1 /* 2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/block.hpp" 38 #include "opto/c2compiler.hpp" 39 #include "opto/callGenerator.hpp" 40 #include "opto/callnode.hpp" 41 #include "opto/cfgnode.hpp" 42 #include "opto/chaitin.hpp" 43 #include "opto/compile.hpp" 44 #include "opto/connode.hpp" 45 #include "opto/convertnode.hpp" 46 #include "opto/divnode.hpp" 47 #include "opto/escape.hpp" 48 #include "opto/idealGraphPrinter.hpp" 49 #include "opto/loopnode.hpp" 50 #include "opto/machnode.hpp" 51 #include "opto/macro.hpp" 52 #include "opto/matcher.hpp" 53 #include "opto/mathexactnode.hpp" 54 #include "opto/memnode.hpp" 55 #include "opto/mulnode.hpp" 56 #include "opto/narrowptrnode.hpp" 57 #include "opto/node.hpp" 58 #include "opto/opcodes.hpp" 59 #include "opto/output.hpp" 60 #include "opto/parse.hpp" 61 #include "opto/phaseX.hpp" 62 #include "opto/rootnode.hpp" 63 #include "opto/runtime.hpp" 64 #include "opto/stringopts.hpp" 65 #include "opto/type.hpp" 66 #include "opto/vectornode.hpp" 67 #include "runtime/arguments.hpp" 68 #include "runtime/sharedRuntime.hpp" 69 #include "runtime/signature.hpp" 70 #include "runtime/stubRoutines.hpp" 71 #include "runtime/timer.hpp" 72 #include "utilities/copy.hpp" 73 74 75 // -------------------- Compile::mach_constant_base_node ----------------------- 76 // Constant table base node singleton. 77 MachConstantBaseNode* Compile::mach_constant_base_node() { 78 if (_mach_constant_base_node == NULL) { 79 _mach_constant_base_node = new MachConstantBaseNode(); 80 _mach_constant_base_node->add_req(C->root()); 81 } 82 return _mach_constant_base_node; 83 } 84 85 86 /// Support for intrinsics. 87 88 // Return the index at which m must be inserted (or already exists). 89 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 90 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 91 #ifdef ASSERT 92 for (int i = 1; i < _intrinsics->length(); i++) { 93 CallGenerator* cg1 = _intrinsics->at(i-1); 94 CallGenerator* cg2 = _intrinsics->at(i); 95 assert(cg1->method() != cg2->method() 96 ? cg1->method() < cg2->method() 97 : cg1->is_virtual() < cg2->is_virtual(), 98 "compiler intrinsics list must stay sorted"); 99 } 100 #endif 101 // Binary search sorted list, in decreasing intervals [lo, hi]. 102 int lo = 0, hi = _intrinsics->length()-1; 103 while (lo <= hi) { 104 int mid = (uint)(hi + lo) / 2; 105 ciMethod* mid_m = _intrinsics->at(mid)->method(); 106 if (m < mid_m) { 107 hi = mid-1; 108 } else if (m > mid_m) { 109 lo = mid+1; 110 } else { 111 // look at minor sort key 112 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 113 if (is_virtual < mid_virt) { 114 hi = mid-1; 115 } else if (is_virtual > mid_virt) { 116 lo = mid+1; 117 } else { 118 return mid; // exact match 119 } 120 } 121 } 122 return lo; // inexact match 123 } 124 125 void Compile::register_intrinsic(CallGenerator* cg) { 126 if (_intrinsics == NULL) { 127 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 128 } 129 // This code is stolen from ciObjectFactory::insert. 130 // Really, GrowableArray should have methods for 131 // insert_at, remove_at, and binary_search. 132 int len = _intrinsics->length(); 133 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 134 if (index == len) { 135 _intrinsics->append(cg); 136 } else { 137 #ifdef ASSERT 138 CallGenerator* oldcg = _intrinsics->at(index); 139 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 140 #endif 141 _intrinsics->append(_intrinsics->at(len-1)); 142 int pos; 143 for (pos = len-2; pos >= index; pos--) { 144 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 145 } 146 _intrinsics->at_put(index, cg); 147 } 148 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 149 } 150 151 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 152 assert(m->is_loaded(), "don't try this on unloaded methods"); 153 if (_intrinsics != NULL) { 154 int index = intrinsic_insertion_index(m, is_virtual); 155 if (index < _intrinsics->length() 156 && _intrinsics->at(index)->method() == m 157 && _intrinsics->at(index)->is_virtual() == is_virtual) { 158 return _intrinsics->at(index); 159 } 160 } 161 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 162 if (m->intrinsic_id() != vmIntrinsics::_none && 163 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 164 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 165 if (cg != NULL) { 166 // Save it for next time: 167 register_intrinsic(cg); 168 return cg; 169 } else { 170 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 171 } 172 } 173 return NULL; 174 } 175 176 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 177 // in library_call.cpp. 178 179 180 #ifndef PRODUCT 181 // statistics gathering... 182 183 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 184 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 185 186 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 187 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 188 int oflags = _intrinsic_hist_flags[id]; 189 assert(flags != 0, "what happened?"); 190 if (is_virtual) { 191 flags |= _intrinsic_virtual; 192 } 193 bool changed = (flags != oflags); 194 if ((flags & _intrinsic_worked) != 0) { 195 juint count = (_intrinsic_hist_count[id] += 1); 196 if (count == 1) { 197 changed = true; // first time 198 } 199 // increment the overall count also: 200 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 201 } 202 if (changed) { 203 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 204 // Something changed about the intrinsic's virtuality. 205 if ((flags & _intrinsic_virtual) != 0) { 206 // This is the first use of this intrinsic as a virtual call. 207 if (oflags != 0) { 208 // We already saw it as a non-virtual, so note both cases. 209 flags |= _intrinsic_both; 210 } 211 } else if ((oflags & _intrinsic_both) == 0) { 212 // This is the first use of this intrinsic as a non-virtual 213 flags |= _intrinsic_both; 214 } 215 } 216 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 217 } 218 // update the overall flags also: 219 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 220 return changed; 221 } 222 223 static char* format_flags(int flags, char* buf) { 224 buf[0] = 0; 225 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 226 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 227 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 228 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 229 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 230 if (buf[0] == 0) strcat(buf, ","); 231 assert(buf[0] == ',', "must be"); 232 return &buf[1]; 233 } 234 235 void Compile::print_intrinsic_statistics() { 236 char flagsbuf[100]; 237 ttyLocker ttyl; 238 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 239 tty->print_cr("Compiler intrinsic usage:"); 240 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 241 if (total == 0) total = 1; // avoid div0 in case of no successes 242 #define PRINT_STAT_LINE(name, c, f) \ 243 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 244 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 245 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 246 int flags = _intrinsic_hist_flags[id]; 247 juint count = _intrinsic_hist_count[id]; 248 if ((flags | count) != 0) { 249 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 250 } 251 } 252 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 253 if (xtty != NULL) xtty->tail("statistics"); 254 } 255 256 void Compile::print_statistics() { 257 { ttyLocker ttyl; 258 if (xtty != NULL) xtty->head("statistics type='opto'"); 259 Parse::print_statistics(); 260 PhaseCCP::print_statistics(); 261 PhaseRegAlloc::print_statistics(); 262 Scheduling::print_statistics(); 263 PhasePeephole::print_statistics(); 264 PhaseIdealLoop::print_statistics(); 265 if (xtty != NULL) xtty->tail("statistics"); 266 } 267 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 268 // put this under its own <statistics> element. 269 print_intrinsic_statistics(); 270 } 271 } 272 #endif //PRODUCT 273 274 // Support for bundling info 275 Bundle* Compile::node_bundling(const Node *n) { 276 assert(valid_bundle_info(n), "oob"); 277 return &_node_bundling_base[n->_idx]; 278 } 279 280 bool Compile::valid_bundle_info(const Node *n) { 281 return (_node_bundling_limit > n->_idx); 282 } 283 284 285 void Compile::gvn_replace_by(Node* n, Node* nn) { 286 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 287 Node* use = n->last_out(i); 288 bool is_in_table = initial_gvn()->hash_delete(use); 289 uint uses_found = 0; 290 for (uint j = 0; j < use->len(); j++) { 291 if (use->in(j) == n) { 292 if (j < use->req()) 293 use->set_req(j, nn); 294 else 295 use->set_prec(j, nn); 296 uses_found++; 297 } 298 } 299 if (is_in_table) { 300 // reinsert into table 301 initial_gvn()->hash_find_insert(use); 302 } 303 record_for_igvn(use); 304 i -= uses_found; // we deleted 1 or more copies of this edge 305 } 306 } 307 308 309 static inline bool not_a_node(const Node* n) { 310 if (n == NULL) return true; 311 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 312 if (*(address*)n == badAddress) return true; // kill by Node::destruct 313 return false; 314 } 315 316 // Identify all nodes that are reachable from below, useful. 317 // Use breadth-first pass that records state in a Unique_Node_List, 318 // recursive traversal is slower. 319 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 320 int estimated_worklist_size = unique(); 321 useful.map( estimated_worklist_size, NULL ); // preallocate space 322 323 // Initialize worklist 324 if (root() != NULL) { useful.push(root()); } 325 // If 'top' is cached, declare it useful to preserve cached node 326 if( cached_top_node() ) { useful.push(cached_top_node()); } 327 328 // Push all useful nodes onto the list, breadthfirst 329 for( uint next = 0; next < useful.size(); ++next ) { 330 assert( next < unique(), "Unique useful nodes < total nodes"); 331 Node *n = useful.at(next); 332 uint max = n->len(); 333 for( uint i = 0; i < max; ++i ) { 334 Node *m = n->in(i); 335 if (not_a_node(m)) continue; 336 useful.push(m); 337 } 338 } 339 } 340 341 // Update dead_node_list with any missing dead nodes using useful 342 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 343 void Compile::update_dead_node_list(Unique_Node_List &useful) { 344 uint max_idx = unique(); 345 VectorSet& useful_node_set = useful.member_set(); 346 347 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 348 // If node with index node_idx is not in useful set, 349 // mark it as dead in dead node list. 350 if (! useful_node_set.test(node_idx) ) { 351 record_dead_node(node_idx); 352 } 353 } 354 } 355 356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 357 int shift = 0; 358 for (int i = 0; i < inlines->length(); i++) { 359 CallGenerator* cg = inlines->at(i); 360 CallNode* call = cg->call_node(); 361 if (shift > 0) { 362 inlines->at_put(i-shift, cg); 363 } 364 if (!useful.member(call)) { 365 shift++; 366 } 367 } 368 inlines->trunc_to(inlines->length()-shift); 369 } 370 371 // Disconnect all useless nodes by disconnecting those at the boundary. 372 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 373 uint next = 0; 374 while (next < useful.size()) { 375 Node *n = useful.at(next++); 376 if (n->is_SafePoint()) { 377 // We're done with a parsing phase. Replaced nodes are not valid 378 // beyond that point. 379 n->as_SafePoint()->delete_replaced_nodes(); 380 } 381 // Use raw traversal of out edges since this code removes out edges 382 int max = n->outcnt(); 383 for (int j = 0; j < max; ++j) { 384 Node* child = n->raw_out(j); 385 if (! useful.member(child)) { 386 assert(!child->is_top() || child != top(), 387 "If top is cached in Compile object it is in useful list"); 388 // Only need to remove this out-edge to the useless node 389 n->raw_del_out(j); 390 --j; 391 --max; 392 } 393 } 394 if (n->outcnt() == 1 && n->has_special_unique_user()) { 395 record_for_igvn(n->unique_out()); 396 } 397 } 398 // Remove useless macro and predicate opaq nodes 399 for (int i = C->macro_count()-1; i >= 0; i--) { 400 Node* n = C->macro_node(i); 401 if (!useful.member(n)) { 402 remove_macro_node(n); 403 } 404 } 405 // Remove useless expensive node 406 for (int i = C->expensive_count()-1; i >= 0; i--) { 407 Node* n = C->expensive_node(i); 408 if (!useful.member(n)) { 409 remove_expensive_node(n); 410 } 411 } 412 // clean up the late inline lists 413 remove_useless_late_inlines(&_string_late_inlines, useful); 414 remove_useless_late_inlines(&_boxing_late_inlines, useful); 415 remove_useless_late_inlines(&_late_inlines, useful); 416 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 417 } 418 419 //------------------------------frame_size_in_words----------------------------- 420 // frame_slots in units of words 421 int Compile::frame_size_in_words() const { 422 // shift is 0 in LP32 and 1 in LP64 423 const int shift = (LogBytesPerWord - LogBytesPerInt); 424 int words = _frame_slots >> shift; 425 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 426 return words; 427 } 428 429 // To bang the stack of this compiled method we use the stack size 430 // that the interpreter would need in case of a deoptimization. This 431 // removes the need to bang the stack in the deoptimization blob which 432 // in turn simplifies stack overflow handling. 433 int Compile::bang_size_in_bytes() const { 434 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 435 } 436 437 // ============================================================================ 438 //------------------------------CompileWrapper--------------------------------- 439 class CompileWrapper : public StackObj { 440 Compile *const _compile; 441 public: 442 CompileWrapper(Compile* compile); 443 444 ~CompileWrapper(); 445 }; 446 447 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 448 // the Compile* pointer is stored in the current ciEnv: 449 ciEnv* env = compile->env(); 450 assert(env == ciEnv::current(), "must already be a ciEnv active"); 451 assert(env->compiler_data() == NULL, "compile already active?"); 452 env->set_compiler_data(compile); 453 assert(compile == Compile::current(), "sanity"); 454 455 compile->set_type_dict(NULL); 456 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 457 compile->clone_map().set_clone_idx(0); 458 compile->set_type_hwm(NULL); 459 compile->set_type_last_size(0); 460 compile->set_last_tf(NULL, NULL); 461 compile->set_indexSet_arena(NULL); 462 compile->set_indexSet_free_block_list(NULL); 463 compile->init_type_arena(); 464 Type::Initialize(compile); 465 _compile->set_scratch_buffer_blob(NULL); 466 _compile->begin_method(); 467 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 468 } 469 CompileWrapper::~CompileWrapper() { 470 _compile->end_method(); 471 if (_compile->scratch_buffer_blob() != NULL) 472 BufferBlob::free(_compile->scratch_buffer_blob()); 473 _compile->env()->set_compiler_data(NULL); 474 } 475 476 477 //----------------------------print_compile_messages--------------------------- 478 void Compile::print_compile_messages() { 479 #ifndef PRODUCT 480 // Check if recompiling 481 if (_subsume_loads == false && PrintOpto) { 482 // Recompiling without allowing machine instructions to subsume loads 483 tty->print_cr("*********************************************************"); 484 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 485 tty->print_cr("*********************************************************"); 486 } 487 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 488 // Recompiling without escape analysis 489 tty->print_cr("*********************************************************"); 490 tty->print_cr("** Bailout: Recompile without escape analysis **"); 491 tty->print_cr("*********************************************************"); 492 } 493 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 494 // Recompiling without boxing elimination 495 tty->print_cr("*********************************************************"); 496 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 497 tty->print_cr("*********************************************************"); 498 } 499 if (C->directive()->BreakAtCompileOption) { 500 // Open the debugger when compiling this method. 501 tty->print("### Breaking when compiling: "); 502 method()->print_short_name(); 503 tty->cr(); 504 BREAKPOINT; 505 } 506 507 if( PrintOpto ) { 508 if (is_osr_compilation()) { 509 tty->print("[OSR]%3d", _compile_id); 510 } else { 511 tty->print("%3d", _compile_id); 512 } 513 } 514 #endif 515 } 516 517 518 //-----------------------init_scratch_buffer_blob------------------------------ 519 // Construct a temporary BufferBlob and cache it for this compile. 520 void Compile::init_scratch_buffer_blob(int const_size) { 521 // If there is already a scratch buffer blob allocated and the 522 // constant section is big enough, use it. Otherwise free the 523 // current and allocate a new one. 524 BufferBlob* blob = scratch_buffer_blob(); 525 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 526 // Use the current blob. 527 } else { 528 if (blob != NULL) { 529 BufferBlob::free(blob); 530 } 531 532 ResourceMark rm; 533 _scratch_const_size = const_size; 534 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 535 blob = BufferBlob::create("Compile::scratch_buffer", size); 536 // Record the buffer blob for next time. 537 set_scratch_buffer_blob(blob); 538 // Have we run out of code space? 539 if (scratch_buffer_blob() == NULL) { 540 // Let CompilerBroker disable further compilations. 541 record_failure("Not enough space for scratch buffer in CodeCache"); 542 return; 543 } 544 } 545 546 // Initialize the relocation buffers 547 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 548 set_scratch_locs_memory(locs_buf); 549 } 550 551 552 //-----------------------scratch_emit_size------------------------------------- 553 // Helper function that computes size by emitting code 554 uint Compile::scratch_emit_size(const Node* n) { 555 // Start scratch_emit_size section. 556 set_in_scratch_emit_size(true); 557 558 // Emit into a trash buffer and count bytes emitted. 