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 _print_inlining_list(NULL), 615 _print_inlining(0) { 616 C = this; 617 618 CompileWrapper cw(this); 619 #ifndef PRODUCT 620 if (TimeCompiler2) { 621 tty->print(" "); 622 target->holder()->name()->print(); 623 tty->print("."); 624 target->print_short_name(); 625 tty->print(" "); 626 } 627 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 628 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 629 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 630 if (!print_opto_assembly) { 631 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 632 if (print_assembly && !Disassembler::can_decode()) { 633 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 634 print_opto_assembly = true; 635 } 636 } 637 set_print_assembly(print_opto_assembly); 638 set_parsed_irreducible_loop(false); 639 #endif 640 641 if (ProfileTraps) { 642 // Make sure the method being compiled gets its own MDO, 643 // so we can at least track the decompile_count(). 644 method()->ensure_method_data(); 645 } 646 647 Init(::AliasLevel); 648 649 650 print_compile_messages(); 651 652 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 653 _ilt = InlineTree::build_inline_tree_root(); 654 else 655 _ilt = NULL; 656 657 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 658 assert(num_alias_types() >= AliasIdxRaw, ""); 659 660 #define MINIMUM_NODE_HASH 1023 661 // Node list that Iterative GVN will start with 662 Unique_Node_List for_igvn(comp_arena()); 663 set_for_igvn(&for_igvn); 664 665 // GVN that will be run immediately on new nodes 666 uint estimated_size = method()->code_size()*4+64; 667 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 668 PhaseGVN gvn(node_arena(), estimated_size); 669 set_initial_gvn(&gvn); 670 671 if (PrintInlining) { 672 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningEvent>(comp_arena(), 1, 1, PrintInliningEvent()); 673 } 674 { // Scope for timing the parser 675 TracePhase t3("parse", &_t_parser, true); 676 677 // Put top into the hash table ASAP. 678 initial_gvn()->transform_no_reclaim(top()); 679 680 // Set up tf(), start(), and find a CallGenerator. 681 CallGenerator* cg = NULL; 682 if (is_osr_compilation()) { 683 const TypeTuple *domain = StartOSRNode::osr_domain(); 684 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 685 init_tf(TypeFunc::make(domain, range)); 686 StartNode* s = new (this) StartOSRNode(root(), domain); 687 initial_gvn()->set_type_bottom(s); 688 init_start(s); 689 cg = CallGenerator::for_osr(method(), entry_bci()); 690 } else { 691 // Normal case. 692 init_tf(TypeFunc::make(method())); 693 StartNode* s = new (this) StartNode(root(), tf()->domain()); 694 initial_gvn()->set_type_bottom(s); 695 init_start(s); 696 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 697 // With java.lang.ref.reference.get() we must go through the 698 // intrinsic when G1 is enabled - even when get() is the root 699 // method of the compile - so that, if necessary, the value in 700 // the referent field of the reference object gets recorded by 701 // the pre-barrier code. 702 // Specifically, if G1 is enabled, the value in the referent 703 // field is recorded by the G1 SATB pre barrier. This will 704 // result in the referent being marked live and the reference 705 // object removed from the list of discovered references during 706 // reference processing. 707 cg = find_intrinsic(method(), false); 708 } 709 if (cg == NULL) { 710 float past_uses = method()->interpreter_invocation_count(); 711 float expected_uses = past_uses; 712 cg = CallGenerator::for_inline(method(), expected_uses); 713 } 714 } 715 if (failing()) return; 716 if (cg == NULL) { 717 record_method_not_compilable_all_tiers("cannot parse method"); 718 return; 719 } 720 JVMState* jvms = build_start_state(start(), tf()); 721 if ((jvms = cg->generate(jvms)) == NULL) { 722 record_method_not_compilable("method parse failed"); 723 return; 724 } 725 GraphKit kit(jvms); 726 727 if (!kit.stopped()) { 728 // Accept return values, and transfer control we know not where. 729 // This is done by a special, unique ReturnNode bound to root. 730 return_values(kit.jvms()); 731 } 732 733 if (kit.has_exceptions()) { 734 // Any exceptions that escape from this call must be rethrown 735 // to whatever caller is dynamically above us on the stack. 736 // This is done by a special, unique RethrowNode bound to root. 737 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 738 } 739 740 if (!failing() && has_stringbuilder()) { 741 { 742 // remove useless nodes to make the usage analysis simpler 743 ResourceMark rm; 744 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 745 } 746 747 { 748 ResourceMark rm; 749 print_method("Before StringOpts", 3); 750 PhaseStringOpts pso(initial_gvn(), &for_igvn); 751 print_method("After StringOpts", 3); 752 } 753 754 // now inline anything that we skipped the first time around 755 while (_late_inlines.length() > 0) { 756 CallGenerator* cg = _late_inlines.pop(); 757 cg->do_late_inline(); 758 if (failing()) return; 759 } 760 } 761 assert(_late_inlines.length() == 0, "should have been processed"); 762 dump_inlining(); 763 764 print_method("Before RemoveUseless", 3); 765 766 // Remove clutter produced by parsing. 767 if (!failing()) { 768 ResourceMark rm; 769 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 770 } 771 } 772 773 // Note: Large methods are capped off in do_one_bytecode(). 774 if (failing()) return; 775 776 // After parsing, node notes are no longer automagic. 777 // They must be propagated by register_new_node_with_optimizer(), 778 // clone(), or the like. 779 set_default_node_notes(NULL); 780 781 for (;;) { 782 int successes = Inline_Warm(); 783 if (failing()) return; 784 if (successes == 0) break; 785 } 786 787 // Drain the list. 788 Finish_Warm(); 789 #ifndef PRODUCT 790 if (_printer) { 791 _printer->print_inlining(this); 792 } 793 #endif 794 795 if (failing()) return; 796 NOT_PRODUCT( verify_graph_edges(); ) 797 798 // Now optimize 799 Optimize(); 800 if (failing()) return; 801 NOT_PRODUCT( verify_graph_edges(); ) 802 803 #ifndef PRODUCT 804 if (PrintIdeal) { 805 ttyLocker ttyl; // keep the following output all in one block 806 // This output goes directly to the tty, not the compiler log. 807 // To enable tools to match it up with the compilation activity, 808 // be sure to tag this tty output with the compile ID. 809 if (xtty != NULL) { 810 xtty->head("ideal compile_id='%d'%s", compile_id(), 811 is_osr_compilation() ? " compile_kind='osr'" : 812 ""); 813 } 814 root()->dump(9999); 815 if (xtty != NULL) { 816 xtty->tail("ideal"); 817 } 818 } 819 #endif 820 821 // Now that we know the size of all the monitors we can add a fixed slot 822 // for the original deopt pc. 823 824 _orig_pc_slot = fixed_slots(); 825 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 826 set_fixed_slots(next_slot); 827 828 // Now generate code 829 Code_Gen(); 830 if (failing()) return; 831 832 // Check if we want to skip execution of all compiled code. 833 { 834 #ifndef PRODUCT 835 if (OptoNoExecute) { 836 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 837 return; 838 } 839 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 840 #endif 841 842 if (is_osr_compilation()) { 843 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 844 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 845 } else { 846 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 847 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 848 } 849 850 env()->register_method(_method, _entry_bci, 851 &_code_offsets, 852 _orig_pc_slot_offset_in_bytes, 853 code_buffer(), 854 frame_size_in_words(), _oop_map_set, 855 &_handler_table, &_inc_table, 856 compiler, 857 env()->comp_level(), 858 has_unsafe_access(), 859 SharedRuntime::is_wide_vector(max_vector_size()) 860 ); 861 862 if (log() != NULL) // Print code cache state into compiler log 863 log()->code_cache_state(); 864 } 865 } 866 867 //------------------------------Compile---------------------------------------- 868 // Compile a runtime stub 869 Compile::Compile( ciEnv* ci_env, 870 TypeFunc_generator generator, 871 address stub_function, 872 const char *stub_name, 873 int is_fancy_jump, 874 bool pass_tls, 875 bool save_arg_registers, 876 bool return_pc ) 877 : Phase(Compiler), 878 _env(ci_env), 879 _log(ci_env->log()), 880 _compile_id(-1), 881 _save_argument_registers(save_arg_registers), 882 _method(NULL), 883 _stub_name(stub_name), 884 _stub_function(stub_function), 885 _stub_entry_point(NULL), 886 _entry_bci(InvocationEntryBci), 887 _initial_gvn(NULL), 888 _for_igvn(NULL), 889 _warm_calls(NULL), 890 _orig_pc_slot(0), 891 _orig_pc_slot_offset_in_bytes(0), 892 _subsume_loads(true), 893 _do_escape_analysis(false), 894 _failure_reason(NULL), 895 _code_buffer("Compile::Fill_buffer"), 896 _has_method_handle_invokes(false), 897 _mach_constant_base_node(NULL), 898 _node_bundling_limit(0), 899 _node_bundling_base(NULL), 900 _java_calls(0), 901 _inner_loops(0), 902 #ifndef PRODUCT 903 _trace_opto_output(TraceOptoOutput), 904 _printer(NULL), 905 #endif 906 _dead_node_list(comp_arena()), 907 _dead_node_count(0), 908 _congraph(NULL), 909 _print_inlining_list(NULL), 910 _print_inlining(0) { 911 C = this; 912 913 #ifndef PRODUCT 914 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 915 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 916 set_print_assembly(PrintFrameConverterAssembly); 917 set_parsed_irreducible_loop(false); 918 #endif 919 CompileWrapper cw(this); 920 Init(/*AliasLevel=*/ 0); 921 init_tf((*generator)()); 922 923 { 924 // The following is a dummy for the sake of GraphKit::gen_stub 925 Unique_Node_List for_igvn(comp_arena()); 926 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 927 PhaseGVN gvn(Thread::current()->resource_area(),255); 928 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 929 gvn.transform_no_reclaim(top()); 930 931 GraphKit kit; 932 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 933 } 934 935 NOT_PRODUCT( verify_graph_edges(); ) 936 Code_Gen(); 937 if (failing()) return; 938 939 940 // Entry point will be accessed using compile->stub_entry_point(); 941 if (code_buffer() == NULL) { 942 Matcher::soft_match_failure(); 943 } else { 944 if (PrintAssembly && (WizardMode || Verbose)) 945 tty->print_cr("### Stub::%s", stub_name); 946 947 if (!failing()) { 948 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 949 950 // Make the NMethod 951 // For now we mark the frame as never safe for profile stackwalking 952 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 953 code_buffer(), 954 CodeOffsets::frame_never_safe, 955 // _code_offsets.value(CodeOffsets::Frame_Complete), 956 frame_size_in_words(), 957 _oop_map_set, 958 save_arg_registers); 959 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 960 961 _stub_entry_point = rs->entry_point(); 962 } 963 } 964 } 965 966 //------------------------------Init------------------------------------------- 967 // Prepare for a single compilation 968 void Compile::Init(int aliaslevel) { 969 _unique = 0; 970 _regalloc = NULL; 971 972 _tf = NULL; // filled in later 973 _top = NULL; // cached later 974 _matcher = NULL; // filled in later 975 _cfg = NULL; // filled in later 976 977 set_24_bit_selection_and_mode(Use24BitFP, false); 978 979 _node_note_array = NULL; 980 _default_node_notes = NULL; 981 982 _immutable_memory = NULL; // filled in at first inquiry 983 984 // Globally visible Nodes 985 // First set TOP to NULL to give safe behavior during creation of RootNode 986 set_cached_top_node(NULL); 987 set_root(new (this) RootNode()); 988 // Now that you have a Root to point to, create the real TOP 989 set_cached_top_node( new (this) ConNode(Type::TOP) ); 990 set_recent_alloc(NULL, NULL); 991 992 // Create Debug Information Recorder to record scopes, oopmaps, etc. 993 env()->set_oop_recorder(new OopRecorder(env()->arena())); 994 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 995 env()->set_dependencies(new Dependencies(env())); 996 997 _fixed_slots = 0; 998 set_has_split_ifs(false); 999 set_has_loops(has_method() && method()->has_loops()); // first approximation 1000 set_has_stringbuilder(false); 1001 _trap_can_recompile = false; // no traps emitted yet 1002 _major_progress = true; // start out assuming good things will happen 1003 set_has_unsafe_access(false); 1004 set_max_vector_size(0); 1005 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1006 set_decompile_count(0); 1007 1008 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1009 set_num_loop_opts(LoopOptsCount); 1010 set_do_inlining(Inline); 1011 set_max_inline_size(MaxInlineSize); 1012 set_freq_inline_size(FreqInlineSize); 1013 set_do_scheduling(OptoScheduling); 1014 set_do_count_invocations(false); 1015 set_do_method_data_update(false); 1016 1017 if (debug_info()->recording_non_safepoints()) { 1018 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1019 (comp_arena(), 8, 0, NULL)); 1020 set_default_node_notes(Node_Notes::make(this)); 1021 } 1022 1023 // // -- Initialize types before each compile -- 1024 // // Update cached type information 1025 // if( _method && _method->constants() ) 1026 // Type::update_loaded_types(_method, _method->constants()); 1027 1028 // Init alias_type map. 1029 if (!