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_idx(0), 658 _preserve_jvm_state(0) { 659 C = this; 660 661 CompileWrapper cw(this); 662 #ifndef PRODUCT 663 if (TimeCompiler2) { 664 tty->print(" "); 665 target->holder()->name()->print(); 666 tty->print("."); 667 target->print_short_name(); 668 tty->print(" "); 669 } 670 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 671 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 672 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 673 if (!print_opto_assembly) { 674 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 675 if (print_assembly && !Disassembler::can_decode()) { 676 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 677 print_opto_assembly = true; 678 } 679 } 680 set_print_assembly(print_opto_assembly); 681 set_parsed_irreducible_loop(false); 682 #endif 683 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 684 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 685 686 if (ProfileTraps) { 687 // Make sure the method being compiled gets its own MDO, 688 // so we can at least track the decompile_count(). 689 method()->ensure_method_data(); 690 } 691 692 Init(::AliasLevel); 693 694 695 print_compile_messages(); 696 697 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 698 _ilt = InlineTree::build_inline_tree_root(); 699 else 700 _ilt = NULL; 701 702 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 703 assert(num_alias_types() >= AliasIdxRaw, ""); 704 705 #define MINIMUM_NODE_HASH 1023 706 // Node list that Iterative GVN will start with 707 Unique_Node_List for_igvn(comp_arena()); 708 set_for_igvn(&for_igvn); 709 710 // GVN that will be run immediately on new nodes 711 uint estimated_size = method()->code_size()*4+64; 712 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 713 PhaseGVN gvn(node_arena(), estimated_size); 714 set_initial_gvn(&gvn); 715 716 if (print_inlining() || print_intrinsics()) { 717 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 718 } 719 { // Scope for timing the parser 720 TracePhase t3("parse", &_t_parser, true); 721 722 // Put top into the hash table ASAP. 723 initial_gvn()->transform_no_reclaim(top()); 724 725 // Set up tf(), start(), and find a CallGenerator. 726 CallGenerator* cg = NULL; 727 if (is_osr_compilation()) { 728 const TypeTuple *domain = StartOSRNode::osr_domain(); 729 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 730 init_tf(TypeFunc::make(domain, range)); 731 StartNode* s = new (this) StartOSRNode(root(), domain); 732 initial_gvn()->set_type_bottom(s); 733 init_start(s); 734 cg = CallGenerator::for_osr(method(), entry_bci()); 735 } else { 736 // Normal case. 737 init_tf(TypeFunc::make(method())); 738 StartNode* s = new (this) StartNode(root(), tf()->domain()); 739 initial_gvn()->set_type_bottom(s); 740 init_start(s); 741 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 742 // With java.lang.ref.reference.get() we must go through the 743 // intrinsic when G1 is enabled - even when get() is the root 744 // method of the compile - so that, if necessary, the value in 745 // the referent field of the reference object gets recorded by 746 // the pre-barrier code. 747 // Specifically, if G1 is enabled, the value in the referent 748 // field is recorded by the G1 SATB pre barrier. This will 749 // result in the referent being marked live and the reference 750 // object removed from the list of discovered references during 751 // reference processing. 752 cg = find_intrinsic(method(), false); 753 } 754 if (cg == NULL) { 755 float past_uses = method()->interpreter_invocation_count(); 756 float expected_uses = past_uses; 757 cg = CallGenerator::for_inline(method(), expected_uses); 758 } 759 } 760 if (failing()) return; 761 if (cg == NULL) { 762 record_method_not_compilable_all_tiers("cannot parse method"); 763 return; 764 } 765 JVMState* jvms = build_start_state(start(), tf()); 766 if ((jvms = cg->generate(jvms, NULL)) == NULL) { 767 record_method_not_compilable("method parse failed"); 768 return; 769 } 770 GraphKit kit(jvms); 771 772 if (!kit.stopped()) { 773 // Accept return values, and transfer control we know not where. 774 // This is done by a special, unique ReturnNode bound to root. 775 return_values(kit.jvms()); 776 } 777 778 if (kit.has_exceptions()) { 779 // Any exceptions that escape from this call must be rethrown 780 // to whatever caller is dynamically above us on the stack. 781 // This is done by a special, unique RethrowNode bound to root. 782 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 783 } 784 785 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 786 787 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 788 inline_string_calls(true); 789 } 790 791 if (failing()) return; 792 793 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 794 795 // Remove clutter produced by parsing. 796 if (!failing()) { 797 ResourceMark rm; 798 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 799 } 800 } 801 802 // Note: Large methods are capped off in do_one_bytecode(). 803 if (failing()) return; 804 805 // After parsing, node notes are no longer automagic. 806 // They must be propagated by register_new_node_with_optimizer(), 807 // clone(), or the like. 808 set_default_node_notes(NULL); 809 810 for (;;) { 811 int successes = Inline_Warm(); 812 if (failing()) return; 813 if (successes == 0) break; 814 } 815 816 // Drain the list. 817 Finish_Warm(); 818 #ifndef PRODUCT 819 if (_printer) { 820 _printer->print_inlining(this); 821 } 822 #endif 823 824 if (failing()) return; 825 NOT_PRODUCT( verify_graph_edges(); ) 826 827 // Now optimize 828 Optimize(); 829 if (failing()) return; 830 NOT_PRODUCT( verify_graph_edges(); ) 831 832 #ifndef PRODUCT 833 if (PrintIdeal) { 834 ttyLocker ttyl; // keep the following output all in one block 835 // This output goes directly to the tty, not the compiler log. 836 // To enable tools to match it up with the compilation activity, 837 // be sure to tag this tty output with the compile ID. 838 if (xtty != NULL) { 839 xtty->head("ideal compile_id='%d'%s", compile_id(), 840 is_osr_compilation() ? " compile_kind='osr'" : 841 ""); 842 } 843 root()->dump(9999); 844 if (xtty != NULL) { 845 xtty->tail("ideal"); 846 } 847 } 848 #endif 849 850 // Now that we know the size of all the monitors we can add a fixed slot 851 // for the original deopt pc. 852 853 _orig_pc_slot = fixed_slots(); 854 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 855 set_fixed_slots(next_slot); 856 857 // Now generate code 858 Code_Gen(); 859 if (failing()) return; 860 861 // Check if we want to skip execution of all compiled code. 862 { 863 #ifndef PRODUCT 864 if (OptoNoExecute) { 865 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 866 return; 867 } 868 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 869 #endif 870 871 if (is_osr_compilation()) { 872 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 873 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 874 } else { 875 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 876 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 877 } 878 879 env()->register_method(_method, _entry_bci, 880 &_code_offsets, 881 _orig_pc_slot_offset_in_bytes, 882 code_buffer(), 883 frame_size_in_words(), _oop_map_set, 884 &_handler_table, &_inc_table, 885 compiler, 886 env()->comp_level(), 887 has_unsafe_access(), 888 SharedRuntime::is_wide_vector(max_vector_size()) 889 ); 890 891 if (log() != NULL) // Print code cache state into compiler log 892 log()->code_cache_state(); 893 } 894 } 895 896 //------------------------------Compile---------------------------------------- 897 // Compile a runtime stub 898 Compile::Compile( ciEnv* ci_env, 899 TypeFunc_generator generator, 900 address stub_function, 901 const char *stub_name, 902 int is_fancy_jump, 903 bool pass_tls, 904 bool save_arg_registers, 905 bool return_pc ) 906 : Phase(Compiler), 907 _env(ci_env), 908 _log(ci_env->log()), 909 _compile_id(0), 910 _save_argument_registers(save_arg_registers), 911 _method(NULL), 912 _stub_name(stub_name), 913 _stub_function(stub_function), 914 _stub_entry_point(NULL), 915 _entry_bci(InvocationEntryBci), 916 _initial_gvn(NULL), 917 _for_igvn(NULL), 918 _warm_calls(NULL), 919 _orig_pc_slot(0), 920 _orig_pc_slot_offset_in_bytes(0), 921 _subsume_loads(true), 922 _do_escape_analysis(false), 923 _eliminate_boxing(false), 924 _failure_reason(NULL), 925 _code_buffer("Compile::Fill_buffer"), 926 _has_method_handle_invokes(false), 927 _mach_constant_base_node(NULL), 928 _node_bundling_limit(0), 929 _node_bundling_base(NULL), 930 _java_calls(0), 931 _inner_loops(0), 932 #ifndef PRODUCT 933 _trace_opto_output(TraceOptoOutput), 934 _printer(NULL), 935 #endif 936 _dead_node_list(comp_arena()), 937 _dead_node_count(0), 938 _congraph(NULL), 939 _number_of_mh_late_inlines(0), 940 _inlining_progress(false), 941 _inlining_incrementally(false), 942 _print_inlining_list(NULL), 943 _print_inlining_idx(0), 944 _preserve_jvm_state(0) { 945 C = this; 946 947 #ifndef PRODUCT 948 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 949 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 950 set_print_assembly(PrintFrameConverterAssembly); 951 set_parsed_irreducible_loop(false); 952 #endif 953 CompileWrapper cw(this); 954 Init(/*AliasLevel=*/ 0); 955 init_tf((*generator)()); 956 957 { 958 // The following is a dummy for the sake of GraphKit::gen_stub 959 Unique_Node_List for_igvn(comp_arena()); 960 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 961 PhaseGVN gvn(Thread::current()->resource_area(),255); 962 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 963 gvn.transform_no_reclaim(top()); 964 965 GraphKit kit; 966 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 967 } 968 969 NOT_PRODUCT( verify_graph_edges(); ) 970 Code_Gen(); 971 if (failing()) return; 972 973 974 // Entry point will be accessed using compile->stub_entry_point(); 975 if (code_buffer() == NULL) { 976 Matcher::soft_match_failure(); 977 } else { 978 if (PrintAssembly && (WizardMode || Verbose)) 979 tty->print_cr("### Stub::%s", stub_name); 980 981 if (!failing()) { 982 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 983 984 // Make the NMethod 985 // For now we mark the frame as never safe for profile stackwalking 986 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 987 code_buffer(), 988 CodeOffsets::frame_never_safe, 989 // _code_offsets.value(CodeOffsets::Frame_Complete), 990 frame_size_in_words(), 991 _oop_map_set, 992 save_arg_registers); 993 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 994 995 _stub_entry_point = rs->entry_point(); 996 } 997 } 998 } 999 1000 //------------------------------Init------------------------------------------- 1001 // Prepare for a single compilation 1002 void Compile::Init(int aliaslevel) { 1003 _unique = 0; 1004 _regalloc = NULL; 1005 1006 _tf = NULL; // filled in later 1007 _top = NULL; // cached later 1008 _matcher = NULL; // filled in later 1009 _cfg = NULL; // filled in later 1010 1011 set_24_bit_selection_and_mode(Use24BitFP, false); 1012 1013 _node_note_array = NULL; 1014 _default_node_notes = NULL; 1015 1016 _immutable_memory = NULL; // filled in at first inquiry 1017 1018 // Globally visible Nodes 1019 // First set TOP to NULL to give safe behavior during creation of RootNode 1020 set_cached_top_node(NULL); 1021 set_root(new (this) RootNode()); 1022 // Now that you have a Root to point to, create the real TOP 1023 set_cached_top_node( new (this) ConNode(Type::TOP) ); 1024 set_recent_alloc(NULL, NULL); 1025 1026 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1027 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1028 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1029 env()->set_dependencies(new Dependencies(env())); 1030 1031 _fixed_slots = 0; 1032 set_has_split_ifs(false); 1033 set_has_loops(has_method() && method()->has_loops()); // first approximation 1034 set_has_stringbuilder(false); 1035 set_has_boxed_value(false); 1036 _trap_can_recompile = false; // no traps emitted yet 1037 _major_progress = true; // start out assuming good things will happen 1038 set_has_unsafe_access(false); 1039 set_max_vector_size(0); 1040 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1041 set_decompile_count(0); 1042 1043 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1044 set_num_loop_opts(LoopOptsCount); 1045 set_do_inlining(Inline); 1046 set_max_inline_size(MaxInlineSize); 1047 set_freq_inline_size(FreqInlineSize); 1048 set_do_scheduling(OptoScheduling); 1049 set_do_count_invocations(false); 1050 set_do_method_data_update(false); 1051 1052 if (debug_info()->recording_non_safepoints()) { 1053 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1054 (comp_arena(), 8, 0, NULL)); 1055 set_default_node_notes(Node_Notes::make(this)); 1056 } 1057 1058 // // -- Initialize types before each compile -- 1059 // // Update cached type information 1060 // if( _method && _method->constants() ) 1061 // Type::update_loaded_types(_method, _method->constants()); 1062 1063 // Init alias_type map. 1064 if (!_do_escape_analysis && aliaslevel == 3) 1065 aliaslevel = 2; // No unique types without escape analysis 1066 _AliasLevel = aliaslevel; 1067 const int grow_ats = 16; 1068 _max_alias_types = grow_ats; 1069 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1070 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1071 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1072 { 1073 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1074 } 1075 // Initialize the first few types. 