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/convertnode.hpp" 46 #include "opto/divnode.hpp" 47 #include "opto/escape.hpp" 48 #include "opto/idealGraphPrinter.hpp" 49 #include "opto/loopnode.hpp" 50 #include "opto/machnode.hpp" 51 #include "opto/macro.hpp" 52 #include "opto/matcher.hpp" 53 #include "opto/mathexactnode.hpp" 54 #include "opto/memnode.hpp" 55 #include "opto/mulnode.hpp" 56 #include "opto/narrowptrnode.hpp" 57 #include "opto/node.hpp" 58 #include "opto/opcodes.hpp" 59 #include "opto/output.hpp" 60 #include "opto/parse.hpp" 61 #include "opto/phaseX.hpp" 62 #include "opto/rootnode.hpp" 63 #include "opto/runtime.hpp" 64 #include "opto/stringopts.hpp" 65 #include "opto/type.hpp" 66 #include "opto/vectornode.hpp" 67 #include "runtime/arguments.hpp" 68 #include "runtime/signature.hpp" 69 #include "runtime/stubRoutines.hpp" 70 #include "runtime/timer.hpp" 71 #include "utilities/copy.hpp" 72 73 74 // -------------------- Compile::mach_constant_base_node ----------------------- 75 // Constant table base node singleton. 76 MachConstantBaseNode* Compile::mach_constant_base_node() { 77 if (_mach_constant_base_node == NULL) { 78 _mach_constant_base_node = new MachConstantBaseNode(); 79 _mach_constant_base_node->add_req(C->root()); 80 } 81 return _mach_constant_base_node; 82 } 83 84 85 /// Support for intrinsics. 86 87 // Return the index at which m must be inserted (or already exists). 88 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 89 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 90 #ifdef ASSERT 91 for (int i = 1; i < _intrinsics->length(); i++) { 92 CallGenerator* cg1 = _intrinsics->at(i-1); 93 CallGenerator* cg2 = _intrinsics->at(i); 94 assert(cg1->method() != cg2->method() 95 ? cg1->method() < cg2->method() 96 : cg1->is_virtual() < cg2->is_virtual(), 97 "compiler intrinsics list must stay sorted"); 98 } 99 #endif 100 // Binary search sorted list, in decreasing intervals [lo, hi]. 101 int lo = 0, hi = _intrinsics->length()-1; 102 while (lo <= hi) { 103 int mid = (uint)(hi + lo) / 2; 104 ciMethod* mid_m = _intrinsics->at(mid)->method(); 105 if (m < mid_m) { 106 hi = mid-1; 107 } else if (m > mid_m) { 108 lo = mid+1; 109 } else { 110 // look at minor sort key 111 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 112 if (is_virtual < mid_virt) { 113 hi = mid-1; 114 } else if (is_virtual > mid_virt) { 115 lo = mid+1; 116 } else { 117 return mid; // exact match 118 } 119 } 120 } 121 return lo; // inexact match 122 } 123 124 void Compile::register_intrinsic(CallGenerator* cg) { 125 if (_intrinsics == NULL) { 126 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 127 } 128 // This code is stolen from ciObjectFactory::insert. 129 // Really, GrowableArray should have methods for 130 // insert_at, remove_at, and binary_search. 131 int len = _intrinsics->length(); 132 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 133 if (index == len) { 134 _intrinsics->append(cg); 135 } else { 136 #ifdef ASSERT 137 CallGenerator* oldcg = _intrinsics->at(index); 138 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 139 #endif 140 _intrinsics->append(_intrinsics->at(len-1)); 141 int pos; 142 for (pos = len-2; pos >= index; pos--) { 143 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 144 } 145 _intrinsics->at_put(index, cg); 146 } 147 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 148 } 149 150 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 151 assert(m->is_loaded(), "don't try this on unloaded methods"); 152 if (_intrinsics != NULL) { 153 int index = intrinsic_insertion_index(m, is_virtual); 154 if (index < _intrinsics->length() 155 && _intrinsics->at(index)->method() == m 156 && _intrinsics->at(index)->is_virtual() == is_virtual) { 157 return _intrinsics->at(index); 158 } 159 } 160 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 161 if (m->intrinsic_id() != vmIntrinsics::_none && 162 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 163 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 164 if (cg != NULL) { 165 // Save it for next time: 166 register_intrinsic(cg); 167 return cg; 168 } else { 169 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 170 } 171 } 172 return NULL; 173 } 174 175 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 176 // in library_call.cpp. 177 178 179 #ifndef PRODUCT 180 // statistics gathering... 181 182 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 183 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 184 185 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 186 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 187 int oflags = _intrinsic_hist_flags[id]; 188 assert(flags != 0, "what happened?"); 189 if (is_virtual) { 190 flags |= _intrinsic_virtual; 191 } 192 bool changed = (flags != oflags); 193 if ((flags & _intrinsic_worked) != 0) { 194 juint count = (_intrinsic_hist_count[id] += 1); 195 if (count == 1) { 196 changed = true; // first time 197 } 198 // increment the overall count also: 199 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 200 } 201 if (changed) { 202 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 203 // Something changed about the intrinsic's virtuality. 204 if ((flags & _intrinsic_virtual) != 0) { 205 // This is the first use of this intrinsic as a virtual call. 206 if (oflags != 0) { 207 // We already saw it as a non-virtual, so note both cases. 208 flags |= _intrinsic_both; 209 } 210 } else if ((oflags & _intrinsic_both) == 0) { 211 // This is the first use of this intrinsic as a non-virtual 212 flags |= _intrinsic_both; 213 } 214 } 215 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 216 } 217 // update the overall flags also: 218 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 219 return changed; 220 } 221 222 static char* format_flags(int flags, char* buf) { 223 buf[0] = 0; 224 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 225 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 226 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 227 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 228 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 229 if (buf[0] == 0) strcat(buf, ","); 230 assert(buf[0] == ',', "must be"); 231 return &buf[1]; 232 } 233 234 void Compile::print_intrinsic_statistics() { 235 char flagsbuf[100]; 236 ttyLocker ttyl; 237 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 238 tty->print_cr("Compiler intrinsic usage:"); 239 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 240 if (total == 0) total = 1; // avoid div0 in case of no successes 241 #define PRINT_STAT_LINE(name, c, f) \ 242 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 243 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 244 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 245 int flags = _intrinsic_hist_flags[id]; 246 juint count = _intrinsic_hist_count[id]; 247 if ((flags | count) != 0) { 248 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 249 } 250 } 251 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 252 if (xtty != NULL) xtty->tail("statistics"); 253 } 254 255 void Compile::print_statistics() { 256 { ttyLocker ttyl; 257 if (xtty != NULL) xtty->head("statistics type='opto'"); 258 Parse::print_statistics(); 259 PhaseCCP::print_statistics(); 260 PhaseRegAlloc::print_statistics(); 261 Scheduling::print_statistics(); 262 PhasePeephole::print_statistics(); 263 PhaseIdealLoop::print_statistics(); 264 if (xtty != NULL) xtty->tail("statistics"); 265 } 266 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 267 // put this under its own <statistics> element. 268 print_intrinsic_statistics(); 269 } 270 } 271 #endif //PRODUCT 272 273 // Support for bundling info 274 Bundle* Compile::node_bundling(const Node *n) { 275 assert(valid_bundle_info(n), "oob"); 276 return &_node_bundling_base[n->_idx]; 277 } 278 279 bool Compile::valid_bundle_info(const Node *n) { 280 return (_node_bundling_limit > n->_idx); 281 } 282 283 284 void Compile::gvn_replace_by(Node* n, Node* nn) { 285 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 286 Node* use = n->last_out(i); 287 bool is_in_table = initial_gvn()->hash_delete(use); 288 uint uses_found = 0; 289 for (uint j = 0; j < use->len(); j++) { 290 if (use->in(j) == n) { 291 if (j < use->req()) 292 use->set_req(j, nn); 293 else 294 use->set_prec(j, nn); 295 uses_found++; 296 } 297 } 298 if (is_in_table) { 299 // reinsert into table 300 initial_gvn()->hash_find_insert(use); 301 } 302 record_for_igvn(use); 303 i -= uses_found; // we deleted 1 or more copies of this edge 304 } 305 } 306 307 308 static inline bool not_a_node(const Node* n) { 309 if (n == NULL) return true; 310 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 311 if (*(address*)n == badAddress) return true; // kill by Node::destruct 312 return false; 313 } 314 315 // Identify all nodes that are reachable from below, useful. 316 // Use breadth-first pass that records state in a Unique_Node_List, 317 // recursive traversal is slower. 318 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 319 int estimated_worklist_size = unique(); 320 useful.map( estimated_worklist_size, NULL ); // preallocate space 321 322 // Initialize worklist 323 if (root() != NULL) { useful.push(root()); } 324 // If 'top' is cached, declare it useful to preserve cached node 325 if( cached_top_node() ) { useful.push(cached_top_node()); } 326 327 // Push all useful nodes onto the list, breadthfirst 328 for( uint next = 0; next < useful.size(); ++next ) { 329 assert( next < unique(), "Unique useful nodes < total nodes"); 330 Node *n = useful.at(next); 331 uint max = n->len(); 332 for( uint i = 0; i < max; ++i ) { 333 Node *m = n->in(i); 334 if (not_a_node(m)) continue; 335 useful.push(m); 336 } 337 } 338 } 339 340 // Update dead_node_list with any missing dead nodes using useful 341 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 342 void Compile::update_dead_node_list(Unique_Node_List &useful) { 343 uint max_idx = unique(); 344 VectorSet& useful_node_set = useful.member_set(); 345 346 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 347 // If node with index node_idx is not in useful set, 348 // mark it as dead in dead node list. 349 if (! useful_node_set.test(node_idx) ) { 350 record_dead_node(node_idx); 351 } 352 } 353 } 354 355 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 356 int shift = 0; 357 for (int i = 0; i < inlines->length(); i++) { 358 CallGenerator* cg = inlines->at(i); 359 CallNode* call = cg->call_node(); 360 if (shift > 0) { 361 inlines->at_put(i-shift, cg); 362 } 363 if (!useful.member(call)) { 364 shift++; 365 } 366 } 367 inlines->trunc_to(inlines->length()-shift); 368 } 369 370 // Disconnect all useless nodes by disconnecting those at the boundary. 371 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 372 uint next = 0; 373 while (next < useful.size()) { 374 Node *n = useful.at(next++); 375 if (n->is_SafePoint()) { 376 // We're done with a parsing phase. Replaced nodes are not valid 377 // beyond that point. 378 n->as_SafePoint()->delete_replaced_nodes(); 379 } 380 // Use raw traversal of out edges since this code removes out edges 381 int max = n->outcnt(); 382 for (int j = 0; j < max; ++j) { 383 Node* child = n->raw_out(j); 384 if (! useful.member(child)) { 385 assert(!child->is_top() || child != top(), 386 "If top is cached in Compile object it is in useful list"); 387 // Only need to remove this out-edge to the useless node 388 n->raw_del_out(j); 389 --j; 390 --max; 391 } 392 } 393 if (n->outcnt() == 1 && n->has_special_unique_user()) { 394 record_for_igvn(n->unique_out()); 395 } 396 } 397 // Remove useless macro and predicate opaq nodes 398 for (int i = C->macro_count()-1; i >= 0; i--) { 399 Node* n = C->macro_node(i); 400 if (!useful.member(n)) { 401 remove_macro_node(n); 402 } 403 } 404 // Remove useless expensive node 405 for (int i = C->expensive_count()-1; i >= 0; i--) { 406 Node* n = C->expensive_node(i); 407 if (!useful.member(n)) { 408 remove_expensive_node(n); 409 } 410 } 411 // clean up the late inline lists 412 remove_useless_late_inlines(&_string_late_inlines, useful); 413 remove_useless_late_inlines(&_boxing_late_inlines, useful); 414 remove_useless_late_inlines(&_late_inlines, useful); 415 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 416 } 417 418 //------------------------------frame_size_in_words----------------------------- 419 // frame_slots in units of words 420 int Compile::frame_size_in_words() const { 421 // shift is 0 in LP32 and 1 in LP64 422 const int shift = (LogBytesPerWord - LogBytesPerInt); 423 int words = _frame_slots >> shift; 424 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 425 return words; 426 } 427 428 // To bang the stack of this compiled method we use the stack size 429 // that the interpreter would need in case of a deoptimization. This 430 // removes the need to bang the stack in the deoptimization blob which 431 // in turn simplifies stack overflow handling. 432 int Compile::bang_size_in_bytes() const { 433 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 434 } 435 436 // ============================================================================ 437 //------------------------------CompileWrapper--------------------------------- 438 class CompileWrapper : public StackObj { 439 Compile *const _compile; 440 public: 441 CompileWrapper(Compile* compile); 442 443 ~CompileWrapper(); 444 }; 445 446 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 447 // the Compile* pointer is stored in the current ciEnv: 448 ciEnv* env = compile->env(); 449 assert(env == ciEnv::current(), "must already be a ciEnv active"); 450 assert(env->compiler_data() == NULL, "compile already active?"); 451 env->set_compiler_data(compile); 452 assert(compile == Compile::current(), "sanity"); 453 454 compile->set_type_dict(NULL); 455 compile->set_type_hwm(NULL); 456 compile->set_type_last_size(0); 457 compile->set_last_tf(NULL, NULL); 458 compile->set_indexSet_arena(NULL); 459 compile->set_indexSet_free_block_list(NULL); 460 compile->init_type_arena(); 461 Type::Initialize(compile); 462 _compile->set_scratch_buffer_blob(NULL); 463 _compile->begin_method(); 464 } 465 CompileWrapper::~CompileWrapper() { 466 _compile->end_method(); 467 if (_compile->scratch_buffer_blob() != NULL) 468 BufferBlob::free(_compile->scratch_buffer_blob()); 469 _compile->env()->set_compiler_data(NULL); 470 } 471 472 473 //----------------------------print_compile_messages--------------------------- 474 void Compile::print_compile_messages() { 475 #ifndef PRODUCT 476 // Check if recompiling 477 if (_subsume_loads == false && PrintOpto) { 478 // Recompiling without allowing machine instructions to subsume loads 479 tty->print_cr("*********************************************************"); 480 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 481 tty->print_cr("*********************************************************"); 482 } 483 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 484 // Recompiling without escape analysis 485 tty->print_cr("*********************************************************"); 486 tty->print_cr("** Bailout: Recompile without escape analysis **"); 487 tty->print_cr("*********************************************************"); 488 } 489 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 490 // Recompiling without boxing elimination 491 tty->print_cr("*********************************************************"); 492 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 493 tty->print_cr("*********************************************************"); 494 } 495 if (env()->break_at_compile()) { 496 // Open the debugger when compiling this method. 497 tty->print("### Breaking when compiling: "); 498 method()->print_short_name(); 499 tty->cr(); 500 BREAKPOINT; 501 } 502 503 if( PrintOpto ) { 504 if (is_osr_compilation()) { 505 tty->print("[OSR]%3d", _compile_id); 506 } else { 507 tty->print("%3d", _compile_id); 508 } 509 } 510 #endif 511 } 512 513 514 //-----------------------init_scratch_buffer_blob------------------------------ 515 // Construct a temporary BufferBlob and cache it for this compile. 516 void Compile::init_scratch_buffer_blob(int const_size) { 517 // If there is already a scratch buffer blob allocated and the 518 // constant section is big enough, use it. Otherwise free the 519 // current and allocate a new one. 520 BufferBlob* blob = scratch_buffer_blob(); 521 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 522 // Use the current blob. 523 } else { 524 if (blob != NULL) { 525 BufferBlob::free(blob); 526 } 527 528 ResourceMark rm; 529 _scratch_const_size = const_size; 530 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 531 blob = BufferBlob::create("Compile::scratch_buffer", size); 532 // Record the buffer blob for next time. 533 set_scratch_buffer_blob(blob); 534 // Have we run out of code space? 535 if (scratch_buffer_blob() == NULL) { 536 // Let CompilerBroker disable further compilations. 537 record_failure("Not enough space for scratch buffer in CodeCache"); 538 return; 539 } 540 } 541 542 // Initialize the relocation buffers 543 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 544 set_scratch_locs_memory(locs_buf); 545 } 546 547 548 //-----------------------scratch_emit_size------------------------------------- 549 // Helper function that computes size by emitting code 550 uint Compile::scratch_emit_size(const Node* n) { 551 // Start scratch_emit_size section. 552 set_in_scratch_emit_size(true); 553 554 // Emit into a trash buffer and count bytes emitted. 