559 // This is a pretty expensive way to compute a size, 560 // but it works well enough if seldom used. 561 // All common fixed-size instructions are given a size 562 // method by the AD file. 563 // Note that the scratch buffer blob and locs memory are 564 // allocated at the beginning of the compile task, and 565 // may be shared by several calls to scratch_emit_size. 566 // The allocation of the scratch buffer blob is particularly 567 // expensive, since it has to grab the code cache lock. 568 BufferBlob* blob = this->scratch_buffer_blob(); 569 assert(blob != NULL, "Initialize BufferBlob at start"); 570 assert(blob->size() > MAX_inst_size, "sanity"); 571 relocInfo* locs_buf = scratch_locs_memory(); 572 address blob_begin = blob->content_begin(); 573 address blob_end = (address)locs_buf; 574 assert(blob->content_contains(blob_end), "sanity"); 575 CodeBuffer buf(blob_begin, blob_end - blob_begin); 576 buf.initialize_consts_size(_scratch_const_size); 577 buf.initialize_stubs_size(MAX_stubs_size); 578 assert(locs_buf != NULL, "sanity"); 579 int lsize = MAX_locs_size / 3; 580 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 581 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 582 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 583 584 // Do the emission. 585 586 Label fakeL; // Fake label for branch instructions. 587 Label* saveL = NULL; 588 uint save_bnum = 0; 589 bool is_branch = n->is_MachBranch(); 590 if (is_branch) { 591 MacroAssembler masm(&buf); 592 masm.bind(fakeL); 593 n->as_MachBranch()->save_label(&saveL, &save_bnum); 594 n->as_MachBranch()->label_set(&fakeL, 0); 595 } 596 n->emit(buf, this->regalloc()); 597 598 // Emitting into the scratch buffer should not fail 599 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 600 601 if (is_branch) // Restore label. 602 n->as_MachBranch()->label_set(saveL, save_bnum); 603 604 // End scratch_emit_size section. 605 set_in_scratch_emit_size(false); 606 607 return buf.insts_size(); 608 } 609 610 611 // ============================================================================ 612 //------------------------------Compile standard------------------------------- 613 debug_only( int Compile::_debug_idx = 100000; ) 614 615 // Compile a method. entry_bci is -1 for normal compilations and indicates 616 // the continuation bci for on stack replacement. 617 618 619 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 620 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 621 : Phase(Compiler), 622 _env(ci_env), 623 _directive(directive), 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(directive->TraceOptoOutputOption), 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 = directive->PrintOptoAssemblyOption; 692 if (!print_opto_assembly) { 693 bool print_assembly = directive->PrintAssemblyOption; 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 (directive->ReplayInlineOption) { 703 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 704 } 705 #endif 706 set_print_inlining(directive->PrintInliningOption NOT_PRODUCT( || PrintOptoInlining)); 707 set_print_intrinsics(directive->PrintIntrinsicsOption); 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 && UseG1GC) { 762 // With java.lang.ref.reference.get() we must go through the 763 // intrinsic when G1 is enabled - even when get() is the root 764 // method of the compile - so that, if necessary, the value in 765 // the referent field of the reference object gets recorded by 766 // the pre-barrier code. 767 // Specifically, if G1 is enabled, the value in the referent 768 // field is recorded by the G1 SATB pre barrier. This will 769 // result in the referent being marked live and the reference 770 // object removed from the list of discovered references during 771 // reference processing. 772 cg = find_intrinsic(method(), false); 773 } 774 if (cg == NULL) { 775 float past_uses = method()->interpreter_invocation_count(); 776 float expected_uses = past_uses; 777 cg = CallGenerator::for_inline(method(), expected_uses); 778 } 779 } 780 if (failing()) return; 781 if (cg == NULL) { 782 record_method_not_compilable_all_tiers("cannot parse method"); 783 return; 784 } 785 JVMState* jvms = build_start_state(start(), tf()); 786 if ((jvms = cg->generate(jvms)) == NULL) { 787 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 788 record_method_not_compilable("method parse failed"); 789 } 790 return; 791 } 792 GraphKit kit(jvms); 793 794 if (!kit.stopped()) { 795 // Accept return values, and transfer control we know not where. 796 // This is done by a special, unique ReturnNode bound to root. 797 return_values(kit.jvms()); 798 } 799 800 if (kit.has_exceptions()) { 801 // Any exceptions that escape from this call must be rethrown 802 // to whatever caller is dynamically above us on the stack. 803 // This is done by a special, unique RethrowNode bound to root. 804 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 805 } 806 807 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 808 809 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 810 inline_string_calls(true); 811 } 812 813 if (failing()) return; 814 815 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 816 817 // Remove clutter produced by parsing. 818 if (!failing()) { 819 ResourceMark rm; 820 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 821 } 822 } 823 824 // Note: Large methods are capped off in do_one_bytecode(). 825 if (failing()) return; 826 827 // After parsing, node notes are no longer automagic. 828 // They must be propagated by register_new_node_with_optimizer(), 829 // clone(), or the like. 830 set_default_node_notes(NULL); 831 832 for (;;) { 833 int successes = Inline_Warm(); 834 if (failing()) return; 835 if (successes == 0) break; 836 } 837 838 // Drain the list. 839 Finish_Warm(); 840 #ifndef PRODUCT 841 if (_printer && _printer->should_print(1)) { 842 _printer->print_inlining(this); 843 } 844 #endif 845 846 if (failing()) return; 847 NOT_PRODUCT( verify_graph_edges(); ) 848 849 // Now optimize 850 Optimize(); 851 if (failing()) return; 852 NOT_PRODUCT( verify_graph_edges(); ) 853 854 #ifndef PRODUCT 855 if (PrintIdeal) { 856 ttyLocker ttyl; // keep the following output all in one block 857 // This output goes directly to the tty, not the compiler log. 858 // To enable tools to match it up with the compilation activity, 859 // be sure to tag this tty output with the compile ID. 860 if (xtty != NULL) { 861 xtty->head("ideal compile_id='%d'%s", compile_id(), 862 is_osr_compilation() ? " compile_kind='osr'" : 863 ""); 864 } 865 root()->dump(9999); 866 if (xtty != NULL) { 867 xtty->tail("ideal"); 868 } 869 } 870 #endif 871 872 NOT_PRODUCT( verify_barriers(); ) 873 874 // Dump compilation data to replay it. 875 if (directive->DumpReplayOption) { 876 env()->dump_replay_data(_compile_id); 877 } 878 if (directive->DumpInlineOption && (ilt() != NULL)) { 879 env()->dump_inline_data(_compile_id); 880 } 881 882 // Now that we know the size of all the monitors we can add a fixed slot 883 // for the original deopt pc. 884 885 _orig_pc_slot = fixed_slots(); 886 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 887 set_fixed_slots(next_slot); 888 889 // Compute when to use implicit null checks. Used by matching trap based 890 // nodes and NullCheck optimization. 891 set_allowed_deopt_reasons(); 892 893 // Now generate code 894 Code_Gen(); 895 if (failing()) return; 896 897 // Check if we want to skip execution of all compiled code. 898 { 899 #ifndef PRODUCT 900 if (OptoNoExecute) { 901 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 902 return; 903 } 904 #endif 905 TracePhase tp("install_code", &timers[_t_registerMethod]); 906 907 if (is_osr_compilation()) { 908 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 909 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 910 } else { 911 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 912 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 913 } 914 915 env()->register_method(_method, _entry_bci, 916 &_code_offsets, 917 _orig_pc_slot_offset_in_bytes, 918 code_buffer(), 919 frame_size_in_words(), _oop_map_set, 920 &_handler_table, &_inc_table, 921 compiler, 922 has_unsafe_access(), 923 SharedRuntime::is_wide_vector(max_vector_size()), 924 _directive, 925 rtm_state() 926 ); 927 928 if (log() != NULL) // Print code cache state into compiler log 929 log()->code_cache_state(); 930 } 931 } 932 933 //------------------------------Compile---------------------------------------- 934 // Compile a runtime stub 935 Compile::Compile( ciEnv* ci_env, 936 TypeFunc_generator generator, 937 address stub_function, 938 const char *stub_name, 939 int is_fancy_jump, 940 bool pass_tls, 941 bool save_arg_registers, 942 bool return_pc, 943 DirectiveSet* directive) 944 : Phase(Compiler), 945 _env(ci_env), 946 _directive(directive), 947 _log(ci_env->log()), 948 _compile_id(0), 949 _save_argument_registers(save_arg_registers), 950 _method(NULL), 951 _stub_name(stub_name), 952 _stub_function(stub_function), 953 _stub_entry_point(NULL), 954 _entry_bci(InvocationEntryBci), 955 _initial_gvn(NULL), 956 _for_igvn(NULL), 957 _warm_calls(NULL), 958 _orig_pc_slot(0), 959 _orig_pc_slot_offset_in_bytes(0), 960 _subsume_loads(true), 961 _do_escape_analysis(false), 962 _eliminate_boxing(false), 963 _failure_reason(NULL), 964 _code_buffer("Compile::Fill_buffer"), 965 _has_method_handle_invokes(false), 966 _mach_constant_base_node(NULL), 967 _node_bundling_limit(0), 968 _node_bundling_base(NULL), 969 _java_calls(0), 970 _inner_loops(0), 971 #ifndef PRODUCT 972 _trace_opto_output(TraceOptoOutput), 973 _in_dump_cnt(0), 974 _printer(NULL), 975 #endif 976 _comp_arena(mtCompiler), 977 _node_arena(mtCompiler), 978 _old_arena(mtCompiler), 979 _Compile_types(mtCompiler), 980 _dead_node_list(comp_arena()), 981 _dead_node_count(0), 982 _congraph(NULL), 983 _replay_inline_data(NULL), 984 _number_of_mh_late_inlines(0), 985 _inlining_progress(false), 986 _inlining_incrementally(false), 987 _print_inlining_list(NULL), 988 _print_inlining_stream(NULL), 989 _print_inlining_idx(0), 990 _print_inlining_output(NULL), 991 _allowed_reasons(0), 992 _interpreter_frame_size(0), 993 _max_node_limit(MaxNodeLimit) { 994 C = this; 995 996 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 997 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 998 999 #ifndef PRODUCT 1000 set_print_assembly(PrintFrameConverterAssembly); 1001 set_parsed_irreducible_loop(false); 1002 #endif 1003 set_has_irreducible_loop(false); // no loops 1004 1005 CompileWrapper cw(this); 1006 Init(/*AliasLevel=*/ 0); 1007 init_tf((*generator)()); 1008 1009 { 1010 // The following is a dummy for the sake of GraphKit::gen_stub 1011 Unique_Node_List for_igvn(comp_arena()); 1012 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1013 PhaseGVN gvn(Thread::current()->resource_area(),255); 1014 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1015 gvn.transform_no_reclaim(top()); 1016 1017 GraphKit kit; 1018 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1019 } 1020 1021 NOT_PRODUCT( verify_graph_edges(); ) 1022 Code_Gen(); 1023 if (failing()) return; 1024 1025 1026 // Entry point will be accessed using compile->stub_entry_point(); 1027 if (code_buffer() == NULL) { 1028 Matcher::soft_match_failure(); 1029 } else { 1030 if (PrintAssembly && (WizardMode || Verbose)) 1031 tty->print_cr("### Stub::%s", stub_name); 1032 1033 if (!failing()) { 1034 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1035 1036 // Make the NMethod 1037 // For now we mark the frame as never safe for profile stackwalking 1038 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1039 code_buffer(), 1040 CodeOffsets::frame_never_safe, 1041 // _code_offsets.value(CodeOffsets::Frame_Complete), 1042 frame_size_in_words(), 1043 _oop_map_set, 1044 save_arg_registers); 1045 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1046 1047 _stub_entry_point = rs->entry_point(); 1048 } 1049 } 1050 } 1051 1052 //------------------------------Init------------------------------------------- 1053 // Prepare for a single compilation 1054 void Compile::Init(int aliaslevel) { 1055 _unique = 0; 1056 _regalloc = NULL; 1057 1058 _tf = NULL; // filled in later 1059 _top = NULL; // cached later 1060 _matcher = NULL; // filled in later 1061 _cfg = NULL; // filled in later 1062 1063 set_24_bit_selection_and_mode(Use24BitFP, false); 1064 1065 _node_note_array = NULL; 1066 _default_node_notes = NULL; 1067 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1068 1069 _immutable_memory = NULL; // filled in at first inquiry 1070 1071 // Globally visible Nodes 1072 // First set TOP to NULL to give safe behavior during creation of RootNode 1073 set_cached_top_node(NULL); 1074 set_root(new RootNode()); 1075 // Now that you have a Root to point to, create the real TOP 1076 set_cached_top_node( new ConNode(Type::TOP) ); 1077 set_recent_alloc(NULL, NULL); 1078 1079 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1080 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1081 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1082 env()->set_dependencies(new Dependencies(env())); 1083 1084 _fixed_slots = 0; 1085 set_has_split_ifs(false); 1086 set_has_loops(has_method() && method()->has_loops()); // first approximation 1087 set_has_stringbuilder(false); 1088 set_has_boxed_value(false); 1089 _trap_can_recompile = false; // no traps emitted yet 1090 _major_progress = true; // start out assuming good things will happen 1091 set_has_unsafe_access(false); 1092 set_max_vector_size(0); 1093 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1094 set_decompile_count(0); 1095 1096 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1097 set_num_loop_opts(LoopOptsCount); 1098 set_do_inlining(Inline); 1099 set_max_inline_size(MaxInlineSize); 1100 set_freq_inline_size(FreqInlineSize); 1101 set_do_scheduling(OptoScheduling); 1102 set_do_count_invocations(false); 1103 set_do_method_data_update(false); 1104 1105 set_do_vector_loop(false); 1106 1107 bool do_vector = false; 1108 if (AllowVectorizeOnDemand) { 1109 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1110 set_do_vector_loop(true); 1111 } else if (has_method() && method()->name() != 0 && 1112 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1113 set_do_vector_loop(true); 1114 } 1115 #ifndef PRODUCT 1116 if (do_vector_loop() && Verbose) { 1117 tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii()); 1118 } 1119 #endif 1120 } 1121 1122 set_age_code(has_method() && method()->profile_aging()); 1123 set_rtm_state(NoRTM); // No RTM lock eliding by default 1124 _max_node_limit = _directive->MaxNodeLimitOption; 1125 1126 #if INCLUDE_RTM_OPT 1127 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1128 int rtm_state = method()->method_data()->rtm_state(); 1129 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1130 // Don't generate RTM lock eliding code. 1131 set_rtm_state(NoRTM); 1132 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1133 // Generate RTM lock eliding code without abort ratio calculation code. 1134 set_rtm_state(UseRTM); 1135 } else if (UseRTMDeopt) { 1136 // Generate RTM lock eliding code and include abort ratio calculation 1137 // code if UseRTMDeopt is on. 1138 set_rtm_state(ProfileRTM); 1139 } 1140 } 1141 #endif 1142 if (debug_info()->recording_non_safepoints()) { 1143 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1144 (comp_arena(), 8, 0, NULL)); 1145 set_default_node_notes(Node_Notes::make(this)); 1146 } 1147 1148 // // -- Initialize types before each compile -- 1149 // // Update cached type information 1150 // if( _method && _method->constants() ) 1151 // Type::update_loaded_types(_method, _method->constants()); 1152 1153 // Init alias_type map. 1154 if (!_do_escape_analysis && aliaslevel == 3) 1155 aliaslevel = 2; // No unique types without escape analysis 1156 _AliasLevel = aliaslevel; 1157 const int grow_ats = 16; 1158 _max_alias_types = grow_ats; 1159 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1160 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1161 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1162 { 1163 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1164 } 1165 // Initialize the first few types. 1166 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1167 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1168 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1169 _num_alias_types = AliasIdxRaw+1; 1170 // Zero out the alias type cache. 1171 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1172 // A NULL adr_type hits in the cache right away. Preload the right answer. 1173 probe_alias_cache(NULL)->_index = AliasIdxTop; 1174 1175 _intrinsics = NULL; 1176 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1177 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1178 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1179 register_library_intrinsics(); 1180 } 1181 1182 //---------------------------init_start---------------------------------------- 1183 // Install the StartNode on this compile object. 1184 void Compile::init_start(StartNode* s) { 1185 if (failing()) 1186 return; // already failing 1187 assert(s == start(), ""); 1188 } 1189 1190 /** 1191 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1192 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1193 * the ideal graph. 1194 */ 1195 StartNode* Compile::start() const { 1196 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 1197 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1198 Node* start = root()->fast_out(i); 1199 if (start->is_Start()) { 1200 return start->as_Start(); 1201 } 1202 } 1203 fatal("Did not find Start node!"); 1204 return NULL; 1205 } 1206 1207 //-------------------------------immutable_memory------------------------------------- 1208 // Access immutable memory 1209 Node* Compile::immutable_memory() { 1210 if (_immutable_memory != NULL) { 1211 return _immutable_memory; 1212 } 1213 StartNode* s = start(); 1214 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1215 Node *p = s->fast_out(i); 1216 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1217 _immutable_memory = p; 1218 return _immutable_memory; 1219 } 1220 } 1221 ShouldNotReachHere(); 1222 return NULL; 1223 } 1224 1225 //----------------------set_cached_top_node------------------------------------ 1226 // Install the cached top node, and make sure Node::is_top works correctly. 