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/mathexactnode.hpp" 51 #include "opto/memnode.hpp" 52 #include "opto/mulnode.hpp" 53 #include "opto/node.hpp" 54 #include "opto/opcodes.hpp" 55 #include "opto/output.hpp" 56 #include "opto/parse.hpp" 57 #include "opto/phaseX.hpp" 58 #include "opto/rootnode.hpp" 59 #include "opto/runtime.hpp" 60 #include "opto/stringopts.hpp" 61 #include "opto/type.hpp" 62 #include "opto/vectornode.hpp" 63 #include "runtime/arguments.hpp" 64 #include "runtime/signature.hpp" 65 #include "runtime/stubRoutines.hpp" 66 #include "runtime/timer.hpp" 67 #include "trace/tracing.hpp" 68 #include "utilities/copy.hpp" 69 #ifdef TARGET_ARCH_MODEL_x86_32 70 # include "adfiles/ad_x86_32.hpp" 71 #endif 72 #ifdef TARGET_ARCH_MODEL_x86_64 73 # include "adfiles/ad_x86_64.hpp" 74 #endif 75 #ifdef TARGET_ARCH_MODEL_sparc 76 # include "adfiles/ad_sparc.hpp" 77 #endif 78 #ifdef TARGET_ARCH_MODEL_zero 79 # include "adfiles/ad_zero.hpp" 80 #endif 81 #ifdef TARGET_ARCH_MODEL_arm 82 # include "adfiles/ad_arm.hpp" 83 #endif 84 #ifdef TARGET_ARCH_MODEL_ppc 85 # include "adfiles/ad_ppc.hpp" 86 #endif 87 88 89 // -------------------- Compile::mach_constant_base_node ----------------------- 90 // Constant table base node singleton. 91 MachConstantBaseNode* Compile::mach_constant_base_node() { 92 if (_mach_constant_base_node == NULL) { 93 _mach_constant_base_node = new (C) MachConstantBaseNode(); 94 _mach_constant_base_node->add_req(C->root()); 95 } 96 return _mach_constant_base_node; 97 } 98 99 100 /// Support for intrinsics. 101 102 // Return the index at which m must be inserted (or already exists). 103 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 104 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 105 #ifdef ASSERT 106 for (int i = 1; i < _intrinsics->length(); i++) { 107 CallGenerator* cg1 = _intrinsics->at(i-1); 108 CallGenerator* cg2 = _intrinsics->at(i); 109 assert(cg1->method() != cg2->method() 110 ? cg1->method() < cg2->method() 111 : cg1->is_virtual() < cg2->is_virtual(), 112 "compiler intrinsics list must stay sorted"); 113 } 114 #endif 115 // Binary search sorted list, in decreasing intervals [lo, hi]. 116 int lo = 0, hi = _intrinsics->length()-1; 117 while (lo <= hi) { 118 int mid = (uint)(hi + lo) / 2; 119 ciMethod* mid_m = _intrinsics->at(mid)->method(); 120 if (m < mid_m) { 121 hi = mid-1; 122 } else if (m > mid_m) { 123 lo = mid+1; 124 } else { 125 // look at minor sort key 126 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 127 if (is_virtual < mid_virt) { 128 hi = mid-1; 129 } else if (is_virtual > mid_virt) { 130 lo = mid+1; 131 } else { 132 return mid; // exact match 133 } 134 } 135 } 136 return lo; // inexact match 137 } 138 139 void Compile::register_intrinsic(CallGenerator* cg) { 140 if (_intrinsics == NULL) { 141 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 142 } 143 // This code is stolen from ciObjectFactory::insert. 144 // Really, GrowableArray should have methods for 145 // insert_at, remove_at, and binary_search. 146 int len = _intrinsics->length(); 147 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 148 if (index == len) { 149 _intrinsics->append(cg); 150 } else { 151 #ifdef ASSERT 152 CallGenerator* oldcg = _intrinsics->at(index); 153 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 154 #endif 155 _intrinsics->append(_intrinsics->at(len-1)); 156 int pos; 157 for (pos = len-2; pos >= index; pos--) { 158 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 159 } 160 _intrinsics->at_put(index, cg); 161 } 162 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 163 } 164 165 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 166 assert(m->is_loaded(), "don't try this on unloaded methods"); 167 if (_intrinsics != NULL) { 168 int index = intrinsic_insertion_index(m, is_virtual); 169 if (index < _intrinsics->length() 170 && _intrinsics->at(index)->method() == m 171 && _intrinsics->at(index)->is_virtual() == is_virtual) { 172 return _intrinsics->at(index); 173 } 174 } 175 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 176 if (m->intrinsic_id() != vmIntrinsics::_none && 177 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 178 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 179 if (cg != NULL) { 180 // Save it for next time: 181 register_intrinsic(cg); 182 return cg; 183 } else { 184 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 185 } 186 } 187 return NULL; 188 } 189 190 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 191 // in library_call.cpp. 192 193 194 #ifndef PRODUCT 195 // statistics gathering... 196 197 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 198 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 199 200 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 201 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 202 int oflags = _intrinsic_hist_flags[id]; 203 assert(flags != 0, "what happened?"); 204 if (is_virtual) { 205 flags |= _intrinsic_virtual; 206 } 207 bool changed = (flags != oflags); 208 if ((flags & _intrinsic_worked) != 0) { 209 juint count = (_intrinsic_hist_count[id] += 1); 210 if (count == 1) { 211 changed = true; // first time 212 } 213 // increment the overall count also: 214 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 215 } 216 if (changed) { 217 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 218 // Something changed about the intrinsic's virtuality. 219 if ((flags & _intrinsic_virtual) != 0) { 220 // This is the first use of this intrinsic as a virtual call. 221 if (oflags != 0) { 222 // We already saw it as a non-virtual, so note both cases. 223 flags |= _intrinsic_both; 224 } 225 } else if ((oflags & _intrinsic_both) == 0) { 226 // This is the first use of this intrinsic as a non-virtual 227 flags |= _intrinsic_both; 228 } 229 } 230 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 231 } 232 // update the overall flags also: 233 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 234 return changed; 235 } 236 237 static char* format_flags(int flags, char* buf) { 238 buf[0] = 0; 239 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 240 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 241 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 242 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 243 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 244 if (buf[0] == 0) strcat(buf, ","); 245 assert(buf[0] == ',', "must be"); 246 return &buf[1]; 247 } 248 249 void Compile::print_intrinsic_statistics() { 250 char flagsbuf[100]; 251 ttyLocker ttyl; 252 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 253 tty->print_cr("Compiler intrinsic usage:"); 254 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 255 if (total == 0) total = 1; // avoid div0 in case of no successes 256 #define PRINT_STAT_LINE(name, c, f) \ 257 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 258 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 259 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 260 int flags = _intrinsic_hist_flags[id]; 261 juint count = _intrinsic_hist_count[id]; 262 if ((flags | count) != 0) { 263 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 264 } 265 } 266 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 267 if (xtty != NULL) xtty->tail("statistics"); 268 } 269 270 void Compile::print_statistics() { 271 { ttyLocker ttyl; 272 if (xtty != NULL) xtty->head("statistics type='opto'"); 273 Parse::print_statistics(); 274 PhaseCCP::print_statistics(); 275 PhaseRegAlloc::print_statistics(); 276 Scheduling::print_statistics(); 277 PhasePeephole::print_statistics(); 278 PhaseIdealLoop::print_statistics(); 279 if (xtty != NULL) xtty->tail("statistics"); 280 } 281 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 282 // put this under its own <statistics> element. 283 print_intrinsic_statistics(); 284 } 285 } 286 #endif //PRODUCT 287 288 // Support for bundling info 289 Bundle* Compile::node_bundling(const Node *n) { 290 assert(valid_bundle_info(n), "oob"); 291 return &_node_bundling_base[n->_idx]; 292 } 293 294 bool Compile::valid_bundle_info(const Node *n) { 295 return (_node_bundling_limit > n->_idx); 296 } 297 298 299 void Compile::gvn_replace_by(Node* n, Node* nn) { 300 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 301 Node* use = n->last_out(i); 302 bool is_in_table = initial_gvn()->hash_delete(use); 303 uint uses_found = 0; 304 for (uint j = 0; j < use->len(); j++) { 305 if (use->in(j) == n) { 306 if (j < use->req()) 307 use->set_req(j, nn); 308 else 309 use->set_prec(j, nn); 310 uses_found++; 311 } 312 } 313 if (is_in_table) { 314 // reinsert into table 315 initial_gvn()->hash_find_insert(use); 316 } 317 record_for_igvn(use); 318 i -= uses_found; // we deleted 1 or more copies of this edge 319 } 320 } 321 322 323 static inline bool not_a_node(const Node* n) { 324 if (n == NULL) return true; 325 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 326 if (*(address*)n == badAddress) return true; // kill by Node::destruct 327 return false; 328 } 329 330 // Identify all nodes that are reachable from below, useful. 331 // Use breadth-first pass that records state in a Unique_Node_List, 332 // recursive traversal is slower. 333 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 334 int estimated_worklist_size = unique(); 335 useful.map( estimated_worklist_size, NULL ); // preallocate space 336 337 // Initialize worklist 338 if (root() != NULL) { useful.push(root()); } 339 // If 'top' is cached, declare it useful to preserve cached node 340 if( cached_top_node() ) { useful.push(cached_top_node()); } 341 342 // Push all useful nodes onto the list, breadthfirst 343 for( uint next = 0; next < useful.size(); ++next ) { 344 assert( next < unique(), "Unique useful nodes < total nodes"); 345 Node *n = useful.at(next); 346 uint max = n->len(); 347 for( uint i = 0; i < max; ++i ) { 348 Node *m = n->in(i); 349 if (not_a_node(m)) continue; 350 useful.push(m); 351 } 352 } 353 } 354 355 // Update dead_node_list with any missing dead nodes using useful 356 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 357 void Compile::update_dead_node_list(Unique_Node_List &useful) { 358 uint max_idx = unique(); 359 VectorSet& useful_node_set = useful.member_set(); 360 361 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 362 // If node with index node_idx is not in useful set, 363 // mark it as dead in dead node list. 364 if (! useful_node_set.test(node_idx) ) { 365 record_dead_node(node_idx); 366 } 367 } 368 } 369 370 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 371 int shift = 0; 372 for (int i = 0; i < inlines->length(); i++) { 373 CallGenerator* cg = inlines->at(i); 374 CallNode* call = cg->call_node(); 375 if (shift > 0) { 376 inlines->at_put(i-shift, cg); 377 } 378 if (!useful.member(call)) { 379 shift++; 380 } 381 } 382 inlines->trunc_to(inlines->length()-shift); 383 } 384 385 // Disconnect all useless nodes by disconnecting those at the boundary. 386 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 387 uint next = 0; 388 while (next < useful.size()) { 389 Node *n = useful.at(next++); 390 // Use raw traversal of out edges since this code removes out edges 391 int max = n->outcnt(); 392 for (int j = 0; j < max; ++j) { 393 Node* child = n->raw_out(j); 394 if (! useful.member(child)) { 395 assert(!child->is_top() || child != top(), 396 "If top is cached in Compile object it is in useful list"); 397 // Only need to remove this out-edge to the useless node 398 n->raw_del_out(j); 399 --j; 400 --max; 401 } 402 } 403 if (n->outcnt() == 1 && n->has_special_unique_user()) { 404 record_for_igvn(n->unique_out()); 405 } 406 } 407 // Remove useless macro and predicate opaq nodes 408 for (int i = C->macro_count()-1; i >= 0; i--) { 409 Node* n = C->macro_node(i); 410 if (!useful.member(n)) { 411 remove_macro_node(n); 412 } 413 } 414 // Remove useless expensive node 415 for (int i = C->expensive_count()-1; i >= 0; i--) { 416 Node* n = C->expensive_node(i); 417 if (!useful.member(n)) { 418 remove_expensive_node(n); 419 } 420 } 421 // clean up the late inline lists 422 remove_useless_late_inlines(&_string_late_inlines, useful); 423 remove_useless_late_inlines(&_boxing_late_inlines, useful); 424 remove_useless_late_inlines(&_late_inlines, useful); 425 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 426 } 427 428 //------------------------------frame_size_in_words----------------------------- 429 // frame_slots in units of words 430 int Compile::frame_size_in_words() const { 431 // shift is 0 in LP32 and 1 in LP64 432 const int shift = (LogBytesPerWord - LogBytesPerInt); 433 int words = _frame_slots >> shift; 434 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 435 return words; 436 } 437 438 // ============================================================================ 439 //------------------------------CompileWrapper--------------------------------- 440 class CompileWrapper : public StackObj { 441 Compile *const _compile; 442 public: 443 CompileWrapper(Compile* compile); 444 445 ~CompileWrapper(); 446 }; 447 448 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 449 // the Compile* pointer is stored in the current ciEnv: 450 ciEnv* env = compile->env(); 451 assert(env == ciEnv::current(), "must already be a ciEnv active"); 452 assert(env->compiler_data() == NULL, "compile already active?"); 453 env->set_compiler_data(compile); 454 assert(compile == Compile::current(), "sanity"); 455 456 compile->set_type_dict(NULL); 457 compile->set_type_hwm(NULL); 458 compile->set_type_last_size(0); 459 compile->set_last_tf(NULL, NULL); 460 compile->set_indexSet_arena(NULL); 461 compile->set_indexSet_free_block_list(NULL); 462 compile->init_type_arena(); 463 Type::Initialize(compile); 464 _compile->set_scratch_buffer_blob(NULL); 465 _compile->begin_method(); 466 } 467 CompileWrapper::~CompileWrapper() { 468 _compile->end_method(); 469 if (_compile->scratch_buffer_blob() != NULL) 470 BufferBlob::free(_compile->scratch_buffer_blob()); 471 _compile->env()->set_compiler_data(NULL); 472 } 473 474 475 //----------------------------print_compile_messages--------------------------- 476 void Compile::print_compile_messages() { 477 #ifndef PRODUCT 478 // Check if recompiling 479 if (_subsume_loads == false && PrintOpto) { 480 // Recompiling without allowing machine instructions to subsume loads 481 tty->print_cr("*********************************************************"); 482 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 483 tty->print_cr("*********************************************************"); 484 } 485 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 486 // Recompiling without escape analysis 487 tty->print_cr("*********************************************************"); 488 tty->print_cr("** Bailout: Recompile without escape analysis **"); 489 tty->print_cr("*********************************************************"); 490 } 491 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 492 // Recompiling without boxing elimination 493 tty->print_cr("*********************************************************"); 494 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 495 tty->print_cr("*********************************************************"); 496 } 497 if (env()->break_at_compile()) { 498 // Open the debugger when compiling this method. 499 tty->print("### Breaking when compiling: "); 500 method()->print_short_name(); 501 tty->cr(); 502 BREAKPOINT; 503 } 504 505 if( PrintOpto ) { 506 if (is_osr_compilation()) { 507 tty->print("[OSR]%3d", _compile_id); 508 } else { 509 tty->print("%3d", _compile_id); 510 } 511 } 512 #endif 513 } 514 515 516 //-----------------------init_scratch_buffer_blob------------------------------ 517 // Construct a temporary BufferBlob and cache it for this compile. 