555 // This is a pretty expensive way to compute a size, 556 // but it works well enough if seldom used. 557 // All common fixed-size instructions are given a size 558 // method by the AD file. 559 // Note that the scratch buffer blob and locs memory are 560 // allocated at the beginning of the compile task, and 561 // may be shared by several calls to scratch_emit_size. 562 // The allocation of the scratch buffer blob is particularly 563 // expensive, since it has to grab the code cache lock. 564 BufferBlob* blob = this->scratch_buffer_blob(); 565 assert(blob != NULL, "Initialize BufferBlob at start"); 566 assert(blob->size() > MAX_inst_size, "sanity"); 567 relocInfo* locs_buf = scratch_locs_memory(); 568 address blob_begin = blob->content_begin(); 569 address blob_end = (address)locs_buf; 570 assert(blob->content_contains(blob_end), "sanity"); 571 CodeBuffer buf(blob_begin, blob_end - blob_begin); 572 buf.initialize_consts_size(_scratch_const_size); 573 buf.initialize_stubs_size(MAX_stubs_size); 574 assert(locs_buf != NULL, "sanity"); 575 int lsize = MAX_locs_size / 3; 576 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 577 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 578 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 579 580 // Do the emission. 581 582 Label fakeL; // Fake label for branch instructions. 583 Label* saveL = NULL; 584 uint save_bnum = 0; 585 bool is_branch = n->is_MachBranch(); 586 if (is_branch) { 587 MacroAssembler masm(&buf); 588 masm.bind(fakeL); 589 n->as_MachBranch()->save_label(&saveL, &save_bnum); 590 n->as_MachBranch()->label_set(&fakeL, 0); 591 } 592 n->emit(buf, this->regalloc()); 593 if (is_branch) // Restore label. 594 n->as_MachBranch()->label_set(saveL, save_bnum); 595 596 // End scratch_emit_size section. 597 set_in_scratch_emit_size(false); 598 599 return buf.insts_size(); 600 } 601 602 603 // ============================================================================ 604 //------------------------------Compile standard------------------------------- 605 debug_only( int Compile::_debug_idx = 100000; ) 606 607 // Compile a method. entry_bci is -1 for normal compilations and indicates 608 // the continuation bci for on stack replacement. 609 610 611 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 612 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 613 : Phase(Compiler), 614 _env(ci_env), 615 _log(ci_env->log()), 616 _compile_id(ci_env->compile_id()), 617 _save_argument_registers(false), 618 _stub_name(NULL), 619 _stub_function(NULL), 620 _stub_entry_point(NULL), 621 _method(target), 622 _entry_bci(osr_bci), 623 _initial_gvn(NULL), 624 _for_igvn(NULL), 625 _warm_calls(NULL), 626 _subsume_loads(subsume_loads), 627 _do_escape_analysis(do_escape_analysis), 628 _eliminate_boxing(eliminate_boxing), 629 _failure_reason(NULL), 630 _code_buffer("Compile::Fill_buffer"), 631 _orig_pc_slot(0), 632 _orig_pc_slot_offset_in_bytes(0), 633 _has_method_handle_invokes(false), 634 _mach_constant_base_node(NULL), 635 _node_bundling_limit(0), 636 _node_bundling_base(NULL), 637 _java_calls(0), 638 _inner_loops(0), 639 _scratch_const_size(-1), 640 _in_scratch_emit_size(false), 641 _dead_node_list(comp_arena()), 642 _dead_node_count(0), 643 #ifndef PRODUCT 644 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 645 _in_dump_cnt(0), 646 _printer(IdealGraphPrinter::printer()), 647 #endif 648 _congraph(NULL), 649 _comp_arena(mtCompiler), 650 _node_arena(mtCompiler), 651 _old_arena(mtCompiler), 652 _Compile_types(mtCompiler), 653 _replay_inline_data(NULL), 654 _late_inlines(comp_arena(), 2, 0, NULL), 655 _string_late_inlines(comp_arena(), 2, 0, NULL), 656 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 657 _late_inlines_pos(0), 658 _number_of_mh_late_inlines(0), 659 _inlining_progress(false), 660 _inlining_incrementally(false), 661 _print_inlining_list(NULL), 662 _print_inlining_stream(NULL), 663 _print_inlining_idx(0), 664 _print_inlining_output(NULL), 665 _interpreter_frame_size(0), 666 _max_node_limit(MaxNodeLimit) { 667 C = this; 668 669 CompileWrapper cw(this); 670 671 if (CITimeVerbose) { 672 tty->print(" "); 673 target->holder()->name()->print(); 674 tty->print("."); 675 target->print_short_name(); 676 tty->print(" "); 677 } 678 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 679 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 680 681 #ifndef PRODUCT 682 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 683 if (!print_opto_assembly) { 684 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 685 if (print_assembly && !Disassembler::can_decode()) { 686 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 687 print_opto_assembly = true; 688 } 689 } 690 set_print_assembly(print_opto_assembly); 691 set_parsed_irreducible_loop(false); 692 693 if (method()->has_option("ReplayInline")) { 694 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 695 } 696 #endif 697 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 698 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 699 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 700 701 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 702 // Make sure the method being compiled gets its own MDO, 703 // so we can at least track the decompile_count(). 704 // Need MDO to record RTM code generation state. 705 method()->ensure_method_data(); 706 } 707 708 Init(::AliasLevel); 709 710 711 print_compile_messages(); 712 713 _ilt = InlineTree::build_inline_tree_root(); 714 715 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 716 assert(num_alias_types() >= AliasIdxRaw, ""); 717 718 #define MINIMUM_NODE_HASH 1023 719 // Node list that Iterative GVN will start with 720 Unique_Node_List for_igvn(comp_arena()); 721 set_for_igvn(&for_igvn); 722 723 // GVN that will be run immediately on new nodes 724 uint estimated_size = method()->code_size()*4+64; 725 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 726 PhaseGVN gvn(node_arena(), estimated_size); 727 set_initial_gvn(&gvn); 728 729 print_inlining_init(); 730 { // Scope for timing the parser 731 TracePhase tp("parse", &timers[_t_parser]); 732 733 // Put top into the hash table ASAP. 734 initial_gvn()->transform_no_reclaim(top()); 735 736 // Set up tf(), start(), and find a CallGenerator. 737 CallGenerator* cg = NULL; 738 if (is_osr_compilation()) { 739 const TypeTuple *domain = StartOSRNode::osr_domain(); 740 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 741 init_tf(TypeFunc::make(domain, range)); 742 StartNode* s = new StartOSRNode(root(), domain); 743 initial_gvn()->set_type_bottom(s); 744 init_start(s); 745 cg = CallGenerator::for_osr(method(), entry_bci()); 746 } else { 747 // Normal case. 748 init_tf(TypeFunc::make(method())); 749 StartNode* s = new StartNode(root(), tf()->domain()); 750 initial_gvn()->set_type_bottom(s); 751 init_start(s); 752 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 753 // With java.lang.ref.reference.get() we must go through the 754 // intrinsic when G1 is enabled - even when get() is the root 755 // method of the compile - so that, if necessary, the value in 756 // the referent field of the reference object gets recorded by 757 // the pre-barrier code. 758 // Specifically, if G1 is enabled, the value in the referent 759 // field is recorded by the G1 SATB pre barrier. This will 760 // result in the referent being marked live and the reference 761 // object removed from the list of discovered references during 762 // reference processing. 763 cg = find_intrinsic(method(), false); 764 } 765 if (cg == NULL) { 766 float past_uses = method()->interpreter_invocation_count(); 767 float expected_uses = past_uses; 768 cg = CallGenerator::for_inline(method(), expected_uses); 769 } 770 } 771 if (failing()) return; 772 if (cg == NULL) { 773 record_method_not_compilable_all_tiers("cannot parse method"); 774 return; 775 } 776 JVMState* jvms = build_start_state(start(), tf()); 777 if ((jvms = cg->generate(jvms)) == NULL) { 778 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 779 record_method_not_compilable("method parse failed"); 780 } 781 return; 782 } 783 GraphKit kit(jvms); 784 785 if (!kit.stopped()) { 786 // Accept return values, and transfer control we know not where. 787 // This is done by a special, unique ReturnNode bound to root. 788 return_values(kit.jvms()); 789 } 790 791 if (kit.has_exceptions()) { 792 // Any exceptions that escape from this call must be rethrown 793 // to whatever caller is dynamically above us on the stack. 794 // This is done by a special, unique RethrowNode bound to root. 795 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 796 } 797 798 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 799 800 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 801 inline_string_calls(true); 802 } 803 804 if (failing()) return; 805 806 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 807 808 // Remove clutter produced by parsing. 809 if (!failing()) { 810 ResourceMark rm; 811 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 812 } 813 } 814 815 // Note: Large methods are capped off in do_one_bytecode(). 816 if (failing()) return; 817 818 // After parsing, node notes are no longer automagic. 819 // They must be propagated by register_new_node_with_optimizer(), 820 // clone(), or the like. 821 set_default_node_notes(NULL); 822 823 for (;;) { 824 int successes = Inline_Warm(); 825 if (failing()) return; 826 if (successes == 0) break; 827 } 828 829 // Drain the list. 830 Finish_Warm(); 831 #ifndef PRODUCT 832 if (_printer && _printer->should_print(_method)) { 833 _printer->print_inlining(this); 834 } 835 #endif 836 837 if (failing()) return; 838 NOT_PRODUCT( verify_graph_edges(); ) 839 840 // Now optimize 841 Optimize(); 842 if (failing()) return; 843 NOT_PRODUCT( verify_graph_edges(); ) 844 845 #ifndef PRODUCT 846 if (PrintIdeal) { 847 ttyLocker ttyl; // keep the following output all in one block 848 // This output goes directly to the tty, not the compiler log. 849 // To enable tools to match it up with the compilation activity, 850 // be sure to tag this tty output with the compile ID. 851 if (xtty != NULL) { 852 xtty->head("ideal compile_id='%d'%s", compile_id(), 853 is_osr_compilation() ? " compile_kind='osr'" : 854 ""); 855 } 856 root()->dump(9999); 857 if (xtty != NULL) { 858 xtty->tail("ideal"); 859 } 860 } 861 #endif 862 863 NOT_PRODUCT( verify_barriers(); ) 864 865 // Dump compilation data to replay it. 866 if (method()->has_option("DumpReplay")) { 867 env()->dump_replay_data(_compile_id); 868 } 869 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 870 env()->dump_inline_data(_compile_id); 871 } 872 873 // Now that we know the size of all the monitors we can add a fixed slot 874 // for the original deopt pc. 875 876 _orig_pc_slot = fixed_slots(); 877 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 878 set_fixed_slots(next_slot); 879 880 // Compute when to use implicit null checks. Used by matching trap based 881 // nodes and NullCheck optimization. 882 set_allowed_deopt_reasons(); 883 884 // Now generate code 885 Code_Gen(); 886 if (failing()) return; 887 888 // Check if we want to skip execution of all compiled code. 889 { 890 #ifndef PRODUCT 891 if (OptoNoExecute) { 892 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 893 return; 894 } 895 #endif 896 TracePhase tp("install_code", &timers[_t_registerMethod]); 897 898 if (is_osr_compilation()) { 899 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 900 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 901 } else { 902 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 903 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 904 } 905 906 env()->register_method(_method, _entry_bci, 907 &_code_offsets, 908 _orig_pc_slot_offset_in_bytes, 909 code_buffer(), 910 frame_size_in_words(), _oop_map_set, 911 &_handler_table, &_inc_table, 912 compiler, 913 env()->comp_level(), 914 has_unsafe_access(), 915 SharedRuntime::is_wide_vector(max_vector_size()), 916 rtm_state() 917 ); 918 919 if (log() != NULL) // Print code cache state into compiler log 920 log()->code_cache_state(); 921 } 922 } 923 924 //------------------------------Compile---------------------------------------- 925 // Compile a runtime stub 926 Compile::Compile( ciEnv* ci_env, 927 TypeFunc_generator generator, 928 address stub_function, 929 const char *stub_name, 930 int is_fancy_jump, 931 bool pass_tls, 932 bool save_arg_registers, 933 bool return_pc ) 934 : Phase(Compiler), 935 _env(ci_env), 936 _log(ci_env->log()), 937 _compile_id(0), 938 _save_argument_registers(save_arg_registers), 939 _method(NULL), 940 _stub_name(stub_name), 941 _stub_function(stub_function), 942 _stub_entry_point(NULL), 943 _entry_bci(InvocationEntryBci), 944 _initial_gvn(NULL), 945 _for_igvn(NULL), 946 _warm_calls(NULL), 947 _orig_pc_slot(0), 948 _orig_pc_slot_offset_in_bytes(0), 949 _subsume_loads(true), 950 _do_escape_analysis(false), 951 _eliminate_boxing(false), 952 _failure_reason(NULL), 953 _code_buffer("Compile::Fill_buffer"), 954 _has_method_handle_invokes(false), 955 _mach_constant_base_node(NULL), 956 _node_bundling_limit(0), 957 _node_bundling_base(NULL), 958 _java_calls(0), 959 _inner_loops(0), 960 #ifndef PRODUCT 961 _trace_opto_output(TraceOptoOutput), 962 _in_dump_cnt(0), 963 _printer(NULL), 964 #endif 965 _comp_arena(mtCompiler), 966 _node_arena(mtCompiler), 967 _old_arena(mtCompiler), 968 _Compile_types(mtCompiler), 969 _dead_node_list(comp_arena()), 970 _dead_node_count(0), 971 _congraph(NULL), 972 _replay_inline_data(NULL), 973 _number_of_mh_late_inlines(0), 974 _inlining_progress(false), 975 _inlining_incrementally(false), 976 _print_inlining_list(NULL), 977 _print_inlining_stream(NULL), 978 _print_inlining_idx(0), 979 _print_inlining_output(NULL), 980 _allowed_reasons(0), 981 _interpreter_frame_size(0), 982 _max_node_limit(MaxNodeLimit) { 983 C = this; 984 985 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 986 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 987 988 #ifndef PRODUCT 989 set_print_assembly(PrintFrameConverterAssembly); 990 set_parsed_irreducible_loop(false); 991 #endif 992 set_has_irreducible_loop(false); // no loops 993 994 CompileWrapper cw(this); 995 Init(/*AliasLevel=*/ 0); 996 init_tf((*generator)()); 997 998 { 999 // The following is a dummy for the sake of GraphKit::gen_stub 1000 Unique_Node_List for_igvn(comp_arena()); 1001 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1002 PhaseGVN gvn(Thread::current()->resource_area(),255); 1003 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1004 gvn.transform_no_reclaim(top()); 1005 1006 GraphKit kit; 1007 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1008 } 1009 1010 NOT_PRODUCT( verify_graph_edges(); ) 1011 Code_Gen(); 1012 if (failing()) return; 1013 1014 1015 // Entry point will be accessed using compile->stub_entry_point(); 1016 if (code_buffer() == NULL) { 1017 Matcher::soft_match_failure(); 1018 } else { 1019 if (PrintAssembly && (WizardMode || Verbose)) 1020 tty->print_cr("### Stub::%s", stub_name); 1021 1022 if (!failing()) { 1023 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1024 1025 // Make the NMethod 1026 // For now we mark the frame as never safe for profile stackwalking 1027 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1028 code_buffer(), 1029 CodeOffsets::frame_never_safe, 1030 // _code_offsets.value(CodeOffsets::Frame_Complete), 1031 frame_size_in_words(), 1032 _oop_map_set, 1033 save_arg_registers); 1034 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1035 1036 _stub_entry_point = rs->entry_point(); 1037 } 1038 } 1039 } 1040 1041 //------------------------------Init------------------------------------------- 1042 // Prepare for a single compilation 1043 void Compile::Init(int aliaslevel) { 1044 _unique = 0; 1045 _regalloc = NULL; 1046 1047 _tf = NULL; // filled in later 1048 _top = NULL; // cached later 1049 _matcher = NULL; // filled in later 1050 _cfg = NULL; // filled in later 1051 1052 set_24_bit_selection_and_mode(Use24BitFP, false); 1053 1054 _node_note_array = NULL; 1055 _default_node_notes = NULL; 1056 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1057 1058 _immutable_memory = NULL; // filled in at first inquiry 1059 1060 // Globally visible Nodes 1061 // First set TOP to NULL to give safe behavior during creation of RootNode 1062 set_cached_top_node(NULL); 1063 set_root(new RootNode()); 1064 // Now that you have a Root to point to, create the real TOP 1065 set_cached_top_node( new ConNode(Type::TOP) ); 1066 set_recent_alloc(NULL, NULL); 1067 1068 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1069 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1070 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1071 env()->set_dependencies(new Dependencies(env())); 1072 1073 _fixed_slots = 0; 1074 set_has_split_ifs(false); 1075 set_has_loops(has_method() && method()->has_loops()); // first approximation 1076 set_has_stringbuilder(false); 1077 set_has_boxed_value(false); 1078 _trap_can_recompile = false; // no traps emitted yet 1079 _major_progress = true; // start out assuming good things will happen 1080 set_has_unsafe_access(false); 1081 set_max_vector_size(0); 1082 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1083 set_decompile_count(0); 1084 1085 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1086 set_num_loop_opts(LoopOptsCount); 1087 set_do_inlining(Inline); 1088 set_max_inline_size(MaxInlineSize); 1089 set_freq_inline_size(FreqInlineSize); 1090 set_do_scheduling(OptoScheduling); 1091 set_do_count_invocations(false); 1092 set_do_method_data_update(false); 1093 set_age_code(has_method() && method()->profile_aging()); 1094 set_rtm_state(NoRTM); // No RTM lock eliding by default 1095 method_has_option_value("MaxNodeLimit", _max_node_limit); 1096 #if INCLUDE_RTM_OPT 1097 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1098 int rtm_state = method()->method_data()->rtm_state(); 1099 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1100 // Don't generate RTM lock eliding code. 1101 set_rtm_state(NoRTM); 1102 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1103 // Generate RTM lock eliding code without abort ratio calculation code. 1104 set_rtm_state(UseRTM); 1105 } else if (UseRTMDeopt) { 1106 // Generate RTM lock eliding code and include abort ratio calculation 1107 // code if UseRTMDeopt is on. 1108 set_rtm_state(ProfileRTM); 1109 } 1110 } 1111 #endif 1112 if (debug_info()->recording_non_safepoints()) { 1113 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1114 (comp_arena(), 8, 0, NULL)); 1115 set_default_node_notes(Node_Notes::make(this)); 1116 } 1117 1118 // // -- Initialize types before each compile -- 1119 // // Update cached type information 1120 // if( _method && _method->constants() ) 1121 // Type::update_loaded_types(_method, _method->constants()); 1122 1123 // Init alias_type map. 1124 if (!