1227 void Compile::set_cached_top_node(Node* tn) { 1228 if (tn != NULL) verify_top(tn); 1229 Node* old_top = _top; 1230 _top = tn; 1231 // Calling Node::setup_is_top allows the nodes the chance to adjust 1232 // their _out arrays. 1233 if (_top != NULL) _top->setup_is_top(); 1234 if (old_top != NULL) old_top->setup_is_top(); 1235 assert(_top == NULL || top()->is_top(), ""); 1236 } 1237 1238 #ifdef ASSERT 1239 uint Compile::count_live_nodes_by_graph_walk() { 1240 Unique_Node_List useful(comp_arena()); 1241 // Get useful node list by walking the graph. 1242 identify_useful_nodes(useful); 1243 return useful.size(); 1244 } 1245 1246 void Compile::print_missing_nodes() { 1247 1248 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1249 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1250 return; 1251 } 1252 1253 // This is an expensive function. It is executed only when the user 1254 // specifies VerifyIdealNodeCount option or otherwise knows the 1255 // additional work that needs to be done to identify reachable nodes 1256 // by walking the flow graph and find the missing ones using 1257 // _dead_node_list. 1258 1259 Unique_Node_List useful(comp_arena()); 1260 // Get useful node list by walking the graph. 1261 identify_useful_nodes(useful); 1262 1263 uint l_nodes = C->live_nodes(); 1264 uint l_nodes_by_walk = useful.size(); 1265 1266 if (l_nodes != l_nodes_by_walk) { 1267 if (_log != NULL) { 1268 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1269 _log->stamp(); 1270 _log->end_head(); 1271 } 1272 VectorSet& useful_member_set = useful.member_set(); 1273 int last_idx = l_nodes_by_walk; 1274 for (int i = 0; i < last_idx; i++) { 1275 if (useful_member_set.test(i)) { 1276 if (_dead_node_list.test(i)) { 1277 if (_log != NULL) { 1278 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1279 } 1280 if (PrintIdealNodeCount) { 1281 // Print the log message to tty 1282 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1283 useful.at(i)->dump(); 1284 } 1285 } 1286 } 1287 else if (! _dead_node_list.test(i)) { 1288 if (_log != NULL) { 1289 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1290 } 1291 if (PrintIdealNodeCount) { 1292 // Print the log message to tty 1293 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1294 } 1295 } 1296 } 1297 if (_log != NULL) { 1298 _log->tail("mismatched_nodes"); 1299 } 1300 } 1301 } 1302 void Compile::record_modified_node(Node* n) { 1303 if (_modified_nodes != NULL && !_inlining_incrementally && 1304 n->outcnt() != 0 && !n->is_Con()) { 1305 _modified_nodes->push(n); 1306 } 1307 } 1308 1309 void Compile::remove_modified_node(Node* n) { 1310 if (_modified_nodes != NULL) { 1311 _modified_nodes->remove(n); 1312 } 1313 } 1314 #endif 1315 1316 #ifndef PRODUCT 1317 void Compile::verify_top(Node* tn) const { 1318 if (tn != NULL) { 1319 assert(tn->is_Con(), "top node must be a constant"); 1320 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1321 assert(tn->in(0) != NULL, "must have live top node"); 1322 } 1323 } 1324 #endif 1325 1326 1327 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1328 1329 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1330 guarantee(arr != NULL, ""); 1331 int num_blocks = arr->length(); 1332 if (grow_by < num_blocks) grow_by = num_blocks; 1333 int num_notes = grow_by * _node_notes_block_size; 1334 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1335 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1336 while (num_notes > 0) { 1337 arr->append(notes); 1338 notes += _node_notes_block_size; 1339 num_notes -= _node_notes_block_size; 1340 } 1341 assert(num_notes == 0, "exact multiple, please"); 1342 } 1343 1344 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1345 if (source == NULL || dest == NULL) return false; 1346 1347 if (dest->is_Con()) 1348 return false; // Do not push debug info onto constants. 1349 1350 #ifdef ASSERT 1351 // Leave a bread crumb trail pointing to the original node: 1352 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1353 dest->set_debug_orig(source); 1354 } 1355 #endif 1356 1357 if (node_note_array() == NULL) 1358 return false; // Not collecting any notes now. 1359 1360 // This is a copy onto a pre-existing node, which may already have notes. 1361 // If both nodes have notes, do not overwrite any pre-existing notes. 1362 Node_Notes* source_notes = node_notes_at(source->_idx); 1363 if (source_notes == NULL || source_notes->is_clear()) return false; 1364 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1365 if (dest_notes == NULL || dest_notes->is_clear()) { 1366 return set_node_notes_at(dest->_idx, source_notes); 1367 } 1368 1369 Node_Notes merged_notes = (*source_notes); 1370 // The order of operations here ensures that dest notes will win... 1371 merged_notes.update_from(dest_notes); 1372 return set_node_notes_at(dest->_idx, &merged_notes); 1373 } 1374 1375 1376 //--------------------------allow_range_check_smearing------------------------- 1377 // Gating condition for coalescing similar range checks. 1378 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1379 // single covering check that is at least as strong as any of them. 1380 // If the optimization succeeds, the simplified (strengthened) range check 1381 // will always succeed. If it fails, we will deopt, and then give up 1382 // on the optimization. 1383 bool Compile::allow_range_check_smearing() const { 1384 // If this method has already thrown a range-check, 1385 // assume it was because we already tried range smearing 1386 // and it failed. 1387 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1388 return !already_trapped; 1389 } 1390 1391 1392 //------------------------------flatten_alias_type----------------------------- 1393 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1394 int offset = tj->offset(); 1395 TypePtr::PTR ptr = tj->ptr(); 1396 1397 // Known instance (scalarizable allocation) alias only with itself. 1398 bool is_known_inst = tj->isa_oopptr() != NULL && 1399 tj->is_oopptr()->is_known_instance(); 1400 1401 // Process weird unsafe references. 1402 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1403 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1404 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1405 tj = TypeOopPtr::BOTTOM; 1406 ptr = tj->ptr(); 1407 offset = tj->offset(); 1408 } 1409 1410 // Array pointers need some flattening 1411 const TypeAryPtr *ta = tj->isa_aryptr(); 1412 if (ta && ta->is_stable()) { 1413 // Erase stability property for alias analysis. 1414 tj = ta = ta->cast_to_stable(false); 1415 } 1416 if( ta && is_known_inst ) { 1417 if ( offset != Type::OffsetBot && 1418 offset > arrayOopDesc::length_offset_in_bytes() ) { 1419 offset = Type::OffsetBot; // Flatten constant access into array body only 1420 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1421 } 1422 } else if( ta && _AliasLevel >= 2 ) { 1423 // For arrays indexed by constant indices, we flatten the alias 1424 // space to include all of the array body. Only the header, klass 1425 // and array length can be accessed un-aliased. 1426 if( offset != Type::OffsetBot ) { 1427 if( ta->const_oop() ) { // MethodData* or Method* 1428 offset = Type::OffsetBot; // Flatten constant access into array body 1429 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1430 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1431 // range is OK as-is. 1432 tj = ta = TypeAryPtr::RANGE; 1433 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1434 tj = TypeInstPtr::KLASS; // all klass loads look alike 1435 ta = TypeAryPtr::RANGE; // generic ignored junk 1436 ptr = TypePtr::BotPTR; 1437 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1438 tj = TypeInstPtr::MARK; 1439 ta = TypeAryPtr::RANGE; // generic ignored junk 1440 ptr = TypePtr::BotPTR; 1441 } else { // Random constant offset into array body 1442 offset = Type::OffsetBot; // Flatten constant access into array body 1443 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1444 } 1445 } 1446 // Arrays of fixed size alias with arrays of unknown size. 1447 if (ta->size() != TypeInt::POS) { 1448 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1449 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1450 } 1451 // Arrays of known objects become arrays of unknown objects. 1452 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1453 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1454 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1455 } 1456 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1457 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1458 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1459 } 1460 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1461 // cannot be distinguished by bytecode alone. 1462 if (ta->elem() == TypeInt::BOOL) { 1463 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1464 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1465 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1466 } 1467 // During the 2nd round of IterGVN, NotNull castings are removed. 1468 // Make sure the Bottom and NotNull variants alias the same. 1469 // Also, make sure exact and non-exact variants alias the same. 1470 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1471 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1472 } 1473 } 1474 1475 // Oop pointers need some flattening 1476 const TypeInstPtr *to = tj->isa_instptr(); 1477 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1478 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1479 if( ptr == TypePtr::Constant ) { 1480 if (to->klass() != ciEnv::current()->Class_klass() || 1481 offset < k->size_helper() * wordSize) { 1482 // No constant oop pointers (such as Strings); they alias with 1483 // unknown strings. 1484 assert(!is_known_inst, "not scalarizable allocation"); 1485 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1486 } 1487 } else if( is_known_inst ) { 1488 tj = to; // Keep NotNull and klass_is_exact for instance type 1489 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1490 // During the 2nd round of IterGVN, NotNull castings are removed. 1491 // Make sure the Bottom and NotNull variants alias the same. 1492 // Also, make sure exact and non-exact variants alias the same. 1493 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1494 } 1495 if (to->speculative() != NULL) { 1496 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1497 } 1498 // Canonicalize the holder of this field 1499 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1500 // First handle header references such as a LoadKlassNode, even if the 1501 // object's klass is unloaded at compile time (4965979). 1502 if (!is_known_inst) { // Do it only for non-instance types 1503 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1504 } 1505 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1506 // Static fields are in the space above the normal instance 1507 // fields in the java.lang.Class instance. 1508 if (to->klass() != ciEnv::current()->Class_klass()) { 1509 to = NULL; 1510 tj = TypeOopPtr::BOTTOM; 1511 offset = tj->offset(); 1512 } 1513 } else { 1514 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1515 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1516 if( is_known_inst ) { 1517 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1518 } else { 1519 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1520 } 1521 } 1522 } 1523 } 1524 1525 // Klass pointers to object array klasses need some flattening 1526 const TypeKlassPtr *tk = tj->isa_klassptr(); 1527 if( tk ) { 1528 // If we are referencing a field within a Klass, we need 1529 // to assume the worst case of an Object. Both exact and 1530 // inexact types must flatten to the same alias class so 1531 // use NotNull as the PTR. 1532 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1533 1534 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1535 TypeKlassPtr::OBJECT->klass(), 1536 offset); 1537 } 1538 1539 ciKlass* klass = tk->klass(); 1540 if( klass->is_obj_array_klass() ) { 1541 ciKlass* k = TypeAryPtr::OOPS->klass(); 1542 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1543 k = TypeInstPtr::BOTTOM->klass(); 1544 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1545 } 1546 1547 // Check for precise loads from the primary supertype array and force them 1548 // to the supertype cache alias index. Check for generic array loads from 1549 // the primary supertype array and also force them to the supertype cache 1550 // alias index. Since the same load can reach both, we need to merge 1551 // these 2 disparate memories into the same alias class. Since the 1552 // primary supertype array is read-only, there's no chance of confusion 1553 // where we bypass an array load and an array store. 1554 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1555 if (offset == Type::OffsetBot || 1556 (offset >= primary_supers_offset && 1557 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1558 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1559 offset = in_bytes(Klass::secondary_super_cache_offset()); 1560 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1561 } 1562 } 1563 1564 // Flatten all Raw pointers together. 1565 if (tj->base() == Type::RawPtr) 1566 tj = TypeRawPtr::BOTTOM; 1567 1568 if (tj->base() == Type::AnyPtr) 1569 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1570 1571 // Flatten all to bottom for now 1572 switch( _AliasLevel ) { 1573 case 0: 1574 tj = TypePtr::BOTTOM; 1575 break; 1576 case 1: // Flatten to: oop, static, field or array 1577 switch (tj->base()) { 1578 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1579 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1580 case Type::AryPtr: // do not distinguish arrays at all 1581 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1582 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1583 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1584 default: ShouldNotReachHere(); 1585 } 1586 break; 1587 case 2: // No collapsing at level 2; keep all splits 1588 case 3: // No collapsing at level 3; keep all splits 1589 break; 1590 default: 1591 Unimplemented(); 1592 } 1593 1594 offset = tj->offset(); 1595 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1596 1597 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1598 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1599 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1600 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1601 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1602 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1603 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1604 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1605 assert( tj->ptr() != TypePtr::TopPTR && 1606 tj->ptr() != TypePtr::AnyNull && 1607 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1608 // assert( tj->ptr() != TypePtr::Constant || 1609 // tj->base() == Type::RawPtr || 1610 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1611 1612 return tj; 1613 } 1614 1615 void Compile::AliasType::Init(int i, const TypePtr* at) { 1616 _index = i; 1617 _adr_type = at; 1618 _field = NULL; 1619 _element = NULL; 1620 _is_rewritable = true; // default 1621 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1622 if (atoop != NULL && atoop->is_known_instance()) { 1623 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1624 _general_index = Compile::current()->get_alias_index(gt); 1625 } else { 1626 _general_index = 0; 1627 } 1628 } 1629 1630 //---------------------------------print_on------------------------------------ 1631 #ifndef PRODUCT 1632 void Compile::AliasType::print_on(outputStream* st) { 1633 if (index() < 10) 1634 st->print("@ <%d> ", index()); 1635 else st->print("@ <%d>", index()); 1636 st->print(is_rewritable() ? " " : " RO"); 1637 int offset = adr_type()->offset(); 1638 if (offset == Type::OffsetBot) 1639 st->print(" +any"); 1640 else st->print(" +%-3d", offset); 1641 st->print(" in "); 1642 adr_type()->dump_on(st); 1643 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1644 if (field() != NULL && tjp) { 1645 if (tjp->klass() != field()->holder() || 1646 tjp->offset() != field()->offset_in_bytes()) { 1647 st->print(" != "); 1648 field()->print(); 1649 st->print(" ***"); 1650 } 1651 } 1652 } 1653 1654 void print_alias_types() { 1655 Compile* C = Compile::current(); 1656 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1657 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1658 C->alias_type(idx)->print_on(tty); 1659 tty->cr(); 1660 } 1661 } 1662 #endif 1663 1664 1665 //----------------------------probe_alias_cache-------------------------------- 1666 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1667 intptr_t key = (intptr_t) adr_type; 1668 key ^= key >> logAliasCacheSize; 1669 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1670 } 1671 1672 1673 //-----------------------------grow_alias_types-------------------------------- 1674 void Compile::grow_alias_types() { 1675 const int old_ats = _max_alias_types; // how many before? 1676 const int new_ats = old_ats; // how many more? 1677 const int grow_ats = old_ats+new_ats; // how many now? 1678 _max_alias_types = grow_ats; 1679 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1680 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1681 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1682 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1683 } 1684 1685 1686 //--------------------------------find_alias_type------------------------------ 1687 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1688 if (_AliasLevel == 0) 1689 return alias_type(AliasIdxBot); 1690 1691 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1692 if (ace->_adr_type == adr_type) { 1693 return alias_type(ace->_index); 1694 } 1695 1696 // Handle special cases. 1697 if (adr_type == NULL) return alias_type(AliasIdxTop); 1698 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1699 1700 // Do it the slow way. 1701 const TypePtr* flat = flatten_alias_type(adr_type); 1702 1703 #ifdef ASSERT 1704 assert(flat == flatten_alias_type(flat), "idempotent"); 1705 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1706 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1707 const TypeOopPtr* foop = flat->is_oopptr(); 1708 // Scalarizable allocations have exact klass always. 