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/shenandoahSupport.hpp" 65 #include "opto/stringopts.hpp" 66 #include "opto/type.hpp" 67 #include "opto/vectornode.hpp" 68 #include "runtime/arguments.hpp" 69 #include "runtime/sharedRuntime.hpp" 70 #include "runtime/signature.hpp" 71 #include "runtime/stubRoutines.hpp" 72 #include "runtime/timer.hpp" 73 #include "utilities/copy.hpp" 74 75 76 // -------------------- Compile::mach_constant_base_node ----------------------- 77 // Constant table base node singleton. 78 MachConstantBaseNode* Compile::mach_constant_base_node() { 79 if (_mach_constant_base_node == NULL) { 80 _mach_constant_base_node = new MachConstantBaseNode(); 81 _mach_constant_base_node->add_req(C->root()); 82 } 83 return _mach_constant_base_node; 84 } 85 86 87 /// Support for intrinsics. 88 89 // Return the index at which m must be inserted (or already exists). 90 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 91 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 92 #ifdef ASSERT 93 for (int i = 1; i < _intrinsics->length(); i++) { 94 CallGenerator* cg1 = _intrinsics->at(i-1); 95 CallGenerator* cg2 = _intrinsics->at(i); 96 assert(cg1->method() != cg2->method() 97 ? cg1->method() < cg2->method() 98 : cg1->is_virtual() < cg2->is_virtual(), 99 "compiler intrinsics list must stay sorted"); 100 } 101 #endif 102 // Binary search sorted list, in decreasing intervals [lo, hi]. 103 int lo = 0, hi = _intrinsics->length()-1; 104 while (lo <= hi) { 105 int mid = (uint)(hi + lo) / 2; 106 ciMethod* mid_m = _intrinsics->at(mid)->method(); 107 if (m < mid_m) { 108 hi = mid-1; 109 } else if (m > mid_m) { 110 lo = mid+1; 111 } else { 112 // look at minor sort key 113 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 114 if (is_virtual < mid_virt) { 115 hi = mid-1; 116 } else if (is_virtual > mid_virt) { 117 lo = mid+1; 118 } else { 119 return mid; // exact match 120 } 121 } 122 } 123 return lo; // inexact match 124 } 125 126 void Compile::register_intrinsic(CallGenerator* cg) { 127 if (_intrinsics == NULL) { 128 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 129 } 130 // This code is stolen from ciObjectFactory::insert. 131 // Really, GrowableArray should have methods for 132 // insert_at, remove_at, and binary_search. 133 int len = _intrinsics->length(); 134 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 135 if (index == len) { 136 _intrinsics->append(cg); 137 } else { 138 #ifdef ASSERT 139 CallGenerator* oldcg = _intrinsics->at(index); 140 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 141 #endif 142 _intrinsics->append(_intrinsics->at(len-1)); 143 int pos; 144 for (pos = len-2; pos >= index; pos--) { 145 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 146 } 147 _intrinsics->at_put(index, cg); 148 } 149 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 150 } 151 152 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 153 assert(m->is_loaded(), "don't try this on unloaded methods"); 154 if (_intrinsics != NULL) { 155 int index = intrinsic_insertion_index(m, is_virtual); 156 if (index < _intrinsics->length() 157 && _intrinsics->at(index)->method() == m 158 && _intrinsics->at(index)->is_virtual() == is_virtual) { 159 return _intrinsics->at(index); 160 } 161 } 162 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 163 if (m->intrinsic_id() != vmIntrinsics::_none && 164 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 165 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 166 if (cg != NULL) { 167 // Save it for next time: 168 register_intrinsic(cg); 169 return cg; 170 } else { 171 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 172 } 173 } 174 return NULL; 175 } 176 177 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 178 // in library_call.cpp. 179 180 181 #ifndef PRODUCT 182 // statistics gathering... 183 184 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 185 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 186 187 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 188 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 189 int oflags = _intrinsic_hist_flags[id]; 190 assert(flags != 0, "what happened?"); 191 if (is_virtual) { 192 flags |= _intrinsic_virtual; 193 } 194 bool changed = (flags != oflags); 195 if ((flags & _intrinsic_worked) != 0) { 196 juint count = (_intrinsic_hist_count[id] += 1); 197 if (count == 1) { 198 changed = true; // first time 199 } 200 // increment the overall count also: 201 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 202 } 203 if (changed) { 204 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 205 // Something changed about the intrinsic's virtuality. 206 if ((flags & _intrinsic_virtual) != 0) { 207 // This is the first use of this intrinsic as a virtual call. 208 if (oflags != 0) { 209 // We already saw it as a non-virtual, so note both cases. 210 flags |= _intrinsic_both; 211 } 212 } else if ((oflags & _intrinsic_both) == 0) { 213 // This is the first use of this intrinsic as a non-virtual 214 flags |= _intrinsic_both; 215 } 216 } 217 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 218 } 219 // update the overall flags also: 220 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 221 return changed; 222 } 223 224 static char* format_flags(int flags, char* buf) { 225 buf[0] = 0; 226 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 227 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 228 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 229 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 230 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 231 if (buf[0] == 0) strcat(buf, ","); 232 assert(buf[0] == ',', "must be"); 233 return &buf[1]; 234 } 235 236 void Compile::print_intrinsic_statistics() { 237 char flagsbuf[100]; 238 ttyLocker ttyl; 239 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 240 tty->print_cr("Compiler intrinsic usage:"); 241 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 242 if (total == 0) total = 1; // avoid div0 in case of no successes 243 #define PRINT_STAT_LINE(name, c, f) \ 244 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 245 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 246 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 247 int flags = _intrinsic_hist_flags[id]; 248 juint count = _intrinsic_hist_count[id]; 249 if ((flags | count) != 0) { 250 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 251 } 252 } 253 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 254 if (xtty != NULL) xtty->tail("statistics"); 255 } 256 257 void Compile::print_statistics() { 258 { ttyLocker ttyl; 259 if (xtty != NULL) xtty->head("statistics type='opto'"); 260 Parse::print_statistics(); 261 PhaseCCP::print_statistics(); 262 PhaseRegAlloc::print_statistics(); 263 Scheduling::print_statistics(); 264 PhasePeephole::print_statistics(); 265 PhaseIdealLoop::print_statistics(); 266 if (xtty != NULL) xtty->tail("statistics"); 267 } 268 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 269 // put this under its own <statistics> element. 270 print_intrinsic_statistics(); 271 } 272 } 273 #endif //PRODUCT 274 275 // Support for bundling info 276 Bundle* Compile::node_bundling(const Node *n) { 277 assert(valid_bundle_info(n), "oob"); 278 return &_node_bundling_base[n->_idx]; 279 } 280 281 bool Compile::valid_bundle_info(const Node *n) { 282 return (_node_bundling_limit > n->_idx); 283 } 284 285 286 void Compile::gvn_replace_by(Node* n, Node* nn) { 287 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 288 Node* use = n->last_out(i); 289 bool is_in_table = initial_gvn()->hash_delete(use); 290 uint uses_found = 0; 291 for (uint j = 0; j < use->len(); j++) { 292 if (use->in(j) == n) { 293 if (j < use->req()) 294 use->set_req(j, nn); 295 else 296 use->set_prec(j, nn); 297 uses_found++; 298 } 299 } 300 if (is_in_table) { 301 // reinsert into table 302 initial_gvn()->hash_find_insert(use); 303 } 304 record_for_igvn(use); 305 i -= uses_found; // we deleted 1 or more copies of this edge 306 } 307 } 308 309 310 static inline bool not_a_node(const Node* n) { 311 if (n == NULL) return true; 312 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 313 if (*(address*)n == badAddress) return true; // kill by Node::destruct 314 return false; 315 } 316 317 // Identify all nodes that are reachable from below, useful. 318 // Use breadth-first pass that records state in a Unique_Node_List, 319 // recursive traversal is slower. 320 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 321 int estimated_worklist_size = unique(); 322 useful.map( estimated_worklist_size, NULL ); // preallocate space 323 324 // Initialize worklist 325 if (root() != NULL) { useful.push(root()); } 326 // If 'top' is cached, declare it useful to preserve cached node 327 if( cached_top_node() ) { useful.push(cached_top_node()); } 328 329 // Push all useful nodes onto the list, breadthfirst 330 for( uint next = 0; next < useful.size(); ++next ) { 331 assert( next < unique(), "Unique useful nodes < total nodes"); 332 Node *n = useful.at(next); 333 uint max = n->len(); 334 for( uint i = 0; i < max; ++i ) { 335 Node *m = n->in(i); 336 if (not_a_node(m)) continue; 337 useful.push(m); 338 } 339 } 340 } 341 342 // Update dead_node_list with any missing dead nodes using useful 343 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 344 void Compile::update_dead_node_list(Unique_Node_List &useful) { 345 uint max_idx = unique(); 346 VectorSet& useful_node_set = useful.member_set(); 347 348 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 349 // If node with index node_idx is not in useful set, 350 // mark it as dead in dead node list. 351 if (! useful_node_set.test(node_idx) ) { 352 record_dead_node(node_idx); 353 } 354 } 355 } 356 357 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 358 int shift = 0; 359 for (int i = 0; i < inlines->length(); i++) { 360 CallGenerator* cg = inlines->at(i); 361 CallNode* call = cg->call_node(); 362 if (shift > 0) { 363 inlines->at_put(i-shift, cg); 364 } 365 if (!useful.member(call)) { 366 shift++; 367 } 368 } 369 inlines->trunc_to(inlines->length()-shift); 370 } 371 372 // Disconnect all useless nodes by disconnecting those at the boundary. 373 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 374 uint next = 0; 375 while (next < useful.size()) { 376 Node *n = useful.at(next++); 377 if (n->is_SafePoint()) { 378 // We're done with a parsing phase. Replaced nodes are not valid 379 // beyond that point. 380 n->as_SafePoint()->delete_replaced_nodes(); 381 } 382 // Use raw traversal of out edges since this code removes out edges 383 int max = n->outcnt(); 384 for (int j = 0; j < max; ++j) { 385 Node* child = n->raw_out(j); 386 if (! useful.member(child)) { 387 assert(!child->is_top() || child != top(), 388 "If top is cached in Compile object it is in useful list"); 389 // Only need to remove this out-edge to the useless node 390 n->raw_del_out(j); 391 --j; 392 --max; 393 } 394 } 395 if (n->outcnt() == 1 && n->has_special_unique_user()) { 396 record_for_igvn(n->unique_out()); 397 } 398 } 399 // Remove useless macro and predicate opaq nodes 400 for (int i = C->macro_count()-1; i >= 0; i--) { 401 Node* n = C->macro_node(i); 402 if (!useful.member(n)) { 403 remove_macro_node(n); 404 } 405 } 406 // Remove useless expensive node 407 for (int i = C->expensive_count()-1; i >= 0; i--) { 408 Node* n = C->expensive_node(i); 409 if (!useful.member(n)) { 410 remove_expensive_node(n); 411 } 412 } 413 // clean up the late inline lists 414 remove_useless_late_inlines(&_string_late_inlines, useful); 415 remove_useless_late_inlines(&_boxing_late_inlines, useful); 416 remove_useless_late_inlines(&_late_inlines, useful); 417 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 418 } 419 420 //------------------------------frame_size_in_words----------------------------- 421 // frame_slots in units of words 422 int Compile::frame_size_in_words() const { 423 // shift is 0 in LP32 and 1 in LP64 424 const int shift = (LogBytesPerWord - LogBytesPerInt); 425 int words = _frame_slots >> shift; 426 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 427 return words; 428 } 429 430 // To bang the stack of this compiled method we use the stack size 431 // that the interpreter would need in case of a deoptimization. This 432 // removes the need to bang the stack in the deoptimization blob which 433 // in turn simplifies stack overflow handling. 434 int Compile::bang_size_in_bytes() const { 435 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 436 } 437 438 // ============================================================================ 439 //------------------------------CompileWrapper--------------------------------- 440 class CompileWrapper : public StackObj { 441 Compile *const _compile; 442 public: 443 CompileWrapper(Compile* compile); 444 445 ~CompileWrapper(); 446 }; 447 448 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 449 // the Compile* pointer is stored in the current ciEnv: 450 ciEnv* env = compile->env(); 451 assert(env == ciEnv::current(), "must already be a ciEnv active"); 452 assert(env->compiler_data() == NULL, "compile already active?"); 453 env->set_compiler_data(compile); 454 assert(compile == Compile::current(), "sanity"); 455 456 compile->set_type_dict(NULL); 457 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 458 compile->clone_map().set_clone_idx(0); 459 compile->set_type_hwm(NULL); 460 compile->set_type_last_size(0); 461 compile->set_last_tf(NULL, NULL); 462 compile->set_indexSet_arena(NULL); 463 compile->set_indexSet_free_block_list(NULL); 464 compile->init_type_arena(); 465 Type::Initialize(compile); 466 _compile->set_scratch_buffer_blob(NULL); 467 _compile->begin_method(); 468 _compile->clone_map().set_debug(_compile->has_method() && _compile->method_has_option(_compile->clone_map().debug_option_name)); 469 } 470 CompileWrapper::~CompileWrapper() { 471 _compile->end_method(); 472 if (_compile->scratch_buffer_blob() != NULL) 473 BufferBlob::free(_compile->scratch_buffer_blob()); 474 _compile->env()->set_compiler_data(NULL); 475 } 476 477 478 //----------------------------print_compile_messages--------------------------- 479 void Compile::print_compile_messages() { 480 #ifndef PRODUCT 481 // Check if recompiling 482 if (_subsume_loads == false && PrintOpto) { 483 // Recompiling without allowing machine instructions to subsume loads 484 tty->print_cr("*********************************************************"); 485 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 486 tty->print_cr("*********************************************************"); 487 } 488 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 489 // Recompiling without escape analysis 490 tty->print_cr("*********************************************************"); 491 tty->print_cr("** Bailout: Recompile without escape analysis **"); 492 tty->print_cr("*********************************************************"); 493 } 494 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 495 // Recompiling without boxing elimination 496 tty->print_cr("*********************************************************"); 497 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 498 tty->print_cr("*********************************************************"); 499 } 500 if (env()->break_at_compile()) { 501 // Open the debugger when compiling this method. 502 tty->print("### Breaking when compiling: "); 503 method()->print_short_name(); 504 tty->cr(); 505 BREAKPOINT; 506 } 507 508 if( PrintOpto ) { 509 if (is_osr_compilation()) { 510 tty->print("[OSR]%3d", _compile_id); 511 } else { 512 tty->print("%3d", _compile_id); 513 } 514 } 515 #endif 516 } 517 518 519 //-----------------------init_scratch_buffer_blob------------------------------ 520 // Construct a temporary BufferBlob and cache it for this compile. 521 void Compile::init_scratch_buffer_blob(int const_size) { 522 // If there is already a scratch buffer blob allocated and the 523 // constant section is big enough, use it. Otherwise free the 524 // current and allocate a new one. 525 BufferBlob* blob = scratch_buffer_blob(); 526 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 527 // Use the current blob. 528 } else { 529 if (blob != NULL) { 530 BufferBlob::free(blob); 531 } 532 533 ResourceMark rm; 534 _scratch_const_size = const_size; 535 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 536 blob = BufferBlob::create("Compile::scratch_buffer", size); 537 // Record the buffer blob for next time. 538 set_scratch_buffer_blob(blob); 539 // Have we run out of code space? 540 if (scratch_buffer_blob() == NULL) { 541 // Let CompilerBroker disable further compilations. 542 record_failure("Not enough space for scratch buffer in CodeCache"); 543 return; 544 } 545 } 546 547 // Initialize the relocation buffers 548 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 549 set_scratch_locs_memory(locs_buf); 550 } 551 552 553 //-----------------------scratch_emit_size------------------------------------- 554 // Helper function that computes size by emitting code 555 uint Compile::scratch_emit_size(const Node* n) { 556 // Start scratch_emit_size section. 557 set_in_scratch_emit_size(true); 558 559 // Emit into a trash buffer and count bytes emitted. 560 // This is a pretty expensive way to compute a size, 561 // but it works well enough if seldom used. 562 // All common fixed-size instructions are given a size 563 // method by the AD file. 564 // Note that the scratch buffer blob and locs memory are 565 // allocated at the beginning of the compile task, and 566 // may be shared by several calls to scratch_emit_size. 567 // The allocation of the scratch buffer blob is particularly 568 // expensive, since it has to grab the code cache lock. 569 BufferBlob* blob = this->scratch_buffer_blob(); 570 assert(blob != NULL, "Initialize BufferBlob at start"); 571 assert(blob->size() > MAX_inst_size, "sanity"); 572 relocInfo* locs_buf = scratch_locs_memory(); 573 address blob_begin = blob->content_begin(); 574 address blob_end = (address)locs_buf; 575 assert(blob->content_contains(blob_end), "sanity"); 576 CodeBuffer buf(blob_begin, blob_end - blob_begin); 577 buf.initialize_consts_size(_scratch_const_size); 578 buf.initialize_stubs_size(MAX_stubs_size); 579 assert(locs_buf != NULL, "sanity"); 580 int lsize = MAX_locs_size / 3; 581 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 582 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 583 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 584 585 // Do the emission. 586 587 Label fakeL; // Fake label for branch instructions. 