1 /* 2 * Copyright (c) 1997, 2014, 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 "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/block.hpp" 38 #include "opto/c2compiler.hpp" 39 #include "opto/callGenerator.hpp" 40 #include "opto/callnode.hpp" 41 #include "opto/cfgnode.hpp" 42 #include "opto/chaitin.hpp" 43 #include "opto/compile.hpp" 44 #include "opto/connode.hpp" 45 #include "opto/divnode.hpp" 46 #include "opto/escape.hpp" 47 #include "opto/idealGraphPrinter.hpp" 48 #include "opto/loopnode.hpp" 49 #include "opto/machnode.hpp" 50 #include "opto/macro.hpp" 51 #include "opto/matcher.hpp" 52 #include "opto/mathexactnode.hpp" 53 #include "opto/memnode.hpp" 54 #include "opto/mulnode.hpp" 55 #include "opto/narrowptrnode.hpp" 56 #include "opto/node.hpp" 57 #include "opto/opcodes.hpp" 58 #include "opto/output.hpp" 59 #include "opto/parse.hpp" 60 #include "opto/phaseX.hpp" 61 #include "opto/rootnode.hpp" 62 #include "opto/runtime.hpp" 63 #include "opto/stringopts.hpp" 64 #include "opto/type.hpp" 65 #include "opto/vectornode.hpp" 66 #include "runtime/arguments.hpp" 67 #include "runtime/signature.hpp" 68 #include "runtime/stubRoutines.hpp" 69 #include "runtime/timer.hpp" 70 #include "utilities/copy.hpp" 71 72 73 // -------------------- Compile::mach_constant_base_node ----------------------- 74 // Constant table base node singleton. 75 MachConstantBaseNode* Compile::mach_constant_base_node() { 76 if (_mach_constant_base_node == NULL) { 77 _mach_constant_base_node = new MachConstantBaseNode(); 78 _mach_constant_base_node->add_req(C->root()); 79 } 80 return _mach_constant_base_node; 81 } 82 83 84 /// Support for intrinsics. 85 86 // Return the index at which m must be inserted (or already exists). 87 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 88 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 89 #ifdef ASSERT 90 for (int i = 1; i < _intrinsics->length(); i++) { 91 CallGenerator* cg1 = _intrinsics->at(i-1); 92 CallGenerator* cg2 = _intrinsics->at(i); 93 assert(cg1->method() != cg2->method() 94 ? cg1->method() < cg2->method() 95 : cg1->is_virtual() < cg2->is_virtual(), 96 "compiler intrinsics list must stay sorted"); 97 } 98 #endif 99 // Binary search sorted list, in decreasing intervals [lo, hi]. 100 int lo = 0, hi = _intrinsics->length()-1; 101 while (lo <= hi) { 102 int mid = (uint)(hi + lo) / 2; 103 ciMethod* mid_m = _intrinsics->at(mid)->method(); 104 if (m < mid_m) { 105 hi = mid-1; 106 } else if (m > mid_m) { 107 lo = mid+1; 108 } else { 109 // look at minor sort key 110 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 111 if (is_virtual < mid_virt) { 112 hi = mid-1; 113 } else if (is_virtual > mid_virt) { 114 lo = mid+1; 115 } else { 116 return mid; // exact match 117 } 118 } 119 } 120 return lo; // inexact match 121 } 122 123 void Compile::register_intrinsic(CallGenerator* cg) { 124 if (_intrinsics == NULL) { 125 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 126 } 127 // This code is stolen from ciObjectFactory::insert. 128 // Really, GrowableArray should have methods for 129 // insert_at, remove_at, and binary_search. 130 int len = _intrinsics->length(); 131 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 132 if (index == len) { 133 _intrinsics->append(cg); 134 } else { 135 #ifdef ASSERT 136 CallGenerator* oldcg = _intrinsics->at(index); 137 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 138 #endif 139 _intrinsics->append(_intrinsics->at(len-1)); 140 int pos; 141 for (pos = len-2; pos >= index; pos--) { 142 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 143 } 144 _intrinsics->at_put(index, cg); 145 } 146 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 147 } 148 149 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 150 assert(m->is_loaded(), "don't try this on unloaded methods"); 151 if (_intrinsics != NULL) { 152 int index = intrinsic_insertion_index(m, is_virtual); 153 if (index < _intrinsics->length() 154 && _intrinsics->at(index)->method() == m 155 && _intrinsics->at(index)->is_virtual() == is_virtual) { 156 return _intrinsics->at(index); 157 } 158 } 159 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 160 if (m->intrinsic_id() != vmIntrinsics::_none && 161 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 162 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 163 if (cg != NULL) { 164 // Save it for next time: 165 register_intrinsic(cg); 166 return cg; 167 } else { 168 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 169 } 170 } 171 return NULL; 172 } 173 174 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 175 // in library_call.cpp. 176 177 178 #ifndef PRODUCT 179 // statistics gathering... 180 181 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 182 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 183 184 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 185 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 186 int oflags = _intrinsic_hist_flags[id]; 187 assert(flags != 0, "what happened?"); 188 if (is_virtual) { 189 flags |= _intrinsic_virtual; 190 } 191 bool changed = (flags != oflags); 192 if ((flags & _intrinsic_worked) != 0) { 193 juint count = (_intrinsic_hist_count[id] += 1); 194 if (count == 1) { 195 changed = true; // first time 196 } 197 // increment the overall count also: 198 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 199 } 200 if (changed) { 201 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 202 // Something changed about the intrinsic's virtuality. 203 if ((flags & _intrinsic_virtual) != 0) { 204 // This is the first use of this intrinsic as a virtual call. 205 if (oflags != 0) { 206 // We already saw it as a non-virtual, so note both cases. 207 flags |= _intrinsic_both; 208 } 209 } else if ((oflags & _intrinsic_both) == 0) { 210 // This is the first use of this intrinsic as a non-virtual 211 flags |= _intrinsic_both; 212 } 213 } 214 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 215 } 216 // update the overall flags also: 217 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 218 return changed; 219 } 220 221 static char* format_flags(int flags, char* buf) { 222 buf[0] = 0; 223 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 224 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 225 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 226 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 227 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 228 if (buf[0] == 0) strcat(buf, ","); 229 assert(buf[0] == ',', "must be"); 230 return &buf[1]; 231 } 232 233 void Compile::print_intrinsic_statistics() { 234 char flagsbuf[100]; 235 ttyLocker ttyl; 236 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 237 tty->print_cr("Compiler intrinsic usage:"); 238 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 239 if (total == 0) total = 1; // avoid div0 in case of no successes 240 #define PRINT_STAT_LINE(name, c, f) \ 241 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 242 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 243 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 244 int flags = _intrinsic_hist_flags[id]; 245 juint count = _intrinsic_hist_count[id]; 246 if ((flags | count) != 0) { 247 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 248 } 249 } 250 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 251 if (xtty != NULL) xtty->tail("statistics"); 252 } 253 254 void Compile::print_statistics() { 255 { ttyLocker ttyl; 256 if (xtty != NULL) xtty->head("statistics type='opto'"); 257 Parse::print_statistics(); 258 PhaseCCP::print_statistics(); 259 PhaseRegAlloc::print_statistics(); 260 Scheduling::print_statistics(); 261 PhasePeephole::print_statistics(); 262 PhaseIdealLoop::print_statistics(); 263 if (xtty != NULL) xtty->tail("statistics"); 264 } 265 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 266 // put this under its own <statistics> element. 267 print_intrinsic_statistics(); 268 } 269 } 270 #endif //PRODUCT 271 272 // Support for bundling info 273 Bundle* Compile::node_bundling(const Node *n) { 274 assert(valid_bundle_info(n), "oob"); 275 return &_node_bundling_base[n->_idx]; 276 } 277 278 bool Compile::valid_bundle_info(const Node *n) { 279 return (_node_bundling_limit > n->_idx); 280 } 281 282 283 void Compile::gvn_replace_by(Node* n, Node* nn) { 284 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 285 Node* use = n->last_out(i); 286 bool is_in_table = initial_gvn()->hash_delete(use); 287 uint uses_found = 0; 288 for (uint j = 0; j < use->len(); j++) { 289 if (use->in(j) == n) { 290 if (j < use->req()) 291 use->set_req(j, nn); 292 else 293 use->set_prec(j, nn); 294 uses_found++; 295 } 296 } 297 if (is_in_table) { 298 // reinsert into table 299 initial_gvn()->hash_find_insert(use); 300 } 301 record_for_igvn(use); 302 i -= uses_found; // we deleted 1 or more copies of this edge 303 } 304 } 305 306 307 static inline bool not_a_node(const Node* n) { 308 if (n == NULL) return true; 309 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 310 if (*(address*)n == badAddress) return true; // kill by Node::destruct 311 return false; 312 } 313 314 // Identify all nodes that are reachable from below, useful. 315 // Use breadth-first pass that records state in a Unique_Node_List, 316 // recursive traversal is slower. 317 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 318 int estimated_worklist_size = unique(); 319 useful.map( estimated_worklist_size, NULL ); // preallocate space 320 321 // Initialize worklist 322 if (root() != NULL) { useful.push(root()); } 323 // If 'top' is cached, declare it useful to preserve cached node 324 if( cached_top_node() ) { useful.push(cached_top_node()); } 325 326 // Push all useful nodes onto the list, breadthfirst 327 for( uint next = 0; next < useful.size(); ++next ) { 328 assert( next < unique(), "Unique useful nodes < total nodes"); 329 Node *n = useful.at(next); 330 uint max = n->len(); 331 for( uint i = 0; i < max; ++i ) { 332 Node *m = n->in(i); 333 if (not_a_node(m)) continue; 334 useful.push(m); 335 } 336 } 337 } 338 339 // Update dead_node_list with any missing dead nodes using useful 340 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 341 void Compile::update_dead_node_list(Unique_Node_List &useful) { 342 uint max_idx = unique(); 343 VectorSet& useful_node_set = useful.member_set(); 344 345 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 346 // If node with index node_idx is not in useful set, 347 // mark it as dead in dead node list. 348 if (! useful_node_set.test(node_idx) ) { 349 record_dead_node(node_idx); 350 } 351 } 352 } 353 354 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 355 int shift = 0; 356 for (int i = 0; i < inlines->length(); i++) { 357 CallGenerator* cg = inlines->at(i); 358 CallNode* call = cg->call_node(); 359 if (shift > 0) { 360 inlines->at_put(i-shift, cg); 361 } 362 if (!useful.member(call)) { 363 shift++; 364 } 365 } 366 inlines->trunc_to(inlines->length()-shift); 367 } 368 369 // Disconnect all useless nodes by disconnecting those at the boundary. 370 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 371 uint next = 0; 372 while (next < useful.size()) { 373 Node *n = useful.at(next++); 374 if (n->is_SafePoint()) { 375 // We're done with a parsing phase. Replaced nodes are not valid 376 // beyond that point. 377 n->as_SafePoint()->delete_replaced_nodes(); 378 } 379 // Use raw traversal of out edges since this code removes out edges 380 int max = n->outcnt(); 381 for (int j = 0; j < max; ++j) { 382 Node* child = n->raw_out(j); 383 if (! useful.member(child)) { 384 assert(!child->is_top() || child != top(), 385 "If top is cached in Compile object it is in useful list"); 386 // Only need to remove this out-edge to the useless node 387 n->raw_del_out(j); 388 --j; 389 --max; 390 } 391 } 392 if (n->outcnt() == 1 && n->has_special_unique_user()) { 393 record_for_igvn(n->unique_out()); 394 } 395 } 396 // Remove useless macro and predicate opaq nodes 397 for (int i = C->macro_count()-1; i >= 0; i--) { 398 Node* n = C->macro_node(i); 399 if (!useful.member(n)) { 400 remove_macro_node(n); 401 } 402 } 403 // Remove useless expensive node 404 for (int i = C->expensive_count()-1; i >= 0; i--) { 405 Node* n = C->expensive_node(i); 406 if (!useful.member(n)) { 407 remove_expensive_node(n); 408 } 409 } 410 // clean up the late inline lists 411 remove_useless_late_inlines(&_string_late_inlines, useful); 412 remove_useless_late_inlines(&_boxing_late_inlines, useful); 413 remove_useless_late_inlines(&_late_inlines, useful); 414 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 415 } 416 417 //------------------------------frame_size_in_words----------------------------- 418 // frame_slots in units of words 419 int Compile::frame_size_in_words() const { 420 // shift is 0 in LP32 and 1 in LP64 421 const int shift = (LogBytesPerWord - LogBytesPerInt); 422 int words = _frame_slots >> shift; 423 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 424 return words; 425 } 426 427 // To bang the stack of this compiled method we use the stack size 428 // that the interpreter would need in case of a deoptimization. This 429 // removes the need to bang the stack in the deoptimization blob which 430 // in turn simplifies stack overflow handling. 431 int Compile::bang_size_in_bytes() const { 432 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 433 } 434 435 // ============================================================================ 436 //------------------------------CompileWrapper--------------------------------- 437 class CompileWrapper : public StackObj { 438 Compile *const _compile; 439 public: 440 CompileWrapper(Compile* compile); 441 442 ~CompileWrapper(); 443 }; 444 445 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 446 // the Compile* pointer is stored in the current ciEnv: 447 ciEnv* env = compile->env(); 448 assert(env == ciEnv::current(), "must already be a ciEnv active"); 449 assert(env->compiler_data() == NULL, "compile already active?"); 450 env->set_compiler_data(compile); 451 assert(compile == Compile::current(), "sanity"); 452 453 compile->set_type_dict(NULL); 454 compile->set_type_hwm(NULL); 455 compile->set_type_last_size(0); 456 compile->set_last_tf(NULL, NULL); 457 compile->set_indexSet_arena(NULL); 458 compile->set_indexSet_free_block_list(NULL); 459 compile->init_type_arena(); 460 Type::Initialize(compile); 461 _compile->set_scratch_buffer_blob(NULL); 462 _compile->begin_method(); 463 } 464 CompileWrapper::~CompileWrapper() { 465 _compile->end_method(); 466 if (_compile->scratch_buffer_blob() != NULL) 467 BufferBlob::free(_compile->scratch_buffer_blob()); 468 _compile->env()->set_compiler_data(NULL); 469 } 470 471 472 //----------------------------print_compile_messages--------------------------- 473 void Compile::print_compile_messages() { 474 #ifndef PRODUCT 475 // Check if recompiling 476 if (_subsume_loads == false && PrintOpto) { 477 // Recompiling without allowing machine instructions to subsume loads 478 tty->print_cr("*********************************************************"); 479 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 480 tty->print_cr("*********************************************************"); 481 } 482 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 483 // Recompiling without escape analysis 484 tty->print_cr("*********************************************************"); 485 tty->print_cr("** Bailout: Recompile without escape analysis **"); 486 tty->print_cr("*********************************************************"); 487 } 488 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 489 // Recompiling without boxing elimination 490 tty->print_cr("*********************************************************"); 491 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 492 tty->print_cr("*********************************************************"); 493 } 494 if (env()->break_at_compile()) { 495 // Open the debugger when compiling this method. 496 tty->print("### Breaking when compiling: "); 497 method()->print_short_name(); 498 tty->cr(); 499 BREAKPOINT; 500 } 501 502 if( PrintOpto ) { 503 if (is_osr_compilation()) { 504 tty->print("[OSR]%3d", _compile_id); 505 } else { 506 tty->print("%3d", _compile_id); 507 } 508 } 509 #endif 510 } 511 512 513 //-----------------------init_scratch_buffer_blob------------------------------ 514 // Construct a temporary BufferBlob and cache it for this compile. 515 void Compile::init_scratch_buffer_blob(int const_size) { 516 // If there is already a scratch buffer blob allocated and the 517 // constant section is big enough, use it. Otherwise free the 518 // current and allocate a new one. 519 BufferBlob* blob = scratch_buffer_blob(); 520 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 521 // Use the current blob. 522 } else { 523 if (blob != NULL) { 524 BufferBlob::free(blob); 525 } 526 527 ResourceMark rm; 528 _scratch_const_size = const_size; 529 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 530 blob = BufferBlob::create("Compile::scratch_buffer", size); 531 // Record the buffer blob for next time. 532 set_scratch_buffer_blob(blob); 533 // Have we run out of code space? 534 if (scratch_buffer_blob() == NULL) { 535 // Let CompilerBroker disable further compilations. 536 record_failure("Not enough space for scratch buffer in CodeCache"); 537 return; 538 } 539 } 540 541 // Initialize the relocation buffers 542 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 543 set_scratch_locs_memory(locs_buf); 544 } 545 546 547 //-----------------------scratch_emit_size------------------------------------- 548 // Helper function that computes size by emitting code 549 uint Compile::scratch_emit_size(const Node* n) { 550 // Start scratch_emit_size section. 551 set_in_scratch_emit_size(true); 552 553 // Emit into a trash buffer and count bytes emitted. 