1 /* 2 * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/block.hpp" 38 #include "opto/c2compiler.hpp" 39 #include "opto/callGenerator.hpp" 40 #include "opto/callnode.hpp" 41 #include "opto/cfgnode.hpp" 42 #include "opto/chaitin.hpp" 43 #include "opto/compile.hpp" 44 #include "opto/connode.hpp" 45 #include "opto/divnode.hpp" 46 #include "opto/escape.hpp" 47 #include "opto/idealGraphPrinter.hpp" 48 #include "opto/loopnode.hpp" 49 #include "opto/machnode.hpp" 50 #include "opto/macro.hpp" 51 #include "opto/matcher.hpp" 52 #include "opto/mathexactnode.hpp" 53 #include "opto/memnode.hpp" 54 #include "opto/mulnode.hpp" 55 #include "opto/narrowptrnode.hpp" 56 #include "opto/node.hpp" 57 #include "opto/opcodes.hpp" 58 #include "opto/output.hpp" 59 #include "opto/parse.hpp" 60 #include "opto/phaseX.hpp" 61 #include "opto/rootnode.hpp" 62 #include "opto/runtime.hpp" 63 #include "opto/stringopts.hpp" 64 #include "opto/type.hpp" 65 #include "opto/vectornode.hpp" 66 #include "runtime/arguments.hpp" 67 #include "runtime/signature.hpp" 68 #include "runtime/stubRoutines.hpp" 69 #include "runtime/timer.hpp" 70 #include "trace/tracing.hpp" 71 #include "utilities/copy.hpp" 72 73 74 // -------------------- Compile::mach_constant_base_node ----------------------- 75 // Constant table base node singleton. 76 MachConstantBaseNode* Compile::mach_constant_base_node() { 77 if (_mach_constant_base_node == NULL) { 78 _mach_constant_base_node = new MachConstantBaseNode(); 79 _mach_constant_base_node->add_req(C->root()); 80 } 81 return _mach_constant_base_node; 82 } 83 84 85 /// Support for intrinsics. 86 87 // Return the index at which m must be inserted (or already exists). 88 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 89 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 90 #ifdef ASSERT 91 for (int i = 1; i < _intrinsics->length(); i++) { 92 CallGenerator* cg1 = _intrinsics->at(i-1); 93 CallGenerator* cg2 = _intrinsics->at(i); 94 assert(cg1->method() != cg2->method() 95 ? cg1->method() < cg2->method() 96 : cg1->is_virtual() < cg2->is_virtual(), 97 "compiler intrinsics list must stay sorted"); 98 } 99 #endif 100 // Binary search sorted list, in decreasing intervals [lo, hi]. 101 int lo = 0, hi = _intrinsics->length()-1; 102 while (lo <= hi) { 103 int mid = (uint)(hi + lo) / 2; 104 ciMethod* mid_m = _intrinsics->at(mid)->method(); 105 if (m < mid_m) { 106 hi = mid-1; 107 } else if (m > mid_m) { 108 lo = mid+1; 109 } else { 110 // look at minor sort key 111 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 112 if (is_virtual < mid_virt) { 113 hi = mid-1; 114 } else if (is_virtual > mid_virt) { 115 lo = mid+1; 116 } else { 117 return mid; // exact match 118 } 119 } 120 } 121 return lo; // inexact match 122 } 123 124 void Compile::register_intrinsic(CallGenerator* cg) { 125 if (_intrinsics == NULL) { 126 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 127 } 128 // This code is stolen from ciObjectFactory::insert. 129 // Really, GrowableArray should have methods for 130 // insert_at, remove_at, and binary_search. 131 int len = _intrinsics->length(); 132 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 133 if (index == len) { 134 _intrinsics->append(cg); 135 } else { 136 #ifdef ASSERT 137 CallGenerator* oldcg = _intrinsics->at(index); 138 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 139 #endif 140 _intrinsics->append(_intrinsics->at(len-1)); 141 int pos; 142 for (pos = len-2; pos >= index; pos--) { 143 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 144 } 145 _intrinsics->at_put(index, cg); 146 } 147 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 148 } 149 150 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 151 assert(m->is_loaded(), "don't try this on unloaded methods"); 152 if (_intrinsics != NULL) { 153 int index = intrinsic_insertion_index(m, is_virtual); 154 if (index < _intrinsics->length() 155 && _intrinsics->at(index)->method() == m 156 && _intrinsics->at(index)->is_virtual() == is_virtual) { 157 return _intrinsics->at(index); 158 } 159 } 160 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 161 if (m->intrinsic_id() != vmIntrinsics::_none && 162 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 163 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 164 if (cg != NULL) { 165 // Save it for next time: 166 register_intrinsic(cg); 167 return cg; 168 } else { 169 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 170 } 171 } 172 return NULL; 173 } 174 175 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 176 // in library_call.cpp. 177 178 179 #ifndef PRODUCT 180 // statistics gathering... 181 182 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 183 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 184 185 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 186 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 187 int oflags = _intrinsic_hist_flags[id]; 188 assert(flags != 0, "what happened?"); 189 if (is_virtual) { 190 flags |= _intrinsic_virtual; 191 } 192 bool changed = (flags != oflags); 193 if ((flags & _intrinsic_worked) != 0) { 194 juint count = (_intrinsic_hist_count[id] += 1); 195 if (count == 1) { 196 changed = true; // first time 197 } 198 // increment the overall count also: 199 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 200 } 201 if (changed) { 202 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 203 // Something changed about the intrinsic's virtuality. 204 if ((flags & _intrinsic_virtual) != 0) { 205 // This is the first use of this intrinsic as a virtual call. 206 if (oflags != 0) { 207 // We already saw it as a non-virtual, so note both cases. 208 flags |= _intrinsic_both; 209 } 210 } else if ((oflags & _intrinsic_both) == 0) { 211 // This is the first use of this intrinsic as a non-virtual 212 flags |= _intrinsic_both; 213 } 214 } 215 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 216 } 217 // update the overall flags also: 218 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 219 return changed; 220 } 221 222 static char* format_flags(int flags, char* buf) { 223 buf[0] = 0; 224 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 225 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 226 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 227 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 228 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 229 if (buf[0] == 0) strcat(buf, ","); 230 assert(buf[0] == ',', "must be"); 231 return &buf[1]; 232 } 233 234 void Compile::print_intrinsic_statistics() { 235 char flagsbuf[100]; 236 ttyLocker ttyl; 237 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 238 tty->print_cr("Compiler intrinsic usage:"); 239 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 240 if (total == 0) total = 1; // avoid div0 in case of no successes 241 #define PRINT_STAT_LINE(name, c, f) \ 242 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 243 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 244 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 245 int flags = _intrinsic_hist_flags[id]; 246 juint count = _intrinsic_hist_count[id]; 247 if ((flags | count) != 0) { 248 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 249 } 250 } 251 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 252 if (xtty != NULL) xtty->tail("statistics"); 253 } 254 255 void Compile::print_statistics() { 256 { ttyLocker ttyl; 257 if (xtty != NULL) xtty->head("statistics type='opto'"); 258 Parse::print_statistics(); 259 PhaseCCP::print_statistics(); 260 PhaseRegAlloc::print_statistics(); 261 Scheduling::print_statistics(); 262 PhasePeephole::print_statistics(); 263 PhaseIdealLoop::print_statistics(); 264 if (xtty != NULL) xtty->tail("statistics"); 265 } 266 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 267 // put this under its own <statistics> element. 268 print_intrinsic_statistics(); 269 } 270 } 271 #endif //PRODUCT 272 273 // Support for bundling info 274 Bundle* Compile::node_bundling(const Node *n) { 275 assert(valid_bundle_info(n), "oob"); 276 return &_node_bundling_base[n->_idx]; 277 } 278 279 bool Compile::valid_bundle_info(const Node *n) { 280 return (_node_bundling_limit > n->_idx); 281 } 282 283 284 void Compile::gvn_replace_by(Node* n, Node* nn) { 285 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 286 Node* use = n->last_out(i); 287 bool is_in_table = initial_gvn()->hash_delete(use); 288 uint uses_found = 0; 289 for (uint j = 0; j < use->len(); j++) { 290 if (use->in(j) == n) { 291 if (j < use->req()) 292 use->set_req(j, nn); 293 else 294 use->set_prec(j, nn); 295 uses_found++; 296 } 297 } 298 if (is_in_table) { 299 // reinsert into table 300 initial_gvn()->hash_find_insert(use); 301 } 302 record_for_igvn(use); 303 i -= uses_found; // we deleted 1 or more copies of this edge 304 } 305 } 306 307 308 static inline bool not_a_node(const Node* n) { 309 if (n == NULL) return true; 310 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 311 if (*(address*)n == badAddress) return true; // kill by Node::destruct 312 return false; 313 } 314 315 // Identify all nodes that are reachable from below, useful. 316 // Use breadth-first pass that records state in a Unique_Node_List, 317 // recursive traversal is slower. 318 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 319 int estimated_worklist_size = unique(); 320 useful.map( estimated_worklist_size, NULL ); // preallocate space 321 322 // Initialize worklist 323 if (root() != NULL) { useful.push(root()); } 324 // If 'top' is cached, declare it useful to preserve cached node 325 if( cached_top_node() ) { useful.push(cached_top_node()); } 326 327 // Push all useful nodes onto the list, breadthfirst 328 for( uint next = 0; next < useful.size(); ++next ) { 329 assert( next < unique(), "Unique useful nodes < total nodes"); 330 Node *n = useful.at(next); 331 uint max = n->len(); 332 for( uint i = 0; i < max; ++i ) { 333 Node *m = n->in(i); 334 if (not_a_node(m)) continue; 335 useful.push(m); 336 } 337 } 338 } 339 340 // Update dead_node_list with any missing dead nodes using useful 341 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 342 void Compile::update_dead_node_list(Unique_Node_List &useful) { 343 uint max_idx = unique(); 344 VectorSet& useful_node_set = useful.member_set(); 345 346 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 347 // If node with index node_idx is not in useful set, 348 // mark it as dead in dead node list. 349 if (! useful_node_set.test(node_idx) ) { 350 record_dead_node(node_idx); 351 } 352 } 353 } 354 355 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 356 int shift = 0; 357 for (int i = 0; i < inlines->length(); i++) { 358 CallGenerator* cg = inlines->at(i); 359 CallNode* call = cg->call_node(); 360 if (shift > 0) { 361 inlines->at_put(i-shift, cg); 362 } 363 if (!useful.member(call)) { 364 shift++; 365 } 366 } 367 inlines->trunc_to(inlines->length()-shift); 368 } 369 370 // Disconnect all useless nodes by disconnecting those at the boundary. 371 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 372 uint next = 0; 373 while (next < useful.size()) { 374 Node *n = useful.at(next++); 375 if (n->is_SafePoint()) { 376 // We're done with a parsing phase. Replaced nodes are not valid 377 // beyond that point. 378 n->as_SafePoint()->delete_replaced_nodes(); 379 } 380 // Use raw traversal of out edges since this code removes out edges 381 int max = n->outcnt(); 382 for (int j = 0; j < max; ++j) { 383 Node* child = n->raw_out(j); 384 if (! useful.member(child)) { 385 assert(!child->is_top() || child != top(), 386 "If top is cached in Compile object it is in useful list"); 387 // Only need to remove this out-edge to the useless node 388 n->raw_del_out(j); 389 --j; 390 --max; 391 } 392 } 393 if (n->outcnt() == 1 && n->has_special_unique_user()) { 394 record_for_igvn(n->unique_out()); 395 } 396 } 397 // Remove useless macro and predicate opaq nodes 398 for (int i = C->macro_count()-1; i >= 0; i--) { 399 Node* n = C->macro_node(i); 400 if (!useful.member(n)) { 401 remove_macro_node(n); 402 } 403 } 404 // Remove useless expensive node 405 for (int i = C->expensive_count()-1; i >= 0; i--) { 406 Node* n = C->expensive_node(i); 407 if (!useful.member(n)) { 408 remove_expensive_node(n); 409 } 410 } 411 // clean up the late inline lists 412 remove_useless_late_inlines(&_string_late_inlines, useful); 413 remove_useless_late_inlines(&_boxing_late_inlines, useful); 414 remove_useless_late_inlines(&_late_inlines, useful); 415 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 416 } 417 418 //------------------------------frame_size_in_words----------------------------- 419 // frame_slots in units of words 420 int Compile::frame_size_in_words() const { 421 // shift is 0 in LP32 and 1 in LP64 422 const int shift = (LogBytesPerWord - LogBytesPerInt); 423 int words = _frame_slots >> shift; 424 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 425 return words; 426 } 427 428 // To bang the stack of this compiled method we use the stack size 429 // that the interpreter would need in case of a deoptimization. This 430 // removes the need to bang the stack in the deoptimization blob which 431 // in turn simplifies stack overflow handling. 432 int Compile::bang_size_in_bytes() const { 433 return MAX2(_interpreter_frame_size, frame_size_in_bytes()); 434 } 435 436 // ============================================================================ 437 //------------------------------CompileWrapper--------------------------------- 438 class CompileWrapper : public StackObj { 439 Compile *const _compile; 440 public: 441 CompileWrapper(Compile* compile); 442 443 ~CompileWrapper(); 444 }; 445 446 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 447 // the Compile* pointer is stored in the current ciEnv: 448 ciEnv* env = compile->env(); 449 assert(env == ciEnv::current(), "must already be a ciEnv active"); 450 assert(env->compiler_data() == NULL, "compile already active?"); 451 env->set_compiler_data(compile); 452 assert(compile == Compile::current(), "sanity"); 453 454 compile->set_type_dict(NULL); 455 compile->set_type_hwm(NULL); 456 compile->set_type_last_size(0); 457 compile->set_last_tf(NULL, NULL); 458 compile->set_indexSet_arena(NULL); 459 compile->set_indexSet_free_block_list(NULL); 460 compile->init_type_arena(); 461 Type::Initialize(compile); 462 _compile->set_scratch_buffer_blob(NULL); 463 _compile->begin_method(); 464 } 465 CompileWrapper::~CompileWrapper() { 466 _compile->end_method(); 467 if (_compile->scratch_buffer_blob() != NULL) 468 BufferBlob::free(_compile->scratch_buffer_blob()); 469 _compile->env()->set_compiler_data(NULL); 470 } 471 472 473 //----------------------------print_compile_messages--------------------------- 474 void Compile::print_compile_messages() { 475 #ifndef PRODUCT 476 // Check if recompiling 477 if (_subsume_loads == false && PrintOpto) { 478 // Recompiling without allowing machine instructions to subsume loads 479 tty->print_cr("*********************************************************"); 480 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 481 tty->print_cr("*********************************************************"); 482 } 483 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 484 // Recompiling without escape analysis 485 tty->print_cr("*********************************************************"); 486 tty->print_cr("** Bailout: Recompile without escape analysis **"); 487 tty->print_cr("*********************************************************"); 488 } 489 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 490 // Recompiling without boxing elimination 491 tty->print_cr("*********************************************************"); 492 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 493 tty->print_cr("*********************************************************"); 494 } 495 if (env()->break_at_compile()) { 496 // Open the debugger when compiling this method. 497 tty->print("### Breaking when compiling: "); 498 method()->print_short_name(); 499 tty->cr(); 500 BREAKPOINT; 501 } 502 503 if( PrintOpto ) { 504 if (is_osr_compilation()) { 505 tty->print("[OSR]%3d", _compile_id); 506 } else { 507 tty->print("%3d", _compile_id); 508 } 509 } 510 #endif 511 } 512 513 514 //-----------------------init_scratch_buffer_blob------------------------------ 515 // Construct a temporary BufferBlob and cache it for this compile. 516 void Compile::init_scratch_buffer_blob(int const_size) { 517 // If there is already a scratch buffer blob allocated and the 518 // constant section is big enough, use it. Otherwise free the 519 // current and allocate a new one. 520 BufferBlob* blob = scratch_buffer_blob(); 521 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 522 // Use the current blob. 523 } else { 524 if (blob != NULL) { 525 BufferBlob::free(blob); 526 } 527 528 ResourceMark rm; 529 _scratch_const_size = const_size; 530 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 531 blob = BufferBlob::create("Compile::scratch_buffer", size); 532 // Record the buffer blob for next time. 533 set_scratch_buffer_blob(blob); 534 // Have we run out of code space? 535 if (scratch_buffer_blob() == NULL) { 536 // Let CompilerBroker disable further compilations. 537 record_failure("Not enough space for scratch buffer in CodeCache"); 538 CompileBroker::handle_full_code_cache(CodeBlobType::NonMethod); 539 return; 540 } 541 } 542 543 // Initialize the relocation buffers 544 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 545 set_scratch_locs_memory(locs_buf); 546 } 547 548 549 //-----------------------scratch_emit_size------------------------------------- 550 // Helper function that computes size by emitting code 551 uint Compile::scratch_emit_size(const Node* n) { 552 // Start scratch_emit_size section. 553 set_in_scratch_emit_size(true); 554 555 // Emit into a trash buffer and count bytes emitted. 