_do_escape_analysis && aliaslevel == 3) 1125 aliaslevel = 2; // No unique types without escape analysis 1126 _AliasLevel = aliaslevel; 1127 const int grow_ats = 16; 1128 _max_alias_types = grow_ats; 1129 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1130 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1131 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1132 { 1133 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1134 } 1135 // Initialize the first few types. 1136 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1137 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1138 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1139 _num_alias_types = AliasIdxRaw+1; 1140 // Zero out the alias type cache. 1141 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1142 // A NULL adr_type hits in the cache right away. Preload the right answer. 1143 probe_alias_cache(NULL)->_index = AliasIdxTop; 1144 1145 _intrinsics = NULL; 1146 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1147 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1148 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1149 register_library_intrinsics(); 1150 } 1151 1152 //---------------------------init_start---------------------------------------- 1153 // Install the StartNode on this compile object. 1154 void Compile::init_start(StartNode* s) { 1155 if (failing()) 1156 return; // already failing 1157 assert(s == start(), ""); 1158 } 1159 1160 /** 1161 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1162 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1163 * the ideal graph. 1164 */ 1165 StartNode* Compile::start() const { 1166 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 1167 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1168 Node* start = root()->fast_out(i); 1169 if (start->is_Start()) { 1170 return start->as_Start(); 1171 } 1172 } 1173 fatal("Did not find Start node!"); 1174 return NULL; 1175 } 1176 1177 //-------------------------------immutable_memory------------------------------------- 1178 // Access immutable memory 1179 Node* Compile::immutable_memory() { 1180 if (_immutable_memory != NULL) { 1181 return _immutable_memory; 1182 } 1183 StartNode* s = start(); 1184 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1185 Node *p = s->fast_out(i); 1186 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1187 _immutable_memory = p; 1188 return _immutable_memory; 1189 } 1190 } 1191 ShouldNotReachHere(); 1192 return NULL; 1193 } 1194 1195 //----------------------set_cached_top_node------------------------------------ 1196 // Install the cached top node, and make sure Node::is_top works correctly. 1197 void Compile::set_cached_top_node(Node* tn) { 1198 if (tn != NULL) verify_top(tn); 1199 Node* old_top = _top; 1200 _top = tn; 1201 // Calling Node::setup_is_top allows the nodes the chance to adjust 1202 // their _out arrays. 1203 if (_top != NULL) _top->setup_is_top(); 1204 if (old_top != NULL) old_top->setup_is_top(); 1205 assert(_top == NULL || top()->is_top(), ""); 1206 } 1207 1208 #ifdef ASSERT 1209 uint Compile::count_live_nodes_by_graph_walk() { 1210 Unique_Node_List useful(comp_arena()); 1211 // Get useful node list by walking the graph. 1212 identify_useful_nodes(useful); 1213 return useful.size(); 1214 } 1215 1216 void Compile::print_missing_nodes() { 1217 1218 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1219 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1220 return; 1221 } 1222 1223 // This is an expensive function. It is executed only when the user 1224 // specifies VerifyIdealNodeCount option or otherwise knows the 1225 // additional work that needs to be done to identify reachable nodes 1226 // by walking the flow graph and find the missing ones using 1227 // _dead_node_list. 1228 1229 Unique_Node_List useful(comp_arena()); 1230 // Get useful node list by walking the graph. 1231 identify_useful_nodes(useful); 1232 1233 uint l_nodes = C->live_nodes(); 1234 uint l_nodes_by_walk = useful.size(); 1235 1236 if (l_nodes != l_nodes_by_walk) { 1237 if (_log != NULL) { 1238 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1239 _log->stamp(); 1240 _log->end_head(); 1241 } 1242 VectorSet& useful_member_set = useful.member_set(); 1243 int last_idx = l_nodes_by_walk; 1244 for (int i = 0; i < last_idx; i++) { 1245 if (useful_member_set.test(i)) { 1246 if (_dead_node_list.test(i)) { 1247 if (_log != NULL) { 1248 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1249 } 1250 if (PrintIdealNodeCount) { 1251 // Print the log message to tty 1252 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1253 useful.at(i)->dump(); 1254 } 1255 } 1256 } 1257 else if (! _dead_node_list.test(i)) { 1258 if (_log != NULL) { 1259 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1260 } 1261 if (PrintIdealNodeCount) { 1262 // Print the log message to tty 1263 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1264 } 1265 } 1266 } 1267 if (_log != NULL) { 1268 _log->tail("mismatched_nodes"); 1269 } 1270 } 1271 } 1272 void Compile::record_modified_node(Node* n) { 1273 if (_modified_nodes != NULL && !_inlining_incrementally && 1274 n->outcnt() != 0 && !n->is_Con()) { 1275 _modified_nodes->push(n); 1276 } 1277 } 1278 1279 void Compile::remove_modified_node(Node* n) { 1280 if (_modified_nodes != NULL) { 1281 _modified_nodes->remove(n); 1282 } 1283 } 1284 #endif 1285 1286 #ifndef PRODUCT 1287 void Compile::verify_top(Node* tn) const { 1288 if (tn != NULL) { 1289 assert(tn->is_Con(), "top node must be a constant"); 1290 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1291 assert(tn->in(0) != NULL, "must have live top node"); 1292 } 1293 } 1294 #endif 1295 1296 1297 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1298 1299 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1300 guarantee(arr != NULL, ""); 1301 int num_blocks = arr->length(); 1302 if (grow_by < num_blocks) grow_by = num_blocks; 1303 int num_notes = grow_by * _node_notes_block_size; 1304 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1305 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1306 while (num_notes > 0) { 1307 arr->append(notes); 1308 notes += _node_notes_block_size; 1309 num_notes -= _node_notes_block_size; 1310 } 1311 assert(num_notes == 0, "exact multiple, please"); 1312 } 1313 1314 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1315 if (source == NULL || dest == NULL) return false; 1316 1317 if (dest->is_Con()) 1318 return false; // Do not push debug info onto constants. 1319 1320 #ifdef ASSERT 1321 // Leave a bread crumb trail pointing to the original node: 1322 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1323 dest->set_debug_orig(source); 1324 } 1325 #endif 1326 1327 if (node_note_array() == NULL) 1328 return false; // Not collecting any notes now. 1329 1330 // This is a copy onto a pre-existing node, which may already have notes. 1331 // If both nodes have notes, do not overwrite any pre-existing notes. 1332 Node_Notes* source_notes = node_notes_at(source->_idx); 1333 if (source_notes == NULL || source_notes->is_clear()) return false; 1334 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1335 if (dest_notes == NULL || dest_notes->is_clear()) { 1336 return set_node_notes_at(dest->_idx, source_notes); 1337 } 1338 1339 Node_Notes merged_notes = (*source_notes); 1340 // The order of operations here ensures that dest notes will win... 1341 merged_notes.update_from(dest_notes); 1342 return set_node_notes_at(dest->_idx, &merged_notes); 1343 } 1344 1345 1346 //--------------------------allow_range_check_smearing------------------------- 1347 // Gating condition for coalescing similar range checks. 1348 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1349 // single covering check that is at least as strong as any of them. 1350 // If the optimization succeeds, the simplified (strengthened) range check 1351 // will always succeed. If it fails, we will deopt, and then give up 1352 // on the optimization. 1353 bool Compile::allow_range_check_smearing() const { 1354 // If this method has already thrown a range-check, 1355 // assume it was because we already tried range smearing 1356 // and it failed. 1357 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1358 return !already_trapped; 1359 } 1360 1361 1362 //------------------------------flatten_alias_type----------------------------- 1363 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1364 int offset = tj->offset(); 1365 TypePtr::PTR ptr = tj->ptr(); 1366 1367 // Known instance (scalarizable allocation) alias only with itself. 1368 bool is_known_inst = tj->isa_oopptr() != NULL && 1369 tj->is_oopptr()->is_known_instance(); 1370 1371 // Process weird unsafe references. 1372 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1373 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1374 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1375 tj = TypeOopPtr::BOTTOM; 1376 ptr = tj->ptr(); 1377 offset = tj->offset(); 1378 } 1379 1380 // Array pointers need some flattening 1381 const TypeAryPtr *ta = tj->isa_aryptr(); 1382 if (ta && ta->is_stable()) { 1383 // Erase stability property for alias analysis. 1384 tj = ta = ta->cast_to_stable(false); 1385 } 1386 if( ta && is_known_inst ) { 1387 if ( offset != Type::OffsetBot && 1388 offset > arrayOopDesc::length_offset_in_bytes() ) { 1389 offset = Type::OffsetBot; // Flatten constant access into array body only 1390 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1391 } 1392 } else if( ta && _AliasLevel >= 2 ) { 1393 // For arrays indexed by constant indices, we flatten the alias 1394 // space to include all of the array body. Only the header, klass 1395 // and array length can be accessed un-aliased. 1396 if( offset != Type::OffsetBot ) { 1397 if( ta->const_oop() ) { // MethodData* or Method* 1398 offset = Type::OffsetBot; // Flatten constant access into array body 1399 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1400 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1401 // range is OK as-is. 1402 tj = ta = TypeAryPtr::RANGE; 1403 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1404 tj = TypeInstPtr::KLASS; // all klass loads look alike 1405 ta = TypeAryPtr::RANGE; // generic ignored junk 1406 ptr = TypePtr::BotPTR; 1407 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1408 tj = TypeInstPtr::MARK; 1409 ta = TypeAryPtr::RANGE; // generic ignored junk 1410 ptr = TypePtr::BotPTR; 1411 } else { // Random constant offset into array body 1412 offset = Type::OffsetBot; // Flatten constant access into array body 1413 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1414 } 1415 } 1416 // Arrays of fixed size alias with arrays of unknown size. 1417 if (ta->size() != TypeInt::POS) { 1418 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1419 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1420 } 1421 // Arrays of known objects become arrays of unknown objects. 1422 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1423 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1424 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1425 } 1426 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1427 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1428 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1429 } 1430 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1431 // cannot be distinguished by bytecode alone. 1432 if (ta->elem() == TypeInt::BOOL) { 1433 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1434 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1435 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1436 } 1437 // During the 2nd round of IterGVN, NotNull castings are removed. 1438 // Make sure the Bottom and NotNull variants alias the same. 1439 // Also, make sure exact and non-exact variants alias the same. 1440 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1441 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1442 } 1443 } 1444 1445 // Oop pointers need some flattening 1446 const TypeInstPtr *to = tj->isa_instptr(); 1447 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1448 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1449 if( ptr == TypePtr::Constant ) { 1450 if (to->klass() != ciEnv::current()->Class_klass() || 1451 offset < k->size_helper() * wordSize) { 1452 // No constant oop pointers (such as Strings); they alias with 1453 // unknown strings. 1454 assert(!is_known_inst, "not scalarizable allocation"); 1455 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1456 } 1457 } else if( is_known_inst ) { 1458 tj = to; // Keep NotNull and klass_is_exact for instance type 1459 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1460 // During the 2nd round of IterGVN, NotNull castings are removed. 1461 // Make sure the Bottom and NotNull variants alias the same. 1462 // Also, make sure exact and non-exact variants alias the same. 1463 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1464 } 1465 if (to->speculative() != NULL) { 1466 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1467 } 1468 // Canonicalize the holder of this field 1469 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1470 // First handle header references such as a LoadKlassNode, even if the 1471 // object's klass is unloaded at compile time (4965979). 1472 if (!is_known_inst) { // Do it only for non-instance types 1473 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1474 } 1475 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1476 // Static fields are in the space above the normal instance 1477 // fields in the java.lang.Class instance. 1478 if (to->klass() != ciEnv::current()->Class_klass()) { 1479 to = NULL; 1480 tj = TypeOopPtr::BOTTOM; 1481 offset = tj->offset(); 1482 } 1483 } else { 1484 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1485 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1486 if( is_known_inst ) { 1487 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1488 } else { 1489 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1490 } 1491 } 1492 } 1493 } 1494 1495 // Klass pointers to object array klasses need some flattening 1496 const TypeKlassPtr *tk = tj->isa_klassptr(); 1497 if( tk ) { 1498 // If we are referencing a field within a Klass, we need 1499 // to assume the worst case of an Object. Both exact and 1500 // inexact types must flatten to the same alias class so 1501 // use NotNull as the PTR. 1502 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1503 1504 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1505 TypeKlassPtr::OBJECT->klass(), 1506 offset); 1507 } 1508 1509 ciKlass* klass = tk->klass(); 1510 if( klass->is_obj_array_klass() ) { 1511 ciKlass* k = TypeAryPtr::OOPS->klass(); 1512 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1513 k = TypeInstPtr::BOTTOM->klass(); 1514 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1515 } 1516 1517 // Check for precise loads from the primary supertype array and force them 1518 // to the supertype cache alias index. Check for generic array loads from 1519 // the primary supertype array and also force them to the supertype cache 1520 // alias index. Since the same load can reach both, we need to merge 1521 // these 2 disparate memories into the same alias class. Since the 1522 // primary supertype array is read-only, there's no chance of confusion 1523 // where we bypass an array load and an array store. 1524 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1525 if (offset == Type::OffsetBot || 1526 (offset >= primary_supers_offset && 1527 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1528 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1529 offset = in_bytes(Klass::secondary_super_cache_offset()); 1530 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1531 } 1532 } 1533 1534 // Flatten all Raw pointers together. 1535 if (tj->base() == Type::RawPtr) 1536 tj = TypeRawPtr::BOTTOM; 1537 1538 if (tj->base() == Type::AnyPtr) 1539 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1540 1541 // Flatten all to bottom for now 1542 switch( _AliasLevel ) { 1543 case 0: 1544 tj = TypePtr::BOTTOM; 1545 break; 1546 case 1: // Flatten to: oop, static, field or array 1547 switch (tj->base()) { 1548 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1549 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1550 case Type::AryPtr: // do not distinguish arrays at all 1551 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1552 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1553 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1554 default: ShouldNotReachHere(); 1555 } 1556 break; 1557 case 2: // No collapsing at level 2; keep all splits 1558 case 3: // No collapsing at level 3; keep all splits 1559 break; 1560 default: 1561 Unimplemented(); 1562 } 1563 1564 offset = tj->offset(); 1565 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1566 1567 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1568 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1569 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1570 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1571 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1572 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1573 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1574 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1575 assert( tj->ptr() != TypePtr::TopPTR && 1576 tj->ptr() != TypePtr::AnyNull && 1577 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1578 // assert( tj->ptr() != TypePtr::Constant || 1579 // tj->base() == Type::RawPtr || 1580 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1581 1582 return tj; 1583 } 1584 1585 void Compile::AliasType::Init(int i, const TypePtr* at) { 1586 _index = i; 1587 _adr_type = at; 1588 _field = NULL; 1589 _element = NULL; 1590 _is_rewritable = true; // default 1591 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1592 if (atoop != NULL && atoop->is_known_instance()) { 1593 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1594 _general_index = Compile::current()->get_alias_index(gt); 1595 } else { 1596 _general_index = 0; 1597 } 1598 } 1599 1600 //---------------------------------print_on------------------------------------ 1601 #ifndef PRODUCT 1602 void Compile::AliasType::print_on(outputStream* st) { 1603 if (index() < 10) 1604 st->print("@ <%d> ", index()); 1605 else st->print("@ <%d>", index()); 1606 st->print(is_rewritable() ? " " : " RO"); 1607 int offset = adr_type()->offset(); 1608 if (offset == Type::OffsetBot) 1609 st->print(" +any"); 1610 else st->print(" +%-3d", offset); 1611 st->print(" in "); 1612 adr_type()->dump_on(st); 1613 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1614 if (field() != NULL && tjp) { 1615 if (tjp->klass() != field()->holder() || 1616 tjp->offset() != field()->offset_in_bytes()) { 1617 st->print(" != "); 1618 field()->print(); 1619 st->print(" ***"); 1620 } 1621 } 1622 } 1623 1624 void print_alias_types() { 1625 Compile* C = Compile::current(); 1626 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1627 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1628 C->alias_type(idx)->print_on(tty); 1629 tty->cr(); 1630 } 1631 } 1632 #endif 1633 1634 1635 //----------------------------probe_alias_cache-------------------------------- 1636 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1637 intptr_t key = (intptr_t) adr_type; 1638 key ^= key >> logAliasCacheSize; 1639 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1640 } 1641 1642 1643 //-----------------------------grow_alias_types-------------------------------- 1644 void Compile::grow_alias_types() { 1645 const int old_ats = _max_alias_types; // how many before? 