1 /* 2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/block.hpp" 38 #include "opto/c2compiler.hpp" 39 #include "opto/callGenerator.hpp" 40 #include "opto/callnode.hpp" 41 #include "opto/cfgnode.hpp" 42 #include "opto/chaitin.hpp" 43 #include "opto/compile.hpp" 44 #include "opto/connode.hpp" 45 #include "opto/convertnode.hpp" 46 #include "opto/divnode.hpp" 47 #include "opto/escape.hpp" 48 #include "opto/idealGraphPrinter.hpp" 49 #include "opto/loopnode.hpp" 50 #include "opto/machnode.hpp" 51 #include "opto/macro.hpp" 52 #include "opto/matcher.hpp" 53 #include "opto/mathexactnode.hpp" 54 #include "opto/memnode.hpp" 55 #include "opto/mulnode.hpp" 56 #include "opto/narrowptrnode.hpp" 57 #include "opto/node.hpp" 58 #include "opto/opcodes.hpp" 59 #include "opto/output.hpp" 60 #include "opto/parse.hpp" 61 #include "opto/phaseX.hpp" 62 #include "opto/rootnode.hpp" 63 #include "opto/runtime.hpp" 64 #include "opto/stringopts.hpp" 65 #include "opto/type.hpp" 66 #include "opto/vectornode.hpp" 67 #include "runtime/arguments.hpp" 68 #include "runtime/sharedRuntime.hpp" 69 #include "runtime/signature.hpp" 70 #include "runtime/stubRoutines.hpp" 71 #include "runtime/timer.hpp" 72 #include "utilities/copy.hpp" 73 74 75 // -------------------- Compile::mach_constant_base_node ----------------------- 76 // Constant table base node singleton. 77 MachConstantBaseNode* Compile::mach_constant_base_node() { 78 if (_mach_constant_base_node == NULL) { 79 _mach_constant_base_node = new MachConstantBaseNode(); 80 _mach_constant_base_node->add_req(C->root()); 81 } 82 return _mach_constant_base_node; 83 } 84 85 86 /// Support for intrinsics. 87 88 // Return the index at which m must be inserted (or already exists). 89 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 90 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 91 #ifdef ASSERT 92 for (int i = 1; i < _intrinsics->length(); i++) { 93 CallGenerator* cg1 = _intrinsics->at(i-1); 94 CallGenerator* cg2 = _intrinsics->at(i); 95 assert(cg1->method() != cg2->method() 96 ? cg1->method() < cg2->method() 97 : cg1->is_virtual() < cg2->is_virtual(), 98 "compiler intrinsics list must stay sorted"); 99 } 100 #endif 101 // Binary search sorted list, in decreasing intervals [lo, hi]. 102 int lo = 0, hi = _intrinsics->length()-1; 103 while (lo <= hi) { 104 int mid = (uint)(hi + lo) / 2; 105 ciMethod* mid_m = _intrinsics->at(mid)->method(); 106 if (m < mid_m) { 107 hi = mid-1; 108 } else if (m > mid_m) { 109 lo = mid+1; 110 } else { 111 // look at minor sort key 112 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 113 if (is_virtual < mid_virt) { 114 hi = mid-1; 115 } else if (is_virtual > mid_virt) { 116 lo = mid+1; 117 } else { 118 return mid; // exact match 119 } 120 } 121 } 122 return lo; // inexact match 123 } 124 125 void Compile::register_intrinsic(CallGenerator* cg) { 126 if (_intrinsics == NULL) { 127 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 128 } 129 // This code is stolen from ciObjectFactory::insert. 130 // Really, GrowableArray should have methods for 131 // insert_at, remove_at, and binary_search. 132 int len = _intrinsics->length(); 133 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 134 if (index == len) { 135 _intrinsics->append(cg); 136 } else { 137 #ifdef ASSERT 138 CallGenerator* oldcg = _intrinsics->at(index); 139 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 140 #endif 141 _intrinsics->append(_intrinsics->at(len-1)); 142 int pos; 143 for (pos = len-2; pos >= index; pos--) { 144 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 145 } 146 _intrinsics->at_put(index, cg); 147 } 148 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 149 } 150 151 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 152 assert(m->is_loaded(), "don't try this on unloaded methods"); 153 if (_intrinsics != NULL) { 154 int index = intrinsic_insertion_index(m, is_virtual); 155 if (index < _intrinsics->length() 156 && _intrinsics->at(index)->method() == m 157 && _intrinsics->at(index)->is_virtual() == is_virtual) { 158 return _intrinsics->at(index); 159 } 160 } 161 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 162 if (m->intrinsic_id() != vmIntrinsics::_none && 163 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 164 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 165 if (cg != NULL) { 166 // Save it for next time: 167 register_intrinsic(cg); 168 return cg; 169 } else { 170 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 171 } 172 } 173 return NULL; 174 } 175 176 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 177 // in library_call.cpp. 178 179 180 #ifndef PRODUCT 181 // statistics gathering... 182 183 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 184 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 185 186 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 187 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 188 int oflags = _intrinsic_hist_flags[id]; 189 assert(flags != 0, "what happened?"); 190 if (is_virtual) { 191 flags |= _intrinsic_virtual; 192 } 193 bool changed = (flags != oflags); 194 if ((flags & _intrinsic_worked) != 0) { 195 juint count = (_intrinsic_hist_count[id] += 1); 196 if (count == 1) { 197 changed = true; // first time 198 } 199 // increment the overall count also: 200 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 201 } 202 if (changed) { 203 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 204 // Something changed about the intrinsic's virtuality. 205 if ((flags & _intrinsic_virtual) != 0) { 206 // This is the first use of this intrinsic as a virtual call. 207 if (oflags != 0) { 208 // We already saw it as a non-virtual, so note both cases. 209 flags |= _intrinsic_both; 210 } 211 } else if ((oflags & _intrinsic_both) == 0) { 212 // This is the first use of this intrinsic as a non-virtual 213 flags |= _intrinsic_both; 214 } 215 } 216 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 217 } 218 // update the overall flags also: 219 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 220 return changed; 221 } 222 223 static char* format_flags(int flags, char* buf) { 224 buf[0] = 0; 225 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 226 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 227 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 228 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 229 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 230 if (buf[0] == 0) strcat(buf, ","); 231 assert(buf[0] == ',', "must be"); 232 return &buf[1]; 233 } 234 235 void Compile::print_intrinsic_statistics() { 236 char flagsbuf[100]; 237 ttyLocker ttyl; 238 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 239 tty->print_cr("Compiler intrinsic usage:"); 240 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 241 if (total == 0) total = 1; // avoid div0 in case of no successes 242 #define PRINT_STAT_LINE(name, c, f) \ 243 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 244 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 245 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 246 int flags = _intrinsic_hist_flags[id]; 247 juint count = _intrinsic_hist_count[id]; 248 if ((flags | count) != 0) { 249 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 250 } 251 } 252 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 253 if (xtty != NULL) xtty->tail("statistics"); 254 } 255 256 void Compile::print_statistics() { 257 { ttyLocker ttyl; 258 if (xtty != NULL) xtty->head("statistics type='opto'"); 259 Parse::print_statistics(); 260 PhaseCCP::print_statistics(); 261 PhaseRegAlloc::print_statistics(); 262 Scheduling::print_statistics(); 263 PhasePeephole::print_statistics(); 264 PhaseIdealLoop::print_statistics(); 265 if (xtty != NULL) xtty->tail("statistics"); 266 } 267 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 268 // put this under its own <statistics> element. 269 print_intrinsic_statistics(); 270 } 271 } 272 #endif //PRODUCT 273 274 // Support for bundling info 275 Bundle* Compile::node_bundling(const Node *n) { 276 assert(valid_bundle_info(n), "oob"); 277 return &_node_bundling_base[n->_idx]; 278 } 279 280 bool Compile::valid_bundle_info(const Node *n) { 281 return (_node_bundling_limit > n->_idx); 282 } 283 284 285 void Compile::gvn_replace_by(Node* n, Node* nn) { 286 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 287 Node* use = n->last_out(i); 288 bool is_in_table = initial_gvn()->hash_delete(use); 289 uint uses_found = 0; 290 for (uint j = 0; j < use->len(); j++) { 291 if (use->in(j) == n) { 292 if (j < use->req()) 293 use->set_req(j, nn); 294 else 295 use->set_prec(j, nn); 296 uses_found++; 297 } 298 } 299 if (is_in_table) { 300 // reinsert into table 301 initial_gvn()->hash_find_insert(use); 302 } 303 record_for_igvn(use); 304 i -= uses_found; // we deleted 1 or more copies of this edge 305 } 306 } 307 308 309 static inline bool not_a_node(const Node* n) { 310 if (n == NULL) return true; 311 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 312 if (*(address*)n == badAddress) return true; // kill by Node::destruct 313 return false; 314 } 315 316 // Identify all nodes that are reachable from below, useful. 317 // Use breadth-first pass that records state in a Unique_Node_List, 318 // recursive traversal is slower. 319 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 320 int estimated_worklist_size = live_nodes(); 321 useful.map( estimated_worklist_size, NULL ); // preallocate space 322 323 // Initialize worklist 324 if (root() != NULL) { useful.push(root()); } 325 // If 'top' is cached, declare it useful to preserve cached node 326 if( cached_top_node() ) { useful.push(cached_top_node()); } 327 328 // Push all useful nodes onto the list, breadthfirst 329 for( uint next = 0; next < useful.size(); ++next ) { 330 assert( next < unique(), "Unique useful nodes < total nodes"); 331 Node *n = useful.at(next); 332 uint max = n->len(); 333 for( uint i = 0; i < max; ++i ) { 334 Node *m = n->in(i); 335 if (not_a_node(m)) continue; 336 useful.push(m); 337 } 338 } 339 } 340 341 // Update dead_node_list with any missing dead nodes using useful 342 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 343 void Compile::update_dead_node_list(Unique_Node_List &useful) { 344 uint max_idx = unique(); 345 VectorSet& useful_node_set = useful.member_set(); 346 347 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 348 // If node with index node_idx is not in useful set, 349 // mark it as dead in dead node list. 350 if (! useful_node_set.test(node_idx) ) { 351 record_dead_node(node_idx); 352 } 353 } 354 } 355 356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 357 int shift = 0; 358 for (int i = 0; i < inlines->length(); i++) { 359 CallGenerator* cg = inlines->at(i); 360 CallNode* call = cg->call_node(); 361 if (shift > 0) { 362 inlines->at_put(i-shift, cg); 363 } 364 if (!useful.member(call)) { 365 shift++; 366 } 367 } 368 inlines->trunc_to(inlines->length()-shift); 369 } 370 371 // Disconnect all useless nodes by disconnecting those at the boundary. 372 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 373 uint next = 0; 374 while (next < useful.size()) { 375 Node *n = useful.at(next++); 376 if (n->is_SafePoint()) { 377 // We're done with a parsing phase. Replaced nodes are not valid 378 // beyond that point. 379 n->as_SafePoint()->delete_replaced_nodes(); 380 } 381 // Use raw traversal of out edges since this code removes out edges 382 int max = n->outcnt(); 383 for (int j = 0; j < max; ++j) { 384 Node* child = n->raw_out(j); 385 if (! useful.member(child)) { 386 assert(!child->is_top() || child != top(), 387 "If top is cached in Compile object it is in useful list"); 388 // Only need to remove this out-edge to the useless node 389 n->raw_del_out(j); 390 --j; 391 --max; 392 } 393 } 394 if (n->outcnt() == 1 && n->has_special_unique_user()) { 395 record_for_igvn(n->unique_out()); 396 } 397 } 398 // Remove useless macro and predicate opaq nodes 399 for (int i = C->macro_count()-1; i >= 0; i--) { 400 Node* n = C->macro_node(i); 401 if (!useful.member(n)) { 402 remove_macro_node(n); 403 } 404 } 405 // Remove useless expensive node 406 for (int i = C->expensive_count()-1; i >= 0; i--) { 407 Node* n = C->expensive_node(i); 408 if (!useful.member(n)) { 409 remove_expensive_node(n); 410 } 411 } 412 // clean up the late inline lists 413 remove_useless_late_inlines(&_string_late_inlines, useful); 414 remove_useless_late_inlines(&_boxing_late_inlines, useful); 415 remove_useless_late_inlines(&_late_inlines, useful); 416 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 417 } 418 419 //------------------------------frame_size_in_words----------------------------- 420 // frame_slots in units of words 421 int Compile::frame_size_in_words() const { 422 // shift is 0 in LP32 and 1 in LP64 423 const int shift = (LogBytesPerWord - LogBytesPerInt); 424 int words = _frame_slots >> shift; 425 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 426 return words; 427 } 428 429 // To bang the stack of this compiled method we use the stack size 430 // that the interpreter would need in case of a deoptimization. This 431 // removes the need to bang the stack in the deoptimization blob which 432 // in turn simplifies stack overflow handling. 433 int Compile::bang_size_in_bytes() const { 434 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 435 } 436 437 // ============================================================================ 438 //------------------------------CompileWrapper--------------------------------- 439 class CompileWrapper : public StackObj { 440 Compile *const _compile; 441 public: 442 CompileWrapper(Compile* compile); 443 444 ~CompileWrapper(); 445 }; 446 447 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 448 // the Compile* pointer is stored in the current ciEnv: 449 ciEnv* env = compile->env(); 450 assert(env == ciEnv::current(), "must already be a ciEnv active"); 451 assert(env->compiler_data() == NULL, "compile already active?"); 452 env->set_compiler_data(compile); 453 assert(compile == Compile::current(), "sanity"); 454 455 compile->set_type_dict(NULL); 456 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 457 compile->clone_map().set_clone_idx(0); 458 compile->set_type_hwm(NULL); 459 compile->set_type_last_size(0); 460 compile->set_last_tf(NULL, NULL); 461 compile->set_indexSet_arena(NULL); 462 compile->set_indexSet_free_block_list(NULL); 463 compile->init_type_arena(); 464 Type::Initialize(compile); 465 _compile->set_scratch_buffer_blob(NULL); 466 _compile->begin_method(); 467 _compile->clone_map().set_debug(_compile->has_method() && _compile->method_has_option(_compile->clone_map().debug_option_name)); 468 } 469 CompileWrapper::~CompileWrapper() { 470 _compile->end_method(); 471 if (_compile->scratch_buffer_blob() != NULL) 472 BufferBlob::free(_compile->scratch_buffer_blob()); 473 _compile->env()->set_compiler_data(NULL); 474 } 475 476 477 //----------------------------print_compile_messages--------------------------- 478 void Compile::print_compile_messages() { 479 #ifndef PRODUCT 480 // Check if recompiling 481 if (_subsume_loads == false && PrintOpto) { 482 // Recompiling without allowing machine instructions to subsume loads 483 tty->print_cr("*********************************************************"); 484 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 485 tty->print_cr("*********************************************************"); 486 } 487 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 488 // Recompiling without escape analysis 489 tty->print_cr("*********************************************************"); 490 tty->print_cr("** Bailout: Recompile without escape analysis **"); 491 tty->print_cr("*********************************************************"); 492 } 493 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 494 // Recompiling without boxing elimination 495 tty->print_cr("*********************************************************"); 496 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 497 tty->print_cr("*********************************************************"); 498 } 499 if (env()->break_at_compile()) { 500 // Open the debugger when compiling this method. 501 tty->print("### Breaking when compiling: "); 502 method()->print_short_name(); 503 tty->cr(); 504 BREAKPOINT; 505 } 506 507 if( PrintOpto ) { 508 if (is_osr_compilation()) { 509 tty->print("[OSR]%3d", _compile_id); 510 } else { 511 tty->print("%3d", _compile_id); 512 } 513 } 514 #endif 515 } 516 517 518 //-----------------------init_scratch_buffer_blob------------------------------ 519 // Construct a temporary BufferBlob and cache it for this compile. 520 void Compile::init_scratch_buffer_blob(int const_size) { 521 // If there is already a scratch buffer blob allocated and the 522 // constant section is big enough, use it. Otherwise free the 523 // current and allocate a new one. 524 BufferBlob* blob = scratch_buffer_blob(); 525 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 526 // Use the current blob. 527 } else { 528 if (blob != NULL) { 529 BufferBlob::free(blob); 530 } 531 532 ResourceMark rm; 533 _scratch_const_size = const_size; 534 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 535 blob = BufferBlob::create("Compile::scratch_buffer", size); 536 // Record the buffer blob for next time. 537 set_scratch_buffer_blob(blob); 538 // Have we run out of code space? 539 if (scratch_buffer_blob() == NULL) { 540 // Let CompilerBroker disable further compilations. 541 record_failure("Not enough space for scratch buffer in CodeCache"); 542 return; 543 } 544 } 545 546 // Initialize the relocation buffers 547 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 548 set_scratch_locs_memory(locs_buf); 549 } 550 551 552 //-----------------------scratch_emit_size------------------------------------- 553 // Helper function that computes size by emitting code 554 uint Compile::scratch_emit_size(const Node* n) { 555 // Start scratch_emit_size section. 556 set_in_scratch_emit_size(true); 557 558 // Emit into a trash buffer and count bytes emitted. 