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/castnode.hpp" 42 #include "opto/cfgnode.hpp" 43 #include "opto/chaitin.hpp" 44 #include "opto/compile.hpp" 45 #include "opto/connode.hpp" 46 #include "opto/convertnode.hpp" 47 #include "opto/divnode.hpp" 48 #include "opto/escape.hpp" 49 #include "opto/idealGraphPrinter.hpp" 50 #include "opto/loopnode.hpp" 51 #include "opto/machnode.hpp" 52 #include "opto/macro.hpp" 53 #include "opto/matcher.hpp" 54 #include "opto/mathexactnode.hpp" 55 #include "opto/memnode.hpp" 56 #include "opto/mulnode.hpp" 57 #include "opto/narrowptrnode.hpp" 58 #include "opto/node.hpp" 59 #include "opto/opcodes.hpp" 60 #include "opto/output.hpp" 61 #include "opto/parse.hpp" 62 #include "opto/phaseX.hpp" 63 #include "opto/rootnode.hpp" 64 #include "opto/runtime.hpp" 65 #include "opto/stringopts.hpp" 66 #include "opto/type.hpp" 67 #include "opto/vectornode.hpp" 68 #include "runtime/arguments.hpp" 69 #include "runtime/sharedRuntime.hpp" 70 #include "runtime/signature.hpp" 71 #include "runtime/stubRoutines.hpp" 72 #include "runtime/timer.hpp" 73 #include "utilities/copy.hpp" 74 75 76 // -------------------- Compile::mach_constant_base_node ----------------------- 77 // Constant table base node singleton. 78 MachConstantBaseNode* Compile::mach_constant_base_node() { 79 if (_mach_constant_base_node == NULL) { 80 _mach_constant_base_node = new MachConstantBaseNode(); 81 _mach_constant_base_node->add_req(C->root()); 82 } 83 return _mach_constant_base_node; 84 } 85 86 87 /// Support for intrinsics. 88 89 // Return the index at which m must be inserted (or already exists). 90 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 91 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 92 #ifdef ASSERT 93 for (int i = 1; i < _intrinsics->length(); i++) { 94 CallGenerator* cg1 = _intrinsics->at(i-1); 95 CallGenerator* cg2 = _intrinsics->at(i); 96 assert(cg1->method() != cg2->method() 97 ? cg1->method() < cg2->method() 98 : cg1->is_virtual() < cg2->is_virtual(), 99 "compiler intrinsics list must stay sorted"); 100 } 101 #endif 102 // Binary search sorted list, in decreasing intervals [lo, hi]. 103 int lo = 0, hi = _intrinsics->length()-1; 104 while (lo <= hi) { 105 int mid = (uint)(hi + lo) / 2; 106 ciMethod* mid_m = _intrinsics->at(mid)->method(); 107 if (m < mid_m) { 108 hi = mid-1; 109 } else if (m > mid_m) { 110 lo = mid+1; 111 } else { 112 // look at minor sort key 113 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 114 if (is_virtual < mid_virt) { 115 hi = mid-1; 116 } else if (is_virtual > mid_virt) { 117 lo = mid+1; 118 } else { 119 return mid; // exact match 120 } 121 } 122 } 123 return lo; // inexact match 124 } 125 126 void Compile::register_intrinsic(CallGenerator* cg) { 127 if (_intrinsics == NULL) { 128 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 129 } 130 // This code is stolen from ciObjectFactory::insert. 131 // Really, GrowableArray should have methods for 132 // insert_at, remove_at, and binary_search. 133 int len = _intrinsics->length(); 134 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 135 if (index == len) { 136 _intrinsics->append(cg); 137 } else { 138 #ifdef ASSERT 139 CallGenerator* oldcg = _intrinsics->at(index); 140 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 141 #endif 142 _intrinsics->append(_intrinsics->at(len-1)); 143 int pos; 144 for (pos = len-2; pos >= index; pos--) { 145 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 146 } 147 _intrinsics->at_put(index, cg); 148 } 149 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 150 } 151 152 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 153 assert(m->is_loaded(), "don't try this on unloaded methods"); 154 if (_intrinsics != NULL) { 155 int index = intrinsic_insertion_index(m, is_virtual); 156 if (index < _intrinsics->length() 157 && _intrinsics->at(index)->method() == m 158 && _intrinsics->at(index)->is_virtual() == is_virtual) { 159 return _intrinsics->at(index); 160 } 161 } 162 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 163 if (m->intrinsic_id() != vmIntrinsics::_none && 164 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 165 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 166 if (cg != NULL) { 167 // Save it for next time: 168 register_intrinsic(cg); 169 return cg; 170 } else { 171 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 172 } 173 } 174 return NULL; 175 } 176 177 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 178 // in library_call.cpp. 179 180 181 #ifndef PRODUCT 182 // statistics gathering... 183 184 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 185 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 186 187 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 188 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 189 int oflags = _intrinsic_hist_flags[id]; 190 assert(flags != 0, "what happened?"); 191 if (is_virtual) { 192 flags |= _intrinsic_virtual; 193 } 194 bool changed = (flags != oflags); 195 if ((flags & _intrinsic_worked) != 0) { 196 juint count = (_intrinsic_hist_count[id] += 1); 197 if (count == 1) { 198 changed = true; // first time 199 } 200 // increment the overall count also: 201 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 202 } 203 if (changed) { 204 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 205 // Something changed about the intrinsic's virtuality. 206 if ((flags & _intrinsic_virtual) != 0) { 207 // This is the first use of this intrinsic as a virtual call. 208 if (oflags != 0) { 209 // We already saw it as a non-virtual, so note both cases. 210 flags |= _intrinsic_both; 211 } 212 } else if ((oflags & _intrinsic_both) == 0) { 213 // This is the first use of this intrinsic as a non-virtual 214 flags |= _intrinsic_both; 215 } 216 } 217 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 218 } 219 // update the overall flags also: 220 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 221 return changed; 222 } 223 224 static char* format_flags(int flags, char* buf) { 225 buf[0] = 0; 226 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 227 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 228 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 229 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 230 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 231 if (buf[0] == 0) strcat(buf, ","); 232 assert(buf[0] == ',', "must be"); 233 return &buf[1]; 234 } 235 236 void Compile::print_intrinsic_statistics() { 237 char flagsbuf[100]; 238 ttyLocker ttyl; 239 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 240 tty->print_cr("Compiler intrinsic usage:"); 241 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 242 if (total == 0) total = 1; // avoid div0 in case of no successes 243 #define PRINT_STAT_LINE(name, c, f) \ 244 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 245 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 246 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 247 int flags = _intrinsic_hist_flags[id]; 248 juint count = _intrinsic_hist_count[id]; 249 if ((flags | count) != 0) { 250 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 251 } 252 } 253 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 254 if (xtty != NULL) xtty->tail("statistics"); 255 } 256 257 void Compile::print_statistics() { 258 { ttyLocker ttyl; 259 if (xtty != NULL) xtty->head("statistics type='opto'"); 260 Parse::print_statistics(); 261 PhaseCCP::print_statistics(); 262 PhaseRegAlloc::print_statistics(); 263 Scheduling::print_statistics(); 264 PhasePeephole::print_statistics(); 265 PhaseIdealLoop::print_statistics(); 266 if (xtty != NULL) xtty->tail("statistics"); 267 } 268 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 269 // put this under its own <statistics> element. 270 print_intrinsic_statistics(); 271 } 272 } 273 #endif //PRODUCT 274 275 // Support for bundling info 276 Bundle* Compile::node_bundling(const Node *n) { 277 assert(valid_bundle_info(n), "oob"); 278 return &_node_bundling_base[n->_idx]; 279 } 280 281 bool Compile::valid_bundle_info(const Node *n) { 282 return (_node_bundling_limit > n->_idx); 283 } 284 285 286 void Compile::gvn_replace_by(Node* n, Node* nn) { 287 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 288 Node* use = n->last_out(i); 289 bool is_in_table = initial_gvn()->hash_delete(use); 290 uint uses_found = 0; 291 for (uint j = 0; j < use->len(); j++) { 292 if (use->in(j) == n) { 293 if (j < use->req()) 294 use->set_req(j, nn); 295 else 296 use->set_prec(j, nn); 297 uses_found++; 298 } 299 } 300 if (is_in_table) { 301 // reinsert into table 302 initial_gvn()->hash_find_insert(use); 303 } 304 record_for_igvn(use); 305 i -= uses_found; // we deleted 1 or more copies of this edge 306 } 307 } 308 309 310 static inline bool not_a_node(const Node* n) { 311 if (n == NULL) return true; 312 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 313 if (*(address*)n == badAddress) return true; // kill by Node::destruct 314 return false; 315 } 316 317 // Identify all nodes that are reachable from below, useful. 318 // Use breadth-first pass that records state in a Unique_Node_List, 319 // recursive traversal is slower. 320 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 321 int estimated_worklist_size = live_nodes(); 322 useful.map( estimated_worklist_size, NULL ); // preallocate space 323 324 // Initialize worklist 325 if (root() != NULL) { useful.push(root()); } 326 // If 'top' is cached, declare it useful to preserve cached node 327 if( cached_top_node() ) { useful.push(cached_top_node()); } 328 329 // Push all useful nodes onto the list, breadthfirst 330 for( uint next = 0; next < useful.size(); ++next ) { 331 assert( next < unique(), "Unique useful nodes < total nodes"); 332 Node *n = useful.at(next); 333 uint max = n->len(); 334 for( uint i = 0; i < max; ++i ) { 335 Node *m = n->in(i); 336 if (not_a_node(m)) continue; 337 useful.push(m); 338 } 339 } 340 } 341 342 // Update dead_node_list with any missing dead nodes using useful 343 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 344 void Compile::update_dead_node_list(Unique_Node_List &useful) { 345 uint max_idx = unique(); 346 VectorSet& useful_node_set = useful.member_set(); 347 348 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 349 // If node with index node_idx is not in useful set, 350 // mark it as dead in dead node list. 351 if (! useful_node_set.test(node_idx) ) { 352 record_dead_node(node_idx); 353 } 354 } 355 } 356 357 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 358 int shift = 0; 359 for (int i = 0; i < inlines->length(); i++) { 360 CallGenerator* cg = inlines->at(i); 361 CallNode* call = cg->call_node(); 362 if (shift > 0) { 363 inlines->at_put(i-shift, cg); 364 } 365 if (!useful.member(call)) { 366 shift++; 367 } 368 } 369 inlines->trunc_to(inlines->length()-shift); 370 } 371 372 // Disconnect all useless nodes by disconnecting those at the boundary. 373 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 374 uint next = 0; 375 while (next < useful.size()) { 376 Node *n = useful.at(next++); 377 if (n->is_SafePoint()) { 378 // We're done with a parsing phase. Replaced nodes are not valid 379 // beyond that point. 380 n->as_SafePoint()->delete_replaced_nodes(); 381 } 382 // Use raw traversal of out edges since this code removes out edges 383 int max = n->outcnt(); 384 for (int j = 0; j < max; ++j) { 385 Node* child = n->raw_out(j); 386 if (! useful.member(child)) { 387 assert(!child->is_top() || child != top(), 388 "If top is cached in Compile object it is in useful list"); 389 // Only need to remove this out-edge to the useless node 390 n->raw_del_out(j); 391 --j; 392 --max; 393 } 394 } 395 if (n->outcnt() == 1 && n->has_special_unique_user()) { 396 record_for_igvn(n->unique_out()); 397 } 398 } 399 // Remove useless macro and predicate opaq nodes 400 for (int i = C->macro_count()-1; i >= 0; i--) { 401 Node* n = C->macro_node(i); 402 if (!useful.member(n)) { 403 remove_macro_node(n); 404 } 405 } 406 #ifdef _LP64 407 // Remove useless CastII nodes with range check dependency 408 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 409 Node* cast = range_check_cast_node(i); 410 if (!useful.member(cast)) { 411 remove_range_check_cast(cast); 412 } 413 } 414 #endif 415 // Remove useless expensive node 416 for (int i = C->expensive_count()-1; i >= 0; i--) { 417 Node* n = C->expensive_node(i); 418 if (!useful.member(n)) { 419 remove_expensive_node(n); 420 } 421 } 422 // clean up the late inline lists 423 remove_useless_late_inlines(&_string_late_inlines, useful); 424 remove_useless_late_inlines(&_boxing_late_inlines, useful); 425 remove_useless_late_inlines(&_late_inlines, useful); 426 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 427 } 428 429 //------------------------------frame_size_in_words----------------------------- 430 // frame_slots in units of words 431 int Compile::frame_size_in_words() const { 432 // shift is 0 in LP32 and 1 in LP64 433 const int shift = (LogBytesPerWord - LogBytesPerInt); 434 int words = _frame_slots >> shift; 435 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 436 return words; 437 } 438 439 // To bang the stack of this compiled method we use the stack size 440 // that the interpreter would need in case of a deoptimization. This 441 // removes the need to bang the stack in the deoptimization blob which 442 // in turn simplifies stack overflow handling. 443 int Compile::bang_size_in_bytes() const { 444 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 445 } 446 447 // ============================================================================ 448 //------------------------------CompileWrapper--------------------------------- 449 class CompileWrapper : public StackObj { 450 Compile *const _compile; 451 public: 452 CompileWrapper(Compile* compile); 453 454 ~CompileWrapper(); 455 }; 456 457 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 458 // the Compile* pointer is stored in the current ciEnv: 459 ciEnv* env = compile->env(); 460 assert(env == ciEnv::current(), "must already be a ciEnv active"); 461 assert(env->compiler_data() == NULL, "compile already active?"); 462 env->set_compiler_data(compile); 463 assert(compile == Compile::current(), "sanity"); 464 465 compile->set_type_dict(NULL); 466 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 467 compile->clone_map().set_clone_idx(0); 468 compile->set_type_hwm(NULL); 469 compile->set_type_last_size(0); 470 compile->set_last_tf(NULL, NULL); 471 compile->set_indexSet_arena(NULL); 472 compile->set_indexSet_free_block_list(NULL); 473 compile->init_type_arena(); 474 Type::Initialize(compile); 475 _compile->set_scratch_buffer_blob(NULL); 476 _compile->begin_method(); 477 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 478 } 479 CompileWrapper::~CompileWrapper() { 480 _compile->end_method(); 481 if (_compile->scratch_buffer_blob() != NULL) 482 BufferBlob::free(_compile->scratch_buffer_blob()); 483 _compile->env()->set_compiler_data(NULL); 484 } 485 486 487 //----------------------------print_compile_messages--------------------------- 488 void Compile::print_compile_messages() { 489 #ifndef PRODUCT 490 // Check if recompiling 491 if (_subsume_loads == false && PrintOpto) { 492 // Recompiling without allowing machine instructions to subsume loads 493 tty->print_cr("*********************************************************"); 494 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 495 tty->print_cr("*********************************************************"); 496 } 497 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 498 // Recompiling without escape analysis 499 tty->print_cr("*********************************************************"); 500 tty->print_cr("** Bailout: Recompile without escape analysis **"); 501 tty->print_cr("*********************************************************"); 502 } 503 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 504 // Recompiling without boxing elimination 505 tty->print_cr("*********************************************************"); 506 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 507 tty->print_cr("*********************************************************"); 508 } 509 if (C->directive()->BreakAtCompileOption) { 510 // Open the debugger when compiling this method. 511 tty->print("### Breaking when compiling: "); 512 method()->print_short_name(); 513 tty->cr(); 514 BREAKPOINT; 515 } 516 517 if( PrintOpto ) { 518 if (is_osr_compilation()) { 519 tty->print("[OSR]%3d", _compile_id); 520 } else { 521 tty->print("%3d", _compile_id); 522 } 523 } 524 #endif 525 } 526 527 528 //-----------------------init_scratch_buffer_blob------------------------------ 529 // Construct a temporary BufferBlob and cache it for this compile. 530 void Compile::init_scratch_buffer_blob(int const_size) { 531 // If there is already a scratch buffer blob allocated and the 532 // constant section is big enough, use it. Otherwise free the 533 // current and allocate a new one. 534 BufferBlob* blob = scratch_buffer_blob(); 535 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 536 // Use the current blob. 537 } else { 538 if (blob != NULL) { 539 BufferBlob::free(blob); 540 } 541 542 ResourceMark rm; 543 _scratch_const_size = const_size; 544 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 545 blob = BufferBlob::create("Compile::scratch_buffer", size); 546 // Record the buffer blob for next time. 547 set_scratch_buffer_blob(blob); 548 // Have we run out of code space? 549 if (scratch_buffer_blob() == NULL) { 550 // Let CompilerBroker disable further compilations. 551 record_failure("Not enough space for scratch buffer in CodeCache"); 552 return; 553 } 554 } 555 556 // Initialize the relocation buffers 557 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 558 set_scratch_locs_memory(locs_buf); 559 } 560 561 562 //-----------------------scratch_emit_size------------------------------------- 563 // Helper function that computes size by emitting code 564 uint Compile::scratch_emit_size(const Node* n) { 565 // Start scratch_emit_size section. 566 set_in_scratch_emit_size(true); 567 568 // Emit into a trash buffer and count bytes emitted. 569 // This is a pretty expensive way to compute a size, 570 // but it works well enough if seldom used. 571 // All common fixed-size instructions are given a size 572 // method by the AD file. 573 // Note that the scratch buffer blob and locs memory are 574 // allocated at the beginning of the compile task, and 575 // may be shared by several calls to scratch_emit_size. 576 // The allocation of the scratch buffer blob is particularly 577 // expensive, since it has to grab the code cache lock. 578 BufferBlob* blob = this->scratch_buffer_blob(); 579 assert(blob != NULL, "Initialize BufferBlob at start"); 580 assert(blob->size() > MAX_inst_size, "sanity"); 581 relocInfo* locs_buf = scratch_locs_memory(); 582 address blob_begin = blob->content_begin(); 583 address blob_end = (address)locs_buf; 584 assert(blob->content_contains(blob_end), "sanity"); 585 CodeBuffer buf(blob_begin, blob_end - blob_begin); 586 buf.initialize_consts_size(_scratch_const_size); 587 buf.initialize_stubs_size(MAX_stubs_size); 588 assert(locs_buf != NULL, "sanity"); 589 int lsize = MAX_locs_size / 3; 590 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 591 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 592 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 593 594 // Do the emission. 