556 // This is a pretty expensive way to compute a size, 557 // but it works well enough if seldom used. 558 // All common fixed-size instructions are given a size 559 // method by the AD file. 560 // Note that the scratch buffer blob and locs memory are 561 // allocated at the beginning of the compile task, and 562 // may be shared by several calls to scratch_emit_size. 563 // The allocation of the scratch buffer blob is particularly 564 // expensive, since it has to grab the code cache lock. 565 BufferBlob* blob = this->scratch_buffer_blob(); 566 assert(blob != NULL, "Initialize BufferBlob at start"); 567 assert(blob->size() > MAX_inst_size, "sanity"); 568 relocInfo* locs_buf = scratch_locs_memory(); 569 address blob_begin = blob->content_begin(); 570 address blob_end = (address)locs_buf; 571 assert(blob->content_contains(blob_end), "sanity"); 572 CodeBuffer buf(blob_begin, blob_end - blob_begin); 573 buf.initialize_consts_size(_scratch_const_size); 574 buf.initialize_stubs_size(MAX_stubs_size); 575 assert(locs_buf != NULL, "sanity"); 576 int lsize = MAX_locs_size / 3; 577 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 578 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 579 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 580 581 // Do the emission. 582 583 Label fakeL; // Fake label for branch instructions. 584 Label* saveL = NULL; 585 uint save_bnum = 0; 586 bool is_branch = n->is_MachBranch(); 587 if (is_branch) { 588 MacroAssembler masm(&buf); 589 masm.bind(fakeL); 590 n->as_MachBranch()->save_label(&saveL, &save_bnum); 591 n->as_MachBranch()->label_set(&fakeL, 0); 592 } 593 n->emit(buf, this->regalloc()); 594 if (is_branch) // Restore label. 595 n->as_MachBranch()->label_set(saveL, save_bnum); 596 597 // End scratch_emit_size section. 598 set_in_scratch_emit_size(false); 599 600 return buf.insts_size(); 601 } 602 603 604 // ============================================================================ 605 //------------------------------Compile standard------------------------------- 606 debug_only( int Compile::_debug_idx = 100000; ) 607 608 // Compile a method. entry_bci is -1 for normal compilations and indicates 609 // the continuation bci for on stack replacement. 610 611 612 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 613 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 614 : Phase(Compiler), 615 _env(ci_env), 616 _log(ci_env->log()), 617 _compile_id(ci_env->compile_id()), 618 _save_argument_registers(false), 619 _stub_name(NULL), 620 _stub_function(NULL), 621 _stub_entry_point(NULL), 622 _method(target), 623 _entry_bci(osr_bci), 624 _initial_gvn(NULL), 625 _for_igvn(NULL), 626 _warm_calls(NULL), 627 _subsume_loads(subsume_loads), 628 _do_escape_analysis(do_escape_analysis), 629 _eliminate_boxing(eliminate_boxing), 630 _failure_reason(NULL), 631 _code_buffer("Compile::Fill_buffer"), 632 _orig_pc_slot(0), 633 _orig_pc_slot_offset_in_bytes(0), 634 _has_method_handle_invokes(false), 635 _mach_constant_base_node(NULL), 636 _node_bundling_limit(0), 637 _node_bundling_base(NULL), 638 _java_calls(0), 639 _inner_loops(0), 640 _scratch_const_size(-1), 641 _in_scratch_emit_size(false), 642 _dead_node_list(comp_arena()), 643 _dead_node_count(0), 644 #ifndef PRODUCT 645 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 646 _in_dump_cnt(0), 647 _printer(IdealGraphPrinter::printer()), 648 #endif 649 _congraph(NULL), 650 _comp_arena(mtCompiler), 651 _node_arena(mtCompiler), 652 _old_arena(mtCompiler), 653 _Compile_types(mtCompiler), 654 _replay_inline_data(NULL), 655 _late_inlines(comp_arena(), 2, 0, NULL), 656 _string_late_inlines(comp_arena(), 2, 0, NULL), 657 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 658 _late_inlines_pos(0), 659 _number_of_mh_late_inlines(0), 660 _inlining_progress(false), 661 _inlining_incrementally(false), 662 _print_inlining_list(NULL), 663 _print_inlining_stream(NULL), 664 _print_inlining_idx(0), 665 _print_inlining_output(NULL), 666 _interpreter_frame_size(0) { 667 C = this; 668 669 CompileWrapper cw(this); 670 #ifndef PRODUCT 671 if (TimeCompiler2) { 672 tty->print(" "); 673 target->holder()->name()->print(); 674 tty->print("."); 675 target->print_short_name(); 676 tty->print(" "); 677 } 678 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 679 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 680 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 681 if (!print_opto_assembly) { 682 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 683 if (print_assembly && !Disassembler::can_decode()) { 684 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 685 print_opto_assembly = true; 686 } 687 } 688 set_print_assembly(print_opto_assembly); 689 set_parsed_irreducible_loop(false); 690 691 if (method()->has_option("ReplayInline")) { 692 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 693 } 694 #endif 695 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 696 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 697 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 698 699 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 700 // Make sure the method being compiled gets its own MDO, 701 // so we can at least track the decompile_count(). 702 // Need MDO to record RTM code generation state. 703 method()->ensure_method_data(); 704 } 705 706 Init(::AliasLevel); 707 708 709 print_compile_messages(); 710 711 _ilt = InlineTree::build_inline_tree_root(); 712 713 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 714 assert(num_alias_types() >= AliasIdxRaw, ""); 715 716 #define MINIMUM_NODE_HASH 1023 717 // Node list that Iterative GVN will start with 718 Unique_Node_List for_igvn(comp_arena()); 719 set_for_igvn(&for_igvn); 720 721 // GVN that will be run immediately on new nodes 722 uint estimated_size = method()->code_size()*4+64; 723 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 724 PhaseGVN gvn(node_arena(), estimated_size); 725 set_initial_gvn(&gvn); 726 727 print_inlining_init(); 728 { // Scope for timing the parser 729 TracePhase t3("parse", &_t_parser, true); 730 731 // Put top into the hash table ASAP. 732 initial_gvn()->transform_no_reclaim(top()); 733 734 // Set up tf(), start(), and find a CallGenerator. 735 CallGenerator* cg = NULL; 736 if (is_osr_compilation()) { 737 const TypeTuple *domain = StartOSRNode::osr_domain(); 738 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 739 init_tf(TypeFunc::make(domain, range)); 740 StartNode* s = new StartOSRNode(root(), domain); 741 initial_gvn()->set_type_bottom(s); 742 init_start(s); 743 cg = CallGenerator::for_osr(method(), entry_bci()); 744 } else { 745 // Normal case. 746 init_tf(TypeFunc::make(method())); 747 StartNode* s = new StartNode(root(), tf()->domain()); 748 initial_gvn()->set_type_bottom(s); 749 init_start(s); 750 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 751 // With java.lang.ref.reference.get() we must go through the 752 // intrinsic when G1 is enabled - even when get() is the root 753 // method of the compile - so that, if necessary, the value in 754 // the referent field of the reference object gets recorded by 755 // the pre-barrier code. 756 // Specifically, if G1 is enabled, the value in the referent 757 // field is recorded by the G1 SATB pre barrier. This will 758 // result in the referent being marked live and the reference 759 // object removed from the list of discovered references during 760 // reference processing. 761 cg = find_intrinsic(method(), false); 762 } 763 if (cg == NULL) { 764 float past_uses = method()->interpreter_invocation_count(); 765 float expected_uses = past_uses; 766 cg = CallGenerator::for_inline(method(), expected_uses); 767 } 768 } 769 if (failing()) return; 770 if (cg == NULL) { 771 record_method_not_compilable_all_tiers("cannot parse method"); 772 return; 773 } 774 JVMState* jvms = build_start_state(start(), tf()); 775 if ((jvms = cg->generate(jvms)) == NULL) { 776 record_method_not_compilable("method parse failed"); 777 return; 778 } 779 GraphKit kit(jvms); 780 781 if (!kit.stopped()) { 782 // Accept return values, and transfer control we know not where. 783 // This is done by a special, unique ReturnNode bound to root. 784 return_values(kit.jvms()); 785 } 786 787 if (kit.has_exceptions()) { 788 // Any exceptions that escape from this call must be rethrown 789 // to whatever caller is dynamically above us on the stack. 790 // This is done by a special, unique RethrowNode bound to root. 791 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 792 } 793 794 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 795 796 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 797 inline_string_calls(true); 798 } 799 800 if (failing()) return; 801 802 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 803 804 // Remove clutter produced by parsing. 805 if (!failing()) { 806 ResourceMark rm; 807 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 808 } 809 } 810 811 // Note: Large methods are capped off in do_one_bytecode(). 812 if (failing()) return; 813 814 // After parsing, node notes are no longer automagic. 815 // They must be propagated by register_new_node_with_optimizer(), 816 // clone(), or the like. 817 set_default_node_notes(NULL); 818 819 for (;;) { 820 int successes = Inline_Warm(); 821 if (failing()) return; 822 if (successes == 0) break; 823 } 824 825 // Drain the list. 826 Finish_Warm(); 827 #ifndef PRODUCT 828 if (_printer) { 829 _printer->print_inlining(this); 830 } 831 #endif 832 833 if (failing()) return; 834 NOT_PRODUCT( verify_graph_edges(); ) 835 836 // Now optimize 837 Optimize(); 838 if (failing()) return; 839 NOT_PRODUCT( verify_graph_edges(); ) 840 841 #ifndef PRODUCT 842 if (PrintIdeal) { 843 ttyLocker ttyl; // keep the following output all in one block 844 // This output goes directly to the tty, not the compiler log. 845 // To enable tools to match it up with the compilation activity, 846 // be sure to tag this tty output with the compile ID. 847 if (xtty != NULL) { 848 xtty->head("ideal compile_id='%d'%s", compile_id(), 849 is_osr_compilation() ? " compile_kind='osr'" : 850 ""); 851 } 852 root()->dump(9999); 853 if (xtty != NULL) { 854 xtty->tail("ideal"); 855 } 856 } 857 #endif 858 859 NOT_PRODUCT( verify_barriers(); ) 860 861 // Dump compilation data to replay it. 862 if (method()->has_option("DumpReplay")) { 863 env()->dump_replay_data(_compile_id); 864 } 865 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 866 env()->dump_inline_data(_compile_id); 867 } 868 869 // Now that we know the size of all the monitors we can add a fixed slot 870 // for the original deopt pc. 871 872 _orig_pc_slot = fixed_slots(); 873 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 874 set_fixed_slots(next_slot); 875 876 // Compute when to use implicit null checks. Used by matching trap based 877 // nodes and NullCheck optimization. 878 set_allowed_deopt_reasons(); 879 880 // Now generate code 881 Code_Gen(); 882 if (failing()) return; 883 884 // Check if we want to skip execution of all compiled code. 885 { 886 #ifndef PRODUCT 887 if (OptoNoExecute) { 888 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 889 return; 890 } 891 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 892 #endif 893 894 if (is_osr_compilation()) { 895 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 896 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 897 } else { 898 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 899 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 900 } 901 902 env()->register_method(_method, _entry_bci, 903 &_code_offsets, 904 _orig_pc_slot_offset_in_bytes, 905 code_buffer(), 906 frame_size_in_words(), _oop_map_set, 907 &_handler_table, &_inc_table, 908 compiler, 909 env()->comp_level(), 910 has_unsafe_access(), 911 SharedRuntime::is_wide_vector(max_vector_size()), 912 rtm_state() 913 ); 914 915 if (log() != NULL) // Print code cache state into compiler log 916 log()->code_cache_state(); 917 } 918 } 919 920 //------------------------------Compile---------------------------------------- 921 // Compile a runtime stub 922 Compile::Compile( ciEnv* ci_env, 923 TypeFunc_generator generator, 924 address stub_function, 925 const char *stub_name, 926 int is_fancy_jump, 927 bool pass_tls, 928 bool save_arg_registers, 929 bool return_pc ) 930 : Phase(Compiler), 931 _env(ci_env), 932 _log(ci_env->log()), 933 _compile_id(0), 934 _save_argument_registers(save_arg_registers), 935 _method(NULL), 936 _stub_name(stub_name), 937 _stub_function(stub_function), 938 _stub_entry_point(NULL), 939 _entry_bci(InvocationEntryBci), 940 _initial_gvn(NULL), 941 _for_igvn(NULL), 942 _warm_calls(NULL), 943 _orig_pc_slot(0), 944 _orig_pc_slot_offset_in_bytes(0), 945 _subsume_loads(true), 946 _do_escape_analysis(false), 947 _eliminate_boxing(false), 948 _failure_reason(NULL), 949 _code_buffer("Compile::Fill_buffer"), 950 _has_method_handle_invokes(false), 951 _mach_constant_base_node(NULL), 952 _node_bundling_limit(0), 953 _node_bundling_base(NULL), 954 _java_calls(0), 955 _inner_loops(0), 956 #ifndef PRODUCT 957 _trace_opto_output(TraceOptoOutput), 958 _in_dump_cnt(0), 959 _printer(NULL), 960 #endif 961 _comp_arena(mtCompiler), 962 _node_arena(mtCompiler), 963 _old_arena(mtCompiler), 964 _Compile_types(mtCompiler), 965 _dead_node_list(comp_arena()), 966 _dead_node_count(0), 967 _congraph(NULL), 968 _replay_inline_data(NULL), 969 _number_of_mh_late_inlines(0), 970 _inlining_progress(false), 971 _inlining_incrementally(false), 972 _print_inlining_list(NULL), 973 _print_inlining_stream(NULL), 974 _print_inlining_idx(0), 975 _print_inlining_output(NULL), 976 _allowed_reasons(0), 977 _interpreter_frame_size(0) { 978 C = this; 979 980 #ifndef PRODUCT 981 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 982 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 983 set_print_assembly(PrintFrameConverterAssembly); 984 set_parsed_irreducible_loop(false); 985 #endif 986 set_has_irreducible_loop(false); // no loops 987 988 CompileWrapper cw(this); 989 Init(/*AliasLevel=*/ 0); 990 init_tf((*generator)()); 991 992 { 993 // The following is a dummy for the sake of GraphKit::gen_stub 994 Unique_Node_List for_igvn(comp_arena()); 995 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 996 PhaseGVN gvn(Thread::current()->resource_area(),255); 997 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 998 gvn.transform_no_reclaim(top()); 999 1000 GraphKit kit; 1001 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1002 } 1003 1004 NOT_PRODUCT( verify_graph_edges(); ) 1005 Code_Gen(); 1006 if (failing()) return; 1007 1008 1009 // Entry point will be accessed using compile->stub_entry_point(); 1010 if (code_buffer() == NULL) { 1011 Matcher::soft_match_failure(); 1012 } else { 1013 if (PrintAssembly && (WizardMode || Verbose)) 1014 tty->print_cr("### Stub::%s", stub_name); 1015 1016 if (!failing()) { 1017 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1018 1019 // Make the NMethod 1020 // For now we mark the frame as never safe for profile stackwalking 1021 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1022 code_buffer(), 1023 CodeOffsets::frame_never_safe, 1024 // _code_offsets.value(CodeOffsets::Frame_Complete), 1025 frame_size_in_words(), 1026 _oop_map_set, 1027 save_arg_registers); 1028 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1029 1030 _stub_entry_point = rs->entry_point(); 1031 } 1032 } 1033 } 1034 1035 //------------------------------Init------------------------------------------- 1036 // Prepare for a single compilation 1037 void Compile::Init(int aliaslevel) { 1038 _unique = 0; 1039 _regalloc = NULL; 1040 1041 _tf = NULL; // filled in later 1042 _top = NULL; // cached later 1043 _matcher = NULL; // filled in later 1044 _cfg = NULL; // filled in later 1045 1046 set_24_bit_selection_and_mode(Use24BitFP, false); 1047 1048 _node_note_array = NULL; 1049 _default_node_notes = NULL; 1050 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1051 1052 _immutable_memory = NULL; // filled in at first inquiry 1053 1054 // Globally visible Nodes 1055 // First set TOP to NULL to give safe behavior during creation of RootNode 1056 set_cached_top_node(NULL); 1057 set_root(new RootNode()); 1058 // Now that you have a Root to point to, create the real TOP 1059 set_cached_top_node( new ConNode(Type::TOP) ); 1060 set_recent_alloc(NULL, NULL); 1061 1062 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1063 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1064 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1065 env()->set_dependencies(new Dependencies(env())); 1066 1067 _fixed_slots = 0; 1068 set_has_split_ifs(false); 1069 set_has_loops(has_method() && method()->has_loops()); // first approximation 1070 set_has_stringbuilder(false); 1071 set_has_boxed_value(false); 1072 _trap_can_recompile = false; // no traps emitted yet 1073 _major_progress = true; // start out assuming good things will happen 1074 set_has_unsafe_access(false); 1075 set_max_vector_size(0); 1076 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1077 set_decompile_count(0); 1078 1079 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1080 set_num_loop_opts(LoopOptsCount); 1081 set_do_inlining(Inline); 1082 set_max_inline_size(MaxInlineSize); 1083 set_freq_inline_size(FreqInlineSize); 1084 set_do_scheduling(OptoScheduling); 1085 set_do_count_invocations(false); 1086 set_do_method_data_update(false); 1087 set_age_code(has_method() && method()->profile_aging()); 1088 set_rtm_state(NoRTM); // No RTM lock eliding by default 1089 #if INCLUDE_RTM_OPT 1090 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1091 int rtm_state = method()->method_data()->rtm_state(); 1092 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1093 // Don't generate RTM lock eliding code. 1094 set_rtm_state(NoRTM); 1095 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1096 // Generate RTM lock eliding code without abort ratio calculation code. 1097 set_rtm_state(UseRTM); 1098 } else if (UseRTMDeopt) { 1099 // Generate RTM lock eliding code and include abort ratio calculation 1100 // code if UseRTMDeopt is on. 1101 set_rtm_state(ProfileRTM); 1102 } 1103 } 1104 #endif 1105 if (debug_info()->recording_non_safepoints()) { 1106 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1107 (comp_arena(), 8, 0, NULL)); 1108 set_default_node_notes(Node_Notes::make(this)); 1109 } 1110 1111 // // -- Initialize types before each compile -- 1112 // // Update cached type information 1113 // if( _method && _method->constants() ) 1114 // Type::update_loaded_types(_method, _method->constants()); 1115 1116 // Init alias_type map. 1117 if (!