559 // This is a pretty expensive way to compute a size, 560 // but it works well enough if seldom used. 561 // All common fixed-size instructions are given a size 562 // method by the AD file. 563 // Note that the scratch buffer blob and locs memory are 564 // allocated at the beginning of the compile task, and 565 // may be shared by several calls to scratch_emit_size. 566 // The allocation of the scratch buffer blob is particularly 567 // expensive, since it has to grab the code cache lock. 568 BufferBlob* blob = this->scratch_buffer_blob(); 569 assert(blob != NULL, "Initialize BufferBlob at start"); 570 assert(blob->size() > MAX_inst_size, "sanity"); 571 relocInfo* locs_buf = scratch_locs_memory(); 572 address blob_begin = blob->content_begin(); 573 address blob_end = (address)locs_buf; 574 assert(blob->content_contains(blob_end), "sanity"); 575 CodeBuffer buf(blob_begin, blob_end - blob_begin); 576 buf.initialize_consts_size(_scratch_const_size); 577 buf.initialize_stubs_size(MAX_stubs_size); 578 assert(locs_buf != NULL, "sanity"); 579 int lsize = MAX_locs_size / 3; 580 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 581 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 582 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 583 584 // Do the emission. 585 586 Label fakeL; // Fake label for branch instructions. 587 Label* saveL = NULL; 588 uint save_bnum = 0; 589 bool is_branch = n->is_MachBranch(); 590 if (is_branch) { 591 MacroAssembler masm(&buf); 592 masm.bind(fakeL); 593 n->as_MachBranch()->save_label(&saveL, &save_bnum); 594 n->as_MachBranch()->label_set(&fakeL, 0); 595 } 596 n->emit(buf, this->regalloc()); 597 598 // Emitting into the scratch buffer should not fail 599 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 600 601 if (is_branch) // Restore label. 602 n->as_MachBranch()->label_set(saveL, save_bnum); 603 604 // End scratch_emit_size section. 605 set_in_scratch_emit_size(false); 606 607 return buf.insts_size(); 608 } 609 610 611 // ============================================================================ 612 //------------------------------Compile standard------------------------------- 613 debug_only( int Compile::_debug_idx = 100000; ) 614 615 // Compile a method. entry_bci is -1 for normal compilations and indicates 616 // the continuation bci for on stack replacement. 617 618 619 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 620 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 621 : Phase(Compiler), 622 _env(ci_env), 623 _log(ci_env->log()), 624 _compile_id(ci_env->compile_id()), 625 _save_argument_registers(false), 626 _stub_name(NULL), 627 _stub_function(NULL), 628 _stub_entry_point(NULL), 629 _method(target), 630 _entry_bci(osr_bci), 631 _initial_gvn(NULL), 632 _for_igvn(NULL), 633 _warm_calls(NULL), 634 _subsume_loads(subsume_loads), 635 _do_escape_analysis(do_escape_analysis), 636 _eliminate_boxing(eliminate_boxing), 637 _failure_reason(NULL), 638 _code_buffer("Compile::Fill_buffer"), 639 _orig_pc_slot(0), 640 _orig_pc_slot_offset_in_bytes(0), 641 _has_method_handle_invokes(false), 642 _mach_constant_base_node(NULL), 643 _node_bundling_limit(0), 644 _node_bundling_base(NULL), 645 _java_calls(0), 646 _inner_loops(0), 647 _scratch_const_size(-1), 648 _in_scratch_emit_size(false), 649 _dead_node_list(comp_arena()), 650 _dead_node_count(0), 651 #ifndef PRODUCT 652 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 653 _in_dump_cnt(0), 654 _printer(IdealGraphPrinter::printer()), 655 #endif 656 _congraph(NULL), 657 _comp_arena(mtCompiler), 658 _node_arena(mtCompiler), 659 _old_arena(mtCompiler), 660 _Compile_types(mtCompiler), 661 _replay_inline_data(NULL), 662 _late_inlines(comp_arena(), 2, 0, NULL), 663 _string_late_inlines(comp_arena(), 2, 0, NULL), 664 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 665 _late_inlines_pos(0), 666 _number_of_mh_late_inlines(0), 667 _inlining_progress(false), 668 _inlining_incrementally(false), 669 _print_inlining_list(NULL), 670 _print_inlining_stream(NULL), 671 _print_inlining_idx(0), 672 _print_inlining_output(NULL), 673 _interpreter_frame_size(0), 674 _max_node_limit(MaxNodeLimit) { 675 C = this; 676 677 CompileWrapper cw(this); 678 679 if (CITimeVerbose) { 680 tty->print(" "); 681 target->holder()->name()->print(); 682 tty->print("."); 683 target->print_short_name(); 684 tty->print(" "); 685 } 686 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 687 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 688 689 #ifndef PRODUCT 690 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 691 if (!print_opto_assembly) { 692 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 693 if (print_assembly && !Disassembler::can_decode()) { 694 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 695 print_opto_assembly = true; 696 } 697 } 698 set_print_assembly(print_opto_assembly); 699 set_parsed_irreducible_loop(false); 700 701 if (method()->has_option("ReplayInline")) { 702 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 703 } 704 #endif 705 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 706 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 707 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 708 709 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 710 // Make sure the method being compiled gets its own MDO, 711 // so we can at least track the decompile_count(). 712 // Need MDO to record RTM code generation state. 713 method()->ensure_method_data(); 714 } 715 716 Init(::AliasLevel); 717 718 719 print_compile_messages(); 720 721 _ilt = InlineTree::build_inline_tree_root(); 722 723 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 724 assert(num_alias_types() >= AliasIdxRaw, ""); 725 726 #define MINIMUM_NODE_HASH 1023 727 // Node list that Iterative GVN will start with 728 Unique_Node_List for_igvn(comp_arena()); 729 set_for_igvn(&for_igvn); 730 731 // GVN that will be run immediately on new nodes 732 uint estimated_size = method()->code_size()*4+64; 733 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 734 PhaseGVN gvn(node_arena(), estimated_size); 735 set_initial_gvn(&gvn); 736 737 print_inlining_init(); 738 { // Scope for timing the parser 739 TracePhase tp("parse", &timers[_t_parser]); 740 741 // Put top into the hash table ASAP. 742 initial_gvn()->transform_no_reclaim(top()); 743 744 // Set up tf(), start(), and find a CallGenerator. 745 CallGenerator* cg = NULL; 746 if (is_osr_compilation()) { 747 const TypeTuple *domain = StartOSRNode::osr_domain(); 748 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 749 init_tf(TypeFunc::make(domain, range)); 750 StartNode* s = new StartOSRNode(root(), domain); 751 initial_gvn()->set_type_bottom(s); 752 init_start(s); 753 cg = CallGenerator::for_osr(method(), entry_bci()); 754 } else { 755 // Normal case. 756 init_tf(TypeFunc::make(method())); 757 StartNode* s = new StartNode(root(), tf()->domain()); 758 initial_gvn()->set_type_bottom(s); 759 init_start(s); 760 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 761 // With java.lang.ref.reference.get() we must go through the 762 // intrinsic when G1 is enabled - even when get() is the root 763 // method of the compile - so that, if necessary, the value in 764 // the referent field of the reference object gets recorded by 765 // the pre-barrier code. 766 // Specifically, if G1 is enabled, the value in the referent 767 // field is recorded by the G1 SATB pre barrier. This will 768 // result in the referent being marked live and the reference 769 // object removed from the list of discovered references during 770 // reference processing. 771 cg = find_intrinsic(method(), false); 772 } 773 if (cg == NULL) { 774 float past_uses = method()->interpreter_invocation_count(); 775 float expected_uses = past_uses; 776 cg = CallGenerator::for_inline(method(), expected_uses); 777 } 778 } 779 if (failing()) return; 780 if (cg == NULL) { 781 record_method_not_compilable_all_tiers("cannot parse method"); 782 return; 783 } 784 JVMState* jvms = build_start_state(start(), tf()); 785 if ((jvms = cg->generate(jvms)) == NULL) { 786 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 787 record_method_not_compilable("method parse failed"); 788 } 789 return; 790 } 791 GraphKit kit(jvms); 792 793 if (!kit.stopped()) { 794 // Accept return values, and transfer control we know not where. 795 // This is done by a special, unique ReturnNode bound to root. 796 return_values(kit.jvms()); 797 } 798 799 if (kit.has_exceptions()) { 800 // Any exceptions that escape from this call must be rethrown 801 // to whatever caller is dynamically above us on the stack. 802 // This is done by a special, unique RethrowNode bound to root. 803 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 804 } 805 806 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 807 808 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 809 inline_string_calls(true); 810 } 811 812 if (failing()) return; 813 814 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 815 816 // Remove clutter produced by parsing. 817 if (!failing()) { 818 ResourceMark rm; 819 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 820 } 821 } 822 823 // Note: Large methods are capped off in do_one_bytecode(). 824 if (failing()) return; 825 826 // After parsing, node notes are no longer automagic. 827 // They must be propagated by register_new_node_with_optimizer(), 828 // clone(), or the like. 829 set_default_node_notes(NULL); 830 831 for (;;) { 832 int successes = Inline_Warm(); 833 if (failing()) return; 834 if (successes == 0) break; 835 } 836 837 // Drain the list. 838 Finish_Warm(); 839 #ifndef PRODUCT 840 if (_printer && _printer->should_print(_method)) { 841 _printer->print_inlining(this); 842 } 843 #endif 844 845 if (failing()) return; 846 NOT_PRODUCT( verify_graph_edges(); ) 847 848 // Now optimize 849 Optimize(); 850 if (failing()) return; 851 NOT_PRODUCT( verify_graph_edges(); ) 852 853 #ifndef PRODUCT 854 if (PrintIdeal) { 855 ttyLocker ttyl; // keep the following output all in one block 856 // This output goes directly to the tty, not the compiler log. 857 // To enable tools to match it up with the compilation activity, 858 // be sure to tag this tty output with the compile ID. 859 if (xtty != NULL) { 860 xtty->head("ideal compile_id='%d'%s", compile_id(), 861 is_osr_compilation() ? " compile_kind='osr'" : 862 ""); 863 } 864 root()->dump(9999); 865 if (xtty != NULL) { 866 xtty->tail("ideal"); 867 } 868 } 869 #endif 870 871 NOT_PRODUCT( verify_barriers(); ) 872 873 // Dump compilation data to replay it. 874 if (method()->has_option("DumpReplay")) { 875 env()->dump_replay_data(_compile_id); 876 } 877 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 878 env()->dump_inline_data(_compile_id); 879 } 880 881 // Now that we know the size of all the monitors we can add a fixed slot 882 // for the original deopt pc. 883 884 _orig_pc_slot = fixed_slots(); 885 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 886 set_fixed_slots(next_slot); 887 888 // Compute when to use implicit null checks. Used by matching trap based 889 // nodes and NullCheck optimization. 890 set_allowed_deopt_reasons(); 891 892 // Now generate code 893 Code_Gen(); 894 if (failing()) return; 895 896 // Check if we want to skip execution of all compiled code. 897 { 898 #ifndef PRODUCT 899 if (OptoNoExecute) { 900 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 901 return; 902 } 903 #endif 904 TracePhase tp("install_code", &timers[_t_registerMethod]); 905 906 if (is_osr_compilation()) { 907 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 908 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 909 } else { 910 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 911 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 912 } 913 914 env()->register_method(_method, _entry_bci, 915 &_code_offsets, 916 _orig_pc_slot_offset_in_bytes, 917 code_buffer(), 918 frame_size_in_words(), _oop_map_set, 919 &_handler_table, &_inc_table, 920 compiler, 921 env()->comp_level(), 922 has_unsafe_access(), 923 SharedRuntime::is_wide_vector(max_vector_size()), 924 rtm_state() 925 ); 926 927 if (log() != NULL) // Print code cache state into compiler log 928 log()->code_cache_state(); 929 } 930 } 931 932 //------------------------------Compile---------------------------------------- 933 // Compile a runtime stub 934 Compile::Compile( ciEnv* ci_env, 935 TypeFunc_generator generator, 936 address stub_function, 937 const char *stub_name, 938 int is_fancy_jump, 939 bool pass_tls, 940 bool save_arg_registers, 941 bool return_pc ) 942 : Phase(Compiler), 943 _env(ci_env), 944 _log(ci_env->log()), 945 _compile_id(0), 946 _save_argument_registers(save_arg_registers), 947 _method(NULL), 948 _stub_name(stub_name), 949 _stub_function(stub_function), 950 _stub_entry_point(NULL), 951 _entry_bci(InvocationEntryBci), 952 _initial_gvn(NULL), 953 _for_igvn(NULL), 954 _warm_calls(NULL), 955 _orig_pc_slot(0), 956 _orig_pc_slot_offset_in_bytes(0), 957 _subsume_loads(true), 958 _do_escape_analysis(false), 959 _eliminate_boxing(false), 960 _failure_reason(NULL), 961 _code_buffer("Compile::Fill_buffer"), 962 _has_method_handle_invokes(false), 963 _mach_constant_base_node(NULL), 964 _node_bundling_limit(0), 965 _node_bundling_base(NULL), 966 _java_calls(0), 967 _inner_loops(0), 968 #ifndef PRODUCT 969 _trace_opto_output(TraceOptoOutput), 970 _in_dump_cnt(0), 971 _printer(NULL), 972 #endif 973 _comp_arena(mtCompiler), 974 _node_arena(mtCompiler), 975 _old_arena(mtCompiler), 976 _Compile_types(mtCompiler), 977 _dead_node_list(comp_arena()), 978 _dead_node_count(0), 979 _congraph(NULL), 980 _replay_inline_data(NULL), 981 _number_of_mh_late_inlines(0), 982 _inlining_progress(false), 983 _inlining_incrementally(false), 984 _print_inlining_list(NULL), 985 _print_inlining_stream(NULL), 986 _print_inlining_idx(0), 987 _print_inlining_output(NULL), 988 _allowed_reasons(0), 989 _interpreter_frame_size(0), 990 _max_node_limit(MaxNodeLimit) { 991 C = this; 992 993 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 994 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 995 996 #ifndef PRODUCT 997 set_print_assembly(PrintFrameConverterAssembly); 998 set_parsed_irreducible_loop(false); 999 #endif 1000 set_has_irreducible_loop(false); // no loops 1001 1002 CompileWrapper cw(this); 1003 Init(/*AliasLevel=*/ 0); 1004 init_tf((*generator)()); 1005 1006 { 1007 // The following is a dummy for the sake of GraphKit::gen_stub 1008 Unique_Node_List for_igvn(comp_arena()); 1009 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1010 PhaseGVN gvn(Thread::current()->resource_area(),255); 1011 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1012 gvn.transform_no_reclaim(top()); 1013 1014 GraphKit kit; 1015 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1016 } 1017 1018 NOT_PRODUCT( verify_graph_edges(); ) 1019 Code_Gen(); 1020 if (failing()) return; 1021 1022 1023 // Entry point will be accessed using compile->stub_entry_point(); 1024 if (code_buffer() == NULL) { 1025 Matcher::soft_match_failure(); 1026 } else { 1027 if (PrintAssembly && (WizardMode || Verbose)) 1028 tty->print_cr("### Stub::%s", stub_name); 1029 1030 if (!failing()) { 1031 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1032 1033 // Make the NMethod 1034 // For now we mark the frame as never safe for profile stackwalking 1035 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1036 code_buffer(), 1037 CodeOffsets::frame_never_safe, 1038 // _code_offsets.value(CodeOffsets::Frame_Complete), 1039 frame_size_in_words(), 1040 _oop_map_set, 1041 save_arg_registers); 1042 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1043 1044 _stub_entry_point = rs->entry_point(); 1045 } 1046 } 1047 } 1048 1049 //------------------------------Init------------------------------------------- 1050 // Prepare for a single compilation 1051 void Compile::Init(int aliaslevel) { 1052 _unique = 0; 1053 _regalloc = NULL; 1054 1055 _tf = NULL; // filled in later 1056 _top = NULL; // cached later 1057 _matcher = NULL; // filled in later 1058 _cfg = NULL; // filled in later 1059 1060 set_24_bit_selection_and_mode(Use24BitFP, false); 1061 1062 _node_note_array = NULL; 1063 _default_node_notes = NULL; 1064 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1065 1066 _immutable_memory = NULL; // filled in at first inquiry 1067 1068 // Globally visible Nodes 1069 // First set TOP to NULL to give safe behavior during creation of RootNode 1070 set_cached_top_node(NULL); 1071 set_root(new RootNode()); 1072 // Now that you have a Root to point to, create the real TOP 1073 set_cached_top_node( new ConNode(Type::TOP) ); 1074 set_recent_alloc(NULL, NULL); 1075 1076 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1077 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1078 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1079 env()->set_dependencies(new Dependencies(env())); 1080 1081 _fixed_slots = 0; 1082 set_has_split_ifs(false); 1083 set_has_loops(has_method() && method()->has_loops()); // first approximation 1084 set_has_stringbuilder(false); 1085 set_has_boxed_value(false); 1086 _trap_can_recompile = false; // no traps emitted yet 1087 _major_progress = true; // start out assuming good things will happen 1088 set_has_unsafe_access(false); 1089 set_max_vector_size(0); 1090 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1091 set_decompile_count(0); 1092 1093 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1094 set_num_loop_opts(LoopOptsCount); 1095 set_do_inlining(Inline); 1096 set_max_inline_size(MaxInlineSize); 1097 set_freq_inline_size(FreqInlineSize); 1098 set_do_scheduling(OptoScheduling); 1099 set_do_count_invocations(false); 1100 set_do_method_data_update(false); 1101 1102 set_do_vector_loop(false); 1103 1104 if (AllowVectorizeOnDemand) { 1105 if (has_method() && (method()->has_option("Vectorize") || method()->has_option("VectorizeDebug"))) { 1106 set_do_vector_loop(true); 1107 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());}) 1108 } else if (has_method() && method()->name() != 0 && 1109 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1110 set_do_vector_loop(true); 1111 } 1112 } 1113 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1114 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());}) 1115 1116 set_age_code(has_method() && method()->profile_aging()); 1117 set_rtm_state(NoRTM); // No RTM lock eliding by default 1118 method_has_option_value("MaxNodeLimit", _max_node_limit); 1119 #if INCLUDE_RTM_OPT 1120 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1121 int rtm_state = method()->method_data()->rtm_state(); 1122 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1123 // Don't generate RTM lock eliding code. 1124 set_rtm_state(NoRTM); 1125 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1126 // Generate RTM lock eliding code without abort ratio calculation code. 1127 set_rtm_state(UseRTM); 1128 } else if (UseRTMDeopt) { 1129 // Generate RTM lock eliding code and include abort ratio calculation 1130 // code if UseRTMDeopt is on. 1131 set_rtm_state(ProfileRTM); 1132 } 1133 } 1134 #endif 1135 if (debug_info()->recording_non_safepoints()) { 1136 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1137 (comp_arena(), 8, 0, NULL)); 1138 set_default_node_notes(Node_Notes::make(this)); 1139 } 1140 1141 // // -- Initialize types before each compile -- 1142 // // Update cached type information 1143 // if( _method && _method->constants() ) 1144 // Type::update_loaded_types(_method, _method->constants()); 1145 1146 // Init alias_type map. 1147 if (!