588 Label* saveL = NULL; 589 uint save_bnum = 0; 590 bool is_branch = n->is_MachBranch(); 591 if (is_branch) { 592 MacroAssembler masm(&buf); 593 masm.bind(fakeL); 594 n->as_MachBranch()->save_label(&saveL, &save_bnum); 595 n->as_MachBranch()->label_set(&fakeL, 0); 596 } 597 n->emit(buf, this->regalloc()); 598 599 // Emitting into the scratch buffer should not fail 600 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 601 602 if (is_branch) // Restore label. 603 n->as_MachBranch()->label_set(saveL, save_bnum); 604 605 // End scratch_emit_size section. 606 set_in_scratch_emit_size(false); 607 608 return buf.insts_size(); 609 } 610 611 612 // ============================================================================ 613 //------------------------------Compile standard------------------------------- 614 debug_only( int Compile::_debug_idx = 100000; ) 615 616 // Compile a method. entry_bci is -1 for normal compilations and indicates 617 // the continuation bci for on stack replacement. 618 619 620 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 621 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 622 : Phase(Compiler), 623 _env(ci_env), 624 _log(ci_env->log()), 625 _compile_id(ci_env->compile_id()), 626 _save_argument_registers(false), 627 _stub_name(NULL), 628 _stub_function(NULL), 629 _stub_entry_point(NULL), 630 _method(target), 631 _entry_bci(osr_bci), 632 _initial_gvn(NULL), 633 _for_igvn(NULL), 634 _warm_calls(NULL), 635 _subsume_loads(subsume_loads), 636 _do_escape_analysis(do_escape_analysis), 637 _eliminate_boxing(eliminate_boxing), 638 _failure_reason(NULL), 639 _code_buffer("Compile::Fill_buffer"), 640 _orig_pc_slot(0), 641 _orig_pc_slot_offset_in_bytes(0), 642 _has_method_handle_invokes(false), 643 _mach_constant_base_node(NULL), 644 _node_bundling_limit(0), 645 _node_bundling_base(NULL), 646 _java_calls(0), 647 _inner_loops(0), 648 _scratch_const_size(-1), 649 _in_scratch_emit_size(false), 650 _dead_node_list(comp_arena()), 651 _dead_node_count(0), 652 #ifndef PRODUCT 653 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 654 _in_dump_cnt(0), 655 _printer(IdealGraphPrinter::printer()), 656 #endif 657 _congraph(NULL), 658 _comp_arena(mtCompiler), 659 _node_arena(mtCompiler), 660 _old_arena(mtCompiler), 661 _Compile_types(mtCompiler), 662 _replay_inline_data(NULL), 663 _late_inlines(comp_arena(), 2, 0, NULL), 664 _string_late_inlines(comp_arena(), 2, 0, NULL), 665 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 666 _late_inlines_pos(0), 667 _number_of_mh_late_inlines(0), 668 _inlining_progress(false), 669 _inlining_incrementally(false), 670 _print_inlining_list(NULL), 671 _print_inlining_stream(NULL), 672 _print_inlining_idx(0), 673 _print_inlining_output(NULL), 674 _interpreter_frame_size(0), 675 _max_node_limit(MaxNodeLimit) { 676 C = this; 677 678 CompileWrapper cw(this); 679 680 if (CITimeVerbose) { 681 tty->print(" "); 682 target->holder()->name()->print(); 683 tty->print("."); 684 target->print_short_name(); 685 tty->print(" "); 686 } 687 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 688 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 689 690 #ifndef PRODUCT 691 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 692 if (!print_opto_assembly) { 693 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 694 if (print_assembly && !Disassembler::can_decode()) { 695 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 696 print_opto_assembly = true; 697 } 698 } 699 set_print_assembly(print_opto_assembly); 700 set_parsed_irreducible_loop(false); 701 702 if (method()->has_option("ReplayInline")) { 703 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 704 } 705 #endif 706 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 707 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 708 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 709 710 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 711 // Make sure the method being compiled gets its own MDO, 712 // so we can at least track the decompile_count(). 713 // Need MDO to record RTM code generation state. 714 method()->ensure_method_data(); 715 } 716 717 Init(::AliasLevel); 718 719 720 print_compile_messages(); 721 722 _ilt = InlineTree::build_inline_tree_root(); 723 724 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 725 assert(num_alias_types() >= AliasIdxRaw, ""); 726 727 #define MINIMUM_NODE_HASH 1023 728 // Node list that Iterative GVN will start with 729 Unique_Node_List for_igvn(comp_arena()); 730 set_for_igvn(&for_igvn); 731 732 // GVN that will be run immediately on new nodes 733 uint estimated_size = method()->code_size()*4+64; 734 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 735 PhaseGVN gvn(node_arena(), estimated_size); 736 set_initial_gvn(&gvn); 737 738 print_inlining_init(); 739 { // Scope for timing the parser 740 TracePhase tp("parse", &timers[_t_parser]); 741 742 // Put top into the hash table ASAP. 743 initial_gvn()->transform_no_reclaim(top()); 744 745 // Set up tf(), start(), and find a CallGenerator. 746 CallGenerator* cg = NULL; 747 if (is_osr_compilation()) { 748 const TypeTuple *domain = StartOSRNode::osr_domain(); 749 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 750 init_tf(TypeFunc::make(domain, range)); 751 StartNode* s = new StartOSRNode(root(), domain); 752 initial_gvn()->set_type_bottom(s); 753 init_start(s); 754 cg = CallGenerator::for_osr(method(), entry_bci()); 755 } else { 756 // Normal case. 757 init_tf(TypeFunc::make(method())); 758 StartNode* s = new StartNode(root(), tf()->domain()); 759 initial_gvn()->set_type_bottom(s); 760 init_start(s); 761 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get 762 && (UseG1GC || UseShenandoahGC)) { 763 // With java.lang.ref.reference.get() we must go through the 764 // intrinsic when G1 is enabled - even when get() is the root 765 // method of the compile - so that, if necessary, the value in 766 // the referent field of the reference object gets recorded by 767 // the pre-barrier code. 768 // Specifically, if G1 is enabled, the value in the referent 769 // field is recorded by the G1 SATB pre barrier. This will 770 // result in the referent being marked live and the reference 771 // object removed from the list of discovered references during 772 // reference processing. 773 cg = find_intrinsic(method(), false); 774 } 775 if (cg == NULL) { 776 float past_uses = method()->interpreter_invocation_count(); 777 float expected_uses = past_uses; 778 cg = CallGenerator::for_inline(method(), expected_uses); 779 } 780 } 781 if (failing()) return; 782 if (cg == NULL) { 783 record_method_not_compilable_all_tiers("cannot parse method"); 784 return; 785 } 786 JVMState* jvms = build_start_state(start(), tf()); 787 if ((jvms = cg->generate(jvms)) == NULL) { 788 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 789 record_method_not_compilable("method parse failed"); 790 } 791 return; 792 } 793 GraphKit kit(jvms); 794 795 if (!kit.stopped()) { 796 // Accept return values, and transfer control we know not where. 797 // This is done by a special, unique ReturnNode bound to root. 798 return_values(kit.jvms()); 799 } 800 801 if (kit.has_exceptions()) { 802 // Any exceptions that escape from this call must be rethrown 803 // to whatever caller is dynamically above us on the stack. 804 // This is done by a special, unique RethrowNode bound to root. 805 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 806 } 807 808 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 809 810 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 811 inline_string_calls(true); 812 } 813 814 if (failing()) return; 815 816 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 817 818 // Remove clutter produced by parsing. 819 if (!failing()) { 820 ResourceMark rm; 821 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 822 } 823 } 824 825 // Note: Large methods are capped off in do_one_bytecode(). 826 if (failing()) return; 827 828 // After parsing, node notes are no longer automagic. 829 // They must be propagated by register_new_node_with_optimizer(), 830 // clone(), or the like. 831 set_default_node_notes(NULL); 832 833 for (;;) { 834 int successes = Inline_Warm(); 835 if (failing()) return; 836 if (successes == 0) break; 837 } 838 839 // Drain the list. 840 Finish_Warm(); 841 #ifndef PRODUCT 842 if (_printer && _printer->should_print(_method)) { 843 _printer->print_inlining(this); 844 } 845 #endif 846 847 if (failing()) return; 848 NOT_PRODUCT( verify_graph_edges(); ) 849 850 // Now optimize 851 Optimize(); 852 if (failing()) return; 853 NOT_PRODUCT( verify_graph_edges(); ) 854 855 #ifndef PRODUCT 856 if (PrintIdeal) { 857 ttyLocker ttyl; // keep the following output all in one block 858 // This output goes directly to the tty, not the compiler log. 859 // To enable tools to match it up with the compilation activity, 860 // be sure to tag this tty output with the compile ID. 861 if (xtty != NULL) { 862 xtty->head("ideal compile_id='%d'%s", compile_id(), 863 is_osr_compilation() ? " compile_kind='osr'" : 864 ""); 865 } 866 root()->dump(9999); 867 if (xtty != NULL) { 868 xtty->tail("ideal"); 869 } 870 } 871 #endif 872 873 NOT_PRODUCT( verify_barriers(); ) 874 875 // Dump compilation data to replay it. 876 if (method()->has_option("DumpReplay")) { 877 env()->dump_replay_data(_compile_id); 878 } 879 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 880 env()->dump_inline_data(_compile_id); 881 } 882 883 // Now that we know the size of all the monitors we can add a fixed slot 884 // for the original deopt pc. 885 886 _orig_pc_slot = fixed_slots(); 887 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 888 set_fixed_slots(next_slot); 889 890 // Compute when to use implicit null checks. Used by matching trap based 891 // nodes and NullCheck optimization. 892 set_allowed_deopt_reasons(); 893 894 // Now generate code 895 Code_Gen(); 896 if (failing()) return; 897 898 // Check if we want to skip execution of all compiled code. 899 { 900 #ifndef PRODUCT 901 if (OptoNoExecute) { 902 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 903 return; 904 } 905 #endif 906 TracePhase tp("install_code", &timers[_t_registerMethod]); 907 908 if (is_osr_compilation()) { 909 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 910 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 911 } else { 912 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 913 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 914 } 915 916 env()->register_method(_method, _entry_bci, 917 &_code_offsets, 918 _orig_pc_slot_offset_in_bytes, 919 code_buffer(), 920 frame_size_in_words(), _oop_map_set, 921 &_handler_table, &_inc_table, 922 compiler, 923 env()->comp_level(), 924 has_unsafe_access(), 925 SharedRuntime::is_wide_vector(max_vector_size()), 926 rtm_state() 927 ); 928 929 if (log() != NULL) // Print code cache state into compiler log 930 log()->code_cache_state(); 931 } 932 } 933 934 //------------------------------Compile---------------------------------------- 935 // Compile a runtime stub 936 Compile::Compile( ciEnv* ci_env, 937 TypeFunc_generator generator, 938 address stub_function, 939 const char *stub_name, 940 int is_fancy_jump, 941 bool pass_tls, 942 bool save_arg_registers, 943 bool return_pc ) 944 : Phase(Compiler), 945 _env(ci_env), 946 _log(ci_env->log()), 947 _compile_id(0), 948 _save_argument_registers(save_arg_registers), 949 _method(NULL), 950 _stub_name(stub_name), 951 _stub_function(stub_function), 952 _stub_entry_point(NULL), 953 _entry_bci(InvocationEntryBci), 954 _initial_gvn(NULL), 955 _for_igvn(NULL), 956 _warm_calls(NULL), 957 _orig_pc_slot(0), 958 _orig_pc_slot_offset_in_bytes(0), 959 _subsume_loads(true), 960 _do_escape_analysis(false), 961 _eliminate_boxing(false), 962 _failure_reason(NULL), 963 _code_buffer("Compile::Fill_buffer"), 964 _has_method_handle_invokes(false), 965 _mach_constant_base_node(NULL), 966 _node_bundling_limit(0), 967 _node_bundling_base(NULL), 968 _java_calls(0), 969 _inner_loops(0), 970 #ifndef PRODUCT 971 _trace_opto_output(TraceOptoOutput), 972 _in_dump_cnt(0), 973 _printer(NULL), 974 #endif 975 _comp_arena(mtCompiler), 976 _node_arena(mtCompiler), 977 _old_arena(mtCompiler), 978 _Compile_types(mtCompiler), 979 _dead_node_list(comp_arena()), 980 _dead_node_count(0), 981 _congraph(NULL), 982 _replay_inline_data(NULL), 983 _number_of_mh_late_inlines(0), 984 _inlining_progress(false), 985 _inlining_incrementally(false), 986 _print_inlining_list(NULL), 987 _print_inlining_stream(NULL), 988 _print_inlining_idx(0), 989 _print_inlining_output(NULL), 990 _allowed_reasons(0), 991 _interpreter_frame_size(0), 992 _max_node_limit(MaxNodeLimit) { 993 C = this; 994 995 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 996 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 997 998 #ifndef PRODUCT 999 set_print_assembly(PrintFrameConverterAssembly); 1000 set_parsed_irreducible_loop(false); 1001 #endif 1002 set_has_irreducible_loop(false); // no loops 1003 1004 CompileWrapper cw(this); 1005 Init(/*AliasLevel=*/ 0); 1006 init_tf((*generator)()); 1007 1008 { 1009 // The following is a dummy for the sake of GraphKit::gen_stub 1010 Unique_Node_List for_igvn(comp_arena()); 1011 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1012 PhaseGVN gvn(Thread::current()->resource_area(),255); 1013 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1014 gvn.transform_no_reclaim(top()); 1015 1016 GraphKit kit; 1017 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1018 } 1019 1020 NOT_PRODUCT( verify_graph_edges(); ) 1021 Code_Gen(); 1022 if (failing()) return; 1023 1024 1025 // Entry point will be accessed using compile->stub_entry_point(); 1026 if (code_buffer() == NULL) { 1027 Matcher::soft_match_failure(); 1028 } else { 1029 if (PrintAssembly && (WizardMode || Verbose)) 1030 tty->print_cr("### Stub::%s", stub_name); 1031 1032 if (!failing()) { 1033 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1034 1035 // Make the NMethod 1036 // For now we mark the frame as never safe for profile stackwalking 1037 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1038 code_buffer(), 1039 CodeOffsets::frame_never_safe, 1040 // _code_offsets.value(CodeOffsets::Frame_Complete), 1041 frame_size_in_words(), 1042 _oop_map_set, 1043 save_arg_registers); 1044 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1045 1046 _stub_entry_point = rs->entry_point(); 1047 } 1048 } 1049 } 1050 1051 //------------------------------Init------------------------------------------- 1052 // Prepare for a single compilation 1053 void Compile::Init(int aliaslevel) { 1054 _unique = 0; 1055 _regalloc = NULL; 1056 1057 _tf = NULL; // filled in later 1058 _top = NULL; // cached later 1059 _matcher = NULL; // filled in later 1060 _cfg = NULL; // filled in later 1061 1062 set_24_bit_selection_and_mode(Use24BitFP, false); 1063 1064 _node_note_array = NULL; 1065 _default_node_notes = NULL; 1066 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1067 1068 _immutable_memory = NULL; // filled in at first inquiry 1069 1070 // Globally visible Nodes 1071 // First set TOP to NULL to give safe behavior during creation of RootNode 1072 set_cached_top_node(NULL); 1073 set_root(new RootNode()); 1074 // Now that you have a Root to point to, create the real TOP 1075 set_cached_top_node( new ConNode(Type::TOP) ); 1076 set_recent_alloc(NULL, NULL); 1077 1078 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1079 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1080 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1081 env()->set_dependencies(new Dependencies(env())); 1082 1083 _fixed_slots = 0; 1084 set_has_split_ifs(false); 1085 set_has_loops(has_method() && method()->has_loops()); // first approximation 1086 set_has_stringbuilder(false); 1087 set_has_boxed_value(false); 1088 _trap_can_recompile = false; // no traps emitted yet 1089 _major_progress = true; // start out assuming good things will happen 1090 set_has_unsafe_access(false); 1091 set_max_vector_size(0); 1092 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1093 set_decompile_count(0); 1094 1095 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1096 set_num_loop_opts(LoopOptsCount); 1097 set_do_inlining(Inline); 1098 set_max_inline_size(MaxInlineSize); 1099 set_freq_inline_size(FreqInlineSize); 1100 set_do_scheduling(OptoScheduling); 1101 set_do_count_invocations(false); 1102 set_do_method_data_update(false); 1103 1104 set_do_vector_loop(false); 1105 1106 bool do_vector = false; 1107 if (AllowVectorizeOnDemand) { 1108 if (has_method() && (method()->has_option("Vectorize") || method()->has_option("VectorizeDebug"))) { 1109 set_do_vector_loop(true); 1110 } else if (has_method() && method()->name() != 0 && 1111 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1112 set_do_vector_loop(true); 1113 } 1114 #ifndef PRODUCT 1115 if (do_vector_loop() && Verbose) { 1116 tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii()); 1117 } 1118 #endif 1119 } 1120 1121 set_age_code(has_method() && method()->profile_aging()); 1122 set_rtm_state(NoRTM); // No RTM lock eliding by default 1123 method_has_option_value("MaxNodeLimit", _max_node_limit); 1124 #if INCLUDE_RTM_OPT 1125 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1126 int rtm_state = method()->method_data()->rtm_state(); 1127 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1128 // Don't generate RTM lock eliding code. 1129 set_rtm_state(NoRTM); 1130 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1131 // Generate RTM lock eliding code without abort ratio calculation code. 1132 set_rtm_state(UseRTM); 1133 } else if (UseRTMDeopt) { 1134 // Generate RTM lock eliding code and include abort ratio calculation 1135 // code if UseRTMDeopt is on. 1136 set_rtm_state(ProfileRTM); 1137 } 1138 } 1139 #endif 1140 if (debug_info()->recording_non_safepoints()) { 1141 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1142 (comp_arena(), 8, 0, NULL)); 1143 set_default_node_notes(Node_Notes::make(this)); 1144 } 1145 1146 // // -- Initialize types before each compile -- 1147 // // Update cached type information 1148 // if( _method && _method->constants() ) 1149 // Type::update_loaded_types(_method, _method->constants()); 1150 1151 // Init alias_type map. 1152 if (!_do_escape_analysis && aliaslevel == 3) 1153 aliaslevel = 2; // No unique types without escape analysis 1154 _AliasLevel = aliaslevel; 1155 const int grow_ats = 16; 1156 _max_alias_types = grow_ats; 1157 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1158 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1159 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1160 { 1161 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1162 } 1163 // Initialize the first few types. 