1 /* 2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/block.hpp" 38 #include "opto/c2compiler.hpp" 39 #include "opto/callGenerator.hpp" 40 #include "opto/callnode.hpp" 41 #include "opto/cfgnode.hpp" 42 #include "opto/chaitin.hpp" 43 #include "opto/compile.hpp" 44 #include "opto/connode.hpp" 45 #include "opto/convertnode.hpp" 46 #include "opto/divnode.hpp" 47 #include "opto/escape.hpp" 48 #include "opto/idealGraphPrinter.hpp" 49 #include "opto/loopnode.hpp" 50 #include "opto/machnode.hpp" 51 #include "opto/macro.hpp" 52 #include "opto/matcher.hpp" 53 #include "opto/mathexactnode.hpp" 54 #include "opto/memnode.hpp" 55 #include "opto/mulnode.hpp" 56 #include "opto/narrowptrnode.hpp" 57 #include "opto/node.hpp" 58 #include "opto/opcodes.hpp" 59 #include "opto/output.hpp" 60 #include "opto/parse.hpp" 61 #include "opto/phaseX.hpp" 62 #include "opto/rootnode.hpp" 63 #include "opto/runtime.hpp" 64 #include "opto/stringopts.hpp" 65 #include "opto/type.hpp" 66 #include "opto/vectornode.hpp" 67 #include "runtime/arguments.hpp" 68 #include "runtime/sharedRuntime.hpp" 69 #include "runtime/signature.hpp" 70 #include "runtime/stubRoutines.hpp" 71 #include "runtime/timer.hpp" 72 #include "utilities/copy.hpp" 73 74 75 // -------------------- Compile::mach_constant_base_node ----------------------- 76 // Constant table base node singleton. 77 MachConstantBaseNode* Compile::mach_constant_base_node() { 78 if (_mach_constant_base_node == NULL) { 79 _mach_constant_base_node = new MachConstantBaseNode(); 80 _mach_constant_base_node->add_req(C->root()); 81 } 82 return _mach_constant_base_node; 83 } 84 85 86 /// Support for intrinsics. 87 88 // Return the index at which m must be inserted (or already exists). 89 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 90 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 91 #ifdef ASSERT 92 for (int i = 1; i < _intrinsics->length(); i++) { 93 CallGenerator* cg1 = _intrinsics->at(i-1); 94 CallGenerator* cg2 = _intrinsics->at(i); 95 assert(cg1->method() != cg2->method() 96 ? cg1->method() < cg2->method() 97 : cg1->is_virtual() < cg2->is_virtual(), 98 "compiler intrinsics list must stay sorted"); 99 } 100 #endif 101 // Binary search sorted list, in decreasing intervals [lo, hi]. 102 int lo = 0, hi = _intrinsics->length()-1; 103 while (lo <= hi) { 104 int mid = (uint)(hi + lo) / 2; 105 ciMethod* mid_m = _intrinsics->at(mid)->method(); 106 if (m < mid_m) { 107 hi = mid-1; 108 } else if (m > mid_m) { 109 lo = mid+1; 110 } else { 111 // look at minor sort key 112 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 113 if (is_virtual < mid_virt) { 114 hi = mid-1; 115 } else if (is_virtual > mid_virt) { 116 lo = mid+1; 117 } else { 118 return mid; // exact match 119 } 120 } 121 } 122 return lo; // inexact match 123 } 124 125 void Compile::register_intrinsic(CallGenerator* cg) { 126 if (_intrinsics == NULL) { 127 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 128 } 129 // This code is stolen from ciObjectFactory::insert. 130 // Really, GrowableArray should have methods for 131 // insert_at, remove_at, and binary_search. 132 int len = _intrinsics->length(); 133 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 134 if (index == len) { 135 _intrinsics->append(cg); 136 } else { 137 #ifdef ASSERT 138 CallGenerator* oldcg = _intrinsics->at(index); 139 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 140 #endif 141 _intrinsics->append(_intrinsics->at(len-1)); 142 int pos; 143 for (pos = len-2; pos >= index; pos--) { 144 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 145 } 146 _intrinsics->at_put(index, cg); 147 } 148 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 149 } 150 151 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 152 assert(m->is_loaded(), "don't try this on unloaded methods"); 153 if (_intrinsics != NULL) { 154 int index = intrinsic_insertion_index(m, is_virtual); 155 if (index < _intrinsics->length() 156 && _intrinsics->at(index)->method() == m 157 && _intrinsics->at(index)->is_virtual() == is_virtual) { 158 return _intrinsics->at(index); 159 } 160 } 161 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 162 if (m->intrinsic_id() != vmIntrinsics::_none && 163 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 164 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 165 if (cg != NULL) { 166 // Save it for next time: 167 register_intrinsic(cg); 168 return cg; 169 } else { 170 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 171 } 172 } 173 return NULL; 174 } 175 176 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 177 // in library_call.cpp. 178 179 180 #ifndef PRODUCT 181 // statistics gathering... 182 183 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 184 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 185 186 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 187 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 188 int oflags = _intrinsic_hist_flags[id]; 189 assert(flags != 0, "what happened?"); 190 if (is_virtual) { 191 flags |= _intrinsic_virtual; 192 } 193 bool changed = (flags != oflags); 194 if ((flags & _intrinsic_worked) != 0) { 195 juint count = (_intrinsic_hist_count[id] += 1); 196 if (count == 1) { 197 changed = true; // first time 198 } 199 // increment the overall count also: 200 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 201 } 202 if (changed) { 203 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 204 // Something changed about the intrinsic's virtuality. 205 if ((flags & _intrinsic_virtual) != 0) { 206 // This is the first use of this intrinsic as a virtual call. 207 if (oflags != 0) { 208 // We already saw it as a non-virtual, so note both cases. 209 flags |= _intrinsic_both; 210 } 211 } else if ((oflags & _intrinsic_both) == 0) { 212 // This is the first use of this intrinsic as a non-virtual 213 flags |= _intrinsic_both; 214 } 215 } 216 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 217 } 218 // update the overall flags also: 219 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 220 return changed; 221 } 222 223 static char* format_flags(int flags, char* buf) { 224 buf[0] = 0; 225 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 226 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 227 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 228 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 229 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 230 if (buf[0] == 0) strcat(buf, ","); 231 assert(buf[0] == ',', "must be"); 232 return &buf[1]; 233 } 234 235 void Compile::print_intrinsic_statistics() { 236 char flagsbuf[100]; 237 ttyLocker ttyl; 238 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 239 tty->print_cr("Compiler intrinsic usage:"); 240 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 241 if (total == 0) total = 1; // avoid div0 in case of no successes 242 #define PRINT_STAT_LINE(name, c, f) \ 243 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 244 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 245 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 246 int flags = _intrinsic_hist_flags[id]; 247 juint count = _intrinsic_hist_count[id]; 248 if ((flags | count) != 0) { 249 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 250 } 251 } 252 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 253 if (xtty != NULL) xtty->tail("statistics"); 254 } 255 256 void Compile::print_statistics() { 257 { ttyLocker ttyl; 258 if (xtty != NULL) xtty->head("statistics type='opto'"); 259 Parse::print_statistics(); 260 PhaseCCP::print_statistics(); 261 PhaseRegAlloc::print_statistics(); 262 Scheduling::print_statistics(); 263 PhasePeephole::print_statistics(); 264 PhaseIdealLoop::print_statistics(); 265 if (xtty != NULL) xtty->tail("statistics"); 266 } 267 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 268 // put this under its own <statistics> element. 269 print_intrinsic_statistics(); 270 } 271 } 272 #endif //PRODUCT 273 274 // Support for bundling info 275 Bundle* Compile::node_bundling(const Node *n) { 276 assert(valid_bundle_info(n), "oob"); 277 return &_node_bundling_base[n->_idx]; 278 } 279 280 bool Compile::valid_bundle_info(const Node *n) { 281 return (_node_bundling_limit > n->_idx); 282 } 283 284 285 void Compile::gvn_replace_by(Node* n, Node* nn) { 286 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 287 Node* use = n->last_out(i); 288 bool is_in_table = initial_gvn()->hash_delete(use); 289 uint uses_found = 0; 290 for (uint j = 0; j < use->len(); j++) { 291 if (use->in(j) == n) { 292 if (j < use->req()) 293 use->set_req(j, nn); 294 else 295 use->set_prec(j, nn); 296 uses_found++; 297 } 298 } 299 if (is_in_table) { 300 // reinsert into table 301 initial_gvn()->hash_find_insert(use); 302 } 303 record_for_igvn(use); 304 i -= uses_found; // we deleted 1 or more copies of this edge 305 } 306 } 307 308 309 static inline bool not_a_node(const Node* n) { 310 if (n == NULL) return true; 311 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 312 if (*(address*)n == badAddress) return true; // kill by Node::destruct 313 return false; 314 } 315 316 // Identify all nodes that are reachable from below, useful. 317 // Use breadth-first pass that records state in a Unique_Node_List, 318 // recursive traversal is slower. 319 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 320 int estimated_worklist_size = unique(); 321 useful.map( estimated_worklist_size, NULL ); // preallocate space 322 323 // Initialize worklist 324 if (root() != NULL) { useful.push(root()); } 325 // If 'top' is cached, declare it useful to preserve cached node 326 if( cached_top_node() ) { useful.push(cached_top_node()); } 327 328 // Push all useful nodes onto the list, breadthfirst 329 for( uint next = 0; next < useful.size(); ++next ) { 330 assert( next < unique(), "Unique useful nodes < total nodes"); 331 Node *n = useful.at(next); 332 uint max = n->len(); 333 for( uint i = 0; i < max; ++i ) { 334 Node *m = n->in(i); 335 if (not_a_node(m)) continue; 336 useful.push(m); 337 } 338 } 339 } 340 341 // Update dead_node_list with any missing dead nodes using useful 342 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 343 void Compile::update_dead_node_list(Unique_Node_List &useful) { 344 uint max_idx = unique(); 345 VectorSet& useful_node_set = useful.member_set(); 346 347 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 348 // If node with index node_idx is not in useful set, 349 // mark it as dead in dead node list. 350 if (! useful_node_set.test(node_idx) ) { 351 record_dead_node(node_idx); 352 } 353 } 354 } 355 356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 357 int shift = 0; 358 for (int i = 0; i < inlines->length(); i++) { 359 CallGenerator* cg = inlines->at(i); 360 CallNode* call = cg->call_node(); 361 if (shift > 0) { 362 inlines->at_put(i-shift, cg); 363 } 364 if (!useful.member(call)) { 365 shift++; 366 } 367 } 368 inlines->trunc_to(inlines->length()-shift); 369 } 370 371 // Disconnect all useless nodes by disconnecting those at the boundary. 372 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 373 uint next = 0; 374 while (next < useful.size()) { 375 Node *n = useful.at(next++); 376 if (n->is_SafePoint()) { 377 // We're done with a parsing phase. Replaced nodes are not valid 378 // beyond that point. 379 n->as_SafePoint()->delete_replaced_nodes(); 380 } 381 // Use raw traversal of out edges since this code removes out edges 382 int max = n->outcnt(); 383 for (int j = 0; j < max; ++j) { 384 Node* child = n->raw_out(j); 385 if (! useful.member(child)) { 386 assert(!child->is_top() || child != top(), 387 "If top is cached in Compile object it is in useful list"); 388 // Only need to remove this out-edge to the useless node 389 n->raw_del_out(j); 390 --j; 391 --max; 392 } 393 } 394 if (n->outcnt() == 1 && n->has_special_unique_user()) { 395 record_for_igvn(n->unique_out()); 396 } 397 } 398 // Remove useless macro and predicate opaq nodes 399 for (int i = C->macro_count()-1; i >= 0; i--) { 400 Node* n = C->macro_node(i); 401 if (!useful.member(n)) { 402 remove_macro_node(n); 403 } 404 } 405 // Remove useless expensive node 406 for (int i = C->expensive_count()-1; i >= 0; i--) { 407 Node* n = C->expensive_node(i); 408 if (!useful.member(n)) { 409 remove_expensive_node(n); 410 } 411 } 412 // clean up the late inline lists 413 remove_useless_late_inlines(&_string_late_inlines, useful); 414 remove_useless_late_inlines(&_boxing_late_inlines, useful); 415 remove_useless_late_inlines(&_late_inlines, useful); 416 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 417 } 418 419 //------------------------------frame_size_in_words----------------------------- 420 // frame_slots in units of words 421 int Compile::frame_size_in_words() const { 422 // shift is 0 in LP32 and 1 in LP64 423 const int shift = (LogBytesPerWord - LogBytesPerInt); 424 int words = _frame_slots >> shift; 425 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 426 return words; 427 } 428 429 // To bang the stack of this compiled method we use the stack size 430 // that the interpreter would need in case of a deoptimization. This 431 // removes the need to bang the stack in the deoptimization blob which 432 // in turn simplifies stack overflow handling. 433 int Compile::bang_size_in_bytes() const { 434 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 435 } 436 437 // ============================================================================ 438 //------------------------------CompileWrapper--------------------------------- 439 class CompileWrapper : public StackObj { 440 Compile *const _compile; 441 public: 442 CompileWrapper(Compile* compile); 443 444 ~CompileWrapper(); 445 }; 446 447 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 448 // the Compile* pointer is stored in the current ciEnv: 449 ciEnv* env = compile->env(); 450 assert(env == ciEnv::current(), "must already be a ciEnv active"); 451 assert(env->compiler_data() == NULL, "compile already active?"); 452 env->set_compiler_data(compile); 453 assert(compile == Compile::current(), "sanity"); 454 455 compile->set_type_dict(NULL); 456 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 457 compile->clone_map().set_clone_idx(0); 458 compile->set_type_hwm(NULL); 459 compile->set_type_last_size(0); 460 compile->set_last_tf(NULL, NULL); 461 compile->set_indexSet_arena(NULL); 462 compile->set_indexSet_free_block_list(NULL); 463 compile->init_type_arena(); 464 Type::Initialize(compile); 465 _compile->set_scratch_buffer_blob(NULL); 466 _compile->begin_method(); 467 _compile->clone_map().set_debug(_compile->has_method() && _compile->method_has_option(_compile->clone_map().debug_option_name)); 468 } 469 CompileWrapper::~CompileWrapper() { 470 _compile->end_method(); 471 if (_compile->scratch_buffer_blob() != NULL) 472 BufferBlob::free(_compile->scratch_buffer_blob()); 473 _compile->env()->set_compiler_data(NULL); 474 } 475 476 477 //----------------------------print_compile_messages--------------------------- 478 void Compile::print_compile_messages() { 479 #ifndef PRODUCT 480 // Check if recompiling 481 if (_subsume_loads == false && PrintOpto) { 482 // Recompiling without allowing machine instructions to subsume loads 483 tty->print_cr("*********************************************************"); 484 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 485 tty->print_cr("*********************************************************"); 486 } 487 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 488 // Recompiling without escape analysis 489 tty->print_cr("*********************************************************"); 490 tty->print_cr("** Bailout: Recompile without escape analysis **"); 491 tty->print_cr("*********************************************************"); 492 } 493 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 494 // Recompiling without boxing elimination 495 tty->print_cr("*********************************************************"); 496 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 497 tty->print_cr("*********************************************************"); 498 } 499 if (env()->break_at_compile()) { 500 // Open the debugger when compiling this method. 501 tty->print("### Breaking when compiling: "); 502 method()->print_short_name(); 503 tty->cr(); 504 BREAKPOINT; 505 } 506 507 if( PrintOpto ) { 508 if (is_osr_compilation()) { 509 tty->print("[OSR]%3d", _compile_id); 510 } else { 511 tty->print("%3d", _compile_id); 512 } 513 } 514 #endif 515 } 516 517 518 //-----------------------init_scratch_buffer_blob------------------------------ 519 // Construct a temporary BufferBlob and cache it for this compile. 520 void Compile::init_scratch_buffer_blob(int const_size) { 521 // If there is already a scratch buffer blob allocated and the 522 // constant section is big enough, use it. Otherwise free the 523 // current and allocate a new one. 524 BufferBlob* blob = scratch_buffer_blob(); 525 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 526 // Use the current blob. 527 } else { 528 if (blob != NULL) { 529 BufferBlob::free(blob); 530 } 531 532 ResourceMark rm; 533 _scratch_const_size = const_size; 534 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 535 blob = BufferBlob::create("Compile::scratch_buffer", size); 536 // Record the buffer blob for next time. 537 set_scratch_buffer_blob(blob); 538 // Have we run out of code space? 539 if (scratch_buffer_blob() == NULL) { 540 // Let CompilerBroker disable further compilations. 541 record_failure("Not enough space for scratch buffer in CodeCache"); 542 return; 543 } 544 } 545 546 // Initialize the relocation buffers 547 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 548 set_scratch_locs_memory(locs_buf); 549 } 550 551 552 //-----------------------scratch_emit_size------------------------------------- 553 // Helper function that computes size by emitting code 554 uint Compile::scratch_emit_size(const Node* n) { 555 // Start scratch_emit_size section. 556 set_in_scratch_emit_size(true); 557 558 // Emit into a trash buffer and count bytes emitted. 