1709 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1710 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1711 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1712 } 1713 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1714 #endif 1715 1716 int idx = AliasIdxTop; 1717 for (int i = 0; i < num_alias_types(); i++) { 1718 if (alias_type(i)->adr_type() == flat) { 1719 idx = i; 1720 break; 1721 } 1722 } 1723 1724 if (idx == AliasIdxTop) { 1725 if (no_create) return NULL; 1726 // Grow the array if necessary. 1727 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1728 // Add a new alias type. 1729 idx = _num_alias_types++; 1730 _alias_types[idx]->Init(idx, flat); 1731 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1732 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1733 if (flat->isa_instptr()) { 1734 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1735 && flat->is_instptr()->klass() == env()->Class_klass()) 1736 alias_type(idx)->set_rewritable(false); 1737 } 1738 if (flat->isa_aryptr()) { 1739 #ifdef ASSERT 1740 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1741 // (T_BYTE has the weakest alignment and size restrictions...) 1742 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1743 #endif 1744 if (flat->offset() == TypePtr::OffsetBot) { 1745 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1746 } 1747 } 1748 if (flat->isa_klassptr()) { 1749 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1750 alias_type(idx)->set_rewritable(false); 1751 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1752 alias_type(idx)->set_rewritable(false); 1753 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1754 alias_type(idx)->set_rewritable(false); 1755 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1756 alias_type(idx)->set_rewritable(false); 1757 } 1758 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1759 // but the base pointer type is not distinctive enough to identify 1760 // references into JavaThread.) 1761 1762 // Check for final fields. 1763 const TypeInstPtr* tinst = flat->isa_instptr(); 1764 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1765 ciField* field; 1766 if (tinst->const_oop() != NULL && 1767 tinst->klass() == ciEnv::current()->Class_klass() && 1768 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1769 // static field 1770 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1771 field = k->get_field_by_offset(tinst->offset(), true); 1772 } else { 1773 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1774 field = k->get_field_by_offset(tinst->offset(), false); 1775 } 1776 assert(field == NULL || 1777 original_field == NULL || 1778 (field->holder() == original_field->holder() && 1779 field->offset() == original_field->offset() && 1780 field->is_static() == original_field->is_static()), "wrong field?"); 1781 // Set field() and is_rewritable() attributes. 1782 if (field != NULL) alias_type(idx)->set_field(field); 1783 } 1784 } 1785 1786 // Fill the cache for next time. 1787 ace->_adr_type = adr_type; 1788 ace->_index = idx; 1789 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1790 1791 // Might as well try to fill the cache for the flattened version, too. 1792 AliasCacheEntry* face = probe_alias_cache(flat); 1793 if (face->_adr_type == NULL) { 1794 face->_adr_type = flat; 1795 face->_index = idx; 1796 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1797 } 1798 1799 return alias_type(idx); 1800 } 1801 1802 1803 Compile::AliasType* Compile::alias_type(ciField* field) { 1804 const TypeOopPtr* t; 1805 if (field->is_static()) 1806 t = TypeInstPtr::make(field->holder()->java_mirror()); 1807 else 1808 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1809 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1810 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1811 return atp; 1812 } 1813 1814 1815 //------------------------------have_alias_type-------------------------------- 1816 bool Compile::have_alias_type(const TypePtr* adr_type) { 1817 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1818 if (ace->_adr_type == adr_type) { 1819 return true; 1820 } 1821 1822 // Handle special cases. 1823 if (adr_type == NULL) return true; 1824 if (adr_type == TypePtr::BOTTOM) return true; 1825 1826 return find_alias_type(adr_type, true, NULL) != NULL; 1827 } 1828 1829 //-----------------------------must_alias-------------------------------------- 1830 // True if all values of the given address type are in the given alias category. 1831 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1832 if (alias_idx == AliasIdxBot) return true; // the universal category 1833 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1834 if (alias_idx == AliasIdxTop) return false; // the empty category 1835 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1836 1837 // the only remaining possible overlap is identity 1838 int adr_idx = get_alias_index(adr_type); 1839 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1840 assert(adr_idx == alias_idx || 1841 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1842 && adr_type != TypeOopPtr::BOTTOM), 1843 "should not be testing for overlap with an unsafe pointer"); 1844 return adr_idx == alias_idx; 1845 } 1846 1847 //------------------------------can_alias-------------------------------------- 1848 // True if any values of the given address type are in the given alias category. 1849 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1850 if (alias_idx == AliasIdxTop) return false; // the empty category 1851 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1852 if (alias_idx == AliasIdxBot) return true; // the universal category 1853 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1854 1855 // the only remaining possible overlap is identity 1856 int adr_idx = get_alias_index(adr_type); 1857 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1858 return adr_idx == alias_idx; 1859 } 1860 1861 1862 1863 //---------------------------pop_warm_call------------------------------------- 1864 WarmCallInfo* Compile::pop_warm_call() { 1865 WarmCallInfo* wci = _warm_calls; 1866 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1867 return wci; 1868 } 1869 1870 //----------------------------Inline_Warm-------------------------------------- 1871 int Compile::Inline_Warm() { 1872 // If there is room, try to inline some more warm call sites. 1873 // %%% Do a graph index compaction pass when we think we're out of space? 1874 if (!InlineWarmCalls) return 0; 1875 1876 int calls_made_hot = 0; 1877 int room_to_grow = NodeCountInliningCutoff - unique(); 1878 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1879 int amount_grown = 0; 1880 WarmCallInfo* call; 1881 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1882 int est_size = (int)call->size(); 1883 if (est_size > (room_to_grow - amount_grown)) { 1884 // This one won't fit anyway. Get rid of it. 1885 call->make_cold(); 1886 continue; 1887 } 1888 call->make_hot(); 1889 calls_made_hot++; 1890 amount_grown += est_size; 1891 amount_to_grow -= est_size; 1892 } 1893 1894 if (calls_made_hot > 0) set_major_progress(); 1895 return calls_made_hot; 1896 } 1897 1898 1899 //----------------------------Finish_Warm-------------------------------------- 1900 void Compile::Finish_Warm() { 1901 if (!InlineWarmCalls) return; 1902 if (failing()) return; 1903 if (warm_calls() == NULL) return; 1904 1905 // Clean up loose ends, if we are out of space for inlining. 1906 WarmCallInfo* call; 1907 while ((call = pop_warm_call()) != NULL) { 1908 call->make_cold(); 1909 } 1910 } 1911 1912 //---------------------cleanup_loop_predicates----------------------- 1913 // Remove the opaque nodes that protect the predicates so that all unused 1914 // checks and uncommon_traps will be eliminated from the ideal graph 1915 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1916 if (predicate_count()==0) return; 1917 for (int i = predicate_count(); i > 0; i--) { 1918 Node * n = predicate_opaque1_node(i-1); 1919 assert(n->Opcode() == Op_Opaque1, "must be"); 1920 igvn.replace_node(n, n->in(1)); 1921 } 1922 assert(predicate_count()==0, "should be clean!"); 1923 } 1924 1925 // StringOpts and late inlining of string methods 1926 void Compile::inline_string_calls(bool parse_time) { 1927 { 1928 // remove useless nodes to make the usage analysis simpler 1929 ResourceMark rm; 1930 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1931 } 1932 1933 { 1934 ResourceMark rm; 1935 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1936 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1937 print_method(PHASE_AFTER_STRINGOPTS, 3); 1938 } 1939 1940 // now inline anything that we skipped the first time around 1941 if (!parse_time) { 1942 _late_inlines_pos = _late_inlines.length(); 1943 } 1944 1945 while (_string_late_inlines.length() > 0) { 1946 CallGenerator* cg = _string_late_inlines.pop(); 1947 cg->do_late_inline(); 1948 if (failing()) return; 1949 } 1950 _string_late_inlines.trunc_to(0); 1951 } 1952 1953 // Late inlining of boxing methods 1954 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1955 if (_boxing_late_inlines.length() > 0) { 1956 assert(has_boxed_value(), "inconsistent"); 1957 1958 PhaseGVN* gvn = initial_gvn(); 1959 set_inlining_incrementally(true); 1960 1961 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1962 for_igvn()->clear(); 1963 gvn->replace_with(&igvn); 1964 1965 _late_inlines_pos = _late_inlines.length(); 1966 1967 while (_boxing_late_inlines.length() > 0) { 1968 CallGenerator* cg = _boxing_late_inlines.pop(); 1969 cg->do_late_inline(); 1970 if (failing()) return; 1971 } 1972 _boxing_late_inlines.trunc_to(0); 1973 1974 { 1975 ResourceMark rm; 1976 PhaseRemoveUseless pru(gvn, for_igvn()); 1977 } 1978 1979 igvn = PhaseIterGVN(gvn); 1980 igvn.optimize(); 1981 1982 set_inlining_progress(false); 1983 set_inlining_incrementally(false); 1984 } 1985 } 1986 1987 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1988 assert(IncrementalInline, "incremental inlining should be on"); 1989 PhaseGVN* gvn = initial_gvn(); 1990 1991 set_inlining_progress(false); 1992 for_igvn()->clear(); 1993 gvn->replace_with(&igvn); 1994 1995 { 1996 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 1997 int i = 0; 1998 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1999 CallGenerator* cg = _late_inlines.at(i); 2000 _late_inlines_pos = i+1; 2001 cg->do_late_inline(); 2002 if (failing()) return; 2003 } 2004 int j = 0; 2005 for (; i < _late_inlines.length(); i++, j++) { 2006 _late_inlines.at_put(j, _late_inlines.at(i)); 2007 } 2008 _late_inlines.trunc_to(j); 2009 } 2010 2011 { 2012 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2013 ResourceMark rm; 2014 PhaseRemoveUseless pru(gvn, for_igvn()); 2015 } 2016 2017 { 2018 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2019 igvn = PhaseIterGVN(gvn); 2020 } 2021 } 2022 2023 // Perform incremental inlining until bound on number of live nodes is reached 2024 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2025 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2026 2027 PhaseGVN* gvn = initial_gvn(); 2028 2029 set_inlining_incrementally(true); 2030 set_inlining_progress(true); 2031 uint low_live_nodes = 0; 2032 2033 while(inlining_progress() && _late_inlines.length() > 0) { 2034 2035 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2036 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2037 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2038 // PhaseIdealLoop is expensive so we only try it once we are 2039 // out of live nodes and we only try it again if the previous 2040 // helped got the number of nodes down significantly 2041 PhaseIdealLoop ideal_loop( igvn, false, true ); 2042 if (failing()) return; 2043 low_live_nodes = live_nodes(); 2044 _major_progress = true; 2045 } 2046 2047 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2048 break; 2049 } 2050 } 2051 2052 inline_incrementally_one(igvn); 2053 2054 if (failing()) return; 2055 2056 { 2057 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2058 igvn.optimize(); 2059 } 2060 2061 if (failing()) return; 2062 } 2063 2064 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2065 2066 if (_string_late_inlines.length() > 0) { 2067 assert(has_stringbuilder(), "inconsistent"); 2068 for_igvn()->clear(); 2069 initial_gvn()->replace_with(&igvn); 2070 2071 inline_string_calls(false); 2072 2073 if (failing()) return; 2074 2075 { 2076 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2077 ResourceMark rm; 2078 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2079 } 2080 2081 { 2082 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2083 igvn = PhaseIterGVN(gvn); 2084 igvn.optimize(); 2085 } 2086 } 2087 2088 set_inlining_incrementally(false); 2089 } 2090 2091 2092 //------------------------------Optimize--------------------------------------- 2093 // Given a graph, optimize it. 2094 void Compile::Optimize() { 2095 TracePhase tp("optimizer", &timers[_t_optimizer]); 2096 2097 #ifndef PRODUCT 2098 if (_directive->BreakAtCompileOption) { 2099 BREAKPOINT; 2100 } 2101 2102 #endif 2103 2104 ResourceMark rm; 2105 int loop_opts_cnt; 2106 2107 print_inlining_reinit(); 2108 2109 NOT_PRODUCT( verify_graph_edges(); ) 2110 2111 print_method(PHASE_AFTER_PARSING); 2112 2113 { 2114 // Iterative Global Value Numbering, including ideal transforms 2115 // Initialize IterGVN with types and values from parse-time GVN 2116 PhaseIterGVN igvn(initial_gvn()); 2117 #ifdef ASSERT 2118 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2119 #endif 2120 { 2121 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2122 igvn.optimize(); 2123 } 2124 2125 print_method(PHASE_ITER_GVN1, 2); 2126 2127 if (failing()) return; 2128 2129 inline_incrementally(igvn); 2130 2131 print_method(PHASE_INCREMENTAL_INLINE, 2); 2132 2133 if (failing()) return; 2134 2135 if (eliminate_boxing()) { 2136 // Inline valueOf() methods now. 2137 inline_boxing_calls(igvn); 2138 2139 if (AlwaysIncrementalInline) { 2140 inline_incrementally(igvn); 2141 } 2142 2143 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2144 2145 if (failing()) return; 2146 } 2147 2148 // Remove the speculative part of types and clean up the graph from 2149 // the extra CastPP nodes whose only purpose is to carry them. Do 2150 // that early so that optimizations are not disrupted by the extra 2151 // CastPP nodes. 2152 remove_speculative_types(igvn); 2153 2154 // No more new expensive nodes will be added to the list from here 2155 // so keep only the actual candidates for optimizations. 2156 cleanup_expensive_nodes(igvn); 2157 2158 // Perform escape analysis 2159 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2160 if (has_loops()) { 2161 // Cleanup graph (remove dead nodes). 2162 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2163 PhaseIdealLoop ideal_loop( igvn, false, true ); 2164 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2165 if (failing()) return; 2166 } 2167 ConnectionGraph::do_analysis(this, &igvn); 2168 2169 if (failing()) return; 2170 2171 // Optimize out fields loads from scalar replaceable allocations. 2172 igvn.optimize(); 2173 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2174 2175 if (failing()) return; 2176 2177 if (congraph() != NULL && macro_count() > 0) { 2178 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2179 PhaseMacroExpand mexp(igvn); 2180 mexp.eliminate_macro_nodes(); 2181 igvn.set_delay_transform(false); 2182 2183 igvn.optimize(); 2184 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2185 2186 if (failing()) return; 2187 } 2188 } 2189 2190 // Loop transforms on the ideal graph. Range Check Elimination, 2191 // peeling, unrolling, etc. 2192 2193 // Set loop opts counter 2194 loop_opts_cnt = num_loop_opts(); 2195 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2196 { 2197 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2198 PhaseIdealLoop ideal_loop( igvn, true ); 2199 loop_opts_cnt--; 2200 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2201 if (failing()) return; 2202 } 2203 // Loop opts pass if partial peeling occurred in previous pass 2204 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2205 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2206 PhaseIdealLoop ideal_loop( igvn, false ); 2207 loop_opts_cnt--; 2208 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2209 if (failing()) return; 2210 } 2211 // Loop opts pass for loop-unrolling before CCP 2212 if(major_progress() && (loop_opts_cnt > 0)) { 2213 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2214 PhaseIdealLoop ideal_loop( igvn, false ); 2215 loop_opts_cnt--; 2216 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2217 } 2218 if (!failing()) { 2219 // Verify that last round of loop opts produced a valid graph 2220 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2221 PhaseIdealLoop::verify(igvn); 2222 } 2223 } 2224 if (failing()) return; 2225 2226 // Conditional Constant Propagation; 2227 PhaseCCP ccp( &igvn ); 2228 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2229 { 2230 TracePhase tp("ccp", &timers[_t_ccp]); 2231 ccp.do_transform(); 2232 } 2233 print_method(PHASE_CPP1, 2); 2234 2235 assert( true, "Break here to ccp.dump_old2new_map()"); 2236 2237 // Iterative Global Value Numbering, including ideal transforms 2238 { 2239 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2240 igvn = ccp; 2241 igvn.optimize(); 2242 } 2243 2244 print_method(PHASE_ITER_GVN2, 2); 2245 2246 if (failing()) return; 2247 2248 // Loop transforms on the ideal graph. Range Check Elimination, 2249 // peeling, unrolling, etc. 2250 if(loop_opts_cnt > 0) { 2251 debug_only( int cnt = 0; ); 2252 while(major_progress() && (loop_opts_cnt > 0)) { 2253 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2254 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2255 PhaseIdealLoop ideal_loop( igvn, true); 2256 loop_opts_cnt--; 2257 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2258 if (failing()) return; 2259 } 2260 } 2261 2262 { 2263 // Verify that all previous optimizations produced a valid graph 2264 // at least to this point, even if no loop optimizations were done. 2265 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2266 PhaseIdealLoop::verify(igvn); 2267 } 2268 2269 { 2270 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2271 PhaseMacroExpand mex(igvn); 2272 if (mex.expand_macro_nodes()) { 2273 assert(failing(), "must bail out w/ explicit message"); 2274 return; 2275 } 2276 } 2277 2278 DEBUG_ONLY( _modified_nodes = NULL; ) 2279 } // (End scope of igvn; run destructor if necessary for asserts.) 