1164 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1165 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1166 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1167 _num_alias_types = AliasIdxRaw+1; 1168 // Zero out the alias type cache. 1169 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1170 // A NULL adr_type hits in the cache right away. Preload the right answer. 1171 probe_alias_cache(NULL)->_index = AliasIdxTop; 1172 1173 _intrinsics = NULL; 1174 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1175 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1176 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1177 register_library_intrinsics(); 1178 } 1179 1180 //---------------------------init_start---------------------------------------- 1181 // Install the StartNode on this compile object. 1182 void Compile::init_start(StartNode* s) { 1183 if (failing()) 1184 return; // already failing 1185 assert(s == start(), ""); 1186 } 1187 1188 /** 1189 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1190 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1191 * the ideal graph. 1192 */ 1193 StartNode* Compile::start() const { 1194 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 1195 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1196 Node* start = root()->fast_out(i); 1197 if (start->is_Start()) { 1198 return start->as_Start(); 1199 } 1200 } 1201 fatal("Did not find Start node!"); 1202 return NULL; 1203 } 1204 1205 //-------------------------------immutable_memory------------------------------------- 1206 // Access immutable memory 1207 Node* Compile::immutable_memory() { 1208 if (_immutable_memory != NULL) { 1209 return _immutable_memory; 1210 } 1211 StartNode* s = start(); 1212 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1213 Node *p = s->fast_out(i); 1214 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1215 _immutable_memory = p; 1216 return _immutable_memory; 1217 } 1218 } 1219 ShouldNotReachHere(); 1220 return NULL; 1221 } 1222 1223 //----------------------set_cached_top_node------------------------------------ 1224 // Install the cached top node, and make sure Node::is_top works correctly. 1225 void Compile::set_cached_top_node(Node* tn) { 1226 if (tn != NULL) verify_top(tn); 1227 Node* old_top = _top; 1228 _top = tn; 1229 // Calling Node::setup_is_top allows the nodes the chance to adjust 1230 // their _out arrays. 1231 if (_top != NULL) _top->setup_is_top(); 1232 if (old_top != NULL) old_top->setup_is_top(); 1233 assert(_top == NULL || top()->is_top(), ""); 1234 } 1235 1236 #ifdef ASSERT 1237 uint Compile::count_live_nodes_by_graph_walk() { 1238 Unique_Node_List useful(comp_arena()); 1239 // Get useful node list by walking the graph. 1240 identify_useful_nodes(useful); 1241 return useful.size(); 1242 } 1243 1244 void Compile::print_missing_nodes() { 1245 1246 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1247 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1248 return; 1249 } 1250 1251 // This is an expensive function. It is executed only when the user 1252 // specifies VerifyIdealNodeCount option or otherwise knows the 1253 // additional work that needs to be done to identify reachable nodes 1254 // by walking the flow graph and find the missing ones using 1255 // _dead_node_list. 1256 1257 Unique_Node_List useful(comp_arena()); 1258 // Get useful node list by walking the graph. 1259 identify_useful_nodes(useful); 1260 1261 uint l_nodes = C->live_nodes(); 1262 uint l_nodes_by_walk = useful.size(); 1263 1264 if (l_nodes != l_nodes_by_walk) { 1265 if (_log != NULL) { 1266 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1267 _log->stamp(); 1268 _log->end_head(); 1269 } 1270 VectorSet& useful_member_set = useful.member_set(); 1271 int last_idx = l_nodes_by_walk; 1272 for (int i = 0; i < last_idx; i++) { 1273 if (useful_member_set.test(i)) { 1274 if (_dead_node_list.test(i)) { 1275 if (_log != NULL) { 1276 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1277 } 1278 if (PrintIdealNodeCount) { 1279 // Print the log message to tty 1280 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1281 useful.at(i)->dump(); 1282 } 1283 } 1284 } 1285 else if (! _dead_node_list.test(i)) { 1286 if (_log != NULL) { 1287 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1288 } 1289 if (PrintIdealNodeCount) { 1290 // Print the log message to tty 1291 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1292 } 1293 } 1294 } 1295 if (_log != NULL) { 1296 _log->tail("mismatched_nodes"); 1297 } 1298 } 1299 } 1300 void Compile::record_modified_node(Node* n) { 1301 if (_modified_nodes != NULL && !_inlining_incrementally && 1302 n->outcnt() != 0 && !n->is_Con()) { 1303 _modified_nodes->push(n); 1304 } 1305 } 1306 1307 void Compile::remove_modified_node(Node* n) { 1308 if (_modified_nodes != NULL) { 1309 _modified_nodes->remove(n); 1310 } 1311 } 1312 #endif 1313 1314 #ifndef PRODUCT 1315 void Compile::verify_top(Node* tn) const { 1316 if (tn != NULL) { 1317 assert(tn->is_Con(), "top node must be a constant"); 1318 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1319 assert(tn->in(0) != NULL, "must have live top node"); 1320 } 1321 } 1322 #endif 1323 1324 1325 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1326 1327 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1328 guarantee(arr != NULL, ""); 1329 int num_blocks = arr->length(); 1330 if (grow_by < num_blocks) grow_by = num_blocks; 1331 int num_notes = grow_by * _node_notes_block_size; 1332 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1333 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1334 while (num_notes > 0) { 1335 arr->append(notes); 1336 notes += _node_notes_block_size; 1337 num_notes -= _node_notes_block_size; 1338 } 1339 assert(num_notes == 0, "exact multiple, please"); 1340 } 1341 1342 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1343 if (source == NULL || dest == NULL) return false; 1344 1345 if (dest->is_Con()) 1346 return false; // Do not push debug info onto constants. 1347 1348 #ifdef ASSERT 1349 // Leave a bread crumb trail pointing to the original node: 1350 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1351 dest->set_debug_orig(source); 1352 } 1353 #endif 1354 1355 if (node_note_array() == NULL) 1356 return false; // Not collecting any notes now. 1357 1358 // This is a copy onto a pre-existing node, which may already have notes. 1359 // If both nodes have notes, do not overwrite any pre-existing notes. 1360 Node_Notes* source_notes = node_notes_at(source->_idx); 1361 if (source_notes == NULL || source_notes->is_clear()) return false; 1362 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1363 if (dest_notes == NULL || dest_notes->is_clear()) { 1364 return set_node_notes_at(dest->_idx, source_notes); 1365 } 1366 1367 Node_Notes merged_notes = (*source_notes); 1368 // The order of operations here ensures that dest notes will win... 1369 merged_notes.update_from(dest_notes); 1370 return set_node_notes_at(dest->_idx, &merged_notes); 1371 } 1372 1373 1374 //--------------------------allow_range_check_smearing------------------------- 1375 // Gating condition for coalescing similar range checks. 1376 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1377 // single covering check that is at least as strong as any of them. 1378 // If the optimization succeeds, the simplified (strengthened) range check 1379 // will always succeed. If it fails, we will deopt, and then give up 1380 // on the optimization. 1381 bool Compile::allow_range_check_smearing() const { 1382 // If this method has already thrown a range-check, 1383 // assume it was because we already tried range smearing 1384 // and it failed. 1385 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1386 return !already_trapped; 1387 } 1388 1389 1390 //------------------------------flatten_alias_type----------------------------- 1391 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1392 int offset = tj->offset(); 1393 TypePtr::PTR ptr = tj->ptr(); 1394 1395 // Known instance (scalarizable allocation) alias only with itself. 1396 bool is_known_inst = tj->isa_oopptr() != NULL && 1397 tj->is_oopptr()->is_known_instance(); 1398 1399 // Process weird unsafe references. 1400 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1401 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1402 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1403 tj = TypeOopPtr::BOTTOM; 1404 ptr = tj->ptr(); 1405 offset = tj->offset(); 1406 } 1407 1408 // Array pointers need some flattening 1409 const TypeAryPtr *ta = tj->isa_aryptr(); 1410 if (ta && ta->is_stable()) { 1411 // Erase stability property for alias analysis. 1412 tj = ta = ta->cast_to_stable(false); 1413 } 1414 if( ta && is_known_inst ) { 1415 if ( offset != Type::OffsetBot && 1416 offset > arrayOopDesc::length_offset_in_bytes() ) { 1417 offset = Type::OffsetBot; // Flatten constant access into array body only 1418 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1419 } 1420 } else if( ta && _AliasLevel >= 2 ) { 1421 // For arrays indexed by constant indices, we flatten the alias 1422 // space to include all of the array body. Only the header, klass 1423 // and array length can be accessed un-aliased. 1424 if( offset != Type::OffsetBot ) { 1425 if( ta->const_oop() ) { // MethodData* or Method* 1426 offset = Type::OffsetBot; // Flatten constant access into array body 1427 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1428 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1429 // range is OK as-is. 1430 tj = ta = TypeAryPtr::RANGE; 1431 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1432 tj = TypeInstPtr::KLASS; // all klass loads look alike 1433 ta = TypeAryPtr::RANGE; // generic ignored junk 1434 ptr = TypePtr::BotPTR; 1435 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1436 tj = TypeInstPtr::MARK; 1437 ta = TypeAryPtr::RANGE; // generic ignored junk 1438 ptr = TypePtr::BotPTR; 1439 } else if (offset == -8) { 1440 // Need to distinguish brooks ptr as is. 1441 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1442 } else { // Random constant offset into array body 1443 offset = Type::OffsetBot; // Flatten constant access into array body 1444 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1445 } 1446 } 1447 // Arrays of fixed size alias with arrays of unknown size. 1448 if (ta->size() != TypeInt::POS) { 1449 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1450 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1451 } 1452 // Arrays of known objects become arrays of unknown objects. 1453 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1454 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1455 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1456 } 1457 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1458 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1459 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1460 } 1461 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1462 // cannot be distinguished by bytecode alone. 1463 if (ta->elem() == TypeInt::BOOL) { 1464 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1465 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1466 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1467 } 1468 // During the 2nd round of IterGVN, NotNull castings are removed. 1469 // Make sure the Bottom and NotNull variants alias the same. 1470 // Also, make sure exact and non-exact variants alias the same. 1471 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1472 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1473 } 1474 } 1475 1476 // Oop pointers need some flattening 1477 const TypeInstPtr *to = tj->isa_instptr(); 1478 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1479 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1480 if( ptr == TypePtr::Constant ) { 1481 if (to->klass() != ciEnv::current()->Class_klass() || 1482 offset < k->size_helper() * wordSize) { 1483 // No constant oop pointers (such as Strings); they alias with 1484 // unknown strings. 1485 assert(!is_known_inst, "not scalarizable allocation"); 1486 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1487 } 1488 } else if( is_known_inst ) { 1489 tj = to; // Keep NotNull and klass_is_exact for instance type 1490 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1491 // During the 2nd round of IterGVN, NotNull castings are removed. 1492 // Make sure the Bottom and NotNull variants alias the same. 1493 // Also, make sure exact and non-exact variants alias the same. 1494 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1495 } 1496 if (to->speculative() != NULL) { 1497 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1498 } 1499 // Canonicalize the holder of this field 1500 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1501 // First handle header references such as a LoadKlassNode, even if the 1502 // object's klass is unloaded at compile time (4965979). 1503 if (!is_known_inst) { // Do it only for non-instance types 1504 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1505 } 1506 } else if ((offset != -8) && (offset < 0 || offset >= k->size_helper() * wordSize)) { 1507 // Static fields are in the space above the normal instance 1508 // fields in the java.lang.Class instance. 1509 if (to->klass() != ciEnv::current()->Class_klass()) { 1510 to = NULL; 1511 tj = TypeOopPtr::BOTTOM; 1512 offset = tj->offset(); 1513 } 1514 } else { 1515 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1516 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1517 if( is_known_inst ) { 1518 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1519 } else { 1520 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1521 } 1522 } 1523 } 1524 } 1525 1526 // Klass pointers to object array klasses need some flattening 1527 const TypeKlassPtr *tk = tj->isa_klassptr(); 1528 if( tk ) { 1529 // If we are referencing a field within a Klass, we need 1530 // to assume the worst case of an Object. Both exact and 1531 // inexact types must flatten to the same alias class so 1532 // use NotNull as the PTR. 1533 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1534 1535 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1536 TypeKlassPtr::OBJECT->klass(), 1537 offset); 1538 } 1539 1540 ciKlass* klass = tk->klass(); 1541 if( klass->is_obj_array_klass() ) { 1542 ciKlass* k = TypeAryPtr::OOPS->klass(); 1543 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1544 k = TypeInstPtr::BOTTOM->klass(); 1545 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1546 } 1547 1548 // Check for precise loads from the primary supertype array and force them 1549 // to the supertype cache alias index. Check for generic array loads from 1550 // the primary supertype array and also force them to the supertype cache 1551 // alias index. Since the same load can reach both, we need to merge 1552 // these 2 disparate memories into the same alias class. Since the 1553 // primary supertype array is read-only, there's no chance of confusion 1554 // where we bypass an array load and an array store. 1555 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1556 if (offset == Type::OffsetBot || 1557 (offset >= primary_supers_offset && 1558 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1559 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1560 offset = in_bytes(Klass::secondary_super_cache_offset()); 1561 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1562 } 1563 } 1564 1565 // Flatten all Raw pointers together. 1566 if (tj->base() == Type::RawPtr) 1567 tj = TypeRawPtr::BOTTOM; 1568 1569 if (tj->base() == Type::AnyPtr) 1570 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1571 1572 // Flatten all to bottom for now 1573 switch( _AliasLevel ) { 1574 case 0: 1575 tj = TypePtr::BOTTOM; 1576 break; 1577 case 1: // Flatten to: oop, static, field or array 1578 switch (tj->base()) { 1579 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1580 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1581 case Type::AryPtr: // do not distinguish arrays at all 1582 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1583 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1584 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1585 default: ShouldNotReachHere(); 1586 } 1587 break; 1588 case 2: // No collapsing at level 2; keep all splits 1589 case 3: // No collapsing at level 3; keep all splits 1590 break; 1591 default: 1592 Unimplemented(); 1593 } 1594 1595 offset = tj->offset(); 1596 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1597 1598 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1599 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1600 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1601 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1602 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1603 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1604 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) || 1605 (offset == -8 && tj->base() == Type::AryPtr && UseShenandoahGC), 1606 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1607 assert( tj->ptr() != TypePtr::TopPTR && 1608 tj->ptr() != TypePtr::AnyNull && 1609 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1610 // assert( tj->ptr() != TypePtr::Constant || 1611 // tj->base() == Type::RawPtr || 1612 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1613 1614 return tj; 1615 } 1616 1617 void Compile::AliasType::Init(int i, const TypePtr* at) { 1618 _index = i; 1619 _adr_type = at; 1620 _field = NULL; 1621 _element = NULL; 1622 _is_rewritable = true; // default 1623 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1624 if (atoop != NULL && atoop->is_known_instance()) { 1625 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1626 _general_index = Compile::current()->get_alias_index(gt); 1627 } else { 1628 _general_index = 0; 1629 } 1630 } 1631 1632 //---------------------------------print_on------------------------------------ 1633 #ifndef PRODUCT 1634 void Compile::AliasType::print_on(outputStream* st) { 1635 if (index() < 10) 1636 st->print("@ <%d> ", index()); 1637 else st->print("@ <%d>", index()); 1638 st->print(is_rewritable() ? " " : " RO"); 1639 int offset = adr_type()->offset(); 1640 if (offset == Type::OffsetBot) 1641 st->print(" +any"); 1642 else st->print(" +%-3d", offset); 1643 st->print(" in "); 1644 adr_type()->dump_on(st); 1645 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1646 if (field() != NULL && tjp) { 1647 if (tjp->klass() != field()->holder() || 1648 tjp->offset() != field()->offset_in_bytes()) { 1649 st->print(" != "); 1650 field()->print(); 1651 st->print(" ***"); 1652 } 1653 } 1654 } 1655 1656 void print_alias_types() { 1657 Compile* C = Compile::current(); 1658 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1659 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1660 C->alias_type(idx)->print_on(tty); 1661 tty->cr(); 1662 } 1663 } 1664 #endif 1665 1666 1667 //----------------------------probe_alias_cache-------------------------------- 1668 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1669 intptr_t key = (intptr_t) adr_type; 1670 key ^= key >> logAliasCacheSize; 1671 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1672 } 1673 1674 1675 //-----------------------------grow_alias_types-------------------------------- 1676 void Compile::grow_alias_types() { 1677 const int old_ats = _max_alias_types; // how many before? 