1646 const int new_ats = old_ats; // how many more? 1647 const int grow_ats = old_ats+new_ats; // how many now? 1648 _max_alias_types = grow_ats; 1649 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1650 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1651 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1652 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1653 } 1654 1655 1656 //--------------------------------find_alias_type------------------------------ 1657 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1658 if (_AliasLevel == 0) 1659 return alias_type(AliasIdxBot); 1660 1661 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1662 if (ace->_adr_type == adr_type) { 1663 return alias_type(ace->_index); 1664 } 1665 1666 // Handle special cases. 1667 if (adr_type == NULL) return alias_type(AliasIdxTop); 1668 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1669 1670 // Do it the slow way. 1671 const TypePtr* flat = flatten_alias_type(adr_type); 1672 1673 #ifdef ASSERT 1674 assert(flat == flatten_alias_type(flat), "idempotent"); 1675 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1676 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1677 const TypeOopPtr* foop = flat->is_oopptr(); 1678 // Scalarizable allocations have exact klass always. 1679 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1680 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1681 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1682 } 1683 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1684 #endif 1685 1686 int idx = AliasIdxTop; 1687 for (int i = 0; i < num_alias_types(); i++) { 1688 if (alias_type(i)->adr_type() == flat) { 1689 idx = i; 1690 break; 1691 } 1692 } 1693 1694 if (idx == AliasIdxTop) { 1695 if (no_create) return NULL; 1696 // Grow the array if necessary. 1697 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1698 // Add a new alias type. 1699 idx = _num_alias_types++; 1700 _alias_types[idx]->Init(idx, flat); 1701 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1702 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1703 if (flat->isa_instptr()) { 1704 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1705 && flat->is_instptr()->klass() == env()->Class_klass()) 1706 alias_type(idx)->set_rewritable(false); 1707 } 1708 if (flat->isa_aryptr()) { 1709 #ifdef ASSERT 1710 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1711 // (T_BYTE has the weakest alignment and size restrictions...) 1712 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1713 #endif 1714 if (flat->offset() == TypePtr::OffsetBot) { 1715 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1716 } 1717 } 1718 if (flat->isa_klassptr()) { 1719 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1720 alias_type(idx)->set_rewritable(false); 1721 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1722 alias_type(idx)->set_rewritable(false); 1723 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1724 alias_type(idx)->set_rewritable(false); 1725 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1726 alias_type(idx)->set_rewritable(false); 1727 } 1728 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1729 // but the base pointer type is not distinctive enough to identify 1730 // references into JavaThread.) 1731 1732 // Check for final fields. 1733 const TypeInstPtr* tinst = flat->isa_instptr(); 1734 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1735 ciField* field; 1736 if (tinst->const_oop() != NULL && 1737 tinst->klass() == ciEnv::current()->Class_klass() && 1738 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1739 // static field 1740 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1741 field = k->get_field_by_offset(tinst->offset(), true); 1742 } else { 1743 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1744 field = k->get_field_by_offset(tinst->offset(), false); 1745 } 1746 assert(field == NULL || 1747 original_field == NULL || 1748 (field->holder() == original_field->holder() && 1749 field->offset() == original_field->offset() && 1750 field->is_static() == original_field->is_static()), "wrong field?"); 1751 // Set field() and is_rewritable() attributes. 1752 if (field != NULL) alias_type(idx)->set_field(field); 1753 } 1754 } 1755 1756 // Fill the cache for next time. 1757 ace->_adr_type = adr_type; 1758 ace->_index = idx; 1759 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1760 1761 // Might as well try to fill the cache for the flattened version, too. 1762 AliasCacheEntry* face = probe_alias_cache(flat); 1763 if (face->_adr_type == NULL) { 1764 face->_adr_type = flat; 1765 face->_index = idx; 1766 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1767 } 1768 1769 return alias_type(idx); 1770 } 1771 1772 1773 Compile::AliasType* Compile::alias_type(ciField* field) { 1774 const TypeOopPtr* t; 1775 if (field->is_static()) 1776 t = TypeInstPtr::make(field->holder()->java_mirror()); 1777 else 1778 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1779 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1780 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1781 return atp; 1782 } 1783 1784 1785 //------------------------------have_alias_type-------------------------------- 1786 bool Compile::have_alias_type(const TypePtr* adr_type) { 1787 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1788 if (ace->_adr_type == adr_type) { 1789 return true; 1790 } 1791 1792 // Handle special cases. 1793 if (adr_type == NULL) return true; 1794 if (adr_type == TypePtr::BOTTOM) return true; 1795 1796 return find_alias_type(adr_type, true, NULL) != NULL; 1797 } 1798 1799 //-----------------------------must_alias-------------------------------------- 1800 // True if all values of the given address type are in the given alias category. 1801 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1802 if (alias_idx == AliasIdxBot) return true; // the universal category 1803 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1804 if (alias_idx == AliasIdxTop) return false; // the empty category 1805 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1806 1807 // the only remaining possible overlap is identity 1808 int adr_idx = get_alias_index(adr_type); 1809 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1810 assert(adr_idx == alias_idx || 1811 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1812 && adr_type != TypeOopPtr::BOTTOM), 1813 "should not be testing for overlap with an unsafe pointer"); 1814 return adr_idx == alias_idx; 1815 } 1816 1817 //------------------------------can_alias-------------------------------------- 1818 // True if any values of the given address type are in the given alias category. 1819 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1820 if (alias_idx == AliasIdxTop) return false; // the empty category 1821 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1822 if (alias_idx == AliasIdxBot) return true; // the universal category 1823 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1824 1825 // the only remaining possible overlap is identity 1826 int adr_idx = get_alias_index(adr_type); 1827 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1828 return adr_idx == alias_idx; 1829 } 1830 1831 1832 1833 //---------------------------pop_warm_call------------------------------------- 1834 WarmCallInfo* Compile::pop_warm_call() { 1835 WarmCallInfo* wci = _warm_calls; 1836 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1837 return wci; 1838 } 1839 1840 //----------------------------Inline_Warm-------------------------------------- 1841 int Compile::Inline_Warm() { 1842 // If there is room, try to inline some more warm call sites. 1843 // %%% Do a graph index compaction pass when we think we're out of space? 1844 if (!InlineWarmCalls) return 0; 1845 1846 int calls_made_hot = 0; 1847 int room_to_grow = NodeCountInliningCutoff - unique(); 1848 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1849 int amount_grown = 0; 1850 WarmCallInfo* call; 1851 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1852 int est_size = (int)call->size(); 1853 if (est_size > (room_to_grow - amount_grown)) { 1854 // This one won't fit anyway. Get rid of it. 1855 call->make_cold(); 1856 continue; 1857 } 1858 call->make_hot(); 1859 calls_made_hot++; 1860 amount_grown += est_size; 1861 amount_to_grow -= est_size; 1862 } 1863 1864 if (calls_made_hot > 0) set_major_progress(); 1865 return calls_made_hot; 1866 } 1867 1868 1869 //----------------------------Finish_Warm-------------------------------------- 1870 void Compile::Finish_Warm() { 1871 if (!InlineWarmCalls) return; 1872 if (failing()) return; 1873 if (warm_calls() == NULL) return; 1874 1875 // Clean up loose ends, if we are out of space for inlining. 1876 WarmCallInfo* call; 1877 while ((call = pop_warm_call()) != NULL) { 1878 call->make_cold(); 1879 } 1880 } 1881 1882 //---------------------cleanup_loop_predicates----------------------- 1883 // Remove the opaque nodes that protect the predicates so that all unused 1884 // checks and uncommon_traps will be eliminated from the ideal graph 1885 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1886 if (predicate_count()==0) return; 1887 for (int i = predicate_count(); i > 0; i--) { 1888 Node * n = predicate_opaque1_node(i-1); 1889 assert(n->Opcode() == Op_Opaque1, "must be"); 1890 igvn.replace_node(n, n->in(1)); 1891 } 1892 assert(predicate_count()==0, "should be clean!"); 1893 } 1894 1895 // StringOpts and late inlining of string methods 1896 void Compile::inline_string_calls(bool parse_time) { 1897 { 1898 // remove useless nodes to make the usage analysis simpler 1899 ResourceMark rm; 1900 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1901 } 1902 1903 { 1904 ResourceMark rm; 1905 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1906 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1907 print_method(PHASE_AFTER_STRINGOPTS, 3); 1908 } 1909 1910 // now inline anything that we skipped the first time around 1911 if (!parse_time) { 1912 _late_inlines_pos = _late_inlines.length(); 1913 } 1914 1915 while (_string_late_inlines.length() > 0) { 1916 CallGenerator* cg = _string_late_inlines.pop(); 1917 cg->do_late_inline(); 1918 if (failing()) return; 1919 } 1920 _string_late_inlines.trunc_to(0); 1921 } 1922 1923 // Late inlining of boxing methods 1924 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1925 if (_boxing_late_inlines.length() > 0) { 1926 assert(has_boxed_value(), "inconsistent"); 1927 1928 PhaseGVN* gvn = initial_gvn(); 1929 set_inlining_incrementally(true); 1930 1931 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1932 for_igvn()->clear(); 1933 gvn->replace_with(&igvn); 1934 1935 _late_inlines_pos = _late_inlines.length(); 1936 1937 while (_boxing_late_inlines.length() > 0) { 1938 CallGenerator* cg = _boxing_late_inlines.pop(); 1939 cg->do_late_inline(); 1940 if (failing()) return; 1941 } 1942 _boxing_late_inlines.trunc_to(0); 1943 1944 { 1945 ResourceMark rm; 1946 PhaseRemoveUseless pru(gvn, for_igvn()); 1947 } 1948 1949 igvn = PhaseIterGVN(gvn); 1950 igvn.optimize(); 1951 1952 set_inlining_progress(false); 1953 set_inlining_incrementally(false); 1954 } 1955 } 1956 1957 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1958 assert(IncrementalInline, "incremental inlining should be on"); 1959 PhaseGVN* gvn = initial_gvn(); 1960 1961 set_inlining_progress(false); 1962 for_igvn()->clear(); 1963 gvn->replace_with(&igvn); 1964 1965 { 1966 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 1967 int i = 0; 1968 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1969 CallGenerator* cg = _late_inlines.at(i); 1970 _late_inlines_pos = i+1; 1971 cg->do_late_inline(); 1972 if (failing()) return; 1973 } 1974 int j = 0; 1975 for (; i < _late_inlines.length(); i++, j++) { 1976 _late_inlines.at_put(j, _late_inlines.at(i)); 1977 } 1978 _late_inlines.trunc_to(j); 1979 } 1980 1981 { 1982 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 1983 ResourceMark rm; 1984 PhaseRemoveUseless pru(gvn, for_igvn()); 1985 } 1986 1987 { 1988 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 1989 igvn = PhaseIterGVN(gvn); 1990 } 1991 } 1992 1993 // Perform incremental inlining until bound on number of live nodes is reached 1994 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1995 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 1996 1997 PhaseGVN* gvn = initial_gvn(); 1998 1999 set_inlining_incrementally(true); 2000 set_inlining_progress(true); 2001 uint low_live_nodes = 0; 2002 2003 while(inlining_progress() && _late_inlines.length() > 0) { 2004 2005 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2006 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2007 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2008 // PhaseIdealLoop is expensive so we only try it once we are 2009 // out of live nodes and we only try it again if the previous 2010 // helped got the number of nodes down significantly 2011 PhaseIdealLoop ideal_loop( igvn, false, true ); 2012 if (failing()) return; 2013 low_live_nodes = live_nodes(); 2014 _major_progress = true; 2015 } 2016 2017 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2018 break; 2019 } 2020 } 2021 2022 inline_incrementally_one(igvn); 2023 2024 if (failing()) return; 2025 2026 { 2027 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2028 igvn.optimize(); 2029 } 2030 2031 if (failing()) return; 2032 } 2033 2034 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2035 2036 if (_string_late_inlines.length() > 0) { 2037 assert(has_stringbuilder(), "inconsistent"); 2038 for_igvn()->clear(); 2039 initial_gvn()->replace_with(&igvn); 2040 2041 inline_string_calls(false); 2042 2043 if (failing()) return; 2044 2045 { 2046 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2047 ResourceMark rm; 2048 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2049 } 2050 2051 { 2052 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2053 igvn = PhaseIterGVN(gvn); 2054 igvn.optimize(); 2055 } 2056 } 2057 2058 set_inlining_incrementally(false); 2059 } 2060 2061 2062 //------------------------------Optimize--------------------------------------- 2063 // Given a graph, optimize it. 2064 void Compile::Optimize() { 2065 TracePhase tp("optimizer", &timers[_t_optimizer]); 2066 2067 #ifndef PRODUCT 2068 if (env()->break_at_compile()) { 2069 BREAKPOINT; 2070 } 2071 2072 #endif 2073 2074 ResourceMark rm; 2075 int loop_opts_cnt; 2076 2077 print_inlining_reinit(); 2078 2079 NOT_PRODUCT( verify_graph_edges(); ) 2080 2081 print_method(PHASE_AFTER_PARSING); 2082 2083 { 2084 // Iterative Global Value Numbering, including ideal transforms 2085 // Initialize IterGVN with types and values from parse-time GVN 2086 PhaseIterGVN igvn(initial_gvn()); 2087 #ifdef ASSERT 2088 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2089 #endif 2090 { 2091 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2092 igvn.optimize(); 2093 } 2094 2095 print_method(PHASE_ITER_GVN1, 2); 2096 2097 if (failing()) return; 2098 2099 inline_incrementally(igvn); 2100 2101 print_method(PHASE_INCREMENTAL_INLINE, 2); 2102 2103 if (failing()) return; 2104 2105 if (eliminate_boxing()) { 2106 // Inline valueOf() methods now. 2107 inline_boxing_calls(igvn); 2108 2109 if (AlwaysIncrementalInline) { 2110 inline_incrementally(igvn); 2111 } 2112 2113 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2114 2115 if (failing()) return; 2116 } 2117 2118 // Remove the speculative part of types and clean up the graph from 2119 // the extra CastPP nodes whose only purpose is to carry them. Do 2120 // that early so that optimizations are not disrupted by the extra 2121 // CastPP nodes. 2122 remove_speculative_types(igvn); 2123 2124 // No more new expensive nodes will be added to the list from here 2125 // so keep only the actual candidates for optimizations. 2126 cleanup_expensive_nodes(igvn); 2127 2128 // Perform escape analysis 2129 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2130 if (has_loops()) { 2131 // Cleanup graph (remove dead nodes). 2132 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2133 PhaseIdealLoop ideal_loop( igvn, false, true ); 2134 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2135 if (failing()) return; 2136 } 2137 ConnectionGraph::do_analysis(this, &igvn); 2138 2139 if (failing()) return; 2140 2141 // Optimize out fields loads from scalar replaceable allocations. 