_do_escape_analysis && aliaslevel == 3) 1148 aliaslevel = 2; // No unique types without escape analysis 1149 _AliasLevel = aliaslevel; 1150 const int grow_ats = 16; 1151 _max_alias_types = grow_ats; 1152 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1153 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1154 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1155 { 1156 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1157 } 1158 // Initialize the first few types. 1159 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1160 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1161 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1162 _num_alias_types = AliasIdxRaw+1; 1163 // Zero out the alias type cache. 1164 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1165 // A NULL adr_type hits in the cache right away. Preload the right answer. 1166 probe_alias_cache(NULL)->_index = AliasIdxTop; 1167 1168 _intrinsics = NULL; 1169 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1170 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1171 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1172 register_library_intrinsics(); 1173 } 1174 1175 //---------------------------init_start---------------------------------------- 1176 // Install the StartNode on this compile object. 1177 void Compile::init_start(StartNode* s) { 1178 if (failing()) 1179 return; // already failing 1180 assert(s == start(), ""); 1181 } 1182 1183 /** 1184 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1185 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1186 * the ideal graph. 1187 */ 1188 StartNode* Compile::start() const { 1189 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 1190 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1191 Node* start = root()->fast_out(i); 1192 if (start->is_Start()) { 1193 return start->as_Start(); 1194 } 1195 } 1196 fatal("Did not find Start node!"); 1197 return NULL; 1198 } 1199 1200 //-------------------------------immutable_memory------------------------------------- 1201 // Access immutable memory 1202 Node* Compile::immutable_memory() { 1203 if (_immutable_memory != NULL) { 1204 return _immutable_memory; 1205 } 1206 StartNode* s = start(); 1207 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1208 Node *p = s->fast_out(i); 1209 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1210 _immutable_memory = p; 1211 return _immutable_memory; 1212 } 1213 } 1214 ShouldNotReachHere(); 1215 return NULL; 1216 } 1217 1218 //----------------------set_cached_top_node------------------------------------ 1219 // Install the cached top node, and make sure Node::is_top works correctly. 1220 void Compile::set_cached_top_node(Node* tn) { 1221 if (tn != NULL) verify_top(tn); 1222 Node* old_top = _top; 1223 _top = tn; 1224 // Calling Node::setup_is_top allows the nodes the chance to adjust 1225 // their _out arrays. 1226 if (_top != NULL) _top->setup_is_top(); 1227 if (old_top != NULL) old_top->setup_is_top(); 1228 assert(_top == NULL || top()->is_top(), ""); 1229 } 1230 1231 #ifdef ASSERT 1232 uint Compile::count_live_nodes_by_graph_walk() { 1233 Unique_Node_List useful(comp_arena()); 1234 // Get useful node list by walking the graph. 1235 identify_useful_nodes(useful); 1236 return useful.size(); 1237 } 1238 1239 void Compile::print_missing_nodes() { 1240 1241 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1242 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1243 return; 1244 } 1245 1246 // This is an expensive function. It is executed only when the user 1247 // specifies VerifyIdealNodeCount option or otherwise knows the 1248 // additional work that needs to be done to identify reachable nodes 1249 // by walking the flow graph and find the missing ones using 1250 // _dead_node_list. 1251 1252 Unique_Node_List useful(comp_arena()); 1253 // Get useful node list by walking the graph. 1254 identify_useful_nodes(useful); 1255 1256 uint l_nodes = C->live_nodes(); 1257 uint l_nodes_by_walk = useful.size(); 1258 1259 if (l_nodes != l_nodes_by_walk) { 1260 if (_log != NULL) { 1261 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1262 _log->stamp(); 1263 _log->end_head(); 1264 } 1265 VectorSet& useful_member_set = useful.member_set(); 1266 int last_idx = l_nodes_by_walk; 1267 for (int i = 0; i < last_idx; i++) { 1268 if (useful_member_set.test(i)) { 1269 if (_dead_node_list.test(i)) { 1270 if (_log != NULL) { 1271 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1272 } 1273 if (PrintIdealNodeCount) { 1274 // Print the log message to tty 1275 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1276 useful.at(i)->dump(); 1277 } 1278 } 1279 } 1280 else if (! _dead_node_list.test(i)) { 1281 if (_log != NULL) { 1282 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1283 } 1284 if (PrintIdealNodeCount) { 1285 // Print the log message to tty 1286 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1287 } 1288 } 1289 } 1290 if (_log != NULL) { 1291 _log->tail("mismatched_nodes"); 1292 } 1293 } 1294 } 1295 void Compile::record_modified_node(Node* n) { 1296 if (_modified_nodes != NULL && !_inlining_incrementally && 1297 n->outcnt() != 0 && !n->is_Con()) { 1298 _modified_nodes->push(n); 1299 } 1300 } 1301 1302 void Compile::remove_modified_node(Node* n) { 1303 if (_modified_nodes != NULL) { 1304 _modified_nodes->remove(n); 1305 } 1306 } 1307 #endif 1308 1309 #ifndef PRODUCT 1310 void Compile::verify_top(Node* tn) const { 1311 if (tn != NULL) { 1312 assert(tn->is_Con(), "top node must be a constant"); 1313 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1314 assert(tn->in(0) != NULL, "must have live top node"); 1315 } 1316 } 1317 #endif 1318 1319 1320 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1321 1322 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1323 guarantee(arr != NULL, ""); 1324 int num_blocks = arr->length(); 1325 if (grow_by < num_blocks) grow_by = num_blocks; 1326 int num_notes = grow_by * _node_notes_block_size; 1327 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1328 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1329 while (num_notes > 0) { 1330 arr->append(notes); 1331 notes += _node_notes_block_size; 1332 num_notes -= _node_notes_block_size; 1333 } 1334 assert(num_notes == 0, "exact multiple, please"); 1335 } 1336 1337 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1338 if (source == NULL || dest == NULL) return false; 1339 1340 if (dest->is_Con()) 1341 return false; // Do not push debug info onto constants. 1342 1343 #ifdef ASSERT 1344 // Leave a bread crumb trail pointing to the original node: 1345 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1346 dest->set_debug_orig(source); 1347 } 1348 #endif 1349 1350 if (node_note_array() == NULL) 1351 return false; // Not collecting any notes now. 1352 1353 // This is a copy onto a pre-existing node, which may already have notes. 1354 // If both nodes have notes, do not overwrite any pre-existing notes. 1355 Node_Notes* source_notes = node_notes_at(source->_idx); 1356 if (source_notes == NULL || source_notes->is_clear()) return false; 1357 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1358 if (dest_notes == NULL || dest_notes->is_clear()) { 1359 return set_node_notes_at(dest->_idx, source_notes); 1360 } 1361 1362 Node_Notes merged_notes = (*source_notes); 1363 // The order of operations here ensures that dest notes will win... 1364 merged_notes.update_from(dest_notes); 1365 return set_node_notes_at(dest->_idx, &merged_notes); 1366 } 1367 1368 1369 //--------------------------allow_range_check_smearing------------------------- 1370 // Gating condition for coalescing similar range checks. 1371 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1372 // single covering check that is at least as strong as any of them. 1373 // If the optimization succeeds, the simplified (strengthened) range check 1374 // will always succeed. If it fails, we will deopt, and then give up 1375 // on the optimization. 1376 bool Compile::allow_range_check_smearing() const { 1377 // If this method has already thrown a range-check, 1378 // assume it was because we already tried range smearing 1379 // and it failed. 1380 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1381 return !already_trapped; 1382 } 1383 1384 1385 //------------------------------flatten_alias_type----------------------------- 1386 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1387 int offset = tj->offset(); 1388 TypePtr::PTR ptr = tj->ptr(); 1389 1390 // Known instance (scalarizable allocation) alias only with itself. 1391 bool is_known_inst = tj->isa_oopptr() != NULL && 1392 tj->is_oopptr()->is_known_instance(); 1393 1394 // Process weird unsafe references. 1395 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1396 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1397 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1398 tj = TypeOopPtr::BOTTOM; 1399 ptr = tj->ptr(); 1400 offset = tj->offset(); 1401 } 1402 1403 // Array pointers need some flattening 1404 const TypeAryPtr *ta = tj->isa_aryptr(); 1405 if (ta && ta->is_stable()) { 1406 // Erase stability property for alias analysis. 1407 tj = ta = ta->cast_to_stable(false); 1408 } 1409 if( ta && is_known_inst ) { 1410 if ( offset != Type::OffsetBot && 1411 offset > arrayOopDesc::length_offset_in_bytes() ) { 1412 offset = Type::OffsetBot; // Flatten constant access into array body only 1413 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1414 } 1415 } else if( ta && _AliasLevel >= 2 ) { 1416 // For arrays indexed by constant indices, we flatten the alias 1417 // space to include all of the array body. Only the header, klass 1418 // and array length can be accessed un-aliased. 1419 if( offset != Type::OffsetBot ) { 1420 if( ta->const_oop() ) { // MethodData* or Method* 1421 offset = Type::OffsetBot; // Flatten constant access into array body 1422 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1423 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1424 // range is OK as-is. 1425 tj = ta = TypeAryPtr::RANGE; 1426 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1427 tj = TypeInstPtr::KLASS; // all klass loads look alike 1428 ta = TypeAryPtr::RANGE; // generic ignored junk 1429 ptr = TypePtr::BotPTR; 1430 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1431 tj = TypeInstPtr::MARK; 1432 ta = TypeAryPtr::RANGE; // generic ignored junk 1433 ptr = TypePtr::BotPTR; 1434 } else { // Random constant offset into array body 1435 offset = Type::OffsetBot; // Flatten constant access into array body 1436 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1437 } 1438 } 1439 // Arrays of fixed size alias with arrays of unknown size. 1440 if (ta->size() != TypeInt::POS) { 1441 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1442 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1443 } 1444 // Arrays of known objects become arrays of unknown objects. 1445 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1446 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1447 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1448 } 1449 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1450 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1451 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1452 } 1453 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1454 // cannot be distinguished by bytecode alone. 1455 if (ta->elem() == TypeInt::BOOL) { 1456 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1457 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1458 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1459 } 1460 // During the 2nd round of IterGVN, NotNull castings are removed. 1461 // Make sure the Bottom and NotNull variants alias the same. 1462 // Also, make sure exact and non-exact variants alias the same. 1463 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1464 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1465 } 1466 } 1467 1468 // Oop pointers need some flattening 1469 const TypeInstPtr *to = tj->isa_instptr(); 1470 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1471 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1472 if( ptr == TypePtr::Constant ) { 1473 if (to->klass() != ciEnv::current()->Class_klass() || 1474 offset < k->size_helper() * wordSize) { 1475 // No constant oop pointers (such as Strings); they alias with 1476 // unknown strings. 1477 assert(!is_known_inst, "not scalarizable allocation"); 1478 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1479 } 1480 } else if( is_known_inst ) { 1481 tj = to; // Keep NotNull and klass_is_exact for instance type 1482 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1483 // During the 2nd round of IterGVN, NotNull castings are removed. 1484 // Make sure the Bottom and NotNull variants alias the same. 1485 // Also, make sure exact and non-exact variants alias the same. 1486 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1487 } 1488 if (to->speculative() != NULL) { 1489 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1490 } 1491 // Canonicalize the holder of this field 1492 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1493 // First handle header references such as a LoadKlassNode, even if the 1494 // object's klass is unloaded at compile time (4965979). 1495 if (!is_known_inst) { // Do it only for non-instance types 1496 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1497 } 1498 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1499 // Static fields are in the space above the normal instance 1500 // fields in the java.lang.Class instance. 1501 if (to->klass() != ciEnv::current()->Class_klass()) { 1502 to = NULL; 1503 tj = TypeOopPtr::BOTTOM; 1504 offset = tj->offset(); 1505 } 1506 } else { 1507 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1508 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1509 if( is_known_inst ) { 1510 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1511 } else { 1512 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1513 } 1514 } 1515 } 1516 } 1517 1518 // Klass pointers to object array klasses need some flattening 1519 const TypeKlassPtr *tk = tj->isa_klassptr(); 1520 if( tk ) { 1521 // If we are referencing a field within a Klass, we need 1522 // to assume the worst case of an Object. Both exact and 1523 // inexact types must flatten to the same alias class so 1524 // use NotNull as the PTR. 1525 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1526 1527 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1528 TypeKlassPtr::OBJECT->klass(), 1529 offset); 1530 } 1531 1532 ciKlass* klass = tk->klass(); 1533 if( klass->is_obj_array_klass() ) { 1534 ciKlass* k = TypeAryPtr::OOPS->klass(); 1535 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1536 k = TypeInstPtr::BOTTOM->klass(); 1537 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1538 } 1539 1540 // Check for precise loads from the primary supertype array and force them 1541 // to the supertype cache alias index. Check for generic array loads from 1542 // the primary supertype array and also force them to the supertype cache 1543 // alias index. Since the same load can reach both, we need to merge 1544 // these 2 disparate memories into the same alias class. Since the 1545 // primary supertype array is read-only, there's no chance of confusion 1546 // where we bypass an array load and an array store. 1547 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1548 if (offset == Type::OffsetBot || 1549 (offset >= primary_supers_offset && 1550 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1551 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1552 offset = in_bytes(Klass::secondary_super_cache_offset()); 1553 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1554 } 1555 } 1556 1557 // Flatten all Raw pointers together. 1558 if (tj->base() == Type::RawPtr) 1559 tj = TypeRawPtr::BOTTOM; 1560 1561 if (tj->base() == Type::AnyPtr) 1562 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1563 1564 // Flatten all to bottom for now 1565 switch( _AliasLevel ) { 1566 case 0: 1567 tj = TypePtr::BOTTOM; 1568 break; 1569 case 1: // Flatten to: oop, static, field or array 1570 switch (tj->base()) { 1571 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1572 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1573 case Type::AryPtr: // do not distinguish arrays at all 1574 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1575 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1576 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1577 default: ShouldNotReachHere(); 1578 } 1579 break; 1580 case 2: // No collapsing at level 2; keep all splits 1581 case 3: // No collapsing at level 3; keep all splits 1582 break; 1583 default: 1584 Unimplemented(); 1585 } 1586 1587 offset = tj->offset(); 1588 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1589 1590 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1591 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1592 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1593 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1594 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1595 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1596 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1597 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1598 assert( tj->ptr() != TypePtr::TopPTR && 1599 tj->ptr() != TypePtr::AnyNull && 1600 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1601 // assert( tj->ptr() != TypePtr::Constant || 1602 // tj->base() == Type::RawPtr || 1603 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1604 1605 return tj; 1606 } 1607 1608 void Compile::AliasType::Init(int i, const TypePtr* at) { 1609 _index = i; 1610 _adr_type = at; 1611 _field = NULL; 1612 _element = NULL; 1613 _is_rewritable = true; // default 1614 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1615 if (atoop != NULL && atoop->is_known_instance()) { 1616 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1617 _general_index = Compile::current()->get_alias_index(gt); 1618 } else { 1619 _general_index = 0; 1620 } 1621 } 1622 1623 //---------------------------------print_on------------------------------------ 1624 #ifndef PRODUCT 1625 void Compile::AliasType::print_on(outputStream* st) { 1626 if (index() < 10) 1627 st->print("@ <%d> ", index()); 1628 else st->print("@ <%d>", index()); 1629 st->print(is_rewritable() ? " " : " RO"); 1630 int offset = adr_type()->offset(); 1631 if (offset == Type::OffsetBot) 1632 st->print(" +any"); 1633 else st->print(" +%-3d", offset); 1634 st->print(" in "); 1635 adr_type()->dump_on(st); 1636 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1637 if (field() != NULL && tjp) { 1638 if (tjp->klass() != field()->holder() || 1639 tjp->offset() != field()->offset_in_bytes()) { 1640 st->print(" != "); 1641 field()->print(); 1642 st->print(" ***"); 1643 } 1644 } 1645 } 1646 1647 void print_alias_types() { 1648 Compile* C = Compile::current(); 1649 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1650 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1651 C->alias_type(idx)->print_on(tty); 1652 tty->cr(); 1653 } 1654 } 1655 #endif 1656 1657 1658 //----------------------------probe_alias_cache-------------------------------- 1659 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1660 intptr_t key = (intptr_t) adr_type; 1661 key ^= key >> logAliasCacheSize; 1662 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1663 } 1664 1665 1666 //-----------------------------grow_alias_types-------------------------------- 1667 void Compile::grow_alias_types() { 1668 const int old_ats = _max_alias_types; // how many before? 