595 596 Label fakeL; // Fake label for branch instructions. 597 Label* saveL = NULL; 598 uint save_bnum = 0; 599 bool is_branch = n->is_MachBranch(); 600 if (is_branch) { 601 MacroAssembler masm(&buf); 602 masm.bind(fakeL); 603 n->as_MachBranch()->save_label(&saveL, &save_bnum); 604 n->as_MachBranch()->label_set(&fakeL, 0); 605 } 606 n->emit(buf, this->regalloc()); 607 608 // Emitting into the scratch buffer should not fail 609 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 610 611 if (is_branch) // Restore label. 612 n->as_MachBranch()->label_set(saveL, save_bnum); 613 614 // End scratch_emit_size section. 615 set_in_scratch_emit_size(false); 616 617 return buf.insts_size(); 618 } 619 620 621 // ============================================================================ 622 //------------------------------Compile standard------------------------------- 623 debug_only( int Compile::_debug_idx = 100000; ) 624 625 // Compile a method. entry_bci is -1 for normal compilations and indicates 626 // the continuation bci for on stack replacement. 627 628 629 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 630 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 631 : Phase(Compiler), 632 _env(ci_env), 633 _directive(directive), 634 _log(ci_env->log()), 635 _compile_id(ci_env->compile_id()), 636 _save_argument_registers(false), 637 _stub_name(NULL), 638 _stub_function(NULL), 639 _stub_entry_point(NULL), 640 _method(target), 641 _entry_bci(osr_bci), 642 _initial_gvn(NULL), 643 _for_igvn(NULL), 644 _warm_calls(NULL), 645 _subsume_loads(subsume_loads), 646 _do_escape_analysis(do_escape_analysis), 647 _eliminate_boxing(eliminate_boxing), 648 _failure_reason(NULL), 649 _code_buffer("Compile::Fill_buffer"), 650 _orig_pc_slot(0), 651 _orig_pc_slot_offset_in_bytes(0), 652 _has_method_handle_invokes(false), 653 _mach_constant_base_node(NULL), 654 _node_bundling_limit(0), 655 _node_bundling_base(NULL), 656 _java_calls(0), 657 _inner_loops(0), 658 _scratch_const_size(-1), 659 _in_scratch_emit_size(false), 660 _dead_node_list(comp_arena()), 661 _dead_node_count(0), 662 #ifndef PRODUCT 663 _trace_opto_output(directive->TraceOptoOutputOption), 664 _in_dump_cnt(0), 665 _printer(IdealGraphPrinter::printer()), 666 #endif 667 _congraph(NULL), 668 _comp_arena(mtCompiler), 669 _node_arena(mtCompiler), 670 _old_arena(mtCompiler), 671 _Compile_types(mtCompiler), 672 _replay_inline_data(NULL), 673 _late_inlines(comp_arena(), 2, 0, NULL), 674 _string_late_inlines(comp_arena(), 2, 0, NULL), 675 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 676 _late_inlines_pos(0), 677 _number_of_mh_late_inlines(0), 678 _inlining_progress(false), 679 _inlining_incrementally(false), 680 _print_inlining_list(NULL), 681 _print_inlining_stream(NULL), 682 _print_inlining_idx(0), 683 _print_inlining_output(NULL), 684 _interpreter_frame_size(0), 685 _max_node_limit(MaxNodeLimit), 686 _has_reserved_stack_access(target->has_reserved_stack_access()) { 687 C = this; 688 #ifndef PRODUCT 689 if (_printer != NULL) { 690 _printer->set_compile(this); 691 } 692 #endif 693 CompileWrapper cw(this); 694 695 if (CITimeVerbose) { 696 tty->print(" "); 697 target->holder()->name()->print(); 698 tty->print("."); 699 target->print_short_name(); 700 tty->print(" "); 701 } 702 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 703 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 704 705 #ifndef PRODUCT 706 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 707 if (!print_opto_assembly) { 708 bool print_assembly = directive->PrintAssemblyOption; 709 if (print_assembly && !Disassembler::can_decode()) { 710 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 711 print_opto_assembly = true; 712 } 713 } 714 set_print_assembly(print_opto_assembly); 715 set_parsed_irreducible_loop(false); 716 717 if (directive->ReplayInlineOption) { 718 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 719 } 720 #endif 721 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 722 set_print_intrinsics(directive->PrintIntrinsicsOption); 723 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 724 725 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 726 // Make sure the method being compiled gets its own MDO, 727 // so we can at least track the decompile_count(). 728 // Need MDO to record RTM code generation state. 729 method()->ensure_method_data(); 730 } 731 732 Init(::AliasLevel); 733 734 735 print_compile_messages(); 736 737 _ilt = InlineTree::build_inline_tree_root(); 738 739 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 740 assert(num_alias_types() >= AliasIdxRaw, ""); 741 742 #define MINIMUM_NODE_HASH 1023 743 // Node list that Iterative GVN will start with 744 Unique_Node_List for_igvn(comp_arena()); 745 set_for_igvn(&for_igvn); 746 747 // GVN that will be run immediately on new nodes 748 uint estimated_size = method()->code_size()*4+64; 749 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 750 PhaseGVN gvn(node_arena(), estimated_size); 751 set_initial_gvn(&gvn); 752 753 print_inlining_init(); 754 { // Scope for timing the parser 755 TracePhase tp("parse", &timers[_t_parser]); 756 757 // Put top into the hash table ASAP. 758 initial_gvn()->transform_no_reclaim(top()); 759 760 // Set up tf(), start(), and find a CallGenerator. 761 CallGenerator* cg = NULL; 762 if (is_osr_compilation()) { 763 const TypeTuple *domain = StartOSRNode::osr_domain(); 764 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 765 init_tf(TypeFunc::make(domain, range)); 766 StartNode* s = new StartOSRNode(root(), domain); 767 initial_gvn()->set_type_bottom(s); 768 init_start(s); 769 cg = CallGenerator::for_osr(method(), entry_bci()); 770 } else { 771 // Normal case. 772 init_tf(TypeFunc::make(method())); 773 StartNode* s = new StartNode(root(), tf()->domain()); 774 initial_gvn()->set_type_bottom(s); 775 init_start(s); 776 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 777 // With java.lang.ref.reference.get() we must go through the 778 // intrinsic when G1 is enabled - even when get() is the root 779 // method of the compile - so that, if necessary, the value in 780 // the referent field of the reference object gets recorded by 781 // the pre-barrier code. 782 // Specifically, if G1 is enabled, the value in the referent 783 // field is recorded by the G1 SATB pre barrier. This will 784 // result in the referent being marked live and the reference 785 // object removed from the list of discovered references during 786 // reference processing. 787 cg = find_intrinsic(method(), false); 788 } 789 if (cg == NULL) { 790 float past_uses = method()->interpreter_invocation_count(); 791 float expected_uses = past_uses; 792 cg = CallGenerator::for_inline(method(), expected_uses); 793 } 794 } 795 if (failing()) return; 796 if (cg == NULL) { 797 record_method_not_compilable_all_tiers("cannot parse method"); 798 return; 799 } 800 JVMState* jvms = build_start_state(start(), tf()); 801 if ((jvms = cg->generate(jvms)) == NULL) { 802 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 803 record_method_not_compilable("method parse failed"); 804 } 805 return; 806 } 807 GraphKit kit(jvms); 808 809 if (!kit.stopped()) { 810 // Accept return values, and transfer control we know not where. 811 // This is done by a special, unique ReturnNode bound to root. 812 return_values(kit.jvms()); 813 } 814 815 if (kit.has_exceptions()) { 816 // Any exceptions that escape from this call must be rethrown 817 // to whatever caller is dynamically above us on the stack. 818 // This is done by a special, unique RethrowNode bound to root. 819 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 820 } 821 822 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 823 824 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 825 inline_string_calls(true); 826 } 827 828 if (failing()) return; 829 830 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 831 832 // Remove clutter produced by parsing. 833 if (!failing()) { 834 ResourceMark rm; 835 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 836 } 837 } 838 839 // Note: Large methods are capped off in do_one_bytecode(). 840 if (failing()) return; 841 842 // After parsing, node notes are no longer automagic. 843 // They must be propagated by register_new_node_with_optimizer(), 844 // clone(), or the like. 845 set_default_node_notes(NULL); 846 847 for (;;) { 848 int successes = Inline_Warm(); 849 if (failing()) return; 850 if (successes == 0) break; 851 } 852 853 // Drain the list. 854 Finish_Warm(); 855 #ifndef PRODUCT 856 if (_printer && _printer->should_print(1)) { 857 _printer->print_inlining(); 858 } 859 #endif 860 861 if (failing()) return; 862 NOT_PRODUCT( verify_graph_edges(); ) 863 864 // Now optimize 865 Optimize(); 866 if (failing()) return; 867 NOT_PRODUCT( verify_graph_edges(); ) 868 869 #ifndef PRODUCT 870 if (PrintIdeal) { 871 ttyLocker ttyl; // keep the following output all in one block 872 // This output goes directly to the tty, not the compiler log. 873 // To enable tools to match it up with the compilation activity, 874 // be sure to tag this tty output with the compile ID. 875 if (xtty != NULL) { 876 xtty->head("ideal compile_id='%d'%s", compile_id(), 877 is_osr_compilation() ? " compile_kind='osr'" : 878 ""); 879 } 880 root()->dump(9999); 881 if (xtty != NULL) { 882 xtty->tail("ideal"); 883 } 884 } 885 #endif 886 887 NOT_PRODUCT( verify_barriers(); ) 888 889 // Dump compilation data to replay it. 890 if (directive->DumpReplayOption) { 891 env()->dump_replay_data(_compile_id); 892 } 893 if (directive->DumpInlineOption && (ilt() != NULL)) { 894 env()->dump_inline_data(_compile_id); 895 } 896 897 // Now that we know the size of all the monitors we can add a fixed slot 898 // for the original deopt pc. 899 900 _orig_pc_slot = fixed_slots(); 901 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 902 set_fixed_slots(next_slot); 903 904 // Compute when to use implicit null checks. Used by matching trap based 905 // nodes and NullCheck optimization. 906 set_allowed_deopt_reasons(); 907 908 // Now generate code 909 Code_Gen(); 910 if (failing()) return; 911 912 // Check if we want to skip execution of all compiled code. 913 { 914 #ifndef PRODUCT 915 if (OptoNoExecute) { 916 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 917 return; 918 } 919 #endif 920 TracePhase tp("install_code", &timers[_t_registerMethod]); 921 922 if (is_osr_compilation()) { 923 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 924 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 925 } else { 926 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 927 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 928 } 929 930 env()->register_method(_method, _entry_bci, 931 &_code_offsets, 932 _orig_pc_slot_offset_in_bytes, 933 code_buffer(), 934 frame_size_in_words(), _oop_map_set, 935 &_handler_table, &_inc_table, 936 compiler, 937 has_unsafe_access(), 938 SharedRuntime::is_wide_vector(max_vector_size()), 939 _directive, 940 rtm_state() 941 ); 942 943 if (log() != NULL) // Print code cache state into compiler log 944 log()->code_cache_state(); 945 } 946 } 947 948 //------------------------------Compile---------------------------------------- 949 // Compile a runtime stub 950 Compile::Compile( ciEnv* ci_env, 951 TypeFunc_generator generator, 952 address stub_function, 953 const char *stub_name, 954 int is_fancy_jump, 955 bool pass_tls, 956 bool save_arg_registers, 957 bool return_pc, 958 DirectiveSet* directive) 959 : Phase(Compiler), 960 _env(ci_env), 961 _directive(directive), 962 _log(ci_env->log()), 963 _compile_id(0), 964 _save_argument_registers(save_arg_registers), 965 _method(NULL), 966 _stub_name(stub_name), 967 _stub_function(stub_function), 968 _stub_entry_point(NULL), 969 _entry_bci(InvocationEntryBci), 970 _initial_gvn(NULL), 971 _for_igvn(NULL), 972 _warm_calls(NULL), 973 _orig_pc_slot(0), 974 _orig_pc_slot_offset_in_bytes(0), 975 _subsume_loads(true), 976 _do_escape_analysis(false), 977 _eliminate_boxing(false), 978 _failure_reason(NULL), 979 _code_buffer("Compile::Fill_buffer"), 980 _has_method_handle_invokes(false), 981 _mach_constant_base_node(NULL), 982 _node_bundling_limit(0), 983 _node_bundling_base(NULL), 984 _java_calls(0), 985 _inner_loops(0), 986 #ifndef PRODUCT 987 _trace_opto_output(TraceOptoOutput), 988 _in_dump_cnt(0), 989 _printer(NULL), 990 #endif 991 _comp_arena(mtCompiler), 992 _node_arena(mtCompiler), 993 _old_arena(mtCompiler), 994 _Compile_types(mtCompiler), 995 _dead_node_list(comp_arena()), 996 _dead_node_count(0), 997 _congraph(NULL), 998 _replay_inline_data(NULL), 999 _number_of_mh_late_inlines(0), 1000 _inlining_progress(false), 1001 _inlining_incrementally(false), 1002 _print_inlining_list(NULL), 1003 _print_inlining_stream(NULL), 1004 _print_inlining_idx(0), 1005 _print_inlining_output(NULL), 1006 _allowed_reasons(0), 1007 _interpreter_frame_size(0), 1008 _max_node_limit(MaxNodeLimit) { 1009 C = this; 1010 1011 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1012 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1013 1014 #ifndef PRODUCT 1015 set_print_assembly(PrintFrameConverterAssembly); 1016 set_parsed_irreducible_loop(false); 1017 #endif 1018 set_has_irreducible_loop(false); // no loops 1019 1020 CompileWrapper cw(this); 1021 Init(/*AliasLevel=*/ 0); 1022 init_tf((*generator)()); 1023 1024 { 1025 // The following is a dummy for the sake of GraphKit::gen_stub 1026 Unique_Node_List for_igvn(comp_arena()); 1027 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1028 PhaseGVN gvn(Thread::current()->resource_area(),255); 1029 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1030 gvn.transform_no_reclaim(top()); 1031 1032 GraphKit kit; 1033 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1034 } 1035 1036 NOT_PRODUCT( verify_graph_edges(); ) 1037 Code_Gen(); 1038 if (failing()) return; 1039 1040 1041 // Entry point will be accessed using compile->stub_entry_point(); 1042 if (code_buffer() == NULL) { 1043 Matcher::soft_match_failure(); 1044 } else { 1045 if (PrintAssembly && (WizardMode || Verbose)) 1046 tty->print_cr("### Stub::%s", stub_name); 1047 1048 if (!failing()) { 1049 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1050 1051 // Make the NMethod 1052 // For now we mark the frame as never safe for profile stackwalking 1053 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1054 code_buffer(), 1055 CodeOffsets::frame_never_safe, 1056 // _code_offsets.value(CodeOffsets::Frame_Complete), 1057 frame_size_in_words(), 1058 _oop_map_set, 1059 save_arg_registers); 1060 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1061 1062 _stub_entry_point = rs->entry_point(); 1063 } 1064 } 1065 } 1066 1067 //------------------------------Init------------------------------------------- 1068 // Prepare for a single compilation 1069 void Compile::Init(int aliaslevel) { 1070 _unique = 0; 1071 _regalloc = NULL; 1072 1073 _tf = NULL; // filled in later 1074 _top = NULL; // cached later 1075 _matcher = NULL; // filled in later 1076 _cfg = NULL; // filled in later 1077 1078 set_24_bit_selection_and_mode(Use24BitFP, false); 1079 1080 _node_note_array = NULL; 1081 _default_node_notes = NULL; 1082 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1083 1084 _immutable_memory = NULL; // filled in at first inquiry 1085 1086 // Globally visible Nodes 1087 // First set TOP to NULL to give safe behavior during creation of RootNode 1088 set_cached_top_node(NULL); 1089 set_root(new RootNode()); 1090 // Now that you have a Root to point to, create the real TOP 1091 set_cached_top_node( new ConNode(Type::TOP) ); 1092 set_recent_alloc(NULL, NULL); 1093 1094 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1095 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1096 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1097 env()->set_dependencies(new Dependencies(env())); 1098 1099 _fixed_slots = 0; 1100 set_has_split_ifs(false); 1101 set_has_loops(has_method() && method()->has_loops()); // first approximation 1102 set_has_stringbuilder(false); 1103 set_has_boxed_value(false); 1104 _trap_can_recompile = false; // no traps emitted yet 1105 _major_progress = true; // start out assuming good things will happen 1106 set_has_unsafe_access(false); 1107 set_max_vector_size(0); 1108 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1109 set_decompile_count(0); 1110 1111 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1112 set_num_loop_opts(LoopOptsCount); 1113 set_do_inlining(Inline); 1114 set_max_inline_size(MaxInlineSize); 1115 set_freq_inline_size(FreqInlineSize); 1116 set_do_scheduling(OptoScheduling); 1117 set_do_count_invocations(false); 1118 set_do_method_data_update(false); 1119 1120 set_do_vector_loop(false); 1121 1122 if (AllowVectorizeOnDemand) { 1123 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1124 set_do_vector_loop(true); 1125 NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii());}) 1126 } else if (has_method() && method()->name() != 0 && 1127 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1128 set_do_vector_loop(true); 1129 } 1130 } 1131 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1132 NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n", method()->name()->as_quoted_ascii());}) 1133 1134 set_age_code(has_method() && method()->profile_aging()); 1135 set_rtm_state(NoRTM); // No RTM lock eliding by default 1136 _max_node_limit = _directive->MaxNodeLimitOption; 1137 1138 #if INCLUDE_RTM_OPT 1139 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1140 int rtm_state = method()->method_data()->rtm_state(); 1141 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1142 // Don't generate RTM lock eliding code. 1143 set_rtm_state(NoRTM); 1144 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1145 // Generate RTM lock eliding code without abort ratio calculation code. 1146 set_rtm_state(UseRTM); 1147 } else if (UseRTMDeopt) { 1148 // Generate RTM lock eliding code and include abort ratio calculation 1149 // code if UseRTMDeopt is on. 1150 set_rtm_state(ProfileRTM); 1151 } 1152 } 1153 #endif 1154 if (debug_info()->recording_non_safepoints()) { 1155 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1156 (comp_arena(), 8, 0, NULL)); 1157 set_default_node_notes(Node_Notes::make(this)); 1158 } 1159 1160 // // -- Initialize types before each compile -- 1161 // // Update cached type information 1162 // if( _method && _method->constants() ) 1163 // Type::update_loaded_types(_method, _method->constants()); 1164 1165 // Init alias_type map. 1166 if (!