518 void Compile::init_scratch_buffer_blob(int const_size) { 519 // If there is already a scratch buffer blob allocated and the 520 // constant section is big enough, use it. Otherwise free the 521 // current and allocate a new one. 522 BufferBlob* blob = scratch_buffer_blob(); 523 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 524 // Use the current blob. 525 } else { 526 if (blob != NULL) { 527 BufferBlob::free(blob); 528 } 529 530 ResourceMark rm; 531 _scratch_const_size = const_size; 532 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 533 blob = BufferBlob::create("Compile::scratch_buffer", size); 534 // Record the buffer blob for next time. 535 set_scratch_buffer_blob(blob); 536 // Have we run out of code space? 537 if (scratch_buffer_blob() == NULL) { 538 // Let CompilerBroker disable further compilations. 539 record_failure("Not enough space for scratch buffer in CodeCache"); 540 return; 541 } 542 } 543 544 // Initialize the relocation buffers 545 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 546 set_scratch_locs_memory(locs_buf); 547 } 548 549 550 //-----------------------scratch_emit_size------------------------------------- 551 // Helper function that computes size by emitting code 552 uint Compile::scratch_emit_size(const Node* n) { 553 // Start scratch_emit_size section. 554 set_in_scratch_emit_size(true); 555 556 // Emit into a trash buffer and count bytes emitted. 557 // This is a pretty expensive way to compute a size, 558 // but it works well enough if seldom used. 559 // All common fixed-size instructions are given a size 560 // method by the AD file. 561 // Note that the scratch buffer blob and locs memory are 562 // allocated at the beginning of the compile task, and 563 // may be shared by several calls to scratch_emit_size. 564 // The allocation of the scratch buffer blob is particularly 565 // expensive, since it has to grab the code cache lock. 566 BufferBlob* blob = this->scratch_buffer_blob(); 567 assert(blob != NULL, "Initialize BufferBlob at start"); 568 assert(blob->size() > MAX_inst_size, "sanity"); 569 relocInfo* locs_buf = scratch_locs_memory(); 570 address blob_begin = blob->content_begin(); 571 address blob_end = (address)locs_buf; 572 assert(blob->content_contains(blob_end), "sanity"); 573 CodeBuffer buf(blob_begin, blob_end - blob_begin); 574 buf.initialize_consts_size(_scratch_const_size); 575 buf.initialize_stubs_size(MAX_stubs_size); 576 assert(locs_buf != NULL, "sanity"); 577 int lsize = MAX_locs_size / 3; 578 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 579 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 580 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 581 582 // Do the emission. 583 584 Label fakeL; // Fake label for branch instructions. 585 Label* saveL = NULL; 586 uint save_bnum = 0; 587 bool is_branch = n->is_MachBranch(); 588 if (is_branch) { 589 MacroAssembler masm(&buf); 590 masm.bind(fakeL); 591 n->as_MachBranch()->save_label(&saveL, &save_bnum); 592 n->as_MachBranch()->label_set(&fakeL, 0); 593 } 594 n->emit(buf, this->regalloc()); 595 if (is_branch) // Restore label. 596 n->as_MachBranch()->label_set(saveL, save_bnum); 597 598 // End scratch_emit_size section. 599 set_in_scratch_emit_size(false); 600 601 return buf.insts_size(); 602 } 603 604 605 // ============================================================================ 606 //------------------------------Compile standard------------------------------- 607 debug_only( int Compile::_debug_idx = 100000; ) 608 609 // Compile a method. entry_bci is -1 for normal compilations and indicates 610 // the continuation bci for on stack replacement. 611 612 613 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 614 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 615 : Phase(Compiler), 616 _env(ci_env), 617 _log(ci_env->log()), 618 _compile_id(ci_env->compile_id()), 619 _save_argument_registers(false), 620 _stub_name(NULL), 621 _stub_function(NULL), 622 _stub_entry_point(NULL), 623 _method(target), 624 _entry_bci(osr_bci), 625 _initial_gvn(NULL), 626 _for_igvn(NULL), 627 _warm_calls(NULL), 628 _subsume_loads(subsume_loads), 629 _do_escape_analysis(do_escape_analysis), 630 _eliminate_boxing(eliminate_boxing), 631 _failure_reason(NULL), 632 _code_buffer("Compile::Fill_buffer"), 633 _orig_pc_slot(0), 634 _orig_pc_slot_offset_in_bytes(0), 635 _has_method_handle_invokes(false), 636 _mach_constant_base_node(NULL), 637 _node_bundling_limit(0), 638 _node_bundling_base(NULL), 639 _java_calls(0), 640 _inner_loops(0), 641 _scratch_const_size(-1), 642 _in_scratch_emit_size(false), 643 _dead_node_list(comp_arena()), 644 _dead_node_count(0), 645 #ifndef PRODUCT 646 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 647 _printer(IdealGraphPrinter::printer()), 648 #endif 649 _congraph(NULL), 650 _late_inlines(comp_arena(), 2, 0, NULL), 651 _string_late_inlines(comp_arena(), 2, 0, NULL), 652 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 653 _late_inlines_pos(0), 654 _number_of_mh_late_inlines(0), 655 _inlining_progress(false), 656 _inlining_incrementally(false), 657 _print_inlining_list(NULL), 658 _print_inlining_idx(0), 659 _preserve_jvm_state(0) { 660 C = this; 661 662 CompileWrapper cw(this); 663 #ifndef PRODUCT 664 if (TimeCompiler2) { 665 tty->print(" "); 666 target->holder()->name()->print(); 667 tty->print("."); 668 target->print_short_name(); 669 tty->print(" "); 670 } 671 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 672 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 673 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 674 if (!print_opto_assembly) { 675 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 676 if (print_assembly && !Disassembler::can_decode()) { 677 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 678 print_opto_assembly = true; 679 } 680 } 681 set_print_assembly(print_opto_assembly); 682 set_parsed_irreducible_loop(false); 683 #endif 684 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 685 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 686 687 if (ProfileTraps) { 688 // Make sure the method being compiled gets its own MDO, 689 // so we can at least track the decompile_count(). 690 method()->ensure_method_data(); 691 } 692 693 Init(::AliasLevel); 694 695 696 print_compile_messages(); 697 698 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 699 _ilt = InlineTree::build_inline_tree_root(); 700 else 701 _ilt = NULL; 702 703 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 704 assert(num_alias_types() >= AliasIdxRaw, ""); 705 706 #define MINIMUM_NODE_HASH 1023 707 // Node list that Iterative GVN will start with 708 Unique_Node_List for_igvn(comp_arena()); 709 set_for_igvn(&for_igvn); 710 711 // GVN that will be run immediately on new nodes 712 uint estimated_size = method()->code_size()*4+64; 713 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 714 PhaseGVN gvn(node_arena(), estimated_size); 715 set_initial_gvn(&gvn); 716 717 if (print_inlining() || print_intrinsics()) { 718 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 719 } 720 { // Scope for timing the parser 721 TracePhase t3("parse", &_t_parser, true); 722 723 // Put top into the hash table ASAP. 724 initial_gvn()->transform_no_reclaim(top()); 725 726 // Set up tf(), start(), and find a CallGenerator. 727 CallGenerator* cg = NULL; 728 if (is_osr_compilation()) { 729 const TypeTuple *domain = StartOSRNode::osr_domain(); 730 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 731 init_tf(TypeFunc::make(domain, range)); 732 StartNode* s = new (this) StartOSRNode(root(), domain); 733 initial_gvn()->set_type_bottom(s); 734 init_start(s); 735 cg = CallGenerator::for_osr(method(), entry_bci()); 736 } else { 737 // Normal case. 738 init_tf(TypeFunc::make(method())); 739 StartNode* s = new (this) StartNode(root(), tf()->domain()); 740 initial_gvn()->set_type_bottom(s); 741 init_start(s); 742 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 743 // With java.lang.ref.reference.get() we must go through the 744 // intrinsic when G1 is enabled - even when get() is the root 745 // method of the compile - so that, if necessary, the value in 746 // the referent field of the reference object gets recorded by 747 // the pre-barrier code. 748 // Specifically, if G1 is enabled, the value in the referent 749 // field is recorded by the G1 SATB pre barrier. This will 750 // result in the referent being marked live and the reference 751 // object removed from the list of discovered references during 752 // reference processing. 753 cg = find_intrinsic(method(), false); 754 } 755 if (cg == NULL) { 756 float past_uses = method()->interpreter_invocation_count(); 757 float expected_uses = past_uses; 758 cg = CallGenerator::for_inline(method(), expected_uses); 759 } 760 } 761 if (failing()) return; 762 if (cg == NULL) { 763 record_method_not_compilable_all_tiers("cannot parse method"); 764 return; 765 } 766 JVMState* jvms = build_start_state(start(), tf()); 767 if ((jvms = cg->generate(jvms, NULL)) == NULL) { 768 record_method_not_compilable("method parse failed"); 769 return; 770 } 771 GraphKit kit(jvms); 772 773 if (!kit.stopped()) { 774 // Accept return values, and transfer control we know not where. 775 // This is done by a special, unique ReturnNode bound to root. 776 return_values(kit.jvms()); 777 } 778 779 if (kit.has_exceptions()) { 780 // Any exceptions that escape from this call must be rethrown 781 // to whatever caller is dynamically above us on the stack. 782 // This is done by a special, unique RethrowNode bound to root. 783 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 784 } 785 786 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 787 788 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 789 inline_string_calls(true); 790 } 791 792 if (failing()) return; 793 794 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 795 796 // Remove clutter produced by parsing. 797 if (!failing()) { 798 ResourceMark rm; 799 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 800 } 801 } 802 803 // Note: Large methods are capped off in do_one_bytecode(). 804 if (failing()) return; 805 806 // After parsing, node notes are no longer automagic. 807 // They must be propagated by register_new_node_with_optimizer(), 808 // clone(), or the like. 809 set_default_node_notes(NULL); 810 811 for (;;) { 812 int successes = Inline_Warm(); 813 if (failing()) return; 814 if (successes == 0) break; 815 } 816 817 // Drain the list. 818 Finish_Warm(); 819 #ifndef PRODUCT 820 if (_printer) { 821 _printer->print_inlining(this); 822 } 823 #endif 824 825 if (failing()) return; 826 NOT_PRODUCT( verify_graph_edges(); ) 827 828 // Now optimize 829 Optimize(); 830 if (failing()) return; 831 NOT_PRODUCT( verify_graph_edges(); ) 832 833 #ifndef PRODUCT 834 if (PrintIdeal) { 835 ttyLocker ttyl; // keep the following output all in one block 836 // This output goes directly to the tty, not the compiler log. 837 // To enable tools to match it up with the compilation activity, 838 // be sure to tag this tty output with the compile ID. 839 if (xtty != NULL) { 840 xtty->head("ideal compile_id='%d'%s", compile_id(), 841 is_osr_compilation() ? " compile_kind='osr'" : 842 ""); 843 } 844 root()->dump(9999); 845 if (xtty != NULL) { 846 xtty->tail("ideal"); 847 } 848 } 849 #endif 850 851 // Now that we know the size of all the monitors we can add a fixed slot 852 // for the original deopt pc. 853 854 _orig_pc_slot = fixed_slots(); 855 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 856 set_fixed_slots(next_slot); 857 858 // Now generate code 859 Code_Gen(); 860 if (failing()) return; 861 862 // Check if we want to skip execution of all compiled code. 863 { 864 #ifndef PRODUCT 865 if (OptoNoExecute) { 866 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 867 return; 868 } 869 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 870 #endif 871 872 if (is_osr_compilation()) { 873 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 874 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 875 } else { 876 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 877 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 878 } 879 880 env()->register_method(_method, _entry_bci, 881 &_code_offsets, 882 _orig_pc_slot_offset_in_bytes, 883 code_buffer(), 884 frame_size_in_words(), _oop_map_set, 885 &_handler_table, &_inc_table, 886 compiler, 887 env()->comp_level(), 888 has_unsafe_access(), 889 SharedRuntime::is_wide_vector(max_vector_size()) 890 ); 891 892 if (log() != NULL) // Print code cache state into compiler log 893 log()->code_cache_state(); 894 } 895 } 896 897 //------------------------------Compile---------------------------------------- 898 // Compile a runtime stub 899 Compile::Compile( ciEnv* ci_env, 900 TypeFunc_generator generator, 901 address stub_function, 902 const char *stub_name, 903 int is_fancy_jump, 904 bool pass_tls, 905 bool save_arg_registers, 906 bool return_pc ) 907 : Phase(Compiler), 908 _env(ci_env), 909 _log(ci_env->log()), 910 _compile_id(0), 911 _save_argument_registers(save_arg_registers), 912 _method(NULL), 913 _stub_name(stub_name), 914 _stub_function(stub_function), 915 _stub_entry_point(NULL), 916 _entry_bci(InvocationEntryBci), 917 _initial_gvn(NULL), 918 _for_igvn(NULL), 919 _warm_calls(NULL), 920 _orig_pc_slot(0), 921 _orig_pc_slot_offset_in_bytes(0), 922 _subsume_loads(true), 923 _do_escape_analysis(false), 924 _eliminate_boxing(false), 925 _failure_reason(NULL), 926 _code_buffer("Compile::Fill_buffer"), 927 _has_method_handle_invokes(false), 928 _mach_constant_base_node(NULL), 929 _node_bundling_limit(0), 930 _node_bundling_base(NULL), 931 _java_calls(0), 932 _inner_loops(0), 933 #ifndef PRODUCT 934 _trace_opto_output(TraceOptoOutput), 935 _printer(NULL), 936 #endif 937 _dead_node_list(comp_arena()), 938 _dead_node_count(0), 939 _congraph(NULL), 940 _number_of_mh_late_inlines(0), 941 _inlining_progress(false), 942 _inlining_incrementally(false), 943 _print_inlining_list(NULL), 944 _print_inlining_idx(0), 945 _preserve_jvm_state(0) { 946 C = this; 947 948 #ifndef PRODUCT 949 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 950 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 951 set_print_assembly(PrintFrameConverterAssembly); 952 set_parsed_irreducible_loop(false); 953 #endif 954 CompileWrapper cw(this); 955 Init(/*AliasLevel=*/ 0); 956 init_tf((*generator)()); 957 958 { 959 // The following is a dummy for the sake of GraphKit::gen_stub 960 Unique_Node_List for_igvn(comp_arena()); 961 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 962 PhaseGVN gvn(Thread::current()->resource_area(),255); 963 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 964 gvn.transform_no_reclaim(top()); 965 966 GraphKit kit; 967 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 968 } 969 970 NOT_PRODUCT( verify_graph_edges(); ) 971 Code_Gen(); 972 if (failing()) return; 973 974 975 // Entry point will be accessed using compile->stub_entry_point(); 976 if (code_buffer() == NULL) { 977 Matcher::soft_match_failure(); 978 } else { 979 if (PrintAssembly && (WizardMode || Verbose)) 980 tty->print_cr("### Stub::%s", stub_name); 981 982 if (!failing()) { 983 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 984 985 // Make the NMethod 986 // For now we mark the frame as never safe for profile stackwalking 987 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 988 code_buffer(), 989 CodeOffsets::frame_never_safe, 990 // _code_offsets.value(CodeOffsets::Frame_Complete), 991 frame_size_in_words(), 992 _oop_map_set, 993 save_arg_registers); 994 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 995 996 _stub_entry_point = rs->entry_point(); 997 } 998 } 999 } 1000 1001 //------------------------------Init------------------------------------------- 1002 // Prepare for a single compilation 1003 void Compile::Init(int aliaslevel) { 1004 _unique = 0; 1005 _regalloc = NULL; 1006 1007 _tf = NULL; // filled in later 1008 _top = NULL; // cached later 1009 _matcher = NULL; // filled in later 1010 _cfg = NULL; // filled in later 1011 1012 set_24_bit_selection_and_mode(Use24BitFP, false); 1013 1014 _node_note_array = NULL; 1015 _default_node_notes = NULL; 1016 1017 _immutable_memory = NULL; // filled in at first inquiry 1018 1019 // Globally visible Nodes 1020 // First set TOP to NULL to give safe behavior during creation of RootNode 1021 set_cached_top_node(NULL); 1022 set_root(new (this) RootNode()); 1023 // Now that you have a Root to point to, create the real TOP 1024 set_cached_top_node( new (this) ConNode(Type::TOP) ); 1025 set_recent_alloc(NULL, NULL); 1026 1027 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1028 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1029 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1030 env()->set_dependencies(new Dependencies(env())); 1031 1032 _fixed_slots = 0; 1033 set_has_split_ifs(false); 1034 set_has_loops(has_method() && method()->has_loops()); // first approximation 1035 set_has_stringbuilder(false); 1036 set_has_boxed_value(false); 1037 _trap_can_recompile = false; // no traps emitted yet 1038 _major_progress = true; // start out assuming good things will happen 1039 set_has_unsafe_access(false); 1040 set_max_vector_size(0); 1041 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1042 set_decompile_count(0); 1043 1044 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1045 set_num_loop_opts(LoopOptsCount); 1046 set_do_inlining(Inline); 1047 set_max_inline_size(MaxInlineSize); 1048 set_freq_inline_size(FreqInlineSize); 1049 set_do_scheduling(OptoScheduling); 1050 set_do_count_invocations(false); 1051 set_do_method_data_update(false); 1052 1053 if (debug_info()->recording_non_safepoints()) { 1054 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1055 (comp_arena(), 8, 0, NULL)); 1056 set_default_node_notes(Node_Notes::make(this)); 1057 } 1058 1059 // // -- Initialize types before each compile -- 1060 // // Update cached type information 1061 // if( _method && _method->constants() ) 1062 // Type::update_loaded_types(_method, _method->constants()); 1063 1064 // Init alias_type map. 1065 if (!