1669 const int new_ats = old_ats; // how many more? 1670 const int grow_ats = old_ats+new_ats; // how many now? 1671 _max_alias_types = grow_ats; 1672 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1673 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1674 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1675 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1676 } 1677 1678 1679 //--------------------------------find_alias_type------------------------------ 1680 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1681 if (_AliasLevel == 0) 1682 return alias_type(AliasIdxBot); 1683 1684 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1685 if (ace->_adr_type == adr_type) { 1686 return alias_type(ace->_index); 1687 } 1688 1689 // Handle special cases. 1690 if (adr_type == NULL) return alias_type(AliasIdxTop); 1691 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1692 1693 // Do it the slow way. 1694 const TypePtr* flat = flatten_alias_type(adr_type); 1695 1696 #ifdef ASSERT 1697 assert(flat == flatten_alias_type(flat), "idempotent"); 1698 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1699 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1700 const TypeOopPtr* foop = flat->is_oopptr(); 1701 // Scalarizable allocations have exact klass always. 1702 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1703 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1704 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1705 } 1706 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1707 #endif 1708 1709 int idx = AliasIdxTop; 1710 for (int i = 0; i < num_alias_types(); i++) { 1711 if (alias_type(i)->adr_type() == flat) { 1712 idx = i; 1713 break; 1714 } 1715 } 1716 1717 if (idx == AliasIdxTop) { 1718 if (no_create) return NULL; 1719 // Grow the array if necessary. 1720 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1721 // Add a new alias type. 1722 idx = _num_alias_types++; 1723 _alias_types[idx]->Init(idx, flat); 1724 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1725 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1726 if (flat->isa_instptr()) { 1727 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1728 && flat->is_instptr()->klass() == env()->Class_klass()) 1729 alias_type(idx)->set_rewritable(false); 1730 } 1731 if (flat->isa_aryptr()) { 1732 #ifdef ASSERT 1733 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1734 // (T_BYTE has the weakest alignment and size restrictions...) 1735 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1736 #endif 1737 if (flat->offset() == TypePtr::OffsetBot) { 1738 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1739 } 1740 } 1741 if (flat->isa_klassptr()) { 1742 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1743 alias_type(idx)->set_rewritable(false); 1744 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1745 alias_type(idx)->set_rewritable(false); 1746 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1747 alias_type(idx)->set_rewritable(false); 1748 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1749 alias_type(idx)->set_rewritable(false); 1750 } 1751 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1752 // but the base pointer type is not distinctive enough to identify 1753 // references into JavaThread.) 1754 1755 // Check for final fields. 1756 const TypeInstPtr* tinst = flat->isa_instptr(); 1757 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1758 ciField* field; 1759 if (tinst->const_oop() != NULL && 1760 tinst->klass() == ciEnv::current()->Class_klass() && 1761 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1762 // static field 1763 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1764 field = k->get_field_by_offset(tinst->offset(), true); 1765 } else { 1766 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1767 field = k->get_field_by_offset(tinst->offset(), false); 1768 } 1769 assert(field == NULL || 1770 original_field == NULL || 1771 (field->holder() == original_field->holder() && 1772 field->offset() == original_field->offset() && 1773 field->is_static() == original_field->is_static()), "wrong field?"); 1774 // Set field() and is_rewritable() attributes. 1775 if (field != NULL) alias_type(idx)->set_field(field); 1776 } 1777 } 1778 1779 // Fill the cache for next time. 1780 ace->_adr_type = adr_type; 1781 ace->_index = idx; 1782 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1783 1784 // Might as well try to fill the cache for the flattened version, too. 1785 AliasCacheEntry* face = probe_alias_cache(flat); 1786 if (face->_adr_type == NULL) { 1787 face->_adr_type = flat; 1788 face->_index = idx; 1789 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1790 } 1791 1792 return alias_type(idx); 1793 } 1794 1795 1796 Compile::AliasType* Compile::alias_type(ciField* field) { 1797 const TypeOopPtr* t; 1798 if (field->is_static()) 1799 t = TypeInstPtr::make(field->holder()->java_mirror()); 1800 else 1801 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1802 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1803 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1804 return atp; 1805 } 1806 1807 1808 //------------------------------have_alias_type-------------------------------- 1809 bool Compile::have_alias_type(const TypePtr* adr_type) { 1810 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1811 if (ace->_adr_type == adr_type) { 1812 return true; 1813 } 1814 1815 // Handle special cases. 1816 if (adr_type == NULL) return true; 1817 if (adr_type == TypePtr::BOTTOM) return true; 1818 1819 return find_alias_type(adr_type, true, NULL) != NULL; 1820 } 1821 1822 //-----------------------------must_alias-------------------------------------- 1823 // True if all values of the given address type are in the given alias category. 1824 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1825 if (alias_idx == AliasIdxBot) return true; // the universal category 1826 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1827 if (alias_idx == AliasIdxTop) return false; // the empty category 1828 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1829 1830 // the only remaining possible overlap is identity 1831 int adr_idx = get_alias_index(adr_type); 1832 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1833 assert(adr_idx == alias_idx || 1834 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1835 && adr_type != TypeOopPtr::BOTTOM), 1836 "should not be testing for overlap with an unsafe pointer"); 1837 return adr_idx == alias_idx; 1838 } 1839 1840 //------------------------------can_alias-------------------------------------- 1841 // True if any values of the given address type are in the given alias category. 1842 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1843 if (alias_idx == AliasIdxTop) return false; // the empty category 1844 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1845 if (alias_idx == AliasIdxBot) return true; // the universal category 1846 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1847 1848 // the only remaining possible overlap is identity 1849 int adr_idx = get_alias_index(adr_type); 1850 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1851 return adr_idx == alias_idx; 1852 } 1853 1854 1855 1856 //---------------------------pop_warm_call------------------------------------- 1857 WarmCallInfo* Compile::pop_warm_call() { 1858 WarmCallInfo* wci = _warm_calls; 1859 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1860 return wci; 1861 } 1862 1863 //----------------------------Inline_Warm-------------------------------------- 1864 int Compile::Inline_Warm() { 1865 // If there is room, try to inline some more warm call sites. 1866 // %%% Do a graph index compaction pass when we think we're out of space? 1867 if (!InlineWarmCalls) return 0; 1868 1869 int calls_made_hot = 0; 1870 int room_to_grow = NodeCountInliningCutoff - unique(); 1871 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1872 int amount_grown = 0; 1873 WarmCallInfo* call; 1874 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1875 int est_size = (int)call->size(); 1876 if (est_size > (room_to_grow - amount_grown)) { 1877 // This one won't fit anyway. Get rid of it. 1878 call->make_cold(); 1879 continue; 1880 } 1881 call->make_hot(); 1882 calls_made_hot++; 1883 amount_grown += est_size; 1884 amount_to_grow -= est_size; 1885 } 1886 1887 if (calls_made_hot > 0) set_major_progress(); 1888 return calls_made_hot; 1889 } 1890 1891 1892 //----------------------------Finish_Warm-------------------------------------- 1893 void Compile::Finish_Warm() { 1894 if (!InlineWarmCalls) return; 1895 if (failing()) return; 1896 if (warm_calls() == NULL) return; 1897 1898 // Clean up loose ends, if we are out of space for inlining. 1899 WarmCallInfo* call; 1900 while ((call = pop_warm_call()) != NULL) { 1901 call->make_cold(); 1902 } 1903 } 1904 1905 //---------------------cleanup_loop_predicates----------------------- 1906 // Remove the opaque nodes that protect the predicates so that all unused 1907 // checks and uncommon_traps will be eliminated from the ideal graph 1908 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1909 if (predicate_count()==0) return; 1910 for (int i = predicate_count(); i > 0; i--) { 1911 Node * n = predicate_opaque1_node(i-1); 1912 assert(n->Opcode() == Op_Opaque1, "must be"); 1913 igvn.replace_node(n, n->in(1)); 1914 } 1915 assert(predicate_count()==0, "should be clean!"); 1916 } 1917 1918 // StringOpts and late inlining of string methods 1919 void Compile::inline_string_calls(bool parse_time) { 1920 { 1921 // remove useless nodes to make the usage analysis simpler 1922 ResourceMark rm; 1923 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1924 } 1925 1926 { 1927 ResourceMark rm; 1928 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1929 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1930 print_method(PHASE_AFTER_STRINGOPTS, 3); 1931 } 1932 1933 // now inline anything that we skipped the first time around 1934 if (!parse_time) { 1935 _late_inlines_pos = _late_inlines.length(); 1936 } 1937 1938 while (_string_late_inlines.length() > 0) { 1939 CallGenerator* cg = _string_late_inlines.pop(); 1940 cg->do_late_inline(); 1941 if (failing()) return; 1942 } 1943 _string_late_inlines.trunc_to(0); 1944 } 1945 1946 // Late inlining of boxing methods 1947 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1948 if (_boxing_late_inlines.length() > 0) { 1949 assert(has_boxed_value(), "inconsistent"); 1950 1951 PhaseGVN* gvn = initial_gvn(); 1952 set_inlining_incrementally(true); 1953 1954 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1955 for_igvn()->clear(); 1956 gvn->replace_with(&igvn); 1957 1958 _late_inlines_pos = _late_inlines.length(); 1959 1960 while (_boxing_late_inlines.length() > 0) { 1961 CallGenerator* cg = _boxing_late_inlines.pop(); 1962 cg->do_late_inline(); 1963 if (failing()) return; 1964 } 1965 _boxing_late_inlines.trunc_to(0); 1966 1967 { 1968 ResourceMark rm; 1969 PhaseRemoveUseless pru(gvn, for_igvn()); 1970 } 1971 1972 igvn = PhaseIterGVN(gvn); 1973 igvn.optimize(); 1974 1975 set_inlining_progress(false); 1976 set_inlining_incrementally(false); 1977 } 1978 } 1979 1980 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1981 assert(IncrementalInline, "incremental inlining should be on"); 1982 PhaseGVN* gvn = initial_gvn(); 1983 1984 set_inlining_progress(false); 1985 for_igvn()->clear(); 1986 gvn->replace_with(&igvn); 1987 1988 { 1989 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 1990 int i = 0; 1991 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1992 CallGenerator* cg = _late_inlines.at(i); 1993 _late_inlines_pos = i+1; 1994 cg->do_late_inline(); 1995 if (failing()) return; 1996 } 1997 int j = 0; 1998 for (; i < _late_inlines.length(); i++, j++) { 1999 _late_inlines.at_put(j, _late_inlines.at(i)); 2000 } 2001 _late_inlines.trunc_to(j); 2002 } 2003 2004 { 2005 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2006 ResourceMark rm; 2007 PhaseRemoveUseless pru(gvn, for_igvn()); 2008 } 2009 2010 { 2011 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2012 igvn = PhaseIterGVN(gvn); 2013 } 2014 } 2015 2016 // Perform incremental inlining until bound on number of live nodes is reached 2017 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2018 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2019 2020 PhaseGVN* gvn = initial_gvn(); 2021 2022 set_inlining_incrementally(true); 2023 set_inlining_progress(true); 2024 uint low_live_nodes = 0; 2025 2026 while(inlining_progress() && _late_inlines.length() > 0) { 2027 2028 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2029 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2030 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2031 // PhaseIdealLoop is expensive so we only try it once we are 2032 // out of live nodes and we only try it again if the previous 2033 // helped got the number of nodes down significantly 2034 PhaseIdealLoop ideal_loop( igvn, false, true ); 2035 if (failing()) return; 2036 low_live_nodes = live_nodes(); 2037 _major_progress = true; 2038 } 2039 2040 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2041 break; 2042 } 2043 } 2044 2045 inline_incrementally_one(igvn); 2046 2047 if (failing()) return; 2048 2049 { 2050 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2051 igvn.optimize(); 2052 } 2053 2054 if (failing()) return; 2055 } 2056 2057 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2058 2059 if (_string_late_inlines.length() > 0) { 2060 assert(has_stringbuilder(), "inconsistent"); 2061 for_igvn()->clear(); 2062 initial_gvn()->replace_with(&igvn); 2063 2064 inline_string_calls(false); 2065 2066 if (failing()) return; 2067 2068 { 2069 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2070 ResourceMark rm; 2071 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2072 } 2073 2074 { 2075 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2076 igvn = PhaseIterGVN(gvn); 2077 igvn.optimize(); 2078 } 2079 } 2080 2081 set_inlining_incrementally(false); 2082 } 2083 2084 2085 //------------------------------Optimize--------------------------------------- 2086 // Given a graph, optimize it. 2087 void Compile::Optimize() { 2088 TracePhase tp("optimizer", &timers[_t_optimizer]); 2089 2090 #ifndef PRODUCT 2091 if (env()->break_at_compile()) { 2092 BREAKPOINT; 2093 } 2094 2095 #endif 2096 2097 ResourceMark rm; 2098 int loop_opts_cnt; 2099 2100 print_inlining_reinit(); 2101 2102 NOT_PRODUCT( verify_graph_edges(); ) 2103 2104 print_method(PHASE_AFTER_PARSING); 2105 2106 { 2107 // Iterative Global Value Numbering, including ideal transforms 2108 // Initialize IterGVN with types and values from parse-time GVN 2109 PhaseIterGVN igvn(initial_gvn()); 2110 #ifdef ASSERT 2111 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2112 #endif 2113 { 2114 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2115 igvn.optimize(); 2116 } 2117 2118 print_method(PHASE_ITER_GVN1, 2); 2119 2120 if (failing()) return; 2121 2122 inline_incrementally(igvn); 2123 2124 print_method(PHASE_INCREMENTAL_INLINE, 2); 2125 2126 if (failing()) return; 2127 2128 if (eliminate_boxing()) { 2129 // Inline valueOf() methods now. 2130 inline_boxing_calls(igvn); 2131 2132 if (AlwaysIncrementalInline) { 2133 inline_incrementally(igvn); 2134 } 2135 2136 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2137 2138 if (failing()) return; 2139 } 2140 2141 // Remove the speculative part of types and clean up the graph from 2142 // the extra CastPP nodes whose only purpose is to carry them. Do 2143 // that early so that optimizations are not disrupted by the extra 2144 // CastPP nodes. 2145 remove_speculative_types(igvn); 2146 2147 // No more new expensive nodes will be added to the list from here 2148 // so keep only the actual candidates for optimizations. 2149 cleanup_expensive_nodes(igvn); 2150 2151 // Perform escape analysis 2152 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2153 if (has_loops()) { 2154 // Cleanup graph (remove dead nodes). 2155 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2156 PhaseIdealLoop ideal_loop( igvn, false, true ); 2157 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2158 if (failing()) return; 2159 } 2160 ConnectionGraph::do_analysis(this, &igvn); 2161 2162 if (failing()) return; 2163 2164 // Optimize out fields loads from scalar replaceable allocations. 