_do_escape_analysis && aliaslevel == 3) 1167 aliaslevel = 2; // No unique types without escape analysis 1168 _AliasLevel = aliaslevel; 1169 const int grow_ats = 16; 1170 _max_alias_types = grow_ats; 1171 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1172 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1173 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1174 { 1175 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1176 } 1177 // Initialize the first few types. 1178 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1179 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1180 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1181 _num_alias_types = AliasIdxRaw+1; 1182 // Zero out the alias type cache. 1183 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1184 // A NULL adr_type hits in the cache right away. Preload the right answer. 1185 probe_alias_cache(NULL)->_index = AliasIdxTop; 1186 1187 _intrinsics = NULL; 1188 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1189 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1190 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1191 #ifdef _LP64 1192 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1193 #endif 1194 register_library_intrinsics(); 1195 } 1196 1197 //---------------------------init_start---------------------------------------- 1198 // Install the StartNode on this compile object. 1199 void Compile::init_start(StartNode* s) { 1200 if (failing()) 1201 return; // already failing 1202 assert(s == start(), ""); 1203 } 1204 1205 /** 1206 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1207 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1208 * the ideal graph. 1209 */ 1210 StartNode* Compile::start() const { 1211 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1212 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1213 Node* start = root()->fast_out(i); 1214 if (start->is_Start()) { 1215 return start->as_Start(); 1216 } 1217 } 1218 fatal("Did not find Start node!"); 1219 return NULL; 1220 } 1221 1222 //-------------------------------immutable_memory------------------------------------- 1223 // Access immutable memory 1224 Node* Compile::immutable_memory() { 1225 if (_immutable_memory != NULL) { 1226 return _immutable_memory; 1227 } 1228 StartNode* s = start(); 1229 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1230 Node *p = s->fast_out(i); 1231 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1232 _immutable_memory = p; 1233 return _immutable_memory; 1234 } 1235 } 1236 ShouldNotReachHere(); 1237 return NULL; 1238 } 1239 1240 //----------------------set_cached_top_node------------------------------------ 1241 // Install the cached top node, and make sure Node::is_top works correctly. 1242 void Compile::set_cached_top_node(Node* tn) { 1243 if (tn != NULL) verify_top(tn); 1244 Node* old_top = _top; 1245 _top = tn; 1246 // Calling Node::setup_is_top allows the nodes the chance to adjust 1247 // their _out arrays. 1248 if (_top != NULL) _top->setup_is_top(); 1249 if (old_top != NULL) old_top->setup_is_top(); 1250 assert(_top == NULL || top()->is_top(), ""); 1251 } 1252 1253 #ifdef ASSERT 1254 uint Compile::count_live_nodes_by_graph_walk() { 1255 Unique_Node_List useful(comp_arena()); 1256 // Get useful node list by walking the graph. 1257 identify_useful_nodes(useful); 1258 return useful.size(); 1259 } 1260 1261 void Compile::print_missing_nodes() { 1262 1263 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1264 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1265 return; 1266 } 1267 1268 // This is an expensive function. It is executed only when the user 1269 // specifies VerifyIdealNodeCount option or otherwise knows the 1270 // additional work that needs to be done to identify reachable nodes 1271 // by walking the flow graph and find the missing ones using 1272 // _dead_node_list. 1273 1274 Unique_Node_List useful(comp_arena()); 1275 // Get useful node list by walking the graph. 1276 identify_useful_nodes(useful); 1277 1278 uint l_nodes = C->live_nodes(); 1279 uint l_nodes_by_walk = useful.size(); 1280 1281 if (l_nodes != l_nodes_by_walk) { 1282 if (_log != NULL) { 1283 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1284 _log->stamp(); 1285 _log->end_head(); 1286 } 1287 VectorSet& useful_member_set = useful.member_set(); 1288 int last_idx = l_nodes_by_walk; 1289 for (int i = 0; i < last_idx; i++) { 1290 if (useful_member_set.test(i)) { 1291 if (_dead_node_list.test(i)) { 1292 if (_log != NULL) { 1293 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1294 } 1295 if (PrintIdealNodeCount) { 1296 // Print the log message to tty 1297 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1298 useful.at(i)->dump(); 1299 } 1300 } 1301 } 1302 else if (! _dead_node_list.test(i)) { 1303 if (_log != NULL) { 1304 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1305 } 1306 if (PrintIdealNodeCount) { 1307 // Print the log message to tty 1308 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1309 } 1310 } 1311 } 1312 if (_log != NULL) { 1313 _log->tail("mismatched_nodes"); 1314 } 1315 } 1316 } 1317 void Compile::record_modified_node(Node* n) { 1318 if (_modified_nodes != NULL && !_inlining_incrementally && 1319 n->outcnt() != 0 && !n->is_Con()) { 1320 _modified_nodes->push(n); 1321 } 1322 } 1323 1324 void Compile::remove_modified_node(Node* n) { 1325 if (_modified_nodes != NULL) { 1326 _modified_nodes->remove(n); 1327 } 1328 } 1329 #endif 1330 1331 #ifndef PRODUCT 1332 void Compile::verify_top(Node* tn) const { 1333 if (tn != NULL) { 1334 assert(tn->is_Con(), "top node must be a constant"); 1335 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1336 assert(tn->in(0) != NULL, "must have live top node"); 1337 } 1338 } 1339 #endif 1340 1341 1342 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1343 1344 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1345 guarantee(arr != NULL, ""); 1346 int num_blocks = arr->length(); 1347 if (grow_by < num_blocks) grow_by = num_blocks; 1348 int num_notes = grow_by * _node_notes_block_size; 1349 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1350 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1351 while (num_notes > 0) { 1352 arr->append(notes); 1353 notes += _node_notes_block_size; 1354 num_notes -= _node_notes_block_size; 1355 } 1356 assert(num_notes == 0, "exact multiple, please"); 1357 } 1358 1359 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1360 if (source == NULL || dest == NULL) return false; 1361 1362 if (dest->is_Con()) 1363 return false; // Do not push debug info onto constants. 1364 1365 #ifdef ASSERT 1366 // Leave a bread crumb trail pointing to the original node: 1367 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1368 dest->set_debug_orig(source); 1369 } 1370 #endif 1371 1372 if (node_note_array() == NULL) 1373 return false; // Not collecting any notes now. 1374 1375 // This is a copy onto a pre-existing node, which may already have notes. 1376 // If both nodes have notes, do not overwrite any pre-existing notes. 1377 Node_Notes* source_notes = node_notes_at(source->_idx); 1378 if (source_notes == NULL || source_notes->is_clear()) return false; 1379 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1380 if (dest_notes == NULL || dest_notes->is_clear()) { 1381 return set_node_notes_at(dest->_idx, source_notes); 1382 } 1383 1384 Node_Notes merged_notes = (*source_notes); 1385 // The order of operations here ensures that dest notes will win... 1386 merged_notes.update_from(dest_notes); 1387 return set_node_notes_at(dest->_idx, &merged_notes); 1388 } 1389 1390 1391 //--------------------------allow_range_check_smearing------------------------- 1392 // Gating condition for coalescing similar range checks. 1393 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1394 // single covering check that is at least as strong as any of them. 1395 // If the optimization succeeds, the simplified (strengthened) range check 1396 // will always succeed. If it fails, we will deopt, and then give up 1397 // on the optimization. 1398 bool Compile::allow_range_check_smearing() const { 1399 // If this method has already thrown a range-check, 1400 // assume it was because we already tried range smearing 1401 // and it failed. 1402 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1403 return !already_trapped; 1404 } 1405 1406 1407 //------------------------------flatten_alias_type----------------------------- 1408 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1409 int offset = tj->offset(); 1410 TypePtr::PTR ptr = tj->ptr(); 1411 1412 // Known instance (scalarizable allocation) alias only with itself. 1413 bool is_known_inst = tj->isa_oopptr() != NULL && 1414 tj->is_oopptr()->is_known_instance(); 1415 1416 // Process weird unsafe references. 1417 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1418 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1419 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1420 tj = TypeOopPtr::BOTTOM; 1421 ptr = tj->ptr(); 1422 offset = tj->offset(); 1423 } 1424 1425 // Array pointers need some flattening 1426 const TypeAryPtr *ta = tj->isa_aryptr(); 1427 if (ta && ta->is_stable()) { 1428 // Erase stability property for alias analysis. 1429 tj = ta = ta->cast_to_stable(false); 1430 } 1431 if( ta && is_known_inst ) { 1432 if ( offset != Type::OffsetBot && 1433 offset > arrayOopDesc::length_offset_in_bytes() ) { 1434 offset = Type::OffsetBot; // Flatten constant access into array body only 1435 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1436 } 1437 } else if( ta && _AliasLevel >= 2 ) { 1438 // For arrays indexed by constant indices, we flatten the alias 1439 // space to include all of the array body. Only the header, klass 1440 // and array length can be accessed un-aliased. 1441 if( offset != Type::OffsetBot ) { 1442 if( ta->const_oop() ) { // MethodData* or Method* 1443 offset = Type::OffsetBot; // Flatten constant access into array body 1444 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1445 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1446 // range is OK as-is. 1447 tj = ta = TypeAryPtr::RANGE; 1448 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1449 tj = TypeInstPtr::KLASS; // all klass loads look alike 1450 ta = TypeAryPtr::RANGE; // generic ignored junk 1451 ptr = TypePtr::BotPTR; 1452 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1453 tj = TypeInstPtr::MARK; 1454 ta = TypeAryPtr::RANGE; // generic ignored junk 1455 ptr = TypePtr::BotPTR; 1456 } else { // Random constant offset into array body 1457 offset = Type::OffsetBot; // Flatten constant access into array body 1458 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1459 } 1460 } 1461 // Arrays of fixed size alias with arrays of unknown size. 1462 if (ta->size() != TypeInt::POS) { 1463 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1464 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1465 } 1466 // Arrays of known objects become arrays of unknown objects. 1467 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1468 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1469 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1470 } 1471 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1472 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1473 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1474 } 1475 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1476 // cannot be distinguished by bytecode alone. 1477 if (ta->elem() == TypeInt::BOOL) { 1478 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1479 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1480 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1481 } 1482 // During the 2nd round of IterGVN, NotNull castings are removed. 1483 // Make sure the Bottom and NotNull variants alias the same. 1484 // Also, make sure exact and non-exact variants alias the same. 1485 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1486 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1487 } 1488 } 1489 1490 // Oop pointers need some flattening 1491 const TypeInstPtr *to = tj->isa_instptr(); 1492 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1493 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1494 if( ptr == TypePtr::Constant ) { 1495 if (to->klass() != ciEnv::current()->Class_klass() || 1496 offset < k->size_helper() * wordSize) { 1497 // No constant oop pointers (such as Strings); they alias with 1498 // unknown strings. 1499 assert(!is_known_inst, "not scalarizable allocation"); 1500 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1501 } 1502 } else if( is_known_inst ) { 1503 tj = to; // Keep NotNull and klass_is_exact for instance type 1504 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1505 // During the 2nd round of IterGVN, NotNull castings are removed. 1506 // Make sure the Bottom and NotNull variants alias the same. 1507 // Also, make sure exact and non-exact variants alias the same. 1508 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1509 } 1510 if (to->speculative() != NULL) { 1511 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1512 } 1513 // Canonicalize the holder of this field 1514 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1515 // First handle header references such as a LoadKlassNode, even if the 1516 // object's klass is unloaded at compile time (4965979). 1517 if (!is_known_inst) { // Do it only for non-instance types 1518 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1519 } 1520 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1521 // Static fields are in the space above the normal instance 1522 // fields in the java.lang.Class instance. 1523 if (to->klass() != ciEnv::current()->Class_klass()) { 1524 to = NULL; 1525 tj = TypeOopPtr::BOTTOM; 1526 offset = tj->offset(); 1527 } 1528 } else { 1529 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1530 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1531 if( is_known_inst ) { 1532 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1533 } else { 1534 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1535 } 1536 } 1537 } 1538 } 1539 1540 // Klass pointers to object array klasses need some flattening 1541 const TypeKlassPtr *tk = tj->isa_klassptr(); 1542 if( tk ) { 1543 // If we are referencing a field within a Klass, we need 1544 // to assume the worst case of an Object. Both exact and 1545 // inexact types must flatten to the same alias class so 1546 // use NotNull as the PTR. 1547 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1548 1549 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1550 TypeKlassPtr::OBJECT->klass(), 1551 offset); 1552 } 1553 1554 ciKlass* klass = tk->klass(); 1555 if( klass->is_obj_array_klass() ) { 1556 ciKlass* k = TypeAryPtr::OOPS->klass(); 1557 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1558 k = TypeInstPtr::BOTTOM->klass(); 1559 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1560 } 1561 1562 // Check for precise loads from the primary supertype array and force them 1563 // to the supertype cache alias index. Check for generic array loads from 1564 // the primary supertype array and also force them to the supertype cache 1565 // alias index. Since the same load can reach both, we need to merge 1566 // these 2 disparate memories into the same alias class. Since the 1567 // primary supertype array is read-only, there's no chance of confusion 1568 // where we bypass an array load and an array store. 1569 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1570 if (offset == Type::OffsetBot || 1571 (offset >= primary_supers_offset && 1572 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1573 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1574 offset = in_bytes(Klass::secondary_super_cache_offset()); 1575 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1576 } 1577 } 1578 1579 // Flatten all Raw pointers together. 1580 if (tj->base() == Type::RawPtr) 1581 tj = TypeRawPtr::BOTTOM; 1582 1583 if (tj->base() == Type::AnyPtr) 1584 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1585 1586 // Flatten all to bottom for now 1587 switch( _AliasLevel ) { 1588 case 0: 1589 tj = TypePtr::BOTTOM; 1590 break; 1591 case 1: // Flatten to: oop, static, field or array 1592 switch (tj->base()) { 1593 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1594 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1595 case Type::AryPtr: // do not distinguish arrays at all 1596 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1597 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1598 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1599 default: ShouldNotReachHere(); 1600 } 1601 break; 1602 case 2: // No collapsing at level 2; keep all splits 1603 case 3: // No collapsing at level 3; keep all splits 1604 break; 1605 default: 1606 Unimplemented(); 1607 } 1608 1609 offset = tj->offset(); 1610 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1611 1612 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1613 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1614 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1615 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1616 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1617 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1618 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1619 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1620 assert( tj->ptr() != TypePtr::TopPTR && 1621 tj->ptr() != TypePtr::AnyNull && 1622 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1623 // assert( tj->ptr() != TypePtr::Constant || 1624 // tj->base() == Type::RawPtr || 1625 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1626 1627 return tj; 1628 } 1629 1630 void Compile::AliasType::Init(int i, const TypePtr* at) { 1631 _index = i; 1632 _adr_type = at; 1633 _field = NULL; 1634 _element = NULL; 1635 _is_rewritable = true; // default 1636 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1637 if (atoop != NULL && atoop->is_known_instance()) { 1638 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1639 _general_index = Compile::current()->get_alias_index(gt); 1640 } else { 1641 _general_index = 0; 1642 } 1643 } 1644 1645 //---------------------------------print_on------------------------------------ 1646 #ifndef PRODUCT 1647 void Compile::AliasType::print_on(outputStream* st) { 1648 if (index() < 10) 1649 st->print("@ <%d> ", index()); 1650 else st->print("@ <%d>", index()); 1651 st->print(is_rewritable() ? " " : " RO"); 1652 int offset = adr_type()->offset(); 1653 if (offset == Type::OffsetBot) 1654 st->print(" +any"); 1655 else st->print(" +%-3d", offset); 1656 st->print(" in "); 1657 adr_type()->dump_on(st); 1658 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1659 if (field() != NULL && tjp) { 1660 if (tjp->klass() != field()->holder() || 1661 tjp->offset() != field()->offset_in_bytes()) { 1662 st->print(" != "); 1663 field()->print(); 1664 st->print(" ***"); 1665 } 1666 } 1667 } 1668 1669 void print_alias_types() { 1670 Compile* C = Compile::current(); 1671 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1672 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1673 C->alias_type(idx)->print_on(tty); 1674 tty->cr(); 1675 } 1676 } 1677 #endif 1678 1679 1680 //----------------------------probe_alias_cache-------------------------------- 1681 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1682 intptr_t key = (intptr_t) adr_type; 1683 key ^= key >> logAliasCacheSize; 1684 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1685 } 1686 1687 1688 //-----------------------------grow_alias_types-------------------------------- 1689 void Compile::grow_alias_types() { 1690 const int old_ats = _max_alias_types; // how many before? 