1 /* 2 * Copyright (c) 1997, 2012, 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 "classfile/systemDictionary.hpp" 29 #include "code/exceptionHandlerTable.hpp" 30 #include "code/nmethod.hpp" 31 #include "compiler/compileLog.hpp" 32 #include "compiler/disassembler.hpp" 33 #include "compiler/oopMap.hpp" 34 #include "opto/addnode.hpp" 35 #include "opto/block.hpp" 36 #include "opto/c2compiler.hpp" 37 #include "opto/callGenerator.hpp" 38 #include "opto/callnode.hpp" 39 #include "opto/cfgnode.hpp" 40 #include "opto/chaitin.hpp" 41 #include "opto/compile.hpp" 42 #include "opto/connode.hpp" 43 #include "opto/divnode.hpp" 44 #include "opto/escape.hpp" 45 #include "opto/idealGraphPrinter.hpp" 46 #include "opto/loopnode.hpp" 47 #include "opto/machnode.hpp" 48 #include "opto/macro.hpp" 49 #include "opto/matcher.hpp" 50 #include "opto/memnode.hpp" 51 #include "opto/mulnode.hpp" 52 #include "opto/node.hpp" 53 #include "opto/opcodes.hpp" 54 #include "opto/output.hpp" 55 #include "opto/parse.hpp" 56 #include "opto/phaseX.hpp" 57 #include "opto/rootnode.hpp" 58 #include "opto/runtime.hpp" 59 #include "opto/stringopts.hpp" 60 #include "opto/type.hpp" 61 #include "opto/vectornode.hpp" 62 #include "runtime/arguments.hpp" 63 #include "runtime/signature.hpp" 64 #include "runtime/stubRoutines.hpp" 65 #include "runtime/timer.hpp" 66 #include "utilities/copy.hpp" 67 #ifdef TARGET_ARCH_MODEL_x86_32 68 # include "adfiles/ad_x86_32.hpp" 69 #endif 70 #ifdef TARGET_ARCH_MODEL_x86_64 71 # include "adfiles/ad_x86_64.hpp" 72 #endif 73 #ifdef TARGET_ARCH_MODEL_sparc 74 # include "adfiles/ad_sparc.hpp" 75 #endif 76 #ifdef TARGET_ARCH_MODEL_zero 77 # include "adfiles/ad_zero.hpp" 78 #endif 79 #ifdef TARGET_ARCH_MODEL_arm 80 # include "adfiles/ad_arm.hpp" 81 #endif 82 #ifdef TARGET_ARCH_MODEL_ppc 83 # include "adfiles/ad_ppc.hpp" 84 #endif 85 86 87 // -------------------- Compile::mach_constant_base_node ----------------------- 88 // Constant table base node singleton. 89 MachConstantBaseNode* Compile::mach_constant_base_node() { 90 if (_mach_constant_base_node == NULL) { 91 _mach_constant_base_node = new (C) MachConstantBaseNode(); 92 _mach_constant_base_node->add_req(C->root()); 93 } 94 return _mach_constant_base_node; 95 } 96 97 98 /// Support for intrinsics. 99 100 // Return the index at which m must be inserted (or already exists). 101 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 102 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 103 #ifdef ASSERT 104 for (int i = 1; i < _intrinsics->length(); i++) { 105 CallGenerator* cg1 = _intrinsics->at(i-1); 106 CallGenerator* cg2 = _intrinsics->at(i); 107 assert(cg1->method() != cg2->method() 108 ? cg1->method() < cg2->method() 109 : cg1->is_virtual() < cg2->is_virtual(), 110 "compiler intrinsics list must stay sorted"); 111 } 112 #endif 113 // Binary search sorted list, in decreasing intervals [lo, hi]. 114 int lo = 0, hi = _intrinsics->length()-1; 115 while (lo <= hi) { 116 int mid = (uint)(hi + lo) / 2; 117 ciMethod* mid_m = _intrinsics->at(mid)->method(); 118 if (m < mid_m) { 119 hi = mid-1; 120 } else if (m > mid_m) { 121 lo = mid+1; 122 } else { 123 // look at minor sort key 124 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 125 if (is_virtual < mid_virt) { 126 hi = mid-1; 127 } else if (is_virtual > mid_virt) { 128 lo = mid+1; 129 } else { 130 return mid; // exact match 131 } 132 } 133 } 134 return lo; // inexact match 135 } 136 137 void Compile::register_intrinsic(CallGenerator* cg) { 138 if (_intrinsics == NULL) { 139 _intrinsics = new GrowableArray<CallGenerator*>(60); 140 } 141 // This code is stolen from ciObjectFactory::insert. 142 // Really, GrowableArray should have methods for 143 // insert_at, remove_at, and binary_search. 144 int len = _intrinsics->length(); 145 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 146 if (index == len) { 147 _intrinsics->append(cg); 148 } else { 149 #ifdef ASSERT 150 CallGenerator* oldcg = _intrinsics->at(index); 151 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 152 #endif 153 _intrinsics->append(_intrinsics->at(len-1)); 154 int pos; 155 for (pos = len-2; pos >= index; pos--) { 156 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 157 } 158 _intrinsics->at_put(index, cg); 159 } 160 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 161 } 162 163 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 164 assert(m->is_loaded(), "don't try this on unloaded methods"); 165 if (_intrinsics != NULL) { 166 int index = intrinsic_insertion_index(m, is_virtual); 167 if (index < _intrinsics->length() 168 && _intrinsics->at(index)->method() == m 169 && _intrinsics->at(index)->is_virtual() == is_virtual) { 170 return _intrinsics->at(index); 171 } 172 } 173 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 174 if (m->intrinsic_id() != vmIntrinsics::_none && 175 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 176 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 177 if (cg != NULL) { 178 // Save it for next time: 179 register_intrinsic(cg); 180 return cg; 181 } else { 182 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 183 } 184 } 185 return NULL; 186 } 187 188 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 189 // in library_call.cpp. 190 191 192 #ifndef PRODUCT 193 // statistics gathering... 194 195 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 196 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 197 198 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 199 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 200 int oflags = _intrinsic_hist_flags[id]; 201 assert(flags != 0, "what happened?"); 202 if (is_virtual) { 203 flags |= _intrinsic_virtual; 204 } 205 bool changed = (flags != oflags); 206 if ((flags & _intrinsic_worked) != 0) { 207 juint count = (_intrinsic_hist_count[id] += 1); 208 if (count == 1) { 209 changed = true; // first time 210 } 211 // increment the overall count also: 212 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 213 } 214 if (changed) { 215 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 216 // Something changed about the intrinsic's virtuality. 217 if ((flags & _intrinsic_virtual) != 0) { 218 // This is the first use of this intrinsic as a virtual call. 219 if (oflags != 0) { 220 // We already saw it as a non-virtual, so note both cases. 221 flags |= _intrinsic_both; 222 } 223 } else if ((oflags & _intrinsic_both) == 0) { 224 // This is the first use of this intrinsic as a non-virtual 225 flags |= _intrinsic_both; 226 } 227 } 228 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 229 } 230 // update the overall flags also: 231 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 232 return changed; 233 } 234 235 static char* format_flags(int flags, char* buf) { 236 buf[0] = 0; 237 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 238 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 239 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 240 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 241 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 242 if (buf[0] == 0) strcat(buf, ","); 243 assert(buf[0] == ',', "must be"); 244 return &buf[1]; 245 } 246 247 void Compile::print_intrinsic_statistics() { 248 char flagsbuf[100]; 249 ttyLocker ttyl; 250 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 251 tty->print_cr("Compiler intrinsic usage:"); 252 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 253 if (total == 0) total = 1; // avoid div0 in case of no successes 254 #define PRINT_STAT_LINE(name, c, f) \ 255 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 256 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 257 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 258 int flags = _intrinsic_hist_flags[id]; 259 juint count = _intrinsic_hist_count[id]; 260 if ((flags | count) != 0) { 261 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 262 } 263 } 264 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 265 if (xtty != NULL) xtty->tail("statistics"); 266 } 267 268 void Compile::print_statistics() { 269 { ttyLocker ttyl; 270 if (xtty != NULL) xtty->head("statistics type='opto'"); 271 Parse::print_statistics(); 272 PhaseCCP::print_statistics(); 273 PhaseRegAlloc::print_statistics(); 274 Scheduling::print_statistics(); 275 PhasePeephole::print_statistics(); 276 PhaseIdealLoop::print_statistics(); 277 if (xtty != NULL) xtty->tail("statistics"); 278 } 279 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 280 // put this under its own <statistics> element. 281 print_intrinsic_statistics(); 282 } 283 } 284 #endif //PRODUCT 285 286 // Support for bundling info 287 Bundle* Compile::node_bundling(const Node *n) { 288 assert(valid_bundle_info(n), "oob"); 289 return &_node_bundling_base[n->_idx]; 290 } 291 292 bool Compile::valid_bundle_info(const Node *n) { 293 return (_node_bundling_limit > n->_idx); 294 } 295 296 297 void Compile::gvn_replace_by(Node* n, Node* nn) { 298 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 299 Node* use = n->last_out(i); 300 bool is_in_table = initial_gvn()->hash_delete(use); 301 uint uses_found = 0; 302 for (uint j = 0; j < use->len(); j++) { 303 if (use->in(j) == n) { 304 if (j < use->req()) 305 use->set_req(j, nn); 306 else 307 use->set_prec(j, nn); 308 uses_found++; 309 } 310 } 311 if (is_in_table) { 312 // reinsert into table 313 initial_gvn()->hash_find_insert(use); 314 } 315 record_for_igvn(use); 316 i -= uses_found; // we deleted 1 or more copies of this edge 317 } 318 } 319 320 321 static inline bool not_a_node(const Node* n) { 322 if (n == NULL) return true; 323 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 324 if (*(address*)n == badAddress) return true; // kill by Node::destruct 325 return false; 326 } 327 328 // Identify all nodes that are reachable from below, useful. 329 // Use breadth-first pass that records state in a Unique_Node_List, 330 // recursive traversal is slower. 331 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 332 int estimated_worklist_size = unique(); 333 useful.map( estimated_worklist_size, NULL ); // preallocate space 334 335 // Initialize worklist 336 if (root() != NULL) { useful.push(root()); } 337 // If 'top' is cached, declare it useful to preserve cached node 338 if( cached_top_node() ) { useful.push(cached_top_node()); } 339 340 // Push all useful nodes onto the list, breadthfirst 341 for( uint next = 0; next < useful.size(); ++next ) { 342 assert( next < unique(), "Unique useful nodes < total nodes"); 343 Node *n = useful.at(next); 344 uint max = n->len(); 345 for( uint i = 0; i < max; ++i ) { 346 Node *m = n->in(i); 347 if (not_a_node(m)) continue; 348 useful.push(m); 349 } 350 } 351 } 352 353 // Update dead_node_list with any missing dead nodes using useful 354 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 355 void Compile::update_dead_node_list(Unique_Node_List &useful) { 356 uint max_idx = unique(); 357 VectorSet& useful_node_set = useful.member_set(); 358 359 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 360 // If node with index node_idx is not in useful set, 361 // mark it as dead in dead node list. 362 if (! useful_node_set.test(node_idx) ) { 363 record_dead_node(node_idx); 364 } 365 } 366 } 367 368 // Disconnect all useless nodes by disconnecting those at the boundary. 369 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 370 uint next = 0; 371 while (next < useful.size()) { 372 Node *n = useful.at(next++); 373 // Use raw traversal of out edges since this code removes out edges 374 int max = n->outcnt(); 375 for (int j = 0; j < max; ++j) { 376 Node* child = n->raw_out(j); 377 if (! useful.member(child)) { 378 assert(!child->is_top() || child != top(), 379 "If top is cached in Compile object it is in useful list"); 380 // Only need to remove this out-edge to the useless node 381 n->raw_del_out(j); 382 --j; 383 --max; 384 } 385 } 386 if (n->outcnt() == 1 && n->has_special_unique_user()) { 387 record_for_igvn(n->unique_out()); 388 } 389 } 390 // Remove useless macro and predicate opaq nodes 391 for (int i = C->macro_count()-1; i >= 0; i--) { 392 Node* n = C->macro_node(i); 393 if (!useful.member(n)) { 394 remove_macro_node(n); 395 } 396 } 397 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 398 } 399 400 //------------------------------frame_size_in_words----------------------------- 401 // frame_slots in units of words 402 int Compile::frame_size_in_words() const { 403 // shift is 0 in LP32 and 1 in LP64 404 const int shift = (LogBytesPerWord - LogBytesPerInt); 405 int words = _frame_slots >> shift; 406 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 407 return words; 408 } 409 410 // ============================================================================ 411 //------------------------------CompileWrapper--------------------------------- 412 class CompileWrapper : public StackObj { 413 Compile *const _compile; 414 public: 415 CompileWrapper(Compile* compile); 416 417 ~CompileWrapper(); 418 }; 419 420 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 421 // the Compile* pointer is stored in the current ciEnv: 422 ciEnv* env = compile->env(); 423 assert(env == ciEnv::current(), "must already be a ciEnv active"); 424 assert(env->compiler_data() == NULL, "compile already active?"); 425 env->set_compiler_data(compile); 426 assert(compile == Compile::current(), "sanity"); 427 428 compile->set_type_dict(NULL); 429 compile->set_type_hwm(NULL); 430 compile->set_type_last_size(0); 431 compile->set_last_tf(NULL, NULL); 432 compile->set_indexSet_arena(NULL); 433 compile->set_indexSet_free_block_list(NULL); 434 compile->init_type_arena(); 435 Type::Initialize(compile); 436 _compile->set_scratch_buffer_blob(NULL); 437 _compile->begin_method(); 438 } 439 CompileWrapper::~CompileWrapper() { 440 _compile->end_method(); 441 if (_compile->scratch_buffer_blob() != NULL) 442 BufferBlob::free(_compile->scratch_buffer_blob()); 443 _compile->env()->set_compiler_data(NULL); 444 } 445 446 447 //----------------------------print_compile_messages--------------------------- 448 void Compile::print_compile_messages() { 449 #ifndef PRODUCT 450 // Check if recompiling 451 if (_subsume_loads == false && PrintOpto) { 452 // Recompiling without allowing machine instructions to subsume loads 453 tty->print_cr("*********************************************************"); 454 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 455 tty->print_cr("*********************************************************"); 456 } 457 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 458 // Recompiling without escape analysis 459 tty->print_cr("*********************************************************"); 460 tty->print_cr("** Bailout: Recompile without escape analysis **"); 461 tty->print_cr("*********************************************************"); 462 } 463 if (env()->break_at_compile()) { 464 // Open the debugger when compiling this method. 