _inner_loop_count; // count loops which need alignment 2631 VectorSet _visited; // Visitation flags 2632 Node_List _tests; // Set of IfNodes & PCTableNodes 2633 2634 Final_Reshape_Counts() : 2635 _call_count(0), _float_count(0), _double_count(0), 2636 _java_call_count(0), _inner_loop_count(0), 2637 _visited( Thread::current()->resource_area() ) { } 2638 2639 void inc_call_count () { _call_count ++; } 2640 void inc_float_count () { _float_count ++; } 2641 void inc_double_count() { _double_count++; } 2642 void inc_java_call_count() { _java_call_count++; } 2643 void inc_inner_loop_count() { _inner_loop_count++; } 2644 2645 int get_call_count () const { return _call_count ; } 2646 int get_float_count () const { return _float_count ; } 2647 int get_double_count() const { return _double_count; } 2648 int get_java_call_count() const { return _java_call_count; } 2649 int get_inner_loop_count() const { return _inner_loop_count; } 2650 }; 2651 2652 #ifdef ASSERT 2653 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2654 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2655 // Make sure the offset goes inside the instance layout. 2656 return k->contains_field_offset(tp->offset()); 2657 // Note that OffsetBot and OffsetTop are very negative. 2658 } 2659 #endif 2660 2661 // Eliminate trivially redundant StoreCMs and accumulate their 2662 // precedence edges. 2663 void Compile::eliminate_redundant_card_marks(Node* n) { 2664 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2665 if (n->in(MemNode::Address)->outcnt() > 1) { 2666 // There are multiple users of the same address so it might be 2667 // possible to eliminate some of the StoreCMs 2668 Node* mem = n->in(MemNode::Memory); 2669 Node* adr = n->in(MemNode::Address); 2670 Node* val = n->in(MemNode::ValueIn); 2671 Node* prev = n; 2672 bool done = false; 2673 // Walk the chain of StoreCMs eliminating ones that match. As 2674 // long as it's a chain of single users then the optimization is 2675 // safe. Eliminating partially redundant StoreCMs would require 2676 // cloning copies down the other paths. 2677 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2678 if (adr == mem->in(MemNode::Address) && 2679 val == mem->in(MemNode::ValueIn)) { 2680 // redundant StoreCM 2681 if (mem->req() > MemNode::OopStore) { 2682 // Hasn't been processed by this code yet. 2683 n->add_prec(mem->in(MemNode::OopStore)); 2684 } else { 2685 // Already converted to precedence edge 2686 for (uint i = mem->req(); i < mem->len(); i++) { 2687 // Accumulate any precedence edges 2688 if (mem->in(i) != NULL) { 2689 n->add_prec(mem->in(i)); 2690 } 2691 } 2692 // Everything above this point has been processed. 2693 done = true; 2694 } 2695 // Eliminate the previous StoreCM 2696 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2697 assert(mem->outcnt() == 0, "should be dead"); 2698 mem->disconnect_inputs(NULL, this); 2699 } else { 2700 prev = mem; 2701 } 2702 mem = prev->in(MemNode::Memory); 2703 } 2704 } 2705 } 2706 2707 2708 //------------------------------final_graph_reshaping_impl---------------------- 2709 // Implement items 1-5 from final_graph_reshaping below. 2710 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2711 2712 if ( n->outcnt() == 0 ) return; // dead node 2713 uint nop = n->Opcode(); 2714 2715 // Check for 2-input instruction with "last use" on right input. 2716 // Swap to left input. Implements item (2). 2717 if( n->req() == 3 && // two-input instruction 2718 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2719 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2720 n->in(2)->outcnt() == 1 &&// right use IS a last use 2721 !n->in(2)->is_Con() ) { // right use is not a constant 2722 // Check for commutative opcode 2723 switch( nop ) { 2724 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2725 case Op_MaxI: case Op_MinI: 2726 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2727 case Op_AndL: case Op_XorL: case Op_OrL: 2728 case Op_AndI: case Op_XorI: case Op_OrI: { 2729 // Move "last use" input to left by swapping inputs 2730 n->swap_edges(1, 2); 2731 break; 2732 } 2733 default: 2734 break; 2735 } 2736 } 2737 2738 #ifdef ASSERT 2739 if( n->is_Mem() ) { 2740 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2741 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2742 // oop will be recorded in oop map if load crosses safepoint 2743 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2744 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2745 "raw memory operations should have control edge"); 2746 } 2747 #endif 2748 // Count FPU ops and common calls, implements item (3) 2749 switch( nop ) { 2750 // Count all float operations that may use FPU 2751 case Op_AddF: 2752 case Op_SubF: 2753 case Op_MulF: 2754 case Op_DivF: 2755 case Op_NegF: 2756 case Op_ModF: 2757 case Op_ConvI2F: 2758 case Op_ConF: 2759 case Op_CmpF: 2760 case Op_CmpF3: 2761 // case Op_ConvL2F: // longs are split into 32-bit halves 2762 frc.inc_float_count(); 2763 break; 2764 2765 case Op_ConvF2D: 2766 case Op_ConvD2F: 2767 frc.inc_float_count(); 2768 frc.inc_double_count(); 2769 break; 2770 2771 // Count all double operations that may use FPU 2772 case Op_AddD: 2773 case Op_SubD: 2774 case Op_MulD: 2775 case Op_DivD: 2776 case Op_NegD: 2777 case Op_ModD: 2778 case Op_ConvI2D: 2779 case Op_ConvD2I: 2780 // case Op_ConvL2D: // handled by leaf call 2781 // case Op_ConvD2L: // handled by leaf call 2782 case Op_ConD: 2783 case Op_CmpD: 2784 case Op_CmpD3: 2785 frc.inc_double_count(); 2786 break; 2787 case Op_Opaque1: // Remove Opaque Nodes before matching 2788 case Op_Opaque2: // Remove Opaque Nodes before matching 2789 case Op_Opaque3: 2790 n->subsume_by(n->in(1), this); 2791 break; 2792 case Op_CallStaticJava: 2793 case Op_CallJava: 2794 case Op_CallDynamicJava: 2795 frc.inc_java_call_count(); // Count java call site; 2796 case Op_CallRuntime: 2797 case Op_CallLeaf: 2798 case Op_CallLeafNoFP: { 2799 assert( n->is_Call(), "" ); 2800 CallNode *call = n->as_Call(); 2801 // Count call sites where the FP mode bit would have to be flipped. 2802 // Do not count uncommon runtime calls: 2803 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2804 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2805 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2806 frc.inc_call_count(); // Count the call site 2807 } else { // See if uncommon argument is shared 2808 Node *n = call->in(TypeFunc::Parms); 2809 int nop = n->Opcode(); 2810 // Clone shared simple arguments to uncommon calls, item (1). 2811 if( n->outcnt() > 1 && 2812 !n->is_Proj() && 2813 nop != Op_CreateEx && 2814 nop != Op_CheckCastPP && 2815 nop != Op_DecodeN && 2816 nop != Op_DecodeNKlass && 2817 !n->is_Mem() ) { 2818 Node *x = n->clone(); 2819 call->set_req( TypeFunc::Parms, x ); 2820 } 2821 } 2822 break; 2823 } 2824 2825 case Op_StoreD: 2826 case Op_LoadD: 2827 case Op_LoadD_unaligned: 2828 frc.inc_double_count(); 2829 goto handle_mem; 2830 case Op_StoreF: 2831 case Op_LoadF: 2832 frc.inc_float_count(); 2833 goto handle_mem; 2834 2835 case Op_StoreCM: 2836 { 2837 // Convert OopStore dependence into precedence edge 2838 Node* prec = n->in(MemNode::OopStore); 2839 n->del_req(MemNode::OopStore); 2840 n->add_prec(prec); 2841 eliminate_redundant_card_marks(n); 2842 } 2843 2844 // fall through 2845 2846 case Op_StoreB: 2847 case Op_StoreC: 2848 case Op_StorePConditional: 2849 case Op_StoreI: 2850 case Op_StoreL: 2851 case Op_StoreIConditional: 2852 case Op_StoreLConditional: 2853 case Op_CompareAndSwapB: 2854 case Op_CompareAndSwapS: 2855 case Op_CompareAndSwapI: 2856 case Op_CompareAndSwapL: 2857 case Op_CompareAndSwapP: 2858 case Op_CompareAndSwapN: 2859 case Op_WeakCompareAndSwapB: 2860 case Op_WeakCompareAndSwapS: 2861 case Op_WeakCompareAndSwapI: 2862 case Op_WeakCompareAndSwapL: 2863 case Op_WeakCompareAndSwapP: 2864 case Op_WeakCompareAndSwapN: 2865 case Op_CompareAndExchangeB: 2866 case Op_CompareAndExchangeS: 2867 case Op_CompareAndExchangeI: 2868 case Op_CompareAndExchangeL: 2869 case Op_CompareAndExchangeP: 2870 case Op_CompareAndExchangeN: 2871 case Op_GetAndAddS: 2872 case Op_GetAndAddB: 2873 case Op_GetAndAddI: 2874 case Op_GetAndAddL: 2875 case Op_GetAndSetS: 2876 case Op_GetAndSetB: 2877 case Op_GetAndSetI: 2878 case Op_GetAndSetL: 2879 case Op_GetAndSetP: 2880 case Op_GetAndSetN: 2881 case Op_StoreP: 2882 case Op_StoreN: 2883 case Op_StoreNKlass: 2884 case Op_LoadB: 2885 case Op_LoadUB: 2886 case Op_LoadUS: 2887 case Op_LoadI: 2888 case Op_LoadKlass: 2889 case Op_LoadNKlass: 2890 case Op_LoadL: 2891 case Op_LoadL_unaligned: 2892 case Op_LoadPLocked: 2893 case Op_LoadP: 2894 case Op_LoadN: 2895 case Op_LoadRange: 2896 case Op_LoadS: { 2897 handle_mem: 2898 #ifdef ASSERT 2899 if( VerifyOptoOopOffsets ) { 2900 assert( n->is_Mem(), "" ); 2901 MemNode *mem = (MemNode*)n; 2902 // Check to see if address types have grounded out somehow. 