1 /* 2 * Copyright (c) 1997, 2012, 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 "classfile/systemDictionary.hpp" 29 #include "code/exceptionHandlerTable.hpp" 30 #include "code/nmethod.hpp" 31 #include "compiler/compileLog.hpp" 32 #include "compiler/disassembler.hpp" 33 #include "compiler/oopMap.hpp" 34 #include "opto/addnode.hpp" 35 #include "opto/block.hpp" 36 #include "opto/c2compiler.hpp" 37 #include "opto/callGenerator.hpp" 38 #include "opto/callnode.hpp" 39 #include "opto/cfgnode.hpp" 40 #include "opto/chaitin.hpp" 41 #include "opto/compile.hpp" 42 #include "opto/connode.hpp" 43 #include "opto/divnode.hpp" 44 #include "opto/escape.hpp" 45 #include "opto/idealGraphPrinter.hpp" 46 #include "opto/loopnode.hpp" 47 #include "opto/machnode.hpp" 48 #include "opto/macro.hpp" 49 #include "opto/matcher.hpp" 50 #include "opto/memnode.hpp" 51 #include "opto/mulnode.hpp" 52 #include "opto/node.hpp" 53 #include "opto/opcodes.hpp" 54 #include "opto/output.hpp" 55 #include "opto/parse.hpp" 56 #include "opto/phaseX.hpp" 57 #include "opto/rootnode.hpp" 58 #include "opto/runtime.hpp" 59 #include "opto/stringopts.hpp" 60 #include "opto/type.hpp" 61 #include "opto/vectornode.hpp" 62 #include "runtime/arguments.hpp" 63 #include "runtime/signature.hpp" 64 #include "runtime/stubRoutines.hpp" 65 #include "runtime/timer.hpp" 66 #include "utilities/copy.hpp" 67 #ifdef TARGET_ARCH_MODEL_x86_32 68 # include "adfiles/ad_x86_32.hpp" 69 #endif 70 #ifdef TARGET_ARCH_MODEL_x86_64 71 # include "adfiles/ad_x86_64.hpp" 72 #endif 73 #ifdef TARGET_ARCH_MODEL_sparc 74 # include "adfiles/ad_sparc.hpp" 75 #endif 76 #ifdef TARGET_ARCH_MODEL_zero 77 # include "adfiles/ad_zero.hpp" 78 #endif 79 #ifdef TARGET_ARCH_MODEL_arm 80 # include "adfiles/ad_arm.hpp" 81 #endif 82 #ifdef TARGET_ARCH_MODEL_ppc 83 # include "adfiles/ad_ppc.hpp" 84 #endif 85 86 87 // -------------------- Compile::mach_constant_base_node ----------------------- 88 // Constant table base node singleton. 89 MachConstantBaseNode* Compile::mach_constant_base_node() { 90 if (_mach_constant_base_node == NULL) { 91 _mach_constant_base_node = new (C) MachConstantBaseNode(); 92 _mach_constant_base_node->add_req(C->root()); 93 } 94 return _mach_constant_base_node; 95 } 96 97 98 /// Support for intrinsics. 99 100 // Return the index at which m must be inserted (or already exists). 101 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 102 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 103 #ifdef ASSERT 104 for (int i = 1; i < _intrinsics->length(); i++) { 105 CallGenerator* cg1 = _intrinsics->at(i-1); 106 CallGenerator* cg2 = _intrinsics->at(i); 107 assert(cg1->method() != cg2->method() 108 ? cg1->method() < cg2->method() 109 : cg1->is_virtual() < cg2->is_virtual(), 110 "compiler intrinsics list must stay sorted"); 111 } 112 #endif 113 // Binary search sorted list, in decreasing intervals [lo, hi]. 114 int lo = 0, hi = _intrinsics->length()-1; 115 while (lo <= hi) { 116 int mid = (uint)(hi + lo) / 2; 117 ciMethod* mid_m = _intrinsics->at(mid)->method(); 118 if (m < mid_m) { 119 hi = mid-1; 120 } else if (m > mid_m) { 121 lo = mid+1; 122 } else { 123 // look at minor sort key 124 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 125 if (is_virtual < mid_virt) { 126 hi = mid-1; 127 } else if (is_virtual > mid_virt) { 128 lo = mid+1; 129 } else { 130 return mid; // exact match 131 } 132 } 133 } 134 return lo; // inexact match 135 } 136 137 void Compile::register_intrinsic(CallGenerator* cg) { 138 if (_intrinsics == NULL) { 139 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 140 } 141 // This code is stolen from ciObjectFactory::insert. 142 // Really, GrowableArray should have methods for 143 // insert_at, remove_at, and binary_search. 144 int len = _intrinsics->length(); 145 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 146 if (index == len) { 147 _intrinsics->append(cg); 148 } else { 149 #ifdef ASSERT 150 CallGenerator* oldcg = _intrinsics->at(index); 151 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 152 #endif 153 _intrinsics->append(_intrinsics->at(len-1)); 154 int pos; 155 for (pos = len-2; pos >= index; pos--) { 156 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 157 } 158 _intrinsics->at_put(index, cg); 159 } 160 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 161 } 162 163 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 164 assert(m->is_loaded(), "don't try this on unloaded methods"); 165 if (_intrinsics != NULL) { 166 int index = intrinsic_insertion_index(m, is_virtual); 167 if (index < _intrinsics->length() 168 && _intrinsics->at(index)->method() == m 169 && _intrinsics->at(index)->is_virtual() == is_virtual) { 170 return _intrinsics->at(index); 171 } 172 } 173 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 174 if (m->intrinsic_id() != vmIntrinsics::_none && 175 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 176 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 177 if (cg != NULL) { 178 // Save it for next time: 179 register_intrinsic(cg); 180 return cg; 181 } else { 182 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 183 } 184 } 185 return NULL; 186 } 187 188 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 189 // in library_call.cpp. 190 191 192 #ifndef PRODUCT 193 // statistics gathering... 194 195 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 196 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 197 198 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 199 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 200 int oflags = _intrinsic_hist_flags[id]; 201 assert(flags != 0, "what happened?"); 202 if (is_virtual) { 203 flags |= _intrinsic_virtual; 204 } 205 bool changed = (flags != oflags); 206 if ((flags & _intrinsic_worked) != 0) { 207 juint count = (_intrinsic_hist_count[id] += 1); 208 if (count == 1) { 209 changed = true; // first time 210 } 211 // increment the overall count also: 212 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 213 } 214 if (changed) { 215 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 216 // Something changed about the intrinsic's virtuality. 217 if ((flags & _intrinsic_virtual) != 0) { 218 // This is the first use of this intrinsic as a virtual call. 219 if (oflags != 0) { 220 // We already saw it as a non-virtual, so note both cases. 221 flags |= _intrinsic_both; 222 } 223 } else if ((oflags & _intrinsic_both) == 0) { 224 // This is the first use of this intrinsic as a non-virtual 225 flags |= _intrinsic_both; 226 } 227 } 228 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 229 } 230 // update the overall flags also: 231 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 232 return changed; 233 } 234 235 static char* format_flags(int flags, char* buf) { 236 buf[0] = 0; 237 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 238 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 239 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 240 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 241 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 242 if (buf[0] == 0) strcat(buf, ","); 243 assert(buf[0] == ',', "must be"); 244 return &buf[1]; 245 } 246 247 void Compile::print_intrinsic_statistics() { 248 char flagsbuf[100]; 249 ttyLocker ttyl; 250 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 251 tty->print_cr("Compiler intrinsic usage:"); 252 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 253 if (total == 0) total = 1; // avoid div0 in case of no successes 254 #define PRINT_STAT_LINE(name, c, f) \ 255 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 256 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 257 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 258 int flags = _intrinsic_hist_flags[id]; 259 juint count = _intrinsic_hist_count[id]; 260 if ((flags | count) != 0) { 261 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 262 } 263 } 264 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 265 if (xtty != NULL) xtty->tail("statistics"); 266 } 267 268 void Compile::print_statistics() { 269 { ttyLocker ttyl; 270 if (xtty != NULL) xtty->head("statistics type='opto'"); 271 Parse::print_statistics(); 272 PhaseCCP::print_statistics(); 273 PhaseRegAlloc::print_statistics(); 274 Scheduling::print_statistics(); 275 PhasePeephole::print_statistics(); 276 PhaseIdealLoop::print_statistics(); 277 if (xtty != NULL) xtty->tail("statistics"); 278 } 279 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 280 // put this under its own <statistics> element. 281 print_intrinsic_statistics(); 282 } 283 } 284 #endif //PRODUCT 285 286 // Support for bundling info 287 Bundle* Compile::node_bundling(const Node *n) { 288 assert(valid_bundle_info(n), "oob"); 289 return &_node_bundling_base[n->_idx]; 290 } 291 292 bool Compile::valid_bundle_info(const Node *n) { 293 return (_node_bundling_limit > n->_idx); 294 } 295 296 297 void Compile::gvn_replace_by(Node* n, Node* nn) { 298 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 299 Node* use = n->last_out(i); 300 bool is_in_table = initial_gvn()->hash_delete(use); 301 uint uses_found = 0; 302 for (uint j = 0; j < use->len(); j++) { 303 if (use->in(j) == n) { 304 if (j < use->req()) 305 use->set_req(j, nn); 306 else 307 use->set_prec(j, nn); 308 uses_found++; 309 } 310 } 311 if (is_in_table) { 312 // reinsert into table 313 initial_gvn()->hash_find_insert(use); 314 } 315 record_for_igvn(use); 316 i -= uses_found; // we deleted 1 or more copies of this edge 317 } 318 } 319 320 321 static inline bool not_a_node(const Node* n) { 322 if (n == NULL) return true; 323 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 324 if (*(address*)n == badAddress) return true; // kill by Node::destruct 325 return false; 326 } 327 328 // Identify all nodes that are reachable from below, useful. 329 // Use breadth-first pass that records state in a Unique_Node_List, 330 // recursive traversal is slower. 331 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 332 int estimated_worklist_size = unique(); 333 useful.map( estimated_worklist_size, NULL ); // preallocate space 334 335 // Initialize worklist 336 if (root() != NULL) { useful.push(root()); } 337 // If 'top' is cached, declare it useful to preserve cached node 338 if( cached_top_node() ) { useful.push(cached_top_node()); } 339 340 // Push all useful nodes onto the list, breadthfirst 341 for( uint next = 0; next < useful.size(); ++next ) { 342 assert( next < unique(), "Unique useful nodes < total nodes"); 343 Node *n = useful.at(next); 344 uint max = n->len(); 345 for( uint i = 0; i < max; ++i ) { 346 Node *m = n->in(i); 347 if (not_a_node(m)) continue; 348 useful.push(m); 349 } 350 } 351 } 352 353 // Update dead_node_list with any missing dead nodes using useful 354 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 355 void Compile::update_dead_node_list(Unique_Node_List &useful) { 356 uint max_idx = unique(); 357 VectorSet& useful_node_set = useful.member_set(); 358 359 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 360 // If node with index node_idx is not in useful set, 361 // mark it as dead in dead node list. 362 if (! useful_node_set.test(node_idx) ) { 363 record_dead_node(node_idx); 364 } 365 } 366 } 367 368 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 369 int shift = 0; 370 for (int i = 0; i < inlines->length(); i++) { 371 CallGenerator* cg = inlines->at(i); 372 CallNode* call = cg->call_node(); 373 if (shift > 0) { 374 inlines->at_put(i-shift, cg); 375 } 376 if (!useful.member(call)) { 377 shift++; 378 } 379 } 380 inlines->trunc_to(inlines->length()-shift); 381 } 382 383 // Disconnect all useless nodes by disconnecting those at the boundary. 384 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 385 uint next = 0; 386 while (next < useful.size()) { 387 Node *n = useful.at(next++); 388 // Use raw traversal of out edges since this code removes out edges 389 int max = n->outcnt(); 390 for (int j = 0; j < max; ++j) { 391 Node* child = n->raw_out(j); 392 if (! useful.member(child)) { 393 assert(!child->is_top() || child != top(), 394 "If top is cached in Compile object it is in useful list"); 395 // Only need to remove this out-edge to the useless node 396 n->raw_del_out(j); 397 --j; 398 --max; 399 } 400 } 401 if (n->outcnt() == 1 && n->has_special_unique_user()) { 402 record_for_igvn(n->unique_out()); 403 } 404 } 405 // Remove useless macro and predicate opaq nodes 406 for (int i = C->macro_count()-1; i >= 0; i--) { 407 Node* n = C->macro_node(i); 408 if (!useful.member(n)) { 409 remove_macro_node(n); 410 } 411 } 412 // Remove useless expensive node 413 for (int i = C->expensive_count()-1; i >= 0; i--) { 414 Node* n = C->expensive_node(i); 415 if (!useful.member(n)) { 416 remove_expensive_node(n); 417 } 418 } 419 // clean up the late inline lists 420 remove_useless_late_inlines(&_string_late_inlines, useful); 421 remove_useless_late_inlines(&_late_inlines, useful); 422 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 423 } 424 425 //------------------------------frame_size_in_words----------------------------- 426 // frame_slots in units of words 427 int Compile::frame_size_in_words() const { 428 // shift is 0 in LP32 and 1 in LP64 429 const int shift = (LogBytesPerWord - LogBytesPerInt); 430 int words = _frame_slots >> shift; 431 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 432 return words; 433 } 434 435 // ============================================================================ 436 //------------------------------CompileWrapper--------------------------------- 437 class CompileWrapper : public StackObj { 438 Compile *const _compile; 439 public: 440 CompileWrapper(Compile* compile); 441 442 ~CompileWrapper(); 443 }; 444 445 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 446 // the Compile* pointer is stored in the current ciEnv: 447 ciEnv* env = compile->env(); 448 assert(env == ciEnv::current(), "must already be a ciEnv active"); 449 assert(env->compiler_data() == NULL, "compile already active?"); 450 env->set_compiler_data(compile); 451 assert(compile == Compile::current(), "sanity"); 452 453 compile->set_type_dict(NULL); 454 compile->set_type_hwm(NULL); 455 compile->set_type_last_size(0); 456 compile->set_last_tf(NULL, NULL); 457 compile->set_indexSet_arena(NULL); 458 compile->set_indexSet_free_block_list(NULL); 459 compile->init_type_arena(); 460 Type::Initialize(compile); 461 _compile->set_scratch_buffer_blob(NULL); 462 _compile->begin_method(); 463 } 464 CompileWrapper::~CompileWrapper() { 465 _compile->end_method(); 466 if (_compile->scratch_buffer_blob() != NULL) 467 BufferBlob::free(_compile->scratch_buffer_blob()); 468 _compile->env()->set_compiler_data(NULL); 469 } 470 471 472 //----------------------------print_compile_messages--------------------------- 473 void Compile::print_compile_messages() { 474 #ifndef PRODUCT 475 // Check if recompiling 476 if (_subsume_loads == false && PrintOpto) { 477 // Recompiling without allowing machine instructions to subsume loads 478 tty->print_cr("*********************************************************"); 479 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 480 tty->print_cr("*********************************************************"); 481 } 482 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 483 // Recompiling without escape analysis 484 tty->print_cr("*********************************************************"); 485 tty->print_cr("** Bailout: Recompile without escape analysis **"); 486 tty->print_cr("*********************************************************"); 487 } 488 if (env()->break_at_compile()) { 489 // Open the debugger when compiling this method. 