1076 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1077 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1078 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1079 _num_alias_types = AliasIdxRaw+1; 1080 // Zero out the alias type cache. 1081 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1082 // A NULL adr_type hits in the cache right away. Preload the right answer. 1083 probe_alias_cache(NULL)->_index = AliasIdxTop; 1084 1085 _intrinsics = NULL; 1086 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1087 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1088 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1089 register_library_intrinsics(); 1090 } 1091 1092 //---------------------------init_start---------------------------------------- 1093 // Install the StartNode on this compile object. 1094 void Compile::init_start(StartNode* s) { 1095 if (failing()) 1096 return; // already failing 1097 assert(s == start(), ""); 1098 } 1099 1100 StartNode* Compile::start() const { 1101 assert(!failing(), ""); 1102 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1103 Node* start = root()->fast_out(i); 1104 if( start->is_Start() ) 1105 return start->as_Start(); 1106 } 1107 ShouldNotReachHere(); 1108 return NULL; 1109 } 1110 1111 //-------------------------------immutable_memory------------------------------------- 1112 // Access immutable memory 1113 Node* Compile::immutable_memory() { 1114 if (_immutable_memory != NULL) { 1115 return _immutable_memory; 1116 } 1117 StartNode* s = start(); 1118 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1119 Node *p = s->fast_out(i); 1120 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1121 _immutable_memory = p; 1122 return _immutable_memory; 1123 } 1124 } 1125 ShouldNotReachHere(); 1126 return NULL; 1127 } 1128 1129 //----------------------set_cached_top_node------------------------------------ 1130 // Install the cached top node, and make sure Node::is_top works correctly. 1131 void Compile::set_cached_top_node(Node* tn) { 1132 if (tn != NULL) verify_top(tn); 1133 Node* old_top = _top; 1134 _top = tn; 1135 // Calling Node::setup_is_top allows the nodes the chance to adjust 1136 // their _out arrays. 1137 if (_top != NULL) _top->setup_is_top(); 1138 if (old_top != NULL) old_top->setup_is_top(); 1139 assert(_top == NULL || top()->is_top(), ""); 1140 } 1141 1142 #ifdef ASSERT 1143 uint Compile::count_live_nodes_by_graph_walk() { 1144 Unique_Node_List useful(comp_arena()); 1145 // Get useful node list by walking the graph. 1146 identify_useful_nodes(useful); 1147 return useful.size(); 1148 } 1149 1150 void Compile::print_missing_nodes() { 1151 1152 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1153 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1154 return; 1155 } 1156 1157 // This is an expensive function. It is executed only when the user 1158 // specifies VerifyIdealNodeCount option or otherwise knows the 1159 // additional work that needs to be done to identify reachable nodes 1160 // by walking the flow graph and find the missing ones using 1161 // _dead_node_list. 1162 1163 Unique_Node_List useful(comp_arena()); 1164 // Get useful node list by walking the graph. 1165 identify_useful_nodes(useful); 1166 1167 uint l_nodes = C->live_nodes(); 1168 uint l_nodes_by_walk = useful.size(); 1169 1170 if (l_nodes != l_nodes_by_walk) { 1171 if (_log != NULL) { 1172 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1173 _log->stamp(); 1174 _log->end_head(); 1175 } 1176 VectorSet& useful_member_set = useful.member_set(); 1177 int last_idx = l_nodes_by_walk; 1178 for (int i = 0; i < last_idx; i++) { 1179 if (useful_member_set.test(i)) { 1180 if (_dead_node_list.test(i)) { 1181 if (_log != NULL) { 1182 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1183 } 1184 if (PrintIdealNodeCount) { 1185 // Print the log message to tty 1186 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1187 useful.at(i)->dump(); 1188 } 1189 } 1190 } 1191 else if (! _dead_node_list.test(i)) { 1192 if (_log != NULL) { 1193 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1194 } 1195 if (PrintIdealNodeCount) { 1196 // Print the log message to tty 1197 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1198 } 1199 } 1200 } 1201 if (_log != NULL) { 1202 _log->tail("mismatched_nodes"); 1203 } 1204 } 1205 } 1206 #endif 1207 1208 #ifndef PRODUCT 1209 void Compile::verify_top(Node* tn) const { 1210 if (tn != NULL) { 1211 assert(tn->is_Con(), "top node must be a constant"); 1212 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1213 assert(tn->in(0) != NULL, "must have live top node"); 1214 } 1215 } 1216 #endif 1217 1218 1219 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1220 1221 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1222 guarantee(arr != NULL, ""); 1223 int num_blocks = arr->length(); 1224 if (grow_by < num_blocks) grow_by = num_blocks; 1225 int num_notes = grow_by * _node_notes_block_size; 1226 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1227 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1228 while (num_notes > 0) { 1229 arr->append(notes); 1230 notes += _node_notes_block_size; 1231 num_notes -= _node_notes_block_size; 1232 } 1233 assert(num_notes == 0, "exact multiple, please"); 1234 } 1235 1236 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1237 if (source == NULL || dest == NULL) return false; 1238 1239 if (dest->is_Con()) 1240 return false; // Do not push debug info onto constants. 1241 1242 #ifdef ASSERT 1243 // Leave a bread crumb trail pointing to the original node: 1244 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1245 dest->set_debug_orig(source); 1246 } 1247 #endif 1248 1249 if (node_note_array() == NULL) 1250 return false; // Not collecting any notes now. 1251 1252 // This is a copy onto a pre-existing node, which may already have notes. 1253 // If both nodes have notes, do not overwrite any pre-existing notes. 1254 Node_Notes* source_notes = node_notes_at(source->_idx); 1255 if (source_notes == NULL || source_notes->is_clear()) return false; 1256 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1257 if (dest_notes == NULL || dest_notes->is_clear()) { 1258 return set_node_notes_at(dest->_idx, source_notes); 1259 } 1260 1261 Node_Notes merged_notes = (*source_notes); 1262 // The order of operations here ensures that dest notes will win... 1263 merged_notes.update_from(dest_notes); 1264 return set_node_notes_at(dest->_idx, &merged_notes); 1265 } 1266 1267 1268 //--------------------------allow_range_check_smearing------------------------- 1269 // Gating condition for coalescing similar range checks. 1270 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1271 // single covering check that is at least as strong as any of them. 1272 // If the optimization succeeds, the simplified (strengthened) range check 1273 // will always succeed. If it fails, we will deopt, and then give up 1274 // on the optimization. 1275 bool Compile::allow_range_check_smearing() const { 1276 // If this method has already thrown a range-check, 1277 // assume it was because we already tried range smearing 1278 // and it failed. 1279 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1280 return !already_trapped; 1281 } 1282 1283 1284 //------------------------------flatten_alias_type----------------------------- 1285 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1286 int offset = tj->offset(); 1287 TypePtr::PTR ptr = tj->ptr(); 1288 1289 // Known instance (scalarizable allocation) alias only with itself. 1290 bool is_known_inst = tj->isa_oopptr() != NULL && 1291 tj->is_oopptr()->is_known_instance(); 1292 1293 // Process weird unsafe references. 1294 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1295 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1296 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1297 tj = TypeOopPtr::BOTTOM; 1298 ptr = tj->ptr(); 1299 offset = tj->offset(); 1300 } 1301 1302 // Array pointers need some flattening 1303 const TypeAryPtr *ta = tj->isa_aryptr(); 1304 if (ta && ta->is_stable()) { 1305 // Erase stability property for alias analysis. 1306 tj = ta = ta->cast_to_stable(false); 1307 } 1308 if( ta && is_known_inst ) { 1309 if ( offset != Type::OffsetBot && 1310 offset > arrayOopDesc::length_offset_in_bytes() ) { 1311 offset = Type::OffsetBot; // Flatten constant access into array body only 1312 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1313 } 1314 } else if( ta && _AliasLevel >= 2 ) { 1315 // For arrays indexed by constant indices, we flatten the alias 1316 // space to include all of the array body. Only the header, klass 1317 // and array length can be accessed un-aliased. 1318 if( offset != Type::OffsetBot ) { 1319 if( ta->const_oop() ) { // MethodData* or Method* 1320 offset = Type::OffsetBot; // Flatten constant access into array body 1321 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1322 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1323 // range is OK as-is. 1324 tj = ta = TypeAryPtr::RANGE; 1325 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1326 tj = TypeInstPtr::KLASS; // all klass loads look alike 1327 ta = TypeAryPtr::RANGE; // generic ignored junk 1328 ptr = TypePtr::BotPTR; 1329 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1330 tj = TypeInstPtr::MARK; 1331 ta = TypeAryPtr::RANGE; // generic ignored junk 1332 ptr = TypePtr::BotPTR; 1333 } else { // Random constant offset into array body 1334 offset = Type::OffsetBot; // Flatten constant access into array body 1335 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1336 } 1337 } 1338 // Arrays of fixed size alias with arrays of unknown size. 1339 if (ta->size() != TypeInt::POS) { 1340 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1341 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1342 } 1343 // Arrays of known objects become arrays of unknown objects. 1344 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1345 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1346 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1347 } 1348 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1349 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1350 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1351 } 1352 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1353 // cannot be distinguished by bytecode alone. 1354 if (ta->elem() == TypeInt::BOOL) { 1355 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1356 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1357 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1358 } 1359 // During the 2nd round of IterGVN, NotNull castings are removed. 1360 // Make sure the Bottom and NotNull variants alias the same. 1361 // Also, make sure exact and non-exact variants alias the same. 1362 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1363 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1364 } 1365 } 1366 1367 // Oop pointers need some flattening 1368 const TypeInstPtr *to = tj->isa_instptr(); 1369 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1370 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1371 if( ptr == TypePtr::Constant ) { 1372 if (to->klass() != ciEnv::current()->Class_klass() || 1373 offset < k->size_helper() * wordSize) { 1374 // No constant oop pointers (such as Strings); they alias with 1375 // unknown strings. 1376 assert(!is_known_inst, "not scalarizable allocation"); 1377 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1378 } 1379 } else if( is_known_inst ) { 1380 tj = to; // Keep NotNull and klass_is_exact for instance type 1381 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1382 // During the 2nd round of IterGVN, NotNull castings are removed. 1383 // Make sure the Bottom and NotNull variants alias the same. 1384 // Also, make sure exact and non-exact variants alias the same. 1385 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1386 } 1387 // Canonicalize the holder of this field 1388 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1389 // First handle header references such as a LoadKlassNode, even if the 1390 // object's klass is unloaded at compile time (4965979). 1391 if (!is_known_inst) { // Do it only for non-instance types 1392 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1393 } 1394 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1395 // Static fields are in the space above the normal instance 1396 // fields in the java.lang.Class instance. 