2903 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2904 assert( !tp || oop_offset_is_sane(tp), "" ); 2905 } 2906 #endif 2907 break; 2908 } 2909 2910 case Op_AddP: { // Assert sane base pointers 2911 Node *addp = n->in(AddPNode::Address); 2912 assert( !addp->is_AddP() || 2913 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2914 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2915 "Base pointers must match (addp %u)", addp->_idx ); 2916 #ifdef _LP64 2917 if ((UseCompressedOops || UseCompressedClassPointers) && 2918 addp->Opcode() == Op_ConP && 2919 addp == n->in(AddPNode::Base) && 2920 n->in(AddPNode::Offset)->is_Con()) { 2921 // If the transformation of ConP to ConN+DecodeN is beneficial depends 2922 // on the platform and on the compressed oops mode. 2923 // Use addressing with narrow klass to load with offset on x86. 2924 // Some platforms can use the constant pool to load ConP. 2925 // Do this transformation here since IGVN will convert ConN back to ConP. 2926 const Type* t = addp->bottom_type(); 2927 bool is_oop = t->isa_oopptr() != NULL; 2928 bool is_klass = t->isa_klassptr() != NULL; 2929 2930 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 2931 (is_klass && Matcher::const_klass_prefer_decode())) { 2932 Node* nn = NULL; 2933 2934 int op = is_oop ? Op_ConN : Op_ConNKlass; 2935 2936 // Look for existing ConN node of the same exact type. 2937 Node* r = root(); 2938 uint cnt = r->outcnt(); 2939 for (uint i = 0; i < cnt; i++) { 2940 Node* m = r->raw_out(i); 2941 if (m!= NULL && m->Opcode() == op && 2942 m->bottom_type()->make_ptr() == t) { 2943 nn = m; 2944 break; 2945 } 2946 } 2947 if (nn != NULL) { 2948 // Decode a narrow oop to match address 2949 // [R12 + narrow_oop_reg<<3 + offset] 2950 if (is_oop) { 2951 nn = new DecodeNNode(nn, t); 2952 } else { 2953 nn = new DecodeNKlassNode(nn, t); 2954 } 2955 // Check for succeeding AddP which uses the same Base. 2956 // Otherwise we will run into the assertion above when visiting that guy. 2957 for (uint i = 0; i < n->outcnt(); ++i) { 2958 Node *out_i = n->raw_out(i); 2959 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 2960 out_i->set_req(AddPNode::Base, nn); 2961 #ifdef ASSERT 2962 for (uint j = 0; j < out_i->outcnt(); ++j) { 2963 Node *out_j = out_i->raw_out(j); 2964 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 2965 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 2966 } 2967 #endif 2968 } 2969 } 2970 n->set_req(AddPNode::Base, nn); 2971 n->set_req(AddPNode::Address, nn); 2972 if (addp->outcnt() == 0) { 2973 addp->disconnect_inputs(NULL, this); 2974 } 2975 } 2976 } 2977 } 2978 #endif 2979 // platform dependent reshaping of the address expression 2980 reshape_address(n->as_AddP()); 2981 break; 2982 } 2983 2984 case Op_CastPP: { 2985 // Remove CastPP nodes to gain more freedom during scheduling but 2986 // keep the dependency they encode as control or precedence edges 2987 // (if control is set already) on memory operations. Some CastPP 2988 // nodes don't have a control (don't carry a dependency): skip 2989 // those. 2990 if (n->in(0) != NULL) { 2991 ResourceMark rm; 2992 Unique_Node_List wq; 2993 wq.push(n); 2994 for (uint next = 0; next < wq.size(); ++next) { 2995 Node *m = wq.at(next); 2996 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 2997 Node* use = m->fast_out(i); 2998 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 2999 use->ensure_control_or_add_prec(n->in(0)); 3000 } else { 3001 switch(use->Opcode()) { 3002 case Op_AddP: 3003 case Op_DecodeN: 3004 case Op_DecodeNKlass: 3005 case Op_CheckCastPP: 3006 case Op_CastPP: 3007 wq.push(use); 3008 break; 3009 } 3010 } 3011 } 3012 } 3013 } 3014 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3015 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3016 Node* in1 = n->in(1); 3017 const Type* t = n->bottom_type(); 3018 Node* new_in1 = in1->clone(); 3019 new_in1->as_DecodeN()->set_type(t); 3020 3021 if (!Matcher::narrow_oop_use_complex_address()) { 3022 // 3023 // x86, ARM and friends can handle 2 adds in addressing mode 3024 // and Matcher can fold a DecodeN node into address by using 3025 // a narrow oop directly and do implicit NULL check in address: 3026 // 3027 // [R12 + narrow_oop_reg<<3 + offset] 3028 // NullCheck narrow_oop_reg 3029 // 3030 // On other platforms (Sparc) we have to keep new DecodeN node and 3031 // use it to do implicit NULL check in address: 3032 // 3033 // decode_not_null narrow_oop_reg, base_reg 3034 // [base_reg + offset] 3035 // NullCheck base_reg 3036 // 3037 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3038 // to keep the information to which NULL check the new DecodeN node 3039 // corresponds to use it as value in implicit_null_check(). 3040 // 3041 new_in1->set_req(0, n->in(0)); 3042 } 3043 3044 n->subsume_by(new_in1, this); 3045 if (in1->outcnt() == 0) { 3046 in1->disconnect_inputs(NULL, this); 3047 } 3048 } else { 3049 n->subsume_by(n->in(1), this); 3050 if (n->outcnt() == 0) { 3051 n->disconnect_inputs(NULL, this); 3052 } 3053 } 3054 break; 3055 } 3056 #ifdef _LP64 3057 case Op_CmpP: 3058 // Do this transformation here to preserve CmpPNode::sub() and 3059 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3060 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3061 Node* in1 = n->in(1); 3062 Node* in2 = n->in(2); 3063 if (!in1->is_DecodeNarrowPtr()) { 3064 in2 = in1; 3065 in1 = n->in(2); 3066 } 3067 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3068 3069 Node* new_in2 = NULL; 3070 if (in2->is_DecodeNarrowPtr()) { 3071 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3072 new_in2 = in2->in(1); 3073 } else if (in2->Opcode() == Op_ConP) { 3074 const Type* t = in2->bottom_type(); 3075 if (t == TypePtr::NULL_PTR) { 3076 assert(in1->is_DecodeN(), "compare klass to null?"); 3077 // Don't convert CmpP null check into CmpN if compressed 3078 // oops implicit null check is not generated. 3079 // This will allow to generate normal oop implicit null check. 3080 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3081 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3082 // 3083 // This transformation together with CastPP transformation above 3084 // will generated code for implicit NULL checks for compressed oops. 3085 // 3086 // The original code after Optimize() 3087 // 3088 // LoadN memory, narrow_oop_reg 3089 // decode narrow_oop_reg, base_reg 3090 // CmpP base_reg, NULL 3091 // CastPP base_reg // NotNull 3092 // Load [base_reg + offset], val_reg 3093 // 3094 // after these transformations will be 3095 // 3096 // LoadN memory, narrow_oop_reg 3097 // CmpN narrow_oop_reg, NULL 3098 // decode_not_null narrow_oop_reg, base_reg 3099 // Load [base_reg + offset], val_reg 3100 // 3101 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3102 // since narrow oops can be used in debug info now (see the code in 3103 // final_graph_reshaping_walk()). 3104 // 3105 // At the end the code will be matched to 3106 // on x86: 3107 // 3108 // Load_narrow_oop memory, narrow_oop_reg 3109 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3110 // NullCheck narrow_oop_reg 3111 // 3112 // and on sparc: 3113 // 3114 // Load_narrow_oop memory, narrow_oop_reg 3115 // decode_not_null narrow_oop_reg, base_reg 3116 // Load [base_reg + offset], val_reg 3117 // NullCheck base_reg 3118 // 3119 } else if (t->isa_oopptr()) { 3120 new_in2 = ConNode::make(t->make_narrowoop()); 3121 } else if (t->isa_klassptr()) { 3122 new_in2 = ConNode::make(t->make_narrowklass()); 3123 } 3124 } 3125 if (new_in2 != NULL) { 3126 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3127 n->subsume_by(cmpN, this); 3128 if (in1->outcnt() == 0) { 3129 in1->disconnect_inputs(NULL, this); 3130 } 3131 if (in2->outcnt() == 0) { 3132 in2->disconnect_inputs(NULL, this); 3133 } 3134 } 3135 } 3136 break; 3137 3138 case Op_DecodeN: 3139 case Op_DecodeNKlass: 3140 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3141 // DecodeN could be pinned when it can't be fold into 3142 // an address expression, see the code for Op_CastPP above. 3143 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3144 break; 3145 3146 case Op_EncodeP: 3147 case Op_EncodePKlass: { 3148 Node* in1 = n->in(1); 3149 if (in1->is_DecodeNarrowPtr()) { 3150 n->subsume_by(in1->in(1), this); 3151 } else if (in1->Opcode() == Op_ConP) { 3152 const Type* t = in1->bottom_type(); 3153 if (t == TypePtr::NULL_PTR) { 3154 assert(t->isa_oopptr(), "null klass?"); 3155 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3156 } else if (t->isa_oopptr()) { 3157 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3158 } else if (t->isa_klassptr()) { 3159 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3160 } 3161 } 3162 if (in1->outcnt() == 0) { 3163 in1->disconnect_inputs(NULL, this); 3164 } 3165 break; 3166 } 3167 3168 case Op_Proj: { 3169 if (OptimizeStringConcat) { 3170 ProjNode* p = n->as_Proj(); 3171 if (p->_is_io_use) { 3172 // Separate projections were used for the exception path which 3173 // are normally removed by a late inline. If it wasn't inlined 3174 // then they will hang around and should just be replaced with 3175 // the original one. 3176 Node* proj = NULL; 3177 // Replace with just one 3178 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3179 Node *use = i.get(); 3180 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3181 proj = use; 3182 break; 3183 } 3184 } 3185 assert(proj != NULL, "must be found"); 3186 p->subsume_by(proj, this); 3187 } 3188 } 3189 break; 3190 } 3191 3192 case Op_Phi: 3193 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3194 // The EncodeP optimization may create Phi with the same edges 3195 // for all paths. It is not handled well by Register Allocator. 