490 tty->print("### Breaking when compiling: "); 491 method()->print_short_name(); 492 tty->cr(); 493 BREAKPOINT; 494 } 495 496 if( PrintOpto ) { 497 if (is_osr_compilation()) { 498 tty->print("[OSR]%3d", _compile_id); 499 } else { 500 tty->print("%3d", _compile_id); 501 } 502 } 503 #endif 504 } 505 506 507 //-----------------------init_scratch_buffer_blob------------------------------ 508 // Construct a temporary BufferBlob and cache it for this compile. 509 void Compile::init_scratch_buffer_blob(int const_size) { 510 // If there is already a scratch buffer blob allocated and the 511 // constant section is big enough, use it. Otherwise free the 512 // current and allocate a new one. 513 BufferBlob* blob = scratch_buffer_blob(); 514 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 515 // Use the current blob. 516 } else { 517 if (blob != NULL) { 518 BufferBlob::free(blob); 519 } 520 521 ResourceMark rm; 522 _scratch_const_size = const_size; 523 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 524 blob = BufferBlob::create("Compile::scratch_buffer", size); 525 // Record the buffer blob for next time. 526 set_scratch_buffer_blob(blob); 527 // Have we run out of code space? 528 if (scratch_buffer_blob() == NULL) { 529 // Let CompilerBroker disable further compilations. 530 record_failure("Not enough space for scratch buffer in CodeCache"); 531 return; 532 } 533 } 534 535 // Initialize the relocation buffers 536 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 537 set_scratch_locs_memory(locs_buf); 538 } 539 540 541 //-----------------------scratch_emit_size------------------------------------- 542 // Helper function that computes size by emitting code 543 uint Compile::scratch_emit_size(const Node* n) { 544 // Start scratch_emit_size section. 545 set_in_scratch_emit_size(true); 546 547 // Emit into a trash buffer and count bytes emitted. 548 // This is a pretty expensive way to compute a size, 549 // but it works well enough if seldom used. 550 // All common fixed-size instructions are given a size 551 // method by the AD file. 552 // Note that the scratch buffer blob and locs memory are 553 // allocated at the beginning of the compile task, and 554 // may be shared by several calls to scratch_emit_size. 555 // The allocation of the scratch buffer blob is particularly 556 // expensive, since it has to grab the code cache lock. 557 BufferBlob* blob = this->scratch_buffer_blob(); 558 assert(blob != NULL, "Initialize BufferBlob at start"); 559 assert(blob->size() > MAX_inst_size, "sanity"); 560 relocInfo* locs_buf = scratch_locs_memory(); 561 address blob_begin = blob->content_begin(); 562 address blob_end = (address)locs_buf; 563 assert(blob->content_contains(blob_end), "sanity"); 564 CodeBuffer buf(blob_begin, blob_end - blob_begin); 565 buf.initialize_consts_size(_scratch_const_size); 566 buf.initialize_stubs_size(MAX_stubs_size); 567 assert(locs_buf != NULL, "sanity"); 568 int lsize = MAX_locs_size / 3; 569 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 570 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 571 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 572 573 // Do the emission. 574 575 Label fakeL; // Fake label for branch instructions. 576 Label* saveL = NULL; 577 uint save_bnum = 0; 578 bool is_branch = n->is_MachBranch(); 579 if (is_branch) { 580 MacroAssembler masm(&buf); 581 masm.bind(fakeL); 582 n->as_MachBranch()->save_label(&saveL, &save_bnum); 583 n->as_MachBranch()->label_set(&fakeL, 0); 584 } 585 n->emit(buf, this->regalloc()); 586 if (is_branch) // Restore label. 587 n->as_MachBranch()->label_set(saveL, save_bnum); 588 589 // End scratch_emit_size section. 590 set_in_scratch_emit_size(false); 591 592 return buf.insts_size(); 593 } 594 595 596 // ============================================================================ 597 //------------------------------Compile standard------------------------------- 598 debug_only( int Compile::_debug_idx = 100000; ) 599 600 // Compile a method. entry_bci is -1 for normal compilations and indicates 601 // the continuation bci for on stack replacement. 602 603 604 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 605 : Phase(Compiler), 606 _env(ci_env), 607 _log(ci_env->log()), 608 _compile_id(ci_env->compile_id()), 609 _save_argument_registers(false), 610 _stub_name(NULL), 611 _stub_function(NULL), 612 _stub_entry_point(NULL), 613 _method(target), 614 _entry_bci(osr_bci), 615 _initial_gvn(NULL), 616 _for_igvn(NULL), 617 _warm_calls(NULL), 618 _subsume_loads(subsume_loads), 619 _do_escape_analysis(do_escape_analysis), 620 _failure_reason(NULL), 621 _code_buffer("Compile::Fill_buffer"), 622 _orig_pc_slot(0), 623 _orig_pc_slot_offset_in_bytes(0), 624 _has_method_handle_invokes(false), 625 _mach_constant_base_node(NULL), 626 _node_bundling_limit(0), 627 _node_bundling_base(NULL), 628 _java_calls(0), 629 _inner_loops(0), 630 _scratch_const_size(-1), 631 _in_scratch_emit_size(false), 632 _dead_node_list(comp_arena()), 633 _dead_node_count(0), 634 #ifndef PRODUCT 635 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 636 _printer(IdealGraphPrinter::printer()), 637 #endif 638 _congraph(NULL), 639 _late_inlines(comp_arena(), 2, 0, NULL), 640 _string_late_inlines(comp_arena(), 2, 0, NULL), 641 _late_inlines_pos(0), 642 _number_of_mh_late_inlines(0), 643 _inlining_progress(false), 644 _inlining_incrementally(false), 645 _print_inlining_list(NULL), 646 _print_inlining(0) { 647 C = this; 648 649 CompileWrapper cw(this); 650 #ifndef PRODUCT 651 if (TimeCompiler2) { 652 tty->print(" "); 653 target->holder()->name()->print(); 654 tty->print("."); 655 target->print_short_name(); 656 tty->print(" "); 657 } 658 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 659 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 660 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 661 if (!print_opto_assembly) { 662 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 663 if (print_assembly && !Disassembler::can_decode()) { 664 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 665 print_opto_assembly = true; 666 } 667 } 668 set_print_assembly(print_opto_assembly); 669 set_parsed_irreducible_loop(false); 670 #endif 671 672 if (ProfileTraps) { 673 // Make sure the method being compiled gets its own MDO, 674 // so we can at least track the decompile_count(). 675 method()->ensure_method_data(); 676 } 677 678 Init(::AliasLevel); 679 680 681 print_compile_messages(); 682 683 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 684 _ilt = InlineTree::build_inline_tree_root(); 685 else 686 _ilt = NULL; 687 688 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 689 assert(num_alias_types() >= AliasIdxRaw, ""); 690 691 #define MINIMUM_NODE_HASH 1023 692 // Node list that Iterative GVN will start with 693 Unique_Node_List for_igvn(comp_arena()); 694 set_for_igvn(&for_igvn); 695 696 // GVN that will be run immediately on new nodes 697 uint estimated_size = method()->code_size()*4+64; 698 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 699 PhaseGVN gvn(node_arena(), estimated_size); 700 set_initial_gvn(&gvn); 701 702 if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) { 703 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 704 } 705 { // Scope for timing the parser 706 TracePhase t3("parse", &_t_parser, true); 707 708 // Put top into the hash table ASAP. 709 initial_gvn()->transform_no_reclaim(top()); 710 711 // Set up tf(), start(), and find a CallGenerator. 712 CallGenerator* cg = NULL; 713 if (is_osr_compilation()) { 714 const TypeTuple *domain = StartOSRNode::osr_domain(); 715 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 716 init_tf(TypeFunc::make(domain, range)); 717 StartNode* s = new (this) StartOSRNode(root(), domain); 718 initial_gvn()->set_type_bottom(s); 719 init_start(s); 720 cg = CallGenerator::for_osr(method(), entry_bci()); 721 } else { 722 // Normal case. 723 init_tf(TypeFunc::make(method())); 724 StartNode* s = new (this) StartNode(root(), tf()->domain()); 725 initial_gvn()->set_type_bottom(s); 726 init_start(s); 727 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 728 // With java.lang.ref.reference.get() we must go through the 729 // intrinsic when G1 is enabled - even when get() is the root 730 // method of the compile - so that, if necessary, the value in 731 // the referent field of the reference object gets recorded by 732 // the pre-barrier code. 733 // Specifically, if G1 is enabled, the value in the referent 734 // field is recorded by the G1 SATB pre barrier. This will 735 // result in the referent being marked live and the reference 736 // object removed from the list of discovered references during 737 // reference processing. 738 cg = find_intrinsic(method(), false); 739 } 740 if (cg == NULL) { 741 float past_uses = method()->interpreter_invocation_count(); 742 float expected_uses = past_uses; 743 cg = CallGenerator::for_inline(method(), expected_uses); 744 } 745 } 746 if (failing()) return; 747 if (cg == NULL) { 748 record_method_not_compilable_all_tiers("cannot parse method"); 749 return; 750 } 751 JVMState* jvms = build_start_state(start(), tf()); 752 if ((jvms = cg->generate(jvms)) == NULL) { 753 record_method_not_compilable("method parse failed"); 754 return; 755 } 756 GraphKit kit(jvms); 757 758 if (!kit.stopped()) { 759 // Accept return values, and transfer control we know not where. 760 // This is done by a special, unique ReturnNode bound to root. 761 return_values(kit.jvms()); 762 } 763 764 if (kit.has_exceptions()) { 765 // Any exceptions that escape from this call must be rethrown 766 // to whatever caller is dynamically above us on the stack. 767 // This is done by a special, unique RethrowNode bound to root. 768 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 769 } 770 771 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 772 773 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 774 inline_string_calls(true); 775 } 776 777 if (failing()) return; 778 779 print_method("Before RemoveUseless", 3); 780 781 // Remove clutter produced by parsing. 782 if (!failing()) { 783 ResourceMark rm; 784 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 785 } 786 } 787 788 // Note: Large methods are capped off in do_one_bytecode(). 789 if (failing()) return; 790 791 // After parsing, node notes are no longer automagic. 792 // They must be propagated by register_new_node_with_optimizer(), 793 // clone(), or the like. 794 set_default_node_notes(NULL); 795 796 for (;;) { 797 int successes = Inline_Warm(); 798 if (failing()) return; 799 if (successes == 0) break; 800 } 801 802 // Drain the list. 803 Finish_Warm(); 804 #ifndef PRODUCT 805 if (_printer) { 806 _printer->print_inlining(this); 807 } 808 #endif 809 810 if (failing()) return; 811 NOT_PRODUCT( verify_graph_edges(); ) 812 813 // Now optimize 814 Optimize(); 815 if (failing()) return; 816 NOT_PRODUCT( verify_graph_edges(); ) 817 818 #ifndef PRODUCT 819 if (PrintIdeal) { 820 ttyLocker ttyl; // keep the following output all in one block 821 // This output goes directly to the tty, not the compiler log. 822 // To enable tools to match it up with the compilation activity, 823 // be sure to tag this tty output with the compile ID. 824 if (xtty != NULL) { 825 xtty->head("ideal compile_id='%d'%s", compile_id(), 826 is_osr_compilation() ? " compile_kind='osr'" : 827 ""); 828 } 829 root()->dump(9999); 830 if (xtty != NULL) { 831 xtty->tail("ideal"); 832 } 833 } 834 #endif 835 836 // Now that we know the size of all the monitors we can add a fixed slot 837 // for the original deopt pc. 838 839 _orig_pc_slot = fixed_slots(); 840 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 841 set_fixed_slots(next_slot); 842 843 // Now generate code 844 Code_Gen(); 845 if (failing()) return; 846 847 // Check if we want to skip execution of all compiled code. 848 { 849 #ifndef PRODUCT 850 if (OptoNoExecute) { 851 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 852 return; 853 } 854 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 855 #endif 856 857 if (is_osr_compilation()) { 858 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 859 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 860 } else { 861 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 862 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 863 } 864 865 env()->register_method(_method, _entry_bci, 866 &_code_offsets, 867 _orig_pc_slot_offset_in_bytes, 868 code_buffer(), 869 frame_size_in_words(), _oop_map_set, 870 &_handler_table, &_inc_table, 871 compiler, 872 env()->comp_level(), 873 has_unsafe_access(), 874 SharedRuntime::is_wide_vector(max_vector_size()) 875 ); 876 877 if (log() != NULL) // Print code cache state into compiler log 878 log()->code_cache_state(); 879 } 880 } 881 882 //------------------------------Compile---------------------------------------- 883 // Compile a runtime stub 884 Compile::Compile( ciEnv* ci_env, 885 TypeFunc_generator generator, 886 address stub_function, 887 const char *stub_name, 888 int is_fancy_jump, 889 bool pass_tls, 890 bool save_arg_registers, 891 bool return_pc ) 892 : Phase(Compiler), 893 _env(ci_env), 894 _log(ci_env->log()), 895 _compile_id(0), 896 _save_argument_registers(save_arg_registers), 897 _method(NULL), 898 _stub_name(stub_name), 899 _stub_function(stub_function), 900 _stub_entry_point(NULL), 901 _entry_bci(InvocationEntryBci), 902 _initial_gvn(NULL), 903 _for_igvn(NULL), 904 _warm_calls(NULL), 905 _orig_pc_slot(0), 906 _orig_pc_slot_offset_in_bytes(0), 907 _subsume_loads(true), 908 _do_escape_analysis(false), 909 _failure_reason(NULL), 910 _code_buffer("Compile::Fill_buffer"), 911 _has_method_handle_invokes(false), 912 _mach_constant_base_node(NULL), 913 _node_bundling_limit(0), 914 _node_bundling_base(NULL), 915 _java_calls(0), 916 _inner_loops(0), 917 #ifndef PRODUCT 918 _trace_opto_output(TraceOptoOutput), 919 _printer(NULL), 920 #endif 921 _dead_node_list(comp_arena()), 922 _dead_node_count(0), 923 _congraph(NULL), 924 _number_of_mh_late_inlines(0), 925 _inlining_progress(false), 926 _inlining_incrementally(false), 927 _print_inlining_list(NULL), 928 _print_inlining(0) { 929 C = this; 930 931 #ifndef PRODUCT 932 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 933 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 934 set_print_assembly(PrintFrameConverterAssembly); 935 set_parsed_irreducible_loop(false); 936 #endif 937 CompileWrapper cw(this); 938 Init(/*AliasLevel=*/ 0); 939 init_tf((*generator)()); 940 941 { 942 // The following is a dummy for the sake of GraphKit::gen_stub 943 Unique_Node_List for_igvn(comp_arena()); 944 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 945 PhaseGVN gvn(Thread::current()->resource_area(),255); 946 