2142 igvn.optimize(); 2143 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2144 2145 if (failing()) return; 2146 2147 if (congraph() != NULL && macro_count() > 0) { 2148 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2149 PhaseMacroExpand mexp(igvn); 2150 mexp.eliminate_macro_nodes(); 2151 igvn.set_delay_transform(false); 2152 2153 igvn.optimize(); 2154 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2155 2156 if (failing()) return; 2157 } 2158 } 2159 2160 // Loop transforms on the ideal graph. Range Check Elimination, 2161 // peeling, unrolling, etc. 2162 2163 // Set loop opts counter 2164 loop_opts_cnt = num_loop_opts(); 2165 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2166 { 2167 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2168 PhaseIdealLoop ideal_loop( igvn, true ); 2169 loop_opts_cnt--; 2170 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2171 if (failing()) return; 2172 } 2173 // Loop opts pass if partial peeling occurred in previous pass 2174 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2175 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2176 PhaseIdealLoop ideal_loop( igvn, false ); 2177 loop_opts_cnt--; 2178 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2179 if (failing()) return; 2180 } 2181 // Loop opts pass for loop-unrolling before CCP 2182 if(major_progress() && (loop_opts_cnt > 0)) { 2183 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2184 PhaseIdealLoop ideal_loop( igvn, false ); 2185 loop_opts_cnt--; 2186 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2187 } 2188 if (!failing()) { 2189 // Verify that last round of loop opts produced a valid graph 2190 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2191 PhaseIdealLoop::verify(igvn); 2192 } 2193 } 2194 if (failing()) return; 2195 2196 // Conditional Constant Propagation; 2197 PhaseCCP ccp( &igvn ); 2198 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2199 { 2200 TracePhase tp("ccp", &timers[_t_ccp]); 2201 ccp.do_transform(); 2202 } 2203 print_method(PHASE_CPP1, 2); 2204 2205 assert( true, "Break here to ccp.dump_old2new_map()"); 2206 2207 // Iterative Global Value Numbering, including ideal transforms 2208 { 2209 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2210 igvn = ccp; 2211 igvn.optimize(); 2212 } 2213 2214 print_method(PHASE_ITER_GVN2, 2); 2215 2216 if (failing()) return; 2217 2218 // Loop transforms on the ideal graph. Range Check Elimination, 2219 // peeling, unrolling, etc. 2220 if(loop_opts_cnt > 0) { 2221 debug_only( int cnt = 0; ); 2222 while(major_progress() && (loop_opts_cnt > 0)) { 2223 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2224 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2225 PhaseIdealLoop ideal_loop( igvn, true); 2226 loop_opts_cnt--; 2227 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2228 if (failing()) return; 2229 } 2230 } 2231 2232 { 2233 // Verify that all previous optimizations produced a valid graph 2234 // at least to this point, even if no loop optimizations were done. 2235 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2236 PhaseIdealLoop::verify(igvn); 2237 } 2238 2239 { 2240 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2241 PhaseMacroExpand mex(igvn); 2242 if (mex.expand_macro_nodes()) { 2243 assert(failing(), "must bail out w/ explicit message"); 2244 return; 2245 } 2246 } 2247 2248 DEBUG_ONLY( _modified_nodes = NULL; ) 2249 } // (End scope of igvn; run destructor if necessary for asserts.) 2250 2251 process_print_inlining(); 2252 // A method with only infinite loops has no edges entering loops from root 2253 { 2254 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2255 if (final_graph_reshaping()) { 2256 assert(failing(), "must bail out w/ explicit message"); 2257 return; 2258 } 2259 } 2260 2261 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2262 } 2263 2264 2265 //------------------------------Code_Gen--------------------------------------- 2266 // Given a graph, generate code for it 2267 void Compile::Code_Gen() { 2268 if (failing()) { 2269 return; 2270 } 2271 2272 // Perform instruction selection. You might think we could reclaim Matcher 2273 // memory PDQ, but actually the Matcher is used in generating spill code. 2274 // Internals of the Matcher (including some VectorSets) must remain live 2275 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2276 // set a bit in reclaimed memory. 2277 2278 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2279 // nodes. Mapping is only valid at the root of each matched subtree. 2280 NOT_PRODUCT( verify_graph_edges(); ) 2281 2282 Matcher matcher; 2283 _matcher = &matcher; 2284 { 2285 TracePhase tp("matcher", &timers[_t_matcher]); 2286 matcher.match(); 2287 } 2288 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2289 // nodes. Mapping is only valid at the root of each matched subtree. 2290 NOT_PRODUCT( verify_graph_edges(); ) 2291 2292 // If you have too many nodes, or if matching has failed, bail out 2293 check_node_count(0, "out of nodes matching instructions"); 2294 if (failing()) { 2295 return; 2296 } 2297 2298 // Build a proper-looking CFG 2299 PhaseCFG cfg(node_arena(), root(), matcher); 2300 _cfg = &cfg; 2301 { 2302 TracePhase tp("scheduler", &timers[_t_scheduler]); 2303 bool success = cfg.do_global_code_motion(); 2304 if (!success) { 2305 return; 2306 } 2307 2308 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2309 NOT_PRODUCT( verify_graph_edges(); ) 2310 debug_only( cfg.verify(); ) 2311 } 2312 2313 PhaseChaitin regalloc(unique(), cfg, matcher); 2314 _regalloc = ®alloc; 2315 { 2316 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2317 // Perform register allocation. After Chaitin, use-def chains are 2318 // no longer accurate (at spill code) and so must be ignored. 2319 // Node->LRG->reg mappings are still accurate. 2320 _regalloc->Register_Allocate(); 2321 2322 // Bail out if the allocator builds too many nodes 2323 if (failing()) { 2324 return; 2325 } 2326 } 2327 2328 // Prior to register allocation we kept empty basic blocks in case the 2329 // the allocator needed a place to spill. After register allocation we 2330 // are not adding any new instructions. If any basic block is empty, we 2331 // can now safely remove it. 2332 { 2333 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2334 cfg.remove_empty_blocks(); 2335 if (do_freq_based_layout()) { 2336 PhaseBlockLayout layout(cfg); 2337 } else { 2338 cfg.set_loop_alignment(); 2339 } 2340 cfg.fixup_flow(); 2341 } 2342 2343 // Apply peephole optimizations 2344 if( OptoPeephole ) { 2345 TracePhase tp("peephole", &timers[_t_peephole]); 2346 PhasePeephole peep( _regalloc, cfg); 2347 peep.do_transform(); 2348 } 2349 2350 // Do late expand if CPU requires this. 2351 if (Matcher::require_postalloc_expand) { 2352 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2353 cfg.postalloc_expand(_regalloc); 2354 } 2355 2356 // Convert Nodes to instruction bits in a buffer 2357 { 2358 TraceTime tp("output", &timers[_t_output], CITime); 2359 Output(); 2360 } 2361 2362 print_method(PHASE_FINAL_CODE); 2363 2364 // He's dead, Jim. 2365 _cfg = (PhaseCFG*)0xdeadbeef; 2366 _regalloc = (PhaseChaitin*)0xdeadbeef; 2367 } 2368 2369 2370 //------------------------------dump_asm--------------------------------------- 2371 // Dump formatted assembly 2372 #ifndef PRODUCT 2373 void Compile::dump_asm(int *pcs, uint pc_limit) { 2374 bool cut_short = false; 2375 tty->print_cr("#"); 2376 tty->print("# "); _tf->dump(); tty->cr(); 2377 tty->print_cr("#"); 2378 2379 // For all blocks 2380 int pc = 0x0; // Program counter 2381 char starts_bundle = ' '; 2382 _regalloc->dump_frame(); 2383 2384 Node *n = NULL; 2385 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2386 if (VMThread::should_terminate()) { 2387 cut_short = true; 2388 break; 2389 } 2390 Block* block = _cfg->get_block(i); 2391 if (block->is_connector() && !Verbose) { 2392 continue; 2393 } 2394 n = block->head(); 2395 if (pcs && n->_idx < pc_limit) { 2396 tty->print("%3.3x ", pcs[n->_idx]); 2397 } else { 2398 tty->print(" "); 2399 } 2400 block->dump_head(_cfg); 2401 if (block->is_connector()) { 2402 tty->print_cr(" # Empty connector block"); 2403 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2404 tty->print_cr(" # Block is sole successor of call"); 2405 } 2406 2407 // For all instructions 2408 Node *delay = NULL; 2409 for (uint j = 0; j < block->number_of_nodes(); j++) { 2410 if (VMThread::should_terminate()) { 2411 cut_short = true; 2412 break; 2413 } 2414 n = block->get_node(j); 2415 if (valid_bundle_info(n)) { 2416 Bundle* bundle = node_bundling(n); 2417 if (bundle->used_in_unconditional_delay()) { 2418 delay = n; 2419 continue; 2420 } 2421 if (bundle->starts_bundle()) { 2422 starts_bundle = '+'; 2423 } 2424 } 2425 2426 if (WizardMode) { 2427 n->dump(); 2428 } 2429 2430 if( !n->is_Region() && // Dont print in the Assembly 2431 !n->is_Phi() && // a few noisely useless nodes 2432 !n->is_Proj() && 2433 !n->is_MachTemp() && 2434 !n->is_SafePointScalarObject() && 2435 !n->is_Catch() && // Would be nice to print exception table targets 2436 !n->is_MergeMem() && // Not very interesting 2437 !n->is_top() && // Debug info table constants 2438 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2439 ) { 2440 if (pcs && n->_idx < pc_limit) 2441 tty->print("%3.3x", pcs[n->_idx]); 2442 else 2443 tty->print(" "); 2444 tty->print(" %c ", starts_bundle); 2445 starts_bundle = ' '; 2446 tty->print("\t"); 2447 n->format(_regalloc, tty); 2448 tty->cr(); 2449 } 2450 2451 // If we have an instruction with a delay slot, and have seen a delay, 2452 // then back up and print it 2453 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2454 assert(delay != NULL, "no unconditional delay instruction"); 2455 if (WizardMode) delay->dump(); 2456 2457 if (node_bundling(delay)->starts_bundle()) 2458 starts_bundle = '+'; 2459 if (pcs && n->_idx < pc_limit) 2460 tty->print("%3.3x", pcs[n->_idx]); 2461 else 2462 tty->print(" "); 2463 tty->print(" %c ", starts_bundle); 2464 starts_bundle = ' '; 2465 tty->print("\t"); 2466 delay->format(_regalloc, tty); 2467 tty->cr(); 2468 delay = NULL; 2469 } 2470 2471 // Dump the exception table as well 2472 if( n->is_Catch() && (Verbose || WizardMode) ) { 2473 // Print the exception table for this offset 2474 _handler_table.print_subtable_for(pc); 2475 } 2476 } 2477 2478 if (pcs && n->_idx < pc_limit) 2479 tty->print_cr("%3.3x", pcs[n->_idx]); 2480 else 2481 tty->cr(); 2482 2483 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2484 2485 } // End of per-block dump 2486 tty->cr(); 2487 2488 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2489 } 2490 #endif 2491 2492 //------------------------------Final_Reshape_Counts--------------------------- 2493 // This class defines counters to help identify when a method 2494 // may/must be executed using hardware with only 24-bit precision. 2495 struct Final_Reshape_Counts : public StackObj { 2496 int _call_count; // count non-inlined 'common' calls 2497 int _float_count; // count float ops requiring 24-bit precision 2498 int _double_count; // count double ops requiring more precision 2499 int _java_call_count; // count non-inlined 'java' calls 2500 int _inner_loop_count; // count loops which need alignment 2501 VectorSet _visited; // Visitation flags 2502 Node_List _tests; // Set of IfNodes & PCTableNodes 2503 2504 Final_Reshape_Counts() : 2505 _call_count(0), _float_count(0), _double_count(0), 2506 _java_call_count(0), _inner_loop_count(0), 2507 _visited( Thread::current()->resource_area() ) { } 2508 2509 void inc_call_count () { _call_count ++; } 2510 void inc_float_count () { _float_count ++; } 2511 void inc_double_count() { _double_count++; } 2512 void inc_java_call_count() { _java_call_count++; } 2513 void inc_inner_loop_count() { _inner_loop_count++; } 2514 2515 int get_call_count () const { return _call_count ; } 2516 int get_float_count () const { return _float_count ; } 2517 int get_double_count() const { return _double_count; } 2518 int get_java_call_count() const { return _java_call_count; } 2519 int get_inner_loop_count() const { return _inner_loop_count; } 2520 }; 2521 2522 #ifdef ASSERT 2523 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2524 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2525 // Make sure the offset goes inside the instance layout. 2526 return k->contains_field_offset(tp->offset()); 2527 // Note that OffsetBot and OffsetTop are very negative. 2528 } 2529 #endif 2530 2531 // Eliminate trivially redundant StoreCMs and accumulate their 2532 // precedence edges. 2533 void Compile::eliminate_redundant_card_marks(Node* n) { 2534 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2535 if (n->in(MemNode::Address)->outcnt() > 1) { 2536 // There are multiple users of the same address so it might be 2537 // possible to eliminate some of the StoreCMs 2538 Node* mem = n->in(MemNode::Memory); 2539 Node* adr = n->in(MemNode::Address); 2540 Node* val = n->in(MemNode::ValueIn); 2541 Node* prev = n; 2542 bool done = false; 2543 // Walk the chain of StoreCMs eliminating ones that match. As 2544 // long as it's a chain of single users then the optimization is 2545 // safe. Eliminating partially redundant StoreCMs would require 2546 // cloning copies down the other paths. 2547 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2548 if (adr == mem->in(MemNode::Address) && 2549 val == mem->in(MemNode::ValueIn)) { 2550 // redundant StoreCM 2551 if (mem->req() > MemNode::OopStore) { 2552 // Hasn't been processed by this code yet. 2553 n->add_prec(mem->in(MemNode::OopStore)); 2554 } else { 2555 // Already converted to precedence edge 2556 for (uint i = mem->req(); i < mem->len(); i++) { 2557 // Accumulate any precedence edges 2558 if (mem->in(i) != NULL) { 2559 n->add_prec(mem->in(i)); 2560 } 2561 } 2562 // Everything above this point has been processed. 2563 done = true; 2564 } 2565 // Eliminate the previous StoreCM 2566 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2567 assert(mem->outcnt() == 0, "should be dead"); 2568 mem->disconnect_inputs(NULL, this); 2569 } else { 2570 prev = mem; 2571 } 2572 mem = prev->in(MemNode::Memory); 2573 } 2574 } 2575 } 2576 2577 //------------------------------final_graph_reshaping_impl---------------------- 2578 // Implement items 1-5 from final_graph_reshaping below. 2579 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2580 2581 if ( n->outcnt() == 0 ) return; // dead node 2582 uint nop = n->Opcode(); 2583 2584 // Check for 2-input instruction with "last use" on right input. 2585 // Swap to left input. Implements item (2). 2586 if( n->req() == 3 && // two-input instruction 2587 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2588 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2589 n->in(2)->outcnt() == 1 &&// right use IS a last use 2590 !n->in(2)->is_Con() ) { // right use is not a constant 2591 // Check for commutative opcode 2592 switch( nop ) { 2593 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2594 case Op_MaxI: case Op_MinI: 2595 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2596 case Op_AndL: case Op_XorL: case Op_OrL: 2597 case Op_AndI: case Op_XorI: case Op_OrI: { 2598 // Move "last use" input to left by swapping inputs 2599 n->swap_edges(1, 2); 2600 break; 2601 } 2602 default: 2603 break; 2604 } 2605 } 2606 2607 #ifdef ASSERT 2608 if( n->is_Mem() ) { 2609 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2610 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2611 // oop will be recorded in oop map if load crosses safepoint 2612 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2613 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2614 "raw memory operations should have control edge"); 2615 } 2616 #endif 2617 // Count FPU ops and common calls, implements item (3) 2618 switch( nop ) { 2619 // Count all float operations that may use FPU 2620 case Op_AddF: 2621 case Op_SubF: 2622 case Op_MulF: 2623 case Op_DivF: 2624 case Op_NegF: 2625 case Op_ModF: 2626 case Op_ConvI2F: 2627 case Op_ConF: 2628 case Op_CmpF: 2629 case Op_CmpF3: 2630 // case Op_ConvL2F: // longs are split into 32-bit halves 2631 frc.inc_float_count(); 2632 break; 2633 2634 case Op_ConvF2D: 2635 case Op_ConvD2F: 2636 frc.inc_float_count(); 2637 frc.inc_double_count(); 2638 break; 2639 2640 // Count all double operations that may use FPU 2641 case Op_AddD: 2642 case Op_SubD: 2643 case Op_MulD: 2644 case Op_DivD: 2645 case Op_NegD: 2646 case Op_ModD: 2647 case Op_ConvI2D: 2648 case Op_ConvD2I: 2649 // case Op_ConvL2D: // handled by leaf call 2650 // case Op_ConvD2L: // handled by leaf call 2651 case Op_ConD: 2652 case Op_CmpD: 2653 case Op_CmpD3: 2654 frc.inc_double_count(); 2655 break; 2656 case Op_Opaque1: // Remove Opaque Nodes before matching 2657 case Op_Opaque2: // Remove Opaque Nodes before matching 2658 case Op_Opaque3: 2659 n->subsume_by(n->in(1), this); 2660 break; 2661 case Op_CallStaticJava: 2662 case Op_CallJava: 2663 case Op_CallDynamicJava: 2664 frc.inc_java_call_count(); // Count java call site; 2665 case Op_CallRuntime: 2666 case Op_CallLeaf: 2667 case Op_CallLeafNoFP: { 2668 assert( n->is_Call(), "" ); 2669 CallNode *call = n->as_Call(); 2670 // Count call sites where the FP mode bit would have to be flipped. 2671 // Do not count uncommon runtime calls: 2672 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2673 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2674 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2675 frc.inc_call_count(); // Count the call site 2676 } else { // See if uncommon argument is shared 2677 Node *n = call->in(TypeFunc::Parms); 2678 int nop = n->Opcode(); 2679 // Clone shared simple arguments to uncommon calls, item (1). 