559 // This is a pretty expensive way to compute a size, 560 // but it works well enough if seldom used. 561 // All common fixed-size instructions are given a size 562 // method by the AD file. 563 // Note that the scratch buffer blob and locs memory are 564 // allocated at the beginning of the compile task, and 565 // may be shared by several calls to scratch_emit_size. 566 // The allocation of the scratch buffer blob is particularly 567 // expensive, since it has to grab the code cache lock. 568 BufferBlob* blob = this->scratch_buffer_blob(); 569 assert(blob != NULL, "Initialize BufferBlob at start"); 570 assert(blob->size() > MAX_inst_size, "sanity"); 571 relocInfo* locs_buf = scratch_locs_memory(); 572 address blob_begin = blob->content_begin(); 573 address blob_end = (address)locs_buf; 574 assert(blob->content_contains(blob_end), "sanity"); 575 CodeBuffer buf(blob_begin, blob_end - blob_begin); 576 buf.initialize_consts_size(_scratch_const_size); 577 buf.initialize_stubs_size(MAX_stubs_size); 578 assert(locs_buf != NULL, "sanity"); 579 int lsize = MAX_locs_size / 3; 580 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 581 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 582 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 583 584 // Do the emission. 585 586 Label fakeL; // Fake label for branch instructions. 587 Label* saveL = NULL; 588 uint save_bnum = 0; 589 bool is_branch = n->is_MachBranch(); 590 if (is_branch) { 591 MacroAssembler masm(&buf); 592 masm.bind(fakeL); 593 n->as_MachBranch()->save_label(&saveL, &save_bnum); 594 n->as_MachBranch()->label_set(&fakeL, 0); 595 } 596 n->emit(buf, this->regalloc()); 597 598 // Emitting into the scratch buffer should not fail 599 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 600 601 if (is_branch) // Restore label. 602 n->as_MachBranch()->label_set(saveL, save_bnum); 603 604 // End scratch_emit_size section. 605 set_in_scratch_emit_size(false); 606 607 return buf.insts_size(); 608 } 609 610 611 // ============================================================================ 612 //------------------------------Compile standard------------------------------- 613 debug_only( int Compile::_debug_idx = 100000; ) 614 615 // Compile a method. entry_bci is -1 for normal compilations and indicates 616 // the continuation bci for on stack replacement. 617 618 619 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 620 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 621 : Phase(Compiler), 622 _env(ci_env), 623 _log(ci_env->log()), 624 _compile_id(ci_env->compile_id()), 625 _save_argument_registers(false), 626 _stub_name(NULL), 627 _stub_function(NULL), 628 _stub_entry_point(NULL), 629 _method(target), 630 _entry_bci(osr_bci), 631 _initial_gvn(NULL), 632 _for_igvn(NULL), 633 _warm_calls(NULL), 634 _subsume_loads(subsume_loads), 635 _do_escape_analysis(do_escape_analysis), 636 _eliminate_boxing(eliminate_boxing), 637 _failure_reason(NULL), 638 _code_buffer("Compile::Fill_buffer"), 639 _orig_pc_slot(0), 640 _orig_pc_slot_offset_in_bytes(0), 641 _has_method_handle_invokes(false), 642 _mach_constant_base_node(NULL), 643 _node_bundling_limit(0), 644 _node_bundling_base(NULL), 645 _java_calls(0), 646 _inner_loops(0), 647 _scratch_const_size(-1), 648 _in_scratch_emit_size(false), 649 _dead_node_list(comp_arena()), 650 _dead_node_count(0), 651 #ifndef PRODUCT 652 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 653 _in_dump_cnt(0), 654 _printer(IdealGraphPrinter::printer()), 655 #endif 656 _congraph(NULL), 657 _comp_arena(mtCompiler), 658 _node_arena(mtCompiler), 659 _old_arena(mtCompiler), 660 _Compile_types(mtCompiler), 661 _replay_inline_data(NULL), 662 _late_inlines(comp_arena(), 2, 0, NULL), 663 _string_late_inlines(comp_arena(), 2, 0, NULL), 664 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 665 _late_inlines_pos(0), 666 _number_of_mh_late_inlines(0), 667 _inlining_progress(false), 668 _inlining_incrementally(false), 669 _print_inlining_list(NULL), 670 _print_inlining_stream(NULL), 671 _print_inlining_idx(0), 672 _print_inlining_output(NULL), 673 _interpreter_frame_size(0), 674 _max_node_limit(MaxNodeLimit) { 675 C = this; 676 677 CompileWrapper cw(this); 678 679 if (CITimeVerbose) { 680 tty->print(" "); 681 target->holder()->name()->print(); 682 tty->print("."); 683 target->print_short_name(); 684 tty->print(" "); 685 } 686 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 687 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 688 689 #ifndef PRODUCT 690 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 691 if (!print_opto_assembly) { 692 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 693 if (print_assembly && !Disassembler::can_decode()) { 694 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 695 print_opto_assembly = true; 696 } 697 } 698 set_print_assembly(print_opto_assembly); 699 set_parsed_irreducible_loop(false); 700 701 if (method()->has_option("ReplayInline")) { 702 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 703 } 704 #endif 705 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 706 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 707 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 708 709 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 710 // Make sure the method being compiled gets its own MDO, 711 // so we can at least track the decompile_count(). 712 // Need MDO to record RTM code generation state. 713 method()->ensure_method_data(); 714 } 715 716 Init(::AliasLevel); 717 718 719 print_compile_messages(); 720 721 _ilt = InlineTree::build_inline_tree_root(); 722 723 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 724 assert(num_alias_types() >= AliasIdxRaw, ""); 725 726 #define MINIMUM_NODE_HASH 1023 727 // Node list that Iterative GVN will start with 728 Unique_Node_List for_igvn(comp_arena()); 729 set_for_igvn(&for_igvn); 730 731 // GVN that will be run immediately on new nodes 732 uint estimated_size = method()->code_size()*4+64; 733 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 734 PhaseGVN gvn(node_arena(), estimated_size); 735 set_initial_gvn(&gvn); 736 737 print_inlining_init(); 738 { // Scope for timing the parser 739 TracePhase tp("parse", &timers[_t_parser]); 740 741 // Put top into the hash table ASAP. 742 initial_gvn()->transform_no_reclaim(top()); 743 744 // Set up tf(), start(), and find a CallGenerator. 745 CallGenerator* cg = NULL; 746 if (is_osr_compilation()) { 747 const TypeTuple *domain = StartOSRNode::osr_domain(); 748 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 749 init_tf(TypeFunc::make(domain, range)); 750 StartNode* s = new StartOSRNode(root(), domain); 751 initial_gvn()->set_type_bottom(s); 752 init_start(s); 753 cg = CallGenerator::for_osr(method(), entry_bci()); 754 } else { 755 // Normal case. 756 init_tf(TypeFunc::make(method())); 757 StartNode* s = new StartNode(root(), tf()->domain()); 758 initial_gvn()->set_type_bottom(s); 759 init_start(s); 760 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 761 // With java.lang.ref.reference.get() we must go through the 762 // intrinsic when G1 is enabled - even when get() is the root 763 // method of the compile - so that, if necessary, the value in 764 // the referent field of the reference object gets recorded by 765 // the pre-barrier code. 766 // Specifically, if G1 is enabled, the value in the referent 767 // field is recorded by the G1 SATB pre barrier. This will 768 // result in the referent being marked live and the reference 769 // object removed from the list of discovered references during 770 // reference processing. 771 cg = find_intrinsic(method(), false); 772 } 773 if (cg == NULL) { 774 float past_uses = method()->interpreter_invocation_count(); 775 float expected_uses = past_uses; 776 cg = CallGenerator::for_inline(method(), expected_uses); 777 } 778 } 779 if (failing()) return; 780 if (cg == NULL) { 781 record_method_not_compilable_all_tiers("cannot parse method"); 782 return; 783 } 784 JVMState* jvms = build_start_state(start(), tf()); 785 if ((jvms = cg->generate(jvms)) == NULL) { 786 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 787 record_method_not_compilable("method parse failed"); 788 } 789 return; 790 } 791 GraphKit kit(jvms); 792 793 if (!kit.stopped()) { 794 // Accept return values, and transfer control we know not where. 795 // This is done by a special, unique ReturnNode bound to root. 796 return_values(kit.jvms()); 797 } 798 799 if (kit.has_exceptions()) { 800 // Any exceptions that escape from this call must be rethrown 801 // to whatever caller is dynamically above us on the stack. 802 // This is done by a special, unique RethrowNode bound to root. 803 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 804 } 805 806 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 807 808 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 809 inline_string_calls(true); 810 } 811 812 if (failing()) return; 813 814 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 815 816 // Remove clutter produced by parsing. 817 if (!failing()) { 818 ResourceMark rm; 819 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 820 } 821 } 822 823 // Note: Large methods are capped off in do_one_bytecode(). 824 if (failing()) return; 825 826 // After parsing, node notes are no longer automagic. 827 // They must be propagated by register_new_node_with_optimizer(), 828 // clone(), or the like. 829 set_default_node_notes(NULL); 830 831 for (;;) { 832 int successes = Inline_Warm(); 833 if (failing()) return; 834 if (successes == 0) break; 835 } 836 837 // Drain the list. 838 Finish_Warm(); 839 #ifndef PRODUCT 840 if (_printer && _printer->should_print(_method)) { 841 _printer->print_inlining(this); 842 } 843 #endif 844 845 if (failing()) return; 846 NOT_PRODUCT( verify_graph_edges(); ) 847 848 // Now optimize 849 Optimize(); 850 if (failing()) return; 851 NOT_PRODUCT( verify_graph_edges(); ) 852 853 #ifndef PRODUCT 854 if (PrintIdeal) { 855 ttyLocker ttyl; // keep the following output all in one block 856 // This output goes directly to the tty, not the compiler log. 857 // To enable tools to match it up with the compilation activity, 858 // be sure to tag this tty output with the compile ID. 859 if (xtty != NULL) { 860 xtty->head("ideal compile_id='%d'%s", compile_id(), 861 is_osr_compilation() ? " compile_kind='osr'" : 862 ""); 863 } 864 root()->dump(9999); 865 if (xtty != NULL) { 866 xtty->tail("ideal"); 867 } 868 } 869 #endif 870 871 NOT_PRODUCT( verify_barriers(); ) 872 873 // Dump compilation data to replay it. 874 if (method()->has_option("DumpReplay")) { 875 env()->dump_replay_data(_compile_id); 876 } 877 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 878 env()->dump_inline_data(_compile_id); 879 } 880 881 // Now that we know the size of all the monitors we can add a fixed slot 882 // for the original deopt pc. 883 884 _orig_pc_slot = fixed_slots(); 885 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 886 set_fixed_slots(next_slot); 887 888 // Compute when to use implicit null checks. Used by matching trap based 889 // nodes and NullCheck optimization. 890 set_allowed_deopt_reasons(); 891 892 // Now generate code 893 Code_Gen(); 894 if (failing()) return; 895 896 // Check if we want to skip execution of all compiled code. 897 { 898 #ifndef PRODUCT 899 if (OptoNoExecute) { 900 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 901 return; 902 } 903 #endif 904 TracePhase tp("install_code", &timers[_t_registerMethod]); 905 906 if (is_osr_compilation()) { 907 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 908 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 909 } else { 910 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 911 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 912 } 913 914 env()->register_method(_method, _entry_bci, 915 &_code_offsets, 916 _orig_pc_slot_offset_in_bytes, 917 code_buffer(), 918 frame_size_in_words(), _oop_map_set, 919 &_handler_table, &_inc_table, 920 compiler, 921 env()->comp_level(), 922 has_unsafe_access(), 923 SharedRuntime::is_wide_vector(max_vector_size()), 924 rtm_state() 925 ); 926 927 if (log() != NULL) // Print code cache state into compiler log 928 log()->code_cache_state(); 929 } 930 } 931 932 //------------------------------Compile---------------------------------------- 933 // Compile a runtime stub 934 Compile::Compile( ciEnv* ci_env, 935 TypeFunc_generator generator, 936 address stub_function, 937 const char *stub_name, 938 int is_fancy_jump, 939 bool pass_tls, 940 bool save_arg_registers, 941 bool return_pc ) 942 : Phase(Compiler), 943 _env(ci_env), 944 _log(ci_env->log()), 945 _compile_id(0), 946 _save_argument_registers(save_arg_registers), 947 _method(NULL), 948 _stub_name(stub_name), 949 _stub_function(stub_function), 950 _stub_entry_point(NULL), 951 _entry_bci(InvocationEntryBci), 952 _initial_gvn(NULL), 953 _for_igvn(NULL), 954 _warm_calls(NULL), 955 _orig_pc_slot(0), 956 _orig_pc_slot_offset_in_bytes(0), 957 _subsume_loads(true), 958 _do_escape_analysis(false), 959 _eliminate_boxing(false), 960 _failure_reason(NULL), 961 _code_buffer("Compile::Fill_buffer"), 962 _has_method_handle_invokes(false), 963 _mach_constant_base_node(NULL), 964 _node_bundling_limit(0), 965 _node_bundling_base(NULL), 966 _java_calls(0), 967 _inner_loops(0), 968 #ifndef PRODUCT 969 _trace_opto_output(TraceOptoOutput), 970 _in_dump_cnt(0), 971 _printer(NULL), 972 #endif 973 _comp_arena(mtCompiler), 974 _node_arena(mtCompiler), 975 _old_arena(mtCompiler), 976 _Compile_types(mtCompiler), 977 _dead_node_list(comp_arena()), 978 _dead_node_count(0), 979 _congraph(NULL), 980 _replay_inline_data(NULL), 981 _number_of_mh_late_inlines(0), 982 _inlining_progress(false), 983 _inlining_incrementally(false), 984 _print_inlining_list(NULL), 985 _print_inlining_stream(NULL), 986 _print_inlining_idx(0), 987 _print_inlining_output(NULL), 988 _allowed_reasons(0), 989 _interpreter_frame_size(0), 990 _max_node_limit(MaxNodeLimit) { 991 C = this; 992 993 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 994 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 995 996 #ifndef PRODUCT 997 set_print_assembly(PrintFrameConverterAssembly); 998 set_parsed_irreducible_loop(false); 999 #endif 1000 set_has_irreducible_loop(false); // no loops 1001 1002 CompileWrapper cw(this); 1003 Init(/*AliasLevel=*/ 0); 1004 init_tf((*generator)()); 1005 1006 { 1007 // The following is a dummy for the sake of GraphKit::gen_stub 1008 Unique_Node_List for_igvn(comp_arena()); 1009 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1010 PhaseGVN gvn(Thread::current()->resource_area(),255); 1011 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1012 gvn.transform_no_reclaim(top()); 1013 1014 GraphKit kit; 1015 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1016 } 1017 1018 NOT_PRODUCT( verify_graph_edges(); ) 1019 Code_Gen(); 1020 if (failing()) return; 1021 1022 1023 // Entry point will be accessed using compile->stub_entry_point(); 1024 if (code_buffer() == NULL) { 1025 Matcher::soft_match_failure(); 1026 } else { 1027 if (PrintAssembly && (WizardMode || Verbose)) 1028 tty->print_cr("### Stub::%s", stub_name); 1029 1030 if (!failing()) { 1031 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1032 1033 // Make the NMethod 1034 // For now we mark the frame as never safe for profile stackwalking 1035 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1036 code_buffer(), 1037 CodeOffsets::frame_never_safe, 1038 // _code_offsets.value(CodeOffsets::Frame_Complete), 1039 frame_size_in_words(), 1040 _oop_map_set, 1041 save_arg_registers); 1042 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1043 1044 _stub_entry_point = rs->entry_point(); 1045 } 1046 } 1047 } 1048 1049 //------------------------------Init------------------------------------------- 1050 // Prepare for a single compilation 1051 void Compile::Init(int aliaslevel) { 1052 _unique = 0; 1053 _regalloc = NULL; 1054 1055 _tf = NULL; // filled in later 1056 _top = NULL; // cached later 1057 _matcher = NULL; // filled in later 1058 _cfg = NULL; // filled in later 1059 1060 set_24_bit_selection_and_mode(Use24BitFP, false); 1061 1062 _node_note_array = NULL; 1063 _default_node_notes = NULL; 1064 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1065 1066 _immutable_memory = NULL; // filled in at first inquiry 1067 1068 // Globally visible Nodes 1069 // First set TOP to NULL to give safe behavior during creation of RootNode 1070 set_cached_top_node(NULL); 1071 set_root(new RootNode()); 1072 // Now that you have a Root to point to, create the real TOP 1073 set_cached_top_node( new ConNode(Type::TOP) ); 1074 set_recent_alloc(NULL, NULL); 1075 1076 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1077 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1078 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1079 env()->set_dependencies(new Dependencies(env())); 1080 1081 _fixed_slots = 0; 1082 set_has_split_ifs(false); 1083 set_has_loops(has_method() && method()->has_loops()); // first approximation 1084 set_has_stringbuilder(false); 1085 set_has_boxed_value(false); 1086 _trap_can_recompile = false; // no traps emitted yet 1087 _major_progress = true; // start out assuming good things will happen 1088 set_has_unsafe_access(false); 1089 set_max_vector_size(0); 1090 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1091 set_decompile_count(0); 1092 1093 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1094 set_num_loop_opts(LoopOptsCount); 1095 set_do_inlining(Inline); 1096 set_max_inline_size(MaxInlineSize); 1097 set_freq_inline_size(FreqInlineSize); 1098 set_do_scheduling(OptoScheduling); 1099 set_do_count_invocations(false); 1100 set_do_method_data_update(false); 1101 1102 set_do_vector_loop(false); 1103 1104 bool do_vector = false; 1105 if (AllowVectorizeOnDemand) { 1106 if (has_method() && (method()->has_option("Vectorize") || method()->has_option("VectorizeDebug"))) { 1107 set_do_vector_loop(true); 1108 } else if (has_method() && method()->name() != 0 && 1109 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1110 set_do_vector_loop(true); 1111 } 1112 #ifndef PRODUCT 1113 if (do_vector_loop() && Verbose) { 1114 tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii()); 1115 } 1116 #endif 1117 } 1118 1119 set_age_code(has_method() && method()->profile_aging()); 1120 set_rtm_state(NoRTM); // No RTM lock eliding by default 1121 method_has_option_value("MaxNodeLimit", _max_node_limit); 1122 #if INCLUDE_RTM_OPT 1123 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1124 int rtm_state = method()->method_data()->rtm_state(); 1125 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1126 // Don't generate RTM lock eliding code. 1127 set_rtm_state(NoRTM); 1128 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1129 // Generate RTM lock eliding code without abort ratio calculation code. 1130 set_rtm_state(UseRTM); 1131 } else if (UseRTMDeopt) { 1132 // Generate RTM lock eliding code and include abort ratio calculation 1133 // code if UseRTMDeopt is on. 1134 set_rtm_state(ProfileRTM); 1135 } 1136 } 1137 #endif 1138 if (debug_info()->recording_non_safepoints()) { 1139 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1140 (comp_arena(), 8, 0, NULL)); 1141 set_default_node_notes(Node_Notes::make(this)); 1142 } 1143 1144 // // -- Initialize types before each compile -- 1145 // // Update cached type information 1146 // if( _method && _method->constants() ) 1147 // Type::update_loaded_types(_method, _method->constants()); 1148 1149 // Init alias_type map. 1150 if (!