_do_escape_analysis && aliaslevel == 3) 1030 aliaslevel = 2; // No unique types without escape analysis 1031 _AliasLevel = aliaslevel; 1032 const int grow_ats = 16; 1033 _max_alias_types = grow_ats; 1034 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1035 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1036 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1037 { 1038 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1039 } 1040 // Initialize the first few types. 1041 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1042 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1043 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1044 _num_alias_types = AliasIdxRaw+1; 1045 // Zero out the alias type cache. 1046 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1047 // A NULL adr_type hits in the cache right away. Preload the right answer. 1048 probe_alias_cache(NULL)->_index = AliasIdxTop; 1049 1050 _intrinsics = NULL; 1051 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1052 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1053 register_library_intrinsics(); 1054 } 1055 1056 //---------------------------init_start---------------------------------------- 1057 // Install the StartNode on this compile object. 1058 void Compile::init_start(StartNode* s) { 1059 if (failing()) 1060 return; // already failing 1061 assert(s == start(), ""); 1062 } 1063 1064 StartNode* Compile::start() const { 1065 assert(!failing(), ""); 1066 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1067 Node* start = root()->fast_out(i); 1068 if( start->is_Start() ) 1069 return start->as_Start(); 1070 } 1071 ShouldNotReachHere(); 1072 return NULL; 1073 } 1074 1075 //-------------------------------immutable_memory------------------------------------- 1076 // Access immutable memory 1077 Node* Compile::immutable_memory() { 1078 if (_immutable_memory != NULL) { 1079 return _immutable_memory; 1080 } 1081 StartNode* s = start(); 1082 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1083 Node *p = s->fast_out(i); 1084 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1085 _immutable_memory = p; 1086 return _immutable_memory; 1087 } 1088 } 1089 ShouldNotReachHere(); 1090 return NULL; 1091 } 1092 1093 //----------------------set_cached_top_node------------------------------------ 1094 // Install the cached top node, and make sure Node::is_top works correctly. 1095 void Compile::set_cached_top_node(Node* tn) { 1096 if (tn != NULL) verify_top(tn); 1097 Node* old_top = _top; 1098 _top = tn; 1099 // Calling Node::setup_is_top allows the nodes the chance to adjust 1100 // their _out arrays. 1101 if (_top != NULL) _top->setup_is_top(); 1102 if (old_top != NULL) old_top->setup_is_top(); 1103 assert(_top == NULL || top()->is_top(), ""); 1104 } 1105 1106 #ifdef ASSERT 1107 uint Compile::count_live_nodes_by_graph_walk() { 1108 Unique_Node_List useful(comp_arena()); 1109 // Get useful node list by walking the graph. 1110 identify_useful_nodes(useful); 1111 return useful.size(); 1112 } 1113 1114 void Compile::print_missing_nodes() { 1115 1116 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1117 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1118 return; 1119 } 1120 1121 // This is an expensive function. It is executed only when the user 1122 // specifies VerifyIdealNodeCount option or otherwise knows the 1123 // additional work that needs to be done to identify reachable nodes 1124 // by walking the flow graph and find the missing ones using 1125 // _dead_node_list. 1126 1127 Unique_Node_List useful(comp_arena()); 1128 // Get useful node list by walking the graph. 1129 identify_useful_nodes(useful); 1130 1131 uint l_nodes = C->live_nodes(); 1132 uint l_nodes_by_walk = useful.size(); 1133 1134 if (l_nodes != l_nodes_by_walk) { 1135 if (_log != NULL) { 1136 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1137 _log->stamp(); 1138 _log->end_head(); 1139 } 1140 VectorSet& useful_member_set = useful.member_set(); 1141 int last_idx = l_nodes_by_walk; 1142 for (int i = 0; i < last_idx; i++) { 1143 if (useful_member_set.test(i)) { 1144 if (_dead_node_list.test(i)) { 1145 if (_log != NULL) { 1146 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1147 } 1148 if (PrintIdealNodeCount) { 1149 // Print the log message to tty 1150 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1151 useful.at(i)->dump(); 1152 } 1153 } 1154 } 1155 else if (! _dead_node_list.test(i)) { 1156 if (_log != NULL) { 1157 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1158 } 1159 if (PrintIdealNodeCount) { 1160 // Print the log message to tty 1161 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1162 } 1163 } 1164 } 1165 if (_log != NULL) { 1166 _log->tail("mismatched_nodes"); 1167 } 1168 } 1169 } 1170 #endif 1171 1172 #ifndef PRODUCT 1173 void Compile::verify_top(Node* tn) const { 1174 if (tn != NULL) { 1175 assert(tn->is_Con(), "top node must be a constant"); 1176 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1177 assert(tn->in(0) != NULL, "must have live top node"); 1178 } 1179 } 1180 #endif 1181 1182 1183 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1184 1185 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1186 guarantee(arr != NULL, ""); 1187 int num_blocks = arr->length(); 1188 if (grow_by < num_blocks) grow_by = num_blocks; 1189 int num_notes = grow_by * _node_notes_block_size; 1190 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1191 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1192 while (num_notes > 0) { 1193 arr->append(notes); 1194 notes += _node_notes_block_size; 1195 num_notes -= _node_notes_block_size; 1196 } 1197 assert(num_notes == 0, "exact multiple, please"); 1198 } 1199 1200 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1201 if (source == NULL || dest == NULL) return false; 1202 1203 if (dest->is_Con()) 1204 return false; // Do not push debug info onto constants. 1205 1206 #ifdef ASSERT 1207 // Leave a bread crumb trail pointing to the original node: 1208 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1209 dest->set_debug_orig(source); 1210 } 1211 #endif 1212 1213 if (node_note_array() == NULL) 1214 return false; // Not collecting any notes now. 1215 1216 // This is a copy onto a pre-existing node, which may already have notes. 1217 // If both nodes have notes, do not overwrite any pre-existing notes. 1218 Node_Notes* source_notes = node_notes_at(source->_idx); 1219 if (source_notes == NULL || source_notes->is_clear()) return false; 1220 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1221 if (dest_notes == NULL || dest_notes->is_clear()) { 1222 return set_node_notes_at(dest->_idx, source_notes); 1223 } 1224 1225 Node_Notes merged_notes = (*source_notes); 1226 // The order of operations here ensures that dest notes will win... 1227 merged_notes.update_from(dest_notes); 1228 return set_node_notes_at(dest->_idx, &merged_notes); 1229 } 1230 1231 1232 //--------------------------allow_range_check_smearing------------------------- 1233 // Gating condition for coalescing similar range checks. 1234 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1235 // single covering check that is at least as strong as any of them. 1236 // If the optimization succeeds, the simplified (strengthened) range check 1237 // will always succeed. If it fails, we will deopt, and then give up 1238 // on the optimization. 1239 bool Compile::allow_range_check_smearing() const { 1240 // If this method has already thrown a range-check, 1241 // assume it was because we already tried range smearing 1242 // and it failed. 1243 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1244 return !already_trapped; 1245 } 1246 1247 1248 //------------------------------flatten_alias_type----------------------------- 1249 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1250 int offset = tj->offset(); 1251 TypePtr::PTR ptr = tj->ptr(); 1252 1253 // Known instance (scalarizable allocation) alias only with itself. 1254 bool is_known_inst = tj->isa_oopptr() != NULL && 1255 tj->is_oopptr()->is_known_instance(); 1256 1257 // Process weird unsafe references. 1258 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1259 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1260 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1261 tj = TypeOopPtr::BOTTOM; 1262 ptr = tj->ptr(); 1263 offset = tj->offset(); 1264 } 1265 1266 // Array pointers need some flattening 1267 const TypeAryPtr *ta = tj->isa_aryptr(); 1268 if( ta && is_known_inst ) { 1269 if ( offset != Type::OffsetBot && 1270 offset > arrayOopDesc::length_offset_in_bytes() ) { 1271 offset = Type::OffsetBot; // Flatten constant access into array body only 1272 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1273 } 1274 } else if( ta && _AliasLevel >= 2 ) { 1275 // For arrays indexed by constant indices, we flatten the alias 1276 // space to include all of the array body. Only the header, klass 1277 // and array length can be accessed un-aliased. 1278 if( offset != Type::OffsetBot ) { 1279 if( ta->const_oop() ) { // MethodData* or Method* 1280 offset = Type::OffsetBot; // Flatten constant access into array body 1281 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1282 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1283 // range is OK as-is. 1284 tj = ta = TypeAryPtr::RANGE; 1285 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1286 tj = TypeInstPtr::KLASS; // all klass loads look alike 1287 ta = TypeAryPtr::RANGE; // generic ignored junk 1288 ptr = TypePtr::BotPTR; 1289 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1290 tj = TypeInstPtr::MARK; 1291 ta = TypeAryPtr::RANGE; // generic ignored junk 1292 ptr = TypePtr::BotPTR; 1293 } else { // Random constant offset into array body 1294 offset = Type::OffsetBot; // Flatten constant access into array body 1295 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1296 } 1297 } 1298 // Arrays of fixed size alias with arrays of unknown size. 1299 if (ta->size() != TypeInt::POS) { 1300 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1301 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1302 } 1303 // Arrays of known objects become arrays of unknown objects. 1304 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1305 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1306 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1307 } 1308 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1309 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1310 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1311 } 1312 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1313 // cannot be distinguished by bytecode alone. 