1 /* 2 * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/block.hpp" 38 #include "opto/c2compiler.hpp" 39 #include "opto/callGenerator.hpp" 40 #include "opto/callnode.hpp" 41 #include "opto/cfgnode.hpp" 42 #include "opto/chaitin.hpp" 43 #include "opto/compile.hpp" 44 #include "opto/connode.hpp" 45 #include "opto/divnode.hpp" 46 #include "opto/escape.hpp" 47 #include "opto/idealGraphPrinter.hpp" 48 #include "opto/loopnode.hpp" 49 #include "opto/machnode.hpp" 50 #include "opto/macro.hpp" 51 #include "opto/matcher.hpp" 52 #include "opto/mathexactnode.hpp" 53 #include "opto/memnode.hpp" 54 #include "opto/mulnode.hpp" 55 #include "opto/narrowptrnode.hpp" 56 #include "opto/node.hpp" 57 #include "opto/opcodes.hpp" 58 #include "opto/output.hpp" 59 #include "opto/parse.hpp" 60 #include "opto/phaseX.hpp" 61 #include "opto/rootnode.hpp" 62 #include "opto/runtime.hpp" 63 #include "opto/stringopts.hpp" 64 #include "opto/type.hpp" 65 #include "opto/vectornode.hpp" 66 #include "runtime/arguments.hpp" 67 #include "runtime/signature.hpp" 68 #include "runtime/stubRoutines.hpp" 69 #include "runtime/timer.hpp" 70 #include "utilities/copy.hpp" 71 72 73 // -------------------- Compile::mach_constant_base_node ----------------------- 74 // Constant table base node singleton. 75 MachConstantBaseNode* Compile::mach_constant_base_node() { 76 if (_mach_constant_base_node == NULL) { 77 _mach_constant_base_node = new MachConstantBaseNode(); 78 _mach_constant_base_node->add_req(C->root()); 79 } 80 return _mach_constant_base_node; 81 } 82 83 84 /// Support for intrinsics. 85 86 // Return the index at which m must be inserted (or already exists). 87 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 88 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 89 #ifdef ASSERT 90 for (int i = 1; i < _intrinsics->length(); i++) { 91 CallGenerator* cg1 = _intrinsics->at(i-1); 92 CallGenerator* cg2 = _intrinsics->at(i); 93 assert(cg1->method() != cg2->method() 94 ? cg1->method() < cg2->method() 95 : cg1->is_virtual() < cg2->is_virtual(), 96 "compiler intrinsics list must stay sorted"); 97 } 98 #endif 99 // Binary search sorted list, in decreasing intervals [lo, hi]. 100 int lo = 0, hi = _intrinsics->length()-1; 101 while (lo <= hi) { 102 int mid = (uint)(hi + lo) / 2; 103 ciMethod* mid_m = _intrinsics->at(mid)->method(); 104 if (m < mid_m) { 105 hi = mid-1; 106 } else if (m > mid_m) { 107 lo = mid+1; 108 } else { 109 // look at minor sort key 110 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 111 if (is_virtual < mid_virt) { 112 hi = mid-1; 113 } else if (is_virtual > mid_virt) { 114 lo = mid+1; 115 } else { 116 return mid; // exact match 117 } 118 } 119 } 120 return lo; // inexact match 121 } 122 123 void Compile::register_intrinsic(CallGenerator* cg) { 124 if (_intrinsics == NULL) { 125 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 126 } 127 // This code is stolen from ciObjectFactory::insert. 128 // Really, GrowableArray should have methods for 129 // insert_at, remove_at, and binary_search. 130 int len = _intrinsics->length(); 131 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 132 if (index == len) { 133 _intrinsics->append(cg); 134 } else { 135 #ifdef ASSERT 136 CallGenerator* oldcg = _intrinsics->at(index); 137 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 138 #endif 139 _intrinsics->append(_intrinsics->at(len-1)); 140 int pos; 141 for (pos = len-2; pos >= index; pos--) { 142 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 143 } 144 _intrinsics->at_put(index, cg); 145 } 146 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 147 } 148 149 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 150 assert(m->is_loaded(), "don't try this on unloaded methods"); 151 if (_intrinsics != NULL) { 152 int index = intrinsic_insertion_index(m, is_virtual); 153 if (index < _intrinsics->length() 154 && _intrinsics->at(index)->method() == m 155 && _intrinsics->at(index)->is_virtual() == is_virtual) { 156 return _intrinsics->at(index); 157 } 158 } 159 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 160 if (m->intrinsic_id() != vmIntrinsics::_none && 161 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 162 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 163 if (cg != NULL) { 164 // Save it for next time: 165 register_intrinsic(cg); 166 return cg; 167 } else { 168 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 169 } 170 } 171 return NULL; 172 } 173 174 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 175 // in library_call.cpp. 176 177 178 #ifndef PRODUCT 179 // statistics gathering... 180 181 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 182 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 183 184 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 185 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 186 int oflags = _intrinsic_hist_flags[id]; 187 assert(flags != 0, "what happened?"); 188 if (is_virtual) { 189 flags |= _intrinsic_virtual; 190 } 191 bool changed = (flags != oflags); 192 if ((flags & _intrinsic_worked) != 0) { 193 juint count = (_intrinsic_hist_count[id] += 1); 194 if (count == 1) { 195 changed = true; // first time 196 } 197 // increment the overall count also: 198 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 199 } 200 if (changed) { 201 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 202 // Something changed about the intrinsic's virtuality. 203 if ((flags & _intrinsic_virtual) != 0) { 204 // This is the first use of this intrinsic as a virtual call. 205 if (oflags != 0) { 206 // We already saw it as a non-virtual, so note both cases. 207 flags |= _intrinsic_both; 208 } 209 } else if ((oflags & _intrinsic_both) == 0) { 210 // This is the first use of this intrinsic as a non-virtual 211 flags |= _intrinsic_both; 212 } 213 } 214 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 215 } 216 // update the overall flags also: 217 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 218 return changed; 219 } 220 221 static char* format_flags(int flags, char* buf) { 222 buf[0] = 0; 223 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 224 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 225 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 226 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 227 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 228 if (buf[0] == 0) strcat(buf, ","); 229 assert(buf[0] == ',', "must be"); 230 return &buf[1]; 231 } 232 233 void Compile::print_intrinsic_statistics() { 234 char flagsbuf[100]; 235 ttyLocker ttyl; 236 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 237 tty->print_cr("Compiler intrinsic usage:"); 238 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 239 if (total == 0) total = 1; // avoid div0 in case of no successes 240 #define PRINT_STAT_LINE(name, c, f) \ 241 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 242 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 243 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 244 int flags = _intrinsic_hist_flags[id]; 245 juint count = _intrinsic_hist_count[id]; 246 if ((flags | count) != 0) { 247 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 248 } 249 } 250 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 251 if (xtty != NULL) xtty->tail("statistics"); 252 } 253 254 void Compile::print_statistics() { 255 { ttyLocker ttyl; 256 if (xtty != NULL) xtty->head("statistics type='opto'"); 257 Parse::print_statistics(); 258 PhaseCCP::print_statistics(); 259 PhaseRegAlloc::print_statistics(); 260 Scheduling::print_statistics(); 261 PhasePeephole::print_statistics(); 262 PhaseIdealLoop::print_statistics(); 263 if (xtty != NULL) xtty->tail("statistics"); 264 } 265 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 266 // put this under its own <statistics> element. 267 print_intrinsic_statistics(); 268 } 269 } 270 #endif //PRODUCT 271 272 // Support for bundling info 273 Bundle* Compile::node_bundling(const Node *n) { 274 assert(valid_bundle_info(n), "oob"); 275 return &_node_bundling_base[n->_idx]; 276 } 277 278 bool Compile::valid_bundle_info(const Node *n) { 279 return (_node_bundling_limit > n->_idx); 280 } 281 282 283 void Compile::gvn_replace_by(Node* n, Node* nn) { 284 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 285 Node* use = n->last_out(i); 286 bool is_in_table = initial_gvn()->hash_delete(use); 287 uint uses_found = 0; 288 for (uint j = 0; j < use->len(); j++) { 289 if (use->in(j) == n) { 290 if (j < use->req()) 291 use->set_req(j, nn); 292 else 293 use->set_prec(j, nn); 294 uses_found++; 295 } 296 } 297 if (is_in_table) { 298 // reinsert into table 299 initial_gvn()->hash_find_insert(use); 300 } 301 record_for_igvn(use); 302 i -= uses_found; // we deleted 1 or more copies of this edge 303 } 304 } 305 306 307 static inline bool not_a_node(const Node* n) { 308 if (n == NULL) return true; 309 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 310 if (*(address*)n == badAddress) return true; // kill by Node::destruct 311 return false; 312 } 313 314 // Identify all nodes that are reachable from below, useful. 315 // Use breadth-first pass that records state in a Unique_Node_List, 316 // recursive traversal is slower. 317 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 318 int estimated_worklist_size = unique(); 319 useful.map( estimated_worklist_size, NULL ); // preallocate space 320 321 // Initialize worklist 322 if (root() != NULL) { useful.push(root()); } 323 // If 'top' is cached, declare it useful to preserve cached node 324 if( cached_top_node() ) { useful.push(cached_top_node()); } 325 326 // Push all useful nodes onto the list, breadthfirst 327 for( uint next = 0; next < useful.size(); ++next ) { 328 assert( next < unique(), "Unique useful nodes < total nodes"); 329 Node *n = useful.at(next); 330 uint max = n->len(); 331 for( uint i = 0; i < max; ++i ) { 332 Node *m = n->in(i); 333 if (not_a_node(m)) continue; 334 useful.push(m); 335 } 336 } 337 } 338 339 // Update dead_node_list with any missing dead nodes using useful 340 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 341 void Compile::update_dead_node_list(Unique_Node_List &useful) { 342 uint max_idx = unique(); 343 VectorSet& useful_node_set = useful.member_set(); 344 345 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 346 // If node with index node_idx is not in useful set, 347 // mark it as dead in dead node list. 348 if (! useful_node_set.test(node_idx) ) { 349 record_dead_node(node_idx); 350 } 351 } 352 } 353 354 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 355 int shift = 0; 356 for (int i = 0; i < inlines->length(); i++) { 357 CallGenerator* cg = inlines->at(i); 358 CallNode* call = cg->call_node(); 359 if (shift > 0) { 360 inlines->at_put(i-shift, cg); 361 } 362 if (!useful.member(call)) { 363 shift++; 364 } 365 } 366 inlines->trunc_to(inlines->length()-shift); 367 } 368 369 // Disconnect all useless nodes by disconnecting those at the boundary. 370 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 371 uint next = 0; 372 while (next < useful.size()) { 373 Node *n = useful.at(next++); 374 if (n->is_SafePoint()) { 375 // We're done with a parsing phase. Replaced nodes are not valid 376 // beyond that point. 377 n->as_SafePoint()->delete_replaced_nodes(); 378 } 379 // Use raw traversal of out edges since this code removes out edges 380 int max = n->outcnt(); 381 for (int j = 0; j < max; ++j) { 382 Node* child = n->raw_out(j); 383 if (! useful.member(child)) { 384 assert(!child->is_top() || child != top(), 385 "If top is cached in Compile object it is in useful list"); 386 // Only need to remove this out-edge to the useless node 387 n->raw_del_out(j); 388 --j; 389 --max; 390 } 391 } 392 if (n->outcnt() == 1 && n->has_special_unique_user()) { 393 record_for_igvn(n->unique_out()); 394 } 395 } 396 // Remove useless macro and predicate opaq nodes 397 for (int i = C->macro_count()-1; i >= 0; i--) { 398 Node* n = C->macro_node(i); 399 if (!useful.member(n)) { 400 remove_macro_node(n); 401 } 402 } 403 // Remove useless expensive node 404 for (int i = C->expensive_count()-1; i >= 0; i--) { 405 Node* n = C->expensive_node(i); 406 if (!useful.member(n)) { 407 remove_expensive_node(n); 408 } 409 } 410 // clean up the late inline lists 411 remove_useless_late_inlines(&_string_late_inlines, useful); 412 remove_useless_late_inlines(&_boxing_late_inlines, useful); 413 remove_useless_late_inlines(&_late_inlines, useful); 414 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 415 } 416 417 //------------------------------frame_size_in_words----------------------------- 418 // frame_slots in units of words 419 int Compile::frame_size_in_words() const { 420 // shift is 0 in LP32 and 1 in LP64 421 const int shift = (LogBytesPerWord - LogBytesPerInt); 422 int words = _frame_slots >> shift; 423 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 424 return words; 425 } 426 427 // To bang the stack of this compiled method we use the stack size 428 // that the interpreter would need in case of a deoptimization. This 429 // removes the need to bang the stack in the deoptimization blob which 430 // in turn simplifies stack overflow handling. 431 int Compile::bang_size_in_bytes() const { 432 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 433 } 434 435 // ============================================================================ 436 //------------------------------CompileWrapper--------------------------------- 437 class CompileWrapper : public StackObj { 438 Compile *const _compile; 439 public: 440 CompileWrapper(Compile* compile); 441 442 ~CompileWrapper(); 443 }; 444 445 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 446 // the Compile* pointer is stored in the current ciEnv: 447 ciEnv* env = compile->env(); 448 assert(env == ciEnv::current(), "must already be a ciEnv active"); 449 assert(env->compiler_data() == NULL, "compile already active?"); 450 env->set_compiler_data(compile); 451 assert(compile == Compile::current(), "sanity"); 452 453 compile->set_type_dict(NULL); 454 compile->set_type_hwm(NULL); 455 compile->set_type_last_size(0); 456 compile->set_last_tf(NULL, NULL); 457 compile->set_indexSet_arena(NULL); 458 compile->set_indexSet_free_block_list(NULL); 459 compile->init_type_arena(); 460 Type::Initialize(compile); 461 _compile->set_scratch_buffer_blob(NULL); 462 _compile->begin_method(); 463 } 464 CompileWrapper::~CompileWrapper() { 465 _compile->end_method(); 466 if (_compile->scratch_buffer_blob() != NULL) 467 BufferBlob::free(_compile->scratch_buffer_blob()); 468 _compile->env()->set_compiler_data(NULL); 469 } 470 471 472 //----------------------------print_compile_messages--------------------------- 473 void Compile::print_compile_messages() { 474 #ifndef PRODUCT 475 // Check if recompiling 476 if (_subsume_loads == false && PrintOpto) { 477 // Recompiling without allowing machine instructions to subsume loads 478 tty->print_cr("*********************************************************"); 479 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 480 tty->print_cr("*********************************************************"); 481 } 482 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 483 // Recompiling without escape analysis 484 tty->print_cr("*********************************************************"); 485 tty->print_cr("** Bailout: Recompile without escape analysis **"); 486 tty->print_cr("*********************************************************"); 487 } 488 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 489 // Recompiling without boxing elimination 490 tty->print_cr("*********************************************************"); 491 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 492 tty->print_cr("*********************************************************"); 493 } 494 if (env()->break_at_compile()) { 495 // Open the debugger when compiling this method. 496 tty->print("### Breaking when compiling: "); 497 method()->print_short_name(); 498 tty->cr(); 499 BREAKPOINT; 500 } 501 502 if( PrintOpto ) { 503 if (is_osr_compilation()) { 504 tty->print("[OSR]%3d", _compile_id); 505 } else { 506 tty->print("%3d", _compile_id); 507 } 508 } 509 #endif 510 } 511 512 513 //-----------------------init_scratch_buffer_blob------------------------------ 514 // Construct a temporary BufferBlob and cache it for this compile. 515 void Compile::init_scratch_buffer_blob(int const_size) { 516 // If there is already a scratch buffer blob allocated and the 517 // constant section is big enough, use it. Otherwise free the 518 // current and allocate a new one. 519 BufferBlob* blob = scratch_buffer_blob(); 520 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 521 // Use the current blob. 522 } else { 523 if (blob != NULL) { 524 BufferBlob::free(blob); 525 } 526 527 ResourceMark rm; 528 _scratch_const_size = const_size; 529 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 530 blob = BufferBlob::create("Compile::scratch_buffer", size); 531 // Record the buffer blob for next time. 532 set_scratch_buffer_blob(blob); 533 // Have we run out of code space? 