3196 Node* unique_in = n->in(1); 3197 assert(unique_in != NULL, ""); 3198 uint cnt = n->req(); 3199 for (uint i = 2; i < cnt; i++) { 3200 Node* m = n->in(i); 3201 assert(m != NULL, ""); 3202 if (unique_in != m) 3203 unique_in = NULL; 3204 } 3205 if (unique_in != NULL) { 3206 n->subsume_by(unique_in, this); 3207 } 3208 } 3209 break; 3210 3211 #endif 3212 3213 #ifdef ASSERT 3214 case Op_CastII: 3215 // Verify that all range check dependent CastII nodes were removed. 3216 if (n->isa_CastII()->has_range_check()) { 3217 n->dump(3); 3218 assert(false, "Range check dependent CastII node was not removed"); 3219 } 3220 break; 3221 #endif 3222 3223 case Op_ModI: 3224 if (UseDivMod) { 3225 // Check if a%b and a/b both exist 3226 Node* d = n->find_similar(Op_DivI); 3227 if (d) { 3228 // Replace them with a fused divmod if supported 3229 if (Matcher::has_match_rule(Op_DivModI)) { 3230 DivModINode* divmod = DivModINode::make(n); 3231 d->subsume_by(divmod->div_proj(), this); 3232 n->subsume_by(divmod->mod_proj(), this); 3233 } else { 3234 // replace a%b with a-((a/b)*b) 3235 Node* mult = new MulINode(d, d->in(2)); 3236 Node* sub = new SubINode(d->in(1), mult); 3237 n->subsume_by(sub, this); 3238 } 3239 } 3240 } 3241 break; 3242 3243 case Op_ModL: 3244 if (UseDivMod) { 3245 // Check if a%b and a/b both exist 3246 Node* d = n->find_similar(Op_DivL); 3247 if (d) { 3248 // Replace them with a fused divmod if supported 3249 if (Matcher::has_match_rule(Op_DivModL)) { 3250 DivModLNode* divmod = DivModLNode::make(n); 3251 d->subsume_by(divmod->div_proj(), this); 3252 n->subsume_by(divmod->mod_proj(), this); 3253 } else { 3254 // replace a%b with a-((a/b)*b) 3255 Node* mult = new MulLNode(d, d->in(2)); 3256 Node* sub = new SubLNode(d->in(1), mult); 3257 n->subsume_by(sub, this); 3258 } 3259 } 3260 } 3261 break; 3262 3263 case Op_LoadVector: 3264 case Op_StoreVector: 3265 break; 3266 3267 case Op_AddReductionVI: 3268 case Op_AddReductionVL: 3269 case Op_AddReductionVF: 3270 case Op_AddReductionVD: 3271 case Op_MulReductionVI: 3272 case Op_MulReductionVL: 3273 case Op_MulReductionVF: 3274 case Op_MulReductionVD: 3275 break; 3276 3277 case Op_PackB: 3278 case Op_PackS: 3279 case Op_PackI: 3280 case Op_PackF: 3281 case Op_PackL: 3282 case Op_PackD: 3283 if (n->req()-1 > 2) { 3284 // Replace many operand PackNodes with a binary tree for matching 3285 PackNode* p = (PackNode*) n; 3286 Node* btp = p->binary_tree_pack(1, n->req()); 3287 n->subsume_by(btp, this); 3288 } 3289 break; 3290 case Op_Loop: 3291 case Op_CountedLoop: 3292 if (n->as_Loop()->is_inner_loop()) { 3293 frc.inc_inner_loop_count(); 3294 } 3295 break; 3296 case Op_LShiftI: 3297 case Op_RShiftI: 3298 case Op_URShiftI: 3299 case Op_LShiftL: 3300 case Op_RShiftL: 3301 case Op_URShiftL: 3302 if (Matcher::need_masked_shift_count) { 3303 // The cpu's shift instructions don't restrict the count to the 3304 // lower 5/6 bits. We need to do the masking ourselves. 3305 Node* in2 = n->in(2); 3306 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3307 const TypeInt* t = in2->find_int_type(); 3308 if (t != NULL && t->is_con()) { 3309 juint shift = t->get_con(); 3310 if (shift > mask) { // Unsigned cmp 3311 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3312 } 3313 } else { 3314 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3315 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3316 n->set_req(2, shift); 3317 } 3318 } 3319 if (in2->outcnt() == 0) { // Remove dead node 3320 in2->disconnect_inputs(NULL, this); 3321 } 3322 } 3323 break; 3324 case Op_MemBarStoreStore: 3325 case Op_MemBarRelease: 3326 // Break the link with AllocateNode: it is no longer useful and 3327 // confuses register allocation. 3328 if (n->req() > MemBarNode::Precedent) { 3329 n->set_req(MemBarNode::Precedent, top()); 3330 } 3331 break; 3332 case Op_RangeCheck: { 3333 RangeCheckNode* rc = n->as_RangeCheck(); 3334 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3335 n->subsume_by(iff, this); 3336 frc._tests.push(iff); 3337 break; 3338 } 3339 case Op_ConvI2L: { 3340 if (!Matcher::convi2l_type_required) { 3341 // Code generation on some platforms doesn't need accurate 3342 // ConvI2L types. Widening the type can help remove redundant 3343 // address computations. 3344 n->as_Type()->set_type(TypeLong::INT); 3345 ResourceMark rm; 3346 Node_List wq; 3347 wq.push(n); 3348 for (uint next = 0; next < wq.size(); next++) { 3349 Node *m = wq.at(next); 3350 3351 for(;;) { 3352 // Loop over all nodes with identical inputs edges as m 3353 Node* k = m->find_similar(m->Opcode()); 3354 if (k == NULL) { 3355 break; 3356 } 3357 // Push their uses so we get a chance to remove node made 3358 // redundant 3359 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3360 Node* u = k->fast_out(i); 3361 assert(!wq.contains(u), "shouldn't process one node several times"); 3362 if (u->Opcode() == Op_LShiftL || 3363 u->Opcode() == Op_AddL || 3364 u->Opcode() == Op_SubL || 3365 u->Opcode() == Op_AddP) { 3366 wq.push(u); 3367 } 3368 } 3369 // Replace all nodes with identical edges as m with m 3370 k->subsume_by(m, this); 3371 } 3372 } 3373 } 3374 break; 3375 } 3376 #ifdef ASSERT 3377 case Op_ValueTypePtr: 3378 case Op_ValueType: { 3379 n->dump(-1); 3380 assert(false, "value type node was not removed"); 3381 break; 3382 } 3383 #endif 3384 default: 3385 assert( !n->is_Call(), "" ); 3386 assert( !n->is_Mem(), "" ); 3387 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3388 break; 3389 } 3390 3391 // Collect CFG split points 3392 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3393 frc._tests.push(n); 3394 } 3395 } 3396 3397 //------------------------------final_graph_reshaping_walk--------------------- 3398 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3399 // requires that the walk visits a node's inputs before visiting the node. 3400 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3401 ResourceArea *area = Thread::current()->resource_area(); 3402 Unique_Node_List sfpt(area); 3403 3404 frc._visited.set(root->_idx); // first, mark node as visited 3405 uint cnt = root->req(); 3406 Node *n = root; 3407 uint i = 0; 3408 while (true) { 3409 if (i < cnt) { 3410 // Place all non-visited non-null inputs onto stack 3411 Node* m = n->in(i); 3412 ++i; 3413 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3414 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3415 // compute worst case interpreter size in case of a deoptimization 3416 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3417 3418 sfpt.push(m); 3419 } 3420 cnt = m->req(); 3421 nstack.push(n, i); // put on stack parent and next input's index 3422 n = m; 3423 i = 0; 3424 } 3425 } else { 3426 // Now do post-visit work 3427 final_graph_reshaping_impl( n, frc ); 3428 if (nstack.is_empty()) 3429 break; // finished 3430 n = nstack.node(); // Get node from stack 3431 cnt = n->req(); 3432 i = nstack.index(); 3433 nstack.pop(); // Shift to the next node on stack 3434 } 3435 } 3436 3437 // Skip next transformation if compressed oops are not used. 3438 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3439 (!UseCompressedOops && !UseCompressedClassPointers)) 3440 return; 3441 3442 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3443 // It could be done for an uncommon traps or any safepoints/calls 3444 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3445 while (sfpt.size() > 0) { 3446 n = sfpt.pop(); 3447 JVMState *jvms = n->as_SafePoint()->jvms(); 3448 assert(jvms != NULL, "sanity"); 3449 int start = jvms->debug_start(); 3450 int end = n->req(); 3451 bool is_uncommon = (n->is_CallStaticJava() && 3452 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3453 for (int j = start; j < end; j++) { 3454 Node* in = n->in(j); 3455 if (in->is_DecodeNarrowPtr()) { 3456 bool safe_to_skip = true; 3457 if (!is_uncommon ) { 3458 // Is it safe to skip? 3459 for (uint i = 0; i < in->outcnt(); i++) { 3460 Node* u = in->raw_out(i); 3461 if (!u->is_SafePoint() || 3462 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3463 safe_to_skip = false; 3464 } 3465 } 3466 } 3467 if (safe_to_skip) { 3468 n->set_req(j, in->in(1)); 3469 } 3470 if (in->outcnt() == 0) { 3471 in->disconnect_inputs(NULL, this); 3472 } 3473 } 3474 } 3475 } 3476 } 3477 3478 //------------------------------final_graph_reshaping-------------------------- 3479 // Final Graph Reshaping. 3480 // 3481 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3482 // and not commoned up and forced early. Must come after regular 3483 // optimizations to avoid GVN undoing the cloning. Clone constant 3484 // inputs to Loop Phis; these will be split by the allocator anyways. 3485 // Remove Opaque nodes. 3486 // (2) Move last-uses by commutative operations to the left input to encourage 3487 // Intel update-in-place two-address operations and better register usage 3488 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3489 // calls canonicalizing them back. 3490 // (3) Count the number of double-precision FP ops, single-precision FP ops 3491 // and call sites. On Intel, we can get correct rounding either by 3492 // forcing singles to memory (requires extra stores and loads after each 3493 // FP bytecode) or we can set a rounding mode bit (requires setting and 3494 // clearing the mode bit around call sites). The mode bit is only used 3495 // if the relative frequency of single FP ops to calls is low enough. 