1 /* 2 * Copyright (c) 1997, 2013, 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 "trace/tracing.hpp" 67 #include "utilities/copy.hpp" 68 #ifdef TARGET_ARCH_MODEL_x86_32 69 # include "adfiles/ad_x86_32.hpp" 70 #endif 71 #ifdef TARGET_ARCH_MODEL_x86_64 72 # include "adfiles/ad_x86_64.hpp" 73 #endif 74 #ifdef TARGET_ARCH_MODEL_sparc 75 # include "adfiles/ad_sparc.hpp" 76 #endif 77 #ifdef TARGET_ARCH_MODEL_zero 78 # include "adfiles/ad_zero.hpp" 79 #endif 80 #ifdef TARGET_ARCH_MODEL_arm 81 # include "adfiles/ad_arm.hpp" 82 #endif 83 #ifdef TARGET_ARCH_MODEL_ppc 84 # include "adfiles/ad_ppc.hpp" 85 #endif 86 87 88 // -------------------- Compile::mach_constant_base_node ----------------------- 89 // Constant table base node singleton. 90 MachConstantBaseNode* Compile::mach_constant_base_node() { 91 if (_mach_constant_base_node == NULL) { 92 _mach_constant_base_node = new (C) MachConstantBaseNode(); 93 _mach_constant_base_node->add_req(C->root()); 94 } 95 return _mach_constant_base_node; 96 } 97 98 99 /// Support for intrinsics. 100 101 // Return the index at which m must be inserted (or already exists). 102 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 103 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 104 #ifdef ASSERT 105 for (int i = 1; i < _intrinsics->length(); i++) { 106 CallGenerator* cg1 = _intrinsics->at(i-1); 107 CallGenerator* cg2 = _intrinsics->at(i); 108 assert(cg1->method() != cg2->method() 109 ? cg1->method() < cg2->method() 110 : cg1->is_virtual() < cg2->is_virtual(), 111 "compiler intrinsics list must stay sorted"); 112 } 113 #endif 114 // Binary search sorted list, in decreasing intervals [lo, hi]. 115 int lo = 0, hi = _intrinsics->length()-1; 116 while (lo <= hi) { 117 int mid = (uint)(hi + lo) / 2; 118 ciMethod* mid_m = _intrinsics->at(mid)->method(); 119 if (m < mid_m) { 120 hi = mid-1; 121 } else if (m > mid_m) { 122 lo = mid+1; 123 } else { 124 // look at minor sort key 125 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 126 if (is_virtual < mid_virt) { 127 hi = mid-1; 128 } else if (is_virtual > mid_virt) { 129 lo = mid+1; 130 } else { 131 return mid; // exact match 132 } 133 } 134 } 135 return lo; // inexact match 136 } 137 138 void Compile::register_intrinsic(CallGenerator* cg) { 139 if (_intrinsics == NULL) { 140 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 141 } 142 // This code is stolen from ciObjectFactory::insert. 143 // Really, GrowableArray should have methods for 144 // insert_at, remove_at, and binary_search. 145 int len = _intrinsics->length(); 146 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 147 if (index == len) { 148 _intrinsics->append(cg); 149 } else { 150 #ifdef ASSERT 151 CallGenerator* oldcg = _intrinsics->at(index); 152 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 153 #endif 154 _intrinsics->append(_intrinsics->at(len-1)); 155 int pos; 156 for (pos = len-2; pos >= index; pos--) { 157 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 158 } 159 _intrinsics->at_put(index, cg); 160 } 161 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 162 } 163 164 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 165 assert(m->is_loaded(), "don't try this on unloaded methods"); 166 if (_intrinsics != NULL) { 167 int index = intrinsic_insertion_index(m, is_virtual); 168 if (index < _intrinsics->length() 169 && _intrinsics->at(index)->method() == m 170 && _intrinsics->at(index)->is_virtual() == is_virtual) { 171 return _intrinsics->at(index); 172 } 173 } 174 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 175 if (m->intrinsic_id() != vmIntrinsics::_none && 176 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 177 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 178 if (cg != NULL) { 179 // Save it for next time: 180 register_intrinsic(cg); 181 return cg; 182 } else { 183 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 184 } 185 } 186 return NULL; 187 } 188 189 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 190 // in library_call.cpp. 191 192 193 #ifndef PRODUCT 194 // statistics gathering... 195 196 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 197 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 198 199 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 200 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 201 int oflags = _intrinsic_hist_flags[id]; 202 assert(flags != 0, "what happened?"); 203 if (is_virtual) { 204 flags |= _intrinsic_virtual; 205 } 206 bool changed = (flags != oflags); 207 if ((flags & _intrinsic_worked) != 0) { 208 juint count = (_intrinsic_hist_count[id] += 1); 209 if (count == 1) { 210 changed = true; // first time 211 } 212 // increment the overall count also: 213 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 214 } 215 if (changed) { 216 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 217 // Something changed about the intrinsic's virtuality. 218 if ((flags & _intrinsic_virtual) != 0) { 219 // This is the first use of this intrinsic as a virtual call. 220 if (oflags != 0) { 221 // We already saw it as a non-virtual, so note both cases. 222 flags |= _intrinsic_both; 223 } 224 } else if ((oflags & _intrinsic_both) == 0) { 225 // This is the first use of this intrinsic as a non-virtual 226 flags |= _intrinsic_both; 227 } 228 } 229 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 230 } 231 // update the overall flags also: 232 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 233 return changed; 234 } 235 236 static char* format_flags(int flags, char* buf) { 237 buf[0] = 0; 238 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 239 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 240 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 241 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 242 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 243 if (buf[0] == 0) strcat(buf, ","); 244 assert(buf[0] == ',', "must be"); 245 return &buf[1]; 246 } 247 248 void Compile::print_intrinsic_statistics() { 249 char flagsbuf[100]; 250 ttyLocker ttyl; 251 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 252 tty->print_cr("Compiler intrinsic usage:"); 253 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 254 if (total == 0) total = 1; // avoid div0 in case of no successes 255 #define PRINT_STAT_LINE(name, c, f) \ 256 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 257 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 258 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 259 int flags = _intrinsic_hist_flags[id]; 260 juint count = _intrinsic_hist_count[id]; 261 if ((flags | count) != 0) { 262 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 263 } 264 } 265 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 266 if (xtty != NULL) xtty->tail("statistics"); 267 } 268 269 void Compile::print_statistics() { 270 { ttyLocker ttyl; 271 if (xtty != NULL) xtty->head("statistics type='opto'"); 272 Parse::print_statistics(); 273 PhaseCCP::print_statistics(); 274 PhaseRegAlloc::print_statistics(); 275 Scheduling::print_statistics(); 276 PhasePeephole::print_statistics(); 277 PhaseIdealLoop::print_statistics(); 278 if (xtty != NULL) xtty->tail("statistics"); 279 } 280 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 281 // put this under its own <statistics> element. 282 print_intrinsic_statistics(); 283 } 284 } 285 #endif //PRODUCT 286 287 // Support for bundling info 288 Bundle* Compile::node_bundling(const Node *n) { 289 assert(valid_bundle_info(n), "oob"); 290 return &_node_bundling_base[n->_idx]; 291 } 292 293 bool Compile::valid_bundle_info(const Node *n) { 294 return (_node_bundling_limit > n->_idx); 295 } 296 297 298 void Compile::gvn_replace_by(Node* n, Node* nn) { 299 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 300 Node* use = n->last_out(i); 301 bool is_in_table = initial_gvn()->hash_delete(use); 302 uint uses_found = 0; 303 for (uint j = 0; j < use->len(); j++) { 304 if (use->in(j) == n) { 305 if (j < use->req()) 306 use->set_req(j, nn); 307 else 308 use->set_prec(j, nn); 309 uses_found++; 310 } 311 } 312 if (is_in_table) { 313 // reinsert into table 314 initial_gvn()->hash_find_insert(use); 315 } 316 record_for_igvn(use); 317 i -= uses_found; // we deleted 1 or more copies of this edge 318 } 319 } 320 321 322 static inline bool not_a_node(const Node* n) { 323 if (n == NULL) return true; 324 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 325 if (*(address*)n == badAddress) return true; // kill by Node::destruct 326 return false; 327 } 328 329 // Identify all nodes that are reachable from below, useful. 330 // Use breadth-first pass that records state in a Unique_Node_List, 331 // recursive traversal is slower. 332 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 333 int estimated_worklist_size = unique(); 334 useful.map( estimated_worklist_size, NULL ); // preallocate space 335 336 // Initialize worklist 337 if (root() != NULL) { useful.push(root()); } 338 // If 'top' is cached, declare it useful to preserve cached node 339 if( cached_top_node() ) { useful.push(cached_top_node()); } 340 341 // Push all useful nodes onto the list, breadthfirst 342 for( uint next = 0; next < useful.size(); ++next ) { 343 assert( next < unique(), "Unique useful nodes < total nodes"); 344 Node *n = useful.at(next); 345 uint max = n->len(); 346 for( uint i = 0; i < max; ++i ) { 347 Node *m = n->in(i); 348 if (not_a_node(m)) continue; 349 useful.push(m); 350 } 351 } 352 } 353 354 // Update dead_node_list with any missing dead nodes using useful 355 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 356 void Compile::update_dead_node_list(Unique_Node_List &useful) { 357 uint max_idx = unique(); 358 VectorSet& useful_node_set = useful.member_set(); 359 360 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 361 // If node with index node_idx is not in useful set, 362 // mark it as dead in dead node list. 363 if (! useful_node_set.test(node_idx) ) { 364 record_dead_node(node_idx); 365 } 366 } 367 } 368 369 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 370 int shift = 0; 371 for (int i = 0; i < inlines->length(); i++) { 372 CallGenerator* cg = inlines->at(i); 373 CallNode* call = cg->call_node(); 374 if (shift > 0) { 375 inlines->at_put(i-shift, cg); 376 } 377 if (!useful.member(call)) { 378 shift++; 379 } 380 } 381 inlines->trunc_to(inlines->length()-shift); 382 } 383 384 // Disconnect all useless nodes by disconnecting those at the boundary. 385 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 386 uint next = 0; 387 while (next < useful.size()) { 388 Node *n = useful.at(next++); 389 // Use raw traversal of out edges since this code removes out edges 390 int max = n->outcnt(); 391 for (int j = 0; j < max; ++j) { 392 Node* child = n->raw_out(j); 393 if (! useful.member(child)) { 394 assert(!child->is_top() || child != top(), 395 "If top is cached in Compile object it is in useful list"); 396 // Only need to remove this out-edge to the useless node 397 n->raw_del_out(j); 398 --j; 399 --max; 400 } 401 } 402 if (n->outcnt() == 1 && n->has_special_unique_user()) { 403 record_for_igvn(n->unique_out()); 404 } 405 } 406 // Remove useless macro and predicate opaq nodes 407 for (int i = C->macro_count()-1; i >= 0; i--) { 408 Node* n = C->macro_node(i); 409 if (!useful.member(n)) { 410 remove_macro_node(n); 411 } 412 } 413 // Remove useless expensive node 414 for (int i = C->expensive_count()-1; i >= 0; i--) { 415 Node* n = C->expensive_node(i); 416 if (!useful.member(n)) { 417 remove_expensive_node(n); 418 } 419 } 420 // clean up the late inline lists 421 remove_useless_late_inlines(&_string_late_inlines, useful); 422 remove_useless_late_inlines(&_boxing_late_inlines, useful); 423 remove_useless_late_inlines(&_late_inlines, useful); 424 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 425 } 426 427 //------------------------------frame_size_in_words----------------------------- 428 // frame_slots in units of words 429 int Compile::frame_size_in_words() const { 430 // shift is 0 in LP32 and 1 in LP64 431 const int shift = (LogBytesPerWord - LogBytesPerInt); 432 int words = _frame_slots >> shift; 433 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 434 return words; 435 } 436 437 // ============================================================================ 438 //------------------------------CompileWrapper--------------------------------- 439 class CompileWrapper : public StackObj { 440 Compile *const _compile; 441 public: 442 CompileWrapper(Compile* compile); 443 444 ~CompileWrapper(); 445 }; 446 447 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 448 // the Compile* pointer is stored in the current ciEnv: 449 ciEnv* env = compile->env(); 450 assert(env == ciEnv::current(), "must already be a ciEnv active"); 451 assert(env->compiler_data() == NULL, "compile already active?"); 452 env->set_compiler_data(compile); 453 assert(compile == Compile::current(), "sanity"); 454 455 compile->set_type_dict(NULL); 456 compile->set_type_hwm(NULL); 457 compile->set_type_last_size(0); 458 compile->set_last_tf(NULL, NULL); 459 compile->set_indexSet_arena(NULL); 460 compile->set_indexSet_free_block_list(NULL); 461 compile->init_type_arena(); 462 Type::Initialize(compile); 463 _compile->set_scratch_buffer_blob(NULL); 464 _compile->begin_method(); 465 } 466 CompileWrapper::~CompileWrapper() { 467 _compile->end_method(); 468 if (_compile->scratch_buffer_blob() != NULL) 469 BufferBlob::free(_compile->scratch_buffer_blob()); 470 _compile->env()->set_compiler_data(NULL); 471 } 472 473 474 //----------------------------print_compile_messages--------------------------- 475 void Compile::print_compile_messages() { 476 #ifndef PRODUCT 477 // Check if recompiling 478 if (_subsume_loads == false && PrintOpto) { 479 // Recompiling without allowing machine instructions to subsume loads 480 tty->print_cr("*********************************************************"); 481 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 482 tty->print_cr("*********************************************************"); 483 } 484 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 485 // Recompiling without escape analysis 486 tty->print_cr("*********************************************************"); 487 tty->print_cr("** Bailout: Recompile without escape analysis **"); 488 tty->print_cr("*********************************************************"); 489 } 490 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 491 // Recompiling without boxing elimination 492 tty->print_cr("*********************************************************"); 493 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 494 tty->print_cr("*********************************************************"); 495 } 496 if (env()->break_at_compile()) { 497 // Open the debugger when compiling this method. 498 tty->print("### Breaking when compiling: "); 499 method()->print_short_name(); 500 tty->cr(); 501 BREAKPOINT; 502 } 503 504 if( PrintOpto ) { 505 if (is_osr_compilation()) { 506 tty->print("[OSR]%3d", _compile_id); 507 } else { 508 tty->print("%3d", _compile_id); 509 } 510 } 511 #endif 512 } 513 514 515 //-----------------------init_scratch_buffer_blob------------------------------ 516 // Construct a temporary BufferBlob and cache it for this compile. 517 void Compile::init_scratch_buffer_blob(int const_size) { 518 // If there is already a scratch buffer blob allocated and the 519 // constant section is big enough, use it. Otherwise free the 520 // current and allocate a new one. 521 BufferBlob* blob = scratch_buffer_blob(); 522 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 523 // Use the current blob. 524 } else { 525 if (blob != NULL) { 526 BufferBlob::free(blob); 527 } 528 529 ResourceMark rm; 530 _scratch_const_size = const_size; 531 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 532 blob = BufferBlob::create("Compile::scratch_buffer", size); 533 // Record the buffer blob for next time. 534 set_scratch_buffer_blob(blob); 535 // Have we run out of code space? 