534 if (scratch_buffer_blob() == NULL) { 535 // Let CompilerBroker disable further compilations. 536 record_failure("Not enough space for scratch buffer in CodeCache"); 537 return; 538 } 539 } 540 541 // Initialize the relocation buffers 542 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 543 set_scratch_locs_memory(locs_buf); 544 } 545 546 547 //-----------------------scratch_emit_size------------------------------------- 548 // Helper function that computes size by emitting code 549 uint Compile::scratch_emit_size(const Node* n) { 550 // Start scratch_emit_size section. 551 set_in_scratch_emit_size(true); 552 553 // Emit into a trash buffer and count bytes emitted. 554 // This is a pretty expensive way to compute a size, 555 // but it works well enough if seldom used. 556 // All common fixed-size instructions are given a size 557 // method by the AD file. 558 // Note that the scratch buffer blob and locs memory are 559 // allocated at the beginning of the compile task, and 560 // may be shared by several calls to scratch_emit_size. 561 // The allocation of the scratch buffer blob is particularly 562 // expensive, since it has to grab the code cache lock. 563 BufferBlob* blob = this->scratch_buffer_blob(); 564 assert(blob != NULL, "Initialize BufferBlob at start"); 565 assert(blob->size() > MAX_inst_size, "sanity"); 566 relocInfo* locs_buf = scratch_locs_memory(); 567 address blob_begin = blob->content_begin(); 568 address blob_end = (address)locs_buf; 569 assert(blob->content_contains(blob_end), "sanity"); 570 CodeBuffer buf(blob_begin, blob_end - blob_begin); 571 buf.initialize_consts_size(_scratch_const_size); 572 buf.initialize_stubs_size(MAX_stubs_size); 573 assert(locs_buf != NULL, "sanity"); 574 int lsize = MAX_locs_size / 3; 575 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 576 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 577 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 578 579 // Do the emission. 580 581 Label fakeL; // Fake label for branch instructions. 582 Label* saveL = NULL; 583 uint save_bnum = 0; 584 bool is_branch = n->is_MachBranch(); 585 if (is_branch) { 586 MacroAssembler masm(&buf); 587 masm.bind(fakeL); 588 n->as_MachBranch()->save_label(&saveL, &save_bnum); 589 n->as_MachBranch()->label_set(&fakeL, 0); 590 } 591 n->emit(buf, this->regalloc()); 592 if (is_branch) // Restore label. 593 n->as_MachBranch()->label_set(saveL, save_bnum); 594 595 // End scratch_emit_size section. 596 set_in_scratch_emit_size(false); 597 598 return buf.insts_size(); 599 } 600 601 602 // ============================================================================ 603 //------------------------------Compile standard------------------------------- 604 debug_only( int Compile::_debug_idx = 100000; ) 605 606 // Compile a method. entry_bci is -1 for normal compilations and indicates 607 // the continuation bci for on stack replacement. 608 609 610 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 611 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 612 : Phase(Compiler), 613 _env(ci_env), 614 _log(ci_env->log()), 615 _compile_id(ci_env->compile_id()), 616 _save_argument_registers(false), 617 _stub_name(NULL), 618 _stub_function(NULL), 619 _stub_entry_point(NULL), 620 _method(target), 621 _entry_bci(osr_bci), 622 _initial_gvn(NULL), 623 _for_igvn(NULL), 624 _warm_calls(NULL), 625 _subsume_loads(subsume_loads), 626 _do_escape_analysis(do_escape_analysis), 627 _eliminate_boxing(eliminate_boxing), 628 _failure_reason(NULL), 629 _code_buffer("Compile::Fill_buffer"), 630 _orig_pc_slot(0), 631 _orig_pc_slot_offset_in_bytes(0), 632 _has_method_handle_invokes(false), 633 _mach_constant_base_node(NULL), 634 _node_bundling_limit(0), 635 _node_bundling_base(NULL), 636 _java_calls(0), 637 _inner_loops(0), 638 _scratch_const_size(-1), 639 _in_scratch_emit_size(false), 640 _dead_node_list(comp_arena()), 641 _dead_node_count(0), 642 #ifndef PRODUCT 643 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 644 _in_dump_cnt(0), 645 _printer(IdealGraphPrinter::printer()), 646 #endif 647 _congraph(NULL), 648 _comp_arena(mtCompiler), 649 _node_arena(mtCompiler), 650 _old_arena(mtCompiler), 651 _Compile_types(mtCompiler), 652 _replay_inline_data(NULL), 653 _late_inlines(comp_arena(), 2, 0, NULL), 654 _string_late_inlines(comp_arena(), 2, 0, NULL), 655 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 656 _late_inlines_pos(0), 657 _number_of_mh_late_inlines(0), 658 _inlining_progress(false), 659 _inlining_incrementally(false), 660 _print_inlining_list(NULL), 661 _print_inlining_stream(NULL), 662 _print_inlining_idx(0), 663 _print_inlining_output(NULL), 664 _interpreter_frame_size(0), 665 _max_node_limit(MaxNodeLimit) { 666 C = this; 667 668 CompileWrapper cw(this); 669 670 if (CITimeVerbose) { 671 tty->print(" "); 672 target->holder()->name()->print(); 673 tty->print("."); 674 target->print_short_name(); 675 tty->print(" "); 676 } 677 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 678 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 679 680 #ifndef PRODUCT 681 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 682 if (!print_opto_assembly) { 683 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 684 if (print_assembly && !Disassembler::can_decode()) { 685 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 686 print_opto_assembly = true; 687 } 688 } 689 set_print_assembly(print_opto_assembly); 690 set_parsed_irreducible_loop(false); 691 692 if (method()->has_option("ReplayInline")) { 693 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 694 } 695 #endif 696 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 697 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 698 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 699 700 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 701 // Make sure the method being compiled gets its own MDO, 702 // so we can at least track the decompile_count(). 703 // Need MDO to record RTM code generation state. 704 method()->ensure_method_data(); 705 } 706 707 Init(::AliasLevel); 708 709 710 print_compile_messages(); 711 712 _ilt = InlineTree::build_inline_tree_root(); 713 714 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 715 assert(num_alias_types() >= AliasIdxRaw, ""); 716 717 #define MINIMUM_NODE_HASH 1023 718 // Node list that Iterative GVN will start with 719 Unique_Node_List for_igvn(comp_arena()); 720 set_for_igvn(&for_igvn); 721 722 // GVN that will be run immediately on new nodes 723 uint estimated_size = method()->code_size()*4+64; 724 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 725 PhaseGVN gvn(node_arena(), estimated_size); 726 set_initial_gvn(&gvn); 727 728 print_inlining_init(); 729 { // Scope for timing the parser 730 TracePhase tp("parse", &timers[_t_parser]); 731 732 // Put top into the hash table ASAP. 733 initial_gvn()->transform_no_reclaim(top()); 734 735 // Set up tf(), start(), and find a CallGenerator. 736 CallGenerator* cg = NULL; 737 if (is_osr_compilation()) { 738 const TypeTuple *domain = StartOSRNode::osr_domain(); 739 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 740 init_tf(TypeFunc::make(domain, range)); 741 StartNode* s = new StartOSRNode(root(), domain); 742 initial_gvn()->set_type_bottom(s); 743 init_start(s); 744 cg = CallGenerator::for_osr(method(), entry_bci()); 745 } else { 746 // Normal case. 747 init_tf(TypeFunc::make(method())); 748 StartNode* s = new StartNode(root(), tf()->domain()); 749 initial_gvn()->set_type_bottom(s); 750 init_start(s); 751 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 752 // With java.lang.ref.reference.get() we must go through the 753 // intrinsic when G1 is enabled - even when get() is the root 754 // method of the compile - so that, if necessary, the value in 755 // the referent field of the reference object gets recorded by 756 // the pre-barrier code. 757 // Specifically, if G1 is enabled, the value in the referent 758 // field is recorded by the G1 SATB pre barrier. This will 759 // result in the referent being marked live and the reference 760 // object removed from the list of discovered references during 761 // reference processing. 762 cg = find_intrinsic(method(), false); 763 } 764 if (cg == NULL) { 765 float past_uses = method()->interpreter_invocation_count(); 766 float expected_uses = past_uses; 767 cg = CallGenerator::for_inline(method(), expected_uses); 768 } 769 } 770 if (failing()) return; 771 if (cg == NULL) { 772 record_method_not_compilable_all_tiers("cannot parse method"); 773 return; 774 } 775 JVMState* jvms = build_start_state(start(), tf()); 776 if ((jvms = cg->generate(jvms)) == NULL) { 777 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 778 record_method_not_compilable("method parse failed"); 779 } 780 return; 781 } 782 GraphKit kit(jvms); 783 784 if (!kit.stopped()) { 785 // Accept return values, and transfer control we know not where. 786 // This is done by a special, unique ReturnNode bound to root. 787 return_values(kit.jvms()); 788 } 789 790 if (kit.has_exceptions()) { 791 // Any exceptions that escape from this call must be rethrown 792 // to whatever caller is dynamically above us on the stack. 793 // This is done by a special, unique RethrowNode bound to root. 794 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 795 } 796 797 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 798 799 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 800 inline_string_calls(true); 801 } 802 803 if (failing()) return; 804 805 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 806 807 // Remove clutter produced by parsing. 808 if (!failing()) { 809 ResourceMark rm; 810 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 811 } 812 } 813 814 // Note: Large methods are capped off in do_one_bytecode(). 815 if (failing()) return; 816 817 // After parsing, node notes are no longer automagic. 818 // They must be propagated by register_new_node_with_optimizer(), 819 // clone(), or the like. 820 set_default_node_notes(NULL); 821 822 for (;;) { 823 int successes = Inline_Warm(); 824 if (failing()) return; 825 if (successes == 0) break; 826 } 827 828 // Drain the list. 829 Finish_Warm(); 830 #ifndef PRODUCT 831 if (_printer && _printer->should_print(_method)) { 832 _printer->print_inlining(this); 833 } 834 #endif 835 836 if (failing()) return; 837 NOT_PRODUCT( verify_graph_edges(); ) 838 839 // Now optimize 840 Optimize(); 841 if (failing()) return; 842 NOT_PRODUCT( verify_graph_edges(); ) 843 844 #ifndef PRODUCT 845 if (PrintIdeal) { 846 ttyLocker ttyl; // keep the following output all in one block 847 // This output goes directly to the tty, not the compiler log. 848 // To enable tools to match it up with the compilation activity, 849 // be sure to tag this tty output with the compile ID. 850 if (xtty != NULL) { 851 xtty->head("ideal compile_id='%d'%s", compile_id(), 852 is_osr_compilation() ? " compile_kind='osr'" : 853 ""); 854 } 855 root()->dump(9999); 856 if (xtty != NULL) { 857 xtty->tail("ideal"); 858 } 859 } 860 #endif 861 862 NOT_PRODUCT( verify_barriers(); ) 863 864 // Dump compilation data to replay it. 865 if (method()->has_option("DumpReplay")) { 866 env()->dump_replay_data(_compile_id); 867 } 868 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 869 env()->dump_inline_data(_compile_id); 870 } 871 872 // Now that we know the size of all the monitors we can add a fixed slot 873 // for the original deopt pc. 874 875 _orig_pc_slot = fixed_slots(); 876 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 877 set_fixed_slots(next_slot); 878 879 // Compute when to use implicit null checks. Used by matching trap based 880 // nodes and NullCheck optimization. 881 set_allowed_deopt_reasons(); 882 883 // Now generate code 884 Code_Gen(); 885 if (failing()) return; 886 887 // Check if we want to skip execution of all compiled code. 888 { 889 #ifndef PRODUCT 890 if (OptoNoExecute) { 891 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 892 return; 893 } 894 #endif 895 TracePhase tp("install_code", &timers[_t_registerMethod]); 896 897 if (is_osr_compilation()) { 898 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 899 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 900 } else { 901 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 902 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 903 } 904 905 env()->register_method(_method, _entry_bci, 906 &_code_offsets, 907 _orig_pc_slot_offset_in_bytes, 908 code_buffer(), 909 frame_size_in_words(), _oop_map_set, 910 &_handler_table, &_inc_table, 911 compiler, 912 env()->comp_level(), 913 has_unsafe_access(), 914 SharedRuntime::is_wide_vector(max_vector_size()), 915 rtm_state() 916 ); 917 918 if (log() != NULL) // Print code cache state into compiler log 919 log()->code_cache_state(); 920 } 921 } 922 923 //------------------------------Compile---------------------------------------- 924 // Compile a runtime stub 925 Compile::Compile( ciEnv* ci_env, 926 TypeFunc_generator generator, 927 address stub_function, 928 const char *stub_name, 929 int is_fancy_jump, 930 bool pass_tls, 931 bool save_arg_registers, 932 bool return_pc ) 933 : Phase(Compiler), 934 _env(ci_env), 935 _log(ci_env->log()), 936 _compile_id(0), 937 _save_argument_registers(save_arg_registers), 938 _method(NULL), 939 _stub_name(stub_name), 940 _stub_function(stub_function), 941 _stub_entry_point(NULL), 942 _entry_bci(InvocationEntryBci), 943 _initial_gvn(NULL), 944 _for_igvn(NULL), 945 _warm_calls(NULL), 946 _orig_pc_slot(0), 947 _orig_pc_slot_offset_in_bytes(0), 948 _subsume_loads(true), 949 _do_escape_analysis(false), 950 _eliminate_boxing(false), 951 _failure_reason(NULL), 952 _code_buffer("Compile::Fill_buffer"), 953 _has_method_handle_invokes(false), 954 _mach_constant_base_node(NULL), 955 _node_bundling_limit(0), 956 _node_bundling_base(NULL), 957 _java_calls(0), 958 _inner_loops(0), 959 #ifndef PRODUCT 960 _trace_opto_output(TraceOptoOutput), 961 _in_dump_cnt(0), 962 _printer(NULL), 963 #endif 964 _comp_arena(mtCompiler), 965 _node_arena(mtCompiler), 966 _old_arena(mtCompiler), 967 _Compile_types(mtCompiler), 968 _dead_node_list(comp_arena()), 969 _dead_node_count(0), 970 _congraph(NULL), 971 _replay_inline_data(NULL), 972 _number_of_mh_late_inlines(0), 973 _inlining_progress(false), 974 _inlining_incrementally(false), 975 _print_inlining_list(NULL), 976 _print_inlining_stream(NULL), 977 _print_inlining_idx(0), 978 _print_inlining_output(NULL), 979 _allowed_reasons(0), 980 _interpreter_frame_size(0), 981 _max_node_limit(MaxNodeLimit) { 982 C = this; 983 984 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 985 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 986 987 #ifndef PRODUCT 988 set_print_assembly(PrintFrameConverterAssembly); 989 set_parsed_irreducible_loop(false); 990 #endif 991 set_has_irreducible_loop(false); // no loops 992 993 CompileWrapper cw(this); 994 Init(/*AliasLevel=*/ 0); 995 init_tf((*generator)()); 996 997 { 998 // The following is a dummy for the sake of GraphKit::gen_stub 999 Unique_Node_List for_igvn(comp_arena()); 1000 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1001 PhaseGVN gvn(Thread::current()->resource_area(),255); 1002 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1003 gvn.transform_no_reclaim(top()); 1004 1005 GraphKit kit; 1006 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1007 } 1008 1009 NOT_PRODUCT( verify_graph_edges(); ) 1010 Code_Gen(); 1011 if (failing()) return; 1012 1013 1014 // Entry point will be accessed using compile->stub_entry_point(); 1015 if (code_buffer() == NULL) { 1016 Matcher::soft_match_failure(); 1017 } else { 1018 if (PrintAssembly && (WizardMode || Verbose)) 1019 tty->print_cr("### Stub::%s", stub_name); 1020 1021 if (!failing()) { 1022 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1023 1024 // Make the NMethod 1025 // For now we mark the frame as never safe for profile stackwalking 1026 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1027 code_buffer(), 1028 CodeOffsets::frame_never_safe, 1029 // _code_offsets.value(CodeOffsets::Frame_Complete), 1030 frame_size_in_words(), 1031 _oop_map_set, 1032 save_arg_registers); 1033 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1034 1035 _stub_entry_point = rs->entry_point(); 1036 } 1037 } 1038 } 1039 1040 //------------------------------Init------------------------------------------- 1041 // Prepare for a single compilation 1042 void Compile::Init(int aliaslevel) { 1043 _unique = 0; 1044 _regalloc = NULL; 1045 1046 _tf = NULL; // filled in later 1047 _top = NULL; // cached later 1048 _matcher = NULL; // filled in later 1049 _cfg = NULL; // filled in later 1050 1051 set_24_bit_selection_and_mode(Use24BitFP, false); 1052 1053 _node_note_array = NULL; 1054 _default_node_notes = NULL; 1055 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1056 1057 _immutable_memory = NULL; // filled in at first inquiry 1058 1059 // Globally visible Nodes 1060 // First set