_do_escape_analysis && aliaslevel == 3) 1151 aliaslevel = 2; // No unique types without escape analysis 1152 _AliasLevel = aliaslevel; 1153 const int grow_ats = 16; 1154 _max_alias_types = grow_ats; 1155 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1156 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1157 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1158 { 1159 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1160 } 1161 // Initialize the first few types. 1162 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1163 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1164 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1165 _num_alias_types = AliasIdxRaw+1; 1166 // Zero out the alias type cache. 1167 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1168 // A NULL adr_type hits in the cache right away. Preload the right answer. 1169 probe_alias_cache(NULL)->_index = AliasIdxTop; 1170 1171 _intrinsics = NULL; 1172 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1173 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1174 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1175 register_library_intrinsics(); 1176 } 1177 1178 //---------------------------init_start---------------------------------------- 1179 // Install the StartNode on this compile object. 1180 void Compile::init_start(StartNode* s) { 1181 if (failing()) 1182 return; // already failing 1183 assert(s == start(), ""); 1184 } 1185 1186 /** 1187 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1188 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1189 * the ideal graph. 1190 */ 1191 StartNode* Compile::start() const { 1192 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 1193 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1194 Node* start = root()->fast_out(i); 1195 if (start->is_Start()) { 1196 return start->as_Start(); 1197 } 1198 } 1199 fatal("Did not find Start node!"); 1200 return NULL; 1201 } 1202 1203 //-------------------------------immutable_memory------------------------------------- 1204 // Access immutable memory 1205 Node* Compile::immutable_memory() { 1206 if (_immutable_memory != NULL) { 1207 return _immutable_memory; 1208 } 1209 StartNode* s = start(); 1210 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1211 Node *p = s->fast_out(i); 1212 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1213 _immutable_memory = p; 1214 return _immutable_memory; 1215 } 1216 } 1217 ShouldNotReachHere(); 1218 return NULL; 1219 } 1220 1221 //----------------------set_cached_top_node------------------------------------ 1222 // Install the cached top node, and make sure Node::is_top works correctly. 1223 void Compile::set_cached_top_node(Node* tn) { 1224 if (tn != NULL) verify_top(tn); 1225 Node* old_top = _top; 1226 _top = tn; 1227 // Calling Node::setup_is_top allows the nodes the chance to adjust 1228 // their _out arrays. 1229 if (_top != NULL) _top->setup_is_top(); 1230 if (old_top != NULL) old_top->setup_is_top(); 1231 assert(_top == NULL || top()->is_top(), ""); 1232 } 1233 1234 #ifdef ASSERT 1235 uint Compile::count_live_nodes_by_graph_walk() { 1236 Unique_Node_List useful(comp_arena()); 1237 // Get useful node list by walking the graph. 1238 identify_useful_nodes(useful); 1239 return useful.size(); 1240 } 1241 1242 void Compile::print_missing_nodes() { 1243 1244 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1245 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1246 return; 1247 } 1248 1249 // This is an expensive function. It is executed only when the user 1250 // specifies VerifyIdealNodeCount option or otherwise knows the 1251 // additional work that needs to be done to identify reachable nodes 1252 // by walking the flow graph and find the missing ones using 1253 // _dead_node_list. 1254 1255 Unique_Node_List useful(comp_arena()); 1256 // Get useful node list by walking the graph. 1257 identify_useful_nodes(useful); 1258 1259 uint l_nodes = C->live_nodes(); 1260 uint l_nodes_by_walk = useful.size(); 1261 1262 if (l_nodes != l_nodes_by_walk) { 1263 if (_log != NULL) { 1264 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1265 _log->stamp(); 1266 _log->end_head(); 1267 } 1268 VectorSet& useful_member_set = useful.member_set(); 1269 int last_idx = l_nodes_by_walk; 1270 for (int i = 0; i < last_idx; i++) { 1271 if (useful_member_set.test(i)) { 1272 if (_dead_node_list.test(i)) { 1273 if (_log != NULL) { 1274 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1275 } 1276 if (PrintIdealNodeCount) { 1277 // Print the log message to tty 1278 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1279 useful.at(i)->dump(); 1280 } 1281 } 1282 } 1283 else if (! _dead_node_list.test(i)) { 1284 if (_log != NULL) { 1285 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1286 } 1287 if (PrintIdealNodeCount) { 1288 // Print the log message to tty 1289 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1290 } 1291 } 1292 } 1293 if (_log != NULL) { 1294 _log->tail("mismatched_nodes"); 1295 } 1296 } 1297 } 1298 void Compile::record_modified_node(Node* n) { 1299 if (_modified_nodes != NULL && !_inlining_incrementally && 1300 n->outcnt() != 0 && !n->is_Con()) { 1301 _modified_nodes->push(n); 1302 } 1303 } 1304 1305 void Compile::remove_modified_node(Node* n) { 1306 if (_modified_nodes != NULL) { 1307 _modified_nodes->remove(n); 1308 } 1309 } 1310 #endif 1311 1312 #ifndef PRODUCT 1313 void Compile::verify_top(Node* tn) const { 1314 if (tn != NULL) { 1315 assert(tn->is_Con(), "top node must be a constant"); 1316 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1317 assert(tn->in(0) != NULL, "must have live top node"); 1318 } 1319 } 1320 #endif 1321 1322 1323 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1324 1325 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1326 guarantee(arr != NULL, ""); 1327 int num_blocks = arr->length(); 1328 if (grow_by < num_blocks) grow_by = num_blocks; 1329 int num_notes = grow_by * _node_notes_block_size; 1330 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1331 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1332 while (num_notes > 0) { 1333 arr->append(notes); 1334 notes += _node_notes_block_size; 1335 num_notes -= _node_notes_block_size; 1336 } 1337 assert(num_notes == 0, "exact multiple, please"); 1338 } 1339 1340 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1341 if (source == NULL || dest == NULL) return false; 1342 1343 if (dest->is_Con()) 1344 return false; // Do not push debug info onto constants. 1345 1346 #ifdef ASSERT 1347 // Leave a bread crumb trail pointing to the original node: 1348 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1349 dest->set_debug_orig(source); 1350 } 1351 #endif 1352 1353 if (node_note_array() == NULL) 1354 return false; // Not collecting any notes now. 1355 1356 // This is a copy onto a pre-existing node, which may already have notes. 1357 // If both nodes have notes, do not overwrite any pre-existing notes. 1358 Node_Notes* source_notes = node_notes_at(source->_idx); 1359 if (source_notes == NULL || source_notes->is_clear()) return false; 1360 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1361 if (dest_notes == NULL || dest_notes->is_clear()) { 1362 return set_node_notes_at(dest->_idx, source_notes); 1363 } 1364 1365 Node_Notes merged_notes = (*source_notes); 1366 // The order of operations here ensures that dest notes will win... 1367 merged_notes.update_from(dest_notes); 1368 return set_node_notes_at(dest->_idx, &merged_notes); 1369 } 1370 1371 1372 //--------------------------allow_range_check_smearing------------------------- 1373 // Gating condition for coalescing similar range checks. 1374 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1375 // single covering check that is at least as strong as any of them. 1376 // If the optimization succeeds, the simplified (strengthened) range check 1377 // will always succeed. If it fails, we will deopt, and then give up 1378 // on the optimization. 1379 bool Compile::allow_range_check_smearing() const { 1380 // If this method has already thrown a range-check, 1381 // assume it was because we already tried range smearing 1382 // and it failed. 1383 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1384 return !already_trapped; 1385 } 1386 1387 1388 //------------------------------flatten_alias_type----------------------------- 1389 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1390 int offset = tj->offset(); 1391 TypePtr::PTR ptr = tj->ptr(); 1392 1393 // Known instance (scalarizable allocation) alias only with itself. 1394 bool is_known_inst = tj->isa_oopptr() != NULL && 1395 tj->is_oopptr()->is_known_instance(); 1396 1397 // Process weird unsafe references. 1398 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1399 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1400 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1401 tj = TypeOopPtr::BOTTOM; 1402 ptr = tj->ptr(); 1403 offset = tj->offset(); 1404 } 1405 1406 // Array pointers need some flattening 1407 const TypeAryPtr *ta = tj->isa_aryptr(); 1408 if (ta && ta->is_stable()) { 1409 // Erase stability property for alias analysis. 1410 tj = ta = ta->cast_to_stable(false); 1411 } 1412 if( ta && is_known_inst ) { 1413 if ( offset != Type::OffsetBot && 1414 offset > arrayOopDesc::length_offset_in_bytes() ) { 1415 offset = Type::OffsetBot; // Flatten constant access into array body only 1416 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1417 } 1418 } else if( ta && _AliasLevel >= 2 ) { 1419 // For arrays indexed by constant indices, we flatten the alias 1420 // space to include all of the array body. Only the header, klass 1421 // and array length can be accessed un-aliased. 1422 if( offset != Type::OffsetBot ) { 1423 if( ta->const_oop() ) { // MethodData* or Method* 1424 offset = Type::OffsetBot; // Flatten constant access into array body 1425 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1426 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1427 // range is OK as-is. 1428 tj = ta = TypeAryPtr::RANGE; 1429 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1430 tj = TypeInstPtr::KLASS; // all klass loads look alike 1431 ta = TypeAryPtr::RANGE; // generic ignored junk 1432 ptr = TypePtr::BotPTR; 1433 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1434 tj = TypeInstPtr::MARK; 1435 ta = TypeAryPtr::RANGE; // generic ignored junk 1436 ptr = TypePtr::BotPTR; 1437 } else { // Random constant offset into array body 1438 offset = Type::OffsetBot; // Flatten constant access into array body 1439 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1440 } 1441 } 1442 // Arrays of fixed size alias with arrays of unknown size. 1443 if (ta->size() != TypeInt::POS) { 1444 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1445 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1446 } 1447 // Arrays of known objects become arrays of unknown objects. 1448 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1449 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1450 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1451 } 1452 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1453 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1454 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1455 } 1456 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1457 // cannot be distinguished by bytecode alone. 1458 if (ta->elem() == TypeInt::BOOL) { 1459 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1460 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1461 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1462 } 1463 // During the 2nd round of IterGVN, NotNull castings are removed. 1464 // Make sure the Bottom and NotNull variants alias the same. 1465 // Also, make sure exact and non-exact variants alias the same. 1466 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1467 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1468 } 1469 } 1470 1471 // Oop pointers need some flattening 1472 const TypeInstPtr *to = tj->isa_instptr(); 1473 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1474 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1475 if( ptr == TypePtr::Constant ) { 1476 if (to->klass() != ciEnv::current()->Class_klass() || 1477 offset < k->size_helper() * wordSize) { 1478 // No constant oop pointers (such as Strings); they alias with 1479 // unknown strings. 1480 assert(!is_known_inst, "not scalarizable allocation"); 1481 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1482 } 1483 } else if( is_known_inst ) { 1484 tj = to; // Keep NotNull and klass_is_exact for instance type 1485 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1486 // During the 2nd round of IterGVN, NotNull castings are removed. 1487 // Make sure the Bottom and NotNull variants alias the same. 1488 // Also, make sure exact and non-exact variants alias the same. 1489 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1490 } 1491 if (to->speculative() != NULL) { 1492 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1493 } 1494 // Canonicalize the holder of this field 1495 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1496 // First handle header references such as a LoadKlassNode, even if the 1497 // object's klass is unloaded at compile time (4965979). 1498 if (!is_known_inst) { // Do it only for non-instance types 1499 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1500 } 1501 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1502 // Static fields are in the space above the normal instance 1503 // fields in the java.lang.Class instance. 1504 if (to->klass() != ciEnv::current()->Class_klass()) { 1505 to = NULL; 1506 tj = TypeOopPtr::BOTTOM; 1507 offset = tj->offset(); 1508 } 1509 } else { 1510 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1511 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1512 if( is_known_inst ) { 1513 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1514 } else { 1515 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1516 } 1517 } 1518 } 1519 } 1520 1521 // Klass pointers to object array klasses need some flattening 1522 const TypeKlassPtr *tk = tj->isa_klassptr(); 1523 if( tk ) { 1524 // If we are referencing a field within a Klass, we need 1525 // to assume the worst case of an Object. Both exact and 1526 // inexact types must flatten to the same alias class so 1527 // use NotNull as the PTR. 1528 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1529 1530 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1531 TypeKlassPtr::OBJECT->klass(), 1532 offset); 1533 } 1534 1535 ciKlass* klass = tk->klass(); 1536 if( klass->is_obj_array_klass() ) { 1537 ciKlass* k = TypeAryPtr::OOPS->klass(); 1538 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1539 k = TypeInstPtr::BOTTOM->klass(); 1540 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1541 } 1542 1543 // Check for precise loads from the primary supertype array and force them 1544 // to the supertype cache alias index. Check for generic array loads from 1545 // the primary supertype array and also force them to the supertype cache 1546 // alias index. Since the same load can reach both, we need to merge 1547 // these 2 disparate memories into the same alias class. Since the 1548 // primary supertype array is read-only, there's no chance of confusion 1549 // where we bypass an array load and an array store. 1550 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1551 if (offset == Type::OffsetBot || 1552 (offset >= primary_supers_offset && 1553 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1554 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1555 offset = in_bytes(Klass::secondary_super_cache_offset()); 1556 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1557 } 1558 } 1559 1560 // Flatten all Raw pointers together. 1561 if (tj->base() == Type::RawPtr) 1562 tj = TypeRawPtr::BOTTOM; 1563 1564 if (tj->base() == Type::AnyPtr) 1565 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1566 1567 // Flatten all to bottom for now 1568 switch( _AliasLevel ) { 1569 case 0: 1570 tj = TypePtr::BOTTOM; 1571 break; 1572 case 1: // Flatten to: oop, static, field or array 1573 switch (tj->base()) { 1574 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1575 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1576 case Type::AryPtr: // do not distinguish arrays at all 1577 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1578 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1579 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1580 default: ShouldNotReachHere(); 1581 } 1582 break; 1583 case 2: // No collapsing at level 2; keep all splits 1584 case 3: // No collapsing at level 3; keep all splits 1585 break; 1586 default: 1587 Unimplemented(); 1588 } 1589 1590 offset = tj->offset(); 1591 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1592 1593 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1594 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1595 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1596 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1597 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1598 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1599 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1600 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1601 assert( tj->ptr() != TypePtr::TopPTR && 1602 tj->ptr() != TypePtr::AnyNull && 1603 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1604 // assert( tj->ptr() != TypePtr::Constant || 1605 // tj->base() == Type::RawPtr || 1606 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1607 1608 return tj; 1609 } 1610 1611 void Compile::AliasType::Init(int i, const TypePtr* at) { 1612 _index = i; 1613 _adr_type = at; 1614 _field = NULL; 1615 _element = NULL; 1616 _is_rewritable = true; // default 1617 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1618 if (atoop != NULL && atoop->is_known_instance()) { 1619 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1620 _general_index = Compile::current()->get_alias_index(gt); 1621 } else { 1622 _general_index = 0; 1623 } 1624 } 1625 1626 //---------------------------------print_on------------------------------------ 1627 #ifndef PRODUCT 1628 void Compile::AliasType::print_on(outputStream* st) { 1629 if (index() < 10) 1630 st->print("@ <%d> ", index()); 1631 else st->print("@ <%d>", index()); 1632 st->print(is_rewritable() ? " " : " RO"); 1633 int offset = adr_type()->offset(); 1634 if (offset == Type::OffsetBot) 1635 st->print(" +any"); 1636 else st->print(" +%-3d", offset); 1637 st->print(" in "); 1638 adr_type()->dump_on(st); 1639 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1640 if (field() != NULL && tjp) { 1641 if (tjp->klass() != field()->holder() || 1642 tjp->offset() != field()->offset_in_bytes()) { 1643 st->print(" != "); 1644 field()->print(); 1645 st->print(" ***"); 1646 } 1647 } 1648 } 1649 1650 void print_alias_types() { 1651 Compile* C = Compile::current(); 1652 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1653 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1654 C->alias_type(idx)->print_on(tty); 1655 tty->cr(); 1656 } 1657 } 1658 #endif 1659 1660 1661 //----------------------------probe_alias_cache-------------------------------- 1662 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1663 intptr_t key = (intptr_t) adr_type; 1664 key ^= key >> logAliasCacheSize; 1665 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1666 } 1667 1668 1669 //-----------------------------grow_alias_types-------------------------------- 1670 void Compile::grow_alias_types() { 1671 const int old_ats = _max_alias_types; // how many before? 