_do_escape_analysis && aliaslevel == 3) 1066 aliaslevel = 2; // No unique types without escape analysis 1067 _AliasLevel = aliaslevel; 1068 const int grow_ats = 16; 1069 _max_alias_types = grow_ats; 1070 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1071 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1072 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1073 { 1074 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1075 } 1076 // Initialize the first few types. 1077 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1078 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1079 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1080 _num_alias_types = AliasIdxRaw+1; 1081 // Zero out the alias type cache. 1082 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1083 // A NULL adr_type hits in the cache right away. Preload the right answer. 1084 probe_alias_cache(NULL)->_index = AliasIdxTop; 1085 1086 _intrinsics = NULL; 1087 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1088 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1089 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1090 register_library_intrinsics(); 1091 } 1092 1093 //---------------------------init_start---------------------------------------- 1094 // Install the StartNode on this compile object. 1095 void Compile::init_start(StartNode* s) { 1096 if (failing()) 1097 return; // already failing 1098 assert(s == start(), ""); 1099 } 1100 1101 StartNode* Compile::start() const { 1102 assert(!failing(), ""); 1103 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1104 Node* start = root()->fast_out(i); 1105 if( start->is_Start() ) 1106 return start->as_Start(); 1107 } 1108 ShouldNotReachHere(); 1109 return NULL; 1110 } 1111 1112 //-------------------------------immutable_memory------------------------------------- 1113 // Access immutable memory 1114 Node* Compile::immutable_memory() { 1115 if (_immutable_memory != NULL) { 1116 return _immutable_memory; 1117 } 1118 StartNode* s = start(); 1119 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1120 Node *p = s->fast_out(i); 1121 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1122 _immutable_memory = p; 1123 return _immutable_memory; 1124 } 1125 } 1126 ShouldNotReachHere(); 1127 return NULL; 1128 } 1129 1130 //----------------------set_cached_top_node------------------------------------ 1131 // Install the cached top node, and make sure Node::is_top works correctly. 1132 void Compile::set_cached_top_node(Node* tn) { 1133 if (tn != NULL) verify_top(tn); 1134 Node* old_top = _top; 1135 _top = tn; 1136 // Calling Node::setup_is_top allows the nodes the chance to adjust 1137 // their _out arrays. 1138 if (_top != NULL) _top->setup_is_top(); 1139 if (old_top != NULL) old_top->setup_is_top(); 1140 assert(_top == NULL || top()->is_top(), ""); 1141 } 1142 1143 #ifdef ASSERT 1144 uint Compile::count_live_nodes_by_graph_walk() { 1145 Unique_Node_List useful(comp_arena()); 1146 // Get useful node list by walking the graph. 1147 identify_useful_nodes(useful); 1148 return useful.size(); 1149 } 1150 1151 void Compile::print_missing_nodes() { 1152 1153 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1154 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1155 return; 1156 } 1157 1158 // This is an expensive function. It is executed only when the user 1159 // specifies VerifyIdealNodeCount option or otherwise knows the 1160 // additional work that needs to be done to identify reachable nodes 1161 // by walking the flow graph and find the missing ones using 1162 // _dead_node_list. 1163 1164 Unique_Node_List useful(comp_arena()); 1165 // Get useful node list by walking the graph. 1166 identify_useful_nodes(useful); 1167 1168 uint l_nodes = C->live_nodes(); 1169 uint l_nodes_by_walk = useful.size(); 1170 1171 if (l_nodes != l_nodes_by_walk) { 1172 if (_log != NULL) { 1173 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1174 _log->stamp(); 1175 _log->end_head(); 1176 } 1177 VectorSet& useful_member_set = useful.member_set(); 1178 int last_idx = l_nodes_by_walk; 1179 for (int i = 0; i < last_idx; i++) { 1180 if (useful_member_set.test(i)) { 1181 if (_dead_node_list.test(i)) { 1182 if (_log != NULL) { 1183 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1184 } 1185 if (PrintIdealNodeCount) { 1186 // Print the log message to tty 1187 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1188 useful.at(i)->dump(); 1189 } 1190 } 1191 } 1192 else if (! _dead_node_list.test(i)) { 1193 if (_log != NULL) { 1194 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1195 } 1196 if (PrintIdealNodeCount) { 1197 // Print the log message to tty 1198 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1199 } 1200 } 1201 } 1202 if (_log != NULL) { 1203 _log->tail("mismatched_nodes"); 1204 } 1205 } 1206 } 1207 #endif 1208 1209 #ifndef PRODUCT 1210 void Compile::verify_top(Node* tn) const { 1211 if (tn != NULL) { 1212 assert(tn->is_Con(), "top node must be a constant"); 1213 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1214 assert(tn->in(0) != NULL, "must have live top node"); 1215 } 1216 } 1217 #endif 1218 1219 1220 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1221 1222 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1223 guarantee(arr != NULL, ""); 1224 int num_blocks = arr->length(); 1225 if (grow_by < num_blocks) grow_by = num_blocks; 1226 int num_notes = grow_by * _node_notes_block_size; 1227 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1228 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1229 while (num_notes > 0) { 1230 arr->append(notes); 1231 notes += _node_notes_block_size; 1232 num_notes -= _node_notes_block_size; 1233 } 1234 assert(num_notes == 0, "exact multiple, please"); 1235 } 1236 1237 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1238 if (source == NULL || dest == NULL) return false; 1239 1240 if (dest->is_Con()) 1241 return false; // Do not push debug info onto constants. 1242 1243 #ifdef ASSERT 1244 // Leave a bread crumb trail pointing to the original node: 1245 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1246 dest->set_debug_orig(source); 1247 } 1248 #endif 1249 1250 if (node_note_array() == NULL) 1251 return false; // Not collecting any notes now. 1252 1253 // This is a copy onto a pre-existing node, which may already have notes. 1254 // If both nodes have notes, do not overwrite any pre-existing notes. 1255 Node_Notes* source_notes = node_notes_at(source->_idx); 1256 if (source_notes == NULL || source_notes->is_clear()) return false; 1257 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1258 if (dest_notes == NULL || dest_notes->is_clear()) { 1259 return set_node_notes_at(dest->_idx, source_notes); 1260 } 1261 1262 Node_Notes merged_notes = (*source_notes); 1263 // The order of operations here ensures that dest notes will win... 1264 merged_notes.update_from(dest_notes); 1265 return set_node_notes_at(dest->_idx, &merged_notes); 1266 } 1267 1268 1269 //--------------------------allow_range_check_smearing------------------------- 1270 // Gating condition for coalescing similar range checks. 1271 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1272 // single covering check that is at least as strong as any of them. 1273 // If the optimization succeeds, the simplified (strengthened) range check 1274 // will always succeed. If it fails, we will deopt, and then give up 1275 // on the optimization. 1276 bool Compile::allow_range_check_smearing() const { 1277 // If this method has already thrown a range-check, 1278 // assume it was because we already tried range smearing 1279 // and it failed. 1280 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1281 return !already_trapped; 1282 } 1283 1284 1285 //------------------------------flatten_alias_type----------------------------- 1286 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1287 int offset = tj->offset(); 1288 TypePtr::PTR ptr = tj->ptr(); 1289 1290 // Known instance (scalarizable allocation) alias only with itself. 1291 bool is_known_inst = tj->isa_oopptr() != NULL && 1292 tj->is_oopptr()->is_known_instance(); 1293 1294 // Process weird unsafe references. 1295 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1296 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1297 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1298 tj = TypeOopPtr::BOTTOM; 1299 ptr = tj->ptr(); 1300 offset = tj->offset(); 1301 } 1302 1303 // Array pointers need some flattening 1304 const TypeAryPtr *ta = tj->isa_aryptr(); 1305 if (ta && ta->is_stable()) { 1306 // Erase stability property for alias analysis. 1307 tj = ta = ta->cast_to_stable(false); 1308 } 1309 if( ta && is_known_inst ) { 1310 if ( offset != Type::OffsetBot && 1311 offset > arrayOopDesc::length_offset_in_bytes() ) { 1312 offset = Type::OffsetBot; // Flatten constant access into array body only 1313 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1314 } 1315 } else if( ta && _AliasLevel >= 2 ) { 1316 // For arrays indexed by constant indices, we flatten the alias 1317 // space to include all of the array body. Only the header, klass 1318 // and array length can be accessed un-aliased. 1319 if( offset != Type::OffsetBot ) { 1320 if( ta->const_oop() ) { // MethodData* or Method* 1321 offset = Type::OffsetBot; // Flatten constant access into array body 1322 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1323 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1324 // range is OK as-is. 1325 tj = ta = TypeAryPtr::RANGE; 1326 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1327 tj = TypeInstPtr::KLASS; // all klass loads look alike 1328 ta = TypeAryPtr::RANGE; // generic ignored junk 1329 ptr = TypePtr::BotPTR; 1330 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1331 tj = TypeInstPtr::MARK; 1332 ta = TypeAryPtr::RANGE; // generic ignored junk 1333 ptr = TypePtr::BotPTR; 1334 } else { // Random constant offset into array body 1335 offset = Type::OffsetBot; // Flatten constant access into array body 1336 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1337 } 1338 } 1339 // Arrays of fixed size alias with arrays of unknown size. 1340 if (ta->size() != TypeInt::POS) { 1341 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1342 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1343 } 1344 // Arrays of known objects become arrays of unknown objects. 1345 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1346 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1347 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1348 } 1349 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1350 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1351 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1352 } 1353 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1354 // cannot be distinguished by bytecode alone. 1355 if (ta->elem() == TypeInt::BOOL) { 1356 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1357 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1358 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1359 } 1360 // During the 2nd round of IterGVN, NotNull castings are removed. 1361 // Make sure the Bottom and NotNull variants alias the same. 1362 // Also, make sure exact and non-exact variants alias the same. 1363 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1364 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1365 } 1366 } 1367 1368 // Oop pointers need some flattening 1369 const TypeInstPtr *to = tj->isa_instptr(); 1370 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1371 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1372 if( ptr == TypePtr::Constant ) { 1373 if (to->klass() != ciEnv::current()->Class_klass() || 1374 offset < k->size_helper() * wordSize) { 1375 // No constant oop pointers (such as Strings); they alias with 1376 // unknown strings. 1377 assert(!is_known_inst, "not scalarizable allocation"); 1378 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1379 } 1380 } else if( is_known_inst ) { 1381 tj = to; // Keep NotNull and klass_is_exact for instance type 1382 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1383 // During the 2nd round of IterGVN, NotNull castings are removed. 1384 // Make sure the Bottom and NotNull variants alias the same. 1385 // Also, make sure exact and non-exact variants alias the same. 1386 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1387 } 1388 // Canonicalize the holder of this field 1389 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1390 // First handle header references such as a LoadKlassNode, even if the 1391 // object's klass is unloaded at compile time (4965979). 1392 if (!is_known_inst) { // Do it only for non-instance types 1393 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1394 } 1395 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1396 // Static fields are in the space above the normal instance 1397 // fields in the java.lang.Class instance. 1398 if (to->klass() != ciEnv::current()->Class_klass()) { 1399 to = NULL; 1400 tj = TypeOopPtr::BOTTOM; 1401 offset = tj->offset(); 1402 } 1403 } else { 1404 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1405 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1406 if( is_known_inst ) { 1407 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1408 } else { 1409 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1410 } 1411 } 1412 } 1413 } 1414 1415 // Klass pointers to object array klasses need some flattening 1416 const TypeKlassPtr *tk = tj->isa_klassptr(); 1417 if( tk ) { 1418 // If we are referencing a field within a Klass, we need 1419 // to assume the worst case of an Object. Both exact and 1420 // inexact types must flatten to the same alias class so 1421 // use NotNull as the PTR. 1422 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1423 1424 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1425 TypeKlassPtr::OBJECT->klass(), 1426 offset); 1427 } 1428 1429 ciKlass* klass = tk->klass(); 1430 if( klass->is_obj_array_klass() ) { 1431 ciKlass* k = TypeAryPtr::OOPS->klass(); 1432 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1433 k = TypeInstPtr::BOTTOM->klass(); 1434 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1435 } 1436 1437 // Check for precise loads from the primary supertype array and force them 1438 // to the supertype cache alias index. Check for generic array loads from 1439 // the primary supertype array and also force them to the supertype cache 1440 // alias index. Since the same load can reach both, we need to merge 1441 // these 2 disparate memories into the same alias class. Since the 1442 // primary supertype array is read-only, there's no chance of confusion 1443 // where we bypass an array load and an array store. 