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 947 gvn.transform_no_reclaim(top()); 948 949 GraphKit kit; 950 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 951 } 952 953 NOT_PRODUCT( verify_graph_edges(); ) 954 Code_Gen(); 955 if (failing()) return; 956 957 958 // Entry point will be accessed using compile->stub_entry_point(); 959 if (code_buffer() == NULL) { 960 Matcher::soft_match_failure(); 961 } else { 962 if (PrintAssembly && (WizardMode || Verbose)) 963 tty->print_cr("### Stub::%s", stub_name); 964 965 if (!failing()) { 966 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 967 968 // Make the NMethod 969 // For now we mark the frame as never safe for profile stackwalking 970 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 971 code_buffer(), 972 CodeOffsets::frame_never_safe, 973 // _code_offsets.value(CodeOffsets::Frame_Complete), 974 frame_size_in_words(), 975 _oop_map_set, 976 save_arg_registers); 977 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 978 979 _stub_entry_point = rs->entry_point(); 980 } 981 } 982 } 983 984 //------------------------------Init------------------------------------------- 985 // Prepare for a single compilation 986 void Compile::Init(int aliaslevel) { 987 _unique = 0; 988 _regalloc = NULL; 989 990 _tf = NULL; // filled in later 991 _top = NULL; // cached later 992 _matcher = NULL; // filled in later 993 _cfg = NULL; // filled in later 994 995 set_24_bit_selection_and_mode(Use24BitFP, false); 996 997 _node_note_array = NULL; 998 _default_node_notes = NULL; 999 1000 _immutable_memory = NULL; // filled in at first inquiry 1001 1002 // Globally visible Nodes 1003 // First set TOP to NULL to give safe behavior during creation of RootNode 1004 set_cached_top_node(NULL); 1005 set_root(new (this) RootNode()); 1006 // Now that you have a Root to point to, create the real TOP 1007 set_cached_top_node( new (this) ConNode(Type::TOP) ); 1008 set_recent_alloc(NULL, NULL); 1009 1010 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1011 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1012 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1013 env()->set_dependencies(new Dependencies(env())); 1014 1015 _fixed_slots = 0; 1016 set_has_split_ifs(false); 1017 set_has_loops(has_method() && method()->has_loops()); // first approximation 1018 set_has_stringbuilder(false); 1019 _trap_can_recompile = false; // no traps emitted yet 1020 _major_progress = true; // start out assuming good things will happen 1021 set_has_unsafe_access(false); 1022 set_max_vector_size(0); 1023 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1024 set_decompile_count(0); 1025 1026 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1027 set_num_loop_opts(LoopOptsCount); 1028 set_do_inlining(Inline); 1029 set_max_inline_size(MaxInlineSize); 1030 set_freq_inline_size(FreqInlineSize); 1031 set_do_scheduling(OptoScheduling); 1032 set_do_count_invocations(false); 1033 set_do_method_data_update(false); 1034 1035 if (debug_info()->recording_non_safepoints()) { 1036 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1037 (comp_arena(), 8, 0, NULL)); 1038 set_default_node_notes(Node_Notes::make(this)); 1039 } 1040 1041 // // -- Initialize types before each compile -- 1042 // // Update cached type information 1043 // if( _method && _method->constants() ) 1044 // Type::update_loaded_types(_method, _method->constants()); 1045 1046 // Init alias_type map. 1047 if (!_do_escape_analysis && aliaslevel == 3) 1048 aliaslevel = 2; // No unique types without escape analysis 1049 _AliasLevel = aliaslevel; 1050 const int grow_ats = 16; 1051 _max_alias_types = grow_ats; 1052 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1053 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1054 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1055 { 1056 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1057 } 1058 // Initialize the first few types. 1059 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1060 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1061 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1062 _num_alias_types = AliasIdxRaw+1; 1063 // Zero out the alias type cache. 1064 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1065 // A NULL adr_type hits in the cache right away. Preload the right answer. 1066 probe_alias_cache(NULL)->_index = AliasIdxTop; 1067 1068 _intrinsics = NULL; 1069 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1070 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1071 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1072 register_library_intrinsics(); 1073 } 1074 1075 //---------------------------init_start---------------------------------------- 1076 // Install the StartNode on this compile object. 1077 void Compile::init_start(StartNode* s) { 1078 if (failing()) 1079 return; // already failing 1080 assert(s == start(), ""); 1081 } 1082 1083 StartNode* Compile::start() const { 1084 assert(!failing(), ""); 1085 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1086 Node* start = root()->fast_out(i); 1087 if( start->is_Start() ) 1088 return start->as_Start(); 1089 } 1090 ShouldNotReachHere(); 1091 return NULL; 1092 } 1093 1094 //-------------------------------immutable_memory------------------------------------- 1095 // Access immutable memory 1096 Node* Compile::immutable_memory() { 1097 if (_immutable_memory != NULL) { 1098 return _immutable_memory; 1099 } 1100 StartNode* s = start(); 1101 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1102 Node *p = s->fast_out(i); 1103 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1104 _immutable_memory = p; 1105 return _immutable_memory; 1106 } 1107 } 1108 ShouldNotReachHere(); 1109 return NULL; 1110 } 1111 1112 //----------------------set_cached_top_node------------------------------------ 1113 // Install the cached top node, and make sure Node::is_top works correctly. 1114 void Compile::set_cached_top_node(Node* tn) { 1115 if (tn != NULL) verify_top(tn); 1116 Node* old_top = _top; 1117 _top = tn; 1118 // Calling Node::setup_is_top allows the nodes the chance to adjust 1119 // their _out arrays. 1120 if (_top != NULL) _top->setup_is_top(); 1121 if (old_top != NULL) old_top->setup_is_top(); 1122 assert(_top == NULL || top()->is_top(), ""); 1123 } 1124 1125 #ifdef ASSERT 1126 uint Compile::count_live_nodes_by_graph_walk() { 1127 Unique_Node_List useful(comp_arena()); 1128 // Get useful node list by walking the graph. 1129 identify_useful_nodes(useful); 1130 return useful.size(); 1131 } 1132 1133 void Compile::print_missing_nodes() { 1134 1135 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1136 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1137 return; 1138 } 1139 1140 // This is an expensive function. It is executed only when the user 1141 // specifies VerifyIdealNodeCount option or otherwise knows the 1142 // additional work that needs to be done to identify reachable nodes 1143 // by walking the flow graph and find the missing ones using 1144 // _dead_node_list. 1145 1146 Unique_Node_List useful(comp_arena()); 1147 // Get useful node list by walking the graph. 1148 identify_useful_nodes(useful); 1149 1150 uint l_nodes = C->live_nodes(); 1151 uint l_nodes_by_walk = useful.size(); 1152 1153 if (l_nodes != l_nodes_by_walk) { 1154 if (_log != NULL) { 1155 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1156 _log->stamp(); 1157 _log->end_head(); 1158 } 1159 VectorSet& useful_member_set = useful.member_set(); 1160 int last_idx = l_nodes_by_walk; 1161 for (int i = 0; i < last_idx; i++) { 1162 if (useful_member_set.test(i)) { 1163 if (_dead_node_list.test(i)) { 1164 if (_log != NULL) { 1165 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1166 } 1167 if (PrintIdealNodeCount) { 1168 // Print the log message to tty 1169 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1170 useful.at(i)->dump(); 1171 } 1172 } 1173 } 1174 else if (! _dead_node_list.test(i)) { 1175 if (_log != NULL) { 1176 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1177 } 1178 if (PrintIdealNodeCount) { 1179 // Print the log message to tty 1180 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1181 } 1182 } 1183 } 1184 if (_log != NULL) { 1185 _log->tail("mismatched_nodes"); 1186 } 1187 } 1188 } 1189 #endif 1190 1191 #ifndef PRODUCT 1192 void Compile::verify_top(Node* tn) const { 1193 if (tn != NULL) { 1194 assert(tn->is_Con(), "top node must be a constant"); 1195 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1196 assert(tn->in(0) != NULL, "must have live top node"); 1197 } 1198 } 1199 #endif 1200 1201 1202 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1203 1204 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1205 guarantee(arr != NULL, ""); 1206 int num_blocks = arr->length(); 1207 if (grow_by < num_blocks) grow_by = num_blocks; 1208 int num_notes = grow_by * _node_notes_block_size; 1209 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1210 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1211 while (num_notes > 0) { 1212 arr->append(notes); 1213 notes += _node_notes_block_size; 1214 num_notes -= _node_notes_block_size; 1215 } 1216 assert(num_notes == 0, "exact multiple, please"); 1217 } 1218 1219 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1220 if (source == NULL || dest == NULL) return false; 1221 1222 if (dest->is_Con()) 1223 return false; // Do not push debug info onto constants. 1224 1225 #ifdef ASSERT 1226 // Leave a bread crumb trail pointing to the original node: 1227 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1228 dest->set_debug_orig(source); 1229 } 1230 #endif 1231 1232 if (node_note_array() == NULL) 1233 return false; // Not collecting any notes now. 1234 1235 // This is a copy onto a pre-existing node, which may already have notes. 1236 // If both nodes have notes, do not overwrite any pre-existing notes. 1237 Node_Notes* source_notes = node_notes_at(source->_idx); 1238 if (source_notes == NULL || source_notes->is_clear()) return false; 1239 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1240 if (dest_notes == NULL || dest_notes->is_clear()) { 1241 return set_node_notes_at(dest->_idx, source_notes); 1242 } 1243 1244 Node_Notes merged_notes = (*source_notes); 1245 // The order of operations here ensures that dest notes will win... 1246 merged_notes.update_from(dest_notes); 1247 return set_node_notes_at(dest->_idx, &merged_notes); 1248 } 1249 1250 1251 //--------------------------allow_range_check_smearing------------------------- 1252 // Gating condition for coalescing similar range checks. 1253 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1254 // single covering check that is at least as strong as any of them. 1255 // If the optimization succeeds, the simplified (strengthened) range check 1256 // will always succeed. If it fails, we will deopt, and then give up 1257 // on the optimization. 1258 bool Compile::allow_range_check_smearing() const { 1259 // If this method has already thrown a range-check, 1260 // assume it was because we already tried range smearing 1261 // and it failed. 1262 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1263 return !already_trapped; 1264 } 1265 1266 1267 //------------------------------flatten_alias_type----------------------------- 1268 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1269 int offset = tj->offset(); 1270 TypePtr::PTR ptr = tj->ptr(); 1271 1272 // Known instance (scalarizable allocation) alias only with itself. 1273 bool is_known_inst = tj->isa_oopptr() != NULL && 1274 tj->is_oopptr()->is_known_instance(); 1275 1276 // Process weird unsafe references. 1277 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1278 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1279 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1280 tj = TypeOopPtr::BOTTOM; 1281 ptr = tj->ptr(); 1282 offset = tj->offset(); 1283 } 1284 1285 // Array pointers need some flattening 1286 const TypeAryPtr *ta = tj->isa_aryptr(); 1287 if( ta && is_known_inst ) { 1288 if ( offset != Type::OffsetBot && 1289 offset > arrayOopDesc::length_offset_in_bytes() ) { 1290 offset = Type::OffsetBot; // Flatten constant access into array body only 1291 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1292 } 1293 } else if( ta && _AliasLevel >= 2 ) { 1294 // For arrays indexed by constant indices, we flatten the alias 1295 // space to include all of the array body. Only the header, klass 1296 // and array length can be accessed un-aliased. 1297 if( offset != Type::OffsetBot ) { 1298 if( ta->const_oop() ) { // MethodData* or Method* 1299 offset = Type::OffsetBot; // Flatten constant access into array body 1300 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1301 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1302 // range is OK as-is. 1303 tj = ta = TypeAryPtr::RANGE; 1304 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1305 tj = TypeInstPtr::KLASS; // all klass loads look alike 1306 ta = TypeAryPtr::RANGE; // generic ignored junk 1307 ptr = TypePtr::BotPTR; 1308 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1309 tj = TypeInstPtr::MARK; 1310 ta = TypeAryPtr::RANGE; // generic ignored junk 1311 ptr = TypePtr::BotPTR; 1312 } else { // Random constant offset into array body 1313 offset = Type::OffsetBot; // Flatten constant access into array body 1314 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1315 } 1316 } 1317 // Arrays of fixed size alias with arrays of unknown size. 1318 if (ta->size() != TypeInt::POS) { 1319 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1320 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1321 } 1322 // Arrays of known objects become arrays of unknown objects. 1323 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1324 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1325 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1326 } 1327 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1328 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1329 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1330 } 1331 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1332 // cannot be distinguished by bytecode alone. 1333 if (ta->elem() == TypeInt::BOOL) { 1334 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1335 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1336 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1337 } 1338 // During the 2nd round of IterGVN, NotNull castings are removed. 1339 // Make sure the Bottom and NotNull variants alias the same. 1340 // Also, make sure exact and non-exact variants alias the same. 1341 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1342 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1343 } 1344 } 1345 1346 // Oop pointers need some flattening 1347 const TypeInstPtr *to = tj->isa_instptr(); 1348 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1349 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1350 if( ptr == TypePtr::Constant ) { 1351 if (to->klass() != ciEnv::current()->Class_klass() || 1352 offset < k->size_helper() * wordSize) { 1353 // No constant oop pointers (such as Strings); they alias with 1354 // unknown strings. 