2680 if( n->outcnt() > 1 && 2681 !n->is_Proj() && 2682 nop != Op_CreateEx && 2683 nop != Op_CheckCastPP && 2684 nop != Op_DecodeN && 2685 nop != Op_DecodeNKlass && 2686 !n->is_Mem() ) { 2687 Node *x = n->clone(); 2688 call->set_req( TypeFunc::Parms, x ); 2689 } 2690 } 2691 break; 2692 } 2693 2694 case Op_StoreD: 2695 case Op_LoadD: 2696 case Op_LoadD_unaligned: 2697 frc.inc_double_count(); 2698 goto handle_mem; 2699 case Op_StoreF: 2700 case Op_LoadF: 2701 frc.inc_float_count(); 2702 goto handle_mem; 2703 2704 case Op_StoreCM: 2705 { 2706 // Convert OopStore dependence into precedence edge 2707 Node* prec = n->in(MemNode::OopStore); 2708 n->del_req(MemNode::OopStore); 2709 n->add_prec(prec); 2710 eliminate_redundant_card_marks(n); 2711 } 2712 2713 // fall through 2714 2715 case Op_StoreB: 2716 case Op_StoreC: 2717 case Op_StorePConditional: 2718 case Op_StoreI: 2719 case Op_StoreL: 2720 case Op_StoreIConditional: 2721 case Op_StoreLConditional: 2722 case Op_CompareAndSwapI: 2723 case Op_CompareAndSwapL: 2724 case Op_CompareAndSwapP: 2725 case Op_CompareAndSwapN: 2726 case Op_GetAndAddI: 2727 case Op_GetAndAddL: 2728 case Op_GetAndSetI: 2729 case Op_GetAndSetL: 2730 case Op_GetAndSetP: 2731 case Op_GetAndSetN: 2732 case Op_StoreP: 2733 case Op_StoreN: 2734 case Op_StoreNKlass: 2735 case Op_LoadB: 2736 case Op_LoadUB: 2737 case Op_LoadUS: 2738 case Op_LoadI: 2739 case Op_LoadKlass: 2740 case Op_LoadNKlass: 2741 case Op_LoadL: 2742 case Op_LoadL_unaligned: 2743 case Op_LoadPLocked: 2744 case Op_LoadP: 2745 case Op_LoadN: 2746 case Op_LoadRange: 2747 case Op_LoadS: { 2748 handle_mem: 2749 #ifdef ASSERT 2750 if( VerifyOptoOopOffsets ) { 2751 assert( n->is_Mem(), "" ); 2752 MemNode *mem = (MemNode*)n; 2753 // Check to see if address types have grounded out somehow. 2754 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2755 assert( !tp || oop_offset_is_sane(tp), "" ); 2756 } 2757 #endif 2758 break; 2759 } 2760 2761 case Op_AddP: { // Assert sane base pointers 2762 Node *addp = n->in(AddPNode::Address); 2763 assert( !addp->is_AddP() || 2764 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2765 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2766 "Base pointers must match" ); 2767 #ifdef _LP64 2768 if ((UseCompressedOops || UseCompressedClassPointers) && 2769 addp->Opcode() == Op_ConP && 2770 addp == n->in(AddPNode::Base) && 2771 n->in(AddPNode::Offset)->is_Con()) { 2772 // Use addressing with narrow klass to load with offset on x86. 2773 // On sparc loading 32-bits constant and decoding it have less 2774 // instructions (4) then load 64-bits constant (7). 2775 // Do this transformation here since IGVN will convert ConN back to ConP. 2776 const Type* t = addp->bottom_type(); 2777 if (t->isa_oopptr() || t->isa_klassptr()) { 2778 Node* nn = NULL; 2779 2780 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2781 2782 // Look for existing ConN node of the same exact type. 2783 Node* r = root(); 2784 uint cnt = r->outcnt(); 2785 for (uint i = 0; i < cnt; i++) { 2786 Node* m = r->raw_out(i); 2787 if (m!= NULL && m->Opcode() == op && 2788 m->bottom_type()->make_ptr() == t) { 2789 nn = m; 2790 break; 2791 } 2792 } 2793 if (nn != NULL) { 2794 // Decode a narrow oop to match address 2795 // [R12 + narrow_oop_reg<<3 + offset] 2796 if (t->isa_oopptr()) { 2797 nn = new DecodeNNode(nn, t); 2798 } else { 2799 nn = new DecodeNKlassNode(nn, t); 2800 } 2801 n->set_req(AddPNode::Base, nn); 2802 n->set_req(AddPNode::Address, nn); 2803 if (addp->outcnt() == 0) { 2804 addp->disconnect_inputs(NULL, this); 2805 } 2806 } 2807 } 2808 } 2809 #endif 2810 break; 2811 } 2812 2813 #ifdef _LP64 2814 case Op_CastPP: 2815 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2816 Node* in1 = n->in(1); 2817 const Type* t = n->bottom_type(); 2818 Node* new_in1 = in1->clone(); 2819 new_in1->as_DecodeN()->set_type(t); 2820 2821 if (!Matcher::narrow_oop_use_complex_address()) { 2822 // 2823 // x86, ARM and friends can handle 2 adds in addressing mode 2824 // and Matcher can fold a DecodeN node into address by using 2825 // a narrow oop directly and do implicit NULL check in address: 2826 // 2827 // [R12 + narrow_oop_reg<<3 + offset] 2828 // NullCheck narrow_oop_reg 2829 // 2830 // On other platforms (Sparc) we have to keep new DecodeN node and 2831 // use it to do implicit NULL check in address: 2832 // 2833 // decode_not_null narrow_oop_reg, base_reg 2834 // [base_reg + offset] 2835 // NullCheck base_reg 2836 // 2837 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2838 // to keep the information to which NULL check the new DecodeN node 2839 // corresponds to use it as value in implicit_null_check(). 2840 // 2841 new_in1->set_req(0, n->in(0)); 2842 } 2843 2844 n->subsume_by(new_in1, this); 2845 if (in1->outcnt() == 0) { 2846 in1->disconnect_inputs(NULL, this); 2847 } 2848 } 2849 break; 2850 2851 case Op_CmpP: 2852 // Do this transformation here to preserve CmpPNode::sub() and 2853 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2854 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2855 Node* in1 = n->in(1); 2856 Node* in2 = n->in(2); 2857 if (!in1->is_DecodeNarrowPtr()) { 2858 in2 = in1; 2859 in1 = n->in(2); 2860 } 2861 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2862 2863 Node* new_in2 = NULL; 2864 if (in2->is_DecodeNarrowPtr()) { 2865 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2866 new_in2 = in2->in(1); 2867 } else if (in2->Opcode() == Op_ConP) { 2868 const Type* t = in2->bottom_type(); 2869 if (t == TypePtr::NULL_PTR) { 2870 assert(in1->is_DecodeN(), "compare klass to null?"); 2871 // Don't convert CmpP null check into CmpN if compressed 2872 // oops implicit null check is not generated. 2873 // This will allow to generate normal oop implicit null check. 2874 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2875 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 2876 // 2877 // This transformation together with CastPP transformation above 2878 // will generated code for implicit NULL checks for compressed oops. 2879 // 2880 // The original code after Optimize() 2881 // 2882 // LoadN memory, narrow_oop_reg 2883 // decode narrow_oop_reg, base_reg 2884 // CmpP base_reg, NULL 2885 // CastPP base_reg // NotNull 2886 // Load [base_reg + offset], val_reg 2887 // 2888 // after these transformations will be 2889 // 2890 // LoadN memory, narrow_oop_reg 2891 // CmpN narrow_oop_reg, NULL 2892 // decode_not_null narrow_oop_reg, base_reg 2893 // Load [base_reg + offset], val_reg 2894 // 2895 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2896 // since narrow oops can be used in debug info now (see the code in 2897 // final_graph_reshaping_walk()). 2898 // 2899 // At the end the code will be matched to 2900 // on x86: 2901 // 2902 // Load_narrow_oop memory, narrow_oop_reg 2903 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2904 // NullCheck narrow_oop_reg 2905 // 2906 // and on sparc: 2907 // 2908 // Load_narrow_oop memory, narrow_oop_reg 2909 // decode_not_null narrow_oop_reg, base_reg 2910 // Load [base_reg + offset], val_reg 2911 // NullCheck base_reg 2912 // 2913 } else if (t->isa_oopptr()) { 2914 new_in2 = ConNode::make(t->make_narrowoop()); 2915 } else if (t->isa_klassptr()) { 2916 new_in2 = ConNode::make(t->make_narrowklass()); 2917 } 2918 } 2919 if (new_in2 != NULL) { 2920 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 2921 n->subsume_by(cmpN, this); 2922 if (in1->outcnt() == 0) { 2923 in1->disconnect_inputs(NULL, this); 2924 } 2925 if (in2->outcnt() == 0) { 2926 in2->disconnect_inputs(NULL, this); 2927 } 2928 } 2929 } 2930 break; 2931 2932 case Op_DecodeN: 2933 case Op_DecodeNKlass: 2934 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2935 // DecodeN could be pinned when it can't be fold into 2936 // an address expression, see the code for Op_CastPP above. 2937 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2938 break; 2939 2940 case Op_EncodeP: 2941 case Op_EncodePKlass: { 2942 Node* in1 = n->in(1); 2943 if (in1->is_DecodeNarrowPtr()) { 2944 n->subsume_by(in1->in(1), this); 2945 } else if (in1->Opcode() == Op_ConP) { 2946 const Type* t = in1->bottom_type(); 2947 if (t == TypePtr::NULL_PTR) { 2948 assert(t->isa_oopptr(), "null klass?"); 2949 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 2950 } else if (t->isa_oopptr()) { 2951 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 2952 } else if (t->isa_klassptr()) { 2953 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 2954 } 2955 } 2956 if (in1->outcnt() == 0) { 2957 in1->disconnect_inputs(NULL, this); 2958 } 2959 break; 2960 } 2961 2962 case Op_Proj: { 2963 if (OptimizeStringConcat) { 2964 ProjNode* p = n->as_Proj(); 2965 if (p->_is_io_use) { 2966 // Separate projections were used for the exception path which 2967 // are normally removed by a late inline. If it wasn't inlined 2968 // then they will hang around and should just be replaced with 2969 // the original one. 2970 Node* proj = NULL; 2971 // Replace with just one 2972 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2973 Node *use = i.get(); 2974 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2975 proj = use; 2976 break; 2977 } 2978 } 2979 assert(proj != NULL, "must be found"); 2980 p->subsume_by(proj, this); 2981 } 2982 } 2983 break; 2984 } 2985 2986 case Op_Phi: 2987 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2988 // The EncodeP optimization may create Phi with the same edges 2989 // for all paths. It is not handled well by Register Allocator. 2990 Node* unique_in = n->in(1); 2991 assert(unique_in != NULL, ""); 2992 uint cnt = n->req(); 2993 for (uint i = 2; i < cnt; i++) { 2994 Node* m = n->in(i); 2995 assert(m != NULL, ""); 2996 if (unique_in != m) 2997 unique_in = NULL; 2998 } 2999 if (unique_in != NULL) { 3000 n->subsume_by(unique_in, this); 3001 } 3002 } 3003 break; 3004 3005 #endif 3006 3007 case Op_ModI: 3008 if (UseDivMod) { 3009 // Check if a%b and a/b both exist 3010 Node* d = n->find_similar(Op_DivI); 3011 if (d) { 3012 // Replace them with a fused divmod if supported 3013 if (Matcher::has_match_rule(Op_DivModI)) { 3014 DivModINode* divmod = DivModINode::make(n); 3015 d->subsume_by(divmod->div_proj(), this); 3016 n->subsume_by(divmod->mod_proj(), this); 3017 } else { 3018 // replace a%b with a-((a/b)*b) 3019 Node* mult = new MulINode(d, d->in(2)); 3020 Node* sub = new SubINode(d->in(1), mult); 3021 n->subsume_by(sub, this); 3022 } 3023 } 3024 } 3025 break; 3026 3027 case Op_ModL: 3028 if (UseDivMod) { 3029 // Check if a%b and a/b both exist 3030 Node* d = n->find_similar(Op_DivL); 3031 if (d) { 3032 // Replace them with a fused divmod if supported 3033 if (Matcher::has_match_rule(Op_DivModL)) { 3034 DivModLNode* divmod = DivModLNode::make(n); 3035 d->subsume_by(divmod->div_proj(), this); 3036 n->subsume_by(divmod->mod_proj(), this); 3037 } else { 3038 // replace a%b with a-((a/b)*b) 3039 Node* mult = new MulLNode(d, d->in(2)); 3040 Node* sub = new SubLNode(d->in(1), mult); 3041 n->subsume_by(sub, this); 3042 } 3043 } 3044 } 3045 break; 3046 3047 case Op_LoadVector: 3048 case Op_StoreVector: 3049 break; 3050 3051 case Op_PackB: 3052 case Op_PackS: 3053 case Op_PackI: 3054 case Op_PackF: 3055 case Op_PackL: 3056 case Op_PackD: 3057 if (n->req()-1 > 2) { 3058 // Replace many operand PackNodes with a binary tree for matching 3059 PackNode* p = (PackNode*) n; 3060 Node* btp = p->binary_tree_pack(1, n->req()); 3061 n->subsume_by(btp, this); 3062 } 3063 break; 3064 case Op_Loop: 3065 case Op_CountedLoop: 3066 if (n->as_Loop()->is_inner_loop()) { 3067 frc.inc_inner_loop_count(); 3068 } 3069 break; 3070 case Op_LShiftI: 3071 case Op_RShiftI: 3072 case Op_URShiftI: 3073 case Op_LShiftL: 3074 case Op_RShiftL: 3075 case Op_URShiftL: 3076 if (Matcher::need_masked_shift_count) { 3077 // The cpu's shift instructions don't restrict the count to the 3078 // lower 5/6 bits. We need to do the masking ourselves. 3079 Node* in2 = n->in(2); 3080 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3081 const TypeInt* t = in2->find_int_type(); 3082 if (t != NULL && t->is_con()) { 3083 juint shift = t->get_con(); 3084 if (shift > mask) { // Unsigned cmp 3085 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3086 } 3087 } else { 3088 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3089 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3090 n->set_req(2, shift); 3091 } 3092 } 3093 if (in2->outcnt() == 0) { // Remove dead node 3094 in2->disconnect_inputs(NULL, this); 3095 } 3096 } 3097 break; 3098 case Op_MemBarStoreStore: 3099 case Op_MemBarRelease: 3100 // Break the link with AllocateNode: it is no longer useful and 3101 // confuses register allocation. 3102 if (n->req() > MemBarNode::Precedent) { 3103 n->set_req(MemBarNode::Precedent, top()); 3104 } 3105 break; 3106 default: 3107 assert( !n->is_Call(), "" ); 3108 assert( !n->is_Mem(), "" ); 3109 break; 3110 } 3111 3112 // Collect CFG split points 3113 if (n->is_MultiBranch()) 3114 frc._tests.push(n); 3115 } 3116 3117 //------------------------------final_graph_reshaping_walk--------------------- 3118 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3119 // requires that the walk visits a node's inputs before visiting the node. 3120 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3121 ResourceArea *area = Thread::current()->resource_area(); 3122 Unique_Node_List sfpt(area); 3123 3124 frc._visited.set(root->_idx); // first, mark node as visited 3125 uint cnt = root->req(); 3126 Node *n = root; 3127 uint i = 0; 3128 while (true) { 3129 if (i < cnt) { 3130 // Place all non-visited non-null inputs onto stack 3131 Node* m = n->in(i); 3132 ++i; 3133 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3134 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3135 // compute worst case interpreter size in case of a deoptimization 3136 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3137 3138 sfpt.push(m); 3139 } 3140 cnt = m->req(); 3141 nstack.push(n, i); // put on stack parent and next input's index 3142 n = m; 3143 i = 0; 3144 } 3145 } else { 3146 // Now do post-visit work 3147 final_graph_reshaping_impl( n, frc ); 3148 if (nstack.is_empty()) 3149 break; // finished 3150 n = nstack.node(); // Get node from stack 3151 cnt = n->req(); 3152 i = nstack.index(); 3153 nstack.pop(); // Shift to the next node on stack 3154 } 3155 } 3156 3157 // Skip next transformation if compressed oops are not used. 3158 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3159 (!UseCompressedOops && !UseCompressedClassPointers)) 3160 return; 3161 3162 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3163 // It could be done for an uncommon traps or any safepoints/calls 3164 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3165 while (sfpt.size() > 0) { 3166 n = sfpt.pop(); 3167 JVMState *jvms = n->as_SafePoint()->jvms(); 3168 assert(jvms != NULL, "sanity"); 3169 int start = jvms->debug_start(); 3170 int end = n->req(); 3171 bool is_uncommon = (n->is_CallStaticJava() && 3172 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3173 for (int j = start; j < end; j++) { 3174 Node* in = n->in(j); 3175 if (in->is_DecodeNarrowPtr()) { 3176 bool safe_to_skip = true; 3177 if (!is_uncommon ) { 3178 // Is it safe to skip? 3179 for (uint i = 0; i < in->outcnt(); i++) { 3180 Node* u = in->raw_out(i); 3181 if (!u->is_SafePoint() || 3182 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3183 safe_to_skip = false; 3184 } 3185 } 3186 } 3187 if (safe_to_skip) { 3188 n->set_req(j, in->in(1)); 3189 } 3190 if (in->outcnt() == 0) { 3191 in->disconnect_inputs(NULL, this); 3192 } 3193 } 3194 } 3195 } 3196 } 3197 3198 //------------------------------final_graph_reshaping-------------------------- 3199 // Final Graph Reshaping. 3200 // 3201 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3202 // and not commoned up and forced early. Must come after regular 3203 // optimizations to avoid GVN undoing the cloning. Clone constant 3204 // inputs to Loop Phis; these will be split by the allocator anyways. 3205 // Remove Opaque nodes. 3206 // (2) Move last-uses by commutative operations to the left input to encourage 3207 // Intel update-in-place two-address operations and better register usage 3208 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3209 // calls canonicalizing them back. 3210 // (3) Count the number of double-precision FP ops, single-precision FP ops 3211 // and call sites. On Intel, we can get correct rounding either by 3212 // forcing singles to memory (requires extra stores and loads after each 3213 // FP bytecode) or we can set a rounding mode bit (requires setting and 3214 // clearing the mode bit around call sites). The mode bit is only used 3215 // if the relative frequency of single FP ops to calls is low enough. 3216 // This is a key transform for SPEC mpeg_audio. 3217 // (4) Detect infinite loops; blobs of code reachable from above but not 3218 // below. Several of the Code_Gen algorithms fail on such code shapes, 3219 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3220 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3221 // Detection is by looking for IfNodes where only 1 projection is 3222 // reachable from below or CatchNodes missing some targets. 3223 // (5) Assert for insane oop offsets in debug mode. 3224 3225 bool Compile::final_graph_reshaping() { 3226 // an infinite loop may have been eliminated by the optimizer, 3227 // in which case the graph will be empty. 3228 if (root()->req() == 1) { 3229 record_method_not_compilable("trivial infinite loop"); 3230 return true; 3231 } 3232 3233 // Expensive nodes have their control input set to prevent the GVN 3234 // from freely commoning them. There's no GVN beyond this point so 3235 // no need to keep the control input. We want the expensive nodes to 3236 // be freely moved to the least frequent code path by gcm. 3237 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3238 for (int i = 0; i < expensive_count(); i++) { 3239 _expensive_nodes->at(i)->set_req(0, NULL); 3240 } 3241 3242 Final_Reshape_Counts frc; 3243 3244 // Visit everybody reachable! 3245 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3246 Node_Stack nstack(unique() >> 1); 3247 final_graph_reshaping_walk(nstack, root(), frc); 3248 3249 // Check for unreachable (from below) code (i.e., infinite loops). 3250 for( uint i = 0; i < frc._tests.size(); i++ ) { 3251 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3252 // Get number of CFG targets. 3253 // Note that PCTables include exception targets after calls. 