2280 2281 process_print_inlining(); 2282 // A method with only infinite loops has no edges entering loops from root 2283 { 2284 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2285 if (final_graph_reshaping()) { 2286 assert(failing(), "must bail out w/ explicit message"); 2287 return; 2288 } 2289 } 2290 2291 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2292 } 2293 2294 2295 //------------------------------Code_Gen--------------------------------------- 2296 // Given a graph, generate code for it 2297 void Compile::Code_Gen() { 2298 if (failing()) { 2299 return; 2300 } 2301 2302 // Perform instruction selection. You might think we could reclaim Matcher 2303 // memory PDQ, but actually the Matcher is used in generating spill code. 2304 // Internals of the Matcher (including some VectorSets) must remain live 2305 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2306 // set a bit in reclaimed memory. 2307 2308 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2309 // nodes. Mapping is only valid at the root of each matched subtree. 2310 NOT_PRODUCT( verify_graph_edges(); ) 2311 2312 Matcher matcher; 2313 _matcher = &matcher; 2314 { 2315 TracePhase tp("matcher", &timers[_t_matcher]); 2316 matcher.match(); 2317 } 2318 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2319 // nodes. Mapping is only valid at the root of each matched subtree. 2320 NOT_PRODUCT( verify_graph_edges(); ) 2321 2322 // If you have too many nodes, or if matching has failed, bail out 2323 check_node_count(0, "out of nodes matching instructions"); 2324 if (failing()) { 2325 return; 2326 } 2327 2328 // Build a proper-looking CFG 2329 PhaseCFG cfg(node_arena(), root(), matcher); 2330 _cfg = &cfg; 2331 { 2332 TracePhase tp("scheduler", &timers[_t_scheduler]); 2333 bool success = cfg.do_global_code_motion(); 2334 if (!success) { 2335 return; 2336 } 2337 2338 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2339 NOT_PRODUCT( verify_graph_edges(); ) 2340 debug_only( cfg.verify(); ) 2341 } 2342 2343 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2344 _regalloc = ®alloc; 2345 { 2346 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2347 // Perform register allocation. After Chaitin, use-def chains are 2348 // no longer accurate (at spill code) and so must be ignored. 2349 // Node->LRG->reg mappings are still accurate. 2350 _regalloc->Register_Allocate(); 2351 2352 // Bail out if the allocator builds too many nodes 2353 if (failing()) { 2354 return; 2355 } 2356 } 2357 2358 // Prior to register allocation we kept empty basic blocks in case the 2359 // the allocator needed a place to spill. After register allocation we 2360 // are not adding any new instructions. If any basic block is empty, we 2361 // can now safely remove it. 2362 { 2363 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2364 cfg.remove_empty_blocks(); 2365 if (do_freq_based_layout()) { 2366 PhaseBlockLayout layout(cfg); 2367 } else { 2368 cfg.set_loop_alignment(); 2369 } 2370 cfg.fixup_flow(); 2371 } 2372 2373 // Apply peephole optimizations 2374 if( OptoPeephole ) { 2375 TracePhase tp("peephole", &timers[_t_peephole]); 2376 PhasePeephole peep( _regalloc, cfg); 2377 peep.do_transform(); 2378 } 2379 2380 // Do late expand if CPU requires this. 2381 if (Matcher::require_postalloc_expand) { 2382 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2383 cfg.postalloc_expand(_regalloc); 2384 } 2385 2386 // Convert Nodes to instruction bits in a buffer 2387 { 2388 TraceTime tp("output", &timers[_t_output], CITime); 2389 Output(); 2390 } 2391 2392 print_method(PHASE_FINAL_CODE); 2393 2394 // He's dead, Jim. 2395 _cfg = (PhaseCFG*)0xdeadbeef; 2396 _regalloc = (PhaseChaitin*)0xdeadbeef; 2397 } 2398 2399 2400 //------------------------------dump_asm--------------------------------------- 2401 // Dump formatted assembly 2402 #ifndef PRODUCT 2403 void Compile::dump_asm(int *pcs, uint pc_limit) { 2404 bool cut_short = false; 2405 tty->print_cr("#"); 2406 tty->print("# "); _tf->dump(); tty->cr(); 2407 tty->print_cr("#"); 2408 2409 // For all blocks 2410 int pc = 0x0; // Program counter 2411 char starts_bundle = ' '; 2412 _regalloc->dump_frame(); 2413 2414 Node *n = NULL; 2415 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2416 if (VMThread::should_terminate()) { 2417 cut_short = true; 2418 break; 2419 } 2420 Block* block = _cfg->get_block(i); 2421 if (block->is_connector() && !Verbose) { 2422 continue; 2423 } 2424 n = block->head(); 2425 if (pcs && n->_idx < pc_limit) { 2426 tty->print("%3.3x ", pcs[n->_idx]); 2427 } else { 2428 tty->print(" "); 2429 } 2430 block->dump_head(_cfg); 2431 if (block->is_connector()) { 2432 tty->print_cr(" # Empty connector block"); 2433 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2434 tty->print_cr(" # Block is sole successor of call"); 2435 } 2436 2437 // For all instructions 2438 Node *delay = NULL; 2439 for (uint j = 0; j < block->number_of_nodes(); j++) { 2440 if (VMThread::should_terminate()) { 2441 cut_short = true; 2442 break; 2443 } 2444 n = block->get_node(j); 2445 if (valid_bundle_info(n)) { 2446 Bundle* bundle = node_bundling(n); 2447 if (bundle->used_in_unconditional_delay()) { 2448 delay = n; 2449 continue; 2450 } 2451 if (bundle->starts_bundle()) { 2452 starts_bundle = '+'; 2453 } 2454 } 2455 2456 if (WizardMode) { 2457 n->dump(); 2458 } 2459 2460 if( !n->is_Region() && // Dont print in the Assembly 2461 !n->is_Phi() && // a few noisely useless nodes 2462 !n->is_Proj() && 2463 !n->is_MachTemp() && 2464 !n->is_SafePointScalarObject() && 2465 !n->is_Catch() && // Would be nice to print exception table targets 2466 !n->is_MergeMem() && // Not very interesting 2467 !n->is_top() && // Debug info table constants 2468 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2469 ) { 2470 if (pcs && n->_idx < pc_limit) 2471 tty->print("%3.3x", pcs[n->_idx]); 2472 else 2473 tty->print(" "); 2474 tty->print(" %c ", starts_bundle); 2475 starts_bundle = ' '; 2476 tty->print("\t"); 2477 n->format(_regalloc, tty); 2478 tty->cr(); 2479 } 2480 2481 // If we have an instruction with a delay slot, and have seen a delay, 2482 // then back up and print it 2483 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2484 assert(delay != NULL, "no unconditional delay instruction"); 2485 if (WizardMode) delay->dump(); 2486 2487 if (node_bundling(delay)->starts_bundle()) 2488 starts_bundle = '+'; 2489 if (pcs && n->_idx < pc_limit) 2490 tty->print("%3.3x", pcs[n->_idx]); 2491 else 2492 tty->print(" "); 2493 tty->print(" %c ", starts_bundle); 2494 starts_bundle = ' '; 2495 tty->print("\t"); 2496 delay->format(_regalloc, tty); 2497 tty->cr(); 2498 delay = NULL; 2499 } 2500 2501 // Dump the exception table as well 2502 if( n->is_Catch() && (Verbose || WizardMode) ) { 2503 // Print the exception table for this offset 2504 _handler_table.print_subtable_for(pc); 2505 } 2506 } 2507 2508 if (pcs && n->_idx < pc_limit) 2509 tty->print_cr("%3.3x", pcs[n->_idx]); 2510 else 2511 tty->cr(); 2512 2513 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2514 2515 } // End of per-block dump 2516 tty->cr(); 2517 2518 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2519 } 2520 #endif 2521 2522 //------------------------------Final_Reshape_Counts--------------------------- 2523 // This class defines counters to help identify when a method 2524 // may/must be executed using hardware with only 24-bit precision. 2525 struct Final_Reshape_Counts : public StackObj { 2526 int _call_count; // count non-inlined 'common' calls 2527 int _float_count; // count float ops requiring 24-bit precision 2528 int _double_count; // count double ops requiring more precision 2529 int _java_call_count; // count non-inlined 'java' calls 2530 int _inner_loop_count; // count loops which need alignment 2531 VectorSet _visited; // Visitation flags 2532 Node_List _tests; // Set of IfNodes & PCTableNodes 2533 2534 Final_Reshape_Counts() : 2535 _call_count(0), _float_count(0), _double_count(0), 2536 _java_call_count(0), _inner_loop_count(0), 2537 _visited( Thread::current()->resource_area() ) { } 2538 2539 void inc_call_count () { _call_count ++; } 2540 void inc_float_count () { _float_count ++; } 2541 void inc_double_count() { _double_count++; } 2542 void inc_java_call_count() { _java_call_count++; } 2543 void inc_inner_loop_count() { _inner_loop_count++; } 2544 2545 int get_call_count () const { return _call_count ; } 2546 int get_float_count () const { return _float_count ; } 2547 int get_double_count() const { return _double_count; } 2548 int get_java_call_count() const { return _java_call_count; } 2549 int get_inner_loop_count() const { return _inner_loop_count; } 2550 }; 2551 2552 #ifdef ASSERT 2553 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2554 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2555 // Make sure the offset goes inside the instance layout. 2556 return k->contains_field_offset(tp->offset()); 2557 // Note that OffsetBot and OffsetTop are very negative. 2558 } 2559 #endif 2560 2561 // Eliminate trivially redundant StoreCMs and accumulate their 2562 // precedence edges. 2563 void Compile::eliminate_redundant_card_marks(Node* n) { 2564 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2565 if (n->in(MemNode::Address)->outcnt() > 1) { 2566 // There are multiple users of the same address so it might be 2567 // possible to eliminate some of the StoreCMs 2568 Node* mem = n->in(MemNode::Memory); 2569 Node* adr = n->in(MemNode::Address); 2570 Node* val = n->in(MemNode::ValueIn); 2571 Node* prev = n; 2572 bool done = false; 2573 // Walk the chain of StoreCMs eliminating ones that match. As 2574 // long as it's a chain of single users then the optimization is 2575 // safe. Eliminating partially redundant StoreCMs would require 2576 // cloning copies down the other paths. 2577 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2578 if (adr == mem->in(MemNode::Address) && 2579 val == mem->in(MemNode::ValueIn)) { 2580 // redundant StoreCM 2581 if (mem->req() > MemNode::OopStore) { 2582 // Hasn't been processed by this code yet. 2583 n->add_prec(mem->in(MemNode::OopStore)); 2584 } else { 2585 // Already converted to precedence edge 2586 for (uint i = mem->req(); i < mem->len(); i++) { 2587 // Accumulate any precedence edges 2588 if (mem->in(i) != NULL) { 2589 n->add_prec(mem->in(i)); 2590 } 2591 } 2592 // Everything above this point has been processed. 2593 done = true; 2594 } 2595 // Eliminate the previous StoreCM 2596 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2597 assert(mem->outcnt() == 0, "should be dead"); 2598 mem->disconnect_inputs(NULL, this); 2599 } else { 2600 prev = mem; 2601 } 2602 mem = prev->in(MemNode::Memory); 2603 } 2604 } 2605 } 2606 2607 //------------------------------final_graph_reshaping_impl---------------------- 2608 // Implement items 1-5 from final_graph_reshaping below. 2609 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2610 2611 if ( n->outcnt() == 0 ) return; // dead node 2612 uint nop = n->Opcode(); 2613 2614 // Check for 2-input instruction with "last use" on right input. 2615 // Swap to left input. Implements item (2). 2616 if( n->req() == 3 && // two-input instruction 2617 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2618 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2619 n->in(2)->outcnt() == 1 &&// right use IS a last use 2620 !n->in(2)->is_Con() ) { // right use is not a constant 2621 // Check for commutative opcode 2622 switch( nop ) { 2623 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2624 case Op_MaxI: case Op_MinI: 2625 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2626 case Op_AndL: case Op_XorL: case Op_OrL: 2627 case Op_AndI: case Op_XorI: case Op_OrI: { 2628 // Move "last use" input to left by swapping inputs 2629 n->swap_edges(1, 2); 2630 break; 2631 } 2632 default: 2633 break; 2634 } 2635 } 2636 2637 #ifdef ASSERT 2638 if( n->is_Mem() ) { 2639 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2640 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2641 // oop will be recorded in oop map if load crosses safepoint 2642 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2643 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2644 "raw memory operations should have control edge"); 2645 } 2646 #endif 2647 // Count FPU ops and common calls, implements item (3) 2648 switch( nop ) { 2649 // Count all float operations that may use FPU 2650 case Op_AddF: 2651 case Op_SubF: 2652 case Op_MulF: 2653 case Op_DivF: 2654 case Op_NegF: 2655 case Op_ModF: 2656 case Op_ConvI2F: 2657 case Op_ConF: 2658 case Op_CmpF: 2659 case Op_CmpF3: 2660 // case Op_ConvL2F: // longs are split into 32-bit halves 2661 frc.inc_float_count(); 2662 break; 2663 2664 case Op_ConvF2D: 2665 case Op_ConvD2F: 2666 frc.inc_float_count(); 2667 frc.inc_double_count(); 2668 break; 2669 2670 // Count all double operations that may use FPU 2671 case Op_AddD: 2672 case Op_SubD: 2673 case Op_MulD: 2674 case Op_DivD: 2675 case Op_NegD: 2676 case Op_ModD: 2677 case Op_ConvI2D: 2678 case Op_ConvD2I: 2679 // case Op_ConvL2D: // handled by leaf call 2680 // case Op_ConvD2L: // handled by leaf call 2681 case Op_ConD: 2682 case Op_CmpD: 2683 case Op_CmpD3: 2684 frc.inc_double_count(); 2685 break; 2686 case Op_Opaque1: // Remove Opaque Nodes before matching 2687 case Op_Opaque2: // Remove Opaque Nodes before matching 2688 case Op_Opaque3: 2689 n->subsume_by(n->in(1), this); 2690 break; 2691 case Op_CallStaticJava: 2692 case Op_CallJava: 2693 case Op_CallDynamicJava: 2694 frc.inc_java_call_count(); // Count java call site; 2695 case Op_CallRuntime: 2696 case Op_CallLeaf: 2697 case Op_CallLeafNoFP: { 2698 assert( n->is_Call(), "" ); 2699 CallNode *call = n->as_Call(); 2700 // Count call sites where the FP mode bit would have to be flipped. 2701 // Do not count uncommon runtime calls: 2702 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2703 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2704 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2705 frc.inc_call_count(); // Count the call site 2706 } else { // See if uncommon argument is shared 2707 Node *n = call->in(TypeFunc::Parms); 2708 int nop = n->Opcode(); 2709 // Clone shared simple arguments to uncommon calls, item (1). 2710 if( n->outcnt() > 1 && 2711 !n->is_Proj() && 2712 nop != Op_CreateEx && 2713 nop != Op_CheckCastPP && 2714 nop != Op_DecodeN && 2715 nop != Op_DecodeNKlass && 2716 !n->is_Mem() ) { 2717 Node *x = n->clone(); 2718 call->set_req( TypeFunc::Parms, x ); 2719 } 2720 } 2721 break; 2722 } 2723 2724 case Op_StoreD: 2725 case Op_LoadD: 2726 case Op_LoadD_unaligned: 2727 frc.inc_double_count(); 2728 goto handle_mem; 2729 case Op_StoreF: 2730 case Op_LoadF: 2731 frc.inc_float_count(); 2732 goto handle_mem; 2733 2734 case Op_StoreCM: 2735 { 2736 // Convert OopStore dependence into precedence edge 2737 Node* prec = n->in(MemNode::OopStore); 2738 n->del_req(MemNode::OopStore); 2739 n->add_prec(prec); 2740 eliminate_redundant_card_marks(n); 2741 } 2742 2743 // fall through 2744 2745 case Op_StoreB: 2746 case Op_StoreC: 2747 case Op_StorePConditional: 2748 case Op_StoreI: 2749 case Op_StoreL: 2750 case Op_StoreIConditional: 2751 case Op_StoreLConditional: 2752 case Op_CompareAndSwapI: 2753 case Op_CompareAndSwapL: 2754 case Op_CompareAndSwapP: 2755 case Op_CompareAndSwapN: 2756 case Op_GetAndAddI: 2757 case Op_GetAndAddL: 2758 case Op_GetAndSetI: 2759 case Op_GetAndSetL: 2760 case Op_GetAndSetP: 2761 case Op_GetAndSetN: 2762 case Op_StoreP: 2763 case Op_StoreN: 2764 case Op_StoreNKlass: 2765 case Op_LoadB: 2766 case Op_LoadUB: 2767 case Op_LoadUS: 2768 case Op_LoadI: 2769 case Op_LoadKlass: 2770 case Op_LoadNKlass: 2771 case Op_LoadL: 2772 case Op_LoadL_unaligned: 2773 case Op_LoadPLocked: 2774 case Op_LoadP: 2775 case Op_LoadN: 2776 case Op_LoadRange: 2777 case Op_LoadS: { 2778 handle_mem: 2779 #ifdef ASSERT 2780 if( VerifyOptoOopOffsets ) { 2781 assert( n->is_Mem(), "" ); 2782 MemNode *mem = (MemNode*)n; 2783 // Check to see if address types have grounded out somehow. 2784 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2785 assert( !tp || oop_offset_is_sane(tp), "" ); 2786 } 2787 #endif 2788 break; 2789 } 2790 2791 case Op_AddP: { // Assert sane base pointers 2792 Node *addp = n->in(AddPNode::Address); 2793 assert( !addp->is_AddP() || 2794 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2795 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2796 "Base pointers must match" ); 2797 #ifdef _LP64 2798 if ((UseCompressedOops || UseCompressedClassPointers) && 2799 addp->Opcode() == Op_ConP && 2800 addp == n->in(AddPNode::Base) && 2801 n->in(AddPNode::Offset)->is_Con()) { 2802 // Use addressing with narrow klass to load with offset on x86. 2803 // On sparc loading 32-bits constant and decoding it have less 2804 // instructions (4) then load 64-bits constant (7). 2805 // Do this transformation here since IGVN will convert ConN back to ConP. 2806 const Type* t = addp->bottom_type(); 2807 if (t->isa_oopptr() || t->isa_klassptr()) { 2808 Node* nn = NULL; 2809 2810 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2811 2812 // Look for existing ConN node of the same exact type. 2813 Node* r = root(); 2814 uint cnt = r->outcnt(); 2815 for (uint i = 0; i < cnt; i++) { 2816 Node* m = r->raw_out(i); 2817 if (m!