_do_escape_analysis && aliaslevel == 3) 1118 aliaslevel = 2; // No unique types without escape analysis 1119 _AliasLevel = aliaslevel; 1120 const int grow_ats = 16; 1121 _max_alias_types = grow_ats; 1122 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1123 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1124 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1125 { 1126 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1127 } 1128 // Initialize the first few types. 1129 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1130 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1131 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1132 _num_alias_types = AliasIdxRaw+1; 1133 // Zero out the alias type cache. 1134 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1135 // A NULL adr_type hits in the cache right away. Preload the right answer. 1136 probe_alias_cache(NULL)->_index = AliasIdxTop; 1137 1138 _intrinsics = NULL; 1139 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1140 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1141 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1142 register_library_intrinsics(); 1143 } 1144 1145 //---------------------------init_start---------------------------------------- 1146 // Install the StartNode on this compile object. 1147 void Compile::init_start(StartNode* s) { 1148 if (failing()) 1149 return; // already failing 1150 assert(s == start(), ""); 1151 } 1152 1153 StartNode* Compile::start() const { 1154 assert(!failing(), ""); 1155 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1156 Node* start = root()->fast_out(i); 1157 if( start->is_Start() ) 1158 return start->as_Start(); 1159 } 1160 fatal("Did not find Start node!"); 1161 return NULL; 1162 } 1163 1164 //-------------------------------immutable_memory------------------------------------- 1165 // Access immutable memory 1166 Node* Compile::immutable_memory() { 1167 if (_immutable_memory != NULL) { 1168 return _immutable_memory; 1169 } 1170 StartNode* s = start(); 1171 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1172 Node *p = s->fast_out(i); 1173 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1174 _immutable_memory = p; 1175 return _immutable_memory; 1176 } 1177 } 1178 ShouldNotReachHere(); 1179 return NULL; 1180 } 1181 1182 //----------------------set_cached_top_node------------------------------------ 1183 // Install the cached top node, and make sure Node::is_top works correctly. 1184 void Compile::set_cached_top_node(Node* tn) { 1185 if (tn != NULL) verify_top(tn); 1186 Node* old_top = _top; 1187 _top = tn; 1188 // Calling Node::setup_is_top allows the nodes the chance to adjust 1189 // their _out arrays. 1190 if (_top != NULL) _top->setup_is_top(); 1191 if (old_top != NULL) old_top->setup_is_top(); 1192 assert(_top == NULL || top()->is_top(), ""); 1193 } 1194 1195 #ifdef ASSERT 1196 uint Compile::count_live_nodes_by_graph_walk() { 1197 Unique_Node_List useful(comp_arena()); 1198 // Get useful node list by walking the graph. 1199 identify_useful_nodes(useful); 1200 return useful.size(); 1201 } 1202 1203 void Compile::print_missing_nodes() { 1204 1205 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1206 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1207 return; 1208 } 1209 1210 // This is an expensive function. It is executed only when the user 1211 // specifies VerifyIdealNodeCount option or otherwise knows the 1212 // additional work that needs to be done to identify reachable nodes 1213 // by walking the flow graph and find the missing ones using 1214 // _dead_node_list. 1215 1216 Unique_Node_List useful(comp_arena()); 1217 // Get useful node list by walking the graph. 1218 identify_useful_nodes(useful); 1219 1220 uint l_nodes = C->live_nodes(); 1221 uint l_nodes_by_walk = useful.size(); 1222 1223 if (l_nodes != l_nodes_by_walk) { 1224 if (_log != NULL) { 1225 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1226 _log->stamp(); 1227 _log->end_head(); 1228 } 1229 VectorSet& useful_member_set = useful.member_set(); 1230 int last_idx = l_nodes_by_walk; 1231 for (int i = 0; i < last_idx; i++) { 1232 if (useful_member_set.test(i)) { 1233 if (_dead_node_list.test(i)) { 1234 if (_log != NULL) { 1235 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1236 } 1237 if (PrintIdealNodeCount) { 1238 // Print the log message to tty 1239 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1240 useful.at(i)->dump(); 1241 } 1242 } 1243 } 1244 else if (! _dead_node_list.test(i)) { 1245 if (_log != NULL) { 1246 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1247 } 1248 if (PrintIdealNodeCount) { 1249 // Print the log message to tty 1250 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1251 } 1252 } 1253 } 1254 if (_log != NULL) { 1255 _log->tail("mismatched_nodes"); 1256 } 1257 } 1258 } 1259 void Compile::record_modified_node(Node* n) { 1260 if (_modified_nodes != NULL && !_inlining_incrementally && 1261 n->outcnt() != 0 && !n->is_Con()) { 1262 _modified_nodes->push(n); 1263 } 1264 } 1265 1266 void Compile::remove_modified_node(Node* n) { 1267 if (_modified_nodes != NULL) { 1268 _modified_nodes->remove(n); 1269 } 1270 } 1271 #endif 1272 1273 #ifndef PRODUCT 1274 void Compile::verify_top(Node* tn) const { 1275 if (tn != NULL) { 1276 assert(tn->is_Con(), "top node must be a constant"); 1277 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1278 assert(tn->in(0) != NULL, "must have live top node"); 1279 } 1280 } 1281 #endif 1282 1283 1284 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1285 1286 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1287 guarantee(arr != NULL, ""); 1288 int num_blocks = arr->length(); 1289 if (grow_by < num_blocks) grow_by = num_blocks; 1290 int num_notes = grow_by * _node_notes_block_size; 1291 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1292 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1293 while (num_notes > 0) { 1294 arr->append(notes); 1295 notes += _node_notes_block_size; 1296 num_notes -= _node_notes_block_size; 1297 } 1298 assert(num_notes == 0, "exact multiple, please"); 1299 } 1300 1301 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1302 if (source == NULL || dest == NULL) return false; 1303 1304 if (dest->is_Con()) 1305 return false; // Do not push debug info onto constants. 1306 1307 #ifdef ASSERT 1308 // Leave a bread crumb trail pointing to the original node: 1309 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1310 dest->set_debug_orig(source); 1311 } 1312 #endif 1313 1314 if (node_note_array() == NULL) 1315 return false; // Not collecting any notes now. 1316 1317 // This is a copy onto a pre-existing node, which may already have notes. 1318 // If both nodes have notes, do not overwrite any pre-existing notes. 1319 Node_Notes* source_notes = node_notes_at(source->_idx); 1320 if (source_notes == NULL || source_notes->is_clear()) return false; 1321 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1322 if (dest_notes == NULL || dest_notes->is_clear()) { 1323 return set_node_notes_at(dest->_idx, source_notes); 1324 } 1325 1326 Node_Notes merged_notes = (*source_notes); 1327 // The order of operations here ensures that dest notes will win... 1328 merged_notes.update_from(dest_notes); 1329 return set_node_notes_at(dest->_idx, &merged_notes); 1330 } 1331 1332 1333 //--------------------------allow_range_check_smearing------------------------- 1334 // Gating condition for coalescing similar range checks. 1335 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1336 // single covering check that is at least as strong as any of them. 1337 // If the optimization succeeds, the simplified (strengthened) range check 1338 // will always succeed. If it fails, we will deopt, and then give up 1339 // on the optimization. 1340 bool Compile::allow_range_check_smearing() const { 1341 // If this method has already thrown a range-check, 1342 // assume it was because we already tried range smearing 1343 // and it failed. 1344 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1345 return !already_trapped; 1346 } 1347 1348 1349 //------------------------------flatten_alias_type----------------------------- 1350 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1351 int offset = tj->offset(); 1352 TypePtr::PTR ptr = tj->ptr(); 1353 1354 // Known instance (scalarizable allocation) alias only with itself. 1355 bool is_known_inst = tj->isa_oopptr() != NULL && 1356 tj->is_oopptr()->is_known_instance(); 1357 1358 // Process weird unsafe references. 1359 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1360 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1361 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1362 tj = TypeOopPtr::BOTTOM; 1363 ptr = tj->ptr(); 1364 offset = tj->offset(); 1365 } 1366 1367 // Array pointers need some flattening 1368 const TypeAryPtr *ta = tj->isa_aryptr(); 1369 if (ta && ta->is_stable()) { 1370 // Erase stability property for alias analysis. 1371 tj = ta = ta->cast_to_stable(false); 1372 } 1373 if( ta && is_known_inst ) { 1374 if ( offset != Type::OffsetBot && 1375 offset > arrayOopDesc::length_offset_in_bytes() ) { 1376 offset = Type::OffsetBot; // Flatten constant access into array body only 1377 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1378 } 1379 } else if( ta && _AliasLevel >= 2 ) { 1380 // For arrays indexed by constant indices, we flatten the alias 1381 // space to include all of the array body. Only the header, klass 1382 // and array length can be accessed un-aliased. 1383 if( offset != Type::OffsetBot ) { 1384 if( ta->const_oop() ) { // MethodData* or Method* 1385 offset = Type::OffsetBot; // Flatten constant access into array body 1386 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1387 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1388 // range is OK as-is. 1389 tj = ta = TypeAryPtr::RANGE; 1390 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1391 tj = TypeInstPtr::KLASS; // all klass loads look alike 1392 ta = TypeAryPtr::RANGE; // generic ignored junk 1393 ptr = TypePtr::BotPTR; 1394 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1395 tj = TypeInstPtr::MARK; 1396 ta = TypeAryPtr::RANGE; // generic ignored junk 1397 ptr = TypePtr::BotPTR; 1398 } else { // Random constant offset into array body 1399 offset = Type::OffsetBot; // Flatten constant access into array body 1400 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1401 } 1402 } 1403 // Arrays of fixed size alias with arrays of unknown size. 1404 if (ta->size() != TypeInt::POS) { 1405 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1406 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1407 } 1408 // Arrays of known objects become arrays of unknown objects. 1409 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1410 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1411 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1412 } 1413 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1414 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1415 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1416 } 1417 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1418 // cannot be distinguished by bytecode alone. 1419 if (ta->elem() == TypeInt::BOOL) { 1420 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1421 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1422 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1423 } 1424 // During the 2nd round of IterGVN, NotNull castings are removed. 1425 // Make sure the Bottom and NotNull variants alias the same. 1426 // Also, make sure exact and non-exact variants alias the same. 1427 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1428 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1429 } 1430 } 1431 1432 // Oop pointers need some flattening 1433 const TypeInstPtr *to = tj->isa_instptr(); 1434 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1435 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1436 if( ptr == TypePtr::Constant ) { 1437 if (to->klass() != ciEnv::current()->Class_klass() || 1438 offset < k->size_helper() * wordSize) { 1439 // No constant oop pointers (such as Strings); they alias with 1440 // unknown strings. 1441 assert(!is_known_inst, "not scalarizable allocation"); 1442 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1443 } 1444 } else if( is_known_inst ) { 1445 tj = to; // Keep NotNull and klass_is_exact for instance type 1446 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1447 // During the 2nd round of IterGVN, NotNull castings are removed. 1448 // Make sure the Bottom and NotNull variants alias the same. 1449 // Also, make sure exact and non-exact variants alias the same. 1450 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1451 } 1452 if (to->speculative() != NULL) { 1453 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1454 } 1455 // Canonicalize the holder of this field 1456 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1457 // First handle header references such as a LoadKlassNode, even if the 1458 // object's klass is unloaded at compile time (4965979). 1459 if (!is_known_inst) { // Do it only for non-instance types 1460 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1461 } 1462 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1463 // Static fields are in the space above the normal instance 1464 // fields in the java.lang.Class instance. 1465 if (to->klass() != ciEnv::current()->Class_klass()) { 1466 to = NULL; 1467 tj = TypeOopPtr::BOTTOM; 1468 offset = tj->offset(); 1469 } 1470 } else { 1471 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1472 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1473 if( is_known_inst ) { 1474 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1475 } else { 1476 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1477 } 1478 } 1479 } 1480 } 1481 1482 // Klass pointers to object array klasses need some flattening 1483 const TypeKlassPtr *tk = tj->isa_klassptr(); 1484 if( tk ) { 1485 // If we are referencing a field within a Klass, we need 1486 // to assume the worst case of an Object. Both exact and 1487 // inexact types must flatten to the same alias class so 1488 // use NotNull as the PTR. 1489 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1490 1491 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1492 TypeKlassPtr::OBJECT->klass(), 1493 offset); 1494 } 1495 1496 ciKlass* klass = tk->klass(); 1497 if( klass->is_obj_array_klass() ) { 1498 ciKlass* k = TypeAryPtr::OOPS->klass(); 1499 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1500 k = TypeInstPtr::BOTTOM->klass(); 1501 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1502 } 1503 1504 // Check for precise loads from the primary supertype array and force them 1505 // to the supertype cache alias index. Check for generic array loads from 1506 // the primary supertype array and also force them to the supertype cache 1507 // alias index. Since the same load can reach both, we need to merge 1508 // these 2 disparate memories into the same alias class. Since the 1509 // primary supertype array is read-only, there's no chance of confusion 1510 // where we bypass an array load and an array store. 1511 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1512 if (offset == Type::OffsetBot || 1513 (offset >= primary_supers_offset && 1514 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1515 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1516 offset = in_bytes(Klass::secondary_super_cache_offset()); 1517 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1518 } 1519 } 1520 1521 // Flatten all Raw pointers together. 