554 // This is a pretty expensive way to compute a size, 555 // but it works well enough if seldom used. 556 // All common fixed-size instructions are given a size 557 // method by the AD file. 558 // Note that the scratch buffer blob and locs memory are 559 // allocated at the beginning of the compile task, and 560 // may be shared by several calls to scratch_emit_size. 561 // The allocation of the scratch buffer blob is particularly 562 // expensive, since it has to grab the code cache lock. 563 BufferBlob* blob = this->scratch_buffer_blob(); 564 assert(blob != NULL, "Initialize BufferBlob at start"); 565 assert(blob->size() > MAX_inst_size, "sanity"); 566 relocInfo* locs_buf = scratch_locs_memory(); 567 address blob_begin = blob->content_begin(); 568 address blob_end = (address)locs_buf; 569 assert(blob->content_contains(blob_end), "sanity"); 570 CodeBuffer buf(blob_begin, blob_end - blob_begin); 571 buf.initialize_consts_size(_scratch_const_size); 572 buf.initialize_stubs_size(MAX_stubs_size); 573 assert(locs_buf != NULL, "sanity"); 574 int lsize = MAX_locs_size / 3; 575 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 576 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 577 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 578 579 // Do the emission. 580 581 Label fakeL; // Fake label for branch instructions. 582 Label* saveL = NULL; 583 uint save_bnum = 0; 584 bool is_branch = n->is_MachBranch(); 585 if (is_branch) { 586 MacroAssembler masm(&buf); 587 masm.bind(fakeL); 588 n->as_MachBranch()->save_label(&saveL, &save_bnum); 589 n->as_MachBranch()->label_set(&fakeL, 0); 590 } 591 n->emit(buf, this->regalloc()); 592 if (is_branch) // Restore label. 593 n->as_MachBranch()->label_set(saveL, save_bnum); 594 595 // End scratch_emit_size section. 596 set_in_scratch_emit_size(false); 597 598 return buf.insts_size(); 599 } 600 601 602 // ============================================================================ 603 //------------------------------Compile standard------------------------------- 604 debug_only( int Compile::_debug_idx = 100000; ) 605 606 // Compile a method. entry_bci is -1 for normal compilations and indicates 607 // the continuation bci for on stack replacement. 608 609 610 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 611 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 612 : Phase(Compiler), 613 _env(ci_env), 614 _log(ci_env->log()), 615 _compile_id(ci_env->compile_id()), 616 _save_argument_registers(false), 617 _stub_name(NULL), 618 _stub_function(NULL), 619 _stub_entry_point(NULL), 620 _method(target), 621 _entry_bci(osr_bci), 622 _initial_gvn(NULL), 623 _for_igvn(NULL), 624 _warm_calls(NULL), 625 _subsume_loads(subsume_loads), 626 _do_escape_analysis(do_escape_analysis), 627 _eliminate_boxing(eliminate_boxing), 628 _failure_reason(NULL), 629 _code_buffer("Compile::Fill_buffer"), 630 _orig_pc_slot(0), 631 _orig_pc_slot_offset_in_bytes(0), 632 _has_method_handle_invokes(false), 633 _mach_constant_base_node(NULL), 634 _node_bundling_limit(0), 635 _node_bundling_base(NULL), 636 _java_calls(0), 637 _inner_loops(0), 638 _scratch_const_size(-1), 639 _in_scratch_emit_size(false), 640 _dead_node_list(comp_arena()), 641 _dead_node_count(0), 642 #ifndef PRODUCT 643 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 644 _in_dump_cnt(0), 645 _printer(IdealGraphPrinter::printer()), 646 #endif 647 _congraph(NULL), 648 _comp_arena(mtCompiler), 649 _node_arena(mtCompiler), 650 _old_arena(mtCompiler), 651 _Compile_types(mtCompiler), 652 _replay_inline_data(NULL), 653 _late_inlines(comp_arena(), 2, 0, NULL), 654 _string_late_inlines(comp_arena(), 2, 0, NULL), 655 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 656 _late_inlines_pos(0), 657 _number_of_mh_late_inlines(0), 658 _inlining_progress(false), 659 _inlining_incrementally(false), 660 _print_inlining_list(NULL), 661 _print_inlining_stream(NULL), 662 _print_inlining_idx(0), 663 _print_inlining_output(NULL), 664 _interpreter_frame_size(0) { 665 C = this; 666 667 CompileWrapper cw(this); 668 669 if (CITimeVerbose) { 670 tty->print(" "); 671 target->holder()->name()->print(); 672 tty->print("."); 673 target->print_short_name(); 674 tty->print(" "); 675 } 676 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 677 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 678 679 #ifndef PRODUCT 680 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 681 if (!print_opto_assembly) { 682 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 683 if (print_assembly && !Disassembler::can_decode()) { 684 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 685 print_opto_assembly = true; 686 } 687 } 688 set_print_assembly(print_opto_assembly); 689 set_parsed_irreducible_loop(false); 690 691 if (method()->has_option("ReplayInline")) { 692 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 693 } 694 #endif 695 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 696 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 697 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 698 699 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 700 // Make sure the method being compiled gets its own MDO, 701 // so we can at least track the decompile_count(). 702 // Need MDO to record RTM code generation state. 703 method()->ensure_method_data(); 704 } 705 706 Init(::AliasLevel); 707 708 709 print_compile_messages(); 710 711 _ilt = InlineTree::build_inline_tree_root(); 712 713 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 714 assert(num_alias_types() >= AliasIdxRaw, ""); 715 716 #define MINIMUM_NODE_HASH 1023 717 // Node list that Iterative GVN will start with 718 Unique_Node_List for_igvn(comp_arena()); 719 set_for_igvn(&for_igvn); 720 721 // GVN that will be run immediately on new nodes 722 uint estimated_size = method()->code_size()*4+64; 723 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 724 PhaseGVN gvn(node_arena(), estimated_size); 725 set_initial_gvn(&gvn); 726 727 print_inlining_init(); 728 { // Scope for timing the parser 729 TracePhase tp("parse", &timers[_t_parser]); 730 731 // Put top into the hash table ASAP. 732 initial_gvn()->transform_no_reclaim(top()); 733 734 // Set up tf(), start(), and find a CallGenerator. 735 CallGenerator* cg = NULL; 736 if (is_osr_compilation()) { 737 const TypeTuple *domain = StartOSRNode::osr_domain(); 738 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 739 init_tf(TypeFunc::make(domain, range)); 740 StartNode* s = new StartOSRNode(root(), domain); 741 initial_gvn()->set_type_bottom(s); 742 init_start(s); 743 cg = CallGenerator::for_osr(method(), entry_bci()); 744 } else { 745 // Normal case. 746 init_tf(TypeFunc::make(method())); 747 StartNode* s = new StartNode(root(), tf()->domain()); 748 initial_gvn()->set_type_bottom(s); 749 init_start(s); 750 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 751 // With java.lang.ref.reference.get() we must go through the 752 // intrinsic when G1 is enabled - even when get() is the root 753 // method of the compile - so that, if necessary, the value in 754 // the referent field of the reference object gets recorded by 755 // the pre-barrier code. 756 // Specifically, if G1 is enabled, the value in the referent 757 // field is recorded by the G1 SATB pre barrier. This will 758 // result in the referent being marked live and the reference 759 // object removed from the list of discovered references during 760 // reference processing. 761 cg = find_intrinsic(method(), false); 762 } 763 if (cg == NULL) { 764 float past_uses = method()->interpreter_invocation_count(); 765 float expected_uses = past_uses; 766 cg = CallGenerator::for_inline(method(), expected_uses); 767 } 768 } 769 if (failing()) return; 770 if (cg == NULL) { 771 record_method_not_compilable_all_tiers("cannot parse method"); 772 return; 773 } 774 JVMState* jvms = build_start_state(start(), tf()); 775 if ((jvms = cg->generate(jvms)) == NULL) { 776 record_method_not_compilable("method parse failed"); 777 return; 778 } 779 GraphKit kit(jvms); 780 781 if (!kit.stopped()) { 782 // Accept return values, and transfer control we know not where. 783 // This is done by a special, unique ReturnNode bound to root. 784 return_values(kit.jvms()); 785 } 786 787 if (kit.has_exceptions()) { 788 // Any exceptions that escape from this call must be rethrown 789 // to whatever caller is dynamically above us on the stack. 790 // This is done by a special, unique RethrowNode bound to root. 791 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 792 } 793 794 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 795 796 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 797 inline_string_calls(true); 798 } 799 800 if (failing()) return; 801 802 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 803 804 // Remove clutter produced by parsing. 805 if (!failing()) { 806 ResourceMark rm; 807 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 808 } 809 } 810 811 // Note: Large methods are capped off in do_one_bytecode(). 812 if (failing()) return; 813 814 // After parsing, node notes are no longer automagic. 815 // They must be propagated by register_new_node_with_optimizer(), 816 // clone(), or the like. 817 set_default_node_notes(NULL); 818 819 for (;;) { 820 int successes = Inline_Warm(); 821 if (failing()) return; 822 if (successes == 0) break; 823 } 824 825 // Drain the list. 826 Finish_Warm(); 827 #ifndef PRODUCT 828 if (_printer && _printer->should_print(_method)) { 829 _printer->print_inlining(this); 830 } 831 #endif 832 833 if (failing()) return; 834 NOT_PRODUCT( verify_graph_edges(); ) 835 836 // Now optimize 837 Optimize(); 838 if (failing()) return; 839 NOT_PRODUCT( verify_graph_edges(); ) 840 841 #ifndef PRODUCT 842 if (PrintIdeal) { 843 ttyLocker ttyl; // keep the following output all in one block 844 // This output goes directly to the tty, not the compiler log. 845 // To enable tools to match it up with the compilation activity, 846 // be sure to tag this tty output with the compile ID. 847 if (xtty != NULL) { 848 xtty->head("ideal compile_id='%d'%s", compile_id(), 849 is_osr_compilation() ? " compile_kind='osr'" : 850 ""); 851 } 852 root()->dump(9999); 853 if (xtty != NULL) { 854 xtty->tail("ideal"); 855 } 856 } 857 #endif 858 859 NOT_PRODUCT( verify_barriers(); ) 860 861 // Dump compilation data to replay it. 862 if (method()->has_option("DumpReplay")) { 863 env()->dump_replay_data(_compile_id); 864 } 865 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 866 env()->dump_inline_data(_compile_id); 867 } 868 869 // Now that we know the size of all the monitors we can add a fixed slot 870 // for the original deopt pc. 871 872 _orig_pc_slot = fixed_slots(); 873 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 874 set_fixed_slots(next_slot); 875 876 // Compute when to use implicit null checks. Used by matching trap based 877 // nodes and NullCheck optimization. 878 set_allowed_deopt_reasons(); 879 880 // Now generate code 881 Code_Gen(); 882 if (failing()) return; 883 884 // Check if we want to skip execution of all compiled code. 885 { 886 #ifndef PRODUCT 887 if (OptoNoExecute) { 888 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 889 return; 890 } 891 #endif 892 TracePhase tp("install_code", &timers[_t_registerMethod]); 893 894 if (is_osr_compilation()) { 895 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 896 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 897 } else { 898 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 899 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 900 } 901 902 env()->register_method(_method, _entry_bci, 903 &_code_offsets, 904 _orig_pc_slot_offset_in_bytes, 905 code_buffer(), 906 frame_size_in_words(), _oop_map_set, 907 &_handler_table, &_inc_table, 908 compiler, 909 env()->comp_level(), 910 has_unsafe_access(), 911 SharedRuntime::is_wide_vector(max_vector_size()), 912 rtm_state() 913 ); 914 915 if (log() != NULL) // Print code cache state into compiler log 916 log()->code_cache_state(); 917 } 918 } 919 920 //------------------------------Compile---------------------------------------- 921 // Compile a runtime stub 922 Compile::Compile( ciEnv* ci_env, 923 TypeFunc_generator generator, 924 address stub_function, 925 const char *stub_name, 926 int is_fancy_jump, 927 bool pass_tls, 928 bool save_arg_registers, 929 bool return_pc ) 930 : Phase(Compiler), 931 _env(ci_env), 932 _log(ci_env->log()), 933 _compile_id(0), 934 _save_argument_registers(save_arg_registers), 935 _method(NULL), 936 _stub_name(stub_name), 937 _stub_function(stub_function), 938 _stub_entry_point(NULL), 939 _entry_bci(InvocationEntryBci), 940 _initial_gvn(NULL), 941 _for_igvn(NULL), 942 _warm_calls(NULL), 943 _orig_pc_slot(0), 944 _orig_pc_slot_offset_in_bytes(0), 945 _subsume_loads(true), 946 _do_escape_analysis(false), 947 _eliminate_boxing(false), 948 _failure_reason(NULL), 949 _code_buffer("Compile::Fill_buffer"), 950 _has_method_handle_invokes(false), 951 _mach_constant_base_node(NULL), 952 _node_bundling_limit(0), 953 _node_bundling_base(NULL), 954 _java_calls(0), 955 _inner_loops(0), 956 #ifndef PRODUCT 957 _trace_opto_output(TraceOptoOutput), 958 _in_dump_cnt(0), 959 _printer(NULL), 960 #endif 961 _comp_arena(mtCompiler), 962 _node_arena(mtCompiler), 963 _old_arena(mtCompiler), 964 _Compile_types(mtCompiler), 965 _dead_node_list(comp_arena()), 966 _dead_node_count(0), 967 _congraph(NULL), 968 _replay_inline_data(NULL), 969 _number_of_mh_late_inlines(0), 970 _inlining_progress(false), 971 _inlining_incrementally(false), 972 _print_inlining_list(NULL), 973 _print_inlining_stream(NULL), 974 _print_inlining_idx(0), 975 _print_inlining_output(NULL), 976 _allowed_reasons(0), 977 _interpreter_frame_size(0) { 978 C = this; 979 980 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 981 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 982 983 #ifndef PRODUCT 984 set_print_assembly(PrintFrameConverterAssembly); 985 set_parsed_irreducible_loop(false); 986 #endif 987 set_has_irreducible_loop(false); // no loops 988 989 CompileWrapper cw(this); 990 Init(/*AliasLevel=*/ 0); 991 init_tf((*generator)()); 992 993 { 994 // The following is a dummy for the sake of GraphKit::gen_stub 995 Unique_Node_List for_igvn(comp_arena()); 996 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 997 PhaseGVN gvn(Thread::current()->resource_area(),255); 998 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 999 gvn.transform_no_reclaim(top()); 1000 1001 GraphKit kit; 1002 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1003 } 1004 1005 NOT_PRODUCT( verify_graph_edges(); ) 1006 Code_Gen(); 1007 if (failing()) return; 1008 1009 1010 // Entry point will be accessed using compile->stub_entry_point(); 1011 if (code_buffer() == NULL) { 1012 Matcher::soft_match_failure(); 1013 } else { 1014 if (PrintAssembly && (WizardMode || Verbose)) 1015 tty->print_cr("### Stub::%s", stub_name); 1016 1017 if (!failing()) { 1018 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1019 1020 // Make the NMethod 1021 // For now we mark the frame as never safe for profile stackwalking 1022 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1023 code_buffer(), 1024 CodeOffsets::frame_never_safe, 1025 // _code_offsets.value(CodeOffsets::Frame_Complete), 1026 frame_size_in_words(), 1027 _oop_map_set, 1028 save_arg_registers); 1029 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1030 1031 _stub_entry_point = rs->entry_point(); 1032 } 1033 } 1034 } 1035 1036 //------------------------------Init------------------------------------------- 1037 // Prepare for a single compilation 1038 void Compile::Init(int aliaslevel) { 1039 _unique = 0; 1040 _regalloc = NULL; 1041 1042 _tf = NULL; // filled in later 1043 _top = NULL; // cached later 1044 _matcher = NULL; // filled in later 1045 _cfg = NULL; // filled in later 1046 1047 set_24_bit_selection_and_mode(Use24BitFP, false); 1048 1049 _node_note_array = NULL; 1050 _default_node_notes = NULL; 1051 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1052 1053 _immutable_memory = NULL; // filled in at first inquiry 1054 1055 // Globally visible Nodes 1056 // First set TOP to NULL to give safe behavior during creation of RootNode 1057 set_cached_top_node(NULL); 1058 set_root(new RootNode()); 1059 // Now that you have a Root to point to, create the real TOP 1060 set_cached_top_node( new ConNode(Type::TOP) ); 1061 set_recent_alloc(NULL, NULL); 1062 1063 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1064 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1065 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1066 env()->set_dependencies(new Dependencies(env())); 1067 1068 _fixed_slots = 0; 1069 set_has_split_ifs(false); 1070 set_has_loops(has_method() && method()->has_loops()); // first approximation 1071 set_has_stringbuilder(false); 1072 set_has_boxed_value(false); 1073 _trap_can_recompile = false; // no traps emitted yet 1074 _major_progress = true; // start out assuming good things will happen 1075 set_has_unsafe_access(false); 1076 set_max_vector_size(0); 1077 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1078 set_decompile_count(0); 1079 1080 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1081 set_num_loop_opts(LoopOptsCount); 1082 set_do_inlining(Inline); 1083 set_max_inline_size(MaxInlineSize); 1084 set_freq_inline_size(FreqInlineSize); 1085 set_do_scheduling(OptoScheduling); 1086 set_do_count_invocations(false); 1087 set_do_method_data_update(false); 1088 set_age_code(has_method() && method()->profile_aging()); 1089 set_rtm_state(NoRTM); // No RTM lock eliding by default 1090 #if INCLUDE_RTM_OPT 1091 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1092 int rtm_state = method()->method_data()->rtm_state(); 1093 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1094 // Don't generate RTM lock eliding code. 1095 set_rtm_state(NoRTM); 1096 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1097 // Generate RTM lock eliding code without abort ratio calculation code. 1098 set_rtm_state(UseRTM); 1099 } else if (UseRTMDeopt) { 1100 // Generate RTM lock eliding code and include abort ratio calculation 1101 // code if UseRTMDeopt is on. 1102 set_rtm_state(ProfileRTM); 1103 } 1104 } 1105 #endif 1106 if (debug_info()->recording_non_safepoints()) { 1107 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1108 (comp_arena(), 8, 0, NULL)); 1109 set_default_node_notes(Node_Notes::make(this)); 1110 } 1111 1112 // // -- Initialize types before each compile -- 1113 // // Update cached type information 1114 // if( _method && _method->constants() ) 1115 // Type::update_loaded_types(_method, _method->constants()); 1116 1117 // Init alias_type map. 1118 if (!