1314 if (ta->elem() == TypeInt::BOOL) { 1315 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1316 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1317 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1318 } 1319 // During the 2nd round of IterGVN, NotNull castings are removed. 1320 // Make sure the Bottom and NotNull variants alias the same. 1321 // Also, make sure exact and non-exact variants alias the same. 1322 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1323 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1324 } 1325 } 1326 1327 // Oop pointers need some flattening 1328 const TypeInstPtr *to = tj->isa_instptr(); 1329 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1330 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1331 if( ptr == TypePtr::Constant ) { 1332 if (to->klass() != ciEnv::current()->Class_klass() || 1333 offset < k->size_helper() * wordSize) { 1334 // No constant oop pointers (such as Strings); they alias with 1335 // unknown strings. 1336 assert(!is_known_inst, "not scalarizable allocation"); 1337 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1338 } 1339 } else if( is_known_inst ) { 1340 tj = to; // Keep NotNull and klass_is_exact for instance type 1341 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1342 // During the 2nd round of IterGVN, NotNull castings are removed. 1343 // Make sure the Bottom and NotNull variants alias the same. 1344 // Also, make sure exact and non-exact variants alias the same. 1345 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1346 } 1347 // Canonicalize the holder of this field 1348 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1349 // First handle header references such as a LoadKlassNode, even if the 1350 // object's klass is unloaded at compile time (4965979). 1351 if (!is_known_inst) { // Do it only for non-instance types 1352 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1353 } 1354 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1355 // Static fields are in the space above the normal instance 1356 // fields in the java.lang.Class instance. 1357 if (to->klass() != ciEnv::current()->Class_klass()) { 1358 to = NULL; 1359 tj = TypeOopPtr::BOTTOM; 1360 offset = tj->offset(); 1361 } 1362 } else { 1363 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1364 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1365 if( is_known_inst ) { 1366 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1367 } else { 1368 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1369 } 1370 } 1371 } 1372 } 1373 1374 // Klass pointers to object array klasses need some flattening 1375 const TypeKlassPtr *tk = tj->isa_klassptr(); 1376 if( tk ) { 1377 // If we are referencing a field within a Klass, we need 1378 // to assume the worst case of an Object. Both exact and 1379 // inexact types must flatten to the same alias class so 1380 // use NotNull as the PTR. 1381 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1382 1383 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1384 TypeKlassPtr::OBJECT->klass(), 1385 offset); 1386 } 1387 1388 ciKlass* klass = tk->klass(); 1389 if( klass->is_obj_array_klass() ) { 1390 ciKlass* k = TypeAryPtr::OOPS->klass(); 1391 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1392 k = TypeInstPtr::BOTTOM->klass(); 1393 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1394 } 1395 1396 // Check for precise loads from the primary supertype array and force them 1397 // to the supertype cache alias index. Check for generic array loads from 1398 // the primary supertype array and also force them to the supertype cache 1399 // alias index. Since the same load can reach both, we need to merge 1400 // these 2 disparate memories into the same alias class. Since the 1401 // primary supertype array is read-only, there's no chance of confusion 1402 // where we bypass an array load and an array store. 1403 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1404 if (offset == Type::OffsetBot || 1405 (offset >= primary_supers_offset && 1406 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1407 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1408 offset = in_bytes(Klass::secondary_super_cache_offset()); 1409 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1410 } 1411 } 1412 1413 // Flatten all Raw pointers together. 1414 if (tj->base() == Type::RawPtr) 1415 tj = TypeRawPtr::BOTTOM; 1416 1417 if (tj->base() == Type::AnyPtr) 1418 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1419 1420 // Flatten all to bottom for now 1421 switch( _AliasLevel ) { 1422 case 0: 1423 tj = TypePtr::BOTTOM; 1424 break; 1425 case 1: // Flatten to: oop, static, field or array 1426 switch (tj->base()) { 1427 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1428 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1429 case Type::AryPtr: // do not distinguish arrays at all 1430 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1431 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1432 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1433 default: ShouldNotReachHere(); 1434 } 1435 break; 1436 case 2: // No collapsing at level 2; keep all splits 1437 case 3: // No collapsing at level 3; keep all splits 1438 break; 1439 default: 1440 Unimplemented(); 1441 } 1442 1443 offset = tj->offset(); 1444 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1445 1446 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1447 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1448 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1449 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1450 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1451 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1452 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1453 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1454 assert( tj->ptr() != TypePtr::TopPTR && 1455 tj->ptr() != TypePtr::AnyNull && 1456 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1457 // assert( tj->ptr() != TypePtr::Constant || 1458 // tj->base() == Type::RawPtr || 1459 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1460 1461 return tj; 1462 } 1463 1464 void Compile::AliasType::Init(int i, const TypePtr* at) { 1465 _index = i; 1466 _adr_type = at; 1467 _field = NULL; 1468 _is_rewritable = true; // default 1469 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1470 if (atoop != NULL && atoop->is_known_instance()) { 1471 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1472 _general_index = Compile::current()->get_alias_index(gt); 1473 } else { 1474 _general_index = 0; 1475 } 1476 } 1477 1478 //---------------------------------print_on------------------------------------ 1479 #ifndef PRODUCT 1480 void Compile::AliasType::print_on(outputStream* st) { 1481 if (index() < 10) 1482 st->print("@ <%d> ", index()); 1483 else st->print("@ <%d>", index()); 1484 st->print(is_rewritable() ? " " : " RO"); 1485 int offset = adr_type()->offset(); 1486 if (offset == Type::OffsetBot) 1487 st->print(" +any"); 1488 else st->print(" +%-3d", offset); 1489 st->print(" in "); 1490 adr_type()->dump_on(st); 1491 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1492 if (field() != NULL && tjp) { 1493 if (tjp->klass() != field()->holder() || 1494 tjp->offset() != field()->offset_in_bytes()) { 1495 st->print(" != "); 1496 field()->print(); 1497 st->print(" ***"); 1498 } 1499 } 1500 } 1501 1502 void print_alias_types() { 1503 Compile* C = Compile::current(); 1504 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1505 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1506 C->alias_type(idx)->print_on(tty); 1507 tty->cr(); 1508 } 1509 } 1510 #endif 1511 1512 1513 //----------------------------probe_alias_cache-------------------------------- 1514 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1515 intptr_t key = (intptr_t) adr_type; 1516 key ^= key >> logAliasCacheSize; 1517 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1518 } 1519 1520 1521 //-----------------------------grow_alias_types-------------------------------- 1522 void Compile::grow_alias_types() { 1523 const int old_ats = _max_alias_types; // how many before? 1524 const int new_ats = old_ats; // how many more? 1525 const int grow_ats = old_ats+new_ats; // how many now? 1526 _max_alias_types = grow_ats; 1527 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1528 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1529 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1530 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1531 } 1532 1533 1534 //--------------------------------find_alias_type------------------------------ 1535 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1536 if (_AliasLevel == 0) 1537 return alias_type(AliasIdxBot); 1538 1539 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1540 if (ace->_adr_type == adr_type) { 1541 return alias_type(ace->_index); 1542 } 1543 1544 // Handle special cases. 1545 if (adr_type == NULL) return alias_type(AliasIdxTop); 1546 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1547 1548 // Do it the slow way. 1549 const TypePtr* flat = flatten_alias_type(adr_type); 1550 1551 #ifdef ASSERT 1552 assert(flat == flatten_alias_type(flat), "idempotent"); 1553 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1554 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1555 const TypeOopPtr* foop = flat->is_oopptr(); 1556 // Scalarizable allocations have exact klass always. 1557 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1558 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1559 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1560 } 1561 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1562 #endif 1563 1564 int idx = AliasIdxTop; 1565 for (int i = 0; i < num_alias_types(); i++) { 1566 if (alias_type(i)->adr_type() == flat) { 1567 idx = i; 1568 break; 1569 } 1570 } 1571 1572 if (idx == AliasIdxTop) { 1573 if (no_create) return NULL; 1574 // Grow the array if necessary. 1575 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1576 // Add a new alias type. 1577 idx = _num_alias_types++; 1578 _alias_types[idx]->Init(idx, flat); 1579 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1580 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1581 if (flat->isa_instptr()) { 1582 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1583 && flat->is_instptr()->klass() == env()->Class_klass()) 1584 alias_type(idx)->set_rewritable(false); 1585 } 1586 if (flat->isa_klassptr()) { 1587 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1588 alias_type(idx)->set_rewritable(false); 1589 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1590 alias_type(idx)->set_rewritable(false); 1591 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1592 alias_type(idx)->set_rewritable(false); 1593 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1594 alias_type(idx)->set_rewritable(false); 1595 } 1596 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1597 // but the base pointer type is not distinctive enough to identify 1598 // references into JavaThread.) 1599 1600 // Check for final fields. 1601 const TypeInstPtr* tinst = flat->isa_instptr(); 1602 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1603 ciField* field; 1604 if (tinst->const_oop() != NULL && 1605 tinst->klass() == ciEnv::current()->Class_klass() && 1606 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1607 // static field 1608 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1609 field = k->get_field_by_offset(tinst->offset(), true); 1610 } else { 1611 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1612 field = k->get_field_by_offset(tinst->offset(), false); 1613 } 1614 assert(field == NULL || 1615 original_field == NULL || 1616 (field->holder() == original_field->holder() && 1617 field->offset() == original_field->offset() && 1618 field->is_static() == original_field->is_static()), "wrong field?"); 1619 // Set field() and is_rewritable() attributes. 