536 if (scratch_buffer_blob() == NULL) { 537 // Let CompilerBroker disable further compilations. 538 record_failure("Not enough space for scratch buffer in CodeCache"); 539 return; 540 } 541 } 542 543 // Initialize the relocation buffers 544 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 545 set_scratch_locs_memory(locs_buf); 546 } 547 548 549 //-----------------------scratch_emit_size------------------------------------- 550 // Helper function that computes size by emitting code 551 uint Compile::scratch_emit_size(const Node* n) { 552 // Start scratch_emit_size section. 553 set_in_scratch_emit_size(true); 554 555 // Emit into a trash buffer and count bytes emitted. 556 // This is a pretty expensive way to compute a size, 557 // but it works well enough if seldom used. 558 // All common fixed-size instructions are given a size 559 // method by the AD file. 560 // Note that the scratch buffer blob and locs memory are 561 // allocated at the beginning of the compile task, and 562 // may be shared by several calls to scratch_emit_size. 563 // The allocation of the scratch buffer blob is particularly 564 // expensive, since it has to grab the code cache lock. 565 BufferBlob* blob = this->scratch_buffer_blob(); 566 assert(blob != NULL, "Initialize BufferBlob at start"); 567 assert(blob->size() > MAX_inst_size, "sanity"); 568 relocInfo* locs_buf = scratch_locs_memory(); 569 address blob_begin = blob->content_begin(); 570 address blob_end = (address)locs_buf; 571 assert(blob->content_contains(blob_end), "sanity"); 572 CodeBuffer buf(blob_begin, blob_end - blob_begin); 573 buf.initialize_consts_size(_scratch_const_size); 574 buf.initialize_stubs_size(MAX_stubs_size); 575 assert(locs_buf != NULL, "sanity"); 576 int lsize = MAX_locs_size / 3; 577 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 578 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 579 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 580 581 // Do the emission. 582 583 Label fakeL; // Fake label for branch instructions. 584 Label* saveL = NULL; 585 uint save_bnum = 0; 586 bool is_branch = n->is_MachBranch(); 587 if (is_branch) { 588 MacroAssembler masm(&buf); 589 masm.bind(fakeL); 590 n->as_MachBranch()->save_label(&saveL, &save_bnum); 591 n->as_MachBranch()->label_set(&fakeL, 0); 592 } 593 n->emit(buf, this->regalloc()); 594 if (is_branch) // Restore label. 595 n->as_MachBranch()->label_set(saveL, save_bnum); 596 597 // End scratch_emit_size section. 598 set_in_scratch_emit_size(false); 599 600 return buf.insts_size(); 601 } 602 603 604 // ============================================================================ 605 //------------------------------Compile standard------------------------------- 606 debug_only( int Compile::_debug_idx = 100000; ) 607 608 // Compile a method. entry_bci is -1 for normal compilations and indicates 609 // the continuation bci for on stack replacement. 610 611 612 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 613 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 614 : Phase(Compiler), 615 _env(ci_env), 616 _log(ci_env->log()), 617 _compile_id(ci_env->compile_id()), 618 _save_argument_registers(false), 619 _stub_name(NULL), 620 _stub_function(NULL), 621 _stub_entry_point(NULL), 622 _method(target), 623 _entry_bci(osr_bci), 624 _initial_gvn(NULL), 625 _for_igvn(NULL), 626 _warm_calls(NULL), 627 _subsume_loads(subsume_loads), 628 _do_escape_analysis(do_escape_analysis), 629 _eliminate_boxing(eliminate_boxing), 630 _failure_reason(NULL), 631 _code_buffer("Compile::Fill_buffer"), 632 _orig_pc_slot(0), 633 _orig_pc_slot_offset_in_bytes(0), 634 _has_method_handle_invokes(false), 635 _mach_constant_base_node(NULL), 636 _node_bundling_limit(0), 637 _node_bundling_base(NULL), 638 _java_calls(0), 639 _inner_loops(0), 640 _scratch_const_size(-1), 641 _in_scratch_emit_size(false), 642 _dead_node_list(comp_arena()), 643 _dead_node_count(0), 644 #ifndef PRODUCT 645 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 646 _printer(IdealGraphPrinter::printer()), 647 #endif 648 _congraph(NULL), 649 _late_inlines(comp_arena(), 2, 0, NULL), 650 _string_late_inlines(comp_arena(), 2, 0, NULL), 651 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 652 _late_inlines_pos(0), 653 _number_of_mh_late_inlines(0), 654 _inlining_progress(false), 655 _inlining_incrementally(false), 656 _print_inlining_list(NULL), 657 _print_inlining(0) { 658 C = this; 659 660 CompileWrapper cw(this); 661 #ifndef PRODUCT 662 if (TimeCompiler2) { 663 tty->print(" "); 664 target->holder()->name()->print(); 665 tty->print("."); 666 target->print_short_name(); 667 tty->print(" "); 668 } 669 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 670 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 671 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 672 if (!print_opto_assembly) { 673 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 674 if (print_assembly && !Disassembler::can_decode()) { 675 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 676 print_opto_assembly = true; 677 } 678 } 679 set_print_assembly(print_opto_assembly); 680 set_parsed_irreducible_loop(false); 681 #endif 682 683 if (ProfileTraps) { 684 // Make sure the method being compiled gets its own MDO, 685 // so we can at least track the decompile_count(). 686 method()->ensure_method_data(); 687 } 688 689 Init(::AliasLevel); 690 691 692 print_compile_messages(); 693 694 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 695 _ilt = InlineTree::build_inline_tree_root(); 696 else 697 _ilt = NULL; 698 699 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 700 assert(num_alias_types() >= AliasIdxRaw, ""); 701 702 #define MINIMUM_NODE_HASH 1023 703 // Node list that Iterative GVN will start with 704 Unique_Node_List for_igvn(comp_arena()); 705 set_for_igvn(&for_igvn); 706 707 // GVN that will be run immediately on new nodes 708 uint estimated_size = method()->code_size()*4+64; 709 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 710 PhaseGVN gvn(node_arena(), estimated_size); 711 set_initial_gvn(&gvn); 712 713 if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) { 714 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 715 } 716 { // Scope for timing the parser 717 TracePhase t3("parse", &_t_parser, true); 718 719 // Put top into the hash table ASAP. 720 initial_gvn()->transform_no_reclaim(top()); 721 722 // Set up tf(), start(), and find a CallGenerator. 723 CallGenerator* cg = NULL; 724 if (is_osr_compilation()) { 725 const TypeTuple *domain = StartOSRNode::osr_domain(); 726 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 727 init_tf(TypeFunc::make(domain, range)); 728 StartNode* s = new (this) StartOSRNode(root(), domain); 729 initial_gvn()->set_type_bottom(s); 730 init_start(s); 731 cg = CallGenerator::for_osr(method(), entry_bci()); 732 } else { 733 // Normal case. 734 init_tf(TypeFunc::make(method())); 735 StartNode* s = new (this) StartNode(root(), tf()->domain()); 736 initial_gvn()->set_type_bottom(s); 737 init_start(s); 738 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 739 // With java.lang.ref.reference.get() we must go through the 740 // intrinsic when G1 is enabled - even when get() is the root 741 // method of the compile - so that, if necessary, the value in 742 // the referent field of the reference object gets recorded by 743 // the pre-barrier code. 744 // Specifically, if G1 is enabled, the value in the referent 745 // field is recorded by the G1 SATB pre barrier. This will 746 // result in the referent being marked live and the reference 747 // object removed from the list of discovered references during 748 // reference processing. 749 cg = find_intrinsic(method(), false); 750 } 751 if (cg == NULL) { 752 float past_uses = method()->interpreter_invocation_count(); 753 float expected_uses = past_uses; 754 cg = CallGenerator::for_inline(method(), expected_uses); 755 } 756 } 757 if (failing()) return; 758 if (cg == NULL) { 759 record_method_not_compilable_all_tiers("cannot parse method"); 760 return; 761 } 762 JVMState* jvms = build_start_state(start(), tf()); 763 if ((jvms = cg->generate(jvms)) == NULL) { 764 record_method_not_compilable("method parse failed"); 765 return; 766 } 767 GraphKit kit(jvms); 768 769 if (!kit.stopped()) { 770 // Accept return values, and transfer control we know not where. 771 // This is done by a special, unique ReturnNode bound to root. 772 return_values(kit.jvms()); 773 } 774 775 if (kit.has_exceptions()) { 776 // Any exceptions that escape from this call must be rethrown 777 // to whatever caller is dynamically above us on the stack. 778 // This is done by a special, unique RethrowNode bound to root. 779 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 780 } 781 782 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 783 784 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 785 inline_string_calls(true); 786 } 787 788 if (failing()) return; 789 790 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 791 792 // Remove clutter produced by parsing. 793 if (!failing()) { 794 ResourceMark rm; 795 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 796 } 797 } 798 799 // Note: Large methods are capped off in do_one_bytecode(). 800 if (failing()) return; 801 802 // After parsing, node notes are no longer automagic. 803 // They must be propagated by register_new_node_with_optimizer(), 804 // clone(), or the like. 805 set_default_node_notes(NULL); 806 807 for (;;) { 808 int successes = Inline_Warm(); 809 if (failing()) return; 810 if (successes == 0) break; 811 } 812 813 // Drain the list. 814 Finish_Warm(); 815 #ifndef PRODUCT 816 if (_printer) { 817 _printer->print_inlining(this); 818 } 819 #endif 820 821 if (failing()) return; 822 NOT_PRODUCT( verify_graph_edges(); ) 823 824 // Now optimize 825 Optimize(); 826 if (failing()) return; 827 NOT_PRODUCT( verify_graph_edges(); ) 828 829 #ifndef PRODUCT 830 if (PrintIdeal) { 831 ttyLocker ttyl; // keep the following output all in one block 832 // This output goes directly to the tty, not the compiler log. 833 // To enable tools to match it up with the compilation activity, 834 // be sure to tag this tty output with the compile ID. 835 if (xtty != NULL) { 836 xtty->head("ideal compile_id='%d'%s", compile_id(), 837 is_osr_compilation() ? " compile_kind='osr'" : 838 ""); 839 } 840 root()->dump(9999); 841 if (xtty != NULL) { 842 xtty->tail("ideal"); 843 } 844 } 845 #endif 846 847 // Now that we know the size of all the monitors we can add a fixed slot 848 // for the original deopt pc. 849 850 _orig_pc_slot = fixed_slots(); 851 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 852 set_fixed_slots(next_slot); 853 854 // Now generate code 855 Code_Gen(); 856 if (failing()) return; 857 858 // Check if we want to skip execution of all compiled code. 859 { 860 #ifndef PRODUCT 861 if (OptoNoExecute) { 862 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 863 return; 864 } 865 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 866 #endif 867 868 if (is_osr_compilation()) { 869 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 870 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 871 } else { 872 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 873 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 874 } 875 876 env()->register_method(_method, _entry_bci, 877 &_code_offsets, 878 _orig_pc_slot_offset_in_bytes, 879 code_buffer(), 880 frame_size_in_words(), _oop_map_set, 881 &_handler_table, &_inc_table, 882 compiler, 883 env()->comp_level(), 884 has_unsafe_access(), 885 SharedRuntime::is_wide_vector(max_vector_size()) 886 ); 887 888 if (log() != NULL) // Print code cache state into compiler log 889 log()->code_cache_state(); 890 } 891 } 892 893 //------------------------------Compile---------------------------------------- 894 // Compile a runtime stub 895 Compile::Compile( ciEnv* ci_env, 896 TypeFunc_generator generator, 897 address stub_function, 898 const char *stub_name, 899 int is_fancy_jump, 900 bool pass_tls, 901 bool save_arg_registers, 902 bool return_pc ) 903 : Phase(Compiler), 904 _env(ci_env), 905 _log(ci_env->log()), 906 _compile_id(0), 907 _save_argument_registers(save_arg_registers), 908 _method(NULL), 909 _stub_name(stub_name), 910 _stub_function(stub_function), 911 _stub_entry_point(NULL), 912 _entry_bci(InvocationEntryBci), 913 _initial_gvn(NULL), 914 _for_igvn(NULL), 915 _warm_calls(NULL), 916 _orig_pc_slot(0), 917 _orig_pc_slot_offset_in_bytes(0), 918 _subsume_loads(true), 919 _do_escape_analysis(false), 920 _eliminate_boxing(false), 921 _failure_reason(NULL), 922 _code_buffer("Compile::Fill_buffer"), 923 _has_method_handle_invokes(false), 924 _mach_constant_base_node(NULL), 925 _node_bundling_limit(0), 926 _node_bundling_base(NULL), 927 _java_calls(0), 928 _inner_loops(0), 929 #ifndef PRODUCT 930 _trace_opto_output(TraceOptoOutput), 931 _printer(NULL), 932 #endif 933 _dead_node_list(comp_arena()), 934 _dead_node_count(0), 935 _congraph(NULL), 936 _number_of_mh_late_inlines(0), 937 _inlining_progress(false), 938 _inlining_incrementally(false), 939 _print_inlining_list(NULL), 940 _print_inlining(0) { 941 C = this; 942 943 #ifndef PRODUCT 944 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 945 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 946 set_print_assembly(PrintFrameConverterAssembly); 947 set_parsed_irreducible_loop(false); 948 #endif 949 CompileWrapper cw(this); 950 Init(/*AliasLevel=*/ 0); 951 init_tf((*generator)()); 952 953 { 954 // The following is a dummy for the sake of GraphKit::gen_stub 955 Unique_Node_List for_igvn(comp_arena()); 956 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 957 PhaseGVN gvn(Thread::current()->resource_area(),255); 958 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 959 gvn.transform_no_reclaim(top()); 960 961 GraphKit kit; 962 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 963 } 964 965 NOT_PRODUCT( verify_graph_edges(); ) 966 Code_Gen(); 967 if (failing()) return; 968 969 970 // Entry point will be accessed using compile->stub_entry_point(); 971 if (code_buffer() == NULL) { 972 Matcher::soft_match_failure(); 973 } else { 974 if (PrintAssembly && (WizardMode || Verbose)) 975 tty->print_cr("### Stub::%s", stub_name); 976 977 if (!failing()) { 978 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 979 980 // Make the NMethod 981 // For now we mark the frame as never safe for profile stackwalking 982 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 983 code_buffer(), 984 CodeOffsets::frame_never_safe, 985 // _code_offsets.value(CodeOffsets::Frame_Complete), 986 frame_size_in_words(), 987 _oop_map_set, 988 save_arg_registers); 989 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 990 991 _stub_entry_point = rs->entry_point(); 992 } 993 } 994 } 995 996 //------------------------------Init------------------------------------------- 997 // Prepare for a single compilation 998 void Compile::Init(int aliaslevel) { 999 _unique = 0; 1000 _regalloc = NULL; 1001 1002 _tf = NULL; // filled in later 1003 _top = NULL; // cached later 1004 _matcher = NULL; // filled in later 1005 _cfg = NULL; // filled in later 1006 1007 set_24_bit_selection_and_mode(Use24BitFP, false); 1008 1009 _node_note_array = NULL; 1010 _default_node_notes = NULL; 1011 1012 _immutable_memory = NULL; // filled in at first inquiry 1013 1014 // Globally visible Nodes 1015 // First set TOP to NULL to give safe behavior during creation of RootNode 1016 set_cached_top_node(NULL); 1017 set_root(new (this) RootNode()); 1018 // Now that you have a Root to point to, create the real TOP 1019 set_cached_top_node( new (this) ConNode(Type::TOP) ); 1020 set_recent_alloc(NULL, NULL); 1021 1022 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1023 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1024 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1025 env()->set_dependencies(new Dependencies(env())); 1026 1027 _fixed_slots = 0; 1028 set_has_split_ifs(false); 1029 set_has_loops(has_method() && method()->has_loops()); // first approximation 1030 set_has_stringbuilder(false); 1031 set_has_boxed_value(false); 1032 _trap_can_recompile = false; // no traps emitted yet 1033 _major_progress = true; // start out assuming good things will happen 1034 set_has_unsafe_access(false); 1035 set_max_vector_size(0); 1036 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1037 set_decompile_count(0); 1038 1039 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1040 set_num_loop_opts(LoopOptsCount); 1041 set_do_inlining(Inline); 1042 set_max_inline_size(MaxInlineSize); 1043 set_freq_inline_size(FreqInlineSize); 1044 set_do_scheduling(OptoScheduling); 1045 set_do_count_invocations(false); 1046 set_do_method_data_update(false); 1047 1048 if (debug_info()->recording_non_safepoints()) { 1049 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1050 (comp_arena(), 8, 0, NULL)); 1051 set_default_node_notes(Node_Notes::make(this)); 1052 } 1053 1054 // // -- Initialize types before each compile -- 1055 // // Update cached type information 1056 // if( _method && _method->constants() ) 1057 // Type::update_loaded_types(_method, _method->constants()); 1058 1059 // Init alias_type map. 1060 if (!