_do_escape_analysis && aliaslevel == 3) 1119 aliaslevel = 2; // No unique types without escape analysis 1120 _AliasLevel = aliaslevel; 1121 const int grow_ats = 16; 1122 _max_alias_types = grow_ats; 1123 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1124 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1125 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1126 { 1127 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1128 } 1129 // Initialize the first few types. 1130 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1131 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1132 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1133 _num_alias_types = AliasIdxRaw+1; 1134 // Zero out the alias type cache. 1135 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1136 // A NULL adr_type hits in the cache right away. Preload the right answer. 1137 probe_alias_cache(NULL)->_index = AliasIdxTop; 1138 1139 _intrinsics = NULL; 1140 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1141 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1142 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1143 register_library_intrinsics(); 1144 } 1145 1146 //---------------------------init_start---------------------------------------- 1147 // Install the StartNode on this compile object. 1148 void Compile::init_start(StartNode* s) { 1149 if (failing()) 1150 return; // already failing 1151 assert(s == start(), ""); 1152 } 1153 1154 /** 1155 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1156 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1157 * the ideal graph. 1158 */ 1159 StartNode* Compile::start() const { 1160 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 1161 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1162 Node* start = root()->fast_out(i); 1163 if (start->is_Start()) { 1164 return start->as_Start(); 1165 } 1166 } 1167 fatal("Did not find Start node!"); 1168 return NULL; 1169 } 1170 1171 //-------------------------------immutable_memory------------------------------------- 1172 // Access immutable memory 1173 Node* Compile::immutable_memory() { 1174 if (_immutable_memory != NULL) { 1175 return _immutable_memory; 1176 } 1177 StartNode* s = start(); 1178 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1179 Node *p = s->fast_out(i); 1180 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1181 _immutable_memory = p; 1182 return _immutable_memory; 1183 } 1184 } 1185 ShouldNotReachHere(); 1186 return NULL; 1187 } 1188 1189 //----------------------set_cached_top_node------------------------------------ 1190 // Install the cached top node, and make sure Node::is_top works correctly. 1191 void Compile::set_cached_top_node(Node* tn) { 1192 if (tn != NULL) verify_top(tn); 1193 Node* old_top = _top; 1194 _top = tn; 1195 // Calling Node::setup_is_top allows the nodes the chance to adjust 1196 // their _out arrays. 1197 if (_top != NULL) _top->setup_is_top(); 1198 if (old_top != NULL) old_top->setup_is_top(); 1199 assert(_top == NULL || top()->is_top(), ""); 1200 } 1201 1202 #ifdef ASSERT 1203 uint Compile::count_live_nodes_by_graph_walk() { 1204 Unique_Node_List useful(comp_arena()); 1205 // Get useful node list by walking the graph. 1206 identify_useful_nodes(useful); 1207 return useful.size(); 1208 } 1209 1210 void Compile::print_missing_nodes() { 1211 1212 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1213 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1214 return; 1215 } 1216 1217 // This is an expensive function. It is executed only when the user 1218 // specifies VerifyIdealNodeCount option or otherwise knows the 1219 // additional work that needs to be done to identify reachable nodes 1220 // by walking the flow graph and find the missing ones using 1221 // _dead_node_list. 1222 1223 Unique_Node_List useful(comp_arena()); 1224 // Get useful node list by walking the graph. 1225 identify_useful_nodes(useful); 1226 1227 uint l_nodes = C->live_nodes(); 1228 uint l_nodes_by_walk = useful.size(); 1229 1230 if (l_nodes != l_nodes_by_walk) { 1231 if (_log != NULL) { 1232 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1233 _log->stamp(); 1234 _log->end_head(); 1235 } 1236 VectorSet& useful_member_set = useful.member_set(); 1237 int last_idx = l_nodes_by_walk; 1238 for (int i = 0; i < last_idx; i++) { 1239 if (useful_member_set.test(i)) { 1240 if (_dead_node_list.test(i)) { 1241 if (_log != NULL) { 1242 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1243 } 1244 if (PrintIdealNodeCount) { 1245 // Print the log message to tty 1246 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1247 useful.at(i)->dump(); 1248 } 1249 } 1250 } 1251 else if (! _dead_node_list.test(i)) { 1252 if (_log != NULL) { 1253 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1254 } 1255 if (PrintIdealNodeCount) { 1256 // Print the log message to tty 1257 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1258 } 1259 } 1260 } 1261 if (_log != NULL) { 1262 _log->tail("mismatched_nodes"); 1263 } 1264 } 1265 } 1266 void Compile::record_modified_node(Node* n) { 1267 if (_modified_nodes != NULL && !_inlining_incrementally && 1268 n->outcnt() != 0 && !n->is_Con()) { 1269 _modified_nodes->push(n); 1270 } 1271 } 1272 1273 void Compile::remove_modified_node(Node* n) { 1274 if (_modified_nodes != NULL) { 1275 _modified_nodes->remove(n); 1276 } 1277 } 1278 #endif 1279 1280 #ifndef PRODUCT 1281 void Compile::verify_top(Node* tn) const { 1282 if (tn != NULL) { 1283 assert(tn->is_Con(), "top node must be a constant"); 1284 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1285 assert(tn->in(0) != NULL, "must have live top node"); 1286 } 1287 } 1288 #endif 1289 1290 1291 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1292 1293 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1294 guarantee(arr != NULL, ""); 1295 int num_blocks = arr->length(); 1296 if (grow_by < num_blocks) grow_by = num_blocks; 1297 int num_notes = grow_by * _node_notes_block_size; 1298 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1299 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1300 while (num_notes > 0) { 1301 arr->append(notes); 1302 notes += _node_notes_block_size; 1303 num_notes -= _node_notes_block_size; 1304 } 1305 assert(num_notes == 0, "exact multiple, please"); 1306 } 1307 1308 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1309 if (source == NULL || dest == NULL) return false; 1310 1311 if (dest->is_Con()) 1312 return false; // Do not push debug info onto constants. 1313 1314 #ifdef ASSERT 1315 // Leave a bread crumb trail pointing to the original node: 1316 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1317 dest->set_debug_orig(source); 1318 } 1319 #endif 1320 1321 if (node_note_array() == NULL) 1322 return false; // Not collecting any notes now. 1323 1324 // This is a copy onto a pre-existing node, which may already have notes. 1325 // If both nodes have notes, do not overwrite any pre-existing notes. 1326 Node_Notes* source_notes = node_notes_at(source->_idx); 1327 if (source_notes == NULL || source_notes->is_clear()) return false; 1328 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1329 if (dest_notes == NULL || dest_notes->is_clear()) { 1330 return set_node_notes_at(dest->_idx, source_notes); 1331 } 1332 1333 Node_Notes merged_notes = (*source_notes); 1334 // The order of operations here ensures that dest notes will win... 1335 merged_notes.update_from(dest_notes); 1336 return set_node_notes_at(dest->_idx, &merged_notes); 1337 } 1338 1339 1340 //--------------------------allow_range_check_smearing------------------------- 1341 // Gating condition for coalescing similar range checks. 1342 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1343 // single covering check that is at least as strong as any of them. 1344 // If the optimization succeeds, the simplified (strengthened) range check 1345 // will always succeed. If it fails, we will deopt, and then give up 1346 // on the optimization. 1347 bool Compile::allow_range_check_smearing() const { 1348 // If this method has already thrown a range-check, 1349 // assume it was because we already tried range smearing 1350 // and it failed. 1351 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1352 return !already_trapped; 1353 } 1354 1355 1356 //------------------------------flatten_alias_type----------------------------- 1357 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1358 int offset = tj->offset(); 1359 TypePtr::PTR ptr = tj->ptr(); 1360 1361 // Known instance (scalarizable allocation) alias only with itself. 1362 bool is_known_inst = tj->isa_oopptr() != NULL && 1363 tj->is_oopptr()->is_known_instance(); 1364 1365 // Process weird unsafe references. 1366 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1367 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1368 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1369 tj = TypeOopPtr::BOTTOM; 1370 ptr = tj->ptr(); 1371 offset = tj->offset(); 1372 } 1373 1374 // Array pointers need some flattening 1375 const TypeAryPtr *ta = tj->isa_aryptr(); 1376 if (ta && ta->is_stable()) { 1377 // Erase stability property for alias analysis. 1378 tj = ta = ta->cast_to_stable(false); 1379 } 1380 if( ta && is_known_inst ) { 1381 if ( offset != Type::OffsetBot && 1382 offset > arrayOopDesc::length_offset_in_bytes() ) { 1383 offset = Type::OffsetBot; // Flatten constant access into array body only 1384 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1385 } 1386 } else if( ta && _AliasLevel >= 2 ) { 1387 // For arrays indexed by constant indices, we flatten the alias 1388 // space to include all of the array body. Only the header, klass 1389 // and array length can be accessed un-aliased. 1390 if( offset != Type::OffsetBot ) { 1391 if( ta->const_oop() ) { // MethodData* or Method* 1392 offset = Type::OffsetBot; // Flatten constant access into array body 1393 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1394 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1395 // range is OK as-is. 1396 tj = ta = TypeAryPtr::RANGE; 1397 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1398 tj = TypeInstPtr::KLASS; // all klass loads look alike 1399 ta = TypeAryPtr::RANGE; // generic ignored junk 1400 ptr = TypePtr::BotPTR; 1401 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1402 tj = TypeInstPtr::MARK; 1403 ta = TypeAryPtr::RANGE; // generic ignored junk 1404 ptr = TypePtr::BotPTR; 1405 } else { // Random constant offset into array body 1406 offset = Type::OffsetBot; // Flatten constant access into array body 1407 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1408 } 1409 } 1410 // Arrays of fixed size alias with arrays of unknown size. 1411 if (ta->size() != TypeInt::POS) { 1412 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1413 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1414 } 1415 // Arrays of known objects become arrays of unknown objects. 1416 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1417 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1418 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1419 } 1420 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1421 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1422 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1423 } 1424 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1425 // cannot be distinguished by bytecode alone. 1426 if (ta->elem() == TypeInt::BOOL) { 1427 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1428 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1429 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1430 } 1431 // During the 2nd round of IterGVN, NotNull castings are removed. 1432 // Make sure the Bottom and NotNull variants alias the same. 1433 // Also, make sure exact and non-exact variants alias the same. 1434 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1435 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1436 } 1437 } 1438 1439 // Oop pointers need some flattening 1440 const TypeInstPtr *to = tj->isa_instptr(); 1441 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1442 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1443 if( ptr == TypePtr::Constant ) { 1444 if (to->klass() != ciEnv::current()->Class_klass() || 1445 offset < k->size_helper() * wordSize) { 1446 // No constant oop pointers (such as Strings); they alias with 1447 // unknown strings. 1448 assert(!is_known_inst, "not scalarizable allocation"); 1449 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1450 } 1451 } else if( is_known_inst ) { 1452 tj = to; // Keep NotNull and klass_is_exact for instance type 1453 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1454 // During the 2nd round of IterGVN, NotNull castings are removed. 1455 // Make sure the Bottom and NotNull variants alias the same. 1456 // Also, make sure exact and non-exact variants alias the same. 1457 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1458 } 1459 if (to->speculative() != NULL) { 1460 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1461 } 1462 // Canonicalize the holder of this field 1463 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1464 // First handle header references such as a LoadKlassNode, even if the 1465 // object's klass is unloaded at compile time (4965979). 1466 if (!is_known_inst) { // Do it only for non-instance types 1467 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1468 } 1469 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1470 // Static fields are in the space above the normal instance 1471 // fields in the java.lang.Class instance. 1472 if (to->klass() != ciEnv::current()->Class_klass()) { 1473 to = NULL; 1474 tj = TypeOopPtr::BOTTOM; 1475 offset = tj->offset(); 1476 } 1477 } else { 1478 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1479 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1480 if( is_known_inst ) { 1481 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1482 } else { 1483 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1484 } 1485 } 1486 } 1487 } 1488 1489 // Klass pointers to object array klasses need some flattening 1490 const TypeKlassPtr *tk = tj->isa_klassptr(); 1491 if( tk ) { 1492 // If we are referencing a field within a Klass, we need 1493 // to assume the worst case of an Object. Both exact and 1494 // inexact types must flatten to the same alias class so 1495 // use NotNull as the PTR. 1496 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1497 1498 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1499 TypeKlassPtr::OBJECT->klass(), 1500 offset); 1501 } 1502 1503 ciKlass* klass = tk->klass(); 1504 if( klass->is_obj_array_klass() ) { 1505 ciKlass* k = TypeAryPtr::OOPS->klass(); 1506 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1507 k = TypeInstPtr::BOTTOM->klass(); 1508 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1509 } 1510 1511 // Check for precise loads from the primary supertype array and force them 1512 // to the supertype cache alias index. Check for generic array loads from 1513 // the primary supertype array and also force them to the supertype cache 1514 // alias index. Since the same load can reach both, we need to merge 1515 // these 2 disparate memories into the same alias class. Since the 1516 // primary supertype array is read-only, there's no chance of confusion 1517 // where we bypass an array load and an array store. 1518 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1519 if (offset == Type::OffsetBot || 1520 (offset >= primary_supers_offset && 1521 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1522 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1523 offset = in_bytes(Klass::secondary_super_cache_offset()); 1524 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1525 } 1526 } 1527 1528 // Flatten all Raw pointers together. 