1620 if (field != NULL) alias_type(idx)->set_field(field); 1621 } 1622 } 1623 1624 // Fill the cache for next time. 1625 ace->_adr_type = adr_type; 1626 ace->_index = idx; 1627 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1628 1629 // Might as well try to fill the cache for the flattened version, too. 1630 AliasCacheEntry* face = probe_alias_cache(flat); 1631 if (face->_adr_type == NULL) { 1632 face->_adr_type = flat; 1633 face->_index = idx; 1634 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1635 } 1636 1637 return alias_type(idx); 1638 } 1639 1640 1641 Compile::AliasType* Compile::alias_type(ciField* field) { 1642 const TypeOopPtr* t; 1643 if (field->is_static()) 1644 t = TypeInstPtr::make(field->holder()->java_mirror()); 1645 else 1646 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1647 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1648 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1649 return atp; 1650 } 1651 1652 1653 //------------------------------have_alias_type-------------------------------- 1654 bool Compile::have_alias_type(const TypePtr* adr_type) { 1655 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1656 if (ace->_adr_type == adr_type) { 1657 return true; 1658 } 1659 1660 // Handle special cases. 1661 if (adr_type == NULL) return true; 1662 if (adr_type == TypePtr::BOTTOM) return true; 1663 1664 return find_alias_type(adr_type, true, NULL) != NULL; 1665 } 1666 1667 //-----------------------------must_alias-------------------------------------- 1668 // True if all values of the given address type are in the given alias category. 1669 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1670 if (alias_idx == AliasIdxBot) return true; // the universal category 1671 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1672 if (alias_idx == AliasIdxTop) return false; // the empty category 1673 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1674 1675 // the only remaining possible overlap is identity 1676 int adr_idx = get_alias_index(adr_type); 1677 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1678 assert(adr_idx == alias_idx || 1679 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1680 && adr_type != TypeOopPtr::BOTTOM), 1681 "should not be testing for overlap with an unsafe pointer"); 1682 return adr_idx == alias_idx; 1683 } 1684 1685 //------------------------------can_alias-------------------------------------- 1686 // True if any values of the given address type are in the given alias category. 1687 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1688 if (alias_idx == AliasIdxTop) return false; // the empty category 1689 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1690 if (alias_idx == AliasIdxBot) return true; // the universal category 1691 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1692 1693 // the only remaining possible overlap is identity 1694 int adr_idx = get_alias_index(adr_type); 1695 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1696 return adr_idx == alias_idx; 1697 } 1698 1699 1700 1701 //---------------------------pop_warm_call------------------------------------- 1702 WarmCallInfo* Compile::pop_warm_call() { 1703 WarmCallInfo* wci = _warm_calls; 1704 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1705 return wci; 1706 } 1707 1708 //----------------------------Inline_Warm-------------------------------------- 1709 int Compile::Inline_Warm() { 1710 // If there is room, try to inline some more warm call sites. 1711 // %%% Do a graph index compaction pass when we think we're out of space? 1712 if (!InlineWarmCalls) return 0; 1713 1714 int calls_made_hot = 0; 1715 int room_to_grow = NodeCountInliningCutoff - unique(); 1716 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1717 int amount_grown = 0; 1718 WarmCallInfo* call; 1719 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1720 int est_size = (int)call->size(); 1721 if (est_size > (room_to_grow - amount_grown)) { 1722 // This one won't fit anyway. Get rid of it. 1723 call->make_cold(); 1724 continue; 1725 } 1726 call->make_hot(); 1727 calls_made_hot++; 1728 amount_grown += est_size; 1729 amount_to_grow -= est_size; 1730 } 1731 1732 if (calls_made_hot > 0) set_major_progress(); 1733 return calls_made_hot; 1734 } 1735 1736 1737 //----------------------------Finish_Warm-------------------------------------- 1738 void Compile::Finish_Warm() { 1739 if (!InlineWarmCalls) return; 1740 if (failing()) return; 1741 if (warm_calls() == NULL) return; 1742 1743 // Clean up loose ends, if we are out of space for inlining. 1744 WarmCallInfo* call; 1745 while ((call = pop_warm_call()) != NULL) { 1746 call->make_cold(); 1747 } 1748 } 1749 1750 //---------------------cleanup_loop_predicates----------------------- 1751 // Remove the opaque nodes that protect the predicates so that all unused 1752 // checks and uncommon_traps will be eliminated from the ideal graph 1753 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1754 if (predicate_count()==0) return; 1755 for (int i = predicate_count(); i > 0; i--) { 1756 Node * n = predicate_opaque1_node(i-1); 1757 assert(n->Opcode() == Op_Opaque1, "must be"); 1758 igvn.replace_node(n, n->in(1)); 1759 } 1760 assert(predicate_count()==0, "should be clean!"); 1761 } 1762 1763 //------------------------------Optimize--------------------------------------- 1764 // Given a graph, optimize it. 1765 void Compile::Optimize() { 1766 TracePhase t1("optimizer", &_t_optimizer, true); 1767 1768 #ifndef PRODUCT 1769 if (env()->break_at_compile()) { 1770 BREAKPOINT; 1771 } 1772 1773 #endif 1774 1775 ResourceMark rm; 1776 int loop_opts_cnt; 1777 1778 NOT_PRODUCT( verify_graph_edges(); ) 1779 1780 print_method("After Parsing"); 1781 1782 { 1783 // Iterative Global Value Numbering, including ideal transforms 1784 // Initialize IterGVN with types and values from parse-time GVN 1785 PhaseIterGVN igvn(initial_gvn()); 1786 { 1787 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1788 igvn.optimize(); 1789 } 1790 1791 print_method("Iter GVN 1", 2); 1792 1793 if (failing()) return; 1794 1795 // Perform escape analysis 1796 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 1797 if (has_loops()) { 1798 // Cleanup graph (remove dead nodes). 1799 TracePhase t2("idealLoop", &_t_idealLoop, true); 1800 PhaseIdealLoop ideal_loop( igvn, false, true ); 1801 if (major_progress()) print_method("PhaseIdealLoop before EA", 2); 1802 if (failing()) return; 1803 } 1804 ConnectionGraph::do_analysis(this, &igvn); 1805 1806 if (failing()) return; 1807 1808 // Optimize out fields loads from scalar replaceable allocations. 1809 igvn.optimize(); 1810 print_method("Iter GVN after EA", 2); 1811 1812 if (failing()) return; 1813 1814 if (congraph() != NULL && macro_count() > 0) { 1815 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 1816 PhaseMacroExpand mexp(igvn); 1817 mexp.eliminate_macro_nodes(); 1818 igvn.set_delay_transform(false); 1819 1820 igvn.optimize(); 1821 print_method("Iter GVN after eliminating allocations and locks", 2); 1822 1823 if (failing()) return; 1824 } 1825 } 1826 1827 // Loop transforms on the ideal graph. Range Check Elimination, 1828 // peeling, unrolling, etc. 1829 1830 // Set loop opts counter 1831 loop_opts_cnt = num_loop_opts(); 1832 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1833 { 1834 TracePhase t2("idealLoop", &_t_idealLoop, true); 1835 PhaseIdealLoop ideal_loop( igvn, true ); 1836 loop_opts_cnt--; 1837 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1838 if (failing()) return; 1839 } 1840 // Loop opts pass if partial peeling occurred in previous pass 1841 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1842 TracePhase t3("idealLoop", &_t_idealLoop, true); 1843 PhaseIdealLoop ideal_loop( igvn, false ); 1844 loop_opts_cnt--; 1845 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1846 if (failing()) return; 1847 } 1848 // Loop opts pass for loop-unrolling before CCP 1849 if(major_progress() && (loop_opts_cnt > 0)) { 1850 TracePhase t4("idealLoop", &_t_idealLoop, true); 1851 PhaseIdealLoop ideal_loop( igvn, false ); 1852 loop_opts_cnt--; 1853 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1854 } 1855 if (!failing()) { 1856 // Verify that last round of loop opts produced a valid graph 1857 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1858 PhaseIdealLoop::verify(igvn); 1859 } 1860 } 1861 if (failing()) return; 1862 1863 // Conditional Constant Propagation; 1864 PhaseCCP ccp( &igvn ); 1865 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1866 { 1867 TracePhase t2("ccp", &_t_ccp, true); 1868 ccp.do_transform(); 1869 } 1870 print_method("PhaseCPP 1", 2); 1871 1872 assert( true, "Break here to ccp.dump_old2new_map()"); 1873 1874 // Iterative Global Value Numbering, including ideal transforms 1875 { 1876 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1877 igvn = ccp; 1878 igvn.optimize(); 1879 } 1880 1881 print_method("Iter GVN 2", 2); 1882 1883 if (failing()) return; 1884 1885 // Loop transforms on the ideal graph. Range Check Elimination, 1886 // peeling, unrolling, etc. 1887 if(loop_opts_cnt > 0) { 1888 debug_only( int cnt = 0; ); 1889 while(major_progress() && (loop_opts_cnt > 0)) { 1890 TracePhase t2("idealLoop", &_t_idealLoop, true); 1891 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1892 PhaseIdealLoop ideal_loop( igvn, true); 1893 loop_opts_cnt--; 1894 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1895 if (failing()) return; 1896 } 1897 } 1898 1899 { 1900 // Verify that all previous optimizations produced a valid graph 1901 // at least to this point, even if no loop optimizations were done. 1902 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1903 PhaseIdealLoop::verify(igvn); 1904 } 1905 1906 { 1907 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1908 PhaseMacroExpand mex(igvn); 1909 if (mex.expand_macro_nodes()) { 1910 assert(failing(), "must bail out w/ explicit message"); 1911 return; 1912 } 1913 } 1914 1915 } // (End scope of igvn; run destructor if necessary for asserts.) 1916 1917 // A method with only infinite loops has no edges entering loops from root 1918 { 1919 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1920 if (final_graph_reshaping()) { 1921 assert(failing(), "must bail out w/ explicit message"); 1922 return; 1923 } 1924 } 1925 1926 print_method("Optimize finished", 2); 1927 } 1928 1929 1930 //------------------------------Code_Gen--------------------------------------- 1931 // Given a graph, generate code for it 1932 void Compile::Code_Gen() { 1933 if (failing()) return; 1934 1935 // Perform instruction selection. You might think we could reclaim Matcher 1936 // memory PDQ, but actually the Matcher is used in generating spill code. 1937 // Internals of the Matcher (including some VectorSets) must remain live 1938 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1939 // set a bit in reclaimed memory. 1940 1941 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1942 // nodes. Mapping is only valid at the root of each matched subtree. 1943 NOT_PRODUCT( verify_graph_edges(); ) 1944 1945 Node_List proj_list; 1946 Matcher m(proj_list); 1947 _matcher = &m; 1948 { 1949 TracePhase t2("matcher", &_t_matcher, true); 1950 m.match(); 1951 } 1952 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1953 // nodes. Mapping is only valid at the root of each matched subtree. 1954 NOT_PRODUCT( verify_graph_edges(); ) 1955 1956 // If you have too many nodes, or if matching has failed, bail out 1957 check_node_count(0, "out of nodes matching instructions"); 1958 if (failing()) return; 1959 1960 // Build a proper-looking CFG 1961 PhaseCFG cfg(node_arena(), root(), m); 1962 _cfg = &cfg; 1963 { 1964 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1965 cfg.Dominators(); 1966 if (failing()) return; 1967 1968 NOT_PRODUCT( verify_graph_edges(); ) 1969 1970 cfg.Estimate_Block_Frequency(); 1971 cfg.GlobalCodeMotion(m,unique(),proj_list); 1972 if (failing()) return; 1973 1974 print_method("Global code motion", 2); 1975 1976 NOT_PRODUCT( verify_graph_edges(); ) 1977 1978 debug_only( cfg.verify(); ) 1979 } 1980 NOT_PRODUCT( verify_graph_edges(); ) 1981 1982 PhaseChaitin regalloc(unique(),cfg,m); 1983 _regalloc = ®alloc; 1984 { 1985 TracePhase t2("regalloc", &_t_registerAllocation, true); 1986 // Perform any platform dependent preallocation actions. This is used, 1987 // for example, to avoid taking an implicit null pointer exception 1988 // using the frame pointer on win95. 