2165 igvn.optimize(); 2166 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2167 2168 if (failing()) return; 2169 2170 if (congraph() != NULL && macro_count() > 0) { 2171 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2172 PhaseMacroExpand mexp(igvn); 2173 mexp.eliminate_macro_nodes(); 2174 igvn.set_delay_transform(false); 2175 2176 igvn.optimize(); 2177 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2178 2179 if (failing()) return; 2180 } 2181 } 2182 2183 // Loop transforms on the ideal graph. Range Check Elimination, 2184 // peeling, unrolling, etc. 2185 2186 // Set loop opts counter 2187 loop_opts_cnt = num_loop_opts(); 2188 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2189 { 2190 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2191 PhaseIdealLoop ideal_loop( igvn, true ); 2192 loop_opts_cnt--; 2193 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2194 if (failing()) return; 2195 } 2196 // Loop opts pass if partial peeling occurred in previous pass 2197 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2198 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2199 PhaseIdealLoop ideal_loop( igvn, false ); 2200 loop_opts_cnt--; 2201 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2202 if (failing()) return; 2203 } 2204 // Loop opts pass for loop-unrolling before CCP 2205 if(major_progress() && (loop_opts_cnt > 0)) { 2206 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2207 PhaseIdealLoop ideal_loop( igvn, false ); 2208 loop_opts_cnt--; 2209 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2210 } 2211 if (!failing()) { 2212 // Verify that last round of loop opts produced a valid graph 2213 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2214 PhaseIdealLoop::verify(igvn); 2215 } 2216 } 2217 if (failing()) return; 2218 2219 // Conditional Constant Propagation; 2220 PhaseCCP ccp( &igvn ); 2221 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2222 { 2223 TracePhase tp("ccp", &timers[_t_ccp]); 2224 ccp.do_transform(); 2225 } 2226 print_method(PHASE_CPP1, 2); 2227 2228 assert( true, "Break here to ccp.dump_old2new_map()"); 2229 2230 // Iterative Global Value Numbering, including ideal transforms 2231 { 2232 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2233 igvn = ccp; 2234 igvn.optimize(); 2235 } 2236 2237 print_method(PHASE_ITER_GVN2, 2); 2238 2239 if (failing()) return; 2240 2241 // Loop transforms on the ideal graph. Range Check Elimination, 2242 // peeling, unrolling, etc. 2243 if(loop_opts_cnt > 0) { 2244 debug_only( int cnt = 0; ); 2245 while(major_progress() && (loop_opts_cnt > 0)) { 2246 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2247 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2248 PhaseIdealLoop ideal_loop( igvn, true); 2249 loop_opts_cnt--; 2250 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2251 if (failing()) return; 2252 } 2253 } 2254 // Ensure that major progress is now clear 2255 C->clear_major_progress(); 2256 2257 { 2258 // Verify that all previous optimizations produced a valid graph 2259 // at least to this point, even if no loop optimizations were done. 2260 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2261 PhaseIdealLoop::verify(igvn); 2262 } 2263 2264 { 2265 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2266 PhaseMacroExpand mex(igvn); 2267 if (mex.expand_macro_nodes()) { 2268 assert(failing(), "must bail out w/ explicit message"); 2269 return; 2270 } 2271 } 2272 2273 DEBUG_ONLY( _modified_nodes = NULL; ) 2274 } // (End scope of igvn; run destructor if necessary for asserts.) 2275 2276 process_print_inlining(); 2277 // A method with only infinite loops has no edges entering loops from root 2278 { 2279 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2280 if (final_graph_reshaping()) { 2281 assert(failing(), "must bail out w/ explicit message"); 2282 return; 2283 } 2284 } 2285 2286 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2287 } 2288 2289 2290 //------------------------------Code_Gen--------------------------------------- 2291 // Given a graph, generate code for it 2292 void Compile::Code_Gen() { 2293 if (failing()) { 2294 return; 2295 } 2296 2297 // Perform instruction selection. You might think we could reclaim Matcher 2298 // memory PDQ, but actually the Matcher is used in generating spill code. 2299 // Internals of the Matcher (including some VectorSets) must remain live 2300 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2301 // set a bit in reclaimed memory. 2302 2303 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2304 // nodes. Mapping is only valid at the root of each matched subtree. 2305 NOT_PRODUCT( verify_graph_edges(); ) 2306 2307 Matcher matcher; 2308 _matcher = &matcher; 2309 { 2310 TracePhase tp("matcher", &timers[_t_matcher]); 2311 matcher.match(); 2312 } 2313 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2314 // nodes. Mapping is only valid at the root of each matched subtree. 2315 NOT_PRODUCT( verify_graph_edges(); ) 2316 2317 // If you have too many nodes, or if matching has failed, bail out 2318 check_node_count(0, "out of nodes matching instructions"); 2319 if (failing()) { 2320 return; 2321 } 2322 2323 // Build a proper-looking CFG 2324 PhaseCFG cfg(node_arena(), root(), matcher); 2325 _cfg = &cfg; 2326 { 2327 TracePhase tp("scheduler", &timers[_t_scheduler]); 2328 bool success = cfg.do_global_code_motion(); 2329 if (!success) { 2330 return; 2331 } 2332 2333 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2334 NOT_PRODUCT( verify_graph_edges(); ) 2335 debug_only( cfg.verify(); ) 2336 } 2337 2338 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2339 _regalloc = ®alloc; 2340 { 2341 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2342 // Perform register allocation. After Chaitin, use-def chains are 2343 // no longer accurate (at spill code) and so must be ignored. 2344 // Node->LRG->reg mappings are still accurate. 2345 _regalloc->Register_Allocate(); 2346 2347 // Bail out if the allocator builds too many nodes 2348 if (failing()) { 2349 return; 2350 } 2351 } 2352 2353 // Prior to register allocation we kept empty basic blocks in case the 2354 // the allocator needed a place to spill. After register allocation we 2355 // are not adding any new instructions. If any basic block is empty, we 2356 // can now safely remove it. 2357 { 2358 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2359 cfg.remove_empty_blocks(); 2360 if (do_freq_based_layout()) { 2361 PhaseBlockLayout layout(cfg); 2362 } else { 2363 cfg.set_loop_alignment(); 2364 } 2365 cfg.fixup_flow(); 2366 } 2367 2368 // Apply peephole optimizations 2369 if( OptoPeephole ) { 2370 TracePhase tp("peephole", &timers[_t_peephole]); 2371 PhasePeephole peep( _regalloc, cfg); 2372 peep.do_transform(); 2373 } 2374 2375 // Do late expand if CPU requires this. 2376 if (Matcher::require_postalloc_expand) { 2377 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2378 cfg.postalloc_expand(_regalloc); 2379 } 2380 2381 // Convert Nodes to instruction bits in a buffer 2382 { 2383 TraceTime tp("output", &timers[_t_output], CITime); 2384 Output(); 2385 } 2386 2387 print_method(PHASE_FINAL_CODE); 2388 2389 // He's dead, Jim. 2390 _cfg = (PhaseCFG*)0xdeadbeef; 2391 _regalloc = (PhaseChaitin*)0xdeadbeef; 2392 } 2393 2394 2395 //------------------------------dump_asm--------------------------------------- 2396 // Dump formatted assembly 2397 #ifndef PRODUCT 2398 void Compile::dump_asm(int *pcs, uint pc_limit) { 2399 bool cut_short = false; 2400 tty->print_cr("#"); 2401 tty->print("# "); _tf->dump(); tty->cr(); 2402 tty->print_cr("#"); 2403 2404 // For all blocks 2405 int pc = 0x0; // Program counter 2406 char starts_bundle = ' '; 2407 _regalloc->dump_frame(); 2408 2409 Node *n = NULL; 2410 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2411 if (VMThread::should_terminate()) { 2412 cut_short = true; 2413 break; 2414 } 2415 Block* block = _cfg->get_block(i); 2416 if (block->is_connector() && !Verbose) { 2417 continue; 2418 } 2419 n = block->head(); 2420 if (pcs && n->_idx < pc_limit) { 2421 tty->print("%3.3x ", pcs[n->_idx]); 2422 } else { 2423 tty->print(" "); 2424 } 2425 block->dump_head(_cfg); 2426 if (block->is_connector()) { 2427 tty->print_cr(" # Empty connector block"); 2428 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2429 tty->print_cr(" # Block is sole successor of call"); 2430 } 2431 2432 // For all instructions 2433 Node *delay = NULL; 2434 for (uint j = 0; j < block->number_of_nodes(); j++) { 2435 if (VMThread::should_terminate()) { 2436 cut_short = true; 2437 break; 2438 } 2439 n = block->get_node(j); 2440 if (valid_bundle_info(n)) { 2441 Bundle* bundle = node_bundling(n); 2442 if (bundle->used_in_unconditional_delay()) { 2443 delay = n; 2444 continue; 2445 } 2446 if (bundle->starts_bundle()) { 2447 starts_bundle = '+'; 2448 } 2449 } 2450 2451 if (WizardMode) { 2452 n->dump(); 2453 } 2454 2455 if( !n->is_Region() && // Dont print in the Assembly 2456 !n->is_Phi() && // a few noisely useless nodes 2457 !n->is_Proj() && 2458 !n->is_MachTemp() && 2459 !n->is_SafePointScalarObject() && 2460 !n->is_Catch() && // Would be nice to print exception table targets 2461 !n->is_MergeMem() && // Not very interesting 2462 !n->is_top() && // Debug info table constants 2463 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2464 ) { 2465 if (pcs && n->_idx < pc_limit) 2466 tty->print("%3.3x", pcs[n->_idx]); 2467 else 2468 tty->print(" "); 2469 tty->print(" %c ", starts_bundle); 2470 starts_bundle = ' '; 2471 tty->print("\t"); 2472 n->format(_regalloc, tty); 2473 tty->cr(); 2474 } 2475 2476 // If we have an instruction with a delay slot, and have seen a delay, 2477 // then back up and print it 2478 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2479 assert(delay != NULL, "no unconditional delay instruction"); 2480 if (WizardMode) delay->dump(); 2481 2482 if (node_bundling(delay)->starts_bundle()) 2483 starts_bundle = '+'; 2484 if (pcs && n->_idx < pc_limit) 2485 tty->print("%3.3x", pcs[n->_idx]); 2486 else 2487 tty->print(" "); 2488 tty->print(" %c ", starts_bundle); 2489 starts_bundle = ' '; 2490 tty->print("\t"); 2491 delay->format(_regalloc, tty); 2492 tty->cr(); 2493 delay = NULL; 2494 } 2495 2496 // Dump the exception table as well 2497 if( n->is_Catch() && (Verbose || WizardMode) ) { 2498 // Print the exception table for this offset 2499 _handler_table.print_subtable_for(pc); 2500 } 2501 } 2502 2503 if (pcs && n->_idx < pc_limit) 2504 tty->print_cr("%3.3x", pcs[n->_idx]); 2505 else 2506 tty->cr(); 2507 2508 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2509 2510 } // End of per-block dump 2511 tty->cr(); 2512 2513 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2514 } 2515 #endif 2516 2517 //------------------------------Final_Reshape_Counts--------------------------- 2518 // This class defines counters to help identify when a method 2519 // may/must be executed using hardware with only 24-bit precision. 2520 struct Final_Reshape_Counts : public StackObj { 2521 int _call_count; // count non-inlined 'common' calls 2522 int _float_count; // count float ops requiring 24-bit precision 2523 int _double_count; // count double ops requiring more precision 2524 int _java_call_count; // count non-inlined 'java' calls 2525 int _inner_loop_count; // count loops which need alignment 2526 VectorSet _visited; // Visitation flags 2527 Node_List _tests; // Set of IfNodes & PCTableNodes 2528 2529 Final_Reshape_Counts() : 2530 _call_count(0), _float_count(0), _double_count(0), 2531 _java_call_count(0), _inner_loop_count(0), 2532 _visited( Thread::current()->resource_area() ) { } 2533 2534 void inc_call_count () { _call_count ++; } 2535 void inc_float_count () { _float_count ++; } 2536 void inc_double_count() { _double_count++; } 2537 void inc_java_call_count() { _java_call_count++; } 2538 void inc_inner_loop_count() { _inner_loop_count++; } 2539 2540 int get_call_count () const { return _call_count ; } 2541 int get_float_count () const { return _float_count ; } 2542 int get_double_count() const { return _double_count; } 2543 int get_java_call_count() const { return _java_call_count; } 2544 int get_inner_loop_count() const { return _inner_loop_count; } 2545 }; 2546 2547 #ifdef ASSERT 2548 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2549 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2550 // Make sure the offset goes inside the instance layout. 2551 return k->contains_field_offset(tp->offset()); 2552 // Note that OffsetBot and OffsetTop are very negative. 2553 } 2554 #endif 2555 2556 // Eliminate trivially redundant StoreCMs and accumulate their 2557 // precedence edges. 2558 void Compile::eliminate_redundant_card_marks(Node* n) { 2559 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2560 if (n->in(MemNode::Address)->outcnt() > 1) { 2561 // There are multiple users of the same address so it might be 2562 // possible to eliminate some of the StoreCMs 2563 Node* mem = n->in(MemNode::Memory); 2564 Node* adr = n->in(MemNode::Address); 2565 Node* val = n->in(MemNode::ValueIn); 2566 Node* prev = n; 2567 bool done = false; 2568 // Walk the chain of StoreCMs eliminating ones that match. As 2569 // long as it's a chain of single users then the optimization is 2570 // safe. Eliminating partially redundant StoreCMs would require 2571 // cloning copies down the other paths. 2572 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2573 if (adr == mem->in(MemNode::Address) && 2574 val == mem->in(MemNode::ValueIn)) { 2575 // redundant StoreCM 2576 if (mem->req() > MemNode::OopStore) { 2577 // Hasn't been processed by this code yet. 2578 n->add_prec(mem->in(MemNode::OopStore)); 2579 } else { 2580 // Already converted to precedence edge 2581 for (uint i = mem->req(); i < mem->len(); i++) { 2582 // Accumulate any precedence edges 2583 if (mem->in(i) != NULL) { 2584 n->add_prec(mem->in(i)); 2585 } 2586 } 2587 // Everything above this point has been processed. 2588 done = true; 2589 } 2590 // Eliminate the previous StoreCM 2591 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2592 assert(mem->outcnt() == 0, "should be dead"); 2593 mem->disconnect_inputs(NULL, this); 2594 } else { 2595 prev = mem; 2596 } 2597 mem = prev->in(MemNode::Memory); 2598 } 2599 } 2600 } 2601 2602 //------------------------------final_graph_reshaping_impl---------------------- 2603 // Implement items 1-5 from final_graph_reshaping below. 2604 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2605 2606 if ( n->outcnt() == 0 ) return; // dead node 2607 uint nop = n->Opcode(); 2608 2609 // Check for 2-input instruction with "last use" on right input. 2610 // Swap to left input. Implements item (2). 2611 if( n->req() == 3 && // two-input instruction 2612 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2613 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2614 n->in(2)->outcnt() == 1 &&// right use IS a last use 2615 !n->in(2)->is_Con() ) { // right use is not a constant 2616 // Check for commutative opcode 2617 switch( nop ) { 2618 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2619 case Op_MaxI: case Op_MinI: 2620 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2621 case Op_AndL: case Op_XorL: case Op_OrL: 2622 case Op_AndI: case Op_XorI: case Op_OrI: { 2623 // Move "last use" input to left by swapping inputs 2624 n->swap_edges(1, 2); 2625 break; 2626 } 2627 default: 2628 break; 2629 } 2630 } 2631 2632 #ifdef ASSERT 2633 if( n->is_Mem() ) { 2634 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2635 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2636 // oop will be recorded in oop map if load crosses safepoint 2637 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2638 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2639 "raw memory operations should have control edge"); 2640 } 2641 #endif 2642 // Count FPU ops and common calls, implements item (3) 2643 switch( nop ) { 2644 // Count all float operations that may use FPU 2645 case Op_AddF: 2646 case Op_SubF: 2647 case Op_MulF: 2648 case Op_DivF: 2649 case Op_NegF: 2650 case Op_ModF: 2651 case Op_ConvI2F: 2652 case Op_ConF: 2653 case Op_CmpF: 2654 case Op_CmpF3: 2655 // case Op_ConvL2F: // longs are split into 32-bit halves 2656 frc.inc_float_count(); 2657 break; 2658 2659 case Op_ConvF2D: 2660 case Op_ConvD2F: 2661 frc.inc_float_count(); 2662 frc.inc_double_count(); 2663 break; 2664 2665 // Count all double operations that may use FPU 2666 case Op_AddD: 2667 case Op_SubD: 2668 case Op_MulD: 2669 case Op_DivD: 2670 case Op_NegD: 2671 case Op_ModD: 2672 case Op_ConvI2D: 2673 case Op_ConvD2I: 2674 // case Op_ConvL2D: // handled by leaf call 2675 // case Op_ConvD2L: // handled by leaf call 2676 case Op_ConD: 2677 case Op_CmpD: 2678 case Op_CmpD3: 2679 frc.inc_double_count(); 2680 break; 2681 case Op_Opaque1: // Remove Opaque Nodes before matching 2682 case Op_Opaque2: // Remove Opaque Nodes before matching 2683 case Op_Opaque3: 2684 n->subsume_by(n->in(1), this); 2685 break; 2686 case Op_CallStaticJava: 2687 case Op_CallJava: 2688 case Op_CallDynamicJava: 2689 frc.inc_java_call_count(); // Count java call site; 2690 case Op_CallRuntime: 2691 case Op_CallLeaf: 2692 case Op_CallLeafNoFP: { 2693 assert( n->is_Call(), "" ); 2694 CallNode *call = n->as_Call(); 2695 // Count call sites where the FP mode bit would have to be flipped. 2696 // Do not count uncommon runtime calls: 2697 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2698 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2699 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2700 frc.inc_call_count(); // Count the call site 2701 } else { // See if uncommon argument is shared 2702 Node *n = call->in(TypeFunc::Parms); 2703 int nop = n->Opcode(); 2704 // Clone shared simple arguments to uncommon calls, item (1). 2705 if( n->outcnt() > 1 && 2706 !n->is_Proj() && 2707 nop != Op_CreateEx && 2708 nop != Op_CheckCastPP && 2709 nop != Op_DecodeN && 2710 nop != Op_DecodeNKlass && 2711 !n->is_Mem() ) { 2712 Node *x = n->clone(); 2713 call->set_req( TypeFunc::Parms, x ); 2714 } 2715 } 2716 break; 2717 } 2718 2719 case Op_StoreD: 2720 case Op_LoadD: 2721 case Op_LoadD_unaligned: 2722 frc.inc_double_count(); 2723 goto handle_mem; 2724 case Op_StoreF: 2725 case Op_LoadF: 2726 frc.inc_float_count(); 2727 goto handle_mem; 2728 2729 case Op_StoreCM: 2730 { 2731 // Convert OopStore dependence into precedence edge 2732 Node* prec = n->in(MemNode::OopStore); 2733 n->del_req(MemNode::OopStore); 2734 n->add_prec(prec); 2735 eliminate_redundant_card_marks(n); 2736 } 2737 2738 // fall through 2739 2740 case Op_StoreB: 2741 case Op_StoreC: 2742 case Op_StorePConditional: 2743 case Op_StoreI: 2744 case Op_StoreL: 2745 case Op_StoreIConditional: 2746 case Op_StoreLConditional: 2747 case Op_CompareAndSwapI: 2748 case Op_CompareAndSwapL: 2749 case Op_CompareAndSwapP: 2750 case Op_CompareAndSwapN: 2751 case Op_GetAndAddI: 2752 case Op_GetAndAddL: 2753 case Op_GetAndSetI: 2754 case Op_GetAndSetL: 2755 case Op_GetAndSetP: 2756 case Op_GetAndSetN: 2757 case Op_StoreP: 2758 case Op_StoreN: 2759 case Op_StoreNKlass: 2760 case Op_LoadB: 2761 case Op_LoadUB: 2762 case Op_LoadUS: 2763 case Op_LoadI: 2764 case Op_LoadKlass: 2765 case Op_LoadNKlass: 2766 case Op_LoadL: 2767 case Op_LoadL_unaligned: 2768 case Op_LoadPLocked: 2769 case Op_LoadP: 2770 case Op_LoadN: 2771 case Op_LoadRange: 2772 case Op_LoadS: { 2773 handle_mem: 2774 #ifdef ASSERT 2775 if( VerifyOptoOopOffsets ) { 2776 assert( n->is_Mem(), "" ); 2777 MemNode *mem = (MemNode*)n; 2778 // Check to see if address types have grounded out somehow. 