465 tty->print("### Breaking when compiling: "); 466 method()->print_short_name(); 467 tty->cr(); 468 BREAKPOINT; 469 } 470 471 if( PrintOpto ) { 472 if (is_osr_compilation()) { 473 tty->print("[OSR]%3d", _compile_id); 474 } else { 475 tty->print("%3d", _compile_id); 476 } 477 } 478 #endif 479 } 480 481 482 //-----------------------init_scratch_buffer_blob------------------------------ 483 // Construct a temporary BufferBlob and cache it for this compile. 484 void Compile::init_scratch_buffer_blob(int const_size) { 485 // If there is already a scratch buffer blob allocated and the 486 // constant section is big enough, use it. Otherwise free the 487 // current and allocate a new one. 488 BufferBlob* blob = scratch_buffer_blob(); 489 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 490 // Use the current blob. 491 } else { 492 if (blob != NULL) { 493 BufferBlob::free(blob); 494 } 495 496 ResourceMark rm; 497 _scratch_const_size = const_size; 498 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 499 blob = BufferBlob::create("Compile::scratch_buffer", size); 500 // Record the buffer blob for next time. 501 set_scratch_buffer_blob(blob); 502 // Have we run out of code space? 503 if (scratch_buffer_blob() == NULL) { 504 // Let CompilerBroker disable further compilations. 505 record_failure("Not enough space for scratch buffer in CodeCache"); 506 return; 507 } 508 } 509 510 // Initialize the relocation buffers 511 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 512 set_scratch_locs_memory(locs_buf); 513 } 514 515 516 //-----------------------scratch_emit_size------------------------------------- 517 // Helper function that computes size by emitting code 518 uint Compile::scratch_emit_size(const Node* n) { 519 // Start scratch_emit_size section. 520 set_in_scratch_emit_size(true); 521 522 // Emit into a trash buffer and count bytes emitted. 523 // This is a pretty expensive way to compute a size, 524 // but it works well enough if seldom used. 525 // All common fixed-size instructions are given a size 526 // method by the AD file. 527 // Note that the scratch buffer blob and locs memory are 528 // allocated at the beginning of the compile task, and 529 // may be shared by several calls to scratch_emit_size. 530 // The allocation of the scratch buffer blob is particularly 531 // expensive, since it has to grab the code cache lock. 532 BufferBlob* blob = this->scratch_buffer_blob(); 533 assert(blob != NULL, "Initialize BufferBlob at start"); 534 assert(blob->size() > MAX_inst_size, "sanity"); 535 relocInfo* locs_buf = scratch_locs_memory(); 536 address blob_begin = blob->content_begin(); 537 address blob_end = (address)locs_buf; 538 assert(blob->content_contains(blob_end), "sanity"); 539 CodeBuffer buf(blob_begin, blob_end - blob_begin); 540 buf.initialize_consts_size(_scratch_const_size); 541 buf.initialize_stubs_size(MAX_stubs_size); 542 assert(locs_buf != NULL, "sanity"); 543 int lsize = MAX_locs_size / 3; 544 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 545 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 546 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 547 548 // Do the emission. 549 550 Label fakeL; // Fake label for branch instructions. 551 Label* saveL = NULL; 552 uint save_bnum = 0; 553 bool is_branch = n->is_MachBranch(); 554 if (is_branch) { 555 MacroAssembler masm(&buf); 556 masm.bind(fakeL); 557 n->as_MachBranch()->save_label(&saveL, &save_bnum); 558 n->as_MachBranch()->label_set(&fakeL, 0); 559 } 560 n->emit(buf, this->regalloc()); 561 if (is_branch) // Restore label. 562 n->as_MachBranch()->label_set(saveL, save_bnum); 563 564 // End scratch_emit_size section. 565 set_in_scratch_emit_size(false); 566 567 return buf.insts_size(); 568 } 569 570 571 // ============================================================================ 572 //------------------------------Compile standard------------------------------- 573 debug_only( int Compile::_debug_idx = 100000; ) 574 575 // Compile a method. entry_bci is -1 for normal compilations and indicates 576 // the continuation bci for on stack replacement. 577 578 579 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 580 : Phase(Compiler), 581 _env(ci_env), 582 _log(ci_env->log()), 583 _compile_id(ci_env->compile_id()), 584 _save_argument_registers(false), 585 _stub_name(NULL), 586 _stub_function(NULL), 587 _stub_entry_point(NULL), 588 _method(target), 589 _entry_bci(osr_bci), 590 _initial_gvn(NULL), 591 _for_igvn(NULL), 592 _warm_calls(NULL), 593 _subsume_loads(subsume_loads), 594 _do_escape_analysis(do_escape_analysis), 595 _failure_reason(NULL), 596 _code_buffer("Compile::Fill_buffer"), 597 _orig_pc_slot(0), 598 _orig_pc_slot_offset_in_bytes(0), 599 _has_method_handle_invokes(false), 600 _mach_constant_base_node(NULL), 601 _node_bundling_limit(0), 602 _node_bundling_base(NULL), 603 _java_calls(0), 604 _inner_loops(0), 605 _scratch_const_size(-1), 606 _in_scratch_emit_size(false), 607 _dead_node_list(comp_arena()), 608 _dead_node_count(0), 609 #ifndef PRODUCT 610 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 611 _printer(IdealGraphPrinter::printer()), 612 #endif 613 _congraph(NULL) { 614 C = this; 615 616 CompileWrapper cw(this); 617 #ifndef PRODUCT 618 if (TimeCompiler2) { 619 tty->print(" "); 620 target->holder()->name()->print(); 621 tty->print("."); 622 target->print_short_name(); 623 tty->print(" "); 624 } 625 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 626 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 627 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 628 if (!print_opto_assembly) { 629 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 630 if (print_assembly && !Disassembler::can_decode()) { 631 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 632 print_opto_assembly = true; 633 } 634 } 635 set_print_assembly(print_opto_assembly); 636 set_parsed_irreducible_loop(false); 637 #endif 638 639 if (ProfileTraps) { 640 // Make sure the method being compiled gets its own MDO, 641 // so we can at least track the decompile_count(). 642 method()->ensure_method_data(); 643 } 644 645 Init(::AliasLevel); 646 647 648 print_compile_messages(); 649 650 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 651 _ilt = InlineTree::build_inline_tree_root(); 652 else 653 _ilt = NULL; 654 655 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 656 assert(num_alias_types() >= AliasIdxRaw, ""); 657 658 #define MINIMUM_NODE_HASH 1023 659 // Node list that Iterative GVN will start with 660 Unique_Node_List for_igvn(comp_arena()); 661 set_for_igvn(&for_igvn); 662 663 // GVN that will be run immediately on new nodes 664 uint estimated_size = method()->code_size()*4+64; 665 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 666 PhaseGVN gvn(node_arena(), estimated_size); 667 set_initial_gvn(&gvn); 668 669 { // Scope for timing the parser 670 TracePhase t3("parse", &_t_parser, true); 671 672 // Put top into the hash table ASAP. 673 initial_gvn()->transform_no_reclaim(top()); 674 675 // Set up tf(), start(), and find a CallGenerator. 676 CallGenerator* cg = NULL; 677 if (is_osr_compilation()) { 678 const TypeTuple *domain = StartOSRNode::osr_domain(); 679 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 680 init_tf(TypeFunc::make(domain, range)); 681 StartNode* s = new (this) StartOSRNode(root(), domain); 682 initial_gvn()->set_type_bottom(s); 683 init_start(s); 684 cg = CallGenerator::for_osr(method(), entry_bci()); 685 } else { 686 // Normal case. 687 init_tf(TypeFunc::make(method())); 688 StartNode* s = new (this) StartNode(root(), tf()->domain()); 689 initial_gvn()->set_type_bottom(s); 690 init_start(s); 691 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 692 // With java.lang.ref.reference.get() we must go through the 693 // intrinsic when G1 is enabled - even when get() is the root 694 // method of the compile - so that, if necessary, the value in 695 // the referent field of the reference object gets recorded by 696 // the pre-barrier code. 697 // Specifically, if G1 is enabled, the value in the referent 698 // field is recorded by the G1 SATB pre barrier. This will 699 // result in the referent being marked live and the reference 700 // object removed from the list of discovered references during 701 // reference processing. 702 cg = find_intrinsic(method(), false); 703 } 704 if (cg == NULL) { 705 float past_uses = method()->interpreter_invocation_count(); 706 float expected_uses = past_uses; 707 cg = CallGenerator::for_inline(method(), expected_uses); 708 } 709 } 710 if (failing()) return; 711 if (cg == NULL) { 712 record_method_not_compilable_all_tiers("cannot parse method"); 713 return; 714 } 715 JVMState* jvms = build_start_state(start(), tf()); 716 if ((jvms = cg->generate(jvms)) == NULL) { 717 record_method_not_compilable("method parse failed"); 718 return; 719 } 720 GraphKit kit(jvms); 721 722 if (!kit.stopped()) { 723 // Accept return values, and transfer control we know not where. 724 // This is done by a special, unique ReturnNode bound to root. 725 return_values(kit.jvms()); 726 } 727 728 if (kit.has_exceptions()) { 729 // Any exceptions that escape from this call must be rethrown 730 // to whatever caller is dynamically above us on the stack. 731 // This is done by a special, unique RethrowNode bound to root. 732 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 733 } 734 735 if (!failing() && has_stringbuilder()) { 736 { 737 // remove useless nodes to make the usage analysis simpler 738 ResourceMark rm; 739 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 740 } 741 742 { 743 ResourceMark rm; 744 print_method("Before StringOpts", 3); 745 PhaseStringOpts pso(initial_gvn(), &for_igvn); 746 print_method("After StringOpts", 3); 747 } 748 749 // now inline anything that we skipped the first time around 750 while (_late_inlines.length() > 0) { 751 CallGenerator* cg = _late_inlines.pop(); 752 cg->do_late_inline(); 753 if (failing()) return; 754 } 755 } 756 assert(_late_inlines.length() == 0, "should have been processed"); 757 758 print_method("Before RemoveUseless", 3); 759 760 // Remove clutter produced by parsing. 761 if (!failing()) { 762 ResourceMark rm; 763 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 764 } 765 } 766 767 // Note: Large methods are capped off in do_one_bytecode(). 768 if (failing()) return; 769 770 // After parsing, node notes are no longer automagic. 771 // They must be propagated by register_new_node_with_optimizer(), 772 // clone(), or the like. 773 set_default_node_notes(NULL); 774 775 for (;;) { 776 int successes = Inline_Warm(); 777 if (failing()) return; 778 if (successes == 0) break; 779 } 780 781 // Drain the list. 782 Finish_Warm(); 783 #ifndef PRODUCT 784 if (_printer) { 785 _printer->print_inlining(this); 786 } 787 #endif 788 789 if (failing()) return; 790 NOT_PRODUCT( verify_graph_edges(); ) 791 792 // Now optimize 793 Optimize(); 794 if (failing()) return; 795 NOT_PRODUCT( verify_graph_edges(); ) 796 797 #ifndef PRODUCT 798 if (PrintIdeal) { 799 ttyLocker ttyl; // keep the following output all in one block 800 // This output goes directly to the tty, not the compiler log. 801 // To enable tools to match it up with the compilation activity, 802 // be sure to tag this tty output with the compile ID. 803 if (xtty != NULL) { 804 xtty->head("ideal compile_id='%d'%s", compile_id(), 805 is_osr_compilation() ? " compile_kind='osr'" : 806 ""); 807 } 808 root()->dump(9999); 809 if (xtty != NULL) { 810 xtty->tail("ideal"); 811 } 812 } 813 #endif 814 815 // Now that we know the size of all the monitors we can add a fixed slot 816 // for the original deopt pc. 817 818 _orig_pc_slot = fixed_slots(); 819 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 820 set_fixed_slots(next_slot); 821 822 // Now generate code 823 Code_Gen(); 824 if (failing()) return; 825 826 // Check if we want to skip execution of all compiled code. 827 { 828 #ifndef PRODUCT 829 if (OptoNoExecute) { 830 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 831 return; 832 } 833 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 834 #endif 835 836 if (is_osr_compilation()) { 837 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 838 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 839 } else { 840 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 841 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 842 } 843 844 env()->register_method(_method, _entry_bci, 845 &_code_offsets, 846 _orig_pc_slot_offset_in_bytes, 847 code_buffer(), 848 frame_size_in_words(), _oop_map_set, 849 &_handler_table, &_inc_table, 850 compiler, 851 env()->comp_level(), 852 has_unsafe_access(), 853 SharedRuntime::is_wide_vector(max_vector_size()) 854 ); 855 856 if (log() != NULL) // Print code cache state into compiler log 857 log()->code_cache_state(); 858 } 859 } 860 861 //------------------------------Compile---------------------------------------- 862 // Compile a runtime stub 863 Compile::Compile( ciEnv* ci_env, 864 TypeFunc_generator generator, 865 address stub_function, 866 const char *stub_name, 867 int is_fancy_jump, 868 bool pass_tls, 869 bool save_arg_registers, 870 bool return_pc ) 871 : Phase(Compiler), 872 _env(ci_env), 873 _log(ci_env->log()), 874 _compile_id(-1), 875 _save_argument_registers(save_arg_registers), 876 _method(NULL), 877 _stub_name(stub_name), 878 _stub_function(stub_function), 879 _stub_entry_point(NULL), 880 _entry_bci(InvocationEntryBci), 881 _initial_gvn(NULL), 882 _for_igvn(NULL), 883 _warm_calls(NULL), 884 _orig_pc_slot(0), 885 _orig_pc_slot_offset_in_bytes(0), 886 _subsume_loads(true), 887 _do_escape_analysis(false), 888 _failure_reason(NULL), 889 _code_buffer("Compile::Fill_buffer"), 890 _has_method_handle_invokes(false), 891 _mach_constant_base_node(NULL), 892 _node_bundling_limit(0), 893 _node_bundling_base(NULL), 894 _java_calls(0), 895 _inner_loops(0), 896 #ifndef