1529 if (tj->base() == Type::RawPtr) 1530 tj = TypeRawPtr::BOTTOM; 1531 1532 if (tj->base() == Type::AnyPtr) 1533 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1534 1535 // Flatten all to bottom for now 1536 switch( _AliasLevel ) { 1537 case 0: 1538 tj = TypePtr::BOTTOM; 1539 break; 1540 case 1: // Flatten to: oop, static, field or array 1541 switch (tj->base()) { 1542 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1543 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1544 case Type::AryPtr: // do not distinguish arrays at all 1545 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1546 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1547 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1548 default: ShouldNotReachHere(); 1549 } 1550 break; 1551 case 2: // No collapsing at level 2; keep all splits 1552 case 3: // No collapsing at level 3; keep all splits 1553 break; 1554 default: 1555 Unimplemented(); 1556 } 1557 1558 offset = tj->offset(); 1559 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1560 1561 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1562 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1563 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1564 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1565 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1566 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1567 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1568 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1569 assert( tj->ptr() != TypePtr::TopPTR && 1570 tj->ptr() != TypePtr::AnyNull && 1571 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1572 // assert( tj->ptr() != TypePtr::Constant || 1573 // tj->base() == Type::RawPtr || 1574 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1575 1576 return tj; 1577 } 1578 1579 void Compile::AliasType::Init(int i, const TypePtr* at) { 1580 _index = i; 1581 _adr_type = at; 1582 _field = NULL; 1583 _element = NULL; 1584 _is_rewritable = true; // default 1585 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1586 if (atoop != NULL && atoop->is_known_instance()) { 1587 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1588 _general_index = Compile::current()->get_alias_index(gt); 1589 } else { 1590 _general_index = 0; 1591 } 1592 } 1593 1594 //---------------------------------print_on------------------------------------ 1595 #ifndef PRODUCT 1596 void Compile::AliasType::print_on(outputStream* st) { 1597 if (index() < 10) 1598 st->print("@ <%d> ", index()); 1599 else st->print("@ <%d>", index()); 1600 st->print(is_rewritable() ? " " : " RO"); 1601 int offset = adr_type()->offset(); 1602 if (offset == Type::OffsetBot) 1603 st->print(" +any"); 1604 else st->print(" +%-3d", offset); 1605 st->print(" in "); 1606 adr_type()->dump_on(st); 1607 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1608 if (field() != NULL && tjp) { 1609 if (tjp->klass() != field()->holder() || 1610 tjp->offset() != field()->offset_in_bytes()) { 1611 st->print(" != "); 1612 field()->print(); 1613 st->print(" ***"); 1614 } 1615 } 1616 } 1617 1618 void print_alias_types() { 1619 Compile* C = Compile::current(); 1620 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1621 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1622 C->alias_type(idx)->print_on(tty); 1623 tty->cr(); 1624 } 1625 } 1626 #endif 1627 1628 1629 //----------------------------probe_alias_cache-------------------------------- 1630 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1631 intptr_t key = (intptr_t) adr_type; 1632 key ^= key >> logAliasCacheSize; 1633 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1634 } 1635 1636 1637 //-----------------------------grow_alias_types-------------------------------- 1638 void Compile::grow_alias_types() { 1639 const int old_ats = _max_alias_types; // how many before? 1640 const int new_ats = old_ats; // how many more? 1641 const int grow_ats = old_ats+new_ats; // how many now? 1642 _max_alias_types = grow_ats; 1643 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1644 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1645 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1646 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1647 } 1648 1649 1650 //--------------------------------find_alias_type------------------------------ 1651 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1652 if (_AliasLevel == 0) 1653 return alias_type(AliasIdxBot); 1654 1655 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1656 if (ace->_adr_type == adr_type) { 1657 return alias_type(ace->_index); 1658 } 1659 1660 // Handle special cases. 1661 if (adr_type == NULL) return alias_type(AliasIdxTop); 1662 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1663 1664 // Do it the slow way. 1665 const TypePtr* flat = flatten_alias_type(adr_type); 1666 1667 #ifdef ASSERT 1668 assert(flat == flatten_alias_type(flat), "idempotent"); 1669 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1670 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1671 const TypeOopPtr* foop = flat->is_oopptr(); 1672 // Scalarizable allocations have exact klass always. 1673 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1674 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1675 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1676 } 1677 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1678 #endif 1679 1680 int idx = AliasIdxTop; 1681 for (int i = 0; i < num_alias_types(); i++) { 1682 if (alias_type(i)->adr_type() == flat) { 1683 idx = i; 1684 break; 1685 } 1686 } 1687 1688 if (idx == AliasIdxTop) { 1689 if (no_create) return NULL; 1690 // Grow the array if necessary. 1691 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1692 // Add a new alias type. 1693 idx = _num_alias_types++; 1694 _alias_types[idx]->Init(idx, flat); 1695 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1696 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1697 if (flat->isa_instptr()) { 1698 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1699 && flat->is_instptr()->klass() == env()->Class_klass()) 1700 alias_type(idx)->set_rewritable(false); 1701 } 1702 if (flat->isa_aryptr()) { 1703 #ifdef ASSERT 1704 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1705 // (T_BYTE has the weakest alignment and size restrictions...) 1706 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1707 #endif 1708 if (flat->offset() == TypePtr::OffsetBot) { 1709 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1710 } 1711 } 1712 if (flat->isa_klassptr()) { 1713 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1714 alias_type(idx)->set_rewritable(false); 1715 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1716 alias_type(idx)->set_rewritable(false); 1717 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1718 alias_type(idx)->set_rewritable(false); 1719 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1720 alias_type(idx)->set_rewritable(false); 1721 } 1722 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1723 // but the base pointer type is not distinctive enough to identify 1724 // references into JavaThread.) 1725 1726 // Check for final fields. 1727 const TypeInstPtr* tinst = flat->isa_instptr(); 1728 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1729 ciField* field; 1730 if (tinst->const_oop() != NULL && 1731 tinst->klass() == ciEnv::current()->Class_klass() && 1732 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1733 // static field 1734 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1735 field = k->get_field_by_offset(tinst->offset(), true); 1736 } else { 1737 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1738 field = k->get_field_by_offset(tinst->offset(), false); 1739 } 1740 assert(field == NULL || 1741 original_field == NULL || 1742 (field->holder() == original_field->holder() && 1743 field->offset() == original_field->offset() && 1744 field->is_static() == original_field->is_static()), "wrong field?"); 1745 // Set field() and is_rewritable() attributes. 1746 if (field != NULL) alias_type(idx)->set_field(field); 1747 } 1748 } 1749 1750 // Fill the cache for next time. 1751 ace->_adr_type = adr_type; 1752 ace->_index = idx; 1753 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1754 1755 // Might as well try to fill the cache for the flattened version, too. 1756 AliasCacheEntry* face = probe_alias_cache(flat); 1757 if (face->_adr_type == NULL) { 1758 face->_adr_type = flat; 1759 face->_index = idx; 1760 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1761 } 1762 1763 return alias_type(idx); 1764 } 1765 1766 1767 Compile::AliasType* Compile::alias_type(ciField* field) { 1768 const TypeOopPtr* t; 1769 if (field->is_static()) 1770 t = TypeInstPtr::make(field->holder()->java_mirror()); 1771 else 1772 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1773 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1774 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1775 return atp; 1776 } 1777 1778 1779 //------------------------------have_alias_type-------------------------------- 1780 bool Compile::have_alias_type(const TypePtr* adr_type) { 1781 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1782 if (ace->_adr_type == adr_type) { 1783 return true; 1784 } 1785 1786 // Handle special cases. 1787 if (adr_type == NULL) return true; 1788 if (adr_type == TypePtr::BOTTOM) return true; 1789 1790 return find_alias_type(adr_type, true, NULL) != NULL; 1791 } 1792 1793 //-----------------------------must_alias-------------------------------------- 1794 // True if all values of the given address type are in the given alias category. 1795 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1796 if (alias_idx == AliasIdxBot) return true; // the universal category 1797 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1798 if (alias_idx == AliasIdxTop) return false; // the empty category 1799 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1800 1801 // the only remaining possible overlap is identity 1802 int adr_idx = get_alias_index(adr_type); 1803 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1804 assert(adr_idx == alias_idx || 1805 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1806 && adr_type != TypeOopPtr::BOTTOM), 1807 "should not be testing for overlap with an unsafe pointer"); 1808 return adr_idx == alias_idx; 1809 } 1810 1811 //------------------------------can_alias-------------------------------------- 1812 // True if any values of the given address type are in the given alias category. 1813 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1814 if (alias_idx == AliasIdxTop) return false; // the empty category 1815 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1816 if (alias_idx == AliasIdxBot) return true; // the universal category 1817 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1818 1819 // the only remaining possible overlap is identity 1820 int adr_idx = get_alias_index(adr_type); 1821 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1822 return adr_idx == alias_idx; 1823 } 1824 1825 1826 1827 //---------------------------pop_warm_call------------------------------------- 1828 WarmCallInfo* Compile::pop_warm_call() { 1829 WarmCallInfo* wci = _warm_calls; 1830 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1831 return wci; 1832 } 1833 1834 //----------------------------Inline_Warm-------------------------------------- 1835 int Compile::Inline_Warm() { 1836 // If there is room, try to inline some more warm call sites. 1837 // %%% Do a graph index compaction pass when we think we're out of space? 1838 if (!InlineWarmCalls) return 0; 1839 1840 int calls_made_hot = 0; 1841 int room_to_grow = NodeCountInliningCutoff - unique(); 1842 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1843 int amount_grown = 0; 1844 WarmCallInfo* call; 1845 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1846 int est_size = (int)call->size(); 1847 if (est_size > (room_to_grow - amount_grown)) { 1848 // This one won't fit anyway. Get rid of it. 1849 call->make_cold(); 1850 continue; 1851 } 1852 call->make_hot(); 1853 calls_made_hot++; 1854 amount_grown += est_size; 1855 amount_to_grow -= est_size; 1856 } 1857 1858 if (calls_made_hot > 0) set_major_progress(); 1859 return calls_made_hot; 1860 } 1861 1862 1863 //----------------------------Finish_Warm-------------------------------------- 1864 void Compile::Finish_Warm() { 1865 if (!InlineWarmCalls) return; 1866 if (failing()) return; 1867 if (warm_calls() == NULL) return; 1868 1869 // Clean up loose ends, if we are out of space for inlining. 1870 WarmCallInfo* call; 1871 while ((call = pop_warm_call()) != NULL) { 1872 call->make_cold(); 1873 } 1874 } 1875 1876 //---------------------cleanup_loop_predicates----------------------- 1877 // Remove the opaque nodes that protect the predicates so that all unused 1878 // checks and uncommon_traps will be eliminated from the ideal graph 1879 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1880 if (predicate_count()==0) return; 1881 for (int i = predicate_count(); i > 0; i--) { 1882 Node * n = predicate_opaque1_node(i-1); 1883 assert(n->Opcode() == Op_Opaque1, "must be"); 1884 igvn.replace_node(n, n->in(1)); 1885 } 1886 assert(predicate_count()==0, "should be clean!"); 1887 } 1888 1889 // StringOpts and late inlining of string methods 1890 void Compile::inline_string_calls(bool parse_time) { 1891 { 1892 // remove useless nodes to make the usage analysis simpler 1893 ResourceMark rm; 1894 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1895 } 1896 1897 { 1898 ResourceMark rm; 1899 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1900 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1901 print_method(PHASE_AFTER_STRINGOPTS, 3); 1902 } 1903 1904 // now inline anything that we skipped the first time around 1905 if (!parse_time) { 1906 _late_inlines_pos = _late_inlines.length(); 1907 } 1908 1909 while (_string_late_inlines.length() > 0) { 1910 CallGenerator* cg = _string_late_inlines.pop(); 1911 cg->do_late_inline(); 1912 if (failing()) return; 1913 } 1914 _string_late_inlines.trunc_to(0); 1915 } 1916 1917 // Late inlining of boxing methods 1918 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1919 if (_boxing_late_inlines.length() > 0) { 1920 assert(has_boxed_value(), "inconsistent"); 1921 1922 PhaseGVN* gvn = initial_gvn(); 1923 set_inlining_incrementally(true); 1924 1925 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1926 for_igvn()->clear(); 1927 gvn->replace_with(&igvn); 1928 1929 _late_inlines_pos = _late_inlines.length(); 1930 1931 while (_boxing_late_inlines.length() > 0) { 1932 CallGenerator* cg = _boxing_late_inlines.pop(); 1933 cg->do_late_inline(); 1934 if (failing()) return; 1935 } 1936 _boxing_late_inlines.trunc_to(0); 1937 1938 { 1939 ResourceMark rm; 1940 PhaseRemoveUseless pru(gvn, for_igvn()); 1941 } 1942 1943 igvn = PhaseIterGVN(gvn); 1944 igvn.optimize(); 1945 1946 set_inlining_progress(false); 1947 set_inlining_incrementally(false); 1948 } 1949 } 1950 1951 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1952 assert(IncrementalInline, "incremental inlining should be on"); 1953 PhaseGVN* gvn = initial_gvn(); 1954 1955 set_inlining_progress(false); 1956 for_igvn()->clear(); 1957 gvn->replace_with(&igvn); 1958 1959 { 1960 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 1961 int i = 0; 1962 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1963 CallGenerator* cg = _late_inlines.at(i); 1964 _late_inlines_pos = i+1; 1965 cg->do_late_inline(); 1966 if (failing()) return; 1967 } 1968 int j = 0; 1969 for (; i < _late_inlines.length(); i++, j++) { 1970 _late_inlines.at_put(j, _late_inlines.at(i)); 1971 } 1972 _late_inlines.trunc_to(j); 1973 } 1974 1975 { 1976 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 1977 ResourceMark rm; 1978 PhaseRemoveUseless pru(gvn, for_igvn()); 1979 } 1980 1981 { 1982 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 1983 igvn = PhaseIterGVN(gvn); 1984 } 1985 } 1986 1987 // Perform incremental inlining until bound on number of live nodes is reached 1988 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1989 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 1990 1991 PhaseGVN* gvn = initial_gvn(); 1992 1993 set_inlining_incrementally(true); 1994 set_inlining_progress(true); 1995 uint low_live_nodes = 0; 1996 1997 while(inlining_progress() && _late_inlines.length() > 0) { 1998 1999 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2000 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2001 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2002 // PhaseIdealLoop is expensive so we only try it once we are 2003 // out of live nodes and we only try it again if the previous 2004 // helped got the number of nodes down significantly 2005 PhaseIdealLoop ideal_loop( igvn, false, true ); 2006 if (failing()) return; 2007 low_live_nodes = live_nodes(); 2008 _major_progress = true; 2009 } 2010 2011 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2012 break; 2013 } 2014 } 2015 2016 inline_incrementally_one(igvn); 2017 2018 if (failing()) return; 2019 2020 { 2021 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2022 igvn.optimize(); 2023 } 2024 2025 if (failing()) return; 2026 } 2027 2028 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2029 2030 if (_string_late_inlines.length() > 0) { 2031 assert(has_stringbuilder(), "inconsistent"); 2032 for_igvn()->clear(); 2033 initial_gvn()->replace_with(&igvn); 2034 2035 inline_string_calls(false); 2036 2037 if (failing()) return; 2038 2039 { 2040 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2041 ResourceMark rm; 2042 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2043 } 2044 2045 { 2046 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2047 igvn = PhaseIterGVN(gvn); 2048 igvn.optimize(); 2049 } 2050 } 2051 2052 set_inlining_incrementally(false); 2053 } 2054 2055 2056 //------------------------------Optimize--------------------------------------- 2057 // Given a graph, optimize it. 2058 void Compile::Optimize() { 2059 TracePhase tp("optimizer", &timers[_t_optimizer]); 2060 2061 #ifndef PRODUCT 2062 if (env()->break_at_compile()) { 2063 BREAKPOINT; 2064 } 2065 2066 #endif 2067 2068 ResourceMark rm; 2069 int loop_opts_cnt; 2070 2071 print_inlining_reinit(); 2072 2073 NOT_PRODUCT( verify_graph_edges(); ) 2074 2075 print_method(PHASE_AFTER_PARSING); 2076 2077 { 2078 // Iterative Global Value Numbering, including ideal transforms 2079 // Initialize IterGVN with types and values from parse-time GVN 2080 PhaseIterGVN igvn(initial_gvn()); 2081 #ifdef ASSERT 2082 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2083 #endif 2084 { 2085 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2086 igvn.optimize(); 2087 } 2088 2089 print_method(PHASE_ITER_GVN1, 2); 2090 2091 if (failing()) return; 2092 2093 inline_incrementally(igvn); 2094 2095 print_method(PHASE_INCREMENTAL_INLINE, 2); 2096 2097 if (failing()) return; 2098 2099 if (eliminate_boxing()) { 2100 // Inline valueOf() methods now. 2101 inline_boxing_calls(igvn); 2102 2103 if (AlwaysIncrementalInline) { 2104 inline_incrementally(igvn); 2105 } 2106 2107 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2108 2109 if (failing()) return; 2110 } 2111 2112 // Remove the speculative part of types and clean up the graph from 2113 // the extra CastPP nodes whose only purpose is to carry them. Do 2114 // that early so that optimizations are not disrupted by the extra 2115 // CastPP nodes. 2116 remove_speculative_types(igvn); 2117 2118 // No more new expensive nodes will be added to the list from here 2119 // so keep only the actual candidates for optimizations. 