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