1522 if (tj->base() == Type::RawPtr) 1523 tj = TypeRawPtr::BOTTOM; 1524 1525 if (tj->base() == Type::AnyPtr) 1526 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1527 1528 // Flatten all to bottom for now 1529 switch( _AliasLevel ) { 1530 case 0: 1531 tj = TypePtr::BOTTOM; 1532 break; 1533 case 1: // Flatten to: oop, static, field or array 1534 switch (tj->base()) { 1535 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1536 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1537 case Type::AryPtr: // do not distinguish arrays at all 1538 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1539 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1540 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1541 default: ShouldNotReachHere(); 1542 } 1543 break; 1544 case 2: // No collapsing at level 2; keep all splits 1545 case 3: // No collapsing at level 3; keep all splits 1546 break; 1547 default: 1548 Unimplemented(); 1549 } 1550 1551 offset = tj->offset(); 1552 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1553 1554 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1555 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1556 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1557 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1558 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1559 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1560 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1561 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1562 assert( tj->ptr() != TypePtr::TopPTR && 1563 tj->ptr() != TypePtr::AnyNull && 1564 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1565 // assert( tj->ptr() != TypePtr::Constant || 1566 // tj->base() == Type::RawPtr || 1567 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1568 1569 return tj; 1570 } 1571 1572 void Compile::AliasType::Init(int i, const TypePtr* at) { 1573 _index = i; 1574 _adr_type = at; 1575 _field = NULL; 1576 _element = NULL; 1577 _is_rewritable = true; // default 1578 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1579 if (atoop != NULL && atoop->is_known_instance()) { 1580 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1581 _general_index = Compile::current()->get_alias_index(gt); 1582 } else { 1583 _general_index = 0; 1584 } 1585 } 1586 1587 //---------------------------------print_on------------------------------------ 1588 #ifndef PRODUCT 1589 void Compile::AliasType::print_on(outputStream* st) { 1590 if (index() < 10) 1591 st->print("@ <%d> ", index()); 1592 else st->print("@ <%d>", index()); 1593 st->print(is_rewritable() ? " " : " RO"); 1594 int offset = adr_type()->offset(); 1595 if (offset == Type::OffsetBot) 1596 st->print(" +any"); 1597 else st->print(" +%-3d", offset); 1598 st->print(" in "); 1599 adr_type()->dump_on(st); 1600 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1601 if (field() != NULL && tjp) { 1602 if (tjp->klass() != field()->holder() || 1603 tjp->offset() != field()->offset_in_bytes()) { 1604 st->print(" != "); 1605 field()->print(); 1606 st->print(" ***"); 1607 } 1608 } 1609 } 1610 1611 void print_alias_types() { 1612 Compile* C = Compile::current(); 1613 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1614 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1615 C->alias_type(idx)->print_on(tty); 1616 tty->cr(); 1617 } 1618 } 1619 #endif 1620 1621 1622 //----------------------------probe_alias_cache-------------------------------- 1623 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1624 intptr_t key = (intptr_t) adr_type; 1625 key ^= key >> logAliasCacheSize; 1626 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1627 } 1628 1629 1630 //-----------------------------grow_alias_types-------------------------------- 1631 void Compile::grow_alias_types() { 1632 const int old_ats = _max_alias_types; // how many before? 1633 const int new_ats = old_ats; // how many more? 1634 const int grow_ats = old_ats+new_ats; // how many now? 1635 _max_alias_types = grow_ats; 1636 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1637 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1638 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1639 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1640 } 1641 1642 1643 //--------------------------------find_alias_type------------------------------ 1644 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1645 if (_AliasLevel == 0) 1646 return alias_type(AliasIdxBot); 1647 1648 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1649 if (ace->_adr_type == adr_type) { 1650 return alias_type(ace->_index); 1651 } 1652 1653 // Handle special cases. 1654 if (adr_type == NULL) return alias_type(AliasIdxTop); 1655 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1656 1657 // Do it the slow way. 1658 const TypePtr* flat = flatten_alias_type(adr_type); 1659 1660 #ifdef ASSERT 1661 assert(flat == flatten_alias_type(flat), "idempotent"); 1662 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1663 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1664 const TypeOopPtr* foop = flat->is_oopptr(); 1665 // Scalarizable allocations have exact klass always. 1666 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1667 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1668 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1669 } 1670 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1671 #endif 1672 1673 int idx = AliasIdxTop; 1674 for (int i = 0; i < num_alias_types(); i++) { 1675 if (alias_type(i)->adr_type() == flat) { 1676 idx = i; 1677 break; 1678 } 1679 } 1680 1681 if (idx == AliasIdxTop) { 1682 if (no_create) return NULL; 1683 // Grow the array if necessary. 1684 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1685 // Add a new alias type. 1686 idx = _num_alias_types++; 1687 _alias_types[idx]->Init(idx, flat); 1688 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1689 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1690 if (flat->isa_instptr()) { 1691 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1692 && flat->is_instptr()->klass() == env()->Class_klass()) 1693 alias_type(idx)->set_rewritable(false); 1694 } 1695 if (flat->isa_aryptr()) { 1696 #ifdef ASSERT 1697 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1698 // (T_BYTE has the weakest alignment and size restrictions...) 1699 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1700 #endif 1701 if (flat->offset() == TypePtr::OffsetBot) { 1702 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1703 } 1704 } 1705 if (flat->isa_klassptr()) { 1706 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1707 alias_type(idx)->set_rewritable(false); 1708 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1709 alias_type(idx)->set_rewritable(false); 1710 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1711 alias_type(idx)->set_rewritable(false); 1712 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1713 alias_type(idx)->set_rewritable(false); 1714 } 1715 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1716 // but the base pointer type is not distinctive enough to identify 1717 // references into JavaThread.) 1718 1719 // Check for final fields. 1720 const TypeInstPtr* tinst = flat->isa_instptr(); 1721 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1722 ciField* field; 1723 if (tinst->const_oop() != NULL && 1724 tinst->klass() == ciEnv::current()->Class_klass() && 1725 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1726 // static field 1727 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1728 field = k->get_field_by_offset(tinst->offset(), true); 1729 } else { 1730 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1731 field = k->get_field_by_offset(tinst->offset(), false); 1732 } 1733 assert(field == NULL || 1734 original_field == NULL || 1735 (field->holder() == original_field->holder() && 1736 field->offset() == original_field->offset() && 1737 field->is_static() == original_field->is_static()), "wrong field?"); 1738 // Set field() and is_rewritable() attributes. 1739 if (field != NULL) alias_type(idx)->set_field(field); 1740 } 1741 } 1742 1743 // Fill the cache for next time. 1744 ace->_adr_type = adr_type; 1745 ace->_index = idx; 1746 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1747 1748 // Might as well try to fill the cache for the flattened version, too. 1749 AliasCacheEntry* face = probe_alias_cache(flat); 1750 if (face->_adr_type == NULL) { 1751 face->_adr_type = flat; 1752 face->_index = idx; 1753 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1754 } 1755 1756 return alias_type(idx); 1757 } 1758 1759 1760 Compile::AliasType* Compile::alias_type(ciField* field) { 1761 const TypeOopPtr* t; 1762 if (field->is_static()) 1763 t = TypeInstPtr::make(field->holder()->java_mirror()); 1764 else 1765 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1766 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1767 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1768 return atp; 1769 } 1770 1771 1772 //------------------------------have_alias_type-------------------------------- 1773 bool Compile::have_alias_type(const TypePtr* adr_type) { 1774 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1775 if (ace->_adr_type == adr_type) { 1776 return true; 1777 } 1778 1779 // Handle special cases. 1780 if (adr_type == NULL) return true; 1781 if (adr_type == TypePtr::BOTTOM) return true; 1782 1783 return find_alias_type(adr_type, true, NULL) != NULL; 1784 } 1785 1786 //-----------------------------must_alias-------------------------------------- 1787 // True if all values of the given address type are in the given alias category. 1788 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1789 if (alias_idx == AliasIdxBot) return true; // the universal category 1790 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1791 if (alias_idx == AliasIdxTop) return false; // the empty category 1792 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1793 1794 // the only remaining possible overlap is identity 1795 int adr_idx = get_alias_index(adr_type); 1796 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1797 assert(adr_idx == alias_idx || 1798 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1799 && adr_type != TypeOopPtr::BOTTOM), 1800 "should not be testing for overlap with an unsafe pointer"); 1801 return adr_idx == alias_idx; 1802 } 1803 1804 //------------------------------can_alias-------------------------------------- 1805 // True if any values of the given address type are in the given alias category. 1806 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1807 if (alias_idx == AliasIdxTop) return false; // the empty category 1808 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1809 if (alias_idx == AliasIdxBot) return true; // the universal category 1810 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1811 1812 // the only remaining possible overlap is identity 1813 int adr_idx = get_alias_index(adr_type); 1814 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1815 return adr_idx == alias_idx; 1816 } 1817 1818 1819 1820 //---------------------------pop_warm_call------------------------------------- 1821 WarmCallInfo* Compile::pop_warm_call() { 1822 WarmCallInfo* wci = _warm_calls; 1823 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1824 return wci; 1825 } 1826 1827 //----------------------------Inline_Warm-------------------------------------- 1828 int Compile::Inline_Warm() { 1829 // If there is room, try to inline some more warm call sites. 1830 // %%% Do a graph index compaction pass when we think we're out of space? 1831 if (!InlineWarmCalls) return 0; 1832 1833 int calls_made_hot = 0; 1834 int room_to_grow = NodeCountInliningCutoff - unique(); 1835 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1836 int amount_grown = 0; 1837 WarmCallInfo* call; 1838 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1839 int est_size = (int)call->size(); 1840 if (est_size > (room_to_grow - amount_grown)) { 1841 // This one won't fit anyway. Get rid of it. 1842 call->make_cold(); 1843 continue; 1844 } 1845 call->make_hot(); 1846 calls_made_hot++; 1847 amount_grown += est_size; 1848 amount_to_grow -= est_size; 1849 } 1850 1851 if (calls_made_hot > 0) set_major_progress(); 1852 return calls_made_hot; 1853 } 1854 1855 1856 //----------------------------Finish_Warm-------------------------------------- 1857 void Compile::Finish_Warm() { 1858 if (!InlineWarmCalls) return; 1859 if (failing()) return; 1860 if (warm_calls() == NULL) return; 1861 1862 // Clean up loose ends, if we are out of space for inlining. 1863 WarmCallInfo* call; 1864 while ((call = pop_warm_call()) != NULL) { 1865 call->make_cold(); 1866 } 1867 } 1868 1869 //---------------------cleanup_loop_predicates----------------------- 1870 // Remove the opaque nodes that protect the predicates so that all unused 1871 // checks and uncommon_traps will be eliminated from the ideal graph 1872 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1873 if (predicate_count()==0) return; 1874 for (int i = predicate_count(); i > 0; i--) { 1875 Node * n = predicate_opaque1_node(i-1); 1876 assert(n->Opcode() == Op_Opaque1, "must be"); 1877 igvn.replace_node(n, n->in(1)); 1878 } 1879 assert(predicate_count()==0, "should be clean!"); 1880 } 1881 1882 // StringOpts and late inlining of string methods 1883 void Compile::inline_string_calls(bool parse_time) { 1884 { 1885 // remove useless nodes to make the usage analysis simpler 1886 ResourceMark rm; 1887 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1888 } 1889 1890 { 1891 ResourceMark rm; 1892 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1893 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1894 print_method(PHASE_AFTER_STRINGOPTS, 3); 1895 } 1896 1897 // now inline anything that we skipped the first time around 1898 if (!parse_time) { 1899 _late_inlines_pos = _late_inlines.length(); 1900 } 1901 1902 while (_string_late_inlines.length() > 0) { 1903 CallGenerator* cg = _string_late_inlines.pop(); 1904 cg->do_late_inline(); 1905 if (failing()) return; 1906 } 1907 _string_late_inlines.trunc_to(0); 1908 } 1909 1910 // Late inlining of boxing methods 1911 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1912 if (_boxing_late_inlines.length() > 0) { 1913 assert(has_boxed_value(), "inconsistent"); 1914 1915 PhaseGVN* gvn = initial_gvn(); 1916 set_inlining_incrementally(true); 1917 1918 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1919 for_igvn()->clear(); 1920 gvn->replace_with(&igvn); 1921 1922 _late_inlines_pos = _late_inlines.length(); 1923 1924 while (_boxing_late_inlines.length() > 0) { 1925 CallGenerator* cg = _boxing_late_inlines.pop(); 1926 cg->do_late_inline(); 1927 if (failing()) return; 1928 } 1929 _boxing_late_inlines.trunc_to(0); 1930 1931 { 1932 ResourceMark rm; 1933 PhaseRemoveUseless pru(gvn, for_igvn()); 1934 } 1935 1936 igvn = PhaseIterGVN(gvn); 1937 igvn.optimize(); 1938 1939 set_inlining_progress(false); 1940 set_inlining_incrementally(false); 1941 } 1942 } 1943 1944 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1945 assert(IncrementalInline, "incremental inlining should be on"); 1946 PhaseGVN* gvn = initial_gvn(); 1947 1948 set_inlining_progress(false); 1949 for_igvn()->clear(); 1950 gvn->replace_with(&igvn); 1951 1952 int i = 0; 1953 1954 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1955 CallGenerator* cg = _late_inlines.at(i); 1956 _late_inlines_pos = i+1; 1957 cg->do_late_inline(); 1958 if (failing()) return; 1959 } 1960 int j = 0; 1961 for (; i < _late_inlines.length(); i++, j++) { 1962 _late_inlines.at_put(j, _late_inlines.at(i)); 1963 } 1964 _late_inlines.trunc_to(j); 1965 1966 { 1967 ResourceMark rm; 1968 PhaseRemoveUseless pru(gvn, for_igvn()); 1969 } 1970 1971 igvn = PhaseIterGVN(gvn); 1972 } 1973 1974 // Perform incremental inlining until bound on number of live nodes is reached 1975 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1976 PhaseGVN* gvn = initial_gvn(); 1977 1978 set_inlining_incrementally(true); 1979 set_inlining_progress(true); 1980 uint low_live_nodes = 0; 1981 1982 while(inlining_progress() && _late_inlines.length() > 0) { 1983 1984 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1985 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1986 // PhaseIdealLoop is expensive so we only try it once we are 1987 // out of live nodes and we only try it again if the previous 1988 // helped got the number of nodes down significantly 1989 PhaseIdealLoop ideal_loop( igvn, false, true ); 1990 if (failing()) return; 1991 low_live_nodes = live_nodes(); 1992 _major_progress = true; 1993 } 1994 1995 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1996 break; 1997 } 1998 } 1999 2000 inline_incrementally_one(igvn); 2001 2002 if (failing()) return; 2003 2004 igvn.optimize(); 2005 2006 if (failing()) return; 2007 } 2008 2009 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2010 2011 if (_string_late_inlines.length() > 0) { 2012 assert(has_stringbuilder(), "inconsistent"); 2013 for_igvn()->clear(); 2014 initial_gvn()->replace_with(&igvn); 2015 2016 inline_string_calls(false); 2017 2018 if (failing()) return; 2019 2020 { 2021 ResourceMark rm; 2022 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2023 } 2024 2025 igvn = PhaseIterGVN(gvn); 2026 2027 igvn.optimize(); 2028 } 2029 2030 set_inlining_incrementally(false); 2031 } 2032 2033 2034 //------------------------------Optimize--------------------------------------- 2035 // Given a graph, optimize it. 2036 void Compile::Optimize() { 2037 TracePhase t1("optimizer", &_t_optimizer, true); 2038 2039 #ifndef PRODUCT 2040 if (env()->break_at_compile()) { 2041 BREAKPOINT; 2042 } 2043 2044 #endif 2045 2046 ResourceMark rm; 2047 int loop_opts_cnt; 2048 2049 print_inlining_reinit(); 2050 2051 NOT_PRODUCT( verify_graph_edges(); ) 2052 2053 print_method(PHASE_AFTER_PARSING); 2054 2055 { 2056 // Iterative Global Value Numbering, including ideal transforms 2057 // Initialize IterGVN with types and values from parse-time GVN 2058 PhaseIterGVN igvn(initial_gvn()); 2059 #ifdef ASSERT 2060 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2061 #endif 2062 { 2063 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 2064 igvn.optimize(); 2065 } 2066 2067 print_method(PHASE_ITER_GVN1, 2); 2068 2069 if (failing()) return; 2070 2071 { 2072 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2073 inline_incrementally(igvn); 2074 } 2075 2076 print_method(PHASE_INCREMENTAL_INLINE, 2); 2077 2078 if (failing()) return; 2079 2080 if (eliminate_boxing()) { 2081 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2082 // Inline valueOf() methods now. 2083 inline_boxing_calls(igvn); 2084 2085 if (AlwaysIncrementalInline) { 2086 inline_incrementally(igvn); 2087 } 2088 2089 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2090 2091 if (failing()) return; 2092 } 2093 2094 // Remove the speculative part of types and clean up the graph from 2095 // the extra CastPP nodes whose only purpose is to carry them. Do 2096 // that early so that optimizations are not disrupted by the extra 2097 // CastPP nodes. 2098 remove_speculative_types(igvn); 2099 2100 // No more new expensive nodes will be added to the list from here 2101 // so keep only the actual candidates for optimizations. 2102 cleanup_expensive_nodes(igvn); 2103 2104 // Perform escape analysis 2105 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2106 if (has_loops()) { 2107 // Cleanup graph (remove dead nodes). 2108 TracePhase t2("idealLoop", &_t_idealLoop, true); 2109 PhaseIdealLoop ideal_loop( igvn, false, true ); 2110 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2111 if (failing()) return; 2112 } 2113 ConnectionGraph::do_analysis(this, &igvn); 2114 2115 if (failing()) return; 2116 2117 // Optimize out fields loads from scalar replaceable allocations. 