TOP to NULL to give safe behavior during creation of RootNode 1061 set_cached_top_node(NULL); 1062 set_root(new RootNode()); 1063 // Now that you have a Root to point to, create the real TOP 1064 set_cached_top_node( new ConNode(Type::TOP) ); 1065 set_recent_alloc(NULL, NULL); 1066 1067 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1068 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1069 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1070 env()->set_dependencies(new Dependencies(env())); 1071 1072 _fixed_slots = 0; 1073 set_has_split_ifs(false); 1074 set_has_loops(has_method() && method()->has_loops()); // first approximation 1075 set_has_stringbuilder(false); 1076 set_has_boxed_value(false); 1077 _trap_can_recompile = false; // no traps emitted yet 1078 _major_progress = true; // start out assuming good things will happen 1079 set_has_unsafe_access(false); 1080 set_max_vector_size(0); 1081 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1082 set_decompile_count(0); 1083 1084 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1085 set_num_loop_opts(LoopOptsCount); 1086 set_do_inlining(Inline); 1087 set_max_inline_size(MaxInlineSize); 1088 set_freq_inline_size(FreqInlineSize); 1089 set_do_scheduling(OptoScheduling); 1090 set_do_count_invocations(false); 1091 set_do_method_data_update(false); 1092 set_age_code(has_method() && method()->profile_aging()); 1093 set_rtm_state(NoRTM); // No RTM lock eliding by default 1094 method_has_option_value("MaxNodeLimit", _max_node_limit); 1095 #if INCLUDE_RTM_OPT 1096 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1097 int rtm_state = method()->method_data()->rtm_state(); 1098 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1099 // Don't generate RTM lock eliding code. 1100 set_rtm_state(NoRTM); 1101 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1102 // Generate RTM lock eliding code without abort ratio calculation code. 1103 set_rtm_state(UseRTM); 1104 } else if (UseRTMDeopt) { 1105 // Generate RTM lock eliding code and include abort ratio calculation 1106 // code if UseRTMDeopt is on. 1107 set_rtm_state(ProfileRTM); 1108 } 1109 } 1110 #endif 1111 if (debug_info()->recording_non_safepoints()) { 1112 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1113 (comp_arena(), 8, 0, NULL)); 1114 set_default_node_notes(Node_Notes::make(this)); 1115 } 1116 1117 // // -- Initialize types before each compile -- 1118 // // Update cached type information 1119 // if( _method && _method->constants() ) 1120 // Type::update_loaded_types(_method, _method->constants()); 1121 1122 // Init alias_type map. 1123 if (!_do_escape_analysis && aliaslevel == 3) 1124 aliaslevel = 2; // No unique types without escape analysis 1125 _AliasLevel = aliaslevel; 1126 const int grow_ats = 16; 1127 _max_alias_types = grow_ats; 1128 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1129 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1130 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1131 { 1132 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1133 } 1134 // Initialize the first few types. 1135 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1136 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1137 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1138 _num_alias_types = AliasIdxRaw+1; 1139 // Zero out the alias type cache. 1140 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1141 // A NULL adr_type hits in the cache right away. Preload the right answer. 1142 probe_alias_cache(NULL)->_index = AliasIdxTop; 1143 1144 _intrinsics = NULL; 1145 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1146 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1147 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1148 register_library_intrinsics(); 1149 } 1150 1151 //---------------------------init_start---------------------------------------- 1152 // Install the StartNode on this compile object. 1153 void Compile::init_start(StartNode* s) { 1154 if (failing()) 1155 return; // already failing 1156 assert(s == start(), ""); 1157 } 1158 1159 /** 1160 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1161 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1162 * the ideal graph. 1163 */ 1164 StartNode* Compile::start() const { 1165 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 1166 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1167 Node* start = root()->fast_out(i); 1168 if (start->is_Start()) { 1169 return start->as_Start(); 1170 } 1171 } 1172 fatal("Did not find Start node!"); 1173 return NULL; 1174 } 1175 1176 //-------------------------------immutable_memory------------------------------------- 1177 // Access immutable memory 1178 Node* Compile::immutable_memory() { 1179 if (_immutable_memory != NULL) { 1180 return _immutable_memory; 1181 } 1182 StartNode* s = start(); 1183 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1184 Node *p = s->fast_out(i); 1185 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1186 _immutable_memory = p; 1187 return _immutable_memory; 1188 } 1189 } 1190 ShouldNotReachHere(); 1191 return NULL; 1192 } 1193 1194 //----------------------set_cached_top_node------------------------------------ 1195 // Install the cached top node, and make sure Node::is_top works correctly. 1196 void Compile::set_cached_top_node(Node* tn) { 1197 if (tn != NULL) verify_top(tn); 1198 Node* old_top = _top; 1199 _top = tn; 1200 // Calling Node::setup_is_top allows the nodes the chance to adjust 1201 // their _out arrays. 1202 if (_top != NULL) _top->setup_is_top(); 1203 if (old_top != NULL) old_top->setup_is_top(); 1204 assert(_top == NULL || top()->is_top(), ""); 1205 } 1206 1207 #ifdef ASSERT 1208 uint Compile::count_live_nodes_by_graph_walk() { 1209 Unique_Node_List useful(comp_arena()); 1210 // Get useful node list by walking the graph. 1211 identify_useful_nodes(useful); 1212 return useful.size(); 1213 } 1214 1215 void Compile::print_missing_nodes() { 1216 1217 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1218 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1219 return; 1220 } 1221 1222 // This is an expensive function. It is executed only when the user 1223 // specifies VerifyIdealNodeCount option or otherwise knows the 1224 // additional work that needs to be done to identify reachable nodes 1225 // by walking the flow graph and find the missing ones using 1226 // _dead_node_list. 1227 1228 Unique_Node_List useful(comp_arena()); 1229 // Get useful node list by walking the graph. 1230 identify_useful_nodes(useful); 1231 1232 uint l_nodes = C->live_nodes(); 1233 uint l_nodes_by_walk = useful.size(); 1234 1235 if (l_nodes != l_nodes_by_walk) { 1236 if (_log != NULL) { 1237 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1238 _log->stamp(); 1239 _log->end_head(); 1240 } 1241 VectorSet& useful_member_set = useful.member_set(); 1242 int last_idx = l_nodes_by_walk; 1243 for (int i = 0; i < last_idx; i++) { 1244 if (useful_member_set.test(i)) { 1245 if (_dead_node_list.test(i)) { 1246 if (_log != NULL) { 1247 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1248 } 1249 if (PrintIdealNodeCount) { 1250 // Print the log message to tty 1251 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1252 useful.at(i)->dump(); 1253 } 1254 } 1255 } 1256 else if (! _dead_node_list.test(i)) { 1257 if (_log != NULL) { 1258 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1259 } 1260 if (PrintIdealNodeCount) { 1261 // Print the log message to tty 1262 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1263 } 1264 } 1265 } 1266 if (_log != NULL) { 1267 _log->tail("mismatched_nodes"); 1268 } 1269 } 1270 } 1271 void Compile::record_modified_node(Node* n) { 1272 if (_modified_nodes != NULL && !_inlining_incrementally && 1273 n->outcnt() != 0 && !n->is_Con()) { 1274 _modified_nodes->push(n); 1275 } 1276 } 1277 1278 void Compile::remove_modified_node(Node* n) { 1279 if (_modified_nodes != NULL) { 1280 _modified_nodes->remove(n); 1281 } 1282 } 1283 #endif 1284 1285 #ifndef PRODUCT 1286 void Compile::verify_top(Node* tn) const { 1287 if (tn != NULL) { 1288 assert(tn->is_Con(), "top node must be a constant"); 1289 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1290 assert(tn->in(0) != NULL, "must have live top node"); 1291 } 1292 } 1293 #endif 1294 1295 1296 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1297 1298 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1299 guarantee(arr != NULL, ""); 1300 int num_blocks = arr->length(); 1301 if (grow_by < num_blocks) grow_by = num_blocks; 1302 int num_notes = grow_by * _node_notes_block_size; 1303 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1304 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1305 while (num_notes > 0) { 1306 arr->append(notes); 1307 notes += _node_notes_block_size; 1308 num_notes -= _node_notes_block_size; 1309 } 1310 assert(num_notes == 0, "exact multiple, please"); 1311 } 1312 1313 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1314 if (source == NULL || dest == NULL) return false; 1315 1316 if (dest->is_Con()) 1317 return false; // Do not push debug info onto constants. 1318 1319 #ifdef ASSERT 1320 // Leave a bread crumb trail pointing to the original node: 1321 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1322 dest->set_debug_orig(source); 1323 } 1324 #endif 1325 1326 if (node_note_array() == NULL) 1327 return false; // Not collecting any notes now. 1328 1329 // This is a copy onto a pre-existing node, which may already have notes. 1330 // If both nodes have notes, do not overwrite any pre-existing notes. 1331 Node_Notes* source_notes = node_notes_at(source->_idx); 1332 if (source_notes == NULL || source_notes->is_clear()) return false; 1333 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1334 if (dest_notes == NULL || dest_notes->is_clear()) { 1335 return set_node_notes_at(dest->_idx, source_notes); 1336 } 1337 1338 Node_Notes merged_notes = (*source_notes); 1339 // The order of operations here ensures that dest notes will win... 1340 merged_notes.update_from(dest_notes); 1341 return set_node_notes_at(dest->_idx, &merged_notes); 1342 } 1343 1344 1345 //--------------------------allow_range_check_smearing------------------------- 1346 // Gating condition for coalescing similar range checks. 1347 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1348 // single covering check that is at least as strong as any of them. 1349 // If the optimization succeeds, the simplified (strengthened) range check 1350 // will always succeed. If it fails, we will deopt, and then give up 1351 // on the optimization. 1352 bool Compile::allow_range_check_smearing() const { 1353 // If this method has already thrown a range-check, 1354 // assume it was because we already tried range smearing 1355 // and it failed. 1356 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1357 return !already_trapped; 1358 } 1359 1360 1361 //------------------------------flatten_alias_type----------------------------- 1362 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1363 int offset = tj->offset(); 1364 TypePtr::PTR ptr = tj->ptr(); 1365 1366 // Known instance (scalarizable allocation) alias only with itself. 1367 bool is_known_inst = tj->isa_oopptr() != NULL && 1368 tj->is_oopptr()->is_known_instance(); 1369 1370 // Process weird unsafe references. 1371 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1372 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1373 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1374 tj = TypeOopPtr::BOTTOM; 1375 ptr = tj->ptr(); 1376 offset = tj->offset(); 1377 } 1378 1379 // Array pointers need some flattening 1380 const TypeAryPtr *ta = tj->isa_aryptr(); 1381 if (ta && ta->is_stable()) { 1382 // Erase stability property for alias analysis. 1383 tj = ta = ta->cast_to_stable(false); 1384 } 1385 if( ta && is_known_inst ) { 1386 if ( offset != Type::OffsetBot && 1387 offset > arrayOopDesc::length_offset_in_bytes() ) { 1388 offset = Type::OffsetBot; // Flatten constant access into array body only 1389 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1390 } 1391 } else if( ta && _AliasLevel >= 2 ) { 1392 // For arrays indexed by constant indices, we flatten the alias 1393 // space to include all of the array body. Only the header, klass 1394 // and array length can be accessed un-aliased. 1395 if( offset != Type::OffsetBot ) { 1396 if( ta->const_oop() ) { // MethodData* or Method* 1397 offset = Type::OffsetBot; // Flatten constant access into array body 1398 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1399 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1400 // range is OK as-is. 1401 tj = ta = TypeAryPtr::RANGE; 1402 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1403 tj = TypeInstPtr::KLASS; // all klass loads look alike 1404 ta = TypeAryPtr::RANGE; // generic ignored junk 1405 ptr = TypePtr::BotPTR; 1406 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1407 tj = TypeInstPtr::MARK; 1408 ta = TypeAryPtr::RANGE; // generic ignored junk 1409 ptr = TypePtr::BotPTR; 1410 } else { // Random constant offset into array body 1411 offset = Type::OffsetBot; // Flatten constant access into array body 1412 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1413 } 1414 } 1415 // Arrays of fixed size alias with arrays of unknown size. 1416 if (ta->size() != TypeInt::POS) { 1417 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1418 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1419 } 1420 // Arrays of known objects become arrays of unknown objects. 1421 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1422 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1423 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1424 } 1425 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1426 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1427 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1428 } 1429 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1430 // cannot be distinguished by bytecode alone. 1431 if (ta->elem() == TypeInt::BOOL) { 1432 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1433 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1434 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1435 } 1436 // During the 2nd round of IterGVN, NotNull castings are removed. 1437 // Make sure the Bottom and NotNull variants alias the same. 1438 // Also, make sure exact and non-exact variants alias the same. 1439 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1440 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1441 } 1442 } 1443 1444 // Oop pointers need some flattening 1445 const TypeInstPtr *to = tj->isa_instptr(); 1446 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1447 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1448 if( ptr == TypePtr::Constant ) { 1449 if (to->klass() != ciEnv::current()->Class_klass() || 1450 offset < k->size_helper() * wordSize) { 1451 // No constant oop pointers (such as Strings); they alias with 1452 // unknown strings. 1453 assert(!is_known_inst, "not scalarizable allocation"); 1454 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1455 } 1456 } else if( is_known_inst ) { 1457 tj = to; // Keep NotNull and klass_is_exact for instance type 1458 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1459 // During the 2nd round of IterGVN, NotNull castings are removed. 1460 // Make sure the Bottom and NotNull variants alias the same. 1461 // Also, make sure exact and non-exact variants alias the same. 1462 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1463 } 1464 if (to->speculative() != NULL) { 1465 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1466 } 1467 // Canonicalize the holder of this field 1468 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1469 // First handle header references such as a LoadKlassNode, even if the 1470 // object's klass is unloaded at compile time (4965979). 1471 if (!is_known_inst) { // Do it only for non-instance types 1472 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1473 } 1474 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1475 // Static fields are in the space above the normal instance 1476 // fields in the java.lang.Class instance. 