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