= NULL && m->Opcode() == op && 2818 m->bottom_type()->make_ptr() == t) { 2819 nn = m; 2820 break; 2821 } 2822 } 2823 if (nn != NULL) { 2824 // Decode a narrow oop to match address 2825 // [R12 + narrow_oop_reg<<3 + offset] 2826 if (t->isa_oopptr()) { 2827 nn = new DecodeNNode(nn, t); 2828 } else { 2829 nn = new DecodeNKlassNode(nn, t); 2830 } 2831 n->set_req(AddPNode::Base, nn); 2832 n->set_req(AddPNode::Address, nn); 2833 if (addp->outcnt() == 0) { 2834 addp->disconnect_inputs(NULL, this); 2835 } 2836 } 2837 } 2838 } 2839 #endif 2840 break; 2841 } 2842 2843 case Op_CastPP: { 2844 // Remove CastPP nodes to gain more freedom during scheduling but 2845 // keep the dependency they encode as control or precedence edges 2846 // (if control is set already) on memory operations. Some CastPP 2847 // nodes don't have a control (don't carry a dependency): skip 2848 // those. 2849 if (n->in(0) != NULL) { 2850 ResourceMark rm; 2851 Unique_Node_List wq; 2852 wq.push(n); 2853 for (uint next = 0; next < wq.size(); ++next) { 2854 Node *m = wq.at(next); 2855 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 2856 Node* use = m->fast_out(i); 2857 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 2858 use->ensure_control_or_add_prec(n->in(0)); 2859 } else if (use->in(0) == NULL) { 2860 switch(use->Opcode()) { 2861 case Op_AddP: 2862 case Op_DecodeN: 2863 case Op_DecodeNKlass: 2864 case Op_CheckCastPP: 2865 case Op_CastPP: 2866 wq.push(use); 2867 break; 2868 } 2869 } 2870 } 2871 } 2872 } 2873 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 2874 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2875 Node* in1 = n->in(1); 2876 const Type* t = n->bottom_type(); 2877 Node* new_in1 = in1->clone(); 2878 new_in1->as_DecodeN()->set_type(t); 2879 2880 if (!Matcher::narrow_oop_use_complex_address()) { 2881 // 2882 // x86, ARM and friends can handle 2 adds in addressing mode 2883 // and Matcher can fold a DecodeN node into address by using 2884 // a narrow oop directly and do implicit NULL check in address: 2885 // 2886 // [R12 + narrow_oop_reg<<3 + offset] 2887 // NullCheck narrow_oop_reg 2888 // 2889 // On other platforms (Sparc) we have to keep new DecodeN node and 2890 // use it to do implicit NULL check in address: 2891 // 2892 // decode_not_null narrow_oop_reg, base_reg 2893 // [base_reg + offset] 2894 // NullCheck base_reg 2895 // 2896 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2897 // to keep the information to which NULL check the new DecodeN node 2898 // corresponds to use it as value in implicit_null_check(). 2899 // 2900 new_in1->set_req(0, n->in(0)); 2901 } 2902 2903 n->subsume_by(new_in1, this); 2904 if (in1->outcnt() == 0) { 2905 in1->disconnect_inputs(NULL, this); 2906 } 2907 } else { 2908 n->subsume_by(n->in(1), this); 2909 if (n->outcnt() == 0) { 2910 n->disconnect_inputs(NULL, this); 2911 } 2912 } 2913 break; 2914 } 2915 #ifdef _LP64 2916 case Op_CmpP: 2917 // Do this transformation here to preserve CmpPNode::sub() and 2918 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2919 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2920 Node* in1 = n->in(1); 2921 Node* in2 = n->in(2); 2922 if (!in1->is_DecodeNarrowPtr()) { 2923 in2 = in1; 2924 in1 = n->in(2); 2925 } 2926 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2927 2928 Node* new_in2 = NULL; 2929 if (in2->is_DecodeNarrowPtr()) { 2930 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2931 new_in2 = in2->in(1); 2932 } else if (in2->Opcode() == Op_ConP) { 2933 const Type* t = in2->bottom_type(); 2934 if (t == TypePtr::NULL_PTR) { 2935 assert(in1->is_DecodeN(), "compare klass to null?"); 2936 // Don't convert CmpP null check into CmpN if compressed 2937 // oops implicit null check is not generated. 2938 // This will allow to generate normal oop implicit null check. 2939 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2940 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 2941 // 2942 // This transformation together with CastPP transformation above 2943 // will generated code for implicit NULL checks for compressed oops. 2944 // 2945 // The original code after Optimize() 2946 // 2947 // LoadN memory, narrow_oop_reg 2948 // decode narrow_oop_reg, base_reg 2949 // CmpP base_reg, NULL 2950 // CastPP base_reg // NotNull 2951 // Load [base_reg + offset], val_reg 2952 // 2953 // after these transformations will be 2954 // 2955 // LoadN memory, narrow_oop_reg 2956 // CmpN narrow_oop_reg, NULL 2957 // decode_not_null narrow_oop_reg, base_reg 2958 // Load [base_reg + offset], val_reg 2959 // 2960 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2961 // since narrow oops can be used in debug info now (see the code in 2962 // final_graph_reshaping_walk()). 2963 // 2964 // At the end the code will be matched to 2965 // on x86: 2966 // 2967 // Load_narrow_oop memory, narrow_oop_reg 2968 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2969 // NullCheck narrow_oop_reg 2970 // 2971 // and on sparc: 2972 // 2973 // Load_narrow_oop memory, narrow_oop_reg 2974 // decode_not_null narrow_oop_reg, base_reg 2975 // Load [base_reg + offset], val_reg 2976 // NullCheck base_reg 2977 // 2978 } else if (t->isa_oopptr()) { 2979 new_in2 = ConNode::make(t->make_narrowoop()); 2980 } else if (t->isa_klassptr()) { 2981 new_in2 = ConNode::make(t->make_narrowklass()); 2982 } 2983 } 2984 if (new_in2 != NULL) { 2985 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 2986 n->subsume_by(cmpN, this); 2987 if (in1->outcnt() == 0) { 2988 in1->disconnect_inputs(NULL, this); 2989 } 2990 if (in2->outcnt() == 0) { 2991 in2->disconnect_inputs(NULL, this); 2992 } 2993 } 2994 } 2995 break; 2996 2997 case Op_DecodeN: 2998 case Op_DecodeNKlass: 2999 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3000 // DecodeN could be pinned when it can't be fold into 3001 // an address expression, see the code for Op_CastPP above. 3002 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3003 break; 3004 3005 case Op_EncodeP: 3006 case Op_EncodePKlass: { 3007 Node* in1 = n->in(1); 3008 if (in1->is_DecodeNarrowPtr()) { 3009 n->subsume_by(in1->in(1), this); 3010 } else if (in1->Opcode() == Op_ConP) { 3011 const Type* t = in1->bottom_type(); 3012 if (t == TypePtr::NULL_PTR) { 3013 assert(t->isa_oopptr(), "null klass?"); 3014 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3015 } else if (t->isa_oopptr()) { 3016 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3017 } else if (t->isa_klassptr()) { 3018 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3019 } 3020 } 3021 if (in1->outcnt() == 0) { 3022 in1->disconnect_inputs(NULL, this); 3023 } 3024 break; 3025 } 3026 3027 case Op_Proj: { 3028 if (OptimizeStringConcat) { 3029 ProjNode* p = n->as_Proj(); 3030 if (p->_is_io_use) { 3031 // Separate projections were used for the exception path which 3032 // are normally removed by a late inline. If it wasn't inlined 3033 // then they will hang around and should just be replaced with 3034 // the original one. 3035 Node* proj = NULL; 3036 // Replace with just one 3037 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3038 Node *use = i.get(); 3039 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3040 proj = use; 3041 break; 3042 } 3043 } 3044 assert(proj != NULL, "must be found"); 3045 p->subsume_by(proj, this); 3046 } 3047 } 3048 break; 3049 } 3050 3051 case Op_Phi: 3052 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3053 // The EncodeP optimization may create Phi with the same edges 3054 // for all paths. It is not handled well by Register Allocator. 3055 Node* unique_in = n->in(1); 3056 assert(unique_in != NULL, ""); 3057 uint cnt = n->req(); 3058 for (uint i = 2; i < cnt; i++) { 3059 Node* m = n->in(i); 3060 assert(m != NULL, ""); 3061 if (unique_in != m) 3062 unique_in = NULL; 3063 } 3064 if (unique_in != NULL) { 3065 n->subsume_by(unique_in, this); 3066 } 3067 } 3068 break; 3069 3070 #endif 3071 3072 case Op_ModI: 3073 if (UseDivMod) { 3074 // Check if a%b and a/b both exist 3075 Node* d = n->find_similar(Op_DivI); 3076 if (d) { 3077 // Replace them with a fused divmod if supported 3078 if (Matcher::has_match_rule(Op_DivModI)) { 3079 DivModINode* divmod = DivModINode::make(n); 3080 d->subsume_by(divmod->div_proj(), this); 3081 n->subsume_by(divmod->mod_proj(), this); 3082 } else { 3083 // replace a%b with a-((a/b)*b) 3084 Node* mult = new MulINode(d, d->in(2)); 3085 Node* sub = new SubINode(d->in(1), mult); 3086 n->subsume_by(sub, this); 3087 } 3088 } 3089 } 3090 break; 3091 3092 case Op_ModL: 3093 if (UseDivMod) { 3094 // Check if a%b and a/b both exist 3095 Node* d = n->find_similar(Op_DivL); 3096 if (d) { 3097 // Replace them with a fused divmod if supported 3098 if (Matcher::has_match_rule(Op_DivModL)) { 3099 DivModLNode* divmod = DivModLNode::make(n); 3100 d->subsume_by(divmod->div_proj(), this); 3101 n->subsume_by(divmod->mod_proj(), this); 3102 } else { 3103 // replace a%b with a-((a/b)*b) 3104 Node* mult = new MulLNode(d, d->in(2)); 3105 Node* sub = new SubLNode(d->in(1), mult); 3106 n->subsume_by(sub, this); 3107 } 3108 } 3109 } 3110 break; 3111 3112 case Op_LoadVector: 3113 case Op_StoreVector: 3114 break; 3115 3116 case Op_AddReductionVI: 3117 case Op_AddReductionVL: 3118 case Op_AddReductionVF: 3119 case Op_AddReductionVD: 3120 case Op_MulReductionVI: 3121 case Op_MulReductionVL: 3122 case Op_MulReductionVF: 3123 case Op_MulReductionVD: 3124 break; 3125 3126 case Op_PackB: 3127 case Op_PackS: 3128 case Op_PackI: 3129 case Op_PackF: 3130 case Op_PackL: 3131 case Op_PackD: 3132 if (n->req()-1 > 2) { 3133 // Replace many operand PackNodes with a binary tree for matching 3134 PackNode* p = (PackNode*) n; 3135 Node* btp = p->binary_tree_pack(1, n->req()); 3136 n->subsume_by(btp, this); 3137 } 3138 break; 3139 case Op_Loop: 3140 case Op_CountedLoop: 3141 if (n->as_Loop()->is_inner_loop()) { 3142 frc.inc_inner_loop_count(); 3143 } 3144 break; 3145 case Op_LShiftI: 3146 case Op_RShiftI: 3147 case Op_URShiftI: 3148 case Op_LShiftL: 3149 case Op_RShiftL: 3150 case Op_URShiftL: 3151 if (Matcher::need_masked_shift_count) { 3152 // The cpu's shift instructions don't restrict the count to the 3153 // lower 5/6 bits. We need to do the masking ourselves. 3154 Node* in2 = n->in(2); 3155 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3156 const TypeInt* t = in2->find_int_type(); 3157 if (t != NULL && t->is_con()) { 3158 juint shift = t->get_con(); 3159 if (shift > mask) { // Unsigned cmp 3160 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3161 } 3162 } else { 3163 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3164 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3165 n->set_req(2, shift); 3166 } 3167 } 3168 if (in2->outcnt() == 0) { // Remove dead node 3169 in2->disconnect_inputs(NULL, this); 3170 } 3171 } 3172 break; 3173 case Op_MemBarStoreStore: 3174 case Op_MemBarRelease: 3175 // Break the link with AllocateNode: it is no longer useful and 3176 // confuses register allocation. 3177 if (n->req() > MemBarNode::Precedent) { 3178 n->set_req(MemBarNode::Precedent, top()); 3179 } 3180 break; 3181 default: 3182 assert( !n->is_Call(), "" ); 3183 assert( !n->is_Mem(), "" ); 3184 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3185 break; 3186 } 3187 3188 // Collect CFG split points 3189 if (n->is_MultiBranch()) 3190 frc._tests.push(n); 3191 } 3192 3193 //------------------------------final_graph_reshaping_walk--------------------- 3194 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3195 // requires that the walk visits a node's inputs before visiting the node. 3196 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3197 ResourceArea *area = Thread::current()->resource_area(); 3198 Unique_Node_List sfpt(area); 3199 3200 frc._visited.set(root->_idx); // first, mark node as visited 3201 uint cnt = root->req(); 3202 Node *n = root; 3203 uint i = 0; 3204 while (true) { 3205 if (i < cnt) { 3206 // Place all non-visited non-null inputs onto stack 3207 Node* m = n->in(i); 3208 ++i; 3209 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3210 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3211 // compute worst case interpreter size in case of a deoptimization 3212 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3213 3214 sfpt.push(m); 3215 } 3216 cnt = m->req(); 3217 nstack.push(n, i); // put on stack parent and next input's index 3218 n = m; 3219 i = 0; 3220 } 3221 } else { 3222 // Now do post-visit work 3223 final_graph_reshaping_impl( n, frc ); 3224 if (nstack.is_empty()) 3225 break; // finished 3226 n = nstack.node(); // Get node from stack 3227 cnt = n->req(); 3228 i = nstack.index(); 3229 nstack.pop(); // Shift to the next node on stack 3230 } 3231 } 3232 3233 // Skip next transformation if compressed oops are not used. 3234 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3235 (!UseCompressedOops && !UseCompressedClassPointers)) 3236 return; 3237 3238 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3239 // It could be done for an uncommon traps or any safepoints/calls 3240 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3241 while (sfpt.size() > 0) { 3242 n = sfpt.pop(); 3243 JVMState *jvms = n->as_SafePoint()->jvms(); 3244 assert(jvms != NULL, "sanity"); 3245 int start = jvms->debug_start(); 3246 int end = n->req(); 3247 bool is_uncommon = (n->is_CallStaticJava() && 3248 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3249 for (int j = start; j < end; j++) { 3250 Node* in = n->in(j); 3251 if (in->is_DecodeNarrowPtr()) { 3252 bool safe_to_skip = true; 3253 if (!is_uncommon ) { 3254 // Is it safe to skip? 3255 for (uint i = 0; i < in->outcnt(); i++) { 3256 Node* u = in->raw_out(i); 3257 if (!u->is_SafePoint() || 3258 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3259 safe_to_skip = false; 3260 } 3261 } 3262 } 3263 if (safe_to_skip) { 3264 n->set_req(j, in->in(1)); 3265 } 3266 if (in->outcnt() == 0) { 3267 in->disconnect_inputs(NULL, this); 3268 } 3269 } 3270 } 3271 } 3272 } 3273 3274 //------------------------------final_graph_reshaping-------------------------- 3275 // Final Graph Reshaping. 3276 // 3277 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3278 // and not commoned up and forced early. Must come after regular 3279 // optimizations to avoid GVN undoing the cloning. Clone constant 3280 // inputs to Loop Phis; these will be split by the allocator anyways. 3281 // Remove Opaque nodes. 3282 // (2) Move last-uses by commutative operations to the left input to encourage 3283 // Intel update-in-place two-address operations and better register usage 3284 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3285 // calls canonicalizing them back. 3286 // (3) Count the number of double-precision FP ops, single-precision FP ops 3287 // and call sites. On Intel, we can get correct rounding either by 3288 // forcing singles to memory (requires extra stores and loads after each 3289 // FP bytecode) or we can set a rounding mode bit (requires setting and 3290 // clearing the mode bit around call sites). The mode bit is only used 3291 // if the relative frequency of single FP ops to calls is low enough. 3292 // This is a key transform for SPEC mpeg_audio. 3293 // (4) Detect infinite loops; blobs of code reachable from above but not 3294 // below. Several of the Code_Gen algorithms fail on such code shapes, 3295 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3296 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3297 // Detection is by looking for IfNodes where only 1 projection is 3298 // reachable from below or CatchNodes missing some targets. 3299 // (5) Assert for insane oop offsets in debug mode. 3300 3301 bool Compile::final_graph_reshaping() { 3302 // an infinite loop may have been eliminated by the optimizer, 3303 // in which case the graph will be empty. 3304 if (root()->req() == 1) { 3305 record_method_not_compilable("trivial infinite loop"); 3306 return true; 3307 } 3308 3309 // Expensive nodes have their control input set to prevent the GVN 3310 // from freely commoning them. There's no GVN beyond this point so 3311 // no need to keep the control input. We want the expensive nodes to 3312 // be freely moved to the least frequent code path by gcm. 3313 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3314 for (int i = 0; i < expensive_count(); i++) { 3315 _expensive_nodes->at(i)->set_req(0, NULL); 3316 } 3317 3318 Final_Reshape_Counts frc; 3319 3320 // Visit everybody reachable! 3321 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3322 Node_Stack nstack(live_nodes() >> 1); 3323 final_graph_reshaping_walk(nstack, root(), frc); 3324 3325 // Check for unreachable (from below) code (i.e., infinite loops). 3326 for( uint i = 0; i < frc._tests.size(); i++ ) { 3327 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3328 // Get number of CFG targets. 3329 // Note that PCTables include exception targets after calls. 3330 uint required_outcnt = n->required_outcnt(); 3331 if (n->outcnt() != required_outcnt) { 3332 // Check for a few special cases. Rethrow Nodes never take the 3333 // 'fall-thru' path, so expected kids is 1 less. 3334 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3335 if (n->in(0)->in(0)->is_Call()) { 3336 CallNode *call = n->in(0)->in(0)->as_Call(); 3337 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3338 required_outcnt--; // Rethrow always has 1 less kid 3339 } else if (call->req() > TypeFunc::Parms && 3340 call->is_CallDynamicJava()) { 3341 // Check for null receiver. In such case, the optimizer has 3342 // detected that the virtual call will always result in a null 3343 // pointer exception. The fall-through projection of this CatchNode 3344 // will not be populated. 3345 Node *arg0 = call->in(TypeFunc::Parms); 3346 if (arg0->is_Type() && 3347 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3348 required_outcnt--; 3349 } 3350 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3351 call->req() > TypeFunc::Parms+1 && 3352 call->is_CallStaticJava()) { 3353 // Check for negative array length. In such case, the optimizer has 3354 // detected that the allocation attempt will always result in an 3355 // exception. There is no fall-through projection of this CatchNode . 3356 Node *arg1 = call->in(TypeFunc::Parms+1); 3357 if (arg1->is_Type() && 3358 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3359 required_outcnt--; 3360 } 3361 } 3362 } 3363 } 3364 // Recheck with a better notion of 'required_outcnt' 3365 if (n->outcnt() != required_outcnt) { 3366 record_method_not_compilable("malformed control flow"); 3367 return true; // Not all targets reachable! 