1477 if (to->klass() != ciEnv::current()->Class_klass()) { 1478 to = NULL; 1479 tj = TypeOopPtr::BOTTOM; 1480 offset = tj->offset(); 1481 } 1482 } else { 1483 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1484 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1485 if( is_known_inst ) { 1486 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1487 } else { 1488 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1489 } 1490 } 1491 } 1492 } 1493 1494 // Klass pointers to object array klasses need some flattening 1495 const TypeKlassPtr *tk = tj->isa_klassptr(); 1496 if( tk ) { 1497 // If we are referencing a field within a Klass, we need 1498 // to assume the worst case of an Object. Both exact and 1499 // inexact types must flatten to the same alias class so 1500 // use NotNull as the PTR. 1501 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1502 1503 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1504 TypeKlassPtr::OBJECT->klass(), 1505 offset); 1506 } 1507 1508 ciKlass* klass = tk->klass(); 1509 if( klass->is_obj_array_klass() ) { 1510 ciKlass* k = TypeAryPtr::OOPS->klass(); 1511 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1512 k = TypeInstPtr::BOTTOM->klass(); 1513 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1514 } 1515 1516 // Check for precise loads from the primary supertype array and force them 1517 // to the supertype cache alias index. Check for generic array loads from 1518 // the primary supertype array and also force them to the supertype cache 1519 // alias index. Since the same load can reach both, we need to merge 1520 // these 2 disparate memories into the same alias class. Since the 1521 // primary supertype array is read-only, there's no chance of confusion 1522 // where we bypass an array load and an array store. 1523 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1524 if (offset == Type::OffsetBot || 1525 (offset >= primary_supers_offset && 1526 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1527 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1528 offset = in_bytes(Klass::secondary_super_cache_offset()); 1529 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1530 } 1531 } 1532 1533 // Flatten all Raw pointers together. 1534 if (tj->base() == Type::RawPtr) 1535 tj = TypeRawPtr::BOTTOM; 1536 1537 if (tj->base() == Type::AnyPtr) 1538 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1539 1540 // Flatten all to bottom for now 1541 switch( _AliasLevel ) { 1542 case 0: 1543 tj = TypePtr::BOTTOM; 1544 break; 1545 case 1: // Flatten to: oop, static, field or array 1546 switch (tj->base()) { 1547 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1548 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1549 case Type::AryPtr: // do not distinguish arrays at all 1550 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1551 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1552 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1553 default: ShouldNotReachHere(); 1554 } 1555 break; 1556 case 2: // No collapsing at level 2; keep all splits 1557 case 3: // No collapsing at level 3; keep all splits 1558 break; 1559 default: 1560 Unimplemented(); 1561 } 1562 1563 offset = tj->offset(); 1564 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1565 1566 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1567 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1568 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1569 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1570 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1571 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1572 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1573 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1574 assert( tj->ptr() != TypePtr::TopPTR && 1575 tj->ptr() != TypePtr::AnyNull && 1576 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1577 // assert( tj->ptr() != TypePtr::Constant || 1578 // tj->base() == Type::RawPtr || 1579 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1580 1581 return tj; 1582 } 1583 1584 void Compile::AliasType::Init(int i, const TypePtr* at) { 1585 _index = i; 1586 _adr_type = at; 1587 _field = NULL; 1588 _element = NULL; 1589 _is_rewritable = true; // default 1590 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1591 if (atoop != NULL && atoop->is_known_instance()) { 1592 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1593 _general_index = Compile::current()->get_alias_index(gt); 1594 } else { 1595 _general_index = 0; 1596 } 1597 } 1598 1599 //---------------------------------print_on------------------------------------ 1600 #ifndef PRODUCT 1601 void Compile::AliasType::print_on(outputStream* st) { 1602 if (index() < 10) 1603 st->print("@ <%d> ", index()); 1604 else st->print("@ <%d>", index()); 1605 st->print(is_rewritable() ? " " : " RO"); 1606 int offset = adr_type()->offset(); 1607 if (offset == Type::OffsetBot) 1608 st->print(" +any"); 1609 else st->print(" +%-3d", offset); 1610 st->print(" in "); 1611 adr_type()->dump_on(st); 1612 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1613 if (field() != NULL && tjp) { 1614 if (tjp->klass() != field()->holder() || 1615 tjp->offset() != field()->offset_in_bytes()) { 1616 st->print(" != "); 1617 field()->print(); 1618 st->print(" ***"); 1619 } 1620 } 1621 } 1622 1623 void print_alias_types() { 1624 Compile* C = Compile::current(); 1625 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1626 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1627 C->alias_type(idx)->print_on(tty); 1628 tty->cr(); 1629 } 1630 } 1631 #endif 1632 1633 1634 //----------------------------probe_alias_cache-------------------------------- 1635 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1636 intptr_t key = (intptr_t) adr_type; 1637 key ^= key >> logAliasCacheSize; 1638 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1639 } 1640 1641 1642 //-----------------------------grow_alias_types-------------------------------- 1643 void Compile::grow_alias_types() { 1644 const int old_ats = _max_alias_types; // how many before? 1645 const int new_ats = old_ats; // how many more? 1646 const int grow_ats = old_ats+new_ats; // how many now? 1647 _max_alias_types = grow_ats; 1648 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1649 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1650 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1651 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1652 } 1653 1654 1655 //--------------------------------find_alias_type------------------------------ 1656 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1657 if (_AliasLevel == 0) 1658 return alias_type(AliasIdxBot); 1659 1660 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1661 if (ace->_adr_type == adr_type) { 1662 return alias_type(ace->_index); 1663 } 1664 1665 // Handle special cases. 1666 if (adr_type == NULL) return alias_type(AliasIdxTop); 1667 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1668 1669 // Do it the slow way. 1670 const TypePtr* flat = flatten_alias_type(adr_type); 1671 1672 #ifdef ASSERT 1673 assert(flat == flatten_alias_type(flat), "idempotent"); 1674 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1675 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1676 const TypeOopPtr* foop = flat->is_oopptr(); 1677 // Scalarizable allocations have exact klass always. 1678 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1679 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1680 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1681 } 1682 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1683 #endif 1684 1685 int idx = AliasIdxTop; 1686 for (int i = 0; i < num_alias_types(); i++) { 1687 if (alias_type(i)->adr_type() == flat) { 1688 idx = i; 1689 break; 1690 } 1691 } 1692 1693 if (idx == AliasIdxTop) { 1694 if (no_create) return NULL; 1695 // Grow the array if necessary. 1696 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1697 // Add a new alias type. 1698 idx = _num_alias_types++; 1699 _alias_types[idx]->Init(idx, flat); 1700 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1701 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1702 if (flat->isa_instptr()) { 1703 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1704 && flat->is_instptr()->klass() == env()->Class_klass()) 1705 alias_type(idx)->set_rewritable(false); 1706 } 1707 if (flat->isa_aryptr()) { 1708 #ifdef ASSERT 1709 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1710 // (T_BYTE has the weakest alignment and size restrictions...) 1711 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1712 #endif 1713 if (flat->offset() == TypePtr::OffsetBot) { 1714 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1715 } 1716 } 1717 if (flat->isa_klassptr()) { 1718 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1719 alias_type(idx)->set_rewritable(false); 1720 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1721 alias_type(idx)->set_rewritable(false); 1722 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1723 alias_type(idx)->set_rewritable(false); 1724 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1725 alias_type(idx)->set_rewritable(false); 1726 } 1727 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1728 // but the base pointer type is not distinctive enough to identify 1729 // references into JavaThread.) 1730 1731 // Check for final fields. 1732 const TypeInstPtr* tinst = flat->isa_instptr(); 1733 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1734 ciField* field; 1735 if (tinst->const_oop() != NULL && 1736 tinst->klass() == ciEnv::current()->Class_klass() && 1737 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1738 // static field 1739 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1740 field = k->get_field_by_offset(tinst->offset(), true); 1741 } else { 1742 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1743 field = k->get_field_by_offset(tinst->offset(), false); 1744 } 1745 assert(field == NULL || 1746 original_field == NULL || 1747 (field->holder() == original_field->holder() && 1748 field->offset() == original_field->offset() && 1749 field->is_static() == original_field->is_static()), "wrong field?"); 1750 // Set field() and is_rewritable() attributes. 1751 if (field != NULL) alias_type(idx)->set_field(field); 1752 } 1753 } 1754 1755 // Fill the cache for next time. 1756 ace->_adr_type = adr_type; 1757 ace->_index = idx; 1758 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1759 1760 // Might as well try to fill the cache for the flattened version, too. 1761 AliasCacheEntry* face = probe_alias_cache(flat); 1762 if (face->_adr_type == NULL) { 1763 face->_adr_type = flat; 1764 face->_index = idx; 1765 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1766 } 1767 1768 return alias_type(idx); 1769 } 1770 1771 1772 Compile::AliasType* Compile::alias_type(ciField* field) { 1773 const TypeOopPtr* t; 1774 if (field->is_static()) 1775 t = TypeInstPtr::make(field->holder()->java_mirror()); 1776 else 1777 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1778 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1779 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1780 return atp; 1781 } 1782 1783 1784 //------------------------------have_alias_type-------------------------------- 1785 bool Compile::have_alias_type(const TypePtr* adr_type) { 1786 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1787 if (ace->_adr_type == adr_type) { 1788 return true; 1789 } 1790 1791 // Handle special cases. 1792 if (adr_type == NULL) return true; 1793 if (adr_type == TypePtr::BOTTOM) return true; 1794 1795 return find_alias_type(adr_type, true, NULL) != NULL; 1796 } 1797 1798 //-----------------------------must_alias-------------------------------------- 1799 // True if all values of the given address type are in the given alias category. 1800 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1801 if (alias_idx == AliasIdxBot) return true; // the universal category 1802 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1803 if (alias_idx == AliasIdxTop) return false; // the empty category 1804 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1805 1806 // the only remaining possible overlap is identity 1807 int adr_idx = get_alias_index(adr_type); 1808 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1809 assert(adr_idx == alias_idx || 1810 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1811 && adr_type != TypeOopPtr::BOTTOM), 1812 "should not be testing for overlap with an unsafe pointer"); 1813 return adr_idx == alias_idx; 1814 } 1815 1816 //------------------------------can_alias-------------------------------------- 1817 // True if any values of the given address type are in the given alias category. 1818 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1819 if (alias_idx == AliasIdxTop) return false; // the empty category 1820 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1821 if (alias_idx == AliasIdxBot) return true; // the universal category 1822 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1823 1824 // the only remaining possible overlap is identity 1825 int adr_idx = get_alias_index(adr_type); 1826 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1827 return adr_idx == alias_idx; 1828 } 1829 1830 1831 1832 //---------------------------pop_warm_call------------------------------------- 1833 WarmCallInfo* Compile::pop_warm_call() { 1834 WarmCallInfo* wci = _warm_calls; 1835 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1836 return wci; 1837 } 1838 1839 //----------------------------Inline_Warm-------------------------------------- 1840 int Compile::Inline_Warm() { 1841 // If there is room, try to inline some more warm call sites. 1842 // %%% Do a graph index compaction pass when we think we're out of space? 1843 if (!InlineWarmCalls) return 0; 1844 1845 int calls_made_hot = 0; 1846 int room_to_grow = NodeCountInliningCutoff - unique(); 1847 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1848 int amount_grown = 0; 1849 WarmCallInfo* call; 1850 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1851 int est_size = (int)call->size(); 1852 if (est_size > (room_to_grow - amount_grown)) { 1853 // This one won't fit anyway. Get rid of it. 1854 call->make_cold(); 1855 continue; 1856 } 1857 call->make_hot(); 1858 calls_made_hot++; 1859 amount_grown += est_size; 1860 amount_to_grow -= est_size; 1861 } 1862 1863 if (calls_made_hot > 0) set_major_progress(); 1864 return calls_made_hot; 1865 } 1866 1867 1868 //----------------------------Finish_Warm-------------------------------------- 1869 void Compile::Finish_Warm() { 1870 if (!InlineWarmCalls) return; 1871 if (failing()) return; 1872 if (warm_calls() == NULL) return; 1873 1874 // Clean up loose ends, if we are out of space for inlining. 1875 WarmCallInfo* call; 1876 while ((call = pop_warm_call()) != NULL) { 1877 call->make_cold(); 1878 } 1879 } 1880 1881 //---------------------cleanup_loop_predicates----------------------- 1882 // Remove the opaque nodes that protect the predicates so that all unused 1883 // checks and uncommon_traps will be eliminated from the ideal graph 1884 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1885 if (predicate_count()==0) return; 1886 for (int i = predicate_count(); i > 0; i--) { 1887 Node * n = predicate_opaque1_node(i-1); 1888 assert(n->Opcode() == Op_Opaque1, "must be"); 1889 igvn.replace_node(n, n->in(1)); 1890 } 1891 assert(predicate_count()==0, "should be clean!"); 1892 } 1893 1894 // StringOpts and late inlining of string methods 1895 void Compile::inline_string_calls(bool parse_time) { 1896 { 1897 // remove useless nodes to make the usage analysis simpler 1898 ResourceMark rm; 1899 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1900 } 1901 1902 { 1903 ResourceMark rm; 1904 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1905 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1906 print_method(PHASE_AFTER_STRINGOPTS, 3); 1907 } 1908 1909 // now inline anything that we skipped the first time around 1910 if (!parse_time) { 1911 _late_inlines_pos = _late_inlines.length(); 1912 } 1913 1914 while (_string_late_inlines.length() > 0) { 1915 CallGenerator* cg = _string_late_inlines.pop(); 1916 cg->do_late_inline(); 1917 if (failing()) return; 1918 } 1919 _string_late_inlines.trunc_to(0); 1920 } 1921 1922 // Late inlining of boxing methods 1923 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1924 if (_boxing_late_inlines.length() > 0) { 1925 assert(has_boxed_value(), "inconsistent"); 1926 1927 PhaseGVN* gvn = initial_gvn(); 1928 set_inlining_incrementally(true); 1929 1930 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1931 for_igvn()->clear(); 1932 gvn->replace_with(&igvn); 1933 1934 _late_inlines_pos = _late_inlines.length(); 1935 1936 while (_boxing_late_inlines.length() > 0) { 1937 CallGenerator* cg = _boxing_late_inlines.pop(); 1938 cg->do_late_inline(); 1939 if (failing()) return; 1940 } 1941 _boxing_late_inlines.trunc_to(0); 1942 1943 { 1944 ResourceMark rm; 1945 PhaseRemoveUseless pru(gvn, for_igvn()); 1946 } 1947 1948 igvn = PhaseIterGVN(gvn); 1949 igvn.optimize(); 1950 1951 set_inlining_progress(false); 1952 set_inlining_incrementally(false); 1953 } 1954 } 1955 1956 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1957 assert(IncrementalInline, "incremental inlining should be on"); 1958 PhaseGVN* gvn = initial_gvn(); 1959 1960 set_inlining_progress(false); 1961 for_igvn()->clear(); 1962 gvn->replace_with(&igvn); 1963 1964 { 1965 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 1966 int i = 0; 1967 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1968 CallGenerator* cg = _late_inlines.at(i); 1969 _late_inlines_pos = i+1; 1970 cg->do_late_inline(); 1971 if (failing()) return; 1972 } 1973 int