2120 cleanup_expensive_nodes(igvn); 2121 2122 // Perform escape analysis 2123 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2124 if (has_loops()) { 2125 // Cleanup graph (remove dead nodes). 2126 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2127 PhaseIdealLoop ideal_loop( igvn, false, true ); 2128 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2129 if (failing()) return; 2130 } 2131 ConnectionGraph::do_analysis(this, &igvn); 2132 2133 if (failing()) return; 2134 2135 // Optimize out fields loads from scalar replaceable allocations. 2136 igvn.optimize(); 2137 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2138 2139 if (failing()) return; 2140 2141 if (congraph() != NULL && macro_count() > 0) { 2142 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2143 PhaseMacroExpand mexp(igvn); 2144 mexp.eliminate_macro_nodes(); 2145 igvn.set_delay_transform(false); 2146 2147 igvn.optimize(); 2148 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2149 2150 if (failing()) return; 2151 } 2152 } 2153 2154 // Loop transforms on the ideal graph. Range Check Elimination, 2155 // peeling, unrolling, etc. 2156 2157 // Set loop opts counter 2158 loop_opts_cnt = num_loop_opts(); 2159 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2160 { 2161 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2162 PhaseIdealLoop ideal_loop( igvn, true ); 2163 loop_opts_cnt--; 2164 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2165 if (failing()) return; 2166 } 2167 // Loop opts pass if partial peeling occurred in previous pass 2168 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2169 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2170 PhaseIdealLoop ideal_loop( igvn, false ); 2171 loop_opts_cnt--; 2172 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2173 if (failing()) return; 2174 } 2175 // Loop opts pass for loop-unrolling before CCP 2176 if(major_progress() && (loop_opts_cnt > 0)) { 2177 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2178 PhaseIdealLoop ideal_loop( igvn, false ); 2179 loop_opts_cnt--; 2180 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2181 } 2182 if (!failing()) { 2183 // Verify that last round of loop opts produced a valid graph 2184 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2185 PhaseIdealLoop::verify(igvn); 2186 } 2187 } 2188 if (failing()) return; 2189 2190 // Conditional Constant Propagation; 2191 PhaseCCP ccp( &igvn ); 2192 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2193 { 2194 TracePhase tp("ccp", &timers[_t_ccp]); 2195 ccp.do_transform(); 2196 } 2197 print_method(PHASE_CPP1, 2); 2198 2199 assert( true, "Break here to ccp.dump_old2new_map()"); 2200 2201 // Iterative Global Value Numbering, including ideal transforms 2202 { 2203 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2204 igvn = ccp; 2205 igvn.optimize(); 2206 } 2207 2208 print_method(PHASE_ITER_GVN2, 2); 2209 2210 if (failing()) return; 2211 2212 // Loop transforms on the ideal graph. Range Check Elimination, 2213 // peeling, unrolling, etc. 2214 if(loop_opts_cnt > 0) { 2215 debug_only( int cnt = 0; ); 2216 while(major_progress() && (loop_opts_cnt > 0)) { 2217 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2218 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2219 PhaseIdealLoop ideal_loop( igvn, true); 2220 loop_opts_cnt--; 2221 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2222 if (failing()) return; 2223 } 2224 } 2225 2226 { 2227 // Verify that all previous optimizations produced a valid graph 2228 // at least to this point, even if no loop optimizations were done. 2229 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2230 PhaseIdealLoop::verify(igvn); 2231 } 2232 2233 { 2234 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2235 PhaseMacroExpand mex(igvn); 2236 if (mex.expand_macro_nodes()) { 2237 assert(failing(), "must bail out w/ explicit message"); 2238 return; 2239 } 2240 } 2241 2242 DEBUG_ONLY( _modified_nodes = NULL; ) 2243 } // (End scope of igvn; run destructor if necessary for asserts.) 2244 2245 process_print_inlining(); 2246 // A method with only infinite loops has no edges entering loops from root 2247 { 2248 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2249 if (final_graph_reshaping()) { 2250 assert(failing(), "must bail out w/ explicit message"); 2251 return; 2252 } 2253 } 2254 2255 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2256 } 2257 2258 2259 //------------------------------Code_Gen--------------------------------------- 2260 // Given a graph, generate code for it 2261 void Compile::Code_Gen() { 2262 if (failing()) { 2263 return; 2264 } 2265 2266 // Perform instruction selection. You might think we could reclaim Matcher 2267 // memory PDQ, but actually the Matcher is used in generating spill code. 2268 // Internals of the Matcher (including some VectorSets) must remain live 2269 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2270 // set a bit in reclaimed memory. 2271 2272 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2273 // nodes. Mapping is only valid at the root of each matched subtree. 2274 NOT_PRODUCT( verify_graph_edges(); ) 2275 2276 Matcher matcher; 2277 _matcher = &matcher; 2278 { 2279 TracePhase tp("matcher", &timers[_t_matcher]); 2280 matcher.match(); 2281 } 2282 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2283 // nodes. Mapping is only valid at the root of each matched subtree. 2284 NOT_PRODUCT( verify_graph_edges(); ) 2285 2286 // If you have too many nodes, or if matching has failed, bail out 2287 check_node_count(0, "out of nodes matching instructions"); 2288 if (failing()) { 2289 return; 2290 } 2291 2292 // Build a proper-looking CFG 2293 PhaseCFG cfg(node_arena(), root(), matcher); 2294 _cfg = &cfg; 2295 { 2296 TracePhase tp("scheduler", &timers[_t_scheduler]); 2297 bool success = cfg.do_global_code_motion(); 2298 if (!success) { 2299 return; 2300 } 2301 2302 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2303 NOT_PRODUCT( verify_graph_edges(); ) 2304 debug_only( cfg.verify(); ) 2305 } 2306 2307 PhaseChaitin regalloc(unique(), cfg, matcher); 2308 _regalloc = ®alloc; 2309 { 2310 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2311 // Perform register allocation. After Chaitin, use-def chains are 2312 // no longer accurate (at spill code) and so must be ignored. 2313 // Node->LRG->reg mappings are still accurate. 2314 _regalloc->Register_Allocate(); 2315 2316 // Bail out if the allocator builds too many nodes 2317 if (failing()) { 2318 return; 2319 } 2320 } 2321 2322 // Prior to register allocation we kept empty basic blocks in case the 2323 // the allocator needed a place to spill. After register allocation we 2324 // are not adding any new instructions. If any basic block is empty, we 2325 // can now safely remove it. 2326 { 2327 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2328 cfg.remove_empty_blocks(); 2329 if (do_freq_based_layout()) { 2330 PhaseBlockLayout layout(cfg); 2331 } else { 2332 cfg.set_loop_alignment(); 2333 } 2334 cfg.fixup_flow(); 2335 } 2336 2337 // Apply peephole optimizations 2338 if( OptoPeephole ) { 2339 TracePhase tp("peephole", &timers[_t_peephole]); 2340 PhasePeephole peep( _regalloc, cfg); 2341 peep.do_transform(); 2342 } 2343 2344 // Do late expand if CPU requires this. 2345 if (Matcher::require_postalloc_expand) { 2346 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2347 cfg.postalloc_expand(_regalloc); 2348 } 2349 2350 // Convert Nodes to instruction bits in a buffer 2351 { 2352 TraceTime tp("output", &timers[_t_output], CITime); 2353 Output(); 2354 } 2355 2356 print_method(PHASE_FINAL_CODE); 2357 2358 // He's dead, Jim. 2359 _cfg = (PhaseCFG*)0xdeadbeef; 2360 _regalloc = (PhaseChaitin*)0xdeadbeef; 2361 } 2362 2363 2364 //------------------------------dump_asm--------------------------------------- 2365 // Dump formatted assembly 2366 #ifndef PRODUCT 2367 void Compile::dump_asm(int *pcs, uint pc_limit) { 2368 bool cut_short = false; 2369 tty->print_cr("#"); 2370 tty->print("# "); _tf->dump(); tty->cr(); 2371 tty->print_cr("#"); 2372 2373 // For all blocks 2374 int pc = 0x0; // Program counter 2375 char starts_bundle = ' '; 2376 _regalloc->dump_frame(); 2377 2378 Node *n = NULL; 2379 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2380 if (VMThread::should_terminate()) { 2381 cut_short = true; 2382 break; 2383 } 2384 Block* block = _cfg->get_block(i); 2385 if (block->is_connector() && !Verbose) { 2386 continue; 2387 } 2388 n = block->head(); 2389 if (pcs && n->_idx < pc_limit) { 2390 tty->print("%3.3x ", pcs[n->_idx]); 2391 } else { 2392 tty->print(" "); 2393 } 2394 block->dump_head(_cfg); 2395 if (block->is_connector()) { 2396 tty->print_cr(" # Empty connector block"); 2397 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2398 tty->print_cr(" # Block is sole successor of call"); 2399 } 2400 2401 // For all instructions 2402 Node *delay = NULL; 2403 for (uint j = 0; j < block->number_of_nodes(); j++) { 2404 if (VMThread::should_terminate()) { 2405 cut_short = true; 2406 break; 2407 } 2408 n = block->get_node(j); 2409 if (valid_bundle_info(n)) { 2410 Bundle* bundle = node_bundling(n); 2411 if (bundle->used_in_unconditional_delay()) { 2412 delay = n; 2413 continue; 2414 } 2415 if (bundle->starts_bundle()) { 2416 starts_bundle = '+'; 2417 } 2418 } 2419 2420 if (WizardMode) { 2421 n->dump(); 2422 } 2423 2424 if( !n->is_Region() && // Dont print in the Assembly 2425 !n->is_Phi() && // a few noisely useless nodes 2426 !n->is_Proj() && 2427 !n->is_MachTemp() && 2428 !n->is_SafePointScalarObject() && 2429 !n->is_Catch() && // Would be nice to print exception table targets 2430 !n->is_MergeMem() && // Not very interesting 2431 !n->is_top() && // Debug info table constants 2432 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2433 ) { 2434 if (pcs && n->_idx < pc_limit) 2435 tty->print("%3.3x", pcs[n->_idx]); 2436 else 2437 tty->print(" "); 2438 tty->print(" %c ", starts_bundle); 2439 starts_bundle = ' '; 2440 tty->print("\t"); 2441 n->format(_regalloc, tty); 2442 tty->cr(); 2443 } 2444 2445 // If we have an instruction with a delay slot, and have seen a delay, 2446 // then back up and print it 2447 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2448 assert(delay != NULL, "no unconditional delay instruction"); 2449 if (WizardMode) delay->dump(); 2450 2451 if (node_bundling(delay)->starts_bundle()) 2452 starts_bundle = '+'; 2453 if (pcs && n->_idx < pc_limit) 2454 tty->print("%3.3x", pcs[n->_idx]); 2455 else 2456 tty->print(" "); 2457 tty->print(" %c ", starts_bundle); 2458 starts_bundle = ' '; 2459 tty->print("\t"); 2460 delay->format(_regalloc, tty); 2461 tty->cr(); 2462 delay = NULL; 2463 } 2464 2465 // Dump the exception table as well 2466 if( n->is_Catch() && (Verbose || WizardMode) ) { 2467 // Print the exception table for this offset 2468 _handler_table.print_subtable_for(pc); 2469 } 2470 } 2471 2472 if (pcs && n->_idx < pc_limit) 2473 tty->print_cr("%3.3x", pcs[n->_idx]); 2474 else 2475 tty->cr(); 2476 2477 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2478 2479 } // End of per-block dump 2480 tty->cr(); 2481 2482 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2483 } 2484 #endif 2485 2486 //------------------------------Final_Reshape_Counts--------------------------- 2487 // This class defines counters to help identify when a method 2488 // may/must be executed using hardware with only 24-bit precision. 2489 struct Final_Reshape_Counts : public StackObj { 2490 int _call_count; // count non-inlined 'common' calls 2491 int _float_count; // count float ops requiring 24-bit precision 2492 int _double_count; // count double ops requiring more precision 2493 int _java_call_count; // count non-inlined 'java' calls 2494 int _inner_loop_count; // count loops which need alignment 2495 VectorSet _visited; // Visitation flags 2496 Node_List _tests; // Set of IfNodes & PCTableNodes 2497 2498 Final_Reshape_Counts() : 2499 _call_count(0), _float_count(0), _double_count(0), 2500 _java_call_count(0), _inner_loop_count(0), 2501 _visited( Thread::current()->resource_area() ) { } 2502 2503 void inc_call_count () { _call_count ++; } 2504 void inc_float_count () { _float_count ++; } 2505 void inc_double_count() { _double_count++; } 2506 void inc_java_call_count() { _java_call_count++; } 2507 void inc_inner_loop_count() { _inner_loop_count++; } 2508 2509 int get_call_count () const { return _call_count ; } 2510 int get_float_count () const { return _float_count ; } 2511 int get_double_count() const { return _double_count; } 2512 int get_java_call_count() const { return _java_call_count; } 2513 int get_inner_loop_count() const { return _inner_loop_count; } 2514 }; 2515 2516 #ifdef ASSERT 2517 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2518 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2519 // Make sure the offset goes inside the instance layout. 2520 return k->contains_field_offset(tp->offset()); 2521 // Note that OffsetBot and OffsetTop are very negative. 2522 } 2523 #endif 2524 2525 // Eliminate trivially redundant StoreCMs and accumulate their 2526 // precedence edges. 2527 void Compile::eliminate_redundant_card_marks(Node* n) { 2528 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2529 if (n->in(MemNode::Address)->outcnt() > 1) { 2530 // There are multiple users of the same address so it might be 2531 // possible to eliminate some of the StoreCMs 2532 Node* mem = n->in(MemNode::Memory); 2533 Node* adr = n->in(MemNode::Address); 2534 Node* val = n->in(MemNode::ValueIn); 2535 Node* prev = n; 2536 bool done = false; 2537 // Walk the chain of StoreCMs eliminating ones that match. As 2538 // long as it's a chain of single users then the optimization is 2539 // safe. Eliminating partially redundant StoreCMs would require 2540 // cloning copies down the other paths. 2541 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2542 if (adr == mem->in(MemNode::Address) && 2543 val == mem->in(MemNode::ValueIn)) { 2544 // redundant StoreCM 2545 if (mem->req() > MemNode::OopStore) { 2546 // Hasn't been processed by this code yet. 2547 n->add_prec(mem->in(MemNode::OopStore)); 2548 } else { 2549 // Already converted to precedence edge 2550 for (uint i = mem->req(); i < mem->len(); i++) { 2551 // Accumulate any precedence edges 2552 if (mem->in(i) != NULL) { 2553 n->add_prec(mem->in(i)); 2554 } 2555 } 2556 // Everything above this point has been processed. 2557 done = true; 2558 } 2559 // Eliminate the previous StoreCM 2560 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2561 assert(mem->outcnt() == 0, "should be dead"); 2562 mem->disconnect_inputs(NULL, this); 2563 } else { 2564 prev = mem; 2565 } 2566 mem = prev->in(MemNode::Memory); 2567 } 2568 } 2569 } 2570 2571 //------------------------------final_graph_reshaping_impl---------------------- 2572 // Implement items 1-5 from final_graph_reshaping below. 2573 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2574 2575 if ( n->outcnt() == 0 ) return; // dead node 2576 uint nop = n->Opcode(); 2577 2578 // Check for 2-input instruction with "last use" on right input. 2579 // Swap to left input. Implements item (2). 2580 if( n->req() == 3 && // two-input instruction 2581 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2582 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2583 n->in(2)->outcnt() == 1 &&// right use IS a last use 2584 !n->in(2)->is_Con() ) { // right use is not a constant 2585 // Check for commutative opcode 2586 switch( nop ) { 2587 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2588 case Op_MaxI: case Op_MinI: 2589 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2590 case Op_AndL: case Op_XorL: case Op_OrL: 2591 case Op_AndI: case Op_XorI: case Op_OrI: { 2592 // Move "last use" input to left by swapping inputs 2593 n->swap_edges(1, 2); 2594 break; 2595 } 2596 default: 2597 break; 2598 } 2599 } 2600 2601 #ifdef ASSERT 2602 if( n->is_Mem() ) { 2603 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2604 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2605 // oop will be recorded in oop map if load crosses safepoint 2606 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2607 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2608 "raw memory operations should have control edge"); 2609 } 2610 #endif 2611 // Count FPU ops and common calls, implements item (3) 2612 switch( nop ) { 2613 // Count all float operations that may use FPU 2614 case Op_AddF: 2615 case Op_SubF: 2616 case Op_MulF: 2617 case Op_DivF: 2618 case Op_NegF: 2619 case Op_ModF: 2620 case Op_ConvI2F: 2621 case Op_ConF: 2622 case Op_CmpF: 2623 case Op_CmpF3: 2624 // case Op_ConvL2F: // longs are split into 32-bit halves 2625 frc.inc_float_count(); 2626 break; 2627 2628 case Op_ConvF2D: 2629 case Op_ConvD2F: 2630 frc.inc_float_count(); 2631 frc.inc_double_count(); 2632 break; 2633 2634 // Count all double operations that may use FPU 2635 case Op_AddD: 2636 case Op_SubD: 2637 case Op_MulD: 2638 case Op_DivD: 2639 case Op_NegD: 2640 case Op_ModD: 2641 case Op_ConvI2D: 2642 case Op_ConvD2I: 2643 // case Op_ConvL2D: // handled by leaf call 2644 // case Op_ConvD2L: // handled by leaf call 2645 case Op_ConD: 2646 case Op_CmpD: 2647 case Op_CmpD3: 2648 frc.inc_double_count(); 2649 break; 2650 case Op_Opaque1: // Remove Opaque Nodes before matching 2651 case Op_Opaque2: // Remove Opaque Nodes before matching 2652 case Op_Opaque3: 2653 n->subsume_by(n->in(1), this); 2654 break; 2655 case Op_CallStaticJava: 2656 case Op_CallJava: 2657 case Op_CallDynamicJava: 2658 frc.inc_java_call_count(); // Count java call site; 2659 case Op_CallRuntime: 2660 case Op_CallLeaf: 2661 case Op_CallLeafNoFP: { 2662 assert( n->is_Call(), "" ); 2663 CallNode *call = n->as_Call(); 2664 // Count call sites where the FP mode bit would have to be flipped. 2665 // Do not count uncommon runtime calls: 2666 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2667 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2668 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2669 frc.inc_call_count(); // Count the call site 2670 } else { // See if uncommon argument is shared 2671 Node *n = call->in(TypeFunc::Parms); 2672 int nop = n->Opcode(); 2673 // Clone shared simple arguments to uncommon calls, item (1). 