3368 } 3369 } 3370 // Check that I actually visited all kids. Unreached kids 3371 // must be infinite loops. 3372 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3373 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3374 record_method_not_compilable("infinite loop"); 3375 return true; // Found unvisited kid; must be unreach 3376 } 3377 } 3378 3379 // If original bytecodes contained a mixture of floats and doubles 3380 // check if the optimizer has made it homogenous, item (3). 3381 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3382 frc.get_float_count() > 32 && 3383 frc.get_double_count() == 0 && 3384 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3385 set_24_bit_selection_and_mode( false, true ); 3386 } 3387 3388 set_java_calls(frc.get_java_call_count()); 3389 set_inner_loops(frc.get_inner_loop_count()); 3390 3391 // No infinite loops, no reason to bail out. 3392 return false; 3393 } 3394 3395 //-----------------------------too_many_traps---------------------------------- 3396 // Report if there are too many traps at the current method and bci. 3397 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3398 bool Compile::too_many_traps(ciMethod* method, 3399 int bci, 3400 Deoptimization::DeoptReason reason) { 3401 ciMethodData* md = method->method_data(); 3402 if (md->is_empty()) { 3403 // Assume the trap has not occurred, or that it occurred only 3404 // because of a transient condition during start-up in the interpreter. 3405 return false; 3406 } 3407 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3408 if (md->has_trap_at(bci, m, reason) != 0) { 3409 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3410 // Also, if there are multiple reasons, or if there is no per-BCI record, 3411 // assume the worst. 3412 if (log()) 3413 log()->elem("observe trap='%s' count='%d'", 3414 Deoptimization::trap_reason_name(reason), 3415 md->trap_count(reason)); 3416 return true; 3417 } else { 3418 // Ignore method/bci and see if there have been too many globally. 3419 return too_many_traps(reason, md); 3420 } 3421 } 3422 3423 // Less-accurate variant which does not require a method and bci. 3424 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3425 ciMethodData* logmd) { 3426 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3427 // Too many traps globally. 3428 // Note that we use cumulative trap_count, not just md->trap_count. 3429 if (log()) { 3430 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3431 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3432 Deoptimization::trap_reason_name(reason), 3433 mcount, trap_count(reason)); 3434 } 3435 return true; 3436 } else { 3437 // The coast is clear. 3438 return false; 3439 } 3440 } 3441 3442 //--------------------------too_many_recompiles-------------------------------- 3443 // Report if there are too many recompiles at the current method and bci. 3444 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3445 // Is not eager to return true, since this will cause the compiler to use 3446 // Action_none for a trap point, to avoid too many recompilations. 3447 bool Compile::too_many_recompiles(ciMethod* method, 3448 int bci, 3449 Deoptimization::DeoptReason reason) { 3450 ciMethodData* md = method->method_data(); 3451 if (md->is_empty()) { 3452 // Assume the trap has not occurred, or that it occurred only 3453 // because of a transient condition during start-up in the interpreter. 3454 return false; 3455 } 3456 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3457 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3458 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3459 Deoptimization::DeoptReason per_bc_reason 3460 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3461 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3462 if ((per_bc_reason == Deoptimization::Reason_none 3463 || md->has_trap_at(bci, m, reason) != 0) 3464 // The trap frequency measure we care about is the recompile count: 3465 && md->trap_recompiled_at(bci, m) 3466 && md->overflow_recompile_count() >= bc_cutoff) { 3467 // Do not emit a trap here if it has already caused recompilations. 3468 // Also, if there are multiple reasons, or if there is no per-BCI record, 3469 // assume the worst. 3470 if (log()) 3471 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3472 Deoptimization::trap_reason_name(reason), 3473 md->trap_count(reason), 3474 md->overflow_recompile_count()); 3475 return true; 3476 } else if (trap_count(reason) != 0 3477 && decompile_count() >= m_cutoff) { 3478 // Too many recompiles globally, and we have seen this sort of trap. 3479 // Use cumulative decompile_count, not just md->decompile_count. 3480 if (log()) 3481 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3482 Deoptimization::trap_reason_name(reason), 3483 md->trap_count(reason), trap_count(reason), 3484 md->decompile_count(), decompile_count()); 3485 return true; 3486 } else { 3487 // The coast is clear. 3488 return false; 3489 } 3490 } 3491 3492 // Compute when not to trap. Used by matching trap based nodes and 3493 // NullCheck optimization. 3494 void Compile::set_allowed_deopt_reasons() { 3495 _allowed_reasons = 0; 3496 if (is_method_compilation()) { 3497 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3498 assert(rs < BitsPerInt, "recode bit map"); 3499 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3500 _allowed_reasons |= nth_bit(rs); 3501 } 3502 } 3503 } 3504 } 3505 3506 #ifndef PRODUCT 3507 //------------------------------verify_graph_edges--------------------------- 3508 // Walk the Graph and verify that there is a one-to-one correspondence 3509 // between Use-Def edges and Def-Use edges in the graph. 3510 void Compile::verify_graph_edges(bool no_dead_code) { 3511 if (VerifyGraphEdges) { 3512 ResourceArea *area = Thread::current()->resource_area(); 3513 Unique_Node_List visited(area); 3514 // Call recursive graph walk to check edges 3515 _root->verify_edges(visited); 3516 if (no_dead_code) { 3517 // Now make sure that no visited node is used by an unvisited node. 3518 bool dead_nodes = false; 3519 Unique_Node_List checked(area); 3520 while (visited.size() > 0) { 3521 Node* n = visited.pop(); 3522 checked.push(n); 3523 for (uint i = 0; i < n->outcnt(); i++) { 3524 Node* use = n->raw_out(i); 3525 if (checked.member(use)) continue; // already checked 3526 if (visited.member(use)) continue; // already in the graph 3527 if (use->is_Con()) continue; // a dead ConNode is OK 3528 // At this point, we have found a dead node which is DU-reachable. 3529 if (!dead_nodes) { 3530 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3531 dead_nodes = true; 3532 } 3533 use->dump(2); 3534 tty->print_cr("---"); 3535 checked.push(use); // No repeats; pretend it is now checked. 3536 } 3537 } 3538 assert(!dead_nodes, "using nodes must be reachable from root"); 3539 } 3540 } 3541 } 3542 3543 // Verify GC barriers consistency 3544 // Currently supported: 3545 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3546 void Compile::verify_barriers() { 3547 if (UseG1GC) { 3548 // Verify G1 pre-barriers 3549 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3550 3551 ResourceArea *area = Thread::current()->resource_area(); 3552 Unique_Node_List visited(area); 3553 Node_List worklist(area); 3554 // We're going to walk control flow backwards starting from the Root 3555 worklist.push(_root); 3556 while (worklist.size() > 0) { 3557 Node* x = worklist.pop(); 3558 if (x == NULL || x == top()) continue; 3559 if (visited.member(x)) { 3560 continue; 3561 } else { 3562 visited.push(x); 3563 } 3564 3565 if (x->is_Region()) { 3566 for (uint i = 1; i < x->req(); i++) { 3567 worklist.push(x->in(i)); 3568 } 3569 } else { 3570 worklist.push(x->in(0)); 3571 // We are looking for the pattern: 3572 // /->ThreadLocal 3573 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3574 // \->ConI(0) 3575 // We want to verify that the If and the LoadB have the same control 3576 // See GraphKit::g1_write_barrier_pre() 3577 if (x->is_If()) { 3578 IfNode *iff = x->as_If(); 3579 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3580 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3581 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3582 && cmp->in(1)->is_Load()) { 3583 LoadNode* load = cmp->in(1)->as_Load(); 3584 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3585 && load->in(2)->in(3)->is_Con() 3586 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3587 3588 Node* if_ctrl = iff->in(0); 3589 Node* load_ctrl = load->in(0); 3590 3591 if (if_ctrl != load_ctrl) { 3592 // Skip possible CProj->NeverBranch in infinite loops 3593 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3594 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3595 if_ctrl = if_ctrl->in(0)->in(0); 3596 } 3597 } 3598 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3599 } 3600 } 3601 } 3602 } 3603 } 3604 } 3605 } 3606 } 3607 3608 #endif 3609 3610 // The Compile object keeps track of failure reasons separately from the ciEnv. 3611 // This is required because there is not quite a 1-1 relation between the 3612 // ciEnv and its compilation task and the Compile object. Note that one 3613 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3614 // to backtrack and retry without subsuming loads. Other than this backtracking 3615 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3616 // by the logic in C2Compiler. 3617 void Compile::record_failure(const char* reason) { 3618 if (log() != NULL) { 3619 log()->elem("failure reason='%s' phase='compile'", reason); 3620 } 3621 if (_failure_reason == NULL) { 3622 // Record the first failure reason. 3623 _failure_reason = reason; 3624 } 3625 3626 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3627 C->print_method(PHASE_FAILURE); 3628 } 3629 _root = NULL; // flush the graph, too 3630 } 3631 3632 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3633 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3634 _phase_name(name), _dolog(CITimeVerbose) 3635 { 3636 if (_dolog) { 3637 C = Compile::current(); 3638 _log = C->log(); 3639 } else { 3640 C = NULL; 3641 _log = NULL; 3642 } 3643 if (_log != NULL) { 3644 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3645 _log->stamp(); 3646 _log->end_head(); 3647 } 3648 } 3649 3650 Compile::TracePhase::~TracePhase() { 3651 3652 C = Compile::current(); 3653 if (_dolog) { 3654 _log = C->log(); 3655 } else { 3656 _log = NULL; 3657 } 3658 3659 #ifdef ASSERT 3660 if (PrintIdealNodeCount) { 3661 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3662 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3663 } 3664 3665 if (VerifyIdealNodeCount) { 3666 Compile::current()->print_missing_nodes(); 3667 } 3668 #endif 3669 3670 if (_log != NULL) { 3671 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3672 } 3673 } 3674 3675 //============================================================================= 3676 // Two Constant's are equal when the type and the value are equal. 3677 bool Compile::Constant::operator==(const Constant& other) { 3678 if (type() != other.type() ) return false; 3679 if (can_be_reused() != other.can_be_reused()) return false; 3680 // For floating point values we compare the bit pattern. 3681 switch (type()) { 3682 case T_FLOAT: return (_v._value.i == other._v._value.i); 3683 case T_LONG: 3684 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3685 case T_OBJECT: 3686 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3687 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3688 case T_METADATA: return (_v._metadata == other._v._metadata); 3689 default: ShouldNotReachHere(); 3690 } 3691 return false; 3692 } 3693 3694 static int type_to_size_in_bytes(BasicType t) { 3695 switch (t) { 3696 case T_LONG: return sizeof(jlong ); 3697 case T_FLOAT: return sizeof(jfloat ); 3698 case T_DOUBLE: return sizeof(jdouble); 3699 case T_METADATA: return sizeof(Metadata*); 3700 // We use T_VOID as marker for jump-table entries (labels) which 3701 // need an internal word relocation. 3702 case T_VOID: 3703 case T_ADDRESS: 3704 case T_OBJECT: return sizeof(jobject); 3705 } 3706 3707 ShouldNotReachHere(); 3708 return -1; 3709 } 3710 3711 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3712 // sort descending 3713 if (a->freq() > b->freq()) return -1; 3714 if (a->freq() < b->freq()) return 1; 3715 return 0; 3716 } 3717 3718 void Compile::ConstantTable::calculate_offsets_and_size() { 3719 // First, sort the array by frequencies. 3720 _constants.sort(qsort_comparator); 3721 3722 #ifdef ASSERT 3723 // Make sure all jump-table entries were sorted to the end of the 3724 // array (they have a negative frequency). 3725 bool found_void = false; 3726 for (int i = 0; i < _constants.length(); i++) { 3727 Constant con = _constants.at(i); 3728 if (con.type() == T_VOID) 3729 found_void = true; // jump-tables 3730 else 3731 assert(!found_void, "wrong sorting"); 3732 } 3733 #endif 3734 3735 int offset = 0; 3736 for (int i = 0; i < _constants.length(); i++) { 3737 Constant* con = _constants.adr_at(i); 3738 3739 // Align offset for type. 3740 int typesize = type_to_size_in_bytes(con->type()); 3741 offset = align_size_up(offset, typesize); 3742 con->set_offset(offset); // set constant's offset 3743 3744 if (con->type() == T_VOID) { 3745 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3746 offset = offset + typesize * n->outcnt(); // expand jump-table 3747 } else { 3748 offset = offset + typesize; 3749 } 3750 } 3751 3752 // Align size up to the next section start (which is insts; see 3753 // CodeBuffer::align_at_start). 3754 assert(_size == -1, "already set?"); 3755 _size = align_size_up(offset, CodeEntryAlignment); 3756 } 3757 3758 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3759 MacroAssembler _masm(&cb); 3760 for (int i = 0; i < _constants.length(); i++) { 3761 Constant con = _constants.at(i); 3762 address constant_addr; 3763 switch (con.type()) { 3764 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3765 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3766 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3767 case T_OBJECT: { 3768 jobject obj = con.get_jobject(); 3769 int oop_index = _masm.oop_recorder()->find_index(obj); 3770 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3771 break; 3772 } 3773 case T_ADDRESS: { 3774 address addr = (address) con.get_jobject(); 3775 constant_addr = _masm.address_constant(addr); 3776 break; 3777 } 3778 // We use T_VOID as marker for jump-table entries (labels) which 3779 // need an internal word relocation. 3780 case T_VOID: { 3781 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3782 // Fill the jump-table with a dummy word. The real value is 3783 // filled in later in fill_jump_table. 3784 address dummy = (address) n; 3785 constant_addr = _masm.address_constant(dummy); 3786 // Expand jump-table 3787 for (uint i = 1; i < n->outcnt(); i++) { 3788 address temp_addr = _masm.address_constant(dummy + i); 3789 assert(temp_addr, "consts section too small"); 3790 } 3791 break; 3792 } 3793 case T_METADATA: { 3794 Metadata* obj = con.get_metadata(); 3795 int metadata_index = _masm.oop_recorder()->find_index(obj); 3796 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3797 break; 3798 } 3799 default: ShouldNotReachHere(); 3800 } 3801 assert(constant_addr, "consts section too small"); 3802 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3803 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3804 } 3805 } 3806 3807 int Compile::ConstantTable::find_offset(Constant& con) const { 3808 int idx = _constants.find(con); 3809 assert(idx != -1, "constant must be in constant table"); 3810 int offset = _constants.at(idx).offset(); 3811 assert(offset != -1, "constant table not emitted yet?"); 3812 return offset; 3813 } 3814 3815 void Compile::ConstantTable::add(Constant& con) { 3816 if (con.can_be_reused()) { 3817 int idx = _constants.find(con); 3818 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3819 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3820 return; 3821 } 3822 } 3823 (void) _constants.append(con); 3824 } 3825 3826 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3827 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3828 Constant con(type, value, b->_freq); 3829 add(con); 3830 return con; 3831 } 3832 3833 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3834 Constant con(metadata); 3835 add(con); 3836 return con; 3837 } 3838 3839 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3840 jvalue value; 3841 BasicType type = oper->type()->basic_type(); 3842 switch (type) { 3843 case T_LONG: value.j = oper->constantL(); break; 3844 case T_FLOAT: value.f = oper->constantF(); break; 3845 case T_DOUBLE: value.d = oper->constantD(); break; 3846 case T_OBJECT: 3847 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3848 case T_METADATA: return add((Metadata*)oper->constant()); break; 3849 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3850 } 3851 return add(n, type, value); 3852 } 3853 3854 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3855 jvalue value; 3856 // We can use the node pointer here to identify the right jump-table 3857 // as this method is called from Compile::Fill_buffer right before 3858 // the MachNodes are emitted and the jump-table is filled (means the 3859 // MachNode pointers do not change anymore). 3860 value.l = (jobject) n; 3861 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3862 add(con); 3863 return con; 3864 } 3865 3866 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3867 // If called from Compile::scratch_emit_size do nothing. 3868 if (Compile::current()->in_scratch_emit_size()) return; 3869 3870 assert(labels.is_nonempty(), "must be"); 3871 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3872 3873 // Since MachConstantNode::constant_offset() also contains 3874 // table_base_offset() we need to subtract the table_base_offset() 3875 // to get the plain offset into the constant table. 