1672 const int new_ats = old_ats; // how many more? 1673 const int grow_ats = old_ats+new_ats; // how many now? 1674 _max_alias_types = grow_ats; 1675 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1676 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1677 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1678 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1679 } 1680 1681 1682 //--------------------------------find_alias_type------------------------------ 1683 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1684 if (_AliasLevel == 0) 1685 return alias_type(AliasIdxBot); 1686 1687 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1688 if (ace->_adr_type == adr_type) { 1689 return alias_type(ace->_index); 1690 } 1691 1692 // Handle special cases. 1693 if (adr_type == NULL) return alias_type(AliasIdxTop); 1694 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1695 1696 // Do it the slow way. 1697 const TypePtr* flat = flatten_alias_type(adr_type); 1698 1699 #ifdef ASSERT 1700 assert(flat == flatten_alias_type(flat), "idempotent"); 1701 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1702 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1703 const TypeOopPtr* foop = flat->is_oopptr(); 1704 // Scalarizable allocations have exact klass always. 1705 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1706 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1707 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1708 } 1709 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1710 #endif 1711 1712 int idx = AliasIdxTop; 1713 for (int i = 0; i < num_alias_types(); i++) { 1714 if (alias_type(i)->adr_type() == flat) { 1715 idx = i; 1716 break; 1717 } 1718 } 1719 1720 if (idx == AliasIdxTop) { 1721 if (no_create) return NULL; 1722 // Grow the array if necessary. 1723 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1724 // Add a new alias type. 1725 idx = _num_alias_types++; 1726 _alias_types[idx]->Init(idx, flat); 1727 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1728 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1729 if (flat->isa_instptr()) { 1730 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1731 && flat->is_instptr()->klass() == env()->Class_klass()) 1732 alias_type(idx)->set_rewritable(false); 1733 } 1734 if (flat->isa_aryptr()) { 1735 #ifdef ASSERT 1736 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1737 // (T_BYTE has the weakest alignment and size restrictions...) 1738 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1739 #endif 1740 if (flat->offset() == TypePtr::OffsetBot) { 1741 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1742 } 1743 } 1744 if (flat->isa_klassptr()) { 1745 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1746 alias_type(idx)->set_rewritable(false); 1747 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1748 alias_type(idx)->set_rewritable(false); 1749 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1750 alias_type(idx)->set_rewritable(false); 1751 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1752 alias_type(idx)->set_rewritable(false); 1753 } 1754 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1755 // but the base pointer type is not distinctive enough to identify 1756 // references into JavaThread.) 1757 1758 // Check for final fields. 1759 const TypeInstPtr* tinst = flat->isa_instptr(); 1760 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1761 ciField* field; 1762 if (tinst->const_oop() != NULL && 1763 tinst->klass() == ciEnv::current()->Class_klass() && 1764 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1765 // static field 1766 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1767 field = k->get_field_by_offset(tinst->offset(), true); 1768 } else { 1769 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1770 field = k->get_field_by_offset(tinst->offset(), false); 1771 } 1772 assert(field == NULL || 1773 original_field == NULL || 1774 (field->holder() == original_field->holder() && 1775 field->offset() == original_field->offset() && 1776 field->is_static() == original_field->is_static()), "wrong field?"); 1777 // Set field() and is_rewritable() attributes. 1778 if (field != NULL) alias_type(idx)->set_field(field); 1779 } 1780 } 1781 1782 // Fill the cache for next time. 1783 ace->_adr_type = adr_type; 1784 ace->_index = idx; 1785 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1786 1787 // Might as well try to fill the cache for the flattened version, too. 1788 AliasCacheEntry* face = probe_alias_cache(flat); 1789 if (face->_adr_type == NULL) { 1790 face->_adr_type = flat; 1791 face->_index = idx; 1792 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1793 } 1794 1795 return alias_type(idx); 1796 } 1797 1798 1799 Compile::AliasType* Compile::alias_type(ciField* field) { 1800 const TypeOopPtr* t; 1801 if (field->is_static()) 1802 t = TypeInstPtr::make(field->holder()->java_mirror()); 1803 else 1804 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1805 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1806 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1807 return atp; 1808 } 1809 1810 1811 //------------------------------have_alias_type-------------------------------- 1812 bool Compile::have_alias_type(const TypePtr* adr_type) { 1813 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1814 if (ace->_adr_type == adr_type) { 1815 return true; 1816 } 1817 1818 // Handle special cases. 1819 if (adr_type == NULL) return true; 1820 if (adr_type == TypePtr::BOTTOM) return true; 1821 1822 return find_alias_type(adr_type, true, NULL) != NULL; 1823 } 1824 1825 //-----------------------------must_alias-------------------------------------- 1826 // True if all values of the given address type are in the given alias category. 1827 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1828 if (alias_idx == AliasIdxBot) return true; // the universal category 1829 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1830 if (alias_idx == AliasIdxTop) return false; // the empty category 1831 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1832 1833 // the only remaining possible overlap is identity 1834 int adr_idx = get_alias_index(adr_type); 1835 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1836 assert(adr_idx == alias_idx || 1837 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1838 && adr_type != TypeOopPtr::BOTTOM), 1839 "should not be testing for overlap with an unsafe pointer"); 1840 return adr_idx == alias_idx; 1841 } 1842 1843 //------------------------------can_alias-------------------------------------- 1844 // True if any values of the given address type are in the given alias category. 1845 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1846 if (alias_idx == AliasIdxTop) return false; // the empty category 1847 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1848 if (alias_idx == AliasIdxBot) return true; // the universal category 1849 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1850 1851 // the only remaining possible overlap is identity 1852 int adr_idx = get_alias_index(adr_type); 1853 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1854 return adr_idx == alias_idx; 1855 } 1856 1857 1858 1859 //---------------------------pop_warm_call------------------------------------- 1860 WarmCallInfo* Compile::pop_warm_call() { 1861 WarmCallInfo* wci = _warm_calls; 1862 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1863 return wci; 1864 } 1865 1866 //----------------------------Inline_Warm-------------------------------------- 1867 int Compile::Inline_Warm() { 1868 // If there is room, try to inline some more warm call sites. 1869 // %%% Do a graph index compaction pass when we think we're out of space? 1870 if (!InlineWarmCalls) return 0; 1871 1872 int calls_made_hot = 0; 1873 int room_to_grow = NodeCountInliningCutoff - unique(); 1874 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1875 int amount_grown = 0; 1876 WarmCallInfo* call; 1877 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1878 int est_size = (int)call->size(); 1879 if (est_size > (room_to_grow - amount_grown)) { 1880 // This one won't fit anyway. Get rid of it. 1881 call->make_cold(); 1882 continue; 1883 } 1884 call->make_hot(); 1885 calls_made_hot++; 1886 amount_grown += est_size; 1887 amount_to_grow -= est_size; 1888 } 1889 1890 if (calls_made_hot > 0) set_major_progress(); 1891 return calls_made_hot; 1892 } 1893 1894 1895 //----------------------------Finish_Warm-------------------------------------- 1896 void Compile::Finish_Warm() { 1897 if (!InlineWarmCalls) return; 1898 if (failing()) return; 1899 if (warm_calls() == NULL) return; 1900 1901 // Clean up loose ends, if we are out of space for inlining. 1902 WarmCallInfo* call; 1903 while ((call = pop_warm_call()) != NULL) { 1904 call->make_cold(); 1905 } 1906 } 1907 1908 //---------------------cleanup_loop_predicates----------------------- 1909 // Remove the opaque nodes that protect the predicates so that all unused 1910 // checks and uncommon_traps will be eliminated from the ideal graph 1911 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1912 if (predicate_count()==0) return; 1913 for (int i = predicate_count(); i > 0; i--) { 1914 Node * n = predicate_opaque1_node(i-1); 1915 assert(n->Opcode() == Op_Opaque1, "must be"); 1916 igvn.replace_node(n, n->in(1)); 1917 } 1918 assert(predicate_count()==0, "should be clean!"); 1919 } 1920 1921 // StringOpts and late inlining of string methods 1922 void Compile::inline_string_calls(bool parse_time) { 1923 { 1924 // remove useless nodes to make the usage analysis simpler 1925 ResourceMark rm; 1926 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1927 } 1928 1929 { 1930 ResourceMark rm; 1931 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1932 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1933 print_method(PHASE_AFTER_STRINGOPTS, 3); 1934 } 1935 1936 // now inline anything that we skipped the first time around 1937 if (!parse_time) { 1938 _late_inlines_pos = _late_inlines.length(); 1939 } 1940 1941 while (_string_late_inlines.length() > 0) { 1942 CallGenerator* cg = _string_late_inlines.pop(); 1943 cg->do_late_inline(); 1944 if (failing()) return; 1945 } 1946 _string_late_inlines.trunc_to(0); 1947 } 1948 1949 // Late inlining of boxing methods 1950 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1951 if (_boxing_late_inlines.length() > 0) { 1952 assert(has_boxed_value(), "inconsistent"); 1953 1954 PhaseGVN* gvn = initial_gvn(); 1955 set_inlining_incrementally(true); 1956 1957 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1958 for_igvn()->clear(); 1959 gvn->replace_with(&igvn); 1960 1961 _late_inlines_pos = _late_inlines.length(); 1962 1963 while (_boxing_late_inlines.length() > 0) { 1964 CallGenerator* cg = _boxing_late_inlines.pop(); 1965 cg->do_late_inline(); 1966 if (failing()) return; 1967 } 1968 _boxing_late_inlines.trunc_to(0); 1969 1970 { 1971 ResourceMark rm; 1972 PhaseRemoveUseless pru(gvn, for_igvn()); 1973 } 1974 1975 igvn = PhaseIterGVN(gvn); 1976 igvn.optimize(); 1977 1978 set_inlining_progress(false); 1979 set_inlining_incrementally(false); 1980 } 1981 } 1982 1983 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1984 assert(IncrementalInline, "incremental inlining should be on"); 1985 PhaseGVN* gvn = initial_gvn(); 1986 1987 set_inlining_progress(false); 1988 for_igvn()->clear(); 1989 gvn->replace_with(&igvn); 1990 1991 { 1992 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 1993 int i = 0; 1994 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1995 CallGenerator* cg = _late_inlines.at(i); 1996 _late_inlines_pos = i+1; 1997 cg->do_late_inline(); 1998 if (failing()) return; 1999 } 2000 int j = 0; 2001 for (; i < _late_inlines.length(); i++, j++) { 2002 _late_inlines.at_put(j, _late_inlines.at(i)); 2003 } 2004 _late_inlines.trunc_to(j); 2005 } 2006 2007 { 2008 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2009 ResourceMark rm; 2010 PhaseRemoveUseless pru(gvn, for_igvn()); 2011 } 2012 2013 { 2014 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2015 igvn = PhaseIterGVN(gvn); 2016 } 2017 } 2018 2019 // Perform incremental inlining until bound on number of live nodes is reached 2020 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2021 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2022 2023 PhaseGVN* gvn = initial_gvn(); 2024 2025 set_inlining_incrementally(true); 2026 set_inlining_progress(true); 2027 uint low_live_nodes = 0; 2028 2029 while(inlining_progress() && _late_inlines.length() > 0) { 2030 2031 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2032 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2033 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2034 // PhaseIdealLoop is expensive so we only try it once we are 2035 // out of live nodes and we only try it again if the previous 2036 // helped got the number of nodes down significantly 2037 PhaseIdealLoop ideal_loop( igvn, false, true ); 2038 if (failing()) return; 2039 low_live_nodes = live_nodes(); 2040 _major_progress = true; 2041 } 2042 2043 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2044 break; 2045 } 2046 } 2047 2048 inline_incrementally_one(igvn); 2049 2050 if (failing()) return; 2051 2052 { 2053 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2054 igvn.optimize(); 2055 } 2056 2057 if (failing()) return; 2058 } 2059 2060 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2061 2062 if (_string_late_inlines.length() > 0) { 2063 assert(has_stringbuilder(), "inconsistent"); 2064 for_igvn()->clear(); 2065 initial_gvn()->replace_with(&igvn); 2066 2067 inline_string_calls(false); 2068 2069 if (failing()) return; 2070 2071 { 2072 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2073 ResourceMark rm; 2074 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2075 } 2076 2077 { 2078 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2079 igvn = PhaseIterGVN(gvn); 2080 igvn.optimize(); 2081 } 2082 } 2083 2084 set_inlining_incrementally(false); 2085 } 2086 2087 2088 //------------------------------Optimize--------------------------------------- 2089 // Given a graph, optimize it. 2090 void Compile::Optimize() { 2091 TracePhase tp("optimizer", &timers[_t_optimizer]); 2092 2093 #ifndef PRODUCT 2094 if (env()->break_at_compile()) { 2095 BREAKPOINT; 2096 } 2097 2098 #endif 2099 2100 ResourceMark rm; 2101 int loop_opts_cnt; 2102 2103 print_inlining_reinit(); 2104 2105 NOT_PRODUCT( verify_graph_edges(); ) 2106 2107 print_method(PHASE_AFTER_PARSING); 2108 2109 { 2110 // Iterative Global Value Numbering, including ideal transforms 2111 // Initialize IterGVN with types and values from parse-time GVN 2112 PhaseIterGVN igvn(initial_gvn()); 2113 #ifdef ASSERT 2114 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2115 #endif 2116 { 2117 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2118 igvn.optimize(); 2119 } 2120 2121 print_method(PHASE_ITER_GVN1, 2); 2122 2123 if (failing()) return; 2124 2125 inline_incrementally(igvn); 2126 2127 print_method(PHASE_INCREMENTAL_INLINE, 2); 2128 2129 if (failing()) return; 2130 2131 if (eliminate_boxing()) { 2132 // Inline valueOf() methods now. 2133 inline_boxing_calls(igvn); 2134 2135 if (AlwaysIncrementalInline) { 2136 inline_incrementally(igvn); 2137 } 2138 2139 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2140 2141 if (failing()) return; 2142 } 2143 2144 // Remove the speculative part of types and clean up the graph from 2145 // the extra CastPP nodes whose only purpose is to carry them. Do 2146 // that early so that optimizations are not disrupted by the extra 2147 // CastPP nodes. 2148 remove_speculative_types(igvn); 2149 2150 // No more new expensive nodes will be added to the list from here 2151 // so keep only the actual candidates for optimizations. 2152 cleanup_expensive_nodes(igvn); 2153 2154 // Perform escape analysis 2155 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2156 if (has_loops()) { 2157 // Cleanup graph (remove dead nodes). 2158 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2159 PhaseIdealLoop ideal_loop( igvn, false, true ); 2160 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2161 if (failing()) return; 2162 } 2163 ConnectionGraph::do_analysis(this, &igvn); 2164 2165 if (failing()) return; 2166 2167 // Optimize out fields loads from scalar replaceable allocations. 