1989 _regalloc->pd_preallocate_hook(); 1990 1991 // Perform register allocation. After Chaitin, use-def chains are 1992 // no longer accurate (at spill code) and so must be ignored. 1993 // Node->LRG->reg mappings are still accurate. 1994 _regalloc->Register_Allocate(); 1995 1996 // Bail out if the allocator builds too many nodes 1997 if (failing()) return; 1998 } 1999 2000 // Prior to register allocation we kept empty basic blocks in case the 2001 // the allocator needed a place to spill. After register allocation we 2002 // are not adding any new instructions. If any basic block is empty, we 2003 // can now safely remove it. 2004 { 2005 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2006 cfg.remove_empty(); 2007 if (do_freq_based_layout()) { 2008 PhaseBlockLayout layout(cfg); 2009 } else { 2010 cfg.set_loop_alignment(); 2011 } 2012 cfg.fixup_flow(); 2013 } 2014 2015 // Perform any platform dependent postallocation verifications. 2016 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 2017 2018 // Apply peephole optimizations 2019 if( OptoPeephole ) { 2020 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2021 PhasePeephole peep( _regalloc, cfg); 2022 peep.do_transform(); 2023 } 2024 2025 // Convert Nodes to instruction bits in a buffer 2026 { 2027 // %%%% workspace merge brought two timers together for one job 2028 TracePhase t2a("output", &_t_output, true); 2029 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2030 Output(); 2031 } 2032 2033 print_method("Final Code"); 2034 2035 // He's dead, Jim. 2036 _cfg = (PhaseCFG*)0xdeadbeef; 2037 _regalloc = (PhaseChaitin*)0xdeadbeef; 2038 } 2039 2040 2041 //------------------------------dump_asm--------------------------------------- 2042 // Dump formatted assembly 2043 #ifndef PRODUCT 2044 void Compile::dump_asm(int *pcs, uint pc_limit) { 2045 bool cut_short = false; 2046 tty->print_cr("#"); 2047 tty->print("# "); _tf->dump(); tty->cr(); 2048 tty->print_cr("#"); 2049 2050 // For all blocks 2051 int pc = 0x0; // Program counter 2052 char starts_bundle = ' '; 2053 _regalloc->dump_frame(); 2054 2055 Node *n = NULL; 2056 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 2057 if (VMThread::should_terminate()) { cut_short = true; break; } 2058 Block *b = _cfg->_blocks[i]; 2059 if (b->is_connector() && !Verbose) continue; 2060 n = b->_nodes[0]; 2061 if (pcs && n->_idx < pc_limit) 2062 tty->print("%3.3x ", pcs[n->_idx]); 2063 else 2064 tty->print(" "); 2065 b->dump_head( &_cfg->_bbs ); 2066 if (b->is_connector()) { 2067 tty->print_cr(" # Empty connector block"); 2068 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2069 tty->print_cr(" # Block is sole successor of call"); 2070 } 2071 2072 // For all instructions 2073 Node *delay = NULL; 2074 for( uint j = 0; j<b->_nodes.size(); j++ ) { 2075 if (VMThread::should_terminate()) { cut_short = true; break; } 2076 n = b->_nodes[j]; 2077 if (valid_bundle_info(n)) { 2078 Bundle *bundle = node_bundling(n); 2079 if (bundle->used_in_unconditional_delay()) { 2080 delay = n; 2081 continue; 2082 } 2083 if (bundle->starts_bundle()) 2084 starts_bundle = '+'; 2085 } 2086 2087 if (WizardMode) n->dump(); 2088 2089 if( !n->is_Region() && // Dont print in the Assembly 2090 !n->is_Phi() && // a few noisely useless nodes 2091 !n->is_Proj() && 2092 !n->is_MachTemp() && 2093 !n->is_SafePointScalarObject() && 2094 !n->is_Catch() && // Would be nice to print exception table targets 2095 !n->is_MergeMem() && // Not very interesting 2096 !n->is_top() && // Debug info table constants 2097 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2098 ) { 2099 if (pcs && n->_idx < pc_limit) 2100 tty->print("%3.3x", pcs[n->_idx]); 2101 else 2102 tty->print(" "); 2103 tty->print(" %c ", starts_bundle); 2104 starts_bundle = ' '; 2105 tty->print("\t"); 2106 n->format(_regalloc, tty); 2107 tty->cr(); 2108 } 2109 2110 // If we have an instruction with a delay slot, and have seen a delay, 2111 // then back up and print it 2112 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2113 assert(delay != NULL, "no unconditional delay instruction"); 2114 if (WizardMode) delay->dump(); 2115 2116 if (node_bundling(delay)->starts_bundle()) 2117 starts_bundle = '+'; 2118 if (pcs && n->_idx < pc_limit) 2119 tty->print("%3.3x", pcs[n->_idx]); 2120 else 2121 tty->print(" "); 2122 tty->print(" %c ", starts_bundle); 2123 starts_bundle = ' '; 2124 tty->print("\t"); 2125 delay->format(_regalloc, tty); 2126 tty->print_cr(""); 2127 delay = NULL; 2128 } 2129 2130 // Dump the exception table as well 2131 if( n->is_Catch() && (Verbose || WizardMode) ) { 2132 // Print the exception table for this offset 2133 _handler_table.print_subtable_for(pc); 2134 } 2135 } 2136 2137 if (pcs && n->_idx < pc_limit) 2138 tty->print_cr("%3.3x", pcs[n->_idx]); 2139 else 2140 tty->print_cr(""); 2141 2142 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2143 2144 } // End of per-block dump 2145 tty->print_cr(""); 2146 2147 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2148 } 2149 #endif 2150 2151 //------------------------------Final_Reshape_Counts--------------------------- 2152 // This class defines counters to help identify when a method 2153 // may/must be executed using hardware with only 24-bit precision. 2154 struct Final_Reshape_Counts : public StackObj { 2155 int _call_count; // count non-inlined 'common' calls 2156 int _float_count; // count float ops requiring 24-bit precision 2157 int _double_count; // count double ops requiring more precision 2158 int _java_call_count; // count non-inlined 'java' calls 2159 int _inner_loop_count; // count loops which need alignment 2160 VectorSet _visited; // Visitation flags 2161 Node_List _tests; // Set of IfNodes & PCTableNodes 2162 2163 Final_Reshape_Counts() : 2164 _call_count(0), _float_count(0), _double_count(0), 2165 _java_call_count(0), _inner_loop_count(0), 2166 _visited( Thread::current()->resource_area() ) { } 2167 2168 void inc_call_count () { _call_count ++; } 2169 void inc_float_count () { _float_count ++; } 2170 void inc_double_count() { _double_count++; } 2171 void inc_java_call_count() { _java_call_count++; } 2172 void inc_inner_loop_count() { _inner_loop_count++; } 2173 2174 int get_call_count () const { return _call_count ; } 2175 int get_float_count () const { return _float_count ; } 2176 int get_double_count() const { return _double_count; } 2177 int get_java_call_count() const { return _java_call_count; } 2178 int get_inner_loop_count() const { return _inner_loop_count; } 2179 }; 2180 2181 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2182 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2183 // Make sure the offset goes inside the instance layout. 2184 return k->contains_field_offset(tp->offset()); 2185 // Note that OffsetBot and OffsetTop are very negative. 2186 } 2187 2188 // Eliminate trivially redundant StoreCMs and accumulate their 2189 // precedence edges. 2190 void Compile::eliminate_redundant_card_marks(Node* n) { 2191 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2192 if (n->in(MemNode::Address)->outcnt() > 1) { 2193 // There are multiple users of the same address so it might be 2194 // possible to eliminate some of the StoreCMs 2195 Node* mem = n->in(MemNode::Memory); 2196 Node* adr = n->in(MemNode::Address); 2197 Node* val = n->in(MemNode::ValueIn); 2198 Node* prev = n; 2199 bool done = false; 2200 // Walk the chain of StoreCMs eliminating ones that match. As 2201 // long as it's a chain of single users then the optimization is 2202 // safe. Eliminating partially redundant StoreCMs would require 2203 // cloning copies down the other paths. 2204 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2205 if (adr == mem->in(MemNode::Address) && 2206 val == mem->in(MemNode::ValueIn)) { 2207 // redundant StoreCM 2208 if (mem->req() > MemNode::OopStore) { 2209 // Hasn't been processed by this code yet. 2210 n->add_prec(mem->in(MemNode::OopStore)); 2211 } else { 2212 // Already converted to precedence edge 2213 for (uint i = mem->req(); i < mem->len(); i++) { 2214 // Accumulate any precedence edges 2215 if (mem->in(i) != NULL) { 2216 n->add_prec(mem->in(i)); 2217 } 2218 } 2219 // Everything above this point has been processed. 2220 done = true; 2221 } 2222 // Eliminate the previous StoreCM 2223 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2224 assert(mem->outcnt() == 0, "should be dead"); 2225 mem->disconnect_inputs(NULL, this); 2226 } else { 2227 prev = mem; 2228 } 2229 mem = prev->in(MemNode::Memory); 2230 } 2231 } 2232 } 2233 2234 //------------------------------final_graph_reshaping_impl---------------------- 2235 // Implement items 1-5 from final_graph_reshaping below. 2236 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2237 2238 if ( n->outcnt() == 0 ) return; // dead node 2239 uint nop = n->Opcode(); 2240 2241 // Check for 2-input instruction with "last use" on right input. 2242 // Swap to left input. Implements item (2). 2243 if( n->req() == 3 && // two-input instruction 2244 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2245 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2246 n->in(2)->outcnt() == 1 &&// right use IS a last use 2247 !n->in(2)->is_Con() ) { // right use is not a constant 2248 // Check for commutative opcode 2249 switch( nop ) { 2250 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2251 case Op_MaxI: case Op_MinI: 2252 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2253 case Op_AndL: case Op_XorL: case Op_OrL: 2254 case Op_AndI: case Op_XorI: case Op_OrI: { 2255 // Move "last use" input to left by swapping inputs 2256 n->swap_edges(1, 2); 2257 break; 2258 } 2259 default: 2260 break; 2261 } 2262 } 2263 2264 #ifdef ASSERT 2265 if( n->is_Mem() ) { 2266 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2267 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2268 // oop will be recorded in oop map if load crosses safepoint 2269 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2270 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2271 "raw memory operations should have control edge"); 2272 } 2273 #endif 2274 // Count FPU ops and common calls, implements item (3) 2275 switch( nop ) { 2276 // Count all float operations that may use FPU 2277 case Op_AddF: 2278 case Op_SubF: 2279 case Op_MulF: 2280 case Op_DivF: 2281 case Op_NegF: 2282 case Op_ModF: 2283 case Op_ConvI2F: 2284 case Op_ConF: 2285 case Op_CmpF: 2286 case Op_CmpF3: 2287 // case Op_ConvL2F: // longs are split into 32-bit halves 2288 frc.inc_float_count(); 2289 break; 2290 2291 case Op_ConvF2D: 2292 case Op_ConvD2F: 2293 frc.inc_float_count(); 2294 frc.inc_double_count(); 2295 break; 2296 2297 // Count all double operations that may use FPU 2298 case Op_AddD: 2299 case Op_SubD: 2300 case Op_MulD: 2301 case Op_DivD: 2302 case Op_NegD: 2303 case Op_ModD: 2304 case Op_ConvI2D: 2305 case Op_ConvD2I: 2306 // case Op_ConvL2D: // handled by leaf call 2307 // case Op_ConvD2L: // handled by leaf call 2308 case Op_ConD: 2309 case Op_CmpD: 2310 case Op_CmpD3: 2311 frc.inc_double_count(); 2312 break; 2313 case Op_Opaque1: // Remove Opaque Nodes before matching 2314 case Op_Opaque2: // Remove Opaque Nodes before matching 2315 n->subsume_by(n->in(1), this); 2316 break; 2317 case Op_CallStaticJava: 2318 case Op_CallJava: 2319 case Op_CallDynamicJava: 2320 frc.inc_java_call_count(); // Count java call site; 2321 case Op_CallRuntime: 2322 case Op_CallLeaf: 2323 case Op_CallLeafNoFP: { 2324 assert( n->is_Call(), "" ); 2325 CallNode *call = n->as_Call(); 2326 // Count call sites where the FP mode bit would have to be flipped. 