1397 if (to->klass() != ciEnv::current()->Class_klass()) { 1398 to = NULL; 1399 tj = TypeOopPtr::BOTTOM; 1400 offset = tj->offset(); 1401 } 1402 } else { 1403 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1404 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1405 if( is_known_inst ) { 1406 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1407 } else { 1408 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1409 } 1410 } 1411 } 1412 } 1413 1414 // Klass pointers to object array klasses need some flattening 1415 const TypeKlassPtr *tk = tj->isa_klassptr(); 1416 if( tk ) { 1417 // If we are referencing a field within a Klass, we need 1418 // to assume the worst case of an Object. Both exact and 1419 // inexact types must flatten to the same alias class so 1420 // use NotNull as the PTR. 1421 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1422 1423 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1424 TypeKlassPtr::OBJECT->klass(), 1425 offset); 1426 } 1427 1428 ciKlass* klass = tk->klass(); 1429 if( klass->is_obj_array_klass() ) { 1430 ciKlass* k = TypeAryPtr::OOPS->klass(); 1431 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1432 k = TypeInstPtr::BOTTOM->klass(); 1433 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1434 } 1435 1436 // Check for precise loads from the primary supertype array and force them 1437 // to the supertype cache alias index. Check for generic array loads from 1438 // the primary supertype array and also force them to the supertype cache 1439 // alias index. Since the same load can reach both, we need to merge 1440 // these 2 disparate memories into the same alias class. Since the 1441 // primary supertype array is read-only, there's no chance of confusion 1442 // where we bypass an array load and an array store. 1443 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1444 if (offset == Type::OffsetBot || 1445 (offset >= primary_supers_offset && 1446 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1447 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1448 offset = in_bytes(Klass::secondary_super_cache_offset()); 1449 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1450 } 1451 } 1452 1453 // Flatten all Raw pointers together. 1454 if (tj->base() == Type::RawPtr) 1455 tj = TypeRawPtr::BOTTOM; 1456 1457 if (tj->base() == Type::AnyPtr) 1458 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1459 1460 // Flatten all to bottom for now 1461 switch( _AliasLevel ) { 1462 case 0: 1463 tj = TypePtr::BOTTOM; 1464 break; 1465 case 1: // Flatten to: oop, static, field or array 1466 switch (tj->base()) { 1467 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1468 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1469 case Type::AryPtr: // do not distinguish arrays at all 1470 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1471 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1472 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1473 default: ShouldNotReachHere(); 1474 } 1475 break; 1476 case 2: // No collapsing at level 2; keep all splits 1477 case 3: // No collapsing at level 3; keep all splits 1478 break; 1479 default: 1480 Unimplemented(); 1481 } 1482 1483 offset = tj->offset(); 1484 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1485 1486 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1487 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1488 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1489 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1490 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1491 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1492 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1493 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1494 assert( tj->ptr() != TypePtr::TopPTR && 1495 tj->ptr() != TypePtr::AnyNull && 1496 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1497 // assert( tj->ptr() != TypePtr::Constant || 1498 // tj->base() == Type::RawPtr || 1499 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1500 1501 return tj; 1502 } 1503 1504 void Compile::AliasType::Init(int i, const TypePtr* at) { 1505 _index = i; 1506 _adr_type = at; 1507 _field = NULL; 1508 _element = NULL; 1509 _is_rewritable = true; // default 1510 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1511 if (atoop != NULL && atoop->is_known_instance()) { 1512 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1513 _general_index = Compile::current()->get_alias_index(gt); 1514 } else { 1515 _general_index = 0; 1516 } 1517 } 1518 1519 //---------------------------------print_on------------------------------------ 1520 #ifndef PRODUCT 1521 void Compile::AliasType::print_on(outputStream* st) { 1522 if (index() < 10) 1523 st->print("@ <%d> ", index()); 1524 else st->print("@ <%d>", index()); 1525 st->print(is_rewritable() ? " " : " RO"); 1526 int offset = adr_type()->offset(); 1527 if (offset == Type::OffsetBot) 1528 st->print(" +any"); 1529 else st->print(" +%-3d", offset); 1530 st->print(" in "); 1531 adr_type()->dump_on(st); 1532 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1533 if (field() != NULL && tjp) { 1534 if (tjp->klass() != field()->holder() || 1535 tjp->offset() != field()->offset_in_bytes()) { 1536 st->print(" != "); 1537 field()->print(); 1538 st->print(" ***"); 1539 } 1540 } 1541 } 1542 1543 void print_alias_types() { 1544 Compile* C = Compile::current(); 1545 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1546 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1547 C->alias_type(idx)->print_on(tty); 1548 tty->cr(); 1549 } 1550 } 1551 #endif 1552 1553 1554 //----------------------------probe_alias_cache-------------------------------- 1555 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1556 intptr_t key = (intptr_t) adr_type; 1557 key ^= key >> logAliasCacheSize; 1558 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1559 } 1560 1561 1562 //-----------------------------grow_alias_types-------------------------------- 1563 void Compile::grow_alias_types() { 1564 const int old_ats = _max_alias_types; // how many before? 1565 const int new_ats = old_ats; // how many more? 1566 const int grow_ats = old_ats+new_ats; // how many now? 1567 _max_alias_types = grow_ats; 1568 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1569 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1570 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1571 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1572 } 1573 1574 1575 //--------------------------------find_alias_type------------------------------ 1576 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1577 if (_AliasLevel == 0) 1578 return alias_type(AliasIdxBot); 1579 1580 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1581 if (ace->_adr_type == adr_type) { 1582 return alias_type(ace->_index); 1583 } 1584 1585 // Handle special cases. 1586 if (adr_type == NULL) return alias_type(AliasIdxTop); 1587 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1588 1589 // Do it the slow way. 1590 const TypePtr* flat = flatten_alias_type(adr_type); 1591 1592 #ifdef ASSERT 1593 assert(flat == flatten_alias_type(flat), "idempotent"); 1594 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1595 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1596 const TypeOopPtr* foop = flat->is_oopptr(); 1597 // Scalarizable allocations have exact klass always. 1598 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1599 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1600 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1601 } 1602 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1603 #endif 1604 1605 int idx = AliasIdxTop; 1606 for (int i = 0; i < num_alias_types(); i++) { 1607 if (alias_type(i)->adr_type() == flat) { 1608 idx = i; 1609 break; 1610 } 1611 } 1612 1613 if (idx == AliasIdxTop) { 1614 if (no_create) return NULL; 1615 // Grow the array if necessary. 1616 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1617 // Add a new alias type. 1618 idx = _num_alias_types++; 1619 _alias_types[idx]->Init(idx, flat); 1620 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1621 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1622 if (flat->isa_instptr()) { 1623 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1624 && flat->is_instptr()->klass() == env()->Class_klass()) 1625 alias_type(idx)->set_rewritable(false); 1626 } 1627 if (flat->isa_aryptr()) { 1628 #ifdef ASSERT 1629 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1630 // (T_BYTE has the weakest alignment and size restrictions...) 1631 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1632 #endif 1633 if (flat->offset() == TypePtr::OffsetBot) { 1634 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1635 } 1636 } 1637 if (flat->isa_klassptr()) { 1638 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1639 alias_type(idx)->set_rewritable(false); 1640 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1641 alias_type(idx)->set_rewritable(false); 1642 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1643 alias_type(idx)->set_rewritable(false); 1644 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1645 alias_type(idx)->set_rewritable(false); 1646 } 1647 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1648 // but the base pointer type is not distinctive enough to identify 1649 // references into JavaThread.) 1650 1651 // Check for final fields. 1652 const TypeInstPtr* tinst = flat->isa_instptr(); 1653 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1654 ciField* field; 1655 if (tinst->const_oop() != NULL && 1656 tinst->klass() == ciEnv::current()->Class_klass() && 1657 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1658 // static field 1659 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1660 field = k->get_field_by_offset(tinst->offset(), true); 1661 } else { 1662 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1663 field = k->get_field_by_offset(tinst->offset(), false); 1664 } 1665 assert(field == NULL || 1666 original_field == NULL || 1667 (field->holder() == original_field->holder() && 1668 field->offset() == original_field->offset() && 1669 field->is_static() == original_field->is_static()), "wrong field?"); 1670 // Set field() and is_rewritable() attributes. 1671 if (field != NULL) alias_type(idx)->set_field(field); 1672 } 1673 } 1674 1675 // Fill the cache for next time. 1676 ace->_adr_type = adr_type; 1677 ace->_index = idx; 1678 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1679 1680 // Might as well try to fill the cache for the flattened version, too. 1681 AliasCacheEntry* face = probe_alias_cache(flat); 1682 if (face->_adr_type == NULL) { 1683 face->_adr_type = flat; 1684 face->_index = idx; 1685 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1686 } 1687 1688 return alias_type(idx); 1689 } 1690 1691 1692 Compile::AliasType* Compile::alias_type(ciField* field) { 1693 const TypeOopPtr* t; 1694 if (field->is_static()) 1695 t = TypeInstPtr::make(field->holder()->java_mirror()); 1696 else 1697 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1698 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1699 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1700 return atp; 1701 } 1702 1703 1704 //------------------------------have_alias_type-------------------------------- 1705 bool Compile::have_alias_type(const TypePtr* adr_type) { 1706 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1707 if (ace->_adr_type == adr_type) { 1708 return true; 1709 } 1710 1711 // Handle special cases. 1712 if (adr_type == NULL) return true; 1713 if (adr_type == TypePtr::BOTTOM) return true; 1714 1715 return find_alias_type(adr_type, true, NULL) != NULL; 1716 } 1717 1718 //-----------------------------must_alias-------------------------------------- 1719 // True if all values of the given address type are in the given alias category. 1720 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1721 if (alias_idx == AliasIdxBot) return true; // the universal category 1722 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1723 if (alias_idx == AliasIdxTop) return false; // the empty category 1724 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1725 1726 // the only remaining possible overlap is identity 1727 int adr_idx = get_alias_index(adr_type); 1728 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1729 assert(adr_idx == alias_idx || 1730 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1731 && adr_type != TypeOopPtr::BOTTOM), 1732 "should not be testing for overlap with an unsafe pointer"); 1733 return adr_idx == alias_idx; 1734 } 1735 1736 //------------------------------can_alias-------------------------------------- 1737 // True if any values of the given address type are in the given alias category. 1738 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1739 if (alias_idx == AliasIdxTop) return false; // the empty category 1740 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1741 if (alias_idx == AliasIdxBot) return true; // the universal category 1742 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1743 1744 // the only remaining possible overlap is identity 1745 int adr_idx = get_alias_index(adr_type); 1746 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1747 return adr_idx == alias_idx; 1748 } 1749 1750 1751 1752 //---------------------------pop_warm_call------------------------------------- 1753 WarmCallInfo* Compile::pop_warm_call() { 1754 WarmCallInfo* wci = _warm_calls; 1755 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1756 return wci; 1757 } 1758 1759 //----------------------------Inline_Warm-------------------------------------- 1760 int Compile::Inline_Warm() { 1761 // If there is room, try to inline some more warm call sites. 1762 // %%% Do a graph index compaction pass when we think we're out of space? 