1691 const int new_ats = old_ats; // how many more? 1692 const int grow_ats = old_ats+new_ats; // how many now? 1693 _max_alias_types = grow_ats; 1694 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1695 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1696 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1697 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1698 } 1699 1700 1701 //--------------------------------find_alias_type------------------------------ 1702 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1703 if (_AliasLevel == 0) 1704 return alias_type(AliasIdxBot); 1705 1706 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1707 if (ace->_adr_type == adr_type) { 1708 return alias_type(ace->_index); 1709 } 1710 1711 // Handle special cases. 1712 if (adr_type == NULL) return alias_type(AliasIdxTop); 1713 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1714 1715 // Do it the slow way. 1716 const TypePtr* flat = flatten_alias_type(adr_type); 1717 1718 #ifdef ASSERT 1719 assert(flat == flatten_alias_type(flat), "idempotent"); 1720 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1721 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1722 const TypeOopPtr* foop = flat->is_oopptr(); 1723 // Scalarizable allocations have exact klass always. 1724 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1725 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1726 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1727 } 1728 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1729 #endif 1730 1731 int idx = AliasIdxTop; 1732 for (int i = 0; i < num_alias_types(); i++) { 1733 if (alias_type(i)->adr_type() == flat) { 1734 idx = i; 1735 break; 1736 } 1737 } 1738 1739 if (idx == AliasIdxTop) { 1740 if (no_create) return NULL; 1741 // Grow the array if necessary. 1742 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1743 // Add a new alias type. 1744 idx = _num_alias_types++; 1745 _alias_types[idx]->Init(idx, flat); 1746 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1747 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1748 if (flat->isa_instptr()) { 1749 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1750 && flat->is_instptr()->klass() == env()->Class_klass()) 1751 alias_type(idx)->set_rewritable(false); 1752 } 1753 if (flat->isa_aryptr()) { 1754 #ifdef ASSERT 1755 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1756 // (T_BYTE has the weakest alignment and size restrictions...) 1757 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1758 #endif 1759 if (flat->offset() == TypePtr::OffsetBot) { 1760 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1761 } 1762 } 1763 if (flat->isa_klassptr()) { 1764 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1765 alias_type(idx)->set_rewritable(false); 1766 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1767 alias_type(idx)->set_rewritable(false); 1768 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1769 alias_type(idx)->set_rewritable(false); 1770 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1771 alias_type(idx)->set_rewritable(false); 1772 } 1773 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1774 // but the base pointer type is not distinctive enough to identify 1775 // references into JavaThread.) 1776 1777 // Check for final fields. 1778 const TypeInstPtr* tinst = flat->isa_instptr(); 1779 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1780 ciField* field; 1781 if (tinst->const_oop() != NULL && 1782 tinst->klass() == ciEnv::current()->Class_klass() && 1783 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1784 // static field 1785 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1786 field = k->get_field_by_offset(tinst->offset(), true); 1787 } else { 1788 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1789 field = k->get_field_by_offset(tinst->offset(), false); 1790 } 1791 assert(field == NULL || 1792 original_field == NULL || 1793 (field->holder() == original_field->holder() && 1794 field->offset() == original_field->offset() && 1795 field->is_static() == original_field->is_static()), "wrong field?"); 1796 // Set field() and is_rewritable() attributes. 1797 if (field != NULL) alias_type(idx)->set_field(field); 1798 } 1799 } 1800 1801 // Fill the cache for next time. 1802 ace->_adr_type = adr_type; 1803 ace->_index = idx; 1804 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1805 1806 // Might as well try to fill the cache for the flattened version, too. 1807 AliasCacheEntry* face = probe_alias_cache(flat); 1808 if (face->_adr_type == NULL) { 1809 face->_adr_type = flat; 1810 face->_index = idx; 1811 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1812 } 1813 1814 return alias_type(idx); 1815 } 1816 1817 1818 Compile::AliasType* Compile::alias_type(ciField* field) { 1819 const TypeOopPtr* t; 1820 if (field->is_static()) 1821 t = TypeInstPtr::make(field->holder()->java_mirror()); 1822 else 1823 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1824 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1825 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1826 return atp; 1827 } 1828 1829 1830 //------------------------------have_alias_type-------------------------------- 1831 bool Compile::have_alias_type(const TypePtr* adr_type) { 1832 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1833 if (ace->_adr_type == adr_type) { 1834 return true; 1835 } 1836 1837 // Handle special cases. 1838 if (adr_type == NULL) return true; 1839 if (adr_type == TypePtr::BOTTOM) return true; 1840 1841 return find_alias_type(adr_type, true, NULL) != NULL; 1842 } 1843 1844 //-----------------------------must_alias-------------------------------------- 1845 // True if all values of the given address type are in the given alias category. 1846 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1847 if (alias_idx == AliasIdxBot) return true; // the universal category 1848 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1849 if (alias_idx == AliasIdxTop) return false; // the empty category 1850 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1851 1852 // the only remaining possible overlap is identity 1853 int adr_idx = get_alias_index(adr_type); 1854 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1855 assert(adr_idx == alias_idx || 1856 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1857 && adr_type != TypeOopPtr::BOTTOM), 1858 "should not be testing for overlap with an unsafe pointer"); 1859 return adr_idx == alias_idx; 1860 } 1861 1862 //------------------------------can_alias-------------------------------------- 1863 // True if any values of the given address type are in the given alias category. 1864 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1865 if (alias_idx == AliasIdxTop) return false; // the empty category 1866 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1867 if (alias_idx == AliasIdxBot) return true; // the universal category 1868 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1869 1870 // the only remaining possible overlap is identity 1871 int adr_idx = get_alias_index(adr_type); 1872 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1873 return adr_idx == alias_idx; 1874 } 1875 1876 1877 1878 //---------------------------pop_warm_call------------------------------------- 1879 WarmCallInfo* Compile::pop_warm_call() { 1880 WarmCallInfo* wci = _warm_calls; 1881 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1882 return wci; 1883 } 1884 1885 //----------------------------Inline_Warm-------------------------------------- 1886 int Compile::Inline_Warm() { 1887 // If there is room, try to inline some more warm call sites. 1888 // %%% Do a graph index compaction pass when we think we're out of space? 1889 if (!InlineWarmCalls) return 0; 1890 1891 int calls_made_hot = 0; 1892 int room_to_grow = NodeCountInliningCutoff - unique(); 1893 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1894 int amount_grown = 0; 1895 WarmCallInfo* call; 1896 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1897 int est_size = (int)call->size(); 1898 if (est_size > (room_to_grow - amount_grown)) { 1899 // This one won't fit anyway. Get rid of it. 1900 call->make_cold(); 1901 continue; 1902 } 1903 call->make_hot(); 1904 calls_made_hot++; 1905 amount_grown += est_size; 1906 amount_to_grow -= est_size; 1907 } 1908 1909 if (calls_made_hot > 0) set_major_progress(); 1910 return calls_made_hot; 1911 } 1912 1913 1914 //----------------------------Finish_Warm-------------------------------------- 1915 void Compile::Finish_Warm() { 1916 if (!InlineWarmCalls) return; 1917 if (failing()) return; 1918 if (warm_calls() == NULL) return; 1919 1920 // Clean up loose ends, if we are out of space for inlining. 1921 WarmCallInfo* call; 1922 while ((call = pop_warm_call()) != NULL) { 1923 call->make_cold(); 1924 } 1925 } 1926 1927 //---------------------cleanup_loop_predicates----------------------- 1928 // Remove the opaque nodes that protect the predicates so that all unused 1929 // checks and uncommon_traps will be eliminated from the ideal graph 1930 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1931 if (predicate_count()==0) return; 1932 for (int i = predicate_count(); i > 0; i--) { 1933 Node * n = predicate_opaque1_node(i-1); 1934 assert(n->Opcode() == Op_Opaque1, "must be"); 1935 igvn.replace_node(n, n->in(1)); 1936 } 1937 assert(predicate_count()==0, "should be clean!"); 1938 } 1939 1940 #ifdef _LP64 1941 void Compile::add_range_check_cast(Node* n) { 1942 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1943 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1944 _range_check_casts->append(n); 1945 } 1946 1947 // Remove all range check dependent CastIINodes. 1948 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1949 for (int i = range_check_cast_count(); i > 0; i--) { 1950 Node* cast = range_check_cast_node(i-1); 1951 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1952 igvn.replace_node(cast, cast->in(1)); 1953 } 1954 assert(range_check_cast_count() == 0, "should be empty"); 1955 } 1956 #endif 1957 1958 // StringOpts and late inlining of string methods 1959 void Compile::inline_string_calls(bool parse_time) { 1960 { 1961 // remove useless nodes to make the usage analysis simpler 1962 ResourceMark rm; 1963 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1964 } 1965 1966 { 1967 ResourceMark rm; 1968 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1969 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1970 print_method(PHASE_AFTER_STRINGOPTS, 3); 1971 } 1972 1973 // now inline anything that we skipped the first time around 1974 if (!parse_time) { 1975 _late_inlines_pos = _late_inlines.length(); 1976 } 1977 1978 while (_string_late_inlines.length() > 0) { 1979 CallGenerator* cg = _string_late_inlines.pop(); 1980 cg->do_late_inline(); 1981 if (failing()) return; 1982 } 1983 _string_late_inlines.trunc_to(0); 1984 } 1985 1986 // Late inlining of boxing methods 1987 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1988 if (_boxing_late_inlines.length() > 0) { 1989 assert(has_boxed_value(), "inconsistent"); 1990 1991 PhaseGVN* gvn = initial_gvn(); 1992 set_inlining_incrementally(true); 1993 1994 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1995 for_igvn()->clear(); 1996 gvn->replace_with(&igvn); 1997 1998 _late_inlines_pos = _late_inlines.length(); 1999 2000 while (_boxing_late_inlines.length() > 0) { 2001 CallGenerator* cg = _boxing_late_inlines.pop(); 2002 cg->do_late_inline(); 2003 if (failing()) return; 2004 } 2005 _boxing_late_inlines.trunc_to(0); 2006 2007 { 2008 ResourceMark rm; 2009 PhaseRemoveUseless pru(gvn, for_igvn()); 2010 } 2011 2012 igvn = PhaseIterGVN(gvn); 2013 igvn.optimize(); 2014 2015 set_inlining_progress(false); 2016 set_inlining_incrementally(false); 2017 } 2018 } 2019 2020 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 2021 assert(IncrementalInline, "incremental inlining should be on"); 2022 PhaseGVN* gvn = initial_gvn(); 2023 2024 set_inlining_progress(false); 2025 for_igvn()->clear(); 2026 gvn->replace_with(&igvn); 2027 2028 { 2029 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2030 int i = 0; 2031 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2032 CallGenerator* cg = _late_inlines.at(i); 2033 _late_inlines_pos = i+1; 2034 cg->do_late_inline(); 2035 if (failing()) return; 2036 } 2037 int j = 0; 2038 for (; i < _late_inlines.length(); i++, j++) { 2039 _late_inlines.at_put(j, _late_inlines.at(i)); 2040 } 2041 _late_inlines.trunc_to(j); 2042 } 2043 2044 { 2045 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2046 ResourceMark rm; 2047 PhaseRemoveUseless pru(gvn, for_igvn()); 2048 } 2049 2050 { 2051 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2052 igvn = PhaseIterGVN(gvn); 2053 } 2054 } 2055 2056 // Perform incremental inlining until bound on number of live nodes is reached 2057 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2058 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2059 2060 PhaseGVN* gvn = initial_gvn(); 2061 2062 set_inlining_incrementally(true); 2063 set_inlining_progress(true); 2064 uint low_live_nodes = 0; 2065 2066 while(inlining_progress() && _late_inlines.length() > 0) { 2067 2068 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2069 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2070 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2071 // PhaseIdealLoop is expensive so we only try it once we are 2072 // out of live nodes and we only try it again if the previous 2073 // helped got the number of nodes down significantly 2074 PhaseIdealLoop ideal_loop( igvn, false, true ); 2075 if (failing()) return; 2076 low_live_nodes = live_nodes(); 2077 _major_progress = true; 2078 } 2079 2080 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2081 break; 2082 } 2083 } 2084 2085 inline_incrementally_one(igvn); 2086 2087 if (failing()) return; 2088 2089 { 2090 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2091 igvn.optimize(); 2092 } 2093 2094 if (failing()) return; 2095 } 2096 2097 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2098 2099 if (_string_late_inlines.length() > 0) { 2100 assert(has_stringbuilder(), "inconsistent"); 2101 for_igvn()->clear(); 2102 initial_gvn()->replace_with(&igvn); 2103 2104 inline_string_calls(false); 2105 2106 if (failing()) return; 2107 2108 { 2109 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2110 ResourceMark rm; 2111 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2112 } 2113 2114 { 2115 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2116 igvn = PhaseIterGVN(gvn); 2117 igvn.optimize(); 2118 } 2119 } 2120 2121 set_inlining_incrementally(false); 2122 } 2123 2124 2125 //------------------------------Optimize--------------------------------------- 2126 // Given a graph, optimize it. 2127 void Compile::Optimize() { 2128 TracePhase tp("optimizer", &timers[_t_optimizer]); 2129 2130 #ifndef PRODUCT 2131 if (_directive->BreakAtCompileOption) { 2132 BREAKPOINT; 2133 } 2134 2135 #endif 2136 2137 ResourceMark rm; 2138 int loop_opts_cnt; 2139 2140 print_inlining_reinit(); 2141 2142 NOT_PRODUCT( verify_graph_edges(); ) 2143 2144 print_method(PHASE_AFTER_PARSING); 2145 2146 { 2147 // Iterative Global Value Numbering, including ideal transforms 2148 // Initialize IterGVN with types and values from parse-time GVN 2149 PhaseIterGVN igvn(initial_gvn()); 2150 #ifdef ASSERT 2151 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2152 #endif 2153 { 2154 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2155 igvn.optimize(); 2156 } 2157 2158 print_method(PHASE_ITER_GVN1, 2); 2159 2160 if (failing()) return; 2161 2162 inline_incrementally(igvn); 2163 2164 print_method(PHASE_INCREMENTAL_INLINE, 2); 2165 2166 if (failing()) return; 2167 2168 if (eliminate_boxing()) { 2169 // Inline valueOf() methods now. 2170 inline_boxing_calls(igvn); 2171 2172 if (AlwaysIncrementalInline) { 2173 inline_incrementally(igvn); 2174 } 2175 2176 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2177 2178 if (failing()) return; 2179 } 2180 2181 // Remove the speculative part of types and clean up the graph from 2182 // the extra CastPP nodes whose only purpose is to carry them. Do 2183 // that early so that optimizations are not disrupted by the extra 2184 // CastPP nodes. 2185 remove_speculative_types(igvn); 2186 2187 // No more new expensive nodes will be added to the list from here 2188 // so keep only the actual candidates for optimizations. 2189 cleanup_expensive_nodes(igvn); 2190 2191 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2192 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2193 initial_gvn()->replace_with(&igvn); 2194 for_igvn()->clear(); 2195 Unique_Node_List new_worklist(C->comp_arena()); 2196 { 2197 ResourceMark rm; 2198 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2199 } 2200 set_for_igvn(&new_worklist); 2201 igvn = PhaseIterGVN(initial_gvn()); 2202 igvn.optimize(); 2203 } 2204 2205 // Perform escape analysis 2206 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2207 if (has_loops()) { 2208 // Cleanup graph (remove dead nodes). 2209 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2210 PhaseIdealLoop ideal_loop( igvn, false, true ); 2211 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2212 if (failing()) return; 2213 } 2214 ConnectionGraph::do_analysis(this, &igvn); 2215 2216 if (failing()) return; 2217 2218 // Optimize out fields loads from scalar replaceable allocations. 