2327 // Do not count uncommon runtime calls: 2328 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2329 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2330 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2331 frc.inc_call_count(); // Count the call site 2332 } else { // See if uncommon argument is shared 2333 Node *n = call->in(TypeFunc::Parms); 2334 int nop = n->Opcode(); 2335 // Clone shared simple arguments to uncommon calls, item (1). 2336 if( n->outcnt() > 1 && 2337 !n->is_Proj() && 2338 nop != Op_CreateEx && 2339 nop != Op_CheckCastPP && 2340 nop != Op_DecodeN && 2341 nop != Op_DecodeNKlass && 2342 !n->is_Mem() ) { 2343 Node *x = n->clone(); 2344 call->set_req( TypeFunc::Parms, x ); 2345 } 2346 } 2347 break; 2348 } 2349 2350 case Op_StoreD: 2351 case Op_LoadD: 2352 case Op_LoadD_unaligned: 2353 frc.inc_double_count(); 2354 goto handle_mem; 2355 case Op_StoreF: 2356 case Op_LoadF: 2357 frc.inc_float_count(); 2358 goto handle_mem; 2359 2360 case Op_StoreCM: 2361 { 2362 // Convert OopStore dependence into precedence edge 2363 Node* prec = n->in(MemNode::OopStore); 2364 n->del_req(MemNode::OopStore); 2365 n->add_prec(prec); 2366 eliminate_redundant_card_marks(n); 2367 } 2368 2369 // fall through 2370 2371 case Op_StoreB: 2372 case Op_StoreC: 2373 case Op_StorePConditional: 2374 case Op_StoreI: 2375 case Op_StoreL: 2376 case Op_StoreIConditional: 2377 case Op_StoreLConditional: 2378 case Op_CompareAndSwapI: 2379 case Op_CompareAndSwapL: 2380 case Op_CompareAndSwapP: 2381 case Op_CompareAndSwapN: 2382 case Op_GetAndAddI: 2383 case Op_GetAndAddL: 2384 case Op_GetAndSetI: 2385 case Op_GetAndSetL: 2386 case Op_GetAndSetP: 2387 case Op_GetAndSetN: 2388 case Op_StoreP: 2389 case Op_StoreN: 2390 case Op_StoreNKlass: 2391 case Op_LoadB: 2392 case Op_LoadUB: 2393 case Op_LoadUS: 2394 case Op_LoadI: 2395 case Op_LoadKlass: 2396 case Op_LoadNKlass: 2397 case Op_LoadL: 2398 case Op_LoadL_unaligned: 2399 case Op_LoadPLocked: 2400 case Op_LoadP: 2401 case Op_LoadN: 2402 case Op_LoadRange: 2403 case Op_LoadS: { 2404 handle_mem: 2405 #ifdef ASSERT 2406 if( VerifyOptoOopOffsets ) { 2407 assert( n->is_Mem(), "" ); 2408 MemNode *mem = (MemNode*)n; 2409 // Check to see if address types have grounded out somehow. 2410 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2411 assert( !tp || oop_offset_is_sane(tp), "" ); 2412 } 2413 #endif 2414 break; 2415 } 2416 2417 case Op_AddP: { // Assert sane base pointers 2418 Node *addp = n->in(AddPNode::Address); 2419 assert( !addp->is_AddP() || 2420 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2421 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2422 "Base pointers must match" ); 2423 #ifdef _LP64 2424 if ((UseCompressedOops || UseCompressedKlassPointers) && 2425 addp->Opcode() == Op_ConP && 2426 addp == n->in(AddPNode::Base) && 2427 n->in(AddPNode::Offset)->is_Con()) { 2428 // Use addressing with narrow klass to load with offset on x86. 2429 // On sparc loading 32-bits constant and decoding it have less 2430 // instructions (4) then load 64-bits constant (7). 2431 // Do this transformation here since IGVN will convert ConN back to ConP. 2432 const Type* t = addp->bottom_type(); 2433 if (t->isa_oopptr() || t->isa_klassptr()) { 2434 Node* nn = NULL; 2435 2436 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2437 2438 // Look for existing ConN node of the same exact type. 2439 Node* r = root(); 2440 uint cnt = r->outcnt(); 2441 for (uint i = 0; i < cnt; i++) { 2442 Node* m = r->raw_out(i); 2443 if (m!= NULL && m->Opcode() == op && 2444 m->bottom_type()->make_ptr() == t) { 2445 nn = m; 2446 break; 2447 } 2448 } 2449 if (nn != NULL) { 2450 // Decode a narrow oop to match address 2451 // [R12 + narrow_oop_reg<<3 + offset] 2452 if (t->isa_oopptr()) { 2453 nn = new (this) DecodeNNode(nn, t); 2454 } else { 2455 nn = new (this) DecodeNKlassNode(nn, t); 2456 } 2457 n->set_req(AddPNode::Base, nn); 2458 n->set_req(AddPNode::Address, nn); 2459 if (addp->outcnt() == 0) { 2460 addp->disconnect_inputs(NULL, this); 2461 } 2462 } 2463 } 2464 } 2465 #endif 2466 break; 2467 } 2468 2469 #ifdef _LP64 2470 case Op_CastPP: 2471 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2472 Node* in1 = n->in(1); 2473 const Type* t = n->bottom_type(); 2474 Node* new_in1 = in1->clone(); 2475 new_in1->as_DecodeN()->set_type(t); 2476 2477 if (!Matcher::narrow_oop_use_complex_address()) { 2478 // 2479 // x86, ARM and friends can handle 2 adds in addressing mode 2480 // and Matcher can fold a DecodeN node into address by using 2481 // a narrow oop directly and do implicit NULL check in address: 2482 // 2483 // [R12 + narrow_oop_reg<<3 + offset] 2484 // NullCheck narrow_oop_reg 2485 // 2486 // On other platforms (Sparc) we have to keep new DecodeN node and 2487 // use it to do implicit NULL check in address: 2488 // 2489 // decode_not_null narrow_oop_reg, base_reg 2490 // [base_reg + offset] 2491 // NullCheck base_reg 2492 // 2493 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2494 // to keep the information to which NULL check the new DecodeN node 2495 // corresponds to use it as value in implicit_null_check(). 2496 // 2497 new_in1->set_req(0, n->in(0)); 2498 } 2499 2500 n->subsume_by(new_in1, this); 2501 if (in1->outcnt() == 0) { 2502 in1->disconnect_inputs(NULL, this); 2503 } 2504 } 2505 break; 2506 2507 case Op_CmpP: 2508 // Do this transformation here to preserve CmpPNode::sub() and 2509 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2510 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2511 Node* in1 = n->in(1); 2512 Node* in2 = n->in(2); 2513 if (!in1->is_DecodeNarrowPtr()) { 2514 in2 = in1; 2515 in1 = n->in(2); 2516 } 2517 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2518 2519 Node* new_in2 = NULL; 2520 if (in2->is_DecodeNarrowPtr()) { 2521 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2522 new_in2 = in2->in(1); 2523 } else if (in2->Opcode() == Op_ConP) { 2524 const Type* t = in2->bottom_type(); 2525 if (t == TypePtr::NULL_PTR) { 2526 assert(in1->is_DecodeN(), "compare klass to null?"); 2527 // Don't convert CmpP null check into CmpN if compressed 2528 // oops implicit null check is not generated. 2529 // This will allow to generate normal oop implicit null check. 2530 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2531 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2532 // 2533 // This transformation together with CastPP transformation above 2534 // will generated code for implicit NULL checks for compressed oops. 2535 // 2536 // The original code after Optimize() 2537 // 2538 // LoadN memory, narrow_oop_reg 2539 // decode narrow_oop_reg, base_reg 2540 // CmpP base_reg, NULL 2541 // CastPP base_reg // NotNull 2542 // Load [base_reg + offset], val_reg 2543 // 2544 // after these transformations will be 2545 // 2546 // LoadN memory, narrow_oop_reg 2547 // CmpN narrow_oop_reg, NULL 2548 // decode_not_null narrow_oop_reg, base_reg 2549 // Load [base_reg + offset], val_reg 2550 // 2551 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2552 // since narrow oops can be used in debug info now (see the code in 2553 // final_graph_reshaping_walk()). 2554 // 2555 // At the end the code will be matched to 2556 // on x86: 2557 // 2558 // Load_narrow_oop memory, narrow_oop_reg 2559 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2560 // NullCheck narrow_oop_reg 2561 // 2562 // and on sparc: 2563 // 2564 // Load_narrow_oop memory, narrow_oop_reg 2565 // decode_not_null narrow_oop_reg, base_reg 2566 // Load [base_reg + offset], val_reg 2567 // NullCheck base_reg 2568 // 2569 } else if (t->isa_oopptr()) { 2570 new_in2 = ConNode::make(this, t->make_narrowoop()); 2571 } else if (t->isa_klassptr()) { 2572 new_in2 = ConNode::make(this, t->make_narrowklass()); 2573 } 2574 } 2575 if (new_in2 != NULL) { 2576 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 2577 n->subsume_by(cmpN, this); 2578 if (in1->outcnt() == 0) { 2579 in1->disconnect_inputs(NULL, this); 2580 } 2581 if (in2->outcnt() == 0) { 2582 in2->disconnect_inputs(NULL, this); 2583 } 2584 } 2585 } 2586 break; 2587 2588 case Op_DecodeN: 2589 case Op_DecodeNKlass: 2590 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2591 // DecodeN could be pinned when it can't be fold into 2592 // an address expression, see the code for Op_CastPP above. 2593 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2594 break; 2595 2596 case Op_EncodeP: 2597 case Op_EncodePKlass: { 2598 Node* in1 = n->in(1); 2599 if (in1->is_DecodeNarrowPtr()) { 2600 n->subsume_by(in1->in(1), this); 2601 } else if (in1->Opcode() == Op_ConP) { 2602 const Type* t = in1->bottom_type(); 2603 if (t == TypePtr::NULL_PTR) { 2604 assert(t->isa_oopptr(), "null klass?"); 2605 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2606 } else if (t->isa_oopptr()) { 2607 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2608 } else if (t->isa_klassptr()) { 2609 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2610 } 2611 } 2612 if (in1->outcnt() == 0) { 2613 in1->disconnect_inputs(NULL, this); 2614 } 2615 break; 2616 } 2617 2618 case Op_Proj: { 2619 if (OptimizeStringConcat) { 2620 ProjNode* p = n->as_Proj(); 2621 if (p->_is_io_use) { 2622 // Separate projections were used for the exception path which 2623 // are normally removed by a late inline. If it wasn't inlined 2624 // then they will hang around and should just be replaced with 2625 // the original one. 2626 Node* proj = NULL; 2627 // Replace with just one 2628 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2629 Node *use = i.get(); 2630 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2631 proj = use; 2632 break; 2633 } 2634 } 2635 assert(proj != NULL, "must be found"); 2636 p->subsume_by(proj, this); 2637 } 2638 } 2639 break; 2640 } 2641 2642 case Op_Phi: 2643 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2644 // The EncodeP optimization may create Phi with the same edges 2645 // for all paths. It is not handled well by Register Allocator. 2646 Node* unique_in = n->in(1); 2647 assert(unique_in != NULL, ""); 2648 uint cnt = n->req(); 2649 for (uint i = 2; i < cnt; i++) { 2650 Node* m = n->in(i); 2651 assert(m != NULL, ""); 2652 if (unique_in != m) 2653 unique_in = NULL; 2654 } 2655 if (unique_in != NULL) { 2656 n->subsume_by(unique_in, this); 2657 } 2658 } 2659 break; 2660 2661 #endif 2662 2663 case Op_ModI: 2664 if (UseDivMod) { 2665 // Check if a%b and a/b both exist 2666 Node* d = n->find_similar(Op_DivI); 2667 if (d) { 2668 // Replace them with a fused divmod if supported 2669 if (Matcher::has_match_rule(Op_DivModI)) { 2670 DivModINode* divmod = DivModINode::make(this, n); 2671 d->subsume_by(divmod->div_proj(), this); 2672 n->subsume_by(divmod->mod_proj(), this); 2673 } else { 2674 // replace a%b with a-((a/b)*b) 2675 Node* mult = new (this) MulINode(d, d->in(2)); 2676 Node* sub = new (this) SubINode(d->in(1), mult); 2677 n->subsume_by(sub, this); 2678 } 2679 } 2680 } 2681 break; 2682 2683 case Op_ModL: 2684 if (UseDivMod) { 2685 // Check if a%b and a/b both exist 2686 Node* d = n->find_similar(Op_DivL); 2687 if (d) { 2688 // Replace them with a fused divmod if supported 2689 if (Matcher::has_match_rule(Op_DivModL)) { 2690 DivModLNode* divmod = DivModLNode::make(this, n); 2691 d->subsume_by(divmod->div_proj(), this); 2692 n->subsume_by(divmod->mod_proj(), this); 2693 } else { 2694 // replace a%b with a-((a/b)*b) 2695 Node* mult = new (this) MulLNode(d, d->in(2)); 2696 Node* sub = new (this) SubLNode(d->in(1), mult); 2697 n->subsume_by(sub, this); 2698 } 2699 } 2700 } 2701 break; 2702 2703 case Op_LoadVector: 2704 case Op_StoreVector: 2705 break; 2706 2707 case Op_PackB: 2708 case Op_PackS: 2709 case Op_PackI: 2710 case Op_PackF: 2711 case Op_PackL: 2712 case Op_PackD: 2713 if (n->req()-1 > 2) { 2714 // Replace many operand PackNodes with a binary tree for matching 2715 PackNode* p = (PackNode*) n; 2716 Node* btp = p->binary_tree_pack(this, 1, n->req()); 2717 n->subsume_by(btp, this); 2718 } 2719 break; 2720 case Op_Loop: 2721 case Op_CountedLoop: 2722 if (n->as_Loop()->is_inner_loop()) { 2723 frc.inc_inner_loop_count(); 2724 } 2725 break; 2726 case Op_LShiftI: 2727 case Op_RShiftI: 2728 case Op_URShiftI: 2729 case Op_LShiftL: 2730 case Op_RShiftL: 2731 case Op_URShiftL: 2732 if (Matcher::need_masked_shift_count) { 2733 // The cpu's shift instructions don't restrict the count to the 2734 // lower 5/6 bits. We need to do the masking ourselves. 