1444 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1445 if (offset == Type::OffsetBot || 1446 (offset >= primary_supers_offset && 1447 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1448 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1449 offset = in_bytes(Klass::secondary_super_cache_offset()); 1450 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1451 } 1452 } 1453 1454 // Flatten all Raw pointers together. 1455 if (tj->base() == Type::RawPtr) 1456 tj = TypeRawPtr::BOTTOM; 1457 1458 if (tj->base() == Type::AnyPtr) 1459 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1460 1461 // Flatten all to bottom for now 1462 switch( _AliasLevel ) { 1463 case 0: 1464 tj = TypePtr::BOTTOM; 1465 break; 1466 case 1: // Flatten to: oop, static, field or array 1467 switch (tj->base()) { 1468 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1469 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1470 case Type::AryPtr: // do not distinguish arrays at all 1471 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1472 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1473 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1474 default: ShouldNotReachHere(); 1475 } 1476 break; 1477 case 2: // No collapsing at level 2; keep all splits 1478 case 3: // No collapsing at level 3; keep all splits 1479 break; 1480 default: 1481 Unimplemented(); 1482 } 1483 1484 offset = tj->offset(); 1485 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1486 1487 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1488 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1489 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1490 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1491 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1492 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1493 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1494 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1495 assert( tj->ptr() != TypePtr::TopPTR && 1496 tj->ptr() != TypePtr::AnyNull && 1497 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1498 // assert( tj->ptr() != TypePtr::Constant || 1499 // tj->base() == Type::RawPtr || 1500 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1501 1502 return tj; 1503 } 1504 1505 void Compile::AliasType::Init(int i, const TypePtr* at) { 1506 _index = i; 1507 _adr_type = at; 1508 _field = NULL; 1509 _element = NULL; 1510 _is_rewritable = true; // default 1511 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1512 if (atoop != NULL && atoop->is_known_instance()) { 1513 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1514 _general_index = Compile::current()->get_alias_index(gt); 1515 } else { 1516 _general_index = 0; 1517 } 1518 } 1519 1520 //---------------------------------print_on------------------------------------ 1521 #ifndef PRODUCT 1522 void Compile::AliasType::print_on(outputStream* st) { 1523 if (index() < 10) 1524 st->print("@ <%d> ", index()); 1525 else st->print("@ <%d>", index()); 1526 st->print(is_rewritable() ? " " : " RO"); 1527 int offset = adr_type()->offset(); 1528 if (offset == Type::OffsetBot) 1529 st->print(" +any"); 1530 else st->print(" +%-3d", offset); 1531 st->print(" in "); 1532 adr_type()->dump_on(st); 1533 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1534 if (field() != NULL && tjp) { 1535 if (tjp->klass() != field()->holder() || 1536 tjp->offset() != field()->offset_in_bytes()) { 1537 st->print(" != "); 1538 field()->print(); 1539 st->print(" ***"); 1540 } 1541 } 1542 } 1543 1544 void print_alias_types() { 1545 Compile* C = Compile::current(); 1546 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1547 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1548 C->alias_type(idx)->print_on(tty); 1549 tty->cr(); 1550 } 1551 } 1552 #endif 1553 1554 1555 //----------------------------probe_alias_cache-------------------------------- 1556 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1557 intptr_t key = (intptr_t) adr_type; 1558 key ^= key >> logAliasCacheSize; 1559 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1560 } 1561 1562 1563 //-----------------------------grow_alias_types-------------------------------- 1564 void Compile::grow_alias_types() { 1565 const int old_ats = _max_alias_types; // how many before? 1566 const int new_ats = old_ats; // how many more? 1567 const int grow_ats = old_ats+new_ats; // how many now? 1568 _max_alias_types = grow_ats; 1569 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1570 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1571 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1572 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1573 } 1574 1575 1576 //--------------------------------find_alias_type------------------------------ 1577 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1578 if (_AliasLevel == 0) 1579 return alias_type(AliasIdxBot); 1580 1581 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1582 if (ace->_adr_type == adr_type) { 1583 return alias_type(ace->_index); 1584 } 1585 1586 // Handle special cases. 1587 if (adr_type == NULL) return alias_type(AliasIdxTop); 1588 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1589 1590 // Do it the slow way. 1591 const TypePtr* flat = flatten_alias_type(adr_type); 1592 1593 #ifdef ASSERT 1594 assert(flat == flatten_alias_type(flat), "idempotent"); 1595 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1596 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1597 const TypeOopPtr* foop = flat->is_oopptr(); 1598 // Scalarizable allocations have exact klass always. 1599 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1600 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1601 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1602 } 1603 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1604 #endif 1605 1606 int idx = AliasIdxTop; 1607 for (int i = 0; i < num_alias_types(); i++) { 1608 if (alias_type(i)->adr_type() == flat) { 1609 idx = i; 1610 break; 1611 } 1612 } 1613 1614 if (idx == AliasIdxTop) { 1615 if (no_create) return NULL; 1616 // Grow the array if necessary. 1617 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1618 // Add a new alias type. 1619 idx = _num_alias_types++; 1620 _alias_types[idx]->Init(idx, flat); 1621 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1622 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1623 if (flat->isa_instptr()) { 1624 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1625 && flat->is_instptr()->klass() == env()->Class_klass()) 1626 alias_type(idx)->set_rewritable(false); 1627 } 1628 if (flat->isa_aryptr()) { 1629 #ifdef ASSERT 1630 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1631 // (T_BYTE has the weakest alignment and size restrictions...) 1632 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1633 #endif 1634 if (flat->offset() == TypePtr::OffsetBot) { 1635 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1636 } 1637 } 1638 if (flat->isa_klassptr()) { 1639 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1640 alias_type(idx)->set_rewritable(false); 1641 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1642 alias_type(idx)->set_rewritable(false); 1643 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1644 alias_type(idx)->set_rewritable(false); 1645 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1646 alias_type(idx)->set_rewritable(false); 1647 } 1648 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1649 // but the base pointer type is not distinctive enough to identify 1650 // references into JavaThread.) 1651 1652 // Check for final fields. 1653 const TypeInstPtr* tinst = flat->isa_instptr(); 1654 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1655 ciField* field; 1656 if (tinst->const_oop() != NULL && 1657 tinst->klass() == ciEnv::current()->Class_klass() && 1658 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1659 // static field 1660 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1661 field = k->get_field_by_offset(tinst->offset(), true); 1662 } else { 1663 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1664 field = k->get_field_by_offset(tinst->offset(), false); 1665 } 1666 assert(field == NULL || 1667 original_field == NULL || 1668 (field->holder() == original_field->holder() && 1669 field->offset() == original_field->offset() && 1670 field->is_static() == original_field->is_static()), "wrong field?"); 1671 // Set field() and is_rewritable() attributes. 1672 if (field != NULL) alias_type(idx)->set_field(field); 1673 } 1674 } 1675 1676 // Fill the cache for next time. 1677 ace->_adr_type = adr_type; 1678 ace->_index = idx; 1679 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1680 1681 // Might as well try to fill the cache for the flattened version, too. 1682 AliasCacheEntry* face = probe_alias_cache(flat); 1683 if (face->_adr_type == NULL) { 1684 face->_adr_type = flat; 1685 face->_index = idx; 1686 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1687 } 1688 1689 return alias_type(idx); 1690 } 1691 1692 1693 Compile::AliasType* Compile::alias_type(ciField* field) { 1694 const TypeOopPtr* t; 1695 if (field->is_static()) 1696 t = TypeInstPtr::make(field->holder()->java_mirror()); 1697 else 1698 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1699 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1700 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1701 return atp; 1702 } 1703 1704 1705 //------------------------------have_alias_type-------------------------------- 1706 bool Compile::have_alias_type(const TypePtr* adr_type) { 1707 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1708 if (ace->_adr_type == adr_type) { 1709 return true; 1710 } 1711 1712 // Handle special cases. 1713 if (adr_type == NULL) return true; 1714 if (adr_type == TypePtr::BOTTOM) return true; 1715 1716 return find_alias_type(adr_type, true, NULL) != NULL; 1717 } 1718 1719 //-----------------------------must_alias-------------------------------------- 1720 // True if all values of the given address type are in the given alias category. 1721 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1722 if (alias_idx == AliasIdxBot) return true; // the universal category 1723 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1724 if (alias_idx == AliasIdxTop) return false; // the empty category 1725 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1726 1727 // the only remaining possible overlap is identity 1728 int adr_idx = get_alias_index(adr_type); 1729 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1730 assert(adr_idx == alias_idx || 1731 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1732 && adr_type != TypeOopPtr::BOTTOM), 1733 "should not be testing for overlap with an unsafe pointer"); 1734 return adr_idx == alias_idx; 1735 } 1736 1737 //------------------------------can_alias-------------------------------------- 1738 // True if any values of the given address type are in the given alias category. 1739 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1740 if (alias_idx == AliasIdxTop) return false; // the empty category 1741 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1742 if (alias_idx == AliasIdxBot) return true; // the universal category 1743 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1744 1745 // the only remaining possible overlap is identity 1746 int adr_idx = get_alias_index(adr_type); 1747 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1748 return adr_idx == alias_idx; 1749 } 1750 1751 1752 1753 //---------------------------pop_warm_call------------------------------------- 1754 WarmCallInfo* Compile::pop_warm_call() { 1755 WarmCallInfo* wci = _warm_calls; 1756 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1757 return wci; 1758 } 1759 1760 //----------------------------Inline_Warm-------------------------------------- 1761 int Compile::Inline_Warm() { 1762 // If there is room, try to inline some more warm call sites. 1763 // %%% Do a graph index compaction pass when we think we're out of space? 1764 if (!InlineWarmCalls) return 0; 1765 1766 int calls_made_hot = 0; 1767 int room_to_grow = NodeCountInliningCutoff - unique(); 1768 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1769 int amount_grown = 0; 1770 WarmCallInfo* call; 1771 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1772 int est_size = (int)call->size(); 1773 if (est_size > (room_to_grow - amount_grown)) { 1774 // This one won't fit anyway. Get rid of it. 1775 call->make_cold(); 1776 continue; 1777 } 1778 call->make_hot(); 1779 calls_made_hot++; 1780 amount_grown += est_size; 1781 amount_to_grow -= est_size; 1782 } 1783 1784 if (calls_made_hot > 0) set_major_progress(); 1785 return calls_made_hot; 1786 } 1787 1788 1789 //----------------------------Finish_Warm-------------------------------------- 1790 void Compile::Finish_Warm() { 1791 if (!InlineWarmCalls) return; 1792 if (failing()) return; 1793 if (warm_calls() == NULL) return; 1794 1795 // Clean up loose ends, if we are out of space for inlining. 1796 WarmCallInfo* call; 1797 while ((call = pop_warm_call()) != NULL) { 1798 call->make_cold(); 1799 } 1800 } 1801 1802 //---------------------cleanup_loop_predicates----------------------- 1803 // Remove the opaque nodes that protect the predicates so that all unused 1804 // checks and uncommon_traps will be eliminated from the ideal graph 1805 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1806 if (predicate_count()==0) return; 1807 for (int i = predicate_count(); i > 0; i--) { 1808 Node * n = predicate_opaque1_node(i-1); 1809 assert(n->Opcode() == Op_Opaque1, "must be"); 1810 igvn.replace_node(n, n->in(1)); 1811 } 1812 assert(predicate_count()==0, "should be clean!"); 1813 } 1814 1815 // StringOpts and late inlining of string methods 1816 void Compile::inline_string_calls(bool parse_time) { 1817 { 1818 // remove useless nodes to make the usage analysis simpler 1819 ResourceMark rm; 1820 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1821 } 1822 1823 { 1824 ResourceMark rm; 1825 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1826 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1827 print_method(PHASE_AFTER_STRINGOPTS, 3); 1828 } 1829 1830 // now inline anything that we skipped the first time around 1831 if (!parse_time) { 1832 _late_inlines_pos = _late_inlines.length(); 1833 } 1834 1835 while (_string_late_inlines.length() > 0) { 1836 CallGenerator* cg = _string_late_inlines.pop(); 1837 cg->do_late_inline(); 1838 if (failing()) return; 1839 } 1840 _string_late_inlines.trunc_to(0); 1841 } 1842 1843 // Late inlining of boxing methods 1844 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1845 if (_boxing_late_inlines.length() > 0) { 1846 assert(has_boxed_value(), "inconsistent"); 1847 1848 PhaseGVN* gvn = initial_gvn(); 1849 set_inlining_incrementally(true); 1850 1851 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1852 for_igvn()->clear(); 1853 gvn->replace_with(&igvn); 1854 1855 while (_boxing_late_inlines.length() > 0) { 1856 CallGenerator* cg = _boxing_late_inlines.pop(); 1857 cg->do_late_inline(); 1858 if (failing()) return; 1859 } 1860 _boxing_late_inlines.trunc_to(0); 1861 1862 { 1863 ResourceMark rm; 1864 PhaseRemoveUseless pru(gvn, for_igvn()); 1865 } 1866 1867 igvn = PhaseIterGVN(gvn); 1868 igvn.optimize(); 1869 1870 set_inlining_progress(false); 1871 set_inlining_incrementally(false); 1872 } 1873 } 1874 1875 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1876 assert(IncrementalInline, "incremental inlining should be on"); 1877 PhaseGVN* gvn = initial_gvn(); 1878 1879 set_inlining_progress(false); 1880 for_igvn()->clear(); 1881 gvn->replace_with(&igvn); 1882 1883 int i = 0; 1884 1885 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1886 CallGenerator* cg = _late_inlines.at(i); 1887 _late_inlines_pos = i+1; 1888 cg->do_late_inline(); 1889 if (failing()) return; 1890 } 1891 int j = 