2118 igvn.optimize(); 2119 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2120 2121 if (failing()) return; 2122 2123 if (congraph() != NULL && macro_count() > 0) { 2124 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2125 PhaseMacroExpand mexp(igvn); 2126 mexp.eliminate_macro_nodes(); 2127 igvn.set_delay_transform(false); 2128 2129 igvn.optimize(); 2130 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2131 2132 if (failing()) return; 2133 } 2134 } 2135 2136 // Loop transforms on the ideal graph. Range Check Elimination, 2137 // peeling, unrolling, etc. 2138 2139 // Set loop opts counter 2140 loop_opts_cnt = num_loop_opts(); 2141 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2142 { 2143 TracePhase t2("idealLoop", &_t_idealLoop, true); 2144 PhaseIdealLoop ideal_loop( igvn, true ); 2145 loop_opts_cnt--; 2146 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2147 if (failing()) return; 2148 } 2149 // Loop opts pass if partial peeling occurred in previous pass 2150 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2151 TracePhase t3("idealLoop", &_t_idealLoop, true); 2152 PhaseIdealLoop ideal_loop( igvn, false ); 2153 loop_opts_cnt--; 2154 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2155 if (failing()) return; 2156 } 2157 // Loop opts pass for loop-unrolling before CCP 2158 if(major_progress() && (loop_opts_cnt > 0)) { 2159 TracePhase t4("idealLoop", &_t_idealLoop, true); 2160 PhaseIdealLoop ideal_loop( igvn, false ); 2161 loop_opts_cnt--; 2162 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2163 } 2164 if (!failing()) { 2165 // Verify that last round of loop opts produced a valid graph 2166 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2167 PhaseIdealLoop::verify(igvn); 2168 } 2169 } 2170 if (failing()) return; 2171 2172 // Conditional Constant Propagation; 2173 PhaseCCP ccp( &igvn ); 2174 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2175 { 2176 TracePhase t2("ccp", &_t_ccp, true); 2177 ccp.do_transform(); 2178 } 2179 print_method(PHASE_CPP1, 2); 2180 2181 assert( true, "Break here to ccp.dump_old2new_map()"); 2182 2183 // Iterative Global Value Numbering, including ideal transforms 2184 { 2185 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2186 igvn = ccp; 2187 igvn.optimize(); 2188 } 2189 2190 print_method(PHASE_ITER_GVN2, 2); 2191 2192 if (failing()) return; 2193 2194 // Loop transforms on the ideal graph. Range Check Elimination, 2195 // peeling, unrolling, etc. 2196 if(loop_opts_cnt > 0) { 2197 debug_only( int cnt = 0; ); 2198 while(major_progress() && (loop_opts_cnt > 0)) { 2199 TracePhase t2("idealLoop", &_t_idealLoop, true); 2200 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2201 PhaseIdealLoop ideal_loop( igvn, true); 2202 loop_opts_cnt--; 2203 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2204 if (failing()) return; 2205 } 2206 } 2207 2208 { 2209 // Verify that all previous optimizations produced a valid graph 2210 // at least to this point, even if no loop optimizations were done. 2211 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2212 PhaseIdealLoop::verify(igvn); 2213 } 2214 2215 { 2216 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2217 PhaseMacroExpand mex(igvn); 2218 if (mex.expand_macro_nodes()) { 2219 assert(failing(), "must bail out w/ explicit message"); 2220 return; 2221 } 2222 } 2223 2224 DEBUG_ONLY( _modified_nodes = NULL; ) 2225 } // (End scope of igvn; run destructor if necessary for asserts.) 2226 2227 process_print_inlining(); 2228 // A method with only infinite loops has no edges entering loops from root 2229 { 2230 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2231 if (final_graph_reshaping()) { 2232 assert(failing(), "must bail out w/ explicit message"); 2233 return; 2234 } 2235 } 2236 2237 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2238 } 2239 2240 2241 //------------------------------Code_Gen--------------------------------------- 2242 // Given a graph, generate code for it 2243 void Compile::Code_Gen() { 2244 if (failing()) { 2245 return; 2246 } 2247 2248 // Perform instruction selection. You might think we could reclaim Matcher 2249 // memory PDQ, but actually the Matcher is used in generating spill code. 2250 // Internals of the Matcher (including some VectorSets) must remain live 2251 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2252 // set a bit in reclaimed memory. 2253 2254 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2255 // nodes. Mapping is only valid at the root of each matched subtree. 2256 NOT_PRODUCT( verify_graph_edges(); ) 2257 2258 Matcher matcher; 2259 _matcher = &matcher; 2260 { 2261 TracePhase t2("matcher", &_t_matcher, true); 2262 matcher.match(); 2263 } 2264 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2265 // nodes. Mapping is only valid at the root of each matched subtree. 2266 NOT_PRODUCT( verify_graph_edges(); ) 2267 2268 // If you have too many nodes, or if matching has failed, bail out 2269 check_node_count(0, "out of nodes matching instructions"); 2270 if (failing()) { 2271 return; 2272 } 2273 2274 // Build a proper-looking CFG 2275 PhaseCFG cfg(node_arena(), root(), matcher); 2276 _cfg = &cfg; 2277 { 2278 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2279 bool success = cfg.do_global_code_motion(); 2280 if (!success) { 2281 return; 2282 } 2283 2284 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2285 NOT_PRODUCT( verify_graph_edges(); ) 2286 debug_only( cfg.verify(); ) 2287 } 2288 2289 PhaseChaitin regalloc(unique(), cfg, matcher); 2290 _regalloc = ®alloc; 2291 { 2292 TracePhase t2("regalloc", &_t_registerAllocation, true); 2293 // Perform register allocation. After Chaitin, use-def chains are 2294 // no longer accurate (at spill code) and so must be ignored. 2295 // Node->LRG->reg mappings are still accurate. 2296 _regalloc->Register_Allocate(); 2297 2298 // Bail out if the allocator builds too many nodes 2299 if (failing()) { 2300 return; 2301 } 2302 } 2303 2304 // Prior to register allocation we kept empty basic blocks in case the 2305 // the allocator needed a place to spill. After register allocation we 2306 // are not adding any new instructions. If any basic block is empty, we 2307 // can now safely remove it. 2308 { 2309 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2310 cfg.remove_empty_blocks(); 2311 if (do_freq_based_layout()) { 2312 PhaseBlockLayout layout(cfg); 2313 } else { 2314 cfg.set_loop_alignment(); 2315 } 2316 cfg.fixup_flow(); 2317 } 2318 2319 // Apply peephole optimizations 2320 if( OptoPeephole ) { 2321 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2322 PhasePeephole peep( _regalloc, cfg); 2323 peep.do_transform(); 2324 } 2325 2326 // Do late expand if CPU requires this. 2327 if (Matcher::require_postalloc_expand) { 2328 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true)); 2329 cfg.postalloc_expand(_regalloc); 2330 } 2331 2332 // Convert Nodes to instruction bits in a buffer 2333 { 2334 // %%%% workspace merge brought two timers together for one job 2335 TracePhase t2a("output", &_t_output, true); 2336 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2337 Output(); 2338 } 2339 2340 print_method(PHASE_FINAL_CODE); 2341 2342 // He's dead, Jim. 2343 _cfg = (PhaseCFG*)0xdeadbeef; 2344 _regalloc = (PhaseChaitin*)0xdeadbeef; 2345 } 2346 2347 2348 //------------------------------dump_asm--------------------------------------- 2349 // Dump formatted assembly 2350 #ifndef PRODUCT 2351 void Compile::dump_asm(int *pcs, uint pc_limit) { 2352 bool cut_short = false; 2353 tty->print_cr("#"); 2354 tty->print("# "); _tf->dump(); tty->cr(); 2355 tty->print_cr("#"); 2356 2357 // For all blocks 2358 int pc = 0x0; // Program counter 2359 char starts_bundle = ' '; 2360 _regalloc->dump_frame(); 2361 2362 Node *n = NULL; 2363 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2364 if (VMThread::should_terminate()) { 2365 cut_short = true; 2366 break; 2367 } 2368 Block* block = _cfg->get_block(i); 2369 if (block->is_connector() && !Verbose) { 2370 continue; 2371 } 2372 n = block->head(); 2373 if (pcs && n->_idx < pc_limit) { 2374 tty->print("%3.3x ", pcs[n->_idx]); 2375 } else { 2376 tty->print(" "); 2377 } 2378 block->dump_head(_cfg); 2379 if (block->is_connector()) { 2380 tty->print_cr(" # Empty connector block"); 2381 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2382 tty->print_cr(" # Block is sole successor of call"); 2383 } 2384 2385 // For all instructions 2386 Node *delay = NULL; 2387 for (uint j = 0; j < block->number_of_nodes(); j++) { 2388 if (VMThread::should_terminate()) { 2389 cut_short = true; 2390 break; 2391 } 2392 n = block->get_node(j); 2393 if (valid_bundle_info(n)) { 2394 Bundle* bundle = node_bundling(n); 2395 if (bundle->used_in_unconditional_delay()) { 2396 delay = n; 2397 continue; 2398 } 2399 if (bundle->starts_bundle()) { 2400 starts_bundle = '+'; 2401 } 2402 } 2403 2404 if (WizardMode) { 2405 n->dump(); 2406 } 2407 2408 if( !n->is_Region() && // Dont print in the Assembly 2409 !n->is_Phi() && // a few noisely useless nodes 2410 !n->is_Proj() && 2411 !n->is_MachTemp() && 2412 !n->is_SafePointScalarObject() && 2413 !n->is_Catch() && // Would be nice to print exception table targets 2414 !n->is_MergeMem() && // Not very interesting 2415 !n->is_top() && // Debug info table constants 2416 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2417 ) { 2418 if (pcs && n->_idx < pc_limit) 2419 tty->print("%3.3x", pcs[n->_idx]); 2420 else 2421 tty->print(" "); 2422 tty->print(" %c ", starts_bundle); 2423 starts_bundle = ' '; 2424 tty->print("\t"); 2425 n->format(_regalloc, tty); 2426 tty->cr(); 2427 } 2428 2429 // If we have an instruction with a delay slot, and have seen a delay, 2430 // then back up and print it 2431 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2432 assert(delay != NULL, "no unconditional delay instruction"); 2433 if (WizardMode) delay->dump(); 2434 2435 if (node_bundling(delay)->starts_bundle()) 2436 starts_bundle = '+'; 2437 if (pcs && n->_idx < pc_limit) 2438 tty->print("%3.3x", pcs[n->_idx]); 2439 else 2440 tty->print(" "); 2441 tty->print(" %c ", starts_bundle); 2442 starts_bundle = ' '; 2443 tty->print("\t"); 2444 delay->format(_regalloc, tty); 2445 tty->cr(); 2446 delay = NULL; 2447 } 2448 2449 // Dump the exception table as well 2450 if( n->is_Catch() && (Verbose || WizardMode) ) { 2451 // Print the exception table for this offset 2452 _handler_table.print_subtable_for(pc); 2453 } 2454 } 2455 2456 if (pcs && n->_idx < pc_limit) 2457 tty->print_cr("%3.3x", pcs[n->_idx]); 2458 else 2459 tty->cr(); 2460 2461 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2462 2463 } // End of per-block dump 2464 tty->cr(); 2465 2466 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2467 } 2468 #endif 2469 2470 //------------------------------Final_Reshape_Counts--------------------------- 2471 // This class defines counters to help identify when a method 2472 // may/must be executed using hardware with only 24-bit precision. 2473 struct Final_Reshape_Counts : public StackObj { 2474 int _call_count; // count non-inlined 'common' calls 2475 int _float_count; // count float ops requiring 24-bit precision 2476 int _double_count; // count double ops requiring more precision 2477 int _java_call_count; // count non-inlined 'java' calls 2478 int _inner_loop_count; // count loops which need alignment 2479 VectorSet _visited; // Visitation flags 2480 Node_List _tests; // Set of IfNodes & PCTableNodes 2481 2482 Final_Reshape_Counts() : 2483 _call_count(0), _float_count(0), _double_count(0), 2484 _java_call_count(0), _inner_loop_count(0), 2485 _visited( Thread::current()->resource_area() ) { } 2486 2487 void inc_call_count () { _call_count ++; } 2488 void inc_float_count () { _float_count ++; } 2489 void inc_double_count() { _double_count++; } 2490 void inc_java_call_count() { _java_call_count++; } 2491 void inc_inner_loop_count() { _inner_loop_count++; } 2492 2493 int get_call_count () const { return _call_count ; } 2494 int get_float_count () const { return _float_count ; } 2495 int get_double_count() const { return _double_count; } 2496 int get_java_call_count() const { return _java_call_count; } 2497 int get_inner_loop_count() const { return _inner_loop_count; } 2498 }; 2499 2500 #ifdef ASSERT 2501 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2502 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2503 // Make sure the offset goes inside the instance layout. 2504 return k->contains_field_offset(tp->offset()); 2505 // Note that OffsetBot and OffsetTop are very negative. 2506 } 2507 #endif 2508 2509 // Eliminate trivially redundant StoreCMs and accumulate their 2510 // precedence edges. 2511 void Compile::eliminate_redundant_card_marks(Node* n) { 2512 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2513 if (n->in(MemNode::Address)->outcnt() > 1) { 2514 // There are multiple users of the same address so it might be 2515 // possible to eliminate some of the StoreCMs 2516 Node* mem = n->in(MemNode::Memory); 2517 Node* adr = n->in(MemNode::Address); 2518 Node* val = n->in(MemNode::ValueIn); 2519 Node* prev = n; 2520 bool done = false; 2521 // Walk the chain of StoreCMs eliminating ones that match. As 2522 // long as it's a chain of single users then the optimization is 2523 // safe. Eliminating partially redundant StoreCMs would require 2524 // cloning copies down the other paths. 2525 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2526 if (adr == mem->in(MemNode::Address) && 2527 val == mem->in(MemNode::ValueIn)) { 2528 // redundant StoreCM 2529 if (mem->req() > MemNode::OopStore) { 2530 // Hasn't been processed by this code yet. 2531 n->add_prec(mem->in(MemNode::OopStore)); 2532 } else { 2533 // Already converted to precedence edge 2534 for (uint i = mem->req(); i < mem->len(); i++) { 2535 // Accumulate any precedence edges 2536 if (mem->in(i) != NULL) { 2537 n->add_prec(mem->in(i)); 2538 } 2539 } 2540 // Everything above this point has been processed. 2541 done = true; 2542 } 2543 // Eliminate the previous StoreCM 2544 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2545 assert(mem->outcnt() == 0, "should be dead"); 2546 mem->disconnect_inputs(NULL, this); 2547 } else { 2548 prev = mem; 2549 } 2550 mem = prev->in(MemNode::Memory); 2551 } 2552 } 2553 } 2554 2555 //------------------------------final_graph_reshaping_impl---------------------- 2556 // Implement items 1-5 from final_graph_reshaping below. 2557 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2558 2559 if ( n->outcnt() == 0 ) return; // dead node 2560 uint nop = n->Opcode(); 2561 2562 // Check for 2-input instruction with "last use" on right input. 2563 // Swap to left input. Implements item (2). 2564 if( n->req() == 3 && // two-input instruction 2565 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2566 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2567 n->in(2)->outcnt() == 1 &&// right use IS a last use 2568 !n->in(2)->is_Con() ) { // right use is not a constant 2569 // Check for commutative opcode 2570 switch( nop ) { 2571 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2572 case Op_MaxI: case Op_MinI: 2573 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2574 case Op_AndL: case Op_XorL: case Op_OrL: 2575 case Op_AndI: case Op_XorI: case Op_OrI: { 2576 // Move "last use" input to left by swapping inputs 2577 n->swap_edges(1, 2); 2578 break; 2579 } 2580 default: 2581 break; 2582 } 2583 } 2584 2585 #ifdef ASSERT 2586 if( n->is_Mem() ) { 2587 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2588 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2589 // oop will be recorded in oop map if load crosses safepoint 2590 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2591 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2592 "raw memory operations should have control edge"); 2593 } 2594 #endif 2595 // Count FPU ops and common calls, implements item (3) 2596 switch( nop ) { 2597 // Count all float operations that may use FPU 2598 case Op_AddF: 2599 case Op_SubF: 2600 case Op_MulF: 2601 case Op_DivF: 2602 case Op_NegF: 2603 case Op_ModF: 2604 case Op_ConvI2F: 2605 case Op_ConF: 2606 case Op_CmpF: 2607 case Op_CmpF3: 2608 // case Op_ConvL2F: // longs are split into 32-bit halves 2609 frc.inc_float_count(); 2610 break; 2611 2612 case Op_ConvF2D: 2613 case Op_ConvD2F: 2614 frc.inc_float_count(); 2615 frc.inc_double_count(); 2616 break; 2617 2618 // Count all double operations that may use FPU 2619 case Op_AddD: 2620 case Op_SubD: 2621 case Op_MulD: 2622 case Op_DivD: 2623 case Op_NegD: 2624 case Op_ModD: 2625 case Op_ConvI2D: 2626 case Op_ConvD2I: 2627 // case Op_ConvL2D: // handled by leaf call 2628 // case Op_ConvD2L: // handled by leaf call 2629 case Op_ConD: 2630 case Op_CmpD: 2631 case Op_CmpD3: 2632 frc.inc_double_count(); 2633 break; 2634 case Op_Opaque1: // Remove Opaque Nodes before matching 2635 case Op_Opaque2: // Remove Opaque Nodes before matching 2636 case Op_Opaque3: 2637 n->subsume_by(n->in(1), this); 2638 break; 2639 case Op_CallStaticJava: 2640 case Op_CallJava: 2641 case Op_CallDynamicJava: 2642 frc.inc_java_call_count(); // Count java call site; 2643 case Op_CallRuntime: 2644 case Op_CallLeaf: 2645 case Op_CallLeafNoFP: { 2646 assert( n->is_Call(), "" ); 2647 CallNode *call = n->as_Call(); 2648 // Count call sites where the FP mode bit would have to be flipped. 2649 // Do not count uncommon runtime calls: 2650 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2651 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2652 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2653 frc.inc_call_count(); // Count the call site 2654 } else { // See if uncommon argument is shared 2655 Node *n = call->in(TypeFunc::Parms); 2656 int nop = n->Opcode(); 2657 // Clone shared simple arguments to uncommon calls, item (1). 