2735 Node* in2 = n->in(2); 2736 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 2737 const TypeInt* t = in2->find_int_type(); 2738 if (t != NULL && t->is_con()) { 2739 juint shift = t->get_con(); 2740 if (shift > mask) { // Unsigned cmp 2741 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 2742 } 2743 } else { 2744 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 2745 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 2746 n->set_req(2, shift); 2747 } 2748 } 2749 if (in2->outcnt() == 0) { // Remove dead node 2750 in2->disconnect_inputs(NULL, this); 2751 } 2752 } 2753 break; 2754 default: 2755 assert( !n->is_Call(), "" ); 2756 assert( !n->is_Mem(), "" ); 2757 break; 2758 } 2759 2760 // Collect CFG split points 2761 if (n->is_MultiBranch()) 2762 frc._tests.push(n); 2763 } 2764 2765 //------------------------------final_graph_reshaping_walk--------------------- 2766 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2767 // requires that the walk visits a node's inputs before visiting the node. 2768 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2769 ResourceArea *area = Thread::current()->resource_area(); 2770 Unique_Node_List sfpt(area); 2771 2772 frc._visited.set(root->_idx); // first, mark node as visited 2773 uint cnt = root->req(); 2774 Node *n = root; 2775 uint i = 0; 2776 while (true) { 2777 if (i < cnt) { 2778 // Place all non-visited non-null inputs onto stack 2779 Node* m = n->in(i); 2780 ++i; 2781 if (m != NULL && !frc._visited.test_set(m->_idx)) { 2782 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2783 sfpt.push(m); 2784 cnt = m->req(); 2785 nstack.push(n, i); // put on stack parent and next input's index 2786 n = m; 2787 i = 0; 2788 } 2789 } else { 2790 // Now do post-visit work 2791 final_graph_reshaping_impl( n, frc ); 2792 if (nstack.is_empty()) 2793 break; // finished 2794 n = nstack.node(); // Get node from stack 2795 cnt = n->req(); 2796 i = nstack.index(); 2797 nstack.pop(); // Shift to the next node on stack 2798 } 2799 } 2800 2801 // Skip next transformation if compressed oops are not used. 2802 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 2803 (!UseCompressedOops && !UseCompressedKlassPointers)) 2804 return; 2805 2806 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 2807 // It could be done for an uncommon traps or any safepoints/calls 2808 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 2809 while (sfpt.size() > 0) { 2810 n = sfpt.pop(); 2811 JVMState *jvms = n->as_SafePoint()->jvms(); 2812 assert(jvms != NULL, "sanity"); 2813 int start = jvms->debug_start(); 2814 int end = n->req(); 2815 bool is_uncommon = (n->is_CallStaticJava() && 2816 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2817 for (int j = start; j < end; j++) { 2818 Node* in = n->in(j); 2819 if (in->is_DecodeNarrowPtr()) { 2820 bool safe_to_skip = true; 2821 if (!is_uncommon ) { 2822 // Is it safe to skip? 2823 for (uint i = 0; i < in->outcnt(); i++) { 2824 Node* u = in->raw_out(i); 2825 if (!u->is_SafePoint() || 2826 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2827 safe_to_skip = false; 2828 } 2829 } 2830 } 2831 if (safe_to_skip) { 2832 n->set_req(j, in->in(1)); 2833 } 2834 if (in->outcnt() == 0) { 2835 in->disconnect_inputs(NULL, this); 2836 } 2837 } 2838 } 2839 } 2840 } 2841 2842 //------------------------------final_graph_reshaping-------------------------- 2843 // Final Graph Reshaping. 2844 // 2845 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2846 // and not commoned up and forced early. Must come after regular 2847 // optimizations to avoid GVN undoing the cloning. Clone constant 2848 // inputs to Loop Phis; these will be split by the allocator anyways. 2849 // Remove Opaque nodes. 2850 // (2) Move last-uses by commutative operations to the left input to encourage 2851 // Intel update-in-place two-address operations and better register usage 2852 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 2853 // calls canonicalizing them back. 2854 // (3) Count the number of double-precision FP ops, single-precision FP ops 2855 // and call sites. On Intel, we can get correct rounding either by 2856 // forcing singles to memory (requires extra stores and loads after each 2857 // FP bytecode) or we can set a rounding mode bit (requires setting and 2858 // clearing the mode bit around call sites). The mode bit is only used 2859 // if the relative frequency of single FP ops to calls is low enough. 2860 // This is a key transform for SPEC mpeg_audio. 2861 // (4) Detect infinite loops; blobs of code reachable from above but not 2862 // below. Several of the Code_Gen algorithms fail on such code shapes, 2863 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 2864 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 2865 // Detection is by looking for IfNodes where only 1 projection is 2866 // reachable from below or CatchNodes missing some targets. 2867 // (5) Assert for insane oop offsets in debug mode. 2868 2869 bool Compile::final_graph_reshaping() { 2870 // an infinite loop may have been eliminated by the optimizer, 2871 // in which case the graph will be empty. 2872 if (root()->req() == 1) { 2873 record_method_not_compilable("trivial infinite loop"); 2874 return true; 2875 } 2876 2877 Final_Reshape_Counts frc; 2878 2879 // Visit everybody reachable! 2880 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2881 Node_Stack nstack(unique() >> 1); 2882 final_graph_reshaping_walk(nstack, root(), frc); 2883 2884 // Check for unreachable (from below) code (i.e., infinite loops). 2885 for( uint i = 0; i < frc._tests.size(); i++ ) { 2886 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 2887 // Get number of CFG targets. 2888 // Note that PCTables include exception targets after calls. 2889 uint required_outcnt = n->required_outcnt(); 2890 if (n->outcnt() != required_outcnt) { 2891 // Check for a few special cases. Rethrow Nodes never take the 2892 // 'fall-thru' path, so expected kids is 1 less. 2893 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2894 if (n->in(0)->in(0)->is_Call()) { 2895 CallNode *call = n->in(0)->in(0)->as_Call(); 2896 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2897 required_outcnt--; // Rethrow always has 1 less kid 2898 } else if (call->req() > TypeFunc::Parms && 2899 call->is_CallDynamicJava()) { 2900 // Check for null receiver. In such case, the optimizer has 2901 // detected that the virtual call will always result in a null 2902 // pointer exception. The fall-through projection of this CatchNode 2903 // will not be populated. 2904 Node *arg0 = call->in(TypeFunc::Parms); 2905 if (arg0->is_Type() && 2906 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2907 required_outcnt--; 2908 } 2909 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2910 call->req() > TypeFunc::Parms+1 && 2911 call->is_CallStaticJava()) { 2912 // Check for negative array length. In such case, the optimizer has 2913 // detected that the allocation attempt will always result in an 2914 // exception. There is no fall-through projection of this CatchNode . 2915 Node *arg1 = call->in(TypeFunc::Parms+1); 2916 if (arg1->is_Type() && 2917 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2918 required_outcnt--; 2919 } 2920 } 2921 } 2922 } 2923 // Recheck with a better notion of 'required_outcnt' 2924 if (n->outcnt() != required_outcnt) { 2925 record_method_not_compilable("malformed control flow"); 2926 return true; // Not all targets reachable! 2927 } 2928 } 2929 // Check that I actually visited all kids. Unreached kids 2930 // must be infinite loops. 2931 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2932 if (!frc._visited.test(n->fast_out(j)->_idx)) { 2933 record_method_not_compilable("infinite loop"); 2934 return true; // Found unvisited kid; must be unreach 2935 } 2936 } 2937 2938 // If original bytecodes contained a mixture of floats and doubles 2939 // check if the optimizer has made it homogenous, item (3). 2940 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 2941 frc.get_float_count() > 32 && 2942 frc.get_double_count() == 0 && 2943 (10 * frc.get_call_count() < frc.get_float_count()) ) { 2944 set_24_bit_selection_and_mode( false, true ); 2945 } 2946 2947 set_java_calls(frc.get_java_call_count()); 2948 set_inner_loops(frc.get_inner_loop_count()); 2949 2950 // No infinite loops, no reason to bail out. 2951 return false; 2952 } 2953 2954 //-----------------------------too_many_traps---------------------------------- 2955 // Report if there are too many traps at the current method and bci. 2956 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2957 bool Compile::too_many_traps(ciMethod* method, 2958 int bci, 2959 Deoptimization::DeoptReason reason) { 2960 ciMethodData* md = method->method_data(); 2961 if (md->is_empty()) { 2962 // Assume the trap has not occurred, or that it occurred only 2963 // because of a transient condition during start-up in the interpreter. 2964 return false; 2965 } 2966 if (md->has_trap_at(bci, reason) != 0) { 2967 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2968 // Also, if there are multiple reasons, or if there is no per-BCI record, 2969 // assume the worst. 2970 if (log()) 2971 log()->elem("observe trap='%s' count='%d'", 2972 Deoptimization::trap_reason_name(reason), 2973 md->trap_count(reason)); 2974 return true; 2975 } else { 2976 // Ignore method/bci and see if there have been too many globally. 2977 return too_many_traps(reason, md); 2978 } 2979 } 2980 2981 // Less-accurate variant which does not require a method and bci. 2982 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2983 ciMethodData* logmd) { 2984 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2985 // Too many traps globally. 2986 // Note that we use cumulative trap_count, not just md->trap_count. 2987 if (log()) { 2988 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2989 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2990 Deoptimization::trap_reason_name(reason), 2991 mcount, trap_count(reason)); 2992 } 2993 return true; 2994 } else { 2995 // The coast is clear. 2996 return false; 2997 } 2998 } 2999 3000 //--------------------------too_many_recompiles-------------------------------- 3001 // Report if there are too many recompiles at the current method and bci. 3002 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3003 // Is not eager to return true, since this will cause the compiler to use 3004 // Action_none for a trap point, to avoid too many recompilations. 3005 bool Compile::too_many_recompiles(ciMethod* method, 3006 int bci, 3007 Deoptimization::DeoptReason reason) { 3008 ciMethodData* md = method->method_data(); 3009 if (md->is_empty()) { 3010 // Assume the trap has not occurred, or that it occurred only 3011 // because of a transient condition during start-up in the interpreter. 3012 return false; 3013 } 3014 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3015 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3016 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3017 Deoptimization::DeoptReason per_bc_reason 3018 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3019 if ((per_bc_reason == Deoptimization::Reason_none 3020 || md->has_trap_at(bci, reason) != 0) 3021 // The trap frequency measure we care about is the recompile count: 3022 && md->trap_recompiled_at(bci) 3023 && md->overflow_recompile_count() >= bc_cutoff) { 3024 // Do not emit a trap here if it has already caused recompilations. 