_do_escape_analysis && aliaslevel == 3) 1061 aliaslevel = 2; // No unique types without escape analysis 1062 _AliasLevel = aliaslevel; 1063 const int grow_ats = 16; 1064 _max_alias_types = grow_ats; 1065 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1066 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1067 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1068 { 1069 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1070 } 1071 // Initialize the first few types. 1072 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1073 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1074 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1075 _num_alias_types = AliasIdxRaw+1; 1076 // Zero out the alias type cache. 1077 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1078 // A NULL adr_type hits in the cache right away. Preload the right answer. 1079 probe_alias_cache(NULL)->_index = AliasIdxTop; 1080 1081 _intrinsics = NULL; 1082 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1083 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1084 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1085 register_library_intrinsics(); 1086 } 1087 1088 //---------------------------init_start---------------------------------------- 1089 // Install the StartNode on this compile object. 1090 void Compile::init_start(StartNode* s) { 1091 if (failing()) 1092 return; // already failing 1093 assert(s == start(), ""); 1094 } 1095 1096 StartNode* Compile::start() const { 1097 assert(!failing(), ""); 1098 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1099 Node* start = root()->fast_out(i); 1100 if( start->is_Start() ) 1101 return start->as_Start(); 1102 } 1103 ShouldNotReachHere(); 1104 return NULL; 1105 } 1106 1107 //-------------------------------immutable_memory------------------------------------- 1108 // Access immutable memory 1109 Node* Compile::immutable_memory() { 1110 if (_immutable_memory != NULL) { 1111 return _immutable_memory; 1112 } 1113 StartNode* s = start(); 1114 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1115 Node *p = s->fast_out(i); 1116 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1117 _immutable_memory = p; 1118 return _immutable_memory; 1119 } 1120 } 1121 ShouldNotReachHere(); 1122 return NULL; 1123 } 1124 1125 //----------------------set_cached_top_node------------------------------------ 1126 // Install the cached top node, and make sure Node::is_top works correctly. 1127 void Compile::set_cached_top_node(Node* tn) { 1128 if (tn != NULL) verify_top(tn); 1129 Node* old_top = _top; 1130 _top = tn; 1131 // Calling Node::setup_is_top allows the nodes the chance to adjust 1132 // their _out arrays. 1133 if (_top != NULL) _top->setup_is_top(); 1134 if (old_top != NULL) old_top->setup_is_top(); 1135 assert(_top == NULL || top()->is_top(), ""); 1136 } 1137 1138 #ifdef ASSERT 1139 uint Compile::count_live_nodes_by_graph_walk() { 1140 Unique_Node_List useful(comp_arena()); 1141 // Get useful node list by walking the graph. 1142 identify_useful_nodes(useful); 1143 return useful.size(); 1144 } 1145 1146 void Compile::print_missing_nodes() { 1147 1148 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1149 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1150 return; 1151 } 1152 1153 // This is an expensive function. It is executed only when the user 1154 // specifies VerifyIdealNodeCount option or otherwise knows the 1155 // additional work that needs to be done to identify reachable nodes 1156 // by walking the flow graph and find the missing ones using 1157 // _dead_node_list. 1158 1159 Unique_Node_List useful(comp_arena()); 1160 // Get useful node list by walking the graph. 1161 identify_useful_nodes(useful); 1162 1163 uint l_nodes = C->live_nodes(); 1164 uint l_nodes_by_walk = useful.size(); 1165 1166 if (l_nodes != l_nodes_by_walk) { 1167 if (_log != NULL) { 1168 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1169 _log->stamp(); 1170 _log->end_head(); 1171 } 1172 VectorSet& useful_member_set = useful.member_set(); 1173 int last_idx = l_nodes_by_walk; 1174 for (int i = 0; i < last_idx; i++) { 1175 if (useful_member_set.test(i)) { 1176 if (_dead_node_list.test(i)) { 1177 if (_log != NULL) { 1178 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1179 } 1180 if (PrintIdealNodeCount) { 1181 // Print the log message to tty 1182 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1183 useful.at(i)->dump(); 1184 } 1185 } 1186 } 1187 else if (! _dead_node_list.test(i)) { 1188 if (_log != NULL) { 1189 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1190 } 1191 if (PrintIdealNodeCount) { 1192 // Print the log message to tty 1193 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1194 } 1195 } 1196 } 1197 if (_log != NULL) { 1198 _log->tail("mismatched_nodes"); 1199 } 1200 } 1201 } 1202 #endif 1203 1204 #ifndef PRODUCT 1205 void Compile::verify_top(Node* tn) const { 1206 if (tn != NULL) { 1207 assert(tn->is_Con(), "top node must be a constant"); 1208 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1209 assert(tn->in(0) != NULL, "must have live top node"); 1210 } 1211 } 1212 #endif 1213 1214 1215 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1216 1217 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1218 guarantee(arr != NULL, ""); 1219 int num_blocks = arr->length(); 1220 if (grow_by < num_blocks) grow_by = num_blocks; 1221 int num_notes = grow_by * _node_notes_block_size; 1222 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1223 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1224 while (num_notes > 0) { 1225 arr->append(notes); 1226 notes += _node_notes_block_size; 1227 num_notes -= _node_notes_block_size; 1228 } 1229 assert(num_notes == 0, "exact multiple, please"); 1230 } 1231 1232 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1233 if (source == NULL || dest == NULL) return false; 1234 1235 if (dest->is_Con()) 1236 return false; // Do not push debug info onto constants. 1237 1238 #ifdef ASSERT 1239 // Leave a bread crumb trail pointing to the original node: 1240 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1241 dest->set_debug_orig(source); 1242 } 1243 #endif 1244 1245 if (node_note_array() == NULL) 1246 return false; // Not collecting any notes now. 1247 1248 // This is a copy onto a pre-existing node, which may already have notes. 1249 // If both nodes have notes, do not overwrite any pre-existing notes. 1250 Node_Notes* source_notes = node_notes_at(source->_idx); 1251 if (source_notes == NULL || source_notes->is_clear()) return false; 1252 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1253 if (dest_notes == NULL || dest_notes->is_clear()) { 1254 return set_node_notes_at(dest->_idx, source_notes); 1255 } 1256 1257 Node_Notes merged_notes = (*source_notes); 1258 // The order of operations here ensures that dest notes will win... 1259 merged_notes.update_from(dest_notes); 1260 return set_node_notes_at(dest->_idx, &merged_notes); 1261 } 1262 1263 1264 //--------------------------allow_range_check_smearing------------------------- 1265 // Gating condition for coalescing similar range checks. 1266 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1267 // single covering check that is at least as strong as any of them. 1268 // If the optimization succeeds, the simplified (strengthened) range check 1269 // will always succeed. If it fails, we will deopt, and then give up 1270 // on the optimization. 1271 bool Compile::allow_range_check_smearing() const { 1272 // If this method has already thrown a range-check, 1273 // assume it was because we already tried range smearing 1274 // and it failed. 1275 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1276 return !already_trapped; 1277 } 1278 1279 1280 //------------------------------flatten_alias_type----------------------------- 1281 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1282 int offset = tj->offset(); 1283 TypePtr::PTR ptr = tj->ptr(); 1284 1285 // Known instance (scalarizable allocation) alias only with itself. 1286 bool is_known_inst = tj->isa_oopptr() != NULL && 1287 tj->is_oopptr()->is_known_instance(); 1288 1289 // Process weird unsafe references. 1290 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1291 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1292 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1293 tj = TypeOopPtr::BOTTOM; 1294 ptr = tj->ptr(); 1295 offset = tj->offset(); 1296 } 1297 1298 // Array pointers need some flattening 1299 const TypeAryPtr *ta = tj->isa_aryptr(); 1300 if (ta && ta->is_stable()) { 1301 // Erase stability property for alias analysis. 1302 tj = ta = ta->cast_to_stable(false); 1303 } 1304 if( ta && is_known_inst ) { 1305 if ( offset != Type::OffsetBot && 1306 offset > arrayOopDesc::length_offset_in_bytes() ) { 1307 offset = Type::OffsetBot; // Flatten constant access into array body only 1308 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1309 } 1310 } else if( ta && _AliasLevel >= 2 ) { 1311 // For arrays indexed by constant indices, we flatten the alias 1312 // space to include all of the array body. Only the header, klass 1313 // and array length can be accessed un-aliased. 1314 if( offset != Type::OffsetBot ) { 1315 if( ta->const_oop() ) { // MethodData* or Method* 1316 offset = Type::OffsetBot; // Flatten constant access into array body 1317 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1318 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1319 // range is OK as-is. 1320 tj = ta = TypeAryPtr::RANGE; 1321 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1322 tj = TypeInstPtr::KLASS; // all klass loads look alike 1323 ta = TypeAryPtr::RANGE; // generic ignored junk 1324 ptr = TypePtr::BotPTR; 1325 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1326 tj = TypeInstPtr::MARK; 1327 ta = TypeAryPtr::RANGE; // generic ignored junk 1328 ptr = TypePtr::BotPTR; 1329 } else { // Random constant offset into array body 1330 offset = Type::OffsetBot; // Flatten constant access into array body 1331 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1332 } 1333 } 1334 // Arrays of fixed size alias with arrays of unknown size. 1335 if (ta->size() != TypeInt::POS) { 1336 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1337 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1338 } 1339 // Arrays of known objects become arrays of unknown objects. 1340 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1341 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1342 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1343 } 1344 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1345 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1346 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1347 } 1348 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1349 // cannot be distinguished by bytecode alone. 1350 if (ta->elem() == TypeInt::BOOL) { 1351 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1352 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1353 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1354 } 1355 // During the 2nd round of IterGVN, NotNull castings are removed. 1356 // Make sure the Bottom and NotNull variants alias the same. 1357 // Also, make sure exact and non-exact variants alias the same. 1358 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1359 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1360 } 1361 } 1362 1363 // Oop pointers need some flattening 1364 const TypeInstPtr *to = tj->isa_instptr(); 1365 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1366 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1367 if( ptr == TypePtr::Constant ) { 1368 if (to->klass() != ciEnv::current()->Class_klass() || 1369 offset < k->size_helper() * wordSize) { 1370 // No constant oop pointers (such as Strings); they alias with 1371 // unknown strings. 1372 assert(!is_known_inst, "not scalarizable allocation"); 1373 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1374 } 1375 } else if( is_known_inst ) { 1376 tj = to; // Keep NotNull and klass_is_exact for instance type 1377 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1378 // During the 2nd round of IterGVN, NotNull castings are removed. 1379 // Make sure the Bottom and NotNull variants alias the same. 1380 // Also, make sure exact and non-exact variants alias the same. 1381 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1382 } 1383 // Canonicalize the holder of this field 1384 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1385 // First handle header references such as a LoadKlassNode, even if the 1386 // object's klass is unloaded at compile time (4965979). 1387 if (!is_known_inst) { // Do it only for non-instance types 1388 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1389 } 1390 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1391 // Static fields are in the space above the normal instance 1392 // fields in the java.lang.Class instance. 1393 if (to->klass() != ciEnv::current()->Class_klass()) { 1394 to = NULL; 1395 tj = TypeOopPtr::BOTTOM; 1396 offset = tj->offset(); 1397 } 1398 } else { 1399 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1400 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1401 if( is_known_inst ) { 1402 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1403 } else { 1404 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1405 } 1406 } 1407 } 1408 } 1409 1410 // Klass pointers to object array klasses need some flattening 1411 const TypeKlassPtr *tk = tj->isa_klassptr(); 1412 if( tk ) { 1413 // If we are referencing a field within a Klass, we need 1414 // to assume the worst case of an Object. Both exact and 1415 // inexact types must flatten to the same alias class so 1416 // use NotNull as the PTR. 1417 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1418 1419 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1420 TypeKlassPtr::OBJECT->klass(), 1421 offset); 1422 } 1423 1424 ciKlass* klass = tk->klass(); 1425 if( klass->is_obj_array_klass() ) { 1426 ciKlass* k = TypeAryPtr::OOPS->klass(); 1427 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1428 k = TypeInstPtr::BOTTOM->klass(); 1429 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1430 } 1431 1432 // Check for precise loads from the primary supertype array and force them 1433 // to the supertype cache alias index. Check for generic array loads from 1434 // the primary supertype array and also force them to the supertype cache 1435 // alias index. Since the same load can reach both, we need to merge 1436 // these 2 disparate memories into the same alias class. Since the 1437 // primary supertype array is read-only, there's no chance of confusion 1438 // where we bypass an array load and an array store. 