PRODUCT 897 _trace_opto_output(TraceOptoOutput), 898 _printer(NULL), 899 #endif 900 _dead_node_list(comp_arena()), 901 _dead_node_count(0), 902 _congraph(NULL) { 903 C = this; 904 905 #ifndef PRODUCT 906 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 907 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 908 set_print_assembly(PrintFrameConverterAssembly); 909 set_parsed_irreducible_loop(false); 910 #endif 911 CompileWrapper cw(this); 912 Init(/*AliasLevel=*/ 0); 913 init_tf((*generator)()); 914 915 { 916 // The following is a dummy for the sake of GraphKit::gen_stub 917 Unique_Node_List for_igvn(comp_arena()); 918 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 919 PhaseGVN gvn(Thread::current()->resource_area(),255); 920 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 921 gvn.transform_no_reclaim(top()); 922 923 GraphKit kit; 924 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 925 } 926 927 NOT_PRODUCT( verify_graph_edges(); ) 928 Code_Gen(); 929 if (failing()) return; 930 931 932 // Entry point will be accessed using compile->stub_entry_point(); 933 if (code_buffer() == NULL) { 934 Matcher::soft_match_failure(); 935 } else { 936 if (PrintAssembly && (WizardMode || Verbose)) 937 tty->print_cr("### Stub::%s", stub_name); 938 939 if (!failing()) { 940 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 941 942 // Make the NMethod 943 // For now we mark the frame as never safe for profile stackwalking 944 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 945 code_buffer(), 946 CodeOffsets::frame_never_safe, 947 // _code_offsets.value(CodeOffsets::Frame_Complete), 948 frame_size_in_words(), 949 _oop_map_set, 950 save_arg_registers); 951 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 952 953 _stub_entry_point = rs->entry_point(); 954 } 955 } 956 } 957 958 //------------------------------Init------------------------------------------- 959 // Prepare for a single compilation 960 void Compile::Init(int aliaslevel) { 961 _unique = 0; 962 _regalloc = NULL; 963 964 _tf = NULL; // filled in later 965 _top = NULL; // cached later 966 _matcher = NULL; // filled in later 967 _cfg = NULL; // filled in later 968 969 set_24_bit_selection_and_mode(Use24BitFP, false); 970 971 _node_note_array = NULL; 972 _default_node_notes = NULL; 973 974 _immutable_memory = NULL; // filled in at first inquiry 975 976 // Globally visible Nodes 977 // First set TOP to NULL to give safe behavior during creation of RootNode 978 set_cached_top_node(NULL); 979 set_root(new (this) RootNode()); 980 // Now that you have a Root to point to, create the real TOP 981 set_cached_top_node( new (this) ConNode(Type::TOP) ); 982 set_recent_alloc(NULL, NULL); 983 984 // Create Debug Information Recorder to record scopes, oopmaps, etc. 985 env()->set_oop_recorder(new OopRecorder(env()->arena())); 986 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 987 env()->set_dependencies(new Dependencies(env())); 988 989 _fixed_slots = 0; 990 set_has_split_ifs(false); 991 set_has_loops(has_method() && method()->has_loops()); // first approximation 992 set_has_stringbuilder(false); 993 _trap_can_recompile = false; // no traps emitted yet 994 _major_progress = true; // start out assuming good things will happen 995 set_has_unsafe_access(false); 996 set_max_vector_size(0); 997 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 998 set_decompile_count(0); 999 1000 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1001 set_num_loop_opts(LoopOptsCount); 1002 set_do_inlining(Inline); 1003 set_max_inline_size(MaxInlineSize); 1004 set_freq_inline_size(FreqInlineSize); 1005 set_do_scheduling(OptoScheduling); 1006 set_do_count_invocations(false); 1007 set_do_method_data_update(false); 1008 1009 if (debug_info()->recording_non_safepoints()) { 1010 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1011 (comp_arena(), 8, 0, NULL)); 1012 set_default_node_notes(Node_Notes::make(this)); 1013 } 1014 1015 // // -- Initialize types before each compile -- 1016 // // Update cached type information 1017 // if( _method && _method->constants() ) 1018 // Type::update_loaded_types(_method, _method->constants()); 1019 1020 // Init alias_type map. 1021 if (!_do_escape_analysis && aliaslevel == 3) 1022 aliaslevel = 2; // No unique types without escape analysis 1023 _AliasLevel = aliaslevel; 1024 const int grow_ats = 16; 1025 _max_alias_types = grow_ats; 1026 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1027 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1028 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1029 { 1030 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1031 } 1032 // Initialize the first few types. 1033 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1034 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1035 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1036 _num_alias_types = AliasIdxRaw+1; 1037 // Zero out the alias type cache. 1038 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1039 // A NULL adr_type hits in the cache right away. Preload the right answer. 1040 probe_alias_cache(NULL)->_index = AliasIdxTop; 1041 1042 _intrinsics = NULL; 1043 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1044 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1045 register_library_intrinsics(); 1046 } 1047 1048 //---------------------------init_start---------------------------------------- 1049 // Install the StartNode on this compile object. 1050 void Compile::init_start(StartNode* s) { 1051 if (failing()) 1052 return; // already failing 1053 assert(s == start(), ""); 1054 } 1055 1056 StartNode* Compile::start() const { 1057 assert(!failing(), ""); 1058 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1059 Node* start = root()->fast_out(i); 1060 if( start->is_Start() ) 1061 return start->as_Start(); 1062 } 1063 ShouldNotReachHere(); 1064 return NULL; 1065 } 1066 1067 //-------------------------------immutable_memory------------------------------------- 1068 // Access immutable memory 1069 Node* Compile::immutable_memory() { 1070 if (_immutable_memory != NULL) { 1071 return _immutable_memory; 1072 } 1073 StartNode* s = start(); 1074 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1075 Node *p = s->fast_out(i); 1076 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1077 _immutable_memory = p; 1078 return _immutable_memory; 1079 } 1080 } 1081 ShouldNotReachHere(); 1082 return NULL; 1083 } 1084 1085 //----------------------set_cached_top_node------------------------------------ 1086 // Install the cached top node, and make sure Node::is_top works correctly. 1087 void Compile::set_cached_top_node(Node* tn) { 1088 if (tn != NULL) verify_top(tn); 1089 Node* old_top = _top; 1090 _top = tn; 1091 // Calling Node::setup_is_top allows the nodes the chance to adjust 1092 // their _out arrays. 1093 if (_top != NULL) _top->setup_is_top(); 1094 if (old_top != NULL) old_top->setup_is_top(); 1095 assert(_top == NULL || top()->is_top(), ""); 1096 } 1097 1098 #ifdef ASSERT 1099 uint Compile::count_live_nodes_by_graph_walk() { 1100 Unique_Node_List useful(comp_arena()); 1101 // Get useful node list by walking the graph. 1102 identify_useful_nodes(useful); 1103 return useful.size(); 1104 } 1105 1106 void Compile::print_missing_nodes() { 1107 1108 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1109 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1110 return; 1111 } 1112 1113 // This is an expensive function. It is executed only when the user 1114 // specifies VerifyIdealNodeCount option or otherwise knows the 1115 // additional work that needs to be done to identify reachable nodes 1116 // by walking the flow graph and find the missing ones using 1117 // _dead_node_list. 1118 1119 Unique_Node_List useful(comp_arena()); 1120 // Get useful node list by walking the graph. 1121 identify_useful_nodes(useful); 1122 1123 uint l_nodes = C->live_nodes(); 1124 uint l_nodes_by_walk = useful.size(); 1125 1126 if (l_nodes != l_nodes_by_walk) { 1127 if (_log != NULL) { 1128 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1129 _log->stamp(); 1130 _log->end_head(); 1131 } 1132 VectorSet& useful_member_set = useful.member_set(); 1133 int last_idx = l_nodes_by_walk; 1134 for (int i = 0; i < last_idx; i++) { 1135 if (useful_member_set.test(i)) { 1136 if (_dead_node_list.test(i)) { 1137 if (_log != NULL) { 1138 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1139 } 1140 if (PrintIdealNodeCount) { 1141 // Print the log message to tty 1142 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1143 useful.at(i)->dump(); 1144 } 1145 } 1146 } 1147 else if (! _dead_node_list.test(i)) { 1148 if (_log != NULL) { 1149 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1150 } 1151 if (PrintIdealNodeCount) { 1152 // Print the log message to tty 1153 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1154 } 1155 } 1156 } 1157 if (_log != NULL) { 1158 _log->tail("mismatched_nodes"); 1159 } 1160 } 1161 } 1162 #endif 1163 1164 #ifndef PRODUCT 1165 void Compile::verify_top(Node* tn) const { 1166 if (tn != NULL) { 1167 assert(tn->is_Con(), "top node must be a constant"); 1168 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1169 assert(tn->in(0) != NULL, "must have live top node"); 1170 } 1171 } 1172 #endif 1173 1174 1175 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1176 1177 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1178 guarantee(arr != NULL, ""); 1179 int num_blocks = arr->length(); 1180 if (grow_by < num_blocks) grow_by = num_blocks; 1181 int num_notes = grow_by * _node_notes_block_size; 1182 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1183 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1184 while (num_notes > 0) { 1185 arr->append(notes); 1186 notes += _node_notes_block_size; 1187 num_notes -= _node_notes_block_size; 1188 } 1189 assert(num_notes == 0, "exact multiple, please"); 1190 } 1191 1192 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1193 if (source == NULL || dest == NULL) return false; 1194 1195 if (dest->is_Con()) 1196 return false; // Do not push debug info onto constants. 1197 1198 #ifdef ASSERT 1199 // Leave a bread crumb trail pointing to the original node: 1200 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1201 dest->set_debug_orig(source); 1202 } 1203 #endif 1204 1205 if (node_note_array() == NULL) 1206 return false; // Not collecting any notes now. 1207 1208 // This is a copy onto a pre-existing node, which may already have notes. 1209 // If both nodes have notes, do not overwrite any pre-existing notes. 1210 Node_Notes* source_notes = node_notes_at(source->_idx); 1211 if (source_notes == NULL || source_notes->is_clear()) return false; 1212 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1213 if (dest_notes == NULL || dest_notes->is_clear()) { 1214 return set_node_notes_at(dest->_idx, source_notes); 1215 } 1216 1217 Node_Notes merged_notes = (*source_notes); 1218 // The order of operations here ensures that dest notes will win... 1219 merged_notes.update_from(dest_notes); 1220 return set_node_notes_at(dest->_idx, &merged_notes); 1221 } 1222 1223 1224 //--------------------------allow_range_check_smearing------------------------- 1225 // Gating condition for coalescing similar range checks. 1226 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1227 // single covering check that is at least as strong as any of them. 1228 // If the optimization succeeds, the simplified (strengthened) range check 1229 // will always succeed. If it fails, we will deopt, and then give up 1230 // on the optimization. 1231 bool Compile::allow_range_check_smearing() const { 1232 // If this method has already thrown a range-check, 1233 // assume it was because we already tried range smearing 1234 // and it failed. 1235 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1236 return !already_trapped; 1237 } 1238 1239 1240 //------------------------------flatten_alias_type----------------------------- 1241 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1242 int offset = tj->offset(); 1243 TypePtr::PTR ptr = tj->ptr(); 1244 1245 // Known instance (scalarizable allocation) alias only with itself. 1246 bool is_known_inst = tj->isa_oopptr() != NULL && 1247 tj->is_oopptr()->is_known_instance(); 1248 1249 // Process weird unsafe references. 1250 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1251 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1252 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1253 tj = TypeOopPtr::BOTTOM; 1254 ptr = tj->ptr(); 1255 offset = tj->offset(); 1256 } 1257 1258 // Array pointers need some flattening 1259 const TypeAryPtr *ta = tj->isa_aryptr(); 1260 if( ta && is_known_inst ) { 1261 if ( offset != Type::OffsetBot && 1262 offset > arrayOopDesc::length_offset_in_bytes() ) { 1263 offset = Type::OffsetBot; // Flatten constant access into array body only 1264 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1265 } 1266 } else if( ta && _AliasLevel >= 2 ) { 1267 // For arrays indexed by constant indices, we flatten the alias 1268 // space to include all of the array body. Only the header, klass 1269 // and array length can be accessed un-aliased. 1270 if( offset != Type::OffsetBot ) { 1271 if( ta->const_oop() ) { // MethodData* or Method* 1272 offset = Type::OffsetBot; // Flatten constant access into array body 1273 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1274 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1275 // range is OK as-is. 1276 tj = ta = TypeAryPtr::RANGE; 1277 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1278 tj = TypeInstPtr::KLASS; // all klass loads look alike 1279 ta = TypeAryPtr::RANGE; // generic ignored junk 1280 ptr = TypePtr::BotPTR; 1281 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1282 tj = TypeInstPtr::MARK; 1283 ta = TypeAryPtr::RANGE; // generic ignored junk 1284 ptr = TypePtr::BotPTR; 1285 } else { // Random constant offset into array body 1286 offset = Type::OffsetBot; // Flatten constant access into array body 1287 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1288 } 1289 } 1290 // Arrays of fixed size alias with arrays of unknown size. 1291 if (ta->size() != TypeInt::POS) { 1292 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1293 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1294 } 1295 // Arrays of known objects become arrays of unknown objects. 