2168 igvn.optimize(); 2169 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2170 2171 if (failing()) return; 2172 2173 if (congraph() != NULL && macro_count() > 0) { 2174 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2175 PhaseMacroExpand mexp(igvn); 2176 mexp.eliminate_macro_nodes(); 2177 igvn.set_delay_transform(false); 2178 2179 igvn.optimize(); 2180 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2181 2182 if (failing()) return; 2183 } 2184 } 2185 2186 // Loop transforms on the ideal graph. Range Check Elimination, 2187 // peeling, unrolling, etc. 2188 2189 // Set loop opts counter 2190 loop_opts_cnt = num_loop_opts(); 2191 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2192 { 2193 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2194 PhaseIdealLoop ideal_loop( igvn, true ); 2195 loop_opts_cnt--; 2196 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2197 if (failing()) return; 2198 } 2199 // Loop opts pass if partial peeling occurred in previous pass 2200 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2201 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2202 PhaseIdealLoop ideal_loop( igvn, false ); 2203 loop_opts_cnt--; 2204 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2205 if (failing()) return; 2206 } 2207 // Loop opts pass for loop-unrolling before CCP 2208 if(major_progress() && (loop_opts_cnt > 0)) { 2209 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2210 PhaseIdealLoop ideal_loop( igvn, false ); 2211 loop_opts_cnt--; 2212 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2213 } 2214 if (!failing()) { 2215 // Verify that last round of loop opts produced a valid graph 2216 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2217 PhaseIdealLoop::verify(igvn); 2218 } 2219 } 2220 if (failing()) return; 2221 2222 // Conditional Constant Propagation; 2223 PhaseCCP ccp( &igvn ); 2224 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2225 { 2226 TracePhase tp("ccp", &timers[_t_ccp]); 2227 ccp.do_transform(); 2228 } 2229 print_method(PHASE_CPP1, 2); 2230 2231 assert( true, "Break here to ccp.dump_old2new_map()"); 2232 2233 // Iterative Global Value Numbering, including ideal transforms 2234 { 2235 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2236 igvn = ccp; 2237 igvn.optimize(); 2238 } 2239 2240 print_method(PHASE_ITER_GVN2, 2); 2241 2242 if (failing()) return; 2243 2244 // Loop transforms on the ideal graph. Range Check Elimination, 2245 // peeling, unrolling, etc. 2246 if(loop_opts_cnt > 0) { 2247 debug_only( int cnt = 0; ); 2248 while(major_progress() && (loop_opts_cnt > 0)) { 2249 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2250 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2251 PhaseIdealLoop ideal_loop( igvn, true); 2252 loop_opts_cnt--; 2253 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2254 if (failing()) return; 2255 } 2256 } 2257 // Ensure that major progress is now clear 2258 C->clear_major_progress(); 2259 2260 { 2261 // Verify that all previous optimizations produced a valid graph 2262 // at least to this point, even if no loop optimizations were done. 2263 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2264 PhaseIdealLoop::verify(igvn); 2265 } 2266 2267 { 2268 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2269 PhaseMacroExpand mex(igvn); 2270 if (mex.expand_macro_nodes()) { 2271 assert(failing(), "must bail out w/ explicit message"); 2272 return; 2273 } 2274 } 2275 2276 DEBUG_ONLY( _modified_nodes = NULL; ) 2277 } // (End scope of igvn; run destructor if necessary for asserts.) 2278 2279 process_print_inlining(); 2280 // A method with only infinite loops has no edges entering loops from root 2281 { 2282 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2283 if (final_graph_reshaping()) { 2284 assert(failing(), "must bail out w/ explicit message"); 2285 return; 2286 } 2287 } 2288 2289 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2290 } 2291 2292 2293 //------------------------------Code_Gen--------------------------------------- 2294 // Given a graph, generate code for it 2295 void Compile::Code_Gen() { 2296 if (failing()) { 2297 return; 2298 } 2299 2300 // Perform instruction selection. You might think we could reclaim Matcher 2301 // memory PDQ, but actually the Matcher is used in generating spill code. 2302 // Internals of the Matcher (including some VectorSets) must remain live 2303 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2304 // set a bit in reclaimed memory. 2305 2306 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2307 // nodes. Mapping is only valid at the root of each matched subtree. 2308 NOT_PRODUCT( verify_graph_edges(); ) 2309 2310 Matcher matcher; 2311 _matcher = &matcher; 2312 { 2313 TracePhase tp("matcher", &timers[_t_matcher]); 2314 matcher.match(); 2315 } 2316 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2317 // nodes. Mapping is only valid at the root of each matched subtree. 2318 NOT_PRODUCT( verify_graph_edges(); ) 2319 2320 // If you have too many nodes, or if matching has failed, bail out 2321 check_node_count(0, "out of nodes matching instructions"); 2322 if (failing()) { 2323 return; 2324 } 2325 2326 // Build a proper-looking CFG 2327 PhaseCFG cfg(node_arena(), root(), matcher); 2328 _cfg = &cfg; 2329 { 2330 TracePhase tp("scheduler", &timers[_t_scheduler]); 2331 bool success = cfg.do_global_code_motion(); 2332 if (!success) { 2333 return; 2334 } 2335 2336 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2337 NOT_PRODUCT( verify_graph_edges(); ) 2338 debug_only( cfg.verify(); ) 2339 } 2340 2341 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2342 _regalloc = ®alloc; 2343 { 2344 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2345 // Perform register allocation. After Chaitin, use-def chains are 2346 // no longer accurate (at spill code) and so must be ignored. 2347 // Node->LRG->reg mappings are still accurate. 2348 _regalloc->Register_Allocate(); 2349 2350 // Bail out if the allocator builds too many nodes 2351 if (failing()) { 2352 return; 2353 } 2354 } 2355 2356 // Prior to register allocation we kept empty basic blocks in case the 2357 // the allocator needed a place to spill. After register allocation we 2358 // are not adding any new instructions. If any basic block is empty, we 2359 // can now safely remove it. 2360 { 2361 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2362 cfg.remove_empty_blocks(); 2363 if (do_freq_based_layout()) { 2364 PhaseBlockLayout layout(cfg); 2365 } else { 2366 cfg.set_loop_alignment(); 2367 } 2368 cfg.fixup_flow(); 2369 } 2370 2371 // Apply peephole optimizations 2372 if( OptoPeephole ) { 2373 TracePhase tp("peephole", &timers[_t_peephole]); 2374 PhasePeephole peep( _regalloc, cfg); 2375 peep.do_transform(); 2376 } 2377 2378 // Do late expand if CPU requires this. 2379 if (Matcher::require_postalloc_expand) { 2380 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2381 cfg.postalloc_expand(_regalloc); 2382 } 2383 2384 // Convert Nodes to instruction bits in a buffer 2385 { 2386 TraceTime tp("output", &timers[_t_output], CITime); 2387 Output(); 2388 } 2389 2390 print_method(PHASE_FINAL_CODE); 2391 2392 // He's dead, Jim. 2393 _cfg = (PhaseCFG*)0xdeadbeef; 2394 _regalloc = (PhaseChaitin*)0xdeadbeef; 2395 } 2396 2397 2398 //------------------------------dump_asm--------------------------------------- 2399 // Dump formatted assembly 2400 #ifndef PRODUCT 2401 void Compile::dump_asm(int *pcs, uint pc_limit) { 2402 bool cut_short = false; 2403 tty->print_cr("#"); 2404 tty->print("# "); _tf->dump(); tty->cr(); 2405 tty->print_cr("#"); 2406 2407 // For all blocks 2408 int pc = 0x0; // Program counter 2409 char starts_bundle = ' '; 2410 _regalloc->dump_frame(); 2411 2412 Node *n = NULL; 2413 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2414 if (VMThread::should_terminate()) { 2415 cut_short = true; 2416 break; 2417 } 2418 Block* block = _cfg->get_block(i); 2419 if (block->is_connector() && !Verbose) { 2420 continue; 2421 } 2422 n = block->head(); 2423 if (pcs && n->_idx < pc_limit) { 2424 tty->print("%3.3x ", pcs[n->_idx]); 2425 } else { 2426 tty->print(" "); 2427 } 2428 block->dump_head(_cfg); 2429 if (block->is_connector()) { 2430 tty->print_cr(" # Empty connector block"); 2431 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2432 tty->print_cr(" # Block is sole successor of call"); 2433 } 2434 2435 // For all instructions 2436 Node *delay = NULL; 2437 for (uint j = 0; j < block->number_of_nodes(); j++) { 2438 if (VMThread::should_terminate()) { 2439 cut_short = true; 2440 break; 2441 } 2442 n = block->get_node(j); 2443 if (valid_bundle_info(n)) { 2444 Bundle* bundle = node_bundling(n); 2445 if (bundle->used_in_unconditional_delay()) { 2446 delay = n; 2447 continue; 2448 } 2449 if (bundle->starts_bundle()) { 2450 starts_bundle = '+'; 2451 } 2452 } 2453 2454 if (WizardMode) { 2455 n->dump(); 2456 } 2457 2458 if( !n->is_Region() && // Dont print in the Assembly 2459 !n->is_Phi() && // a few noisely useless nodes 2460 !n->is_Proj() && 2461 !n->is_MachTemp() && 2462 !n->is_SafePointScalarObject() && 2463 !n->is_Catch() && // Would be nice to print exception table targets 2464 !n->is_MergeMem() && // Not very interesting 2465 !n->is_top() && // Debug info table constants 2466 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2467 ) { 2468 if (pcs && n->_idx < pc_limit) 2469 tty->print("%3.3x", pcs[n->_idx]); 2470 else 2471 tty->print(" "); 2472 tty->print(" %c ", starts_bundle); 2473 starts_bundle = ' '; 2474 tty->print("\t"); 2475 n->format(_regalloc, tty); 2476 tty->cr(); 2477 } 2478 2479 // If we have an instruction with a delay slot, and have seen a delay, 2480 // then back up and print it 2481 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2482 assert(delay != NULL, "no unconditional delay instruction"); 2483 if (WizardMode) delay->dump(); 2484 2485 if (node_bundling(delay)->starts_bundle()) 2486 starts_bundle = '+'; 2487 if (pcs && n->_idx < pc_limit) 2488 tty->print("%3.3x", pcs[n->_idx]); 2489 else 2490 tty->print(" "); 2491 tty->print(" %c ", starts_bundle); 2492 starts_bundle = ' '; 2493 tty->print("\t"); 2494 delay->format(_regalloc, tty); 2495 tty->cr(); 2496 delay = NULL; 2497 } 2498 2499 // Dump the exception table as well 2500 if( n->is_Catch() && (Verbose || WizardMode) ) { 2501 // Print the exception table for this offset 2502 _handler_table.print_subtable_for(pc); 2503 } 2504 } 2505 2506 if (pcs && n->_idx < pc_limit) 2507 tty->print_cr("%3.3x", pcs[n->_idx]); 2508 else 2509 tty->cr(); 2510 2511 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2512 2513 } // End of per-block dump 2514 tty->cr(); 2515 2516 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2517 } 2518 #endif 2519 2520 //------------------------------Final_Reshape_Counts--------------------------- 2521 // This class defines counters to help identify when a method 2522 // may/must be executed using hardware with only 24-bit precision. 2523 struct Final_Reshape_Counts : public StackObj { 2524 int _call_count; // count non-inlined 'common' calls 2525 int _float_count; // count float ops requiring 24-bit precision 2526 int _double_count; // count double ops requiring more precision 2527 int _java_call_count; // count non-inlined 'java' calls 2528 int _inner_loop_count; // count loops which need alignment 2529 VectorSet _visited; // Visitation flags 2530 Node_List _tests; // Set of IfNodes & PCTableNodes 2531 2532 Final_Reshape_Counts() : 2533 _call_count(0), _float_count(0), _double_count(0), 2534 _java_call_count(0), _inner_loop_count(0), 2535 _visited( Thread::current()->resource_area() ) { } 2536 2537 void inc_call_count () { _call_count ++; } 2538 void inc_float_count () { _float_count ++; } 2539 void inc_double_count() { _double_count++; } 2540 void inc_java_call_count() { _java_call_count++; } 2541 void inc_inner_loop_count() { _inner_loop_count++; } 2542 2543 int get_call_count () const { return _call_count ; } 2544 int get_float_count () const { return _float_count ; } 2545 int get_double_count() const { return _double_count; } 2546 int get_java_call_count() const { return _java_call_count; } 2547 int get_inner_loop_count() const { return _inner_loop_count; } 2548 }; 2549 2550 #ifdef ASSERT 2551 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2552 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2553 // Make sure the offset goes inside the instance layout. 2554 return k->contains_field_offset(tp->offset()); 2555 // Note that OffsetBot and OffsetTop are very negative. 2556 } 2557 #endif 2558 2559 // Eliminate trivially redundant StoreCMs and accumulate their 2560 // precedence edges. 2561 void Compile::eliminate_redundant_card_marks(Node* n) { 2562 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2563 if (n->in(MemNode::Address)->outcnt() > 1) { 2564 // There are multiple users of the same address so it might be 2565 // possible to eliminate some of the StoreCMs 2566 Node* mem = n->in(MemNode::Memory); 2567 Node* adr = n->in(MemNode::Address); 2568 Node* val = n->in(MemNode::ValueIn); 2569 Node* prev = n; 2570 bool done = false; 2571 // Walk the chain of StoreCMs eliminating ones that match. As 2572 // long as it's a chain of single users then the optimization is 2573 // safe. Eliminating partially redundant StoreCMs would require 2574 // cloning copies down the other paths. 2575 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2576 if (adr == mem->in(MemNode::Address) && 2577 val == mem->in(MemNode::ValueIn)) { 2578 // redundant StoreCM 2579 if (mem->req() > MemNode::OopStore) { 2580 // Hasn't been processed by this code yet. 2581 n->add_prec(mem->in(MemNode::OopStore)); 2582 } else { 2583 // Already converted to precedence edge 2584 for (uint i = mem->req(); i < mem->len(); i++) { 2585 // Accumulate any precedence edges 2586 if (mem->in(i) != NULL) { 2587 n->add_prec(mem->in(i)); 2588 } 2589 } 2590 // Everything above this point has been processed. 2591 done = true; 2592 } 2593 // Eliminate the previous StoreCM 2594 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2595 assert(mem->outcnt() == 0, "should be dead"); 2596 mem->disconnect_inputs(NULL, this); 2597 } else { 2598 prev = mem; 2599 } 2600 mem = prev->in(MemNode::Memory); 2601 } 2602 } 2603 } 2604 2605 //------------------------------final_graph_reshaping_impl---------------------- 2606 // Implement items 1-5 from final_graph_reshaping below. 2607 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2608 2609 if ( n->outcnt() == 0 ) return; // dead node 2610 uint nop = n->Opcode(); 2611 2612 // Check for 2-input instruction with "last use" on right input. 2613 // Swap to left input. Implements item (2). 2614 if( n->req() == 3 && // two-input instruction 2615 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2616 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2617 n->in(2)->outcnt() == 1 &&// right use IS a last use 2618 !n->in(2)->is_Con() ) { // right use is not a constant 2619 // Check for commutative opcode 2620 switch( nop ) { 2621 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2622 case Op_MaxI: case Op_MinI: 2623 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2624 case Op_AndL: case Op_XorL: case Op_OrL: 2625 case Op_AndI: case Op_XorI: case Op_OrI: { 2626 // Move "last use" input to left by swapping inputs 2627 n->swap_edges(1, 2); 2628 break; 2629 } 2630 default: 2631 break; 2632 } 2633 } 2634 2635 #ifdef ASSERT 2636 if( n->is_Mem() ) { 2637 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2638 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2639 // oop will be recorded in oop map if load crosses safepoint 2640 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2641 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2642 "raw memory operations should have control edge"); 2643 } 2644 #endif 2645 // Count FPU ops and common calls, implements item (3) 2646 switch( nop ) { 2647 // Count all float operations that may use FPU 2648 case Op_AddF: 2649 case Op_SubF: 2650 case Op_MulF: 2651 case Op_DivF: 2652 case Op_NegF: 2653 case Op_ModF: 2654 case Op_ConvI2F: 2655 case Op_ConF: 2656 case Op_CmpF: 2657 case Op_CmpF3: 2658 // case Op_ConvL2F: // longs are split into 32-bit halves 2659 frc.inc_float_count(); 2660 break; 2661 2662 case Op_ConvF2D: 2663 case Op_ConvD2F: 2664 frc.inc_float_count(); 2665 frc.inc_double_count(); 2666 break; 2667 2668 // Count all double operations that may use FPU 2669 case Op_AddD: 2670 case Op_SubD: 2671 case Op_MulD: 2672 case Op_DivD: 2673 case Op_NegD: 2674 case Op_ModD: 2675 case Op_ConvI2D: 2676 case Op_ConvD2I: 2677 // case Op_ConvL2D: // handled by leaf call 2678 // case Op_ConvD2L: // handled by leaf call 2679 case Op_ConD: 2680 case Op_CmpD: 2681 case Op_CmpD3: 2682 frc.inc_double_count(); 2683 break; 2684 case Op_Opaque1: // Remove Opaque Nodes before matching 2685 case Op_Opaque2: // Remove Opaque Nodes before matching 2686 case Op_Opaque3: 2687 n->subsume_by(n->in(1), this); 2688 break; 2689 case Op_CallStaticJava: 2690 case Op_CallJava: 2691 case Op_CallDynamicJava: 2692 frc.inc_java_call_count(); // Count java call site; 2693 case Op_CallRuntime: 2694 case Op_CallLeaf: 2695 case Op_CallLeafNoFP: { 2696 assert( n->is_Call(), "" ); 2697 CallNode *call = n->as_Call(); 2698 // Count call sites where the FP mode bit would have to be flipped. 2699 // Do not count uncommon runtime calls: 2700 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2701 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2702 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2703 frc.inc_call_count(); // Count the call site 2704 } else { // See if uncommon argument is shared 2705 Node *n = call->in(TypeFunc::Parms); 2706 int nop = n->Opcode(); 2707 // Clone shared simple arguments to uncommon calls, item (1). 2708 if( n->outcnt() > 1 && 2709 !n->is_Proj() && 2710 nop != Op_CreateEx && 2711 nop != Op_CheckCastPP && 2712 nop != Op_DecodeN && 2713 nop != Op_DecodeNKlass && 2714 !n->is_Mem() ) { 2715 Node *x = n->clone(); 2716 call->set_req( TypeFunc::Parms, x ); 2717 } 2718 } 2719 break; 2720 } 2721 2722 case Op_StoreD: 2723 case Op_LoadD: 2724 case Op_LoadD_unaligned: 2725 frc.inc_double_count(); 2726 goto handle_mem; 2727 case Op_StoreF: 2728 case Op_LoadF: 2729 frc.inc_float_count(); 2730 goto handle_mem; 2731 2732 case Op_StoreCM: 2733 { 2734 // Convert OopStore dependence into precedence edge 2735 Node* prec = n->in(MemNode::OopStore); 2736 n->del_req(MemNode::OopStore); 2737 n->add_prec(prec); 2738 eliminate_redundant_card_marks(n); 2739 } 2740 2741 // fall through 2742 2743 case Op_StoreB: 2744 case Op_StoreC: 2745 case Op_StorePConditional: 2746 case Op_StoreI: 2747 case Op_StoreL: 2748 case Op_StoreIConditional: 2749 case Op_StoreLConditional: 2750 case Op_CompareAndSwapI: 2751 case Op_CompareAndSwapL: 2752 case Op_CompareAndSwapP: 2753 case Op_CompareAndSwapN: 2754 case Op_GetAndAddI: 2755 case Op_GetAndAddL: 2756 case Op_GetAndSetI: 2757 case Op_GetAndSetL: 2758 case Op_GetAndSetP: 2759 case Op_GetAndSetN: 2760 case Op_StoreP: 2761 case Op_StoreN: 2762 case Op_StoreNKlass: 2763 case Op_LoadB: 2764 case Op_LoadUB: 2765 case Op_LoadUS: 2766 case Op_LoadI: 2767 case Op_LoadKlass: 2768 case Op_LoadNKlass: 2769 case Op_LoadL: 2770 case Op_LoadL_unaligned: 2771 case Op_LoadPLocked: 2772 case Op_LoadP: 2773 case Op_LoadN: 2774 case Op_LoadRange: 2775 case Op_LoadS: { 2776 handle_mem: 2777 #ifdef ASSERT 2778 if( VerifyOptoOopOffsets ) { 2779 assert( n->is_Mem(), "" ); 2780 MemNode *mem = (MemNode*)n; 2781 // Check to see if address types have grounded out somehow. 