2779 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2780 assert( !tp || oop_offset_is_sane(tp), "" ); 2781 } 2782 #endif 2783 break; 2784 } 2785 2786 case Op_AddP: { // Assert sane base pointers 2787 Node *addp = n->in(AddPNode::Address); 2788 assert( !addp->is_AddP() || 2789 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2790 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2791 "Base pointers must match" ); 2792 #ifdef _LP64 2793 if ((UseCompressedOops || UseCompressedClassPointers) && 2794 addp->Opcode() == Op_ConP && 2795 addp == n->in(AddPNode::Base) && 2796 n->in(AddPNode::Offset)->is_Con()) { 2797 // Use addressing with narrow klass to load with offset on x86. 2798 // On sparc loading 32-bits constant and decoding it have less 2799 // instructions (4) then load 64-bits constant (7). 2800 // Do this transformation here since IGVN will convert ConN back to ConP. 2801 const Type* t = addp->bottom_type(); 2802 if (t->isa_oopptr() || t->isa_klassptr()) { 2803 Node* nn = NULL; 2804 2805 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2806 2807 // Look for existing ConN node of the same exact type. 2808 Node* r = root(); 2809 uint cnt = r->outcnt(); 2810 for (uint i = 0; i < cnt; i++) { 2811 Node* m = r->raw_out(i); 2812 if (m!= NULL && m->Opcode() == op && 2813 m->bottom_type()->make_ptr() == t) { 2814 nn = m; 2815 break; 2816 } 2817 } 2818 if (nn != NULL) { 2819 // Decode a narrow oop to match address 2820 // [R12 + narrow_oop_reg<<3 + offset] 2821 if (t->isa_oopptr()) { 2822 nn = new DecodeNNode(nn, t); 2823 } else { 2824 nn = new DecodeNKlassNode(nn, t); 2825 } 2826 n->set_req(AddPNode::Base, nn); 2827 n->set_req(AddPNode::Address, nn); 2828 if (addp->outcnt() == 0) { 2829 addp->disconnect_inputs(NULL, this); 2830 } 2831 } 2832 } 2833 } 2834 #endif 2835 break; 2836 } 2837 2838 case Op_CastPP: { 2839 // Remove CastPP nodes to gain more freedom during scheduling but 2840 // keep the dependency they encode as control or precedence edges 2841 // (if control is set already) on memory operations. Some CastPP 2842 // nodes don't have a control (don't carry a dependency): skip 2843 // those. 2844 if (n->in(0) != NULL) { 2845 ResourceMark rm; 2846 Unique_Node_List wq; 2847 wq.push(n); 2848 for (uint next = 0; next < wq.size(); ++next) { 2849 Node *m = wq.at(next); 2850 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 2851 Node* use = m->fast_out(i); 2852 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 2853 use->ensure_control_or_add_prec(n->in(0)); 2854 } else if (use->in(0) == NULL) { 2855 switch(use->Opcode()) { 2856 case Op_AddP: 2857 case Op_DecodeN: 2858 case Op_DecodeNKlass: 2859 case Op_CheckCastPP: 2860 case Op_CastPP: 2861 wq.push(use); 2862 break; 2863 } 2864 } 2865 } 2866 } 2867 } 2868 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 2869 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2870 Node* in1 = n->in(1); 2871 const Type* t = n->bottom_type(); 2872 Node* new_in1 = in1->clone(); 2873 new_in1->as_DecodeN()->set_type(t); 2874 2875 if (!Matcher::narrow_oop_use_complex_address()) { 2876 // 2877 // x86, ARM and friends can handle 2 adds in addressing mode 2878 // and Matcher can fold a DecodeN node into address by using 2879 // a narrow oop directly and do implicit NULL check in address: 2880 // 2881 // [R12 + narrow_oop_reg<<3 + offset] 2882 // NullCheck narrow_oop_reg 2883 // 2884 // On other platforms (Sparc) we have to keep new DecodeN node and 2885 // use it to do implicit NULL check in address: 2886 // 2887 // decode_not_null narrow_oop_reg, base_reg 2888 // [base_reg + offset] 2889 // NullCheck base_reg 2890 // 2891 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2892 // to keep the information to which NULL check the new DecodeN node 2893 // corresponds to use it as value in implicit_null_check(). 2894 // 2895 new_in1->set_req(0, n->in(0)); 2896 } 2897 2898 n->subsume_by(new_in1, this); 2899 if (in1->outcnt() == 0) { 2900 in1->disconnect_inputs(NULL, this); 2901 } 2902 } else { 2903 n->subsume_by(n->in(1), this); 2904 if (n->outcnt() == 0) { 2905 n->disconnect_inputs(NULL, this); 2906 } 2907 } 2908 break; 2909 } 2910 #ifdef _LP64 2911 case Op_CmpP: 2912 // Do this transformation here to preserve CmpPNode::sub() and 2913 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2914 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2915 Node* in1 = n->in(1); 2916 Node* in2 = n->in(2); 2917 if (!in1->is_DecodeNarrowPtr()) { 2918 in2 = in1; 2919 in1 = n->in(2); 2920 } 2921 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2922 2923 Node* new_in2 = NULL; 2924 if (in2->is_DecodeNarrowPtr()) { 2925 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2926 new_in2 = in2->in(1); 2927 } else if (in2->Opcode() == Op_ConP) { 2928 const Type* t = in2->bottom_type(); 2929 if (t == TypePtr::NULL_PTR) { 2930 assert(in1->is_DecodeN(), "compare klass to null?"); 2931 // Don't convert CmpP null check into CmpN if compressed 2932 // oops implicit null check is not generated. 2933 // This will allow to generate normal oop implicit null check. 2934 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2935 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 2936 // 2937 // This transformation together with CastPP transformation above 2938 // will generated code for implicit NULL checks for compressed oops. 2939 // 2940 // The original code after Optimize() 2941 // 2942 // LoadN memory, narrow_oop_reg 2943 // decode narrow_oop_reg, base_reg 2944 // CmpP base_reg, NULL 2945 // CastPP base_reg // NotNull 2946 // Load [base_reg + offset], val_reg 2947 // 2948 // after these transformations will be 2949 // 2950 // LoadN memory, narrow_oop_reg 2951 // CmpN narrow_oop_reg, NULL 2952 // decode_not_null narrow_oop_reg, base_reg 2953 // Load [base_reg + offset], val_reg 2954 // 2955 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2956 // since narrow oops can be used in debug info now (see the code in 2957 // final_graph_reshaping_walk()). 2958 // 2959 // At the end the code will be matched to 2960 // on x86: 2961 // 2962 // Load_narrow_oop memory, narrow_oop_reg 2963 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2964 // NullCheck narrow_oop_reg 2965 // 2966 // and on sparc: 2967 // 2968 // Load_narrow_oop memory, narrow_oop_reg 2969 // decode_not_null narrow_oop_reg, base_reg 2970 // Load [base_reg + offset], val_reg 2971 // NullCheck base_reg 2972 // 2973 } else if (t->isa_oopptr()) { 2974 new_in2 = ConNode::make(t->make_narrowoop()); 2975 } else if (t->isa_klassptr()) { 2976 new_in2 = ConNode::make(t->make_narrowklass()); 2977 } 2978 } 2979 if (new_in2 != NULL) { 2980 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 2981 n->subsume_by(cmpN, this); 2982 if (in1->outcnt() == 0) { 2983 in1->disconnect_inputs(NULL, this); 2984 } 2985 if (in2->outcnt() == 0) { 2986 in2->disconnect_inputs(NULL, this); 2987 } 2988 } 2989 } 2990 break; 2991 2992 case Op_DecodeN: 2993 case Op_DecodeNKlass: 2994 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2995 // DecodeN could be pinned when it can't be fold into 2996 // an address expression, see the code for Op_CastPP above. 2997 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2998 break; 2999 3000 case Op_EncodeP: 3001 case Op_EncodePKlass: { 3002 Node* in1 = n->in(1); 3003 if (in1->is_DecodeNarrowPtr()) { 3004 n->subsume_by(in1->in(1), this); 3005 } else if (in1->Opcode() == Op_ConP) { 3006 const Type* t = in1->bottom_type(); 3007 if (t == TypePtr::NULL_PTR) { 3008 assert(t->isa_oopptr(), "null klass?"); 3009 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3010 } else if (t->isa_oopptr()) { 3011 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3012 } else if (t->isa_klassptr()) { 3013 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3014 } 3015 } 3016 if (in1->outcnt() == 0) { 3017 in1->disconnect_inputs(NULL, this); 3018 } 3019 break; 3020 } 3021 3022 case Op_Proj: { 3023 if (OptimizeStringConcat) { 3024 ProjNode* p = n->as_Proj(); 3025 if (p->_is_io_use) { 3026 // Separate projections were used for the exception path which 3027 // are normally removed by a late inline. If it wasn't inlined 3028 // then they will hang around and should just be replaced with 3029 // the original one. 3030 Node* proj = NULL; 3031 // Replace with just one 3032 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3033 Node *use = i.get(); 3034 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3035 proj = use; 3036 break; 3037 } 3038 } 3039 assert(proj != NULL, "must be found"); 3040 p->subsume_by(proj, this); 3041 } 3042 } 3043 break; 3044 } 3045 3046 case Op_Phi: 3047 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3048 // The EncodeP optimization may create Phi with the same edges 3049 // for all paths. It is not handled well by Register Allocator. 3050 Node* unique_in = n->in(1); 3051 assert(unique_in != NULL, ""); 3052 uint cnt = n->req(); 3053 for (uint i = 2; i < cnt; i++) { 3054 Node* m = n->in(i); 3055 assert(m != NULL, ""); 3056 if (unique_in != m) 3057 unique_in = NULL; 3058 } 3059 if (unique_in != NULL) { 3060 n->subsume_by(unique_in, this); 3061 } 3062 } 3063 break; 3064 3065 #endif 3066 3067 case Op_ModI: 3068 if (UseDivMod) { 3069 // Check if a%b and a/b both exist 3070 Node* d = n->find_similar(Op_DivI); 3071 if (d) { 3072 // Replace them with a fused divmod if supported 3073 if (Matcher::has_match_rule(Op_DivModI)) { 3074 DivModINode* divmod = DivModINode::make(n); 3075 d->subsume_by(divmod->div_proj(), this); 3076 n->subsume_by(divmod->mod_proj(), this); 3077 } else { 3078 // replace a%b with a-((a/b)*b) 3079 Node* mult = new MulINode(d, d->in(2)); 3080 Node* sub = new SubINode(d->in(1), mult); 3081 n->subsume_by(sub, this); 3082 } 3083 } 3084 } 3085 break; 3086 3087 case Op_ModL: 3088 if (UseDivMod) { 3089 // Check if a%b and a/b both exist 3090 Node* d = n->find_similar(Op_DivL); 3091 if (d) { 3092 // Replace them with a fused divmod if supported 3093 if (Matcher::has_match_rule(Op_DivModL)) { 3094 DivModLNode* divmod = DivModLNode::make(n); 3095 d->subsume_by(divmod->div_proj(), this); 3096 n->subsume_by(divmod->mod_proj(), this); 3097 } else { 3098 // replace a%b with a-((a/b)*b) 3099 Node* mult = new MulLNode(d, d->in(2)); 3100 Node* sub = new SubLNode(d->in(1), mult); 3101 n->subsume_by(sub, this); 3102 } 3103 } 3104 } 3105 break; 3106 3107 case Op_LoadVector: 3108 case Op_StoreVector: 3109 break; 3110 3111 case Op_AddReductionVI: 3112 case Op_AddReductionVL: 3113 case Op_AddReductionVF: 3114 case Op_AddReductionVD: 3115 case Op_MulReductionVI: 3116 case Op_MulReductionVL: 3117 case Op_MulReductionVF: 3118 case Op_MulReductionVD: 3119 break; 3120 3121 case Op_PackB: 3122 case Op_PackS: 3123 case Op_PackI: 3124 case Op_PackF: 3125 case Op_PackL: 3126 case Op_PackD: 3127 if (n->req()-1 > 2) { 3128 // Replace many operand PackNodes with a binary tree for matching 3129 PackNode* p = (PackNode*) n; 3130 Node* btp = p->binary_tree_pack(1, n->req()); 3131 n->subsume_by(btp, this); 3132 } 3133 break; 3134 case Op_Loop: 3135 case Op_CountedLoop: 3136 if (n->as_Loop()->is_inner_loop()) { 3137 frc.inc_inner_loop_count(); 3138 } 3139 break; 3140 case Op_LShiftI: 3141 case Op_RShiftI: 3142 case Op_URShiftI: 3143 case Op_LShiftL: 3144 case Op_RShiftL: 3145 case Op_URShiftL: 3146 if (Matcher::need_masked_shift_count) { 3147 // The cpu's shift instructions don't restrict the count to the 3148 // lower 5/6 bits. We need to do the masking ourselves. 3149 Node* in2 = n->in(2); 3150 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3151 const TypeInt* t = in2->find_int_type(); 3152 if (t != NULL && t->is_con()) { 3153 juint shift = t->get_con(); 3154 if (shift > mask) { // Unsigned cmp 3155 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3156 } 3157 } else { 3158 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3159 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3160 n->set_req(2, shift); 3161 } 3162 } 3163 if (in2->outcnt() == 0) { // Remove dead node 3164 in2->disconnect_inputs(NULL, this); 3165 } 3166 } 3167 break; 3168 case Op_MemBarStoreStore: 3169 case Op_MemBarRelease: 3170 // Break the link with AllocateNode: it is no longer useful and 3171 // confuses register allocation. 3172 if (n->req() > MemBarNode::Precedent) { 3173 n->set_req(MemBarNode::Precedent, top()); 3174 } 3175 break; 3176 default: 3177 assert( !n->is_Call(), "" ); 3178 assert( !n->is_Mem(), "" ); 3179 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3180 break; 3181 } 3182 3183 // Collect CFG split points 3184 if (n->is_MultiBranch()) 3185 frc._tests.push(n); 3186 } 3187 3188 //------------------------------final_graph_reshaping_walk--------------------- 3189 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3190 // requires that the walk visits a node's inputs before visiting the node. 3191 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3192 ResourceArea *area = Thread::current()->resource_area(); 3193 Unique_Node_List sfpt(area); 3194 3195 frc._visited.set(root->_idx); // first, mark node as visited 3196 uint cnt = root->req(); 3197 Node *n = root; 3198 uint i = 0; 3199 while (true) { 3200 if (i < cnt) { 3201 // Place all non-visited non-null inputs onto stack 3202 Node* m = n->in(i); 3203 ++i; 3204 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3205 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3206 // compute worst case interpreter size in case of a deoptimization 3207 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3208 3209 sfpt.push(m); 3210 } 3211 cnt = m->req(); 3212 nstack.push(n, i); // put on stack parent and next input's index 3213 n = m; 3214 i = 0; 3215 } 3216 } else { 3217 // Now do post-visit work 3218 final_graph_reshaping_impl( n, frc ); 3219 if (nstack.is_empty()) 3220 break; // finished 3221 n = nstack.node(); // Get node from stack 3222 cnt = n->req(); 3223 i = nstack.index(); 3224 nstack.pop(); // Shift to the next node on stack 3225 } 3226 } 3227 3228 // Skip next transformation if compressed oops are not used. 3229 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3230 (!UseCompressedOops && !UseCompressedClassPointers)) 3231 return; 3232 3233 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3234 // It could be done for an uncommon traps or any safepoints/calls 3235 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3236 while (sfpt.size() > 0) { 3237 n = sfpt.pop(); 3238 JVMState *jvms = n->as_SafePoint()->jvms(); 3239 assert(jvms != NULL, "sanity"); 3240 int start = jvms->debug_start(); 3241 int end = n->req(); 3242 bool is_uncommon = (n->is_CallStaticJava() && 3243 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3244 for (int j = start; j < end; j++) { 3245 Node* in = n->in(j); 3246 if (in->is_DecodeNarrowPtr()) { 3247 bool safe_to_skip = true; 3248 if (!is_uncommon ) { 3249 // Is it safe to skip? 3250 for (uint i = 0; i < in->outcnt(); i++) { 3251 Node* u = in->raw_out(i); 3252 if (!u->is_SafePoint() || 3253 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3254 safe_to_skip = false; 3255 } 3256 } 3257 } 3258 if (safe_to_skip) { 3259 n->set_req(j, in->in(1)); 3260 } 3261 if (in->outcnt() == 0) { 3262 in->disconnect_inputs(NULL, this); 3263 } 3264 } 3265 } 3266 } 3267 } 3268 3269 //------------------------------final_graph_reshaping-------------------------- 3270 // Final Graph Reshaping. 3271 // 3272 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3273 // and not commoned up and forced early. Must come after regular 3274 // optimizations to avoid GVN undoing the cloning. Clone constant 3275 // inputs to Loop Phis; these will be split by the allocator anyways. 3276 // Remove Opaque nodes. 3277 // (2) Move last-uses by commutative operations to the left input to encourage 3278 // Intel update-in-place two-address operations and better register usage 3279 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3280 // calls canonicalizing them back. 3281 // (3) Count the number of double-precision FP ops, single-precision FP ops 3282 // and call sites. On Intel, we can get correct rounding either by 3283 // forcing singles to memory (requires extra stores and loads after each 3284 // FP bytecode) or we can set a rounding mode bit (requires setting and 3285 // clearing the mode bit around call sites). The mode bit is only used 3286 // if the relative frequency of single FP ops to calls is low enough. 3287 // This is a key transform for SPEC mpeg_audio. 3288 // (4) Detect infinite loops; blobs of code reachable from above but not 3289 // below. Several of the Code_Gen algorithms fail on such code shapes, 3290 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3291 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3292 // Detection is by looking for IfNodes where only 1 projection is 3293 // reachable from below or CatchNodes missing some targets. 3294 // (5) Assert for insane oop offsets in debug mode. 3295 3296 bool Compile::final_graph_reshaping() { 3297 // an infinite loop may have been eliminated by the optimizer, 3298 // in which case the graph will be empty. 3299 if (root()->req() == 1) { 3300 record_method_not_compilable("trivial infinite loop"); 3301 return true; 3302 } 3303 3304 // Expensive nodes have their control input set to prevent the GVN 3305 // from freely commoning them. There's no GVN beyond this point so 3306 // no need to keep the control input. We want the expensive nodes to 3307 // be freely moved to the least frequent code path by gcm. 3308 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3309 for (int i = 0; i < expensive_count(); i++) { 3310 _expensive_nodes->at(i)->set_req(0, NULL); 3311 } 3312 3313 Final_Reshape_Counts frc; 3314 3315 // Visit everybody reachable! 3316 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3317 Node_Stack nstack(live_nodes() >> 1); 3318 final_graph_reshaping_walk(nstack, root(), frc); 3319 3320 // Check for unreachable (from below) code (i.e., infinite loops). 3321 for( uint i = 0; i < frc._tests.size(); i++ ) { 3322 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3323 // Get number of CFG targets. 3324 // Note that PCTables include exception targets after calls. 3325 uint required_outcnt = n->required_outcnt(); 3326 if (n->outcnt() != required_outcnt) { 3327 // Check for a few special cases. Rethrow Nodes never take the 3328 // 'fall-thru' path, so expected kids is 1 less. 3329 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3330 if (n->in(0)->in(0)->is_Call()) { 3331 CallNode *call = n->in(0)->in(0)->as_Call(); 3332 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3333 required_outcnt--; // Rethrow always has 1 less kid 3334 } else if (call->req() > TypeFunc::Parms && 3335 call->is_CallDynamicJava()) { 3336 // Check for null receiver. In such case, the optimizer has 3337 // detected that the virtual call will always result in a null 3338 // pointer exception. The fall-through projection of this CatchNode 3339 // will not be populated. 