2219 igvn.optimize(); 2220 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2221 2222 if (failing()) return; 2223 2224 if (congraph() != NULL && macro_count() > 0) { 2225 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2226 PhaseMacroExpand mexp(igvn); 2227 mexp.eliminate_macro_nodes(); 2228 igvn.set_delay_transform(false); 2229 2230 igvn.optimize(); 2231 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2232 2233 if (failing()) return; 2234 } 2235 } 2236 2237 // Loop transforms on the ideal graph. Range Check Elimination, 2238 // peeling, unrolling, etc. 2239 2240 // Set loop opts counter 2241 loop_opts_cnt = num_loop_opts(); 2242 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2243 { 2244 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2245 PhaseIdealLoop ideal_loop( igvn, true ); 2246 loop_opts_cnt--; 2247 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2248 if (failing()) return; 2249 } 2250 // Loop opts pass if partial peeling occurred in previous pass 2251 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2252 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2253 PhaseIdealLoop ideal_loop( igvn, false ); 2254 loop_opts_cnt--; 2255 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2256 if (failing()) return; 2257 } 2258 // Loop opts pass for loop-unrolling before CCP 2259 if(major_progress() && (loop_opts_cnt > 0)) { 2260 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2261 PhaseIdealLoop ideal_loop( igvn, false ); 2262 loop_opts_cnt--; 2263 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2264 } 2265 if (!failing()) { 2266 // Verify that last round of loop opts produced a valid graph 2267 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2268 PhaseIdealLoop::verify(igvn); 2269 } 2270 } 2271 if (failing()) return; 2272 2273 // Conditional Constant Propagation; 2274 PhaseCCP ccp( &igvn ); 2275 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2276 { 2277 TracePhase tp("ccp", &timers[_t_ccp]); 2278 ccp.do_transform(); 2279 } 2280 print_method(PHASE_CPP1, 2); 2281 2282 assert( true, "Break here to ccp.dump_old2new_map()"); 2283 2284 // Iterative Global Value Numbering, including ideal transforms 2285 { 2286 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2287 igvn = ccp; 2288 igvn.optimize(); 2289 } 2290 2291 print_method(PHASE_ITER_GVN2, 2); 2292 2293 if (failing()) return; 2294 2295 // Loop transforms on the ideal graph. Range Check Elimination, 2296 // peeling, unrolling, etc. 2297 if(loop_opts_cnt > 0) { 2298 debug_only( int cnt = 0; ); 2299 while(major_progress() && (loop_opts_cnt > 0)) { 2300 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2301 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2302 PhaseIdealLoop ideal_loop( igvn, true); 2303 loop_opts_cnt--; 2304 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2305 if (failing()) return; 2306 } 2307 } 2308 // Ensure that major progress is now clear 2309 C->clear_major_progress(); 2310 2311 { 2312 // Verify that all previous optimizations produced a valid graph 2313 // at least to this point, even if no loop optimizations were done. 2314 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2315 PhaseIdealLoop::verify(igvn); 2316 } 2317 2318 #ifdef _LP64 2319 if (range_check_cast_count() > 0) { 2320 // No more loop optimizations. Remove all range check dependent CastIINodes. 2321 C->remove_range_check_casts(igvn); 2322 igvn.optimize(); 2323 } 2324 #endif 2325 2326 { 2327 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2328 PhaseMacroExpand mex(igvn); 2329 if (mex.expand_macro_nodes()) { 2330 assert(failing(), "must bail out w/ explicit message"); 2331 return; 2332 } 2333 } 2334 2335 DEBUG_ONLY( _modified_nodes = NULL; ) 2336 } // (End scope of igvn; run destructor if necessary for asserts.) 2337 2338 process_print_inlining(); 2339 // A method with only infinite loops has no edges entering loops from root 2340 { 2341 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2342 if (final_graph_reshaping()) { 2343 assert(failing(), "must bail out w/ explicit message"); 2344 return; 2345 } 2346 } 2347 2348 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2349 } 2350 2351 2352 //------------------------------Code_Gen--------------------------------------- 2353 // Given a graph, generate code for it 2354 void Compile::Code_Gen() { 2355 if (failing()) { 2356 return; 2357 } 2358 2359 // Perform instruction selection. You might think we could reclaim Matcher 2360 // memory PDQ, but actually the Matcher is used in generating spill code. 2361 // Internals of the Matcher (including some VectorSets) must remain live 2362 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2363 // set a bit in reclaimed memory. 2364 2365 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2366 // nodes. Mapping is only valid at the root of each matched subtree. 2367 NOT_PRODUCT( verify_graph_edges(); ) 2368 2369 Matcher matcher; 2370 _matcher = &matcher; 2371 { 2372 TracePhase tp("matcher", &timers[_t_matcher]); 2373 matcher.match(); 2374 } 2375 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2376 // nodes. Mapping is only valid at the root of each matched subtree. 2377 NOT_PRODUCT( verify_graph_edges(); ) 2378 2379 // If you have too many nodes, or if matching has failed, bail out 2380 check_node_count(0, "out of nodes matching instructions"); 2381 if (failing()) { 2382 return; 2383 } 2384 2385 // Build a proper-looking CFG 2386 PhaseCFG cfg(node_arena(), root(), matcher); 2387 _cfg = &cfg; 2388 { 2389 TracePhase tp("scheduler", &timers[_t_scheduler]); 2390 bool success = cfg.do_global_code_motion(); 2391 if (!success) { 2392 return; 2393 } 2394 2395 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2396 NOT_PRODUCT( verify_graph_edges(); ) 2397 debug_only( cfg.verify(); ) 2398 } 2399 2400 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2401 _regalloc = ®alloc; 2402 { 2403 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2404 // Perform register allocation. After Chaitin, use-def chains are 2405 // no longer accurate (at spill code) and so must be ignored. 2406 // Node->LRG->reg mappings are still accurate. 2407 _regalloc->Register_Allocate(); 2408 2409 // Bail out if the allocator builds too many nodes 2410 if (failing()) { 2411 return; 2412 } 2413 } 2414 2415 // Prior to register allocation we kept empty basic blocks in case the 2416 // the allocator needed a place to spill. After register allocation we 2417 // are not adding any new instructions. If any basic block is empty, we 2418 // can now safely remove it. 2419 { 2420 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2421 cfg.remove_empty_blocks(); 2422 if (do_freq_based_layout()) { 2423 PhaseBlockLayout layout(cfg); 2424 } else { 2425 cfg.set_loop_alignment(); 2426 } 2427 cfg.fixup_flow(); 2428 } 2429 2430 // Apply peephole optimizations 2431 if( OptoPeephole ) { 2432 TracePhase tp("peephole", &timers[_t_peephole]); 2433 PhasePeephole peep( _regalloc, cfg); 2434 peep.do_transform(); 2435 } 2436 2437 // Do late expand if CPU requires this. 2438 if (Matcher::require_postalloc_expand) { 2439 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2440 cfg.postalloc_expand(_regalloc); 2441 } 2442 2443 // Convert Nodes to instruction bits in a buffer 2444 { 2445 TraceTime tp("output", &timers[_t_output], CITime); 2446 Output(); 2447 } 2448 2449 print_method(PHASE_FINAL_CODE); 2450 2451 // He's dead, Jim. 2452 _cfg = (PhaseCFG*)0xdeadbeef; 2453 _regalloc = (PhaseChaitin*)0xdeadbeef; 2454 } 2455 2456 2457 //------------------------------dump_asm--------------------------------------- 2458 // Dump formatted assembly 2459 #ifndef PRODUCT 2460 void Compile::dump_asm(int *pcs, uint pc_limit) { 2461 bool cut_short = false; 2462 tty->print_cr("#"); 2463 tty->print("# "); _tf->dump(); tty->cr(); 2464 tty->print_cr("#"); 2465 2466 // For all blocks 2467 int pc = 0x0; // Program counter 2468 char starts_bundle = ' '; 2469 _regalloc->dump_frame(); 2470 2471 Node *n = NULL; 2472 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2473 if (VMThread::should_terminate()) { 2474 cut_short = true; 2475 break; 2476 } 2477 Block* block = _cfg->get_block(i); 2478 if (block->is_connector() && !Verbose) { 2479 continue; 2480 } 2481 n = block->head(); 2482 if (pcs && n->_idx < pc_limit) { 2483 tty->print("%3.3x ", pcs[n->_idx]); 2484 } else { 2485 tty->print(" "); 2486 } 2487 block->dump_head(_cfg); 2488 if (block->is_connector()) { 2489 tty->print_cr(" # Empty connector block"); 2490 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2491 tty->print_cr(" # Block is sole successor of call"); 2492 } 2493 2494 // For all instructions 2495 Node *delay = NULL; 2496 for (uint j = 0; j < block->number_of_nodes(); j++) { 2497 if (VMThread::should_terminate()) { 2498 cut_short = true; 2499 break; 2500 } 2501 n = block->get_node(j); 2502 if (valid_bundle_info(n)) { 2503 Bundle* bundle = node_bundling(n); 2504 if (bundle->used_in_unconditional_delay()) { 2505 delay = n; 2506 continue; 2507 } 2508 if (bundle->starts_bundle()) { 2509 starts_bundle = '+'; 2510 } 2511 } 2512 2513 if (WizardMode) { 2514 n->dump(); 2515 } 2516 2517 if( !n->is_Region() && // Dont print in the Assembly 2518 !n->is_Phi() && // a few noisely useless nodes 2519 !n->is_Proj() && 2520 !n->is_MachTemp() && 2521 !n->is_SafePointScalarObject() && 2522 !n->is_Catch() && // Would be nice to print exception table targets 2523 !n->is_MergeMem() && // Not very interesting 2524 !n->is_top() && // Debug info table constants 2525 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2526 ) { 2527 if (pcs && n->_idx < pc_limit) 2528 tty->print("%3.3x", pcs[n->_idx]); 2529 else 2530 tty->print(" "); 2531 tty->print(" %c ", starts_bundle); 2532 starts_bundle = ' '; 2533 tty->print("\t"); 2534 n->format(_regalloc, tty); 2535 tty->cr(); 2536 } 2537 2538 // If we have an instruction with a delay slot, and have seen a delay, 2539 // then back up and print it 2540 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2541 assert(delay != NULL, "no unconditional delay instruction"); 2542 if (WizardMode) delay->dump(); 2543 2544 if (node_bundling(delay)->starts_bundle()) 2545 starts_bundle = '+'; 2546 if (pcs && n->_idx < pc_limit) 2547 tty->print("%3.3x", pcs[n->_idx]); 2548 else 2549 tty->print(" "); 2550 tty->print(" %c ", starts_bundle); 2551 starts_bundle = ' '; 2552 tty->print("\t"); 2553 delay->format(_regalloc, tty); 2554 tty->cr(); 2555 delay = NULL; 2556 } 2557 2558 // Dump the exception table as well 2559 if( n->is_Catch() && (Verbose || WizardMode) ) { 2560 // Print the exception table for this offset 2561 _handler_table.print_subtable_for(pc); 2562 } 2563 } 2564 2565 if (pcs && n->_idx < pc_limit) 2566 tty->print_cr("%3.3x", pcs[n->_idx]); 2567 else 2568 tty->cr(); 2569 2570 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2571 2572 } // End of per-block dump 2573 tty->cr(); 2574 2575 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2576 } 2577 #endif 2578 2579 //------------------------------Final_Reshape_Counts--------------------------- 2580 // This class defines counters to help identify when a method 2581 // may/must be executed using hardware with only 24-bit precision. 2582 struct Final_Reshape_Counts : public StackObj { 2583 int _call_count; // count non-inlined 'common' calls 2584 int _float_count; // count float ops requiring 24-bit precision 2585 int _double_count; // count double ops requiring more precision 2586 int _java_call_count; // count non-inlined 'java' calls 2587 int _inner_loop_count; // count loops which need alignment 2588 VectorSet _visited; // Visitation flags 2589 Node_List _tests; // Set of IfNodes & PCTableNodes 2590 2591 Final_Reshape_Counts() : 2592 _call_count(0), _float_count(0), _double_count(0), 2593 _java_call_count(0), _inner_loop_count(0), 2594 _visited( Thread::current()->resource_area() ) { } 2595 2596 void inc_call_count () { _call_count ++; } 2597 void inc_float_count () { _float_count ++; } 2598 void inc_double_count() { _double_count++; } 2599 void inc_java_call_count() { _java_call_count++; } 2600 void inc_inner_loop_count() { _inner_loop_count++; } 2601 2602 int get_call_count () const { return _call_count ; } 2603 int get_float_count () const { return _float_count ; } 2604 int get_double_count() const { return _double_count; } 2605 int get_java_call_count() const { return _java_call_count; } 2606 int get_inner_loop_count() const { return _inner_loop_count; } 2607 }; 2608 2609 #ifdef ASSERT 2610 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2611 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2612 // Make sure the offset goes inside the instance layout. 2613 return k->contains_field_offset(tp->offset()); 2614 // Note that OffsetBot and OffsetTop are very negative. 2615 } 2616 #endif 2617 2618 // Eliminate trivially redundant StoreCMs and accumulate their 2619 // precedence edges. 2620 void Compile::eliminate_redundant_card_marks(Node* n) { 2621 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2622 if (n->in(MemNode::Address)->outcnt() > 1) { 2623 // There are multiple users of the same address so it might be 2624 // possible to eliminate some of the StoreCMs 2625 Node* mem = n->in(MemNode::Memory); 2626 Node* adr = n->in(MemNode::Address); 2627 Node* val = n->in(MemNode::ValueIn); 2628 Node* prev = n; 2629 bool done = false; 2630 // Walk the chain of StoreCMs eliminating ones that match. As 2631 // long as it's a chain of single users then the optimization is 2632 // safe. Eliminating partially redundant StoreCMs would require 2633 // cloning copies down the other paths. 2634 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2635 if (adr == mem->in(MemNode::Address) && 2636 val == mem->in(MemNode::ValueIn)) { 2637 // redundant StoreCM 2638 if (mem->req() > MemNode::OopStore) { 2639 // Hasn't been processed by this code yet. 2640 n->add_prec(mem->in(MemNode::OopStore)); 2641 } else { 2642 // Already converted to precedence edge 2643 for (uint i = mem->req(); i < mem->len(); i++) { 2644 // Accumulate any precedence edges 2645 if (mem->in(i) != NULL) { 2646 n->add_prec(mem->in(i)); 2647 } 2648 } 2649 // Everything above this point has been processed. 2650 done = true; 2651 } 2652 // Eliminate the previous StoreCM 2653 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2654 assert(mem->outcnt() == 0, "should be dead"); 2655 mem->disconnect_inputs(NULL, this); 2656 } else { 2657 prev = mem; 2658 } 2659 mem = prev->in(MemNode::Memory); 2660 } 2661 } 2662 } 2663 2664 //------------------------------final_graph_reshaping_impl---------------------- 2665 // Implement items 1-5 from final_graph_reshaping below. 2666 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2667 2668 if ( n->outcnt() == 0 ) return; // dead node 2669 uint nop = n->Opcode(); 2670 2671 // Check for 2-input instruction with "last use" on right input. 2672 // Swap to left input. Implements item (2). 2673 if( n->req() == 3 && // two-input instruction 2674 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2675 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2676 n->in(2)->outcnt() == 1 &&// right use IS a last use 2677 !n->in(2)->is_Con() ) { // right use is not a constant 2678 // Check for commutative opcode 2679 switch( nop ) { 2680 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2681 case Op_MaxI: case Op_MinI: 2682 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2683 case Op_AndL: case Op_XorL: case Op_OrL: 2684 case Op_AndI: case Op_XorI: case Op_OrI: { 2685 // Move "last use" input to left by swapping inputs 2686 n->swap_edges(1, 2); 2687 break; 2688 } 2689 default: 2690 break; 2691 } 2692 } 2693 2694 #ifdef ASSERT 2695 if( n->is_Mem() ) { 2696 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2697 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2698 // oop will be recorded in oop map if load crosses safepoint 2699 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2700 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2701 "raw memory operations should have control edge"); 2702 } 2703 #endif 2704 // Count FPU ops and common calls, implements item (3) 2705 switch( nop ) { 2706 // Count all float operations that may use FPU 2707 case Op_AddF: 2708 case Op_SubF: 2709 case Op_MulF: 2710 case Op_DivF: 2711 case Op_NegF: 2712 case Op_ModF: 2713 case Op_ConvI2F: 2714 case Op_ConF: 2715 case Op_CmpF: 2716 case Op_CmpF3: 2717 // case Op_ConvL2F: // longs are split into 32-bit halves 2718 frc.inc_float_count(); 2719 break; 2720 2721 case Op_ConvF2D: 2722 case Op_ConvD2F: 2723 frc.inc_float_count(); 2724 frc.inc_double_count(); 2725 break; 2726 2727 // Count all double operations that may use FPU 2728 case Op_AddD: 2729 case Op_SubD: 2730 case Op_MulD: 2731 case Op_DivD: 2732 case Op_NegD: 2733 case Op_ModD: 2734 case Op_ConvI2D: 2735 case Op_ConvD2I: 2736 // case Op_ConvL2D: // handled by leaf call 2737 // case Op_ConvD2L: // handled by leaf call 2738 case Op_ConD: 2739 case Op_CmpD: 2740 case Op_CmpD3: 2741 frc.inc_double_count(); 2742 break; 2743 case Op_Opaque1: // Remove Opaque Nodes before matching 2744 case Op_Opaque2: // Remove Opaque Nodes before matching 2745 case Op_Opaque3: 2746 n->subsume_by(n->in(1), this); 2747 break; 2748 case Op_CallStaticJava: 2749 case Op_CallJava: 2750 case Op_CallDynamicJava: 2751 frc.inc_java_call_count(); // Count java call site; 2752 case Op_CallRuntime: 2753 case Op_CallLeaf: 2754 case Op_CallLeafNoFP: { 2755 assert( n->is_Call(), "" ); 2756 CallNode *call = n->as_Call(); 2757 // Count call sites where the FP mode bit would have to be flipped. 2758 // Do not count uncommon runtime calls: 2759 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2760 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2761 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2762 frc.inc_call_count(); // Count the call site 2763 } else { // See if uncommon argument is shared 2764 Node *n = call->in(TypeFunc::Parms); 2765 int nop = n->Opcode(); 2766 // Clone shared simple arguments to uncommon calls, item (1). 2767 if( n->outcnt() > 1 && 2768 !n->is_Proj() && 2769 nop != Op_CreateEx && 2770 nop != Op_CheckCastPP && 2771 nop != Op_DecodeN && 2772 nop != Op_DecodeNKlass && 2773 !n->is_Mem() ) { 2774 Node *x = n->clone(); 2775 call->set_req( TypeFunc::Parms, x ); 2776 } 2777 } 2778 break; 2779 } 2780 2781 case Op_StoreD: 2782 case Op_LoadD: 2783 case Op_LoadD_unaligned: 2784 frc.inc_double_count(); 2785 goto handle_mem; 2786 case Op_StoreF: 2787 case Op_LoadF: 2788 frc.inc_float_count(); 2789 goto handle_mem; 2790 2791 case Op_StoreCM: 2792 { 2793 // Convert OopStore dependence into precedence edge 2794 Node* prec = n->in(MemNode::OopStore); 2795 n->del_req(MemNode::OopStore); 2796 n->add_prec(prec); 2797 eliminate_redundant_card_marks(n); 2798 } 2799 2800 // fall through 2801 2802 case Op_StoreB: 2803 case Op_StoreC: 2804 case Op_StorePConditional: 2805 case Op_StoreI: 2806 case Op_StoreL: 2807 case Op_StoreIConditional: 2808 case Op_StoreLConditional: 2809 case Op_CompareAndSwapI: 2810 case Op_CompareAndSwapL: 2811 case Op_CompareAndSwapP: 2812 case Op_CompareAndSwapN: 2813 case Op_GetAndAddI: 2814 case Op_GetAndAddL: 2815 case Op_GetAndSetI: 2816 case Op_GetAndSetL: 2817 case Op_GetAndSetP: 2818 case Op_GetAndSetN: 2819 case Op_StoreP: 2820 case Op_StoreN: 2821 case Op_StoreNKlass: 2822 case Op_LoadB: 2823 case Op_LoadUB: 2824 case Op_LoadUS: 2825 case Op_LoadI: 2826 case Op_LoadKlass: 2827 case Op_LoadNKlass: 2828 case Op_LoadL: 2829 case Op_LoadL_unaligned: 2830 case Op_LoadPLocked: 2831 case Op_LoadP: 2832 case Op_LoadN: 2833 case Op_LoadRange: 2834 case Op_LoadS: { 2835 handle_mem: 2836 #ifdef ASSERT 2837 if( VerifyOptoOopOffsets ) { 2838 assert( n->is_Mem(), "" ); 2839 MemNode *mem = (MemNode*)n; 2840 // Check to see if address types have grounded out somehow. 