2782 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2783 assert( !tp || oop_offset_is_sane(tp), "" ); 2784 } 2785 #endif 2786 break; 2787 } 2788 2789 case Op_AddP: { // Assert sane base pointers 2790 Node *addp = n->in(AddPNode::Address); 2791 assert( !addp->is_AddP() || 2792 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2793 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2794 "Base pointers must match" ); 2795 #ifdef _LP64 2796 if ((UseCompressedOops || UseCompressedClassPointers) && 2797 addp->Opcode() == Op_ConP && 2798 addp == n->in(AddPNode::Base) && 2799 n->in(AddPNode::Offset)->is_Con()) { 2800 // Use addressing with narrow klass to load with offset on x86. 2801 // On sparc loading 32-bits constant and decoding it have less 2802 // instructions (4) then load 64-bits constant (7). 2803 // Do this transformation here since IGVN will convert ConN back to ConP. 2804 const Type* t = addp->bottom_type(); 2805 if (t->isa_oopptr() || t->isa_klassptr()) { 2806 Node* nn = NULL; 2807 2808 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2809 2810 // Look for existing ConN node of the same exact type. 2811 Node* r = root(); 2812 uint cnt = r->outcnt(); 2813 for (uint i = 0; i < cnt; i++) { 2814 Node* m = r->raw_out(i); 2815 if (m!= NULL && m->Opcode() == op && 2816 m->bottom_type()->make_ptr() == t) { 2817 nn = m; 2818 break; 2819 } 2820 } 2821 if (nn != NULL) { 2822 // Decode a narrow oop to match address 2823 // [R12 + narrow_oop_reg<<3 + offset] 2824 if (t->isa_oopptr()) { 2825 nn = new DecodeNNode(nn, t); 2826 } else { 2827 nn = new DecodeNKlassNode(nn, t); 2828 } 2829 n->set_req(AddPNode::Base, nn); 2830 n->set_req(AddPNode::Address, nn); 2831 if (addp->outcnt() == 0) { 2832 addp->disconnect_inputs(NULL, this); 2833 } 2834 } 2835 } 2836 } 2837 #endif 2838 break; 2839 } 2840 2841 case Op_CastPP: { 2842 // Remove CastPP nodes to gain more freedom during scheduling but 2843 // keep the dependency they encode as control or precedence edges 2844 // (if control is set already) on memory operations. Some CastPP 2845 // nodes don't have a control (don't carry a dependency): skip 2846 // those. 2847 if (n->in(0) != NULL) { 2848 ResourceMark rm; 2849 Unique_Node_List wq; 2850 wq.push(n); 2851 for (uint next = 0; next < wq.size(); ++next) { 2852 Node *m = wq.at(next); 2853 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 2854 Node* use = m->fast_out(i); 2855 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 2856 use->ensure_control_or_add_prec(n->in(0)); 2857 } else if (use->in(0) == NULL) { 2858 switch(use->Opcode()) { 2859 case Op_AddP: 2860 case Op_DecodeN: 2861 case Op_DecodeNKlass: 2862 case Op_CheckCastPP: 2863 case Op_CastPP: 2864 wq.push(use); 2865 break; 2866 } 2867 } 2868 } 2869 } 2870 } 2871 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 2872 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2873 Node* in1 = n->in(1); 2874 const Type* t = n->bottom_type(); 2875 Node* new_in1 = in1->clone(); 2876 new_in1->as_DecodeN()->set_type(t); 2877 2878 if (!Matcher::narrow_oop_use_complex_address()) { 2879 // 2880 // x86, ARM and friends can handle 2 adds in addressing mode 2881 // and Matcher can fold a DecodeN node into address by using 2882 // a narrow oop directly and do implicit NULL check in address: 2883 // 2884 // [R12 + narrow_oop_reg<<3 + offset] 2885 // NullCheck narrow_oop_reg 2886 // 2887 // On other platforms (Sparc) we have to keep new DecodeN node and 2888 // use it to do implicit NULL check in address: 2889 // 2890 // decode_not_null narrow_oop_reg, base_reg 2891 // [base_reg + offset] 2892 // NullCheck base_reg 2893 // 2894 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2895 // to keep the information to which NULL check the new DecodeN node 2896 // corresponds to use it as value in implicit_null_check(). 2897 // 2898 new_in1->set_req(0, n->in(0)); 2899 } 2900 2901 n->subsume_by(new_in1, this); 2902 if (in1->outcnt() == 0) { 2903 in1->disconnect_inputs(NULL, this); 2904 } 2905 } else { 2906 n->subsume_by(n->in(1), this); 2907 if (n->outcnt() == 0) { 2908 n->disconnect_inputs(NULL, this); 2909 } 2910 } 2911 break; 2912 } 2913 #ifdef _LP64 2914 case Op_CmpP: 2915 // Do this transformation here to preserve CmpPNode::sub() and 2916 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2917 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2918 Node* in1 = n->in(1); 2919 Node* in2 = n->in(2); 2920 if (!in1->is_DecodeNarrowPtr()) { 2921 in2 = in1; 2922 in1 = n->in(2); 2923 } 2924 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2925 2926 Node* new_in2 = NULL; 2927 if (in2->is_DecodeNarrowPtr()) { 2928 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2929 new_in2 = in2->in(1); 2930 } else if (in2->Opcode() == Op_ConP) { 2931 const Type* t = in2->bottom_type(); 2932 if (t == TypePtr::NULL_PTR) { 2933 assert(in1->is_DecodeN(), "compare klass to null?"); 2934 // Don't convert CmpP null check into CmpN if compressed 2935 // oops implicit null check is not generated. 2936 // This will allow to generate normal oop implicit null check. 2937 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2938 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 2939 // 2940 // This transformation together with CastPP transformation above 2941 // will generated code for implicit NULL checks for compressed oops. 2942 // 2943 // The original code after Optimize() 2944 // 2945 // LoadN memory, narrow_oop_reg 2946 // decode narrow_oop_reg, base_reg 2947 // CmpP base_reg, NULL 2948 // CastPP base_reg // NotNull 2949 // Load [base_reg + offset], val_reg 2950 // 2951 // after these transformations will be 2952 // 2953 // LoadN memory, narrow_oop_reg 2954 // CmpN narrow_oop_reg, NULL 2955 // decode_not_null narrow_oop_reg, base_reg 2956 // Load [base_reg + offset], val_reg 2957 // 2958 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2959 // since narrow oops can be used in debug info now (see the code in 2960 // final_graph_reshaping_walk()). 2961 // 2962 // At the end the code will be matched to 2963 // on x86: 2964 // 2965 // Load_narrow_oop memory, narrow_oop_reg 2966 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2967 // NullCheck narrow_oop_reg 2968 // 2969 // and on sparc: 2970 // 2971 // Load_narrow_oop memory, narrow_oop_reg 2972 // decode_not_null narrow_oop_reg, base_reg 2973 // Load [base_reg + offset], val_reg 2974 // NullCheck base_reg 2975 // 2976 } else if (t->isa_oopptr()) { 2977 new_in2 = ConNode::make(t->make_narrowoop()); 2978 } else if (t->isa_klassptr()) { 2979 new_in2 = ConNode::make(t->make_narrowklass()); 2980 } 2981 } 2982 if (new_in2 != NULL) { 2983 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 2984 n->subsume_by(cmpN, this); 2985 if (in1->outcnt() == 0) { 2986 in1->disconnect_inputs(NULL, this); 2987 } 2988 if (in2->outcnt() == 0) { 2989 in2->disconnect_inputs(NULL, this); 2990 } 2991 } 2992 } 2993 break; 2994 2995 case Op_DecodeN: 2996 case Op_DecodeNKlass: 2997 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2998 // DecodeN could be pinned when it can't be fold into 2999 // an address expression, see the code for Op_CastPP above. 3000 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3001 break; 3002 3003 case Op_EncodeP: 3004 case Op_EncodePKlass: { 3005 Node* in1 = n->in(1); 3006 if (in1->is_DecodeNarrowPtr()) { 3007 n->subsume_by(in1->in(1), this); 3008 } else if (in1->Opcode() == Op_ConP) { 3009 const Type* t = in1->bottom_type(); 3010 if (t == TypePtr::NULL_PTR) { 3011 assert(t->isa_oopptr(), "null klass?"); 3012 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3013 } else if (t->isa_oopptr()) { 3014 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3015 } else if (t->isa_klassptr()) { 3016 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3017 } 3018 } 3019 if (in1->outcnt() == 0) { 3020 in1->disconnect_inputs(NULL, this); 3021 } 3022 break; 3023 } 3024 3025 case Op_Proj: { 3026 if (OptimizeStringConcat) { 3027 ProjNode* p = n->as_Proj(); 3028 if (p->_is_io_use) { 3029 // Separate projections were used for the exception path which 3030 // are normally removed by a late inline. If it wasn't inlined 3031 // then they will hang around and should just be replaced with 3032 // the original one. 3033 Node* proj = NULL; 3034 // Replace with just one 3035 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3036 Node *use = i.get(); 3037 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3038 proj = use; 3039 break; 3040 } 3041 } 3042 assert(proj != NULL, "must be found"); 3043 p->subsume_by(proj, this); 3044 } 3045 } 3046 break; 3047 } 3048 3049 case Op_Phi: 3050 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3051 // The EncodeP optimization may create Phi with the same edges 3052 // for all paths. It is not handled well by Register Allocator. 3053 Node* unique_in = n->in(1); 3054 assert(unique_in != NULL, ""); 3055 uint cnt = n->req(); 3056 for (uint i = 2; i < cnt; i++) { 3057 Node* m = n->in(i); 3058 assert(m != NULL, ""); 3059 if (unique_in != m) 3060 unique_in = NULL; 3061 } 3062 if (unique_in != NULL) { 3063 n->subsume_by(unique_in, this); 3064 } 3065 } 3066 break; 3067 3068 #endif 3069 3070 case Op_ModI: 3071 if (UseDivMod) { 3072 // Check if a%b and a/b both exist 3073 Node* d = n->find_similar(Op_DivI); 3074 if (d) { 3075 // Replace them with a fused divmod if supported 3076 if (Matcher::has_match_rule(Op_DivModI)) { 3077 DivModINode* divmod = DivModINode::make(n); 3078 d->subsume_by(divmod->div_proj(), this); 3079 n->subsume_by(divmod->mod_proj(), this); 3080 } else { 3081 // replace a%b with a-((a/b)*b) 3082 Node* mult = new MulINode(d, d->in(2)); 3083 Node* sub = new SubINode(d->in(1), mult); 3084 n->subsume_by(sub, this); 3085 } 3086 } 3087 } 3088 break; 3089 3090 case Op_ModL: 3091 if (UseDivMod) { 3092 // Check if a%b and a/b both exist 3093 Node* d = n->find_similar(Op_DivL); 3094 if (d) { 3095 // Replace them with a fused divmod if supported 3096 if (Matcher::has_match_rule(Op_DivModL)) { 3097 DivModLNode* divmod = DivModLNode::make(n); 3098 d->subsume_by(divmod->div_proj(), this); 3099 n->subsume_by(divmod->mod_proj(), this); 3100 } else { 3101 // replace a%b with a-((a/b)*b) 3102 Node* mult = new MulLNode(d, d->in(2)); 3103 Node* sub = new SubLNode(d->in(1), mult); 3104 n->subsume_by(sub, this); 3105 } 3106 } 3107 } 3108 break; 3109 3110 case Op_LoadVector: 3111 case Op_StoreVector: 3112 break; 3113 3114 case Op_AddReductionVI: 3115 case Op_AddReductionVL: 3116 case Op_AddReductionVF: 3117 case Op_AddReductionVD: 3118 case Op_MulReductionVI: 3119 case Op_MulReductionVL: 3120 case Op_MulReductionVF: 3121 case Op_MulReductionVD: 3122 break; 3123 3124 case Op_PackB: 3125 case Op_PackS: 3126 case Op_PackI: 3127 case Op_PackF: 3128 case Op_PackL: 3129 case Op_PackD: 3130 if (n->req()-1 > 2) { 3131 // Replace many operand PackNodes with a binary tree for matching 3132 PackNode* p = (PackNode*) n; 3133 Node* btp = p->binary_tree_pack(1, n->req()); 3134 n->subsume_by(btp, this); 3135 } 3136 break; 3137 case Op_Loop: 3138 case Op_CountedLoop: 3139 if (n->as_Loop()->is_inner_loop()) { 3140 frc.inc_inner_loop_count(); 3141 } 3142 break; 3143 case Op_LShiftI: 3144 case Op_RShiftI: 3145 case Op_URShiftI: 3146 case Op_LShiftL: 3147 case Op_RShiftL: 3148 case Op_URShiftL: 3149 if (Matcher::need_masked_shift_count) { 3150 // The cpu's shift instructions don't restrict the count to the 3151 // lower 5/6 bits. We need to do the masking ourselves. 3152 Node* in2 = n->in(2); 3153 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3154 const TypeInt* t = in2->find_int_type(); 3155 if (t != NULL && t->is_con()) { 3156 juint shift = t->get_con(); 3157 if (shift > mask) { // Unsigned cmp 3158 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3159 } 3160 } else { 3161 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3162 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3163 n->set_req(2, shift); 3164 } 3165 } 3166 if (in2->outcnt() == 0) { // Remove dead node 3167 in2->disconnect_inputs(NULL, this); 3168 } 3169 } 3170 break; 3171 case Op_MemBarStoreStore: 3172 case Op_MemBarRelease: 3173 // Break the link with AllocateNode: it is no longer useful and 3174 // confuses register allocation. 3175 if (n->req() > MemBarNode::Precedent) { 3176 n->set_req(MemBarNode::Precedent, top()); 3177 } 3178 break; 3179 default: 3180 assert( !n->is_Call(), "" ); 3181 assert( !n->is_Mem(), "" ); 3182 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3183 break; 3184 } 3185 3186 // Collect CFG split points 3187 if (n->is_MultiBranch()) 3188 frc._tests.push(n); 3189 } 3190 3191 //------------------------------final_graph_reshaping_walk--------------------- 3192 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3193 // requires that the walk visits a node's inputs before visiting the node. 3194 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3195 ResourceArea *area = Thread::current()->resource_area(); 3196 Unique_Node_List sfpt(area); 3197 3198 frc._visited.set(root->_idx); // first, mark node as visited 3199 uint cnt = root->req(); 3200 Node *n = root; 3201 uint i = 0; 3202 while (true) { 3203 if (i < cnt) { 3204 // Place all non-visited non-null inputs onto stack 3205 Node* m = n->in(i); 3206 ++i; 3207 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3208 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3209 // compute worst case interpreter size in case of a deoptimization 3210 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3211 3212 sfpt.push(m); 3213 } 3214 cnt = m->req(); 3215 nstack.push(n, i); // put on stack parent and next input's index 3216 n = m; 3217 i = 0; 3218 } 3219 } else { 3220 // Now do post-visit work 3221 final_graph_reshaping_impl( n, frc ); 3222 if (nstack.is_empty()) 3223 break; // finished 3224 n = nstack.node(); // Get node from stack 3225 cnt = n->req(); 3226 i = nstack.index(); 3227 nstack.pop(); // Shift to the next node on stack 3228 } 3229 } 3230 3231 // Skip next transformation if compressed oops are not used. 3232 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3233 (!UseCompressedOops && !UseCompressedClassPointers)) 3234 return; 3235 3236 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3237 // It could be done for an uncommon traps or any safepoints/calls 3238 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3239 while (sfpt.size() > 0) { 3240 n = sfpt.pop(); 3241 JVMState *jvms = n->as_SafePoint()->jvms(); 3242 assert(jvms != NULL, "sanity"); 3243 int start = jvms->debug_start(); 3244 int end = n->req(); 3245 bool is_uncommon = (n->is_CallStaticJava() && 3246 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3247 for (int j = start; j < end; j++) { 3248 Node* in = n->in(j); 3249 if (in->is_DecodeNarrowPtr()) { 3250 bool safe_to_skip = true; 3251 if (!is_uncommon ) { 3252 // Is it safe to skip? 3253 for (uint i = 0; i < in->outcnt(); i++) { 3254 Node* u = in->raw_out(i); 3255 if (!u->is_SafePoint() || 3256 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3257 safe_to_skip = false; 3258 } 3259 } 3260 } 3261 if (safe_to_skip) { 3262 n->set_req(j, in->in(1)); 3263 } 3264 if (in->outcnt() == 0) { 3265 in->disconnect_inputs(NULL, this); 3266 } 3267 } 3268 } 3269 } 3270 } 3271 3272 //------------------------------final_graph_reshaping-------------------------- 3273 // Final Graph Reshaping. 3274 // 3275 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3276 // and not commoned up and forced early. Must come after regular 3277 // optimizations to avoid GVN undoing the cloning. Clone constant 3278 // inputs to Loop Phis; these will be split by the allocator anyways. 3279 // Remove Opaque nodes. 3280 // (2) Move last-uses by commutative operations to the left input to encourage 3281 // Intel update-in-place two-address operations and better register usage 3282 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3283 // calls canonicalizing them back. 3284 // (3) Count the number of double-precision FP ops, single-precision FP ops 3285 // and call sites. On Intel, we can get correct rounding either by 3286 // forcing singles to memory (requires extra stores and loads after each 3287 // FP bytecode) or we can set a rounding mode bit (requires setting and 3288 // clearing the mode bit around call sites). The mode bit is only used 3289 // if the relative frequency of single FP ops to calls is low enough. 3290 // This is a key transform for SPEC mpeg_audio. 3291 // (4) Detect infinite loops; blobs of code reachable from above but not 3292 // below. Several of the Code_Gen algorithms fail on such code shapes, 3293 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3294 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3295 // Detection is by looking for IfNodes where only 1 projection is 3296 // reachable from below or CatchNodes missing some targets. 3297 // (5) Assert for insane oop offsets in debug mode. 3298 3299 bool Compile::final_graph_reshaping() { 3300 // an infinite loop may have been eliminated by the optimizer, 3301 // in which case the graph will be empty. 3302 if (root()->req() == 1) { 3303 record_method_not_compilable("trivial infinite loop"); 3304 return true; 3305 } 3306 3307 // Expensive nodes have their control input set to prevent the GVN 3308 // from freely commoning them. There's no GVN beyond this point so 3309 // no need to keep the control input. We want the expensive nodes to 3310 // be freely moved to the least frequent code path by gcm. 3311 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3312 for (int i = 0; i < expensive_count(); i++) { 3313 _expensive_nodes->at(i)->set_req(0, NULL); 3314 } 3315 3316 Final_Reshape_Counts frc; 3317 3318 // Visit everybody reachable! 3319 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3320 Node_Stack nstack(live_nodes() >> 1); 3321 final_graph_reshaping_walk(nstack, root(), frc); 3322 3323 // Check for unreachable (from below) code (i.e., infinite loops). 3324 for( uint i = 0; i < frc._tests.size(); i++ ) { 3325 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3326 // Get number of CFG targets. 3327 // Note that PCTables include exception targets after calls. 3328 uint required_outcnt = n->required_outcnt(); 3329 if (n->outcnt() != required_outcnt) { 3330 // Check for a few special cases. Rethrow Nodes never take the 3331 // 'fall-thru' path, so expected kids is 1 less. 3332 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3333 if (n->in(0)->in(0)->is_Call()) { 3334 CallNode *call = n->in(0)->in(0)->as_Call(); 3335 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3336 required_outcnt--; // Rethrow always has 1 less kid 3337 } else if (call->req() > TypeFunc::Parms && 3338 call->is_CallDynamicJava()) { 3339 // Check for null receiver. In such case, the optimizer has 3340 // detected that the virtual call will always result in a null 3341 // pointer exception. The fall-through projection of this CatchNode 3342 // will not be populated. 