3496 // This is a key transform for SPEC mpeg_audio. 3497 // (4) Detect infinite loops; blobs of code reachable from above but not 3498 // below. Several of the Code_Gen algorithms fail on such code shapes, 3499 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3500 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3501 // Detection is by looking for IfNodes where only 1 projection is 3502 // reachable from below or CatchNodes missing some targets. 3503 // (5) Assert for insane oop offsets in debug mode. 3504 3505 bool Compile::final_graph_reshaping() { 3506 // an infinite loop may have been eliminated by the optimizer, 3507 // in which case the graph will be empty. 3508 if (root()->req() == 1) { 3509 record_method_not_compilable("trivial infinite loop"); 3510 return true; 3511 } 3512 3513 // Expensive nodes have their control input set to prevent the GVN 3514 // from freely commoning them. There's no GVN beyond this point so 3515 // no need to keep the control input. We want the expensive nodes to 3516 // be freely moved to the least frequent code path by gcm. 3517 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3518 for (int i = 0; i < expensive_count(); i++) { 3519 _expensive_nodes->at(i)->set_req(0, NULL); 3520 } 3521 3522 Final_Reshape_Counts frc; 3523 3524 // Visit everybody reachable! 3525 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3526 Node_Stack nstack(live_nodes() >> 1); 3527 final_graph_reshaping_walk(nstack, root(), frc); 3528 3529 // Check for unreachable (from below) code (i.e., infinite loops). 3530 for( uint i = 0; i < frc._tests.size(); i++ ) { 3531 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3532 // Get number of CFG targets. 3533 // Note that PCTables include exception targets after calls. 3534 uint required_outcnt = n->required_outcnt(); 3535 if (n->outcnt() != required_outcnt) { 3536 // Check for a few special cases. Rethrow Nodes never take the 3537 // 'fall-thru' path, so expected kids is 1 less. 3538 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3539 if (n->in(0)->in(0)->is_Call()) { 3540 CallNode *call = n->in(0)->in(0)->as_Call(); 3541 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3542 required_outcnt--; // Rethrow always has 1 less kid 3543 } else if (call->req() > TypeFunc::Parms && 3544 call->is_CallDynamicJava()) { 3545 // Check for null receiver. In such case, the optimizer has 3546 // detected that the virtual call will always result in a null 3547 // pointer exception. The fall-through projection of this CatchNode 3548 // will not be populated. 3549 Node *arg0 = call->in(TypeFunc::Parms); 3550 if (arg0->is_Type() && 3551 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3552 required_outcnt--; 3553 } 3554 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3555 call->req() > TypeFunc::Parms+1 && 3556 call->is_CallStaticJava()) { 3557 // Check for negative array length. In such case, the optimizer has 3558 // detected that the allocation attempt will always result in an 3559 // exception. There is no fall-through projection of this CatchNode . 3560 Node *arg1 = call->in(TypeFunc::Parms+1); 3561 if (arg1->is_Type() && 3562 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3563 required_outcnt--; 3564 } 3565 } 3566 } 3567 } 3568 // Recheck with a better notion of 'required_outcnt' 3569 if (n->outcnt() != required_outcnt) { 3570 assert(false, "malformed control flow"); 3571 record_method_not_compilable("malformed control flow"); 3572 return true; // Not all targets reachable! 3573 } 3574 } 3575 // Check that I actually visited all kids. Unreached kids 3576 // must be infinite loops. 3577 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3578 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3579 record_method_not_compilable("infinite loop"); 3580 assert(false, "infinite loop"); 3581 return true; // Found unvisited kid; must be unreach 3582 } 3583 } 3584 3585 // If original bytecodes contained a mixture of floats and doubles 3586 // check if the optimizer has made it homogenous, item (3). 3587 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3588 frc.get_float_count() > 32 && 3589 frc.get_double_count() == 0 && 3590 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3591 set_24_bit_selection_and_mode( false, true ); 3592 } 3593 3594 set_java_calls(frc.get_java_call_count()); 3595 set_inner_loops(frc.get_inner_loop_count()); 3596 3597 // No infinite loops, no reason to bail out. 3598 return false; 3599 } 3600 3601 //-----------------------------too_many_traps---------------------------------- 3602 // Report if there are too many traps at the current method and bci. 3603 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3604 bool Compile::too_many_traps(ciMethod* method, 3605 int bci, 3606 Deoptimization::DeoptReason reason) { 3607 ciMethodData* md = method->method_data(); 3608 if (md->is_empty()) { 3609 // Assume the trap has not occurred, or that it occurred only 3610 // because of a transient condition during start-up in the interpreter. 3611 return false; 3612 } 3613 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3614 if (md->has_trap_at(bci, m, reason) != 0) { 3615 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3616 // Also, if there are multiple reasons, or if there is no per-BCI record, 3617 // assume the worst. 3618 if (log()) 3619 log()->elem("observe trap='%s' count='%d'", 3620 Deoptimization::trap_reason_name(reason), 3621 md->trap_count(reason)); 3622 return true; 3623 } else { 3624 // Ignore method/bci and see if there have been too many globally. 3625 return too_many_traps(reason, md); 3626 } 3627 } 3628 3629 // Less-accurate variant which does not require a method and bci. 3630 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3631 ciMethodData* logmd) { 3632 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3633 // Too many traps globally. 3634 // Note that we use cumulative trap_count, not just md->trap_count. 3635 if (log()) { 3636 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3637 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3638 Deoptimization::trap_reason_name(reason), 3639 mcount, trap_count(reason)); 3640 } 3641 return true; 3642 } else { 3643 // The coast is clear. 3644 return false; 3645 } 3646 } 3647 3648 //--------------------------too_many_recompiles-------------------------------- 3649 // Report if there are too many recompiles at the current method and bci. 3650 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3651 // Is not eager to return true, since this will cause the compiler to use 3652 // Action_none for a trap point, to avoid too many recompilations. 3653 bool Compile::too_many_recompiles(ciMethod* method, 3654 int bci, 3655 Deoptimization::DeoptReason reason) { 3656 ciMethodData* md = method->method_data(); 3657 if (md->is_empty()) { 3658 // Assume the trap has not occurred, or that it occurred only 3659 // because of a transient condition during start-up in the interpreter. 3660 return false; 3661 } 3662 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3663 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3664 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3665 Deoptimization::DeoptReason per_bc_reason 3666 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3667 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3668 if ((per_bc_reason == Deoptimization::Reason_none 3669 || md->has_trap_at(bci, m, reason) != 0) 3670 // The trap frequency measure we care about is the recompile count: 3671 && md->trap_recompiled_at(bci, m) 3672 && md->overflow_recompile_count() >= bc_cutoff) { 3673 // Do not emit a trap here if it has already caused recompilations. 3674 // Also, if there are multiple reasons, or if there is no per-BCI record, 3675 // assume the worst. 3676 if (log()) 3677 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3678 Deoptimization::trap_reason_name(reason), 3679 md->trap_count(reason), 3680 md->overflow_recompile_count()); 3681 return true; 3682 } else if (trap_count(reason) != 0 3683 && decompile_count() >= m_cutoff) { 3684 // Too many recompiles globally, and we have seen this sort of trap. 3685 // Use cumulative decompile_count, not just md->decompile_count. 3686 if (log()) 3687 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3688 Deoptimization::trap_reason_name(reason), 3689 md->trap_count(reason), trap_count(reason), 3690 md->decompile_count(), decompile_count()); 3691 return true; 3692 } else { 3693 // The coast is clear. 3694 return false; 3695 } 3696 } 3697 3698 // Compute when not to trap. Used by matching trap based nodes and 3699 // NullCheck optimization. 3700 void Compile::set_allowed_deopt_reasons() { 3701 _allowed_reasons = 0; 3702 if (is_method_compilation()) { 3703 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3704 assert(rs < BitsPerInt, "recode bit map"); 3705 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3706 _allowed_reasons |= nth_bit(rs); 3707 } 3708 } 3709 } 3710 } 3711 3712 #ifndef PRODUCT 3713 //------------------------------verify_graph_edges--------------------------- 3714 // Walk the Graph and verify that there is a one-to-one correspondence 3715 // between Use-Def edges and Def-Use edges in the graph. 