1678 const int new_ats = old_ats; // how many more? 1679 const int grow_ats = old_ats+new_ats; // how many now? 1680 _max_alias_types = grow_ats; 1681 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1682 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1683 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1684 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1685 } 1686 1687 1688 //--------------------------------find_alias_type------------------------------ 1689 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1690 if (_AliasLevel == 0) 1691 return alias_type(AliasIdxBot); 1692 1693 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1694 if (ace->_adr_type == adr_type) { 1695 return alias_type(ace->_index); 1696 } 1697 1698 // Handle special cases. 1699 if (adr_type == NULL) return alias_type(AliasIdxTop); 1700 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1701 1702 // Do it the slow way. 1703 const TypePtr* flat = flatten_alias_type(adr_type); 1704 1705 #ifdef ASSERT 1706 assert(flat == flatten_alias_type(flat), "idempotent"); 1707 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1708 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1709 const TypeOopPtr* foop = flat->is_oopptr(); 1710 // Scalarizable allocations have exact klass always. 1711 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1712 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1713 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1714 } 1715 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1716 #endif 1717 1718 int idx = AliasIdxTop; 1719 for (int i = 0; i < num_alias_types(); i++) { 1720 if (alias_type(i)->adr_type() == flat) { 1721 idx = i; 1722 break; 1723 } 1724 } 1725 1726 if (idx == AliasIdxTop) { 1727 if (no_create) return NULL; 1728 // Grow the array if necessary. 1729 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1730 // Add a new alias type. 1731 idx = _num_alias_types++; 1732 _alias_types[idx]->Init(idx, flat); 1733 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1734 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1735 if (flat->isa_instptr()) { 1736 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1737 && flat->is_instptr()->klass() == env()->Class_klass()) 1738 alias_type(idx)->set_rewritable(false); 1739 } 1740 if (flat->isa_aryptr()) { 1741 #ifdef ASSERT 1742 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1743 // (T_BYTE has the weakest alignment and size restrictions...) 1744 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1745 #endif 1746 if (flat->offset() == TypePtr::OffsetBot) { 1747 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1748 } 1749 } 1750 if (flat->isa_klassptr()) { 1751 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1752 alias_type(idx)->set_rewritable(false); 1753 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1754 alias_type(idx)->set_rewritable(false); 1755 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1756 alias_type(idx)->set_rewritable(false); 1757 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1758 alias_type(idx)->set_rewritable(false); 1759 } 1760 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1761 // but the base pointer type is not distinctive enough to identify 1762 // references into JavaThread.) 1763 1764 // Check for final fields. 1765 const TypeInstPtr* tinst = flat->isa_instptr(); 1766 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1767 ciField* field; 1768 if (tinst->const_oop() != NULL && 1769 tinst->klass() == ciEnv::current()->Class_klass() && 1770 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1771 // static field 1772 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1773 field = k->get_field_by_offset(tinst->offset(), true); 1774 } else { 1775 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1776 field = k->get_field_by_offset(tinst->offset(), false); 1777 } 1778 assert(field == NULL || 1779 original_field == NULL || 1780 (field->holder() == original_field->holder() && 1781 field->offset() == original_field->offset() && 1782 field->is_static() == original_field->is_static()), "wrong field?"); 1783 // Set field() and is_rewritable() attributes. 1784 if (field != NULL) alias_type(idx)->set_field(field); 1785 } 1786 } 1787 1788 // Fill the cache for next time. 1789 ace->_adr_type = adr_type; 1790 ace->_index = idx; 1791 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1792 1793 // Might as well try to fill the cache for the flattened version, too. 1794 AliasCacheEntry* face = probe_alias_cache(flat); 1795 if (face->_adr_type == NULL) { 1796 face->_adr_type = flat; 1797 face->_index = idx; 1798 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1799 } 1800 1801 return alias_type(idx); 1802 } 1803 1804 1805 Compile::AliasType* Compile::alias_type(ciField* field) { 1806 const TypeOopPtr* t; 1807 if (field->is_static()) 1808 t = TypeInstPtr::make(field->holder()->java_mirror()); 1809 else 1810 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1811 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1812 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1813 return atp; 1814 } 1815 1816 1817 //------------------------------have_alias_type-------------------------------- 1818 bool Compile::have_alias_type(const TypePtr* adr_type) { 1819 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1820 if (ace->_adr_type == adr_type) { 1821 return true; 1822 } 1823 1824 // Handle special cases. 1825 if (adr_type == NULL) return true; 1826 if (adr_type == TypePtr::BOTTOM) return true; 1827 1828 return find_alias_type(adr_type, true, NULL) != NULL; 1829 } 1830 1831 //-----------------------------must_alias-------------------------------------- 1832 // True if all values of the given address type are in the given alias category. 1833 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1834 if (alias_idx == AliasIdxBot) return true; // the universal category 1835 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1836 if (alias_idx == AliasIdxTop) return false; // the empty category 1837 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1838 1839 // the only remaining possible overlap is identity 1840 int adr_idx = get_alias_index(adr_type); 1841 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1842 assert(adr_idx == alias_idx || 1843 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1844 && adr_type != TypeOopPtr::BOTTOM), 1845 "should not be testing for overlap with an unsafe pointer"); 1846 return adr_idx == alias_idx; 1847 } 1848 1849 //------------------------------can_alias-------------------------------------- 1850 // True if any values of the given address type are in the given alias category. 1851 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1852 if (alias_idx == AliasIdxTop) return false; // the empty category 1853 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1854 if (alias_idx == AliasIdxBot) return true; // the universal category 1855 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1856 1857 // the only remaining possible overlap is identity 1858 int adr_idx = get_alias_index(adr_type); 1859 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1860 return adr_idx == alias_idx; 1861 } 1862 1863 1864 1865 //---------------------------pop_warm_call------------------------------------- 1866 WarmCallInfo* Compile::pop_warm_call() { 1867 WarmCallInfo* wci = _warm_calls; 1868 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1869 return wci; 1870 } 1871 1872 //----------------------------Inline_Warm-------------------------------------- 1873 int Compile::Inline_Warm() { 1874 // If there is room, try to inline some more warm call sites. 1875 // %%% Do a graph index compaction pass when we think we're out of space? 1876 if (!InlineWarmCalls) return 0; 1877 1878 int calls_made_hot = 0; 1879 int room_to_grow = NodeCountInliningCutoff - unique(); 1880 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1881 int amount_grown = 0; 1882 WarmCallInfo* call; 1883 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1884 int est_size = (int)call->size(); 1885 if (est_size > (room_to_grow - amount_grown)) { 1886 // This one won't fit anyway. Get rid of it. 1887 call->make_cold(); 1888 continue; 1889 } 1890 call->make_hot(); 1891 calls_made_hot++; 1892 amount_grown += est_size; 1893 amount_to_grow -= est_size; 1894 } 1895 1896 if (calls_made_hot > 0) set_major_progress(); 1897 return calls_made_hot; 1898 } 1899 1900 1901 //----------------------------Finish_Warm-------------------------------------- 1902 void Compile::Finish_Warm() { 1903 if (!InlineWarmCalls) return; 1904 if (failing()) return; 1905 if (warm_calls() == NULL) return; 1906 1907 // Clean up loose ends, if we are out of space for inlining. 1908 WarmCallInfo* call; 1909 while ((call = pop_warm_call()) != NULL) { 1910 call->make_cold(); 1911 } 1912 } 1913 1914 //---------------------cleanup_loop_predicates----------------------- 1915 // Remove the opaque nodes that protect the predicates so that all unused 1916 // checks and uncommon_traps will be eliminated from the ideal graph 1917 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1918 if (predicate_count()==0) return; 1919 for (int i = predicate_count(); i > 0; i--) { 1920 Node * n = predicate_opaque1_node(i-1); 1921 assert(n->Opcode() == Op_Opaque1, "must be"); 1922 igvn.replace_node(n, n->in(1)); 1923 } 1924 assert(predicate_count()==0, "should be clean!"); 1925 } 1926 1927 // StringOpts and late inlining of string methods 1928 void Compile::inline_string_calls(bool parse_time) { 1929 { 1930 // remove useless nodes to make the usage analysis simpler 1931 ResourceMark rm; 1932 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1933 } 1934 1935 { 1936 ResourceMark rm; 1937 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1938 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1939 print_method(PHASE_AFTER_STRINGOPTS, 3); 1940 } 1941 1942 // now inline anything that we skipped the first time around 1943 if (!parse_time) { 1944 _late_inlines_pos = _late_inlines.length(); 1945 } 1946 1947 while (_string_late_inlines.length() > 0) { 1948 CallGenerator* cg = _string_late_inlines.pop(); 1949 cg->do_late_inline(); 1950 if (failing()) return; 1951 } 1952 _string_late_inlines.trunc_to(0); 1953 } 1954 1955 // Late inlining of boxing methods 1956 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1957 if (_boxing_late_inlines.length() > 0) { 1958 assert(has_boxed_value(), "inconsistent"); 1959 1960 PhaseGVN* gvn = initial_gvn(); 1961 set_inlining_incrementally(true); 1962 1963 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1964 for_igvn()->clear(); 1965 gvn->replace_with(&igvn); 1966 1967 _late_inlines_pos = _late_inlines.length(); 1968 1969 while (_boxing_late_inlines.length() > 0) { 1970 CallGenerator* cg = _boxing_late_inlines.pop(); 1971 cg->do_late_inline(); 1972 if (failing()) return; 1973 } 1974 _boxing_late_inlines.trunc_to(0); 1975 1976 { 1977 ResourceMark rm; 1978 PhaseRemoveUseless pru(gvn, for_igvn()); 1979 } 1980 1981 igvn = PhaseIterGVN(gvn); 1982 igvn.optimize(); 1983 1984 set_inlining_progress(false); 1985 set_inlining_incrementally(false); 1986 } 1987 } 1988 1989 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1990 assert(IncrementalInline, "incremental inlining should be on"); 1991 PhaseGVN* gvn = initial_gvn(); 1992 1993 set_inlining_progress(false); 1994 for_igvn()->clear(); 1995 gvn->replace_with(&igvn); 1996 1997 { 1998 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 1999 int i = 0; 2000 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2001 CallGenerator* cg = _late_inlines.at(i); 2002 _late_inlines_pos = i+1; 2003 cg->do_late_inline(); 2004 if (failing()) return; 2005 } 2006 int j = 0; 2007 for (; i < _late_inlines.length(); i++, j++) { 2008 _late_inlines.at_put(j, _late_inlines.at(i)); 2009 } 2010 _late_inlines.trunc_to(j); 2011 } 2012 2013 { 2014 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2015 ResourceMark rm; 2016 PhaseRemoveUseless pru(gvn, for_igvn()); 2017 } 2018 2019 { 2020 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2021 igvn = PhaseIterGVN(gvn); 2022 } 2023 } 2024 2025 // Perform incremental inlining until bound on number of live nodes is reached 2026 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2027 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2028 2029 PhaseGVN* gvn = initial_gvn(); 2030 2031 set_inlining_incrementally(true); 2032 set_inlining_progress(true); 2033 uint low_live_nodes = 0; 2034 2035 while(inlining_progress() && _late_inlines.length() > 0) { 2036 2037 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2038 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2039 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2040 // PhaseIdealLoop is expensive so we only try it once we are 2041 // out of live nodes and we only try it again if the previous 2042 // helped got the number of nodes down significantly 2043 PhaseIdealLoop ideal_loop( igvn, false, true ); 2044 if (failing()) return; 2045 low_live_nodes = live_nodes(); 2046 _major_progress = true; 2047 } 2048 2049 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2050 break; 2051 } 2052 } 2053 2054 inline_incrementally_one(igvn); 2055 2056 if (failing()) return; 2057 2058 { 2059 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2060 igvn.optimize(); 2061 } 2062 2063 if (failing()) return; 2064 } 2065 2066 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2067 2068 if (_string_late_inlines.length() > 0) { 2069 assert(has_stringbuilder(), "inconsistent"); 2070 for_igvn()->clear(); 2071 initial_gvn()->replace_with(&igvn); 2072 2073 inline_string_calls(false); 2074 2075 if (failing()) return; 2076 2077 { 2078 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2079 ResourceMark rm; 2080 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2081 } 2082 2083 { 2084 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2085 igvn = PhaseIterGVN(gvn); 2086 igvn.optimize(); 2087 } 2088 } 2089 2090 set_inlining_incrementally(false); 2091 } 2092 2093 2094 //------------------------------Optimize--------------------------------------- 2095 // Given a graph, optimize it. 2096 void Compile::Optimize() { 2097 TracePhase tp("optimizer", &timers[_t_optimizer]); 2098 2099 #ifndef PRODUCT 2100 if (env()->break_at_compile()) { 2101 BREAKPOINT; 2102 } 2103 2104 #endif 2105 2106 ResourceMark rm; 2107 int loop_opts_cnt; 2108 2109 print_inlining_reinit(); 2110 2111 NOT_PRODUCT( verify_graph_edges(); ) 2112 2113 print_method(PHASE_AFTER_PARSING); 2114 2115 { 2116 // Iterative Global Value Numbering, including ideal transforms 2117 // Initialize IterGVN with types and values from parse-time GVN 2118 PhaseIterGVN igvn(initial_gvn()); 2119 #ifdef ASSERT 2120 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2121 #endif 2122 { 2123 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2124 igvn.optimize(); 2125 } 2126 2127 print_method(PHASE_ITER_GVN1, 2); 2128 2129 if (failing()) return; 2130 2131 inline_incrementally(igvn); 2132 2133 print_method(PHASE_INCREMENTAL_INLINE, 2); 2134 2135 if (failing()) return; 2136 2137 if (eliminate_boxing()) { 2138 // Inline valueOf() methods now. 2139 inline_boxing_calls(igvn); 2140 2141 if (AlwaysIncrementalInline) { 2142 inline_incrementally(igvn); 2143 } 2144 2145 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2146 2147 if (failing()) return; 2148 } 2149 2150 // Remove the speculative part of types and clean up the graph from 2151 // the extra CastPP nodes whose only purpose is to carry them. Do 2152 // that early so that optimizations are not disrupted by the extra 2153 // CastPP nodes. 2154 remove_speculative_types(igvn); 2155 2156 // No more new expensive nodes will be added to the list from here 2157 // so keep only the actual candidates for optimizations. 2158 cleanup_expensive_nodes(igvn); 2159 2160 // Perform escape analysis 2161 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2162 if (has_loops()) { 2163 // Cleanup graph (remove dead nodes). 2164 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2165 PhaseIdealLoop ideal_loop( igvn, false, true ); 2166 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2167 if (failing()) return; 2168 } 2169 ConnectionGraph::do_analysis(this, &igvn); 2170 2171 if (failing()) return; 2172 2173 // Optimize out fields loads from scalar replaceable allocations. 