2674 if( n->outcnt() > 1 && 2675 !n->is_Proj() && 2676 nop != Op_CreateEx && 2677 nop != Op_CheckCastPP && 2678 nop != Op_DecodeN && 2679 nop != Op_DecodeNKlass && 2680 !n->is_Mem() ) { 2681 Node *x = n->clone(); 2682 call->set_req( TypeFunc::Parms, x ); 2683 } 2684 } 2685 break; 2686 } 2687 2688 case Op_StoreD: 2689 case Op_LoadD: 2690 case Op_LoadD_unaligned: 2691 frc.inc_double_count(); 2692 goto handle_mem; 2693 case Op_StoreF: 2694 case Op_LoadF: 2695 frc.inc_float_count(); 2696 goto handle_mem; 2697 2698 case Op_StoreCM: 2699 { 2700 // Convert OopStore dependence into precedence edge 2701 Node* prec = n->in(MemNode::OopStore); 2702 n->del_req(MemNode::OopStore); 2703 n->add_prec(prec); 2704 eliminate_redundant_card_marks(n); 2705 } 2706 2707 // fall through 2708 2709 case Op_StoreB: 2710 case Op_StoreC: 2711 case Op_StorePConditional: 2712 case Op_StoreI: 2713 case Op_StoreL: 2714 case Op_StoreIConditional: 2715 case Op_StoreLConditional: 2716 case Op_CompareAndSwapI: 2717 case Op_CompareAndSwapL: 2718 case Op_CompareAndSwapP: 2719 case Op_CompareAndSwapN: 2720 case Op_GetAndAddI: 2721 case Op_GetAndAddL: 2722 case Op_GetAndSetI: 2723 case Op_GetAndSetL: 2724 case Op_GetAndSetP: 2725 case Op_GetAndSetN: 2726 case Op_StoreP: 2727 case Op_StoreN: 2728 case Op_StoreNKlass: 2729 case Op_LoadB: 2730 case Op_LoadUB: 2731 case Op_LoadUS: 2732 case Op_LoadI: 2733 case Op_LoadKlass: 2734 case Op_LoadNKlass: 2735 case Op_LoadL: 2736 case Op_LoadL_unaligned: 2737 case Op_LoadPLocked: 2738 case Op_LoadP: 2739 case Op_LoadN: 2740 case Op_LoadRange: 2741 case Op_LoadS: { 2742 handle_mem: 2743 #ifdef ASSERT 2744 if( VerifyOptoOopOffsets ) { 2745 assert( n->is_Mem(), "" ); 2746 MemNode *mem = (MemNode*)n; 2747 // Check to see if address types have grounded out somehow. 2748 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2749 assert( !tp || oop_offset_is_sane(tp), "" ); 2750 } 2751 #endif 2752 break; 2753 } 2754 2755 case Op_AddP: { // Assert sane base pointers 2756 Node *addp = n->in(AddPNode::Address); 2757 assert( !addp->is_AddP() || 2758 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2759 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2760 "Base pointers must match" ); 2761 #ifdef _LP64 2762 if ((UseCompressedOops || UseCompressedClassPointers) && 2763 addp->Opcode() == Op_ConP && 2764 addp == n->in(AddPNode::Base) && 2765 n->in(AddPNode::Offset)->is_Con()) { 2766 // Use addressing with narrow klass to load with offset on x86. 2767 // On sparc loading 32-bits constant and decoding it have less 2768 // instructions (4) then load 64-bits constant (7). 2769 // Do this transformation here since IGVN will convert ConN back to ConP. 2770 const Type* t = addp->bottom_type(); 2771 if (t->isa_oopptr() || t->isa_klassptr()) { 2772 Node* nn = NULL; 2773 2774 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2775 2776 // Look for existing ConN node of the same exact type. 2777 Node* r = root(); 2778 uint cnt = r->outcnt(); 2779 for (uint i = 0; i < cnt; i++) { 2780 Node* m = r->raw_out(i); 2781 if (m!= NULL && m->Opcode() == op && 2782 m->bottom_type()->make_ptr() == t) { 2783 nn = m; 2784 break; 2785 } 2786 } 2787 if (nn != NULL) { 2788 // Decode a narrow oop to match address 2789 // [R12 + narrow_oop_reg<<3 + offset] 2790 if (t->isa_oopptr()) { 2791 nn = new DecodeNNode(nn, t); 2792 } else { 2793 nn = new DecodeNKlassNode(nn, t); 2794 } 2795 n->set_req(AddPNode::Base, nn); 2796 n->set_req(AddPNode::Address, nn); 2797 if (addp->outcnt() == 0) { 2798 addp->disconnect_inputs(NULL, this); 2799 } 2800 } 2801 } 2802 } 2803 #endif 2804 break; 2805 } 2806 2807 #ifdef _LP64 2808 case Op_CastPP: 2809 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2810 Node* in1 = n->in(1); 2811 const Type* t = n->bottom_type(); 2812 Node* new_in1 = in1->clone(); 2813 new_in1->as_DecodeN()->set_type(t); 2814 2815 if (!Matcher::narrow_oop_use_complex_address()) { 2816 // 2817 // x86, ARM and friends can handle 2 adds in addressing mode 2818 // and Matcher can fold a DecodeN node into address by using 2819 // a narrow oop directly and do implicit NULL check in address: 2820 // 2821 // [R12 + narrow_oop_reg<<3 + offset] 2822 // NullCheck narrow_oop_reg 2823 // 2824 // On other platforms (Sparc) we have to keep new DecodeN node and 2825 // use it to do implicit NULL check in address: 2826 // 2827 // decode_not_null narrow_oop_reg, base_reg 2828 // [base_reg + offset] 2829 // NullCheck base_reg 2830 // 2831 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2832 // to keep the information to which NULL check the new DecodeN node 2833 // corresponds to use it as value in implicit_null_check(). 2834 // 2835 new_in1->set_req(0, n->in(0)); 2836 } 2837 2838 n->subsume_by(new_in1, this); 2839 if (in1->outcnt() == 0) { 2840 in1->disconnect_inputs(NULL, this); 2841 } 2842 } 2843 break; 2844 2845 case Op_CmpP: 2846 // Do this transformation here to preserve CmpPNode::sub() and 2847 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2848 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2849 Node* in1 = n->in(1); 2850 Node* in2 = n->in(2); 2851 if (!in1->is_DecodeNarrowPtr()) { 2852 in2 = in1; 2853 in1 = n->in(2); 2854 } 2855 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2856 2857 Node* new_in2 = NULL; 2858 if (in2->is_DecodeNarrowPtr()) { 2859 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2860 new_in2 = in2->in(1); 2861 } else if (in2->Opcode() == Op_ConP) { 2862 const Type* t = in2->bottom_type(); 2863 if (t == TypePtr::NULL_PTR) { 2864 assert(in1->is_DecodeN(), "compare klass to null?"); 2865 // Don't convert CmpP null check into CmpN if compressed 2866 // oops implicit null check is not generated. 2867 // This will allow to generate normal oop implicit null check. 2868 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2869 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 2870 // 2871 // This transformation together with CastPP transformation above 2872 // will generated code for implicit NULL checks for compressed oops. 2873 // 2874 // The original code after Optimize() 2875 // 2876 // LoadN memory, narrow_oop_reg 2877 // decode narrow_oop_reg, base_reg 2878 // CmpP base_reg, NULL 2879 // CastPP base_reg // NotNull 2880 // Load [base_reg + offset], val_reg 2881 // 2882 // after these transformations will be 2883 // 2884 // LoadN memory, narrow_oop_reg 2885 // CmpN narrow_oop_reg, NULL 2886 // decode_not_null narrow_oop_reg, base_reg 2887 // Load [base_reg + offset], val_reg 2888 // 2889 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2890 // since narrow oops can be used in debug info now (see the code in 2891 // final_graph_reshaping_walk()). 2892 // 2893 // At the end the code will be matched to 2894 // on x86: 2895 // 2896 // Load_narrow_oop memory, narrow_oop_reg 2897 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2898 // NullCheck narrow_oop_reg 2899 // 2900 // and on sparc: 2901 // 2902 // Load_narrow_oop memory, narrow_oop_reg 2903 // decode_not_null narrow_oop_reg, base_reg 2904 // Load [base_reg + offset], val_reg 2905 // NullCheck base_reg 2906 // 2907 } else if (t->isa_oopptr()) { 2908 new_in2 = ConNode::make(t->make_narrowoop()); 2909 } else if (t->isa_klassptr()) { 2910 new_in2 = ConNode::make(t->make_narrowklass()); 2911 } 2912 } 2913 if (new_in2 != NULL) { 2914 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 2915 n->subsume_by(cmpN, this); 2916 if (in1->outcnt() == 0) { 2917 in1->disconnect_inputs(NULL, this); 2918 } 2919 if (in2->outcnt() == 0) { 2920 in2->disconnect_inputs(NULL, this); 2921 } 2922 } 2923 } 2924 break; 2925 2926 case Op_DecodeN: 2927 case Op_DecodeNKlass: 2928 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2929 // DecodeN could be pinned when it can't be fold into 2930 // an address expression, see the code for Op_CastPP above. 2931 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2932 break; 2933 2934 case Op_EncodeP: 2935 case Op_EncodePKlass: { 2936 Node* in1 = n->in(1); 2937 if (in1->is_DecodeNarrowPtr()) { 2938 n->subsume_by(in1->in(1), this); 2939 } else if (in1->Opcode() == Op_ConP) { 2940 const Type* t = in1->bottom_type(); 2941 if (t == TypePtr::NULL_PTR) { 2942 assert(t->isa_oopptr(), "null klass?"); 2943 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 2944 } else if (t->isa_oopptr()) { 2945 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 2946 } else if (t->isa_klassptr()) { 2947 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 2948 } 2949 } 2950 if (in1->outcnt() == 0) { 2951 in1->disconnect_inputs(NULL, this); 2952 } 2953 break; 2954 } 2955 2956 case Op_Proj: { 2957 if (OptimizeStringConcat) { 2958 ProjNode* p = n->as_Proj(); 2959 if (p->_is_io_use) { 2960 // Separate projections were used for the exception path which 2961 // are normally removed by a late inline. If it wasn't inlined 2962 // then they will hang around and should just be replaced with 2963 // the original one. 2964 Node* proj = NULL; 2965 // Replace with just one 2966 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2967 Node *use = i.get(); 2968 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2969 proj = use; 2970 break; 2971 } 2972 } 2973 assert(proj != NULL, "must be found"); 2974 p->subsume_by(proj, this); 2975 } 2976 } 2977 break; 2978 } 2979 2980 case Op_Phi: 2981 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2982 // The EncodeP optimization may create Phi with the same edges 2983 // for all paths. It is not handled well by Register Allocator. 2984 Node* unique_in = n->in(1); 2985 assert(unique_in != NULL, ""); 2986 uint cnt = n->req(); 2987 for (uint i = 2; i < cnt; i++) { 2988 Node* m = n->in(i); 2989 assert(m != NULL, ""); 2990 if (unique_in != m) 2991 unique_in = NULL; 2992 } 2993 if (unique_in != NULL) { 2994 n->subsume_by(unique_in, this); 2995 } 2996 } 2997 break; 2998 2999 #endif 3000 3001 case Op_ModI: 3002 if (UseDivMod) { 3003 // Check if a%b and a/b both exist 3004 Node* d = n->find_similar(Op_DivI); 3005 if (d) { 3006 // Replace them with a fused divmod if supported 3007 if (Matcher::has_match_rule(Op_DivModI)) { 3008 DivModINode* divmod = DivModINode::make(n); 3009 d->subsume_by(divmod->div_proj(), this); 3010 n->subsume_by(divmod->mod_proj(), this); 3011 } else { 3012 // replace a%b with a-((a/b)*b) 3013 Node* mult = new MulINode(d, d->in(2)); 3014 Node* sub = new SubINode(d->in(1), mult); 3015 n->subsume_by(sub, this); 3016 } 3017 } 3018 } 3019 break; 3020 3021 case Op_ModL: 3022 if (UseDivMod) { 3023 // Check if a%b and a/b both exist 3024 Node* d = n->find_similar(Op_DivL); 3025 if (d) { 3026 // Replace them with a fused divmod if supported 3027 if (Matcher::has_match_rule(Op_DivModL)) { 3028 DivModLNode* divmod = DivModLNode::make(n); 3029 d->subsume_by(divmod->div_proj(), this); 3030 n->subsume_by(divmod->mod_proj(), this); 3031 } else { 3032 // replace a%b with a-((a/b)*b) 3033 Node* mult = new MulLNode(d, d->in(2)); 3034 Node* sub = new SubLNode(d->in(1), mult); 3035 n->subsume_by(sub, this); 3036 } 3037 } 3038 } 3039 break; 3040 3041 case Op_LoadVector: 3042 case Op_StoreVector: 3043 break; 3044 3045 case Op_PackB: 3046 case Op_PackS: 3047 case Op_PackI: 3048 case Op_PackF: 3049 case Op_PackL: 3050 case Op_PackD: 3051 if (n->req()-1 > 2) { 3052 // Replace many operand PackNodes with a binary tree for matching 3053 PackNode* p = (PackNode*) n; 3054 Node* btp = p->binary_tree_pack(1, n->req()); 3055 n->subsume_by(btp, this); 3056 } 3057 break; 3058 case Op_Loop: 3059 case Op_CountedLoop: 3060 if (n->as_Loop()->is_inner_loop()) { 3061 frc.inc_inner_loop_count(); 3062 } 3063 break; 3064 case Op_LShiftI: 3065 case Op_RShiftI: 3066 case Op_URShiftI: 3067 case Op_LShiftL: 3068 case Op_RShiftL: 3069 case Op_URShiftL: 3070 if (Matcher::need_masked_shift_count) { 3071 // The cpu's shift instructions don't restrict the count to the 3072 // lower 5/6 bits. We need to do the masking ourselves. 3073 Node* in2 = n->in(2); 3074 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3075 const TypeInt* t = in2->find_int_type(); 3076 if (t != NULL && t->is_con()) { 3077 juint shift = t->get_con(); 3078 if (shift > mask) { // Unsigned cmp 3079 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3080 } 3081 } else { 3082 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3083 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3084 n->set_req(2, shift); 3085 } 3086 } 3087 if (in2->outcnt() == 0) { // Remove dead node 3088 in2->disconnect_inputs(NULL, this); 3089 } 3090 } 3091 break; 3092 case Op_MemBarStoreStore: 3093 case Op_MemBarRelease: 3094 // Break the link with AllocateNode: it is no longer useful and 3095 // confuses register allocation. 3096 if (n->req() > MemBarNode::Precedent) { 3097 n->set_req(MemBarNode::Precedent, top()); 3098 } 3099 break; 3100 default: 3101 assert( !n->is_Call(), "" ); 3102 assert( !n->is_Mem(), "" ); 3103 break; 3104 } 3105 3106 // Collect CFG split points 3107 if (n->is_MultiBranch()) 3108 frc._tests.push(n); 3109 } 3110 3111 //------------------------------final_graph_reshaping_walk--------------------- 3112 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3113 // requires that the walk visits a node's inputs before visiting the node. 3114 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3115 ResourceArea *area = Thread::current()->resource_area(); 3116 Unique_Node_List sfpt(area); 3117 3118 frc._visited.set(root->_idx); // first, mark node as visited 3119 uint cnt = root->req(); 3120 Node *n = root; 3121 uint i = 0; 3122 while (true) { 3123 if (i < cnt) { 3124 // Place all non-visited non-null inputs onto stack 3125 Node* m = n->in(i); 3126 ++i; 3127 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3128 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3129 // compute worst case interpreter size in case of a deoptimization 3130 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3131 3132 sfpt.push(m); 3133 } 3134 cnt = m->req(); 3135 nstack.push(n, i); // put on stack parent and next input's index 3136 n = m; 3137 i = 0; 3138 } 3139 } else { 3140 // Now do post-visit work 3141 final_graph_reshaping_impl( n, frc ); 3142 if (nstack.is_empty()) 3143 break; // finished 3144 n = nstack.node(); // Get node from stack 3145 cnt = n->req(); 3146 i = nstack.index(); 3147 nstack.pop(); // Shift to the next node on stack 3148 } 3149 } 3150 3151 // Skip next transformation if compressed oops are not used. 3152 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3153 (!UseCompressedOops && !UseCompressedClassPointers)) 3154 return; 3155 3156 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3157 // It could be done for an uncommon traps or any safepoints/calls 3158 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3159 while (sfpt.size() > 0) { 3160 n = sfpt.pop(); 3161 JVMState *jvms = n->as_SafePoint()->jvms(); 3162 assert(jvms != NULL, "sanity"); 3163 int start = jvms->debug_start(); 3164 int end = n->req(); 3165 bool is_uncommon = (n->is_CallStaticJava() && 3166 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3167 for (int j = start; j < end; j++) { 3168 Node* in = n->in(j); 3169 if (in->is_DecodeNarrowPtr()) { 3170 bool safe_to_skip = true; 3171 if (!is_uncommon ) { 3172 // Is it safe to skip? 3173 for (uint i = 0; i < in->outcnt(); i++) { 3174 Node* u = in->raw_out(i); 3175 if (!u->is_SafePoint() || 3176 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3177 safe_to_skip = false; 3178 } 3179 } 3180 } 3181 if (safe_to_skip) { 3182 n->set_req(j, in->in(1)); 3183 } 3184 if (in->outcnt() == 0) { 3185 in->disconnect_inputs(NULL, this); 3186 } 3187 } 3188 } 3189 } 3190 } 3191 3192 //------------------------------final_graph_reshaping-------------------------- 3193 // Final Graph Reshaping. 3194 // 3195 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3196 // and not commoned up and forced early. Must come after regular 3197 // optimizations to avoid GVN undoing the cloning. Clone constant 3198 // inputs to Loop Phis; these will be split by the allocator anyways. 3199 // Remove Opaque nodes. 3200 // (2) Move last-uses by commutative operations to the left input to encourage 3201 // Intel update-in-place two-address operations and better register usage 3202 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3203 // calls canonicalizing them back. 3204 // (3) Count the number of double-precision FP ops, single-precision FP ops 3205 // and call sites. On Intel, we can get correct rounding either by 3206 // forcing singles to memory (requires extra stores and loads after each 3207 // FP bytecode) or we can set a rounding mode bit (requires setting and 3208 // clearing the mode bit around call sites). The mode bit is only used 3209 // if the relative frequency of single FP ops to calls is low enough. 3210 // This is a key transform for SPEC mpeg_audio. 3211 // (4) Detect infinite loops; blobs of code reachable from above but not 3212 // below. Several of the Code_Gen algorithms fail on such code shapes, 3213 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3214 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3215 // Detection is by looking for IfNodes where only 1 projection is 3216 // reachable from below or CatchNodes missing some targets. 3217 // (5) Assert for insane oop offsets in debug mode. 3218 3219 bool Compile::final_graph_reshaping() { 3220 // an infinite loop may have been eliminated by the optimizer, 3221 // in which case the graph will be empty. 3222 if (root()->req() == 1) { 3223 record_method_not_compilable("trivial infinite loop"); 3224 return true; 3225 } 3226 3227 // Expensive nodes have their control input set to prevent the GVN 3228 // from freely commoning them. There's no GVN beyond this point so 3229 // no need to keep the control input. We want the expensive nodes to 3230 // be freely moved to the least frequent code path by gcm. 3231 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3232 for (int i = 0; i < expensive_count(); i++) { 3233 _expensive_nodes->at(i)->set_req(0, NULL); 3234 } 3235 3236 Final_Reshape_Counts frc; 3237 3238 // Visit everybody reachable! 