3716 void Compile::verify_graph_edges(bool no_dead_code) { 3717 if (VerifyGraphEdges) { 3718 ResourceArea *area = Thread::current()->resource_area(); 3719 Unique_Node_List visited(area); 3720 // Call recursive graph walk to check edges 3721 _root->verify_edges(visited); 3722 if (no_dead_code) { 3723 // Now make sure that no visited node is used by an unvisited node. 3724 bool dead_nodes = false; 3725 Unique_Node_List checked(area); 3726 while (visited.size() > 0) { 3727 Node* n = visited.pop(); 3728 checked.push(n); 3729 for (uint i = 0; i < n->outcnt(); i++) { 3730 Node* use = n->raw_out(i); 3731 if (checked.member(use)) continue; // already checked 3732 if (visited.member(use)) continue; // already in the graph 3733 if (use->is_Con()) continue; // a dead ConNode is OK 3734 // At this point, we have found a dead node which is DU-reachable. 3735 if (!dead_nodes) { 3736 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3737 dead_nodes = true; 3738 } 3739 use->dump(2); 3740 tty->print_cr("---"); 3741 checked.push(use); // No repeats; pretend it is now checked. 3742 } 3743 } 3744 assert(!dead_nodes, "using nodes must be reachable from root"); 3745 } 3746 } 3747 } 3748 3749 // Verify GC barriers consistency 3750 // Currently supported: 3751 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3752 void Compile::verify_barriers() { 3753 if (UseG1GC) { 3754 // Verify G1 pre-barriers 3755 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + SATBMarkQueue::byte_offset_of_active()); 3756 3757 ResourceArea *area = Thread::current()->resource_area(); 3758 Unique_Node_List visited(area); 3759 Node_List worklist(area); 3760 // We're going to walk control flow backwards starting from the Root 3761 worklist.push(_root); 3762 while (worklist.size() > 0) { 3763 Node* x = worklist.pop(); 3764 if (x == NULL || x == top()) continue; 3765 if (visited.member(x)) { 3766 continue; 3767 } else { 3768 visited.push(x); 3769 } 3770 3771 if (x->is_Region()) { 3772 for (uint i = 1; i < x->req(); i++) { 3773 worklist.push(x->in(i)); 3774 } 3775 } else { 3776 worklist.push(x->in(0)); 3777 // We are looking for the pattern: 3778 // /->ThreadLocal 3779 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3780 // \->ConI(0) 3781 // We want to verify that the If and the LoadB have the same control 3782 // See GraphKit::g1_write_barrier_pre() 3783 if (x->is_If()) { 3784 IfNode *iff = x->as_If(); 3785 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3786 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3787 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3788 && cmp->in(1)->is_Load()) { 3789 LoadNode* load = cmp->in(1)->as_Load(); 3790 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3791 && load->in(2)->in(3)->is_Con() 3792 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3793 3794 Node* if_ctrl = iff->in(0); 3795 Node* load_ctrl = load->in(0); 3796 3797 if (if_ctrl != load_ctrl) { 3798 // Skip possible CProj->NeverBranch in infinite loops 3799 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3800 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3801 if_ctrl = if_ctrl->in(0)->in(0); 3802 } 3803 } 3804 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3805 } 3806 } 3807 } 3808 } 3809 } 3810 } 3811 } 3812 } 3813 3814 #endif 3815 3816 // The Compile object keeps track of failure reasons separately from the ciEnv. 3817 // This is required because there is not quite a 1-1 relation between the 3818 // ciEnv and its compilation task and the Compile object. Note that one 3819 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3820 // to backtrack and retry without subsuming loads. Other than this backtracking 3821 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3822 // by the logic in C2Compiler. 3823 void Compile::record_failure(const char* reason) { 3824 if (log() != NULL) { 3825 log()->elem("failure reason='%s' phase='compile'", reason); 3826 } 3827 if (_failure_reason == NULL) { 3828 // Record the first failure reason. 3829 _failure_reason = reason; 3830 } 3831 3832 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3833 C->print_method(PHASE_FAILURE); 3834 } 3835 _root = NULL; // flush the graph, too 3836 } 3837 3838 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3839 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3840 _phase_name(name), _dolog(CITimeVerbose) 3841 { 3842 if (_dolog) { 3843 C = Compile::current(); 3844 _log = C->log(); 3845 } else { 3846 C = NULL; 3847 _log = NULL; 3848 } 3849 if (_log != NULL) { 3850 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3851 _log->stamp(); 3852 _log->end_head(); 3853 } 3854 } 3855 3856 Compile::TracePhase::~TracePhase() { 3857 3858 C = Compile::current(); 3859 if (_dolog) { 3860 _log = C->log(); 3861 } else { 3862 _log = NULL; 3863 } 3864 3865 #ifdef ASSERT 3866 if (PrintIdealNodeCount) { 3867 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3868 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3869 } 3870 3871 if (VerifyIdealNodeCount) { 3872 Compile::current()->print_missing_nodes(); 3873 } 3874 #endif 3875 3876 if (_log != NULL) { 3877 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3878 } 3879 } 3880 3881 //============================================================================= 3882 // Two Constant's are equal when the type and the value are equal. 3883 bool Compile::Constant::operator==(const Constant& other) { 3884 if (type() != other.type() ) return false; 3885 if (can_be_reused() != other.can_be_reused()) return false; 3886 // For floating point values we compare the bit pattern. 3887 switch (type()) { 3888 case T_INT: 3889 case T_FLOAT: return (_v._value.i == other._v._value.i); 3890 case T_LONG: 3891 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3892 case T_OBJECT: 3893 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3894 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3895 case T_METADATA: return (_v._metadata == other._v._metadata); 3896 default: ShouldNotReachHere(); return false; 3897 } 3898 } 3899 3900 static int type_to_size_in_bytes(BasicType t) { 3901 switch (t) { 3902 case T_INT: return sizeof(jint ); 3903 case T_LONG: return sizeof(jlong ); 3904 case T_FLOAT: return sizeof(jfloat ); 3905 case T_DOUBLE: return sizeof(jdouble); 3906 case T_METADATA: return sizeof(Metadata*); 3907 // We use T_VOID as marker for jump-table entries (labels) which 3908 // need an internal word relocation. 3909 case T_VOID: 3910 case T_ADDRESS: 3911 case T_OBJECT: return sizeof(jobject); 3912 default: 3913 ShouldNotReachHere(); 3914 return -1; 3915 } 3916 } 3917 3918 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3919 // sort descending 3920 if (a->freq() > b->freq()) return -1; 3921 if (a->freq() < b->freq()) return 1; 3922 return 0; 3923 } 3924 3925 void Compile::ConstantTable::calculate_offsets_and_size() { 3926 // First, sort the array by frequencies. 3927 _constants.sort(qsort_comparator); 3928 3929 #ifdef ASSERT 3930 // Make sure all jump-table entries were sorted to the end of the 3931 // array (they have a negative frequency). 3932 bool found_void = false; 3933 for (int i = 0; i < _constants.length(); i++) { 3934 Constant con = _constants.at(i); 3935 if (con.type() == T_VOID) 3936 found_void = true; // jump-tables 3937 else 3938 assert(!found_void, "wrong sorting"); 3939 } 3940 #endif 3941 3942 int offset = 0; 3943 for (int i = 0; i < _constants.length(); i++) { 3944 Constant* con = _constants.adr_at(i); 3945 3946 // Align offset for type. 3947 int typesize = type_to_size_in_bytes(con->type()); 3948 offset = align_up(offset, typesize); 3949 con->set_offset(offset); // set constant's offset 3950 3951 if (con->type() == T_VOID) { 3952 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3953 offset = offset + typesize * n->outcnt(); // expand jump-table 3954 } else { 3955 offset = offset + typesize; 3956 } 3957 } 3958 3959 // Align size up to the next section start (which is insts; see 3960 // CodeBuffer::align_at_start). 3961 assert(_size == -1, "already set?"); 3962 _size = align_up(offset, (int)CodeEntryAlignment); 3963 } 3964 3965 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3966 MacroAssembler _masm(&cb); 3967 for (int i = 0; i < _constants.length(); i++) { 3968 Constant con = _constants.at(i); 3969 address constant_addr = NULL; 3970 switch (con.type()) { 3971 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 3972 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3973 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3974 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3975 case T_OBJECT: { 3976 jobject obj = con.get_jobject(); 3977 int oop_index = _masm.oop_recorder()->find_index(obj); 3978 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3979 break; 3980 } 3981 case T_ADDRESS: { 3982 address addr = (address) con.get_jobject(); 3983 constant_addr = _masm.address_constant(addr); 3984 break; 3985 } 3986 // We use T_VOID as marker for jump-table entries (labels) which 3987 // need an internal word relocation. 3988 case T_VOID: { 3989 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3990 // Fill the jump-table with a dummy word. The real value is 3991 // filled in later in fill_jump_table. 