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