3239 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3240 Node_Stack nstack(unique() >> 1); 3241 final_graph_reshaping_walk(nstack, root(), frc); 3242 3243 // Check for unreachable (from below) code (i.e., infinite loops). 3244 for( uint i = 0; i < frc._tests.size(); i++ ) { 3245 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3246 // Get number of CFG targets. 3247 // Note that PCTables include exception targets after calls. 3248 uint required_outcnt = n->required_outcnt(); 3249 if (n->outcnt() != required_outcnt) { 3250 // Check for a few special cases. Rethrow Nodes never take the 3251 // 'fall-thru' path, so expected kids is 1 less. 3252 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3253 if (n->in(0)->in(0)->is_Call()) { 3254 CallNode *call = n->in(0)->in(0)->as_Call(); 3255 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3256 required_outcnt--; // Rethrow always has 1 less kid 3257 } else if (call->req() > TypeFunc::Parms && 3258 call->is_CallDynamicJava()) { 3259 // Check for null receiver. In such case, the optimizer has 3260 // detected that the virtual call will always result in a null 3261 // pointer exception. The fall-through projection of this CatchNode 3262 // will not be populated. 3263 Node *arg0 = call->in(TypeFunc::Parms); 3264 if (arg0->is_Type() && 3265 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3266 required_outcnt--; 3267 } 3268 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3269 call->req() > TypeFunc::Parms+1 && 3270 call->is_CallStaticJava()) { 3271 // Check for negative array length. In such case, the optimizer has 3272 // detected that the allocation attempt will always result in an 3273 // exception. There is no fall-through projection of this CatchNode . 3274 Node *arg1 = call->in(TypeFunc::Parms+1); 3275 if (arg1->is_Type() && 3276 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3277 required_outcnt--; 3278 } 3279 } 3280 } 3281 } 3282 // Recheck with a better notion of 'required_outcnt' 3283 if (n->outcnt() != required_outcnt) { 3284 record_method_not_compilable("malformed control flow"); 3285 return true; // Not all targets reachable! 3286 } 3287 } 3288 // Check that I actually visited all kids. Unreached kids 3289 // must be infinite loops. 3290 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3291 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3292 record_method_not_compilable("infinite loop"); 3293 return true; // Found unvisited kid; must be unreach 3294 } 3295 } 3296 3297 // If original bytecodes contained a mixture of floats and doubles 3298 // check if the optimizer has made it homogenous, item (3). 3299 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3300 frc.get_float_count() > 32 && 3301 frc.get_double_count() == 0 && 3302 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3303 set_24_bit_selection_and_mode( false, true ); 3304 } 3305 3306 set_java_calls(frc.get_java_call_count()); 3307 set_inner_loops(frc.get_inner_loop_count()); 3308 3309 // No infinite loops, no reason to bail out. 3310 return false; 3311 } 3312 3313 //-----------------------------too_many_traps---------------------------------- 3314 // Report if there are too many traps at the current method and bci. 3315 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3316 bool Compile::too_many_traps(ciMethod* method, 3317 int bci, 3318 Deoptimization::DeoptReason reason) { 3319 ciMethodData* md = method->method_data(); 3320 if (md->is_empty()) { 3321 // Assume the trap has not occurred, or that it occurred only 3322 // because of a transient condition during start-up in the interpreter. 3323 return false; 3324 } 3325 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3326 if (md->has_trap_at(bci, m, reason) != 0) { 3327 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3328 // Also, if there are multiple reasons, or if there is no per-BCI record, 3329 // assume the worst. 3330 if (log()) 3331 log()->elem("observe trap='%s' count='%d'", 3332 Deoptimization::trap_reason_name(reason), 3333 md->trap_count(reason)); 3334 return true; 3335 } else { 3336 // Ignore method/bci and see if there have been too many globally. 3337 return too_many_traps(reason, md); 3338 } 3339 } 3340 3341 // Less-accurate variant which does not require a method and bci. 3342 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3343 ciMethodData* logmd) { 3344 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3345 // Too many traps globally. 3346 // Note that we use cumulative trap_count, not just md->trap_count. 3347 if (log()) { 3348 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3349 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3350 Deoptimization::trap_reason_name(reason), 3351 mcount, trap_count(reason)); 3352 } 3353 return true; 3354 } else { 3355 // The coast is clear. 3356 return false; 3357 } 3358 } 3359 3360 //--------------------------too_many_recompiles-------------------------------- 3361 // Report if there are too many recompiles at the current method and bci. 3362 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3363 // Is not eager to return true, since this will cause the compiler to use 3364 // Action_none for a trap point, to avoid too many recompilations. 3365 bool Compile::too_many_recompiles(ciMethod* method, 3366 int bci, 3367 Deoptimization::DeoptReason reason) { 3368 ciMethodData* md = method->method_data(); 3369 if (md->is_empty()) { 3370 // Assume the trap has not occurred, or that it occurred only 3371 // because of a transient condition during start-up in the interpreter. 3372 return false; 3373 } 3374 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3375 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3376 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3377 Deoptimization::DeoptReason per_bc_reason 3378 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3379 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3380 if ((per_bc_reason == Deoptimization::Reason_none 3381 || md->has_trap_at(bci, m, reason) != 0) 3382 // The trap frequency measure we care about is the recompile count: 3383 && md->trap_recompiled_at(bci, m) 3384 && md->overflow_recompile_count() >= bc_cutoff) { 3385 // Do not emit a trap here if it has already caused recompilations. 3386 // Also, if there are multiple reasons, or if there is no per-BCI record, 3387 // assume the worst. 3388 if (log()) 3389 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3390 Deoptimization::trap_reason_name(reason), 3391 md->trap_count(reason), 3392 md->overflow_recompile_count()); 3393 return true; 3394 } else if (trap_count(reason) != 0 3395 && decompile_count() >= m_cutoff) { 3396 // Too many recompiles globally, and we have seen this sort of trap. 3397 // Use cumulative decompile_count, not just md->decompile_count. 3398 if (log()) 3399 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3400 Deoptimization::trap_reason_name(reason), 3401 md->trap_count(reason), trap_count(reason), 3402 md->decompile_count(), decompile_count()); 3403 return true; 3404 } else { 3405 // The coast is clear. 3406 return false; 3407 } 3408 } 3409 3410 // Compute when not to trap. Used by matching trap based nodes and 3411 // NullCheck optimization. 3412 void Compile::set_allowed_deopt_reasons() { 3413 _allowed_reasons = 0; 3414 if (is_method_compilation()) { 3415 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3416 assert(rs < BitsPerInt, "recode bit map"); 3417 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3418 _allowed_reasons |= nth_bit(rs); 3419 } 3420 } 3421 } 3422 } 3423 3424 #ifndef PRODUCT 3425 //------------------------------verify_graph_edges--------------------------- 3426 // Walk the Graph and verify that there is a one-to-one correspondence 3427 // between Use-Def edges and Def-Use edges in the graph. 3428 void Compile::verify_graph_edges(bool no_dead_code) { 3429 if (VerifyGraphEdges) { 3430 ResourceArea *area = Thread::current()->resource_area(); 3431 Unique_Node_List visited(area); 3432 // Call recursive graph walk to check edges 3433 _root->verify_edges(visited); 3434 if (no_dead_code) { 3435 // Now make sure that no visited node is used by an unvisited node. 3436 bool dead_nodes = false; 3437 Unique_Node_List checked(area); 3438 while (visited.size() > 0) { 3439 Node* n = visited.pop(); 3440 checked.push(n); 3441 for (uint i = 0; i < n->outcnt(); i++) { 3442 Node* use = n->raw_out(i); 3443 if (checked.member(use)) continue; // already checked 3444 if (visited.member(use)) continue; // already in the graph 3445 if (use->is_Con()) continue; // a dead ConNode is OK 3446 // At this point, we have found a dead node which is DU-reachable. 3447 if (!dead_nodes) { 3448 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3449 dead_nodes = true; 3450 } 3451 use->dump(2); 3452 tty->print_cr("---"); 3453 checked.push(use); // No repeats; pretend it is now checked. 3454 } 3455 } 3456 assert(!dead_nodes, "using nodes must be reachable from root"); 3457 } 3458 } 3459 } 3460 3461 // Verify GC barriers consistency 3462 // Currently supported: 3463 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3464 void Compile::verify_barriers() { 3465 if (UseG1GC) { 3466 // Verify G1 pre-barriers 3467 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3468 3469 ResourceArea *area = Thread::current()->resource_area(); 3470 Unique_Node_List visited(area); 3471 Node_List worklist(area); 3472 // We're going to walk control flow backwards starting from the Root 3473 worklist.push(_root); 3474 while (worklist.size() > 0) { 3475 Node* x = worklist.pop(); 3476 if (x == NULL || x == top()) continue; 3477 if (visited.member(x)) { 3478 continue; 3479 } else { 3480 visited.push(x); 3481 } 3482 3483 if (x->is_Region()) { 3484 for (uint i = 1; i < x->req(); i++) { 3485 worklist.push(x->in(i)); 3486 } 3487 } else { 3488 worklist.push(x->in(0)); 3489 // We are looking for the pattern: 3490 // /->ThreadLocal 3491 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3492 // \->ConI(0) 3493 // We want to verify that the If and the LoadB have the same control 3494 // See GraphKit::g1_write_barrier_pre() 3495 if (x->is_If()) { 3496 IfNode *iff = x->as_If(); 3497 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3498 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3499 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3500 && cmp->in(1)->is_Load()) { 3501 LoadNode* load = cmp->in(1)->as_Load(); 3502 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3503 && load->in(2)->in(3)->is_Con() 3504 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3505 3506 Node* if_ctrl = iff->in(0); 3507 Node* load_ctrl = load->in(0); 3508 3509 if (if_ctrl != load_ctrl) { 3510 // Skip possible CProj->NeverBranch in infinite loops 3511 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3512 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3513 if_ctrl = if_ctrl->in(0)->in(0); 3514 } 3515 } 3516 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3517 } 3518 } 3519 } 3520 } 3521 } 3522 } 3523 } 3524 } 3525 3526 #endif 3527 3528 // The Compile object keeps track of failure reasons separately from the ciEnv. 3529 // This is required because there is not quite a 1-1 relation between the 3530 // ciEnv and its compilation task and the Compile object. Note that one 3531 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3532 // to backtrack and retry without subsuming loads. Other than this backtracking 3533 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3534 // by the logic in C2Compiler. 3535 void Compile::record_failure(const char* reason) { 3536 if (log() != NULL) { 3537 log()->elem("failure reason='%s' phase='compile'", reason); 3538 } 3539 if (_failure_reason == NULL) { 3540 // Record the first failure reason. 3541 _failure_reason = reason; 3542 } 3543 3544 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3545 C->print_method(PHASE_FAILURE); 3546 } 3547 _root = NULL; // flush the graph, too 3548 } 3549 3550 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3551 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3552 _phase_name(name), _dolog(CITimeVerbose) 3553 { 3554 if (_dolog) { 3555 C = Compile::current(); 3556 _log = C->log(); 3557 } else { 3558 C = NULL; 3559 _log = NULL; 3560 } 3561 if (_log != NULL) { 3562 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3563 _log->stamp(); 3564 _log->end_head(); 3565 } 3566 } 3567 3568 Compile::TracePhase::~TracePhase() { 3569 3570 C = Compile::current(); 3571 if (_dolog) { 3572 _log = C->log(); 3573 } else { 3574 _log = NULL; 3575 } 3576 3577 #ifdef ASSERT 3578 if (PrintIdealNodeCount) { 3579 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3580 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3581 } 3582 3583 if (VerifyIdealNodeCount) { 3584 Compile::current()->print_missing_nodes(); 3585 } 3586 #endif 3587 3588 if (_log != NULL) { 3589 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3590 } 3591 } 3592 3593 //============================================================================= 3594 // Two Constant's are equal when the type and the value are equal. 3595 bool Compile::Constant::operator==(const Constant& other) { 3596 if (type() != other.type() ) return false; 3597 if (can_be_reused() != other.can_be_reused()) return false; 3598 // For floating point values we compare the bit pattern. 3599 switch (type()) { 3600 case T_FLOAT: return (_v._value.i == other._v._value.i); 3601 case T_LONG: 3602 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3603 case T_OBJECT: 3604 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3605 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3606 case T_METADATA: return (_v._metadata == other._v._metadata); 3607 default: ShouldNotReachHere(); 3608 } 3609 return false; 3610 } 3611 3612 static int type_to_size_in_bytes(BasicType t) { 3613 switch (t) { 3614 case T_LONG: return sizeof(jlong ); 3615 case T_FLOAT: return sizeof(jfloat ); 3616 case T_DOUBLE: return sizeof(jdouble); 3617 case T_METADATA: return sizeof(Metadata*); 3618 // We use T_VOID as marker for jump-table entries (labels) which 3619 // need an internal word relocation. 3620 case T_VOID: 3621 case T_ADDRESS: 3622 case T_OBJECT: return sizeof(jobject); 3623 } 3624 3625 ShouldNotReachHere(); 3626 return -1; 3627 } 3628 3629 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3630 // sort descending 3631 if (a->freq() > b->freq()) return -1; 3632 if (a->freq() < b->freq()) return 1; 3633 return 0; 3634 } 3635 3636 void Compile::ConstantTable::calculate_offsets_and_size() { 3637 // First, sort the array by frequencies. 3638 _constants.sort(qsort_comparator); 3639 3640 #ifdef ASSERT 3641 // Make sure all jump-table entries were sorted to the end of the 3642 // array (they have a negative frequency). 3643 bool found_void = false; 3644 for (int i = 0; i < _constants.length(); i++) { 3645 Constant con = _constants.at(i); 3646 if (con.type() == T_VOID) 3647 found_void = true; // jump-tables 3648 else 3649 assert(!found_void, "wrong sorting"); 3650 } 3651 #endif 3652 3653 int offset = 0; 3654 for (int i = 0; i < _constants.length(); i++) { 3655 Constant* con = _constants.adr_at(i); 3656 3657 // Align offset for type. 3658 int typesize = type_to_size_in_bytes(con->type()); 3659 offset = align_size_up(offset, typesize); 3660 con->set_offset(offset); // set constant's offset 3661 3662 if (con->type() == T_VOID) { 3663 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3664 offset = offset + typesize * n->outcnt(); // expand jump-table 3665 } else { 3666 offset = offset + typesize; 3667 } 3668 } 3669 3670 // Align size up to the next section start (which is insts; see 3671 // CodeBuffer::align_at_start). 3672 assert(_size == -1, "already set?"); 3673 _size = align_size_up(offset, CodeEntryAlignment); 3674 } 3675 3676 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3677 MacroAssembler _masm(&cb); 3678 for (int i = 0; i < _constants.length(); i++) { 3679 Constant con = _constants.at(i); 3680 address constant_addr; 3681 switch (con.type()) { 3682 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3683 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3684 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3685 case T_OBJECT: { 3686 jobject obj = con.get_jobject(); 3687 int oop_index = _masm.oop_recorder()->find_index(obj); 3688 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3689 break; 3690 } 3691 case T_ADDRESS: { 3692 address addr = (address) con.get_jobject(); 3693 constant_addr = _masm.address_constant(addr); 3694 break; 3695 } 3696 // We use T_VOID as marker for jump-table entries (labels) which 3697 // need an internal word relocation. 3698 case T_VOID: { 3699 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3700 // Fill the jump-table with a dummy word. The real value is 3701 // filled in later in fill_jump_table. 