2174 igvn.optimize(); 2175 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2176 2177 if (failing()) return; 2178 2179 if (congraph() != NULL && macro_count() > 0) { 2180 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2181 PhaseMacroExpand mexp(igvn); 2182 mexp.eliminate_macro_nodes(); 2183 igvn.set_delay_transform(false); 2184 2185 igvn.optimize(); 2186 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2187 2188 if (failing()) return; 2189 } 2190 } 2191 2192 // Loop transforms on the ideal graph. Range Check Elimination, 2193 // peeling, unrolling, etc. 2194 2195 // Set loop opts counter 2196 loop_opts_cnt = num_loop_opts(); 2197 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2198 { 2199 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2200 PhaseIdealLoop ideal_loop( igvn, true ); 2201 loop_opts_cnt--; 2202 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2203 if (failing()) return; 2204 } 2205 // Loop opts pass if partial peeling occurred in previous pass 2206 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2207 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2208 PhaseIdealLoop ideal_loop( igvn, false ); 2209 loop_opts_cnt--; 2210 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2211 if (failing()) return; 2212 } 2213 // Loop opts pass for loop-unrolling before CCP 2214 if(major_progress() && (loop_opts_cnt > 0)) { 2215 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2216 PhaseIdealLoop ideal_loop( igvn, false ); 2217 loop_opts_cnt--; 2218 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2219 } 2220 if (!failing()) { 2221 // Verify that last round of loop opts produced a valid graph 2222 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2223 PhaseIdealLoop::verify(igvn); 2224 } 2225 } 2226 if (failing()) return; 2227 2228 // Conditional Constant Propagation; 2229 PhaseCCP ccp( &igvn ); 2230 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2231 { 2232 TracePhase tp("ccp", &timers[_t_ccp]); 2233 ccp.do_transform(); 2234 } 2235 print_method(PHASE_CPP1, 2); 2236 2237 assert( true, "Break here to ccp.dump_old2new_map()"); 2238 2239 // Iterative Global Value Numbering, including ideal transforms 2240 { 2241 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2242 igvn = ccp; 2243 igvn.optimize(); 2244 } 2245 2246 print_method(PHASE_ITER_GVN2, 2); 2247 2248 if (failing()) return; 2249 2250 // Loop transforms on the ideal graph. Range Check Elimination, 2251 // peeling, unrolling, etc. 2252 if(loop_opts_cnt > 0) { 2253 debug_only( int cnt = 0; ); 2254 while(major_progress() && (loop_opts_cnt > 0)) { 2255 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2256 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2257 PhaseIdealLoop ideal_loop( igvn, true); 2258 loop_opts_cnt--; 2259 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2260 if (failing()) return; 2261 } 2262 } 2263 2264 { 2265 // Verify that all previous optimizations produced a valid graph 2266 // at least to this point, even if no loop optimizations were done. 2267 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2268 PhaseIdealLoop::verify(igvn); 2269 } 2270 2271 { 2272 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2273 PhaseMacroExpand mex(igvn); 2274 if (mex.expand_macro_nodes()) { 2275 assert(failing(), "must bail out w/ explicit message"); 2276 return; 2277 } 2278 } 2279 2280 DEBUG_ONLY( _modified_nodes = NULL; ) 2281 } // (End scope of igvn; run destructor if necessary for asserts.) 2282 2283 process_print_inlining(); 2284 // A method with only infinite loops has no edges entering loops from root 2285 { 2286 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2287 if (final_graph_reshaping()) { 2288 assert(failing(), "must bail out w/ explicit message"); 2289 return; 2290 } 2291 } 2292 2293 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2294 } 2295 2296 2297 //------------------------------Code_Gen--------------------------------------- 2298 // Given a graph, generate code for it 2299 void Compile::Code_Gen() { 2300 if (failing()) { 2301 return; 2302 } 2303 2304 // Perform instruction selection. You might think we could reclaim Matcher 2305 // memory PDQ, but actually the Matcher is used in generating spill code. 2306 // Internals of the Matcher (including some VectorSets) must remain live 2307 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2308 // set a bit in reclaimed memory. 2309 2310 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2311 // nodes. Mapping is only valid at the root of each matched subtree. 2312 NOT_PRODUCT( verify_graph_edges(); ) 2313 2314 Matcher matcher; 2315 _matcher = &matcher; 2316 { 2317 TracePhase tp("matcher", &timers[_t_matcher]); 2318 matcher.match(); 2319 } 2320 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2321 // nodes. Mapping is only valid at the root of each matched subtree. 2322 NOT_PRODUCT( verify_graph_edges(); ) 2323 2324 // If you have too many nodes, or if matching has failed, bail out 2325 check_node_count(0, "out of nodes matching instructions"); 2326 if (failing()) { 2327 return; 2328 } 2329 2330 // Build a proper-looking CFG 2331 PhaseCFG cfg(node_arena(), root(), matcher); 2332 _cfg = &cfg; 2333 { 2334 TracePhase tp("scheduler", &timers[_t_scheduler]); 2335 bool success = cfg.do_global_code_motion(); 2336 if (!success) { 2337 return; 2338 } 2339 2340 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2341 NOT_PRODUCT( verify_graph_edges(); ) 2342 debug_only( cfg.verify(); ) 2343 } 2344 2345 PhaseChaitin regalloc(unique(), cfg, matcher); 2346 _regalloc = ®alloc; 2347 { 2348 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2349 // Perform register allocation. After Chaitin, use-def chains are 2350 // no longer accurate (at spill code) and so must be ignored. 2351 // Node->LRG->reg mappings are still accurate. 2352 _regalloc->Register_Allocate(); 2353 2354 // Bail out if the allocator builds too many nodes 2355 if (failing()) { 2356 return; 2357 } 2358 } 2359 2360 // Prior to register allocation we kept empty basic blocks in case the 2361 // the allocator needed a place to spill. After register allocation we 2362 // are not adding any new instructions. If any basic block is empty, we 2363 // can now safely remove it. 2364 { 2365 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2366 cfg.remove_empty_blocks(); 2367 if (do_freq_based_layout()) { 2368 PhaseBlockLayout layout(cfg); 2369 } else { 2370 cfg.set_loop_alignment(); 2371 } 2372 cfg.fixup_flow(); 2373 } 2374 2375 // Apply peephole optimizations 2376 if( OptoPeephole ) { 2377 TracePhase tp("peephole", &timers[_t_peephole]); 2378 PhasePeephole peep( _regalloc, cfg); 2379 peep.do_transform(); 2380 } 2381 2382 // Do late expand if CPU requires this. 2383 if (Matcher::require_postalloc_expand) { 2384 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2385 cfg.postalloc_expand(_regalloc); 2386 } 2387 2388 // Convert Nodes to instruction bits in a buffer 2389 { 2390 TraceTime tp("output", &timers[_t_output], CITime); 2391 Output(); 2392 } 2393 2394 print_method(PHASE_FINAL_CODE); 2395 2396 // He's dead, Jim. 2397 _cfg = (PhaseCFG*)0xdeadbeef; 2398 _regalloc = (PhaseChaitin*)0xdeadbeef; 2399 } 2400 2401 2402 //------------------------------dump_asm--------------------------------------- 2403 // Dump formatted assembly 2404 #ifndef PRODUCT 2405 void Compile::dump_asm(int *pcs, uint pc_limit) { 2406 bool cut_short = false; 2407 tty->print_cr("#"); 2408 tty->print("# "); _tf->dump(); tty->cr(); 2409 tty->print_cr("#"); 2410 2411 // For all blocks 2412 int pc = 0x0; // Program counter 2413 char starts_bundle = ' '; 2414 _regalloc->dump_frame(); 2415 2416 Node *n = NULL; 2417 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2418 if (VMThread::should_terminate()) { 2419 cut_short = true; 2420 break; 2421 } 2422 Block* block = _cfg->get_block(i); 2423 if (block->is_connector() && !Verbose) { 2424 continue; 2425 } 2426 n = block->head(); 2427 if (pcs && n->_idx < pc_limit) { 2428 tty->print("%3.3x ", pcs[n->_idx]); 2429 } else { 2430 tty->print(" "); 2431 } 2432 block->dump_head(_cfg); 2433 if (block->is_connector()) { 2434 tty->print_cr(" # Empty connector block"); 2435 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2436 tty->print_cr(" # Block is sole successor of call"); 2437 } 2438 2439 // For all instructions 2440 Node *delay = NULL; 2441 for (uint j = 0; j < block->number_of_nodes(); j++) { 2442 if (VMThread::should_terminate()) { 2443 cut_short = true; 2444 break; 2445 } 2446 n = block->get_node(j); 2447 if (valid_bundle_info(n)) { 2448 Bundle* bundle = node_bundling(n); 2449 if (bundle->used_in_unconditional_delay()) { 2450 delay = n; 2451 continue; 2452 } 2453 if (bundle->starts_bundle()) { 2454 starts_bundle = '+'; 2455 } 2456 } 2457 2458 if (WizardMode) { 2459 n->dump(); 2460 } 2461 2462 if( !n->is_Region() && // Dont print in the Assembly 2463 !n->is_Phi() && // a few noisely useless nodes 2464 !n->is_Proj() && 2465 !n->is_MachTemp() && 2466 !n->is_SafePointScalarObject() && 2467 !n->is_Catch() && // Would be nice to print exception table targets 2468 !n->is_MergeMem() && // Not very interesting 2469 !n->is_top() && // Debug info table constants 2470 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2471 ) { 2472 if (pcs && n->_idx < pc_limit) 2473 tty->print("%3.3x", pcs[n->_idx]); 2474 else 2475 tty->print(" "); 2476 tty->print(" %c ", starts_bundle); 2477 starts_bundle = ' '; 2478 tty->print("\t"); 2479 n->format(_regalloc, tty); 2480 tty->cr(); 2481 } 2482 2483 // If we have an instruction with a delay slot, and have seen a delay, 2484 // then back up and print it 2485 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2486 assert(delay != NULL, "no unconditional delay instruction"); 2487 if (WizardMode) delay->dump(); 2488 2489 if (node_bundling(delay)->starts_bundle()) 2490 starts_bundle = '+'; 2491 if (pcs && n->_idx < pc_limit) 2492 tty->print("%3.3x", pcs[n->_idx]); 2493 else 2494 tty->print(" "); 2495 tty->print(" %c ", starts_bundle); 2496 starts_bundle = ' '; 2497 tty->print("\t"); 2498 delay->format(_regalloc, tty); 2499 tty->cr(); 2500 delay = NULL; 2501 } 2502 2503 // Dump the exception table as well 2504 if( n->is_Catch() && (Verbose || WizardMode) ) { 2505 // Print the exception table for this offset 2506 _handler_table.print_subtable_for(pc); 2507 } 2508 } 2509 2510 if (pcs && n->_idx < pc_limit) 2511 tty->print_cr("%3.3x", pcs[n->_idx]); 2512 else 2513 tty->cr(); 2514 2515 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2516 2517 } // End of per-block dump 2518 tty->cr(); 2519 2520 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2521 } 2522 #endif 2523 2524 //------------------------------Final_Reshape_Counts--------------------------- 2525 // This class defines counters to help identify when a method 2526 // may/must be executed using hardware with only 24-bit precision. 2527 struct Final_Reshape_Counts : public StackObj { 2528 int _call_count; // count non-inlined 'common' calls 2529 int _float_count; // count float ops requiring 24-bit precision 2530 int _double_count; // count double ops requiring more precision 2531 int _java_call_count; // count non-inlined 'java' calls 2532 int _inner_loop_count; // count loops which need alignment 2533 VectorSet _visited; // Visitation flags 2534 Node_List _tests; // Set of IfNodes & PCTableNodes 2535 2536 Final_Reshape_Counts() : 2537 _call_count(0), _float_count(0), _double_count(0), 2538 _java_call_count(0), _inner_loop_count(0), 2539 _visited( Thread::current()->resource_area() ) { } 2540 2541 void inc_call_count () { _call_count ++; } 2542 void inc_float_count () { _float_count ++; } 2543 void inc_double_count() { _double_count++; } 2544 void inc_java_call_count() { _java_call_count++; } 2545 void inc_inner_loop_count() { _inner_loop_count++; } 2546 2547 int get_call_count () const { return _call_count ; } 2548 int get_float_count () const { return _float_count ; } 2549 int get_double_count() const { return _double_count; } 2550 int get_java_call_count() const { return _java_call_count; } 2551 int get_inner_loop_count() const { return _inner_loop_count; } 2552 }; 2553 2554 #ifdef ASSERT 2555 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2556 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2557 // Make sure the offset goes inside the instance layout. 2558 return k->contains_field_offset(tp->offset()); 2559 // Note that OffsetBot and OffsetTop are very negative. 2560 } 2561 #endif 2562 2563 // Eliminate trivially redundant StoreCMs and accumulate their 2564 // precedence edges. 2565 void Compile::eliminate_redundant_card_marks(Node* n) { 2566 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2567 if (n->in(MemNode::Address)->outcnt() > 1) { 2568 // There are multiple users of the same address so it might be 2569 // possible to eliminate some of the StoreCMs 2570 Node* mem = n->in(MemNode::Memory); 2571 Node* adr = n->in(MemNode::Address); 2572 Node* val = n->in(MemNode::ValueIn); 2573 Node* prev = n; 2574 bool done = false; 2575 // Walk the chain of StoreCMs eliminating ones that match. As 2576 // long as it's a chain of single users then the optimization is 2577 // safe. Eliminating partially redundant StoreCMs would require 2578 // cloning copies down the other paths. 2579 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2580 if (adr == mem->in(MemNode::Address) && 2581 val == mem->in(MemNode::ValueIn)) { 2582 // redundant StoreCM 2583 if (mem->req() > MemNode::OopStore) { 2584 // Hasn't been processed by this code yet. 2585 n->add_prec(mem->in(MemNode::OopStore)); 2586 } else { 2587 // Already converted to precedence edge 2588 for (uint i = mem->req(); i < mem->len(); i++) { 2589 // Accumulate any precedence edges 2590 if (mem->in(i) != NULL) { 2591 n->add_prec(mem->in(i)); 2592 } 2593 } 2594 // Everything above this point has been processed. 2595 done = true; 2596 } 2597 // Eliminate the previous StoreCM 2598 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2599 assert(mem->outcnt() == 0, "should be dead"); 2600 mem->disconnect_inputs(NULL, this); 2601 } else { 2602 prev = mem; 2603 } 2604 mem = prev->in(MemNode::Memory); 2605 } 2606 } 2607 } 2608 2609 //------------------------------final_graph_reshaping_impl---------------------- 2610 // Implement items 1-5 from final_graph_reshaping below. 2611 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2612 2613 if ( n->outcnt() == 0 ) return; // dead node 2614 uint nop = n->Opcode(); 2615 2616 // Check for 2-input instruction with "last use" on right input. 2617 // Swap to left input. Implements item (2). 2618 if( n->req() == 3 && // two-input instruction 2619 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2620 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2621 n->in(2)->outcnt() == 1 &&// right use IS a last use 2622 !n->in(2)->is_Con() ) { // right use is not a constant 2623 // Check for commutative opcode 2624 switch( nop ) { 2625 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2626 case Op_MaxI: case Op_MinI: 2627 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2628 case Op_AndL: case Op_XorL: case Op_OrL: 2629 case Op_AndI: case Op_XorI: case Op_OrI: { 2630 // Move "last use" input to left by swapping inputs 2631 n->swap_edges(1, 2); 2632 break; 2633 } 2634 default: 2635 break; 2636 } 2637 } 2638 2639 #ifdef ASSERT 2640 if( n->is_Mem() ) { 2641 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2642 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2643 // oop will be recorded in oop map if load crosses safepoint 2644 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2645 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2646 "raw memory operations should have control edge"); 2647 } 2648 #endif 2649 // Count FPU ops and common calls, implements item (3) 2650 switch( nop ) { 2651 // Count all float operations that may use FPU 2652 case Op_AddF: 2653 case Op_SubF: 2654 case Op_MulF: 2655 case Op_DivF: 2656 case Op_NegF: 2657 case Op_ModF: 2658 case Op_ConvI2F: 2659 case Op_ConF: 2660 case Op_CmpF: 2661 case Op_CmpF3: 2662 // case Op_ConvL2F: // longs are split into 32-bit halves 2663 frc.inc_float_count(); 2664 break; 2665 2666 case Op_ConvF2D: 2667 case Op_ConvD2F: 2668 frc.inc_float_count(); 2669 frc.inc_double_count(); 2670 break; 2671 2672 // Count all double operations that may use FPU 2673 case Op_AddD: 2674 case Op_SubD: 2675 case Op_MulD: 2676 case Op_DivD: 2677 case Op_NegD: 2678 case Op_ModD: 2679 case Op_ConvI2D: 2680 case Op_ConvD2I: 2681 // case Op_ConvL2D: // handled by leaf call 2682 // case Op_ConvD2L: // handled by leaf call 2683 case Op_ConD: 2684 case Op_CmpD: 2685 case Op_CmpD3: 2686 frc.inc_double_count(); 2687 break; 2688 case Op_Opaque1: // Remove Opaque Nodes before matching 2689 case Op_Opaque2: // Remove Opaque Nodes before matching 2690 case Op_Opaque3: 2691 n->subsume_by(n->in(1), this); 2692 break; 2693 case Op_CallStaticJava: 2694 case Op_CallJava: 2695 case Op_CallDynamicJava: 2696 frc.inc_java_call_count(); // Count java call site; 2697 case Op_CallRuntime: 2698 case Op_CallLeaf: 2699 case Op_CallLeafNoFP: { 2700 assert( n->is_Call(), "" ); 2701 CallNode *call = n->as_Call(); 2702 // Count call sites where the FP mode bit would have to be flipped. 2703 // Do not count uncommon runtime calls: 2704 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2705 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2706 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2707 frc.inc_call_count(); // Count the call site 2708 } else { // See if uncommon argument is shared 2709 Node *n = call->in(TypeFunc::Parms); 2710 int nop = n->Opcode(); 2711 // Clone shared simple arguments to uncommon calls, item (1). 2712 if( n->outcnt() > 1 && 2713 !n->is_Proj() && 2714 nop != Op_CreateEx && 2715 nop != Op_CheckCastPP && 2716 nop != Op_DecodeN && 2717 nop != Op_DecodeNKlass && 2718 !n->is_Mem() ) { 2719 Node *x = n->clone(); 2720 call->set_req( TypeFunc::Parms, x ); 2721 } 2722 } 2723 break; 2724 } 2725 2726 case Op_StoreD: 2727 case Op_LoadD: 2728 case Op_LoadD_unaligned: 2729 frc.inc_double_count(); 2730 goto handle_mem; 2731 case Op_StoreF: 2732 case Op_LoadF: 2733 frc.inc_float_count(); 2734 goto handle_mem; 2735 2736 case Op_StoreCM: 2737 { 2738 // Convert OopStore dependence into precedence edge 2739 Node* prec = n->in(MemNode::OopStore); 2740 n->del_req(MemNode::OopStore); 2741 n->add_prec(prec); 2742 eliminate_redundant_card_marks(n); 2743 } 2744 2745 // fall through 2746 2747 case Op_StoreB: 2748 case Op_StoreC: 2749 case Op_StorePConditional: 2750 case Op_StoreI: 2751 case Op_StoreL: 2752 case Op_StoreIConditional: 2753 case Op_StoreLConditional: 2754 case Op_CompareAndSwapI: 2755 case Op_CompareAndSwapL: 2756 case Op_CompareAndSwapP: 2757 case Op_CompareAndSwapN: 2758 case Op_GetAndAddI: 2759 case Op_GetAndAddL: 2760 case Op_GetAndSetI: 2761 case Op_GetAndSetL: 2762 case Op_GetAndSetP: 2763 case Op_GetAndSetN: 2764 case Op_StoreP: 2765 case Op_StoreN: 2766 case Op_StoreNKlass: 2767 case Op_LoadB: 2768 case Op_LoadUB: 2769 case Op_LoadUS: 2770 case Op_LoadI: 2771 case Op_LoadKlass: 2772 case Op_LoadNKlass: 2773 case Op_LoadL: 2774 case Op_LoadL_unaligned: 2775 case Op_LoadPLocked: 2776 case Op_LoadP: 2777 case Op_LoadN: 2778 case Op_LoadRange: 2779 case Op_LoadS: { 2780 handle_mem: 2781 #ifdef ASSERT 2782 if( VerifyOptoOopOffsets ) { 2783 assert( n->is_Mem(), "" ); 2784 MemNode *mem = (MemNode*)n; 2785 // Check to see if address types have grounded out somehow. 2786 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2787 assert( !tp || oop_offset_is_sane(tp), "" ); 2788 } 2789 #endif 2790 break; 2791 } 2792 2793 case Op_AddP: { // Assert sane base pointers 2794 Node *addp = n->in(AddPNode::Address); 2795 assert( !addp->is_AddP() || 2796 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2797 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2798 "Base pointers must match" ); 2799 #ifdef _LP64 2800 if ((UseCompressedOops || UseCompressedClassPointers) && 2801 addp->Opcode() == Op_ConP && 2802 addp == n->in(AddPNode::Base) && 2803 n->in(AddPNode::Offset)->is_Con()) { 2804 // Use addressing with narrow klass to load with offset on x86. 