2658 if( n->outcnt() > 1 && 2659 !n->is_Proj() && 2660 nop != Op_CreateEx && 2661 nop != Op_CheckCastPP && 2662 nop != Op_DecodeN && 2663 nop != Op_DecodeNKlass && 2664 !n->is_Mem() ) { 2665 Node *x = n->clone(); 2666 call->set_req( TypeFunc::Parms, x ); 2667 } 2668 } 2669 break; 2670 } 2671 2672 case Op_StoreD: 2673 case Op_LoadD: 2674 case Op_LoadD_unaligned: 2675 frc.inc_double_count(); 2676 goto handle_mem; 2677 case Op_StoreF: 2678 case Op_LoadF: 2679 frc.inc_float_count(); 2680 goto handle_mem; 2681 2682 case Op_StoreCM: 2683 { 2684 // Convert OopStore dependence into precedence edge 2685 Node* prec = n->in(MemNode::OopStore); 2686 n->del_req(MemNode::OopStore); 2687 n->add_prec(prec); 2688 eliminate_redundant_card_marks(n); 2689 } 2690 2691 // fall through 2692 2693 case Op_StoreB: 2694 case Op_StoreC: 2695 case Op_StorePConditional: 2696 case Op_StoreI: 2697 case Op_StoreL: 2698 case Op_StoreIConditional: 2699 case Op_StoreLConditional: 2700 case Op_CompareAndSwapI: 2701 case Op_CompareAndSwapL: 2702 case Op_CompareAndSwapP: 2703 case Op_CompareAndSwapN: 2704 case Op_GetAndAddI: 2705 case Op_GetAndAddL: 2706 case Op_GetAndSetI: 2707 case Op_GetAndSetL: 2708 case Op_GetAndSetP: 2709 case Op_GetAndSetN: 2710 case Op_StoreP: 2711 case Op_StoreN: 2712 case Op_StoreNKlass: 2713 case Op_LoadB: 2714 case Op_LoadUB: 2715 case Op_LoadUS: 2716 case Op_LoadI: 2717 case Op_LoadKlass: 2718 case Op_LoadNKlass: 2719 case Op_LoadL: 2720 case Op_LoadL_unaligned: 2721 case Op_LoadPLocked: 2722 case Op_LoadP: 2723 case Op_LoadN: 2724 case Op_LoadRange: 2725 case Op_LoadS: { 2726 handle_mem: 2727 #ifdef ASSERT 2728 if( VerifyOptoOopOffsets ) { 2729 assert( n->is_Mem(), "" ); 2730 MemNode *mem = (MemNode*)n; 2731 // Check to see if address types have grounded out somehow. 2732 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2733 assert( !tp || oop_offset_is_sane(tp), "" ); 2734 } 2735 #endif 2736 break; 2737 } 2738 2739 case Op_AddP: { // Assert sane base pointers 2740 Node *addp = n->in(AddPNode::Address); 2741 assert( !addp->is_AddP() || 2742 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2743 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2744 "Base pointers must match" ); 2745 #ifdef _LP64 2746 if ((UseCompressedOops || UseCompressedClassPointers) && 2747 addp->Opcode() == Op_ConP && 2748 addp == n->in(AddPNode::Base) && 2749 n->in(AddPNode::Offset)->is_Con()) { 2750 // Use addressing with narrow klass to load with offset on x86. 2751 // On sparc loading 32-bits constant and decoding it have less 2752 // instructions (4) then load 64-bits constant (7). 2753 // Do this transformation here since IGVN will convert ConN back to ConP. 2754 const Type* t = addp->bottom_type(); 2755 if (t->isa_oopptr() || t->isa_klassptr()) { 2756 Node* nn = NULL; 2757 2758 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2759 2760 // Look for existing ConN node of the same exact type. 2761 Node* r = root(); 2762 uint cnt = r->outcnt(); 2763 for (uint i = 0; i < cnt; i++) { 2764 Node* m = r->raw_out(i); 2765 if (m!= NULL && m->Opcode() == op && 2766 m->bottom_type()->make_ptr() == t) { 2767 nn = m; 2768 break; 2769 } 2770 } 2771 if (nn != NULL) { 2772 // Decode a narrow oop to match address 2773 // [R12 + narrow_oop_reg<<3 + offset] 2774 if (t->isa_oopptr()) { 2775 nn = new DecodeNNode(nn, t); 2776 } else { 2777 nn = new DecodeNKlassNode(nn, t); 2778 } 2779 n->set_req(AddPNode::Base, nn); 2780 n->set_req(AddPNode::Address, nn); 2781 if (addp->outcnt() == 0) { 2782 addp->disconnect_inputs(NULL, this); 2783 } 2784 } 2785 } 2786 } 2787 #endif 2788 break; 2789 } 2790 2791 #ifdef _LP64 2792 case Op_CastPP: 2793 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2794 Node* in1 = n->in(1); 2795 const Type* t = n->bottom_type(); 2796 Node* new_in1 = in1->clone(); 2797 new_in1->as_DecodeN()->set_type(t); 2798 2799 if (!Matcher::narrow_oop_use_complex_address()) { 2800 // 2801 // x86, ARM and friends can handle 2 adds in addressing mode 2802 // and Matcher can fold a DecodeN node into address by using 2803 // a narrow oop directly and do implicit NULL check in address: 2804 // 2805 // [R12 + narrow_oop_reg<<3 + offset] 2806 // NullCheck narrow_oop_reg 2807 // 2808 // On other platforms (Sparc) we have to keep new DecodeN node and 2809 // use it to do implicit NULL check in address: 2810 // 2811 // decode_not_null narrow_oop_reg, base_reg 2812 // [base_reg + offset] 2813 // NullCheck base_reg 2814 // 2815 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2816 // to keep the information to which NULL check the new DecodeN node 2817 // corresponds to use it as value in implicit_null_check(). 2818 // 2819 new_in1->set_req(0, n->in(0)); 2820 } 2821 2822 n->subsume_by(new_in1, this); 2823 if (in1->outcnt() == 0) { 2824 in1->disconnect_inputs(NULL, this); 2825 } 2826 } 2827 break; 2828 2829 case Op_CmpP: 2830 // Do this transformation here to preserve CmpPNode::sub() and 2831 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2832 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2833 Node* in1 = n->in(1); 2834 Node* in2 = n->in(2); 2835 if (!in1->is_DecodeNarrowPtr()) { 2836 in2 = in1; 2837 in1 = n->in(2); 2838 } 2839 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2840 2841 Node* new_in2 = NULL; 2842 if (in2->is_DecodeNarrowPtr()) { 2843 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2844 new_in2 = in2->in(1); 2845 } else if (in2->Opcode() == Op_ConP) { 2846 const Type* t = in2->bottom_type(); 2847 if (t == TypePtr::NULL_PTR) { 2848 assert(in1->is_DecodeN(), "compare klass to null?"); 2849 // Don't convert CmpP null check into CmpN if compressed 2850 // oops implicit null check is not generated. 2851 // This will allow to generate normal oop implicit null check. 2852 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2853 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 2854 // 2855 // This transformation together with CastPP transformation above 2856 // will generated code for implicit NULL checks for compressed oops. 2857 // 2858 // The original code after Optimize() 2859 // 2860 // LoadN memory, narrow_oop_reg 2861 // decode narrow_oop_reg, base_reg 2862 // CmpP base_reg, NULL 2863 // CastPP base_reg // NotNull 2864 // Load [base_reg + offset], val_reg 2865 // 2866 // after these transformations will be 2867 // 2868 // LoadN memory, narrow_oop_reg 2869 // CmpN narrow_oop_reg, NULL 2870 // decode_not_null narrow_oop_reg, base_reg 2871 // Load [base_reg + offset], val_reg 2872 // 2873 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2874 // since narrow oops can be used in debug info now (see the code in 2875 // final_graph_reshaping_walk()). 2876 // 2877 // At the end the code will be matched to 2878 // on x86: 2879 // 2880 // Load_narrow_oop memory, narrow_oop_reg 2881 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2882 // NullCheck narrow_oop_reg 2883 // 2884 // and on sparc: 2885 // 2886 // Load_narrow_oop memory, narrow_oop_reg 2887 // decode_not_null narrow_oop_reg, base_reg 2888 // Load [base_reg + offset], val_reg 2889 // NullCheck base_reg 2890 // 2891 } else if (t->isa_oopptr()) { 2892 new_in2 = ConNode::make(t->make_narrowoop()); 2893 } else if (t->isa_klassptr()) { 2894 new_in2 = ConNode::make(t->make_narrowklass()); 2895 } 2896 } 2897 if (new_in2 != NULL) { 2898 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 2899 n->subsume_by(cmpN, this); 2900 if (in1->outcnt() == 0) { 2901 in1->disconnect_inputs(NULL, this); 2902 } 2903 if (in2->outcnt() == 0) { 2904 in2->disconnect_inputs(NULL, this); 2905 } 2906 } 2907 } 2908 break; 2909 2910 case Op_DecodeN: 2911 case Op_DecodeNKlass: 2912 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2913 // DecodeN could be pinned when it can't be fold into 2914 // an address expression, see the code for Op_CastPP above. 2915 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2916 break; 2917 2918 case Op_EncodeP: 2919 case Op_EncodePKlass: { 2920 Node* in1 = n->in(1); 2921 if (in1->is_DecodeNarrowPtr()) { 2922 n->subsume_by(in1->in(1), this); 2923 } else if (in1->Opcode() == Op_ConP) { 2924 const Type* t = in1->bottom_type(); 2925 if (t == TypePtr::NULL_PTR) { 2926 assert(t->isa_oopptr(), "null klass?"); 2927 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 2928 } else if (t->isa_oopptr()) { 2929 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 2930 } else if (t->isa_klassptr()) { 2931 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 2932 } 2933 } 2934 if (in1->outcnt() == 0) { 2935 in1->disconnect_inputs(NULL, this); 2936 } 2937 break; 2938 } 2939 2940 case Op_Proj: { 2941 if (OptimizeStringConcat) { 2942 ProjNode* p = n->as_Proj(); 2943 if (p->_is_io_use) { 2944 // Separate projections were used for the exception path which 2945 // are normally removed by a late inline. If it wasn't inlined 2946 // then they will hang around and should just be replaced with 2947 // the original one. 2948 Node* proj = NULL; 2949 // Replace with just one 2950 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2951 Node *use = i.get(); 2952 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2953 proj = use; 2954 break; 2955 } 2956 } 2957 assert(proj != NULL, "must be found"); 2958 p->subsume_by(proj, this); 2959 } 2960 } 2961 break; 2962 } 2963 2964 case Op_Phi: 2965 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2966 // The EncodeP optimization may create Phi with the same edges 2967 // for all paths. It is not handled well by Register Allocator. 2968 Node* unique_in = n->in(1); 2969 assert(unique_in != NULL, ""); 2970 uint cnt = n->req(); 2971 for (uint i = 2; i < cnt; i++) { 2972 Node* m = n->in(i); 2973 assert(m != NULL, ""); 2974 if (unique_in != m) 2975 unique_in = NULL; 2976 } 2977 if (unique_in != NULL) { 2978 n->subsume_by(unique_in, this); 2979 } 2980 } 2981 break; 2982 2983 #endif 2984 2985 case Op_ModI: 2986 if (UseDivMod) { 2987 // Check if a%b and a/b both exist 2988 Node* d = n->find_similar(Op_DivI); 2989 if (d) { 2990 // Replace them with a fused divmod if supported 2991 if (Matcher::has_match_rule(Op_DivModI)) { 2992 DivModINode* divmod = DivModINode::make(n); 2993 d->subsume_by(divmod->div_proj(), this); 2994 n->subsume_by(divmod->mod_proj(), this); 2995 } else { 2996 // replace a%b with a-((a/b)*b) 2997 Node* mult = new MulINode(d, d->in(2)); 2998 Node* sub = new SubINode(d->in(1), mult); 2999 n->subsume_by(sub, this); 3000 } 3001 } 3002 } 3003 break; 3004 3005 case Op_ModL: 3006 if (UseDivMod) { 3007 // Check if a%b and a/b both exist 3008 Node* d = n->find_similar(Op_DivL); 3009 if (d) { 3010 // Replace them with a fused divmod if supported 3011 if (Matcher::has_match_rule(Op_DivModL)) { 3012 DivModLNode* divmod = DivModLNode::make(n); 3013 d->subsume_by(divmod->div_proj(), this); 3014 n->subsume_by(divmod->mod_proj(), this); 3015 } else { 3016 // replace a%b with a-((a/b)*b) 3017 Node* mult = new MulLNode(d, d->in(2)); 3018 Node* sub = new SubLNode(d->in(1), mult); 3019 n->subsume_by(sub, this); 3020 } 3021 } 3022 } 3023 break; 3024 3025 case Op_LoadVector: 3026 case Op_StoreVector: 3027 break; 3028 3029 case Op_PackB: 3030 case Op_PackS: 3031 case Op_PackI: 3032 case Op_PackF: 3033 case Op_PackL: 3034 case Op_PackD: 3035 if (n->req()-1 > 2) { 3036 // Replace many operand PackNodes with a binary tree for matching 3037 PackNode* p = (PackNode*) n; 3038 Node* btp = p->binary_tree_pack(1, n->req()); 3039 n->subsume_by(btp, this); 3040 } 3041 break; 3042 case Op_Loop: 3043 case Op_CountedLoop: 3044 if (n->as_Loop()->is_inner_loop()) { 3045 frc.inc_inner_loop_count(); 3046 } 3047 break; 3048 case Op_LShiftI: 3049 case Op_RShiftI: 3050 case Op_URShiftI: 3051 case Op_LShiftL: 3052 case Op_RShiftL: 3053 case Op_URShiftL: 3054 if (Matcher::need_masked_shift_count) { 3055 // The cpu's shift instructions don't restrict the count to the 3056 // lower 5/6 bits. We need to do the masking ourselves. 3057 Node* in2 = n->in(2); 3058 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3059 const TypeInt* t = in2->find_int_type(); 3060 if (t != NULL && t->is_con()) { 3061 juint shift = t->get_con(); 3062 if (shift > mask) { // Unsigned cmp 3063 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3064 } 3065 } else { 3066 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3067 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3068 n->set_req(2, shift); 3069 } 3070 } 3071 if (in2->outcnt() == 0) { // Remove dead node 3072 in2->disconnect_inputs(NULL, this); 3073 } 3074 } 3075 break; 3076 case Op_MemBarStoreStore: 3077 case Op_MemBarRelease: 3078 // Break the link with AllocateNode: it is no longer useful and 3079 // confuses register allocation. 3080 if (n->req() > MemBarNode::Precedent) { 3081 n->set_req(MemBarNode::Precedent, top()); 3082 } 3083 break; 3084 default: 3085 assert( !n->is_Call(), "" ); 3086 assert( !n->is_Mem(), "" ); 3087 break; 3088 } 3089 3090 // Collect CFG split points 3091 if (n->is_MultiBranch()) 3092 frc._tests.push(n); 3093 } 3094 3095 //------------------------------final_graph_reshaping_walk--------------------- 3096 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3097 // requires that the walk visits a node's inputs before visiting the node. 3098 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3099 ResourceArea *area = Thread::current()->resource_area(); 3100 Unique_Node_List sfpt(area); 3101 3102 frc._visited.set(root->_idx); // first, mark node as visited 3103 uint cnt = root->req(); 3104 Node *n = root; 3105 uint i = 0; 3106 while (true) { 3107 if (i < cnt) { 3108 // Place all non-visited non-null inputs onto stack 3109 Node* m = n->in(i); 3110 ++i; 3111 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3112 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3113 // compute worst case interpreter size in case of a deoptimization 3114 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3115 3116 sfpt.push(m); 3117 } 3118 cnt = m->req(); 3119 nstack.push(n, i); // put on stack parent and next input's index 3120 n = m; 3121 i = 0; 3122 } 3123 } else { 3124 // Now do post-visit work 3125 final_graph_reshaping_impl( n, frc ); 3126 if (nstack.is_empty()) 3127 break; // finished 3128 n = nstack.node(); // Get node from stack 3129 cnt = n->req(); 3130 i = nstack.index(); 3131 nstack.pop(); // Shift to the next node on stack 3132 } 3133 } 3134 3135 // Skip next transformation if compressed oops are not used. 3136 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3137 (!UseCompressedOops && !UseCompressedClassPointers)) 3138 return; 3139 3140 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3141 // It could be done for an uncommon traps or any safepoints/calls 3142 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3143 while (sfpt.size() > 0) { 3144 n = sfpt.pop(); 3145 JVMState *jvms = n->as_SafePoint()->jvms(); 3146 assert(jvms != NULL, "sanity"); 3147 int start = jvms->debug_start(); 3148 int end = n->req(); 3149 bool is_uncommon = (n->is_CallStaticJava() && 3150 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3151 for (int j = start; j < end; j++) { 3152 Node* in = n->in(j); 3153 if (in->is_DecodeNarrowPtr()) { 3154 bool safe_to_skip = true; 3155 if (!is_uncommon ) { 3156 // Is it safe to skip? 3157 for (uint i = 0; i < in->outcnt(); i++) { 3158 Node* u = in->raw_out(i); 3159 if (!u->is_SafePoint() || 3160 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3161 safe_to_skip = false; 3162 } 3163 } 3164 } 3165 if (safe_to_skip) { 3166 n->set_req(j, in->in(1)); 3167 } 3168 if (in->outcnt() == 0) { 3169 in->disconnect_inputs(NULL, this); 3170 } 3171 } 3172 } 3173 } 3174 } 3175 3176 //------------------------------final_graph_reshaping-------------------------- 3177 // Final Graph Reshaping. 3178 // 3179 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3180 // and not commoned up and forced early. Must come after regular 3181 // optimizations to avoid GVN undoing the cloning. Clone constant 3182 // inputs to Loop Phis; these will be split by the allocator anyways. 3183 // Remove Opaque nodes. 3184 // (2) Move last-uses by commutative operations to the left input to encourage 3185 // Intel update-in-place two-address operations and better register usage 3186 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3187 // calls canonicalizing them back. 3188 // (3) Count the number of double-precision FP ops, single-precision FP ops 3189 // and call sites. On Intel, we can get correct rounding either by 3190 // forcing singles to memory (requires extra stores and loads after each 3191 // FP bytecode) or we can set a rounding mode bit (requires setting and 3192 // clearing the mode bit around call sites). The mode bit is only used 3193 // if the relative frequency of single FP ops to calls is low enough. 3194 // This is a key transform for SPEC mpeg_audio. 3195 // (4) Detect infinite loops; blobs of code reachable from above but not 3196 // below. Several of the Code_Gen algorithms fail on such code shapes, 3197 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3198 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3199 // Detection is by looking for IfNodes where only 1 projection is 3200 // reachable from below or CatchNodes missing some targets. 3201 // (5) Assert for insane oop offsets in debug mode. 3202 3203 bool Compile::final_graph_reshaping() { 3204 // an infinite loop may have been eliminated by the optimizer, 3205 // in which case the graph will be empty. 3206 if (root()->req() == 1) { 3207 record_method_not_compilable("trivial infinite loop"); 3208 return true; 3209 } 3210 3211 // Expensive nodes have their control input set to prevent the GVN 3212 // from freely commoning them. There's no GVN beyond this point so 3213 // no need to keep the control input. We want the expensive nodes to 3214 // be freely moved to the least frequent code path by gcm. 3215 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3216 for (int i = 0; i < expensive_count(); i++) { 3217 _expensive_nodes->at(i)->set_req(0, NULL); 3218 } 3219 3220 Final_Reshape_Counts frc; 3221 3222 // Visit everybody reachable! 