1439 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1440 if (offset == Type::OffsetBot || 1441 (offset >= primary_supers_offset && 1442 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1443 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1444 offset = in_bytes(Klass::secondary_super_cache_offset()); 1445 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1446 } 1447 } 1448 1449 // Flatten all Raw pointers together. 1450 if (tj->base() == Type::RawPtr) 1451 tj = TypeRawPtr::BOTTOM; 1452 1453 if (tj->base() == Type::AnyPtr) 1454 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1455 1456 // Flatten all to bottom for now 1457 switch( _AliasLevel ) { 1458 case 0: 1459 tj = TypePtr::BOTTOM; 1460 break; 1461 case 1: // Flatten to: oop, static, field or array 1462 switch (tj->base()) { 1463 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1464 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1465 case Type::AryPtr: // do not distinguish arrays at all 1466 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1467 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1468 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1469 default: ShouldNotReachHere(); 1470 } 1471 break; 1472 case 2: // No collapsing at level 2; keep all splits 1473 case 3: // No collapsing at level 3; keep all splits 1474 break; 1475 default: 1476 Unimplemented(); 1477 } 1478 1479 offset = tj->offset(); 1480 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1481 1482 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1483 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1484 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1485 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1486 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1487 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1488 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1489 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1490 assert( tj->ptr() != TypePtr::TopPTR && 1491 tj->ptr() != TypePtr::AnyNull && 1492 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1493 // assert( tj->ptr() != TypePtr::Constant || 1494 // tj->base() == Type::RawPtr || 1495 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1496 1497 return tj; 1498 } 1499 1500 void Compile::AliasType::Init(int i, const TypePtr* at) { 1501 _index = i; 1502 _adr_type = at; 1503 _field = NULL; 1504 _element = NULL; 1505 _is_rewritable = true; // default 1506 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1507 if (atoop != NULL && atoop->is_known_instance()) { 1508 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1509 _general_index = Compile::current()->get_alias_index(gt); 1510 } else { 1511 _general_index = 0; 1512 } 1513 } 1514 1515 //---------------------------------print_on------------------------------------ 1516 #ifndef PRODUCT 1517 void Compile::AliasType::print_on(outputStream* st) { 1518 if (index() < 10) 1519 st->print("@ <%d> ", index()); 1520 else st->print("@ <%d>", index()); 1521 st->print(is_rewritable() ? " " : " RO"); 1522 int offset = adr_type()->offset(); 1523 if (offset == Type::OffsetBot) 1524 st->print(" +any"); 1525 else st->print(" +%-3d", offset); 1526 st->print(" in "); 1527 adr_type()->dump_on(st); 1528 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1529 if (field() != NULL && tjp) { 1530 if (tjp->klass() != field()->holder() || 1531 tjp->offset() != field()->offset_in_bytes()) { 1532 st->print(" != "); 1533 field()->print(); 1534 st->print(" ***"); 1535 } 1536 } 1537 } 1538 1539 void print_alias_types() { 1540 Compile* C = Compile::current(); 1541 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1542 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1543 C->alias_type(idx)->print_on(tty); 1544 tty->cr(); 1545 } 1546 } 1547 #endif 1548 1549 1550 //----------------------------probe_alias_cache-------------------------------- 1551 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1552 intptr_t key = (intptr_t) adr_type; 1553 key ^= key >> logAliasCacheSize; 1554 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1555 } 1556 1557 1558 //-----------------------------grow_alias_types-------------------------------- 1559 void Compile::grow_alias_types() { 1560 const int old_ats = _max_alias_types; // how many before? 1561 const int new_ats = old_ats; // how many more? 1562 const int grow_ats = old_ats+new_ats; // how many now? 1563 _max_alias_types = grow_ats; 1564 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1565 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1566 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1567 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1568 } 1569 1570 1571 //--------------------------------find_alias_type------------------------------ 1572 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1573 if (_AliasLevel == 0) 1574 return alias_type(AliasIdxBot); 1575 1576 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1577 if (ace->_adr_type == adr_type) { 1578 return alias_type(ace->_index); 1579 } 1580 1581 // Handle special cases. 1582 if (adr_type == NULL) return alias_type(AliasIdxTop); 1583 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1584 1585 // Do it the slow way. 1586 const TypePtr* flat = flatten_alias_type(adr_type); 1587 1588 #ifdef ASSERT 1589 assert(flat == flatten_alias_type(flat), "idempotent"); 1590 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1591 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1592 const TypeOopPtr* foop = flat->is_oopptr(); 1593 // Scalarizable allocations have exact klass always. 1594 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1595 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1596 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1597 } 1598 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1599 #endif 1600 1601 int idx = AliasIdxTop; 1602 for (int i = 0; i < num_alias_types(); i++) { 1603 if (alias_type(i)->adr_type() == flat) { 1604 idx = i; 1605 break; 1606 } 1607 } 1608 1609 if (idx == AliasIdxTop) { 1610 if (no_create) return NULL; 1611 // Grow the array if necessary. 1612 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1613 // Add a new alias type. 1614 idx = _num_alias_types++; 1615 _alias_types[idx]->Init(idx, flat); 1616 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1617 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1618 if (flat->isa_instptr()) { 1619 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1620 && flat->is_instptr()->klass() == env()->Class_klass()) 1621 alias_type(idx)->set_rewritable(false); 1622 } 1623 if (flat->isa_aryptr()) { 1624 #ifdef ASSERT 1625 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1626 // (T_BYTE has the weakest alignment and size restrictions...) 1627 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1628 #endif 1629 if (flat->offset() == TypePtr::OffsetBot) 1630 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1631 } 1632 if (flat->isa_klassptr()) { 1633 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1634 alias_type(idx)->set_rewritable(false); 1635 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1636 alias_type(idx)->set_rewritable(false); 1637 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1638 alias_type(idx)->set_rewritable(false); 1639 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1640 alias_type(idx)->set_rewritable(false); 1641 } 1642 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1643 // but the base pointer type is not distinctive enough to identify 1644 // references into JavaThread.) 1645 1646 // Check for final fields. 1647 const TypeInstPtr* tinst = flat->isa_instptr(); 1648 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1649 ciField* field; 1650 if (tinst->const_oop() != NULL && 1651 tinst->klass() == ciEnv::current()->Class_klass() && 1652 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1653 // static field 1654 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1655 field = k->get_field_by_offset(tinst->offset(), true); 1656 } else { 1657 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1658 field = k->get_field_by_offset(tinst->offset(), false); 1659 } 1660 assert(field == NULL || 1661 original_field == NULL || 1662 (field->holder() == original_field->holder() && 1663 field->offset() == original_field->offset() && 1664 field->is_static() == original_field->is_static()), "wrong field?"); 1665 // Set field() and is_rewritable() attributes. 1666 if (field != NULL) alias_type(idx)->set_field(field); 1667 } 1668 } 1669 1670 // Fill the cache for next time. 1671 ace->_adr_type = adr_type; 1672 ace->_index = idx; 1673 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1674 1675 // Might as well try to fill the cache for the flattened version, too. 1676 AliasCacheEntry* face = probe_alias_cache(flat); 1677 if (face->_adr_type == NULL) { 1678 face->_adr_type = flat; 1679 face->_index = idx; 1680 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1681 } 1682 1683 return alias_type(idx); 1684 } 1685 1686 1687 Compile::AliasType* Compile::alias_type(ciField* field) { 1688 const TypeOopPtr* t; 1689 if (field->is_static()) 1690 t = TypeInstPtr::make(field->holder()->java_mirror()); 1691 else 1692 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1693 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1694 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1695 return atp; 1696 } 1697 1698 1699 //------------------------------have_alias_type-------------------------------- 1700 bool Compile::have_alias_type(const TypePtr* adr_type) { 1701 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1702 if (ace->_adr_type == adr_type) { 1703 return true; 1704 } 1705 1706 // Handle special cases. 1707 if (adr_type == NULL) return true; 1708 if (adr_type == TypePtr::BOTTOM) return true; 1709 1710 return find_alias_type(adr_type, true, NULL) != NULL; 1711 } 1712 1713 //-----------------------------must_alias-------------------------------------- 1714 // True if all values of the given address type are in the given alias category. 1715 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1716 if (alias_idx == AliasIdxBot) return true; // the universal category 1717 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1718 if (alias_idx == AliasIdxTop) return false; // the empty category 1719 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1720 1721 // the only remaining possible overlap is identity 1722 int adr_idx = get_alias_index(adr_type); 1723 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1724 assert(adr_idx == alias_idx || 1725 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1726 && adr_type != TypeOopPtr::BOTTOM), 1727 "should not be testing for overlap with an unsafe pointer"); 1728 return adr_idx == alias_idx; 1729 } 1730 1731 //------------------------------can_alias-------------------------------------- 1732 // True if any values of the given address type are in the given alias category. 1733 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1734 if (alias_idx == AliasIdxTop) return false; // the empty category 1735 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1736 if (alias_idx == AliasIdxBot) return true; // the universal category 1737 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1738 1739 // the only remaining possible overlap is identity 1740 int adr_idx = get_alias_index(adr_type); 1741 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1742 return adr_idx == alias_idx; 1743 } 1744 1745 1746 1747 //---------------------------pop_warm_call------------------------------------- 1748 WarmCallInfo* Compile::pop_warm_call() { 1749 WarmCallInfo* wci = _warm_calls; 1750 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1751 return wci; 1752 } 1753 1754 //----------------------------Inline_Warm-------------------------------------- 1755 int Compile::Inline_Warm() { 1756 // If there is room, try to inline some more warm call sites. 1757 // %%% Do a graph index compaction pass when we think we're out of space? 1758 if (!InlineWarmCalls) return 0; 1759 1760 int calls_made_hot = 0; 1761 int room_to_grow = NodeCountInliningCutoff - unique(); 1762 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1763 int amount_grown = 0; 1764 WarmCallInfo* call; 1765 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1766 int est_size = (int)call->size(); 1767 if (est_size > (room_to_grow - amount_grown)) { 1768 // This one won't fit anyway. Get rid of it. 1769 call->make_cold(); 1770 continue; 1771 } 1772 call->make_hot(); 1773 calls_made_hot++; 1774 amount_grown += est_size; 1775 amount_to_grow -= est_size; 1776 } 1777 1778 if (calls_made_hot > 0) set_major_progress(); 1779 return calls_made_hot; 1780 } 1781 1782 1783 //----------------------------Finish_Warm-------------------------------------- 1784 void Compile::Finish_Warm() { 1785 if (!InlineWarmCalls) return; 1786 if (failing()) return; 1787 if (warm_calls() == NULL) return; 1788 1789 // Clean up loose ends, if we are out of space for inlining. 1790 WarmCallInfo* call; 1791 while ((call = pop_warm_call()) != NULL) { 1792 call->make_cold(); 1793 } 1794 } 1795 1796 //---------------------cleanup_loop_predicates----------------------- 1797 // Remove the opaque nodes that protect the predicates so that all unused 1798 // checks and uncommon_traps will be eliminated from the ideal graph 1799 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1800 if (predicate_count()==0) return; 1801 for (int i = predicate_count(); i > 0; i--) { 1802 Node * n = predicate_opaque1_node(i-1); 1803 assert(n->Opcode() == Op_Opaque1, "must be"); 1804 igvn.replace_node(n, n->in(1)); 1805 } 1806 assert(predicate_count()==0, "should be clean!"); 1807 } 1808 1809 // StringOpts and late inlining of string methods 1810 void Compile::inline_string_calls(bool parse_time) { 1811 { 1812 // remove useless nodes to make the usage analysis simpler 1813 ResourceMark rm; 1814 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1815 } 1816 1817 { 1818 ResourceMark rm; 1819 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1820 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1821 print_method(PHASE_AFTER_STRINGOPTS, 3); 1822 } 1823 1824 // now inline anything that we skipped the first time around 1825 if (!parse_time) { 1826 _late_inlines_pos = _late_inlines.length(); 1827 } 1828 1829 while (_string_late_inlines.length() > 0) { 1830 CallGenerator* cg = _string_late_inlines.pop(); 1831 cg->do_late_inline(); 1832 if (failing()) return; 1833 } 1834 _string_late_inlines.trunc_to(0); 1835 } 1836 1837 // Late inlining of boxing methods 1838 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1839 if (_boxing_late_inlines.length() > 0) { 1840 assert(has_boxed_value(), "inconsistent"); 1841 1842 PhaseGVN* gvn = initial_gvn(); 1843 set_inlining_incrementally(true); 1844 1845 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1846 for_igvn()->clear(); 1847 gvn->replace_with(&igvn); 1848 1849 while (_boxing_late_inlines.length() > 0) { 1850 CallGenerator* cg = _boxing_late_inlines.pop(); 1851 cg->do_late_inline(); 1852 if (failing()) return; 1853 } 1854 _boxing_late_inlines.trunc_to(0); 1855 1856 { 1857 ResourceMark rm; 1858 PhaseRemoveUseless pru(gvn, for_igvn()); 1859 } 1860 1861 igvn = PhaseIterGVN(gvn); 1862 igvn.optimize(); 1863 1864 set_inlining_progress(false); 1865 set_inlining_incrementally(false); 1866 } 1867 } 1868 1869 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1870 assert(IncrementalInline, "incremental inlining should be on"); 1871 PhaseGVN* gvn = initial_gvn(); 1872 1873 set_inlining_progress(false); 1874 for_igvn()->clear(); 1875 gvn->replace_with(&igvn); 1876 1877 int i = 0; 1878 1879 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1880 CallGenerator* cg = _late_inlines.at(i); 1881 _late_inlines_pos = i+1; 1882 cg->do_late_inline(); 1883 if (failing()) return; 1884 } 1885 int j = 0; 1886 for (; i < _late_inlines.length(); i++, j++) { 1887 _late_inlines.at_put(j, _late_inlines.at(i)); 1888 } 1889 _late_inlines.trunc_to(j); 1890 1891 { 1892 ResourceMark rm; 1893 PhaseRemoveUseless pru(gvn, for_igvn()); 1894 } 1895 1896 igvn = PhaseIterGVN(gvn); 1897 } 1898 1899 // Perform incremental inlining until bound on number of live nodes is reached 1900 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1901 PhaseGVN* gvn = initial_gvn(); 1902 1903 set_inlining_incrementally(true); 1904 set_inlining_progress(true); 1905 uint low_live_nodes = 0; 1906 1907 while(inlining_progress() && _late_inlines.length() > 0) { 1908 1909 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1910 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1911 // PhaseIdealLoop is expensive so we only try it once we are 1912 // out of loop and we only try it again if the previous helped 1913 // got the number of nodes down significantly 1914 PhaseIdealLoop ideal_loop( igvn, false, true ); 1915 if (failing()) return; 1916 low_live_nodes = live_nodes(); 1917 _major_progress = true; 1918 } 1919 1920 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1921 break; 1922 } 1923 } 1924 1925 inline_incrementally_one(igvn); 1926 1927 if (failing()) return; 1928 1929 igvn.optimize(); 1930 1931 if (failing()) return; 1932 } 1933 1934 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1935 1936 if (_string_late_inlines.length() > 0) { 1937 assert(has_stringbuilder(), "inconsistent"); 1938 for_igvn()->clear(); 1939 initial_gvn()->replace_with(&igvn); 1940 1941 inline_string_calls(false); 1942 1943 if (failing()) return; 1944 1945 { 1946 ResourceMark rm; 1947 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1948 } 1949 1950 igvn = PhaseIterGVN(gvn); 1951 1952 igvn.optimize(); 1953 } 1954 1955 set_inlining_incrementally(false); 1956 } 1957 1958 1959 //------------------------------Optimize--------------------------------------- 1960 // Given a graph, optimize it. 1961 void Compile::Optimize() { 1962 TracePhase t1("optimizer", &_t_optimizer, true); 1963 1964 #ifndef PRODUCT 1965 if (env()->break_at_compile()) { 1966 BREAKPOINT; 1967 } 1968 1969 #endif 1970 1971 ResourceMark rm; 1972 int loop_opts_cnt; 1973 1974 NOT_PRODUCT( verify_graph_edges(); ) 1975 1976 print_method(PHASE_AFTER_PARSING); 1977 1978 { 1979 // Iterative Global Value Numbering, including ideal transforms 1980 // Initialize IterGVN with types and values from parse-time GVN 1981 PhaseIterGVN igvn(initial_gvn()); 1982 { 1983 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1984 igvn.optimize(); 1985 } 1986 1987 print_method(PHASE_ITER_GVN1, 2); 1988 1989 if (failing()) return; 1990 1991 { 1992 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 1993 inline_incrementally(igvn); 1994 } 1995 1996 print_method(PHASE_INCREMENTAL_INLINE, 2); 1997 1998 if (failing()) return; 1999 2000 if (eliminate_boxing()) { 2001 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2002 // Inline valueOf() methods now. 2003 inline_boxing_calls(igvn); 2004 2005 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2006 2007 if (failing()) return; 2008 } 2009 2010 // No more new expensive nodes will be added to the list from here 2011 // so keep only the actual candidates for optimizations. 2012 cleanup_expensive_nodes(igvn); 2013 2014 // Perform escape analysis 2015 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2016 if (has_loops()) { 2017 // Cleanup graph (remove dead nodes). 2018 TracePhase t2("idealLoop", &_t_idealLoop, true); 2019 PhaseIdealLoop ideal_loop( igvn, false, true ); 2020 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2021 if (failing()) return; 2022 } 2023 ConnectionGraph::do_analysis(this, &igvn); 2024 2025 if (failing()) return; 2026 2027 // Optimize out fields loads from scalar replaceable allocations. 2028 igvn.optimize(); 2029 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2030 2031 if (failing()) return; 2032 2033 if (congraph() != NULL && macro_count() > 0) { 2034 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2035 PhaseMacroExpand mexp(igvn); 2036 mexp.eliminate_macro_nodes(); 2037 igvn.set_delay_transform(false); 2038 2039 igvn.optimize(); 2040 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2041 2042 if (failing()) return; 2043 } 2044 } 2045 2046 // Loop transforms on the ideal graph. Range Check Elimination, 2047 // peeling, unrolling, etc. 2048 2049 // Set loop opts counter 2050 loop_opts_cnt = num_loop_opts(); 2051 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2052 { 2053 TracePhase t2("idealLoop", &_t_idealLoop, true); 2054 PhaseIdealLoop ideal_loop( igvn, true ); 2055 loop_opts_cnt--; 2056 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2057 if (failing()) return; 2058 } 2059 // Loop opts pass if partial peeling occurred in previous pass 2060 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2061 TracePhase t3("idealLoop", &_t_idealLoop, true); 2062 PhaseIdealLoop ideal_loop( igvn, false ); 2063 loop_opts_cnt--; 2064 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2065 if (failing()) return; 2066 } 2067 // Loop opts pass for loop-unrolling before CCP 2068 if(major_progress() && (loop_opts_cnt > 0)) { 2069 TracePhase t4("idealLoop", &_t_idealLoop, true); 2070 PhaseIdealLoop ideal_loop( igvn, false ); 2071 loop_opts_cnt--; 2072 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2073 } 2074 if (!failing()) { 2075 // Verify that last round of loop opts produced a valid graph 2076 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2077 PhaseIdealLoop::verify(igvn); 2078 } 2079 } 2080 if (failing()) return; 2081 2082 // Conditional Constant Propagation; 2083 PhaseCCP ccp( &igvn ); 2084 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2085 { 2086 TracePhase t2("ccp", &_t_ccp, true); 2087 ccp.do_transform(); 2088 } 2089 print_method(PHASE_CPP1, 2); 2090 2091 assert( true, "Break here to ccp.dump_old2new_map()"); 2092 2093 // Iterative Global Value Numbering, including ideal transforms 2094 { 2095 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2096 igvn = ccp; 2097 igvn.optimize(); 2098 } 2099 2100 print_method(PHASE_ITER_GVN2, 2); 2101 2102 if (failing()) return; 2103 2104 // Loop transforms on the ideal graph. Range Check Elimination, 2105 // peeling, unrolling, etc. 2106 if(loop_opts_cnt > 0) { 2107 debug_only( int cnt = 0; ); 2108 while(major_progress() && (loop_opts_cnt > 0)) { 2109 TracePhase t2("idealLoop", &_t_idealLoop, true); 2110 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2111 PhaseIdealLoop ideal_loop( igvn, true); 2112 loop_opts_cnt--; 2113 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2114 if (failing()) return; 2115 } 2116 } 2117 2118 { 2119 // Verify that all previous optimizations produced a valid graph 2120 // at least to this point, even if no loop optimizations were done. 2121 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2122 PhaseIdealLoop::verify(igvn); 2123 } 2124 2125 { 2126 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2127 PhaseMacroExpand mex(igvn); 2128 if (mex.expand_macro_nodes()) { 2129 assert(failing(), "must bail out w/ explicit message"); 2130 return; 2131 } 2132 } 2133 2134 } // (End scope of igvn; run destructor if necessary for asserts.) 2135 2136 dump_inlining(); 2137 // A method with only infinite loops has no edges entering loops from root 2138 { 2139 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2140 if (final_graph_reshaping()) { 2141 assert(failing(), "must bail out w/ explicit message"); 2142 return; 2143 } 2144 } 2145 2146 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2147 } 2148 2149 2150 //------------------------------Code_Gen--------------------------------------- 2151 // Given a graph, generate code for it 2152 void Compile::Code_Gen() { 2153 if (failing()) return; 2154 2155 // Perform instruction selection. You might think we could reclaim Matcher 2156 // memory PDQ, but actually the Matcher is used in generating spill code. 2157 // Internals of the Matcher (including some VectorSets) must remain live 2158 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2159 // set a bit in reclaimed memory. 2160 2161 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2162 // nodes. Mapping is only valid at the root of each matched subtree. 2163 NOT_PRODUCT( verify_graph_edges(); ) 2164 2165 Node_List proj_list; 2166 Matcher m(proj_list); 2167 _matcher = &m; 2168 { 2169 TracePhase t2("matcher", &_t_matcher, true); 2170 m.match(); 2171 } 2172 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2173 // nodes. Mapping is only valid at the root of each matched subtree. 2174 NOT_PRODUCT( verify_graph_edges(); ) 2175 2176 // If you have too many nodes, or if matching has failed, bail out 2177 check_node_count(0, "out of nodes matching instructions"); 2178 if (failing()) return; 2179 2180 // Build a proper-looking CFG 2181 PhaseCFG cfg(node_arena(), root(), m); 2182 _cfg = &cfg; 2183 { 2184 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2185 cfg.Dominators(); 2186 if (failing()) return; 2187 2188 NOT_PRODUCT( verify_graph_edges(); ) 2189 2190 cfg.Estimate_Block_Frequency(); 2191 cfg.GlobalCodeMotion(m,unique(),proj_list); 2192 if (failing()) return; 2193 2194 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2195 2196 NOT_PRODUCT( verify_graph_edges(); ) 2197 2198 debug_only( cfg.verify(); ) 2199 } 2200 NOT_PRODUCT( verify_graph_edges(); ) 2201 2202 PhaseChaitin regalloc(unique(), cfg, m); 2203 _regalloc = ®alloc; 2204 { 2205 TracePhase t2("regalloc", &_t_registerAllocation, true); 2206 // Perform register allocation. After Chaitin, use-def chains are 2207 // no longer accurate (at spill code) and so must be ignored. 2208 // Node->LRG->reg mappings are still accurate. 2209 _regalloc->Register_Allocate(); 2210 2211 // Bail out if the allocator builds too many nodes 2212 if (failing()) { 2213 return; 2214 } 2215 } 2216 2217 // Prior to register allocation we kept empty basic blocks in case the 2218 // the allocator needed a place to spill. After register allocation we 2219 // are not adding any new instructions. If any basic block is empty, we 2220 // can now safely remove it. 2221 { 2222 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2223 cfg.remove_empty(); 2224 if (do_freq_based_layout()) { 2225 PhaseBlockLayout layout(cfg); 2226 } else { 2227 cfg.set_loop_alignment(); 2228 } 2229 cfg.fixup_flow(); 2230 } 2231 2232 // Apply peephole optimizations 2233 if( OptoPeephole ) { 2234 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2235 PhasePeephole peep( _regalloc, cfg); 2236 peep.do_transform(); 2237 } 2238 2239 // Convert Nodes to instruction bits in a buffer 2240 { 2241 // %%%% workspace merge brought two timers together for one job 2242 TracePhase t2a("output", &_t_output, true); 2243 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2244 Output(); 2245 } 2246 2247 print_method(PHASE_FINAL_CODE); 2248 2249 // He's dead, Jim. 2250 _cfg = (PhaseCFG*)0xdeadbeef; 2251 _regalloc = (PhaseChaitin*)0xdeadbeef; 2252 } 2253 2254 2255 //------------------------------dump_asm--------------------------------------- 2256 // Dump formatted assembly 2257 #ifndef PRODUCT 2258 void Compile::dump_asm(int *pcs, uint pc_limit) { 2259 bool cut_short = false; 2260 tty->print_cr("#"); 2261 tty->print("# "); _tf->dump(); tty->cr(); 2262 tty->print_cr("#"); 2263 2264 // For all blocks 2265 int pc = 0x0; // Program counter 2266 char starts_bundle = ' '; 2267 _regalloc->dump_frame(); 2268 2269 Node *n = NULL; 2270 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 2271 if (VMThread::should_terminate()) { cut_short = true; break; } 2272 Block *b = _cfg->_blocks[i]; 2273 if (b->is_connector() && !Verbose) continue; 2274 n = b->_nodes[0]; 2275 if (pcs && n->_idx < pc_limit) 2276 tty->print("%3.3x ", pcs[n->_idx]); 2277 else 2278 tty->print(" "); 2279 b->dump_head( &_cfg->_bbs ); 2280 if (b->is_connector()) { 2281 tty->print_cr(" # Empty connector block"); 2282 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2283 tty->print_cr(" # Block is sole successor of call"); 2284 } 2285 2286 // For all instructions 2287 Node *delay = NULL; 2288 for( uint j = 0; j<b->_nodes.size(); j++ ) { 2289 if (VMThread::should_terminate()) { cut_short = true; break; } 2290 n = b->_nodes[j]; 2291 if (valid_bundle_info(n)) { 2292 Bundle *bundle = node_bundling(n); 2293 if (bundle->used_in_unconditional_delay()) { 2294 delay = n; 2295 continue; 2296 } 2297 if (bundle->starts_bundle()) 2298 starts_bundle = '+'; 2299 } 2300 2301 if (WizardMode) n->dump(); 2302 2303 if( !n->is_Region() && // Dont print in the Assembly 2304 !n->is_Phi() && // a few noisely useless nodes 2305 !n->is_Proj() && 2306 !n->is_MachTemp() && 2307 !n->is_SafePointScalarObject() && 2308 !n->is_Catch() && // Would be nice to print exception table targets 2309 !n->is_MergeMem() && // Not very interesting 2310 !n->is_top() && // Debug info table constants 2311 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2312 ) { 2313 if (pcs && n->_idx < pc_limit) 2314 tty->print("%3.3x", pcs[n->_idx]); 2315 else 2316 tty->print(" "); 2317 tty->print(" %c ", starts_bundle); 2318 starts_bundle = ' '; 2319 tty->print("\t"); 2320 n->format(_regalloc, tty); 2321 tty->cr(); 2322 } 2323 2324 // If we have an instruction with a delay slot, and have seen a delay, 2325 // then back up and print it 2326 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2327 assert(delay != NULL, "no unconditional delay instruction"); 2328 if (WizardMode) delay->dump(); 2329 2330 if (node_bundling(delay)->starts_bundle()) 2331 starts_bundle = '+'; 2332 if (pcs && n->_idx < pc_limit) 2333 tty->print("%3.3x", pcs[n->_idx]); 2334 else 2335 tty->print(" "); 2336 tty->print(" %c ", starts_bundle); 2337 starts_bundle = ' '; 2338 tty->print("\t"); 2339 delay->format(_regalloc, tty); 2340 tty->print_cr(""); 2341 delay = NULL; 2342 } 2343 2344 // Dump the exception table as well 2345 if( n->is_Catch() && (Verbose || WizardMode) ) { 2346 // Print the exception table for this offset 2347 _handler_table.print_subtable_for(pc); 2348 } 2349 } 2350 2351 if (pcs && n->_idx < pc_limit) 2352 tty->print_cr("%3.3x", pcs[n->_idx]); 2353 else 2354 tty->print_cr(""); 2355 2356 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2357 2358 } // End of per-block dump 2359 tty->print_cr(""); 2360 2361 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2362 } 2363 #endif 2364 2365 //------------------------------Final_Reshape_Counts--------------------------- 2366 // This class defines counters to help identify when a method 2367 // may/must be executed using hardware with only 24-bit precision. 2368 struct Final_Reshape_Counts : public StackObj { 2369 int _call_count; // count non-inlined 'common' calls 2370 int _float_count; // count float ops requiring 24-bit precision 2371 int _double_count; // count double ops requiring more precision 2372 int _java_call_count; // count non-inlined 'java' calls 2373 int _inner_loop_count; // count loops which need alignment 2374 VectorSet _visited; // Visitation flags 2375 Node_List _tests; // Set of IfNodes & PCTableNodes 2376 2377 Final_Reshape_Counts() : 2378 _call_count(0), _float_count(0), _double_count(0), 2379 _java_call_count(0), _inner_loop_count(0), 2380 _visited( Thread::current()->resource_area() ) { } 2381 2382 void inc_call_count () { _call_count ++; } 2383 void inc_float_count () { _float_count ++; } 2384 void inc_double_count() { _double_count++; } 2385 void inc_java_call_count() { _java_call_count++; } 2386 void inc_inner_loop_count() { _inner_loop_count++; } 2387 2388 int get_call_count () const { return _call_count ; } 2389 int get_float_count () const { return _float_count ; } 2390 int get_double_count() const { return _double_count; } 2391 int get_java_call_count() const { return _java_call_count; } 2392 int get_inner_loop_count() const { return _inner_loop_count; } 2393 }; 2394 2395 #ifdef ASSERT 2396 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2397 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2398 // Make sure the offset goes inside the instance layout. 