3992 address dummy = (address) n; 3993 constant_addr = _masm.address_constant(dummy); 3994 // Expand jump-table 3995 for (uint i = 1; i < n->outcnt(); i++) { 3996 address temp_addr = _masm.address_constant(dummy + i); 3997 assert(temp_addr, "consts section too small"); 3998 } 3999 break; 4000 } 4001 case T_METADATA: { 4002 Metadata* obj = con.get_metadata(); 4003 int metadata_index = _masm.oop_recorder()->find_index(obj); 4004 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4005 break; 4006 } 4007 default: ShouldNotReachHere(); 4008 } 4009 assert(constant_addr, "consts section too small"); 4010 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4011 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4012 } 4013 } 4014 4015 int Compile::ConstantTable::find_offset(Constant& con) const { 4016 int idx = _constants.find(con); 4017 assert(idx != -1, "constant must be in constant table"); 4018 int offset = _constants.at(idx).offset(); 4019 assert(offset != -1, "constant table not emitted yet?"); 4020 return offset; 4021 } 4022 4023 void Compile::ConstantTable::add(Constant& con) { 4024 if (con.can_be_reused()) { 4025 int idx = _constants.find(con); 4026 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4027 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4028 return; 4029 } 4030 } 4031 (void) _constants.append(con); 4032 } 4033 4034 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4035 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4036 Constant con(type, value, b->_freq); 4037 add(con); 4038 return con; 4039 } 4040 4041 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4042 Constant con(metadata); 4043 add(con); 4044 return con; 4045 } 4046 4047 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4048 jvalue value; 4049 BasicType type = oper->type()->basic_type(); 4050 switch (type) { 4051 case T_LONG: value.j = oper->constantL(); break; 4052 case T_FLOAT: value.f = oper->constantF(); break; 4053 case T_DOUBLE: value.d = oper->constantD(); break; 4054 case T_OBJECT: 4055 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4056 case T_METADATA: return add((Metadata*)oper->constant()); break; 4057 default: guarantee(false, "unhandled type: %s", type2name(type)); 4058 } 4059 return add(n, type, value); 4060 } 4061 4062 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4063 jvalue value; 4064 // We can use the node pointer here to identify the right jump-table 4065 // as this method is called from Compile::Fill_buffer right before 4066 // the MachNodes are emitted and the jump-table is filled (means the 4067 // MachNode pointers do not change anymore). 4068 value.l = (jobject) n; 4069 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4070 add(con); 4071 return con; 4072 } 4073 4074 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4075 // If called from Compile::scratch_emit_size do nothing. 4076 if (Compile::current()->in_scratch_emit_size()) return; 4077 4078 assert(labels.is_nonempty(), "must be"); 4079 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4080 4081 // Since MachConstantNode::constant_offset() also contains 4082 // table_base_offset() we need to subtract the table_base_offset() 4083 // to get the plain offset into the constant table. 4084 int offset = n->constant_offset() - table_base_offset(); 4085 4086 MacroAssembler _masm(&cb); 4087 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4088 4089 for (uint i = 0; i < n->outcnt(); i++) { 4090 address* constant_addr = &jump_table_base[i]; 4091 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)); 4092 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4093 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4094 } 4095 } 4096 4097 //----------------------------static_subtype_check----------------------------- 4098 // Shortcut important common cases when superklass is exact: 4099 // (0) superklass is java.lang.Object (can occur in reflective code) 4100 // (1) subklass is already limited to a subtype of superklass => always ok 4101 // (2) subklass does not overlap with superklass => always fail 4102 // (3) superklass has NO subtypes and we can check with a simple compare. 4103 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4104 if (StressReflectiveCode || superk == NULL || subk == NULL) { 4105 return SSC_full_test; // Let caller generate the general case. 4106 } 4107 4108 if (!EnableMVT && !EnableValhalla && superk == env()->Object_klass()) { 4109 return SSC_always_true; // (0) this test cannot fail 4110 } 4111 4112 ciType* superelem = superk; 4113 if (superelem->is_array_klass()) 4114 superelem = superelem->as_array_klass()->base_element_type(); 4115 4116 if (!subk->is_interface()) { // cannot trust static interface types yet 4117 if (subk->is_subtype_of(superk)) { 4118 return SSC_always_true; // (1) false path dead; no dynamic test needed 4119 } 4120 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4121 !superk->is_subtype_of(subk)) { 4122 return SSC_always_false; 4123 } 4124 } 4125 4126 // If casting to an instance klass, it must have no subtypes 4127 if (superk->is_interface()) { 4128 // Cannot trust interfaces yet. 4129 // %%% S.B. superk->nof_implementors() == 1 4130 } else if (superelem->is_instance_klass()) { 4131 ciInstanceKlass* ik = superelem->as_instance_klass(); 4132 if (!ik->has_subklass() && !ik->is_interface()) { 4133 if (!ik->is_final()) { 4134 // Add a dependency if there is a chance of a later subclass. 4135 dependencies()->assert_leaf_type(ik); 4136 } 4137 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4138 } 4139 } else { 4140 // A primitive array type has no subtypes. 4141 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4142 } 4143 4144 return SSC_full_test; 4145 } 4146 4147 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4148 #ifdef _LP64 4149 // The scaled index operand to AddP must be a clean 64-bit value. 4150 // Java allows a 32-bit int to be incremented to a negative 4151 // value, which appears in a 64-bit register as a large 4152 // positive number. Using that large positive number as an 4153 // operand in pointer arithmetic has bad consequences. 4154 // On the other hand, 32-bit overflow is rare, and the possibility 4155 // can often be excluded, if we annotate the ConvI2L node with 4156 // a type assertion that its value is known to be a small positive 4157 // number. (The prior range check has ensured this.) 4158 // This assertion is used by ConvI2LNode::Ideal. 4159 int index_max = max_jint - 1; // array size is max_jint, index is one less 4160 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4161 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4162 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4163 #endif 4164 return idx; 4165 } 4166 4167 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4168 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4169 if (ctrl != NULL) { 4170 // Express control dependency by a CastII node with a narrow type. 4171 value = new CastIINode(value, itype, false, true /* range check dependency */); 4172 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4173 // node from floating above the range check during loop optimizations. Otherwise, the 4174 // ConvI2L node may be eliminated independently of the range check, causing the data path 4175 // to become TOP while the control path is still there (although it's unreachable). 4176 value->set_req(0, ctrl); 4177 // Save CastII node to remove it after loop optimizations. 4178 phase->C->add_range_check_cast(value); 4179 value = phase->transform(value); 4180 } 4181 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4182 return phase->transform(new ConvI2LNode(value, ltype)); 4183 } 4184 4185 // The message about the current inlining is accumulated in 4186 // _print_inlining_stream and transfered into the _print_inlining_list 4187 // once we know whether inlining succeeds or not. For regular 4188 // inlining, messages are appended to the buffer pointed by 4189 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4190 // a new buffer is added after _print_inlining_idx in the list. This 4191 // way we can update the inlining message for late inlining call site 4192 // when the inlining is attempted again. 4193 void Compile::print_inlining_init() { 4194 if (print_inlining() || print_intrinsics()) { 4195 _print_inlining_stream = new stringStream(); 4196 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4197 } 4198 } 4199 4200 void Compile::print_inlining_reinit() { 4201 if (print_inlining() || print_intrinsics()) { 4202 // Re allocate buffer when we change ResourceMark 4203 _print_inlining_stream = new stringStream(); 4204 } 4205 } 4206 4207 void Compile::print_inlining_reset() { 4208 _print_inlining_stream->reset(); 4209 } 4210 4211 void Compile::print_inlining_commit() { 4212 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4213 // Transfer the message from _print_inlining_stream to the current 4214 // _print_inlining_list buffer and clear _print_inlining_stream. 