1763 if (!InlineWarmCalls) return 0; 1764 1765 int calls_made_hot = 0; 1766 int room_to_grow = NodeCountInliningCutoff - unique(); 1767 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1768 int amount_grown = 0; 1769 WarmCallInfo* call; 1770 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1771 int est_size = (int)call->size(); 1772 if (est_size > (room_to_grow - amount_grown)) { 1773 // This one won't fit anyway. Get rid of it. 1774 call->make_cold(); 1775 continue; 1776 } 1777 call->make_hot(); 1778 calls_made_hot++; 1779 amount_grown += est_size; 1780 amount_to_grow -= est_size; 1781 } 1782 1783 if (calls_made_hot > 0) set_major_progress(); 1784 return calls_made_hot; 1785 } 1786 1787 1788 //----------------------------Finish_Warm-------------------------------------- 1789 void Compile::Finish_Warm() { 1790 if (!InlineWarmCalls) return; 1791 if (failing()) return; 1792 if (warm_calls() == NULL) return; 1793 1794 // Clean up loose ends, if we are out of space for inlining. 1795 WarmCallInfo* call; 1796 while ((call = pop_warm_call()) != NULL) { 1797 call->make_cold(); 1798 } 1799 } 1800 1801 //---------------------cleanup_loop_predicates----------------------- 1802 // Remove the opaque nodes that protect the predicates so that all unused 1803 // checks and uncommon_traps will be eliminated from the ideal graph 1804 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1805 if (predicate_count()==0) return; 1806 for (int i = predicate_count(); i > 0; i--) { 1807 Node * n = predicate_opaque1_node(i-1); 1808 assert(n->Opcode() == Op_Opaque1, "must be"); 1809 igvn.replace_node(n, n->in(1)); 1810 } 1811 assert(predicate_count()==0, "should be clean!"); 1812 } 1813 1814 // StringOpts and late inlining of string methods 1815 void Compile::inline_string_calls(bool parse_time) { 1816 { 1817 // remove useless nodes to make the usage analysis simpler 1818 ResourceMark rm; 1819 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1820 } 1821 1822 { 1823 ResourceMark rm; 1824 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1825 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1826 print_method(PHASE_AFTER_STRINGOPTS, 3); 1827 } 1828 1829 // now inline anything that we skipped the first time around 1830 if (!parse_time) { 1831 _late_inlines_pos = _late_inlines.length(); 1832 } 1833 1834 while (_string_late_inlines.length() > 0) { 1835 CallGenerator* cg = _string_late_inlines.pop(); 1836 cg->do_late_inline(); 1837 if (failing()) return; 1838 } 1839 _string_late_inlines.trunc_to(0); 1840 } 1841 1842 // Late inlining of boxing methods 1843 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1844 if (_boxing_late_inlines.length() > 0) { 1845 assert(has_boxed_value(), "inconsistent"); 1846 1847 PhaseGVN* gvn = initial_gvn(); 1848 set_inlining_incrementally(true); 1849 1850 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1851 for_igvn()->clear(); 1852 gvn->replace_with(&igvn); 1853 1854 while (_boxing_late_inlines.length() > 0) { 1855 CallGenerator* cg = _boxing_late_inlines.pop(); 1856 cg->do_late_inline(); 1857 if (failing()) return; 1858 } 1859 _boxing_late_inlines.trunc_to(0); 1860 1861 { 1862 ResourceMark rm; 1863 PhaseRemoveUseless pru(gvn, for_igvn()); 1864 } 1865 1866 igvn = PhaseIterGVN(gvn); 1867 igvn.optimize(); 1868 1869 set_inlining_progress(false); 1870 set_inlining_incrementally(false); 1871 } 1872 } 1873 1874 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1875 assert(IncrementalInline, "incremental inlining should be on"); 1876 PhaseGVN* gvn = initial_gvn(); 1877 1878 set_inlining_progress(false); 1879 for_igvn()->clear(); 1880 gvn->replace_with(&igvn); 1881 1882 int i = 0; 1883 1884 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1885 CallGenerator* cg = _late_inlines.at(i); 1886 _late_inlines_pos = i+1; 1887 cg->do_late_inline(); 1888 if (failing()) return; 1889 } 1890 int j = 0; 1891 for (; i < _late_inlines.length(); i++, j++) { 1892 _late_inlines.at_put(j, _late_inlines.at(i)); 1893 } 1894 _late_inlines.trunc_to(j); 1895 1896 { 1897 ResourceMark rm; 1898 PhaseRemoveUseless pru(gvn, for_igvn()); 1899 } 1900 1901 igvn = PhaseIterGVN(gvn); 1902 } 1903 1904 // Perform incremental inlining until bound on number of live nodes is reached 1905 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1906 PhaseGVN* gvn = initial_gvn(); 1907 1908 set_inlining_incrementally(true); 1909 set_inlining_progress(true); 1910 uint low_live_nodes = 0; 1911 1912 while(inlining_progress() && _late_inlines.length() > 0) { 1913 1914 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1915 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1916 // PhaseIdealLoop is expensive so we only try it once we are 1917 // out of loop and we only try it again if the previous helped 1918 // got the number of nodes down significantly 1919 PhaseIdealLoop ideal_loop( igvn, false, true ); 1920 if (failing()) return; 1921 low_live_nodes = live_nodes(); 1922 _major_progress = true; 1923 } 1924 1925 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1926 break; 1927 } 1928 } 1929 1930 inline_incrementally_one(igvn); 1931 1932 if (failing()) return; 1933 1934 igvn.optimize(); 1935 1936 if (failing()) return; 1937 } 1938 1939 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1940 1941 if (_string_late_inlines.length() > 0) { 1942 assert(has_stringbuilder(), "inconsistent"); 1943 for_igvn()->clear(); 1944 initial_gvn()->replace_with(&igvn); 1945 1946 inline_string_calls(false); 1947 1948 if (failing()) return; 1949 1950 { 1951 ResourceMark rm; 1952 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1953 } 1954 1955 igvn = PhaseIterGVN(gvn); 1956 1957 igvn.optimize(); 1958 } 1959 1960 set_inlining_incrementally(false); 1961 } 1962 1963 1964 //------------------------------Optimize--------------------------------------- 1965 // Given a graph, optimize it. 1966 void Compile::Optimize() { 1967 TracePhase t1("optimizer", &_t_optimizer, true); 1968 1969 #ifndef PRODUCT 1970 if (env()->break_at_compile()) { 1971 BREAKPOINT; 1972 } 1973 1974 #endif 1975 1976 ResourceMark rm; 1977 int loop_opts_cnt; 1978 1979 NOT_PRODUCT( verify_graph_edges(); ) 1980 1981 print_method(PHASE_AFTER_PARSING); 1982 1983 { 1984 // Iterative Global Value Numbering, including ideal transforms 1985 // Initialize IterGVN with types and values from parse-time GVN 1986 PhaseIterGVN igvn(initial_gvn()); 1987 { 1988 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1989 igvn.optimize(); 1990 } 1991 1992 print_method(PHASE_ITER_GVN1, 2); 1993 1994 if (failing()) return; 1995 1996 { 1997 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 1998 inline_incrementally(igvn); 1999 } 2000 2001 print_method(PHASE_INCREMENTAL_INLINE, 2); 2002 2003 if (failing()) return; 2004 2005 if (eliminate_boxing()) { 2006 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2007 // Inline valueOf() methods now. 2008 inline_boxing_calls(igvn); 2009 2010 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2011 2012 if (failing()) return; 2013 } 2014 2015 // No more new expensive nodes will be added to the list from here 2016 // so keep only the actual candidates for optimizations. 2017 cleanup_expensive_nodes(igvn); 2018 2019 // Perform escape analysis 2020 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2021 if (has_loops()) { 2022 // Cleanup graph (remove dead nodes). 2023 TracePhase t2("idealLoop", &_t_idealLoop, true); 2024 PhaseIdealLoop ideal_loop( igvn, false, true ); 2025 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2026 if (failing()) return; 2027 } 2028 ConnectionGraph::do_analysis(this, &igvn); 2029 2030 if (failing()) return; 2031 2032 // Optimize out fields loads from scalar replaceable allocations. 2033 igvn.optimize(); 2034 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2035 2036 if (failing()) return; 2037 2038 if (congraph() != NULL && macro_count() > 0) { 2039 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2040 PhaseMacroExpand mexp(igvn); 2041 mexp.eliminate_macro_nodes(); 2042 igvn.set_delay_transform(false); 2043 2044 igvn.optimize(); 2045 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2046 2047 if (failing()) return; 2048 } 2049 } 2050 2051 // Loop transforms on the ideal graph. Range Check Elimination, 2052 // peeling, unrolling, etc. 2053 2054 // Set loop opts counter 2055 loop_opts_cnt = num_loop_opts(); 2056 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2057 { 2058 TracePhase t2("idealLoop", &_t_idealLoop, true); 2059 PhaseIdealLoop ideal_loop( igvn, true ); 2060 loop_opts_cnt--; 2061 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2062 if (failing()) return; 2063 } 2064 // Loop opts pass if partial peeling occurred in previous pass 2065 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2066 TracePhase t3("idealLoop", &_t_idealLoop, true); 2067 PhaseIdealLoop ideal_loop( igvn, false ); 2068 loop_opts_cnt--; 2069 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2070 if (failing()) return; 2071 } 2072 // Loop opts pass for loop-unrolling before CCP 2073 if(major_progress() && (loop_opts_cnt > 0)) { 2074 TracePhase t4("idealLoop", &_t_idealLoop, true); 2075 PhaseIdealLoop ideal_loop( igvn, false ); 2076 loop_opts_cnt--; 2077 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2078 } 2079 if (!failing()) { 2080 // Verify that last round of loop opts produced a valid graph 2081 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2082 PhaseIdealLoop::verify(igvn); 2083 } 2084 } 2085 if (failing()) return; 2086 2087 // Conditional Constant Propagation; 2088 PhaseCCP ccp( &igvn ); 2089 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2090 { 2091 TracePhase t2("ccp", &_t_ccp, true); 2092 ccp.do_transform(); 2093 } 2094 print_method(PHASE_CPP1, 2); 2095 2096 assert( true, "Break here to ccp.dump_old2new_map()"); 2097 2098 // Iterative Global Value Numbering, including ideal transforms 2099 { 2100 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2101 igvn = ccp; 2102 igvn.optimize(); 2103 } 2104 2105 print_method(PHASE_ITER_GVN2, 2); 2106 2107 if (failing()) return; 2108 2109 // Loop transforms on the ideal graph. Range Check Elimination, 2110 // peeling, unrolling, etc. 2111 if(loop_opts_cnt > 0) { 2112 debug_only( int cnt = 0; ); 2113 while(major_progress() && (loop_opts_cnt > 0)) { 2114 TracePhase t2("idealLoop", &_t_idealLoop, true); 2115 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2116 PhaseIdealLoop ideal_loop( igvn, true); 2117 loop_opts_cnt--; 2118 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2119 if (failing()) return; 2120 } 2121 } 2122 2123 { 2124 // Verify that all previous optimizations produced a valid graph 2125 // at least to this point, even if no loop optimizations were done. 2126 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2127 PhaseIdealLoop::verify(igvn); 2128 } 2129 2130 { 2131 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2132 PhaseMacroExpand mex(igvn); 2133 if (mex.expand_macro_nodes()) { 2134 assert(failing(), "must bail out w/ explicit message"); 2135 return; 2136 } 2137 } 2138 2139 } // (End scope of igvn; run destructor if necessary for asserts.) 2140 2141 dump_inlining(); 2142 // A method with only infinite loops has no edges entering loops from root 2143 { 2144 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2145 if (final_graph_reshaping()) { 2146 assert(failing(), "must bail out w/ explicit message"); 2147 return; 2148 } 2149 } 2150 2151 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2152 } 2153 2154 2155 //------------------------------Code_Gen--------------------------------------- 2156 // Given a graph, generate code for it 2157 void Compile::Code_Gen() { 2158 if (failing()) { 2159 return; 2160 } 2161 2162 // Perform instruction selection. You might think we could reclaim Matcher 2163 // memory PDQ, but actually the Matcher is used in generating spill code. 2164 // Internals of the Matcher (including some VectorSets) must remain live 2165 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2166 // set a bit in reclaimed memory. 2167 2168 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2169 // nodes. Mapping is only valid at the root of each matched subtree. 2170 NOT_PRODUCT( verify_graph_edges(); ) 2171 2172 Matcher matcher; 2173 _matcher = &matcher; 2174 { 2175 TracePhase t2("matcher", &_t_matcher, true); 2176 matcher.match(); 2177 } 2178 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2179 // nodes. Mapping is only valid at the root of each matched subtree. 2180 NOT_PRODUCT( verify_graph_edges(); ) 2181 2182 // If you have too many nodes, or if matching has failed, bail out 2183 check_node_count(0, "out of nodes matching instructions"); 2184 if (failing()) { 2185 return; 2186 } 2187 2188 // Build a proper-looking CFG 2189 PhaseCFG cfg(node_arena(), root(), matcher); 2190 _cfg = &cfg; 2191 { 2192 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2193 bool success = cfg.do_global_code_motion(); 2194 if (!success) { 2195 return; 2196 } 2197 2198 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2199 NOT_PRODUCT( verify_graph_edges(); ) 2200 debug_only( cfg.verify(); ) 2201 } 2202 2203 PhaseChaitin regalloc(unique(), cfg, matcher); 2204 _regalloc = ®alloc; 2205 { 2206 TracePhase t2("regalloc", &_t_registerAllocation, true); 2207 // Perform register allocation. After Chaitin, use-def chains are 2208 // no longer accurate (at spill code) and so must be ignored. 2209 // Node->LRG->reg mappings are still accurate. 