3223 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3224 Node_Stack nstack(unique() >> 1); 3225 final_graph_reshaping_walk(nstack, root(), frc); 3226 3227 // Check for unreachable (from below) code (i.e., infinite loops). 3228 for( uint i = 0; i < frc._tests.size(); i++ ) { 3229 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3230 // Get number of CFG targets. 3231 // Note that PCTables include exception targets after calls. 3232 uint required_outcnt = n->required_outcnt(); 3233 if (n->outcnt() != required_outcnt) { 3234 // Check for a few special cases. Rethrow Nodes never take the 3235 // 'fall-thru' path, so expected kids is 1 less. 3236 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3237 if (n->in(0)->in(0)->is_Call()) { 3238 CallNode *call = n->in(0)->in(0)->as_Call(); 3239 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3240 required_outcnt--; // Rethrow always has 1 less kid 3241 } else if (call->req() > TypeFunc::Parms && 3242 call->is_CallDynamicJava()) { 3243 // Check for null receiver. In such case, the optimizer has 3244 // detected that the virtual call will always result in a null 3245 // pointer exception. The fall-through projection of this CatchNode 3246 // will not be populated. 3247 Node *arg0 = call->in(TypeFunc::Parms); 3248 if (arg0->is_Type() && 3249 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3250 required_outcnt--; 3251 } 3252 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3253 call->req() > TypeFunc::Parms+1 && 3254 call->is_CallStaticJava()) { 3255 // Check for negative array length. In such case, the optimizer has 3256 // detected that the allocation attempt will always result in an 3257 // exception. There is no fall-through projection of this CatchNode . 3258 Node *arg1 = call->in(TypeFunc::Parms+1); 3259 if (arg1->is_Type() && 3260 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3261 required_outcnt--; 3262 } 3263 } 3264 } 3265 } 3266 // Recheck with a better notion of 'required_outcnt' 3267 if (n->outcnt() != required_outcnt) { 3268 record_method_not_compilable("malformed control flow"); 3269 return true; // Not all targets reachable! 3270 } 3271 } 3272 // Check that I actually visited all kids. Unreached kids 3273 // must be infinite loops. 3274 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3275 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3276 record_method_not_compilable("infinite loop"); 3277 return true; // Found unvisited kid; must be unreach 3278 } 3279 } 3280 3281 // If original bytecodes contained a mixture of floats and doubles 3282 // check if the optimizer has made it homogenous, item (3). 3283 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3284 frc.get_float_count() > 32 && 3285 frc.get_double_count() == 0 && 3286 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3287 set_24_bit_selection_and_mode( false, true ); 3288 } 3289 3290 set_java_calls(frc.get_java_call_count()); 3291 set_inner_loops(frc.get_inner_loop_count()); 3292 3293 // No infinite loops, no reason to bail out. 3294 return false; 3295 } 3296 3297 //-----------------------------too_many_traps---------------------------------- 3298 // Report if there are too many traps at the current method and bci. 3299 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3300 bool Compile::too_many_traps(ciMethod* method, 3301 int bci, 3302 Deoptimization::DeoptReason reason) { 3303 ciMethodData* md = method->method_data(); 3304 if (md->is_empty()) { 3305 // Assume the trap has not occurred, or that it occurred only 3306 // because of a transient condition during start-up in the interpreter. 3307 return false; 3308 } 3309 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3310 if (md->has_trap_at(bci, m, reason) != 0) { 3311 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3312 // Also, if there are multiple reasons, or if there is no per-BCI record, 3313 // assume the worst. 3314 if (log()) 3315 log()->elem("observe trap='%s' count='%d'", 3316 Deoptimization::trap_reason_name(reason), 3317 md->trap_count(reason)); 3318 return true; 3319 } else { 3320 // Ignore method/bci and see if there have been too many globally. 3321 return too_many_traps(reason, md); 3322 } 3323 } 3324 3325 // Less-accurate variant which does not require a method and bci. 3326 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3327 ciMethodData* logmd) { 3328 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3329 // Too many traps globally. 3330 // Note that we use cumulative trap_count, not just md->trap_count. 3331 if (log()) { 3332 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3333 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3334 Deoptimization::trap_reason_name(reason), 3335 mcount, trap_count(reason)); 3336 } 3337 return true; 3338 } else { 3339 // The coast is clear. 3340 return false; 3341 } 3342 } 3343 3344 //--------------------------too_many_recompiles-------------------------------- 3345 // Report if there are too many recompiles at the current method and bci. 3346 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3347 // Is not eager to return true, since this will cause the compiler to use 3348 // Action_none for a trap point, to avoid too many recompilations. 3349 bool Compile::too_many_recompiles(ciMethod* method, 3350 int bci, 3351 Deoptimization::DeoptReason reason) { 3352 ciMethodData* md = method->method_data(); 3353 if (md->is_empty()) { 3354 // Assume the trap has not occurred, or that it occurred only 3355 // because of a transient condition during start-up in the interpreter. 3356 return false; 3357 } 3358 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3359 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3360 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3361 Deoptimization::DeoptReason per_bc_reason 3362 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3363 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3364 if ((per_bc_reason == Deoptimization::Reason_none 3365 || md->has_trap_at(bci, m, reason) != 0) 3366 // The trap frequency measure we care about is the recompile count: 3367 && md->trap_recompiled_at(bci, m) 3368 && md->overflow_recompile_count() >= bc_cutoff) { 3369 // Do not emit a trap here if it has already caused recompilations. 3370 // Also, if there are multiple reasons, or if there is no per-BCI record, 3371 // assume the worst. 3372 if (log()) 3373 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3374 Deoptimization::trap_reason_name(reason), 3375 md->trap_count(reason), 3376 md->overflow_recompile_count()); 3377 return true; 3378 } else if (trap_count(reason) != 0 3379 && decompile_count() >= m_cutoff) { 3380 // Too many recompiles globally, and we have seen this sort of trap. 3381 // Use cumulative decompile_count, not just md->decompile_count. 3382 if (log()) 3383 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3384 Deoptimization::trap_reason_name(reason), 3385 md->trap_count(reason), trap_count(reason), 3386 md->decompile_count(), decompile_count()); 3387 return true; 3388 } else { 3389 // The coast is clear. 3390 return false; 3391 } 3392 } 3393 3394 // Compute when not to trap. Used by matching trap based nodes and 3395 // NullCheck optimization. 3396 void Compile::set_allowed_deopt_reasons() { 3397 _allowed_reasons = 0; 3398 if (is_method_compilation()) { 3399 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3400 assert(rs < BitsPerInt, "recode bit map"); 3401 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3402 _allowed_reasons |= nth_bit(rs); 3403 } 3404 } 3405 } 3406 } 3407 3408 #ifndef PRODUCT 3409 //------------------------------verify_graph_edges--------------------------- 3410 // Walk the Graph and verify that there is a one-to-one correspondence 3411 // between Use-Def edges and Def-Use edges in the graph. 3412 void Compile::verify_graph_edges(bool no_dead_code) { 3413 if (VerifyGraphEdges) { 3414 ResourceArea *area = Thread::current()->resource_area(); 3415 Unique_Node_List visited(area); 3416 // Call recursive graph walk to check edges 3417 _root->verify_edges(visited); 3418 if (no_dead_code) { 3419 // Now make sure that no visited node is used by an unvisited node. 3420 bool dead_nodes = false; 3421 Unique_Node_List checked(area); 3422 while (visited.size() > 0) { 3423 Node* n = visited.pop(); 3424 checked.push(n); 3425 for (uint i = 0; i < n->outcnt(); i++) { 3426 Node* use = n->raw_out(i); 3427 if (checked.member(use)) continue; // already checked 3428 if (visited.member(use)) continue; // already in the graph 3429 if (use->is_Con()) continue; // a dead ConNode is OK 3430 // At this point, we have found a dead node which is DU-reachable. 3431 if (!dead_nodes) { 3432 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3433 dead_nodes = true; 3434 } 3435 use->dump(2); 3436 tty->print_cr("---"); 3437 checked.push(use); // No repeats; pretend it is now checked. 3438 } 3439 } 3440 assert(!dead_nodes, "using nodes must be reachable from root"); 3441 } 3442 } 3443 } 3444 3445 // Verify GC barriers consistency 3446 // Currently supported: 3447 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3448 void Compile::verify_barriers() { 3449 if (UseG1GC) { 3450 // Verify G1 pre-barriers 3451 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3452 3453 ResourceArea *area = Thread::current()->resource_area(); 3454 Unique_Node_List visited(area); 3455 Node_List worklist(area); 3456 // We're going to walk control flow backwards starting from the Root 3457 worklist.push(_root); 3458 while (worklist.size() > 0) { 3459 Node* x = worklist.pop(); 3460 if (x == NULL || x == top()) continue; 3461 if (visited.member(x)) { 3462 continue; 3463 } else { 3464 visited.push(x); 3465 } 3466 3467 if (x->is_Region()) { 3468 for (uint i = 1; i < x->req(); i++) { 3469 worklist.push(x->in(i)); 3470 } 3471 } else { 3472 worklist.push(x->in(0)); 3473 // We are looking for the pattern: 3474 // /->ThreadLocal 3475 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3476 // \->ConI(0) 3477 // We want to verify that the If and the LoadB have the same control 3478 // See GraphKit::g1_write_barrier_pre() 3479 if (x->is_If()) { 3480 IfNode *iff = x->as_If(); 3481 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3482 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3483 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3484 && cmp->in(1)->is_Load()) { 3485 LoadNode* load = cmp->in(1)->as_Load(); 3486 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3487 && load->in(2)->in(3)->is_Con() 3488 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3489 3490 Node* if_ctrl = iff->in(0); 3491 Node* load_ctrl = load->in(0); 3492 3493 if (if_ctrl != load_ctrl) { 3494 // Skip possible CProj->NeverBranch in infinite loops 3495 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3496 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3497 if_ctrl = if_ctrl->in(0)->in(0); 3498 } 3499 } 3500 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3501 } 3502 } 3503 } 3504 } 3505 } 3506 } 3507 } 3508 } 3509 3510 #endif 3511 3512 // The Compile object keeps track of failure reasons separately from the ciEnv. 3513 // This is required because there is not quite a 1-1 relation between the 3514 // ciEnv and its compilation task and the Compile object. Note that one 3515 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3516 // to backtrack and retry without subsuming loads. Other than this backtracking 3517 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3518 // by the logic in C2Compiler. 3519 void Compile::record_failure(const char* reason) { 3520 if (log() != NULL) { 3521 log()->elem("failure reason='%s' phase='compile'", reason); 3522 } 3523 if (_failure_reason == NULL) { 3524 // Record the first failure reason. 3525 _failure_reason = reason; 3526 } 3527 3528 EventCompilerFailure event; 3529 if (event.should_commit()) { 3530 event.set_compileID(Compile::compile_id()); 3531 event.set_failure(reason); 3532 event.commit(); 3533 } 3534 3535 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3536 C->print_method(PHASE_FAILURE); 3537 } 3538 _root = NULL; // flush the graph, too 3539 } 3540 3541 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3542 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3543 _phase_name(name), _dolog(dolog) 3544 { 3545 if (dolog) { 3546 C = Compile::current(); 3547 _log = C->log(); 3548 } else { 3549 C = NULL; 3550 _log = NULL; 3551 } 3552 if (_log != NULL) { 3553 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3554 _log->stamp(); 3555 _log->end_head(); 3556 } 3557 } 3558 3559 Compile::TracePhase::~TracePhase() { 3560 3561 C = Compile::current(); 3562 if (_dolog) { 3563 _log = C->log(); 3564 } else { 3565 _log = NULL; 3566 } 3567 3568 #ifdef ASSERT 3569 if (PrintIdealNodeCount) { 3570 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3571 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3572 } 3573 3574 if (VerifyIdealNodeCount) { 3575 Compile::current()->print_missing_nodes(); 3576 } 3577 #endif 3578 3579 if (_log != NULL) { 3580 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3581 } 3582 } 3583 3584 //============================================================================= 3585 // Two Constant's are equal when the type and the value are equal. 3586 bool Compile::Constant::operator==(const Constant& other) { 3587 if (type() != other.type() ) return false; 3588 if (can_be_reused() != other.can_be_reused()) return false; 3589 // For floating point values we compare the bit pattern. 3590 switch (type()) { 3591 case T_FLOAT: return (_v._value.i == other._v._value.i); 3592 case T_LONG: 3593 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3594 case T_OBJECT: 3595 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3596 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3597 case T_METADATA: return (_v._metadata == other._v._metadata); 3598 default: ShouldNotReachHere(); 3599 } 3600 return false; 3601 } 3602 3603 static int type_to_size_in_bytes(BasicType t) { 3604 switch (t) { 3605 case T_LONG: return sizeof(jlong ); 3606 case T_FLOAT: return sizeof(jfloat ); 3607 case T_DOUBLE: return sizeof(jdouble); 3608 case T_METADATA: return sizeof(Metadata*); 3609 // We use T_VOID as marker for jump-table entries (labels) which 3610 // need an internal word relocation. 3611 case T_VOID: 3612 case T_ADDRESS: 3613 case T_OBJECT: return sizeof(jobject); 3614 } 3615 3616 ShouldNotReachHere(); 3617 return -1; 3618 } 3619 3620 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3621 // sort descending 3622 if (a->freq() > b->freq()) return -1; 3623 if (a->freq() < b->freq()) return 1; 3624 return 0; 3625 } 3626 3627 void Compile::ConstantTable::calculate_offsets_and_size() { 3628 // First, sort the array by frequencies. 3629 _constants.sort(qsort_comparator); 3630 3631 #ifdef ASSERT 3632 // Make sure all jump-table entries were sorted to the end of the 3633 // array (they have a negative frequency). 3634 bool found_void = false; 3635 for (int i = 0; i < _constants.length(); i++) { 3636 Constant con = _constants.at(i); 3637 if (con.type() == T_VOID) 3638 found_void = true; // jump-tables 3639 else 3640 assert(!found_void, "wrong sorting"); 3641 } 3642 #endif 3643 3644 int offset = 0; 3645 for (int i = 0; i < _constants.length(); i++) { 3646 Constant* con = _constants.adr_at(i); 3647 3648 // Align offset for type. 3649 int typesize = type_to_size_in_bytes(con->type()); 3650 offset = align_size_up(offset, typesize); 3651 con->set_offset(offset); // set constant's offset 3652 3653 if (con->type() == T_VOID) { 3654 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3655 offset = offset + typesize * n->outcnt(); // expand jump-table 3656 } else { 3657 offset = offset + typesize; 3658 } 3659 } 3660 3661 // Align size up to the next section start (which is insts; see 3662 // CodeBuffer::align_at_start). 3663 assert(_size == -1, "already set?"); 3664 _size = align_size_up(offset, CodeEntryAlignment); 3665 } 3666 3667 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3668 MacroAssembler _masm(&cb); 3669 for (int i = 0; i < _constants.length(); i++) { 3670 Constant con = _constants.at(i); 3671 address constant_addr; 3672 switch (con.type()) { 3673 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3674 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3675 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3676 case T_OBJECT: { 3677 jobject obj = con.get_jobject(); 3678 int oop_index = _masm.oop_recorder()->find_index(obj); 3679 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3680 break; 3681 } 3682 case T_ADDRESS: { 3683 address addr = (address) con.get_jobject(); 3684 constant_addr = _masm.address_constant(addr); 3685 break; 3686 } 3687 // We use T_VOID as marker for jump-table entries (labels) which 3688 // need an internal word relocation. 3689 case T_VOID: { 3690 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3691 // Fill the jump-table with a dummy word. The real value is 3692 // filled in later in fill_jump_table. 