4215 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4216 print_inlining_reset(); 4217 } 4218 4219 void Compile::print_inlining_push() { 4220 // Add new buffer to the _print_inlining_list at current position 4221 _print_inlining_idx++; 4222 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4223 } 4224 4225 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4226 return _print_inlining_list->at(_print_inlining_idx); 4227 } 4228 4229 void Compile::print_inlining_update(CallGenerator* cg) { 4230 if (print_inlining() || print_intrinsics()) { 4231 if (!cg->is_late_inline()) { 4232 if (print_inlining_current().cg() != NULL) { 4233 print_inlining_push(); 4234 } 4235 print_inlining_commit(); 4236 } else { 4237 if (print_inlining_current().cg() != cg && 4238 (print_inlining_current().cg() != NULL || 4239 print_inlining_current().ss()->size() != 0)) { 4240 print_inlining_push(); 4241 } 4242 print_inlining_commit(); 4243 print_inlining_current().set_cg(cg); 4244 } 4245 } 4246 } 4247 4248 void Compile::print_inlining_move_to(CallGenerator* cg) { 4249 // We resume inlining at a late inlining call site. Locate the 4250 // corresponding inlining buffer so that we can update it. 4251 if (print_inlining()) { 4252 for (int i = 0; i < _print_inlining_list->length(); i++) { 4253 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4254 _print_inlining_idx = i; 4255 return; 4256 } 4257 } 4258 ShouldNotReachHere(); 4259 } 4260 } 4261 4262 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4263 if (print_inlining()) { 4264 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4265 assert(print_inlining_current().cg() == cg, "wrong entry"); 4266 // replace message with new message 4267 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4268 print_inlining_commit(); 4269 print_inlining_current().set_cg(cg); 4270 } 4271 } 4272 4273 void Compile::print_inlining_assert_ready() { 4274 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4275 } 4276 4277 void Compile::process_print_inlining() { 4278 bool do_print_inlining = print_inlining() || print_intrinsics(); 4279 if (do_print_inlining || log() != NULL) { 4280 // Print inlining message for candidates that we couldn't inline 4281 // for lack of space 4282 for (int i = 0; i < _late_inlines.length(); i++) { 4283 CallGenerator* cg = _late_inlines.at(i); 4284 if (!cg->is_mh_late_inline()) { 4285 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4286 if (do_print_inlining) { 4287 cg->print_inlining_late(msg); 4288 } 4289 log_late_inline_failure(cg, msg); 4290 } 4291 } 4292 } 4293 if (do_print_inlining) { 4294 ResourceMark rm; 4295 stringStream ss; 4296 for (int i = 0; i < _print_inlining_list->length(); i++) { 4297 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4298 } 4299 size_t end = ss.size(); 4300 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4301 strncpy(_print_inlining_output, ss.base(), end+1); 4302 _print_inlining_output[end] = 0; 4303 } 4304 } 4305 4306 void Compile::dump_print_inlining() { 4307 if (_print_inlining_output != NULL) { 4308 tty->print_raw(_print_inlining_output); 4309 } 4310 } 4311 4312 void Compile::log_late_inline(CallGenerator* cg) { 4313 if (log() != NULL) { 4314 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4315 cg->unique_id()); 4316 JVMState* p = cg->call_node()->jvms(); 4317 while (p != NULL) { 4318 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4319 p = p->caller(); 4320 } 4321 log()->tail("late_inline"); 4322 } 4323 } 4324 4325 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4326 log_late_inline(cg); 4327 if (log() != NULL) { 4328 log()->inline_fail(msg); 4329 } 4330 } 4331 4332 void Compile::log_inline_id(CallGenerator* cg) { 4333 if (log() != NULL) { 4334 // The LogCompilation tool needs a unique way to identify late 4335 // inline call sites. This id must be unique for this call site in 4336 // this compilation. Try to have it unique across compilations as 4337 // well because it can be convenient when grepping through the log 4338 // file. 4339 // Distinguish OSR compilations from others in case CICountOSR is 4340 // on. 4341 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4342 cg->set_unique_id(id); 4343 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4344 } 4345 } 4346 4347 void Compile::log_inline_failure(const char* msg) { 4348 if (C->log() != NULL) { 4349 C->log()->inline_fail(msg); 4350 } 4351 } 4352 4353 4354 // Dump inlining replay data to the stream. 4355 // Don't change thread state and acquire any locks. 4356 void Compile::dump_inline_data(outputStream* out) { 4357 InlineTree* inl_tree = ilt(); 4358 if (inl_tree != NULL) { 4359 out->print(" inline %d", inl_tree->count()); 4360 inl_tree->dump_replay_data(out); 4361 } 4362 } 4363 4364 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4365 if (n1->Opcode() < n2->Opcode()) return -1; 4366 else if (n1->Opcode() > n2->Opcode()) return 1; 4367 4368 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4369 for (uint i = 1; i < n1->req(); i++) { 4370 if (n1->in(i) < n2->in(i)) return -1; 4371 else if (n1->in(i) > n2->in(i)) return 1; 4372 } 4373 4374 return 0; 4375 } 4376 4377 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4378 Node* n1 = *n1p; 4379 Node* n2 = *n2p; 4380 4381 return cmp_expensive_nodes(n1, n2); 4382 } 4383 4384 void Compile::sort_expensive_nodes() { 4385 if (!expensive_nodes_sorted()) { 4386 _expensive_nodes->sort(cmp_expensive_nodes); 4387 } 4388 } 4389 4390 bool Compile::expensive_nodes_sorted() const { 4391 for (int i = 1; i < _expensive_nodes->length(); i++) { 4392 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4393 return false; 4394 } 4395 } 4396 return true; 4397 } 4398 4399 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4400 if (_expensive_nodes->length() == 0) { 4401 return false; 4402 } 4403 4404 assert(OptimizeExpensiveOps, "optimization off?"); 4405 4406 // Take this opportunity to remove dead nodes from the list 4407 int j = 0; 4408 for (int i = 0; i < _expensive_nodes->length(); i++) { 4409 Node* n = _expensive_nodes->at(i); 4410 if (!n->is_unreachable(igvn)) { 4411 assert(n->is_expensive(), "should be expensive"); 4412 _expensive_nodes->at_put(j, n); 4413 j++; 4414 } 4415 } 4416 _expensive_nodes->trunc_to(j); 4417 4418 // Then sort the list so that similar nodes are next to each other 4419 // and check for at least two nodes of identical kind with same data 4420 // inputs. 4421 sort_expensive_nodes(); 4422 4423 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4424 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4425 return true; 4426 } 4427 } 4428 4429 return false; 4430 } 4431 4432 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4433 if (_expensive_nodes->length() == 0) { 4434 return; 4435 } 4436 4437 assert(OptimizeExpensiveOps, "optimization off?"); 4438 4439 // Sort to bring similar nodes next to each other and clear the 4440 // control input of nodes for which there's only a single copy. 4441 sort_expensive_nodes(); 4442 4443 int j = 0; 4444 int identical = 0; 4445 int i = 0; 4446 bool modified = false; 4447 for (; i < _expensive_nodes->length()-1; i++) { 4448 assert(j <= i, "can't write beyond current index"); 4449 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4450 identical++; 4451 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4452 continue; 4453 } 4454 if (identical > 0) { 4455 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4456 identical = 0; 4457 } else { 4458 Node* n = _expensive_nodes->at(i); 4459 igvn.replace_input_of(n, 0, NULL); 4460 igvn.hash_insert(n); 4461 modified = true; 4462 } 4463 } 4464 if (identical > 0) { 4465 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4466 } else if (_expensive_nodes->length() >= 1) { 4467 Node* n = _expensive_nodes->at(i); 4468 igvn.replace_input_of(n, 0, NULL); 4469 igvn.hash_insert(n); 4470 modified = true; 4471 } 4472 _expensive_nodes->trunc_to(j); 4473 if (modified) { 4474 igvn.optimize(); 4475 } 4476 } 4477 4478 void Compile::add_expensive_node(Node * n) { 4479 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4480 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4481 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4482 if (OptimizeExpensiveOps) { 4483 _expensive_nodes->append(n); 4484 } else { 4485 // Clear control input and let IGVN optimize expensive nodes if 4486 // OptimizeExpensiveOps is off. 4487 n->set_req(0, NULL); 4488 } 4489 } 4490 4491 /** 4492 * Remove the speculative part of types and clean up the graph 4493 */ 4494 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4495 if (UseTypeSpeculation) { 4496 Unique_Node_List worklist; 4497 worklist.push(root()); 4498 int modified = 0; 4499 // Go over all type nodes that carry a speculative type, drop the 4500 // speculative part of the type and enqueue the node for an igvn 4501 // which may optimize it out. 4502 for (uint next = 0; next < worklist.size(); ++next) { 4503 Node *n = worklist.at(next); 4504 if (n->is_Type()) { 4505 TypeNode* tn = n->as_Type(); 4506 const Type* t = tn->type(); 4507 const Type* t_no_spec = t->remove_speculative(); 4508 if (t_no_spec != t) { 4509 bool in_hash = igvn.hash_delete(n); 4510 assert(in_hash, "node should be in igvn hash table"); 4511 tn->set_type(t_no_spec); 4512 igvn.hash_insert(n); 4513 igvn._worklist.push(n); // give it a chance to go away 4514 modified++; 4515 } 4516 } 4517 uint max = n->len(); 4518 for( uint i = 0; i < max; ++i ) { 4519 Node *m = n->in(i); 4520 if (not_a_node(m)) continue; 4521 worklist.push(m); 4522 } 4523 } 4524 // Drop the speculative part of all types in the igvn's type table 4525 igvn.remove_speculative_types(); 4526 if (modified > 0) { 4527 igvn.optimize(); 4528 } 4529 #ifdef ASSERT 4530 // Verify that after the IGVN is over no speculative type has resurfaced 4531 worklist.clear(); 4532 worklist.push(root()); 4533 for (uint next = 0; next < worklist.size(); ++next) { 4534 Node *n = worklist.at(next); 4535 const Type* t = igvn.type_or_null(n); 4536 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4537 if (n->is_Type()) { 4538 t = n->as_Type()->type(); 4539 assert(t == t->remove_speculative(), "no more speculative types"); 4540 } 4541 uint max = n->len(); 4542 for( uint i = 0; i < max; ++i ) { 4543 Node *m = n->in(i); 4544 if (not_a_node(m)) continue; 4545 worklist.push(m); 4546 } 4547 } 4548 igvn.check_no_speculative_types(); 4549 #endif 4550 } 4551 } 4552 4553 // Auxiliary method to support randomized stressing/fuzzing. 