2805 // On sparc loading 32-bits constant and decoding it have less 2806 // instructions (4) then load 64-bits constant (7). 2807 // Do this transformation here since IGVN will convert ConN back to ConP. 2808 const Type* t = addp->bottom_type(); 2809 if (t->isa_oopptr() || t->isa_klassptr()) { 2810 Node* nn = NULL; 2811 2812 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2813 2814 // Look for existing ConN node of the same exact type. 2815 Node* r = root(); 2816 uint cnt = r->outcnt(); 2817 for (uint i = 0; i < cnt; i++) { 2818 Node* m = r->raw_out(i); 2819 if (m!= NULL && m->Opcode() == op && 2820 m->bottom_type()->make_ptr() == t) { 2821 nn = m; 2822 break; 2823 } 2824 } 2825 if (nn != NULL) { 2826 // Decode a narrow oop to match address 2827 // [R12 + narrow_oop_reg<<3 + offset] 2828 if (t->isa_oopptr()) { 2829 nn = new DecodeNNode(nn, t); 2830 } else { 2831 nn = new DecodeNKlassNode(nn, t); 2832 } 2833 n->set_req(AddPNode::Base, nn); 2834 n->set_req(AddPNode::Address, nn); 2835 if (addp->outcnt() == 0) { 2836 addp->disconnect_inputs(NULL, this); 2837 } 2838 } 2839 } 2840 } 2841 #endif 2842 break; 2843 } 2844 2845 case Op_CastPP: { 2846 // Remove CastPP nodes to gain more freedom during scheduling but 2847 // keep the dependency they encode as control or precedence edges 2848 // (if control is set already) on memory operations. Some CastPP 2849 // nodes don't have a control (don't carry a dependency): skip 2850 // those. 2851 if (n->in(0) != NULL) { 2852 ResourceMark rm; 2853 Unique_Node_List wq; 2854 wq.push(n); 2855 for (uint next = 0; next < wq.size(); ++next) { 2856 Node *m = wq.at(next); 2857 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 2858 Node* use = m->fast_out(i); 2859 if (use->is_Mem() || use->is_EncodeNarrowPtr() || use->is_ShenandoahBarrier()) { 2860 use->ensure_control_or_add_prec(n->in(0)); 2861 } else if (use->in(0) == NULL) { 2862 switch(use->Opcode()) { 2863 case Op_AddP: 2864 case Op_DecodeN: 2865 case Op_DecodeNKlass: 2866 case Op_CheckCastPP: 2867 case Op_CastPP: 2868 wq.push(use); 2869 break; 2870 } 2871 } 2872 } 2873 } 2874 } 2875 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 2876 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2877 Node* in1 = n->in(1); 2878 const Type* t = n->bottom_type(); 2879 Node* new_in1 = in1->clone(); 2880 new_in1->as_DecodeN()->set_type(t); 2881 2882 if (!Matcher::narrow_oop_use_complex_address()) { 2883 // 2884 // x86, ARM and friends can handle 2 adds in addressing mode 2885 // and Matcher can fold a DecodeN node into address by using 2886 // a narrow oop directly and do implicit NULL check in address: 2887 // 2888 // [R12 + narrow_oop_reg<<3 + offset] 2889 // NullCheck narrow_oop_reg 2890 // 2891 // On other platforms (Sparc) we have to keep new DecodeN node and 2892 // use it to do implicit NULL check in address: 2893 // 2894 // decode_not_null narrow_oop_reg, base_reg 2895 // [base_reg + offset] 2896 // NullCheck base_reg 2897 // 2898 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2899 // to keep the information to which NULL check the new DecodeN node 2900 // corresponds to use it as value in implicit_null_check(). 2901 // 2902 new_in1->set_req(0, n->in(0)); 2903 } 2904 2905 n->subsume_by(new_in1, this); 2906 if (in1->outcnt() == 0) { 2907 in1->disconnect_inputs(NULL, this); 2908 } 2909 } else { 2910 n->subsume_by(n->in(1), this); 2911 if (n->outcnt() == 0) { 2912 n->disconnect_inputs(NULL, this); 2913 } 2914 } 2915 break; 2916 } 2917 #ifdef _LP64 2918 case Op_CmpP: 2919 // Do this transformation here to preserve CmpPNode::sub() and 2920 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2921 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2922 Node* in1 = n->in(1); 2923 Node* in2 = n->in(2); 2924 if (!in1->is_DecodeNarrowPtr()) { 2925 in2 = in1; 2926 in1 = n->in(2); 2927 } 2928 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2929 2930 Node* new_in2 = NULL; 2931 if (in2->is_DecodeNarrowPtr()) { 2932 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2933 new_in2 = in2->in(1); 2934 } else if (in2->Opcode() == Op_ConP) { 2935 const Type* t = in2->bottom_type(); 2936 if (t == TypePtr::NULL_PTR) { 2937 assert(in1->is_DecodeN(), "compare klass to null?"); 2938 // Don't convert CmpP null check into CmpN if compressed 2939 // oops implicit null check is not generated. 2940 // This will allow to generate normal oop implicit null check. 2941 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2942 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 2943 // 2944 // This transformation together with CastPP transformation above 2945 // will generated code for implicit NULL checks for compressed oops. 2946 // 2947 // The original code after Optimize() 2948 // 2949 // LoadN memory, narrow_oop_reg 2950 // decode narrow_oop_reg, base_reg 2951 // CmpP base_reg, NULL 2952 // CastPP base_reg // NotNull 2953 // Load [base_reg + offset], val_reg 2954 // 2955 // after these transformations will be 2956 // 2957 // LoadN memory, narrow_oop_reg 2958 // CmpN narrow_oop_reg, NULL 2959 // decode_not_null narrow_oop_reg, base_reg 2960 // Load [base_reg + offset], val_reg 2961 // 2962 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2963 // since narrow oops can be used in debug info now (see the code in 2964 // final_graph_reshaping_walk()). 2965 // 2966 // At the end the code will be matched to 2967 // on x86: 2968 // 2969 // Load_narrow_oop memory, narrow_oop_reg 2970 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2971 // NullCheck narrow_oop_reg 2972 // 2973 // and on sparc: 2974 // 2975 // Load_narrow_oop memory, narrow_oop_reg 2976 // decode_not_null narrow_oop_reg, base_reg 2977 // Load [base_reg + offset], val_reg 2978 // NullCheck base_reg 2979 // 2980 } else if (t->isa_oopptr()) { 2981 new_in2 = ConNode::make(t->make_narrowoop()); 2982 } else if (t->isa_klassptr()) { 2983 new_in2 = ConNode::make(t->make_narrowklass()); 2984 } 2985 } 2986 if (new_in2 != NULL) { 2987 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 2988 n->subsume_by(cmpN, this); 2989 if (in1->outcnt() == 0) { 2990 in1->disconnect_inputs(NULL, this); 2991 } 2992 if (in2->outcnt() == 0) { 2993 in2->disconnect_inputs(NULL, this); 2994 } 2995 } 2996 } 2997 break; 2998 2999 case Op_DecodeN: 3000 case Op_DecodeNKlass: 3001 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3002 // DecodeN could be pinned when it can't be fold into 3003 // an address expression, see the code for Op_CastPP above. 3004 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3005 break; 3006 3007 case Op_EncodeP: 3008 case Op_EncodePKlass: { 3009 Node* in1 = n->in(1); 3010 if (in1->is_DecodeNarrowPtr()) { 3011 n->subsume_by(in1->in(1), this); 3012 } else if (in1->Opcode() == Op_ConP) { 3013 const Type* t = in1->bottom_type(); 3014 if (t == TypePtr::NULL_PTR) { 3015 assert(t->isa_oopptr(), "null klass?"); 3016 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3017 } else if (t->isa_oopptr()) { 3018 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3019 } else if (t->isa_klassptr()) { 3020 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3021 } 3022 } 3023 if (in1->outcnt() == 0) { 3024 in1->disconnect_inputs(NULL, this); 3025 } 3026 break; 3027 } 3028 3029 case Op_Proj: { 3030 if (OptimizeStringConcat) { 3031 ProjNode* p = n->as_Proj(); 3032 if (p->_is_io_use) { 3033 // Separate projections were used for the exception path which 3034 // are normally removed by a late inline. If it wasn't inlined 3035 // then they will hang around and should just be replaced with 3036 // the original one. 3037 Node* proj = NULL; 3038 // Replace with just one 3039 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3040 Node *use = i.get(); 3041 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3042 proj = use; 3043 break; 3044 } 3045 } 3046 assert(proj != NULL, "must be found"); 3047 p->subsume_by(proj, this); 3048 } 3049 } 3050 break; 3051 } 3052 3053 case Op_Phi: 3054 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3055 // The EncodeP optimization may create Phi with the same edges 3056 // for all paths. It is not handled well by Register Allocator. 3057 Node* unique_in = n->in(1); 3058 assert(unique_in != NULL, ""); 3059 uint cnt = n->req(); 3060 for (uint i = 2; i < cnt; i++) { 3061 Node* m = n->in(i); 3062 assert(m != NULL, ""); 3063 if (unique_in != m) 3064 unique_in = NULL; 3065 } 3066 if (unique_in != NULL) { 3067 n->subsume_by(unique_in, this); 3068 } 3069 } 3070 break; 3071 3072 #endif 3073 3074 case Op_ModI: 3075 if (UseDivMod) { 3076 // Check if a%b and a/b both exist 3077 Node* d = n->find_similar(Op_DivI); 3078 if (d) { 3079 // Replace them with a fused divmod if supported 3080 if (Matcher::has_match_rule(Op_DivModI)) { 3081 DivModINode* divmod = DivModINode::make(n); 3082 d->subsume_by(divmod->div_proj(), this); 3083 n->subsume_by(divmod->mod_proj(), this); 3084 } else { 3085 // replace a%b with a-((a/b)*b) 3086 Node* mult = new MulINode(d, d->in(2)); 3087 Node* sub = new SubINode(d->in(1), mult); 3088 n->subsume_by(sub, this); 3089 } 3090 } 3091 } 3092 break; 3093 3094 case Op_ModL: 3095 if (UseDivMod) { 3096 // Check if a%b and a/b both exist 3097 Node* d = n->find_similar(Op_DivL); 3098 if (d) { 3099 // Replace them with a fused divmod if supported 3100 if (Matcher::has_match_rule(Op_DivModL)) { 3101 DivModLNode* divmod = DivModLNode::make(n); 3102 d->subsume_by(divmod->div_proj(), this); 3103 n->subsume_by(divmod->mod_proj(), this); 3104 } else { 3105 // replace a%b with a-((a/b)*b) 3106 Node* mult = new MulLNode(d, d->in(2)); 3107 Node* sub = new SubLNode(d->in(1), mult); 3108 n->subsume_by(sub, this); 3109 } 3110 } 3111 } 3112 break; 3113 3114 case Op_LoadVector: 3115 case Op_StoreVector: 3116 break; 3117 3118 case Op_AddReductionVI: 3119 case Op_AddReductionVL: 3120 case Op_AddReductionVF: 3121 case Op_AddReductionVD: 3122 case Op_MulReductionVI: 3123 case Op_MulReductionVL: 3124 case Op_MulReductionVF: 3125 case Op_MulReductionVD: 3126 break; 3127 3128 case Op_PackB: 3129 case Op_PackS: 3130 case Op_PackI: 3131 case Op_PackF: 3132 case Op_PackL: 3133 case Op_PackD: 3134 if (n->req()-1 > 2) { 3135 // Replace many operand PackNodes with a binary tree for matching 3136 PackNode* p = (PackNode*) n; 3137 Node* btp = p->binary_tree_pack(1, n->req()); 3138 n->subsume_by(btp, this); 3139 } 3140 break; 3141 case Op_Loop: 3142 case Op_CountedLoop: 3143 if (n->as_Loop()->is_inner_loop()) { 3144 frc.inc_inner_loop_count(); 3145 } 3146 break; 3147 case Op_LShiftI: 3148 case Op_RShiftI: 3149 case Op_URShiftI: 3150 case Op_LShiftL: 3151 case Op_RShiftL: 3152 case Op_URShiftL: 3153 if (Matcher::need_masked_shift_count) { 3154 // The cpu's shift instructions don't restrict the count to the 3155 // lower 5/6 bits. We need to do the masking ourselves. 3156 Node* in2 = n->in(2); 3157 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3158 const TypeInt* t = in2->find_int_type(); 3159 if (t != NULL && t->is_con()) { 3160 juint shift = t->get_con(); 3161 if (shift > mask) { // Unsigned cmp 3162 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3163 } 3164 } else { 3165 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3166 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3167 n->set_req(2, shift); 3168 } 3169 } 3170 if (in2->outcnt() == 0) { // Remove dead node 3171 in2->disconnect_inputs(NULL, this); 3172 } 3173 } 3174 break; 3175 case Op_MemBarStoreStore: 3176 case Op_MemBarRelease: 3177 // Break the link with AllocateNode: it is no longer useful and 3178 // confuses register allocation. 3179 if (n->req() > MemBarNode::Precedent) { 3180 n->set_req(MemBarNode::Precedent, top()); 3181 } 3182 break; 3183 case Op_ShenandoahReadBarrier: 3184 break; 3185 case Op_ShenandoahWriteBarrier: 3186 n->set_req(ShenandoahBarrierNode::Memory, immutable_memory()); 3187 break; 3188 default: 3189 assert( !n->is_Call(), "" ); 3190 assert( !n->is_Mem(), "" ); 3191 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3192 break; 3193 } 3194 3195 // Collect CFG split points 3196 if (n->is_MultiBranch()) 3197 frc._tests.push(n); 3198 } 3199 3200 //------------------------------final_graph_reshaping_walk--------------------- 3201 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3202 // requires that the walk visits a node's inputs before visiting the node. 3203 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3204 ResourceArea *area = Thread::current()->resource_area(); 3205 Unique_Node_List sfpt(area); 3206 3207 frc._visited.set(root->_idx); // first, mark node as visited 3208 uint cnt = root->req(); 3209 Node *n = root; 3210 uint i = 0; 3211 while (true) { 3212 if (i < cnt) { 3213 // Place all non-visited non-null inputs onto stack 3214 Node* m = n->in(i); 3215 ++i; 3216 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3217 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3218 // compute worst case interpreter size in case of a deoptimization 3219 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3220 3221 sfpt.push(m); 3222 } 3223 cnt = m->req(); 3224 nstack.push(n, i); // put on stack parent and next input's index 3225 n = m; 3226 i = 0; 3227 } 3228 } else { 3229 // Now do post-visit work 3230 final_graph_reshaping_impl( n, frc ); 3231 if (nstack.is_empty()) 3232 break; // finished 3233 n = nstack.node(); // Get node from stack 3234 cnt = n->req(); 3235 i = nstack.index(); 3236 nstack.pop(); // Shift to the next node on stack 3237 } 3238 } 3239 3240 // Skip next transformation if compressed oops are not used. 3241 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3242 (!UseCompressedOops && !UseCompressedClassPointers)) 3243 return; 3244 3245 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3246 // It could be done for an uncommon traps or any safepoints/calls 3247 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3248 while (sfpt.size() > 0) { 3249 n = sfpt.pop(); 3250 JVMState *jvms = n->as_SafePoint()->jvms(); 3251 assert(jvms != NULL, "sanity"); 3252 int start = jvms->debug_start(); 3253 int end = n->req(); 3254 bool is_uncommon = (n->is_CallStaticJava() && 3255 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3256 for (int j = start; j < end; j++) { 3257 Node* in = n->in(j); 3258 if (in->is_DecodeNarrowPtr()) { 3259 bool safe_to_skip = true; 3260 if (!is_uncommon ) { 3261 // Is it safe to skip? 3262 for (uint i = 0; i < in->outcnt(); i++) { 3263 Node* u = in->raw_out(i); 3264 if (!u->is_SafePoint() || 3265 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3266 safe_to_skip = false; 3267 } 3268 } 3269 } 3270 if (safe_to_skip) { 3271 n->set_req(j, in->in(1)); 3272 } 3273 if (in->outcnt() == 0) { 3274 in->disconnect_inputs(NULL, this); 3275 } 3276 } 3277 } 3278 } 3279 } 3280 3281 //------------------------------final_graph_reshaping-------------------------- 3282 // Final Graph Reshaping. 3283 // 3284 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3285 // and not commoned up and forced early. Must come after regular 3286 // optimizations to avoid GVN undoing the cloning. Clone constant 3287 // inputs to Loop Phis; these will be split by the allocator anyways. 3288 // Remove Opaque nodes. 3289 // (2) Move last-uses by commutative operations to the left input to encourage 3290 // Intel update-in-place two-address operations and better register usage 3291 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3292 // calls canonicalizing them back. 3293 // (3) Count the number of double-precision FP ops, single-precision FP ops 3294 // and call sites. On Intel, we can get correct rounding either by 3295 // forcing singles to memory (requires extra stores and loads after each 3296 // FP bytecode) or we can set a rounding mode bit (requires setting and 3297 // clearing the mode bit around call sites). The mode bit is only used 3298 // if the relative frequency of single FP ops to calls is low enough. 3299 // This is a key transform for SPEC mpeg_audio. 3300 // (4) Detect infinite loops; blobs of code reachable from above but not 3301 // below. Several of the Code_Gen algorithms fail on such code shapes, 3302 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3303 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3304 // Detection is by looking for IfNodes where only 1 projection is 3305 // reachable from below or CatchNodes missing some targets. 3306 // (5) Assert for insane oop offsets in debug mode. 3307 3308 bool Compile::final_graph_reshaping() { 3309 // an infinite loop may have been eliminated by the optimizer, 3310 // in which case the graph will be empty. 3311 if (root()->req() == 1) { 3312 record_method_not_compilable("trivial infinite loop"); 3313 return true; 3314 } 3315 3316 // Expensive nodes have their control input set to prevent the GVN 3317 // from freely commoning them. There's no GVN beyond this point so 3318 // no need to keep the control input. We want the expensive nodes to 3319 // be freely moved to the least frequent code path by gcm. 3320 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3321 for (int i = 0; i < expensive_count(); i++) { 3322 _expensive_nodes->at(i)->set_req(0, NULL); 3323 } 3324 3325 Final_Reshape_Counts frc; 3326 3327 // Visit everybody reachable! 3328 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3329 Node_Stack nstack(live_nodes() >> 1); 3330 final_graph_reshaping_walk(nstack, root(), frc); 3331 3332 // Check for unreachable (from below) code (i.e., infinite loops). 3333 for( uint i = 0; i < frc._tests.size(); i++ ) { 3334 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3335 // Get number of CFG targets. 3336 // Note that PCTables include exception targets after calls. 3337 uint required_outcnt = n->required_outcnt(); 3338 if (n->outcnt() != required_outcnt) { 3339 // Check for a few special cases. Rethrow Nodes never take the 3340 // 'fall-thru' path, so expected kids is 1 less. 3341 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3342 if (n->in(0)->in(0)->is_Call()) { 3343 CallNode *call = n->in(0)->in(0)->as_Call(); 3344 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3345 required_outcnt--; // Rethrow always has 1 less kid 3346 } else if (call->req() > TypeFunc::Parms && 3347 call->is_CallDynamicJava()) { 3348 // Check for null receiver. In such case, the optimizer has 3349 // detected that the virtual call will always result in a null 3350 // pointer exception. The fall-through projection of this CatchNode 3351 // will not be populated. 