j = 0; 1974 for (; i < _late_inlines.length(); i++, j++) { 1975 _late_inlines.at_put(j, _late_inlines.at(i)); 1976 } 1977 _late_inlines.trunc_to(j); 1978 } 1979 1980 { 1981 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 1982 ResourceMark rm; 1983 PhaseRemoveUseless pru(gvn, for_igvn()); 1984 } 1985 1986 { 1987 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 1988 igvn = PhaseIterGVN(gvn); 1989 } 1990 } 1991 1992 // Perform incremental inlining until bound on number of live nodes is reached 1993 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1994 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 1995 1996 PhaseGVN* gvn = initial_gvn(); 1997 1998 set_inlining_incrementally(true); 1999 set_inlining_progress(true); 2000 uint low_live_nodes = 0; 2001 2002 while(inlining_progress() && _late_inlines.length() > 0) { 2003 2004 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2005 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2006 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2007 // PhaseIdealLoop is expensive so we only try it once we are 2008 // out of live nodes and we only try it again if the previous 2009 // helped got the number of nodes down significantly 2010 PhaseIdealLoop ideal_loop( igvn, false, true ); 2011 if (failing()) return; 2012 low_live_nodes = live_nodes(); 2013 _major_progress = true; 2014 } 2015 2016 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2017 break; 2018 } 2019 } 2020 2021 inline_incrementally_one(igvn); 2022 2023 if (failing()) return; 2024 2025 { 2026 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2027 igvn.optimize(); 2028 } 2029 2030 if (failing()) return; 2031 } 2032 2033 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2034 2035 if (_string_late_inlines.length() > 0) { 2036 assert(has_stringbuilder(), "inconsistent"); 2037 for_igvn()->clear(); 2038 initial_gvn()->replace_with(&igvn); 2039 2040 inline_string_calls(false); 2041 2042 if (failing()) return; 2043 2044 { 2045 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2046 ResourceMark rm; 2047 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2048 } 2049 2050 { 2051 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2052 igvn = PhaseIterGVN(gvn); 2053 igvn.optimize(); 2054 } 2055 } 2056 2057 set_inlining_incrementally(false); 2058 } 2059 2060 2061 //------------------------------Optimize--------------------------------------- 2062 // Given a graph, optimize it. 2063 void Compile::Optimize() { 2064 TracePhase tp("optimizer", &timers[_t_optimizer]); 2065 2066 #ifndef PRODUCT 2067 if (env()->break_at_compile()) { 2068 BREAKPOINT; 2069 } 2070 2071 #endif 2072 2073 ResourceMark rm; 2074 int loop_opts_cnt; 2075 2076 print_inlining_reinit(); 2077 2078 NOT_PRODUCT( verify_graph_edges(); ) 2079 2080 print_method(PHASE_AFTER_PARSING); 2081 2082 { 2083 // Iterative Global Value Numbering, including ideal transforms 2084 // Initialize IterGVN with types and values from parse-time GVN 2085 PhaseIterGVN igvn(initial_gvn()); 2086 #ifdef ASSERT 2087 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2088 #endif 2089 { 2090 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2091 igvn.optimize(); 2092 } 2093 2094 print_method(PHASE_ITER_GVN1, 2); 2095 2096 if (failing()) return; 2097 2098 inline_incrementally(igvn); 2099 2100 print_method(PHASE_INCREMENTAL_INLINE, 2); 2101 2102 if (failing()) return; 2103 2104 if (eliminate_boxing()) { 2105 // Inline valueOf() methods now. 2106 inline_boxing_calls(igvn); 2107 2108 if (AlwaysIncrementalInline) { 2109 inline_incrementally(igvn); 2110 } 2111 2112 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2113 2114 if (failing()) return; 2115 } 2116 2117 // Remove the speculative part of types and clean up the graph from 2118 // the extra CastPP nodes whose only purpose is to carry them. Do 2119 // that early so that optimizations are not disrupted by the extra 2120 // CastPP nodes. 2121 remove_speculative_types(igvn); 2122 2123 // No more new expensive nodes will be added to the list from here 2124 // so keep only the actual candidates for optimizations. 2125 cleanup_expensive_nodes(igvn); 2126 2127 // Perform escape analysis 2128 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2129 if (has_loops()) { 2130 // Cleanup graph (remove dead nodes). 2131 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2132 PhaseIdealLoop ideal_loop( igvn, false, true ); 2133 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2134 if (failing()) return; 2135 } 2136 ConnectionGraph::do_analysis(this, &igvn); 2137 2138 if (failing()) return; 2139 2140 // Optimize out fields loads from scalar replaceable allocations. 2141 igvn.optimize(); 2142 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2143 2144 if (failing()) return; 2145 2146 if (congraph() != NULL && macro_count() > 0) { 2147 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2148 PhaseMacroExpand mexp(igvn); 2149 mexp.eliminate_macro_nodes(); 2150 igvn.set_delay_transform(false); 2151 2152 igvn.optimize(); 2153 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2154 2155 if (failing()) return; 2156 } 2157 } 2158 2159 // Loop transforms on the ideal graph. Range Check Elimination, 2160 // peeling, unrolling, etc. 2161 2162 // Set loop opts counter 2163 loop_opts_cnt = num_loop_opts(); 2164 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2165 { 2166 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2167 PhaseIdealLoop ideal_loop( igvn, true ); 2168 loop_opts_cnt--; 2169 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2170 if (failing()) return; 2171 } 2172 // Loop opts pass if partial peeling occurred in previous pass 2173 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2174 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2175 PhaseIdealLoop ideal_loop( igvn, false ); 2176 loop_opts_cnt--; 2177 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2178 if (failing()) return; 2179 } 2180 // Loop opts pass for loop-unrolling before CCP 2181 if(major_progress() && (loop_opts_cnt > 0)) { 2182 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2183 PhaseIdealLoop ideal_loop( igvn, false ); 2184 loop_opts_cnt--; 2185 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2186 } 2187 if (!failing()) { 2188 // Verify that last round of loop opts produced a valid graph 2189 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2190 PhaseIdealLoop::verify(igvn); 2191 } 2192 } 2193 if (failing()) return; 2194 2195 // Conditional Constant Propagation; 2196 PhaseCCP ccp( &igvn ); 2197 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2198 { 2199 TracePhase tp("ccp", &timers[_t_ccp]); 2200 ccp.do_transform(); 2201 } 2202 print_method(PHASE_CPP1, 2); 2203 2204 assert( true, "Break here to ccp.dump_old2new_map()"); 2205 2206 // Iterative Global Value Numbering, including ideal transforms 2207 { 2208 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2209 igvn = ccp; 2210 igvn.optimize(); 2211 } 2212 2213 print_method(PHASE_ITER_GVN2, 2); 2214 2215 if (failing()) return; 2216 2217 // Loop transforms on the ideal graph. Range Check Elimination, 2218 // peeling, unrolling, etc. 2219 if(loop_opts_cnt > 0) { 2220 debug_only( int cnt = 0; ); 2221 while(major_progress() && (loop_opts_cnt > 0)) { 2222 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2223 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2224 PhaseIdealLoop ideal_loop( igvn, true); 2225 loop_opts_cnt--; 2226 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2227 if (failing()) return; 2228 } 2229 } 2230 2231 { 2232 // Verify that all previous optimizations produced a valid graph 2233 // at least to this point, even if no loop optimizations were done. 2234 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2235 PhaseIdealLoop::verify(igvn); 2236 } 2237 2238 { 2239 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2240 PhaseMacroExpand mex(igvn); 2241 if (mex.expand_macro_nodes()) { 2242 assert(failing(), "must bail out w/ explicit message"); 2243 return; 2244 } 2245 } 2246 2247 DEBUG_ONLY( _modified_nodes = NULL; ) 2248 } // (End scope of igvn; run destructor if necessary for asserts.) 2249 2250 process_print_inlining(); 2251 // A method with only infinite loops has no edges entering loops from root 2252 { 2253 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2254 if (final_graph_reshaping()) { 2255 assert(failing(), "must bail out w/ explicit message"); 2256 return; 2257 } 2258 } 2259 2260 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2261 } 2262 2263 2264 //------------------------------Code_Gen--------------------------------------- 2265 // Given a graph, generate code for it 2266 void Compile::Code_Gen() { 2267 if (failing()) { 2268 return; 2269 } 2270 2271 // Perform instruction selection. You might think we could reclaim Matcher 2272 // memory PDQ, but actually the Matcher is used in generating spill code. 2273 // Internals of the Matcher (including some VectorSets) must remain live 2274 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2275 // set a bit in reclaimed memory. 2276 2277 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2278 // nodes. Mapping is only valid at the root of each matched subtree. 2279 NOT_PRODUCT( verify_graph_edges(); ) 2280 2281 Matcher matcher; 2282 _matcher = &matcher; 2283 { 2284 TracePhase tp("matcher", &timers[_t_matcher]); 2285 matcher.match(); 2286 } 2287 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2288 // nodes. Mapping is only valid at the root of each matched subtree. 2289 NOT_PRODUCT( verify_graph_edges(); ) 2290 2291 // If you have too many nodes, or if matching has failed, bail out 2292 check_node_count(0, "out of nodes matching instructions"); 2293 if (failing()) { 2294 return; 2295 } 2296 2297 // Build a proper-looking CFG 2298 PhaseCFG cfg(node_arena(), root(), matcher); 2299 _cfg = &cfg; 2300 { 2301 TracePhase tp("scheduler", &timers[_t_scheduler]); 2302 bool success = cfg.do_global_code_motion(); 2303 if (!success) { 2304 return; 2305 } 2306 2307 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2308 NOT_PRODUCT( verify_graph_edges(); ) 2309 debug_only( cfg.verify(); ) 2310 } 2311 2312 PhaseChaitin regalloc(unique(), cfg, matcher); 2313 _regalloc = ®alloc; 2314 { 2315 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2316 // Perform register allocation. After Chaitin, use-def chains are 2317 // no longer accurate (at spill code) and so must be ignored. 2318 // Node->LRG->reg mappings are still accurate. 2319 _regalloc->Register_Allocate(); 2320 2321 // Bail out if the allocator builds too many nodes 2322 if (failing()) { 2323 return; 2324 } 2325 } 2326 2327 // Prior to register allocation we kept empty basic blocks in case the 2328 // the allocator needed a place to spill. After register allocation we 2329 // are not adding any new instructions. If any basic block is empty, we 2330 // can now safely remove it. 2331 { 2332 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2333 cfg.remove_empty_blocks(); 2334 if (do_freq_based_layout()) { 2335 PhaseBlockLayout layout(cfg); 2336 } else { 2337 cfg.set_loop_alignment(); 2338 } 2339 cfg.fixup_flow(); 2340 } 2341 2342 // Apply peephole optimizations 2343 if( OptoPeephole ) { 2344 TracePhase tp("peephole", &timers[_t_peephole]); 2345 PhasePeephole peep( _regalloc, cfg); 2346 peep.do_transform(); 2347 } 2348 2349 // Do late expand if CPU requires this. 2350 if (Matcher::require_postalloc_expand) { 2351 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2352 cfg.postalloc_expand(_regalloc); 2353 } 2354 2355 // Convert Nodes to instruction bits in a buffer 2356 { 2357 TraceTime tp("output", &timers[_t_output], CITime); 2358 Output(); 2359 } 2360 2361 print_method(PHASE_FINAL_CODE); 2362 2363 // He's dead, Jim. 2364 _cfg = (PhaseCFG*)0xdeadbeef; 2365 _regalloc = (PhaseChaitin*)0xdeadbeef; 2366 } 2367 2368 2369 //------------------------------dump_asm--------------------------------------- 2370 // Dump formatted assembly 2371 #ifndef PRODUCT 2372 void Compile::dump_asm(int *pcs, uint pc_limit) { 2373 bool cut_short = false; 2374 tty->print_cr("#"); 2375 tty->print("# "); _tf->dump(); tty->cr(); 2376 tty->print_cr("#"); 2377 2378 // For all blocks 2379 int pc = 0x0; // Program counter 2380 char starts_bundle = ' '; 2381 _regalloc->dump_frame(); 2382 2383 Node *n = NULL; 2384 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2385 if (VMThread::should_terminate()) { 2386 cut_short = true; 2387 break; 2388 } 2389 Block* block = _cfg->get_block(i); 2390 if (block->is_connector() && !Verbose) { 2391 continue; 2392 } 2393 n = block->head(); 2394 if (pcs && n->_idx < pc_limit) { 2395 tty->print("%3.3x ", pcs[n->_idx]); 2396 } else { 2397 tty->print(" "); 2398 } 2399 block->dump_head(_cfg); 2400 if (block->is_connector()) { 2401 tty->print_cr(" # Empty connector block"); 2402 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2403 tty->print_cr(" # Block is sole successor of call"); 2404 } 2405 2406 // For all instructions 2407 Node *delay = NULL; 2408 for (uint j = 0; j < block->number_of_nodes(); j++) { 2409 if (VMThread::should_terminate()) { 2410 cut_short = true; 2411 break; 2412 } 2413 n = block->get_node(j); 2414 if (valid_bundle_info(n)) { 2415 Bundle* bundle = node_bundling(n); 2416 if (bundle->used_in_unconditional_delay()) { 2417 delay = n; 2418 continue; 2419 } 2420 if (bundle->starts_bundle()) { 2421 starts_bundle = '+'; 2422 } 2423 } 2424 2425 if (WizardMode) { 2426 n->dump(); 2427 } 2428 2429 if( !n->is_Region() && // Dont print in the Assembly 2430 !n->is_Phi() && // a few noisely useless nodes 2431 !n->is_Proj() && 2432 !n->is_MachTemp() && 2433 !n->is_SafePointScalarObject() && 2434 !n->is_Catch() && // Would be nice to print exception table targets 2435 !n->is_MergeMem() && // Not very interesting 2436 !n->is_top() && // Debug info table constants 2437 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2438 ) { 2439 if (pcs && n->_idx < pc_limit) 2440 tty->print("%3.3x", pcs[n->_idx]); 2441 else 2442 tty->print(" "); 2443 tty->print(" %c ", starts_bundle); 2444 starts_bundle = ' '; 2445 tty->print("\t"); 2446 n->format(_regalloc, tty); 2447 tty->cr(); 2448 } 2449 2450 // If we have an instruction with a delay slot, and have seen a delay, 2451 // then back up and print it 2452 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2453 assert(delay != NULL, "no unconditional delay instruction"); 2454 if (WizardMode) delay->dump(); 2455 2456 if (node_bundling(delay)->starts_bundle()) 2457 starts_bundle = '+'; 2458 if (pcs && n->_idx < pc_limit) 2459 tty->print("%3.3x", pcs[n->_idx]); 2460 else 2461 tty->print(" "); 2462 tty->print(" %c ", starts_bundle); 2463 starts_bundle = ' '; 2464 tty->print("\t"); 2465 delay->format(_regalloc, tty); 2466 tty->cr(); 2467 delay = NULL; 2468 } 2469 2470 // Dump the exception table as well 2471 if( n->is_Catch() && (Verbose || WizardMode) ) { 2472 // Print the exception table for this offset 2473 _handler_table.print_subtable_for(pc); 2474 } 2475 } 2476 2477 if (pcs && n->_idx < pc_limit) 2478 tty->print_cr("%3.3x", pcs[n->_idx]); 2479 else 2480 tty->cr(); 2481 2482 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2483 2484 } // End of per-block dump 2485 tty->cr(); 2486 2487 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2488 } 2489 #endif 2490 2491 //------------------------------Final_Reshape_Counts--------------------------- 2492 // This class defines counters to help identify when a method 2493 // may/must be executed using hardware with only 24-bit precision. 2494 struct Final_Reshape_Counts : public StackObj { 2495 int _call_count; // count non-inlined 'common' calls 2496 int _float_count; // count float ops requiring 24-bit precision 2497 int _double_count; // count double ops requiring more precision 2498 int _java_call_count; // count non-inlined 'java' calls 2499 int _inner_loop_count; // count loops which need alignment 2500 VectorSet _visited; // Visitation flags 2501 Node_List _tests; // Set of IfNodes & PCTableNodes 2502 2503 Final_Reshape_Counts() : 2504 _call_count(0), _float_count(0), _double_count(0), 2505 _java_call_count(0), _inner_loop_count(0), 2506 _visited( Thread::current()->resource_area() ) { } 2507 2508 void inc_call_count () { _call_count ++; } 2509 void inc_float_count () { _float_count ++; } 2510 void inc_double_count() { _double_count++; } 2511 void inc_java_call_count() { _java_call_count++; } 2512 void inc_inner_loop_count() { _inner_loop_count++; } 2513 2514 int get_call_count () const { return _call_count ; } 2515 int get_float_count () const { return _float_count ; } 2516 int get_double_count() const { return _double_count; } 2517 int get_java_call_count() const { return _java_call_count; } 2518 int get_inner_loop_count() const { return _inner_loop_count; } 2519 }; 2520 2521 #ifdef ASSERT 2522 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2523 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2524 // Make sure the offset goes inside the instance layout. 