1355 assert(!is_known_inst, "not scalarizable allocation"); 1356 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1357 } 1358 } else if( is_known_inst ) { 1359 tj = to; // Keep NotNull and klass_is_exact for instance type 1360 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1361 // During the 2nd round of IterGVN, NotNull castings are removed. 1362 // Make sure the Bottom and NotNull variants alias the same. 1363 // Also, make sure exact and non-exact variants alias the same. 1364 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1365 } 1366 // Canonicalize the holder of this field 1367 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1368 // First handle header references such as a LoadKlassNode, even if the 1369 // object's klass is unloaded at compile time (4965979). 1370 if (!is_known_inst) { // Do it only for non-instance types 1371 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1372 } 1373 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1374 // Static fields are in the space above the normal instance 1375 // fields in the java.lang.Class instance. 1376 if (to->klass() != ciEnv::current()->Class_klass()) { 1377 to = NULL; 1378 tj = TypeOopPtr::BOTTOM; 1379 offset = tj->offset(); 1380 } 1381 } else { 1382 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1383 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1384 if( is_known_inst ) { 1385 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1386 } else { 1387 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1388 } 1389 } 1390 } 1391 } 1392 1393 // Klass pointers to object array klasses need some flattening 1394 const TypeKlassPtr *tk = tj->isa_klassptr(); 1395 if( tk ) { 1396 // If we are referencing a field within a Klass, we need 1397 // to assume the worst case of an Object. Both exact and 1398 // inexact types must flatten to the same alias class so 1399 // use NotNull as the PTR. 1400 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1401 1402 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1403 TypeKlassPtr::OBJECT->klass(), 1404 offset); 1405 } 1406 1407 ciKlass* klass = tk->klass(); 1408 if( klass->is_obj_array_klass() ) { 1409 ciKlass* k = TypeAryPtr::OOPS->klass(); 1410 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1411 k = TypeInstPtr::BOTTOM->klass(); 1412 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1413 } 1414 1415 // Check for precise loads from the primary supertype array and force them 1416 // to the supertype cache alias index. Check for generic array loads from 1417 // the primary supertype array and also force them to the supertype cache 1418 // alias index. Since the same load can reach both, we need to merge 1419 // these 2 disparate memories into the same alias class. Since the 1420 // primary supertype array is read-only, there's no chance of confusion 1421 // where we bypass an array load and an array store. 1422 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1423 if (offset == Type::OffsetBot || 1424 (offset >= primary_supers_offset && 1425 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1426 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1427 offset = in_bytes(Klass::secondary_super_cache_offset()); 1428 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1429 } 1430 } 1431 1432 // Flatten all Raw pointers together. 1433 if (tj->base() == Type::RawPtr) 1434 tj = TypeRawPtr::BOTTOM; 1435 1436 if (tj->base() == Type::AnyPtr) 1437 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1438 1439 // Flatten all to bottom for now 1440 switch( _AliasLevel ) { 1441 case 0: 1442 tj = TypePtr::BOTTOM; 1443 break; 1444 case 1: // Flatten to: oop, static, field or array 1445 switch (tj->base()) { 1446 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1447 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1448 case Type::AryPtr: // do not distinguish arrays at all 1449 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1450 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1451 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1452 default: ShouldNotReachHere(); 1453 } 1454 break; 1455 case 2: // No collapsing at level 2; keep all splits 1456 case 3: // No collapsing at level 3; keep all splits 1457 break; 1458 default: 1459 Unimplemented(); 1460 } 1461 1462 offset = tj->offset(); 1463 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1464 1465 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1466 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1467 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1468 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1469 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1470 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1471 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1472 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1473 assert( tj->ptr() != TypePtr::TopPTR && 1474 tj->ptr() != TypePtr::AnyNull && 1475 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1476 // assert( tj->ptr() != TypePtr::Constant || 1477 // tj->base() == Type::RawPtr || 1478 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1479 1480 return tj; 1481 } 1482 1483 void Compile::AliasType::Init(int i, const TypePtr* at) { 1484 _index = i; 1485 _adr_type = at; 1486 _field = NULL; 1487 _is_rewritable = true; // default 1488 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1489 if (atoop != NULL && atoop->is_known_instance()) { 1490 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1491 _general_index = Compile::current()->get_alias_index(gt); 1492 } else { 1493 _general_index = 0; 1494 } 1495 } 1496 1497 //---------------------------------print_on------------------------------------ 1498 #ifndef PRODUCT 1499 void Compile::AliasType::print_on(outputStream* st) { 1500 if (index() < 10) 1501 st->print("@ <%d> ", index()); 1502 else st->print("@ <%d>", index()); 1503 st->print(is_rewritable() ? " " : " RO"); 1504 int offset = adr_type()->offset(); 1505 if (offset == Type::OffsetBot) 1506 st->print(" +any"); 1507 else st->print(" +%-3d", offset); 1508 st->print(" in "); 1509 adr_type()->dump_on(st); 1510 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1511 if (field() != NULL && tjp) { 1512 if (tjp->klass() != field()->holder() || 1513 tjp->offset() != field()->offset_in_bytes()) { 1514 st->print(" != "); 1515 field()->print(); 1516 st->print(" ***"); 1517 } 1518 } 1519 } 1520 1521 void print_alias_types() { 1522 Compile* C = Compile::current(); 1523 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1524 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1525 C->alias_type(idx)->print_on(tty); 1526 tty->cr(); 1527 } 1528 } 1529 #endif 1530 1531 1532 //----------------------------probe_alias_cache-------------------------------- 1533 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1534 intptr_t key = (intptr_t) adr_type; 1535 key ^= key >> logAliasCacheSize; 1536 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1537 } 1538 1539 1540 //-----------------------------grow_alias_types-------------------------------- 1541 void Compile::grow_alias_types() { 1542 const int old_ats = _max_alias_types; // how many before? 1543 const int new_ats = old_ats; // how many more? 1544 const int grow_ats = old_ats+new_ats; // how many now? 1545 _max_alias_types = grow_ats; 1546 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1547 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1548 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1549 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1550 } 1551 1552 1553 //--------------------------------find_alias_type------------------------------ 1554 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1555 if (_AliasLevel == 0) 1556 return alias_type(AliasIdxBot); 1557 1558 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1559 if (ace->_adr_type == adr_type) { 1560 return alias_type(ace->_index); 1561 } 1562 1563 // Handle special cases. 1564 if (adr_type == NULL) return alias_type(AliasIdxTop); 1565 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1566 1567 // Do it the slow way. 1568 const TypePtr* flat = flatten_alias_type(adr_type); 1569 1570 #ifdef ASSERT 1571 assert(flat == flatten_alias_type(flat), "idempotent"); 1572 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1573 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1574 const TypeOopPtr* foop = flat->is_oopptr(); 1575 // Scalarizable allocations have exact klass always. 1576 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1577 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1578 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1579 } 1580 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1581 #endif 1582 1583 int idx = AliasIdxTop; 1584 for (int i = 0; i < num_alias_types(); i++) { 1585 if (alias_type(i)->adr_type() == flat) { 1586 idx = i; 1587 break; 1588 } 1589 } 1590 1591 if (idx == AliasIdxTop) { 1592 if (no_create) return NULL; 1593 // Grow the array if necessary. 1594 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1595 // Add a new alias type. 1596 idx = _num_alias_types++; 1597 _alias_types[idx]->Init(idx, flat); 1598 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1599 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1600 if (flat->isa_instptr()) { 1601 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1602 && flat->is_instptr()->klass() == env()->Class_klass()) 1603 alias_type(idx)->set_rewritable(false); 1604 } 1605 if (flat->isa_klassptr()) { 1606 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1607 alias_type(idx)->set_rewritable(false); 1608 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1609 alias_type(idx)->set_rewritable(false); 1610 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1611 alias_type(idx)->set_rewritable(false); 1612 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1613 alias_type(idx)->set_rewritable(false); 1614 } 1615 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1616 // but the base pointer type is not distinctive enough to identify 1617 // references into JavaThread.) 1618 1619 // Check for final fields. 1620 const TypeInstPtr* tinst = flat->isa_instptr(); 1621 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1622 ciField* field; 1623 if (tinst->const_oop() != NULL && 1624 tinst->klass() == ciEnv::current()->Class_klass() && 1625 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1626 // static field 1627 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1628 field = k->get_field_by_offset(tinst->offset(), true); 1629 } else { 1630 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1631 field = k->get_field_by_offset(tinst->offset(), false); 1632 } 1633 assert(field == NULL || 1634 original_field == NULL || 1635 (field->holder() == original_field->holder() && 1636 field->offset() == original_field->offset() && 1637 field->is_static() == original_field->is_static()), "wrong field?"); 1638 // Set field() and is_rewritable() attributes. 1639 if (field != NULL) alias_type(idx)->set_field(field); 1640 } 1641 } 1642 1643 // Fill the cache for next time. 1644 ace->_adr_type = adr_type; 1645 ace->_index = idx; 1646 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1647 1648 // Might as well try to fill the cache for the flattened version, too. 1649 AliasCacheEntry* face = probe_alias_cache(flat); 1650 if (face->_adr_type == NULL) { 1651 face->_adr_type = flat; 1652 face->_index = idx; 1653 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1654 } 1655 1656 return alias_type(idx); 1657 } 1658 1659 1660 Compile::AliasType* Compile::alias_type(ciField* field) { 1661 const TypeOopPtr* t; 1662 if (field->is_static()) 1663 t = TypeInstPtr::make(field->holder()->java_mirror()); 1664 else 1665 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1666 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1667 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1668 return atp; 1669 } 1670 1671 1672 //------------------------------have_alias_type-------------------------------- 1673 bool Compile::have_alias_type(const TypePtr* adr_type) { 1674 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1675 if (ace->_adr_type == adr_type) { 1676 return true; 1677 } 1678 1679 // Handle special cases. 1680 if (adr_type == NULL) return true; 1681 if (adr_type == TypePtr::BOTTOM) return true; 1682 1683 return find_alias_type(adr_type, true, NULL) != NULL; 1684 } 1685 1686 //-----------------------------must_alias-------------------------------------- 1687 // True if all values of the given address type are in the given alias category. 1688 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1689 if (alias_idx == AliasIdxBot) return true; // the universal category 1690 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1691 if (alias_idx == AliasIdxTop) return false; // the empty category 1692 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1693 1694 // the only remaining possible overlap is identity 1695 int adr_idx = get_alias_index(adr_type); 1696 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1697 assert(adr_idx == alias_idx || 1698 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1699 && adr_type != TypeOopPtr::BOTTOM), 1700 "should not be testing for overlap with an unsafe pointer"); 1701 return adr_idx == alias_idx; 1702 } 1703 1704 //------------------------------can_alias-------------------------------------- 1705 // True if any values of the given address type are in the given alias category. 1706 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1707 if (alias_idx == AliasIdxTop) return false; // the empty category 1708 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1709 if (alias_idx == AliasIdxBot) return true; // the universal category 1710 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1711 1712 // the only remaining possible overlap is identity 1713 int adr_idx = get_alias_index(adr_type); 1714 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1715 return adr_idx == alias_idx; 1716 } 1717 1718 1719 1720 //---------------------------pop_warm_call------------------------------------- 1721 WarmCallInfo* Compile::pop_warm_call() { 1722 WarmCallInfo* wci = _warm_calls; 1723 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1724 return wci; 1725 } 1726 1727 //----------------------------Inline_Warm-------------------------------------- 1728 int Compile::Inline_Warm() { 1729 // If there is room, try to inline some more warm call sites. 1730 // %%% Do a graph index compaction pass when we think we're out of space? 1731 if (!InlineWarmCalls) return 0; 1732 1733 int calls_made_hot = 0; 1734 int room_to_grow = NodeCountInliningCutoff - unique(); 1735 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1736 int amount_grown = 0; 1737 WarmCallInfo* call; 1738 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1739 int est_size = (int)call->size(); 1740 if (est_size > (room_to_grow - amount_grown)) { 1741 // This one won't fit anyway. Get rid of it. 1742 call->make_cold(); 1743 continue; 1744 } 1745 call->make_hot(); 1746 calls_made_hot++; 1747 amount_grown += est_size; 1748 amount_to_grow -= est_size; 1749 } 1750 1751 if (calls_made_hot > 0) set_major_progress(); 1752 return calls_made_hot; 1753 } 1754 1755 1756 //----------------------------Finish_Warm-------------------------------------- 1757 void Compile::Finish_Warm() { 1758 if (!InlineWarmCalls) return; 1759 if (failing()) return; 1760 if (warm_calls() == NULL) return; 1761 1762 // Clean up loose ends, if we are out of space for inlining. 