3025 // Also, if there are multiple reasons, or if there is no per-BCI record, 3026 // assume the worst. 3027 if (log()) 3028 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3029 Deoptimization::trap_reason_name(reason), 3030 md->trap_count(reason), 3031 md->overflow_recompile_count()); 3032 return true; 3033 } else if (trap_count(reason) != 0 3034 && decompile_count() >= m_cutoff) { 3035 // Too many recompiles globally, and we have seen this sort of trap. 3036 // Use cumulative decompile_count, not just md->decompile_count. 3037 if (log()) 3038 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3039 Deoptimization::trap_reason_name(reason), 3040 md->trap_count(reason), trap_count(reason), 3041 md->decompile_count(), decompile_count()); 3042 return true; 3043 } else { 3044 // The coast is clear. 3045 return false; 3046 } 3047 } 3048 3049 3050 #ifndef PRODUCT 3051 //------------------------------verify_graph_edges--------------------------- 3052 // Walk the Graph and verify that there is a one-to-one correspondence 3053 // between Use-Def edges and Def-Use edges in the graph. 3054 void Compile::verify_graph_edges(bool no_dead_code) { 3055 if (VerifyGraphEdges) { 3056 ResourceArea *area = Thread::current()->resource_area(); 3057 Unique_Node_List visited(area); 3058 // Call recursive graph walk to check edges 3059 _root->verify_edges(visited); 3060 if (no_dead_code) { 3061 // Now make sure that no visited node is used by an unvisited node. 3062 bool dead_nodes = 0; 3063 Unique_Node_List checked(area); 3064 while (visited.size() > 0) { 3065 Node* n = visited.pop(); 3066 checked.push(n); 3067 for (uint i = 0; i < n->outcnt(); i++) { 3068 Node* use = n->raw_out(i); 3069 if (checked.member(use)) continue; // already checked 3070 if (visited.member(use)) continue; // already in the graph 3071 if (use->is_Con()) continue; // a dead ConNode is OK 3072 // At this point, we have found a dead node which is DU-reachable. 3073 if (dead_nodes++ == 0) 3074 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3075 use->dump(2); 3076 tty->print_cr("---"); 3077 checked.push(use); // No repeats; pretend it is now checked. 3078 } 3079 } 3080 assert(dead_nodes == 0, "using nodes must be reachable from root"); 3081 } 3082 } 3083 } 3084 #endif 3085 3086 // The Compile object keeps track of failure reasons separately from the ciEnv. 3087 // This is required because there is not quite a 1-1 relation between the 3088 // ciEnv and its compilation task and the Compile object. Note that one 3089 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3090 // to backtrack and retry without subsuming loads. Other than this backtracking 3091 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3092 // by the logic in C2Compiler. 3093 void Compile::record_failure(const char* reason) { 3094 if (log() != NULL) { 3095 log()->elem("failure reason='%s' phase='compile'", reason); 3096 } 3097 if (_failure_reason == NULL) { 3098 // Record the first failure reason. 3099 _failure_reason = reason; 3100 } 3101 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3102 C->print_method(_failure_reason); 3103 } 3104 _root = NULL; // flush the graph, too 3105 } 3106 3107 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3108 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3109 _phase_name(name), _dolog(dolog) 3110 { 3111 if (dolog) { 3112 C = Compile::current(); 3113 _log = C->log(); 3114 } else { 3115 C = NULL; 3116 _log = NULL; 3117 } 3118 if (_log != NULL) { 3119 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3120 _log->stamp(); 3121 _log->end_head(); 3122 } 3123 } 3124 3125 Compile::TracePhase::~TracePhase() { 3126 3127 C = Compile::current(); 3128 if (_dolog) { 3129 _log = C->log(); 3130 } else { 3131 _log = NULL; 3132 } 3133 3134 #ifdef ASSERT 3135 if (PrintIdealNodeCount) { 3136 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3137 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3138 } 3139 3140 if (VerifyIdealNodeCount) { 3141 Compile::current()->print_missing_nodes(); 3142 } 3143 #endif 3144 3145 if (_log != NULL) { 3146 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3147 } 3148 } 3149 3150 //============================================================================= 3151 // Two Constant's are equal when the type and the value are equal. 3152 bool Compile::Constant::operator==(const Constant& other) { 3153 if (type() != other.type() ) return false; 3154 if (can_be_reused() != other.can_be_reused()) return false; 3155 // For floating point values we compare the bit pattern. 3156 switch (type()) { 3157 case T_FLOAT: return (_v._value.i == other._v._value.i); 3158 case T_LONG: 3159 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3160 case T_OBJECT: 3161 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3162 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3163 case T_METADATA: return (_v._metadata == other._v._metadata); 3164 default: ShouldNotReachHere(); 3165 } 3166 return false; 3167 } 3168 3169 static int type_to_size_in_bytes(BasicType t) { 3170 switch (t) { 3171 case T_LONG: return sizeof(jlong ); 3172 case T_FLOAT: return sizeof(jfloat ); 3173 case T_DOUBLE: return sizeof(jdouble); 3174 case T_METADATA: return sizeof(Metadata*); 3175 // We use T_VOID as marker for jump-table entries (labels) which 3176 // need an internal word relocation. 3177 case T_VOID: 3178 case T_ADDRESS: 3179 case T_OBJECT: return sizeof(jobject); 3180 } 3181 3182 ShouldNotReachHere(); 3183 return -1; 3184 } 3185 3186 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3187 // sort descending 3188 if (a->freq() > b->freq()) return -1; 3189 if (a->freq() < b->freq()) return 1; 3190 return 0; 3191 } 3192 3193 void Compile::ConstantTable::calculate_offsets_and_size() { 3194 // First, sort the array by frequencies. 3195 _constants.sort(qsort_comparator); 3196 3197 #ifdef ASSERT 3198 // Make sure all jump-table entries were sorted to the end of the 3199 // array (they have a negative frequency). 3200 bool found_void = false; 3201 for (int i = 0; i < _constants.length(); i++) { 3202 Constant con = _constants.at(i); 3203 if (con.type() == T_VOID) 3204 found_void = true; // jump-tables 3205 else 3206 assert(!found_void, "wrong sorting"); 3207 } 3208 #endif 3209 3210 int offset = 0; 3211 for (int i = 0; i < _constants.length(); i++) { 3212 Constant* con = _constants.adr_at(i); 3213 3214 // Align offset for type. 3215 int typesize = type_to_size_in_bytes(con->type()); 3216 offset = align_size_up(offset, typesize); 3217 con->set_offset(offset); // set constant's offset 3218 3219 if (con->type() == T_VOID) { 3220 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3221 offset = offset + typesize * n->outcnt(); // expand jump-table 3222 } else { 3223 offset = offset + typesize; 3224 } 3225 } 3226 3227 // Align size up to the next section start (which is insts; see 3228 // CodeBuffer::align_at_start). 3229 assert(_size == -1, "already set?"); 3230 _size = align_size_up(offset, CodeEntryAlignment); 3231 } 3232 3233 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3234 MacroAssembler _masm(&cb); 3235 for (int i = 0; i < _constants.length(); i++) { 3236 Constant con = _constants.at(i); 3237 address constant_addr; 3238 switch (con.type()) { 3239 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3240 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3241 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3242 case T_OBJECT: { 3243 jobject obj = con.get_jobject(); 3244 int oop_index = _masm.oop_recorder()->find_index(obj); 3245 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3246 break; 3247 } 3248 case T_ADDRESS: { 3249 address addr = (address) con.get_jobject(); 3250 constant_addr = _masm.address_constant(addr); 3251 break; 3252 } 3253 // We use T_VOID as marker for jump-table entries (labels) which 3254 // need an internal word relocation. 3255 case T_VOID: { 3256 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3257 // Fill the jump-table with a dummy word. The real value is 3258 // filled in later in fill_jump_table. 3259 address dummy = (address) n; 3260 constant_addr = _masm.address_constant(dummy); 3261 // Expand jump-table 3262 for (uint i = 1; i < n->outcnt(); i++) { 3263 address temp_addr = _masm.address_constant(dummy + i); 3264 assert(temp_addr, "consts section too small"); 3265 } 3266 break; 3267 } 3268 case T_METADATA: { 3269 Metadata* obj = con.get_metadata(); 3270 int metadata_index = _masm.oop_recorder()->find_index(obj); 3271 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3272 break; 3273 } 3274 default: ShouldNotReachHere(); 3275 } 3276 assert(constant_addr, "consts section too small"); 3277 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3278 } 3279 } 3280 3281 int Compile::ConstantTable::find_offset(Constant& con) const { 3282 int idx = _constants.find(con); 3283 assert(idx != -1, "constant must be in constant table"); 3284 int offset = _constants.at(idx).offset(); 3285 assert(offset != -1, "constant table not emitted yet?"); 3286 return offset; 3287 } 3288 3289 void Compile::ConstantTable::add(Constant& con) { 3290 if (con.can_be_reused()) { 3291 int idx = _constants.find(con); 3292 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3293 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3294 return; 3295 } 3296 } 3297 (void) _constants.append(con); 3298 } 3299 3300 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3301 Block* b = Compile::current()->cfg()->_bbs[n->_idx]; 3302 Constant con(type, value, b->_freq); 3303 add(con); 3304 return con; 3305 } 3306 3307 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3308 Constant con(metadata); 3309 add(con); 3310 return con; 3311 } 3312 3313 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3314 jvalue value; 3315 BasicType type = oper->type()->basic_type(); 3316 switch (type) { 3317 case T_LONG: value.j = oper->constantL(); break; 3318 case T_FLOAT: value.f = oper->constantF(); break; 3319 case T_DOUBLE: value.d = oper->constantD(); break; 3320 case T_OBJECT: 3321 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3322 case T_METADATA: return add((Metadata*)oper->constant()); break; 3323 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3324 } 3325 return add(n, type, value); 3326 } 3327 3328 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3329 jvalue value; 3330 // We can use the node pointer here to identify the right jump-table 3331 // as this method is called from Compile::Fill_buffer right before 3332 // the MachNodes are emitted and the jump-table is filled (means the 3333 // MachNode pointers do not change anymore). 3334 value.l = (jobject) n; 3335 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3336 add(con); 3337 return con; 3338 } 3339 3340 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3341 // If called from Compile::scratch_emit_size do nothing. 3342 if (Compile::current()->in_scratch_emit_size()) return; 3343 3344 assert(labels.is_nonempty(), "must be"); 3345 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3346 3347 // Since MachConstantNode::constant_offset() also contains 3348 // table_base_offset() we need to subtract the table_base_offset() 3349 // to get the plain offset into the constant table. 3350 int offset = n->constant_offset() - table_base_offset(); 3351 3352 MacroAssembler _masm(&cb); 3353 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3354 3355 for (uint i = 0; i < n->outcnt(); i++) { 3356 address* constant_addr = &jump_table_base[i]; 3357 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))); 3358 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3359 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3360 } 3361 } 3362 3363 void Compile::dump_inlining() { 3364 if (PrintInlining) { 3365 for (int i = 0; i < _print_inlining_list->length(); i++) { 3366 tty->print(_print_inlining_list->at(i).ss()->as_string()); 3367 } 3368 } 3369 }