2525 return k->contains_field_offset(tp->offset()); 2526 // Note that OffsetBot and OffsetTop are very negative. 2527 } 2528 #endif 2529 2530 // Eliminate trivially redundant StoreCMs and accumulate their 2531 // precedence edges. 2532 void Compile::eliminate_redundant_card_marks(Node* n) { 2533 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2534 if (n->in(MemNode::Address)->outcnt() > 1) { 2535 // There are multiple users of the same address so it might be 2536 // possible to eliminate some of the StoreCMs 2537 Node* mem = n->in(MemNode::Memory); 2538 Node* adr = n->in(MemNode::Address); 2539 Node* val = n->in(MemNode::ValueIn); 2540 Node* prev = n; 2541 bool done = false; 2542 // Walk the chain of StoreCMs eliminating ones that match. As 2543 // long as it's a chain of single users then the optimization is 2544 // safe. Eliminating partially redundant StoreCMs would require 2545 // cloning copies down the other paths. 2546 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2547 if (adr == mem->in(MemNode::Address) && 2548 val == mem->in(MemNode::ValueIn)) { 2549 // redundant StoreCM 2550 if (mem->req() > MemNode::OopStore) { 2551 // Hasn't been processed by this code yet. 2552 n->add_prec(mem->in(MemNode::OopStore)); 2553 } else { 2554 // Already converted to precedence edge 2555 for (uint i = mem->req(); i < mem->len(); i++) { 2556 // Accumulate any precedence edges 2557 if (mem->in(i) != NULL) { 2558 n->add_prec(mem->in(i)); 2559 } 2560 } 2561 // Everything above this point has been processed. 2562 done = true; 2563 } 2564 // Eliminate the previous StoreCM 2565 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2566 assert(mem->outcnt() == 0, "should be dead"); 2567 mem->disconnect_inputs(NULL, this); 2568 } else { 2569 prev = mem; 2570 } 2571 mem = prev->in(MemNode::Memory); 2572 } 2573 } 2574 } 2575 2576 //------------------------------final_graph_reshaping_impl---------------------- 2577 // Implement items 1-5 from final_graph_reshaping below. 2578 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2579 2580 if ( n->outcnt() == 0 ) return; // dead node 2581 uint nop = n->Opcode(); 2582 2583 // Check for 2-input instruction with "last use" on right input. 2584 // Swap to left input. Implements item (2). 2585 if( n->req() == 3 && // two-input instruction 2586 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2587 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2588 n->in(2)->outcnt() == 1 &&// right use IS a last use 2589 !n->in(2)->is_Con() ) { // right use is not a constant 2590 // Check for commutative opcode 2591 switch( nop ) { 2592 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2593 case Op_MaxI: case Op_MinI: 2594 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2595 case Op_AndL: case Op_XorL: case Op_OrL: 2596 case Op_AndI: case Op_XorI: case Op_OrI: { 2597 // Move "last use" input to left by swapping inputs 2598 n->swap_edges(1, 2); 2599 break; 2600 } 2601 default: 2602 break; 2603 } 2604 } 2605 2606 #ifdef ASSERT 2607 if( n->is_Mem() ) { 2608 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2609 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2610 // oop will be recorded in oop map if load crosses safepoint 2611 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2612 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2613 "raw memory operations should have control edge"); 2614 } 2615 #endif 2616 // Count FPU ops and common calls, implements item (3) 2617 switch( nop ) { 2618 // Count all float operations that may use FPU 2619 case Op_AddF: 2620 case Op_SubF: 2621 case Op_MulF: 2622 case Op_DivF: 2623 case Op_NegF: 2624 case Op_ModF: 2625 case Op_ConvI2F: 2626 case Op_ConF: 2627 case Op_CmpF: 2628 case Op_CmpF3: 2629 // case Op_ConvL2F: // longs are split into 32-bit halves 2630 frc.inc_float_count(); 2631 break; 2632 2633 case Op_ConvF2D: 2634 case Op_ConvD2F: 2635 frc.inc_float_count(); 2636 frc.inc_double_count(); 2637 break; 2638 2639 // Count all double operations that may use FPU 2640 case Op_AddD: 2641 case Op_SubD: 2642 case Op_MulD: 2643 case Op_DivD: 2644 case Op_NegD: 2645 case Op_ModD: 2646 case Op_ConvI2D: 2647 case Op_ConvD2I: 2648 // case Op_ConvL2D: // handled by leaf call 2649 // case Op_ConvD2L: // handled by leaf call 2650 case Op_ConD: 2651 case Op_CmpD: 2652 case Op_CmpD3: 2653 frc.inc_double_count(); 2654 break; 2655 case Op_Opaque1: // Remove Opaque Nodes before matching 2656 case Op_Opaque2: // Remove Opaque Nodes before matching 2657 case Op_Opaque3: 2658 n->subsume_by(n->in(1), this); 2659 break; 2660 case Op_CallStaticJava: 2661 case Op_CallJava: 2662 case Op_CallDynamicJava: 2663 frc.inc_java_call_count(); // Count java call site; 2664 case Op_CallRuntime: 2665 case Op_CallLeaf: 2666 case Op_CallLeafNoFP: { 2667 assert( n->is_Call(), "" ); 2668 CallNode *call = n->as_Call(); 2669 // Count call sites where the FP mode bit would have to be flipped. 2670 // Do not count uncommon runtime calls: 2671 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2672 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2673 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2674 frc.inc_call_count(); // Count the call site 2675 } else { // See if uncommon argument is shared 2676 Node *n = call->in(TypeFunc::Parms); 2677 int nop = n->Opcode(); 2678 // Clone shared simple arguments to uncommon calls, item (1). 2679 if( n->outcnt() > 1 && 2680 !n->is_Proj() && 2681 nop != Op_CreateEx && 2682 nop != Op_CheckCastPP && 2683 nop != Op_DecodeN && 2684 nop != Op_DecodeNKlass && 2685 !n->is_Mem() ) { 2686 Node *x = n->clone(); 2687 call->set_req( TypeFunc::Parms, x ); 2688 } 2689 } 2690 break; 2691 } 2692 2693 case Op_StoreD: 2694 case Op_LoadD: 2695 case Op_LoadD_unaligned: 2696 frc.inc_double_count(); 2697 goto handle_mem; 2698 case Op_StoreF: 2699 case Op_LoadF: 2700 frc.inc_float_count(); 2701 goto handle_mem; 2702 2703 case Op_StoreCM: 2704 { 2705 // Convert OopStore dependence into precedence edge 2706 Node* prec = n->in(MemNode::OopStore); 2707 n->del_req(MemNode::OopStore); 2708 n->add_prec(prec); 2709 eliminate_redundant_card_marks(n); 2710 } 2711 2712 // fall through 2713 2714 case Op_StoreB: 2715 case Op_StoreC: 2716 case Op_StorePConditional: 2717 case Op_StoreI: 2718 case Op_StoreL: 2719 case Op_StoreIConditional: 2720 case Op_StoreLConditional: 2721 case Op_CompareAndSwapI: 2722 case Op_CompareAndSwapL: 2723 case Op_CompareAndSwapP: 2724 case Op_CompareAndSwapN: 2725 case Op_GetAndAddI: 2726 case Op_GetAndAddL: 2727 case Op_GetAndSetI: 2728 case Op_GetAndSetL: 2729 case Op_GetAndSetP: 2730 case Op_GetAndSetN: 2731 case Op_StoreP: 2732 case Op_StoreN: 2733 case Op_StoreNKlass: 2734 case Op_LoadB: 2735 case Op_LoadUB: 2736 case Op_LoadUS: 2737 case Op_LoadI: 2738 case Op_LoadKlass: 2739 case Op_LoadNKlass: 2740 case Op_LoadL: 2741 case Op_LoadL_unaligned: 2742 case Op_LoadPLocked: 2743 case Op_LoadP: 2744 case Op_LoadN: 2745 case Op_LoadRange: 2746 case Op_LoadS: { 2747 handle_mem: 2748 #ifdef ASSERT 2749 if( VerifyOptoOopOffsets ) { 2750 assert( n->is_Mem(), "" ); 2751 MemNode *mem = (MemNode*)n; 2752 // Check to see if address types have grounded out somehow. 2753 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2754 assert( !tp || oop_offset_is_sane(tp), "" ); 2755 } 2756 #endif 2757 break; 2758 } 2759 2760 case Op_AddP: { // Assert sane base pointers 2761 Node *addp = n->in(AddPNode::Address); 2762 assert( !addp->is_AddP() || 2763 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2764 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2765 "Base pointers must match" ); 2766 #ifdef _LP64 2767 if ((UseCompressedOops || UseCompressedClassPointers) && 2768 addp->Opcode() == Op_ConP && 2769 addp == n->in(AddPNode::Base) && 2770 n->in(AddPNode::Offset)->is_Con()) { 2771 // Use addressing with narrow klass to load with offset on x86. 2772 // On sparc loading 32-bits constant and decoding it have less 2773 // instructions (4) then load 64-bits constant (7). 2774 // Do this transformation here since IGVN will convert ConN back to ConP. 2775 const Type* t = addp->bottom_type(); 2776 if (t->isa_oopptr() || t->isa_klassptr()) { 2777 Node* nn = NULL; 2778 2779 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2780 2781 // Look for existing ConN node of the same exact type. 2782 Node* r = root(); 2783 uint cnt = r->outcnt(); 2784 for (uint i = 0; i < cnt; i++) { 2785 Node* m = r->raw_out(i); 2786 if (m!= NULL && m->Opcode() == op && 2787 m->bottom_type()->make_ptr() == t) { 2788 nn = m; 2789 break; 2790 } 2791 } 2792 if (nn != NULL) { 2793 // Decode a narrow oop to match address 2794 // [R12 + narrow_oop_reg<<3 + offset] 2795 if (t->isa_oopptr()) { 2796 nn = new DecodeNNode(nn, t); 2797 } else { 2798 nn = new DecodeNKlassNode(nn, t); 2799 } 2800 n->set_req(AddPNode::Base, nn); 2801 n->set_req(AddPNode::Address, nn); 2802 if (addp->outcnt() == 0) { 2803 addp->disconnect_inputs(NULL, this); 2804 } 2805 } 2806 } 2807 } 2808 #endif 2809 break; 2810 } 2811 2812 #ifdef _LP64 2813 case Op_CastPP: 2814 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2815 Node* in1 = n->in(1); 2816 const Type* t = n->bottom_type(); 2817 Node* new_in1 = in1->clone(); 2818 new_in1->as_DecodeN()->set_type(t); 2819 2820 if (!Matcher::narrow_oop_use_complex_address()) { 2821 // 2822 // x86, ARM and friends can handle 2 adds in addressing mode 2823 // and Matcher can fold a DecodeN node into address by using 2824 // a narrow oop directly and do implicit NULL check in address: 2825 // 2826 // [R12 + narrow_oop_reg<<3 + offset] 2827 // NullCheck narrow_oop_reg 2828 // 2829 // On other platforms (Sparc) we have to keep new DecodeN node and 2830 // use it to do implicit NULL check in address: 2831 // 2832 // decode_not_null narrow_oop_reg, base_reg 2833 // [base_reg + offset] 2834 // NullCheck base_reg 2835 // 2836 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2837 // to keep the information to which NULL check the new DecodeN node 2838 // corresponds to use it as value in implicit_null_check(). 2839 // 2840 new_in1->set_req(0, n->in(0)); 2841 } 2842 2843 n->subsume_by(new_in1, this); 2844 if (in1->outcnt() == 0) { 2845 in1->disconnect_inputs(NULL, this); 2846 } 2847 } 2848 break; 2849 2850 case Op_CmpP: 2851 // Do this transformation here to preserve CmpPNode::sub() and 2852 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2853 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2854 Node* in1 = n->in(1); 2855 Node* in2 = n->in(2); 2856 if (!in1->is_DecodeNarrowPtr()) { 2857 in2 = in1; 2858 in1 = n->in(2); 2859 } 2860 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2861 2862 Node* new_in2 = NULL; 2863 if (in2->is_DecodeNarrowPtr()) { 2864 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2865 new_in2 = in2->in(1); 2866 } else if (in2->Opcode() == Op_ConP) { 2867 const Type* t = in2->bottom_type(); 2868 if (t == TypePtr::NULL_PTR) { 2869 assert(in1->is_DecodeN(), "compare klass to null?"); 2870 // Don't convert CmpP null check into CmpN if compressed 2871 // oops implicit null check is not generated. 2872 // This will allow to generate normal oop implicit null check. 2873 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2874 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 2875 // 2876 // This transformation together with CastPP transformation above 2877 // will generated code for implicit NULL checks for compressed oops. 2878 // 2879 // The original code after Optimize() 2880 // 2881 // LoadN memory, narrow_oop_reg 2882 // decode narrow_oop_reg, base_reg 2883 // CmpP base_reg, NULL 2884 // CastPP base_reg // NotNull 2885 // Load [base_reg + offset], val_reg 2886 // 2887 // after these transformations will be 2888 // 2889 // LoadN memory, narrow_oop_reg 2890 // CmpN narrow_oop_reg, NULL 2891 // decode_not_null narrow_oop_reg, base_reg 2892 // Load [base_reg + offset], val_reg 2893 // 2894 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2895 // since narrow oops can be used in debug info now (see the code in 2896 // final_graph_reshaping_walk()). 2897 // 2898 // At the end the code will be matched to 2899 // on x86: 2900 // 2901 // Load_narrow_oop memory, narrow_oop_reg 2902 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2903 // NullCheck narrow_oop_reg 2904 // 2905 // and on sparc: 2906 // 2907 // Load_narrow_oop memory, narrow_oop_reg 2908 // decode_not_null narrow_oop_reg, base_reg 2909 // Load [base_reg + offset], val_reg 2910 // NullCheck base_reg 2911 // 2912 } else if (t->isa_oopptr()) { 2913 new_in2 = ConNode::make(t->make_narrowoop()); 2914 } else if (t->isa_klassptr()) { 2915 new_in2 = ConNode::make(t->make_narrowklass()); 2916 } 2917 } 2918 if (new_in2 != NULL) { 2919 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 2920 n->subsume_by(cmpN, this); 2921 if (in1->outcnt() == 0) { 2922 in1->disconnect_inputs(NULL, this); 2923 } 2924 if (in2->outcnt() == 0) { 2925 in2->disconnect_inputs(NULL, this); 2926 } 2927 } 2928 } 2929 break; 2930 2931 case Op_DecodeN: 2932 case Op_DecodeNKlass: 2933 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2934 // DecodeN could be pinned when it can't be fold into 2935 // an address expression, see the code for Op_CastPP above. 2936 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2937 break; 2938 2939 case Op_EncodeP: 2940 case Op_EncodePKlass: { 2941 Node* in1 = n->in(1); 2942 if (in1->is_DecodeNarrowPtr()) { 2943 n->subsume_by(in1->in(1), this); 2944 } else if (in1->Opcode() == Op_ConP) { 2945 const Type* t = in1->bottom_type(); 2946 if (t == TypePtr::NULL_PTR) { 2947 assert(t->isa_oopptr(), "null klass?"); 2948 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 2949 } else if (t->isa_oopptr()) { 2950 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 2951 } else if (t->isa_klassptr()) { 2952 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 2953 } 2954 } 2955 if (in1->outcnt() == 0) { 2956 in1->disconnect_inputs(NULL, this); 2957 } 2958 break; 2959 } 2960 2961 case Op_Proj: { 2962 if (OptimizeStringConcat) { 2963 ProjNode* p = n->as_Proj(); 2964 if (p->_is_io_use) { 2965 // Separate projections were used for the exception path which 2966 // are normally removed by a late inline. If it wasn't inlined 2967 // then they will hang around and should just be replaced with 2968 // the original one. 2969 Node* proj = NULL; 2970 // Replace with just one 2971 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2972 Node *use = i.get(); 2973 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2974 proj = use; 2975 break; 2976 } 2977 } 2978 assert(proj != NULL, "must be found"); 2979 p->subsume_by(proj, this); 2980 } 2981 } 2982 break; 2983 } 2984 2985 case Op_Phi: 2986 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2987 // The EncodeP optimization may create Phi with the same edges 2988 // for all paths. It is not handled well by Register Allocator. 2989 Node* unique_in = n->in(1); 2990 assert(unique_in != NULL, ""); 2991 uint cnt = n->req(); 2992 for (uint i = 2; i < cnt; i++) { 2993 Node* m = n->in(i); 2994 assert(m != NULL, ""); 2995 if (unique_in != m) 2996 unique_in = NULL; 2997 } 2998 if (unique_in != NULL) { 2999 n->subsume_by(unique_in, this); 3000 } 3001 } 3002 break; 3003 3004 #endif 3005 3006 case Op_ModI: 3007 if (UseDivMod) { 3008 // Check if a%b and a/b both exist 3009 Node* d = n->find_similar(Op_DivI); 3010 if (d) { 3011 // Replace them with a fused divmod if supported 3012 if (Matcher::has_match_rule(Op_DivModI)) { 3013 DivModINode* divmod = DivModINode::make(n); 3014 d->subsume_by(divmod->div_proj(), this); 3015 n->subsume_by(divmod->mod_proj(), this); 3016 } else { 3017 // replace a%b with a-((a/b)*b) 3018 Node* mult = new MulINode(d, d->in(2)); 3019 Node* sub = new SubINode(d->in(1), mult); 3020 n->subsume_by(sub, this); 3021 } 3022 } 3023 } 3024 break; 3025 3026 case Op_ModL: 3027 if (UseDivMod) { 3028 // Check if a%b and a/b both exist 3029 Node* d = n->find_similar(Op_DivL); 3030 if (d) { 3031 // Replace them with a fused divmod if supported 3032 if (Matcher::has_match_rule(Op_DivModL)) { 3033 DivModLNode* divmod = DivModLNode::make(n); 3034 d->subsume_by(divmod->div_proj(), this); 3035 n->subsume_by(divmod->mod_proj(), this); 3036 } else { 3037 // replace a%b with a-((a/b)*b) 3038 Node* mult = new MulLNode(d, d->in(2)); 3039 Node* sub = new SubLNode(d->in(1), mult); 3040 n->subsume_by(sub, this); 3041 } 3042 } 3043 } 3044 break; 3045 3046 case Op_LoadVector: 3047 case Op_StoreVector: 3048 break; 3049 3050 case Op_PackB: 3051 case Op_PackS: 3052 case Op_PackI: 3053 case Op_PackF: 3054 case Op_PackL: 3055 case Op_PackD: 3056 if (n->req()-1 > 2) { 3057 // Replace many operand PackNodes with a binary tree for matching 3058 PackNode* p = (PackNode*) n; 3059 Node* btp = p->binary_tree_pack(1, n->req()); 3060 n->subsume_by(btp, this); 3061 } 3062 break; 3063 case Op_Loop: 3064 case Op_CountedLoop: 3065 if (n->as_Loop()->is_inner_loop()) { 3066 frc.inc_inner_loop_count(); 3067 } 3068 break; 3069 case Op_LShiftI: 3070 case Op_RShiftI: 3071 case Op_URShiftI: 3072 case Op_LShiftL: 3073 case Op_RShiftL: 3074 case Op_URShiftL: 3075 if (Matcher::need_masked_shift_count) { 3076 // The cpu's shift instructions don't restrict the count to the 3077 // lower 5/6 bits. We need to do the masking ourselves. 