2210 _regalloc->Register_Allocate(); 2211 2212 // Bail out if the allocator builds too many nodes 2213 if (failing()) { 2214 return; 2215 } 2216 } 2217 2218 // Prior to register allocation we kept empty basic blocks in case the 2219 // the allocator needed a place to spill. After register allocation we 2220 // are not adding any new instructions. If any basic block is empty, we 2221 // can now safely remove it. 2222 { 2223 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2224 cfg.remove_empty_blocks(); 2225 if (do_freq_based_layout()) { 2226 PhaseBlockLayout layout(cfg); 2227 } else { 2228 cfg.set_loop_alignment(); 2229 } 2230 cfg.fixup_flow(); 2231 } 2232 2233 // Apply peephole optimizations 2234 if( OptoPeephole ) { 2235 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2236 PhasePeephole peep( _regalloc, cfg); 2237 peep.do_transform(); 2238 } 2239 2240 // Convert Nodes to instruction bits in a buffer 2241 { 2242 // %%%% workspace merge brought two timers together for one job 2243 TracePhase t2a("output", &_t_output, true); 2244 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2245 Output(); 2246 } 2247 2248 print_method(PHASE_FINAL_CODE); 2249 2250 // He's dead, Jim. 2251 _cfg = (PhaseCFG*)0xdeadbeef; 2252 _regalloc = (PhaseChaitin*)0xdeadbeef; 2253 } 2254 2255 2256 //------------------------------dump_asm--------------------------------------- 2257 // Dump formatted assembly 2258 #ifndef PRODUCT 2259 void Compile::dump_asm(int *pcs, uint pc_limit) { 2260 bool cut_short = false; 2261 tty->print_cr("#"); 2262 tty->print("# "); _tf->dump(); tty->cr(); 2263 tty->print_cr("#"); 2264 2265 // For all blocks 2266 int pc = 0x0; // Program counter 2267 char starts_bundle = ' '; 2268 _regalloc->dump_frame(); 2269 2270 Node *n = NULL; 2271 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2272 if (VMThread::should_terminate()) { 2273 cut_short = true; 2274 break; 2275 } 2276 Block* block = _cfg->get_block(i); 2277 if (block->is_connector() && !Verbose) { 2278 continue; 2279 } 2280 n = block->head(); 2281 if (pcs && n->_idx < pc_limit) { 2282 tty->print("%3.3x ", pcs[n->_idx]); 2283 } else { 2284 tty->print(" "); 2285 } 2286 block->dump_head(_cfg); 2287 if (block->is_connector()) { 2288 tty->print_cr(" # Empty connector block"); 2289 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2290 tty->print_cr(" # Block is sole successor of call"); 2291 } 2292 2293 // For all instructions 2294 Node *delay = NULL; 2295 for (uint j = 0; j < block->number_of_nodes(); j++) { 2296 if (VMThread::should_terminate()) { 2297 cut_short = true; 2298 break; 2299 } 2300 n = block->get_node(j); 2301 if (valid_bundle_info(n)) { 2302 Bundle* bundle = node_bundling(n); 2303 if (bundle->used_in_unconditional_delay()) { 2304 delay = n; 2305 continue; 2306 } 2307 if (bundle->starts_bundle()) { 2308 starts_bundle = '+'; 2309 } 2310 } 2311 2312 if (WizardMode) { 2313 n->dump(); 2314 } 2315 2316 if( !n->is_Region() && // Dont print in the Assembly 2317 !n->is_Phi() && // a few noisely useless nodes 2318 !n->is_Proj() && 2319 !n->is_MachTemp() && 2320 !n->is_SafePointScalarObject() && 2321 !n->is_Catch() && // Would be nice to print exception table targets 2322 !n->is_MergeMem() && // Not very interesting 2323 !n->is_top() && // Debug info table constants 2324 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2325 ) { 2326 if (pcs && n->_idx < pc_limit) 2327 tty->print("%3.3x", pcs[n->_idx]); 2328 else 2329 tty->print(" "); 2330 tty->print(" %c ", starts_bundle); 2331 starts_bundle = ' '; 2332 tty->print("\t"); 2333 n->format(_regalloc, tty); 2334 tty->cr(); 2335 } 2336 2337 // If we have an instruction with a delay slot, and have seen a delay, 2338 // then back up and print it 2339 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2340 assert(delay != NULL, "no unconditional delay instruction"); 2341 if (WizardMode) delay->dump(); 2342 2343 if (node_bundling(delay)->starts_bundle()) 2344 starts_bundle = '+'; 2345 if (pcs && n->_idx < pc_limit) 2346 tty->print("%3.3x", pcs[n->_idx]); 2347 else 2348 tty->print(" "); 2349 tty->print(" %c ", starts_bundle); 2350 starts_bundle = ' '; 2351 tty->print("\t"); 2352 delay->format(_regalloc, tty); 2353 tty->print_cr(""); 2354 delay = NULL; 2355 } 2356 2357 // Dump the exception table as well 2358 if( n->is_Catch() && (Verbose || WizardMode) ) { 2359 // Print the exception table for this offset 2360 _handler_table.print_subtable_for(pc); 2361 } 2362 } 2363 2364 if (pcs && n->_idx < pc_limit) 2365 tty->print_cr("%3.3x", pcs[n->_idx]); 2366 else 2367 tty->print_cr(""); 2368 2369 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2370 2371 } // End of per-block dump 2372 tty->print_cr(""); 2373 2374 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2375 } 2376 #endif 2377 2378 //------------------------------Final_Reshape_Counts--------------------------- 2379 // This class defines counters to help identify when a method 2380 // may/must be executed using hardware with only 24-bit precision. 2381 struct Final_Reshape_Counts : public StackObj { 2382 int _call_count; // count non-inlined 'common' calls 2383 int _float_count; // count float ops requiring 24-bit precision 2384 int _double_count; // count double ops requiring more precision 2385 int _java_call_count; // count non-inlined 'java' calls 2386 int _inner_loop_count; // count loops which need alignment 2387 VectorSet _visited; // Visitation flags 2388 Node_List _tests; // Set of IfNodes & PCTableNodes 2389 2390 Final_Reshape_Counts() : 2391 _call_count(0), _float_count(0), _double_count(0), 2392 _java_call_count(0), _inner_loop_count(0), 2393 _visited( Thread::current()->resource_area() ) { } 2394 2395 void inc_call_count () { _call_count ++; } 2396 void inc_float_count () { _float_count ++; } 2397 void inc_double_count() { _double_count++; } 2398 void inc_java_call_count() { _java_call_count++; } 2399 void inc_inner_loop_count() { _inner_loop_count++; } 2400 2401 int get_call_count () const { return _call_count ; } 2402 int get_float_count () const { return _float_count ; } 2403 int get_double_count() const { return _double_count; } 2404 int get_java_call_count() const { return _java_call_count; } 2405 int get_inner_loop_count() const { return _inner_loop_count; } 2406 }; 2407 2408 #ifdef ASSERT 2409 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2410 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2411 // Make sure the offset goes inside the instance layout. 2412 return k->contains_field_offset(tp->offset()); 2413 // Note that OffsetBot and OffsetTop are very negative. 2414 } 2415 #endif 2416 2417 // Eliminate trivially redundant StoreCMs and accumulate their 2418 // precedence edges. 2419 void Compile::eliminate_redundant_card_marks(Node* n) { 2420 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2421 if (n->in(MemNode::Address)->outcnt() > 1) { 2422 // There are multiple users of the same address so it might be 2423 // possible to eliminate some of the StoreCMs 2424 Node* mem = n->in(MemNode::Memory); 2425 Node* adr = n->in(MemNode::Address); 2426 Node* val = n->in(MemNode::ValueIn); 2427 Node* prev = n; 2428 bool done = false; 2429 // Walk the chain of StoreCMs eliminating ones that match. As 2430 // long as it's a chain of single users then the optimization is 2431 // safe. Eliminating partially redundant StoreCMs would require 2432 // cloning copies down the other paths. 2433 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2434 if (adr == mem->in(MemNode::Address) && 2435 val == mem->in(MemNode::ValueIn)) { 2436 // redundant StoreCM 2437 if (mem->req() > MemNode::OopStore) { 2438 // Hasn't been processed by this code yet. 2439 n->add_prec(mem->in(MemNode::OopStore)); 2440 } else { 2441 // Already converted to precedence edge 2442 for (uint i = mem->req(); i < mem->len(); i++) { 2443 // Accumulate any precedence edges 2444 if (mem->in(i) != NULL) { 2445 n->add_prec(mem->in(i)); 2446 } 2447 } 2448 // Everything above this point has been processed. 2449 done = true; 2450 } 2451 // Eliminate the previous StoreCM 2452 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2453 assert(mem->outcnt() == 0, "should be dead"); 2454 mem->disconnect_inputs(NULL, this); 2455 } else { 2456 prev = mem; 2457 } 2458 mem = prev->in(MemNode::Memory); 2459 } 2460 } 2461 } 2462 2463 //------------------------------final_graph_reshaping_impl---------------------- 2464 // Implement items 1-5 from final_graph_reshaping below. 2465 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2466 2467 if ( n->outcnt() == 0 ) return; // dead node 2468 uint nop = n->Opcode(); 2469 2470 // Check for 2-input instruction with "last use" on right input. 2471 // Swap to left input. Implements item (2). 2472 if( n->req() == 3 && // two-input instruction 2473 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2474 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2475 n->in(2)->outcnt() == 1 &&// right use IS a last use 2476 !n->in(2)->is_Con() ) { // right use is not a constant 2477 // Check for commutative opcode 2478 switch( nop ) { 2479 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2480 case Op_MaxI: case Op_MinI: 2481 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2482 case Op_AndL: case Op_XorL: case Op_OrL: 2483 case Op_AndI: case Op_XorI: case Op_OrI: { 2484 // Move "last use" input to left by swapping inputs 2485 n->swap_edges(1, 2); 2486 break; 2487 } 2488 default: 2489 break; 2490 } 2491 } 2492 2493 #ifdef ASSERT 2494 if( n->is_Mem() ) { 2495 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2496 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2497 // oop will be recorded in oop map if load crosses safepoint 2498 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2499 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2500 "raw memory operations should have control edge"); 2501 } 2502 #endif 2503 // Count FPU ops and common calls, implements item (3) 2504 switch( nop ) { 2505 // Count all float operations that may use FPU 2506 case Op_AddF: 2507 case Op_SubF: 2508 case Op_MulF: 2509 case Op_DivF: 2510 case Op_NegF: 2511 case Op_ModF: 2512 case Op_ConvI2F: 2513 case Op_ConF: 2514 case Op_CmpF: 2515 case Op_CmpF3: 2516 // case Op_ConvL2F: // longs are split into 32-bit halves 2517 frc.inc_float_count(); 2518 break; 2519 2520 case Op_ConvF2D: 2521 case Op_ConvD2F: 2522 frc.inc_float_count(); 2523 frc.inc_double_count(); 2524 break; 2525 2526 // Count all double operations that may use FPU 2527 case Op_AddD: 2528 case Op_SubD: 2529 case Op_MulD: 2530 case Op_DivD: 2531 case Op_NegD: 2532 case Op_ModD: 2533 case Op_ConvI2D: 2534 case Op_ConvD2I: 2535 // case Op_ConvL2D: // handled by leaf call 2536 // case Op_ConvD2L: // handled by leaf call 2537 case Op_ConD: 2538 case Op_CmpD: 2539 case Op_CmpD3: 2540 frc.inc_double_count(); 2541 break; 2542 case Op_Opaque1: // Remove Opaque Nodes before matching 2543 case Op_Opaque2: // Remove Opaque Nodes before matching 2544 n->subsume_by(n->in(1), this); 2545 break; 2546 case Op_CallStaticJava: 2547 case Op_CallJava: 2548 case Op_CallDynamicJava: 2549 frc.inc_java_call_count(); // Count java call site; 2550 case Op_CallRuntime: 2551 case Op_CallLeaf: 2552 case Op_CallLeafNoFP: { 2553 assert( n->is_Call(), "" ); 2554 CallNode *call = n->as_Call(); 2555 // Count call sites where the FP mode bit would have to be flipped. 2556 // Do not count uncommon runtime calls: 2557 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2558 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2559 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2560 frc.inc_call_count(); // Count the call site 2561 } else { // See if uncommon argument is shared 2562 Node *n = call->in(TypeFunc::Parms); 2563 int nop = n->Opcode(); 2564 // Clone shared simple arguments to uncommon calls, item (1). 2565 if( n->outcnt() > 1 && 2566 !n->is_Proj() && 2567 nop != Op_CreateEx && 2568 nop != Op_CheckCastPP && 2569 nop != Op_DecodeN && 2570 nop != Op_DecodeNKlass && 2571 !n->is_Mem() ) { 2572 Node *x = n->clone(); 2573 call->set_req( TypeFunc::Parms, x ); 2574 } 2575 } 2576 break; 2577 } 2578 2579 case Op_StoreD: 2580 case Op_LoadD: 2581 case Op_LoadD_unaligned: 2582 frc.inc_double_count(); 2583 goto handle_mem; 2584 case Op_StoreF: 2585 case Op_LoadF: 2586 frc.inc_float_count(); 2587 goto handle_mem; 2588 2589 case Op_StoreCM: 2590 { 2591 // Convert OopStore dependence into precedence edge 2592 Node* prec = n->in(MemNode::OopStore); 2593 n->del_req(MemNode::OopStore); 2594 n->add_prec(prec); 2595 eliminate_redundant_card_marks(n); 2596 } 2597 2598 // fall through 2599 2600 case Op_StoreB: 2601 case Op_StoreC: 2602 case Op_StorePConditional: 2603 case Op_StoreI: 2604 case Op_StoreL: 2605 case Op_StoreIConditional: 2606 case Op_StoreLConditional: 2607 case Op_CompareAndSwapI: 2608 case Op_CompareAndSwapL: 2609 case Op_CompareAndSwapP: 2610 case Op_CompareAndSwapN: 2611 case Op_GetAndAddI: 2612 case Op_GetAndAddL: 2613 case Op_GetAndSetI: 2614 case Op_GetAndSetL: 2615 case Op_GetAndSetP: 2616 case Op_GetAndSetN: 2617 case Op_StoreP: 2618 case Op_StoreN: 2619 case Op_StoreNKlass: 2620 case Op_LoadB: 2621 case Op_LoadUB: 2622 case Op_LoadUS: 2623 case Op_LoadI: 2624 case Op_LoadKlass: 2625 case Op_LoadNKlass: 2626 case Op_LoadL: 2627 case Op_LoadL_unaligned: 2628 case Op_LoadPLocked: 2629 case Op_LoadP: 2630 case Op_LoadN: 2631 case Op_LoadRange: 2632 case Op_LoadS: { 2633 handle_mem: 2634 #ifdef ASSERT 2635 if( VerifyOptoOopOffsets ) { 2636 assert( n->is_Mem(), "" ); 2637 MemNode *mem = (MemNode*)n; 2638 // Check to see if address types have grounded out somehow. 