0; 1892 for (; i < _late_inlines.length(); i++, j++) { 1893 _late_inlines.at_put(j, _late_inlines.at(i)); 1894 } 1895 _late_inlines.trunc_to(j); 1896 1897 { 1898 ResourceMark rm; 1899 PhaseRemoveUseless pru(gvn, for_igvn()); 1900 } 1901 1902 igvn = PhaseIterGVN(gvn); 1903 } 1904 1905 // Perform incremental inlining until bound on number of live nodes is reached 1906 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1907 PhaseGVN* gvn = initial_gvn(); 1908 1909 set_inlining_incrementally(true); 1910 set_inlining_progress(true); 1911 uint low_live_nodes = 0; 1912 1913 while(inlining_progress() && _late_inlines.length() > 0) { 1914 1915 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1916 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1917 // PhaseIdealLoop is expensive so we only try it once we are 1918 // out of loop and we only try it again if the previous helped 1919 // got the number of nodes down significantly 1920 PhaseIdealLoop ideal_loop( igvn, false, true ); 1921 if (failing()) return; 1922 low_live_nodes = live_nodes(); 1923 _major_progress = true; 1924 } 1925 1926 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1927 break; 1928 } 1929 } 1930 1931 inline_incrementally_one(igvn); 1932 1933 if (failing()) return; 1934 1935 igvn.optimize(); 1936 1937 if (failing()) return; 1938 } 1939 1940 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1941 1942 if (_string_late_inlines.length() > 0) { 1943 assert(has_stringbuilder(), "inconsistent"); 1944 for_igvn()->clear(); 1945 initial_gvn()->replace_with(&igvn); 1946 1947 inline_string_calls(false); 1948 1949 if (failing()) return; 1950 1951 { 1952 ResourceMark rm; 1953 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1954 } 1955 1956 igvn = PhaseIterGVN(gvn); 1957 1958 igvn.optimize(); 1959 } 1960 1961 set_inlining_incrementally(false); 1962 } 1963 1964 1965 //------------------------------Optimize--------------------------------------- 1966 // Given a graph, optimize it. 1967 void Compile::Optimize() { 1968 TracePhase t1("optimizer", &_t_optimizer, true); 1969 1970 #ifndef PRODUCT 1971 if (env()->break_at_compile()) { 1972 BREAKPOINT; 1973 } 1974 1975 #endif 1976 1977 ResourceMark rm; 1978 int loop_opts_cnt; 1979 1980 NOT_PRODUCT( verify_graph_edges(); ) 1981 1982 print_method(PHASE_AFTER_PARSING); 1983 1984 { 1985 // Iterative Global Value Numbering, including ideal transforms 1986 // Initialize IterGVN with types and values from parse-time GVN 1987 PhaseIterGVN igvn(initial_gvn()); 1988 { 1989 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1990 igvn.optimize(); 1991 } 1992 1993 print_method(PHASE_ITER_GVN1, 2); 1994 1995 if (failing()) return; 1996 1997 { 1998 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 1999 inline_incrementally(igvn); 2000 } 2001 2002 print_method(PHASE_INCREMENTAL_INLINE, 2); 2003 2004 if (failing()) return; 2005 2006 if (eliminate_boxing()) { 2007 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2008 // Inline valueOf() methods now. 2009 inline_boxing_calls(igvn); 2010 2011 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2012 2013 if (failing()) return; 2014 } 2015 2016 // No more new expensive nodes will be added to the list from here 2017 // so keep only the actual candidates for optimizations. 2018 cleanup_expensive_nodes(igvn); 2019 2020 // Perform escape analysis 2021 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2022 if (has_loops()) { 2023 // Cleanup graph (remove dead nodes). 2024 TracePhase t2("idealLoop", &_t_idealLoop, true); 2025 PhaseIdealLoop ideal_loop( igvn, false, true ); 2026 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2027 if (failing()) return; 2028 } 2029 ConnectionGraph::do_analysis(this, &igvn); 2030 2031 if (failing()) return; 2032 2033 // Optimize out fields loads from scalar replaceable allocations. 2034 igvn.optimize(); 2035 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2036 2037 if (failing()) return; 2038 2039 if (congraph() != NULL && macro_count() > 0) { 2040 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2041 PhaseMacroExpand mexp(igvn); 2042 mexp.eliminate_macro_nodes(); 2043 igvn.set_delay_transform(false); 2044 2045 igvn.optimize(); 2046 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2047 2048 if (failing()) return; 2049 } 2050 } 2051 2052 // Loop transforms on the ideal graph. Range Check Elimination, 2053 // peeling, unrolling, etc. 2054 2055 // Set loop opts counter 2056 loop_opts_cnt = num_loop_opts(); 2057 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2058 { 2059 TracePhase t2("idealLoop", &_t_idealLoop, true); 2060 PhaseIdealLoop ideal_loop( igvn, true ); 2061 loop_opts_cnt--; 2062 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2063 if (failing()) return; 2064 } 2065 // Loop opts pass if partial peeling occurred in previous pass 2066 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2067 TracePhase t3("idealLoop", &_t_idealLoop, true); 2068 PhaseIdealLoop ideal_loop( igvn, false ); 2069 loop_opts_cnt--; 2070 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2071 if (failing()) return; 2072 } 2073 // Loop opts pass for loop-unrolling before CCP 2074 if(major_progress() && (loop_opts_cnt > 0)) { 2075 TracePhase t4("idealLoop", &_t_idealLoop, true); 2076 PhaseIdealLoop ideal_loop( igvn, false ); 2077 loop_opts_cnt--; 2078 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2079 } 2080 if (!failing()) { 2081 // Verify that last round of loop opts produced a valid graph 2082 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2083 PhaseIdealLoop::verify(igvn); 2084 } 2085 } 2086 if (failing()) return; 2087 2088 // Conditional Constant Propagation; 2089 PhaseCCP ccp( &igvn ); 2090 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2091 { 2092 TracePhase t2("ccp", &_t_ccp, true); 2093 ccp.do_transform(); 2094 } 2095 print_method(PHASE_CPP1, 2); 2096 2097 assert( true, "Break here to ccp.dump_old2new_map()"); 2098 2099 // Iterative Global Value Numbering, including ideal transforms 2100 { 2101 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2102 igvn = ccp; 2103 igvn.optimize(); 2104 } 2105 2106 print_method(PHASE_ITER_GVN2, 2); 2107 2108 if (failing()) return; 2109 2110 // Loop transforms on the ideal graph. Range Check Elimination, 2111 // peeling, unrolling, etc. 2112 if(loop_opts_cnt > 0) { 2113 debug_only( int cnt = 0; ); 2114 while(major_progress() && (loop_opts_cnt > 0)) { 2115 TracePhase t2("idealLoop", &_t_idealLoop, true); 2116 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2117 PhaseIdealLoop ideal_loop( igvn, true); 2118 loop_opts_cnt--; 2119 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2120 if (failing()) return; 2121 } 2122 } 2123 2124 { 2125 // Verify that all previous optimizations produced a valid graph 2126 // at least to this point, even if no loop optimizations were done. 2127 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2128 PhaseIdealLoop::verify(igvn); 2129 } 2130 2131 { 2132 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2133 PhaseMacroExpand mex(igvn); 2134 if (mex.expand_macro_nodes()) { 2135 assert(failing(), "must bail out w/ explicit message"); 2136 return; 2137 } 2138 } 2139 2140 } // (End scope of igvn; run destructor if necessary for asserts.) 2141 2142 dump_inlining(); 2143 // A method with only infinite loops has no edges entering loops from root 2144 { 2145 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2146 if (final_graph_reshaping()) { 2147 assert(failing(), "must bail out w/ explicit message"); 2148 return; 2149 } 2150 } 2151 2152 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2153 } 2154 2155 2156 //------------------------------Code_Gen--------------------------------------- 2157 // Given a graph, generate code for it 2158 void Compile::Code_Gen() { 2159 if (failing()) { 2160 return; 2161 } 2162 2163 // Perform instruction selection. You might think we could reclaim Matcher 2164 // memory PDQ, but actually the Matcher is used in generating spill code. 2165 // Internals of the Matcher (including some VectorSets) must remain live 2166 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2167 // set a bit in reclaimed memory. 2168 2169 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2170 // nodes. Mapping is only valid at the root of each matched subtree. 2171 NOT_PRODUCT( verify_graph_edges(); ) 2172 2173 Matcher matcher; 2174 _matcher = &matcher; 2175 { 2176 TracePhase t2("matcher", &_t_matcher, true); 2177 matcher.match(); 2178 } 2179 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2180 // nodes. Mapping is only valid at the root of each matched subtree. 2181 NOT_PRODUCT( verify_graph_edges(); ) 2182 2183 // If you have too many nodes, or if matching has failed, bail out 2184 check_node_count(0, "out of nodes matching instructions"); 2185 if (failing()) { 2186 return; 2187 } 2188 2189 // Build a proper-looking CFG 2190 PhaseCFG cfg(node_arena(), root(), matcher); 2191 _cfg = &cfg; 2192 { 2193 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2194 bool success = cfg.do_global_code_motion(); 2195 if (!success) { 2196 return; 2197 } 2198 2199 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2200 NOT_PRODUCT( verify_graph_edges(); ) 2201 debug_only( cfg.verify(); ) 2202 } 2203 2204 PhaseChaitin regalloc(unique(), cfg, matcher); 2205 _regalloc = ®alloc; 2206 { 2207 TracePhase t2("regalloc", &_t_registerAllocation, true); 2208 // Perform register allocation. After Chaitin, use-def chains are 2209 // no longer accurate (at spill code) and so must be ignored. 2210 // Node->LRG->reg mappings are still accurate. 2211 _regalloc->Register_Allocate(); 2212 2213 // Bail out if the allocator builds too many nodes 2214 if (failing()) { 2215 return; 2216 } 2217 } 2218 2219 // Prior to register allocation we kept empty basic blocks in case the 2220 // the allocator needed a place to spill. After register allocation we 2221 // are not adding any new instructions. If any basic block is empty, we 2222 // can now safely remove it. 2223 { 2224 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2225 cfg.remove_empty_blocks(); 2226 if (do_freq_based_layout()) { 2227 PhaseBlockLayout layout(cfg); 2228 } else { 2229 cfg.set_loop_alignment(); 2230 } 2231 cfg.fixup_flow(); 2232 } 2233 2234 // Apply peephole optimizations 2235 if( OptoPeephole ) { 2236 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2237 PhasePeephole peep( _regalloc, cfg); 2238 peep.do_transform(); 2239 } 2240 2241 // Convert Nodes to instruction bits in a buffer 2242 { 2243 // %%%% workspace merge brought two timers together for one job 2244 TracePhase t2a("output", &_t_output, true); 2245 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2246 Output(); 2247 } 2248 2249 print_method(PHASE_FINAL_CODE); 2250 2251 // He's dead, Jim. 2252 _cfg = (PhaseCFG*)0xdeadbeef; 2253 _regalloc = (PhaseChaitin*)0xdeadbeef; 2254 } 2255 2256 2257 //------------------------------dump_asm--------------------------------------- 2258 // Dump formatted assembly 2259 #ifndef PRODUCT 2260 void Compile::dump_asm(int *pcs, uint pc_limit) { 2261 bool cut_short = false; 2262 tty->print_cr("#"); 2263 tty->print("# "); _tf->dump(); tty->cr(); 2264 tty->print_cr("#"); 2265 2266 // For all blocks 2267 int pc = 0x0; // Program counter 2268 char starts_bundle = ' '; 2269 _regalloc->dump_frame(); 2270 2271 Node *n = NULL; 2272 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2273 if (VMThread::should_terminate()) { 2274 cut_short = true; 2275 break; 2276 } 2277 Block* block = _cfg->get_block(i); 2278 if (block->is_connector() && !Verbose) { 2279 continue; 2280 } 2281 n = block->head(); 2282 if (pcs && n->_idx < pc_limit) { 2283 tty->print("%3.3x ", pcs[n->_idx]); 2284 } else { 2285 tty->print(" "); 2286 } 2287 block->dump_head(_cfg); 2288 if (block->is_connector()) { 2289 tty->print_cr(" # Empty connector block"); 2290 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2291 tty->print_cr(" # Block is sole successor of call"); 2292 } 2293 2294 // For all instructions 2295 Node *delay = NULL; 2296 for (uint j = 0; j < block->number_of_nodes(); j++) { 2297 if (VMThread::should_terminate()) { 2298 cut_short = true; 2299 break; 2300 } 2301 n = block->get_node(j); 2302 if (valid_bundle_info(n)) { 2303 Bundle* bundle = node_bundling(n); 2304 if (bundle->used_in_unconditional_delay()) { 2305 delay = n; 2306 continue; 2307 } 2308 if (bundle->starts_bundle()) { 2309 starts_bundle = '+'; 2310 } 2311 } 2312 2313 if (WizardMode) { 2314 n->dump(); 2315 } 2316 2317 if( !n->is_Region() && // Dont print in the Assembly 2318 !n->is_Phi() && // a few noisely useless nodes 2319 !n->is_Proj() && 2320 !n->is_MachTemp() && 2321 !n->is_SafePointScalarObject() && 2322 !n->is_Catch() && // Would be nice to print exception table targets 2323 !n->is_MergeMem() && // Not very interesting 2324 !n->is_top() && // Debug info table constants 2325 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2326 ) { 2327 if (pcs && n->_idx < pc_limit) 2328 tty->print("%3.3x", pcs[n->_idx]); 2329 else 2330 tty->print(" "); 2331 tty->print(" %c ", starts_bundle); 2332 starts_bundle = ' '; 2333 tty->print("\t"); 2334 n->format(_regalloc, tty); 2335 tty->cr(); 2336 } 2337 2338 // If we have an instruction with a delay slot, and have seen a delay, 2339 // then back up and print it 2340 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2341 assert(delay != NULL, "no unconditional delay instruction"); 2342 if (WizardMode) delay->dump(); 2343 2344 if (node_bundling(delay)->starts_bundle()) 2345 starts_bundle = '+'; 2346 if (pcs && n->_idx < pc_limit) 2347 tty->print("%3.3x", pcs[n->_idx]); 2348 else 2349 tty->print(" "); 2350 tty->print(" %c ", starts_bundle); 2351 starts_bundle = ' '; 2352 tty->print("\t"); 2353 delay->format(_regalloc, tty); 2354 tty->print_cr(""); 2355 delay = NULL; 2356 } 2357 2358 // Dump the exception table as well 2359 if( n->is_Catch() && (Verbose || WizardMode) ) { 2360 // Print the exception table for this offset 2361 _handler_table.print_subtable_for(pc); 2362 } 2363 } 2364 2365 if (pcs && n->_idx < pc_limit) 2366 tty->print_cr("%3.3x", pcs[n->_idx]); 2367 else 2368 tty->print_cr(""); 2369 2370 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2371 2372 } // End of per-block dump 2373 tty->print_cr(""); 2374 2375 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2376 } 2377 #endif 2378 2379 //------------------------------Final_Reshape_Counts--------------------------- 2380 // This class defines counters to help identify when a method 2381 // may/must be executed using hardware with only 24-bit precision. 2382 struct Final_Reshape_Counts : public StackObj { 2383 int _call_count; // count non-inlined 'common' calls 2384 int _float_count; // count float ops requiring 24-bit precision 2385 int _double_count; // count double ops requiring more precision 2386 int _java_call_count; // count non-inlined 'java' calls 2387 int _inner_loop_count; // count loops which need alignment 2388 VectorSet _visited; // Visitation flags 2389 Node_List _tests; // Set of IfNodes & PCTableNodes 2390 2391 Final_Reshape_Counts() : 2392 _call_count(0), _float_count(0), _double_count(0), 2393 _java_call_count(0), _inner_loop_count(0), 2394 _visited( Thread::current()->resource_area() ) { } 2395 2396 void inc_call_count () { _call_count ++; } 2397 void inc_float_count () { _float_count ++; } 2398 void inc_double_count() { _double_count++; } 2399 void inc_java_call_count() { _java_call_count++; } 2400 void inc_inner_loop_count() { _inner_loop_count++; } 2401 2402 int get_call_count () const { return _call_count ; } 2403 int get_float_count () const { return _float_count ; } 2404 int get_double_count() const { return _double_count; } 2405 int get_java_call_count() const { return _java_call_count; } 2406 int get_inner_loop_count() const { return _inner_loop_count; } 2407 }; 2408 2409 #ifdef ASSERT 2410 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2411 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2412 // Make sure the offset goes inside the instance layout. 2413 return k->contains_field_offset(tp->offset()); 2414 // Note that OffsetBot and OffsetTop are very negative. 