3340 Node *arg0 = call->in(TypeFunc::Parms); 3341 if (arg0->is_Type() && 3342 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3343 required_outcnt--; 3344 } 3345 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3346 call->req() > TypeFunc::Parms+1 && 3347 call->is_CallStaticJava()) { 3348 // Check for negative array length. In such case, the optimizer has 3349 // detected that the allocation attempt will always result in an 3350 // exception. There is no fall-through projection of this CatchNode . 3351 Node *arg1 = call->in(TypeFunc::Parms+1); 3352 if (arg1->is_Type() && 3353 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3354 required_outcnt--; 3355 } 3356 } 3357 } 3358 } 3359 // Recheck with a better notion of 'required_outcnt' 3360 if (n->outcnt() != required_outcnt) { 3361 record_method_not_compilable("malformed control flow"); 3362 return true; // Not all targets reachable! 3363 } 3364 } 3365 // Check that I actually visited all kids. Unreached kids 3366 // must be infinite loops. 3367 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3368 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3369 record_method_not_compilable("infinite loop"); 3370 return true; // Found unvisited kid; must be unreach 3371 } 3372 } 3373 3374 // If original bytecodes contained a mixture of floats and doubles 3375 // check if the optimizer has made it homogenous, item (3). 3376 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3377 frc.get_float_count() > 32 && 3378 frc.get_double_count() == 0 && 3379 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3380 set_24_bit_selection_and_mode( false, true ); 3381 } 3382 3383 set_java_calls(frc.get_java_call_count()); 3384 set_inner_loops(frc.get_inner_loop_count()); 3385 3386 // No infinite loops, no reason to bail out. 3387 return false; 3388 } 3389 3390 //-----------------------------too_many_traps---------------------------------- 3391 // Report if there are too many traps at the current method and bci. 3392 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3393 bool Compile::too_many_traps(ciMethod* method, 3394 int bci, 3395 Deoptimization::DeoptReason reason) { 3396 ciMethodData* md = method->method_data(); 3397 if (md->is_empty()) { 3398 // Assume the trap has not occurred, or that it occurred only 3399 // because of a transient condition during start-up in the interpreter. 3400 return false; 3401 } 3402 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3403 if (md->has_trap_at(bci, m, reason) != 0) { 3404 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3405 // Also, if there are multiple reasons, or if there is no per-BCI record, 3406 // assume the worst. 3407 if (log()) 3408 log()->elem("observe trap='%s' count='%d'", 3409 Deoptimization::trap_reason_name(reason), 3410 md->trap_count(reason)); 3411 return true; 3412 } else { 3413 // Ignore method/bci and see if there have been too many globally. 3414 return too_many_traps(reason, md); 3415 } 3416 } 3417 3418 // Less-accurate variant which does not require a method and bci. 3419 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3420 ciMethodData* logmd) { 3421 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3422 // Too many traps globally. 3423 // Note that we use cumulative trap_count, not just md->trap_count. 3424 if (log()) { 3425 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3426 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3427 Deoptimization::trap_reason_name(reason), 3428 mcount, trap_count(reason)); 3429 } 3430 return true; 3431 } else { 3432 // The coast is clear. 3433 return false; 3434 } 3435 } 3436 3437 //--------------------------too_many_recompiles-------------------------------- 3438 // Report if there are too many recompiles at the current method and bci. 3439 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3440 // Is not eager to return true, since this will cause the compiler to use 3441 // Action_none for a trap point, to avoid too many recompilations. 3442 bool Compile::too_many_recompiles(ciMethod* method, 3443 int bci, 3444 Deoptimization::DeoptReason reason) { 3445 ciMethodData* md = method->method_data(); 3446 if (md->is_empty()) { 3447 // Assume the trap has not occurred, or that it occurred only 3448 // because of a transient condition during start-up in the interpreter. 3449 return false; 3450 } 3451 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3452 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3453 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3454 Deoptimization::DeoptReason per_bc_reason 3455 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3456 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3457 if ((per_bc_reason == Deoptimization::Reason_none 3458 || md->has_trap_at(bci, m, reason) != 0) 3459 // The trap frequency measure we care about is the recompile count: 3460 && md->trap_recompiled_at(bci, m) 3461 && md->overflow_recompile_count() >= bc_cutoff) { 3462 // Do not emit a trap here if it has already caused recompilations. 3463 // Also, if there are multiple reasons, or if there is no per-BCI record, 3464 // assume the worst. 3465 if (log()) 3466 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3467 Deoptimization::trap_reason_name(reason), 3468 md->trap_count(reason), 3469 md->overflow_recompile_count()); 3470 return true; 3471 } else if (trap_count(reason) != 0 3472 && decompile_count() >= m_cutoff) { 3473 // Too many recompiles globally, and we have seen this sort of trap. 3474 // Use cumulative decompile_count, not just md->decompile_count. 3475 if (log()) 3476 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3477 Deoptimization::trap_reason_name(reason), 3478 md->trap_count(reason), trap_count(reason), 3479 md->decompile_count(), decompile_count()); 3480 return true; 3481 } else { 3482 // The coast is clear. 3483 return false; 3484 } 3485 } 3486 3487 // Compute when not to trap. Used by matching trap based nodes and 3488 // NullCheck optimization. 3489 void Compile::set_allowed_deopt_reasons() { 3490 _allowed_reasons = 0; 3491 if (is_method_compilation()) { 3492 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3493 assert(rs < BitsPerInt, "recode bit map"); 3494 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3495 _allowed_reasons |= nth_bit(rs); 3496 } 3497 } 3498 } 3499 } 3500 3501 #ifndef PRODUCT 3502 //------------------------------verify_graph_edges--------------------------- 3503 // Walk the Graph and verify that there is a one-to-one correspondence 3504 // between Use-Def edges and Def-Use edges in the graph. 3505 void Compile::verify_graph_edges(bool no_dead_code) { 3506 if (VerifyGraphEdges) { 3507 ResourceArea *area = Thread::current()->resource_area(); 3508 Unique_Node_List visited(area); 3509 // Call recursive graph walk to check edges 3510 _root->verify_edges(visited); 3511 if (no_dead_code) { 3512 // Now make sure that no visited node is used by an unvisited node. 3513 bool dead_nodes = false; 3514 Unique_Node_List checked(area); 3515 while (visited.size() > 0) { 3516 Node* n = visited.pop(); 3517 checked.push(n); 3518 for (uint i = 0; i < n->outcnt(); i++) { 3519 Node* use = n->raw_out(i); 3520 if (checked.member(use)) continue; // already checked 3521 if (visited.member(use)) continue; // already in the graph 3522 if (use->is_Con()) continue; // a dead ConNode is OK 3523 // At this point, we have found a dead node which is DU-reachable. 3524 if (!dead_nodes) { 3525 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3526 dead_nodes = true; 3527 } 3528 use->dump(2); 3529 tty->print_cr("---"); 3530 checked.push(use); // No repeats; pretend it is now checked. 3531 } 3532 } 3533 assert(!dead_nodes, "using nodes must be reachable from root"); 3534 } 3535 } 3536 } 3537 3538 // Verify GC barriers consistency 3539 // Currently supported: 3540 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3541 void Compile::verify_barriers() { 3542 if (UseG1GC) { 3543 // Verify G1 pre-barriers 3544 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3545 3546 ResourceArea *area = Thread::current()->resource_area(); 3547 Unique_Node_List visited(area); 3548 Node_List worklist(area); 3549 // We're going to walk control flow backwards starting from the Root 3550 worklist.push(_root); 3551 while (worklist.size() > 0) { 3552 Node* x = worklist.pop(); 3553 if (x == NULL || x == top()) continue; 3554 if (visited.member(x)) { 3555 continue; 3556 } else { 3557 visited.push(x); 3558 } 3559 3560 if (x->is_Region()) { 3561 for (uint i = 1; i < x->req(); i++) { 3562 worklist.push(x->in(i)); 3563 } 3564 } else { 3565 worklist.push(x->in(0)); 3566 // We are looking for the pattern: 3567 // /->ThreadLocal 3568 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3569 // \->ConI(0) 3570 // We want to verify that the If and the LoadB have the same control 3571 // See GraphKit::g1_write_barrier_pre() 3572 if (x->is_If()) { 3573 IfNode *iff = x->as_If(); 3574 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3575 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3576 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3577 && cmp->in(1)->is_Load()) { 3578 LoadNode* load = cmp->in(1)->as_Load(); 3579 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3580 && load->in(2)->in(3)->is_Con() 3581 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3582 3583 Node* if_ctrl = iff->in(0); 3584 Node* load_ctrl = load->in(0); 3585 3586 if (if_ctrl != load_ctrl) { 3587 // Skip possible CProj->NeverBranch in infinite loops 3588 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3589 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3590 if_ctrl = if_ctrl->in(0)->in(0); 3591 } 3592 } 3593 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3594 } 3595 } 3596 } 3597 } 3598 } 3599 } 3600 } 3601 } 3602 3603 #endif 3604 3605 // The Compile object keeps track of failure reasons separately from the ciEnv. 3606 // This is required because there is not quite a 1-1 relation between the 3607 // ciEnv and its compilation task and the Compile object. Note that one 3608 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3609 // to backtrack and retry without subsuming loads. Other than this backtracking 3610 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3611 // by the logic in C2Compiler. 3612 void Compile::record_failure(const char* reason) { 3613 if (log() != NULL) { 3614 log()->elem("failure reason='%s' phase='compile'", reason); 3615 } 3616 if (_failure_reason == NULL) { 3617 // Record the first failure reason. 3618 _failure_reason = reason; 3619 } 3620 3621 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3622 C->print_method(PHASE_FAILURE); 3623 } 3624 _root = NULL; // flush the graph, too 3625 } 3626 3627 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3628 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3629 _phase_name(name), _dolog(CITimeVerbose) 3630 { 3631 if (_dolog) { 3632 C = Compile::current(); 3633 _log = C->log(); 3634 } else { 3635 C = NULL; 3636 _log = NULL; 3637 } 3638 if (_log != NULL) { 3639 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3640 _log->stamp(); 3641 _log->end_head(); 3642 } 3643 } 3644 3645 Compile::TracePhase::~TracePhase() { 3646 3647 C = Compile::current(); 3648 if (_dolog) { 3649 _log = C->log(); 3650 } else { 3651 _log = NULL; 3652 } 3653 3654 #ifdef ASSERT 3655 if (PrintIdealNodeCount) { 3656 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3657 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3658 } 3659 3660 if (VerifyIdealNodeCount) { 3661 Compile::current()->print_missing_nodes(); 3662 } 3663 #endif 3664 3665 if (_log != NULL) { 3666 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3667 } 3668 } 3669 3670 //============================================================================= 3671 // Two Constant's are equal when the type and the value are equal. 3672 bool Compile::Constant::operator==(const Constant& other) { 3673 if (type() != other.type() ) return false; 3674 if (can_be_reused() != other.can_be_reused()) return false; 3675 // For floating point values we compare the bit pattern. 3676 switch (type()) { 3677 case T_FLOAT: return (_v._value.i == other._v._value.i); 3678 case T_LONG: 3679 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3680 case T_OBJECT: 3681 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3682 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3683 case T_METADATA: return (_v._metadata == other._v._metadata); 3684 default: ShouldNotReachHere(); 3685 } 3686 return false; 3687 } 3688 3689 static int type_to_size_in_bytes(BasicType t) { 3690 switch (t) { 3691 case T_LONG: return sizeof(jlong ); 3692 case T_FLOAT: return sizeof(jfloat ); 3693 case T_DOUBLE: return sizeof(jdouble); 3694 case T_METADATA: return sizeof(Metadata*); 3695 // We use T_VOID as marker for jump-table entries (labels) which 3696 // need an internal word relocation. 3697 case T_VOID: 3698 case T_ADDRESS: 3699 case T_OBJECT: return sizeof(jobject); 3700 } 3701 3702 ShouldNotReachHere(); 3703 return -1; 3704 } 3705 3706 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3707 // sort descending 3708 if (a->freq() > b->freq()) return -1; 3709 if (a->freq() < b->freq()) return 1; 3710 return 0; 3711 } 3712 3713 void Compile::ConstantTable::calculate_offsets_and_size() { 3714 // First, sort the array by frequencies. 3715 _constants.sort(qsort_comparator); 3716 3717 #ifdef ASSERT 3718 // Make sure all jump-table entries were sorted to the end of the 3719 // array (they have a negative frequency). 3720 bool found_void = false; 3721 for (int i = 0; i < _constants.length(); i++) { 3722 Constant con = _constants.at(i); 3723 if (con.type() == T_VOID) 3724 found_void = true; // jump-tables 3725 else 3726 assert(!found_void, "wrong sorting"); 3727 } 3728 #endif 3729 3730 int offset = 0; 3731 for (int i = 0; i < _constants.length(); i++) { 3732 Constant* con = _constants.adr_at(i); 3733 3734 // Align offset for type. 3735 int typesize = type_to_size_in_bytes(con->type()); 3736 offset = align_size_up(offset, typesize); 3737 con->set_offset(offset); // set constant's offset 3738 3739 if (con->type() == T_VOID) { 3740 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3741 offset = offset + typesize * n->outcnt(); // expand jump-table 3742 } else { 3743 offset = offset + typesize; 3744 } 3745 } 3746 3747 // Align size up to the next section start (which is insts; see 3748 // CodeBuffer::align_at_start). 3749 assert(_size == -1, "already set?"); 3750 _size = align_size_up(offset, CodeEntryAlignment); 3751 } 3752 3753 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3754 MacroAssembler _masm(&cb); 3755 for (int i = 0; i < _constants.length(); i++) { 3756 Constant con = _constants.at(i); 3757 address constant_addr; 3758 switch (con.type()) { 3759 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3760 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3761 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3762 case T_OBJECT: { 3763 jobject obj = con.get_jobject(); 3764 int oop_index = _masm.oop_recorder()->find_index(obj); 3765 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3766 break; 3767 } 3768 case T_ADDRESS: { 3769 address addr = (address) con.get_jobject(); 3770 constant_addr = _masm.address_constant(addr); 3771 break; 3772 } 3773 // We use T_VOID as marker for jump-table entries (labels) which 3774 // need an internal word relocation. 3775 case T_VOID: { 3776 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3777 // Fill the jump-table with a dummy word. The real value is 3778 // filled in later in fill_jump_table. 3779 address dummy = (address) n; 3780 constant_addr = _masm.address_constant(dummy); 3781 // Expand jump-table 3782 for (uint i = 1; i < n->outcnt(); i++) { 3783 address temp_addr = _masm.address_constant(dummy + i); 3784 assert(temp_addr, "consts section too small"); 3785 } 3786 break; 3787 } 3788 case T_METADATA: { 3789 Metadata* obj = con.get_metadata(); 3790 int metadata_index = _masm.oop_recorder()->find_index(obj); 3791 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3792 break; 3793 } 3794 default: ShouldNotReachHere(); 3795 } 3796 assert(constant_addr, "consts section too small"); 3797 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3798 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3799 } 3800 } 3801 3802 int Compile::ConstantTable::find_offset(Constant& con) const { 3803 int idx = _constants.find(con); 3804 assert(idx != -1, "constant must be in constant table"); 3805 int offset = _constants.at(idx).offset(); 3806 assert(offset != -1, "constant table not emitted yet?"); 3807 return offset; 3808 } 3809 3810 void Compile::ConstantTable::add(Constant& con) { 3811 if (con.can_be_reused()) { 3812 int idx = _constants.find(con); 3813 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3814 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3815 return; 3816 } 3817 } 3818 (void) _constants.append(con); 3819 } 3820 3821 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3822 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3823 Constant con(type, value, b->_freq); 3824 add(con); 3825 return con; 3826 } 3827 3828 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3829 Constant con(metadata); 3830 add(con); 3831 return con; 3832 } 3833 3834 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3835 jvalue value; 3836 BasicType type = oper->type()->basic_type(); 3837 switch (type) { 3838 case T_LONG: value.j = oper->constantL(); break; 3839 case T_FLOAT: value.f = oper->constantF(); break; 3840 case T_DOUBLE: value.d = oper->constantD(); break; 3841 case T_OBJECT: 3842 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3843 case T_METADATA: return add((Metadata*)oper->constant()); break; 3844 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3845 } 3846 return add(n, type, value); 3847 } 3848 3849 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3850 jvalue value; 3851 // We can use the node pointer here to identify the right jump-table 3852 // as this method is called from Compile::Fill_buffer right before 3853 // the MachNodes are emitted and the jump-table is filled (means the 3854 // MachNode pointers do not change anymore). 3855 value.l = (jobject) n; 3856 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3857 add(con); 3858 return con; 3859 } 3860 3861 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3862 // If called from Compile::scratch_emit_size do nothing. 