3352 Node *arg0 = call->in(TypeFunc::Parms); 3353 if (arg0->is_Type() && 3354 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3355 required_outcnt--; 3356 } 3357 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3358 call->req() > TypeFunc::Parms+1 && 3359 call->is_CallStaticJava()) { 3360 // Check for negative array length. In such case, the optimizer has 3361 // detected that the allocation attempt will always result in an 3362 // exception. There is no fall-through projection of this CatchNode . 3363 Node *arg1 = call->in(TypeFunc::Parms+1); 3364 if (arg1->is_Type() && 3365 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3366 required_outcnt--; 3367 } 3368 } 3369 } 3370 } 3371 // Recheck with a better notion of 'required_outcnt' 3372 if (n->outcnt() != required_outcnt) { 3373 record_method_not_compilable("malformed control flow"); 3374 return true; // Not all targets reachable! 3375 } 3376 } 3377 // Check that I actually visited all kids. Unreached kids 3378 // must be infinite loops. 3379 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3380 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3381 record_method_not_compilable("infinite loop"); 3382 return true; // Found unvisited kid; must be unreach 3383 } 3384 } 3385 3386 // If original bytecodes contained a mixture of floats and doubles 3387 // check if the optimizer has made it homogenous, item (3). 3388 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3389 frc.get_float_count() > 32 && 3390 frc.get_double_count() == 0 && 3391 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3392 set_24_bit_selection_and_mode( false, true ); 3393 } 3394 3395 set_java_calls(frc.get_java_call_count()); 3396 set_inner_loops(frc.get_inner_loop_count()); 3397 3398 // No infinite loops, no reason to bail out. 3399 return false; 3400 } 3401 3402 //-----------------------------too_many_traps---------------------------------- 3403 // Report if there are too many traps at the current method and bci. 3404 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3405 bool Compile::too_many_traps(ciMethod* method, 3406 int bci, 3407 Deoptimization::DeoptReason reason) { 3408 ciMethodData* md = method->method_data(); 3409 if (md->is_empty()) { 3410 // Assume the trap has not occurred, or that it occurred only 3411 // because of a transient condition during start-up in the interpreter. 3412 return false; 3413 } 3414 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3415 if (md->has_trap_at(bci, m, reason) != 0) { 3416 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3417 // Also, if there are multiple reasons, or if there is no per-BCI record, 3418 // assume the worst. 3419 if (log()) 3420 log()->elem("observe trap='%s' count='%d'", 3421 Deoptimization::trap_reason_name(reason), 3422 md->trap_count(reason)); 3423 return true; 3424 } else { 3425 // Ignore method/bci and see if there have been too many globally. 3426 return too_many_traps(reason, md); 3427 } 3428 } 3429 3430 // Less-accurate variant which does not require a method and bci. 3431 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3432 ciMethodData* logmd) { 3433 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3434 // Too many traps globally. 3435 // Note that we use cumulative trap_count, not just md->trap_count. 3436 if (log()) { 3437 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3438 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3439 Deoptimization::trap_reason_name(reason), 3440 mcount, trap_count(reason)); 3441 } 3442 return true; 3443 } else { 3444 // The coast is clear. 3445 return false; 3446 } 3447 } 3448 3449 //--------------------------too_many_recompiles-------------------------------- 3450 // Report if there are too many recompiles at the current method and bci. 3451 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3452 // Is not eager to return true, since this will cause the compiler to use 3453 // Action_none for a trap point, to avoid too many recompilations. 3454 bool Compile::too_many_recompiles(ciMethod* method, 3455 int bci, 3456 Deoptimization::DeoptReason reason) { 3457 ciMethodData* md = method->method_data(); 3458 if (md->is_empty()) { 3459 // Assume the trap has not occurred, or that it occurred only 3460 // because of a transient condition during start-up in the interpreter. 3461 return false; 3462 } 3463 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3464 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3465 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3466 Deoptimization::DeoptReason per_bc_reason 3467 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3468 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3469 if ((per_bc_reason == Deoptimization::Reason_none 3470 || md->has_trap_at(bci, m, reason) != 0) 3471 // The trap frequency measure we care about is the recompile count: 3472 && md->trap_recompiled_at(bci, m) 3473 && md->overflow_recompile_count() >= bc_cutoff) { 3474 // Do not emit a trap here if it has already caused recompilations. 3475 // Also, if there are multiple reasons, or if there is no per-BCI record, 3476 // assume the worst. 3477 if (log()) 3478 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3479 Deoptimization::trap_reason_name(reason), 3480 md->trap_count(reason), 3481 md->overflow_recompile_count()); 3482 return true; 3483 } else if (trap_count(reason) != 0 3484 && decompile_count() >= m_cutoff) { 3485 // Too many recompiles globally, and we have seen this sort of trap. 3486 // Use cumulative decompile_count, not just md->decompile_count. 3487 if (log()) 3488 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3489 Deoptimization::trap_reason_name(reason), 3490 md->trap_count(reason), trap_count(reason), 3491 md->decompile_count(), decompile_count()); 3492 return true; 3493 } else { 3494 // The coast is clear. 3495 return false; 3496 } 3497 } 3498 3499 // Compute when not to trap. Used by matching trap based nodes and 3500 // NullCheck optimization. 3501 void Compile::set_allowed_deopt_reasons() { 3502 _allowed_reasons = 0; 3503 if (is_method_compilation()) { 3504 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3505 assert(rs < BitsPerInt, "recode bit map"); 3506 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3507 _allowed_reasons |= nth_bit(rs); 3508 } 3509 } 3510 } 3511 } 3512 3513 #ifndef PRODUCT 3514 //------------------------------verify_graph_edges--------------------------- 3515 // Walk the Graph and verify that there is a one-to-one correspondence 3516 // between Use-Def edges and Def-Use edges in the graph. 3517 void Compile::verify_graph_edges(bool no_dead_code) { 3518 if (VerifyGraphEdges) { 3519 ResourceArea *area = Thread::current()->resource_area(); 3520 Unique_Node_List visited(area); 3521 // Call recursive graph walk to check edges 3522 _root->verify_edges(visited); 3523 if (no_dead_code) { 3524 // Now make sure that no visited node is used by an unvisited node. 3525 bool dead_nodes = false; 3526 Unique_Node_List checked(area); 3527 while (visited.size() > 0) { 3528 Node* n = visited.pop(); 3529 checked.push(n); 3530 for (uint i = 0; i < n->outcnt(); i++) { 3531 Node* use = n->raw_out(i); 3532 if (checked.member(use)) continue; // already checked 3533 if (visited.member(use)) continue; // already in the graph 3534 if (use->is_Con()) continue; // a dead ConNode is OK 3535 // At this point, we have found a dead node which is DU-reachable. 3536 if (!dead_nodes) { 3537 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3538 dead_nodes = true; 3539 } 3540 use->dump(2); 3541 tty->print_cr("---"); 3542 checked.push(use); // No repeats; pretend it is now checked. 3543 } 3544 } 3545 assert(!dead_nodes, "using nodes must be reachable from root"); 3546 } 3547 } 3548 } 3549 3550 // Verify GC barriers consistency 3551 // Currently supported: 3552 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3553 void Compile::verify_barriers() { 3554 if (UseG1GC || UseShenandoahGC) { 3555 // Verify G1 pre-barriers 3556 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3557 3558 ResourceArea *area = Thread::current()->resource_area(); 3559 Unique_Node_List visited(area); 3560 Node_List worklist(area); 3561 // We're going to walk control flow backwards starting from the Root 3562 worklist.push(_root); 3563 while (worklist.size() > 0) { 3564 Node* x = worklist.pop(); 3565 if (x == NULL || x == top()) continue; 3566 if (visited.member(x)) { 3567 continue; 3568 } else { 3569 visited.push(x); 3570 } 3571 3572 if (x->is_Region()) { 3573 for (uint i = 1; i < x->req(); i++) { 3574 worklist.push(x->in(i)); 3575 } 3576 } else { 3577 worklist.push(x->in(0)); 3578 // We are looking for the pattern: 3579 // /->ThreadLocal 3580 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3581 // \->ConI(0) 3582 // We want to verify that the If and the LoadB have the same control 3583 // See GraphKit::g1_write_barrier_pre() 3584 if (x->is_If()) { 3585 IfNode *iff = x->as_If(); 3586 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3587 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3588 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3589 && cmp->in(1)->is_Load()) { 3590 LoadNode* load = cmp->in(1)->as_Load(); 3591 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3592 && load->in(2)->in(3)->is_Con() 3593 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3594 3595 Node* if_ctrl = iff->in(0); 3596 Node* load_ctrl = load->in(0); 3597 3598 if (if_ctrl != load_ctrl) { 3599 // Skip possible CProj->NeverBranch in infinite loops 3600 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3601 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3602 if_ctrl = if_ctrl->in(0)->in(0); 3603 } 3604 } 3605 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3606 } 3607 } 3608 } 3609 } 3610 } 3611 } 3612 } 3613 } 3614 3615 #endif 3616 3617 // The Compile object keeps track of failure reasons separately from the ciEnv. 3618 // This is required because there is not quite a 1-1 relation between the 3619 // ciEnv and its compilation task and the Compile object. Note that one 3620 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3621 // to backtrack and retry without subsuming loads. Other than this backtracking 3622 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3623 // by the logic in C2Compiler. 3624 void Compile::record_failure(const char* reason) { 3625 if (log() != NULL) { 3626 log()->elem("failure reason='%s' phase='compile'", reason); 3627 } 3628 if (_failure_reason == NULL) { 3629 // Record the first failure reason. 3630 _failure_reason = reason; 3631 } 3632 3633 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3634 C->print_method(PHASE_FAILURE); 3635 } 3636 _root = NULL; // flush the graph, too 3637 } 3638 3639 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3640 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3641 _phase_name(name), _dolog(CITimeVerbose) 3642 { 3643 if (_dolog) { 3644 C = Compile::current(); 3645 _log = C->log(); 3646 } else { 3647 C = NULL; 3648 _log = NULL; 3649 } 3650 if (_log != NULL) { 3651 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3652 _log->stamp(); 3653 _log->end_head(); 3654 } 3655 } 3656 3657 Compile::TracePhase::~TracePhase() { 3658 3659 C = Compile::current(); 3660 if (_dolog) { 3661 _log = C->log(); 3662 } else { 3663 _log = NULL; 3664 } 3665 3666 #ifdef ASSERT 3667 if (PrintIdealNodeCount) { 3668 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3669 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3670 } 3671 3672 if (VerifyIdealNodeCount) { 3673 Compile::current()->print_missing_nodes(); 3674 } 3675 #endif 3676 3677 if (_log != NULL) { 3678 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3679 } 3680 } 3681 3682 //============================================================================= 3683 // Two Constant's are equal when the type and the value are equal. 3684 bool Compile::Constant::operator==(const Constant& other) { 3685 if (type() != other.type() ) return false; 3686 if (can_be_reused() != other.can_be_reused()) return false; 3687 // For floating point values we compare the bit pattern. 3688 switch (type()) { 3689 case T_FLOAT: return (_v._value.i == other._v._value.i); 3690 case T_LONG: 3691 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3692 case T_OBJECT: 3693 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3694 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3695 case T_METADATA: return (_v._metadata == other._v._metadata); 3696 default: ShouldNotReachHere(); 3697 } 3698 return false; 3699 } 3700 3701 static int type_to_size_in_bytes(BasicType t) { 3702 switch (t) { 3703 case T_LONG: return sizeof(jlong ); 3704 case T_FLOAT: return sizeof(jfloat ); 3705 case T_DOUBLE: return sizeof(jdouble); 3706 case T_METADATA: return sizeof(Metadata*); 3707 // We use T_VOID as marker for jump-table entries (labels) which 3708 // need an internal word relocation. 3709 case T_VOID: 3710 case T_ADDRESS: 3711 case T_OBJECT: return sizeof(jobject); 3712 } 3713 3714 ShouldNotReachHere(); 3715 return -1; 3716 } 3717 3718 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3719 // sort descending 3720 if (a->freq() > b->freq()) return -1; 3721 if (a->freq() < b->freq()) return 1; 3722 return 0; 3723 } 3724 3725 void Compile::ConstantTable::calculate_offsets_and_size() { 3726 // First, sort the array by frequencies. 3727 _constants.sort(qsort_comparator); 3728 3729 #ifdef ASSERT 3730 // Make sure all jump-table entries were sorted to the end of the 3731 // array (they have a negative frequency). 3732 bool found_void = false; 3733 for (int i = 0; i < _constants.length(); i++) { 3734 Constant con = _constants.at(i); 3735 if (con.type() == T_VOID) 3736 found_void = true; // jump-tables 3737 else 3738 assert(!found_void, "wrong sorting"); 3739 } 3740 #endif 3741 3742 int offset = 0; 3743 for (int i = 0; i < _constants.length(); i++) { 3744 Constant* con = _constants.adr_at(i); 3745 3746 // Align offset for type. 3747 int typesize = type_to_size_in_bytes(con->type()); 3748 offset = align_size_up(offset, typesize); 3749 con->set_offset(offset); // set constant's offset 3750 3751 if (con->type() == T_VOID) { 3752 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3753 offset = offset + typesize * n->outcnt(); // expand jump-table 3754 } else { 3755 offset = offset + typesize; 3756 } 3757 } 3758 3759 // Align size up to the next section start (which is insts; see 3760 // CodeBuffer::align_at_start). 3761 assert(_size == -1, "already set?"); 3762 _size = align_size_up(offset, CodeEntryAlignment); 3763 } 3764 3765 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3766 MacroAssembler _masm(&cb); 3767 for (int i = 0; i < _constants.length(); i++) { 3768 Constant con = _constants.at(i); 3769 address constant_addr; 3770 switch (con.type()) { 3771 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3772 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3773 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3774 case T_OBJECT: { 3775 jobject obj = con.get_jobject(); 3776 int oop_index = _masm.oop_recorder()->find_index(obj); 3777 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3778 break; 3779 } 3780 case T_ADDRESS: { 3781 address addr = (address) con.get_jobject(); 3782 constant_addr = _masm.address_constant(addr); 3783 break; 3784 } 3785 // We use T_VOID as marker for jump-table entries (labels) which 3786 // need an internal word relocation. 3787 case T_VOID: { 3788 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3789 // Fill the jump-table with a dummy word. The real value is 3790 // filled in later in fill_jump_table. 3791 address dummy = (address) n; 3792 constant_addr = _masm.address_constant(dummy); 3793 // Expand jump-table 3794 for (uint i = 1; i < n->outcnt(); i++) { 3795 address temp_addr = _masm.address_constant(dummy + i); 3796 assert(temp_addr, "consts section too small"); 3797 } 3798 break; 3799 } 3800 case T_METADATA: { 3801 Metadata* obj = con.get_metadata(); 3802 int metadata_index = _masm.oop_recorder()->find_index(obj); 3803 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3804 break; 3805 } 3806 default: ShouldNotReachHere(); 3807 } 3808 assert(constant_addr, "consts section too small"); 3809 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3810 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3811 } 3812 } 3813 3814 int Compile::ConstantTable::find_offset(Constant& con) const { 3815 int idx = _constants.find(con); 3816 assert(idx != -1, "constant must be in constant table"); 3817 int offset = _constants.at(idx).offset(); 3818 assert(offset != -1, "constant table not emitted yet?"); 3819 return offset; 3820 } 3821 3822 void Compile::ConstantTable::add(Constant& con) { 3823 if (con.can_be_reused()) { 3824 int idx = _constants.find(con); 3825 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3826 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3827 return; 3828 } 3829 } 3830 (void) _constants.append(con); 3831 } 3832 3833 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3834 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3835 Constant con(type, value, b->_freq); 3836 add(con); 3837 return con; 3838 } 3839 3840 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3841 Constant con(metadata); 3842 add(con); 3843 return con; 3844 } 3845 3846 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3847 jvalue value; 3848 BasicType type = oper->type()->basic_type(); 3849 switch (type) { 3850 case T_LONG: value.j = oper->constantL(); break; 3851 case T_FLOAT: value.f = oper->constantF(); break; 3852 case T_DOUBLE: value.d = oper->constantD(); break; 3853 case T_OBJECT: 3854 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3855 case T_METADATA: return add((Metadata*)oper->constant()); break; 3856 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3857 } 3858 return add(n, type, value); 3859 } 3860 3861 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3862 jvalue value; 3863 // We can use the node pointer here to identify the right jump-table 3864 // as this method is called from Compile::Fill_buffer right before 3865 // the MachNodes are emitted and the jump-table is filled (means the 3866 // MachNode pointers do not change anymore). 3867 value.l = (jobject) n; 3868 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3869 add(con); 3870 return con; 3871 } 3872 3873 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3874 // If called from Compile::scratch_emit_size do nothing. 