2415 } 2416 #endif 2417 2418 // Eliminate trivially redundant StoreCMs and accumulate their 2419 // precedence edges. 2420 void Compile::eliminate_redundant_card_marks(Node* n) { 2421 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2422 if (n->in(MemNode::Address)->outcnt() > 1) { 2423 // There are multiple users of the same address so it might be 2424 // possible to eliminate some of the StoreCMs 2425 Node* mem = n->in(MemNode::Memory); 2426 Node* adr = n->in(MemNode::Address); 2427 Node* val = n->in(MemNode::ValueIn); 2428 Node* prev = n; 2429 bool done = false; 2430 // Walk the chain of StoreCMs eliminating ones that match. As 2431 // long as it's a chain of single users then the optimization is 2432 // safe. Eliminating partially redundant StoreCMs would require 2433 // cloning copies down the other paths. 2434 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2435 if (adr == mem->in(MemNode::Address) && 2436 val == mem->in(MemNode::ValueIn)) { 2437 // redundant StoreCM 2438 if (mem->req() > MemNode::OopStore) { 2439 // Hasn't been processed by this code yet. 2440 n->add_prec(mem->in(MemNode::OopStore)); 2441 } else { 2442 // Already converted to precedence edge 2443 for (uint i = mem->req(); i < mem->len(); i++) { 2444 // Accumulate any precedence edges 2445 if (mem->in(i) != NULL) { 2446 n->add_prec(mem->in(i)); 2447 } 2448 } 2449 // Everything above this point has been processed. 2450 done = true; 2451 } 2452 // Eliminate the previous StoreCM 2453 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2454 assert(mem->outcnt() == 0, "should be dead"); 2455 mem->disconnect_inputs(NULL, this); 2456 } else { 2457 prev = mem; 2458 } 2459 mem = prev->in(MemNode::Memory); 2460 } 2461 } 2462 } 2463 2464 //------------------------------final_graph_reshaping_impl---------------------- 2465 // Implement items 1-5 from final_graph_reshaping below. 2466 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2467 2468 if ( n->outcnt() == 0 ) return; // dead node 2469 uint nop = n->Opcode(); 2470 2471 // Check for 2-input instruction with "last use" on right input. 2472 // Swap to left input. Implements item (2). 2473 if( n->req() == 3 && // two-input instruction 2474 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2475 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2476 n->in(2)->outcnt() == 1 &&// right use IS a last use 2477 !n->in(2)->is_Con() ) { // right use is not a constant 2478 // Check for commutative opcode 2479 switch( nop ) { 2480 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2481 case Op_MaxI: case Op_MinI: 2482 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2483 case Op_AndL: case Op_XorL: case Op_OrL: 2484 case Op_AndI: case Op_XorI: case Op_OrI: { 2485 // Move "last use" input to left by swapping inputs 2486 n->swap_edges(1, 2); 2487 break; 2488 } 2489 default: 2490 break; 2491 } 2492 } 2493 2494 #ifdef ASSERT 2495 if( n->is_Mem() ) { 2496 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2497 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2498 // oop will be recorded in oop map if load crosses safepoint 2499 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2500 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2501 "raw memory operations should have control edge"); 2502 } 2503 #endif 2504 // Count FPU ops and common calls, implements item (3) 2505 switch( nop ) { 2506 // Count all float operations that may use FPU 2507 case Op_AddF: 2508 case Op_SubF: 2509 case Op_MulF: 2510 case Op_DivF: 2511 case Op_NegF: 2512 case Op_ModF: 2513 case Op_ConvI2F: 2514 case Op_ConF: 2515 case Op_CmpF: 2516 case Op_CmpF3: 2517 // case Op_ConvL2F: // longs are split into 32-bit halves 2518 frc.inc_float_count(); 2519 break; 2520 2521 case Op_ConvF2D: 2522 case Op_ConvD2F: 2523 frc.inc_float_count(); 2524 frc.inc_double_count(); 2525 break; 2526 2527 // Count all double operations that may use FPU 2528 case Op_AddD: 2529 case Op_SubD: 2530 case Op_MulD: 2531 case Op_DivD: 2532 case Op_NegD: 2533 case Op_ModD: 2534 case Op_ConvI2D: 2535 case Op_ConvD2I: 2536 // case Op_ConvL2D: // handled by leaf call 2537 // case Op_ConvD2L: // handled by leaf call 2538 case Op_ConD: 2539 case Op_CmpD: 2540 case Op_CmpD3: 2541 frc.inc_double_count(); 2542 break; 2543 case Op_Opaque1: // Remove Opaque Nodes before matching 2544 case Op_Opaque2: // Remove Opaque Nodes before matching 2545 n->subsume_by(n->in(1), this); 2546 break; 2547 case Op_CallStaticJava: 2548 case Op_CallJava: 2549 case Op_CallDynamicJava: 2550 frc.inc_java_call_count(); // Count java call site; 2551 case Op_CallRuntime: 2552 case Op_CallLeaf: 2553 case Op_CallLeafNoFP: { 2554 assert( n->is_Call(), "" ); 2555 CallNode *call = n->as_Call(); 2556 // Count call sites where the FP mode bit would have to be flipped. 2557 // Do not count uncommon runtime calls: 2558 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2559 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2560 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2561 frc.inc_call_count(); // Count the call site 2562 } else { // See if uncommon argument is shared 2563 Node *n = call->in(TypeFunc::Parms); 2564 int nop = n->Opcode(); 2565 // Clone shared simple arguments to uncommon calls, item (1). 2566 if( n->outcnt() > 1 && 2567 !n->is_Proj() && 2568 nop != Op_CreateEx && 2569 nop != Op_CheckCastPP && 2570 nop != Op_DecodeN && 2571 nop != Op_DecodeNKlass && 2572 !n->is_Mem() ) { 2573 Node *x = n->clone(); 2574 call->set_req( TypeFunc::Parms, x ); 2575 } 2576 } 2577 break; 2578 } 2579 2580 case Op_StoreD: 2581 case Op_LoadD: 2582 case Op_LoadD_unaligned: 2583 frc.inc_double_count(); 2584 goto handle_mem; 2585 case Op_StoreF: 2586 case Op_LoadF: 2587 frc.inc_float_count(); 2588 goto handle_mem; 2589 2590 case Op_StoreCM: 2591 { 2592 // Convert OopStore dependence into precedence edge 2593 Node* prec = n->in(MemNode::OopStore); 2594 n->del_req(MemNode::OopStore); 2595 n->add_prec(prec); 2596 eliminate_redundant_card_marks(n); 2597 } 2598 2599 // fall through 2600 2601 case Op_StoreB: 2602 case Op_StoreC: 2603 case Op_StorePConditional: 2604 case Op_StoreI: 2605 case Op_StoreL: 2606 case Op_StoreIConditional: 2607 case Op_StoreLConditional: 2608 case Op_CompareAndSwapI: 2609 case Op_CompareAndSwapL: 2610 case Op_CompareAndSwapP: 2611 case Op_CompareAndSwapN: 2612 case Op_GetAndAddI: 2613 case Op_GetAndAddL: 2614 case Op_GetAndSetI: 2615 case Op_GetAndSetL: 2616 case Op_GetAndSetP: 2617 case Op_GetAndSetN: 2618 case Op_StoreP: 2619 case Op_StoreN: 2620 case Op_StoreNKlass: 2621 case Op_LoadB: 2622 case Op_LoadUB: 2623 case Op_LoadUS: 2624 case Op_LoadI: 2625 case Op_LoadKlass: 2626 case Op_LoadNKlass: 2627 case Op_LoadL: 2628 case Op_LoadL_unaligned: 2629 case Op_LoadPLocked: 2630 case Op_LoadP: 2631 case Op_LoadN: 2632 case Op_LoadRange: 2633 case Op_LoadS: { 2634 handle_mem: 2635 #ifdef ASSERT 2636 if( VerifyOptoOopOffsets ) { 2637 assert( n->is_Mem(), "" ); 2638 MemNode *mem = (MemNode*)n; 2639 // Check to see if address types have grounded out somehow. 2640 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2641 assert( !tp || oop_offset_is_sane(tp), "" ); 2642 } 2643 #endif 2644 break; 2645 } 2646 2647 case Op_AddP: { // Assert sane base pointers 2648 Node *addp = n->in(AddPNode::Address); 2649 assert( !addp->is_AddP() || 2650 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2651 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2652 "Base pointers must match" ); 2653 #ifdef _LP64 2654 if ((UseCompressedOops || UseCompressedClassPointers) && 2655 addp->Opcode() == Op_ConP && 2656 addp == n->in(AddPNode::Base) && 2657 n->in(AddPNode::Offset)->is_Con()) { 2658 // Use addressing with narrow klass to load with offset on x86. 2659 // On sparc loading 32-bits constant and decoding it have less 2660 // instructions (4) then load 64-bits constant (7). 2661 // Do this transformation here since IGVN will convert ConN back to ConP. 2662 const Type* t = addp->bottom_type(); 2663 if (t->isa_oopptr() || t->isa_klassptr()) { 2664 Node* nn = NULL; 2665 2666 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2667 2668 // Look for existing ConN node of the same exact type. 2669 Node* r = root(); 2670 uint cnt = r->outcnt(); 2671 for (uint i = 0; i < cnt; i++) { 2672 Node* m = r->raw_out(i); 2673 if (m!= NULL && m->Opcode() == op && 2674 m->bottom_type()->make_ptr() == t) { 2675 nn = m; 2676 break; 2677 } 2678 } 2679 if (nn != NULL) { 2680 // Decode a narrow oop to match address 2681 // [R12 + narrow_oop_reg<<3 + offset] 2682 if (t->isa_oopptr()) { 2683 nn = new (this) DecodeNNode(nn, t); 2684 } else { 2685 nn = new (this) DecodeNKlassNode(nn, t); 2686 } 2687 n->set_req(AddPNode::Base, nn); 2688 n->set_req(AddPNode::Address, nn); 2689 if (addp->outcnt() == 0) { 2690 addp->disconnect_inputs(NULL, this); 2691 } 2692 } 2693 } 2694 } 2695 #endif 2696 break; 2697 } 2698 2699 #ifdef _LP64 2700 case Op_CastPP: 2701 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2702 Node* in1 = n->in(1); 2703 const Type* t = n->bottom_type(); 2704 Node* new_in1 = in1->clone(); 2705 new_in1->as_DecodeN()->set_type(t); 2706 2707 if (!Matcher::narrow_oop_use_complex_address()) { 2708 // 2709 // x86, ARM and friends can handle 2 adds in addressing mode 2710 // and Matcher can fold a DecodeN node into address by using 2711 // a narrow oop directly and do implicit NULL check in address: 2712 // 2713 // [R12 + narrow_oop_reg<<3 + offset] 2714 // NullCheck narrow_oop_reg 2715 // 2716 // On other platforms (Sparc) we have to keep new DecodeN node and 2717 // use it to do implicit NULL check in address: 2718 // 2719 // decode_not_null narrow_oop_reg, base_reg 2720 // [base_reg + offset] 2721 // NullCheck base_reg 2722 // 2723 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2724 // to keep the information to which NULL check the new DecodeN node 2725 // corresponds to use it as value in implicit_null_check(). 2726 // 2727 new_in1->set_req(0, n->in(0)); 2728 } 2729 2730 n->subsume_by(new_in1, this); 2731 if (in1->outcnt() == 0) { 2732 in1->disconnect_inputs(NULL, this); 2733 } 2734 } 2735 break; 2736 2737 case Op_CmpP: 2738 // Do this transformation here to preserve CmpPNode::sub() and 2739 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2740 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2741 Node* in1 = n->in(1); 2742 Node* in2 = n->in(2); 2743 if (!in1->is_DecodeNarrowPtr()) { 2744 in2 = in1; 2745 in1 = n->in(2); 2746 } 2747 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2748 2749 Node* new_in2 = NULL; 2750 if (in2->is_DecodeNarrowPtr()) { 2751 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2752 new_in2 = in2->in(1); 2753 } else if (in2->Opcode() == Op_ConP) { 2754 const Type* t = in2->bottom_type(); 2755 if (t == TypePtr::NULL_PTR) { 2756 assert(in1->is_DecodeN(), "compare klass to null?"); 2757 // Don't convert CmpP null check into CmpN if compressed 2758 // oops implicit null check is not generated. 2759 // This will allow to generate normal oop implicit null check. 2760 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2761 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2762 // 2763 // This transformation together with CastPP transformation above 2764 // will generated code for implicit NULL checks for compressed oops. 2765 // 2766 // The original code after Optimize() 2767 // 2768 // LoadN memory, narrow_oop_reg 2769 // decode narrow_oop_reg, base_reg 2770 // CmpP base_reg, NULL 2771 // CastPP base_reg // NotNull 2772 // Load [base_reg + offset], val_reg 2773 // 2774 // after these transformations will be 2775 // 2776 // LoadN memory, narrow_oop_reg 2777 // CmpN narrow_oop_reg, NULL 2778 // decode_not_null narrow_oop_reg, base_reg 2779 // Load [base_reg + offset], val_reg 2780 // 2781 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2782 // since narrow oops can be used in debug info now (see the code in 2783 // final_graph_reshaping_walk()). 2784 // 2785 // At the end the code will be matched to 2786 // on x86: 2787 // 2788 // Load_narrow_oop memory, narrow_oop_reg 2789 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2790 // NullCheck narrow_oop_reg 2791 // 2792 // and on sparc: 2793 // 2794 // Load_narrow_oop memory, narrow_oop_reg 2795 // decode_not_null narrow_oop_reg, base_reg 2796 // Load [base_reg + offset], val_reg 2797 // NullCheck base_reg 2798 // 2799 } else if (t->isa_oopptr()) { 2800 new_in2 = ConNode::make(this, t->make_narrowoop()); 2801 } else if (t->isa_klassptr()) { 2802 new_in2 = ConNode::make(this, t->make_narrowklass()); 2803 } 2804 } 2805 if (new_in2 != NULL) { 2806 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 2807 n->subsume_by(cmpN, this); 2808 if (in1->outcnt() == 0) { 2809 in1->disconnect_inputs(NULL, this); 2810 } 2811 if (in2->outcnt() == 0) { 2812 in2->disconnect_inputs(NULL, this); 2813 } 2814 } 2815 } 2816 break; 2817 2818 case Op_DecodeN: 2819 case Op_DecodeNKlass: 2820 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2821 // DecodeN could be pinned when it can't be fold into 2822 // an address expression, see the code for Op_CastPP above. 2823 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2824 break; 2825 2826 case Op_EncodeP: 2827 case Op_EncodePKlass: { 2828 Node* in1 = n->in(1); 2829 if (in1->is_DecodeNarrowPtr()) { 2830 n->subsume_by(in1->in(1), this); 2831 } else if (in1->Opcode() == Op_ConP) { 2832 const Type* t = in1->bottom_type(); 2833 if (t == TypePtr::NULL_PTR) { 2834 assert(t->isa_oopptr(), "null klass?"); 2835 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2836 } else if (t->isa_oopptr()) { 2837 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2838 } else if (t->isa_klassptr()) { 2839 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2840 } 2841 } 2842 if (in1->outcnt() == 0) { 2843 in1->disconnect_inputs(NULL, this); 2844 } 2845 break; 2846 } 2847 2848 case Op_Proj: { 2849 if (OptimizeStringConcat) { 2850 ProjNode* p = n->as_Proj(); 2851 if (p->_is_io_use) { 2852 // Separate projections were used for the exception path which 2853 // are normally removed by a late inline. If it wasn't inlined 2854 // then they will hang around and should just be replaced with 2855 // the original one. 2856 Node* proj = NULL; 2857 // Replace with just one 2858 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2859 Node *use = i.get(); 2860 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2861 proj = use; 2862 break; 2863 } 2864 } 2865 assert(proj != NULL, "must be found"); 2866 p->subsume_by(proj, this); 2867 } 2868 } 2869 break; 2870 } 2871 2872 case Op_Phi: 2873 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2874 // The EncodeP optimization may create Phi with the same edges 2875 // for all paths. It is not handled well by Register Allocator. 2876 Node* unique_in = n->in(1); 2877 assert(unique_in != NULL, ""); 2878 uint cnt = n->req(); 2879 for (uint i = 2; i < cnt; i++) { 2880 Node* m = n->in(i); 2881 assert(m != NULL, ""); 2882 if (unique_in != m) 2883 unique_in = NULL; 2884 } 2885 if (unique_in != NULL) { 2886 n->subsume_by(unique_in, this); 2887 } 2888 } 2889 break; 2890 2891 #endif 2892 2893 case Op_ModI: 2894 if (UseDivMod) { 2895 // Check if a%b and a/b both exist 2896 Node* d = n->find_similar(Op_DivI); 2897 if (d) { 2898 // Replace them with a fused divmod if supported 2899 if (Matcher::has_match_rule(Op_DivModI)) { 2900 DivModINode* divmod = DivModINode::make(this, n); 2901 d->subsume_by(divmod->div_proj(), this); 2902 n->subsume_by(divmod->mod_proj(), this); 2903 } else { 2904 // replace a%b with a-((a/b)*b) 2905 Node* mult = new (this) MulINode(d, d->in(2)); 2906 Node* sub = new (this) SubINode(d->in(1), mult); 2907 n->subsume_by(sub, this); 2908 } 2909 } 2910 } 2911 break; 2912 2913 case Op_ModL: 2914 if (UseDivMod) { 2915 // Check if a%b and a/b both exist 2916 Node* d = n->find_similar(Op_DivL); 2917 if (d) { 2918 // Replace them with a fused divmod if supported 2919 if (Matcher::has_match_rule(Op_DivModL)) { 2920 DivModLNode* divmod = DivModLNode::make(this, n); 2921 d->subsume_by(divmod->div_proj(), this); 2922 n->subsume_by(divmod->mod_proj(), this); 2923 } else { 2924 // replace a%b with a-((a/b)*b) 2925 Node* mult = new (this) MulLNode(d, d->in(2)); 2926 Node* sub = new (this) SubLNode(d->in(1), mult); 2927 n->subsume_by(sub, this); 2928 } 2929 } 2930 } 2931 break; 2932 2933 case Op_LoadVector: 2934 case Op_StoreVector: 2935 break; 2936 2937 case Op_PackB: 2938 case Op_PackS: 2939 case Op_PackI: 2940 case Op_PackF: 2941 case Op_PackL: 2942 case Op_PackD: 2943 if (n->req()-1 > 2) { 2944 // Replace many operand PackNodes with a binary tree for matching 2945 PackNode* p = (PackNode*) n; 2946 Node* btp = p->binary_tree_pack(this, 1, n->req()); 2947 n->subsume_by(btp, this); 2948 } 2949 break; 2950 case Op_Loop: 2951 case Op_CountedLoop: 2952 if (n->as_Loop()->is_inner_loop()) { 2953 frc.inc_inner_loop_count(); 2954 } 2955 break; 2956 case Op_LShiftI: 2957 case Op_RShiftI: 2958 case Op_URShiftI: 2959 case Op_LShiftL: 2960 case Op_RShiftL: 2961 case Op_URShiftL: 2962 if (Matcher::need_masked_shift_count) { 2963 // The cpu's shift instructions don't restrict the count to the 2964 // lower 5/6 bits. We need to do the masking ourselves. 