2639 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2640 assert( !tp || oop_offset_is_sane(tp), "" ); 2641 } 2642 #endif 2643 break; 2644 } 2645 2646 case Op_AddP: { // Assert sane base pointers 2647 Node *addp = n->in(AddPNode::Address); 2648 assert( !addp->is_AddP() || 2649 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2650 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2651 "Base pointers must match" ); 2652 #ifdef _LP64 2653 if ((UseCompressedOops || UseCompressedClassPointers) && 2654 addp->Opcode() == Op_ConP && 2655 addp == n->in(AddPNode::Base) && 2656 n->in(AddPNode::Offset)->is_Con()) { 2657 // Use addressing with narrow klass to load with offset on x86. 2658 // On sparc loading 32-bits constant and decoding it have less 2659 // instructions (4) then load 64-bits constant (7). 2660 // Do this transformation here since IGVN will convert ConN back to ConP. 2661 const Type* t = addp->bottom_type(); 2662 if (t->isa_oopptr() || t->isa_klassptr()) { 2663 Node* nn = NULL; 2664 2665 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2666 2667 // Look for existing ConN node of the same exact type. 2668 Node* r = root(); 2669 uint cnt = r->outcnt(); 2670 for (uint i = 0; i < cnt; i++) { 2671 Node* m = r->raw_out(i); 2672 if (m!= NULL && m->Opcode() == op && 2673 m->bottom_type()->make_ptr() == t) { 2674 nn = m; 2675 break; 2676 } 2677 } 2678 if (nn != NULL) { 2679 // Decode a narrow oop to match address 2680 // [R12 + narrow_oop_reg<<3 + offset] 2681 if (t->isa_oopptr()) { 2682 nn = new (this) DecodeNNode(nn, t); 2683 } else { 2684 nn = new (this) DecodeNKlassNode(nn, t); 2685 } 2686 n->set_req(AddPNode::Base, nn); 2687 n->set_req(AddPNode::Address, nn); 2688 if (addp->outcnt() == 0) { 2689 addp->disconnect_inputs(NULL, this); 2690 } 2691 } 2692 } 2693 } 2694 #endif 2695 break; 2696 } 2697 2698 #ifdef _LP64 2699 case Op_CastPP: 2700 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2701 Node* in1 = n->in(1); 2702 const Type* t = n->bottom_type(); 2703 Node* new_in1 = in1->clone(); 2704 new_in1->as_DecodeN()->set_type(t); 2705 2706 if (!Matcher::narrow_oop_use_complex_address()) { 2707 // 2708 // x86, ARM and friends can handle 2 adds in addressing mode 2709 // and Matcher can fold a DecodeN node into address by using 2710 // a narrow oop directly and do implicit NULL check in address: 2711 // 2712 // [R12 + narrow_oop_reg<<3 + offset] 2713 // NullCheck narrow_oop_reg 2714 // 2715 // On other platforms (Sparc) we have to keep new DecodeN node and 2716 // use it to do implicit NULL check in address: 2717 // 2718 // decode_not_null narrow_oop_reg, base_reg 2719 // [base_reg + offset] 2720 // NullCheck base_reg 2721 // 2722 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2723 // to keep the information to which NULL check the new DecodeN node 2724 // corresponds to use it as value in implicit_null_check(). 2725 // 2726 new_in1->set_req(0, n->in(0)); 2727 } 2728 2729 n->subsume_by(new_in1, this); 2730 if (in1->outcnt() == 0) { 2731 in1->disconnect_inputs(NULL, this); 2732 } 2733 } 2734 break; 2735 2736 case Op_CmpP: 2737 // Do this transformation here to preserve CmpPNode::sub() and 2738 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2739 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2740 Node* in1 = n->in(1); 2741 Node* in2 = n->in(2); 2742 if (!in1->is_DecodeNarrowPtr()) { 2743 in2 = in1; 2744 in1 = n->in(2); 2745 } 2746 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2747 2748 Node* new_in2 = NULL; 2749 if (in2->is_DecodeNarrowPtr()) { 2750 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2751 new_in2 = in2->in(1); 2752 } else if (in2->Opcode() == Op_ConP) { 2753 const Type* t = in2->bottom_type(); 2754 if (t == TypePtr::NULL_PTR) { 2755 assert(in1->is_DecodeN(), "compare klass to null?"); 2756 // Don't convert CmpP null check into CmpN if compressed 2757 // oops implicit null check is not generated. 2758 // This will allow to generate normal oop implicit null check. 2759 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2760 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2761 // 2762 // This transformation together with CastPP transformation above 2763 // will generated code for implicit NULL checks for compressed oops. 2764 // 2765 // The original code after Optimize() 2766 // 2767 // LoadN memory, narrow_oop_reg 2768 // decode narrow_oop_reg, base_reg 2769 // CmpP base_reg, NULL 2770 // CastPP base_reg // NotNull 2771 // Load [base_reg + offset], val_reg 2772 // 2773 // after these transformations will be 2774 // 2775 // LoadN memory, narrow_oop_reg 2776 // CmpN narrow_oop_reg, NULL 2777 // decode_not_null narrow_oop_reg, base_reg 2778 // Load [base_reg + offset], val_reg 2779 // 2780 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2781 // since narrow oops can be used in debug info now (see the code in 2782 // final_graph_reshaping_walk()). 2783 // 2784 // At the end the code will be matched to 2785 // on x86: 2786 // 2787 // Load_narrow_oop memory, narrow_oop_reg 2788 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2789 // NullCheck narrow_oop_reg 2790 // 2791 // and on sparc: 2792 // 2793 // Load_narrow_oop memory, narrow_oop_reg 2794 // decode_not_null narrow_oop_reg, base_reg 2795 // Load [base_reg + offset], val_reg 2796 // NullCheck base_reg 2797 // 2798 } else if (t->isa_oopptr()) { 2799 new_in2 = ConNode::make(this, t->make_narrowoop()); 2800 } else if (t->isa_klassptr()) { 2801 new_in2 = ConNode::make(this, t->make_narrowklass()); 2802 } 2803 } 2804 if (new_in2 != NULL) { 2805 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 2806 n->subsume_by(cmpN, this); 2807 if (in1->outcnt() == 0) { 2808 in1->disconnect_inputs(NULL, this); 2809 } 2810 if (in2->outcnt() == 0) { 2811 in2->disconnect_inputs(NULL, this); 2812 } 2813 } 2814 } 2815 break; 2816 2817 case Op_DecodeN: 2818 case Op_DecodeNKlass: 2819 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2820 // DecodeN could be pinned when it can't be fold into 2821 // an address expression, see the code for Op_CastPP above. 2822 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2823 break; 2824 2825 case Op_EncodeP: 2826 case Op_EncodePKlass: { 2827 Node* in1 = n->in(1); 2828 if (in1->is_DecodeNarrowPtr()) { 2829 n->subsume_by(in1->in(1), this); 2830 } else if (in1->Opcode() == Op_ConP) { 2831 const Type* t = in1->bottom_type(); 2832 if (t == TypePtr::NULL_PTR) { 2833 assert(t->isa_oopptr(), "null klass?"); 2834 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2835 } else if (t->isa_oopptr()) { 2836 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2837 } else if (t->isa_klassptr()) { 2838 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2839 } 2840 } 2841 if (in1->outcnt() == 0) { 2842 in1->disconnect_inputs(NULL, this); 2843 } 2844 break; 2845 } 2846 2847 case Op_Proj: { 2848 if (OptimizeStringConcat) { 2849 ProjNode* p = n->as_Proj(); 2850 if (p->_is_io_use) { 2851 // Separate projections were used for the exception path which 2852 // are normally removed by a late inline. If it wasn't inlined 2853 // then they will hang around and should just be replaced with 2854 // the original one. 2855 Node* proj = NULL; 2856 // Replace with just one 2857 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2858 Node *use = i.get(); 2859 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2860 proj = use; 2861 break; 2862 } 2863 } 2864 assert(proj != NULL, "must be found"); 2865 p->subsume_by(proj, this); 2866 } 2867 } 2868 break; 2869 } 2870 2871 case Op_Phi: 2872 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2873 // The EncodeP optimization may create Phi with the same edges 2874 // for all paths. It is not handled well by Register Allocator. 2875 Node* unique_in = n->in(1); 2876 assert(unique_in != NULL, ""); 2877 uint cnt = n->req(); 2878 for (uint i = 2; i < cnt; i++) { 2879 Node* m = n->in(i); 2880 assert(m != NULL, ""); 2881 if (unique_in != m) 2882 unique_in = NULL; 2883 } 2884 if (unique_in != NULL) { 2885 n->subsume_by(unique_in, this); 2886 } 2887 } 2888 break; 2889 2890 #endif 2891 2892 case Op_ModI: 2893 if (UseDivMod) { 2894 // Check if a%b and a/b both exist 2895 Node* d = n->find_similar(Op_DivI); 2896 if (d) { 2897 // Replace them with a fused divmod if supported 2898 if (Matcher::has_match_rule(Op_DivModI)) { 2899 DivModINode* divmod = DivModINode::make(this, n); 2900 d->subsume_by(divmod->div_proj(), this); 2901 n->subsume_by(divmod->mod_proj(), this); 2902 } else { 2903 // replace a%b with a-((a/b)*b) 2904 Node* mult = new (this) MulINode(d, d->in(2)); 2905 Node* sub = new (this) SubINode(d->in(1), mult); 2906 n->subsume_by(sub, this); 2907 } 2908 } 2909 } 2910 break; 2911 2912 case Op_ModL: 2913 if (UseDivMod) { 2914 // Check if a%b and a/b both exist 2915 Node* d = n->find_similar(Op_DivL); 2916 if (d) { 2917 // Replace them with a fused divmod if supported 2918 if (Matcher::has_match_rule(Op_DivModL)) { 2919 DivModLNode* divmod = DivModLNode::make(this, n); 2920 d->subsume_by(divmod->div_proj(), this); 2921 n->subsume_by(divmod->mod_proj(), this); 2922 } else { 2923 // replace a%b with a-((a/b)*b) 2924 Node* mult = new (this) MulLNode(d, d->in(2)); 2925 Node* sub = new (this) SubLNode(d->in(1), mult); 2926 n->subsume_by(sub, this); 2927 } 2928 } 2929 } 2930 break; 2931 2932 case Op_LoadVector: 2933 case Op_StoreVector: 2934 break; 2935 2936 case Op_PackB: 2937 case Op_PackS: 2938 case Op_PackI: 2939 case Op_PackF: 2940 case Op_PackL: 2941 case Op_PackD: 2942 if (n->req()-1 > 2) { 2943 // Replace many operand PackNodes with a binary tree for matching 2944 PackNode* p = (PackNode*) n; 2945 Node* btp = p->binary_tree_pack(this, 1, n->req()); 2946 n->subsume_by(btp, this); 2947 } 2948 break; 2949 case Op_Loop: 2950 case Op_CountedLoop: 2951 if (n->as_Loop()->is_inner_loop()) { 2952 frc.inc_inner_loop_count(); 2953 } 2954 break; 2955 case Op_LShiftI: 2956 case Op_RShiftI: 2957 case Op_URShiftI: 2958 case Op_LShiftL: 2959 case Op_RShiftL: 2960 case Op_URShiftL: 2961 if (Matcher::need_masked_shift_count) { 2962 // The cpu's shift instructions don't restrict the count to the 2963 // lower 5/6 bits. We need to do the masking ourselves. 2964 Node* in2 = n->in(2); 2965 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 2966 const TypeInt* t = in2->find_int_type(); 2967 if (t != NULL && t->is_con()) { 2968 juint shift = t->get_con(); 2969 if (shift > mask) { // Unsigned cmp 2970 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 2971 } 2972 } else { 2973 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 2974 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 2975 n->set_req(2, shift); 2976 } 2977 } 2978 if (in2->outcnt() == 0) { // Remove dead node 2979 in2->disconnect_inputs(NULL, this); 2980 } 2981 } 2982 break; 2983 case Op_MemBarStoreStore: 2984 case Op_MemBarRelease: 2985 // Break the link with AllocateNode: it is no longer useful and 2986 // confuses register allocation. 2987 if (n->req() > MemBarNode::Precedent) { 2988 n->set_req(MemBarNode::Precedent, top()); 2989 } 2990 break; 2991 default: 2992 assert( !n->is_Call(), "" ); 2993 assert( !n->is_Mem(), "" ); 2994 break; 2995 } 2996 2997 // Collect CFG split points 2998 if (n->is_MultiBranch()) 2999 frc._tests.push(n); 3000 } 3001 3002 //------------------------------final_graph_reshaping_walk--------------------- 3003 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3004 // requires that the walk visits a node's inputs before visiting the node. 3005 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3006 ResourceArea *area = Thread::current()->resource_area(); 3007 Unique_Node_List sfpt(area); 3008 3009 frc._visited.set(root->_idx); // first, mark node as visited 3010 uint cnt = root->req(); 3011 Node *n = root; 3012 uint i = 0; 3013 while (true) { 3014 if (i < cnt) { 3015 // Place all non-visited non-null inputs onto stack 3016 Node* m = n->in(i); 3017 ++i; 3018 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3019 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 3020 sfpt.push(m); 3021 cnt = m->req(); 3022 nstack.push(n, i); // put on stack parent and next input's index 3023 n = m; 3024 i = 0; 3025 } 3026 } else { 3027 // Now do post-visit work 3028 final_graph_reshaping_impl( n, frc ); 3029 if (nstack.is_empty()) 3030 break; // finished 3031 n = nstack.node(); // Get node from stack 3032 cnt = n->req(); 3033 i = nstack.index(); 3034 nstack.pop(); // Shift to the next node on stack 3035 } 3036 } 3037 3038 // Skip next transformation if compressed oops are not used. 3039 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3040 (!UseCompressedOops && !UseCompressedClassPointers)) 3041 return; 3042 3043 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3044 // It could be done for an uncommon traps or any safepoints/calls 3045 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3046 while (sfpt.size() > 0) { 3047 n = sfpt.pop(); 3048 JVMState *jvms = n->as_SafePoint()->jvms(); 3049 assert(jvms != NULL, "sanity"); 3050 int start = jvms->debug_start(); 3051 int end = n->req(); 3052 bool is_uncommon = (n->is_CallStaticJava() && 3053 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3054 for (int j = start; j < end; j++) { 3055 Node* in = n->in(j); 3056 if (in->is_DecodeNarrowPtr()) { 3057 bool safe_to_skip = true; 3058 if (!is_uncommon ) { 3059 // Is it safe to skip? 3060 for (uint i = 0; i < in->outcnt(); i++) { 3061 Node* u = in->raw_out(i); 3062 if (!u->is_SafePoint() || 3063 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3064 safe_to_skip = false; 3065 } 3066 } 3067 } 3068 if (safe_to_skip) { 3069 n->set_req(j, in->in(1)); 3070 } 3071 if (in->outcnt() == 0) { 3072 in->disconnect_inputs(NULL, this); 3073 } 3074 } 3075 } 3076 } 3077 } 3078 3079 //------------------------------final_graph_reshaping-------------------------- 3080 // Final Graph Reshaping. 