3078 Node* in2 = n->in(2); 3079 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3080 const TypeInt* t = in2->find_int_type(); 3081 if (t != NULL && t->is_con()) { 3082 juint shift = t->get_con(); 3083 if (shift > mask) { // Unsigned cmp 3084 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3085 } 3086 } else { 3087 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3088 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3089 n->set_req(2, shift); 3090 } 3091 } 3092 if (in2->outcnt() == 0) { // Remove dead node 3093 in2->disconnect_inputs(NULL, this); 3094 } 3095 } 3096 break; 3097 case Op_MemBarStoreStore: 3098 case Op_MemBarRelease: 3099 // Break the link with AllocateNode: it is no longer useful and 3100 // confuses register allocation. 3101 if (n->req() > MemBarNode::Precedent) { 3102 n->set_req(MemBarNode::Precedent, top()); 3103 } 3104 break; 3105 default: 3106 assert( !n->is_Call(), "" ); 3107 assert( !n->is_Mem(), "" ); 3108 assert( nop != Op_ProfileBranch, "should be eliminated during IGVN"); 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 if (method->is_shared()) { 3326 // Don't trust gathered profile for shared LambdaForms (marked w/ @LambdaForm.Shared). 3327 return false; 3328 } 3329 ciMethodData* md = method->method_data(); 3330 if (md->is_empty()) { 3331 // Assume the trap has not occurred, or that it occurred only 3332 // because of a transient condition during start-up in the interpreter. 3333 return false; 3334 } 3335 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3336 if (md->has_trap_at(bci, m, reason) != 0) { 3337 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3338 // Also, if there are multiple reasons, or if there is no per-BCI record, 3339 // assume the worst. 3340 if (log()) 3341 log()->elem("observe trap='%s' count='%d'", 3342 Deoptimization::trap_reason_name(reason), 3343 md->trap_count(reason)); 3344 return true; 3345 } else { 3346 // Ignore method/bci and see if there have been too many globally. 3347 return too_many_traps(reason, md); 3348 } 3349 } 3350 3351 // Less-accurate variant which does not require a method and bci. 3352 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3353 ciMethodData* logmd) { 3354 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3355 // Too many traps globally. 3356 // Note that we use cumulative trap_count, not just md->trap_count. 3357 if (log()) { 3358 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3359 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3360 Deoptimization::trap_reason_name(reason), 3361 mcount, trap_count(reason)); 3362 } 3363 return true; 3364 } else { 3365 // The coast is clear. 3366 return false; 3367 } 3368 } 3369 3370 //--------------------------too_many_recompiles-------------------------------- 3371 // Report if there are too many recompiles at the current method and bci. 3372 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3373 // Is not eager to return true, since this will cause the compiler to use 3374 // Action_none for a trap point, to avoid too many recompilations. 3375 bool Compile::too_many_recompiles(ciMethod* method, 3376 int bci, 3377 Deoptimization::DeoptReason reason) { 3378 if (method->is_shared()) { 3379 // Don't trust gathered profile for shared LambdaForms (marked w/ @LambdaForm.Shared). 3380 return false; 3381 } 3382 ciMethodData* md = method->method_data(); 3383 if (md->is_empty()) { 3384 // Assume the trap has not occurred, or that it occurred only 3385 // because of a transient condition during start-up in the interpreter. 3386 return false; 3387 } 3388 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3389 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3390 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3391 Deoptimization::DeoptReason per_bc_reason 3392 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3393 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3394 if ((per_bc_reason == Deoptimization::Reason_none 3395 || md->has_trap_at(bci, m, reason) != 0) 3396 // The trap frequency measure we care about is the recompile count: 3397 && md->trap_recompiled_at(bci, m) 3398 && md->overflow_recompile_count() >= bc_cutoff) { 3399 // Do not emit a trap here if it has already caused recompilations. 3400 // Also, if there are multiple reasons, or if there is no per-BCI record, 3401 // assume the worst. 3402 if (log()) 3403 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3404 Deoptimization::trap_reason_name(reason), 3405 md->trap_count(reason), 3406 md->overflow_recompile_count()); 3407 return true; 3408 } else if (trap_count(reason) != 0 3409 && decompile_count() >= m_cutoff) { 3410 // Too many recompiles globally, and we have seen this sort of trap. 3411 // Use cumulative decompile_count, not just md->decompile_count. 3412 if (log()) 3413 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3414 Deoptimization::trap_reason_name(reason), 3415 md->trap_count(reason), trap_count(reason), 3416 md->decompile_count(), decompile_count()); 3417 return true; 3418 } else { 3419 // The coast is clear. 3420 return false; 3421 } 3422 } 3423 3424 // Compute when not to trap. Used by matching trap based nodes and 3425 // NullCheck optimization. 3426 void Compile::set_allowed_deopt_reasons() { 3427 _allowed_reasons = 0; 3428 if (is_method_compilation()) { 3429 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3430 assert(rs < BitsPerInt, "recode bit map"); 3431 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3432 _allowed_reasons |= nth_bit(rs); 3433 } 3434 } 3435 } 3436 } 3437 3438 #ifndef PRODUCT 3439 //------------------------------verify_graph_edges--------------------------- 3440 // Walk the Graph and verify that there is a one-to-one correspondence 3441 // between Use-Def edges and Def-Use edges in the graph. 3442 void Compile::verify_graph_edges(bool no_dead_code) { 3443 if (VerifyGraphEdges) { 3444 ResourceArea *area = Thread::current()->resource_area(); 3445 Unique_Node_List visited(area); 3446 // Call recursive graph walk to check edges 3447 _root->verify_edges(visited); 3448 if (no_dead_code) { 3449 // Now make sure that no visited node is used by an unvisited node. 3450 bool dead_nodes = false; 3451 Unique_Node_List checked(area); 3452 while (visited.size() > 0) { 3453 Node* n = visited.pop(); 3454 checked.push(n); 3455 for (uint i = 0; i < n->outcnt(); i++) { 3456 Node* use = n->raw_out(i); 3457 if (checked.member(use)) continue; // already checked 3458 if (visited.member(use)) continue; // already in the graph 3459 if (use->is_Con()) continue; // a dead ConNode is OK 3460 // At this point, we have found a dead node which is DU-reachable. 3461 if (!dead_nodes) { 3462 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3463 dead_nodes = true; 3464 } 3465 use->dump(2); 3466 tty->print_cr("---"); 3467 checked.push(use); // No repeats; pretend it is now checked. 3468 } 3469 } 3470 assert(!dead_nodes, "using nodes must be reachable from root"); 3471 } 3472 } 3473 } 3474 3475 // Verify GC barriers consistency 3476 // Currently supported: 3477 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3478 void Compile::verify_barriers() { 3479 if (UseG1GC) { 3480 // Verify G1 pre-barriers 3481 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3482 3483 ResourceArea *area = Thread::current()->resource_area(); 3484 Unique_Node_List visited(area); 3485 Node_List worklist(area); 3486 // We're going to walk control flow backwards starting from the Root 3487 worklist.push(_root); 3488 while (worklist.size() > 0) { 3489 Node* x = worklist.pop(); 3490 if (x == NULL || x == top()) continue; 3491 if (visited.member(x)) { 3492 continue; 3493 } else { 3494 visited.push(x); 3495 } 3496 3497 if (x->is_Region()) { 3498 for (uint i = 1; i < x->req(); i++) { 3499 worklist.push(x->in(i)); 3500 } 3501 } else { 3502 worklist.push(x->in(0)); 3503 // We are looking for the pattern: 3504 // /->ThreadLocal 3505 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3506 // \->ConI(0) 3507 // We want to verify that the If and the LoadB have the same control 3508 // See GraphKit::g1_write_barrier_pre() 3509 if (x->is_If()) { 3510 IfNode *iff = x->as_If(); 3511 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3512 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3513 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3514 && cmp->in(1)->is_Load()) { 3515 LoadNode* load = cmp->in(1)->as_Load(); 3516 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3517 && load->in(2)->in(3)->is_Con() 3518 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3519 3520 Node* if_ctrl = iff->in(0); 3521 Node* load_ctrl = load->in(0); 3522 3523 if (if_ctrl != load_ctrl) { 3524 // Skip possible CProj->NeverBranch in infinite loops 3525 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3526 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3527 if_ctrl = if_ctrl->in(0)->in(0); 3528 } 3529 } 3530 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3531 } 3532 } 3533 } 3534 } 3535 } 3536 } 3537 } 3538 } 3539 3540 #endif 3541 3542 // The Compile object keeps track of failure reasons separately from the ciEnv. 3543 // This is required because there is not quite a 1-1 relation between the 3544 // ciEnv and its compilation task and the Compile object. Note that one 3545 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3546 // to backtrack and retry without subsuming loads. Other than this backtracking 3547 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3548 // by the logic in C2Compiler. 3549 void Compile::record_failure(const char* reason) { 3550 if (log() != NULL) { 3551 log()->elem("failure reason='%s' phase='compile'", reason); 3552 } 3553 if (_failure_reason == NULL) { 3554 // Record the first failure reason. 3555 _failure_reason = reason; 3556 } 3557 3558 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3559 C->print_method(PHASE_FAILURE); 3560 } 3561 _root = NULL; // flush the graph, too 3562 } 3563 3564 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3565 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3566 _phase_name(name), _dolog(CITimeVerbose) 3567 { 3568 if (_dolog) { 3569 C = Compile::current(); 3570 _log = C->log(); 3571 } else { 3572 C = NULL; 3573 _log = NULL; 3574 } 3575 if (_log != NULL) { 3576 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3577 _log->stamp(); 3578 _log->end_head(); 3579 } 3580 } 3581 3582 Compile::TracePhase::~TracePhase() { 3583 3584 C = Compile::current(); 3585 if (_dolog) { 3586 _log = C->log(); 3587 } else { 3588 _log = NULL; 3589 } 3590 3591 #ifdef ASSERT 3592 if (PrintIdealNodeCount) { 3593 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3594 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3595 } 3596 3597 if (VerifyIdealNodeCount) { 3598 Compile::current()->print_missing_nodes(); 3599 } 3600 #endif 3601 3602 if (_log != NULL) { 3603 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3604 } 3605 } 3606 3607 //============================================================================= 3608 // Two Constant's are equal when the type and the value are equal. 3609 bool Compile::Constant::operator==(const Constant& other) { 3610 if (type() != other.type() ) return false; 3611 if (can_be_reused() != other.can_be_reused()) return false; 3612 // For floating point values we compare the bit pattern. 3613 switch (type()) { 3614 case T_FLOAT: return (_v._value.i == other._v._value.i); 3615 case T_LONG: 3616 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3617 case T_OBJECT: 3618 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3619 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3620 case T_METADATA: return (_v._metadata == other._v._metadata); 3621 default: ShouldNotReachHere(); 3622 } 3623 return false; 3624 } 3625 3626 static int type_to_size_in_bytes(BasicType t) { 3627 switch (t) { 3628 case T_LONG: return sizeof(jlong ); 3629 case T_FLOAT: return sizeof(jfloat ); 3630 case T_DOUBLE: return sizeof(jdouble); 3631 case T_METADATA: return sizeof(Metadata*); 3632 // We use T_VOID as marker for jump-table entries (labels) which 3633 // need an internal word relocation. 3634 case T_VOID: 3635 case T_ADDRESS: 3636 case T_OBJECT: return sizeof(jobject); 3637 } 3638 3639 ShouldNotReachHere(); 3640 return -1; 3641 } 3642 3643 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3644 // sort descending 3645 if (a->freq() > b->freq()) return -1; 3646 if (a->freq() < b->freq()) return 1; 3647 return 0; 3648 } 3649 3650 void Compile::ConstantTable::calculate_offsets_and_size() { 3651 // First, sort the array by frequencies. 3652 _constants.sort(qsort_comparator); 3653 3654 #ifdef ASSERT 3655 // Make sure all jump-table entries were sorted to the end of the 3656 // array (they have a negative frequency). 3657 bool found_void = false; 3658 for (int i = 0; i < _constants.length(); i++) { 3659 Constant con = _constants.at(i); 3660 if (con.type() == T_VOID) 3661 found_void = true; // jump-tables 3662 else 3663 assert(!found_void, "wrong sorting"); 3664 } 3665 #endif 3666 3667 int offset = 0; 3668 for (int i = 0; i < _constants.length(); i++) { 3669 Constant* con = _constants.adr_at(i); 3670 3671 // Align offset for type. 3672 int typesize = type_to_size_in_bytes(con->type()); 3673 offset = align_size_up(offset, typesize); 3674 con->set_offset(offset); // set constant's offset 3675 3676 if (con->type() == T_VOID) { 3677 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3678 offset = offset + typesize * n->outcnt(); // expand jump-table 3679 } else { 3680 offset = offset + typesize; 3681 } 3682 } 3683 3684 // Align size up to the next section start (which is insts; see 3685 // CodeBuffer::align_at_start). 3686 assert(_size == -1, "already set?"); 3687 _size = align_size_up(offset, CodeEntryAlignment); 3688 } 3689 3690 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3691 MacroAssembler _masm(&cb); 3692 for (int i = 0; i < _constants.length(); i++) { 3693 Constant con = _constants.at(i); 3694 address constant_addr; 3695 switch (con.type()) { 3696 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3697 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3698 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3699 case T_OBJECT: { 3700 jobject obj = con.get_jobject(); 3701 int oop_index = _masm.oop_recorder()->find_index(obj); 3702 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3703 break; 3704 } 3705 case T_ADDRESS: { 3706 address addr = (address) con.get_jobject(); 3707 constant_addr = _masm.address_constant(addr); 3708 break; 3709 } 3710 // We use T_VOID as marker for jump-table entries (labels) which 3711 // need an internal word relocation. 3712 case T_VOID: { 3713 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3714 // Fill the jump-table with a dummy word. The real value is 3715 // filled in later in fill_jump_table. 3716 address dummy = (address) n; 3717 constant_addr = _masm.address_constant(dummy); 3718 // Expand jump-table 3719 for (uint i = 1; i < n->outcnt(); i++) { 3720 address temp_addr = _masm.address_constant(dummy + i); 3721 assert(temp_addr, "consts section too small"); 3722 } 3723 break; 3724 } 3725 case T_METADATA: { 3726 Metadata* obj = con.get_metadata(); 3727 int metadata_index = _masm.oop_recorder()->find_index(obj); 3728 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3729 break; 3730 } 3731 default: ShouldNotReachHere(); 3732 } 3733 assert(constant_addr, "consts section too small"); 3734 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3735 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3736 } 3737 } 3738 3739 int Compile::ConstantTable::find_offset(Constant& con) const { 3740 int idx = _constants.find(con); 3741 assert(idx != -1, "constant must be in constant table"); 3742 int offset = _constants.at(idx).offset(); 3743 assert(offset != -1, "constant table not emitted yet?"); 3744 return offset; 3745 } 3746 3747 void Compile::ConstantTable::add(Constant& con) { 3748 if (con.can_be_reused()) { 3749 int idx = _constants.find(con); 3750 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3751 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3752 return; 3753 } 3754 } 3755 (void) _constants.append(con); 3756 } 3757 3758 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3759 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3760 Constant con(type, value, b->_freq); 3761 add(con); 3762 return con; 3763 } 3764 3765 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3766 Constant con(metadata); 3767 add(con); 3768 return con; 3769 } 3770 3771 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3772 jvalue value; 3773 BasicType type = oper->type()->basic_type(); 3774 switch (type) { 3775 case T_LONG: value.j = oper->constantL(); break; 3776 case T_FLOAT: value.f = oper->constantF(); break; 3777 case T_DOUBLE: value.d = oper->constantD(); break; 3778 case T_OBJECT: 3779 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3780 case T_METADATA: return add((Metadata*)oper->constant()); break; 3781 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3782 } 3783 return add(n, type, value); 3784 } 3785 3786 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3787 jvalue value; 3788 // We can use the node pointer here to identify the right jump-table 3789 // as this method is called from Compile::Fill_buffer right before 3790 // the MachNodes are emitted and the jump-table is filled (means the 3791 // MachNode pointers do not change anymore). 