2965 Node* in2 = n->in(2); 2966 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 2967 const TypeInt* t = in2->find_int_type(); 2968 if (t != NULL && t->is_con()) { 2969 juint shift = t->get_con(); 2970 if (shift > mask) { // Unsigned cmp 2971 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 2972 } 2973 } else { 2974 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 2975 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 2976 n->set_req(2, shift); 2977 } 2978 } 2979 if (in2->outcnt() == 0) { // Remove dead node 2980 in2->disconnect_inputs(NULL, this); 2981 } 2982 } 2983 break; 2984 case Op_MemBarStoreStore: 2985 case Op_MemBarRelease: 2986 // Break the link with AllocateNode: it is no longer useful and 2987 // confuses register allocation. 2988 if (n->req() > MemBarNode::Precedent) { 2989 n->set_req(MemBarNode::Precedent, top()); 2990 } 2991 break; 2992 // Must set a control edge on all nodes that produce a FlagsProj 2993 // so they can't escape the block that consumes the flags. 2994 // Must also set the non throwing branch as the control 2995 // for all nodes that depends on the result. Unless the node 2996 // already have a control that isn't the control of the 2997 // flag producer 2998 case Op_FlagsProj: 2999 { 3000 MathExactNode* math = (MathExactNode*) n->in(0); 3001 Node* ctrl = math->control_node(); 3002 Node* non_throwing = math->non_throwing_branch(); 3003 math->set_req(0, ctrl); 3004 3005 Node* result = math->result_node(); 3006 if (result != NULL) { 3007 for (DUIterator_Fast jmax, j = result->fast_outs(jmax); j < jmax; j++) { 3008 Node* out = result->fast_out(j); 3009 if (out->in(0) == NULL) { 3010 out->set_req(0, non_throwing); 3011 } else if (out->in(0) == ctrl) { 3012 out->set_req(0, non_throwing); 3013 } 3014 } 3015 } 3016 } 3017 break; 3018 default: 3019 assert( !n->is_Call(), "" ); 3020 assert( !n->is_Mem(), "" ); 3021 break; 3022 } 3023 3024 // Collect CFG split points 3025 if (n->is_MultiBranch()) 3026 frc._tests.push(n); 3027 } 3028 3029 //------------------------------final_graph_reshaping_walk--------------------- 3030 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3031 // requires that the walk visits a node's inputs before visiting the node. 3032 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3033 ResourceArea *area = Thread::current()->resource_area(); 3034 Unique_Node_List sfpt(area); 3035 3036 frc._visited.set(root->_idx); // first, mark node as visited 3037 uint cnt = root->req(); 3038 Node *n = root; 3039 uint i = 0; 3040 while (true) { 3041 if (i < cnt) { 3042 // Place all non-visited non-null inputs onto stack 3043 Node* m = n->in(i); 3044 ++i; 3045 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3046 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 3047 sfpt.push(m); 3048 cnt = m->req(); 3049 nstack.push(n, i); // put on stack parent and next input's index 3050 n = m; 3051 i = 0; 3052 } 3053 } else { 3054 // Now do post-visit work 3055 final_graph_reshaping_impl( n, frc ); 3056 if (nstack.is_empty()) 3057 break; // finished 3058 n = nstack.node(); // Get node from stack 3059 cnt = n->req(); 3060 i = nstack.index(); 3061 nstack.pop(); // Shift to the next node on stack 3062 } 3063 } 3064 3065 // Skip next transformation if compressed oops are not used. 3066 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3067 (!UseCompressedOops && !UseCompressedClassPointers)) 3068 return; 3069 3070 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3071 // It could be done for an uncommon traps or any safepoints/calls 3072 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3073 while (sfpt.size() > 0) { 3074 n = sfpt.pop(); 3075 JVMState *jvms = n->as_SafePoint()->jvms(); 3076 assert(jvms != NULL, "sanity"); 3077 int start = jvms->debug_start(); 3078 int end = n->req(); 3079 bool is_uncommon = (n->is_CallStaticJava() && 3080 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3081 for (int j = start; j < end; j++) { 3082 Node* in = n->in(j); 3083 if (in->is_DecodeNarrowPtr()) { 3084 bool safe_to_skip = true; 3085 if (!is_uncommon ) { 3086 // Is it safe to skip? 3087 for (uint i = 0; i < in->outcnt(); i++) { 3088 Node* u = in->raw_out(i); 3089 if (!u->is_SafePoint() || 3090 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3091 safe_to_skip = false; 3092 } 3093 } 3094 } 3095 if (safe_to_skip) { 3096 n->set_req(j, in->in(1)); 3097 } 3098 if (in->outcnt() == 0) { 3099 in->disconnect_inputs(NULL, this); 3100 } 3101 } 3102 } 3103 } 3104 } 3105 3106 //------------------------------final_graph_reshaping-------------------------- 3107 // Final Graph Reshaping. 3108 // 3109 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3110 // and not commoned up and forced early. Must come after regular 3111 // optimizations to avoid GVN undoing the cloning. Clone constant 3112 // inputs to Loop Phis; these will be split by the allocator anyways. 3113 // Remove Opaque nodes. 3114 // (2) Move last-uses by commutative operations to the left input to encourage 3115 // Intel update-in-place two-address operations and better register usage 3116 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3117 // calls canonicalizing them back. 3118 // (3) Count the number of double-precision FP ops, single-precision FP ops 3119 // and call sites. On Intel, we can get correct rounding either by 3120 // forcing singles to memory (requires extra stores and loads after each 3121 // FP bytecode) or we can set a rounding mode bit (requires setting and 3122 // clearing the mode bit around call sites). The mode bit is only used 3123 // if the relative frequency of single FP ops to calls is low enough. 3124 // This is a key transform for SPEC mpeg_audio. 3125 // (4) Detect infinite loops; blobs of code reachable from above but not 3126 // below. Several of the Code_Gen algorithms fail on such code shapes, 3127 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3128 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3129 // Detection is by looking for IfNodes where only 1 projection is 3130 // reachable from below or CatchNodes missing some targets. 3131 // (5) Assert for insane oop offsets in debug mode. 3132 3133 bool Compile::final_graph_reshaping() { 3134 // an infinite loop may have been eliminated by the optimizer, 3135 // in which case the graph will be empty. 3136 if (root()->req() == 1) { 3137 record_method_not_compilable("trivial infinite loop"); 3138 return true; 3139 } 3140 3141 // Expensive nodes have their control input set to prevent the GVN 3142 // from freely commoning them. There's no GVN beyond this point so 3143 // no need to keep the control input. We want the expensive nodes to 3144 // be freely moved to the least frequent code path by gcm. 3145 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3146 for (int i = 0; i < expensive_count(); i++) { 3147 _expensive_nodes->at(i)->set_req(0, NULL); 3148 } 3149 3150 Final_Reshape_Counts frc; 3151 3152 // Visit everybody reachable! 3153 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3154 Node_Stack nstack(unique() >> 1); 3155 final_graph_reshaping_walk(nstack, root(), frc); 3156 3157 // Check for unreachable (from below) code (i.e., infinite loops). 3158 for( uint i = 0; i < frc._tests.size(); i++ ) { 3159 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3160 // Get number of CFG targets. 3161 // Note that PCTables include exception targets after calls. 3162 uint required_outcnt = n->required_outcnt(); 3163 if (n->outcnt() != required_outcnt) { 3164 // Check for a few special cases. Rethrow Nodes never take the 3165 // 'fall-thru' path, so expected kids is 1 less. 3166 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3167 if (n->in(0)->in(0)->is_Call()) { 3168 CallNode *call = n->in(0)->in(0)->as_Call(); 3169 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3170 required_outcnt--; // Rethrow always has 1 less kid 3171 } else if (call->req() > TypeFunc::Parms && 3172 call->is_CallDynamicJava()) { 3173 // Check for null receiver. In such case, the optimizer has 3174 // detected that the virtual call will always result in a null 3175 // pointer exception. The fall-through projection of this CatchNode 3176 // will not be populated. 3177 Node *arg0 = call->in(TypeFunc::Parms); 3178 if (arg0->is_Type() && 3179 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3180 required_outcnt--; 3181 } 3182 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3183 call->req() > TypeFunc::Parms+1 && 3184 call->is_CallStaticJava()) { 3185 // Check for negative array length. In such case, the optimizer has 3186 // detected that the allocation attempt will always result in an 3187 // exception. There is no fall-through projection of this CatchNode . 3188 Node *arg1 = call->in(TypeFunc::Parms+1); 3189 if (arg1->is_Type() && 3190 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3191 required_outcnt--; 3192 } 3193 } 3194 } 3195 } 3196 // Recheck with a better notion of 'required_outcnt' 3197 if (n->outcnt() != required_outcnt) { 3198 record_method_not_compilable("malformed control flow"); 3199 return true; // Not all targets reachable! 3200 } 3201 } 3202 // Check that I actually visited all kids. Unreached kids 3203 // must be infinite loops. 3204 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3205 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3206 record_method_not_compilable("infinite loop"); 3207 return true; // Found unvisited kid; must be unreach 3208 } 3209 } 3210 3211 // If original bytecodes contained a mixture of floats and doubles 3212 // check if the optimizer has made it homogenous, item (3). 3213 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3214 frc.get_float_count() > 32 && 3215 frc.get_double_count() == 0 && 3216 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3217 set_24_bit_selection_and_mode( false, true ); 3218 } 3219 3220 set_java_calls(frc.get_java_call_count()); 3221 set_inner_loops(frc.get_inner_loop_count()); 3222 3223 // No infinite loops, no reason to bail out. 3224 return false; 3225 } 3226 3227 //-----------------------------too_many_traps---------------------------------- 3228 // Report if there are too many traps at the current method and bci. 3229 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3230 bool Compile::too_many_traps(ciMethod* method, 3231 int bci, 3232 Deoptimization::DeoptReason reason) { 3233 ciMethodData* md = method->method_data(); 3234 if (md->is_empty()) { 3235 // Assume the trap has not occurred, or that it occurred only 3236 // because of a transient condition during start-up in the interpreter. 3237 return false; 3238 } 3239 if (md->has_trap_at(bci, reason) != 0) { 3240 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3241 // Also, if there are multiple reasons, or if there is no per-BCI record, 3242 // assume the worst. 3243 if (log()) 3244 log()->elem("observe trap='%s' count='%d'", 3245 Deoptimization::trap_reason_name(reason), 3246 md->trap_count(reason)); 3247 return true; 3248 } else { 3249 // Ignore method/bci and see if there have been too many globally. 3250 return too_many_traps(reason, md); 3251 } 3252 } 3253 3254 // Less-accurate variant which does not require a method and bci. 3255 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3256 ciMethodData* logmd) { 3257 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 3258 // Too many traps globally. 3259 // Note that we use cumulative trap_count, not just md->trap_count. 3260 if (log()) { 3261 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3262 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3263 Deoptimization::trap_reason_name(reason), 3264 mcount, trap_count(reason)); 3265 } 3266 return true; 3267 } else { 3268 // The coast is clear. 3269 return false; 3270 } 3271 } 3272 3273 //--------------------------too_many_recompiles-------------------------------- 3274 // Report if there are too many recompiles at the current method and bci. 3275 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3276 // Is not eager to return true, since this will cause the compiler to use 3277 // Action_none for a trap point, to avoid too many recompilations. 3278 bool Compile::too_many_recompiles(ciMethod* method, 3279 int bci, 3280 Deoptimization::DeoptReason reason) { 3281 ciMethodData* md = method->method_data(); 3282 if (md->is_empty()) { 3283 // Assume the trap has not occurred, or that it occurred only 3284 // because of a transient condition during start-up in the interpreter. 3285 return false; 3286 } 3287 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3288 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3289 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3290 Deoptimization::DeoptReason per_bc_reason 3291 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3292 if ((per_bc_reason == Deoptimization::Reason_none 3293 || md->has_trap_at(bci, reason) != 0) 3294 // The trap frequency measure we care about is the recompile count: 3295 && md->trap_recompiled_at(bci) 3296 && md->overflow_recompile_count() >= bc_cutoff) { 3297 // Do not emit a trap here if it has already caused recompilations. 3298 // Also, if there are multiple reasons, or if there is no per-BCI record, 3299 // assume the worst. 3300 if (log()) 3301 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3302 Deoptimization::trap_reason_name(reason), 3303 md->trap_count(reason), 3304 md->overflow_recompile_count()); 3305 return true; 3306 } else if (trap_count(reason) != 0 3307 && decompile_count() >= m_cutoff) { 3308 // Too many recompiles globally, and we have seen this sort of trap. 3309 // Use cumulative decompile_count, not just md->decompile_count. 3310 if (log()) 3311 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3312 Deoptimization::trap_reason_name(reason), 3313 md->trap_count(reason), trap_count(reason), 3314 md->decompile_count(), decompile_count()); 3315 return true; 3316 } else { 3317 // The coast is clear. 3318 return false; 3319 } 3320 } 3321 3322 3323 #ifndef PRODUCT 3324 //------------------------------verify_graph_edges--------------------------- 3325 // Walk the Graph and verify that there is a one-to-one correspondence 3326 // between Use-Def edges and Def-Use edges in the graph. 3327 void Compile::verify_graph_edges(bool no_dead_code) { 3328 if (VerifyGraphEdges) { 3329 ResourceArea *area = Thread::current()->resource_area(); 3330 Unique_Node_List visited(area); 3331 // Call recursive graph walk to check edges 3332 _root->verify_edges(visited); 3333 if (no_dead_code) { 3334 // Now make sure that no visited node is used by an unvisited node. 3335 bool dead_nodes = 0; 3336 Unique_Node_List checked(area); 3337 while (visited.size() > 0) { 3338 Node* n = visited.pop(); 3339 checked.push(n); 3340 for (uint i = 0; i < n->outcnt(); i++) { 3341 Node* use = n->raw_out(i); 3342 if (checked.member(use)) continue; // already checked 3343 if (visited.member(use)) continue; // already in the graph 3344 if (use->is_Con()) continue; // a dead ConNode is OK 3345 // At this point, we have found a dead node which is DU-reachable. 