1763 WarmCallInfo* call; 1764 while ((call = pop_warm_call()) != NULL) { 1765 call->make_cold(); 1766 } 1767 } 1768 1769 //---------------------cleanup_loop_predicates----------------------- 1770 // Remove the opaque nodes that protect the predicates so that all unused 1771 // checks and uncommon_traps will be eliminated from the ideal graph 1772 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1773 if (predicate_count()==0) return; 1774 for (int i = predicate_count(); i > 0; i--) { 1775 Node * n = predicate_opaque1_node(i-1); 1776 assert(n->Opcode() == Op_Opaque1, "must be"); 1777 igvn.replace_node(n, n->in(1)); 1778 } 1779 assert(predicate_count()==0, "should be clean!"); 1780 } 1781 1782 // StringOpts and late inlining of string methods 1783 void Compile::inline_string_calls(bool parse_time) { 1784 { 1785 // remove useless nodes to make the usage analysis simpler 1786 ResourceMark rm; 1787 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1788 } 1789 1790 { 1791 ResourceMark rm; 1792 print_method("Before StringOpts", 3); 1793 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1794 print_method("After StringOpts", 3); 1795 } 1796 1797 // now inline anything that we skipped the first time around 1798 if (!parse_time) { 1799 _late_inlines_pos = _late_inlines.length(); 1800 } 1801 1802 while (_string_late_inlines.length() > 0) { 1803 CallGenerator* cg = _string_late_inlines.pop(); 1804 cg->do_late_inline(); 1805 if (failing()) return; 1806 } 1807 _string_late_inlines.trunc_to(0); 1808 } 1809 1810 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1811 assert(IncrementalInline, "incremental inlining should be on"); 1812 PhaseGVN* gvn = initial_gvn(); 1813 1814 set_inlining_progress(false); 1815 for_igvn()->clear(); 1816 gvn->replace_with(&igvn); 1817 1818 int i = 0; 1819 1820 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1821 CallGenerator* cg = _late_inlines.at(i); 1822 _late_inlines_pos = i+1; 1823 cg->do_late_inline(); 1824 if (failing()) return; 1825 } 1826 int j = 0; 1827 for (; i < _late_inlines.length(); i++, j++) { 1828 _late_inlines.at_put(j, _late_inlines.at(i)); 1829 } 1830 _late_inlines.trunc_to(j); 1831 1832 { 1833 ResourceMark rm; 1834 PhaseRemoveUseless pru(C->initial_gvn(), C->for_igvn()); 1835 } 1836 1837 igvn = PhaseIterGVN(gvn); 1838 } 1839 1840 // Perform incremental inlining until bound on number of live nodes is reached 1841 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1842 PhaseGVN* gvn = initial_gvn(); 1843 1844 set_inlining_incrementally(true); 1845 set_inlining_progress(true); 1846 uint low_live_nodes = 0; 1847 1848 while(inlining_progress() && _late_inlines.length() > 0) { 1849 1850 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1851 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1852 // PhaseIdealLoop is expensive so we only try it once we are 1853 // out of loop and we only try it again if the previous helped 1854 // got the number of nodes down significantly 1855 PhaseIdealLoop ideal_loop( igvn, false, true ); 1856 if (failing()) return; 1857 low_live_nodes = live_nodes(); 1858 _major_progress = true; 1859 } 1860 1861 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1862 break; 1863 } 1864 } 1865 1866 inline_incrementally_one(igvn); 1867 1868 if (failing()) return; 1869 1870 igvn.optimize(); 1871 1872 if (failing()) return; 1873 } 1874 1875 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1876 1877 if (_string_late_inlines.length() > 0) { 1878 assert(has_stringbuilder(), "inconsistent"); 1879 for_igvn()->clear(); 1880 initial_gvn()->replace_with(&igvn); 1881 1882 inline_string_calls(false); 1883 1884 if (failing()) return; 1885 1886 { 1887 ResourceMark rm; 1888 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1889 } 1890 1891 igvn = PhaseIterGVN(gvn); 1892 1893 igvn.optimize(); 1894 } 1895 1896 set_inlining_incrementally(false); 1897 } 1898 1899 1900 //------------------------------Optimize--------------------------------------- 1901 // Given a graph, optimize it. 1902 void Compile::Optimize() { 1903 TracePhase t1("optimizer", &_t_optimizer, true); 1904 1905 #ifndef PRODUCT 1906 if (env()->break_at_compile()) { 1907 BREAKPOINT; 1908 } 1909 1910 #endif 1911 1912 ResourceMark rm; 1913 int loop_opts_cnt; 1914 1915 NOT_PRODUCT( verify_graph_edges(); ) 1916 1917 print_method("After Parsing"); 1918 1919 { 1920 // Iterative Global Value Numbering, including ideal transforms 1921 // Initialize IterGVN with types and values from parse-time GVN 1922 PhaseIterGVN igvn(initial_gvn()); 1923 { 1924 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1925 igvn.optimize(); 1926 } 1927 1928 print_method("Iter GVN 1", 2); 1929 1930 if (failing()) return; 1931 1932 inline_incrementally(igvn); 1933 1934 print_method("Incremental Inline", 2); 1935 1936 if (failing()) return; 1937 1938 // No more new expensive nodes will be added to the list from here 1939 // so keep only the actual candidates for optimizations. 1940 cleanup_expensive_nodes(igvn); 1941 1942 // Perform escape analysis 1943 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 1944 if (has_loops()) { 1945 // Cleanup graph (remove dead nodes). 1946 TracePhase t2("idealLoop", &_t_idealLoop, true); 1947 PhaseIdealLoop ideal_loop( igvn, false, true ); 1948 if (major_progress()) print_method("PhaseIdealLoop before EA", 2); 1949 if (failing()) return; 1950 } 1951 ConnectionGraph::do_analysis(this, &igvn); 1952 1953 if (failing()) return; 1954 1955 // Optimize out fields loads from scalar replaceable allocations. 1956 igvn.optimize(); 1957 print_method("Iter GVN after EA", 2); 1958 1959 if (failing()) return; 1960 1961 if (congraph() != NULL && macro_count() > 0) { 1962 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 1963 PhaseMacroExpand mexp(igvn); 1964 mexp.eliminate_macro_nodes(); 1965 igvn.set_delay_transform(false); 1966 1967 igvn.optimize(); 1968 print_method("Iter GVN after eliminating allocations and locks", 2); 1969 1970 if (failing()) return; 1971 } 1972 } 1973 1974 // Loop transforms on the ideal graph. Range Check Elimination, 1975 // peeling, unrolling, etc. 1976 1977 // Set loop opts counter 1978 loop_opts_cnt = num_loop_opts(); 1979 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1980 { 1981 TracePhase t2("idealLoop", &_t_idealLoop, true); 1982 PhaseIdealLoop ideal_loop( igvn, true ); 1983 loop_opts_cnt--; 1984 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1985 if (failing()) return; 1986 } 1987 // Loop opts pass if partial peeling occurred in previous pass 1988 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1989 TracePhase t3("idealLoop", &_t_idealLoop, true); 1990 PhaseIdealLoop ideal_loop( igvn, false ); 1991 loop_opts_cnt--; 1992 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1993 if (failing()) return; 1994 } 1995 // Loop opts pass for loop-unrolling before CCP 1996 if(major_progress() && (loop_opts_cnt > 0)) { 1997 TracePhase t4("idealLoop", &_t_idealLoop, true); 1998 PhaseIdealLoop ideal_loop( igvn, false ); 1999 loop_opts_cnt--; 2000 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 2001 } 2002 if (!failing()) { 2003 // Verify that last round of loop opts produced a valid graph 2004 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2005 PhaseIdealLoop::verify(igvn); 2006 } 2007 } 2008 if (failing()) return; 2009 2010 // Conditional Constant Propagation; 2011 PhaseCCP ccp( &igvn ); 2012 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2013 { 2014 TracePhase t2("ccp", &_t_ccp, true); 2015 ccp.do_transform(); 2016 } 2017 print_method("PhaseCPP 1", 2); 2018 2019 assert( true, "Break here to ccp.dump_old2new_map()"); 2020 2021 // Iterative Global Value Numbering, including ideal transforms 2022 { 2023 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2024 igvn = ccp; 2025 igvn.optimize(); 2026 } 2027 2028 print_method("Iter GVN 2", 2); 2029 2030 if (failing()) return; 2031 2032 // Loop transforms on the ideal graph. Range Check Elimination, 2033 // peeling, unrolling, etc. 2034 if(loop_opts_cnt > 0) { 2035 debug_only( int cnt = 0; ); 2036 while(major_progress() && (loop_opts_cnt > 0)) { 2037 TracePhase t2("idealLoop", &_t_idealLoop, true); 2038 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2039 PhaseIdealLoop ideal_loop( igvn, true); 2040 loop_opts_cnt--; 2041 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 2042 if (failing()) return; 2043 } 2044 } 2045 2046 { 2047 // Verify that all previous optimizations produced a valid graph 2048 // at least to this point, even if no loop optimizations were done. 2049 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2050 PhaseIdealLoop::verify(igvn); 2051 } 2052 2053 { 2054 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2055 PhaseMacroExpand mex(igvn); 2056 if (mex.expand_macro_nodes()) { 2057 assert(failing(), "must bail out w/ explicit message"); 2058 return; 2059 } 2060 } 2061 2062 } // (End scope of igvn; run destructor if necessary for asserts.) 2063 2064 dump_inlining(); 2065 // A method with only infinite loops has no edges entering loops from root 2066 { 2067 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2068 if (final_graph_reshaping()) { 2069 assert(failing(), "must bail out w/ explicit message"); 2070 return; 2071 } 2072 } 2073 2074 print_method("Optimize finished", 2); 2075 } 2076 2077 2078 //------------------------------Code_Gen--------------------------------------- 2079 // Given a graph, generate code for it 2080 void Compile::Code_Gen() { 2081 if (failing()) return; 2082 2083 // Perform instruction selection. You might think we could reclaim Matcher 2084 // memory PDQ, but actually the Matcher is used in generating spill code. 2085 // Internals of the Matcher (including some VectorSets) must remain live 2086 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2087 // set a bit in reclaimed memory. 2088 2089 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2090 // nodes. Mapping is only valid at the root of each matched subtree. 2091 NOT_PRODUCT( verify_graph_edges(); ) 2092 2093 Node_List proj_list; 2094 Matcher m(proj_list); 2095 _matcher = &m; 2096 { 2097 TracePhase t2("matcher", &_t_matcher, true); 2098 m.match(); 2099 } 2100 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2101 // nodes. Mapping is only valid at the root of each matched subtree. 2102 NOT_PRODUCT( verify_graph_edges(); ) 2103 2104 // If you have too many nodes, or if matching has failed, bail out 2105 check_node_count(0, "out of nodes matching instructions"); 2106 if (failing()) return; 2107 2108 // Build a proper-looking CFG 2109 PhaseCFG cfg(node_arena(), root(), m); 2110 _cfg = &cfg; 2111 { 2112 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2113 cfg.Dominators(); 2114 if (failing()) return; 2115 2116 NOT_PRODUCT( verify_graph_edges(); ) 2117 2118 cfg.Estimate_Block_Frequency(); 2119 cfg.GlobalCodeMotion(m,unique(),proj_list); 2120 if (failing()) return; 2121 2122 print_method("Global code motion", 2); 2123 2124 NOT_PRODUCT( verify_graph_edges(); ) 2125 2126 debug_only( cfg.verify(); ) 2127 } 2128 NOT_PRODUCT( verify_graph_edges(); ) 2129 2130 PhaseChaitin regalloc(unique(), cfg, m); 2131 _regalloc = ®alloc; 2132 { 2133 TracePhase t2("regalloc", &_t_registerAllocation, true); 2134 // Perform register allocation. After Chaitin, use-def chains are 2135 // no longer accurate (at spill code) and so must be ignored. 2136 // Node->LRG->reg mappings are still accurate. 2137 _regalloc->Register_Allocate(); 2138 2139 // Bail out if the allocator builds too many nodes 2140 if (failing()) { 2141 return; 2142 } 2143 } 2144 2145 // Prior to register allocation we kept empty basic blocks in case the 2146 // the allocator needed a place to spill. After register allocation we 2147 // are not adding any new instructions. If any basic block is empty, we 2148 // can now safely remove it. 2149 { 2150 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2151 cfg.remove_empty(); 2152 if (do_freq_based_layout()) { 2153 PhaseBlockLayout layout(cfg); 2154 } else { 2155 cfg.set_loop_alignment(); 2156 } 2157 cfg.fixup_flow(); 2158 } 2159 2160 // Apply peephole optimizations 2161 if( OptoPeephole ) { 2162 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2163 PhasePeephole peep( _regalloc, cfg); 2164 peep.do_transform(); 2165 } 2166 2167 // Convert Nodes to instruction bits in a buffer 2168 { 2169 // %%%% workspace merge brought two timers together for one job 2170 TracePhase t2a("output", &_t_output, true); 2171 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2172 Output(); 2173 } 2174 2175 print_method("Final Code"); 2176 2177 // He's dead, Jim. 2178 _cfg = (PhaseCFG*)0xdeadbeef; 2179 _regalloc = (PhaseChaitin*)0xdeadbeef; 2180 } 2181 2182 2183 //------------------------------dump_asm--------------------------------------- 2184 // Dump formatted assembly 2185 #ifndef PRODUCT 2186 void Compile::dump_asm(int *pcs, uint pc_limit) { 2187 bool cut_short = false; 2188 tty->print_cr("#"); 2189 tty->print("# "); _tf->dump(); tty->cr(); 2190 tty->print_cr("#"); 2191 2192 // For all blocks 2193 int pc = 0x0; // Program counter 2194 char starts_bundle = ' '; 2195 _regalloc->dump_frame(); 2196 2197 Node *n = NULL; 2198 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 2199 if (VMThread::should_terminate()) { cut_short = true; break; } 2200 Block *b = _cfg->_blocks[i]; 2201 if (b->is_connector() && !Verbose) continue; 2202 n = b->_nodes[0]; 2203 if (pcs && n->_idx < pc_limit) 2204 tty->print("%3.3x ", pcs[n->_idx]); 2205 else 2206 tty->print(" "); 2207 b->dump_head( &_cfg->_bbs ); 2208 if (b->is_connector()) { 2209 tty->print_cr(" # Empty connector block"); 2210 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2211 tty->print_cr(" # Block is sole successor of call"); 2212 } 2213 2214 // For all instructions 2215 Node *delay = NULL; 2216 for( uint j = 0; j<b->_nodes.size(); j++ ) { 2217 if (VMThread::should_terminate()) { cut_short = true; break; } 2218 n = b->_nodes[j]; 2219 if (valid_bundle_info(n)) { 2220 Bundle *bundle = node_bundling(n); 2221 if (bundle->used_in_unconditional_delay()) { 2222 delay = n; 2223 continue; 2224 } 2225 if (bundle->starts_bundle()) 2226 starts_bundle = '+'; 2227 } 2228 2229 if (WizardMode) n->dump(); 2230 2231 if( !n->is_Region() && // Dont print in the Assembly 2232 !n->is_Phi() && // a few noisely useless nodes 2233 !n->is_Proj() && 2234 !n->is_MachTemp() && 2235 !n->is_SafePointScalarObject() && 2236 !n->is_Catch() && // Would be nice to print exception table targets 2237 !n->is_MergeMem() && // Not very interesting 2238 !n->is_top() && // Debug info table constants 2239 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2240 ) { 2241 if (pcs && n->_idx < pc_limit) 2242 tty->print("%3.3x", pcs[n->_idx]); 2243 else 2244 tty->print(" "); 2245 tty->print(" %c ", starts_bundle); 2246 starts_bundle = ' '; 2247 tty->print("\t"); 2248 n->format(_regalloc, tty); 2249 tty->cr(); 2250 } 2251 2252 // If we have an instruction with a delay slot, and have seen a delay, 2253 // then back up and print it 2254 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2255 assert(delay != NULL, "no unconditional delay instruction"); 2256 if (WizardMode) delay->dump(); 2257 2258 if (node_bundling(delay)->starts_bundle()) 2259 starts_bundle = '+'; 2260 if (pcs && n->_idx < pc_limit) 2261 tty->print("%3.3x", pcs[n->_idx]); 2262 else 2263 tty->print(" "); 2264 tty->print(" %c ", starts_bundle); 2265 starts_bundle = ' '; 2266 tty->print("\t"); 2267 delay->format(_regalloc, tty); 2268 tty->print_cr(""); 2269 delay = NULL; 2270 } 2271 2272 // Dump the exception table as well 2273 if( n->is_Catch() && (Verbose || WizardMode) ) { 2274 // Print the exception table for this offset 2275 _handler_table.print_subtable_for(pc); 2276 } 2277 } 2278 2279 if (pcs && n->_idx < pc_limit) 2280 tty->print_cr("%3.3x", pcs[n->_idx]); 2281 else 2282 tty->print_cr(""); 2283 2284 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2285 2286 } // End of per-block dump 2287 tty->print_cr(""); 2288 2289 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2290 } 2291 #endif 2292 2293 //------------------------------Final_Reshape_Counts--------------------------- 2294 // This class defines counters to help identify when a method 2295 // may/must be executed using hardware with only 24-bit precision. 