3875 if (Compile::current()->in_scratch_emit_size()) return; 3876 3877 assert(labels.is_nonempty(), "must be"); 3878 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3879 3880 // Since MachConstantNode::constant_offset() also contains 3881 // table_base_offset() we need to subtract the table_base_offset() 3882 // to get the plain offset into the constant table. 3883 int offset = n->constant_offset() - table_base_offset(); 3884 3885 MacroAssembler _masm(&cb); 3886 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3887 3888 for (uint i = 0; i < n->outcnt(); i++) { 3889 address* constant_addr = &jump_table_base[i]; 3890 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))); 3891 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3892 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3893 } 3894 } 3895 3896 //----------------------------static_subtype_check----------------------------- 3897 // Shortcut important common cases when superklass is exact: 3898 // (0) superklass is java.lang.Object (can occur in reflective code) 3899 // (1) subklass is already limited to a subtype of superklass => always ok 3900 // (2) subklass does not overlap with superklass => always fail 3901 // (3) superklass has NO subtypes and we can check with a simple compare. 3902 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 3903 if (StressReflectiveCode) { 3904 return SSC_full_test; // Let caller generate the general case. 3905 } 3906 3907 if (superk == env()->Object_klass()) { 3908 return SSC_always_true; // (0) this test cannot fail 3909 } 3910 3911 ciType* superelem = superk; 3912 if (superelem->is_array_klass()) 3913 superelem = superelem->as_array_klass()->base_element_type(); 3914 3915 if (!subk->is_interface()) { // cannot trust static interface types yet 3916 if (subk->is_subtype_of(superk)) { 3917 return SSC_always_true; // (1) false path dead; no dynamic test needed 3918 } 3919 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 3920 !superk->is_subtype_of(subk)) { 3921 return SSC_always_false; 3922 } 3923 } 3924 3925 // If casting to an instance klass, it must have no subtypes 3926 if (superk->is_interface()) { 3927 // Cannot trust interfaces yet. 3928 // %%% S.B. superk->nof_implementors() == 1 3929 } else if (superelem->is_instance_klass()) { 3930 ciInstanceKlass* ik = superelem->as_instance_klass(); 3931 if (!ik->has_subklass() && !ik->is_interface()) { 3932 if (!ik->is_final()) { 3933 // Add a dependency if there is a chance of a later subclass. 3934 dependencies()->assert_leaf_type(ik); 3935 } 3936 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3937 } 3938 } else { 3939 // A primitive array type has no subtypes. 3940 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3941 } 3942 3943 return SSC_full_test; 3944 } 3945 3946 Node* Compile::conv_I2X_index(PhaseGVN *phase, Node* idx, const TypeInt* sizetype) { 3947 #ifdef _LP64 3948 // The scaled index operand to AddP must be a clean 64-bit value. 3949 // Java allows a 32-bit int to be incremented to a negative 3950 // value, which appears in a 64-bit register as a large 3951 // positive number. Using that large positive number as an 3952 // operand in pointer arithmetic has bad consequences. 3953 // On the other hand, 32-bit overflow is rare, and the possibility 3954 // can often be excluded, if we annotate the ConvI2L node with 3955 // a type assertion that its value is known to be a small positive 3956 // number. (The prior range check has ensured this.) 3957 // This assertion is used by ConvI2LNode::Ideal. 3958 int index_max = max_jint - 1; // array size is max_jint, index is one less 3959 if (sizetype != NULL) index_max = sizetype->_hi - 1; 3960 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 3961 idx = phase->transform(new ConvI2LNode(idx, lidxtype)); 3962 #endif 3963 return idx; 3964 } 3965 3966 // The message about the current inlining is accumulated in 3967 // _print_inlining_stream and transfered into the _print_inlining_list 3968 // once we know whether inlining succeeds or not. For regular 3969 // inlining, messages are appended to the buffer pointed by 3970 // _print_inlining_idx in the _print_inlining_list. For late inlining, 3971 // a new buffer is added after _print_inlining_idx in the list. This 3972 // way we can update the inlining message for late inlining call site 3973 // when the inlining is attempted again. 3974 void Compile::print_inlining_init() { 3975 if (print_inlining() || print_intrinsics()) { 3976 _print_inlining_stream = new stringStream(); 3977 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 3978 } 3979 } 3980 3981 void Compile::print_inlining_reinit() { 3982 if (print_inlining() || print_intrinsics()) { 3983 // Re allocate buffer when we change ResourceMark 3984 _print_inlining_stream = new stringStream(); 3985 } 3986 } 3987 3988 void Compile::print_inlining_reset() { 3989 _print_inlining_stream->reset(); 3990 } 3991 3992 void Compile::print_inlining_commit() { 3993 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 3994 // Transfer the message from _print_inlining_stream to the current 3995 // _print_inlining_list buffer and clear _print_inlining_stream. 3996 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 3997 print_inlining_reset(); 3998 } 3999 4000 void Compile::print_inlining_push() { 4001 // Add new buffer to the _print_inlining_list at current position 4002 _print_inlining_idx++; 4003 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4004 } 4005 4006 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4007 return _print_inlining_list->at(_print_inlining_idx); 4008 } 4009 4010 void Compile::print_inlining_update(CallGenerator* cg) { 4011 if (print_inlining() || print_intrinsics()) { 4012 if (!cg->is_late_inline()) { 4013 if (print_inlining_current().cg() != NULL) { 4014 print_inlining_push(); 4015 } 4016 print_inlining_commit(); 4017 } else { 4018 if (print_inlining_current().cg() != cg && 4019 (print_inlining_current().cg() != NULL || 4020 print_inlining_current().ss()->size() != 0)) { 4021 print_inlining_push(); 4022 } 4023 print_inlining_commit(); 4024 print_inlining_current().set_cg(cg); 4025 } 4026 } 4027 } 4028 4029 void Compile::print_inlining_move_to(CallGenerator* cg) { 4030 // We resume inlining at a late inlining call site. Locate the 4031 // corresponding inlining buffer so that we can update it. 4032 if (print_inlining()) { 4033 for (int i = 0; i < _print_inlining_list->length(); i++) { 4034 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4035 _print_inlining_idx = i; 4036 return; 4037 } 4038 } 4039 ShouldNotReachHere(); 4040 } 4041 } 4042 4043 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4044 if (print_inlining()) { 4045 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4046 assert(print_inlining_current().cg() == cg, "wrong entry"); 4047 // replace message with new message 4048 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4049 print_inlining_commit(); 4050 print_inlining_current().set_cg(cg); 4051 } 4052 } 4053 4054 void Compile::print_inlining_assert_ready() { 4055 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4056 } 4057 4058 void Compile::process_print_inlining() { 4059 bool do_print_inlining = print_inlining() || print_intrinsics(); 4060 if (do_print_inlining || log() != NULL) { 4061 // Print inlining message for candidates that we couldn't inline 4062 // for lack of space 4063 for (int i = 0; i < _late_inlines.length(); i++) { 4064 CallGenerator* cg = _late_inlines.at(i); 4065 if (!cg->is_mh_late_inline()) { 4066 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4067 if (do_print_inlining) { 4068 cg->print_inlining_late(msg); 4069 } 4070 log_late_inline_failure(cg, msg); 4071 } 4072 } 4073 } 4074 if (do_print_inlining) { 4075 ResourceMark rm; 4076 stringStream ss; 4077 for (int i = 0; i < _print_inlining_list->length(); i++) { 4078 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4079 } 4080 size_t end = ss.size(); 4081 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4082 strncpy(_print_inlining_output, ss.base(), end+1); 4083 _print_inlining_output[end] = 0; 4084 } 4085 } 4086 4087 void Compile::dump_print_inlining() { 4088 if (_print_inlining_output != NULL) { 4089 tty->print_raw(_print_inlining_output); 4090 } 4091 } 4092 4093 void Compile::log_late_inline(CallGenerator* cg) { 4094 if (log() != NULL) { 4095 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4096 cg->unique_id()); 4097 JVMState* p = cg->call_node()->jvms(); 4098 while (p != NULL) { 4099 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4100 p = p->caller(); 4101 } 4102 log()->tail("late_inline"); 4103 } 4104 } 4105 4106 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4107 log_late_inline(cg); 4108 if (log() != NULL) { 4109 log()->inline_fail(msg); 4110 } 4111 } 4112 4113 void Compile::log_inline_id(CallGenerator* cg) { 4114 if (log() != NULL) { 4115 // The LogCompilation tool needs a unique way to identify late 4116 // inline call sites. This id must be unique for this call site in 4117 // this compilation. Try to have it unique across compilations as 4118 // well because it can be convenient when grepping through the log 4119 // file. 4120 // Distinguish OSR compilations from others in case CICountOSR is 4121 // on. 4122 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4123 cg->set_unique_id(id); 4124 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4125 } 4126 } 4127 4128 void Compile::log_inline_failure(const char* msg) { 4129 if (C->log() != NULL) { 4130 C->log()->inline_fail(msg); 4131 } 4132 } 4133 4134 4135 // Dump inlining replay data to the stream. 4136 // Don't change thread state and acquire any locks. 4137 void Compile::dump_inline_data(outputStream* out) { 4138 InlineTree* inl_tree = ilt(); 4139 if (inl_tree != NULL) { 4140 out->print(" inline %d", inl_tree->count()); 4141 inl_tree->dump_replay_data(out); 4142 } 4143 } 4144 4145 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4146 if (n1->Opcode() < n2->Opcode()) return -1; 4147 else if (n1->Opcode() > n2->Opcode()) return 1; 4148 4149 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())); 4150 for (uint i = 1; i < n1->req(); i++) { 4151 if (n1->in(i) < n2->in(i)) return -1; 4152 else if (n1->in(i) > n2->in(i)) return 1; 4153 } 4154 4155 return 0; 4156 } 4157 4158 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4159 Node* n1 = *n1p; 4160 Node* n2 = *n2p; 4161 4162 return cmp_expensive_nodes(n1, n2); 4163 } 4164 4165 void Compile::sort_expensive_nodes() { 4166 if (!expensive_nodes_sorted()) { 4167 _expensive_nodes->sort(cmp_expensive_nodes); 4168 } 4169 } 4170 4171 bool Compile::expensive_nodes_sorted() const { 4172 for (int i = 1; i < _expensive_nodes->length(); i++) { 4173 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4174 return false; 4175 } 4176 } 4177 return true; 4178 } 4179 4180 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4181 if (_expensive_nodes->length() == 0) { 4182 return false; 4183 } 4184 4185 assert(OptimizeExpensiveOps, "optimization off?"); 4186 4187 // Take this opportunity to remove dead nodes from the list 4188 int j = 0; 4189 for (int i = 0; i < _expensive_nodes->length(); i++) { 4190 Node* n = _expensive_nodes->at(i); 4191 if (!n->is_unreachable(igvn)) { 4192 assert(n->is_expensive(), "should be expensive"); 4193 _expensive_nodes->at_put(j, n); 4194 j++; 4195 } 4196 } 4197 _expensive_nodes->trunc_to(j); 4198 4199 // Then sort the list so that similar nodes are next to each other 4200 // and check for at least two nodes of identical kind with same data 4201 // inputs. 4202 sort_expensive_nodes(); 4203 4204 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4205 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4206 return true; 4207 } 4208 } 4209 4210 return false; 4211 } 4212 4213 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4214 if (_expensive_nodes->length() == 0) { 4215 return; 4216 } 4217 4218 assert(OptimizeExpensiveOps, "optimization off?"); 4219 4220 // Sort to bring similar nodes next to each other and clear the 4221 // control input of nodes for which there's only a single copy. 4222 sort_expensive_nodes(); 4223 4224 int j = 0; 4225 int identical = 0; 4226 int i = 0; 4227 bool modified = false; 4228 for (; i < _expensive_nodes->length()-1; i++) { 4229 assert(j <= i, "can't write beyond current index"); 4230 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4231 identical++; 4232 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4233 continue; 4234 } 4235 if (identical > 0) { 4236 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4237 identical = 0; 4238 } else { 4239 Node* n = _expensive_nodes->at(i); 4240 igvn.replace_input_of(n, 0, NULL); 4241 igvn.hash_insert(n); 4242 modified = true; 4243 } 4244 } 4245 if (identical > 0) { 4246 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4247 } else if (_expensive_nodes->length() >= 1) { 4248 Node* n = _expensive_nodes->at(i); 4249 igvn.replace_input_of(n, 0, NULL); 4250 igvn.hash_insert(n); 4251 modified = true; 4252 } 4253 _expensive_nodes->trunc_to(j); 4254 if (modified) { 4255 igvn.optimize(); 4256 } 4257 } 4258 4259 void Compile::add_expensive_node(Node * n) { 4260 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4261 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4262 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4263 if (OptimizeExpensiveOps) { 4264 _expensive_nodes->append(n); 4265 } else { 4266 // Clear control input and let IGVN optimize expensive nodes if 4267 // OptimizeExpensiveOps is off. 4268 n->set_req(0, NULL); 4269 } 4270 } 4271 4272 /** 4273 * Remove the speculative part of types and clean up the graph 4274 */ 4275 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4276 if (UseTypeSpeculation) { 4277 Unique_Node_List worklist; 4278 worklist.push(root()); 4279 int modified = 0; 4280 // Go over all type nodes that carry a speculative type, drop the 4281 // speculative part of the type and enqueue the node for an igvn 4282 // which may optimize it out. 4283 for (uint next = 0; next < worklist.size(); ++next) { 4284 Node *n = worklist.at(next); 4285 if (n->is_Type()) { 4286 TypeNode* tn = n->as_Type(); 4287 const Type* t = tn->type(); 4288 const Type* t_no_spec = t->remove_speculative(); 4289 if (t_no_spec != t) { 4290 bool in_hash = igvn.hash_delete(n); 4291 assert(in_hash, "node should be in igvn hash table"); 4292 tn->set_type(t_no_spec); 4293 igvn.hash_insert(n); 4294 igvn._worklist.push(n); // give it a chance to go away 4295 modified++; 4296 } 4297 } 4298 uint max = n->len(); 4299 for( uint i = 0; i < max; ++i ) { 4300 Node *m = n->in(i); 4301 if (not_a_node(m)) continue; 4302 worklist.push(m); 4303 } 4304 } 4305 // Drop the speculative part of all types in the igvn's type table 4306 igvn.remove_speculative_types(); 4307 if (modified > 0) { 4308 igvn.optimize(); 4309 } 4310 #ifdef ASSERT 4311 // Verify that after the IGVN is over no speculative type has resurfaced 4312 worklist.clear(); 4313 worklist.push(root()); 4314 for (uint next = 0; next < worklist.size(); ++next) { 4315 Node *n = worklist.at(next); 4316 const Type* t = igvn.type_or_null(n); 4317 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4318 if (n->is_Type()) { 4319 t = n->as_Type()->type(); 4320 assert(t == t->remove_speculative(), "no more speculative types"); 4321 } 4322 uint max = n->len(); 4323 for( uint i = 0; i < max; ++i ) { 4324 Node *m = n->in(i); 4325 if (not_a_node(m)) continue; 4326 worklist.push(m); 4327 } 4328 } 4329 igvn.check_no_speculative_types(); 4330 #endif 4331 } 4332 } 4333 4334 // Auxiliary method to support randomized stressing/fuzzing. 4335 // 4336 // This method can be called the arbitrary number of times, with current count 4337 // as the argument. The logic allows selecting a single candidate from the 4338 // running list of candidates as follows: 4339 // int count = 0; 4340 // Cand* selected = null; 4341 // while(cand = cand->next()) { 4342 // if (randomized_select(++count)) { 4343 // selected = cand; 4344 // } 4345 // } 4346 // 4347 // Including count equalizes the chances any candidate is "selected". 4348 // This is useful when we don't have the complete list of candidates to choose 4349 // from uniformly. In this case, we need to adjust the randomicity of the 4350 // selection, or else we will end up biasing the selection towards the latter 4351 // candidates. 4352 // 4353 // Quick back-envelope calculation shows that for the list of n candidates 4354 // the equal probability for the candidate to persist as "best" can be 4355 // achieved by replacing it with "next" k-th candidate with the probability 4356 // of 1/k. It can be easily shown that by the end of the run, the 4357 // probability for any candidate is converged to 1/n, thus giving the 4358 // uniform distribution among all the candidates. 4359 // 4360 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4361 #define RANDOMIZED_DOMAIN_POW 29 4362 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4363 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4364 bool Compile::randomized_select(int count) { 4365 assert(count > 0, "only positive"); 4366 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4367 } 4368 4369 const char* CloneMap::debug_option_name = "CloneMapDebug"; 4370 CloneMap& Compile::clone_map() { return _clone_map; } 4371 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4372 4373 void NodeCloneInfo::dump() const { 4374 tty->print(" {%d:%d} ", idx(), gen()); 4375 } 4376 4377 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4378 uint64_t val = value(old->_idx); 4379 NodeCloneInfo cio(val); 4380 assert(val != 0, "old node should be in the map"); 4381 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4382 insert(nnn->_idx, cin.get()); 4383 #ifndef PRODUCT 4384 if (is_debug()) { 4385 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4386 } 4387 #endif 4388 } 4389 4390 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4391 NodeCloneInfo cio(value(old->_idx)); 4392 if (cio.get() == 0) { 4393 cio.set(old->_idx, 0); 4394 insert(old->_idx, cio.get()); 4395 #ifndef PRODUCT 4396 if (is_debug()) { 4397 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4398 } 4399 #endif 4400 } 4401 clone(old, nnn, gen); 4402 } 4403 4404 int CloneMap::max_gen() const { 4405 int g = 0; 4406 DictI di(_dict); 4407 for(; di.test(); ++di) { 4408 int t = gen(di._key); 4409 if (g < t) { 4410 g = t; 4411 #ifndef PRODUCT 4412 if (is_debug()) { 4413 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4414 } 4415 #endif 4416 } 4417 } 4418 return g; 4419 } 4420 4421 void CloneMap::dump(node_idx_t key) const { 4422 uint64_t val = value(key); 4423 if (val != 0) { 4424 NodeCloneInfo ni(val); 4425 ni.dump(); 4426 } 4427 }