2399 return k->contains_field_offset(tp->offset()); 2400 // Note that OffsetBot and OffsetTop are very negative. 2401 } 2402 #endif 2403 2404 // Eliminate trivially redundant StoreCMs and accumulate their 2405 // precedence edges. 2406 void Compile::eliminate_redundant_card_marks(Node* n) { 2407 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2408 if (n->in(MemNode::Address)->outcnt() > 1) { 2409 // There are multiple users of the same address so it might be 2410 // possible to eliminate some of the StoreCMs 2411 Node* mem = n->in(MemNode::Memory); 2412 Node* adr = n->in(MemNode::Address); 2413 Node* val = n->in(MemNode::ValueIn); 2414 Node* prev = n; 2415 bool done = false; 2416 // Walk the chain of StoreCMs eliminating ones that match. As 2417 // long as it's a chain of single users then the optimization is 2418 // safe. Eliminating partially redundant StoreCMs would require 2419 // cloning copies down the other paths. 2420 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2421 if (adr == mem->in(MemNode::Address) && 2422 val == mem->in(MemNode::ValueIn)) { 2423 // redundant StoreCM 2424 if (mem->req() > MemNode::OopStore) { 2425 // Hasn't been processed by this code yet. 2426 n->add_prec(mem->in(MemNode::OopStore)); 2427 } else { 2428 // Already converted to precedence edge 2429 for (uint i = mem->req(); i < mem->len(); i++) { 2430 // Accumulate any precedence edges 2431 if (mem->in(i) != NULL) { 2432 n->add_prec(mem->in(i)); 2433 } 2434 } 2435 // Everything above this point has been processed. 2436 done = true; 2437 } 2438 // Eliminate the previous StoreCM 2439 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2440 assert(mem->outcnt() == 0, "should be dead"); 2441 mem->disconnect_inputs(NULL, this); 2442 } else { 2443 prev = mem; 2444 } 2445 mem = prev->in(MemNode::Memory); 2446 } 2447 } 2448 } 2449 2450 //------------------------------final_graph_reshaping_impl---------------------- 2451 // Implement items 1-5 from final_graph_reshaping below. 2452 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2453 2454 if ( n->outcnt() == 0 ) return; // dead node 2455 uint nop = n->Opcode(); 2456 2457 // Check for 2-input instruction with "last use" on right input. 2458 // Swap to left input. Implements item (2). 2459 if( n->req() == 3 && // two-input instruction 2460 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2461 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2462 n->in(2)->outcnt() == 1 &&// right use IS a last use 2463 !n->in(2)->is_Con() ) { // right use is not a constant 2464 // Check for commutative opcode 2465 switch( nop ) { 2466 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2467 case Op_MaxI: case Op_MinI: 2468 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2469 case Op_AndL: case Op_XorL: case Op_OrL: 2470 case Op_AndI: case Op_XorI: case Op_OrI: { 2471 // Move "last use" input to left by swapping inputs 2472 n->swap_edges(1, 2); 2473 break; 2474 } 2475 default: 2476 break; 2477 } 2478 } 2479 2480 #ifdef ASSERT 2481 if( n->is_Mem() ) { 2482 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2483 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2484 // oop will be recorded in oop map if load crosses safepoint 2485 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2486 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2487 "raw memory operations should have control edge"); 2488 } 2489 #endif 2490 // Count FPU ops and common calls, implements item (3) 2491 switch( nop ) { 2492 // Count all float operations that may use FPU 2493 case Op_AddF: 2494 case Op_SubF: 2495 case Op_MulF: 2496 case Op_DivF: 2497 case Op_NegF: 2498 case Op_ModF: 2499 case Op_ConvI2F: 2500 case Op_ConF: 2501 case Op_CmpF: 2502 case Op_CmpF3: 2503 // case Op_ConvL2F: // longs are split into 32-bit halves 2504 frc.inc_float_count(); 2505 break; 2506 2507 case Op_ConvF2D: 2508 case Op_ConvD2F: 2509 frc.inc_float_count(); 2510 frc.inc_double_count(); 2511 break; 2512 2513 // Count all double operations that may use FPU 2514 case Op_AddD: 2515 case Op_SubD: 2516 case Op_MulD: 2517 case Op_DivD: 2518 case Op_NegD: 2519 case Op_ModD: 2520 case Op_ConvI2D: 2521 case Op_ConvD2I: 2522 // case Op_ConvL2D: // handled by leaf call 2523 // case Op_ConvD2L: // handled by leaf call 2524 case Op_ConD: 2525 case Op_CmpD: 2526 case Op_CmpD3: 2527 frc.inc_double_count(); 2528 break; 2529 case Op_Opaque1: // Remove Opaque Nodes before matching 2530 case Op_Opaque2: // Remove Opaque Nodes before matching 2531 n->subsume_by(n->in(1), this); 2532 break; 2533 case Op_CallStaticJava: 2534 case Op_CallJava: 2535 case Op_CallDynamicJava: 2536 frc.inc_java_call_count(); // Count java call site; 2537 case Op_CallRuntime: 2538 case Op_CallLeaf: 2539 case Op_CallLeafNoFP: { 2540 assert( n->is_Call(), "" ); 2541 CallNode *call = n->as_Call(); 2542 // Count call sites where the FP mode bit would have to be flipped. 2543 // Do not count uncommon runtime calls: 2544 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2545 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2546 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2547 frc.inc_call_count(); // Count the call site 2548 } else { // See if uncommon argument is shared 2549 Node *n = call->in(TypeFunc::Parms); 2550 int nop = n->Opcode(); 2551 // Clone shared simple arguments to uncommon calls, item (1). 2552 if( n->outcnt() > 1 && 2553 !n->is_Proj() && 2554 nop != Op_CreateEx && 2555 nop != Op_CheckCastPP && 2556 nop != Op_DecodeN && 2557 nop != Op_DecodeNKlass && 2558 !n->is_Mem() ) { 2559 Node *x = n->clone(); 2560 call->set_req( TypeFunc::Parms, x ); 2561 } 2562 } 2563 break; 2564 } 2565 2566 case Op_StoreD: 2567 case Op_LoadD: 2568 case Op_LoadD_unaligned: 2569 frc.inc_double_count(); 2570 goto handle_mem; 2571 case Op_StoreF: 2572 case Op_LoadF: 2573 frc.inc_float_count(); 2574 goto handle_mem; 2575 2576 case Op_StoreCM: 2577 { 2578 // Convert OopStore dependence into precedence edge 2579 Node* prec = n->in(MemNode::OopStore); 2580 n->del_req(MemNode::OopStore); 2581 n->add_prec(prec); 2582 eliminate_redundant_card_marks(n); 2583 } 2584 2585 // fall through 2586 2587 case Op_StoreB: 2588 case Op_StoreC: 2589 case Op_StorePConditional: 2590 case Op_StoreI: 2591 case Op_StoreL: 2592 case Op_StoreIConditional: 2593 case Op_StoreLConditional: 2594 case Op_CompareAndSwapI: 2595 case Op_CompareAndSwapL: 2596 case Op_CompareAndSwapP: 2597 case Op_CompareAndSwapN: 2598 case Op_GetAndAddI: 2599 case Op_GetAndAddL: 2600 case Op_GetAndSetI: 2601 case Op_GetAndSetL: 2602 case Op_GetAndSetP: 2603 case Op_GetAndSetN: 2604 case Op_StoreP: 2605 case Op_StoreN: 2606 case Op_StoreNKlass: 2607 case Op_LoadB: 2608 case Op_LoadUB: 2609 case Op_LoadUS: 2610 case Op_LoadI: 2611 case Op_LoadKlass: 2612 case Op_LoadNKlass: 2613 case Op_LoadL: 2614 case Op_LoadL_unaligned: 2615 case Op_LoadPLocked: 2616 case Op_LoadP: 2617 case Op_LoadN: 2618 case Op_LoadRange: 2619 case Op_LoadS: { 2620 handle_mem: 2621 #ifdef ASSERT 2622 if( VerifyOptoOopOffsets ) { 2623 assert( n->is_Mem(), "" ); 2624 MemNode *mem = (MemNode*)n; 2625 // Check to see if address types have grounded out somehow. 2626 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2627 assert( !tp || oop_offset_is_sane(tp), "" ); 2628 } 2629 #endif 2630 break; 2631 } 2632 2633 case Op_AddP: { // Assert sane base pointers 2634 Node *addp = n->in(AddPNode::Address); 2635 assert( !addp->is_AddP() || 2636 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2637 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2638 "Base pointers must match" ); 2639 #ifdef _LP64 2640 if ((UseCompressedOops || UseCompressedKlassPointers) && 2641 addp->Opcode() == Op_ConP && 2642 addp == n->in(AddPNode::Base) && 2643 n->in(AddPNode::Offset)->is_Con()) { 2644 // Use addressing with narrow klass to load with offset on x86. 2645 // On sparc loading 32-bits constant and decoding it have less 2646 // instructions (4) then load 64-bits constant (7). 2647 // Do this transformation here since IGVN will convert ConN back to ConP. 2648 const Type* t = addp->bottom_type(); 2649 if (t->isa_oopptr() || t->isa_klassptr()) { 2650 Node* nn = NULL; 2651 2652 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2653 2654 // Look for existing ConN node of the same exact type. 2655 Node* r = root(); 2656 uint cnt = r->outcnt(); 2657 for (uint i = 0; i < cnt; i++) { 2658 Node* m = r->raw_out(i); 2659 if (m!= NULL && m->Opcode() == op && 2660 m->bottom_type()->make_ptr() == t) { 2661 nn = m; 2662 break; 2663 } 2664 } 2665 if (nn != NULL) { 2666 // Decode a narrow oop to match address 2667 // [R12 + narrow_oop_reg<<3 + offset] 2668 if (t->isa_oopptr()) { 2669 nn = new (this) DecodeNNode(nn, t); 2670 } else { 2671 nn = new (this) DecodeNKlassNode(nn, t); 2672 } 2673 n->set_req(AddPNode::Base, nn); 2674 n->set_req(AddPNode::Address, nn); 2675 if (addp->outcnt() == 0) { 2676 addp->disconnect_inputs(NULL, this); 2677 } 2678 } 2679 } 2680 } 2681 #endif 2682 break; 2683 } 2684 2685 #ifdef _LP64 2686 case Op_CastPP: 2687 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2688 Node* in1 = n->in(1); 2689 const Type* t = n->bottom_type(); 2690 Node* new_in1 = in1->clone(); 2691 new_in1->as_DecodeN()->set_type(t); 2692 2693 if (!Matcher::narrow_oop_use_complex_address()) { 2694 // 2695 // x86, ARM and friends can handle 2 adds in addressing mode 2696 // and Matcher can fold a DecodeN node into address by using 2697 // a narrow oop directly and do implicit NULL check in address: 2698 // 2699 // [R12 + narrow_oop_reg<<3 + offset] 2700 // NullCheck narrow_oop_reg 2701 // 2702 // On other platforms (Sparc) we have to keep new DecodeN node and 2703 // use it to do implicit NULL check in address: 2704 // 2705 // decode_not_null narrow_oop_reg, base_reg 2706 // [base_reg + offset] 2707 // NullCheck base_reg 2708 // 2709 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2710 // to keep the information to which NULL check the new DecodeN node 2711 // corresponds to use it as value in implicit_null_check(). 2712 // 2713 new_in1->set_req(0, n->in(0)); 2714 } 2715 2716 n->subsume_by(new_in1, this); 2717 if (in1->outcnt() == 0) { 2718 in1->disconnect_inputs(NULL, this); 2719 } 2720 } 2721 break; 2722 2723 case Op_CmpP: 2724 // Do this transformation here to preserve CmpPNode::sub() and 2725 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2726 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2727 Node* in1 = n->in(1); 2728 Node* in2 = n->in(2); 2729 if (!in1->is_DecodeNarrowPtr()) { 2730 in2 = in1; 2731 in1 = n->in(2); 2732 } 2733 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2734 2735 Node* new_in2 = NULL; 2736 if (in2->is_DecodeNarrowPtr()) { 2737 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2738 new_in2 = in2->in(1); 2739 } else if (in2->Opcode() == Op_ConP) { 2740 const Type* t = in2->bottom_type(); 2741 if (t == TypePtr::NULL_PTR) { 2742 assert(in1->is_DecodeN(), "compare klass to null?"); 2743 // Don't convert CmpP null check into CmpN if compressed 2744 // oops implicit null check is not generated. 2745 // This will allow to generate normal oop implicit null check. 2746 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2747 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2748 // 2749 // This transformation together with CastPP transformation above 2750 // will generated code for implicit NULL checks for compressed oops. 2751 // 2752 // The original code after Optimize() 2753 // 2754 // LoadN memory, narrow_oop_reg 2755 // decode narrow_oop_reg, base_reg 2756 // CmpP base_reg, NULL 2757 // CastPP base_reg // NotNull 2758 // Load [base_reg + offset], val_reg 2759 // 2760 // after these transformations will be 2761 // 2762 // LoadN memory, narrow_oop_reg 2763 // CmpN narrow_oop_reg, NULL 2764 // decode_not_null narrow_oop_reg, base_reg 2765 // Load [base_reg + offset], val_reg 2766 // 2767 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2768 // since narrow oops can be used in debug info now (see the code in 2769 // final_graph_reshaping_walk()). 2770 // 2771 // At the end the code will be matched to 2772 // on x86: 2773 // 2774 // Load_narrow_oop memory, narrow_oop_reg 2775 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2776 // NullCheck narrow_oop_reg 2777 // 2778 // and on sparc: 2779 // 2780 // Load_narrow_oop memory, narrow_oop_reg 2781 // decode_not_null narrow_oop_reg, base_reg 2782 // Load [base_reg + offset], val_reg 2783 // NullCheck base_reg 2784 // 2785 } else if (t->isa_oopptr()) { 2786 new_in2 = ConNode::make(this, t->make_narrowoop()); 2787 } else if (t->isa_klassptr()) { 2788 new_in2 = ConNode::make(this, t->make_narrowklass()); 2789 } 2790 } 2791 if (new_in2 != NULL) { 2792 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 2793 n->subsume_by(cmpN, this); 2794 if (in1->outcnt() == 0) { 2795 in1->disconnect_inputs(NULL, this); 2796 } 2797 if (in2->outcnt() == 0) { 2798 in2->disconnect_inputs(NULL, this); 2799 } 2800 } 2801 } 2802 break; 2803 2804 case Op_DecodeN: 2805 case Op_DecodeNKlass: 2806 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2807 // DecodeN could be pinned when it can't be fold into 2808 // an address expression, see the code for Op_CastPP above. 2809 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2810 break; 2811 2812 case Op_EncodeP: 2813 case Op_EncodePKlass: { 2814 Node* in1 = n->in(1); 2815 if (in1->is_DecodeNarrowPtr()) { 2816 n->subsume_by(in1->in(1), this); 2817 } else if (in1->Opcode() == Op_ConP) { 2818 const Type* t = in1->bottom_type(); 2819 if (t == TypePtr::NULL_PTR) { 2820 assert(t->isa_oopptr(), "null klass?"); 2821 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2822 } else if (t->isa_oopptr()) { 2823 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2824 } else if (t->isa_klassptr()) { 2825 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2826 } 2827 } 2828 if (in1->outcnt() == 0) { 2829 in1->disconnect_inputs(NULL, this); 2830 } 2831 break; 2832 } 2833 2834 case Op_Proj: { 2835 if (OptimizeStringConcat) { 2836 ProjNode* p = n->as_Proj(); 2837 if (p->_is_io_use) { 2838 // Separate projections were used for the exception path which 2839 // are normally removed by a late inline. If it wasn't inlined 2840 // then they will hang around and should just be replaced with 2841 // the original one. 2842 Node* proj = NULL; 2843 // Replace with just one 2844 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2845 Node *use = i.get(); 2846 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2847 proj = use; 2848 break; 2849 } 2850 } 2851 assert(proj != NULL, "must be found"); 2852 p->subsume_by(proj, this); 2853 } 2854 } 2855 break; 2856 } 2857 2858 case Op_Phi: 2859 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2860 // The EncodeP optimization may create Phi with the same edges 2861 // for all paths. It is not handled well by Register Allocator. 