3346 if (dead_nodes++ == 0) 3347 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3348 use->dump(2); 3349 tty->print_cr("---"); 3350 checked.push(use); // No repeats; pretend it is now checked. 3351 } 3352 } 3353 assert(dead_nodes == 0, "using nodes must be reachable from root"); 3354 } 3355 } 3356 } 3357 #endif 3358 3359 // The Compile object keeps track of failure reasons separately from the ciEnv. 3360 // This is required because there is not quite a 1-1 relation between the 3361 // ciEnv and its compilation task and the Compile object. Note that one 3362 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3363 // to backtrack and retry without subsuming loads. Other than this backtracking 3364 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3365 // by the logic in C2Compiler. 3366 void Compile::record_failure(const char* reason) { 3367 if (log() != NULL) { 3368 log()->elem("failure reason='%s' phase='compile'", reason); 3369 } 3370 if (_failure_reason == NULL) { 3371 // Record the first failure reason. 3372 _failure_reason = reason; 3373 } 3374 3375 EventCompilerFailure event; 3376 if (event.should_commit()) { 3377 event.set_compileID(Compile::compile_id()); 3378 event.set_failure(reason); 3379 event.commit(); 3380 } 3381 3382 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3383 C->print_method(PHASE_FAILURE); 3384 } 3385 _root = NULL; // flush the graph, too 3386 } 3387 3388 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3389 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3390 _phase_name(name), _dolog(dolog) 3391 { 3392 if (dolog) { 3393 C = Compile::current(); 3394 _log = C->log(); 3395 } else { 3396 C = NULL; 3397 _log = NULL; 3398 } 3399 if (_log != NULL) { 3400 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3401 _log->stamp(); 3402 _log->end_head(); 3403 } 3404 } 3405 3406 Compile::TracePhase::~TracePhase() { 3407 3408 C = Compile::current(); 3409 if (_dolog) { 3410 _log = C->log(); 3411 } else { 3412 _log = NULL; 3413 } 3414 3415 #ifdef ASSERT 3416 if (PrintIdealNodeCount) { 3417 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3418 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3419 } 3420 3421 if (VerifyIdealNodeCount) { 3422 Compile::current()->print_missing_nodes(); 3423 } 3424 #endif 3425 3426 if (_log != NULL) { 3427 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3428 } 3429 } 3430 3431 //============================================================================= 3432 // Two Constant's are equal when the type and the value are equal. 3433 bool Compile::Constant::operator==(const Constant& other) { 3434 if (type() != other.type() ) return false; 3435 if (can_be_reused() != other.can_be_reused()) return false; 3436 // For floating point values we compare the bit pattern. 3437 switch (type()) { 3438 case T_FLOAT: return (_v._value.i == other._v._value.i); 3439 case T_LONG: 3440 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3441 case T_OBJECT: 3442 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3443 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3444 case T_METADATA: return (_v._metadata == other._v._metadata); 3445 default: ShouldNotReachHere(); 3446 } 3447 return false; 3448 } 3449 3450 static int type_to_size_in_bytes(BasicType t) { 3451 switch (t) { 3452 case T_LONG: return sizeof(jlong ); 3453 case T_FLOAT: return sizeof(jfloat ); 3454 case T_DOUBLE: return sizeof(jdouble); 3455 case T_METADATA: return sizeof(Metadata*); 3456 // We use T_VOID as marker for jump-table entries (labels) which 3457 // need an internal word relocation. 3458 case T_VOID: 3459 case T_ADDRESS: 3460 case T_OBJECT: return sizeof(jobject); 3461 } 3462 3463 ShouldNotReachHere(); 3464 return -1; 3465 } 3466 3467 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3468 // sort descending 3469 if (a->freq() > b->freq()) return -1; 3470 if (a->freq() < b->freq()) return 1; 3471 return 0; 3472 } 3473 3474 void Compile::ConstantTable::calculate_offsets_and_size() { 3475 // First, sort the array by frequencies. 3476 _constants.sort(qsort_comparator); 3477 3478 #ifdef ASSERT 3479 // Make sure all jump-table entries were sorted to the end of the 3480 // array (they have a negative frequency). 3481 bool found_void = false; 3482 for (int i = 0; i < _constants.length(); i++) { 3483 Constant con = _constants.at(i); 3484 if (con.type() == T_VOID) 3485 found_void = true; // jump-tables 3486 else 3487 assert(!found_void, "wrong sorting"); 3488 } 3489 #endif 3490 3491 int offset = 0; 3492 for (int i = 0; i < _constants.length(); i++) { 3493 Constant* con = _constants.adr_at(i); 3494 3495 // Align offset for type. 3496 int typesize = type_to_size_in_bytes(con->type()); 3497 offset = align_size_up(offset, typesize); 3498 con->set_offset(offset); // set constant's offset 3499 3500 if (con->type() == T_VOID) { 3501 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3502 offset = offset + typesize * n->outcnt(); // expand jump-table 3503 } else { 3504 offset = offset + typesize; 3505 } 3506 } 3507 3508 // Align size up to the next section start (which is insts; see 3509 // CodeBuffer::align_at_start). 3510 assert(_size == -1, "already set?"); 3511 _size = align_size_up(offset, CodeEntryAlignment); 3512 } 3513 3514 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3515 MacroAssembler _masm(&cb); 3516 for (int i = 0; i < _constants.length(); i++) { 3517 Constant con = _constants.at(i); 3518 address constant_addr; 3519 switch (con.type()) { 3520 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3521 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3522 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3523 case T_OBJECT: { 3524 jobject obj = con.get_jobject(); 3525 int oop_index = _masm.oop_recorder()->find_index(obj); 3526 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3527 break; 3528 } 3529 case T_ADDRESS: { 3530 address addr = (address) con.get_jobject(); 3531 constant_addr = _masm.address_constant(addr); 3532 break; 3533 } 3534 // We use T_VOID as marker for jump-table entries (labels) which 3535 // need an internal word relocation. 3536 case T_VOID: { 3537 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3538 // Fill the jump-table with a dummy word. The real value is 3539 // filled in later in fill_jump_table. 3540 address dummy = (address) n; 3541 constant_addr = _masm.address_constant(dummy); 3542 // Expand jump-table 3543 for (uint i = 1; i < n->outcnt(); i++) { 3544 address temp_addr = _masm.address_constant(dummy + i); 3545 assert(temp_addr, "consts section too small"); 3546 } 3547 break; 3548 } 3549 case T_METADATA: { 3550 Metadata* obj = con.get_metadata(); 3551 int metadata_index = _masm.oop_recorder()->find_index(obj); 3552 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3553 break; 3554 } 3555 default: ShouldNotReachHere(); 3556 } 3557 assert(constant_addr, "consts section too small"); 3558 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3559 } 3560 } 3561 3562 int Compile::ConstantTable::find_offset(Constant& con) const { 3563 int idx = _constants.find(con); 3564 assert(idx != -1, "constant must be in constant table"); 3565 int offset = _constants.at(idx).offset(); 3566 assert(offset != -1, "constant table not emitted yet?"); 3567 return offset; 3568 } 3569 3570 void Compile::ConstantTable::add(Constant& con) { 3571 if (con.can_be_reused()) { 3572 int idx = _constants.find(con); 3573 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3574 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3575 return; 3576 } 3577 } 3578 (void) _constants.append(con); 3579 } 3580 3581 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3582 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3583 Constant con(type, value, b->_freq); 3584 add(con); 3585 return con; 3586 } 3587 3588 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3589 Constant con(metadata); 3590 add(con); 3591 return con; 3592 } 3593 3594 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3595 jvalue value; 3596 BasicType type = oper->type()->basic_type(); 3597 switch (type) { 3598 case T_LONG: value.j = oper->constantL(); break; 3599 case T_FLOAT: value.f = oper->constantF(); break; 3600 case T_DOUBLE: value.d = oper->constantD(); break; 3601 case T_OBJECT: 3602 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3603 case T_METADATA: return add((Metadata*)oper->constant()); break; 3604 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3605 } 3606 return add(n, type, value); 3607 } 3608 3609 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3610 jvalue value; 3611 // We can use the node pointer here to identify the right jump-table 3612 // as this method is called from Compile::Fill_buffer right before 3613 // the MachNodes are emitted and the jump-table is filled (means the 3614 // MachNode pointers do not change anymore). 3615 value.l = (jobject) n; 3616 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3617 add(con); 3618 return con; 3619 } 3620 3621 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3622 // If called from Compile::scratch_emit_size do nothing. 3623 if (Compile::current()->in_scratch_emit_size()) return; 3624 3625 assert(labels.is_nonempty(), "must be"); 3626 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3627 3628 // Since MachConstantNode::constant_offset() also contains 3629 // table_base_offset() we need to subtract the table_base_offset() 3630 // to get the plain offset into the constant table. 3631 int offset = n->constant_offset() - table_base_offset(); 3632 3633 MacroAssembler _masm(&cb); 3634 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3635 3636 for (uint i = 0; i < n->outcnt(); i++) { 3637 address* constant_addr = &jump_table_base[i]; 3638 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))); 3639 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3640 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3641 } 3642 } 3643 3644 void Compile::dump_inlining() { 3645 if (print_inlining() || print_intrinsics()) { 3646 // Print inlining message for candidates that we couldn't inline 3647 // for lack of space or non constant receiver 3648 for (int i = 0; i < _late_inlines.length(); i++) { 3649 CallGenerator* cg = _late_inlines.at(i); 3650 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); 3651 } 3652 Unique_Node_List useful; 3653 useful.push(root()); 3654 for (uint next = 0; next < useful.size(); ++next) { 3655 Node* n = useful.at(next); 3656 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { 3657 CallNode* call = n->as_Call(); 3658 CallGenerator* cg = call->generator(); 3659 cg->print_inlining_late("receiver not constant"); 3660 } 3661 uint max = n->len(); 3662 for ( uint i = 0; i < max; ++i ) { 3663 Node *m = n->in(i); 3664 if ( m == NULL ) continue; 3665 useful.push(m); 3666 } 3667 } 3668 for (int i = 0; i < _print_inlining_list->length(); i++) { 3669 tty->print(_print_inlining_list->adr_at(i)->ss()->as_string()); 3670 } 3671 } 3672 } 3673 3674 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 3675 if (n1->Opcode() < n2->Opcode()) return -1; 3676 else if (n1->Opcode() > n2->Opcode()) return 1; 3677 3678 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())); 3679 for (uint i = 1; i < n1->req(); i++) { 3680 if (n1->in(i) < n2->in(i)) return -1; 3681 else if (n1->in(i) > n2->in(i)) return 1; 3682 } 3683 3684 return 0; 3685 } 3686 3687 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 3688 Node* n1 = *n1p; 3689 Node* n2 = *n2p; 3690 3691 return cmp_expensive_nodes(n1, n2); 3692 } 3693 3694 void Compile::sort_expensive_nodes() { 3695 if (!expensive_nodes_sorted()) { 3696 _expensive_nodes->sort(cmp_expensive_nodes); 3697 } 3698 } 3699 3700 bool Compile::expensive_nodes_sorted() const { 3701 for (int i = 1; i < _expensive_nodes->length(); i++) { 3702 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 3703 return false; 3704 } 3705 } 3706 return true; 3707 } 3708 3709 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 3710 if (_expensive_nodes->length() == 0) { 3711 return false; 3712 } 3713 3714 assert(OptimizeExpensiveOps, "optimization off?"); 3715 3716 // Take this opportunity to remove dead nodes from the list 3717 int j = 0; 3718 for (int i = 0; i < _expensive_nodes->length(); i++) { 3719 Node* n = _expensive_nodes->at(i); 3720 if (!n->is_unreachable(igvn)) { 3721 assert(n->is_expensive(), "should be expensive"); 3722 _expensive_nodes->at_put(j, n); 3723 j++; 3724 } 3725 } 3726 _expensive_nodes->trunc_to(j); 3727 3728 // Then sort the list so that similar nodes are next to each other 3729 // and check for at least two nodes of identical kind with same data 3730 // inputs. 3731 sort_expensive_nodes(); 3732 3733 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 3734 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 3735 return true; 3736 } 3737 } 3738 3739 return false; 3740 } 3741 3742 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 3743 if (_expensive_nodes->length() == 0) { 3744 return; 3745 } 3746 3747 assert(OptimizeExpensiveOps, "optimization off?"); 3748 3749 // Sort to bring similar nodes next to each other and clear the 3750 // control input of nodes for which there's only a single copy. 3751 sort_expensive_nodes(); 3752 3753 int j = 0; 3754 int identical = 0; 3755 int i = 0; 3756 for (; i < _expensive_nodes->length()-1; i++) { 3757 assert(j <= i, "can't write beyond current index"); 3758 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 3759 identical++; 3760 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3761 continue; 3762 } 3763 if (identical > 0) { 3764 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3765 identical = 0; 3766 } else { 3767 Node* n = _expensive_nodes->at(i); 3768 igvn.hash_delete(n); 3769 n->set_req(0, NULL); 3770 igvn.hash_insert(n); 3771 } 3772 } 3773 if (identical > 0) { 3774 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3775 } else if (_expensive_nodes->length() >= 1) { 3776 Node* n = _expensive_nodes->at(i); 3777 igvn.hash_delete(n); 3778 n->set_req(0, NULL); 3779 igvn.hash_insert(n); 3780 } 3781 _expensive_nodes->trunc_to(j); 3782 } 3783 3784 void Compile::add_expensive_node(Node * n) { 3785 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 3786 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 3787 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 3788 if (OptimizeExpensiveOps) { 3789 _expensive_nodes->append(n); 3790 } else { 3791 // Clear control input and let IGVN optimize expensive nodes if 3792 // OptimizeExpensiveOps is off. 3793 n->set_req(0, NULL); 3794 } 3795 } 3796 3797 // Auxiliary method to support randomized stressing/fuzzing. 3798 // 3799 // This method can be called the arbitrary number of times, with current count 3800 // as the argument. The logic allows selecting a single candidate from the 3801 // running list of candidates as follows: 3802 // int count = 0; 3803 // Cand* selected = null; 3804 // while(cand = cand->next()) { 3805 // if (randomized_select(++count)) { 3806 // selected = cand; 3807 // } 3808 // } 3809 // 3810 // Including count equalizes the chances any candidate is "selected". 3811 // This is useful when we don't have the complete list of candidates to choose 3812 // from uniformly. In this case, we need to adjust the randomicity of the 3813 // selection, or else we will end up biasing the selection towards the latter 3814 // candidates. 3815 // 3816 // Quick back-envelope calculation shows that for the list of n candidates 3817 // the equal probability for the candidate to persist as "best" can be 3818 // achieved by replacing it with "next" k-th candidate with the probability 3819 // of 1/k. It can be easily shown that by the end of the run, the 3820 // probability for any candidate is converged to 1/n, thus giving the 3821 // uniform distribution among all the candidates. 3822 // 3823 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 3824 #define RANDOMIZED_DOMAIN_POW 29 3825 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 3826 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 3827 bool Compile::randomized_select(int count) { 3828 assert(count > 0, "only positive"); 3829 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 3830 }