2841 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2842 assert( !tp || oop_offset_is_sane(tp), "" ); 2843 } 2844 #endif 2845 break; 2846 } 2847 2848 case Op_AddP: { // Assert sane base pointers 2849 Node *addp = n->in(AddPNode::Address); 2850 assert( !addp->is_AddP() || 2851 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2852 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2853 "Base pointers must match" ); 2854 #ifdef _LP64 2855 if ((UseCompressedOops || UseCompressedClassPointers) && 2856 addp->Opcode() == Op_ConP && 2857 addp == n->in(AddPNode::Base) && 2858 n->in(AddPNode::Offset)->is_Con()) { 2859 // Use addressing with narrow klass to load with offset on x86. 2860 // On sparc loading 32-bits constant and decoding it have less 2861 // instructions (4) then load 64-bits constant (7). 2862 // Do this transformation here since IGVN will convert ConN back to ConP. 2863 const Type* t = addp->bottom_type(); 2864 if (t->isa_oopptr() || t->isa_klassptr()) { 2865 Node* nn = NULL; 2866 2867 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2868 2869 // Look for existing ConN node of the same exact type. 2870 Node* r = root(); 2871 uint cnt = r->outcnt(); 2872 for (uint i = 0; i < cnt; i++) { 2873 Node* m = r->raw_out(i); 2874 if (m!= NULL && m->Opcode() == op && 2875 m->bottom_type()->make_ptr() == t) { 2876 nn = m; 2877 break; 2878 } 2879 } 2880 if (nn != NULL) { 2881 // Decode a narrow oop to match address 2882 // [R12 + narrow_oop_reg<<3 + offset] 2883 if (t->isa_oopptr()) { 2884 nn = new DecodeNNode(nn, t); 2885 } else { 2886 nn = new DecodeNKlassNode(nn, t); 2887 } 2888 n->set_req(AddPNode::Base, nn); 2889 n->set_req(AddPNode::Address, nn); 2890 if (addp->outcnt() == 0) { 2891 addp->disconnect_inputs(NULL, this); 2892 } 2893 } 2894 } 2895 } 2896 #endif 2897 break; 2898 } 2899 2900 case Op_CastPP: { 2901 // Remove CastPP nodes to gain more freedom during scheduling but 2902 // keep the dependency they encode as control or precedence edges 2903 // (if control is set already) on memory operations. Some CastPP 2904 // nodes don't have a control (don't carry a dependency): skip 2905 // those. 2906 if (n->in(0) != NULL) { 2907 ResourceMark rm; 2908 Unique_Node_List wq; 2909 wq.push(n); 2910 for (uint next = 0; next < wq.size(); ++next) { 2911 Node *m = wq.at(next); 2912 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 2913 Node* use = m->fast_out(i); 2914 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 2915 use->ensure_control_or_add_prec(n->in(0)); 2916 } else { 2917 switch(use->Opcode()) { 2918 case Op_AddP: 2919 case Op_DecodeN: 2920 case Op_DecodeNKlass: 2921 case Op_CheckCastPP: 2922 case Op_CastPP: 2923 wq.push(use); 2924 break; 2925 } 2926 } 2927 } 2928 } 2929 } 2930 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 2931 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2932 Node* in1 = n->in(1); 2933 const Type* t = n->bottom_type(); 2934 Node* new_in1 = in1->clone(); 2935 new_in1->as_DecodeN()->set_type(t); 2936 2937 if (!Matcher::narrow_oop_use_complex_address()) { 2938 // 2939 // x86, ARM and friends can handle 2 adds in addressing mode 2940 // and Matcher can fold a DecodeN node into address by using 2941 // a narrow oop directly and do implicit NULL check in address: 2942 // 2943 // [R12 + narrow_oop_reg<<3 + offset] 2944 // NullCheck narrow_oop_reg 2945 // 2946 // On other platforms (Sparc) we have to keep new DecodeN node and 2947 // use it to do implicit NULL check in address: 2948 // 2949 // decode_not_null narrow_oop_reg, base_reg 2950 // [base_reg + offset] 2951 // NullCheck base_reg 2952 // 2953 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2954 // to keep the information to which NULL check the new DecodeN node 2955 // corresponds to use it as value in implicit_null_check(). 2956 // 2957 new_in1->set_req(0, n->in(0)); 2958 } 2959 2960 n->subsume_by(new_in1, this); 2961 if (in1->outcnt() == 0) { 2962 in1->disconnect_inputs(NULL, this); 2963 } 2964 } else { 2965 n->subsume_by(n->in(1), this); 2966 if (n->outcnt() == 0) { 2967 n->disconnect_inputs(NULL, this); 2968 } 2969 } 2970 break; 2971 } 2972 #ifdef _LP64 2973 case Op_CmpP: 2974 // Do this transformation here to preserve CmpPNode::sub() and 2975 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2976 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2977 Node* in1 = n->in(1); 2978 Node* in2 = n->in(2); 2979 if (!in1->is_DecodeNarrowPtr()) { 2980 in2 = in1; 2981 in1 = n->in(2); 2982 } 2983 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2984 2985 Node* new_in2 = NULL; 2986 if (in2->is_DecodeNarrowPtr()) { 2987 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2988 new_in2 = in2->in(1); 2989 } else if (in2->Opcode() == Op_ConP) { 2990 const Type* t = in2->bottom_type(); 2991 if (t == TypePtr::NULL_PTR) { 2992 assert(in1->is_DecodeN(), "compare klass to null?"); 2993 // Don't convert CmpP null check into CmpN if compressed 2994 // oops implicit null check is not generated. 2995 // This will allow to generate normal oop implicit null check. 2996 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2997 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 2998 // 2999 // This transformation together with CastPP transformation above 3000 // will generated code for implicit NULL checks for compressed oops. 3001 // 3002 // The original code after Optimize() 3003 // 3004 // LoadN memory, narrow_oop_reg 3005 // decode narrow_oop_reg, base_reg 3006 // CmpP base_reg, NULL 3007 // CastPP base_reg // NotNull 3008 // Load [base_reg + offset], val_reg 3009 // 3010 // after these transformations will be 3011 // 3012 // LoadN memory, narrow_oop_reg 3013 // CmpN narrow_oop_reg, NULL 3014 // decode_not_null narrow_oop_reg, base_reg 3015 // Load [base_reg + offset], val_reg 3016 // 3017 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3018 // since narrow oops can be used in debug info now (see the code in 3019 // final_graph_reshaping_walk()). 3020 // 3021 // At the end the code will be matched to 3022 // on x86: 3023 // 3024 // Load_narrow_oop memory, narrow_oop_reg 3025 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3026 // NullCheck narrow_oop_reg 3027 // 3028 // and on sparc: 3029 // 3030 // Load_narrow_oop memory, narrow_oop_reg 3031 // decode_not_null narrow_oop_reg, base_reg 3032 // Load [base_reg + offset], val_reg 3033 // NullCheck base_reg 3034 // 3035 } else if (t->isa_oopptr()) { 3036 new_in2 = ConNode::make(t->make_narrowoop()); 3037 } else if (t->isa_klassptr()) { 3038 new_in2 = ConNode::make(t->make_narrowklass()); 3039 } 3040 } 3041 if (new_in2 != NULL) { 3042 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3043 n->subsume_by(cmpN, this); 3044 if (in1->outcnt() == 0) { 3045 in1->disconnect_inputs(NULL, this); 3046 } 3047 if (in2->outcnt() == 0) { 3048 in2->disconnect_inputs(NULL, this); 3049 } 3050 } 3051 } 3052 break; 3053 3054 case Op_DecodeN: 3055 case Op_DecodeNKlass: 3056 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3057 // DecodeN could be pinned when it can't be fold into 3058 // an address expression, see the code for Op_CastPP above. 3059 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3060 break; 3061 3062 case Op_EncodeP: 3063 case Op_EncodePKlass: { 3064 Node* in1 = n->in(1); 3065 if (in1->is_DecodeNarrowPtr()) { 3066 n->subsume_by(in1->in(1), this); 3067 } else if (in1->Opcode() == Op_ConP) { 3068 const Type* t = in1->bottom_type(); 3069 if (t == TypePtr::NULL_PTR) { 3070 assert(t->isa_oopptr(), "null klass?"); 3071 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3072 } else if (t->isa_oopptr()) { 3073 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3074 } else if (t->isa_klassptr()) { 3075 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3076 } 3077 } 3078 if (in1->outcnt() == 0) { 3079 in1->disconnect_inputs(NULL, this); 3080 } 3081 break; 3082 } 3083 3084 case Op_Proj: { 3085 if (OptimizeStringConcat) { 3086 ProjNode* p = n->as_Proj(); 3087 if (p->_is_io_use) { 3088 // Separate projections were used for the exception path which 3089 // are normally removed by a late inline. If it wasn't inlined 3090 // then they will hang around and should just be replaced with 3091 // the original one. 3092 Node* proj = NULL; 3093 // Replace with just one 3094 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3095 Node *use = i.get(); 3096 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3097 proj = use; 3098 break; 3099 } 3100 } 3101 assert(proj != NULL, "must be found"); 3102 p->subsume_by(proj, this); 3103 } 3104 } 3105 break; 3106 } 3107 3108 case Op_Phi: 3109 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3110 // The EncodeP optimization may create Phi with the same edges 3111 // for all paths. It is not handled well by Register Allocator. 3112 Node* unique_in = n->in(1); 3113 assert(unique_in != NULL, ""); 3114 uint cnt = n->req(); 3115 for (uint i = 2; i < cnt; i++) { 3116 Node* m = n->in(i); 3117 assert(m != NULL, ""); 3118 if (unique_in != m) 3119 unique_in = NULL; 3120 } 3121 if (unique_in != NULL) { 3122 n->subsume_by(unique_in, this); 3123 } 3124 } 3125 break; 3126 3127 #ifdef ASSERT 3128 case Op_CastII: 3129 // Verify that all range check dependent CastII nodes were removed. 3130 if (n->isa_CastII()->has_range_check()) { 3131 n->dump(3); 3132 assert(false, "Range check dependent CastII node was not removed"); 3133 } 3134 break; 3135 #endif 3136 3137 #endif // _LP64 3138 3139 case Op_ModI: 3140 if (UseDivMod) { 3141 // Check if a%b and a/b both exist 3142 Node* d = n->find_similar(Op_DivI); 3143 if (d) { 3144 // Replace them with a fused divmod if supported 3145 if (Matcher::has_match_rule(Op_DivModI)) { 3146 DivModINode* divmod = DivModINode::make(n); 3147 d->subsume_by(divmod->div_proj(), this); 3148 n->subsume_by(divmod->mod_proj(), this); 3149 } else { 3150 // replace a%b with a-((a/b)*b) 3151 Node* mult = new MulINode(d, d->in(2)); 3152 Node* sub = new SubINode(d->in(1), mult); 3153 n->subsume_by(sub, this); 3154 } 3155 } 3156 } 3157 break; 3158 3159 case Op_ModL: 3160 if (UseDivMod) { 3161 // Check if a%b and a/b both exist 3162 Node* d = n->find_similar(Op_DivL); 3163 if (d) { 3164 // Replace them with a fused divmod if supported 3165 if (Matcher::has_match_rule(Op_DivModL)) { 3166 DivModLNode* divmod = DivModLNode::make(n); 3167 d->subsume_by(divmod->div_proj(), this); 3168 n->subsume_by(divmod->mod_proj(), this); 3169 } else { 3170 // replace a%b with a-((a/b)*b) 3171 Node* mult = new MulLNode(d, d->in(2)); 3172 Node* sub = new SubLNode(d->in(1), mult); 3173 n->subsume_by(sub, this); 3174 } 3175 } 3176 } 3177 break; 3178 3179 case Op_LoadVector: 3180 case Op_StoreVector: 3181 break; 3182 3183 case Op_AddReductionVI: 3184 case Op_AddReductionVL: 3185 case Op_AddReductionVF: 3186 case Op_AddReductionVD: 3187 case Op_MulReductionVI: 3188 case Op_MulReductionVL: 3189 case Op_MulReductionVF: 3190 case Op_MulReductionVD: 3191 break; 3192 3193 case Op_PackB: 3194 case Op_PackS: 3195 case Op_PackI: 3196 case Op_PackF: 3197 case Op_PackL: 3198 case Op_PackD: 3199 if (n->req()-1 > 2) { 3200 // Replace many operand PackNodes with a binary tree for matching 3201 PackNode* p = (PackNode*) n; 3202 Node* btp = p->binary_tree_pack(1, n->req()); 3203 n->subsume_by(btp, this); 3204 } 3205 break; 3206 case Op_Loop: 3207 case Op_CountedLoop: 3208 if (n->as_Loop()->is_inner_loop()) { 3209 frc.inc_inner_loop_count(); 3210 } 3211 break; 3212 case Op_LShiftI: 3213 case Op_RShiftI: 3214 case Op_URShiftI: 3215 case Op_LShiftL: 3216 case Op_RShiftL: 3217 case Op_URShiftL: 3218 if (Matcher::need_masked_shift_count) { 3219 // The cpu's shift instructions don't restrict the count to the 3220 // lower 5/6 bits. We need to do the masking ourselves. 3221 Node* in2 = n->in(2); 3222 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3223 const TypeInt* t = in2->find_int_type(); 3224 if (t != NULL && t->is_con()) { 3225 juint shift = t->get_con(); 3226 if (shift > mask) { // Unsigned cmp 3227 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3228 } 3229 } else { 3230 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3231 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3232 n->set_req(2, shift); 3233 } 3234 } 3235 if (in2->outcnt() == 0) { // Remove dead node 3236 in2->disconnect_inputs(NULL, this); 3237 } 3238 } 3239 break; 3240 case Op_MemBarStoreStore: 3241 case Op_MemBarRelease: 3242 // Break the link with AllocateNode: it is no longer useful and 3243 // confuses register allocation. 3244 if (n->req() > MemBarNode::Precedent) { 3245 n->set_req(MemBarNode::Precedent, top()); 3246 } 3247 break; 3248 case Op_RangeCheck: { 3249 RangeCheckNode* rc = n->as_RangeCheck(); 3250 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3251 n->subsume_by(iff, this); 3252 frc._tests.push(iff); 3253 break; 3254 } 3255 default: 3256 assert( !n->is_Call(), "" ); 3257 assert( !n->is_Mem(), "" ); 3258 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3259 break; 3260 } 3261 3262 // Collect CFG split points 3263 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3264 frc._tests.push(n); 3265 } 3266 } 3267 3268 //------------------------------final_graph_reshaping_walk--------------------- 3269 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3270 // requires that the walk visits a node's inputs before visiting the node. 3271 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3272 ResourceArea *area = Thread::current()->resource_area(); 3273 Unique_Node_List sfpt(area); 3274 3275 frc._visited.set(root->_idx); // first, mark node as visited 3276 uint cnt = root->req(); 3277 Node *n = root; 3278 uint i = 0; 3279 while (true) { 3280 if (i < cnt) { 3281 // Place all non-visited non-null inputs onto stack 3282 Node* m = n->in(i); 3283 ++i; 3284 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3285 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3286 // compute worst case interpreter size in case of a deoptimization 3287 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3288 3289 sfpt.push(m); 3290 } 3291 cnt = m->req(); 3292 nstack.push(n, i); // put on stack parent and next input's index 3293 n = m; 3294 i = 0; 3295 } 3296 } else { 3297 // Now do post-visit work 3298 final_graph_reshaping_impl( n, frc ); 3299 if (nstack.is_empty()) 3300 break; // finished 3301 n = nstack.node(); // Get node from stack 3302 cnt = n->req(); 3303 i = nstack.index(); 3304 nstack.pop(); // Shift to the next node on stack 3305 } 3306 } 3307 3308 // Skip next transformation if compressed oops are not used. 3309 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3310 (!UseCompressedOops && !UseCompressedClassPointers)) 3311 return; 3312 3313 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3314 // It could be done for an uncommon traps or any safepoints/calls 3315 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3316 while (sfpt.size() > 0) { 3317 n = sfpt.pop(); 3318 JVMState *jvms = n->as_SafePoint()->jvms(); 3319 assert(jvms != NULL, "sanity"); 3320 int start = jvms->debug_start(); 3321 int end = n->req(); 3322 bool is_uncommon = (n->is_CallStaticJava() && 3323 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3324 for (int j = start; j < end; j++) { 3325 Node* in = n->in(j); 3326 if (in->is_DecodeNarrowPtr()) { 3327 bool safe_to_skip = true; 3328 if (!is_uncommon ) { 3329 // Is it safe to skip? 3330 for (uint i = 0; i < in->outcnt(); i++) { 3331 Node* u = in->raw_out(i); 3332 if (!u->is_SafePoint() || 3333 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3334 safe_to_skip = false; 3335 } 3336 } 3337 } 3338 if (safe_to_skip) { 3339 n->set_req(j, in->in(1)); 3340 } 3341 if (in->outcnt() == 0) { 3342 in->disconnect_inputs(NULL, this); 3343 } 3344 } 3345 } 3346 } 3347 } 3348 3349 //------------------------------final_graph_reshaping-------------------------- 3350 // Final Graph Reshaping. 3351 // 3352 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3353 // and not commoned up and forced early. Must come after regular 3354 // optimizations to avoid GVN undoing the cloning. Clone constant 3355 // inputs to Loop Phis; these will be split by the allocator anyways. 3356 // Remove Opaque nodes. 3357 // (2) Move last-uses by commutative operations to the left input to encourage 3358 // Intel update-in-place two-address operations and better register usage 3359 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3360 // calls canonicalizing them back. 3361 // (3) Count the number of double-precision FP ops, single-precision FP ops 3362 // and call sites. On Intel, we can get correct rounding either by 3363 // forcing singles to memory (requires extra stores and loads after each 3364 // FP bytecode) or we can set a rounding mode bit (requires setting and 3365 // clearing the mode bit around call sites). The mode bit is only used 3366 // if the relative frequency of single FP ops to calls is low enough. 3367 // This is a key transform for SPEC mpeg_audio. 3368 // (4) Detect infinite loops; blobs of code reachable from above but not 3369 // below. Several of the Code_Gen algorithms fail on such code shapes, 3370 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3371 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3372 // Detection is by looking for IfNodes where only 1 projection is 3373 // reachable from below or CatchNodes missing some targets. 3374 // (5) Assert for insane oop offsets in debug mode. 3375 3376 bool Compile::final_graph_reshaping() { 3377 // an infinite loop may have been eliminated by the optimizer, 3378 // in which case the graph will be empty. 3379 if (root()->req() == 1) { 3380 record_method_not_compilable("trivial infinite loop"); 3381 return true; 3382 } 3383 3384 // Expensive nodes have their control input set to prevent the GVN 3385 // from freely commoning them. There's no GVN beyond this point so 3386 // no need to keep the control input. We want the expensive nodes to 3387 // be freely moved to the least frequent code path by gcm. 3388 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3389 for (int i = 0; i < expensive_count(); i++) { 3390 _expensive_nodes->at(i)->set_req(0, NULL); 3391 } 3392 3393 Final_Reshape_Counts frc; 3394 3395 // Visit everybody reachable! 3396 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3397 Node_Stack nstack(live_nodes() >> 1); 3398 final_graph_reshaping_walk(nstack, root(), frc); 3399 3400 // Check for unreachable (from below) code (i.e., infinite loops). 3401 for( uint i = 0; i < frc._tests.size(); i++ ) { 3402 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3403 // Get number of CFG targets. 3404 // Note that PCTables include exception targets after calls. 3405 uint required_outcnt = n->required_outcnt(); 3406 if (n->outcnt() != required_outcnt) { 3407 // Check for a few special cases. Rethrow Nodes never take the 3408 // 'fall-thru' path, so expected kids is 1 less. 3409 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3410 if (n->in(0)->in(0)->is_Call()) { 3411 CallNode *call = n->in(0)->in(0)->as_Call(); 3412 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3413 required_outcnt--; // Rethrow always has 1 less kid 3414 } else if (call->req() > TypeFunc::Parms && 3415 call->is_CallDynamicJava()) { 3416 // Check for null receiver. In such case, the optimizer has 3417 // detected that the virtual call will always result in a null 3418 // pointer exception. The fall-through projection of this CatchNode 3419 // will not be populated. 