1296 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1297 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1298 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1299 } 1300 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1301 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1302 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1303 } 1304 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1305 // cannot be distinguished by bytecode alone. 1306 if (ta->elem() == TypeInt::BOOL) { 1307 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1308 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1309 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1310 } 1311 // During the 2nd round of IterGVN, NotNull castings are removed. 1312 // Make sure the Bottom and NotNull variants alias the same. 1313 // Also, make sure exact and non-exact variants alias the same. 1314 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1315 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1316 } 1317 } 1318 1319 // Oop pointers need some flattening 1320 const TypeInstPtr *to = tj->isa_instptr(); 1321 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1322 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1323 if( ptr == TypePtr::Constant ) { 1324 if (to->klass() != ciEnv::current()->Class_klass() || 1325 offset < k->size_helper() * wordSize) { 1326 // No constant oop pointers (such as Strings); they alias with 1327 // unknown strings. 1328 assert(!is_known_inst, "not scalarizable allocation"); 1329 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1330 } 1331 } else if( is_known_inst ) { 1332 tj = to; // Keep NotNull and klass_is_exact for instance type 1333 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1334 // During the 2nd round of IterGVN, NotNull castings are removed. 1335 // Make sure the Bottom and NotNull variants alias the same. 1336 // Also, make sure exact and non-exact variants alias the same. 1337 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1338 } 1339 // Canonicalize the holder of this field 1340 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1341 // First handle header references such as a LoadKlassNode, even if the 1342 // object's klass is unloaded at compile time (4965979). 1343 if (!is_known_inst) { // Do it only for non-instance types 1344 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1345 } 1346 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1347 // Static fields are in the space above the normal instance 1348 // fields in the java.lang.Class instance. 1349 if (to->klass() != ciEnv::current()->Class_klass()) { 1350 to = NULL; 1351 tj = TypeOopPtr::BOTTOM; 1352 offset = tj->offset(); 1353 } 1354 } else { 1355 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1356 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1357 if( is_known_inst ) { 1358 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1359 } else { 1360 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1361 } 1362 } 1363 } 1364 } 1365 1366 // Klass pointers to object array klasses need some flattening 1367 const TypeKlassPtr *tk = tj->isa_klassptr(); 1368 if( tk ) { 1369 // If we are referencing a field within a Klass, we need 1370 // to assume the worst case of an Object. Both exact and 1371 // inexact types must flatten to the same alias class so 1372 // use NotNull as the PTR. 1373 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1374 1375 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1376 TypeKlassPtr::OBJECT->klass(), 1377 offset); 1378 } 1379 1380 ciKlass* klass = tk->klass(); 1381 if( klass->is_obj_array_klass() ) { 1382 ciKlass* k = TypeAryPtr::OOPS->klass(); 1383 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1384 k = TypeInstPtr::BOTTOM->klass(); 1385 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1386 } 1387 1388 // Check for precise loads from the primary supertype array and force them 1389 // to the supertype cache alias index. Check for generic array loads from 1390 // the primary supertype array and also force them to the supertype cache 1391 // alias index. Since the same load can reach both, we need to merge 1392 // these 2 disparate memories into the same alias class. Since the 1393 // primary supertype array is read-only, there's no chance of confusion 1394 // where we bypass an array load and an array store. 1395 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1396 if (offset == Type::OffsetBot || 1397 (offset >= primary_supers_offset && 1398 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1399 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1400 offset = in_bytes(Klass::secondary_super_cache_offset()); 1401 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1402 } 1403 } 1404 1405 // Flatten all Raw pointers together. 1406 if (tj->base() == Type::RawPtr) 1407 tj = TypeRawPtr::BOTTOM; 1408 1409 if (tj->base() == Type::AnyPtr) 1410 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1411 1412 // Flatten all to bottom for now 1413 switch( _AliasLevel ) { 1414 case 0: 1415 tj = TypePtr::BOTTOM; 1416 break; 1417 case 1: // Flatten to: oop, static, field or array 1418 switch (tj->base()) { 1419 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1420 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1421 case Type::AryPtr: // do not distinguish arrays at all 1422 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1423 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1424 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1425 default: ShouldNotReachHere(); 1426 } 1427 break; 1428 case 2: // No collapsing at level 2; keep all splits 1429 case 3: // No collapsing at level 3; keep all splits 1430 break; 1431 default: 1432 Unimplemented(); 1433 } 1434 1435 offset = tj->offset(); 1436 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1437 1438 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1439 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1440 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1441 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1442 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1443 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1444 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1445 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1446 assert( tj->ptr() != TypePtr::TopPTR && 1447 tj->ptr() != TypePtr::AnyNull && 1448 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1449 // assert( tj->ptr() != TypePtr::Constant || 1450 // tj->base() == Type::RawPtr || 1451 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1452 1453 return tj; 1454 } 1455 1456 void Compile::AliasType::Init(int i, const TypePtr* at) { 1457 _index = i; 1458 _adr_type = at; 1459 _field = NULL; 1460 _is_rewritable = true; // default 1461 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1462 if (atoop != NULL && atoop->is_known_instance()) { 1463 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1464 _general_index = Compile::current()->get_alias_index(gt); 1465 } else { 1466 _general_index = 0; 1467 } 1468 } 1469 1470 //---------------------------------print_on------------------------------------ 1471 #ifndef PRODUCT 1472 void Compile::AliasType::print_on(outputStream* st) { 1473 if (index() < 10) 1474 st->print("@ <%d> ", index()); 1475 else st->print("@ <%d>", index()); 1476 st->print(is_rewritable() ? " " : " RO"); 1477 int offset = adr_type()->offset(); 1478 if (offset == Type::OffsetBot) 1479 st->print(" +any"); 1480 else st->print(" +%-3d", offset); 1481 st->print(" in "); 1482 adr_type()->dump_on(st); 1483 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1484 if (field() != NULL && tjp) { 1485 if (tjp->klass() != field()->holder() || 1486 tjp->offset() != field()->offset_in_bytes()) { 1487 st->print(" != "); 1488 field()->print(); 1489 st->print(" ***"); 1490 } 1491 } 1492 } 1493 1494 void print_alias_types() { 1495 Compile* C = Compile::current(); 1496 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1497 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1498 C->alias_type(idx)->print_on(tty); 1499 tty->cr(); 1500 } 1501 } 1502 #endif 1503 1504 1505 //----------------------------probe_alias_cache-------------------------------- 1506 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1507 intptr_t key = (intptr_t) adr_type; 1508 key ^= key >> logAliasCacheSize; 1509 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1510 } 1511 1512 1513 //-----------------------------grow_alias_types-------------------------------- 1514 void Compile::grow_alias_types() { 1515 const int old_ats = _max_alias_types; // how many before? 1516 const int new_ats = old_ats; // how many more? 1517 const int grow_ats = old_ats+new_ats; // how many now? 1518 _max_alias_types = grow_ats; 1519 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1520 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1521 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1522 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1523 } 1524 1525 1526 //--------------------------------find_alias_type------------------------------ 1527 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1528 if (_AliasLevel == 0) 1529 return alias_type(AliasIdxBot); 1530 1531 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1532 if (ace->_adr_type == adr_type) { 1533 return alias_type(ace->_index); 1534 } 1535 1536 // Handle special cases. 1537 if (adr_type == NULL) return alias_type(AliasIdxTop); 1538 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1539 1540 // Do it the slow way. 1541 const TypePtr* flat = flatten_alias_type(adr_type); 1542 1543 #ifdef ASSERT 1544 assert(flat == flatten_alias_type(flat), "idempotent"); 1545 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1546 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1547 const TypeOopPtr* foop = flat->is_oopptr(); 1548 // Scalarizable allocations have exact klass always. 1549 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1550 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1551 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1552 } 1553 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1554 #endif 1555 1556 int idx = AliasIdxTop; 1557 for (int i = 0; i < num_alias_types(); i++) { 1558 if (alias_type(i)->adr_type() == flat) { 1559 idx = i; 1560 break; 1561 } 1562 } 1563 1564 if (idx == AliasIdxTop) { 1565 if (no_create) return NULL; 1566 // Grow the array if necessary. 1567 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1568 // Add a new alias type. 1569 idx = _num_alias_types++; 1570 _alias_types[idx]->Init(idx, flat); 1571 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1572 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1573 if (flat->isa_instptr()) { 1574 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1575 && flat->is_instptr()->klass() == env()->Class_klass()) 1576 alias_type(idx)->set_rewritable(false); 1577 } 1578 if (flat->isa_klassptr()) { 1579 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1580 alias_type(idx)->set_rewritable(false); 1581 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1582 alias_type(idx)->set_rewritable(false); 1583 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1584 alias_type(idx)->set_rewritable(false); 1585 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1586 alias_type(idx)->set_rewritable(false); 1587 } 1588 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1589 // but the base pointer type is not distinctive enough to identify 1590 // references into JavaThread.) 1591 1592 // Check for final fields. 1593 const TypeInstPtr* tinst = flat->isa_instptr(); 1594 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1595 ciField* field; 1596 if (tinst->const_oop() != NULL && 1597 tinst->klass() == ciEnv::current()->Class_klass() && 1598 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1599 // static field 1600 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1601 field = k->get_field_by_offset(tinst->offset(), true); 1602 } else { 1603 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1604 field = k->get_field_by_offset(tinst->offset(), false); 1605 } 1606 assert(field == NULL || 1607 original_field == NULL || 1608 (field->holder() == original_field->holder() && 1609 field->offset() == original_field->offset() && 1610 field->is_static() == original_field->is_static()), "wrong field?"); 1611 // Set field() and is_rewritable() attributes. 1612 if (field != NULL) alias_type(idx)->set_field(field); 1613 } 1614 } 1615 1616 // Fill the cache for next time. 1617 ace->_adr_type = adr_type; 1618 ace->_index = idx; 1619 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1620 1621 // Might as well try to fill the cache for the flattened version, too. 1622 AliasCacheEntry* face = probe_alias_cache(flat); 1623 if (face->_adr_type == NULL) { 1624 face->_adr_type = flat; 1625 face->_index = idx; 1626 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1627 } 1628 1629 return alias_type(idx); 1630 } 1631 1632 1633 Compile::AliasType* Compile::alias_type(ciField* field) { 1634 const TypeOopPtr* t; 1635 if (field->is_static()) 1636 t = TypeInstPtr::make(field->holder()->java_mirror()); 1637 else 1638 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1639 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1640 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1641 return atp; 1642 } 1643 1644 1645 //------------------------------have_alias_type-------------------------------- 1646 bool Compile::have_alias_type(const TypePtr* adr_type) { 1647 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1648 if (ace->_adr_type == adr_type) { 1649 return true; 1650 } 1651 1652 // Handle special cases. 1653 if (adr_type == NULL) return true; 1654 if (adr_type == TypePtr::BOTTOM) return true; 1655 1656 return find_alias_type(adr_type, true, NULL) != NULL; 1657 } 1658 1659 //-----------------------------must_alias-------------------------------------- 1660 // True if all values of the given address type are in the given alias category. 