3254 uint required_outcnt = n->required_outcnt(); 3255 if (n->outcnt() != required_outcnt) { 3256 // Check for a few special cases. Rethrow Nodes never take the 3257 // 'fall-thru' path, so expected kids is 1 less. 3258 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3259 if (n->in(0)->in(0)->is_Call()) { 3260 CallNode *call = n->in(0)->in(0)->as_Call(); 3261 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3262 required_outcnt--; // Rethrow always has 1 less kid 3263 } else if (call->req() > TypeFunc::Parms && 3264 call->is_CallDynamicJava()) { 3265 // Check for null receiver. In such case, the optimizer has 3266 // detected that the virtual call will always result in a null 3267 // pointer exception. The fall-through projection of this CatchNode 3268 // will not be populated. 3269 Node *arg0 = call->in(TypeFunc::Parms); 3270 if (arg0->is_Type() && 3271 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3272 required_outcnt--; 3273 } 3274 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3275 call->req() > TypeFunc::Parms+1 && 3276 call->is_CallStaticJava()) { 3277 // Check for negative array length. In such case, the optimizer has 3278 // detected that the allocation attempt will always result in an 3279 // exception. There is no fall-through projection of this CatchNode . 3280 Node *arg1 = call->in(TypeFunc::Parms+1); 3281 if (arg1->is_Type() && 3282 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3283 required_outcnt--; 3284 } 3285 } 3286 } 3287 } 3288 // Recheck with a better notion of 'required_outcnt' 3289 if (n->outcnt() != required_outcnt) { 3290 record_method_not_compilable("malformed control flow"); 3291 return true; // Not all targets reachable! 3292 } 3293 } 3294 // Check that I actually visited all kids. Unreached kids 3295 // must be infinite loops. 3296 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3297 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3298 record_method_not_compilable("infinite loop"); 3299 return true; // Found unvisited kid; must be unreach 3300 } 3301 } 3302 3303 // If original bytecodes contained a mixture of floats and doubles 3304 // check if the optimizer has made it homogenous, item (3). 3305 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3306 frc.get_float_count() > 32 && 3307 frc.get_double_count() == 0 && 3308 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3309 set_24_bit_selection_and_mode( false, true ); 3310 } 3311 3312 set_java_calls(frc.get_java_call_count()); 3313 set_inner_loops(frc.get_inner_loop_count()); 3314 3315 // No infinite loops, no reason to bail out. 3316 return false; 3317 } 3318 3319 //-----------------------------too_many_traps---------------------------------- 3320 // Report if there are too many traps at the current method and bci. 3321 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3322 bool Compile::too_many_traps(ciMethod* method, 3323 int bci, 3324 Deoptimization::DeoptReason reason) { 3325 ciMethodData* md = method->method_data(); 3326 if (md->is_empty()) { 3327 // Assume the trap has not occurred, or that it occurred only 3328 // because of a transient condition during start-up in the interpreter. 3329 return false; 3330 } 3331 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3332 if (md->has_trap_at(bci, m, reason) != 0) { 3333 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3334 // Also, if there are multiple reasons, or if there is no per-BCI record, 3335 // assume the worst. 3336 if (log()) 3337 log()->elem("observe trap='%s' count='%d'", 3338 Deoptimization::trap_reason_name(reason), 3339 md->trap_count(reason)); 3340 return true; 3341 } else { 3342 // Ignore method/bci and see if there have been too many globally. 3343 return too_many_traps(reason, md); 3344 } 3345 } 3346 3347 // Less-accurate variant which does not require a method and bci. 3348 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3349 ciMethodData* logmd) { 3350 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3351 // Too many traps globally. 3352 // Note that we use cumulative trap_count, not just md->trap_count. 3353 if (log()) { 3354 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3355 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3356 Deoptimization::trap_reason_name(reason), 3357 mcount, trap_count(reason)); 3358 } 3359 return true; 3360 } else { 3361 // The coast is clear. 3362 return false; 3363 } 3364 } 3365 3366 //--------------------------too_many_recompiles-------------------------------- 3367 // Report if there are too many recompiles at the current method and bci. 3368 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3369 // Is not eager to return true, since this will cause the compiler to use 3370 // Action_none for a trap point, to avoid too many recompilations. 3371 bool Compile::too_many_recompiles(ciMethod* method, 3372 int bci, 3373 Deoptimization::DeoptReason reason) { 3374 ciMethodData* md = method->method_data(); 3375 if (md->is_empty()) { 3376 // Assume the trap has not occurred, or that it occurred only 3377 // because of a transient condition during start-up in the interpreter. 3378 return false; 3379 } 3380 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3381 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3382 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3383 Deoptimization::DeoptReason per_bc_reason 3384 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3385 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3386 if ((per_bc_reason == Deoptimization::Reason_none 3387 || md->has_trap_at(bci, m, reason) != 0) 3388 // The trap frequency measure we care about is the recompile count: 3389 && md->trap_recompiled_at(bci, m) 3390 && md->overflow_recompile_count() >= bc_cutoff) { 3391 // Do not emit a trap here if it has already caused recompilations. 3392 // Also, if there are multiple reasons, or if there is no per-BCI record, 3393 // assume the worst. 3394 if (log()) 3395 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3396 Deoptimization::trap_reason_name(reason), 3397 md->trap_count(reason), 3398 md->overflow_recompile_count()); 3399 return true; 3400 } else if (trap_count(reason) != 0 3401 && decompile_count() >= m_cutoff) { 3402 // Too many recompiles globally, and we have seen this sort of trap. 3403 // Use cumulative decompile_count, not just md->decompile_count. 3404 if (log()) 3405 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3406 Deoptimization::trap_reason_name(reason), 3407 md->trap_count(reason), trap_count(reason), 3408 md->decompile_count(), decompile_count()); 3409 return true; 3410 } else { 3411 // The coast is clear. 3412 return false; 3413 } 3414 } 3415 3416 // Compute when not to trap. Used by matching trap based nodes and 3417 // NullCheck optimization. 3418 void Compile::set_allowed_deopt_reasons() { 3419 _allowed_reasons = 0; 3420 if (is_method_compilation()) { 3421 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3422 assert(rs < BitsPerInt, "recode bit map"); 3423 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3424 _allowed_reasons |= nth_bit(rs); 3425 } 3426 } 3427 } 3428 } 3429 3430 #ifndef PRODUCT 3431 //------------------------------verify_graph_edges--------------------------- 3432 // Walk the Graph and verify that there is a one-to-one correspondence 3433 // between Use-Def edges and Def-Use edges in the graph. 3434 void Compile::verify_graph_edges(bool no_dead_code) { 3435 if (VerifyGraphEdges) { 3436 ResourceArea *area = Thread::current()->resource_area(); 3437 Unique_Node_List visited(area); 3438 // Call recursive graph walk to check edges 3439 _root->verify_edges(visited); 3440 if (no_dead_code) { 3441 // Now make sure that no visited node is used by an unvisited node. 3442 bool dead_nodes = false; 3443 Unique_Node_List checked(area); 3444 while (visited.size() > 0) { 3445 Node* n = visited.pop(); 3446 checked.push(n); 3447 for (uint i = 0; i < n->outcnt(); i++) { 3448 Node* use = n->raw_out(i); 3449 if (checked.member(use)) continue; // already checked 3450 if (visited.member(use)) continue; // already in the graph 3451 if (use->is_Con()) continue; // a dead ConNode is OK 3452 // At this point, we have found a dead node which is DU-reachable. 3453 if (!dead_nodes) { 3454 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3455 dead_nodes = true; 3456 } 3457 use->dump(2); 3458 tty->print_cr("---"); 3459 checked.push(use); // No repeats; pretend it is now checked. 3460 } 3461 } 3462 assert(!dead_nodes, "using nodes must be reachable from root"); 3463 } 3464 } 3465 } 3466 3467 // Verify GC barriers consistency 3468 // Currently supported: 3469 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3470 void Compile::verify_barriers() { 3471 if (UseG1GC) { 3472 // Verify G1 pre-barriers 3473 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3474 3475 ResourceArea *area = Thread::current()->resource_area(); 3476 Unique_Node_List visited(area); 3477 Node_List worklist(area); 3478 // We're going to walk control flow backwards starting from the Root 3479 worklist.push(_root); 3480 while (worklist.size() > 0) { 3481 Node* x = worklist.pop(); 3482 if (x == NULL || x == top()) continue; 3483 if (visited.member(x)) { 3484 continue; 3485 } else { 3486 visited.push(x); 3487 } 3488 3489 if (x->is_Region()) { 3490 for (uint i = 1; i < x->req(); i++) { 3491 worklist.push(x->in(i)); 3492 } 3493 } else { 3494 worklist.push(x->in(0)); 3495 // We are looking for the pattern: 3496 // /->ThreadLocal 3497 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3498 // \->ConI(0) 3499 // We want to verify that the If and the LoadB have the same control 3500 // See GraphKit::g1_write_barrier_pre() 3501 if (x->is_If()) { 3502 IfNode *iff = x->as_If(); 3503 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3504 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3505 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3506 && cmp->in(1)->is_Load()) { 3507 LoadNode* load = cmp->in(1)->as_Load(); 3508 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3509 && load->in(2)->in(3)->is_Con() 3510 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3511 3512 Node* if_ctrl = iff->in(0); 3513 Node* load_ctrl = load->in(0); 3514 3515 if (if_ctrl != load_ctrl) { 3516 // Skip possible CProj->NeverBranch in infinite loops 3517 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3518 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3519 if_ctrl = if_ctrl->in(0)->in(0); 3520 } 3521 } 3522 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3523 } 3524 } 3525 } 3526 } 3527 } 3528 } 3529 } 3530 } 3531 3532 #endif 3533 3534 // The Compile object keeps track of failure reasons separately from the ciEnv. 3535 // This is required because there is not quite a 1-1 relation between the 3536 // ciEnv and its compilation task and the Compile object. Note that one 3537 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3538 // to backtrack and retry without subsuming loads. Other than this backtracking 3539 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3540 // by the logic in C2Compiler. 3541 void Compile::record_failure(const char* reason) { 3542 if (log() != NULL) { 3543 log()->elem("failure reason='%s' phase='compile'", reason); 3544 } 3545 if (_failure_reason == NULL) { 3546 // Record the first failure reason. 3547 _failure_reason = reason; 3548 } 3549 3550 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3551 C->print_method(PHASE_FAILURE); 3552 } 3553 _root = NULL; // flush the graph, too 3554 } 3555 3556 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3557 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3558 _phase_name(name), _dolog(CITimeVerbose) 3559 { 3560 if (_dolog) { 3561 C = Compile::current(); 3562 _log = C->log(); 3563 } else { 3564 C = NULL; 3565 _log = NULL; 3566 } 3567 if (_log != NULL) { 3568 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3569 _log->stamp(); 3570 _log->end_head(); 3571 } 3572 } 3573 3574 Compile::TracePhase::~TracePhase() { 3575 3576 C = Compile::current(); 3577 if (_dolog) { 3578 _log = C->log(); 3579 } else { 3580 _log = NULL; 3581 } 3582 3583 #ifdef ASSERT 3584 if (PrintIdealNodeCount) { 3585 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3586 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3587 } 3588 3589 if (VerifyIdealNodeCount) { 3590 Compile::current()->print_missing_nodes(); 3591 } 3592 #endif 3593 3594 if (_log != NULL) { 3595 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3596 } 3597 } 3598 3599 //============================================================================= 3600 // Two Constant's are equal when the type and the value are equal. 3601 bool Compile::Constant::operator==(const Constant& other) { 3602 if (type() != other.type() ) return false; 3603 if (can_be_reused() != other.can_be_reused()) return false; 3604 // For floating point values we compare the bit pattern. 3605 switch (type()) { 3606 case T_FLOAT: return (_v._value.i == other._v._value.i); 3607 case T_LONG: 3608 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3609 case T_OBJECT: 3610 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3611 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3612 case T_METADATA: return (_v._metadata == other._v._metadata); 3613 default: ShouldNotReachHere(); 3614 } 3615 return false; 3616 } 3617 3618 static int type_to_size_in_bytes(BasicType t) { 3619 switch (t) { 3620 case T_LONG: return sizeof(jlong ); 3621 case T_FLOAT: return sizeof(jfloat ); 3622 case T_DOUBLE: return sizeof(jdouble); 3623 case T_METADATA: return sizeof(Metadata*); 3624 // We use T_VOID as marker for jump-table entries (labels) which 3625 // need an internal word relocation. 3626 case T_VOID: 3627 case T_ADDRESS: 3628 case T_OBJECT: return sizeof(jobject); 3629 } 3630 3631 ShouldNotReachHere(); 3632 return -1; 3633 } 3634 3635 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3636 // sort descending 3637 if (a->freq() > b->freq()) return -1; 3638 if (a->freq() < b->freq()) return 1; 3639 return 0; 3640 } 3641 3642 void Compile::ConstantTable::calculate_offsets_and_size() { 3643 // First, sort the array by frequencies. 3644 _constants.sort(qsort_comparator); 3645 3646 #ifdef ASSERT 3647 // Make sure all jump-table entries were sorted to the end of the 3648 // array (they have a negative frequency). 3649 bool found_void = false; 3650 for (int i = 0; i < _constants.length(); i++) { 3651 Constant con = _constants.at(i); 3652 if (con.type() == T_VOID) 3653 found_void = true; // jump-tables 3654 else 3655 assert(!found_void, "wrong sorting"); 3656 } 3657 #endif 3658 3659 int offset = 0; 3660 for (int i = 0; i < _constants.length(); i++) { 3661 Constant* con = _constants.adr_at(i); 3662 3663 // Align offset for type. 3664 int typesize = type_to_size_in_bytes(con->type()); 3665 offset = align_size_up(offset, typesize); 3666 con->set_offset(offset); // set constant's offset 3667 3668 if (con->type() == T_VOID) { 3669 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3670 offset = offset + typesize * n->outcnt(); // expand jump-table 3671 } else { 3672 offset = offset + typesize; 3673 } 3674 } 3675 3676 // Align size up to the next section start (which is insts; see 3677 // CodeBuffer::align_at_start). 3678 assert(_size == -1, "already set?"); 3679 _size = align_size_up(offset, CodeEntryAlignment); 3680 } 3681 3682 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3683 MacroAssembler _masm(&cb); 3684 for (int i = 0; i < _constants.length(); i++) { 3685 Constant con = _constants.at(i); 3686 address constant_addr; 3687 switch (con.type()) { 3688 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3689 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3690 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3691 case T_OBJECT: { 3692 jobject obj = con.get_jobject(); 3693 int oop_index = _masm.oop_recorder()->find_index(obj); 3694 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3695 break; 3696 } 3697 case T_ADDRESS: { 3698 address addr = (address) con.get_jobject(); 3699 constant_addr = _masm.address_constant(addr); 3700 break; 3701 } 3702 // We use T_VOID as marker for jump-table entries (labels) which 3703 // need an internal word relocation. 3704 case T_VOID: { 3705 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3706 // Fill the jump-table with a dummy word. The real value is 3707 // filled in later in fill_jump_table. 3708 address dummy = (address) n; 3709 constant_addr = _masm.address_constant(dummy); 3710 // Expand jump-table 3711 for (uint i = 1; i < n->outcnt(); i++) { 3712 address temp_addr = _masm.address_constant(dummy + i); 3713 assert(temp_addr, "consts section too small"); 3714 } 3715 break; 3716 } 3717 case T_METADATA: { 3718 Metadata* obj = con.get_metadata(); 3719 int metadata_index = _masm.oop_recorder()->find_index(obj); 3720 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3721 break; 3722 } 3723 default: ShouldNotReachHere(); 3724 } 3725 assert(constant_addr, "consts section too small"); 3726 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3727 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3728 } 3729 } 3730 3731 int Compile::ConstantTable::find_offset(Constant& con) const { 3732 int idx = _constants.find(con); 3733 assert(idx != -1, "constant must be in constant table"); 3734 int offset = _constants.at(idx).offset(); 3735 assert(offset != -1, "constant table not emitted yet?"); 3736 return offset; 3737 } 3738 3739 void Compile::ConstantTable::add(Constant& con) { 3740 if (con.can_be_reused()) { 3741 int idx = _constants.find(con); 3742 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3743 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3744 return; 3745 } 3746 } 3747 (void) _constants.append(con); 3748 } 3749 3750 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3751 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3752 Constant con(type, value, b->_freq); 3753 add(con); 3754 return con; 3755 } 3756 3757 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3758 Constant con(metadata); 3759 add(con); 3760 return con; 3761 } 3762 3763 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3764 jvalue value; 3765 BasicType type = oper->type()->basic_type(); 3766 switch (type) { 3767 case T_LONG: value.j = oper->constantL(); break; 3768 case T_FLOAT: value.f = oper->constantF(); break; 3769 case T_DOUBLE: value.d = oper->constantD(); break; 3770 case T_OBJECT: 3771 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3772 case T_METADATA: return add((Metadata*)oper->constant()); break; 3773 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3774 } 3775 return add(n, type, value); 3776 } 3777 3778 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3779 jvalue value; 3780 // We can use the node pointer here to identify the right jump-table 3781 // as this method is called from Compile::Fill_buffer right before 3782 // the MachNodes are emitted and the jump-table is filled (means the 3783 // MachNode pointers do not change anymore). 