3693 address dummy = (address) n; 3694 constant_addr = _masm.address_constant(dummy); 3695 // Expand jump-table 3696 for (uint i = 1; i < n->outcnt(); i++) { 3697 address temp_addr = _masm.address_constant(dummy + i); 3698 assert(temp_addr, "consts section too small"); 3699 } 3700 break; 3701 } 3702 case T_METADATA: { 3703 Metadata* obj = con.get_metadata(); 3704 int metadata_index = _masm.oop_recorder()->find_index(obj); 3705 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3706 break; 3707 } 3708 default: ShouldNotReachHere(); 3709 } 3710 assert(constant_addr, "consts section too small"); 3711 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3712 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3713 } 3714 } 3715 3716 int Compile::ConstantTable::find_offset(Constant& con) const { 3717 int idx = _constants.find(con); 3718 assert(idx != -1, "constant must be in constant table"); 3719 int offset = _constants.at(idx).offset(); 3720 assert(offset != -1, "constant table not emitted yet?"); 3721 return offset; 3722 } 3723 3724 void Compile::ConstantTable::add(Constant& con) { 3725 if (con.can_be_reused()) { 3726 int idx = _constants.find(con); 3727 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3728 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3729 return; 3730 } 3731 } 3732 (void) _constants.append(con); 3733 } 3734 3735 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3736 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3737 Constant con(type, value, b->_freq); 3738 add(con); 3739 return con; 3740 } 3741 3742 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3743 Constant con(metadata); 3744 add(con); 3745 return con; 3746 } 3747 3748 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3749 jvalue value; 3750 BasicType type = oper->type()->basic_type(); 3751 switch (type) { 3752 case T_LONG: value.j = oper->constantL(); break; 3753 case T_FLOAT: value.f = oper->constantF(); break; 3754 case T_DOUBLE: value.d = oper->constantD(); break; 3755 case T_OBJECT: 3756 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3757 case T_METADATA: return add((Metadata*)oper->constant()); break; 3758 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3759 } 3760 return add(n, type, value); 3761 } 3762 3763 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3764 jvalue value; 3765 // We can use the node pointer here to identify the right jump-table 3766 // as this method is called from Compile::Fill_buffer right before 3767 // the MachNodes are emitted and the jump-table is filled (means the 3768 // MachNode pointers do not change anymore). 3769 value.l = (jobject) n; 3770 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3771 add(con); 3772 return con; 3773 } 3774 3775 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3776 // If called from Compile::scratch_emit_size do nothing. 3777 if (Compile::current()->in_scratch_emit_size()) return; 3778 3779 assert(labels.is_nonempty(), "must be"); 3780 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3781 3782 // Since MachConstantNode::constant_offset() also contains 3783 // table_base_offset() we need to subtract the table_base_offset() 3784 // to get the plain offset into the constant table. 3785 int offset = n->constant_offset() - table_base_offset(); 3786 3787 MacroAssembler _masm(&cb); 3788 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3789 3790 for (uint i = 0; i < n->outcnt(); i++) { 3791 address* constant_addr = &jump_table_base[i]; 3792 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i))); 3793 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3794 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3795 } 3796 } 3797 3798 //----------------------------static_subtype_check----------------------------- 3799 // Shortcut important common cases when superklass is exact: 3800 // (0) superklass is java.lang.Object (can occur in reflective code) 3801 // (1) subklass is already limited to a subtype of superklass => always ok 3802 // (2) subklass does not overlap with superklass => always fail 3803 // (3) superklass has NO subtypes and we can check with a simple compare. 3804 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 3805 if (StressReflectiveCode) { 3806 return SSC_full_test; // Let caller generate the general case. 3807 } 3808 3809 if (superk == env()->Object_klass()) { 3810 return SSC_always_true; // (0) this test cannot fail 3811 } 3812 3813 ciType* superelem = superk; 3814 if (superelem->is_array_klass()) 3815 superelem = superelem->as_array_klass()->base_element_type(); 3816 3817 if (!subk->is_interface()) { // cannot trust static interface types yet 3818 if (subk->is_subtype_of(superk)) { 3819 return SSC_always_true; // (1) false path dead; no dynamic test needed 3820 } 3821 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 3822 !superk->is_subtype_of(subk)) { 3823 return SSC_always_false; 3824 } 3825 } 3826 3827 // If casting to an instance klass, it must have no subtypes 3828 if (superk->is_interface()) { 3829 // Cannot trust interfaces yet. 3830 // %%% S.B. superk->nof_implementors() == 1 3831 } else if (superelem->is_instance_klass()) { 3832 ciInstanceKlass* ik = superelem->as_instance_klass(); 3833 if (!ik->has_subklass() && !ik->is_interface()) { 3834 if (!ik->is_final()) { 3835 // Add a dependency if there is a chance of a later subclass. 3836 dependencies()->assert_leaf_type(ik); 3837 } 3838 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3839 } 3840 } else { 3841 // A primitive array type has no subtypes. 3842 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3843 } 3844 3845 return SSC_full_test; 3846 } 3847 3848 // The message about the current inlining is accumulated in 3849 // _print_inlining_stream and transfered into the _print_inlining_list 3850 // once we know whether inlining succeeds or not. For regular 3851 // inlining, messages are appended to the buffer pointed by 3852 // _print_inlining_idx in the _print_inlining_list. For late inlining, 3853 // a new buffer is added after _print_inlining_idx in the list. This 3854 // way we can update the inlining message for late inlining call site 3855 // when the inlining is attempted again. 3856 void Compile::print_inlining_init() { 3857 if (print_inlining() || print_intrinsics()) { 3858 _print_inlining_stream = new stringStream(); 3859 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 3860 } 3861 } 3862 3863 void Compile::print_inlining_reinit() { 3864 if (print_inlining() || print_intrinsics()) { 3865 // Re allocate buffer when we change ResourceMark 3866 _print_inlining_stream = new stringStream(); 3867 } 3868 } 3869 3870 void Compile::print_inlining_reset() { 3871 _print_inlining_stream->reset(); 3872 } 3873 3874 void Compile::print_inlining_commit() { 3875 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 3876 // Transfer the message from _print_inlining_stream to the current 3877 // _print_inlining_list buffer and clear _print_inlining_stream. 3878 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 3879 print_inlining_reset(); 3880 } 3881 3882 void Compile::print_inlining_push() { 3883 // Add new buffer to the _print_inlining_list at current position 3884 _print_inlining_idx++; 3885 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 3886 } 3887 3888 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 3889 return _print_inlining_list->at(_print_inlining_idx); 3890 } 3891 3892 void Compile::print_inlining_update(CallGenerator* cg) { 3893 if (print_inlining() || print_intrinsics()) { 3894 if (!cg->is_late_inline()) { 3895 if (print_inlining_current().cg() != NULL) { 3896 print_inlining_push(); 3897 } 3898 print_inlining_commit(); 3899 } else { 3900 if (print_inlining_current().cg() != cg && 3901 (print_inlining_current().cg() != NULL || 3902 print_inlining_current().ss()->size() != 0)) { 3903 print_inlining_push(); 3904 } 3905 print_inlining_commit(); 3906 print_inlining_current().set_cg(cg); 3907 } 3908 } 3909 } 3910 3911 void Compile::print_inlining_move_to(CallGenerator* cg) { 3912 // We resume inlining at a late inlining call site. Locate the 3913 // corresponding inlining buffer so that we can update it. 3914 if (print_inlining()) { 3915 for (int i = 0; i < _print_inlining_list->length(); i++) { 3916 if (_print_inlining_list->adr_at(i)->cg() == cg) { 3917 _print_inlining_idx = i; 3918 return; 3919 } 3920 } 3921 ShouldNotReachHere(); 3922 } 3923 } 3924 3925 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 3926 if (print_inlining()) { 3927 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 3928 assert(print_inlining_current().cg() == cg, "wrong entry"); 3929 // replace message with new message 3930 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 3931 print_inlining_commit(); 3932 print_inlining_current().set_cg(cg); 3933 } 3934 } 3935 3936 void Compile::print_inlining_assert_ready() { 3937 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 3938 } 3939 3940 void Compile::process_print_inlining() { 3941 bool do_print_inlining = print_inlining() || print_intrinsics(); 3942 if (do_print_inlining || log() != NULL) { 3943 // Print inlining message for candidates that we couldn't inline 3944 // for lack of space 3945 for (int i = 0; i < _late_inlines.length(); i++) { 3946 CallGenerator* cg = _late_inlines.at(i); 3947 if (!cg->is_mh_late_inline()) { 3948 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 3949 if (do_print_inlining) { 3950 cg->print_inlining_late(msg); 3951 } 3952 log_late_inline_failure(cg, msg); 3953 } 3954 } 3955 } 3956 if (do_print_inlining) { 3957 ResourceMark rm; 3958 stringStream ss; 3959 for (int i = 0; i < _print_inlining_list->length(); i++) { 3960 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 3961 } 3962 size_t end = ss.size(); 3963 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 3964 strncpy(_print_inlining_output, ss.base(), end+1); 3965 _print_inlining_output[end] = 0; 3966 } 3967 } 3968 3969 void Compile::dump_print_inlining() { 3970 if (_print_inlining_output != NULL) { 3971 tty->print_raw(_print_inlining_output); 3972 } 3973 } 3974 3975 void Compile::log_late_inline(CallGenerator* cg) { 3976 if (log() != NULL) { 3977 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 3978 cg->unique_id()); 3979 JVMState* p = cg->call_node()->jvms(); 3980 while (p != NULL) { 3981 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 3982 p = p->caller(); 3983 } 3984 log()->tail("late_inline"); 3985 } 3986 } 3987 3988 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 3989 log_late_inline(cg); 3990 if (log() != NULL) { 3991 log()->inline_fail(msg); 3992 } 3993 } 3994 3995 void Compile::log_inline_id(CallGenerator* cg) { 3996 if (log() != NULL) { 3997 // The LogCompilation tool needs a unique way to identify late 3998 // inline call sites. This id must be unique for this call site in 3999 // this compilation. Try to have it unique across compilations as 4000 // well because it can be convenient when grepping through the log 4001 // file. 4002 // Distinguish OSR compilations from others in case CICountOSR is 4003 // on. 4004 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4005 cg->set_unique_id(id); 4006 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4007 } 4008 } 4009 4010 void Compile::log_inline_failure(const char* msg) { 4011 if (C->log() != NULL) { 4012 C->log()->inline_fail(msg); 4013 } 4014 } 4015 4016 4017 // Dump inlining replay data to the stream. 4018 // Don't change thread state and acquire any locks. 4019 void Compile::dump_inline_data(outputStream* out) { 4020 InlineTree* inl_tree = ilt(); 4021 if (inl_tree != NULL) { 4022 out->print(" inline %d", inl_tree->count()); 4023 inl_tree->dump_replay_data(out); 4024 } 4025 } 4026 4027 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4028 if (n1->Opcode() < n2->Opcode()) return -1; 4029 else if (n1->Opcode() > n2->Opcode()) return 1; 4030 4031 assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req())); 4032 for (uint i = 1; i < n1->req(); i++) { 4033 if (n1->in(i) < n2->in(i)) return -1; 4034 else if (n1->in(i) > n2->in(i)) return 1; 4035 } 4036 4037 return 0; 4038 } 4039 4040 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4041 Node* n1 = *n1p; 4042 Node* n2 = *n2p; 4043 4044 return cmp_expensive_nodes(n1, n2); 4045 } 4046 4047 void Compile::sort_expensive_nodes() { 4048 if (!expensive_nodes_sorted()) { 4049 _expensive_nodes->sort(cmp_expensive_nodes); 4050 } 4051 } 4052 4053 bool Compile::expensive_nodes_sorted() const { 4054 for (int i = 1; i < _expensive_nodes->length(); i++) { 4055 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4056 return false; 4057 } 4058 } 4059 return true; 4060 } 4061 4062 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4063 if (_expensive_nodes->length() == 0) { 4064 return false; 4065 } 4066 4067 assert(OptimizeExpensiveOps, "optimization off?"); 4068 4069 // Take this opportunity to remove dead nodes from the list 4070 int j = 0; 4071 for (int i = 0; i < _expensive_nodes->length(); i++) { 4072 Node* n = _expensive_nodes->at(i); 4073 if (!n->is_unreachable(igvn)) { 4074 assert(n->is_expensive(), "should be expensive"); 4075 _expensive_nodes->at_put(j, n); 4076 j++; 4077 } 4078 } 4079 _expensive_nodes->trunc_to(j); 4080 4081 // Then sort the list so that similar nodes are next to each other 4082 // and check for at least two nodes of identical kind with same data 4083 // inputs. 4084 sort_expensive_nodes(); 4085 4086 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4087 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4088 return true; 4089 } 4090 } 4091 4092 return false; 4093 } 4094 4095 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4096 if (_expensive_nodes->length() == 0) { 4097 return; 4098 } 4099 4100 assert(OptimizeExpensiveOps, "optimization off?"); 4101 4102 // Sort to bring similar nodes next to each other and clear the 4103 // control input of nodes for which there's only a single copy. 4104 sort_expensive_nodes(); 4105 4106 int j = 0; 4107 int identical = 0; 4108 int i = 0; 4109 bool modified = false; 4110 for (; i < _expensive_nodes->length()-1; i++) { 4111 assert(j <= i, "can't write beyond current index"); 4112 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4113 identical++; 4114 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4115 continue; 4116 } 4117 if (identical > 0) { 4118 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4119 identical = 0; 4120 } else { 4121 Node* n = _expensive_nodes->at(i); 4122 igvn.replace_input_of(n, 0, NULL); 4123 igvn.hash_insert(n); 4124 modified = true; 4125 } 4126 } 4127 if (identical > 0) { 4128 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4129 } else if (_expensive_nodes->length() >= 1) { 4130 Node* n = _expensive_nodes->at(i); 4131 igvn.replace_input_of(n, 0, NULL); 4132 igvn.hash_insert(n); 4133 modified = true; 4134 } 4135 _expensive_nodes->trunc_to(j); 4136 if (modified) { 4137 igvn.optimize(); 4138 } 4139 } 4140 4141 void Compile::add_expensive_node(Node * n) { 4142 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4143 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4144 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4145 if (OptimizeExpensiveOps) { 4146 _expensive_nodes->append(n); 4147 } else { 4148 // Clear control input and let IGVN optimize expensive nodes if 4149 // OptimizeExpensiveOps is off. 4150 n->set_req(0, NULL); 4151 } 4152 } 4153 4154 /** 4155 * Remove the speculative part of types and clean up the graph 4156 */ 4157 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4158 if (UseTypeSpeculation) { 4159 Unique_Node_List worklist; 4160 worklist.push(root()); 4161 int modified = 0; 4162 // Go over all type nodes that carry a speculative type, drop the 4163 // speculative part of the type and enqueue the node for an igvn 4164 // which may optimize it out. 4165 for (uint next = 0; next < worklist.size(); ++next) { 4166 Node *n = worklist.at(next); 4167 if (n->is_Type()) { 4168 TypeNode* tn = n->as_Type(); 4169 const Type* t = tn->type(); 4170 const Type* t_no_spec = t->remove_speculative(); 4171 if (t_no_spec != t) { 4172 bool in_hash = igvn.hash_delete(n); 4173 assert(in_hash, "node should be in igvn hash table"); 4174 tn->set_type(t_no_spec); 4175 igvn.hash_insert(n); 4176 igvn._worklist.push(n); // give it a chance to go away 4177 modified++; 4178 } 4179 } 4180 uint max = n->len(); 4181 for( uint i = 0; i < max; ++i ) { 4182 Node *m = n->in(i); 4183 if (not_a_node(m)) continue; 4184 worklist.push(m); 4185 } 4186 } 4187 // Drop the speculative part of all types in the igvn's type table 4188 igvn.remove_speculative_types(); 4189 if (modified > 0) { 4190 igvn.optimize(); 4191 } 4192 #ifdef ASSERT 4193 // Verify that after the IGVN is over no speculative type has resurfaced 4194 worklist.clear(); 4195 worklist.push(root()); 4196 for (uint next = 0; next < worklist.size(); ++next) { 4197 Node *n = worklist.at(next); 4198 const Type* t = igvn.type_or_null(n); 4199 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4200 if (n->is_Type()) { 4201 t = n->as_Type()->type(); 4202 assert(t == t->remove_speculative(), "no more speculative types"); 4203 } 4204 uint max = n->len(); 4205 for( uint i = 0; i < max; ++i ) { 4206 Node *m = n->in(i); 4207 if (not_a_node(m)) continue; 4208 worklist.push(m); 4209 } 4210 } 4211 igvn.check_no_speculative_types(); 4212 #endif 4213 } 4214 } 4215 4216 // Auxiliary method to support randomized stressing/fuzzing. 4217 // 4218 // This method can be called the arbitrary number of times, with current count 4219 // as the argument. The logic allows selecting a single candidate from the 4220 // running list of candidates as follows: 4221 // int count = 0; 4222 // Cand* selected = null; 4223 // while(cand = cand->next()) { 4224 // if (randomized_select(++count)) { 4225 // selected = cand; 4226 // } 4227 // } 4228 // 4229 // Including count equalizes the chances any candidate is "selected". 4230 // This is useful when we don't have the complete list of candidates to choose 4231 // from uniformly. In this case, we need to adjust the randomicity of the 4232 // selection, or else we will end up biasing the selection towards the latter 4233 // candidates. 4234 // 4235 // Quick back-envelope calculation shows that for the list of n candidates 4236 // the equal probability for the candidate to persist as "best" can be 4237 // achieved by replacing it with "next" k-th candidate with the probability 4238 // of 1/k. It can be easily shown that by the end of the run, the 4239 // probability for any candidate is converged to 1/n, thus giving the 4240 // uniform distribution among all the candidates. 4241 // 4242 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4243 #define RANDOMIZED_DOMAIN_POW 29 4244 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4245 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4246 bool Compile::randomized_select(int count) { 4247 assert(count > 0, "only positive"); 4248 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4249 }