3343 Node *arg0 = call->in(TypeFunc::Parms); 3344 if (arg0->is_Type() && 3345 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3346 required_outcnt--; 3347 } 3348 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3349 call->req() > TypeFunc::Parms+1 && 3350 call->is_CallStaticJava()) { 3351 // Check for negative array length. In such case, the optimizer has 3352 // detected that the allocation attempt will always result in an 3353 // exception. There is no fall-through projection of this CatchNode . 3354 Node *arg1 = call->in(TypeFunc::Parms+1); 3355 if (arg1->is_Type() && 3356 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3357 required_outcnt--; 3358 } 3359 } 3360 } 3361 } 3362 // Recheck with a better notion of 'required_outcnt' 3363 if (n->outcnt() != required_outcnt) { 3364 record_method_not_compilable("malformed control flow"); 3365 return true; // Not all targets reachable! 3366 } 3367 } 3368 // Check that I actually visited all kids. Unreached kids 3369 // must be infinite loops. 3370 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3371 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3372 record_method_not_compilable("infinite loop"); 3373 return true; // Found unvisited kid; must be unreach 3374 } 3375 } 3376 3377 // If original bytecodes contained a mixture of floats and doubles 3378 // check if the optimizer has made it homogenous, item (3). 3379 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3380 frc.get_float_count() > 32 && 3381 frc.get_double_count() == 0 && 3382 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3383 set_24_bit_selection_and_mode( false, true ); 3384 } 3385 3386 set_java_calls(frc.get_java_call_count()); 3387 set_inner_loops(frc.get_inner_loop_count()); 3388 3389 // No infinite loops, no reason to bail out. 3390 return false; 3391 } 3392 3393 //-----------------------------too_many_traps---------------------------------- 3394 // Report if there are too many traps at the current method and bci. 3395 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3396 bool Compile::too_many_traps(ciMethod* method, 3397 int bci, 3398 Deoptimization::DeoptReason reason) { 3399 ciMethodData* md = method->method_data(); 3400 if (md->is_empty()) { 3401 // Assume the trap has not occurred, or that it occurred only 3402 // because of a transient condition during start-up in the interpreter. 3403 return false; 3404 } 3405 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3406 if (md->has_trap_at(bci, m, reason) != 0) { 3407 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3408 // Also, if there are multiple reasons, or if there is no per-BCI record, 3409 // assume the worst. 3410 if (log()) 3411 log()->elem("observe trap='%s' count='%d'", 3412 Deoptimization::trap_reason_name(reason), 3413 md->trap_count(reason)); 3414 return true; 3415 } else { 3416 // Ignore method/bci and see if there have been too many globally. 3417 return too_many_traps(reason, md); 3418 } 3419 } 3420 3421 // Less-accurate variant which does not require a method and bci. 3422 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3423 ciMethodData* logmd) { 3424 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3425 // Too many traps globally. 3426 // Note that we use cumulative trap_count, not just md->trap_count. 3427 if (log()) { 3428 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3429 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3430 Deoptimization::trap_reason_name(reason), 3431 mcount, trap_count(reason)); 3432 } 3433 return true; 3434 } else { 3435 // The coast is clear. 3436 return false; 3437 } 3438 } 3439 3440 //--------------------------too_many_recompiles-------------------------------- 3441 // Report if there are too many recompiles at the current method and bci. 3442 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3443 // Is not eager to return true, since this will cause the compiler to use 3444 // Action_none for a trap point, to avoid too many recompilations. 3445 bool Compile::too_many_recompiles(ciMethod* method, 3446 int bci, 3447 Deoptimization::DeoptReason reason) { 3448 ciMethodData* md = method->method_data(); 3449 if (md->is_empty()) { 3450 // Assume the trap has not occurred, or that it occurred only 3451 // because of a transient condition during start-up in the interpreter. 3452 return false; 3453 } 3454 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3455 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3456 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3457 Deoptimization::DeoptReason per_bc_reason 3458 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3459 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3460 if ((per_bc_reason == Deoptimization::Reason_none 3461 || md->has_trap_at(bci, m, reason) != 0) 3462 // The trap frequency measure we care about is the recompile count: 3463 && md->trap_recompiled_at(bci, m) 3464 && md->overflow_recompile_count() >= bc_cutoff) { 3465 // Do not emit a trap here if it has already caused recompilations. 3466 // Also, if there are multiple reasons, or if there is no per-BCI record, 3467 // assume the worst. 3468 if (log()) 3469 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3470 Deoptimization::trap_reason_name(reason), 3471 md->trap_count(reason), 3472 md->overflow_recompile_count()); 3473 return true; 3474 } else if (trap_count(reason) != 0 3475 && decompile_count() >= m_cutoff) { 3476 // Too many recompiles globally, and we have seen this sort of trap. 3477 // Use cumulative decompile_count, not just md->decompile_count. 3478 if (log()) 3479 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3480 Deoptimization::trap_reason_name(reason), 3481 md->trap_count(reason), trap_count(reason), 3482 md->decompile_count(), decompile_count()); 3483 return true; 3484 } else { 3485 // The coast is clear. 3486 return false; 3487 } 3488 } 3489 3490 // Compute when not to trap. Used by matching trap based nodes and 3491 // NullCheck optimization. 3492 void Compile::set_allowed_deopt_reasons() { 3493 _allowed_reasons = 0; 3494 if (is_method_compilation()) { 3495 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3496 assert(rs < BitsPerInt, "recode bit map"); 3497 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3498 _allowed_reasons |= nth_bit(rs); 3499 } 3500 } 3501 } 3502 } 3503 3504 #ifndef PRODUCT 3505 //------------------------------verify_graph_edges--------------------------- 3506 // Walk the Graph and verify that there is a one-to-one correspondence 3507 // between Use-Def edges and Def-Use edges in the graph. 3508 void Compile::verify_graph_edges(bool no_dead_code) { 3509 if (VerifyGraphEdges) { 3510 ResourceArea *area = Thread::current()->resource_area(); 3511 Unique_Node_List visited(area); 3512 // Call recursive graph walk to check edges 3513 _root->verify_edges(visited); 3514 if (no_dead_code) { 3515 // Now make sure that no visited node is used by an unvisited node. 3516 bool dead_nodes = false; 3517 Unique_Node_List checked(area); 3518 while (visited.size() > 0) { 3519 Node* n = visited.pop(); 3520 checked.push(n); 3521 for (uint i = 0; i < n->outcnt(); i++) { 3522 Node* use = n->raw_out(i); 3523 if (checked.member(use)) continue; // already checked 3524 if (visited.member(use)) continue; // already in the graph 3525 if (use->is_Con()) continue; // a dead ConNode is OK 3526 // At this point, we have found a dead node which is DU-reachable. 3527 if (!dead_nodes) { 3528 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3529 dead_nodes = true; 3530 } 3531 use->dump(2); 3532 tty->print_cr("---"); 3533 checked.push(use); // No repeats; pretend it is now checked. 3534 } 3535 } 3536 assert(!dead_nodes, "using nodes must be reachable from root"); 3537 } 3538 } 3539 } 3540 3541 // Verify GC barriers consistency 3542 // Currently supported: 3543 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3544 void Compile::verify_barriers() { 3545 if (UseG1GC) { 3546 // Verify G1 pre-barriers 3547 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3548 3549 ResourceArea *area = Thread::current()->resource_area(); 3550 Unique_Node_List visited(area); 3551 Node_List worklist(area); 3552 // We're going to walk control flow backwards starting from the Root 3553 worklist.push(_root); 3554 while (worklist.size() > 0) { 3555 Node* x = worklist.pop(); 3556 if (x == NULL || x == top()) continue; 3557 if (visited.member(x)) { 3558 continue; 3559 } else { 3560 visited.push(x); 3561 } 3562 3563 if (x->is_Region()) { 3564 for (uint i = 1; i < x->req(); i++) { 3565 worklist.push(x->in(i)); 3566 } 3567 } else { 3568 worklist.push(x->in(0)); 3569 // We are looking for the pattern: 3570 // /->ThreadLocal 3571 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3572 // \->ConI(0) 3573 // We want to verify that the If and the LoadB have the same control 3574 // See GraphKit::g1_write_barrier_pre() 3575 if (x->is_If()) { 3576 IfNode *iff = x->as_If(); 3577 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3578 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3579 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3580 && cmp->in(1)->is_Load()) { 3581 LoadNode* load = cmp->in(1)->as_Load(); 3582 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3583 && load->in(2)->in(3)->is_Con() 3584 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3585 3586 Node* if_ctrl = iff->in(0); 3587 Node* load_ctrl = load->in(0); 3588 3589 if (if_ctrl != load_ctrl) { 3590 // Skip possible CProj->NeverBranch in infinite loops 3591 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3592 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3593 if_ctrl = if_ctrl->in(0)->in(0); 3594 } 3595 } 3596 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3597 } 3598 } 3599 } 3600 } 3601 } 3602 } 3603 } 3604 } 3605 3606 #endif 3607 3608 // The Compile object keeps track of failure reasons separately from the ciEnv. 3609 // This is required because there is not quite a 1-1 relation between the 3610 // ciEnv and its compilation task and the Compile object. Note that one 3611 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3612 // to backtrack and retry without subsuming loads. Other than this backtracking 3613 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3614 // by the logic in C2Compiler. 3615 void Compile::record_failure(const char* reason) { 3616 if (log() != NULL) { 3617 log()->elem("failure reason='%s' phase='compile'", reason); 3618 } 3619 if (_failure_reason == NULL) { 3620 // Record the first failure reason. 3621 _failure_reason = reason; 3622 } 3623 3624 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3625 C->print_method(PHASE_FAILURE); 3626 } 3627 _root = NULL; // flush the graph, too 3628 } 3629 3630 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3631 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3632 _phase_name(name), _dolog(CITimeVerbose) 3633 { 3634 if (_dolog) { 3635 C = Compile::current(); 3636 _log = C->log(); 3637 } else { 3638 C = NULL; 3639 _log = NULL; 3640 } 3641 if (_log != NULL) { 3642 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3643 _log->stamp(); 3644 _log->end_head(); 3645 } 3646 } 3647 3648 Compile::TracePhase::~TracePhase() { 3649 3650 C = Compile::current(); 3651 if (_dolog) { 3652 _log = C->log(); 3653 } else { 3654 _log = NULL; 3655 } 3656 3657 #ifdef ASSERT 3658 if (PrintIdealNodeCount) { 3659 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3660 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3661 } 3662 3663 if (VerifyIdealNodeCount) { 3664 Compile::current()->print_missing_nodes(); 3665 } 3666 #endif 3667 3668 if (_log != NULL) { 3669 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3670 } 3671 } 3672 3673 //============================================================================= 3674 // Two Constant's are equal when the type and the value are equal. 3675 bool Compile::Constant::operator==(const Constant& other) { 3676 if (type() != other.type() ) return false; 3677 if (can_be_reused() != other.can_be_reused()) return false; 3678 // For floating point values we compare the bit pattern. 3679 switch (type()) { 3680 case T_FLOAT: return (_v._value.i == other._v._value.i); 3681 case T_LONG: 3682 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3683 case T_OBJECT: 3684 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3685 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3686 case T_METADATA: return (_v._metadata == other._v._metadata); 3687 default: ShouldNotReachHere(); 3688 } 3689 return false; 3690 } 3691 3692 static int type_to_size_in_bytes(BasicType t) { 3693 switch (t) { 3694 case T_LONG: return sizeof(jlong ); 3695 case T_FLOAT: return sizeof(jfloat ); 3696 case T_DOUBLE: return sizeof(jdouble); 3697 case T_METADATA: return sizeof(Metadata*); 3698 // We use T_VOID as marker for jump-table entries (labels) which 3699 // need an internal word relocation. 3700 case T_VOID: 3701 case T_ADDRESS: 3702 case T_OBJECT: return sizeof(jobject); 3703 } 3704 3705 ShouldNotReachHere(); 3706 return -1; 3707 } 3708 3709 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3710 // sort descending 3711 if (a->freq() > b->freq()) return -1; 3712 if (a->freq() < b->freq()) return 1; 3713 return 0; 3714 } 3715 3716 void Compile::ConstantTable::calculate_offsets_and_size() { 3717 // First, sort the array by frequencies. 3718 _constants.sort(qsort_comparator); 3719 3720 #ifdef ASSERT 3721 // Make sure all jump-table entries were sorted to the end of the 3722 // array (they have a negative frequency). 3723 bool found_void = false; 3724 for (int i = 0; i < _constants.length(); i++) { 3725 Constant con = _constants.at(i); 3726 if (con.type() == T_VOID) 3727 found_void = true; // jump-tables 3728 else 3729 assert(!found_void, "wrong sorting"); 3730 } 3731 #endif 3732 3733 int offset = 0; 3734 for (int i = 0; i < _constants.length(); i++) { 3735 Constant* con = _constants.adr_at(i); 3736 3737 // Align offset for type. 3738 int typesize = type_to_size_in_bytes(con->type()); 3739 offset = align_size_up(offset, typesize); 3740 con->set_offset(offset); // set constant's offset 3741 3742 if (con->type() == T_VOID) { 3743 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3744 offset = offset + typesize * n->outcnt(); // expand jump-table 3745 } else { 3746 offset = offset + typesize; 3747 } 3748 } 3749 3750 // Align size up to the next section start (which is insts; see 3751 // CodeBuffer::align_at_start). 3752 assert(_size == -1, "already set?"); 3753 _size = align_size_up(offset, CodeEntryAlignment); 3754 } 3755 3756 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3757 MacroAssembler _masm(&cb); 3758 for (int i = 0; i < _constants.length(); i++) { 3759 Constant con = _constants.at(i); 3760 address constant_addr; 3761 switch (con.type()) { 3762 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3763 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3764 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3765 case T_OBJECT: { 3766 jobject obj = con.get_jobject(); 3767 int oop_index = _masm.oop_recorder()->find_index(obj); 3768 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3769 break; 3770 } 3771 case T_ADDRESS: { 3772 address addr = (address) con.get_jobject(); 3773 constant_addr = _masm.address_constant(addr); 3774 break; 3775 } 3776 // We use T_VOID as marker for jump-table entries (labels) which 3777 // need an internal word relocation. 3778 case T_VOID: { 3779 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3780 // Fill the jump-table with a dummy word. The real value is 3781 // filled in later in fill_jump_table. 3782 address dummy = (address) n; 3783 constant_addr = _masm.address_constant(dummy); 3784 // Expand jump-table 3785 for (uint i = 1; i < n->outcnt(); i++) { 3786 address temp_addr = _masm.address_constant(dummy + i); 3787 assert(temp_addr, "consts section too small"); 3788 } 3789 break; 3790 } 3791 case T_METADATA: { 3792 Metadata* obj = con.get_metadata(); 3793 int metadata_index = _masm.oop_recorder()->find_index(obj); 3794 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3795 break; 3796 } 3797 default: ShouldNotReachHere(); 3798 } 3799 assert(constant_addr, "consts section too small"); 3800 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3801 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3802 } 3803 } 3804 3805 int Compile::ConstantTable::find_offset(Constant& con) const { 3806 int idx = _constants.find(con); 3807 assert(idx != -1, "constant must be in constant table"); 3808 int offset = _constants.at(idx).offset(); 3809 assert(offset != -1, "constant table not emitted yet?"); 3810 return offset; 3811 } 3812 3813 void Compile::ConstantTable::add(Constant& con) { 3814 if (con.can_be_reused()) { 3815 int idx = _constants.find(con); 3816 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3817 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3818 return; 3819 } 3820 } 3821 (void) _constants.append(con); 3822 } 3823 3824 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3825 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3826 Constant con(type, value, b->_freq); 3827 add(con); 3828 return con; 3829 } 3830 3831 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3832 Constant con(metadata); 3833 add(con); 3834 return con; 3835 } 3836 3837 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3838 jvalue value; 3839 BasicType type = oper->type()->basic_type(); 3840 switch (type) { 3841 case T_LONG: value.j = oper->constantL(); break; 3842 case T_FLOAT: value.f = oper->constantF(); break; 3843 case T_DOUBLE: value.d = oper->constantD(); break; 3844 case T_OBJECT: 3845 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3846 case T_METADATA: return add((Metadata*)oper->constant()); break; 3847 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3848 } 3849 return add(n, type, value); 3850 } 3851 3852 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3853 jvalue value; 3854 // We can use the node pointer here to identify the right jump-table 3855 // as this method is called from Compile::Fill_buffer right before 3856 // the MachNodes are emitted and the jump-table is filled (means the 3857 // MachNode pointers do not change anymore). 3858 value.l = (jobject) n; 3859 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3860 add(con); 3861 return con; 3862 } 3863 3864 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3865 // If called from Compile::scratch_emit_size do nothing. 