2296 struct Final_Reshape_Counts : public StackObj { 2297 int _call_count; // count non-inlined 'common' calls 2298 int _float_count; // count float ops requiring 24-bit precision 2299 int _double_count; // count double ops requiring more precision 2300 int _java_call_count; // count non-inlined 'java' calls 2301 int _inner_loop_count; // count loops which need alignment 2302 VectorSet _visited; // Visitation flags 2303 Node_List _tests; // Set of IfNodes & PCTableNodes 2304 2305 Final_Reshape_Counts() : 2306 _call_count(0), _float_count(0), _double_count(0), 2307 _java_call_count(0), _inner_loop_count(0), 2308 _visited( Thread::current()->resource_area() ) { } 2309 2310 void inc_call_count () { _call_count ++; } 2311 void inc_float_count () { _float_count ++; } 2312 void inc_double_count() { _double_count++; } 2313 void inc_java_call_count() { _java_call_count++; } 2314 void inc_inner_loop_count() { _inner_loop_count++; } 2315 2316 int get_call_count () const { return _call_count ; } 2317 int get_float_count () const { return _float_count ; } 2318 int get_double_count() const { return _double_count; } 2319 int get_java_call_count() const { return _java_call_count; } 2320 int get_inner_loop_count() const { return _inner_loop_count; } 2321 }; 2322 2323 #ifdef ASSERT 2324 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2325 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2326 // Make sure the offset goes inside the instance layout. 2327 return k->contains_field_offset(tp->offset()); 2328 // Note that OffsetBot and OffsetTop are very negative. 2329 } 2330 #endif 2331 2332 // Eliminate trivially redundant StoreCMs and accumulate their 2333 // precedence edges. 2334 void Compile::eliminate_redundant_card_marks(Node* n) { 2335 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2336 if (n->in(MemNode::Address)->outcnt() > 1) { 2337 // There are multiple users of the same address so it might be 2338 // possible to eliminate some of the StoreCMs 2339 Node* mem = n->in(MemNode::Memory); 2340 Node* adr = n->in(MemNode::Address); 2341 Node* val = n->in(MemNode::ValueIn); 2342 Node* prev = n; 2343 bool done = false; 2344 // Walk the chain of StoreCMs eliminating ones that match. As 2345 // long as it's a chain of single users then the optimization is 2346 // safe. Eliminating partially redundant StoreCMs would require 2347 // cloning copies down the other paths. 2348 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2349 if (adr == mem->in(MemNode::Address) && 2350 val == mem->in(MemNode::ValueIn)) { 2351 // redundant StoreCM 2352 if (mem->req() > MemNode::OopStore) { 2353 // Hasn't been processed by this code yet. 2354 n->add_prec(mem->in(MemNode::OopStore)); 2355 } else { 2356 // Already converted to precedence edge 2357 for (uint i = mem->req(); i < mem->len(); i++) { 2358 // Accumulate any precedence edges 2359 if (mem->in(i) != NULL) { 2360 n->add_prec(mem->in(i)); 2361 } 2362 } 2363 // Everything above this point has been processed. 2364 done = true; 2365 } 2366 // Eliminate the previous StoreCM 2367 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2368 assert(mem->outcnt() == 0, "should be dead"); 2369 mem->disconnect_inputs(NULL, this); 2370 } else { 2371 prev = mem; 2372 } 2373 mem = prev->in(MemNode::Memory); 2374 } 2375 } 2376 } 2377 2378 //------------------------------final_graph_reshaping_impl---------------------- 2379 // Implement items 1-5 from final_graph_reshaping below. 2380 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2381 2382 if ( n->outcnt() == 0 ) return; // dead node 2383 uint nop = n->Opcode(); 2384 2385 // Check for 2-input instruction with "last use" on right input. 2386 // Swap to left input. Implements item (2). 2387 if( n->req() == 3 && // two-input instruction 2388 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2389 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2390 n->in(2)->outcnt() == 1 &&// right use IS a last use 2391 !n->in(2)->is_Con() ) { // right use is not a constant 2392 // Check for commutative opcode 2393 switch( nop ) { 2394 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2395 case Op_MaxI: case Op_MinI: 2396 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2397 case Op_AndL: case Op_XorL: case Op_OrL: 2398 case Op_AndI: case Op_XorI: case Op_OrI: { 2399 // Move "last use" input to left by swapping inputs 2400 n->swap_edges(1, 2); 2401 break; 2402 } 2403 default: 2404 break; 2405 } 2406 } 2407 2408 #ifdef ASSERT 2409 if( n->is_Mem() ) { 2410 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2411 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2412 // oop will be recorded in oop map if load crosses safepoint 2413 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2414 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2415 "raw memory operations should have control edge"); 2416 } 2417 #endif 2418 // Count FPU ops and common calls, implements item (3) 2419 switch( nop ) { 2420 // Count all float operations that may use FPU 2421 case Op_AddF: 2422 case Op_SubF: 2423 case Op_MulF: 2424 case Op_DivF: 2425 case Op_NegF: 2426 case Op_ModF: 2427 case Op_ConvI2F: 2428 case Op_ConF: 2429 case Op_CmpF: 2430 case Op_CmpF3: 2431 // case Op_ConvL2F: // longs are split into 32-bit halves 2432 frc.inc_float_count(); 2433 break; 2434 2435 case Op_ConvF2D: 2436 case Op_ConvD2F: 2437 frc.inc_float_count(); 2438 frc.inc_double_count(); 2439 break; 2440 2441 // Count all double operations that may use FPU 2442 case Op_AddD: 2443 case Op_SubD: 2444 case Op_MulD: 2445 case Op_DivD: 2446 case Op_NegD: 2447 case Op_ModD: 2448 case Op_ConvI2D: 2449 case Op_ConvD2I: 2450 // case Op_ConvL2D: // handled by leaf call 2451 // case Op_ConvD2L: // handled by leaf call 2452 case Op_ConD: 2453 case Op_CmpD: 2454 case Op_CmpD3: 2455 frc.inc_double_count(); 2456 break; 2457 case Op_Opaque1: // Remove Opaque Nodes before matching 2458 case Op_Opaque2: // Remove Opaque Nodes before matching 2459 n->subsume_by(n->in(1), this); 2460 break; 2461 case Op_CallStaticJava: 2462 case Op_CallJava: 2463 case Op_CallDynamicJava: 2464 frc.inc_java_call_count(); // Count java call site; 2465 case Op_CallRuntime: 2466 case Op_CallLeaf: 2467 case Op_CallLeafNoFP: { 2468 assert( n->is_Call(), "" ); 2469 CallNode *call = n->as_Call(); 2470 // Count call sites where the FP mode bit would have to be flipped. 2471 // Do not count uncommon runtime calls: 2472 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2473 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2474 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2475 frc.inc_call_count(); // Count the call site 2476 } else { // See if uncommon argument is shared 2477 Node *n = call->in(TypeFunc::Parms); 2478 int nop = n->Opcode(); 2479 // Clone shared simple arguments to uncommon calls, item (1). 2480 if( n->outcnt() > 1 && 2481 !n->is_Proj() && 2482 nop != Op_CreateEx && 2483 nop != Op_CheckCastPP && 2484 nop != Op_DecodeN && 2485 nop != Op_DecodeNKlass && 2486 !n->is_Mem() ) { 2487 Node *x = n->clone(); 2488 call->set_req( TypeFunc::Parms, x ); 2489 } 2490 } 2491 break; 2492 } 2493 2494 case Op_StoreD: 2495 case Op_LoadD: 2496 case Op_LoadD_unaligned: 2497 frc.inc_double_count(); 2498 goto handle_mem; 2499 case Op_StoreF: 2500 case Op_LoadF: 2501 frc.inc_float_count(); 2502 goto handle_mem; 2503 2504 case Op_StoreCM: 2505 { 2506 // Convert OopStore dependence into precedence edge 2507 Node* prec = n->in(MemNode::OopStore); 2508 n->del_req(MemNode::OopStore); 2509 n->add_prec(prec); 2510 eliminate_redundant_card_marks(n); 2511 } 2512 2513 // fall through 2514 2515 case Op_StoreB: 2516 case Op_StoreC: 2517 case Op_StorePConditional: 2518 case Op_StoreI: 2519 case Op_StoreL: 2520 case Op_StoreIConditional: 2521 case Op_StoreLConditional: 2522 case Op_CompareAndSwapI: 2523 case Op_CompareAndSwapL: 2524 case Op_CompareAndSwapP: 2525 case Op_CompareAndSwapN: 2526 case Op_GetAndAddI: 2527 case Op_GetAndAddL: 2528 case Op_GetAndSetI: 2529 case Op_GetAndSetL: 2530 case Op_GetAndSetP: 2531 case Op_GetAndSetN: 2532 case Op_StoreP: 2533 case Op_StoreN: 2534 case Op_StoreNKlass: 2535 case Op_LoadB: 2536 case Op_LoadUB: 2537 case Op_LoadUS: 2538 case Op_LoadI: 2539 case Op_LoadKlass: 2540 case Op_LoadNKlass: 2541 case Op_LoadL: 2542 case Op_LoadL_unaligned: 2543 case Op_LoadPLocked: 2544 case Op_LoadP: 2545 case Op_LoadN: 2546 case Op_LoadRange: 2547 case Op_LoadS: { 2548 handle_mem: 2549 #ifdef ASSERT 2550 if( VerifyOptoOopOffsets ) { 2551 assert( n->is_Mem(), "" ); 2552 MemNode *mem = (MemNode*)n; 2553 // Check to see if address types have grounded out somehow. 2554 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2555 assert( !tp || oop_offset_is_sane(tp), "" ); 2556 } 2557 #endif 2558 break; 2559 } 2560 2561 case Op_AddP: { // Assert sane base pointers 2562 Node *addp = n->in(AddPNode::Address); 2563 assert( !addp->is_AddP() || 2564 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2565 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2566 "Base pointers must match" ); 2567 #ifdef _LP64 2568 if ((UseCompressedOops || UseCompressedKlassPointers) && 2569 addp->Opcode() == Op_ConP && 2570 addp == n->in(AddPNode::Base) && 2571 n->in(AddPNode::Offset)->is_Con()) { 2572 // Use addressing with narrow klass to load with offset on x86. 2573 // On sparc loading 32-bits constant and decoding it have less 2574 // instructions (4) then load 64-bits constant (7). 2575 // Do this transformation here since IGVN will convert ConN back to ConP. 2576 const Type* t = addp->bottom_type(); 2577 if (t->isa_oopptr() || t->isa_klassptr()) { 2578 Node* nn = NULL; 2579 2580 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2581 2582 // Look for existing ConN node of the same exact type. 2583 Node* r = root(); 2584 uint cnt = r->outcnt(); 2585 for (uint i = 0; i < cnt; i++) { 2586 Node* m = r->raw_out(i); 2587 if (m!= NULL && m->Opcode() == op && 2588 m->bottom_type()->make_ptr() == t) { 2589 nn = m; 2590 break; 2591 } 2592 } 2593 if (nn != NULL) { 2594 // Decode a narrow oop to match address 2595 // [R12 + narrow_oop_reg<<3 + offset] 2596 if (t->isa_oopptr()) { 2597 nn = new (this) DecodeNNode(nn, t); 2598 } else { 2599 nn = new (this) DecodeNKlassNode(nn, t); 2600 } 2601 n->set_req(AddPNode::Base, nn); 2602 n->set_req(AddPNode::Address, nn); 2603 if (addp->outcnt() == 0) { 2604 addp->disconnect_inputs(NULL, this); 2605 } 2606 } 2607 } 2608 } 2609 #endif 2610 break; 2611 } 2612 2613 #ifdef _LP64 2614 case Op_CastPP: 2615 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2616 Node* in1 = n->in(1); 2617 const Type* t = n->bottom_type(); 2618 Node* new_in1 = in1->clone(); 2619 new_in1->as_DecodeN()->set_type(t); 2620 2621 if (!Matcher::narrow_oop_use_complex_address()) { 2622 // 2623 // x86, ARM and friends can handle 2 adds in addressing mode 2624 // and Matcher can fold a DecodeN node into address by using 2625 // a narrow oop directly and do implicit NULL check in address: 2626 // 2627 // [R12 + narrow_oop_reg<<3 + offset] 2628 // NullCheck narrow_oop_reg 2629 // 2630 // On other platforms (Sparc) we have to keep new DecodeN node and 2631 // use it to do implicit NULL check in address: 2632 // 2633 // decode_not_null narrow_oop_reg, base_reg 2634 // [base_reg + offset] 2635 // NullCheck base_reg 2636 // 2637 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2638 // to keep the information to which NULL check the new DecodeN node 2639 // corresponds to use it as value in implicit_null_check(). 2640 // 2641 new_in1->set_req(0, n->in(0)); 2642 } 2643 2644 n->subsume_by(new_in1, this); 2645 if (in1->outcnt() == 0) { 2646 in1->disconnect_inputs(NULL, this); 2647 } 2648 } 2649 break; 2650 2651 case Op_CmpP: 2652 // Do this transformation here to preserve CmpPNode::sub() and 2653 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2654 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2655 Node* in1 = n->in(1); 2656 Node* in2 = n->in(2); 2657 if (!in1->is_DecodeNarrowPtr()) { 2658 in2 = in1; 2659 in1 = n->in(2); 2660 } 2661 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2662 2663 Node* new_in2 = NULL; 2664 if (in2->is_DecodeNarrowPtr()) { 2665 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2666 new_in2 = in2->in(1); 2667 } else if (in2->Opcode() == Op_ConP) { 2668 const Type* t = in2->bottom_type(); 2669 if (t == TypePtr::NULL_PTR) { 2670 assert(in1->is_DecodeN(), "compare klass to null?"); 2671 // Don't convert CmpP null check into CmpN if compressed 2672 // oops implicit null check is not generated. 2673 // This will allow to generate normal oop implicit null check. 2674 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2675 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2676 // 2677 // This transformation together with CastPP transformation above 2678 // will generated code for implicit NULL checks for compressed oops. 2679 // 2680 // The original code after Optimize() 2681 // 2682 // LoadN memory, narrow_oop_reg 2683 // decode narrow_oop_reg, base_reg 2684 // CmpP base_reg, NULL 2685 // CastPP base_reg // NotNull 2686 // Load [base_reg + offset], val_reg 2687 // 2688 // after these transformations will be 2689 // 2690 // LoadN memory, narrow_oop_reg 2691 // CmpN narrow_oop_reg, NULL 2692 // decode_not_null narrow_oop_reg, base_reg 2693 // Load [base_reg + offset], val_reg 2694 // 2695 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2696 // since narrow oops can be used in debug info now (see the code in 2697 // final_graph_reshaping_walk()). 