2862 Node* unique_in = n->in(1); 2863 assert(unique_in != NULL, ""); 2864 uint cnt = n->req(); 2865 for (uint i = 2; i < cnt; i++) { 2866 Node* m = n->in(i); 2867 assert(m != NULL, ""); 2868 if (unique_in != m) 2869 unique_in = NULL; 2870 } 2871 if (unique_in != NULL) { 2872 n->subsume_by(unique_in, this); 2873 } 2874 } 2875 break; 2876 2877 #endif 2878 2879 case Op_ModI: 2880 if (UseDivMod) { 2881 // Check if a%b and a/b both exist 2882 Node* d = n->find_similar(Op_DivI); 2883 if (d) { 2884 // Replace them with a fused divmod if supported 2885 if (Matcher::has_match_rule(Op_DivModI)) { 2886 DivModINode* divmod = DivModINode::make(this, n); 2887 d->subsume_by(divmod->div_proj(), this); 2888 n->subsume_by(divmod->mod_proj(), this); 2889 } else { 2890 // replace a%b with a-((a/b)*b) 2891 Node* mult = new (this) MulINode(d, d->in(2)); 2892 Node* sub = new (this) SubINode(d->in(1), mult); 2893 n->subsume_by(sub, this); 2894 } 2895 } 2896 } 2897 break; 2898 2899 case Op_ModL: 2900 if (UseDivMod) { 2901 // Check if a%b and a/b both exist 2902 Node* d = n->find_similar(Op_DivL); 2903 if (d) { 2904 // Replace them with a fused divmod if supported 2905 if (Matcher::has_match_rule(Op_DivModL)) { 2906 DivModLNode* divmod = DivModLNode::make(this, n); 2907 d->subsume_by(divmod->div_proj(), this); 2908 n->subsume_by(divmod->mod_proj(), this); 2909 } else { 2910 // replace a%b with a-((a/b)*b) 2911 Node* mult = new (this) MulLNode(d, d->in(2)); 2912 Node* sub = new (this) SubLNode(d->in(1), mult); 2913 n->subsume_by(sub, this); 2914 } 2915 } 2916 } 2917 break; 2918 2919 case Op_LoadVector: 2920 case Op_StoreVector: 2921 break; 2922 2923 case Op_PackB: 2924 case Op_PackS: 2925 case Op_PackI: 2926 case Op_PackF: 2927 case Op_PackL: 2928 case Op_PackD: 2929 if (n->req()-1 > 2) { 2930 // Replace many operand PackNodes with a binary tree for matching 2931 PackNode* p = (PackNode*) n; 2932 Node* btp = p->binary_tree_pack(this, 1, n->req()); 2933 n->subsume_by(btp, this); 2934 } 2935 break; 2936 case Op_Loop: 2937 case Op_CountedLoop: 2938 if (n->as_Loop()->is_inner_loop()) { 2939 frc.inc_inner_loop_count(); 2940 } 2941 break; 2942 case Op_LShiftI: 2943 case Op_RShiftI: 2944 case Op_URShiftI: 2945 case Op_LShiftL: 2946 case Op_RShiftL: 2947 case Op_URShiftL: 2948 if (Matcher::need_masked_shift_count) { 2949 // The cpu's shift instructions don't restrict the count to the 2950 // lower 5/6 bits. We need to do the masking ourselves. 2951 Node* in2 = n->in(2); 2952 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 2953 const TypeInt* t = in2->find_int_type(); 2954 if (t != NULL && t->is_con()) { 2955 juint shift = t->get_con(); 2956 if (shift > mask) { // Unsigned cmp 2957 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 2958 } 2959 } else { 2960 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 2961 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 2962 n->set_req(2, shift); 2963 } 2964 } 2965 if (in2->outcnt() == 0) { // Remove dead node 2966 in2->disconnect_inputs(NULL, this); 2967 } 2968 } 2969 break; 2970 case Op_MemBarStoreStore: 2971 case Op_MemBarRelease: 2972 // Break the link with AllocateNode: it is no longer useful and 2973 // confuses register allocation. 2974 if (n->req() > MemBarNode::Precedent) { 2975 n->set_req(MemBarNode::Precedent, top()); 2976 } 2977 break; 2978 default: 2979 assert( !n->is_Call(), "" ); 2980 assert( !n->is_Mem(), "" ); 2981 break; 2982 } 2983 2984 // Collect CFG split points 2985 if (n->is_MultiBranch()) 2986 frc._tests.push(n); 2987 } 2988 2989 //------------------------------final_graph_reshaping_walk--------------------- 2990 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2991 // requires that the walk visits a node's inputs before visiting the node. 2992 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2993 ResourceArea *area = Thread::current()->resource_area(); 2994 Unique_Node_List sfpt(area); 2995 2996 frc._visited.set(root->_idx); // first, mark node as visited 2997 uint cnt = root->req(); 2998 Node *n = root; 2999 uint i = 0; 3000 while (true) { 3001 if (i < cnt) { 3002 // Place all non-visited non-null inputs onto stack 3003 Node* m = n->in(i); 3004 ++i; 3005 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3006 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 3007 sfpt.push(m); 3008 cnt = m->req(); 3009 nstack.push(n, i); // put on stack parent and next input's index 3010 n = m; 3011 i = 0; 3012 } 3013 } else { 3014 // Now do post-visit work 3015 final_graph_reshaping_impl( n, frc ); 3016 if (nstack.is_empty()) 3017 break; // finished 3018 n = nstack.node(); // Get node from stack 3019 cnt = n->req(); 3020 i = nstack.index(); 3021 nstack.pop(); // Shift to the next node on stack 3022 } 3023 } 3024 3025 // Skip next transformation if compressed oops are not used. 3026 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3027 (!UseCompressedOops && !UseCompressedKlassPointers)) 3028 return; 3029 3030 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3031 // It could be done for an uncommon traps or any safepoints/calls 3032 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3033 while (sfpt.size() > 0) { 3034 n = sfpt.pop(); 3035 JVMState *jvms = n->as_SafePoint()->jvms(); 3036 assert(jvms != NULL, "sanity"); 3037 int start = jvms->debug_start(); 3038 int end = n->req(); 3039 bool is_uncommon = (n->is_CallStaticJava() && 3040 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3041 for (int j = start; j < end; j++) { 3042 Node* in = n->in(j); 3043 if (in->is_DecodeNarrowPtr()) { 3044 bool safe_to_skip = true; 3045 if (!is_uncommon ) { 3046 // Is it safe to skip? 3047 for (uint i = 0; i < in->outcnt(); i++) { 3048 Node* u = in->raw_out(i); 3049 if (!u->is_SafePoint() || 3050 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3051 safe_to_skip = false; 3052 } 3053 } 3054 } 3055 if (safe_to_skip) { 3056 n->set_req(j, in->in(1)); 3057 } 3058 if (in->outcnt() == 0) { 3059 in->disconnect_inputs(NULL, this); 3060 } 3061 } 3062 } 3063 } 3064 } 3065 3066 //------------------------------final_graph_reshaping-------------------------- 3067 // Final Graph Reshaping. 3068 // 3069 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3070 // and not commoned up and forced early. Must come after regular 3071 // optimizations to avoid GVN undoing the cloning. Clone constant 3072 // inputs to Loop Phis; these will be split by the allocator anyways. 3073 // Remove Opaque nodes. 3074 // (2) Move last-uses by commutative operations to the left input to encourage 3075 // Intel update-in-place two-address operations and better register usage 3076 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3077 // calls canonicalizing them back. 3078 // (3) Count the number of double-precision FP ops, single-precision FP ops 3079 // and call sites. On Intel, we can get correct rounding either by 3080 // forcing singles to memory (requires extra stores and loads after each 3081 // FP bytecode) or we can set a rounding mode bit (requires setting and 3082 // clearing the mode bit around call sites). The mode bit is only used 3083 // if the relative frequency of single FP ops to calls is low enough. 3084 // This is a key transform for SPEC mpeg_audio. 3085 // (4) Detect infinite loops; blobs of code reachable from above but not 3086 // below. Several of the Code_Gen algorithms fail on such code shapes, 3087 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3088 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3089 // Detection is by looking for IfNodes where only 1 projection is 3090 // reachable from below or CatchNodes missing some targets. 3091 // (5) Assert for insane oop offsets in debug mode. 3092 3093 bool Compile::final_graph_reshaping() { 3094 // an infinite loop may have been eliminated by the optimizer, 3095 // in which case the graph will be empty. 3096 if (root()->req() == 1) { 3097 record_method_not_compilable("trivial infinite loop"); 3098 return true; 3099 } 3100 3101 // Expensive nodes have their control input set to prevent the GVN 3102 // from freely commoning them. There's no GVN beyond this point so 3103 // no need to keep the control input. We want the expensive nodes to 3104 // be freely moved to the least frequent code path by gcm. 3105 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3106 for (int i = 0; i < expensive_count(); i++) { 3107 _expensive_nodes->at(i)->set_req(0, NULL); 3108 } 3109 3110 Final_Reshape_Counts frc; 3111 3112 // Visit everybody reachable! 3113 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3114 Node_Stack nstack(unique() >> 1); 3115 final_graph_reshaping_walk(nstack, root(), frc); 3116 3117 // Check for unreachable (from below) code (i.e., infinite loops). 3118 for( uint i = 0; i < frc._tests.size(); i++ ) { 3119 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3120 // Get number of CFG targets. 3121 // Note that PCTables include exception targets after calls. 3122 uint required_outcnt = n->required_outcnt(); 3123 if (n->outcnt() != required_outcnt) { 3124 // Check for a few special cases. Rethrow Nodes never take the 3125 // 'fall-thru' path, so expected kids is 1 less. 3126 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3127 if (n->in(0)->in(0)->is_Call()) { 3128 CallNode *call = n->in(0)->in(0)->as_Call(); 3129 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3130 required_outcnt--; // Rethrow always has 1 less kid 3131 } else if (call->req() > TypeFunc::Parms && 3132 call->is_CallDynamicJava()) { 3133 // Check for null receiver. In such case, the optimizer has 3134 // detected that the virtual call will always result in a null 3135 // pointer exception. The fall-through projection of this CatchNode 3136 // will not be populated. 3137 Node *arg0 = call->in(TypeFunc::Parms); 3138 if (arg0->is_Type() && 3139 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3140 required_outcnt--; 3141 } 3142 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3143 call->req() > TypeFunc::Parms+1 && 3144 call->is_CallStaticJava()) { 3145 // Check for negative array length. In such case, the optimizer has 3146 // detected that the allocation attempt will always result in an 3147 // exception. There is no fall-through projection of this CatchNode . 3148 Node *arg1 = call->in(TypeFunc::Parms+1); 3149 if (arg1->is_Type() && 3150 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3151 required_outcnt--; 3152 } 3153 } 3154 } 3155 } 3156 // Recheck with a better notion of 'required_outcnt' 3157 if (n->outcnt() != required_outcnt) { 3158 record_method_not_compilable("malformed control flow"); 3159 return true; // Not all targets reachable! 3160 } 3161 } 3162 // Check that I actually visited all kids. Unreached kids 3163 // must be infinite loops. 3164 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3165 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3166 record_method_not_compilable("infinite loop"); 3167 return true; // Found unvisited kid; must be unreach 3168 } 3169 } 3170 3171 // If original bytecodes contained a mixture of floats and doubles 3172 // check if the optimizer has made it homogenous, item (3). 3173 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3174 frc.get_float_count() > 32 && 3175 frc.get_double_count() == 0 && 3176 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3177 set_24_bit_selection_and_mode( false, true ); 3178 } 3179 3180 set_java_calls(frc.get_java_call_count()); 3181 set_inner_loops(frc.get_inner_loop_count()); 3182 3183 // No infinite loops, no reason to bail out. 3184 return false; 3185 } 3186 3187 //-----------------------------too_many_traps---------------------------------- 3188 // Report if there are too many traps at the current method and bci. 3189 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3190 bool Compile::too_many_traps(ciMethod* method, 3191 int bci, 3192 Deoptimization::DeoptReason reason) { 3193 ciMethodData* md = method->method_data(); 3194 if (md->is_empty()) { 3195 // Assume the trap has not occurred, or that it occurred only 3196 // because of a transient condition during start-up in the interpreter. 3197 return false; 3198 } 3199 if (md->has_trap_at(bci, reason) != 0) { 3200 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3201 // Also, if there are multiple reasons, or if there is no per-BCI record, 3202 // assume the worst. 3203 if (log()) 3204 log()->elem("observe trap='%s' count='%d'", 3205 Deoptimization::trap_reason_name(reason), 3206 md->trap_count(reason)); 3207 return true; 3208 } else { 3209 // Ignore method/bci and see if there have been too many globally. 3210 return too_many_traps(reason, md); 3211 } 3212 } 3213 3214 // Less-accurate variant which does not require a method and bci. 3215 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3216 ciMethodData* logmd) { 3217 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 3218 // Too many traps globally. 3219 // Note that we use cumulative trap_count, not just md->trap_count. 3220 if (log()) { 3221 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3222 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3223 Deoptimization::trap_reason_name(reason), 3224 mcount, trap_count(reason)); 3225 } 3226 return true; 3227 } else { 3228 // The coast is clear. 3229 return false; 3230 } 3231 } 3232 3233 //--------------------------too_many_recompiles-------------------------------- 3234 // Report if there are too many recompiles at the current method and bci. 3235 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3236 // Is not eager to return true, since this will cause the compiler to use 3237 // Action_none for a trap point, to avoid too many recompilations. 3238 bool Compile::too_many_recompiles(ciMethod* method, 3239 int bci, 3240 Deoptimization::DeoptReason reason) { 3241 ciMethodData* md = method->method_data(); 3242 if (md->is_empty()) { 3243 // Assume the trap has not occurred, or that it occurred only 3244 // because of a transient condition during start-up in the interpreter. 3245 return false; 3246 } 3247 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3248 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3249 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3250 Deoptimization::DeoptReason per_bc_reason 3251 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3252 if ((per_bc_reason == Deoptimization::Reason_none 3253 || md->has_trap_at(bci, reason) != 0) 3254 // The trap frequency measure we care about is the recompile count: 3255 && md->trap_recompiled_at(bci) 3256 && md->overflow_recompile_count() >= bc_cutoff) { 3257 // Do not emit a trap here if it has already caused recompilations. 3258 // Also, if there are multiple reasons, or if there is no per-BCI record, 3259 // assume the worst. 3260 if (log()) 3261 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3262 Deoptimization::trap_reason_name(reason), 3263 md->trap_count(reason), 3264 md->overflow_recompile_count()); 3265 return true; 3266 } else if (trap_count(reason) != 0 3267 && decompile_count() >= m_cutoff) { 3268 // Too many recompiles globally, and we have seen this sort of trap. 3269 // Use cumulative decompile_count, not just md->decompile_count. 3270 if (log()) 3271 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3272 Deoptimization::trap_reason_name(reason), 3273 md->trap_count(reason), trap_count(reason), 3274 md->decompile_count(), decompile_count()); 3275 return true; 3276 } else { 3277 // The coast is clear. 3278 return false; 3279 } 3280 } 3281 3282 3283 #ifndef PRODUCT 3284 //------------------------------verify_graph_edges--------------------------- 3285 // Walk the Graph and verify that there is a one-to-one correspondence 3286 // between Use-Def edges and Def-Use edges in the graph. 3287 void Compile::verify_graph_edges(bool no_dead_code) { 3288 if (VerifyGraphEdges) { 3289 ResourceArea *area = Thread::current()->resource_area(); 3290 Unique_Node_List visited(area); 3291 // Call recursive graph walk to check edges 3292 _root->verify_edges(visited); 3293 if (no_dead_code) { 3294 // Now make sure that no visited node is used by an unvisited node. 3295 bool dead_nodes = 0; 3296 Unique_Node_List checked(area); 3297 while (visited.size() > 0) { 3298 Node* n = visited.pop(); 3299 checked.push(n); 3300 for (uint i = 0; i < n->outcnt(); i++) { 3301 Node* use = n->raw_out(i); 3302 if (checked.member(use)) continue; // already checked 3303 if (visited.member(use)) continue; // already in the graph 3304 if (use->is_Con()) continue; // a dead ConNode is OK 3305 // At this point, we have found a dead node which is DU-reachable. 3306 if (dead_nodes++ == 0) 3307 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3308 use->dump(2); 3309 tty->print_cr("---"); 3310 checked.push(use); // No repeats; pretend it is now checked. 