3792 value.l = (jobject) n; 3793 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3794 add(con); 3795 return con; 3796 } 3797 3798 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3799 // If called from Compile::scratch_emit_size do nothing. 3800 if (Compile::current()->in_scratch_emit_size()) return; 3801 3802 assert(labels.is_nonempty(), "must be"); 3803 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3804 3805 // Since MachConstantNode::constant_offset() also contains 3806 // table_base_offset() we need to subtract the table_base_offset() 3807 // to get the plain offset into the constant table. 3808 int offset = n->constant_offset() - table_base_offset(); 3809 3810 MacroAssembler _masm(&cb); 3811 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3812 3813 for (uint i = 0; i < n->outcnt(); i++) { 3814 address* constant_addr = &jump_table_base[i]; 3815 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))); 3816 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3817 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3818 } 3819 } 3820 3821 //----------------------------static_subtype_check----------------------------- 3822 // Shortcut important common cases when superklass is exact: 3823 // (0) superklass is java.lang.Object (can occur in reflective code) 3824 // (1) subklass is already limited to a subtype of superklass => always ok 3825 // (2) subklass does not overlap with superklass => always fail 3826 // (3) superklass has NO subtypes and we can check with a simple compare. 3827 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 3828 if (StressReflectiveCode) { 3829 return SSC_full_test; // Let caller generate the general case. 3830 } 3831 3832 if (superk == env()->Object_klass()) { 3833 return SSC_always_true; // (0) this test cannot fail 3834 } 3835 3836 ciType* superelem = superk; 3837 if (superelem->is_array_klass()) 3838 superelem = superelem->as_array_klass()->base_element_type(); 3839 3840 if (!subk->is_interface()) { // cannot trust static interface types yet 3841 if (subk->is_subtype_of(superk)) { 3842 return SSC_always_true; // (1) false path dead; no dynamic test needed 3843 } 3844 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 3845 !superk->is_subtype_of(subk)) { 3846 return SSC_always_false; 3847 } 3848 } 3849 3850 // If casting to an instance klass, it must have no subtypes 3851 if (superk->is_interface()) { 3852 // Cannot trust interfaces yet. 3853 // %%% S.B. superk->nof_implementors() == 1 3854 } else if (superelem->is_instance_klass()) { 3855 ciInstanceKlass* ik = superelem->as_instance_klass(); 3856 if (!ik->has_subklass() && !ik->is_interface()) { 3857 if (!ik->is_final()) { 3858 // Add a dependency if there is a chance of a later subclass. 3859 dependencies()->assert_leaf_type(ik); 3860 } 3861 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3862 } 3863 } else { 3864 // A primitive array type has no subtypes. 3865 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3866 } 3867 3868 return SSC_full_test; 3869 } 3870 3871 // The message about the current inlining is accumulated in 3872 // _print_inlining_stream and transfered into the _print_inlining_list 3873 // once we know whether inlining succeeds or not. For regular 3874 // inlining, messages are appended to the buffer pointed by 3875 // _print_inlining_idx in the _print_inlining_list. For late inlining, 3876 // a new buffer is added after _print_inlining_idx in the list. This 3877 // way we can update the inlining message for late inlining call site 3878 // when the inlining is attempted again. 3879 void Compile::print_inlining_init() { 3880 if (print_inlining() || print_intrinsics()) { 3881 _print_inlining_stream = new stringStream(); 3882 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 3883 } 3884 } 3885 3886 void Compile::print_inlining_reinit() { 3887 if (print_inlining() || print_intrinsics()) { 3888 // Re allocate buffer when we change ResourceMark 3889 _print_inlining_stream = new stringStream(); 3890 } 3891 } 3892 3893 void Compile::print_inlining_reset() { 3894 _print_inlining_stream->reset(); 3895 } 3896 3897 void Compile::print_inlining_commit() { 3898 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 3899 // Transfer the message from _print_inlining_stream to the current 3900 // _print_inlining_list buffer and clear _print_inlining_stream. 3901 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 3902 print_inlining_reset(); 3903 } 3904 3905 void Compile::print_inlining_push() { 3906 // Add new buffer to the _print_inlining_list at current position 3907 _print_inlining_idx++; 3908 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 3909 } 3910 3911 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 3912 return _print_inlining_list->at(_print_inlining_idx); 3913 } 3914 3915 void Compile::print_inlining_update(CallGenerator* cg) { 3916 if (print_inlining() || print_intrinsics()) { 3917 if (!cg->is_late_inline()) { 3918 if (print_inlining_current().cg() != NULL) { 3919 print_inlining_push(); 3920 } 3921 print_inlining_commit(); 3922 } else { 3923 if (print_inlining_current().cg() != cg && 3924 (print_inlining_current().cg() != NULL || 3925 print_inlining_current().ss()->size() != 0)) { 3926 print_inlining_push(); 3927 } 3928 print_inlining_commit(); 3929 print_inlining_current().set_cg(cg); 3930 } 3931 } 3932 } 3933 3934 void Compile::print_inlining_move_to(CallGenerator* cg) { 3935 // We resume inlining at a late inlining call site. Locate the 3936 // corresponding inlining buffer so that we can update it. 3937 if (print_inlining()) { 3938 for (int i = 0; i < _print_inlining_list->length(); i++) { 3939 if (_print_inlining_list->adr_at(i)->cg() == cg) { 3940 _print_inlining_idx = i; 3941 return; 3942 } 3943 } 3944 ShouldNotReachHere(); 3945 } 3946 } 3947 3948 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 3949 if (print_inlining()) { 3950 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 3951 assert(print_inlining_current().cg() == cg, "wrong entry"); 3952 // replace message with new message 3953 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 3954 print_inlining_commit(); 3955 print_inlining_current().set_cg(cg); 3956 } 3957 } 3958 3959 void Compile::print_inlining_assert_ready() { 3960 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 3961 } 3962 3963 void Compile::process_print_inlining() { 3964 bool do_print_inlining = print_inlining() || print_intrinsics(); 3965 if (do_print_inlining || log() != NULL) { 3966 // Print inlining message for candidates that we couldn't inline 3967 // for lack of space 3968 for (int i = 0; i < _late_inlines.length(); i++) { 3969 CallGenerator* cg = _late_inlines.at(i); 3970 if (!cg->is_mh_late_inline()) { 3971 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 3972 if (do_print_inlining) { 3973 cg->print_inlining_late(msg); 3974 } 3975 log_late_inline_failure(cg, msg); 3976 } 3977 } 3978 } 3979 if (do_print_inlining) { 3980 ResourceMark rm; 3981 stringStream ss; 3982 for (int i = 0; i < _print_inlining_list->length(); i++) { 3983 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 3984 } 3985 size_t end = ss.size(); 3986 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 3987 strncpy(_print_inlining_output, ss.base(), end+1); 3988 _print_inlining_output[end] = 0; 3989 } 3990 } 3991 3992 void Compile::dump_print_inlining() { 3993 if (_print_inlining_output != NULL) { 3994 tty->print_raw(_print_inlining_output); 3995 } 3996 } 3997 3998 void Compile::log_late_inline(CallGenerator* cg) { 3999 if (log() != NULL) { 4000 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4001 cg->unique_id()); 4002 JVMState* p = cg->call_node()->jvms(); 4003 while (p != NULL) { 4004 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4005 p = p->caller(); 4006 } 4007 log()->tail("late_inline"); 4008 } 4009 } 4010 4011 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4012 log_late_inline(cg); 4013 if (log() != NULL) { 4014 log()->inline_fail(msg); 4015 } 4016 } 4017 4018 void Compile::log_inline_id(CallGenerator* cg) { 4019 if (log() != NULL) { 4020 // The LogCompilation tool needs a unique way to identify late 4021 // inline call sites. This id must be unique for this call site in 4022 // this compilation. Try to have it unique across compilations as 4023 // well because it can be convenient when grepping through the log 4024 // file. 4025 // Distinguish OSR compilations from others in case CICountOSR is 4026 // on. 4027 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4028 cg->set_unique_id(id); 4029 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4030 } 4031 } 4032 4033 void Compile::log_inline_failure(const char* msg) { 4034 if (C->log() != NULL) { 4035 C->log()->inline_fail(msg); 4036 } 4037 } 4038 4039 4040 // Dump inlining replay data to the stream. 4041 // Don't change thread state and acquire any locks. 4042 void Compile::dump_inline_data(outputStream* out) { 4043 InlineTree* inl_tree = ilt(); 4044 if (inl_tree != NULL) { 4045 out->print(" inline %d", inl_tree->count()); 4046 inl_tree->dump_replay_data(out); 4047 } 4048 } 4049 4050 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4051 if (n1->Opcode() < n2->Opcode()) return -1; 4052 else if (n1->Opcode() > n2->Opcode()) return 1; 4053 4054 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())); 4055 for (uint i = 1; i < n1->req(); i++) { 4056 if (n1->in(i) < n2->in(i)) return -1; 4057 else if (n1->in(i) > n2->in(i)) return 1; 4058 } 4059 4060 return 0; 4061 } 4062 4063 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4064 Node* n1 = *n1p; 4065 Node* n2 = *n2p; 4066 4067 return cmp_expensive_nodes(n1, n2); 4068 } 4069 4070 void Compile::sort_expensive_nodes() { 4071 if (!expensive_nodes_sorted()) { 4072 _expensive_nodes->sort(cmp_expensive_nodes); 4073 } 4074 } 4075 4076 bool Compile::expensive_nodes_sorted() const { 4077 for (int i = 1; i < _expensive_nodes->length(); i++) { 4078 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4079 return false; 4080 } 4081 } 4082 return true; 4083 } 4084 4085 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4086 if (_expensive_nodes->length() == 0) { 4087 return false; 4088 } 4089 4090 assert(OptimizeExpensiveOps, "optimization off?"); 4091 4092 // Take this opportunity to remove dead nodes from the list 4093 int j = 0; 4094 for (int i = 0; i < _expensive_nodes->length(); i++) { 4095 Node* n = _expensive_nodes->at(i); 4096 if (!n->is_unreachable(igvn)) { 4097 assert(n->is_expensive(), "should be expensive"); 4098 _expensive_nodes->at_put(j, n); 4099 j++; 4100 } 4101 } 4102 _expensive_nodes->trunc_to(j); 4103 4104 // Then sort the list so that similar nodes are next to each other 4105 // and check for at least two nodes of identical kind with same data 4106 // inputs. 4107 sort_expensive_nodes(); 4108 4109 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4110 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4111 return true; 4112 } 4113 } 4114 4115 return false; 4116 } 4117 4118 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4119 if (_expensive_nodes->length() == 0) { 4120 return; 4121 } 4122 4123 assert(OptimizeExpensiveOps, "optimization off?"); 4124 4125 // Sort to bring similar nodes next to each other and clear the 4126 // control input of nodes for which there's only a single copy. 4127 sort_expensive_nodes(); 4128 4129 int j = 0; 4130 int identical = 0; 4131 int i = 0; 4132 bool modified = false; 4133 for (; i < _expensive_nodes->length()-1; i++) { 4134 assert(j <= i, "can't write beyond current index"); 4135 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4136 identical++; 4137 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4138 continue; 4139 } 4140 if (identical > 0) { 4141 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4142 identical = 0; 4143 } else { 4144 Node* n = _expensive_nodes->at(i); 4145 igvn.replace_input_of(n, 0, NULL); 4146 igvn.hash_insert(n); 4147 modified = true; 4148 } 4149 } 4150 if (identical > 0) { 4151 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4152 } else if (_expensive_nodes->length() >= 1) { 4153 Node* n = _expensive_nodes->at(i); 4154 igvn.replace_input_of(n, 0, NULL); 4155 igvn.hash_insert(n); 4156 modified = true; 4157 } 4158 _expensive_nodes->trunc_to(j); 4159 if (modified) { 4160 igvn.optimize(); 4161 } 4162 } 4163 4164 void Compile::add_expensive_node(Node * n) { 4165 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4166 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4167 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4168 if (OptimizeExpensiveOps) { 4169 _expensive_nodes->append(n); 4170 } else { 4171 // Clear control input and let IGVN optimize expensive nodes if 4172 // OptimizeExpensiveOps is off. 4173 n->set_req(0, NULL); 4174 } 4175 } 4176 4177 /** 4178 * Remove the speculative part of types and clean up the graph 4179 */ 4180 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4181 if (UseTypeSpeculation) { 4182 Unique_Node_List worklist; 4183 worklist.push(root()); 4184 int modified = 0; 4185 // Go over all type nodes that carry a speculative type, drop the 4186 // speculative part of the type and enqueue the node for an igvn 4187 // which may optimize it out. 4188 for (uint next = 0; next < worklist.size(); ++next) { 4189 Node *n = worklist.at(next); 4190 if (n->is_Type()) { 4191 TypeNode* tn = n->as_Type(); 4192 const Type* t = tn->type(); 4193 const Type* t_no_spec = t->remove_speculative(); 4194 if (t_no_spec != t) { 4195 bool in_hash = igvn.hash_delete(n); 4196 assert(in_hash, "node should be in igvn hash table"); 4197 tn->set_type(t_no_spec); 4198 igvn.hash_insert(n); 4199 igvn._worklist.push(n); // give it a chance to go away 4200 modified++; 4201 } 4202 } 4203 uint max = n->len(); 4204 for( uint i = 0; i < max; ++i ) { 4205 Node *m = n->in(i); 4206 if (not_a_node(m)) continue; 4207 worklist.push(m); 4208 } 4209 } 4210 // Drop the speculative part of all types in the igvn's type table 4211 igvn.remove_speculative_types(); 4212 if (modified > 0) { 4213 igvn.optimize(); 4214 } 4215 #ifdef ASSERT 4216 // Verify that after the IGVN is over no speculative type has resurfaced 4217 worklist.clear(); 4218 worklist.push(root()); 4219 for (uint next = 0; next < worklist.size(); ++next) { 4220 Node *n = worklist.at(next); 4221 const Type* t = igvn.type_or_null(n); 4222 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4223 if (n->is_Type()) { 4224 t = n->as_Type()->type(); 4225 assert(t == t->remove_speculative(), "no more speculative types"); 4226 } 4227 uint max = n->len(); 4228 for( uint i = 0; i < max; ++i ) { 4229 Node *m = n->in(i); 4230 if (not_a_node(m)) continue; 4231 worklist.push(m); 4232 } 4233 } 4234 igvn.check_no_speculative_types(); 4235 #endif 4236 } 4237 } 4238 4239 // Auxiliary method to support randomized stressing/fuzzing. 4240 // 4241 // This method can be called the arbitrary number of times, with current count 4242 // as the argument. The logic allows selecting a single candidate from the 4243 // running list of candidates as follows: 4244 // int count = 0; 4245 // Cand* selected = null; 4246 // while(cand = cand->next()) { 4247 // if (randomized_select(++count)) { 4248 // selected = cand; 4249 // } 4250 // } 4251 // 4252 // Including count equalizes the chances any candidate is "selected". 4253 // This is useful when we don't have the complete list of candidates to choose 4254 // from uniformly. In this case, we need to adjust the randomicity of the 4255 // selection, or else we will end up biasing the selection towards the latter 4256 // candidates. 4257 // 4258 // Quick back-envelope calculation shows that for the list of n candidates 4259 // the equal probability for the candidate to persist as "best" can be 4260 // achieved by replacing it with "next" k-th candidate with the probability 4261 // of 1/k. It can be easily shown that by the end of the run, the 4262 // probability for any candidate is converged to 1/n, thus giving the 4263 // uniform distribution among all the candidates. 4264 // 4265 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4266 #define RANDOMIZED_DOMAIN_POW 29 4267 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4268 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4269 bool Compile::randomized_select(int count) { 4270 assert(count > 0, "only positive"); 4271 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4272 }