1661 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1662 if (alias_idx == AliasIdxBot) return true; // the universal category 1663 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1664 if (alias_idx == AliasIdxTop) return false; // the empty category 1665 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1666 1667 // the only remaining possible overlap is identity 1668 int adr_idx = get_alias_index(adr_type); 1669 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1670 assert(adr_idx == alias_idx || 1671 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1672 && adr_type != TypeOopPtr::BOTTOM), 1673 "should not be testing for overlap with an unsafe pointer"); 1674 return adr_idx == alias_idx; 1675 } 1676 1677 //------------------------------can_alias-------------------------------------- 1678 // True if any values of the given address type are in the given alias category. 1679 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1680 if (alias_idx == AliasIdxTop) return false; // the empty category 1681 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1682 if (alias_idx == AliasIdxBot) return true; // the universal category 1683 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1684 1685 // the only remaining possible overlap is identity 1686 int adr_idx = get_alias_index(adr_type); 1687 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1688 return adr_idx == alias_idx; 1689 } 1690 1691 1692 1693 //---------------------------pop_warm_call------------------------------------- 1694 WarmCallInfo* Compile::pop_warm_call() { 1695 WarmCallInfo* wci = _warm_calls; 1696 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1697 return wci; 1698 } 1699 1700 //----------------------------Inline_Warm-------------------------------------- 1701 int Compile::Inline_Warm() { 1702 // If there is room, try to inline some more warm call sites. 1703 // %%% Do a graph index compaction pass when we think we're out of space? 1704 if (!InlineWarmCalls) return 0; 1705 1706 int calls_made_hot = 0; 1707 int room_to_grow = NodeCountInliningCutoff - unique(); 1708 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1709 int amount_grown = 0; 1710 WarmCallInfo* call; 1711 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1712 int est_size = (int)call->size(); 1713 if (est_size > (room_to_grow - amount_grown)) { 1714 // This one won't fit anyway. Get rid of it. 1715 call->make_cold(); 1716 continue; 1717 } 1718 call->make_hot(); 1719 calls_made_hot++; 1720 amount_grown += est_size; 1721 amount_to_grow -= est_size; 1722 } 1723 1724 if (calls_made_hot > 0) set_major_progress(); 1725 return calls_made_hot; 1726 } 1727 1728 1729 //----------------------------Finish_Warm-------------------------------------- 1730 void Compile::Finish_Warm() { 1731 if (!InlineWarmCalls) return; 1732 if (failing()) return; 1733 if (warm_calls() == NULL) return; 1734 1735 // Clean up loose ends, if we are out of space for inlining. 1736 WarmCallInfo* call; 1737 while ((call = pop_warm_call()) != NULL) { 1738 call->make_cold(); 1739 } 1740 } 1741 1742 //---------------------cleanup_loop_predicates----------------------- 1743 // Remove the opaque nodes that protect the predicates so that all unused 1744 // checks and uncommon_traps will be eliminated from the ideal graph 1745 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1746 if (predicate_count()==0) return; 1747 for (int i = predicate_count(); i > 0; i--) { 1748 Node * n = predicate_opaque1_node(i-1); 1749 assert(n->Opcode() == Op_Opaque1, "must be"); 1750 igvn.replace_node(n, n->in(1)); 1751 } 1752 assert(predicate_count()==0, "should be clean!"); 1753 } 1754 1755 //------------------------------Optimize--------------------------------------- 1756 // Given a graph, optimize it. 1757 void Compile::Optimize() { 1758 TracePhase t1("optimizer", &_t_optimizer, true); 1759 1760 #ifndef PRODUCT 1761 if (env()->break_at_compile()) { 1762 BREAKPOINT; 1763 } 1764 1765 #endif 1766 1767 ResourceMark rm; 1768 int loop_opts_cnt; 1769 1770 NOT_PRODUCT( verify_graph_edges(); ) 1771 1772 print_method("After Parsing"); 1773 1774 { 1775 // Iterative Global Value Numbering, including ideal transforms 1776 // Initialize IterGVN with types and values from parse-time GVN 1777 PhaseIterGVN igvn(initial_gvn()); 1778 { 1779 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1780 igvn.optimize(); 1781 } 1782 1783 print_method("Iter GVN 1", 2); 1784 1785 if (failing()) return; 1786 1787 // Perform escape analysis 1788 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 1789 if (has_loops()) { 1790 // Cleanup graph (remove dead nodes). 1791 TracePhase t2("idealLoop", &_t_idealLoop, true); 1792 PhaseIdealLoop ideal_loop( igvn, false, true ); 1793 if (major_progress()) print_method("PhaseIdealLoop before EA", 2); 1794 if (failing()) return; 1795 } 1796 ConnectionGraph::do_analysis(this, &igvn); 1797 1798 if (failing()) return; 1799 1800 // Optimize out fields loads from scalar replaceable allocations. 1801 igvn.optimize(); 1802 print_method("Iter GVN after EA", 2); 1803 1804 if (failing()) return; 1805 1806 if (congraph() != NULL && macro_count() > 0) { 1807 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 1808 PhaseMacroExpand mexp(igvn); 1809 mexp.eliminate_macro_nodes(); 1810 igvn.set_delay_transform(false); 1811 1812 igvn.optimize(); 1813 print_method("Iter GVN after eliminating allocations and locks", 2); 1814 1815 if (failing()) return; 1816 } 1817 } 1818 1819 // Loop transforms on the ideal graph. Range Check Elimination, 1820 // peeling, unrolling, etc. 1821 1822 // Set loop opts counter 1823 loop_opts_cnt = num_loop_opts(); 1824 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1825 { 1826 TracePhase t2("idealLoop", &_t_idealLoop, true); 1827 PhaseIdealLoop ideal_loop( igvn, true ); 1828 loop_opts_cnt--; 1829 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1830 if (failing()) return; 1831 } 1832 // Loop opts pass if partial peeling occurred in previous pass 1833 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1834 TracePhase t3("idealLoop", &_t_idealLoop, true); 1835 PhaseIdealLoop ideal_loop( igvn, false ); 1836 loop_opts_cnt--; 1837 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1838 if (failing()) return; 1839 } 1840 // Loop opts pass for loop-unrolling before CCP 1841 if(major_progress() && (loop_opts_cnt > 0)) { 1842 TracePhase t4("idealLoop", &_t_idealLoop, true); 1843 PhaseIdealLoop ideal_loop( igvn, false ); 1844 loop_opts_cnt--; 1845 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1846 } 1847 if (!failing()) { 1848 // Verify that last round of loop opts produced a valid graph 1849 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1850 PhaseIdealLoop::verify(igvn); 1851 } 1852 } 1853 if (failing()) return; 1854 1855 // Conditional Constant Propagation; 1856 PhaseCCP ccp( &igvn ); 1857 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1858 { 1859 TracePhase t2("ccp", &_t_ccp, true); 1860 ccp.do_transform(); 1861 } 1862 print_method("PhaseCPP 1", 2); 1863 1864 assert( true, "Break here to ccp.dump_old2new_map()"); 1865 1866 // Iterative Global Value Numbering, including ideal transforms 1867 { 1868 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1869 igvn = ccp; 1870 igvn.optimize(); 1871 } 1872 1873 print_method("Iter GVN 2", 2); 1874 1875 if (failing()) return; 1876 1877 // Loop transforms on the ideal graph. Range Check Elimination, 1878 // peeling, unrolling, etc. 1879 if(loop_opts_cnt > 0) { 1880 debug_only( int cnt = 0; ); 1881 while(major_progress() && (loop_opts_cnt > 0)) { 1882 TracePhase t2("idealLoop", &_t_idealLoop, true); 1883 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1884 PhaseIdealLoop ideal_loop( igvn, true); 1885 loop_opts_cnt--; 1886 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1887 if (failing()) return; 1888 } 1889 } 1890 1891 { 1892 // Verify that all previous optimizations produced a valid graph 1893 // at least to this point, even if no loop optimizations were done. 1894 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1895 PhaseIdealLoop::verify(igvn); 1896 } 1897 1898 { 1899 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1900 PhaseMacroExpand mex(igvn); 1901 if (mex.expand_macro_nodes()) { 1902 assert(failing(), "must bail out w/ explicit message"); 1903 return; 1904 } 1905 } 1906 1907 } // (End scope of igvn; run destructor if necessary for asserts.) 1908 1909 // A method with only infinite loops has no edges entering loops from root 1910 { 1911 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1912 if (final_graph_reshaping()) { 1913 assert(failing(), "must bail out w/ explicit message"); 1914 return; 1915 } 1916 } 1917 1918 print_method("Optimize finished", 2); 1919 } 1920 1921 1922 //------------------------------Code_Gen--------------------------------------- 1923 // Given a graph, generate code for it 1924 void Compile::Code_Gen() { 1925 if (failing()) return; 1926 1927 // Perform instruction selection. You might think we could reclaim Matcher 1928 // memory PDQ, but actually the Matcher is used in generating spill code. 1929 // Internals of the Matcher (including some VectorSets) must remain live 1930 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1931 // set a bit in reclaimed memory. 1932 1933 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1934 // nodes. Mapping is only valid at the root of each matched subtree. 1935 NOT_PRODUCT( verify_graph_edges(); ) 1936 1937 Node_List proj_list; 1938 Matcher m(proj_list); 1939 _matcher = &m; 1940 { 1941 TracePhase t2("matcher", &_t_matcher, true); 1942 m.match(); 1943 } 1944 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1945 // nodes. Mapping is only valid at the root of each matched subtree. 1946 NOT_PRODUCT( verify_graph_edges(); ) 1947 1948 // If you have too many nodes, or if matching has failed, bail out 1949 check_node_count(0, "out of nodes matching instructions"); 1950 if (failing()) return; 1951 1952 // Build a proper-looking CFG 1953 PhaseCFG cfg(node_arena(), root(), m); 1954 _cfg = &cfg; 1955 { 1956 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1957 cfg.Dominators(); 1958 if (failing()) return; 1959 1960 NOT_PRODUCT( verify_graph_edges(); ) 1961 1962 cfg.Estimate_Block_Frequency(); 1963 cfg.GlobalCodeMotion(m,unique(),proj_list); 1964 if (failing()) return; 1965 1966 print_method("Global code motion", 2); 1967 1968 NOT_PRODUCT( verify_graph_edges(); ) 1969 1970 debug_only( cfg.verify(); ) 1971 } 1972 NOT_PRODUCT( verify_graph_edges(); ) 1973 1974 PhaseChaitin regalloc(unique(),cfg,m); 1975 _regalloc = ®alloc; 1976 { 1977 TracePhase t2("regalloc", &_t_registerAllocation, true); 1978 // Perform any platform dependent preallocation actions. This is used, 1979 // for example, to avoid taking an implicit null pointer exception 1980 // using the frame pointer on win95. 1981 _regalloc->pd_preallocate_hook(); 1982 1983 // Perform register allocation. After Chaitin, use-def chains are 1984 // no longer accurate (at spill code) and so must be ignored. 1985 // Node->LRG->reg mappings are still accurate. 1986 _regalloc->Register_Allocate(); 1987 1988 // Bail out if the allocator builds too many nodes 1989 if (failing()) return; 1990 } 1991 1992 // Prior to register allocation we kept empty basic blocks in case the 1993 // the allocator needed a place to spill. After register allocation we 1994 // are not adding any new instructions. If any basic block is empty, we 1995 // can now safely remove it. 1996 { 1997 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1998 cfg.remove_empty(); 1999 if (do_freq_based_layout()) { 2000 PhaseBlockLayout layout(cfg); 2001 } else { 2002 cfg.set_loop_alignment(); 2003 } 2004 cfg.fixup_flow(); 2005 } 2006 2007 // Perform any platform dependent postallocation verifications. 2008 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 2009 2010 // Apply peephole optimizations 2011 if( OptoPeephole ) { 2012 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2013 PhasePeephole peep( _regalloc, cfg); 2014 peep.do_transform(); 2015 } 2016 2017 // Convert Nodes to instruction bits in a buffer 2018 { 2019 // %%%% workspace merge brought two timers together for one job 2020 TracePhase t2a("output", &_t_output, true); 2021 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2022 Output(); 2023 } 2024 2025 print_method("Final Code"); 2026 2027 // He's dead, Jim. 2028 _cfg = (PhaseCFG*)0xdeadbeef; 2029 _regalloc = (PhaseChaitin*)0xdeadbeef; 2030 } 2031 2032 2033 //------------------------------dump_asm--------------------------------------- 2034 // Dump formatted assembly 2035 #ifndef PRODUCT 2036 void Compile::dump_asm(int *pcs, uint pc_limit) { 2037 bool cut_short = false; 2038 tty->print_cr("#"); 2039 tty->print("# "); _tf->dump(); tty->cr(); 2040 tty->print_cr("#"); 2041 2042 // For all blocks 2043 int pc = 0x0; // Program counter 2044 char starts_bundle = ' '; 2045 _regalloc->dump_frame(); 2046 2047 Node *n = NULL; 2048 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 2049 if (VMThread::should_terminate()) { cut_short = true; break; } 2050 Block *b = _cfg->_blocks[i]; 2051 if (b->is_connector() && !Verbose) continue; 2052 n = b->_nodes[0]; 2053 if (pcs && n->_idx < pc_limit) 2054 tty->print("%3.3x ", pcs[n->_idx]); 2055 else 2056 tty->print(" "); 2057 b->dump_head( &_cfg->_bbs ); 2058 if (b->is_connector()) { 2059 tty->print_cr(" # Empty connector block"); 2060 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2061 tty->print_cr(" # Block is sole successor of call"); 2062 } 2063 2064 // For all instructions 2065 Node *delay = NULL; 2066 for( uint j = 0; j<b->_nodes.size(); j++ ) { 2067 if (VMThread::should_terminate()) { cut_short = true; break; } 2068 n = b->_nodes[j]; 2069 if (valid_bundle_info(n)) { 2070 Bundle *bundle = node_bundling(n); 2071 if (bundle->used_in_unconditional_delay()) { 2072 delay = n; 2073 continue; 2074 } 2075 if (bundle->starts_bundle()) 2076 starts_bundle = '+'; 2077 } 2078 2079 if (WizardMode) n->dump(); 2080 2081 if( !n->is_Region() && // Dont print in the Assembly 2082 !n->is_Phi() && // a few noisely useless nodes 2083 !n->is_Proj() && 2084 !n->is_MachTemp() && 2085 !n->is_SafePointScalarObject() && 2086 !n->is_Catch() && // Would be nice to print exception table targets 2087 !n->is_MergeMem() && // Not very interesting 2088 !n->is_top() && // Debug info table constants 2089 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2090 ) { 2091 if (pcs && n->_idx < pc_limit) 2092 tty->print("%3.3x", pcs[n->_idx]); 2093 else 2094 tty->print(" "); 2095 tty->print(" %c ", starts_bundle); 2096 starts_bundle = ' '; 2097 tty->print("\t"); 2098 n->format(_regalloc, tty); 2099 tty->cr(); 2100 } 2101 2102 // If we have an instruction with a delay slot, and have seen a delay, 2103 // then back up and print it 2104 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2105 assert(delay != NULL, "no unconditional delay instruction"); 2106 if (WizardMode) delay->dump(); 2107 2108 if (node_bundling(delay)->starts_bundle()) 2109 starts_bundle = '+'; 2110 if (pcs && n->_idx < pc_limit) 2111 tty->print("%3.3x", pcs[n->_idx]); 2112 else 2113 tty->print(" "); 2114 tty->print(" %c ", starts_bundle); 2115 starts_bundle = ' '; 2116 tty->print("\t"); 2117 delay->format(_regalloc, tty); 2118 tty->print_cr(""); 2119 delay = NULL; 2120 } 2121 2122 // Dump the exception table as well 2123 if( n->is_Catch() && (Verbose || WizardMode) ) { 2124 // Print the exception table for this offset 2125 _handler_table.print_subtable_for(pc); 2126 } 2127 } 2128 2129 if (pcs && n->_idx < pc_limit) 2130 tty->print_cr("%3.3x", pcs[n->_idx]); 2131 else 2132 tty->print_cr(""); 2133 2134 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2135 2136 } // End of per-block dump 2137 tty->print_cr(""); 2138 2139 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2140 } 2141 #endif 2142 2143 //------------------------------Final_Reshape_Counts--------------------------- 2144 // This class defines counters to help identify when a method 2145 // may/must be executed using hardware with only 24-bit precision. 