3866 if (Compile::current()->in_scratch_emit_size()) return; 3867 3868 assert(labels.is_nonempty(), "must be"); 3869 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3870 3871 // Since MachConstantNode::constant_offset() also contains 3872 // table_base_offset() we need to subtract the table_base_offset() 3873 // to get the plain offset into the constant table. 3874 int offset = n->constant_offset() - table_base_offset(); 3875 3876 MacroAssembler _masm(&cb); 3877 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3878 3879 for (uint i = 0; i < n->outcnt(); i++) { 3880 address* constant_addr = &jump_table_base[i]; 3881 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))); 3882 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3883 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3884 } 3885 } 3886 3887 //----------------------------static_subtype_check----------------------------- 3888 // Shortcut important common cases when superklass is exact: 3889 // (0) superklass is java.lang.Object (can occur in reflective code) 3890 // (1) subklass is already limited to a subtype of superklass => always ok 3891 // (2) subklass does not overlap with superklass => always fail 3892 // (3) superklass has NO subtypes and we can check with a simple compare. 3893 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 3894 if (StressReflectiveCode) { 3895 return SSC_full_test; // Let caller generate the general case. 3896 } 3897 3898 if (superk == env()->Object_klass()) { 3899 return SSC_always_true; // (0) this test cannot fail 3900 } 3901 3902 ciType* superelem = superk; 3903 if (superelem->is_array_klass()) 3904 superelem = superelem->as_array_klass()->base_element_type(); 3905 3906 if (!subk->is_interface()) { // cannot trust static interface types yet 3907 if (subk->is_subtype_of(superk)) { 3908 return SSC_always_true; // (1) false path dead; no dynamic test needed 3909 } 3910 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 3911 !superk->is_subtype_of(subk)) { 3912 return SSC_always_false; 3913 } 3914 } 3915 3916 // If casting to an instance klass, it must have no subtypes 3917 if (superk->is_interface()) { 3918 // Cannot trust interfaces yet. 3919 // %%% S.B. superk->nof_implementors() == 1 3920 } else if (superelem->is_instance_klass()) { 3921 ciInstanceKlass* ik = superelem->as_instance_klass(); 3922 if (!ik->has_subklass() && !ik->is_interface()) { 3923 if (!ik->is_final()) { 3924 // Add a dependency if there is a chance of a later subclass. 3925 dependencies()->assert_leaf_type(ik); 3926 } 3927 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3928 } 3929 } else { 3930 // A primitive array type has no subtypes. 3931 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3932 } 3933 3934 return SSC_full_test; 3935 } 3936 3937 Node* Compile::conv_I2X_index(PhaseGVN *phase, Node* idx, const TypeInt* sizetype) { 3938 #ifdef _LP64 3939 // The scaled index operand to AddP must be a clean 64-bit value. 3940 // Java allows a 32-bit int to be incremented to a negative 3941 // value, which appears in a 64-bit register as a large 3942 // positive number. Using that large positive number as an 3943 // operand in pointer arithmetic has bad consequences. 3944 // On the other hand, 32-bit overflow is rare, and the possibility 3945 // can often be excluded, if we annotate the ConvI2L node with 3946 // a type assertion that its value is known to be a small positive 3947 // number. (The prior range check has ensured this.) 3948 // This assertion is used by ConvI2LNode::Ideal. 3949 int index_max = max_jint - 1; // array size is max_jint, index is one less 3950 if (sizetype != NULL) index_max = sizetype->_hi - 1; 3951 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 3952 idx = phase->transform(new ConvI2LNode(idx, lidxtype)); 3953 #endif 3954 return idx; 3955 } 3956 3957 // The message about the current inlining is accumulated in 3958 // _print_inlining_stream and transfered into the _print_inlining_list 3959 // once we know whether inlining succeeds or not. For regular 3960 // inlining, messages are appended to the buffer pointed by 3961 // _print_inlining_idx in the _print_inlining_list. For late inlining, 3962 // a new buffer is added after _print_inlining_idx in the list. This 3963 // way we can update the inlining message for late inlining call site 3964 // when the inlining is attempted again. 3965 void Compile::print_inlining_init() { 3966 if (print_inlining() || print_intrinsics()) { 3967 _print_inlining_stream = new stringStream(); 3968 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 3969 } 3970 } 3971 3972 void Compile::print_inlining_reinit() { 3973 if (print_inlining() || print_intrinsics()) { 3974 // Re allocate buffer when we change ResourceMark 3975 _print_inlining_stream = new stringStream(); 3976 } 3977 } 3978 3979 void Compile::print_inlining_reset() { 3980 _print_inlining_stream->reset(); 3981 } 3982 3983 void Compile::print_inlining_commit() { 3984 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 3985 // Transfer the message from _print_inlining_stream to the current 3986 // _print_inlining_list buffer and clear _print_inlining_stream. 3987 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 3988 print_inlining_reset(); 3989 } 3990 3991 void Compile::print_inlining_push() { 3992 // Add new buffer to the _print_inlining_list at current position 3993 _print_inlining_idx++; 3994 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 3995 } 3996 3997 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 3998 return _print_inlining_list->at(_print_inlining_idx); 3999 } 4000 4001 void Compile::print_inlining_update(CallGenerator* cg) { 4002 if (print_inlining() || print_intrinsics()) { 4003 if (!cg->is_late_inline()) { 4004 if (print_inlining_current().cg() != NULL) { 4005 print_inlining_push(); 4006 } 4007 print_inlining_commit(); 4008 } else { 4009 if (print_inlining_current().cg() != cg && 4010 (print_inlining_current().cg() != NULL || 4011 print_inlining_current().ss()->size() != 0)) { 4012 print_inlining_push(); 4013 } 4014 print_inlining_commit(); 4015 print_inlining_current().set_cg(cg); 4016 } 4017 } 4018 } 4019 4020 void Compile::print_inlining_move_to(CallGenerator* cg) { 4021 // We resume inlining at a late inlining call site. Locate the 4022 // corresponding inlining buffer so that we can update it. 4023 if (print_inlining()) { 4024 for (int i = 0; i < _print_inlining_list->length(); i++) { 4025 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4026 _print_inlining_idx = i; 4027 return; 4028 } 4029 } 4030 ShouldNotReachHere(); 4031 } 4032 } 4033 4034 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4035 if (print_inlining()) { 4036 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4037 assert(print_inlining_current().cg() == cg, "wrong entry"); 4038 // replace message with new message 4039 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4040 print_inlining_commit(); 4041 print_inlining_current().set_cg(cg); 4042 } 4043 } 4044 4045 void Compile::print_inlining_assert_ready() { 4046 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4047 } 4048 4049 void Compile::process_print_inlining() { 4050 bool do_print_inlining = print_inlining() || print_intrinsics(); 4051 if (do_print_inlining || log() != NULL) { 4052 // Print inlining message for candidates that we couldn't inline 4053 // for lack of space 4054 for (int i = 0; i < _late_inlines.length(); i++) { 4055 CallGenerator* cg = _late_inlines.at(i); 4056 if (!cg->is_mh_late_inline()) { 4057 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4058 if (do_print_inlining) { 4059 cg->print_inlining_late(msg); 4060 } 4061 log_late_inline_failure(cg, msg); 4062 } 4063 } 4064 } 4065 if (do_print_inlining) { 4066 ResourceMark rm; 4067 stringStream ss; 4068 for (int i = 0; i < _print_inlining_list->length(); i++) { 4069 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4070 } 4071 size_t end = ss.size(); 4072 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4073 strncpy(_print_inlining_output, ss.base(), end+1); 4074 _print_inlining_output[end] = 0; 4075 } 4076 } 4077 4078 void Compile::dump_print_inlining() { 4079 if (_print_inlining_output != NULL) { 4080 tty->print_raw(_print_inlining_output); 4081 } 4082 } 4083 4084 void Compile::log_late_inline(CallGenerator* cg) { 4085 if (log() != NULL) { 4086 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4087 cg->unique_id()); 4088 JVMState* p = cg->call_node()->jvms(); 4089 while (p != NULL) { 4090 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4091 p = p->caller(); 4092 } 4093 log()->tail("late_inline"); 4094 } 4095 } 4096 4097 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4098 log_late_inline(cg); 4099 if (log() != NULL) { 4100 log()->inline_fail(msg); 4101 } 4102 } 4103 4104 void Compile::log_inline_id(CallGenerator* cg) { 4105 if (log() != NULL) { 4106 // The LogCompilation tool needs a unique way to identify late 4107 // inline call sites. This id must be unique for this call site in 4108 // this compilation. Try to have it unique across compilations as 4109 // well because it can be convenient when grepping through the log 4110 // file. 4111 // Distinguish OSR compilations from others in case CICountOSR is 4112 // on. 4113 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4114 cg->set_unique_id(id); 4115 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4116 } 4117 } 4118 4119 void Compile::log_inline_failure(const char* msg) { 4120 if (C->log() != NULL) { 4121 C->log()->inline_fail(msg); 4122 } 4123 } 4124 4125 4126 // Dump inlining replay data to the stream. 4127 // Don't change thread state and acquire any locks. 4128 void Compile::dump_inline_data(outputStream* out) { 4129 InlineTree* inl_tree = ilt(); 4130 if (inl_tree != NULL) { 4131 out->print(" inline %d", inl_tree->count()); 4132 inl_tree->dump_replay_data(out); 4133 } 4134 } 4135 4136 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4137 if (n1->Opcode() < n2->Opcode()) return -1; 4138 else if (n1->Opcode() > n2->Opcode()) return 1; 4139 4140 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())); 4141 for (uint i = 1; i < n1->req(); i++) { 4142 if (n1->in(i) < n2->in(i)) return -1; 4143 else if (n1->in(i) > n2->in(i)) return 1; 4144 } 4145 4146 return 0; 4147 } 4148 4149 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4150 Node* n1 = *n1p; 4151 Node* n2 = *n2p; 4152 4153 return cmp_expensive_nodes(n1, n2); 4154 } 4155 4156 void Compile::sort_expensive_nodes() { 4157 if (!expensive_nodes_sorted()) { 4158 _expensive_nodes->sort(cmp_expensive_nodes); 4159 } 4160 } 4161 4162 bool Compile::expensive_nodes_sorted() const { 4163 for (int i = 1; i < _expensive_nodes->length(); i++) { 4164 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4165 return false; 4166 } 4167 } 4168 return true; 4169 } 4170 4171 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4172 if (_expensive_nodes->length() == 0) { 4173 return false; 4174 } 4175 4176 assert(OptimizeExpensiveOps, "optimization off?"); 4177 4178 // Take this opportunity to remove dead nodes from the list 4179 int j = 0; 4180 for (int i = 0; i < _expensive_nodes->length(); i++) { 4181 Node* n = _expensive_nodes->at(i); 4182 if (!n->is_unreachable(igvn)) { 4183 assert(n->is_expensive(), "should be expensive"); 4184 _expensive_nodes->at_put(j, n); 4185 j++; 4186 } 4187 } 4188 _expensive_nodes->trunc_to(j); 4189 4190 // Then sort the list so that similar nodes are next to each other 4191 // and check for at least two nodes of identical kind with same data 4192 // inputs. 4193 sort_expensive_nodes(); 4194 4195 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4196 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4197 return true; 4198 } 4199 } 4200 4201 return false; 4202 } 4203 4204 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4205 if (_expensive_nodes->length() == 0) { 4206 return; 4207 } 4208 4209 assert(OptimizeExpensiveOps, "optimization off?"); 4210 4211 // Sort to bring similar nodes next to each other and clear the 4212 // control input of nodes for which there's only a single copy. 4213 sort_expensive_nodes(); 4214 4215 int j = 0; 4216 int identical = 0; 4217 int i = 0; 4218 bool modified = false; 4219 for (; i < _expensive_nodes->length()-1; i++) { 4220 assert(j <= i, "can't write beyond current index"); 4221 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4222 identical++; 4223 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4224 continue; 4225 } 4226 if (identical > 0) { 4227 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4228 identical = 0; 4229 } else { 4230 Node* n = _expensive_nodes->at(i); 4231 igvn.replace_input_of(n, 0, NULL); 4232 igvn.hash_insert(n); 4233 modified = true; 4234 } 4235 } 4236 if (identical > 0) { 4237 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4238 } else if (_expensive_nodes->length() >= 1) { 4239 Node* n = _expensive_nodes->at(i); 4240 igvn.replace_input_of(n, 0, NULL); 4241 igvn.hash_insert(n); 4242 modified = true; 4243 } 4244 _expensive_nodes->trunc_to(j); 4245 if (modified) { 4246 igvn.optimize(); 4247 } 4248 } 4249 4250 void Compile::add_expensive_node(Node * n) { 4251 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4252 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4253 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4254 if (OptimizeExpensiveOps) { 4255 _expensive_nodes->append(n); 4256 } else { 4257 // Clear control input and let IGVN optimize expensive nodes if 4258 // OptimizeExpensiveOps is off. 4259 n->set_req(0, NULL); 4260 } 4261 } 4262 4263 /** 4264 * Remove the speculative part of types and clean up the graph 4265 */ 4266 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4267 if (UseTypeSpeculation) { 4268 Unique_Node_List worklist; 4269 worklist.push(root()); 4270 int modified = 0; 4271 // Go over all type nodes that carry a speculative type, drop the 4272 // speculative part of the type and enqueue the node for an igvn 4273 // which may optimize it out. 4274 for (uint next = 0; next < worklist.size(); ++next) { 4275 Node *n = worklist.at(next); 4276 if (n->is_Type()) { 4277 TypeNode* tn = n->as_Type(); 4278 const Type* t = tn->type(); 4279 const Type* t_no_spec = t->remove_speculative(); 4280 if (t_no_spec != t) { 4281 bool in_hash = igvn.hash_delete(n); 4282 assert(in_hash, "node should be in igvn hash table"); 4283 tn->set_type(t_no_spec); 4284 igvn.hash_insert(n); 4285 igvn._worklist.push(n); // give it a chance to go away 4286 modified++; 4287 } 4288 } 4289 uint max = n->len(); 4290 for( uint i = 0; i < max; ++i ) { 4291 Node *m = n->in(i); 4292 if (not_a_node(m)) continue; 4293 worklist.push(m); 4294 } 4295 } 4296 // Drop the speculative part of all types in the igvn's type table 4297 igvn.remove_speculative_types(); 4298 if (modified > 0) { 4299 igvn.optimize(); 4300 } 4301 #ifdef ASSERT 4302 // Verify that after the IGVN is over no speculative type has resurfaced 4303 worklist.clear(); 4304 worklist.push(root()); 4305 for (uint next = 0; next < worklist.size(); ++next) { 4306 Node *n = worklist.at(next); 4307 const Type* t = igvn.type_or_null(n); 4308 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4309 if (n->is_Type()) { 4310 t = n->as_Type()->type(); 4311 assert(t == t->remove_speculative(), "no more speculative types"); 4312 } 4313 uint max = n->len(); 4314 for( uint i = 0; i < max; ++i ) { 4315 Node *m = n->in(i); 4316 if (not_a_node(m)) continue; 4317 worklist.push(m); 4318 } 4319 } 4320 igvn.check_no_speculative_types(); 4321 #endif 4322 } 4323 } 4324 4325 // Auxiliary method to support randomized stressing/fuzzing. 4326 // 4327 // This method can be called the arbitrary number of times, with current count 4328 // as the argument. The logic allows selecting a single candidate from the 4329 // running list of candidates as follows: 4330 // int count = 0; 4331 // Cand* selected = null; 4332 // while(cand = cand->next()) { 4333 // if (randomized_select(++count)) { 4334 // selected = cand; 4335 // } 4336 // } 4337 // 4338 // Including count equalizes the chances any candidate is "selected". 4339 // This is useful when we don't have the complete list of candidates to choose 4340 // from uniformly. In this case, we need to adjust the randomicity of the 4341 // selection, or else we will end up biasing the selection towards the latter 4342 // candidates. 4343 // 4344 // Quick back-envelope calculation shows that for the list of n candidates 4345 // the equal probability for the candidate to persist as "best" can be 4346 // achieved by replacing it with "next" k-th candidate with the probability 4347 // of 1/k. It can be easily shown that by the end of the run, the 4348 // probability for any candidate is converged to 1/n, thus giving the 4349 // uniform distribution among all the candidates. 4350 // 4351 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4352 #define RANDOMIZED_DOMAIN_POW 29 4353 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4354 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4355 bool Compile::randomized_select(int count) { 4356 assert(count > 0, "only positive"); 4357 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4358 } 4359 4360 const char* CloneMap::debug_option_name = "CloneMapDebug"; 4361 CloneMap& Compile::clone_map() { return _clone_map; } 4362 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4363 4364 void NodeCloneInfo::dump() const { 4365 tty->print(" {%d:%d} ", idx(), gen()); 4366 } 4367 4368 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4369 uint64_t val = value(old->_idx); 4370 NodeCloneInfo cio(val); 4371 assert(val != 0, "old node should be in the map"); 4372 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4373 insert(nnn->_idx, cin.get()); 4374 #ifndef PRODUCT 4375 if (is_debug()) { 4376 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4377 } 4378 #endif 4379 } 4380 4381 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4382 NodeCloneInfo cio(value(old->_idx)); 4383 if (cio.get() == 0) { 4384 cio.set(old->_idx, 0); 4385 insert(old->_idx, cio.get()); 4386 #ifndef PRODUCT 4387 if (is_debug()) { 4388 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4389 } 4390 #endif 4391 } 4392 clone(old, nnn, gen); 4393 } 4394 4395 int CloneMap::max_gen() const { 4396 int g = 0; 4397 DictI di(_dict); 4398 for(; di.test(); ++di) { 4399 int t = gen(di._key); 4400 if (g < t) { 4401 g = t; 4402 #ifndef PRODUCT 4403 if (is_debug()) { 4404 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4405 } 4406 #endif 4407 } 4408 } 4409 return g; 4410 } 4411 4412 void CloneMap::dump(node_idx_t key) const { 4413 uint64_t val = value(key); 4414 if (val != 0) { 4415 NodeCloneInfo ni(val); 4416 ni.dump(); 4417 } 4418 }