2698 // 2699 // At the end the code will be matched to 2700 // on x86: 2701 // 2702 // Load_narrow_oop memory, narrow_oop_reg 2703 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2704 // NullCheck narrow_oop_reg 2705 // 2706 // and on sparc: 2707 // 2708 // Load_narrow_oop memory, narrow_oop_reg 2709 // decode_not_null narrow_oop_reg, base_reg 2710 // Load [base_reg + offset], val_reg 2711 // NullCheck base_reg 2712 // 2713 } else if (t->isa_oopptr()) { 2714 new_in2 = ConNode::make(this, t->make_narrowoop()); 2715 } else if (t->isa_klassptr()) { 2716 new_in2 = ConNode::make(this, t->make_narrowklass()); 2717 } 2718 } 2719 if (new_in2 != NULL) { 2720 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 2721 n->subsume_by(cmpN, this); 2722 if (in1->outcnt() == 0) { 2723 in1->disconnect_inputs(NULL, this); 2724 } 2725 if (in2->outcnt() == 0) { 2726 in2->disconnect_inputs(NULL, this); 2727 } 2728 } 2729 } 2730 break; 2731 2732 case Op_DecodeN: 2733 case Op_DecodeNKlass: 2734 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2735 // DecodeN could be pinned when it can't be fold into 2736 // an address expression, see the code for Op_CastPP above. 2737 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2738 break; 2739 2740 case Op_EncodeP: 2741 case Op_EncodePKlass: { 2742 Node* in1 = n->in(1); 2743 if (in1->is_DecodeNarrowPtr()) { 2744 n->subsume_by(in1->in(1), this); 2745 } else if (in1->Opcode() == Op_ConP) { 2746 const Type* t = in1->bottom_type(); 2747 if (t == TypePtr::NULL_PTR) { 2748 assert(t->isa_oopptr(), "null klass?"); 2749 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2750 } else if (t->isa_oopptr()) { 2751 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2752 } else if (t->isa_klassptr()) { 2753 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2754 } 2755 } 2756 if (in1->outcnt() == 0) { 2757 in1->disconnect_inputs(NULL, this); 2758 } 2759 break; 2760 } 2761 2762 case Op_Proj: { 2763 if (OptimizeStringConcat) { 2764 ProjNode* p = n->as_Proj(); 2765 if (p->_is_io_use) { 2766 // Separate projections were used for the exception path which 2767 // are normally removed by a late inline. If it wasn't inlined 2768 // then they will hang around and should just be replaced with 2769 // the original one. 2770 Node* proj = NULL; 2771 // Replace with just one 2772 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2773 Node *use = i.get(); 2774 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2775 proj = use; 2776 break; 2777 } 2778 } 2779 assert(proj != NULL, "must be found"); 2780 p->subsume_by(proj, this); 2781 } 2782 } 2783 break; 2784 } 2785 2786 case Op_Phi: 2787 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2788 // The EncodeP optimization may create Phi with the same edges 2789 // for all paths. It is not handled well by Register Allocator. 2790 Node* unique_in = n->in(1); 2791 assert(unique_in != NULL, ""); 2792 uint cnt = n->req(); 2793 for (uint i = 2; i < cnt; i++) { 2794 Node* m = n->in(i); 2795 assert(m != NULL, ""); 2796 if (unique_in != m) 2797 unique_in = NULL; 2798 } 2799 if (unique_in != NULL) { 2800 n->subsume_by(unique_in, this); 2801 } 2802 } 2803 break; 2804 2805 #endif 2806 2807 case Op_ModI: 2808 if (UseDivMod) { 2809 // Check if a%b and a/b both exist 2810 Node* d = n->find_similar(Op_DivI); 2811 if (d) { 2812 // Replace them with a fused divmod if supported 2813 if (Matcher::has_match_rule(Op_DivModI)) { 2814 DivModINode* divmod = DivModINode::make(this, n); 2815 d->subsume_by(divmod->div_proj(), this); 2816 n->subsume_by(divmod->mod_proj(), this); 2817 } else { 2818 // replace a%b with a-((a/b)*b) 2819 Node* mult = new (this) MulINode(d, d->in(2)); 2820 Node* sub = new (this) SubINode(d->in(1), mult); 2821 n->subsume_by(sub, this); 2822 } 2823 } 2824 } 2825 break; 2826 2827 case Op_ModL: 2828 if (UseDivMod) { 2829 // Check if a%b and a/b both exist 2830 Node* d = n->find_similar(Op_DivL); 2831 if (d) { 2832 // Replace them with a fused divmod if supported 2833 if (Matcher::has_match_rule(Op_DivModL)) { 2834 DivModLNode* divmod = DivModLNode::make(this, n); 2835 d->subsume_by(divmod->div_proj(), this); 2836 n->subsume_by(divmod->mod_proj(), this); 2837 } else { 2838 // replace a%b with a-((a/b)*b) 2839 Node* mult = new (this) MulLNode(d, d->in(2)); 2840 Node* sub = new (this) SubLNode(d->in(1), mult); 2841 n->subsume_by(sub, this); 2842 } 2843 } 2844 } 2845 break; 2846 2847 case Op_LoadVector: 2848 case Op_StoreVector: 2849 break; 2850 2851 case Op_PackB: 2852 case Op_PackS: 2853 case Op_PackI: 2854 case Op_PackF: 2855 case Op_PackL: 2856 case Op_PackD: 2857 if (n->req()-1 > 2) { 2858 // Replace many operand PackNodes with a binary tree for matching 2859 PackNode* p = (PackNode*) n; 2860 Node* btp = p->binary_tree_pack(this, 1, n->req()); 2861 n->subsume_by(btp, this); 2862 } 2863 break; 2864 case Op_Loop: 2865 case Op_CountedLoop: 2866 if (n->as_Loop()->is_inner_loop()) { 2867 frc.inc_inner_loop_count(); 2868 } 2869 break; 2870 case Op_LShiftI: 2871 case Op_RShiftI: 2872 case Op_URShiftI: 2873 case Op_LShiftL: 2874 case Op_RShiftL: 2875 case Op_URShiftL: 2876 if (Matcher::need_masked_shift_count) { 2877 // The cpu's shift instructions don't restrict the count to the 2878 // lower 5/6 bits. We need to do the masking ourselves. 2879 Node* in2 = n->in(2); 2880 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 2881 const TypeInt* t = in2->find_int_type(); 2882 if (t != NULL && t->is_con()) { 2883 juint shift = t->get_con(); 2884 if (shift > mask) { // Unsigned cmp 2885 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 2886 } 2887 } else { 2888 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 2889 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 2890 n->set_req(2, shift); 2891 } 2892 } 2893 if (in2->outcnt() == 0) { // Remove dead node 2894 in2->disconnect_inputs(NULL, this); 2895 } 2896 } 2897 break; 2898 case Op_MemBarStoreStore: 2899 // Break the link with AllocateNode: it is no longer useful and 2900 // confuses register allocation. 2901 if (n->req() > MemBarNode::Precedent) { 2902 n->set_req(MemBarNode::Precedent, top()); 2903 } 2904 break; 2905 default: 2906 assert( !n->is_Call(), "" ); 2907 assert( !n->is_Mem(), "" ); 2908 break; 2909 } 2910 2911 // Collect CFG split points 2912 if (n->is_MultiBranch()) 2913 frc._tests.push(n); 2914 } 2915 2916 //------------------------------final_graph_reshaping_walk--------------------- 2917 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2918 // requires that the walk visits a node's inputs before visiting the node. 2919 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2920 ResourceArea *area = Thread::current()->resource_area(); 2921 Unique_Node_List sfpt(area); 2922 2923 frc._visited.set(root->_idx); // first, mark node as visited 2924 uint cnt = root->req(); 2925 Node *n = root; 2926 uint i = 0; 2927 while (true) { 2928 if (i < cnt) { 2929 // Place all non-visited non-null inputs onto stack 2930 Node* m = n->in(i); 2931 ++i; 2932 if (m != NULL && !frc._visited.test_set(m->_idx)) { 2933 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2934 sfpt.push(m); 2935 cnt = m->req(); 2936 nstack.push(n, i); // put on stack parent and next input's index 2937 n = m; 2938 i = 0; 2939 } 2940 } else { 2941 // Now do post-visit work 2942 final_graph_reshaping_impl( n, frc ); 2943 if (nstack.is_empty()) 2944 break; // finished 2945 n = nstack.node(); // Get node from stack 2946 cnt = n->req(); 2947 i = nstack.index(); 2948 nstack.pop(); // Shift to the next node on stack 2949 } 2950 } 2951 2952 // Skip next transformation if compressed oops are not used. 2953 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 2954 (!UseCompressedOops && !UseCompressedKlassPointers)) 2955 return; 2956 2957 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 2958 // It could be done for an uncommon traps or any safepoints/calls 2959 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 2960 while (sfpt.size() > 0) { 2961 n = sfpt.pop(); 2962 JVMState *jvms = n->as_SafePoint()->jvms(); 2963 assert(jvms != NULL, "sanity"); 2964 int start = jvms->debug_start(); 2965 int end = n->req(); 2966 bool is_uncommon = (n->is_CallStaticJava() && 2967 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2968 for (int j = start; j < end; j++) { 2969 Node* in = n->in(j); 2970 if (in->is_DecodeNarrowPtr()) { 2971 bool safe_to_skip = true; 2972 if (!is_uncommon ) { 2973 // Is it safe to skip? 2974 for (uint i = 0; i < in->outcnt(); i++) { 2975 Node* u = in->raw_out(i); 2976 if (!u->is_SafePoint() || 2977 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2978 safe_to_skip = false; 2979 } 2980 } 2981 } 2982 if (safe_to_skip) { 2983 n->set_req(j, in->in(1)); 2984 } 2985 if (in->outcnt() == 0) { 2986 in->disconnect_inputs(NULL, this); 2987 } 2988 } 2989 } 2990 } 2991 } 2992 2993 //------------------------------final_graph_reshaping-------------------------- 2994 // Final Graph Reshaping. 2995 // 2996 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2997 // and not commoned up and forced early. Must come after regular 2998 // optimizations to avoid GVN undoing the cloning. Clone constant 2999 // inputs to Loop Phis; these will be split by the allocator anyways. 3000 // Remove Opaque nodes. 3001 // (2) Move last-uses by commutative operations to the left input to encourage 3002 // Intel update-in-place two-address operations and better register usage 3003 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3004 // calls canonicalizing them back. 3005 // (3) Count the number of double-precision FP ops, single-precision FP ops 3006 // and call sites. On Intel, we can get correct rounding either by 3007 // forcing singles to memory (requires extra stores and loads after each 3008 // FP bytecode) or we can set a rounding mode bit (requires setting and 3009 // clearing the mode bit around call sites). The mode bit is only used 3010 // if the relative frequency of single FP ops to calls is low enough. 3011 // This is a key transform for SPEC mpeg_audio. 3012 // (4) Detect infinite loops; blobs of code reachable from above but not 3013 // below. Several of the Code_Gen algorithms fail on such code shapes, 3014 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3015 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3016 // Detection is by looking for IfNodes where only 1 projection is 3017 // reachable from below or CatchNodes missing some targets. 3018 // (5) Assert for insane oop offsets in debug mode. 3019 3020 bool Compile::final_graph_reshaping() { 3021 // an infinite loop may have been eliminated by the optimizer, 3022 // in which case the graph will be empty. 3023 if (root()->req() == 1) { 3024 record_method_not_compilable("trivial infinite loop"); 3025 return true; 3026 } 3027 3028 // Expensive nodes have their control input set to prevent the GVN 3029 // from freely commoning them. There's no GVN beyond this point so 3030 // no need to keep the control input. We want the expensive nodes to 3031 // be freely moved to the least frequent code path by gcm. 3032 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3033 for (int i = 0; i < expensive_count(); i++) { 3034 _expensive_nodes->at(i)->set_req(0, NULL); 3035 } 3036 3037 Final_Reshape_Counts frc; 3038 3039 // Visit everybody reachable! 3040 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3041 Node_Stack nstack(unique() >> 1); 3042 final_graph_reshaping_walk(nstack, root(), frc); 3043 3044 // Check for unreachable (from below) code (i.e., infinite loops). 3045 for( uint i = 0; i < frc._tests.size(); i++ ) { 3046 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3047 // Get number of CFG targets. 3048 // Note that PCTables include exception targets after calls. 3049 uint required_outcnt = n->required_outcnt(); 3050 if (n->outcnt() != required_outcnt) { 3051 // Check for a few special cases. Rethrow Nodes never take the 3052 // 'fall-thru' path, so expected kids is 1 less. 3053 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3054 if (n->in(0)->in(0)->is_Call()) { 3055 CallNode *call = n->in(0)->in(0)->as_Call(); 3056 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3057 required_outcnt--; // Rethrow always has 1 less kid 3058 } else if (call->req() > TypeFunc::Parms && 3059 call->is_CallDynamicJava()) { 3060 // Check for null receiver. In such case, the optimizer has 3061 // detected that the virtual call will always result in a null 3062 // pointer exception. The fall-through projection of this CatchNode 3063 // will not be populated. 3064 Node *arg0 = call->in(TypeFunc::Parms); 3065 if (arg0->is_Type() && 3066 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3067 required_outcnt--; 3068 } 3069 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3070 call->req() > TypeFunc::Parms+1 && 3071 call->is_CallStaticJava()) { 3072 // Check for negative array length. In such case, the optimizer has 3073 // detected that the allocation attempt will always result in an 3074 // exception. There is no fall-through projection of this CatchNode . 3075 Node *arg1 = call->in(TypeFunc::Parms+1); 3076 if (arg1->is_Type() && 3077 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3078 required_outcnt--; 3079 } 3080 } 3081 } 3082 } 3083 // Recheck with a better notion of 'required_outcnt' 3084 if (n->outcnt() != required_outcnt) { 3085 record_method_not_compilable("malformed control flow"); 3086 return true; // Not all targets reachable! 3087 } 3088 } 3089 // Check that I actually visited all kids. Unreached kids 3090 // must be infinite loops. 3091 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3092 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3093 record_method_not_compilable("infinite loop"); 3094 return true; // Found unvisited kid; must be unreach 3095 } 3096 } 3097 3098 // If original bytecodes contained a mixture of floats and doubles 3099 // check if the optimizer has made it homogenous, item (3). 3100 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3101 frc.get_float_count() > 32 && 3102 frc.get_double_count() == 0 && 3103 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3104 set_24_bit_selection_and_mode( false, true ); 3105 } 3106 3107 set_java_calls(frc.get_java_call_count()); 3108 set_inner_loops(frc.get_inner_loop_count()); 3109 3110 // No infinite loops, no reason to bail out. 3111 return false; 3112 } 3113 3114 //-----------------------------too_many_traps---------------------------------- 3115 // Report if there are too many traps at the current method and bci. 3116 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3117 bool Compile::too_many_traps(ciMethod* method, 3118 int bci, 3119 Deoptimization::DeoptReason reason) { 3120 ciMethodData* md = method->method_data(); 3121 if (md->is_empty()) { 3122 // Assume the trap has not occurred, or that it occurred only 3123 // because of a transient condition during start-up in the interpreter. 3124 return false; 3125 } 3126 if (md->has_trap_at(bci, reason) != 0) { 3127 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3128 // Also, if there are multiple reasons, or if there is no per-BCI record, 3129 // assume the worst. 3130 if (log()) 3131 log()->elem("observe trap='%s' count='%d'", 3132 Deoptimization::trap_reason_name(reason), 3133 md->trap_count(reason)); 3134 return true; 3135 } else { 3136 // Ignore method/bci and see if there have been too many globally. 