2146 struct Final_Reshape_Counts : public StackObj { 2147 int _call_count; // count non-inlined 'common' calls 2148 int _float_count; // count float ops requiring 24-bit precision 2149 int _double_count; // count double ops requiring more precision 2150 int _java_call_count; // count non-inlined 'java' calls 2151 int _inner_loop_count; // count loops which need alignment 2152 VectorSet _visited; // Visitation flags 2153 Node_List _tests; // Set of IfNodes & PCTableNodes 2154 2155 Final_Reshape_Counts() : 2156 _call_count(0), _float_count(0), _double_count(0), 2157 _java_call_count(0), _inner_loop_count(0), 2158 _visited( Thread::current()->resource_area() ) { } 2159 2160 void inc_call_count () { _call_count ++; } 2161 void inc_float_count () { _float_count ++; } 2162 void inc_double_count() { _double_count++; } 2163 void inc_java_call_count() { _java_call_count++; } 2164 void inc_inner_loop_count() { _inner_loop_count++; } 2165 2166 int get_call_count () const { return _call_count ; } 2167 int get_float_count () const { return _float_count ; } 2168 int get_double_count() const { return _double_count; } 2169 int get_java_call_count() const { return _java_call_count; } 2170 int get_inner_loop_count() const { return _inner_loop_count; } 2171 }; 2172 2173 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2174 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2175 // Make sure the offset goes inside the instance layout. 2176 return k->contains_field_offset(tp->offset()); 2177 // Note that OffsetBot and OffsetTop are very negative. 2178 } 2179 2180 // Eliminate trivially redundant StoreCMs and accumulate their 2181 // precedence edges. 2182 void Compile::eliminate_redundant_card_marks(Node* n) { 2183 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2184 if (n->in(MemNode::Address)->outcnt() > 1) { 2185 // There are multiple users of the same address so it might be 2186 // possible to eliminate some of the StoreCMs 2187 Node* mem = n->in(MemNode::Memory); 2188 Node* adr = n->in(MemNode::Address); 2189 Node* val = n->in(MemNode::ValueIn); 2190 Node* prev = n; 2191 bool done = false; 2192 // Walk the chain of StoreCMs eliminating ones that match. As 2193 // long as it's a chain of single users then the optimization is 2194 // safe. Eliminating partially redundant StoreCMs would require 2195 // cloning copies down the other paths. 2196 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2197 if (adr == mem->in(MemNode::Address) && 2198 val == mem->in(MemNode::ValueIn)) { 2199 // redundant StoreCM 2200 if (mem->req() > MemNode::OopStore) { 2201 // Hasn't been processed by this code yet. 2202 n->add_prec(mem->in(MemNode::OopStore)); 2203 } else { 2204 // Already converted to precedence edge 2205 for (uint i = mem->req(); i < mem->len(); i++) { 2206 // Accumulate any precedence edges 2207 if (mem->in(i) != NULL) { 2208 n->add_prec(mem->in(i)); 2209 } 2210 } 2211 // Everything above this point has been processed. 2212 done = true; 2213 } 2214 // Eliminate the previous StoreCM 2215 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2216 assert(mem->outcnt() == 0, "should be dead"); 2217 mem->disconnect_inputs(NULL, this); 2218 } else { 2219 prev = mem; 2220 } 2221 mem = prev->in(MemNode::Memory); 2222 } 2223 } 2224 } 2225 2226 //------------------------------final_graph_reshaping_impl---------------------- 2227 // Implement items 1-5 from final_graph_reshaping below. 2228 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2229 2230 if ( n->outcnt() == 0 ) return; // dead node 2231 uint nop = n->Opcode(); 2232 2233 // Check for 2-input instruction with "last use" on right input. 2234 // Swap to left input. Implements item (2). 2235 if( n->req() == 3 && // two-input instruction 2236 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2237 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2238 n->in(2)->outcnt() == 1 &&// right use IS a last use 2239 !n->in(2)->is_Con() ) { // right use is not a constant 2240 // Check for commutative opcode 2241 switch( nop ) { 2242 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2243 case Op_MaxI: case Op_MinI: 2244 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2245 case Op_AndL: case Op_XorL: case Op_OrL: 2246 case Op_AndI: case Op_XorI: case Op_OrI: { 2247 // Move "last use" input to left by swapping inputs 2248 n->swap_edges(1, 2); 2249 break; 2250 } 2251 default: 2252 break; 2253 } 2254 } 2255 2256 #ifdef ASSERT 2257 if( n->is_Mem() ) { 2258 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2259 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2260 // oop will be recorded in oop map if load crosses safepoint 2261 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2262 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2263 "raw memory operations should have control edge"); 2264 } 2265 #endif 2266 // Count FPU ops and common calls, implements item (3) 2267 switch( nop ) { 2268 // Count all float operations that may use FPU 2269 case Op_AddF: 2270 case Op_SubF: 2271 case Op_MulF: 2272 case Op_DivF: 2273 case Op_NegF: 2274 case Op_ModF: 2275 case Op_ConvI2F: 2276 case Op_ConF: 2277 case Op_CmpF: 2278 case Op_CmpF3: 2279 // case Op_ConvL2F: // longs are split into 32-bit halves 2280 frc.inc_float_count(); 2281 break; 2282 2283 case Op_ConvF2D: 2284 case Op_ConvD2F: 2285 frc.inc_float_count(); 2286 frc.inc_double_count(); 2287 break; 2288 2289 // Count all double operations that may use FPU 2290 case Op_AddD: 2291 case Op_SubD: 2292 case Op_MulD: 2293 case Op_DivD: 2294 case Op_NegD: 2295 case Op_ModD: 2296 case Op_ConvI2D: 2297 case Op_ConvD2I: 2298 // case Op_ConvL2D: // handled by leaf call 2299 // case Op_ConvD2L: // handled by leaf call 2300 case Op_ConD: 2301 case Op_CmpD: 2302 case Op_CmpD3: 2303 frc.inc_double_count(); 2304 break; 2305 case Op_Opaque1: // Remove Opaque Nodes before matching 2306 case Op_Opaque2: // Remove Opaque Nodes before matching 2307 n->subsume_by(n->in(1), this); 2308 break; 2309 case Op_CallStaticJava: 2310 case Op_CallJava: 2311 case Op_CallDynamicJava: 2312 frc.inc_java_call_count(); // Count java call site; 2313 case Op_CallRuntime: 2314 case Op_CallLeaf: 2315 case Op_CallLeafNoFP: { 2316 assert( n->is_Call(), "" ); 2317 CallNode *call = n->as_Call(); 2318 // Count call sites where the FP mode bit would have to be flipped. 2319 // Do not count uncommon runtime calls: 2320 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2321 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2322 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2323 frc.inc_call_count(); // Count the call site 2324 } else { // See if uncommon argument is shared 2325 Node *n = call->in(TypeFunc::Parms); 2326 int nop = n->Opcode(); 2327 // Clone shared simple arguments to uncommon calls, item (1). 2328 if( n->outcnt() > 1 && 2329 !n->is_Proj() && 2330 nop != Op_CreateEx && 2331 nop != Op_CheckCastPP && 2332 nop != Op_DecodeN && 2333 nop != Op_DecodeNKlass && 2334 !n->is_Mem() ) { 2335 Node *x = n->clone(); 2336 call->set_req( TypeFunc::Parms, x ); 2337 } 2338 } 2339 break; 2340 } 2341 2342 case Op_StoreD: 2343 case Op_LoadD: 2344 case Op_LoadD_unaligned: 2345 frc.inc_double_count(); 2346 goto handle_mem; 2347 case Op_StoreF: 2348 case Op_LoadF: 2349 frc.inc_float_count(); 2350 goto handle_mem; 2351 2352 case Op_StoreCM: 2353 { 2354 // Convert OopStore dependence into precedence edge 2355 Node* prec = n->in(MemNode::OopStore); 2356 n->del_req(MemNode::OopStore); 2357 n->add_prec(prec); 2358 eliminate_redundant_card_marks(n); 2359 } 2360 2361 // fall through 2362 2363 case Op_StoreB: 2364 case Op_StoreC: 2365 case Op_StorePConditional: 2366 case Op_StoreI: 2367 case Op_StoreL: 2368 case Op_StoreIConditional: 2369 case Op_StoreLConditional: 2370 case Op_CompareAndSwapI: 2371 case Op_CompareAndSwapL: 2372 case Op_CompareAndSwapP: 2373 case Op_CompareAndSwapN: 2374 case Op_GetAndAddI: 2375 case Op_GetAndAddL: 2376 case Op_GetAndSetI: 2377 case Op_GetAndSetL: 2378 case Op_GetAndSetP: 2379 case Op_GetAndSetN: 2380 case Op_StoreP: 2381 case Op_StoreN: 2382 case Op_StoreNKlass: 2383 case Op_LoadB: 2384 case Op_LoadUB: 2385 case Op_LoadUS: 2386 case Op_LoadI: 2387 case Op_LoadKlass: 2388 case Op_LoadNKlass: 2389 case Op_LoadL: 2390 case Op_LoadL_unaligned: 2391 case Op_LoadPLocked: 2392 case Op_LoadP: 2393 case Op_LoadN: 2394 case Op_LoadRange: 2395 case Op_LoadS: { 2396 handle_mem: 2397 #ifdef ASSERT 2398 if( VerifyOptoOopOffsets ) { 2399 assert( n->is_Mem(), "" ); 2400 MemNode *mem = (MemNode*)n; 2401 // Check to see if address types have grounded out somehow. 2402 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2403 assert( !tp || oop_offset_is_sane(tp), "" ); 2404 } 2405 #endif 2406 break; 2407 } 2408 2409 case Op_AddP: { // Assert sane base pointers 2410 Node *addp = n->in(AddPNode::Address); 2411 assert( !addp->is_AddP() || 2412 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2413 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2414 "Base pointers must match" ); 2415 #ifdef _LP64 2416 if ((UseCompressedOops || UseCompressedKlassPointers) && 2417 addp->Opcode() == Op_ConP && 2418 addp == n->in(AddPNode::Base) && 2419 n->in(AddPNode::Offset)->is_Con()) { 2420 // Use addressing with narrow klass to load with offset on x86. 2421 // On sparc loading 32-bits constant and decoding it have less 2422 // instructions (4) then load 64-bits constant (7). 2423 // Do this transformation here since IGVN will convert ConN back to ConP. 2424 const Type* t = addp->bottom_type(); 2425 if (t->isa_oopptr() || t->isa_klassptr()) { 2426 Node* nn = NULL; 2427 2428 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2429 2430 // Look for existing ConN node of the same exact type. 2431 Node* r = root(); 2432 uint cnt = r->outcnt(); 2433 for (uint i = 0; i < cnt; i++) { 2434 Node* m = r->raw_out(i); 2435 if (m!= NULL && m->Opcode() == op && 2436 m->bottom_type()->make_ptr() == t) { 2437 nn = m; 2438 break; 2439 } 2440 } 2441 if (nn != NULL) { 2442 // Decode a narrow oop to match address 2443 // [R12 + narrow_oop_reg<<3 + offset] 2444 if (t->isa_oopptr()) { 2445 nn = new (this) DecodeNNode(nn, t); 2446 } else { 2447 nn = new (this) DecodeNKlassNode(nn, t); 2448 } 2449 n->set_req(AddPNode::Base, nn); 2450 n->set_req(AddPNode::Address, nn); 2451 if (addp->outcnt() == 0) { 2452 addp->disconnect_inputs(NULL, this); 2453 } 2454 } 2455 } 2456 } 2457 #endif 2458 break; 2459 } 2460 2461 #ifdef _LP64 2462 case Op_CastPP: 2463 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2464 Node* in1 = n->in(1); 2465 const Type* t = n->bottom_type(); 2466 Node* new_in1 = in1->clone(); 2467 new_in1->as_DecodeN()->set_type(t); 2468 2469 if (!Matcher::narrow_oop_use_complex_address()) { 2470 // 2471 // x86, ARM and friends can handle 2 adds in addressing mode 2472 // and Matcher can fold a DecodeN node into address by using 2473 // a narrow oop directly and do implicit NULL check in address: 2474 // 2475 // [R12 + narrow_oop_reg<<3 + offset] 2476 // NullCheck narrow_oop_reg 2477 // 2478 // On other platforms (Sparc) we have to keep new DecodeN node and 2479 // use it to do implicit NULL check in address: 2480 // 2481 // decode_not_null narrow_oop_reg, base_reg 2482 // [base_reg + offset] 2483 // NullCheck base_reg 2484 // 2485 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2486 // to keep the information to which NULL check the new DecodeN node 2487 // corresponds to use it as value in implicit_null_check(). 2488 // 2489 new_in1->set_req(0, n->in(0)); 2490 } 2491 2492 n->subsume_by(new_in1, this); 2493 if (in1->outcnt() == 0) { 2494 in1->disconnect_inputs(NULL, this); 2495 } 2496 } 2497 break; 2498 2499 case Op_CmpP: 2500 // Do this transformation here to preserve CmpPNode::sub() and 2501 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2502 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2503 Node* in1 = n->in(1); 2504 Node* in2 = n->in(2); 2505 if (!in1->is_DecodeNarrowPtr()) { 2506 in2 = in1; 2507 in1 = n->in(2); 2508 } 2509 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2510 2511 Node* new_in2 = NULL; 2512 if (in2->is_DecodeNarrowPtr()) { 2513 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2514 new_in2 = in2->in(1); 2515 } else if (in2->Opcode() == Op_ConP) { 2516 const Type* t = in2->bottom_type(); 2517 if (t == TypePtr::NULL_PTR) { 2518 assert(in1->is_DecodeN(), "compare klass to null?"); 2519 // Don't convert CmpP null check into CmpN if compressed 2520 // oops implicit null check is not generated. 2521 // This will allow to generate normal oop implicit null check. 2522 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2523 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2524 // 2525 // This transformation together with CastPP transformation above 2526 // will generated code for implicit NULL checks for compressed oops. 2527 // 2528 // The original code after Optimize() 2529 // 2530 // LoadN memory, narrow_oop_reg 2531 // decode narrow_oop_reg, base_reg 2532 // CmpP base_reg, NULL 2533 // CastPP base_reg // NotNull 2534 // Load [base_reg + offset], val_reg 2535 // 2536 // after these transformations will be 2537 // 2538 // LoadN memory, narrow_oop_reg 2539 // CmpN narrow_oop_reg, NULL 2540 // decode_not_null narrow_oop_reg, base_reg 2541 // Load [base_reg + offset], val_reg 2542 // 2543 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2544 // since narrow oops can be used in debug info now (see the code in 2545 // final_graph_reshaping_walk()). 