3876 int offset = n->constant_offset() - table_base_offset(); 3877 3878 MacroAssembler _masm(&cb); 3879 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3880 3881 for (uint i = 0; i < n->outcnt(); i++) { 3882 address* constant_addr = &jump_table_base[i]; 3883 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))); 3884 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3885 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3886 } 3887 } 3888 3889 //----------------------------static_subtype_check----------------------------- 3890 // Shortcut important common cases when superklass is exact: 3891 // (0) superklass is java.lang.Object (can occur in reflective code) 3892 // (1) subklass is already limited to a subtype of superklass => always ok 3893 // (2) subklass does not overlap with superklass => always fail 3894 // (3) superklass has NO subtypes and we can check with a simple compare. 3895 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 3896 if (StressReflectiveCode) { 3897 return SSC_full_test; // Let caller generate the general case. 3898 } 3899 3900 if (superk == env()->Object_klass()) { 3901 return SSC_always_true; // (0) this test cannot fail 3902 } 3903 3904 ciType* superelem = superk; 3905 if (superelem->is_array_klass()) 3906 superelem = superelem->as_array_klass()->base_element_type(); 3907 3908 if (!subk->is_interface()) { // cannot trust static interface types yet 3909 if (subk->is_subtype_of(superk)) { 3910 return SSC_always_true; // (1) false path dead; no dynamic test needed 3911 } 3912 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 3913 !superk->is_subtype_of(subk)) { 3914 return SSC_always_false; 3915 } 3916 } 3917 3918 // If casting to an instance klass, it must have no subtypes 3919 if (superk->is_interface()) { 3920 // Cannot trust interfaces yet. 3921 // %%% S.B. superk->nof_implementors() == 1 3922 } else if (superelem->is_instance_klass()) { 3923 ciInstanceKlass* ik = superelem->as_instance_klass(); 3924 if (!ik->has_subklass() && !ik->is_interface()) { 3925 if (!ik->is_final()) { 3926 // Add a dependency if there is a chance of a later subclass. 3927 dependencies()->assert_leaf_type(ik); 3928 } 3929 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3930 } 3931 } else { 3932 // A primitive array type has no subtypes. 3933 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3934 } 3935 3936 return SSC_full_test; 3937 } 3938 3939 Node* Compile::conv_I2X_index(PhaseGVN *phase, Node* idx, const TypeInt* sizetype) { 3940 #ifdef _LP64 3941 // The scaled index operand to AddP must be a clean 64-bit value. 3942 // Java allows a 32-bit int to be incremented to a negative 3943 // value, which appears in a 64-bit register as a large 3944 // positive number. Using that large positive number as an 3945 // operand in pointer arithmetic has bad consequences. 3946 // On the other hand, 32-bit overflow is rare, and the possibility 3947 // can often be excluded, if we annotate the ConvI2L node with 3948 // a type assertion that its value is known to be a small positive 3949 // number. (The prior range check has ensured this.) 3950 // This assertion is used by ConvI2LNode::Ideal. 3951 int index_max = max_jint - 1; // array size is max_jint, index is one less 3952 if (sizetype != NULL) index_max = sizetype->_hi - 1; 3953 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 3954 idx = phase->transform(new ConvI2LNode(idx, lidxtype)); 3955 #endif 3956 return idx; 3957 } 3958 3959 // The message about the current inlining is accumulated in 3960 // _print_inlining_stream and transfered into the _print_inlining_list 3961 // once we know whether inlining succeeds or not. For regular 3962 // inlining, messages are appended to the buffer pointed by 3963 // _print_inlining_idx in the _print_inlining_list. For late inlining, 3964 // a new buffer is added after _print_inlining_idx in the list. This 3965 // way we can update the inlining message for late inlining call site 3966 // when the inlining is attempted again. 3967 void Compile::print_inlining_init() { 3968 if (print_inlining() || print_intrinsics()) { 3969 _print_inlining_stream = new stringStream(); 3970 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 3971 } 3972 } 3973 3974 void Compile::print_inlining_reinit() { 3975 if (print_inlining() || print_intrinsics()) { 3976 // Re allocate buffer when we change ResourceMark 3977 _print_inlining_stream = new stringStream(); 3978 } 3979 } 3980 3981 void Compile::print_inlining_reset() { 3982 _print_inlining_stream->reset(); 3983 } 3984 3985 void Compile::print_inlining_commit() { 3986 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 3987 // Transfer the message from _print_inlining_stream to the current 3988 // _print_inlining_list buffer and clear _print_inlining_stream. 3989 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 3990 print_inlining_reset(); 3991 } 3992 3993 void Compile::print_inlining_push() { 3994 // Add new buffer to the _print_inlining_list at current position 3995 _print_inlining_idx++; 3996 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 3997 } 3998 3999 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4000 return _print_inlining_list->at(_print_inlining_idx); 4001 } 4002 4003 void Compile::print_inlining_update(CallGenerator* cg) { 4004 if (print_inlining() || print_intrinsics()) { 4005 if (!cg->is_late_inline()) { 4006 if (print_inlining_current().cg() != NULL) { 4007 print_inlining_push(); 4008 } 4009 print_inlining_commit(); 4010 } else { 4011 if (print_inlining_current().cg() != cg && 4012 (print_inlining_current().cg() != NULL || 4013 print_inlining_current().ss()->size() != 0)) { 4014 print_inlining_push(); 4015 } 4016 print_inlining_commit(); 4017 print_inlining_current().set_cg(cg); 4018 } 4019 } 4020 } 4021 4022 void Compile::print_inlining_move_to(CallGenerator* cg) { 4023 // We resume inlining at a late inlining call site. Locate the 4024 // corresponding inlining buffer so that we can update it. 4025 if (print_inlining()) { 4026 for (int i = 0; i < _print_inlining_list->length(); i++) { 4027 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4028 _print_inlining_idx = i; 4029 return; 4030 } 4031 } 4032 ShouldNotReachHere(); 4033 } 4034 } 4035 4036 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4037 if (print_inlining()) { 4038 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4039 assert(print_inlining_current().cg() == cg, "wrong entry"); 4040 // replace message with new message 4041 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4042 print_inlining_commit(); 4043 print_inlining_current().set_cg(cg); 4044 } 4045 } 4046 4047 void Compile::print_inlining_assert_ready() { 4048 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4049 } 4050 4051 void Compile::process_print_inlining() { 4052 bool do_print_inlining = print_inlining() || print_intrinsics(); 4053 if (do_print_inlining || log() != NULL) { 4054 // Print inlining message for candidates that we couldn't inline 4055 // for lack of space 4056 for (int i = 0; i < _late_inlines.length(); i++) { 4057 CallGenerator* cg = _late_inlines.at(i); 4058 if (!cg->is_mh_late_inline()) { 4059 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4060 if (do_print_inlining) { 4061 cg->print_inlining_late(msg); 4062 } 4063 log_late_inline_failure(cg, msg); 4064 } 4065 } 4066 } 4067 if (do_print_inlining) { 4068 ResourceMark rm; 4069 stringStream ss; 4070 for (int i = 0; i < _print_inlining_list->length(); i++) { 4071 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4072 } 4073 size_t end = ss.size(); 4074 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4075 strncpy(_print_inlining_output, ss.base(), end+1); 4076 _print_inlining_output[end] = 0; 4077 } 4078 } 4079 4080 void Compile::dump_print_inlining() { 4081 if (_print_inlining_output != NULL) { 4082 tty->print_raw(_print_inlining_output); 4083 } 4084 } 4085 4086 void Compile::log_late_inline(CallGenerator* cg) { 4087 if (log() != NULL) { 4088 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4089 cg->unique_id()); 4090 JVMState* p = cg->call_node()->jvms(); 4091 while (p != NULL) { 4092 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4093 p = p->caller(); 4094 } 4095 log()->tail("late_inline"); 4096 } 4097 } 4098 4099 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4100 log_late_inline(cg); 4101 if (log() != NULL) { 4102 log()->inline_fail(msg); 4103 } 4104 } 4105 4106 void Compile::log_inline_id(CallGenerator* cg) { 4107 if (log() != NULL) { 4108 // The LogCompilation tool needs a unique way to identify late 4109 // inline call sites. This id must be unique for this call site in 4110 // this compilation. Try to have it unique across compilations as 4111 // well because it can be convenient when grepping through the log 4112 // file. 4113 // Distinguish OSR compilations from others in case CICountOSR is 4114 // on. 4115 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4116 cg->set_unique_id(id); 4117 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4118 } 4119 } 4120 4121 void Compile::log_inline_failure(const char* msg) { 4122 if (C->log() != NULL) { 4123 C->log()->inline_fail(msg); 4124 } 4125 } 4126 4127 4128 // Dump inlining replay data to the stream. 4129 // Don't change thread state and acquire any locks. 4130 void Compile::dump_inline_data(outputStream* out) { 4131 InlineTree* inl_tree = ilt(); 4132 if (inl_tree != NULL) { 4133 out->print(" inline %d", inl_tree->count()); 4134 inl_tree->dump_replay_data(out); 4135 } 4136 } 4137 4138 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4139 if (n1->Opcode() < n2->Opcode()) return -1; 4140 else if (n1->Opcode() > n2->Opcode()) return 1; 4141 4142 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())); 4143 for (uint i = 1; i < n1->req(); i++) { 4144 if (n1->in(i) < n2->in(i)) return -1; 4145 else if (n1->in(i) > n2->in(i)) return 1; 4146 } 4147 4148 return 0; 4149 } 4150 4151 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4152 Node* n1 = *n1p; 4153 Node* n2 = *n2p; 4154 4155 return cmp_expensive_nodes(n1, n2); 4156 } 4157 4158 void Compile::sort_expensive_nodes() { 4159 if (!expensive_nodes_sorted()) { 4160 _expensive_nodes->sort(cmp_expensive_nodes); 4161 } 4162 } 4163 4164 bool Compile::expensive_nodes_sorted() const { 4165 for (int i = 1; i < _expensive_nodes->length(); i++) { 4166 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4167 return false; 4168 } 4169 } 4170 return true; 4171 } 4172 4173 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4174 if (_expensive_nodes->length() == 0) { 4175 return false; 4176 } 4177 4178 assert(OptimizeExpensiveOps, "optimization off?"); 4179 4180 // Take this opportunity to remove dead nodes from the list 4181 int j = 0; 4182 for (int i = 0; i < _expensive_nodes->length(); i++) { 4183 Node* n = _expensive_nodes->at(i); 4184 if (!n->is_unreachable(igvn)) { 4185 assert(n->is_expensive(), "should be expensive"); 4186 _expensive_nodes->at_put(j, n); 4187 j++; 4188 } 4189 } 4190 _expensive_nodes->trunc_to(j); 4191 4192 // Then sort the list so that similar nodes are next to each other 4193 // and check for at least two nodes of identical kind with same data 4194 // inputs. 4195 sort_expensive_nodes(); 4196 4197 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4198 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4199 return true; 4200 } 4201 } 4202 4203 return false; 4204 } 4205 4206 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4207 if (_expensive_nodes->length() == 0) { 4208 return; 4209 } 4210 4211 assert(OptimizeExpensiveOps, "optimization off?"); 4212 4213 // Sort to bring similar nodes next to each other and clear the 4214 // control input of nodes for which there's only a single copy. 4215 sort_expensive_nodes(); 4216 4217 int j = 0; 4218 int identical = 0; 4219 int i = 0; 4220 bool modified = false; 4221 for (; i < _expensive_nodes->length()-1; i++) { 4222 assert(j <= i, "can't write beyond current index"); 4223 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4224 identical++; 4225 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4226 continue; 4227 } 4228 if (identical > 0) { 4229 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4230 identical = 0; 4231 } else { 4232 Node* n = _expensive_nodes->at(i); 4233 igvn.replace_input_of(n, 0, NULL); 4234 igvn.hash_insert(n); 4235 modified = true; 4236 } 4237 } 4238 if (identical > 0) { 4239 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4240 } else if (_expensive_nodes->length() >= 1) { 4241 Node* n = _expensive_nodes->at(i); 4242 igvn.replace_input_of(n, 0, NULL); 4243 igvn.hash_insert(n); 4244 modified = true; 4245 } 4246 _expensive_nodes->trunc_to(j); 4247 if (modified) { 4248 igvn.optimize(); 4249 } 4250 } 4251 4252 void Compile::add_expensive_node(Node * n) { 4253 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4254 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4255 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4256 if (OptimizeExpensiveOps) { 4257 _expensive_nodes->append(n); 4258 } else { 4259 // Clear control input and let IGVN optimize expensive nodes if 4260 // OptimizeExpensiveOps is off. 4261 n->set_req(0, NULL); 4262 } 4263 } 4264 4265 /** 4266 * Remove the speculative part of types and clean up the graph 4267 */ 4268 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4269 if (UseTypeSpeculation) { 4270 Unique_Node_List worklist; 4271 worklist.push(root()); 4272 int modified = 0; 4273 // Go over all type nodes that carry a speculative type, drop the 4274 // speculative part of the type and enqueue the node for an igvn 4275 // which may optimize it out. 4276 for (uint next = 0; next < worklist.size(); ++next) { 4277 Node *n = worklist.at(next); 4278 if (n->is_Type()) { 4279 TypeNode* tn = n->as_Type(); 4280 const Type* t = tn->type(); 4281 const Type* t_no_spec = t->remove_speculative(); 4282 if (t_no_spec != t) { 4283 bool in_hash = igvn.hash_delete(n); 4284 assert(in_hash, "node should be in igvn hash table"); 4285 tn->set_type(t_no_spec); 4286 igvn.hash_insert(n); 4287 igvn._worklist.push(n); // give it a chance to go away 4288 modified++; 4289 } 4290 } 4291 uint max = n->len(); 4292 for( uint i = 0; i < max; ++i ) { 4293 Node *m = n->in(i); 4294 if (not_a_node(m)) continue; 4295 worklist.push(m); 4296 } 4297 } 4298 // Drop the speculative part of all types in the igvn's type table 4299 igvn.remove_speculative_types(); 4300 if (modified > 0) { 4301 igvn.optimize(); 4302 } 4303 #ifdef ASSERT 4304 // Verify that after the IGVN is over no speculative type has resurfaced 4305 worklist.clear(); 4306 worklist.push(root()); 4307 for (uint next = 0; next < worklist.size(); ++next) { 4308 Node *n = worklist.at(next); 4309 const Type* t = igvn.type_or_null(n); 4310 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4311 if (n->is_Type()) { 4312 t = n->as_Type()->type(); 4313 assert(t == t->remove_speculative(), "no more speculative types"); 4314 } 4315 uint max = n->len(); 4316 for( uint i = 0; i < max; ++i ) { 4317 Node *m = n->in(i); 4318 if (not_a_node(m)) continue; 4319 worklist.push(m); 4320 } 4321 } 4322 igvn.check_no_speculative_types(); 4323 #endif 4324 } 4325 } 4326 4327 // Auxiliary method to support randomized stressing/fuzzing. 4328 // 4329 // This method can be called the arbitrary number of times, with current count 4330 // as the argument. The logic allows selecting a single candidate from the 4331 // running list of candidates as follows: 4332 // int count = 0; 4333 // Cand* selected = null; 4334 // while(cand = cand->next()) { 4335 // if (randomized_select(++count)) { 4336 // selected = cand; 4337 // } 4338 // } 4339 // 4340 // Including count equalizes the chances any candidate is "selected". 4341 // This is useful when we don't have the complete list of candidates to choose 4342 // from uniformly. In this case, we need to adjust the randomicity of the 4343 // selection, or else we will end up biasing the selection towards the latter 4344 // candidates. 4345 // 4346 // Quick back-envelope calculation shows that for the list of n candidates 4347 // the equal probability for the candidate to persist as "best" can be 4348 // achieved by replacing it with "next" k-th candidate with the probability 4349 // of 1/k. It can be easily shown that by the end of the run, the 4350 // probability for any candidate is converged to 1/n, thus giving the 4351 // uniform distribution among all the candidates. 4352 // 4353 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4354 #define RANDOMIZED_DOMAIN_POW 29 4355 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4356 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4357 bool Compile::randomized_select(int count) { 4358 assert(count > 0, "only positive"); 4359 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4360 } 4361 4362 CloneMap& Compile::clone_map() { return _clone_map; } 4363 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4364 4365 void NodeCloneInfo::dump() const { 4366 tty->print(" {%d:%d} ", idx(), gen()); 4367 } 4368 4369 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4370 uint64_t val = value(old->_idx); 4371 NodeCloneInfo cio(val); 4372 assert(val != 0, "old node should be in the map"); 4373 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4374 insert(nnn->_idx, cin.get()); 4375 #ifndef PRODUCT 4376 if (is_debug()) { 4377 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4378 } 4379 #endif 4380 } 4381 4382 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4383 NodeCloneInfo cio(value(old->_idx)); 4384 if (cio.get() == 0) { 4385 cio.set(old->_idx, 0); 4386 insert(old->_idx, cio.get()); 4387 #ifndef PRODUCT 4388 if (is_debug()) { 4389 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4390 } 4391 #endif 4392 } 4393 clone(old, nnn, gen); 4394 } 4395 4396 int CloneMap::max_gen() const { 4397 int g = 0; 4398 DictI di(_dict); 4399 for(; di.test(); ++di) { 4400 int t = gen(di._key); 4401 if (g < t) { 4402 g = t; 4403 #ifndef PRODUCT 4404 if (is_debug()) { 4405 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4406 } 4407 #endif 4408 } 4409 } 4410 return g; 4411 } 4412 4413 void CloneMap::dump(node_idx_t key) const { 4414 uint64_t val = value(key); 4415 if (val != 0) { 4416 NodeCloneInfo ni(val); 4417 ni.dump(); 4418 } 4419 }