3784 value.l = (jobject) n; 3785 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3786 add(con); 3787 return con; 3788 } 3789 3790 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3791 // If called from Compile::scratch_emit_size do nothing. 3792 if (Compile::current()->in_scratch_emit_size()) return; 3793 3794 assert(labels.is_nonempty(), "must be"); 3795 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3796 3797 // Since MachConstantNode::constant_offset() also contains 3798 // table_base_offset() we need to subtract the table_base_offset() 3799 // to get the plain offset into the constant table. 3800 int offset = n->constant_offset() - table_base_offset(); 3801 3802 MacroAssembler _masm(&cb); 3803 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3804 3805 for (uint i = 0; i < n->outcnt(); i++) { 3806 address* constant_addr = &jump_table_base[i]; 3807 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))); 3808 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3809 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3810 } 3811 } 3812 3813 //----------------------------static_subtype_check----------------------------- 3814 // Shortcut important common cases when superklass is exact: 3815 // (0) superklass is java.lang.Object (can occur in reflective code) 3816 // (1) subklass is already limited to a subtype of superklass => always ok 3817 // (2) subklass does not overlap with superklass => always fail 3818 // (3) superklass has NO subtypes and we can check with a simple compare. 3819 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 3820 if (StressReflectiveCode) { 3821 return SSC_full_test; // Let caller generate the general case. 3822 } 3823 3824 if (superk == env()->Object_klass()) { 3825 return SSC_always_true; // (0) this test cannot fail 3826 } 3827 3828 ciType* superelem = superk; 3829 if (superelem->is_array_klass()) 3830 superelem = superelem->as_array_klass()->base_element_type(); 3831 3832 if (!subk->is_interface()) { // cannot trust static interface types yet 3833 if (subk->is_subtype_of(superk)) { 3834 return SSC_always_true; // (1) false path dead; no dynamic test needed 3835 } 3836 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 3837 !superk->is_subtype_of(subk)) { 3838 return SSC_always_false; 3839 } 3840 } 3841 3842 // If casting to an instance klass, it must have no subtypes 3843 if (superk->is_interface()) { 3844 // Cannot trust interfaces yet. 3845 // %%% S.B. superk->nof_implementors() == 1 3846 } else if (superelem->is_instance_klass()) { 3847 ciInstanceKlass* ik = superelem->as_instance_klass(); 3848 if (!ik->has_subklass() && !ik->is_interface()) { 3849 if (!ik->is_final()) { 3850 // Add a dependency if there is a chance of a later subclass. 3851 dependencies()->assert_leaf_type(ik); 3852 } 3853 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3854 } 3855 } else { 3856 // A primitive array type has no subtypes. 3857 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3858 } 3859 3860 return SSC_full_test; 3861 } 3862 3863 Node* Compile::conv_I2X_index(PhaseGVN *phase, Node* idx, const TypeInt* sizetype) { 3864 #ifdef _LP64 3865 // The scaled index operand to AddP must be a clean 64-bit value. 3866 // Java allows a 32-bit int to be incremented to a negative 3867 // value, which appears in a 64-bit register as a large 3868 // positive number. Using that large positive number as an 3869 // operand in pointer arithmetic has bad consequences. 3870 // On the other hand, 32-bit overflow is rare, and the possibility 3871 // can often be excluded, if we annotate the ConvI2L node with 3872 // a type assertion that its value is known to be a small positive 3873 // number. (The prior range check has ensured this.) 3874 // This assertion is used by ConvI2LNode::Ideal. 3875 int index_max = max_jint - 1; // array size is max_jint, index is one less 3876 if (sizetype != NULL) index_max = sizetype->_hi - 1; 3877 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 3878 idx = phase->transform(new ConvI2LNode(idx, lidxtype)); 3879 #endif 3880 return idx; 3881 } 3882 3883 // The message about the current inlining is accumulated in 3884 // _print_inlining_stream and transfered into the _print_inlining_list 3885 // once we know whether inlining succeeds or not. For regular 3886 // inlining, messages are appended to the buffer pointed by 3887 // _print_inlining_idx in the _print_inlining_list. For late inlining, 3888 // a new buffer is added after _print_inlining_idx in the list. This 3889 // way we can update the inlining message for late inlining call site 3890 // when the inlining is attempted again. 3891 void Compile::print_inlining_init() { 3892 if (print_inlining() || print_intrinsics()) { 3893 _print_inlining_stream = new stringStream(); 3894 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 3895 } 3896 } 3897 3898 void Compile::print_inlining_reinit() { 3899 if (print_inlining() || print_intrinsics()) { 3900 // Re allocate buffer when we change ResourceMark 3901 _print_inlining_stream = new stringStream(); 3902 } 3903 } 3904 3905 void Compile::print_inlining_reset() { 3906 _print_inlining_stream->reset(); 3907 } 3908 3909 void Compile::print_inlining_commit() { 3910 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 3911 // Transfer the message from _print_inlining_stream to the current 3912 // _print_inlining_list buffer and clear _print_inlining_stream. 3913 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 3914 print_inlining_reset(); 3915 } 3916 3917 void Compile::print_inlining_push() { 3918 // Add new buffer to the _print_inlining_list at current position 3919 _print_inlining_idx++; 3920 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 3921 } 3922 3923 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 3924 return _print_inlining_list->at(_print_inlining_idx); 3925 } 3926 3927 void Compile::print_inlining_update(CallGenerator* cg) { 3928 if (print_inlining() || print_intrinsics()) { 3929 if (!cg->is_late_inline()) { 3930 if (print_inlining_current().cg() != NULL) { 3931 print_inlining_push(); 3932 } 3933 print_inlining_commit(); 3934 } else { 3935 if (print_inlining_current().cg() != cg && 3936 (print_inlining_current().cg() != NULL || 3937 print_inlining_current().ss()->size() != 0)) { 3938 print_inlining_push(); 3939 } 3940 print_inlining_commit(); 3941 print_inlining_current().set_cg(cg); 3942 } 3943 } 3944 } 3945 3946 void Compile::print_inlining_move_to(CallGenerator* cg) { 3947 // We resume inlining at a late inlining call site. Locate the 3948 // corresponding inlining buffer so that we can update it. 3949 if (print_inlining()) { 3950 for (int i = 0; i < _print_inlining_list->length(); i++) { 3951 if (_print_inlining_list->adr_at(i)->cg() == cg) { 3952 _print_inlining_idx = i; 3953 return; 3954 } 3955 } 3956 ShouldNotReachHere(); 3957 } 3958 } 3959 3960 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 3961 if (print_inlining()) { 3962 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 3963 assert(print_inlining_current().cg() == cg, "wrong entry"); 3964 // replace message with new message 3965 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 3966 print_inlining_commit(); 3967 print_inlining_current().set_cg(cg); 3968 } 3969 } 3970 3971 void Compile::print_inlining_assert_ready() { 3972 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 3973 } 3974 3975 void Compile::process_print_inlining() { 3976 bool do_print_inlining = print_inlining() || print_intrinsics(); 3977 if (do_print_inlining || log() != NULL) { 3978 // Print inlining message for candidates that we couldn't inline 3979 // for lack of space 3980 for (int i = 0; i < _late_inlines.length(); i++) { 3981 CallGenerator* cg = _late_inlines.at(i); 3982 if (!cg->is_mh_late_inline()) { 3983 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 3984 if (do_print_inlining) { 3985 cg->print_inlining_late(msg); 3986 } 3987 log_late_inline_failure(cg, msg); 3988 } 3989 } 3990 } 3991 if (do_print_inlining) { 3992 ResourceMark rm; 3993 stringStream ss; 3994 for (int i = 0; i < _print_inlining_list->length(); i++) { 3995 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 3996 } 3997 size_t end = ss.size(); 3998 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 3999 strncpy(_print_inlining_output, ss.base(), end+1); 4000 _print_inlining_output[end] = 0; 4001 } 4002 } 4003 4004 void Compile::dump_print_inlining() { 4005 if (_print_inlining_output != NULL) { 4006 tty->print_raw(_print_inlining_output); 4007 } 4008 } 4009 4010 void Compile::log_late_inline(CallGenerator* cg) { 4011 if (log() != NULL) { 4012 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4013 cg->unique_id()); 4014 JVMState* p = cg->call_node()->jvms(); 4015 while (p != NULL) { 4016 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4017 p = p->caller(); 4018 } 4019 log()->tail("late_inline"); 4020 } 4021 } 4022 4023 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4024 log_late_inline(cg); 4025 if (log() != NULL) { 4026 log()->inline_fail(msg); 4027 } 4028 } 4029 4030 void Compile::log_inline_id(CallGenerator* cg) { 4031 if (log() != NULL) { 4032 // The LogCompilation tool needs a unique way to identify late 4033 // inline call sites. This id must be unique for this call site in 4034 // this compilation. Try to have it unique across compilations as 4035 // well because it can be convenient when grepping through the log 4036 // file. 4037 // Distinguish OSR compilations from others in case CICountOSR is 4038 // on. 4039 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4040 cg->set_unique_id(id); 4041 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4042 } 4043 } 4044 4045 void Compile::log_inline_failure(const char* msg) { 4046 if (C->log() != NULL) { 4047 C->log()->inline_fail(msg); 4048 } 4049 } 4050 4051 4052 // Dump inlining replay data to the stream. 4053 // Don't change thread state and acquire any locks. 4054 void Compile::dump_inline_data(outputStream* out) { 4055 InlineTree* inl_tree = ilt(); 4056 if (inl_tree != NULL) { 4057 out->print(" inline %d", inl_tree->count()); 4058 inl_tree->dump_replay_data(out); 4059 } 4060 } 4061 4062 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4063 if (n1->Opcode() < n2->Opcode()) return -1; 4064 else if (n1->Opcode() > n2->Opcode()) return 1; 4065 4066 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())); 4067 for (uint i = 1; i < n1->req(); i++) { 4068 if (n1->in(i) < n2->in(i)) return -1; 4069 else if (n1->in(i) > n2->in(i)) return 1; 4070 } 4071 4072 return 0; 4073 } 4074 4075 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4076 Node* n1 = *n1p; 4077 Node* n2 = *n2p; 4078 4079 return cmp_expensive_nodes(n1, n2); 4080 } 4081 4082 void Compile::sort_expensive_nodes() { 4083 if (!expensive_nodes_sorted()) { 4084 _expensive_nodes->sort(cmp_expensive_nodes); 4085 } 4086 } 4087 4088 bool Compile::expensive_nodes_sorted() const { 4089 for (int i = 1; i < _expensive_nodes->length(); i++) { 4090 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4091 return false; 4092 } 4093 } 4094 return true; 4095 } 4096 4097 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4098 if (_expensive_nodes->length() == 0) { 4099 return false; 4100 } 4101 4102 assert(OptimizeExpensiveOps, "optimization off?"); 4103 4104 // Take this opportunity to remove dead nodes from the list 4105 int j = 0; 4106 for (int i = 0; i < _expensive_nodes->length(); i++) { 4107 Node* n = _expensive_nodes->at(i); 4108 if (!n->is_unreachable(igvn)) { 4109 assert(n->is_expensive(), "should be expensive"); 4110 _expensive_nodes->at_put(j, n); 4111 j++; 4112 } 4113 } 4114 _expensive_nodes->trunc_to(j); 4115 4116 // Then sort the list so that similar nodes are next to each other 4117 // and check for at least two nodes of identical kind with same data 4118 // inputs. 4119 sort_expensive_nodes(); 4120 4121 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4122 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4123 return true; 4124 } 4125 } 4126 4127 return false; 4128 } 4129 4130 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4131 if (_expensive_nodes->length() == 0) { 4132 return; 4133 } 4134 4135 assert(OptimizeExpensiveOps, "optimization off?"); 4136 4137 // Sort to bring similar nodes next to each other and clear the 4138 // control input of nodes for which there's only a single copy. 4139 sort_expensive_nodes(); 4140 4141 int j = 0; 4142 int identical = 0; 4143 int i = 0; 4144 bool modified = false; 4145 for (; i < _expensive_nodes->length()-1; i++) { 4146 assert(j <= i, "can't write beyond current index"); 4147 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4148 identical++; 4149 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4150 continue; 4151 } 4152 if (identical > 0) { 4153 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4154 identical = 0; 4155 } else { 4156 Node* n = _expensive_nodes->at(i); 4157 igvn.replace_input_of(n, 0, NULL); 4158 igvn.hash_insert(n); 4159 modified = true; 4160 } 4161 } 4162 if (identical > 0) { 4163 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4164 } else if (_expensive_nodes->length() >= 1) { 4165 Node* n = _expensive_nodes->at(i); 4166 igvn.replace_input_of(n, 0, NULL); 4167 igvn.hash_insert(n); 4168 modified = true; 4169 } 4170 _expensive_nodes->trunc_to(j); 4171 if (modified) { 4172 igvn.optimize(); 4173 } 4174 } 4175 4176 void Compile::add_expensive_node(Node * n) { 4177 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4178 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4179 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4180 if (OptimizeExpensiveOps) { 4181 _expensive_nodes->append(n); 4182 } else { 4183 // Clear control input and let IGVN optimize expensive nodes if 4184 // OptimizeExpensiveOps is off. 4185 n->set_req(0, NULL); 4186 } 4187 } 4188 4189 /** 4190 * Remove the speculative part of types and clean up the graph 4191 */ 4192 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4193 if (UseTypeSpeculation) { 4194 Unique_Node_List worklist; 4195 worklist.push(root()); 4196 int modified = 0; 4197 // Go over all type nodes that carry a speculative type, drop the 4198 // speculative part of the type and enqueue the node for an igvn 4199 // which may optimize it out. 4200 for (uint next = 0; next < worklist.size(); ++next) { 4201 Node *n = worklist.at(next); 4202 if (n->is_Type()) { 4203 TypeNode* tn = n->as_Type(); 4204 const Type* t = tn->type(); 4205 const Type* t_no_spec = t->remove_speculative(); 4206 if (t_no_spec != t) { 4207 bool in_hash = igvn.hash_delete(n); 4208 assert(in_hash, "node should be in igvn hash table"); 4209 tn->set_type(t_no_spec); 4210 igvn.hash_insert(n); 4211 igvn._worklist.push(n); // give it a chance to go away 4212 modified++; 4213 } 4214 } 4215 uint max = n->len(); 4216 for( uint i = 0; i < max; ++i ) { 4217 Node *m = n->in(i); 4218 if (not_a_node(m)) continue; 4219 worklist.push(m); 4220 } 4221 } 4222 // Drop the speculative part of all types in the igvn's type table 4223 igvn.remove_speculative_types(); 4224 if (modified > 0) { 4225 igvn.optimize(); 4226 } 4227 #ifdef ASSERT 4228 // Verify that after the IGVN is over no speculative type has resurfaced 4229 worklist.clear(); 4230 worklist.push(root()); 4231 for (uint next = 0; next < worklist.size(); ++next) { 4232 Node *n = worklist.at(next); 4233 const Type* t = igvn.type_or_null(n); 4234 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4235 if (n->is_Type()) { 4236 t = n->as_Type()->type(); 4237 assert(t == t->remove_speculative(), "no more speculative types"); 4238 } 4239 uint max = n->len(); 4240 for( uint i = 0; i < max; ++i ) { 4241 Node *m = n->in(i); 4242 if (not_a_node(m)) continue; 4243 worklist.push(m); 4244 } 4245 } 4246 igvn.check_no_speculative_types(); 4247 #endif 4248 } 4249 } 4250 4251 // Auxiliary method to support randomized stressing/fuzzing. 4252 // 4253 // This method can be called the arbitrary number of times, with current count 4254 // as the argument. The logic allows selecting a single candidate from the 4255 // running list of candidates as follows: 4256 // int count = 0; 4257 // Cand* selected = null; 4258 // while(cand = cand->next()) { 4259 // if (randomized_select(++count)) { 4260 // selected = cand; 4261 // } 4262 // } 4263 // 4264 // Including count equalizes the chances any candidate is "selected". 4265 // This is useful when we don't have the complete list of candidates to choose 4266 // from uniformly. In this case, we need to adjust the randomicity of the 4267 // selection, or else we will end up biasing the selection towards the latter 4268 // candidates. 4269 // 4270 // Quick back-envelope calculation shows that for the list of n candidates 4271 // the equal probability for the candidate to persist as "best" can be 4272 // achieved by replacing it with "next" k-th candidate with the probability 4273 // of 1/k. It can be easily shown that by the end of the run, the 4274 // probability for any candidate is converged to 1/n, thus giving the 4275 // uniform distribution among all the candidates. 4276 // 4277 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4278 #define RANDOMIZED_DOMAIN_POW 29 4279 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4280 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4281 bool Compile::randomized_select(int count) { 4282 assert(count > 0, "only positive"); 4283 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4284 }