3081 // 3082 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3083 // and not commoned up and forced early. Must come after regular 3084 // optimizations to avoid GVN undoing the cloning. Clone constant 3085 // inputs to Loop Phis; these will be split by the allocator anyways. 3086 // Remove Opaque nodes. 3087 // (2) Move last-uses by commutative operations to the left input to encourage 3088 // Intel update-in-place two-address operations and better register usage 3089 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3090 // calls canonicalizing them back. 3091 // (3) Count the number of double-precision FP ops, single-precision FP ops 3092 // and call sites. On Intel, we can get correct rounding either by 3093 // forcing singles to memory (requires extra stores and loads after each 3094 // FP bytecode) or we can set a rounding mode bit (requires setting and 3095 // clearing the mode bit around call sites). The mode bit is only used 3096 // if the relative frequency of single FP ops to calls is low enough. 3097 // This is a key transform for SPEC mpeg_audio. 3098 // (4) Detect infinite loops; blobs of code reachable from above but not 3099 // below. Several of the Code_Gen algorithms fail on such code shapes, 3100 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3101 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3102 // Detection is by looking for IfNodes where only 1 projection is 3103 // reachable from below or CatchNodes missing some targets. 3104 // (5) Assert for insane oop offsets in debug mode. 3105 3106 bool Compile::final_graph_reshaping() { 3107 // an infinite loop may have been eliminated by the optimizer, 3108 // in which case the graph will be empty. 3109 if (root()->req() == 1) { 3110 record_method_not_compilable("trivial infinite loop"); 3111 return true; 3112 } 3113 3114 // Expensive nodes have their control input set to prevent the GVN 3115 // from freely commoning them. There's no GVN beyond this point so 3116 // no need to keep the control input. We want the expensive nodes to 3117 // be freely moved to the least frequent code path by gcm. 3118 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3119 for (int i = 0; i < expensive_count(); i++) { 3120 _expensive_nodes->at(i)->set_req(0, NULL); 3121 } 3122 3123 Final_Reshape_Counts frc; 3124 3125 // Visit everybody reachable! 3126 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3127 Node_Stack nstack(unique() >> 1); 3128 final_graph_reshaping_walk(nstack, root(), frc); 3129 3130 // Check for unreachable (from below) code (i.e., infinite loops). 3131 for( uint i = 0; i < frc._tests.size(); i++ ) { 3132 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3133 // Get number of CFG targets. 3134 // Note that PCTables include exception targets after calls. 3135 uint required_outcnt = n->required_outcnt(); 3136 if (n->outcnt() != required_outcnt) { 3137 // Check for a few special cases. Rethrow Nodes never take the 3138 // 'fall-thru' path, so expected kids is 1 less. 3139 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3140 if (n->in(0)->in(0)->is_Call()) { 3141 CallNode *call = n->in(0)->in(0)->as_Call(); 3142 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3143 required_outcnt--; // Rethrow always has 1 less kid 3144 } else if (call->req() > TypeFunc::Parms && 3145 call->is_CallDynamicJava()) { 3146 // Check for null receiver. In such case, the optimizer has 3147 // detected that the virtual call will always result in a null 3148 // pointer exception. The fall-through projection of this CatchNode 3149 // will not be populated. 3150 Node *arg0 = call->in(TypeFunc::Parms); 3151 if (arg0->is_Type() && 3152 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3153 required_outcnt--; 3154 } 3155 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3156 call->req() > TypeFunc::Parms+1 && 3157 call->is_CallStaticJava()) { 3158 // Check for negative array length. In such case, the optimizer has 3159 // detected that the allocation attempt will always result in an 3160 // exception. There is no fall-through projection of this CatchNode . 3161 Node *arg1 = call->in(TypeFunc::Parms+1); 3162 if (arg1->is_Type() && 3163 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3164 required_outcnt--; 3165 } 3166 } 3167 } 3168 } 3169 // Recheck with a better notion of 'required_outcnt' 3170 if (n->outcnt() != required_outcnt) { 3171 record_method_not_compilable("malformed control flow"); 3172 return true; // Not all targets reachable! 3173 } 3174 } 3175 // Check that I actually visited all kids. Unreached kids 3176 // must be infinite loops. 3177 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3178 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3179 record_method_not_compilable("infinite loop"); 3180 return true; // Found unvisited kid; must be unreach 3181 } 3182 } 3183 3184 // If original bytecodes contained a mixture of floats and doubles 3185 // check if the optimizer has made it homogenous, item (3). 3186 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3187 frc.get_float_count() > 32 && 3188 frc.get_double_count() == 0 && 3189 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3190 set_24_bit_selection_and_mode( false, true ); 3191 } 3192 3193 set_java_calls(frc.get_java_call_count()); 3194 set_inner_loops(frc.get_inner_loop_count()); 3195 3196 // No infinite loops, no reason to bail out. 3197 return false; 3198 } 3199 3200 //-----------------------------too_many_traps---------------------------------- 3201 // Report if there are too many traps at the current method and bci. 3202 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3203 bool Compile::too_many_traps(ciMethod* method, 3204 int bci, 3205 Deoptimization::DeoptReason reason) { 3206 ciMethodData* md = method->method_data(); 3207 if (md->is_empty()) { 3208 // Assume the trap has not occurred, or that it occurred only 3209 // because of a transient condition during start-up in the interpreter. 3210 return false; 3211 } 3212 if (md->has_trap_at(bci, reason) != 0) { 3213 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3214 // Also, if there are multiple reasons, or if there is no per-BCI record, 3215 // assume the worst. 3216 if (log()) 3217 log()->elem("observe trap='%s' count='%d'", 3218 Deoptimization::trap_reason_name(reason), 3219 md->trap_count(reason)); 3220 return true; 3221 } else { 3222 // Ignore method/bci and see if there have been too many globally. 3223 return too_many_traps(reason, md); 3224 } 3225 } 3226 3227 // Less-accurate variant which does not require a method and bci. 3228 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3229 ciMethodData* logmd) { 3230 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 3231 // Too many traps globally. 3232 // Note that we use cumulative trap_count, not just md->trap_count. 3233 if (log()) { 3234 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3235 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3236 Deoptimization::trap_reason_name(reason), 3237 mcount, trap_count(reason)); 3238 } 3239 return true; 3240 } else { 3241 // The coast is clear. 3242 return false; 3243 } 3244 } 3245 3246 //--------------------------too_many_recompiles-------------------------------- 3247 // Report if there are too many recompiles at the current method and bci. 3248 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3249 // Is not eager to return true, since this will cause the compiler to use 3250 // Action_none for a trap point, to avoid too many recompilations. 3251 bool Compile::too_many_recompiles(ciMethod* method, 3252 int bci, 3253 Deoptimization::DeoptReason reason) { 3254 ciMethodData* md = method->method_data(); 3255 if (md->is_empty()) { 3256 // Assume the trap has not occurred, or that it occurred only 3257 // because of a transient condition during start-up in the interpreter. 3258 return false; 3259 } 3260 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3261 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3262 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3263 Deoptimization::DeoptReason per_bc_reason 3264 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3265 if ((per_bc_reason == Deoptimization::Reason_none 3266 || md->has_trap_at(bci, reason) != 0) 3267 // The trap frequency measure we care about is the recompile count: 3268 && md->trap_recompiled_at(bci) 3269 && md->overflow_recompile_count() >= bc_cutoff) { 3270 // Do not emit a trap here if it has already caused recompilations. 3271 // Also, if there are multiple reasons, or if there is no per-BCI record, 3272 // assume the worst. 3273 if (log()) 3274 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3275 Deoptimization::trap_reason_name(reason), 3276 md->trap_count(reason), 3277 md->overflow_recompile_count()); 3278 return true; 3279 } else if (trap_count(reason) != 0 3280 && decompile_count() >= m_cutoff) { 3281 // Too many recompiles globally, and we have seen this sort of trap. 3282 // Use cumulative decompile_count, not just md->decompile_count. 3283 if (log()) 3284 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3285 Deoptimization::trap_reason_name(reason), 3286 md->trap_count(reason), trap_count(reason), 3287 md->decompile_count(), decompile_count()); 3288 return true; 3289 } else { 3290 // The coast is clear. 3291 return false; 3292 } 3293 } 3294 3295 3296 #ifndef PRODUCT 3297 //------------------------------verify_graph_edges--------------------------- 3298 // Walk the Graph and verify that there is a one-to-one correspondence 3299 // between Use-Def edges and Def-Use edges in the graph. 3300 void Compile::verify_graph_edges(bool no_dead_code) { 3301 if (VerifyGraphEdges) { 3302 ResourceArea *area = Thread::current()->resource_area(); 3303 Unique_Node_List visited(area); 3304 // Call recursive graph walk to check edges 3305 _root->verify_edges(visited); 3306 if (no_dead_code) { 3307 // Now make sure that no visited node is used by an unvisited node. 3308 bool dead_nodes = 0; 3309 Unique_Node_List checked(area); 3310 while (visited.size() > 0) { 3311 Node* n = visited.pop(); 3312 checked.push(n); 3313 for (uint i = 0; i < n->outcnt(); i++) { 3314 Node* use = n->raw_out(i); 3315 if (checked.member(use)) continue; // already checked 3316 if (visited.member(use)) continue; // already in the graph 3317 if (use->is_Con()) continue; // a dead ConNode is OK 3318 // At this point, we have found a dead node which is DU-reachable. 3319 if (dead_nodes++ == 0) 3320 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3321 use->dump(2); 3322 tty->print_cr("---"); 3323 checked.push(use); // No repeats; pretend it is now checked. 3324 } 3325 } 3326 assert(dead_nodes == 0, "using nodes must be reachable from root"); 3327 } 3328 } 3329 } 3330 #endif 3331 3332 // The Compile object keeps track of failure reasons separately from the ciEnv. 3333 // This is required because there is not quite a 1-1 relation between the 3334 // ciEnv and its compilation task and the Compile object. Note that one 3335 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3336 // to backtrack and retry without subsuming loads. Other than this backtracking 3337 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3338 // by the logic in C2Compiler. 3339 void Compile::record_failure(const char* reason) { 3340 if (log() != NULL) { 3341 log()->elem("failure reason='%s' phase='compile'", reason); 3342 } 3343 if (_failure_reason == NULL) { 3344 // Record the first failure reason. 3345 _failure_reason = reason; 3346 } 3347 3348 EventCompilerFailure event; 3349 if (event.should_commit()) { 3350 event.set_compileID(Compile::compile_id()); 3351 event.set_failure(reason); 3352 event.commit(); 3353 } 3354 3355 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3356 C->print_method(PHASE_FAILURE); 3357 } 3358 _root = NULL; // flush the graph, too 3359 } 3360 3361 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3362 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3363 _phase_name(name), _dolog(dolog) 3364 { 3365 if (dolog) { 3366 C = Compile::current(); 3367 _log = C->log(); 3368 } else { 3369 C = NULL; 3370 _log = NULL; 3371 } 3372 if (_log != NULL) { 3373 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3374 _log->stamp(); 3375 _log->end_head(); 3376 } 3377 } 3378 3379 Compile::TracePhase::~TracePhase() { 3380 3381 C = Compile::current(); 3382 if (_dolog) { 3383 _log = C->log(); 3384 } else { 3385 _log = NULL; 3386 } 3387 3388 #ifdef ASSERT 3389 if (PrintIdealNodeCount) { 3390 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3391 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3392 } 3393 3394 if (VerifyIdealNodeCount) { 3395 Compile::current()->print_missing_nodes(); 3396 } 3397 #endif 3398 3399 if (_log != NULL) { 3400 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3401 } 3402 } 3403 3404 //============================================================================= 3405 // Two Constant's are equal when the type and the value are equal. 3406 bool Compile::Constant::operator==(const Constant& other) { 3407 if (type() != other.type() ) return false; 3408 if (can_be_reused() != other.can_be_reused()) return false; 3409 // For floating point values we compare the bit pattern. 