3702 address dummy = (address) n; 3703 constant_addr = _masm.address_constant(dummy); 3704 // Expand jump-table 3705 for (uint i = 1; i < n->outcnt(); i++) { 3706 address temp_addr = _masm.address_constant(dummy + i); 3707 assert(temp_addr, "consts section too small"); 3708 } 3709 break; 3710 } 3711 case T_METADATA: { 3712 Metadata* obj = con.get_metadata(); 3713 int metadata_index = _masm.oop_recorder()->find_index(obj); 3714 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3715 break; 3716 } 3717 default: ShouldNotReachHere(); 3718 } 3719 assert(constant_addr, "consts section too small"); 3720 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3721 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3722 } 3723 } 3724 3725 int Compile::ConstantTable::find_offset(Constant& con) const { 3726 int idx = _constants.find(con); 3727 assert(idx != -1, "constant must be in constant table"); 3728 int offset = _constants.at(idx).offset(); 3729 assert(offset != -1, "constant table not emitted yet?"); 3730 return offset; 3731 } 3732 3733 void Compile::ConstantTable::add(Constant& con) { 3734 if (con.can_be_reused()) { 3735 int idx = _constants.find(con); 3736 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3737 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3738 return; 3739 } 3740 } 3741 (void) _constants.append(con); 3742 } 3743 3744 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3745 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3746 Constant con(type, value, b->_freq); 3747 add(con); 3748 return con; 3749 } 3750 3751 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3752 Constant con(metadata); 3753 add(con); 3754 return con; 3755 } 3756 3757 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3758 jvalue value; 3759 BasicType type = oper->type()->basic_type(); 3760 switch (type) { 3761 case T_LONG: value.j = oper->constantL(); break; 3762 case T_FLOAT: value.f = oper->constantF(); break; 3763 case T_DOUBLE: value.d = oper->constantD(); break; 3764 case T_OBJECT: 3765 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3766 case T_METADATA: return add((Metadata*)oper->constant()); break; 3767 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3768 } 3769 return add(n, type, value); 3770 } 3771 3772 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3773 jvalue value; 3774 // We can use the node pointer here to identify the right jump-table 3775 // as this method is called from Compile::Fill_buffer right before 3776 // the MachNodes are emitted and the jump-table is filled (means the 3777 // MachNode pointers do not change anymore). 3778 value.l = (jobject) n; 3779 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3780 add(con); 3781 return con; 3782 } 3783 3784 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3785 // If called from Compile::scratch_emit_size do nothing. 3786 if (Compile::current()->in_scratch_emit_size()) return; 3787 3788 assert(labels.is_nonempty(), "must be"); 3789 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3790 3791 // Since MachConstantNode::constant_offset() also contains 3792 // table_base_offset() we need to subtract the table_base_offset() 3793 // to get the plain offset into the constant table. 3794 int offset = n->constant_offset() - table_base_offset(); 3795 3796 MacroAssembler _masm(&cb); 3797 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3798 3799 for (uint i = 0; i < n->outcnt(); i++) { 3800 address* constant_addr = &jump_table_base[i]; 3801 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i))); 3802 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3803 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3804 } 3805 } 3806 3807 //----------------------------static_subtype_check----------------------------- 3808 // Shortcut important common cases when superklass is exact: 3809 // (0) superklass is java.lang.Object (can occur in reflective code) 3810 // (1) subklass is already limited to a subtype of superklass => always ok 3811 // (2) subklass does not overlap with superklass => always fail 3812 // (3) superklass has NO subtypes and we can check with a simple compare. 3813 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 3814 if (StressReflectiveCode) { 3815 return SSC_full_test; // Let caller generate the general case. 3816 } 3817 3818 if (superk == env()->Object_klass()) { 3819 return SSC_always_true; // (0) this test cannot fail 3820 } 3821 3822 ciType* superelem = superk; 3823 if (superelem->is_array_klass()) 3824 superelem = superelem->as_array_klass()->base_element_type(); 3825 3826 if (!subk->is_interface()) { // cannot trust static interface types yet 3827 if (subk->is_subtype_of(superk)) { 3828 return SSC_always_true; // (1) false path dead; no dynamic test needed 3829 } 3830 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 3831 !superk->is_subtype_of(subk)) { 3832 return SSC_always_false; 3833 } 3834 } 3835 3836 // If casting to an instance klass, it must have no subtypes 3837 if (superk->is_interface()) { 3838 // Cannot trust interfaces yet. 3839 // %%% S.B. superk->nof_implementors() == 1 3840 } else if (superelem->is_instance_klass()) { 3841 ciInstanceKlass* ik = superelem->as_instance_klass(); 3842 if (!ik->has_subklass() && !ik->is_interface()) { 3843 if (!ik->is_final()) { 3844 // Add a dependency if there is a chance of a later subclass. 3845 dependencies()->assert_leaf_type(ik); 3846 } 3847 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3848 } 3849 } else { 3850 // A primitive array type has no subtypes. 3851 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3852 } 3853 3854 return SSC_full_test; 3855 } 3856 3857 // The message about the current inlining is accumulated in 3858 // _print_inlining_stream and transfered into the _print_inlining_list 3859 // once we know whether inlining succeeds or not. For regular 3860 // inlining, messages are appended to the buffer pointed by 3861 // _print_inlining_idx in the _print_inlining_list. For late inlining, 3862 // a new buffer is added after _print_inlining_idx in the list. This 3863 // way we can update the inlining message for late inlining call site 3864 // when the inlining is attempted again. 3865 void Compile::print_inlining_init() { 3866 if (print_inlining() || print_intrinsics()) { 3867 _print_inlining_stream = new stringStream(); 3868 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 3869 } 3870 } 3871 3872 void Compile::print_inlining_reinit() { 3873 if (print_inlining() || print_intrinsics()) { 3874 // Re allocate buffer when we change ResourceMark 3875 _print_inlining_stream = new stringStream(); 3876 } 3877 } 3878 3879 void Compile::print_inlining_reset() { 3880 _print_inlining_stream->reset(); 3881 } 3882 3883 void Compile::print_inlining_commit() { 3884 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 3885 // Transfer the message from _print_inlining_stream to the current 3886 // _print_inlining_list buffer and clear _print_inlining_stream. 3887 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 3888 print_inlining_reset(); 3889 } 3890 3891 void Compile::print_inlining_push() { 3892 // Add new buffer to the _print_inlining_list at current position 3893 _print_inlining_idx++; 3894 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 3895 } 3896 3897 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 3898 return _print_inlining_list->at(_print_inlining_idx); 3899 } 3900 3901 void Compile::print_inlining_update(CallGenerator* cg) { 3902 if (print_inlining() || print_intrinsics()) { 3903 if (!cg->is_late_inline()) { 3904 if (print_inlining_current().cg() != NULL) { 3905 print_inlining_push(); 3906 } 3907 print_inlining_commit(); 3908 } else { 3909 if (print_inlining_current().cg() != cg && 3910 (print_inlining_current().cg() != NULL || 3911 print_inlining_current().ss()->size() != 0)) { 3912 print_inlining_push(); 3913 } 3914 print_inlining_commit(); 3915 print_inlining_current().set_cg(cg); 3916 } 3917 } 3918 } 3919 3920 void Compile::print_inlining_move_to(CallGenerator* cg) { 3921 // We resume inlining at a late inlining call site. Locate the 3922 // corresponding inlining buffer so that we can update it. 3923 if (print_inlining()) { 3924 for (int i = 0; i < _print_inlining_list->length(); i++) { 3925 if (_print_inlining_list->adr_at(i)->cg() == cg) { 3926 _print_inlining_idx = i; 3927 return; 3928 } 3929 } 3930 ShouldNotReachHere(); 3931 } 3932 } 3933 3934 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 3935 if (print_inlining()) { 3936 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 3937 assert(print_inlining_current().cg() == cg, "wrong entry"); 3938 // replace message with new message 3939 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 3940 print_inlining_commit(); 3941 print_inlining_current().set_cg(cg); 3942 } 3943 } 3944 3945 void Compile::print_inlining_assert_ready() { 3946 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 3947 } 3948 3949 void Compile::process_print_inlining() { 3950 bool do_print_inlining = print_inlining() || print_intrinsics(); 3951 if (do_print_inlining || log() != NULL) { 3952 // Print inlining message for candidates that we couldn't inline 3953 // for lack of space 3954 for (int i = 0; i < _late_inlines.length(); i++) { 3955 CallGenerator* cg = _late_inlines.at(i); 3956 if (!cg->is_mh_late_inline()) { 3957 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 3958 if (do_print_inlining) { 3959 cg->print_inlining_late(msg); 3960 } 3961 log_late_inline_failure(cg, msg); 3962 } 3963 } 3964 } 3965 if (do_print_inlining) { 3966 ResourceMark rm; 3967 stringStream ss; 3968 for (int i = 0; i < _print_inlining_list->length(); i++) { 3969 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 3970 } 3971 size_t end = ss.size(); 3972 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 3973 strncpy(_print_inlining_output, ss.base(), end+1); 3974 _print_inlining_output[end] = 0; 3975 } 3976 } 3977 3978 void Compile::dump_print_inlining() { 3979 if (_print_inlining_output != NULL) { 3980 tty->print_raw(_print_inlining_output); 3981 } 3982 } 3983 3984 void Compile::log_late_inline(CallGenerator* cg) { 3985 if (log() != NULL) { 3986 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 3987 cg->unique_id()); 3988 JVMState* p = cg->call_node()->jvms(); 3989 while (p != NULL) { 3990 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 3991 p = p->caller(); 3992 } 3993 log()->tail("late_inline"); 3994 } 3995 } 3996 3997 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 3998 log_late_inline(cg); 3999 if (log() != NULL) { 4000 log()->inline_fail(msg); 4001 } 4002 } 4003 4004 void Compile::log_inline_id(CallGenerator* cg) { 4005 if (log() != NULL) { 4006 // The LogCompilation tool needs a unique way to identify late 4007 // inline call sites. This id must be unique for this call site in 4008 // this compilation. Try to have it unique across compilations as 4009 // well because it can be convenient when grepping through the log 4010 // file. 4011 // Distinguish OSR compilations from others in case CICountOSR is 4012 // on. 4013 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4014 cg->set_unique_id(id); 4015 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4016 } 4017 } 4018 4019 void Compile::log_inline_failure(const char* msg) { 4020 if (C->log() != NULL) { 4021 C->log()->inline_fail(msg); 4022 } 4023 } 4024 4025 4026 // Dump inlining replay data to the stream. 4027 // Don't change thread state and acquire any locks. 4028 void Compile::dump_inline_data(outputStream* out) { 4029 InlineTree* inl_tree = ilt(); 4030 if (inl_tree != NULL) { 4031 out->print(" inline %d", inl_tree->count()); 4032 inl_tree->dump_replay_data(out); 4033 } 4034 } 4035 4036 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4037 if (n1->Opcode() < n2->Opcode()) return -1; 4038 else if (n1->Opcode() > n2->Opcode()) return 1; 4039 4040 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())); 4041 for (uint i = 1; i < n1->req(); i++) { 4042 if (n1->in(i) < n2->in(i)) return -1; 4043 else if (n1->in(i) > n2->in(i)) return 1; 4044 } 4045 4046 return 0; 4047 } 4048 4049 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4050 Node* n1 = *n1p; 4051 Node* n2 = *n2p; 4052 4053 return cmp_expensive_nodes(n1, n2); 4054 } 4055 4056 void Compile::sort_expensive_nodes() { 4057 if (!expensive_nodes_sorted()) { 4058 _expensive_nodes->sort(cmp_expensive_nodes); 4059 } 4060 } 4061 4062 bool Compile::expensive_nodes_sorted() const { 4063 for (int i = 1; i < _expensive_nodes->length(); i++) { 4064 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4065 return false; 4066 } 4067 } 4068 return true; 4069 } 4070 4071 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4072 if (_expensive_nodes->length() == 0) { 4073 return false; 4074 } 4075 4076 assert(OptimizeExpensiveOps, "optimization off?"); 4077 4078 // Take this opportunity to remove dead nodes from the list 4079 int j = 0; 4080 for (int i = 0; i < _expensive_nodes->length(); i++) { 4081 Node* n = _expensive_nodes->at(i); 4082 if (!n->is_unreachable(igvn)) { 4083 assert(n->is_expensive(), "should be expensive"); 4084 _expensive_nodes->at_put(j, n); 4085 j++; 4086 } 4087 } 4088 _expensive_nodes->trunc_to(j); 4089 4090 // Then sort the list so that similar nodes are next to each other 4091 // and check for at least two nodes of identical kind with same data 4092 // inputs. 4093 sort_expensive_nodes(); 4094 4095 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4096 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4097 return true; 4098 } 4099 } 4100 4101 return false; 4102 } 4103 4104 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4105 if (_expensive_nodes->length() == 0) { 4106 return; 4107 } 4108 4109 assert(OptimizeExpensiveOps, "optimization off?"); 4110 4111 // Sort to bring similar nodes next to each other and clear the 4112 // control input of nodes for which there's only a single copy. 4113 sort_expensive_nodes(); 4114 4115 int j = 0; 4116 int identical = 0; 4117 int i = 0; 4118 bool modified = false; 4119 for (; i < _expensive_nodes->length()-1; i++) { 4120 assert(j <= i, "can't write beyond current index"); 4121 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4122 identical++; 4123 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4124 continue; 4125 } 4126 if (identical > 0) { 4127 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4128 identical = 0; 4129 } else { 4130 Node* n = _expensive_nodes->at(i); 4131 igvn.replace_input_of(n, 0, NULL); 4132 igvn.hash_insert(n); 4133 modified = true; 4134 } 4135 } 4136 if (identical > 0) { 4137 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4138 } else if (_expensive_nodes->length() >= 1) { 4139 Node* n = _expensive_nodes->at(i); 4140 igvn.replace_input_of(n, 0, NULL); 4141 igvn.hash_insert(n); 4142 modified = true; 4143 } 4144 _expensive_nodes->trunc_to(j); 4145 if (modified) { 4146 igvn.optimize(); 4147 } 4148 } 4149 4150 void Compile::add_expensive_node(Node * n) { 4151 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4152 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4153 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4154 if (OptimizeExpensiveOps) { 4155 _expensive_nodes->append(n); 4156 } else { 4157 // Clear control input and let IGVN optimize expensive nodes if 4158 // OptimizeExpensiveOps is off. 4159 n->set_req(0, NULL); 4160 } 4161 } 4162 4163 /** 4164 * Remove the speculative part of types and clean up the graph 4165 */ 4166 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4167 if (UseTypeSpeculation) { 4168 Unique_Node_List worklist; 4169 worklist.push(root()); 4170 int modified = 0; 4171 // Go over all type nodes that carry a speculative type, drop the 4172 // speculative part of the type and enqueue the node for an igvn 4173 // which may optimize it out. 4174 for (uint next = 0; next < worklist.size(); ++next) { 4175 Node *n = worklist.at(next); 4176 if (n->is_Type()) { 4177 TypeNode* tn = n->as_Type(); 4178 const Type* t = tn->type(); 4179 const Type* t_no_spec = t->remove_speculative(); 4180 if (t_no_spec != t) { 4181 bool in_hash = igvn.hash_delete(n); 4182 assert(in_hash, "node should be in igvn hash table"); 4183 tn->set_type(t_no_spec); 4184 igvn.hash_insert(n); 4185 igvn._worklist.push(n); // give it a chance to go away 4186 modified++; 4187 } 4188 } 4189 uint max = n->len(); 4190 for( uint i = 0; i < max; ++i ) { 4191 Node *m = n->in(i); 4192 if (not_a_node(m)) continue; 4193 worklist.push(m); 4194 } 4195 } 4196 // Drop the speculative part of all types in the igvn's type table 4197 igvn.remove_speculative_types(); 4198 if (modified > 0) { 4199 igvn.optimize(); 4200 } 4201 #ifdef ASSERT 4202 // Verify that after the IGVN is over no speculative type has resurfaced 4203 worklist.clear(); 4204 worklist.push(root()); 4205 for (uint next = 0; next < worklist.size(); ++next) { 4206 Node *n = worklist.at(next); 4207 const Type* t = igvn.type_or_null(n); 4208 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4209 if (n->is_Type()) { 4210 t = n->as_Type()->type(); 4211 assert(t == t->remove_speculative(), "no more speculative types"); 4212 } 4213 uint max = n->len(); 4214 for( uint i = 0; i < max; ++i ) { 4215 Node *m = n->in(i); 4216 if (not_a_node(m)) continue; 4217 worklist.push(m); 4218 } 4219 } 4220 igvn.check_no_speculative_types(); 4221 #endif 4222 } 4223 } 4224 4225 // Auxiliary method to support randomized stressing/fuzzing. 4226 // 4227 // This method can be called the arbitrary number of times, with current count 4228 // as the argument. The logic allows selecting a single candidate from the 4229 // running list of candidates as follows: 4230 // int count = 0; 4231 // Cand* selected = null; 4232 // while(cand = cand->next()) { 4233 // if (randomized_select(++count)) { 4234 // selected = cand; 4235 // } 4236 // } 4237 // 4238 // Including count equalizes the chances any candidate is "selected". 4239 // This is useful when we don't have the complete list of candidates to choose 4240 // from uniformly. In this case, we need to adjust the randomicity of the 4241 // selection, or else we will end up biasing the selection towards the latter 4242 // candidates. 4243 // 4244 // Quick back-envelope calculation shows that for the list of n candidates 4245 // the equal probability for the candidate to persist as "best" can be 4246 // achieved by replacing it with "next" k-th candidate with the probability 4247 // of 1/k. It can be easily shown that by the end of the run, the 4248 // probability for any candidate is converged to 1/n, thus giving the 4249 // uniform distribution among all the candidates. 4250 // 4251 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4252 #define RANDOMIZED_DOMAIN_POW 29 4253 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4254 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4255 bool Compile::randomized_select(int count) { 4256 assert(count > 0, "only positive"); 4257 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4258 }