3863 if (Compile::current()->in_scratch_emit_size()) return; 3864 3865 assert(labels.is_nonempty(), "must be"); 3866 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3867 3868 // Since MachConstantNode::constant_offset() also contains 3869 // table_base_offset() we need to subtract the table_base_offset() 3870 // to get the plain offset into the constant table. 3871 int offset = n->constant_offset() - table_base_offset(); 3872 3873 MacroAssembler _masm(&cb); 3874 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3875 3876 for (uint i = 0; i < n->outcnt(); i++) { 3877 address* constant_addr = &jump_table_base[i]; 3878 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i))); 3879 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3880 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3881 } 3882 } 3883 3884 //----------------------------static_subtype_check----------------------------- 3885 // Shortcut important common cases when superklass is exact: 3886 // (0) superklass is java.lang.Object (can occur in reflective code) 3887 // (1) subklass is already limited to a subtype of superklass => always ok 3888 // (2) subklass does not overlap with superklass => always fail 3889 // (3) superklass has NO subtypes and we can check with a simple compare. 3890 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 3891 if (StressReflectiveCode) { 3892 return SSC_full_test; // Let caller generate the general case. 3893 } 3894 3895 if (superk == env()->Object_klass()) { 3896 return SSC_always_true; // (0) this test cannot fail 3897 } 3898 3899 ciType* superelem = superk; 3900 if (superelem->is_array_klass()) 3901 superelem = superelem->as_array_klass()->base_element_type(); 3902 3903 if (!subk->is_interface()) { // cannot trust static interface types yet 3904 if (subk->is_subtype_of(superk)) { 3905 return SSC_always_true; // (1) false path dead; no dynamic test needed 3906 } 3907 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 3908 !superk->is_subtype_of(subk)) { 3909 return SSC_always_false; 3910 } 3911 } 3912 3913 // If casting to an instance klass, it must have no subtypes 3914 if (superk->is_interface()) { 3915 // Cannot trust interfaces yet. 3916 // %%% S.B. superk->nof_implementors() == 1 3917 } else if (superelem->is_instance_klass()) { 3918 ciInstanceKlass* ik = superelem->as_instance_klass(); 3919 if (!ik->has_subklass() && !ik->is_interface()) { 3920 if (!ik->is_final()) { 3921 // Add a dependency if there is a chance of a later subclass. 3922 dependencies()->assert_leaf_type(ik); 3923 } 3924 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3925 } 3926 } else { 3927 // A primitive array type has no subtypes. 3928 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3929 } 3930 3931 return SSC_full_test; 3932 } 3933 3934 Node* Compile::conv_I2X_index(PhaseGVN *phase, Node* idx, const TypeInt* sizetype) { 3935 #ifdef _LP64 3936 // The scaled index operand to AddP must be a clean 64-bit value. 3937 // Java allows a 32-bit int to be incremented to a negative 3938 // value, which appears in a 64-bit register as a large 3939 // positive number. Using that large positive number as an 3940 // operand in pointer arithmetic has bad consequences. 3941 // On the other hand, 32-bit overflow is rare, and the possibility 3942 // can often be excluded, if we annotate the ConvI2L node with 3943 // a type assertion that its value is known to be a small positive 3944 // number. (The prior range check has ensured this.) 3945 // This assertion is used by ConvI2LNode::Ideal. 3946 int index_max = max_jint - 1; // array size is max_jint, index is one less 3947 if (sizetype != NULL) index_max = sizetype->_hi - 1; 3948 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 3949 idx = phase->transform(new ConvI2LNode(idx, lidxtype)); 3950 #endif 3951 return idx; 3952 } 3953 3954 // The message about the current inlining is accumulated in 3955 // _print_inlining_stream and transfered into the _print_inlining_list 3956 // once we know whether inlining succeeds or not. For regular 3957 // inlining, messages are appended to the buffer pointed by 3958 // _print_inlining_idx in the _print_inlining_list. For late inlining, 3959 // a new buffer is added after _print_inlining_idx in the list. This 3960 // way we can update the inlining message for late inlining call site 3961 // when the inlining is attempted again. 3962 void Compile::print_inlining_init() { 3963 if (print_inlining() || print_intrinsics()) { 3964 _print_inlining_stream = new stringStream(); 3965 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 3966 } 3967 } 3968 3969 void Compile::print_inlining_reinit() { 3970 if (print_inlining() || print_intrinsics()) { 3971 // Re allocate buffer when we change ResourceMark 3972 _print_inlining_stream = new stringStream(); 3973 } 3974 } 3975 3976 void Compile::print_inlining_reset() { 3977 _print_inlining_stream->reset(); 3978 } 3979 3980 void Compile::print_inlining_commit() { 3981 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 3982 // Transfer the message from _print_inlining_stream to the current 3983 // _print_inlining_list buffer and clear _print_inlining_stream. 3984 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 3985 print_inlining_reset(); 3986 } 3987 3988 void Compile::print_inlining_push() { 3989 // Add new buffer to the _print_inlining_list at current position 3990 _print_inlining_idx++; 3991 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 3992 } 3993 3994 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 3995 return _print_inlining_list->at(_print_inlining_idx); 3996 } 3997 3998 void Compile::print_inlining_update(CallGenerator* cg) { 3999 if (print_inlining() || print_intrinsics()) { 4000 if (!cg->is_late_inline()) { 4001 if (print_inlining_current().cg() != NULL) { 4002 print_inlining_push(); 4003 } 4004 print_inlining_commit(); 4005 } else { 4006 if (print_inlining_current().cg() != cg && 4007 (print_inlining_current().cg() != NULL || 4008 print_inlining_current().ss()->size() != 0)) { 4009 print_inlining_push(); 4010 } 4011 print_inlining_commit(); 4012 print_inlining_current().set_cg(cg); 4013 } 4014 } 4015 } 4016 4017 void Compile::print_inlining_move_to(CallGenerator* cg) { 4018 // We resume inlining at a late inlining call site. Locate the 4019 // corresponding inlining buffer so that we can update it. 4020 if (print_inlining()) { 4021 for (int i = 0; i < _print_inlining_list->length(); i++) { 4022 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4023 _print_inlining_idx = i; 4024 return; 4025 } 4026 } 4027 ShouldNotReachHere(); 4028 } 4029 } 4030 4031 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4032 if (print_inlining()) { 4033 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4034 assert(print_inlining_current().cg() == cg, "wrong entry"); 4035 // replace message with new message 4036 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4037 print_inlining_commit(); 4038 print_inlining_current().set_cg(cg); 4039 } 4040 } 4041 4042 void Compile::print_inlining_assert_ready() { 4043 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4044 } 4045 4046 void Compile::process_print_inlining() { 4047 bool do_print_inlining = print_inlining() || print_intrinsics(); 4048 if (do_print_inlining || log() != NULL) { 4049 // Print inlining message for candidates that we couldn't inline 4050 // for lack of space 4051 for (int i = 0; i < _late_inlines.length(); i++) { 4052 CallGenerator* cg = _late_inlines.at(i); 4053 if (!cg->is_mh_late_inline()) { 4054 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4055 if (do_print_inlining) { 4056 cg->print_inlining_late(msg); 4057 } 4058 log_late_inline_failure(cg, msg); 4059 } 4060 } 4061 } 4062 if (do_print_inlining) { 4063 ResourceMark rm; 4064 stringStream ss; 4065 for (int i = 0; i < _print_inlining_list->length(); i++) { 4066 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4067 } 4068 size_t end = ss.size(); 4069 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4070 strncpy(_print_inlining_output, ss.base(), end+1); 4071 _print_inlining_output[end] = 0; 4072 } 4073 } 4074 4075 void Compile::dump_print_inlining() { 4076 if (_print_inlining_output != NULL) { 4077 tty->print_raw(_print_inlining_output); 4078 } 4079 } 4080 4081 void Compile::log_late_inline(CallGenerator* cg) { 4082 if (log() != NULL) { 4083 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4084 cg->unique_id()); 4085 JVMState* p = cg->call_node()->jvms(); 4086 while (p != NULL) { 4087 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4088 p = p->caller(); 4089 } 4090 log()->tail("late_inline"); 4091 } 4092 } 4093 4094 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4095 log_late_inline(cg); 4096 if (log() != NULL) { 4097 log()->inline_fail(msg); 4098 } 4099 } 4100 4101 void Compile::log_inline_id(CallGenerator* cg) { 4102 if (log() != NULL) { 4103 // The LogCompilation tool needs a unique way to identify late 4104 // inline call sites. This id must be unique for this call site in 4105 // this compilation. Try to have it unique across compilations as 4106 // well because it can be convenient when grepping through the log 4107 // file. 4108 // Distinguish OSR compilations from others in case CICountOSR is 4109 // on. 4110 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4111 cg->set_unique_id(id); 4112 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4113 } 4114 } 4115 4116 void Compile::log_inline_failure(const char* msg) { 4117 if (C->log() != NULL) { 4118 C->log()->inline_fail(msg); 4119 } 4120 } 4121 4122 4123 // Dump inlining replay data to the stream. 4124 // Don't change thread state and acquire any locks. 4125 void Compile::dump_inline_data(outputStream* out) { 4126 InlineTree* inl_tree = ilt(); 4127 if (inl_tree != NULL) { 4128 out->print(" inline %d", inl_tree->count()); 4129 inl_tree->dump_replay_data(out); 4130 } 4131 } 4132 4133 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4134 if (n1->Opcode() < n2->Opcode()) return -1; 4135 else if (n1->Opcode() > n2->Opcode()) return 1; 4136 4137 assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req())); 4138 for (uint i = 1; i < n1->req(); i++) { 4139 if (n1->in(i) < n2->in(i)) return -1; 4140 else if (n1->in(i) > n2->in(i)) return 1; 4141 } 4142 4143 return 0; 4144 } 4145 4146 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4147 Node* n1 = *n1p; 4148 Node* n2 = *n2p; 4149 4150 return cmp_expensive_nodes(n1, n2); 4151 } 4152 4153 void Compile::sort_expensive_nodes() { 4154 if (!expensive_nodes_sorted()) { 4155 _expensive_nodes->sort(cmp_expensive_nodes); 4156 } 4157 } 4158 4159 bool Compile::expensive_nodes_sorted() const { 4160 for (int i = 1; i < _expensive_nodes->length(); i++) { 4161 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4162 return false; 4163 } 4164 } 4165 return true; 4166 } 4167 4168 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4169 if (_expensive_nodes->length() == 0) { 4170 return false; 4171 } 4172 4173 assert(OptimizeExpensiveOps, "optimization off?"); 4174 4175 // Take this opportunity to remove dead nodes from the list 4176 int j = 0; 4177 for (int i = 0; i < _expensive_nodes->length(); i++) { 4178 Node* n = _expensive_nodes->at(i); 4179 if (!n->is_unreachable(igvn)) { 4180 assert(n->is_expensive(), "should be expensive"); 4181 _expensive_nodes->at_put(j, n); 4182 j++; 4183 } 4184 } 4185 _expensive_nodes->trunc_to(j); 4186 4187 // Then sort the list so that similar nodes are next to each other 4188 // and check for at least two nodes of identical kind with same data 4189 // inputs. 4190 sort_expensive_nodes(); 4191 4192 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4193 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4194 return true; 4195 } 4196 } 4197 4198 return false; 4199 } 4200 4201 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4202 if (_expensive_nodes->length() == 0) { 4203 return; 4204 } 4205 4206 assert(OptimizeExpensiveOps, "optimization off?"); 4207 4208 // Sort to bring similar nodes next to each other and clear the 4209 // control input of nodes for which there's only a single copy. 4210 sort_expensive_nodes(); 4211 4212 int j = 0; 4213 int identical = 0; 4214 int i = 0; 4215 bool modified = false; 4216 for (; i < _expensive_nodes->length()-1; i++) { 4217 assert(j <= i, "can't write beyond current index"); 4218 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4219 identical++; 4220 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4221 continue; 4222 } 4223 if (identical > 0) { 4224 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4225 identical = 0; 4226 } else { 4227 Node* n = _expensive_nodes->at(i); 4228 igvn.replace_input_of(n, 0, NULL); 4229 igvn.hash_insert(n); 4230 modified = true; 4231 } 4232 } 4233 if (identical > 0) { 4234 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4235 } else if (_expensive_nodes->length() >= 1) { 4236 Node* n = _expensive_nodes->at(i); 4237 igvn.replace_input_of(n, 0, NULL); 4238 igvn.hash_insert(n); 4239 modified = true; 4240 } 4241 _expensive_nodes->trunc_to(j); 4242 if (modified) { 4243 igvn.optimize(); 4244 } 4245 } 4246 4247 void Compile::add_expensive_node(Node * n) { 4248 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4249 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4250 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4251 if (OptimizeExpensiveOps) { 4252 _expensive_nodes->append(n); 4253 } else { 4254 // Clear control input and let IGVN optimize expensive nodes if 4255 // OptimizeExpensiveOps is off. 4256 n->set_req(0, NULL); 4257 } 4258 } 4259 4260 /** 4261 * Remove the speculative part of types and clean up the graph 4262 */ 4263 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4264 if (UseTypeSpeculation) { 4265 Unique_Node_List worklist; 4266 worklist.push(root()); 4267 int modified = 0; 4268 // Go over all type nodes that carry a speculative type, drop the 4269 // speculative part of the type and enqueue the node for an igvn 4270 // which may optimize it out. 4271 for (uint next = 0; next < worklist.size(); ++next) { 4272 Node *n = worklist.at(next); 4273 if (n->is_Type()) { 4274 TypeNode* tn = n->as_Type(); 4275 const Type* t = tn->type(); 4276 const Type* t_no_spec = t->remove_speculative(); 4277 if (t_no_spec != t) { 4278 bool in_hash = igvn.hash_delete(n); 4279 assert(in_hash, "node should be in igvn hash table"); 4280 tn->set_type(t_no_spec); 4281 igvn.hash_insert(n); 4282 igvn._worklist.push(n); // give it a chance to go away 4283 modified++; 4284 } 4285 } 4286 uint max = n->len(); 4287 for( uint i = 0; i < max; ++i ) { 4288 Node *m = n->in(i); 4289 if (not_a_node(m)) continue; 4290 worklist.push(m); 4291 } 4292 } 4293 // Drop the speculative part of all types in the igvn's type table 4294 igvn.remove_speculative_types(); 4295 if (modified > 0) { 4296 igvn.optimize(); 4297 } 4298 #ifdef ASSERT 4299 // Verify that after the IGVN is over no speculative type has resurfaced 4300 worklist.clear(); 4301 worklist.push(root()); 4302 for (uint next = 0; next < worklist.size(); ++next) { 4303 Node *n = worklist.at(next); 4304 const Type* t = igvn.type_or_null(n); 4305 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4306 if (n->is_Type()) { 4307 t = n->as_Type()->type(); 4308 assert(t == t->remove_speculative(), "no more speculative types"); 4309 } 4310 uint max = n->len(); 4311 for( uint i = 0; i < max; ++i ) { 4312 Node *m = n->in(i); 4313 if (not_a_node(m)) continue; 4314 worklist.push(m); 4315 } 4316 } 4317 igvn.check_no_speculative_types(); 4318 #endif 4319 } 4320 } 4321 4322 // Auxiliary method to support randomized stressing/fuzzing. 4323 // 4324 // This method can be called the arbitrary number of times, with current count 4325 // as the argument. The logic allows selecting a single candidate from the 4326 // running list of candidates as follows: 4327 // int count = 0; 4328 // Cand* selected = null; 4329 // while(cand = cand->next()) { 4330 // if (randomized_select(++count)) { 4331 // selected = cand; 4332 // } 4333 // } 4334 // 4335 // Including count equalizes the chances any candidate is "selected". 4336 // This is useful when we don't have the complete list of candidates to choose 4337 // from uniformly. In this case, we need to adjust the randomicity of the 4338 // selection, or else we will end up biasing the selection towards the latter 4339 // candidates. 4340 // 4341 // Quick back-envelope calculation shows that for the list of n candidates 4342 // the equal probability for the candidate to persist as "best" can be 4343 // achieved by replacing it with "next" k-th candidate with the probability 4344 // of 1/k. It can be easily shown that by the end of the run, the 4345 // probability for any candidate is converged to 1/n, thus giving the 4346 // uniform distribution among all the candidates. 4347 // 4348 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4349 #define RANDOMIZED_DOMAIN_POW 29 4350 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4351 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4352 bool Compile::randomized_select(int count) { 4353 assert(count > 0, "only positive"); 4354 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4355 } 4356 4357 const char* CloneMap::debug_option_name = "CloneMapDebug"; 4358 CloneMap& Compile::clone_map() { return _clone_map; } 4359 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4360 4361 void NodeCloneInfo::dump() const { 4362 tty->print(" {%d:%d} ", idx(), gen()); 4363 } 4364 4365 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4366 uint64_t val = value(old->_idx); 4367 NodeCloneInfo cio(val); 4368 assert(val != 0, "old node should be in the map"); 4369 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4370 insert(nnn->_idx, cin.get()); 4371 #ifndef PRODUCT 4372 if (is_debug()) { 4373 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4374 } 4375 #endif 4376 } 4377 4378 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4379 NodeCloneInfo cio(value(old->_idx)); 4380 if (cio.get() == 0) { 4381 cio.set(old->_idx, 0); 4382 insert(old->_idx, cio.get()); 4383 #ifndef PRODUCT 4384 if (is_debug()) { 4385 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4386 } 4387 #endif 4388 } 4389 clone(old, nnn, gen); 4390 } 4391 4392 int CloneMap::max_gen() const { 4393 int g = 0; 4394 DictI di(_dict); 4395 for(; di.test(); ++di) { 4396 int t = gen(di._key); 4397 if (g < t) { 4398 g = t; 4399 #ifndef PRODUCT 4400 if (is_debug()) { 4401 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4402 } 4403 #endif 4404 } 4405 } 4406 return g; 4407 } 4408 4409 void CloneMap::dump(node_idx_t key) const { 4410 uint64_t val = value(key); 4411 if (val != 0) { 4412 NodeCloneInfo ni(val); 4413 ni.dump(); 4414 } 4415 }