3410 switch (type()) { 3411 case T_FLOAT: return (_v._value.i == other._v._value.i); 3412 case T_LONG: 3413 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3414 case T_OBJECT: 3415 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3416 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3417 case T_METADATA: return (_v._metadata == other._v._metadata); 3418 default: ShouldNotReachHere(); 3419 } 3420 return false; 3421 } 3422 3423 static int type_to_size_in_bytes(BasicType t) { 3424 switch (t) { 3425 case T_LONG: return sizeof(jlong ); 3426 case T_FLOAT: return sizeof(jfloat ); 3427 case T_DOUBLE: return sizeof(jdouble); 3428 case T_METADATA: return sizeof(Metadata*); 3429 // We use T_VOID as marker for jump-table entries (labels) which 3430 // need an internal word relocation. 3431 case T_VOID: 3432 case T_ADDRESS: 3433 case T_OBJECT: return sizeof(jobject); 3434 } 3435 3436 ShouldNotReachHere(); 3437 return -1; 3438 } 3439 3440 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3441 // sort descending 3442 if (a->freq() > b->freq()) return -1; 3443 if (a->freq() < b->freq()) return 1; 3444 return 0; 3445 } 3446 3447 void Compile::ConstantTable::calculate_offsets_and_size() { 3448 // First, sort the array by frequencies. 3449 _constants.sort(qsort_comparator); 3450 3451 #ifdef ASSERT 3452 // Make sure all jump-table entries were sorted to the end of the 3453 // array (they have a negative frequency). 3454 bool found_void = false; 3455 for (int i = 0; i < _constants.length(); i++) { 3456 Constant con = _constants.at(i); 3457 if (con.type() == T_VOID) 3458 found_void = true; // jump-tables 3459 else 3460 assert(!found_void, "wrong sorting"); 3461 } 3462 #endif 3463 3464 int offset = 0; 3465 for (int i = 0; i < _constants.length(); i++) { 3466 Constant* con = _constants.adr_at(i); 3467 3468 // Align offset for type. 3469 int typesize = type_to_size_in_bytes(con->type()); 3470 offset = align_size_up(offset, typesize); 3471 con->set_offset(offset); // set constant's offset 3472 3473 if (con->type() == T_VOID) { 3474 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3475 offset = offset + typesize * n->outcnt(); // expand jump-table 3476 } else { 3477 offset = offset + typesize; 3478 } 3479 } 3480 3481 // Align size up to the next section start (which is insts; see 3482 // CodeBuffer::align_at_start). 3483 assert(_size == -1, "already set?"); 3484 _size = align_size_up(offset, CodeEntryAlignment); 3485 } 3486 3487 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3488 MacroAssembler _masm(&cb); 3489 for (int i = 0; i < _constants.length(); i++) { 3490 Constant con = _constants.at(i); 3491 address constant_addr; 3492 switch (con.type()) { 3493 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3494 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3495 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3496 case T_OBJECT: { 3497 jobject obj = con.get_jobject(); 3498 int oop_index = _masm.oop_recorder()->find_index(obj); 3499 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3500 break; 3501 } 3502 case T_ADDRESS: { 3503 address addr = (address) con.get_jobject(); 3504 constant_addr = _masm.address_constant(addr); 3505 break; 3506 } 3507 // We use T_VOID as marker for jump-table entries (labels) which 3508 // need an internal word relocation. 3509 case T_VOID: { 3510 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3511 // Fill the jump-table with a dummy word. The real value is 3512 // filled in later in fill_jump_table. 3513 address dummy = (address) n; 3514 constant_addr = _masm.address_constant(dummy); 3515 // Expand jump-table 3516 for (uint i = 1; i < n->outcnt(); i++) { 3517 address temp_addr = _masm.address_constant(dummy + i); 3518 assert(temp_addr, "consts section too small"); 3519 } 3520 break; 3521 } 3522 case T_METADATA: { 3523 Metadata* obj = con.get_metadata(); 3524 int metadata_index = _masm.oop_recorder()->find_index(obj); 3525 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3526 break; 3527 } 3528 default: ShouldNotReachHere(); 3529 } 3530 assert(constant_addr, "consts section too small"); 3531 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3532 } 3533 } 3534 3535 int Compile::ConstantTable::find_offset(Constant& con) const { 3536 int idx = _constants.find(con); 3537 assert(idx != -1, "constant must be in constant table"); 3538 int offset = _constants.at(idx).offset(); 3539 assert(offset != -1, "constant table not emitted yet?"); 3540 return offset; 3541 } 3542 3543 void Compile::ConstantTable::add(Constant& con) { 3544 if (con.can_be_reused()) { 3545 int idx = _constants.find(con); 3546 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3547 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3548 return; 3549 } 3550 } 3551 (void) _constants.append(con); 3552 } 3553 3554 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3555 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3556 Constant con(type, value, b->_freq); 3557 add(con); 3558 return con; 3559 } 3560 3561 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3562 Constant con(metadata); 3563 add(con); 3564 return con; 3565 } 3566 3567 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3568 jvalue value; 3569 BasicType type = oper->type()->basic_type(); 3570 switch (type) { 3571 case T_LONG: value.j = oper->constantL(); break; 3572 case T_FLOAT: value.f = oper->constantF(); break; 3573 case T_DOUBLE: value.d = oper->constantD(); break; 3574 case T_OBJECT: 3575 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3576 case T_METADATA: return add((Metadata*)oper->constant()); break; 3577 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3578 } 3579 return add(n, type, value); 3580 } 3581 3582 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3583 jvalue value; 3584 // We can use the node pointer here to identify the right jump-table 3585 // as this method is called from Compile::Fill_buffer right before 3586 // the MachNodes are emitted and the jump-table is filled (means the 3587 // MachNode pointers do not change anymore). 3588 value.l = (jobject) n; 3589 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3590 add(con); 3591 return con; 3592 } 3593 3594 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3595 // If called from Compile::scratch_emit_size do nothing. 3596 if (Compile::current()->in_scratch_emit_size()) return; 3597 3598 assert(labels.is_nonempty(), "must be"); 3599 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3600 3601 // Since MachConstantNode::constant_offset() also contains 3602 // table_base_offset() we need to subtract the table_base_offset() 3603 // to get the plain offset into the constant table. 3604 int offset = n->constant_offset() - table_base_offset(); 3605 3606 MacroAssembler _masm(&cb); 3607 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3608 3609 for (uint i = 0; i < n->outcnt(); i++) { 3610 address* constant_addr = &jump_table_base[i]; 3611 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))); 3612 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3613 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3614 } 3615 } 3616 3617 void Compile::dump_inlining() { 3618 if (print_inlining() || print_intrinsics()) { 3619 // Print inlining message for candidates that we couldn't inline 3620 // for lack of space or non constant receiver 3621 for (int i = 0; i < _late_inlines.length(); i++) { 3622 CallGenerator* cg = _late_inlines.at(i); 3623 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); 3624 } 3625 Unique_Node_List useful; 3626 useful.push(root()); 3627 for (uint next = 0; next < useful.size(); ++next) { 3628 Node* n = useful.at(next); 3629 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { 3630 CallNode* call = n->as_Call(); 3631 CallGenerator* cg = call->generator(); 3632 cg->print_inlining_late("receiver not constant"); 3633 } 3634 uint max = n->len(); 3635 for ( uint i = 0; i < max; ++i ) { 3636 Node *m = n->in(i); 3637 if ( m == NULL ) continue; 3638 useful.push(m); 3639 } 3640 } 3641 for (int i = 0; i < _print_inlining_list->length(); i++) { 3642 tty->print(_print_inlining_list->adr_at(i)->ss()->as_string()); 3643 } 3644 } 3645 } 3646 3647 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 3648 if (n1->Opcode() < n2->Opcode()) return -1; 3649 else if (n1->Opcode() > n2->Opcode()) return 1; 3650 3651 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())); 3652 for (uint i = 1; i < n1->req(); i++) { 3653 if (n1->in(i) < n2->in(i)) return -1; 3654 else if (n1->in(i) > n2->in(i)) return 1; 3655 } 3656 3657 return 0; 3658 } 3659 3660 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 3661 Node* n1 = *n1p; 3662 Node* n2 = *n2p; 3663 3664 return cmp_expensive_nodes(n1, n2); 3665 } 3666 3667 void Compile::sort_expensive_nodes() { 3668 if (!expensive_nodes_sorted()) { 3669 _expensive_nodes->sort(cmp_expensive_nodes); 3670 } 3671 } 3672 3673 bool Compile::expensive_nodes_sorted() const { 3674 for (int i = 1; i < _expensive_nodes->length(); i++) { 3675 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 3676 return false; 3677 } 3678 } 3679 return true; 3680 } 3681 3682 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 3683 if (_expensive_nodes->length() == 0) { 3684 return false; 3685 } 3686 3687 assert(OptimizeExpensiveOps, "optimization off?"); 3688 3689 // Take this opportunity to remove dead nodes from the list 3690 int j = 0; 3691 for (int i = 0; i < _expensive_nodes->length(); i++) { 3692 Node* n = _expensive_nodes->at(i); 3693 if (!n->is_unreachable(igvn)) { 3694 assert(n->is_expensive(), "should be expensive"); 3695 _expensive_nodes->at_put(j, n); 3696 j++; 3697 } 3698 } 3699 _expensive_nodes->trunc_to(j); 3700 3701 // Then sort the list so that similar nodes are next to each other 3702 // and check for at least two nodes of identical kind with same data 3703 // inputs. 3704 sort_expensive_nodes(); 3705 3706 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 3707 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 3708 return true; 3709 } 3710 } 3711 3712 return false; 3713 } 3714 3715 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 3716 if (_expensive_nodes->length() == 0) { 3717 return; 3718 } 3719 3720 assert(OptimizeExpensiveOps, "optimization off?"); 3721 3722 // Sort to bring similar nodes next to each other and clear the 3723 // control input of nodes for which there's only a single copy. 3724 sort_expensive_nodes(); 3725 3726 int j = 0; 3727 int identical = 0; 3728 int i = 0; 3729 for (; i < _expensive_nodes->length()-1; i++) { 3730 assert(j <= i, "can't write beyond current index"); 3731 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 3732 identical++; 3733 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3734 continue; 3735 } 3736 if (identical > 0) { 3737 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3738 identical = 0; 3739 } else { 3740 Node* n = _expensive_nodes->at(i); 3741 igvn.hash_delete(n); 3742 n->set_req(0, NULL); 3743 igvn.hash_insert(n); 3744 } 3745 } 3746 if (identical > 0) { 3747 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3748 } else if (_expensive_nodes->length() >= 1) { 3749 Node* n = _expensive_nodes->at(i); 3750 igvn.hash_delete(n); 3751 n->set_req(0, NULL); 3752 igvn.hash_insert(n); 3753 } 3754 _expensive_nodes->trunc_to(j); 3755 } 3756 3757 void Compile::add_expensive_node(Node * n) { 3758 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 3759 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 3760 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 3761 if (OptimizeExpensiveOps) { 3762 _expensive_nodes->append(n); 3763 } else { 3764 // Clear control input and let IGVN optimize expensive nodes if 3765 // OptimizeExpensiveOps is off. 3766 n->set_req(0, NULL); 3767 } 3768 } 3769 3770 // Auxiliary method to support randomized stressing/fuzzing. 3771 // 3772 // This method can be called the arbitrary number of times, with current count 3773 // as the argument. The logic allows selecting a single candidate from the 3774 // running list of candidates as follows: 3775 // int count = 0; 3776 // Cand* selected = null; 3777 // while(cand = cand->next()) { 3778 // if (randomized_select(++count)) { 3779 // selected = cand; 3780 // } 3781 // } 3782 // 3783 // Including count equalizes the chances any candidate is "selected". 3784 // This is useful when we don't have the complete list of candidates to choose 3785 // from uniformly. In this case, we need to adjust the randomicity of the 3786 // selection, or else we will end up biasing the selection towards the latter 3787 // candidates. 3788 // 3789 // Quick back-envelope calculation shows that for the list of n candidates 3790 // the equal probability for the candidate to persist as "best" can be 3791 // achieved by replacing it with "next" k-th candidate with the probability 3792 // of 1/k. It can be easily shown that by the end of the run, the 3793 // probability for any candidate is converged to 1/n, thus giving the 3794 // uniform distribution among all the candidates. 3795 // 3796 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 3797 #define RANDOMIZED_DOMAIN_POW 29 3798 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 3799 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 3800 bool Compile::randomized_select(int count) { 3801 assert(count > 0, "only positive"); 3802 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 3803 }