4554 // 4555 // This method can be called the arbitrary number of times, with current count 4556 // as the argument. The logic allows selecting a single candidate from the 4557 // running list of candidates as follows: 4558 // int count = 0; 4559 // Cand* selected = null; 4560 // while(cand = cand->next()) { 4561 // if (randomized_select(++count)) { 4562 // selected = cand; 4563 // } 4564 // } 4565 // 4566 // Including count equalizes the chances any candidate is "selected". 4567 // This is useful when we don't have the complete list of candidates to choose 4568 // from uniformly. In this case, we need to adjust the randomicity of the 4569 // selection, or else we will end up biasing the selection towards the latter 4570 // candidates. 4571 // 4572 // Quick back-envelope calculation shows that for the list of n candidates 4573 // the equal probability for the candidate to persist as "best" can be 4574 // achieved by replacing it with "next" k-th candidate with the probability 4575 // of 1/k. It can be easily shown that by the end of the run, the 4576 // probability for any candidate is converged to 1/n, thus giving the 4577 // uniform distribution among all the candidates. 4578 // 4579 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4580 #define RANDOMIZED_DOMAIN_POW 29 4581 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4582 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4583 bool Compile::randomized_select(int count) { 4584 assert(count > 0, "only positive"); 4585 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4586 } 4587 4588 CloneMap& Compile::clone_map() { return _clone_map; } 4589 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4590 4591 void NodeCloneInfo::dump() const { 4592 tty->print(" {%d:%d} ", idx(), gen()); 4593 } 4594 4595 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4596 uint64_t val = value(old->_idx); 4597 NodeCloneInfo cio(val); 4598 assert(val != 0, "old node should be in the map"); 4599 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4600 insert(nnn->_idx, cin.get()); 4601 #ifndef PRODUCT 4602 if (is_debug()) { 4603 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4604 } 4605 #endif 4606 } 4607 4608 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4609 NodeCloneInfo cio(value(old->_idx)); 4610 if (cio.get() == 0) { 4611 cio.set(old->_idx, 0); 4612 insert(old->_idx, cio.get()); 4613 #ifndef PRODUCT 4614 if (is_debug()) { 4615 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4616 } 4617 #endif 4618 } 4619 clone(old, nnn, gen); 4620 } 4621 4622 int CloneMap::max_gen() const { 4623 int g = 0; 4624 DictI di(_dict); 4625 for(; di.test(); ++di) { 4626 int t = gen(di._key); 4627 if (g < t) { 4628 g = t; 4629 #ifndef PRODUCT 4630 if (is_debug()) { 4631 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4632 } 4633 #endif 4634 } 4635 } 4636 return g; 4637 } 4638 4639 void CloneMap::dump(node_idx_t key) const { 4640 uint64_t val = value(key); 4641 if (val != 0) { 4642 NodeCloneInfo ni(val); 4643 ni.dump(); 4644 } 4645 } 4646 4647 // Helper function for enforcing certain bytecodes to reexecute if 4648 // deoptimization happens 4649 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 4650 ciMethod* cur_method = jvms->method(); 4651 int cur_bci = jvms->bci(); 4652 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 4653 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 4654 return Interpreter::bytecode_should_reexecute(code) || 4655 (is_anewarray && (code == Bytecodes::_multianewarray)); 4656 // Reexecute _multianewarray bytecode which was replaced with 4657 // sequence of [a]newarray. See Parse::do_multianewarray(). 4658 // 4659 // Note: interpreter should not have it set since this optimization 4660 // is limited by dimensions and guarded by flag so in some cases 4661 // multianewarray() runtime calls will be generated and 4662 // the bytecode should not be reexecutes (stack will not be reset). 4663 } else 4664 return false; 4665 } 4666 4667 void Compile::add_safepoint_edges(SafePointNode* call, JVMState* youngest_jvms, bool can_prune_locals, uint stack_slots_not_pruned) { 4668 // do not scribble on the input jvms 4669 JVMState* out_jvms = youngest_jvms->clone_deep(C); 4670 call->set_jvms(out_jvms); // Start jvms list for call node 4671 4672 // For a known set of bytecodes, the interpreter should reexecute them if 4673 // deoptimization happens. We set the reexecute state for them here 4674 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 4675 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 4676 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 4677 } 4678 4679 // Presize the call: 4680 DEBUG_ONLY(uint non_debug_edges = call->req()); 4681 call->add_req_batch(top(), youngest_jvms->debug_depth()); 4682 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 4683 4684 // Set up edges so that the call looks like this: 4685 // Call [state:] ctl io mem fptr retadr 4686 // [parms:] parm0 ... parmN 4687 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 4688 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 4689 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 4690 // Note that caller debug info precedes callee debug info. 4691 4692 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 4693 uint debug_ptr = call->req(); 4694 4695 // Loop over the map input edges associated with jvms, add them 4696 // to the call node, & reset all offsets to match call node array. 4697 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 4698 uint debug_end = debug_ptr; 4699 uint debug_start = debug_ptr - in_jvms->debug_size(); 4700 debug_ptr = debug_start; // back up the ptr 4701 4702 uint p = debug_start; // walks forward in [debug_start, debug_end) 4703 uint j, k, l; 4704 SafePointNode* in_map = in_jvms->map(); 4705 out_jvms->set_map(call); 4706 4707 if (can_prune_locals) { 4708 assert(in_jvms->method() == out_jvms->method(), "sanity"); 4709 // If the current throw can reach an exception handler in this JVMS, 4710 // then we must keep everything live that can reach that handler. 4711 // As a quick and dirty approximation, we look for any handlers at all. 4712 if (in_jvms->method()->has_exception_handlers()) { 4713 can_prune_locals = false; 4714 } 4715 } 4716 4717 // Add the Locals 4718 k = in_jvms->locoff(); 4719 l = in_jvms->loc_size(); 4720 out_jvms->set_locoff(p); 4721 if (!can_prune_locals) { 4722 for (j = 0; j < l; j++) 4723 call->set_req(p++, in_map->in(k+j)); 4724 } else { 4725 p += l; // already set to top above by add_req_batch 4726 } 4727 4728 // Add the Expression Stack 4729 k = in_jvms->stkoff(); 4730 l = in_jvms->sp(); 4731 out_jvms->set_stkoff(p); 4732 if (!can_prune_locals) { 4733 for (j = 0; j < l; j++) 4734 call->set_req(p++, in_map->in(k+j)); 4735 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 4736 // Divide stack into {S0,...,S1}, where S0 is set to top. 4737 uint s1 = stack_slots_not_pruned; 4738 stack_slots_not_pruned = 0; // for next iteration 4739 if (s1 > l) s1 = l; 4740 uint s0 = l - s1; 4741 p += s0; // skip the tops preinstalled by add_req_batch 4742 for (j = s0; j < l; j++) 4743 call->set_req(p++, in_map->in(k+j)); 4744 } else { 4745 p += l; // already set to top above by add_req_batch 4746 } 4747 4748 // Add the Monitors 4749 k = in_jvms->monoff(); 4750 l = in_jvms->mon_size(); 4751 out_jvms->set_monoff(p); 4752 for (j = 0; j < l; j++) 4753 call->set_req(p++, in_map->in(k+j)); 4754 4755 // Copy any scalar object fields. 4756 k = in_jvms->scloff(); 4757 l = in_jvms->scl_size(); 4758 out_jvms->set_scloff(p); 4759 for (j = 0; j < l; j++) 4760 call->set_req(p++, in_map->in(k+j)); 4761 4762 // Finish the new jvms. 4763 out_jvms->set_endoff(p); 4764 4765 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 4766 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 4767 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 4768 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 4769 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 4770 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 4771 4772 // Update the two tail pointers in parallel. 4773 out_jvms = out_jvms->caller(); 4774 in_jvms = in_jvms->caller(); 4775 } 4776 4777 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 4778 4779 // Test the correctness of JVMState::debug_xxx accessors: 4780 assert(call->jvms()->debug_start() == non_debug_edges, ""); 4781 assert(call->jvms()->debug_end() == call->req(), ""); 4782 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 4783 }