2546 // 2547 // At the end the code will be matched to 2548 // on x86: 2549 // 2550 // Load_narrow_oop memory, narrow_oop_reg 2551 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2552 // NullCheck narrow_oop_reg 2553 // 2554 // and on sparc: 2555 // 2556 // Load_narrow_oop memory, narrow_oop_reg 2557 // decode_not_null narrow_oop_reg, base_reg 2558 // Load [base_reg + offset], val_reg 2559 // NullCheck base_reg 2560 // 2561 } else if (t->isa_oopptr()) { 2562 new_in2 = ConNode::make(this, t->make_narrowoop()); 2563 } else if (t->isa_klassptr()) { 2564 new_in2 = ConNode::make(this, t->make_narrowklass()); 2565 } 2566 } 2567 if (new_in2 != NULL) { 2568 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 2569 n->subsume_by(cmpN, this); 2570 if (in1->outcnt() == 0) { 2571 in1->disconnect_inputs(NULL, this); 2572 } 2573 if (in2->outcnt() == 0) { 2574 in2->disconnect_inputs(NULL, this); 2575 } 2576 } 2577 } 2578 break; 2579 2580 case Op_DecodeN: 2581 case Op_DecodeNKlass: 2582 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2583 // DecodeN could be pinned when it can't be fold into 2584 // an address expression, see the code for Op_CastPP above. 2585 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2586 break; 2587 2588 case Op_EncodeP: 2589 case Op_EncodePKlass: { 2590 Node* in1 = n->in(1); 2591 if (in1->is_DecodeNarrowPtr()) { 2592 n->subsume_by(in1->in(1), this); 2593 } else if (in1->Opcode() == Op_ConP) { 2594 const Type* t = in1->bottom_type(); 2595 if (t == TypePtr::NULL_PTR) { 2596 assert(t->isa_oopptr(), "null klass?"); 2597 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2598 } else if (t->isa_oopptr()) { 2599 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2600 } else if (t->isa_klassptr()) { 2601 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2602 } 2603 } 2604 if (in1->outcnt() == 0) { 2605 in1->disconnect_inputs(NULL, this); 2606 } 2607 break; 2608 } 2609 2610 case Op_Proj: { 2611 if (OptimizeStringConcat) { 2612 ProjNode* p = n->as_Proj(); 2613 if (p->_is_io_use) { 2614 // Separate projections were used for the exception path which 2615 // are normally removed by a late inline. If it wasn't inlined 2616 // then they will hang around and should just be replaced with 2617 // the original one. 2618 Node* proj = NULL; 2619 // Replace with just one 2620 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2621 Node *use = i.get(); 2622 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2623 proj = use; 2624 break; 2625 } 2626 } 2627 assert(proj != NULL, "must be found"); 2628 p->subsume_by(proj, this); 2629 } 2630 } 2631 break; 2632 } 2633 2634 case Op_Phi: 2635 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2636 // The EncodeP optimization may create Phi with the same edges 2637 // for all paths. It is not handled well by Register Allocator. 2638 Node* unique_in = n->in(1); 2639 assert(unique_in != NULL, ""); 2640 uint cnt = n->req(); 2641 for (uint i = 2; i < cnt; i++) { 2642 Node* m = n->in(i); 2643 assert(m != NULL, ""); 2644 if (unique_in != m) 2645 unique_in = NULL; 2646 } 2647 if (unique_in != NULL) { 2648 n->subsume_by(unique_in, this); 2649 } 2650 } 2651 break; 2652 2653 #endif 2654 2655 case Op_ModI: 2656 if (UseDivMod) { 2657 // Check if a%b and a/b both exist 2658 Node* d = n->find_similar(Op_DivI); 2659 if (d) { 2660 // Replace them with a fused divmod if supported 2661 if (Matcher::has_match_rule(Op_DivModI)) { 2662 DivModINode* divmod = DivModINode::make(this, n); 2663 d->subsume_by(divmod->div_proj(), this); 2664 n->subsume_by(divmod->mod_proj(), this); 2665 } else { 2666 // replace a%b with a-((a/b)*b) 2667 Node* mult = new (this) MulINode(d, d->in(2)); 2668 Node* sub = new (this) SubINode(d->in(1), mult); 2669 n->subsume_by(sub, this); 2670 } 2671 } 2672 } 2673 break; 2674 2675 case Op_ModL: 2676 if (UseDivMod) { 2677 // Check if a%b and a/b both exist 2678 Node* d = n->find_similar(Op_DivL); 2679 if (d) { 2680 // Replace them with a fused divmod if supported 2681 if (Matcher::has_match_rule(Op_DivModL)) { 2682 DivModLNode* divmod = DivModLNode::make(this, n); 2683 d->subsume_by(divmod->div_proj(), this); 2684 n->subsume_by(divmod->mod_proj(), this); 2685 } else { 2686 // replace a%b with a-((a/b)*b) 2687 Node* mult = new (this) MulLNode(d, d->in(2)); 2688 Node* sub = new (this) SubLNode(d->in(1), mult); 2689 n->subsume_by(sub, this); 2690 } 2691 } 2692 } 2693 break; 2694 2695 case Op_LoadVector: 2696 case Op_StoreVector: 2697 break; 2698 2699 case Op_PackB: 2700 case Op_PackS: 2701 case Op_PackI: 2702 case Op_PackF: 2703 case Op_PackL: 2704 case Op_PackD: 2705 if (n->req()-1 > 2) { 2706 // Replace many operand PackNodes with a binary tree for matching 2707 PackNode* p = (PackNode*) n; 2708 Node* btp = p->binary_tree_pack(this, 1, n->req()); 2709 n->subsume_by(btp, this); 2710 } 2711 break; 2712 case Op_Loop: 2713 case Op_CountedLoop: 2714 if (n->as_Loop()->is_inner_loop()) { 2715 frc.inc_inner_loop_count(); 2716 } 2717 break; 2718 case Op_LShiftI: 2719 case Op_RShiftI: 2720 case Op_URShiftI: 2721 case Op_LShiftL: 2722 case Op_RShiftL: 2723 case Op_URShiftL: 2724 if (Matcher::need_masked_shift_count) { 2725 // The cpu's shift instructions don't restrict the count to the 2726 // lower 5/6 bits. We need to do the masking ourselves. 2727 Node* in2 = n->in(2); 2728 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 2729 const TypeInt* t = in2->find_int_type(); 2730 if (t != NULL && t->is_con()) { 2731 juint shift = t->get_con(); 2732 if (shift > mask) { // Unsigned cmp 2733 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 2734 } 2735 } else { 2736 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 2737 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 2738 n->set_req(2, shift); 2739 } 2740 } 2741 if (in2->outcnt() == 0) { // Remove dead node 2742 in2->disconnect_inputs(NULL, this); 2743 } 2744 } 2745 break; 2746 default: 2747 assert( !n->is_Call(), "" ); 2748 assert( !n->is_Mem(), "" ); 2749 break; 2750 } 2751 2752 // Collect CFG split points 2753 if (n->is_MultiBranch()) 2754 frc._tests.push(n); 2755 } 2756 2757 //------------------------------final_graph_reshaping_walk--------------------- 2758 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2759 // requires that the walk visits a node's inputs before visiting the node. 2760 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2761 ResourceArea *area = Thread::current()->resource_area(); 2762 Unique_Node_List sfpt(area); 2763 2764 frc._visited.set(root->_idx); // first, mark node as visited 2765 uint cnt = root->req(); 2766 Node *n = root; 2767 uint i = 0; 2768 while (true) { 2769 if (i < cnt) { 2770 // Place all non-visited non-null inputs onto stack 2771 Node* m = n->in(i); 2772 ++i; 2773 if (m != NULL && !frc._visited.test_set(m->_idx)) { 2774 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2775 sfpt.push(m); 2776 cnt = m->req(); 2777 nstack.push(n, i); // put on stack parent and next input's index 2778 n = m; 2779 i = 0; 2780 } 2781 } else { 2782 // Now do post-visit work 2783 final_graph_reshaping_impl( n, frc ); 2784 if (nstack.is_empty()) 2785 break; // finished 2786 n = nstack.node(); // Get node from stack 2787 cnt = n->req(); 2788 i = nstack.index(); 2789 nstack.pop(); // Shift to the next node on stack 2790 } 2791 } 2792 2793 // Skip next transformation if compressed oops are not used. 2794 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 2795 (!UseCompressedOops && !UseCompressedKlassPointers)) 2796 return; 2797 2798 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 2799 // It could be done for an uncommon traps or any safepoints/calls 2800 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 2801 while (sfpt.size() > 0) { 2802 n = sfpt.pop(); 2803 JVMState *jvms = n->as_SafePoint()->jvms(); 2804 assert(jvms != NULL, "sanity"); 2805 int start = jvms->debug_start(); 2806 int end = n->req(); 2807 bool is_uncommon = (n->is_CallStaticJava() && 2808 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2809 for (int j = start; j < end; j++) { 2810 Node* in = n->in(j); 2811 if (in->is_DecodeNarrowPtr()) { 2812 bool safe_to_skip = true; 2813 if (!is_uncommon ) { 2814 // Is it safe to skip? 2815 for (uint i = 0; i < in->outcnt(); i++) { 2816 Node* u = in->raw_out(i); 2817 if (!u->is_SafePoint() || 2818 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2819 safe_to_skip = false; 2820 } 2821 } 2822 } 2823 if (safe_to_skip) { 2824 n->set_req(j, in->in(1)); 2825 } 2826 if (in->outcnt() == 0) { 2827 in->disconnect_inputs(NULL, this); 2828 } 2829 } 2830 } 2831 } 2832 } 2833 2834 //------------------------------final_graph_reshaping-------------------------- 2835 // Final Graph Reshaping. 2836 // 2837 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2838 // and not commoned up and forced early. Must come after regular 2839 // optimizations to avoid GVN undoing the cloning. Clone constant 2840 // inputs to Loop Phis; these will be split by the allocator anyways. 2841 // Remove Opaque nodes. 2842 // (2) Move last-uses by commutative operations to the left input to encourage 2843 // Intel update-in-place two-address operations and better register usage 2844 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 2845 // calls canonicalizing them back. 2846 // (3) Count the number of double-precision FP ops, single-precision FP ops 2847 // and call sites. On Intel, we can get correct rounding either by 2848 // forcing singles to memory (requires extra stores and loads after each 2849 // FP bytecode) or we can set a rounding mode bit (requires setting and 2850 // clearing the mode bit around call sites). The mode bit is only used 2851 // if the relative frequency of single FP ops to calls is low enough. 2852 // This is a key transform for SPEC mpeg_audio. 2853 // (4) Detect infinite loops; blobs of code reachable from above but not 2854 // below. Several of the Code_Gen algorithms fail on such code shapes, 2855 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 2856 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 2857 // Detection is by looking for IfNodes where only 1 projection is 2858 // reachable from below or CatchNodes missing some targets. 2859 // (5) Assert for insane oop offsets in debug mode. 2860 2861 bool Compile::final_graph_reshaping() { 2862 // an infinite loop may have been eliminated by the optimizer, 2863 // in which case the graph will be empty. 2864 if (root()->req() == 1) { 2865 record_method_not_compilable("trivial infinite loop"); 2866 return true; 2867 } 2868 2869 Final_Reshape_Counts frc; 2870 2871 // Visit everybody reachable! 2872 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2873 Node_Stack nstack(unique() >> 1); 2874 final_graph_reshaping_walk(nstack, root(), frc); 2875 2876 // Check for unreachable (from below) code (i.e., infinite loops). 2877 for( uint i = 0; i < frc._tests.size(); i++ ) { 2878 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 2879 // Get number of CFG targets. 2880 // Note that PCTables include exception targets after calls. 2881 uint required_outcnt = n->required_outcnt(); 2882 if (n->outcnt() != required_outcnt) { 2883 // Check for a few special cases. Rethrow Nodes never take the 2884 // 'fall-thru' path, so expected kids is 1 less. 2885 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2886 if (n->in(0)->in(0)->is_Call()) { 2887 CallNode *call = n->in(0)->in(0)->as_Call(); 2888 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2889 required_outcnt--; // Rethrow always has 1 less kid 2890 } else if (call->req() > TypeFunc::Parms && 2891 call->is_CallDynamicJava()) { 2892 // Check for null receiver. In such case, the optimizer has 2893 // detected that the virtual call will always result in a null 2894 // pointer exception. The fall-through projection of this CatchNode 2895 // will not be populated. 2896 Node *arg0 = call->in(TypeFunc::Parms); 2897 if (arg0->is_Type() && 2898 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2899 required_outcnt--; 2900 } 2901 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2902 call->req() > TypeFunc::Parms+1 && 2903 call->is_CallStaticJava()) { 2904 // Check for negative array length. In such case, the optimizer has 2905 // detected that the allocation attempt will always result in an 2906 // exception. There is no fall-through projection of this CatchNode . 2907 Node *arg1 = call->in(TypeFunc::Parms+1); 2908 if (arg1->is_Type() && 2909 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2910 required_outcnt--; 2911 } 2912 } 2913 } 2914 } 2915 // Recheck with a better notion of 'required_outcnt' 2916 if (n->outcnt() != required_outcnt) { 2917 record_method_not_compilable("malformed control flow"); 2918 return true; // Not all targets reachable! 2919 } 2920 } 2921 // Check that I actually visited all kids. Unreached kids 2922 // must be infinite loops. 2923 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2924 if (!frc._visited.test(n->fast_out(j)->_idx)) { 2925 record_method_not_compilable("infinite loop"); 2926 return true; // Found unvisited kid; must be unreach 2927 } 2928 } 2929 2930 // If original bytecodes contained a mixture of floats and doubles 2931 // check if the optimizer has made it homogenous, item (3). 2932 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 2933 frc.get_float_count() > 32 && 2934 frc.get_double_count() == 0 && 2935 (10 * frc.get_call_count() < frc.get_float_count()) ) { 2936 set_24_bit_selection_and_mode( false, true ); 2937 } 2938 2939 set_java_calls(frc.get_java_call_count()); 2940 set_inner_loops(frc.get_inner_loop_count()); 2941 2942 // No infinite loops, no reason to bail out. 2943 return false; 2944 } 2945 2946 //-----------------------------too_many_traps---------------------------------- 2947 // Report if there are too many traps at the current method and bci. 2948 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2949 bool Compile::too_many_traps(ciMethod* method, 2950 int bci, 2951 Deoptimization::DeoptReason reason) { 2952 ciMethodData* md = method->method_data(); 2953 if (md->is_empty()) { 2954 // Assume the trap has not occurred, or that it occurred only 2955 // because of a transient condition during start-up in the interpreter. 