3420 Node *arg0 = call->in(TypeFunc::Parms); 3421 if (arg0->is_Type() && 3422 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3423 required_outcnt--; 3424 } 3425 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3426 call->req() > TypeFunc::Parms+1 && 3427 call->is_CallStaticJava()) { 3428 // Check for negative array length. In such case, the optimizer has 3429 // detected that the allocation attempt will always result in an 3430 // exception. There is no fall-through projection of this CatchNode . 3431 Node *arg1 = call->in(TypeFunc::Parms+1); 3432 if (arg1->is_Type() && 3433 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3434 required_outcnt--; 3435 } 3436 } 3437 } 3438 } 3439 // Recheck with a better notion of 'required_outcnt' 3440 if (n->outcnt() != required_outcnt) { 3441 record_method_not_compilable("malformed control flow"); 3442 return true; // Not all targets reachable! 3443 } 3444 } 3445 // Check that I actually visited all kids. Unreached kids 3446 // must be infinite loops. 3447 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3448 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3449 record_method_not_compilable("infinite loop"); 3450 return true; // Found unvisited kid; must be unreach 3451 } 3452 } 3453 3454 // If original bytecodes contained a mixture of floats and doubles 3455 // check if the optimizer has made it homogenous, item (3). 3456 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3457 frc.get_float_count() > 32 && 3458 frc.get_double_count() == 0 && 3459 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3460 set_24_bit_selection_and_mode( false, true ); 3461 } 3462 3463 set_java_calls(frc.get_java_call_count()); 3464 set_inner_loops(frc.get_inner_loop_count()); 3465 3466 // No infinite loops, no reason to bail out. 3467 return false; 3468 } 3469 3470 //-----------------------------too_many_traps---------------------------------- 3471 // Report if there are too many traps at the current method and bci. 3472 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3473 bool Compile::too_many_traps(ciMethod* method, 3474 int bci, 3475 Deoptimization::DeoptReason reason) { 3476 ciMethodData* md = method->method_data(); 3477 if (md->is_empty()) { 3478 // Assume the trap has not occurred, or that it occurred only 3479 // because of a transient condition during start-up in the interpreter. 3480 return false; 3481 } 3482 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3483 if (md->has_trap_at(bci, m, reason) != 0) { 3484 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3485 // Also, if there are multiple reasons, or if there is no per-BCI record, 3486 // assume the worst. 3487 if (log()) 3488 log()->elem("observe trap='%s' count='%d'", 3489 Deoptimization::trap_reason_name(reason), 3490 md->trap_count(reason)); 3491 return true; 3492 } else { 3493 // Ignore method/bci and see if there have been too many globally. 3494 return too_many_traps(reason, md); 3495 } 3496 } 3497 3498 // Less-accurate variant which does not require a method and bci. 3499 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3500 ciMethodData* logmd) { 3501 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3502 // Too many traps globally. 3503 // Note that we use cumulative trap_count, not just md->trap_count. 3504 if (log()) { 3505 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3506 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3507 Deoptimization::trap_reason_name(reason), 3508 mcount, trap_count(reason)); 3509 } 3510 return true; 3511 } else { 3512 // The coast is clear. 3513 return false; 3514 } 3515 } 3516 3517 //--------------------------too_many_recompiles-------------------------------- 3518 // Report if there are too many recompiles at the current method and bci. 3519 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3520 // Is not eager to return true, since this will cause the compiler to use 3521 // Action_none for a trap point, to avoid too many recompilations. 3522 bool Compile::too_many_recompiles(ciMethod* method, 3523 int bci, 3524 Deoptimization::DeoptReason reason) { 3525 ciMethodData* md = method->method_data(); 3526 if (md->is_empty()) { 3527 // Assume the trap has not occurred, or that it occurred only 3528 // because of a transient condition during start-up in the interpreter. 3529 return false; 3530 } 3531 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3532 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3533 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3534 Deoptimization::DeoptReason per_bc_reason 3535 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3536 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3537 if ((per_bc_reason == Deoptimization::Reason_none 3538 || md->has_trap_at(bci, m, reason) != 0) 3539 // The trap frequency measure we care about is the recompile count: 3540 && md->trap_recompiled_at(bci, m) 3541 && md->overflow_recompile_count() >= bc_cutoff) { 3542 // Do not emit a trap here if it has already caused recompilations. 3543 // Also, if there are multiple reasons, or if there is no per-BCI record, 3544 // assume the worst. 3545 if (log()) 3546 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3547 Deoptimization::trap_reason_name(reason), 3548 md->trap_count(reason), 3549 md->overflow_recompile_count()); 3550 return true; 3551 } else if (trap_count(reason) != 0 3552 && decompile_count() >= m_cutoff) { 3553 // Too many recompiles globally, and we have seen this sort of trap. 3554 // Use cumulative decompile_count, not just md->decompile_count. 3555 if (log()) 3556 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3557 Deoptimization::trap_reason_name(reason), 3558 md->trap_count(reason), trap_count(reason), 3559 md->decompile_count(), decompile_count()); 3560 return true; 3561 } else { 3562 // The coast is clear. 3563 return false; 3564 } 3565 } 3566 3567 // Compute when not to trap. Used by matching trap based nodes and 3568 // NullCheck optimization. 3569 void Compile::set_allowed_deopt_reasons() { 3570 _allowed_reasons = 0; 3571 if (is_method_compilation()) { 3572 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3573 assert(rs < BitsPerInt, "recode bit map"); 3574 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3575 _allowed_reasons |= nth_bit(rs); 3576 } 3577 } 3578 } 3579 } 3580 3581 #ifndef PRODUCT 3582 //------------------------------verify_graph_edges--------------------------- 3583 // Walk the Graph and verify that there is a one-to-one correspondence 3584 // between Use-Def edges and Def-Use edges in the graph. 3585 void Compile::verify_graph_edges(bool no_dead_code) { 3586 if (VerifyGraphEdges) { 3587 ResourceArea *area = Thread::current()->resource_area(); 3588 Unique_Node_List visited(area); 3589 // Call recursive graph walk to check edges 3590 _root->verify_edges(visited); 3591 if (no_dead_code) { 3592 // Now make sure that no visited node is used by an unvisited node. 3593 bool dead_nodes = false; 3594 Unique_Node_List checked(area); 3595 while (visited.size() > 0) { 3596 Node* n = visited.pop(); 3597 checked.push(n); 3598 for (uint i = 0; i < n->outcnt(); i++) { 3599 Node* use = n->raw_out(i); 3600 if (checked.member(use)) continue; // already checked 3601 if (visited.member(use)) continue; // already in the graph 3602 if (use->is_Con()) continue; // a dead ConNode is OK 3603 // At this point, we have found a dead node which is DU-reachable. 3604 if (!dead_nodes) { 3605 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3606 dead_nodes = true; 3607 } 3608 use->dump(2); 3609 tty->print_cr("---"); 3610 checked.push(use); // No repeats; pretend it is now checked. 3611 } 3612 } 3613 assert(!dead_nodes, "using nodes must be reachable from root"); 3614 } 3615 } 3616 } 3617 3618 // Verify GC barriers consistency 3619 // Currently supported: 3620 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3621 void Compile::verify_barriers() { 3622 if (UseG1GC) { 3623 // Verify G1 pre-barriers 3624 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + SATBMarkQueue::byte_offset_of_active()); 3625 3626 ResourceArea *area = Thread::current()->resource_area(); 3627 Unique_Node_List visited(area); 3628 Node_List worklist(area); 3629 // We're going to walk control flow backwards starting from the Root 3630 worklist.push(_root); 3631 while (worklist.size() > 0) { 3632 Node* x = worklist.pop(); 3633 if (x == NULL || x == top()) continue; 3634 if (visited.member(x)) { 3635 continue; 3636 } else { 3637 visited.push(x); 3638 } 3639 3640 if (x->is_Region()) { 3641 for (uint i = 1; i < x->req(); i++) { 3642 worklist.push(x->in(i)); 3643 } 3644 } else { 3645 worklist.push(x->in(0)); 3646 // We are looking for the pattern: 3647 // /->ThreadLocal 3648 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3649 // \->ConI(0) 3650 // We want to verify that the If and the LoadB have the same control 3651 // See GraphKit::g1_write_barrier_pre() 3652 if (x->is_If()) { 3653 IfNode *iff = x->as_If(); 3654 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3655 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3656 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3657 && cmp->in(1)->is_Load()) { 3658 LoadNode* load = cmp->in(1)->as_Load(); 3659 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3660 && load->in(2)->in(3)->is_Con() 3661 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3662 3663 Node* if_ctrl = iff->in(0); 3664 Node* load_ctrl = load->in(0); 3665 3666 if (if_ctrl != load_ctrl) { 3667 // Skip possible CProj->NeverBranch in infinite loops 3668 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3669 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3670 if_ctrl = if_ctrl->in(0)->in(0); 3671 } 3672 } 3673 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3674 } 3675 } 3676 } 3677 } 3678 } 3679 } 3680 } 3681 } 3682 3683 #endif 3684 3685 // The Compile object keeps track of failure reasons separately from the ciEnv. 3686 // This is required because there is not quite a 1-1 relation between the 3687 // ciEnv and its compilation task and the Compile object. Note that one 3688 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3689 // to backtrack and retry without subsuming loads. Other than this backtracking 3690 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3691 // by the logic in C2Compiler. 3692 void Compile::record_failure(const char* reason) { 3693 if (log() != NULL) { 3694 log()->elem("failure reason='%s' phase='compile'", reason); 3695 } 3696 if (_failure_reason == NULL) { 3697 // Record the first failure reason. 3698 _failure_reason = reason; 3699 } 3700 3701 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3702 C->print_method(PHASE_FAILURE); 3703 } 3704 _root = NULL; // flush the graph, too 3705 } 3706 3707 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3708 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3709 _phase_name(name), _dolog(CITimeVerbose) 3710 { 3711 if (_dolog) { 3712 C = Compile::current(); 3713 _log = C->log(); 3714 } else { 3715 C = NULL; 3716 _log = NULL; 3717 } 3718 if (_log != NULL) { 3719 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3720 _log->stamp(); 3721 _log->end_head(); 3722 } 3723 } 3724 3725 Compile::TracePhase::~TracePhase() { 3726 3727 C = Compile::current(); 3728 if (_dolog) { 3729 _log = C->log(); 3730 } else { 3731 _log = NULL; 3732 } 3733 3734 #ifdef ASSERT 3735 if (PrintIdealNodeCount) { 3736 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3737 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3738 } 3739 3740 if (VerifyIdealNodeCount) { 3741 Compile::current()->print_missing_nodes(); 3742 } 3743 #endif 3744 3745 if (_log != NULL) { 3746 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3747 } 3748 } 3749 3750 //============================================================================= 3751 // Two Constant's are equal when the type and the value are equal. 3752 bool Compile::Constant::operator==(const Constant& other) { 3753 if (type() != other.type() ) return false; 3754 if (can_be_reused() != other.can_be_reused()) return false; 3755 // For floating point values we compare the bit pattern. 3756 switch (type()) { 3757 case T_FLOAT: return (_v._value.i == other._v._value.i); 3758 case T_LONG: 3759 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3760 case T_OBJECT: 3761 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3762 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3763 case T_METADATA: return (_v._metadata == other._v._metadata); 3764 default: ShouldNotReachHere(); 3765 } 3766 return false; 3767 } 3768 3769 static int type_to_size_in_bytes(BasicType t) { 3770 switch (t) { 3771 case T_LONG: return sizeof(jlong ); 3772 case T_FLOAT: return sizeof(jfloat ); 3773 case T_DOUBLE: return sizeof(jdouble); 3774 case T_METADATA: return sizeof(Metadata*); 3775 // We use T_VOID as marker for jump-table entries (labels) which 3776 // need an internal word relocation. 3777 case T_VOID: 3778 case T_ADDRESS: 3779 case T_OBJECT: return sizeof(jobject); 3780 } 3781 3782 ShouldNotReachHere(); 3783 return -1; 3784 } 3785 3786 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3787 // sort descending 3788 if (a->freq() > b->freq()) return -1; 3789 if (a->freq() < b->freq()) return 1; 3790 return 0; 3791 } 3792 3793 void Compile::ConstantTable::calculate_offsets_and_size() { 3794 // First, sort the array by frequencies. 3795 _constants.sort(qsort_comparator); 3796 3797 #ifdef ASSERT 3798 // Make sure all jump-table entries were sorted to the end of the 3799 // array (they have a negative frequency). 3800 bool found_void = false; 3801 for (int i = 0; i < _constants.length(); i++) { 3802 Constant con = _constants.at(i); 3803 if (con.type() == T_VOID) 3804 found_void = true; // jump-tables 3805 else 3806 assert(!found_void, "wrong sorting"); 3807 } 3808 #endif 3809 3810 int offset = 0; 3811 for (int i = 0; i < _constants.length(); i++) { 3812 Constant* con = _constants.adr_at(i); 3813 3814 // Align offset for type. 3815 int typesize = type_to_size_in_bytes(con->type()); 3816 offset = align_size_up(offset, typesize); 3817 con->set_offset(offset); // set constant's offset 3818 3819 if (con->type() == T_VOID) { 3820 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3821 offset = offset + typesize * n->outcnt(); // expand jump-table 3822 } else { 3823 offset = offset + typesize; 3824 } 3825 } 3826 3827 // Align size up to the next section start (which is insts; see 3828 // CodeBuffer::align_at_start). 3829 assert(_size == -1, "already set?"); 3830 _size = align_size_up(offset, CodeEntryAlignment); 3831 } 3832 3833 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3834 MacroAssembler _masm(&cb); 3835 for (int i = 0; i < _constants.length(); i++) { 3836 Constant con = _constants.at(i); 3837 address constant_addr = NULL; 3838 switch (con.type()) { 3839 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3840 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3841 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3842 case T_OBJECT: { 3843 jobject obj = con.get_jobject(); 3844 int oop_index = _masm.oop_recorder()->find_index(obj); 3845 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3846 break; 3847 } 3848 case T_ADDRESS: { 3849 address addr = (address) con.get_jobject(); 3850 constant_addr = _masm.address_constant(addr); 3851 break; 3852 } 3853 // We use T_VOID as marker for jump-table entries (labels) which 3854 // need an internal word relocation. 3855 case T_VOID: { 3856 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3857 // Fill the jump-table with a dummy word. The real value is 3858 // filled in later in fill_jump_table. 3859 address dummy = (address) n; 3860 constant_addr = _masm.address_constant(dummy); 3861 // Expand jump-table 3862 for (uint i = 1; i < n->outcnt(); i++) { 3863 address temp_addr = _masm.address_constant(dummy + i); 3864 assert(temp_addr, "consts section too small"); 3865 } 3866 break; 3867 } 3868 case T_METADATA: { 3869 Metadata* obj = con.get_metadata(); 3870 int metadata_index = _masm.oop_recorder()->find_index(obj); 3871 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3872 break; 3873 } 3874 default: ShouldNotReachHere(); 3875 } 3876 assert(constant_addr, "consts section too small"); 3877 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3878 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 3879 } 3880 } 3881 3882 int Compile::ConstantTable::find_offset(Constant& con) const { 3883 int idx = _constants.find(con); 3884 assert(idx != -1, "constant must be in constant table"); 3885 int offset = _constants.at(idx).offset(); 3886 assert(offset != -1, "constant table not emitted yet?"); 3887 return offset; 3888 } 3889 3890 void Compile::ConstantTable::add(Constant& con) { 3891 if (con.can_be_reused()) { 3892 int idx = _constants.find(con); 3893 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3894 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3895 return; 3896 } 3897 } 3898 (void) _constants.append(con); 3899 } 3900 3901 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3902 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3903 Constant con(type, value, b->_freq); 3904 add(con); 3905 return con; 3906 } 3907 3908 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3909 Constant con(metadata); 3910 add(con); 3911 return con; 3912 } 3913 3914 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3915 jvalue value; 3916 BasicType type = oper->type()->basic_type(); 3917 switch (type) { 3918 case T_LONG: value.j = oper->constantL(); break; 3919 case T_FLOAT: value.f = oper->constantF(); break; 3920 case T_DOUBLE: value.d = oper->constantD(); break; 3921 case T_OBJECT: 3922 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3923 case T_METADATA: return add((Metadata*)oper->constant()); break; 3924 default: guarantee(false, "unhandled type: %s", type2name(type)); 3925 } 3926 return add(n, type, value); 3927 } 3928 3929 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3930 jvalue value; 3931 // We can use the node pointer here to identify the right jump-table 3932 // as this method is called from Compile::Fill_buffer right before 3933 // the MachNodes are emitted and the jump-table is filled (means the 3934 // MachNode pointers do not change anymore). 3935 value.l = (jobject) n; 3936 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3937 add(con); 3938 return con; 3939 } 3940 3941 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3942 // If called from Compile::scratch_emit_size do nothing. 3943 if (Compile::current()->in_scratch_emit_size()) return; 3944 3945 assert(labels.is_nonempty(), "must be"); 3946 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 3947 3948 // Since MachConstantNode::constant_offset() also contains 3949 // table_base_offset() we need to subtract the table_base_offset() 3950 // to get the plain offset into the constant table. 