3311 } 3312 } 3313 assert(dead_nodes == 0, "using nodes must be reachable from root"); 3314 } 3315 } 3316 } 3317 #endif 3318 3319 // The Compile object keeps track of failure reasons separately from the ciEnv. 3320 // This is required because there is not quite a 1-1 relation between the 3321 // ciEnv and its compilation task and the Compile object. Note that one 3322 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3323 // to backtrack and retry without subsuming loads. Other than this backtracking 3324 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3325 // by the logic in C2Compiler. 3326 void Compile::record_failure(const char* reason) { 3327 if (log() != NULL) { 3328 log()->elem("failure reason='%s' phase='compile'", reason); 3329 } 3330 if (_failure_reason == NULL) { 3331 // Record the first failure reason. 3332 _failure_reason = reason; 3333 } 3334 3335 EventCompilerFailure event; 3336 if (event.should_commit()) { 3337 event.set_compileID(Compile::compile_id()); 3338 event.set_failure(reason); 3339 event.commit(); 3340 } 3341 3342 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3343 C->print_method(PHASE_FAILURE); 3344 } 3345 _root = NULL; // flush the graph, too 3346 } 3347 3348 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3349 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3350 _phase_name(name), _dolog(dolog) 3351 { 3352 if (dolog) { 3353 C = Compile::current(); 3354 _log = C->log(); 3355 } else { 3356 C = NULL; 3357 _log = NULL; 3358 } 3359 if (_log != NULL) { 3360 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3361 _log->stamp(); 3362 _log->end_head(); 3363 } 3364 } 3365 3366 Compile::TracePhase::~TracePhase() { 3367 3368 C = Compile::current(); 3369 if (_dolog) { 3370 _log = C->log(); 3371 } else { 3372 _log = NULL; 3373 } 3374 3375 #ifdef ASSERT 3376 if (PrintIdealNodeCount) { 3377 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3378 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3379 } 3380 3381 if (VerifyIdealNodeCount) { 3382 Compile::current()->print_missing_nodes(); 3383 } 3384 #endif 3385 3386 if (_log != NULL) { 3387 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3388 } 3389 } 3390 3391 //============================================================================= 3392 // Two Constant's are equal when the type and the value are equal. 3393 bool Compile::Constant::operator==(const Constant& other) { 3394 if (type() != other.type() ) return false; 3395 if (can_be_reused() != other.can_be_reused()) return false; 3396 // For floating point values we compare the bit pattern. 3397 switch (type()) { 3398 case T_FLOAT: return (_v._value.i == other._v._value.i); 3399 case T_LONG: 3400 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3401 case T_OBJECT: 3402 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3403 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3404 case T_METADATA: return (_v._metadata == other._v._metadata); 3405 default: ShouldNotReachHere(); 3406 } 3407 return false; 3408 } 3409 3410 static int type_to_size_in_bytes(BasicType t) { 3411 switch (t) { 3412 case T_LONG: return sizeof(jlong ); 3413 case T_FLOAT: return sizeof(jfloat ); 3414 case T_DOUBLE: return sizeof(jdouble); 3415 case T_METADATA: return sizeof(Metadata*); 3416 // We use T_VOID as marker for jump-table entries (labels) which 3417 // need an internal word relocation. 3418 case T_VOID: 3419 case T_ADDRESS: 3420 case T_OBJECT: return sizeof(jobject); 3421 } 3422 3423 ShouldNotReachHere(); 3424 return -1; 3425 } 3426 3427 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3428 // sort descending 3429 if (a->freq() > b->freq()) return -1; 3430 if (a->freq() < b->freq()) return 1; 3431 return 0; 3432 } 3433 3434 void Compile::ConstantTable::calculate_offsets_and_size() { 3435 // First, sort the array by frequencies. 3436 _constants.sort(qsort_comparator); 3437 3438 #ifdef ASSERT 3439 // Make sure all jump-table entries were sorted to the end of the 3440 // array (they have a negative frequency). 3441 bool found_void = false; 3442 for (int i = 0; i < _constants.length(); i++) { 3443 Constant con = _constants.at(i); 3444 if (con.type() == T_VOID) 3445 found_void = true; // jump-tables 3446 else 3447 assert(!found_void, "wrong sorting"); 3448 } 3449 #endif 3450 3451 int offset = 0; 3452 for (int i = 0; i < _constants.length(); i++) { 3453 Constant* con = _constants.adr_at(i); 3454 3455 // Align offset for type. 3456 int typesize = type_to_size_in_bytes(con->type()); 3457 offset = align_size_up(offset, typesize); 3458 con->set_offset(offset); // set constant's offset 3459 3460 if (con->type() == T_VOID) { 3461 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3462 offset = offset + typesize * n->outcnt(); // expand jump-table 3463 } else { 3464 offset = offset + typesize; 3465 } 3466 } 3467 3468 // Align size up to the next section start (which is insts; see 3469 // CodeBuffer::align_at_start). 3470 assert(_size == -1, "already set?"); 3471 _size = align_size_up(offset, CodeEntryAlignment); 3472 } 3473 3474 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3475 MacroAssembler _masm(&cb); 3476 for (int i = 0; i < _constants.length(); i++) { 3477 Constant con = _constants.at(i); 3478 address constant_addr; 3479 switch (con.type()) { 3480 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3481 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3482 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3483 case T_OBJECT: { 3484 jobject obj = con.get_jobject(); 3485 int oop_index = _masm.oop_recorder()->find_index(obj); 3486 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3487 break; 3488 } 3489 case T_ADDRESS: { 3490 address addr = (address) con.get_jobject(); 3491 constant_addr = _masm.address_constant(addr); 3492 break; 3493 } 3494 // We use T_VOID as marker for jump-table entries (labels) which 3495 // need an internal word relocation. 3496 case T_VOID: { 3497 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3498 // Fill the jump-table with a dummy word. The real value is 3499 // filled in later in fill_jump_table. 3500 address dummy = (address) n; 3501 constant_addr = _masm.address_constant(dummy); 3502 // Expand jump-table 3503 for (uint i = 1; i < n->outcnt(); i++) { 3504 address temp_addr = _masm.address_constant(dummy + i); 3505 assert(temp_addr, "consts section too small"); 3506 } 3507 break; 3508 } 3509 case T_METADATA: { 3510 Metadata* obj = con.get_metadata(); 3511 int metadata_index = _masm.oop_recorder()->find_index(obj); 3512 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3513 break; 3514 } 3515 default: ShouldNotReachHere(); 3516 } 3517 assert(constant_addr, "consts section too small"); 3518 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3519 } 3520 } 3521 3522 int Compile::ConstantTable::find_offset(Constant& con) const { 3523 int idx = _constants.find(con); 3524 assert(idx != -1, "constant must be in constant table"); 3525 int offset = _constants.at(idx).offset(); 3526 assert(offset != -1, "constant table not emitted yet?"); 3527 return offset; 3528 } 3529 3530 void Compile::ConstantTable::add(Constant& con) { 3531 if (con.can_be_reused()) { 3532 int idx = _constants.find(con); 3533 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3534 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3535 return; 3536 } 3537 } 3538 (void) _constants.append(con); 3539 } 3540 3541 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3542 Block* b = Compile::current()->cfg()->_bbs[n->_idx]; 3543 Constant con(type, value, b->_freq); 3544 add(con); 3545 return con; 3546 } 3547 3548 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3549 Constant con(metadata); 3550 add(con); 3551 return con; 3552 } 3553 3554 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3555 jvalue value; 3556 BasicType type = oper->type()->basic_type(); 3557 switch (type) { 3558 case T_LONG: value.j = oper->constantL(); break; 3559 case T_FLOAT: value.f = oper->constantF(); break; 3560 case T_DOUBLE: value.d = oper->constantD(); break; 3561 case T_OBJECT: 3562 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3563 case T_METADATA: return add((Metadata*)oper->constant()); break; 3564 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3565 } 3566 return add(n, type, value); 3567 } 3568 3569 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3570 jvalue value; 3571 // We can use the node pointer here to identify the right jump-table 3572 // as this method is called from Compile::Fill_buffer right before 3573 // the MachNodes are emitted and the jump-table is filled (means the 3574 // MachNode pointers do not change anymore). 3575 value.l = (jobject) n; 3576 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3577 add(con); 3578 return con; 3579 } 3580 3581 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3582 // If called from Compile::scratch_emit_size do nothing. 3583 if (Compile::current()->in_scratch_emit_size()) return; 3584 3585 assert(labels.is_nonempty(), "must be"); 3586 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3587 3588 // Since MachConstantNode::constant_offset() also contains 3589 // table_base_offset() we need to subtract the table_base_offset() 3590 // to get the plain offset into the constant table. 3591 int offset = n->constant_offset() - table_base_offset(); 3592 3593 MacroAssembler _masm(&cb); 3594 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3595 3596 for (uint i = 0; i < n->outcnt(); i++) { 3597 address* constant_addr = &jump_table_base[i]; 3598 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))); 3599 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3600 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3601 } 3602 } 3603 3604 void Compile::dump_inlining() { 3605 if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) { 3606 // Print inlining message for candidates that we couldn't inline 3607 // for lack of space or non constant receiver 3608 for (int i = 0; i < _late_inlines.length(); i++) { 3609 CallGenerator* cg = _late_inlines.at(i); 3610 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); 3611 } 3612 Unique_Node_List useful; 3613 useful.push(root()); 3614 for (uint next = 0; next < useful.size(); ++next) { 3615 Node* n = useful.at(next); 3616 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { 3617 CallNode* call = n->as_Call(); 3618 CallGenerator* cg = call->generator(); 3619 cg->print_inlining_late("receiver not constant"); 3620 } 3621 uint max = n->len(); 3622 for ( uint i = 0; i < max; ++i ) { 3623 Node *m = n->in(i); 3624 if ( m == NULL ) continue; 3625 useful.push(m); 3626 } 3627 } 3628 for (int i = 0; i < _print_inlining_list->length(); i++) { 3629 tty->print(_print_inlining_list->at(i).ss()->as_string()); 3630 } 3631 } 3632 } 3633 3634 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 3635 if (n1->Opcode() < n2->Opcode()) return -1; 3636 else if (n1->Opcode() > n2->Opcode()) return 1; 3637 3638 assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req())); 3639 for (uint i = 1; i < n1->req(); i++) { 3640 if (n1->in(i) < n2->in(i)) return -1; 3641 else if (n1->in(i) > n2->in(i)) return 1; 3642 } 3643 3644 return 0; 3645 } 3646 3647 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 3648 Node* n1 = *n1p; 3649 Node* n2 = *n2p; 3650 3651 return cmp_expensive_nodes(n1, n2); 3652 } 3653 3654 void Compile::sort_expensive_nodes() { 3655 if (!expensive_nodes_sorted()) { 3656 _expensive_nodes->sort(cmp_expensive_nodes); 3657 } 3658 } 3659 3660 bool Compile::expensive_nodes_sorted() const { 3661 for (int i = 1; i < _expensive_nodes->length(); i++) { 3662 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 3663 return false; 3664 } 3665 } 3666 return true; 3667 } 3668 3669 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 3670 if (_expensive_nodes->length() == 0) { 3671 return false; 3672 } 3673 3674 assert(OptimizeExpensiveOps, "optimization off?"); 3675 3676 // Take this opportunity to remove dead nodes from the list 3677 int j = 0; 3678 for (int i = 0; i < _expensive_nodes->length(); i++) { 3679 Node* n = _expensive_nodes->at(i); 3680 if (!n->is_unreachable(igvn)) { 3681 assert(n->is_expensive(), "should be expensive"); 3682 _expensive_nodes->at_put(j, n); 3683 j++; 3684 } 3685 } 3686 _expensive_nodes->trunc_to(j); 3687 3688 // Then sort the list so that similar nodes are next to each other 3689 // and check for at least two nodes of identical kind with same data 3690 // inputs. 3691 sort_expensive_nodes(); 3692 3693 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 3694 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 3695 return true; 3696 } 3697 } 3698 3699 return false; 3700 } 3701 3702 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 3703 if (_expensive_nodes->length() == 0) { 3704 return; 3705 } 3706 3707 assert(OptimizeExpensiveOps, "optimization off?"); 3708 3709 // Sort to bring similar nodes next to each other and clear the 3710 // control input of nodes for which there's only a single copy. 3711 sort_expensive_nodes(); 3712 3713 int j = 0; 3714 int identical = 0; 3715 int i = 0; 3716 for (; i < _expensive_nodes->length()-1; i++) { 3717 assert(j <= i, "can't write beyond current index"); 3718 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 3719 identical++; 3720 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3721 continue; 3722 } 3723 if (identical > 0) { 3724 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3725 identical = 0; 3726 } else { 3727 Node* n = _expensive_nodes->at(i); 3728 igvn.hash_delete(n); 3729 n->set_req(0, NULL); 3730 igvn.hash_insert(n); 3731 } 3732 } 3733 if (identical > 0) { 3734 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3735 } else if (_expensive_nodes->length() >= 1) { 3736 Node* n = _expensive_nodes->at(i); 3737 igvn.hash_delete(n); 3738 n->set_req(0, NULL); 3739 igvn.hash_insert(n); 3740 } 3741 _expensive_nodes->trunc_to(j); 3742 } 3743 3744 void Compile::add_expensive_node(Node * n) { 3745 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 3746 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 3747 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 3748 if (OptimizeExpensiveOps) { 3749 _expensive_nodes->append(n); 3750 } else { 3751 // Clear control input and let IGVN optimize expensive nodes if 3752 // OptimizeExpensiveOps is off. 3753 n->set_req(0, NULL); 3754 } 3755 } 3756 3757 // Auxiliary method to support randomized stressing/fuzzing. 3758 // 3759 // This method can be called the arbitrary number of times, with current count 3760 // as the argument. The logic allows selecting a single candidate from the 3761 // running list of candidates as follows: 3762 // int count = 0; 3763 // Cand* selected = null; 3764 // while(cand = cand->next()) { 3765 // if (randomized_select(++count)) { 3766 // selected = cand; 3767 // } 3768 // } 3769 // 3770 // Including count equalizes the chances any candidate is "selected". 3771 // This is useful when we don't have the complete list of candidates to choose 3772 // from uniformly. In this case, we need to adjust the randomicity of the 3773 // selection, or else we will end up biasing the selection towards the latter 3774 // candidates. 3775 // 3776 // Quick back-envelope calculation shows that for the list of n candidates 3777 // the equal probability for the candidate to persist as "best" can be 3778 // achieved by replacing it with "next" k-th candidate with the probability 3779 // of 1/k. It can be easily shown that by the end of the run, the 3780 // probability for any candidate is converged to 1/n, thus giving the 3781 // uniform distribution among all the candidates. 3782 // 3783 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 3784 #define RANDOMIZED_DOMAIN_POW 29 3785 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 3786 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 3787 bool Compile::randomized_select(int count) { 3788 assert(count > 0, "only positive"); 3789 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 3790 }