2956 return false; 2957 } 2958 if (md->has_trap_at(bci, reason) != 0) { 2959 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2960 // Also, if there are multiple reasons, or if there is no per-BCI record, 2961 // assume the worst. 2962 if (log()) 2963 log()->elem("observe trap='%s' count='%d'", 2964 Deoptimization::trap_reason_name(reason), 2965 md->trap_count(reason)); 2966 return true; 2967 } else { 2968 // Ignore method/bci and see if there have been too many globally. 2969 return too_many_traps(reason, md); 2970 } 2971 } 2972 2973 // Less-accurate variant which does not require a method and bci. 2974 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2975 ciMethodData* logmd) { 2976 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2977 // Too many traps globally. 2978 // Note that we use cumulative trap_count, not just md->trap_count. 2979 if (log()) { 2980 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2981 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2982 Deoptimization::trap_reason_name(reason), 2983 mcount, trap_count(reason)); 2984 } 2985 return true; 2986 } else { 2987 // The coast is clear. 2988 return false; 2989 } 2990 } 2991 2992 //--------------------------too_many_recompiles-------------------------------- 2993 // Report if there are too many recompiles at the current method and bci. 2994 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2995 // Is not eager to return true, since this will cause the compiler to use 2996 // Action_none for a trap point, to avoid too many recompilations. 2997 bool Compile::too_many_recompiles(ciMethod* method, 2998 int bci, 2999 Deoptimization::DeoptReason reason) { 3000 ciMethodData* md = method->method_data(); 3001 if (md->is_empty()) { 3002 // Assume the trap has not occurred, or that it occurred only 3003 // because of a transient condition during start-up in the interpreter. 3004 return false; 3005 } 3006 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3007 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3008 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3009 Deoptimization::DeoptReason per_bc_reason 3010 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3011 if ((per_bc_reason == Deoptimization::Reason_none 3012 || md->has_trap_at(bci, reason) != 0) 3013 // The trap frequency measure we care about is the recompile count: 3014 && md->trap_recompiled_at(bci) 3015 && md->overflow_recompile_count() >= bc_cutoff) { 3016 // Do not emit a trap here if it has already caused recompilations. 3017 // Also, if there are multiple reasons, or if there is no per-BCI record, 3018 // assume the worst. 3019 if (log()) 3020 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3021 Deoptimization::trap_reason_name(reason), 3022 md->trap_count(reason), 3023 md->overflow_recompile_count()); 3024 return true; 3025 } else if (trap_count(reason) != 0 3026 && decompile_count() >= m_cutoff) { 3027 // Too many recompiles globally, and we have seen this sort of trap. 3028 // Use cumulative decompile_count, not just md->decompile_count. 3029 if (log()) 3030 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3031 Deoptimization::trap_reason_name(reason), 3032 md->trap_count(reason), trap_count(reason), 3033 md->decompile_count(), decompile_count()); 3034 return true; 3035 } else { 3036 // The coast is clear. 3037 return false; 3038 } 3039 } 3040 3041 3042 #ifndef PRODUCT 3043 //------------------------------verify_graph_edges--------------------------- 3044 // Walk the Graph and verify that there is a one-to-one correspondence 3045 // between Use-Def edges and Def-Use edges in the graph. 3046 void Compile::verify_graph_edges(bool no_dead_code) { 3047 if (VerifyGraphEdges) { 3048 ResourceArea *area = Thread::current()->resource_area(); 3049 Unique_Node_List visited(area); 3050 // Call recursive graph walk to check edges 3051 _root->verify_edges(visited); 3052 if (no_dead_code) { 3053 // Now make sure that no visited node is used by an unvisited node. 3054 bool dead_nodes = 0; 3055 Unique_Node_List checked(area); 3056 while (visited.size() > 0) { 3057 Node* n = visited.pop(); 3058 checked.push(n); 3059 for (uint i = 0; i < n->outcnt(); i++) { 3060 Node* use = n->raw_out(i); 3061 if (checked.member(use)) continue; // already checked 3062 if (visited.member(use)) continue; // already in the graph 3063 if (use->is_Con()) continue; // a dead ConNode is OK 3064 // At this point, we have found a dead node which is DU-reachable. 3065 if (dead_nodes++ == 0) 3066 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3067 use->dump(2); 3068 tty->print_cr("---"); 3069 checked.push(use); // No repeats; pretend it is now checked. 3070 } 3071 } 3072 assert(dead_nodes == 0, "using nodes must be reachable from root"); 3073 } 3074 } 3075 } 3076 #endif 3077 3078 // The Compile object keeps track of failure reasons separately from the ciEnv. 3079 // This is required because there is not quite a 1-1 relation between the 3080 // ciEnv and its compilation task and the Compile object. Note that one 3081 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3082 // to backtrack and retry without subsuming loads. Other than this backtracking 3083 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3084 // by the logic in C2Compiler. 3085 void Compile::record_failure(const char* reason) { 3086 if (log() != NULL) { 3087 log()->elem("failure reason='%s' phase='compile'", reason); 3088 } 3089 if (_failure_reason == NULL) { 3090 // Record the first failure reason. 3091 _failure_reason = reason; 3092 } 3093 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3094 C->print_method(_failure_reason); 3095 } 3096 _root = NULL; // flush the graph, too 3097 } 3098 3099 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3100 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3101 _phase_name(name), _dolog(dolog) 3102 { 3103 if (dolog) { 3104 C = Compile::current(); 3105 _log = C->log(); 3106 } else { 3107 C = NULL; 3108 _log = NULL; 3109 } 3110 if (_log != NULL) { 3111 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3112 _log->stamp(); 3113 _log->end_head(); 3114 } 3115 } 3116 3117 Compile::TracePhase::~TracePhase() { 3118 3119 C = Compile::current(); 3120 if (_dolog) { 3121 _log = C->log(); 3122 } else { 3123 _log = NULL; 3124 } 3125 3126 #ifdef ASSERT 3127 if (PrintIdealNodeCount) { 3128 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3129 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3130 } 3131 3132 if (VerifyIdealNodeCount) { 3133 Compile::current()->print_missing_nodes(); 3134 } 3135 #endif 3136 3137 if (_log != NULL) { 3138 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3139 } 3140 } 3141 3142 //============================================================================= 3143 // Two Constant's are equal when the type and the value are equal. 3144 bool Compile::Constant::operator==(const Constant& other) { 3145 if (type() != other.type() ) return false; 3146 if (can_be_reused() != other.can_be_reused()) return false; 3147 // For floating point values we compare the bit pattern. 3148 switch (type()) { 3149 case T_FLOAT: return (_v._value.i == other._v._value.i); 3150 case T_LONG: 3151 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3152 case T_OBJECT: 3153 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3154 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3155 case T_METADATA: return (_v._metadata == other._v._metadata); 3156 default: ShouldNotReachHere(); 3157 } 3158 return false; 3159 } 3160 3161 static int type_to_size_in_bytes(BasicType t) { 3162 switch (t) { 3163 case T_LONG: return sizeof(jlong ); 3164 case T_FLOAT: return sizeof(jfloat ); 3165 case T_DOUBLE: return sizeof(jdouble); 3166 case T_METADATA: return sizeof(Metadata*); 3167 // We use T_VOID as marker for jump-table entries (labels) which 3168 // need an internal word relocation. 3169 case T_VOID: 3170 case T_ADDRESS: 3171 case T_OBJECT: return sizeof(jobject); 3172 } 3173 3174 ShouldNotReachHere(); 3175 return -1; 3176 } 3177 3178 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3179 // sort descending 3180 if (a->freq() > b->freq()) return -1; 3181 if (a->freq() < b->freq()) return 1; 3182 return 0; 3183 } 3184 3185 void Compile::ConstantTable::calculate_offsets_and_size() { 3186 // First, sort the array by frequencies. 3187 _constants.sort(qsort_comparator); 3188 3189 #ifdef ASSERT 3190 // Make sure all jump-table entries were sorted to the end of the 3191 // array (they have a negative frequency). 3192 bool found_void = false; 3193 for (int i = 0; i < _constants.length(); i++) { 3194 Constant con = _constants.at(i); 3195 if (con.type() == T_VOID) 3196 found_void = true; // jump-tables 3197 else 3198 assert(!found_void, "wrong sorting"); 3199 } 3200 #endif 3201 3202 int offset = 0; 3203 for (int i = 0; i < _constants.length(); i++) { 3204 Constant* con = _constants.adr_at(i); 3205 3206 // Align offset for type. 3207 int typesize = type_to_size_in_bytes(con->type()); 3208 offset = align_size_up(offset, typesize); 3209 con->set_offset(offset); // set constant's offset 3210 3211 if (con->type() == T_VOID) { 3212 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3213 offset = offset + typesize * n->outcnt(); // expand jump-table 3214 } else { 3215 offset = offset + typesize; 3216 } 3217 } 3218 3219 // Align size up to the next section start (which is insts; see 3220 // CodeBuffer::align_at_start). 3221 assert(_size == -1, "already set?"); 3222 _size = align_size_up(offset, CodeEntryAlignment); 3223 } 3224 3225 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3226 MacroAssembler _masm(&cb); 3227 for (int i = 0; i < _constants.length(); i++) { 3228 Constant con = _constants.at(i); 3229 address constant_addr; 3230 switch (con.type()) { 3231 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3232 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3233 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3234 case T_OBJECT: { 3235 jobject obj = con.get_jobject(); 3236 int oop_index = _masm.oop_recorder()->find_index(obj); 3237 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3238 break; 3239 } 3240 case T_ADDRESS: { 3241 address addr = (address) con.get_jobject(); 3242 constant_addr = _masm.address_constant(addr); 3243 break; 3244 } 3245 // We use T_VOID as marker for jump-table entries (labels) which 3246 // need an internal word relocation. 3247 case T_VOID: { 3248 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3249 // Fill the jump-table with a dummy word. The real value is 3250 // filled in later in fill_jump_table. 3251 address dummy = (address) n; 3252 constant_addr = _masm.address_constant(dummy); 3253 // Expand jump-table 3254 for (uint i = 1; i < n->outcnt(); i++) { 3255 address temp_addr = _masm.address_constant(dummy + i); 3256 assert(temp_addr, "consts section too small"); 3257 } 3258 break; 3259 } 3260 case T_METADATA: { 3261 Metadata* obj = con.get_metadata(); 3262 int metadata_index = _masm.oop_recorder()->find_index(obj); 3263 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3264 break; 3265 } 3266 default: ShouldNotReachHere(); 3267 } 3268 assert(constant_addr, "consts section too small"); 3269 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3270 } 3271 } 3272 3273 int Compile::ConstantTable::find_offset(Constant& con) const { 3274 int idx = _constants.find(con); 3275 assert(idx != -1, "constant must be in constant table"); 3276 int offset = _constants.at(idx).offset(); 3277 assert(offset != -1, "constant table not emitted yet?"); 3278 return offset; 3279 } 3280 3281 void Compile::ConstantTable::add(Constant& con) { 3282 if (con.can_be_reused()) { 3283 int idx = _constants.find(con); 3284 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3285 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3286 return; 3287 } 3288 } 3289 (void) _constants.append(con); 3290 } 3291 3292 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3293 Block* b = Compile::current()->cfg()->_bbs[n->_idx]; 3294 Constant con(type, value, b->_freq); 3295 add(con); 3296 return con; 3297 } 3298 3299 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3300 Constant con(metadata); 3301 add(con); 3302 return con; 3303 } 3304 3305 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3306 jvalue value; 3307 BasicType type = oper->type()->basic_type(); 3308 switch (type) { 3309 case T_LONG: value.j = oper->constantL(); break; 3310 case T_FLOAT: value.f = oper->constantF(); break; 3311 case T_DOUBLE: value.d = oper->constantD(); break; 3312 case T_OBJECT: 3313 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3314 case T_METADATA: return add((Metadata*)oper->constant()); break; 3315 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3316 } 3317 return add(n, type, value); 3318 } 3319 3320 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3321 jvalue value; 3322 // We can use the node pointer here to identify the right jump-table 3323 // as this method is called from Compile::Fill_buffer right before 3324 // the MachNodes are emitted and the jump-table is filled (means the 3325 // MachNode pointers do not change anymore). 3326 value.l = (jobject) n; 3327 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3328 add(con); 3329 return con; 3330 } 3331 3332 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3333 // If called from Compile::scratch_emit_size do nothing. 3334 if (Compile::current()->in_scratch_emit_size()) return; 3335 3336 assert(labels.is_nonempty(), "must be"); 3337 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3338 3339 // Since MachConstantNode::constant_offset() also contains 3340 // table_base_offset() we need to subtract the table_base_offset() 3341 // to get the plain offset into the constant table. 3342 int offset = n->constant_offset() - table_base_offset(); 3343 3344 MacroAssembler _masm(&cb); 3345 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3346 3347 for (uint i = 0; i < n->outcnt(); i++) { 3348 address* constant_addr = &jump_table_base[i]; 3349 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i))); 3350 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3351 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3352 } 3353 }