3951 int offset = n->constant_offset() - table_base_offset(); 3952 3953 MacroAssembler _masm(&cb); 3954 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3955 3956 for (uint i = 0; i < n->outcnt(); i++) { 3957 address* constant_addr = &jump_table_base[i]; 3958 assert(*constant_addr == (((address) n) + i), "all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i)); 3959 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3960 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3961 } 3962 } 3963 3964 //----------------------------static_subtype_check----------------------------- 3965 // Shortcut important common cases when superklass is exact: 3966 // (0) superklass is java.lang.Object (can occur in reflective code) 3967 // (1) subklass is already limited to a subtype of superklass => always ok 3968 // (2) subklass does not overlap with superklass => always fail 3969 // (3) superklass has NO subtypes and we can check with a simple compare. 3970 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 3971 if (StressReflectiveCode) { 3972 return SSC_full_test; // Let caller generate the general case. 3973 } 3974 3975 if (superk == env()->Object_klass()) { 3976 return SSC_always_true; // (0) this test cannot fail 3977 } 3978 3979 ciType* superelem = superk; 3980 if (superelem->is_array_klass()) 3981 superelem = superelem->as_array_klass()->base_element_type(); 3982 3983 if (!subk->is_interface()) { // cannot trust static interface types yet 3984 if (subk->is_subtype_of(superk)) { 3985 return SSC_always_true; // (1) false path dead; no dynamic test needed 3986 } 3987 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 3988 !superk->is_subtype_of(subk)) { 3989 return SSC_always_false; 3990 } 3991 } 3992 3993 // If casting to an instance klass, it must have no subtypes 3994 if (superk->is_interface()) { 3995 // Cannot trust interfaces yet. 3996 // %%% S.B. superk->nof_implementors() == 1 3997 } else if (superelem->is_instance_klass()) { 3998 ciInstanceKlass* ik = superelem->as_instance_klass(); 3999 if (!ik->has_subklass() && !ik->is_interface()) { 4000 if (!ik->is_final()) { 4001 // Add a dependency if there is a chance of a later subclass. 4002 dependencies()->assert_leaf_type(ik); 4003 } 4004 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4005 } 4006 } else { 4007 // A primitive array type has no subtypes. 4008 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4009 } 4010 4011 return SSC_full_test; 4012 } 4013 4014 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4015 #ifdef _LP64 4016 // The scaled index operand to AddP must be a clean 64-bit value. 4017 // Java allows a 32-bit int to be incremented to a negative 4018 // value, which appears in a 64-bit register as a large 4019 // positive number. Using that large positive number as an 4020 // operand in pointer arithmetic has bad consequences. 4021 // On the other hand, 32-bit overflow is rare, and the possibility 4022 // can often be excluded, if we annotate the ConvI2L node with 4023 // a type assertion that its value is known to be a small positive 4024 // number. (The prior range check has ensured this.) 4025 // This assertion is used by ConvI2LNode::Ideal. 4026 int index_max = max_jint - 1; // array size is max_jint, index is one less 4027 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4028 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4029 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4030 #endif 4031 return idx; 4032 } 4033 4034 #ifdef _LP64 4035 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4036 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4037 if (ctrl != NULL) { 4038 // Express control dependency by a CastII node with a narrow type. 4039 value = new CastIINode(value, itype, false, true /* range check dependency */); 4040 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4041 // node from floating above the range check during loop optimizations. Otherwise, the 4042 // ConvI2L node may be eliminated independently of the range check, causing the data path 4043 // to become TOP while the control path is still there (although it's unreachable). 4044 value->set_req(0, ctrl); 4045 // Save CastII node to remove it after loop optimizations. 4046 phase->C->add_range_check_cast(value); 4047 value = phase->transform(value); 4048 } 4049 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4050 return phase->transform(new ConvI2LNode(value, ltype)); 4051 } 4052 #endif 4053 4054 // The message about the current inlining is accumulated in 4055 // _print_inlining_stream and transfered into the _print_inlining_list 4056 // once we know whether inlining succeeds or not. For regular 4057 // inlining, messages are appended to the buffer pointed by 4058 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4059 // a new buffer is added after _print_inlining_idx in the list. This 4060 // way we can update the inlining message for late inlining call site 4061 // when the inlining is attempted again. 4062 void Compile::print_inlining_init() { 4063 if (print_inlining() || print_intrinsics()) { 4064 _print_inlining_stream = new stringStream(); 4065 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4066 } 4067 } 4068 4069 void Compile::print_inlining_reinit() { 4070 if (print_inlining() || print_intrinsics()) { 4071 // Re allocate buffer when we change ResourceMark 4072 _print_inlining_stream = new stringStream(); 4073 } 4074 } 4075 4076 void Compile::print_inlining_reset() { 4077 _print_inlining_stream->reset(); 4078 } 4079 4080 void Compile::print_inlining_commit() { 4081 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4082 // Transfer the message from _print_inlining_stream to the current 4083 // _print_inlining_list buffer and clear _print_inlining_stream. 4084 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4085 print_inlining_reset(); 4086 } 4087 4088 void Compile::print_inlining_push() { 4089 // Add new buffer to the _print_inlining_list at current position 4090 _print_inlining_idx++; 4091 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4092 } 4093 4094 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4095 return _print_inlining_list->at(_print_inlining_idx); 4096 } 4097 4098 void Compile::print_inlining_update(CallGenerator* cg) { 4099 if (print_inlining() || print_intrinsics()) { 4100 if (!cg->is_late_inline()) { 4101 if (print_inlining_current().cg() != NULL) { 4102 print_inlining_push(); 4103 } 4104 print_inlining_commit(); 4105 } else { 4106 if (print_inlining_current().cg() != cg && 4107 (print_inlining_current().cg() != NULL || 4108 print_inlining_current().ss()->size() != 0)) { 4109 print_inlining_push(); 4110 } 4111 print_inlining_commit(); 4112 print_inlining_current().set_cg(cg); 4113 } 4114 } 4115 } 4116 4117 void Compile::print_inlining_move_to(CallGenerator* cg) { 4118 // We resume inlining at a late inlining call site. Locate the 4119 // corresponding inlining buffer so that we can update it. 4120 if (print_inlining()) { 4121 for (int i = 0; i < _print_inlining_list->length(); i++) { 4122 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4123 _print_inlining_idx = i; 4124 return; 4125 } 4126 } 4127 ShouldNotReachHere(); 4128 } 4129 } 4130 4131 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4132 if (print_inlining()) { 4133 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4134 assert(print_inlining_current().cg() == cg, "wrong entry"); 4135 // replace message with new message 4136 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4137 print_inlining_commit(); 4138 print_inlining_current().set_cg(cg); 4139 } 4140 } 4141 4142 void Compile::print_inlining_assert_ready() { 4143 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4144 } 4145 4146 void Compile::process_print_inlining() { 4147 bool do_print_inlining = print_inlining() || print_intrinsics(); 4148 if (do_print_inlining || log() != NULL) { 4149 // Print inlining message for candidates that we couldn't inline 4150 // for lack of space 4151 for (int i = 0; i < _late_inlines.length(); i++) { 4152 CallGenerator* cg = _late_inlines.at(i); 4153 if (!cg->is_mh_late_inline()) { 4154 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4155 if (do_print_inlining) { 4156 cg->print_inlining_late(msg); 4157 } 4158 log_late_inline_failure(cg, msg); 4159 } 4160 } 4161 } 4162 if (do_print_inlining) { 4163 ResourceMark rm; 4164 stringStream ss; 4165 for (int i = 0; i < _print_inlining_list->length(); i++) { 4166 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4167 } 4168 size_t end = ss.size(); 4169 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4170 strncpy(_print_inlining_output, ss.base(), end+1); 4171 _print_inlining_output[end] = 0; 4172 } 4173 } 4174 4175 void Compile::dump_print_inlining() { 4176 if (_print_inlining_output != NULL) { 4177 tty->print_raw(_print_inlining_output); 4178 } 4179 } 4180 4181 void Compile::log_late_inline(CallGenerator* cg) { 4182 if (log() != NULL) { 4183 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4184 cg->unique_id()); 4185 JVMState* p = cg->call_node()->jvms(); 4186 while (p != NULL) { 4187 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4188 p = p->caller(); 4189 } 4190 log()->tail("late_inline"); 4191 } 4192 } 4193 4194 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4195 log_late_inline(cg); 4196 if (log() != NULL) { 4197 log()->inline_fail(msg); 4198 } 4199 } 4200 4201 void Compile::log_inline_id(CallGenerator* cg) { 4202 if (log() != NULL) { 4203 // The LogCompilation tool needs a unique way to identify late 4204 // inline call sites. This id must be unique for this call site in 4205 // this compilation. Try to have it unique across compilations as 4206 // well because it can be convenient when grepping through the log 4207 // file. 4208 // Distinguish OSR compilations from others in case CICountOSR is 4209 // on. 4210 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4211 cg->set_unique_id(id); 4212 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4213 } 4214 } 4215 4216 void Compile::log_inline_failure(const char* msg) { 4217 if (C->log() != NULL) { 4218 C->log()->inline_fail(msg); 4219 } 4220 } 4221 4222 4223 // Dump inlining replay data to the stream. 4224 // Don't change thread state and acquire any locks. 4225 void Compile::dump_inline_data(outputStream* out) { 4226 InlineTree* inl_tree = ilt(); 4227 if (inl_tree != NULL) { 4228 out->print(" inline %d", inl_tree->count()); 4229 inl_tree->dump_replay_data(out); 4230 } 4231 } 4232 4233 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4234 if (n1->Opcode() < n2->Opcode()) return -1; 4235 else if (n1->Opcode() > n2->Opcode()) return 1; 4236 4237 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4238 for (uint i = 1; i < n1->req(); i++) { 4239 if (n1->in(i) < n2->in(i)) return -1; 4240 else if (n1->in(i) > n2->in(i)) return 1; 4241 } 4242 4243 return 0; 4244 } 4245 4246 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4247 Node* n1 = *n1p; 4248 Node* n2 = *n2p; 4249 4250 return cmp_expensive_nodes(n1, n2); 4251 } 4252 4253 void Compile::sort_expensive_nodes() { 4254 if (!expensive_nodes_sorted()) { 4255 _expensive_nodes->sort(cmp_expensive_nodes); 4256 } 4257 } 4258 4259 bool Compile::expensive_nodes_sorted() const { 4260 for (int i = 1; i < _expensive_nodes->length(); i++) { 4261 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4262 return false; 4263 } 4264 } 4265 return true; 4266 } 4267 4268 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4269 if (_expensive_nodes->length() == 0) { 4270 return false; 4271 } 4272 4273 assert(OptimizeExpensiveOps, "optimization off?"); 4274 4275 // Take this opportunity to remove dead nodes from the list 4276 int j = 0; 4277 for (int i = 0; i < _expensive_nodes->length(); i++) { 4278 Node* n = _expensive_nodes->at(i); 4279 if (!n->is_unreachable(igvn)) { 4280 assert(n->is_expensive(), "should be expensive"); 4281 _expensive_nodes->at_put(j, n); 4282 j++; 4283 } 4284 } 4285 _expensive_nodes->trunc_to(j); 4286 4287 // Then sort the list so that similar nodes are next to each other 4288 // and check for at least two nodes of identical kind with same data 4289 // inputs. 4290 sort_expensive_nodes(); 4291 4292 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4293 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4294 return true; 4295 } 4296 } 4297 4298 return false; 4299 } 4300 4301 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4302 if (_expensive_nodes->length() == 0) { 4303 return; 4304 } 4305 4306 assert(OptimizeExpensiveOps, "optimization off?"); 4307 4308 // Sort to bring similar nodes next to each other and clear the 4309 // control input of nodes for which there's only a single copy. 4310 sort_expensive_nodes(); 4311 4312 int j = 0; 4313 int identical = 0; 4314 int i = 0; 4315 bool modified = false; 4316 for (; i < _expensive_nodes->length()-1; i++) { 4317 assert(j <= i, "can't write beyond current index"); 4318 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4319 identical++; 4320 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4321 continue; 4322 } 4323 if (identical > 0) { 4324 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4325 identical = 0; 4326 } else { 4327 Node* n = _expensive_nodes->at(i); 4328 igvn.replace_input_of(n, 0, NULL); 4329 igvn.hash_insert(n); 4330 modified = true; 4331 } 4332 } 4333 if (identical > 0) { 4334 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4335 } else if (_expensive_nodes->length() >= 1) { 4336 Node* n = _expensive_nodes->at(i); 4337 igvn.replace_input_of(n, 0, NULL); 4338 igvn.hash_insert(n); 4339 modified = true; 4340 } 4341 _expensive_nodes->trunc_to(j); 4342 if (modified) { 4343 igvn.optimize(); 4344 } 4345 } 4346 4347 void Compile::add_expensive_node(Node * n) { 4348 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4349 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4350 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4351 if (OptimizeExpensiveOps) { 4352 _expensive_nodes->append(n); 4353 } else { 4354 // Clear control input and let IGVN optimize expensive nodes if 4355 // OptimizeExpensiveOps is off. 4356 n->set_req(0, NULL); 4357 } 4358 } 4359 4360 /** 4361 * Remove the speculative part of types and clean up the graph 4362 */ 4363 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4364 if (UseTypeSpeculation) { 4365 Unique_Node_List worklist; 4366 worklist.push(root()); 4367 int modified = 0; 4368 // Go over all type nodes that carry a speculative type, drop the 4369 // speculative part of the type and enqueue the node for an igvn 4370 // which may optimize it out. 4371 for (uint next = 0; next < worklist.size(); ++next) { 4372 Node *n = worklist.at(next); 4373 if (n->is_Type()) { 4374 TypeNode* tn = n->as_Type(); 4375 const Type* t = tn->type(); 4376 const Type* t_no_spec = t->remove_speculative(); 4377 if (t_no_spec != t) { 4378 bool in_hash = igvn.hash_delete(n); 4379 assert(in_hash, "node should be in igvn hash table"); 4380 tn->set_type(t_no_spec); 4381 igvn.hash_insert(n); 4382 igvn._worklist.push(n); // give it a chance to go away 4383 modified++; 4384 } 4385 } 4386 uint max = n->len(); 4387 for( uint i = 0; i < max; ++i ) { 4388 Node *m = n->in(i); 4389 if (not_a_node(m)) continue; 4390 worklist.push(m); 4391 } 4392 } 4393 // Drop the speculative part of all types in the igvn's type table 4394 igvn.remove_speculative_types(); 4395 if (modified > 0) { 4396 igvn.optimize(); 4397 } 4398 #ifdef ASSERT 4399 // Verify that after the IGVN is over no speculative type has resurfaced 4400 worklist.clear(); 4401 worklist.push(root()); 4402 for (uint next = 0; next < worklist.size(); ++next) { 4403 Node *n = worklist.at(next); 4404 const Type* t = igvn.type_or_null(n); 4405 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4406 if (n->is_Type()) { 4407 t = n->as_Type()->type(); 4408 assert(t == t->remove_speculative(), "no more speculative types"); 4409 } 4410 uint max = n->len(); 4411 for( uint i = 0; i < max; ++i ) { 4412 Node *m = n->in(i); 4413 if (not_a_node(m)) continue; 4414 worklist.push(m); 4415 } 4416 } 4417 igvn.check_no_speculative_types(); 4418 #endif 4419 } 4420 } 4421 4422 // Auxiliary method to support randomized stressing/fuzzing. 4423 // 4424 // This method can be called the arbitrary number of times, with current count 4425 // as the argument. The logic allows selecting a single candidate from the 4426 // running list of candidates as follows: 4427 // int count = 0; 4428 // Cand* selected = null; 4429 // while(cand = cand->next()) { 4430 // if (randomized_select(++count)) { 4431 // selected = cand; 4432 // } 4433 // } 4434 // 4435 // Including count equalizes the chances any candidate is "selected". 4436 // This is useful when we don't have the complete list of candidates to choose 4437 // from uniformly. In this case, we need to adjust the randomicity of the 4438 // selection, or else we will end up biasing the selection towards the latter 4439 // candidates. 4440 // 4441 // Quick back-envelope calculation shows that for the list of n candidates 4442 // the equal probability for the candidate to persist as "best" can be 4443 // achieved by replacing it with "next" k-th candidate with the probability 4444 // of 1/k. It can be easily shown that by the end of the run, the 4445 // probability for any candidate is converged to 1/n, thus giving the 4446 // uniform distribution among all the candidates. 4447 // 4448 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4449 #define RANDOMIZED_DOMAIN_POW 29 4450 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4451 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4452 bool Compile::randomized_select(int count) { 4453 assert(count > 0, "only positive"); 4454 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4455 } 4456 4457 CloneMap& Compile::clone_map() { return _clone_map; } 4458 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4459 4460 void NodeCloneInfo::dump() const { 4461 tty->print(" {%d:%d} ", idx(), gen()); 4462 } 4463 4464 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4465 uint64_t val = value(old->_idx); 4466 NodeCloneInfo cio(val); 4467 assert(val != 0, "old node should be in the map"); 4468 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4469 insert(nnn->_idx, cin.get()); 4470 #ifndef PRODUCT 4471 if (is_debug()) { 4472 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4473 } 4474 #endif 4475 } 4476 4477 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4478 NodeCloneInfo cio(value(old->_idx)); 4479 if (cio.get() == 0) { 4480 cio.set(old->_idx, 0); 4481 insert(old->_idx, cio.get()); 4482 #ifndef PRODUCT 4483 if (is_debug()) { 4484 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4485 } 4486 #endif 4487 } 4488 clone(old, nnn, gen); 4489 } 4490 4491 int CloneMap::max_gen() const { 4492 int g = 0; 4493 DictI di(_dict); 4494 for(; di.test(); ++di) { 4495 int t = gen(di._key); 4496 if (g < t) { 4497 g = t; 4498 #ifndef PRODUCT 4499 if (is_debug()) { 4500 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4501 } 4502 #endif 4503 } 4504 } 4505 return g; 4506 } 4507 4508 void CloneMap::dump(node_idx_t key) const { 4509 uint64_t val = value(key); 4510 if (val != 0) { 4511 NodeCloneInfo ni(val); 4512 ni.dump(); 4513 } 4514 }