3137 return too_many_traps(reason, md); 3138 } 3139 } 3140 3141 // Less-accurate variant which does not require a method and bci. 3142 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3143 ciMethodData* logmd) { 3144 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 3145 // Too many traps globally. 3146 // Note that we use cumulative trap_count, not just md->trap_count. 3147 if (log()) { 3148 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3149 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3150 Deoptimization::trap_reason_name(reason), 3151 mcount, trap_count(reason)); 3152 } 3153 return true; 3154 } else { 3155 // The coast is clear. 3156 return false; 3157 } 3158 } 3159 3160 //--------------------------too_many_recompiles-------------------------------- 3161 // Report if there are too many recompiles at the current method and bci. 3162 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3163 // Is not eager to return true, since this will cause the compiler to use 3164 // Action_none for a trap point, to avoid too many recompilations. 3165 bool Compile::too_many_recompiles(ciMethod* method, 3166 int bci, 3167 Deoptimization::DeoptReason reason) { 3168 ciMethodData* md = method->method_data(); 3169 if (md->is_empty()) { 3170 // Assume the trap has not occurred, or that it occurred only 3171 // because of a transient condition during start-up in the interpreter. 3172 return false; 3173 } 3174 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3175 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3176 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3177 Deoptimization::DeoptReason per_bc_reason 3178 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3179 if ((per_bc_reason == Deoptimization::Reason_none 3180 || md->has_trap_at(bci, reason) != 0) 3181 // The trap frequency measure we care about is the recompile count: 3182 && md->trap_recompiled_at(bci) 3183 && md->overflow_recompile_count() >= bc_cutoff) { 3184 // Do not emit a trap here if it has already caused recompilations. 3185 // Also, if there are multiple reasons, or if there is no per-BCI record, 3186 // assume the worst. 3187 if (log()) 3188 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3189 Deoptimization::trap_reason_name(reason), 3190 md->trap_count(reason), 3191 md->overflow_recompile_count()); 3192 return true; 3193 } else if (trap_count(reason) != 0 3194 && decompile_count() >= m_cutoff) { 3195 // Too many recompiles globally, and we have seen this sort of trap. 3196 // Use cumulative decompile_count, not just md->decompile_count. 3197 if (log()) 3198 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3199 Deoptimization::trap_reason_name(reason), 3200 md->trap_count(reason), trap_count(reason), 3201 md->decompile_count(), decompile_count()); 3202 return true; 3203 } else { 3204 // The coast is clear. 3205 return false; 3206 } 3207 } 3208 3209 3210 #ifndef PRODUCT 3211 //------------------------------verify_graph_edges--------------------------- 3212 // Walk the Graph and verify that there is a one-to-one correspondence 3213 // between Use-Def edges and Def-Use edges in the graph. 3214 void Compile::verify_graph_edges(bool no_dead_code) { 3215 if (VerifyGraphEdges) { 3216 ResourceArea *area = Thread::current()->resource_area(); 3217 Unique_Node_List visited(area); 3218 // Call recursive graph walk to check edges 3219 _root->verify_edges(visited); 3220 if (no_dead_code) { 3221 // Now make sure that no visited node is used by an unvisited node. 3222 bool dead_nodes = 0; 3223 Unique_Node_List checked(area); 3224 while (visited.size() > 0) { 3225 Node* n = visited.pop(); 3226 checked.push(n); 3227 for (uint i = 0; i < n->outcnt(); i++) { 3228 Node* use = n->raw_out(i); 3229 if (checked.member(use)) continue; // already checked 3230 if (visited.member(use)) continue; // already in the graph 3231 if (use->is_Con()) continue; // a dead ConNode is OK 3232 // At this point, we have found a dead node which is DU-reachable. 3233 if (dead_nodes++ == 0) 3234 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3235 use->dump(2); 3236 tty->print_cr("---"); 3237 checked.push(use); // No repeats; pretend it is now checked. 3238 } 3239 } 3240 assert(dead_nodes == 0, "using nodes must be reachable from root"); 3241 } 3242 } 3243 } 3244 #endif 3245 3246 // The Compile object keeps track of failure reasons separately from the ciEnv. 3247 // This is required because there is not quite a 1-1 relation between the 3248 // ciEnv and its compilation task and the Compile object. Note that one 3249 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3250 // to backtrack and retry without subsuming loads. Other than this backtracking 3251 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3252 // by the logic in C2Compiler. 3253 void Compile::record_failure(const char* reason) { 3254 if (log() != NULL) { 3255 log()->elem("failure reason='%s' phase='compile'", reason); 3256 } 3257 if (_failure_reason == NULL) { 3258 // Record the first failure reason. 3259 _failure_reason = reason; 3260 } 3261 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3262 C->print_method(_failure_reason); 3263 } 3264 _root = NULL; // flush the graph, too 3265 } 3266 3267 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3268 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3269 _phase_name(name), _dolog(dolog) 3270 { 3271 if (dolog) { 3272 C = Compile::current(); 3273 _log = C->log(); 3274 } else { 3275 C = NULL; 3276 _log = NULL; 3277 } 3278 if (_log != NULL) { 3279 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3280 _log->stamp(); 3281 _log->end_head(); 3282 } 3283 } 3284 3285 Compile::TracePhase::~TracePhase() { 3286 3287 C = Compile::current(); 3288 if (_dolog) { 3289 _log = C->log(); 3290 } else { 3291 _log = NULL; 3292 } 3293 3294 #ifdef ASSERT 3295 if (PrintIdealNodeCount) { 3296 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3297 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3298 } 3299 3300 if (VerifyIdealNodeCount) { 3301 Compile::current()->print_missing_nodes(); 3302 } 3303 #endif 3304 3305 if (_log != NULL) { 3306 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3307 } 3308 } 3309 3310 //============================================================================= 3311 // Two Constant's are equal when the type and the value are equal. 3312 bool Compile::Constant::operator==(const Constant& other) { 3313 if (type() != other.type() ) return false; 3314 if (can_be_reused() != other.can_be_reused()) return false; 3315 // For floating point values we compare the bit pattern. 3316 switch (type()) { 3317 case T_FLOAT: return (_v._value.i == other._v._value.i); 3318 case T_LONG: 3319 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3320 case T_OBJECT: 3321 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3322 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3323 case T_METADATA: return (_v._metadata == other._v._metadata); 3324 default: ShouldNotReachHere(); 3325 } 3326 return false; 3327 } 3328 3329 static int type_to_size_in_bytes(BasicType t) { 3330 switch (t) { 3331 case T_LONG: return sizeof(jlong ); 3332 case T_FLOAT: return sizeof(jfloat ); 3333 case T_DOUBLE: return sizeof(jdouble); 3334 case T_METADATA: return sizeof(Metadata*); 3335 // We use T_VOID as marker for jump-table entries (labels) which 3336 // need an internal word relocation. 3337 case T_VOID: 3338 case T_ADDRESS: 3339 case T_OBJECT: return sizeof(jobject); 3340 } 3341 3342 ShouldNotReachHere(); 3343 return -1; 3344 } 3345 3346 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3347 // sort descending 3348 if (a->freq() > b->freq()) return -1; 3349 if (a->freq() < b->freq()) return 1; 3350 return 0; 3351 } 3352 3353 void Compile::ConstantTable::calculate_offsets_and_size() { 3354 // First, sort the array by frequencies. 3355 _constants.sort(qsort_comparator); 3356 3357 #ifdef ASSERT 3358 // Make sure all jump-table entries were sorted to the end of the 3359 // array (they have a negative frequency). 3360 bool found_void = false; 3361 for (int i = 0; i < _constants.length(); i++) { 3362 Constant con = _constants.at(i); 3363 if (con.type() == T_VOID) 3364 found_void = true; // jump-tables 3365 else 3366 assert(!found_void, "wrong sorting"); 3367 } 3368 #endif 3369 3370 int offset = 0; 3371 for (int i = 0; i < _constants.length(); i++) { 3372 Constant* con = _constants.adr_at(i); 3373 3374 // Align offset for type. 3375 int typesize = type_to_size_in_bytes(con->type()); 3376 offset = align_size_up(offset, typesize); 3377 con->set_offset(offset); // set constant's offset 3378 3379 if (con->type() == T_VOID) { 3380 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3381 offset = offset + typesize * n->outcnt(); // expand jump-table 3382 } else { 3383 offset = offset + typesize; 3384 } 3385 } 3386 3387 // Align size up to the next section start (which is insts; see 3388 // CodeBuffer::align_at_start). 3389 assert(_size == -1, "already set?"); 3390 _size = align_size_up(offset, CodeEntryAlignment); 3391 } 3392 3393 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3394 MacroAssembler _masm(&cb); 3395 for (int i = 0; i < _constants.length(); i++) { 3396 Constant con = _constants.at(i); 3397 address constant_addr; 3398 switch (con.type()) { 3399 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3400 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3401 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3402 case T_OBJECT: { 3403 jobject obj = con.get_jobject(); 3404 int oop_index = _masm.oop_recorder()->find_index(obj); 3405 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3406 break; 3407 } 3408 case T_ADDRESS: { 3409 address addr = (address) con.get_jobject(); 3410 constant_addr = _masm.address_constant(addr); 3411 break; 3412 } 3413 // We use T_VOID as marker for jump-table entries (labels) which 3414 // need an internal word relocation. 3415 case T_VOID: { 3416 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3417 // Fill the jump-table with a dummy word. The real value is 3418 // filled in later in fill_jump_table. 3419 address dummy = (address) n; 3420 constant_addr = _masm.address_constant(dummy); 3421 // Expand jump-table 3422 for (uint i = 1; i < n->outcnt(); i++) { 3423 address temp_addr = _masm.address_constant(dummy + i); 3424 assert(temp_addr, "consts section too small"); 3425 } 3426 break; 3427 } 3428 case T_METADATA: { 3429 Metadata* obj = con.get_metadata(); 3430 int metadata_index = _masm.oop_recorder()->find_index(obj); 3431 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3432 break; 3433 } 3434 default: ShouldNotReachHere(); 3435 } 3436 assert(constant_addr, "consts section too small"); 3437 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3438 } 3439 } 3440 3441 int Compile::ConstantTable::find_offset(Constant& con) const { 3442 int idx = _constants.find(con); 3443 assert(idx != -1, "constant must be in constant table"); 3444 int offset = _constants.at(idx).offset(); 3445 assert(offset != -1, "constant table not emitted yet?"); 3446 return offset; 3447 } 3448 3449 void Compile::ConstantTable::add(Constant& con) { 3450 if (con.can_be_reused()) { 3451 int idx = _constants.find(con); 3452 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3453 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3454 return; 3455 } 3456 } 3457 (void) _constants.append(con); 3458 } 3459 3460 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3461 Block* b = Compile::current()->cfg()->_bbs[n->_idx]; 3462 Constant con(type, value, b->_freq); 3463 add(con); 3464 return con; 3465 } 3466 3467 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3468 Constant con(metadata); 3469 add(con); 3470 return con; 3471 } 3472 3473 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3474 jvalue value; 3475 BasicType type = oper->type()->basic_type(); 3476 switch (type) { 3477 case T_LONG: value.j = oper->constantL(); break; 3478 case T_FLOAT: value.f = oper->constantF(); break; 3479 case T_DOUBLE: value.d = oper->constantD(); break; 3480 case T_OBJECT: 3481 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3482 case T_METADATA: return add((Metadata*)oper->constant()); break; 3483 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3484 } 3485 return add(n, type, value); 3486 } 3487 3488 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3489 jvalue value; 3490 // We can use the node pointer here to identify the right jump-table 3491 // as this method is called from Compile::Fill_buffer right before 3492 // the MachNodes are emitted and the jump-table is filled (means the 3493 // MachNode pointers do not change anymore). 3494 value.l = (jobject) n; 3495 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3496 add(con); 3497 return con; 3498 } 3499 3500 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3501 // If called from Compile::scratch_emit_size do nothing. 3502 if (Compile::current()->in_scratch_emit_size()) return; 3503 3504 assert(labels.is_nonempty(), "must be"); 3505 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3506 3507 // Since MachConstantNode::constant_offset() also contains 3508 // table_base_offset() we need to subtract the table_base_offset() 3509 // to get the plain offset into the constant table. 3510 int offset = n->constant_offset() - table_base_offset(); 3511 3512 MacroAssembler _masm(&cb); 3513 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3514 3515 for (uint i = 0; i < n->outcnt(); i++) { 3516 address* constant_addr = &jump_table_base[i]; 3517 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i))); 3518 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3519 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3520 } 3521 } 3522 3523 void Compile::dump_inlining() { 3524 if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) { 3525 // Print inlining message for candidates that we couldn't inline 3526 // for lack of space or non constant receiver 3527 for (int i = 0; i < _late_inlines.length(); i++) { 3528 CallGenerator* cg = _late_inlines.at(i); 3529 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); 3530 } 3531 Unique_Node_List useful; 3532 useful.push(root()); 3533 for (uint next = 0; next < useful.size(); ++next) { 3534 Node* n = useful.at(next); 3535 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { 3536 CallNode* call = n->as_Call(); 3537 CallGenerator* cg = call->generator(); 3538 cg->print_inlining_late("receiver not constant"); 3539 } 3540 uint max = n->len(); 3541 for ( uint i = 0; i < max; ++i ) { 3542 Node *m = n->in(i); 3543 if ( m == NULL ) continue; 3544 useful.push(m); 3545 } 3546 } 3547 for (int i = 0; i < _print_inlining_list->length(); i++) { 3548 tty->print(_print_inlining_list->at(i).ss()->as_string()); 3549 } 3550 } 3551 } 3552 3553 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 3554 if (n1->Opcode() < n2->Opcode()) return -1; 3555 else if (n1->Opcode() > n2->Opcode()) return 1; 3556 3557 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())); 3558 for (uint i = 1; i < n1->req(); i++) { 3559 if (n1->in(i) < n2->in(i)) return -1; 3560 else if (n1->in(i) > n2->in(i)) return 1; 3561 } 3562 3563 return 0; 3564 } 3565 3566 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 3567 Node* n1 = *n1p; 3568 Node* n2 = *n2p; 3569 3570 return cmp_expensive_nodes(n1, n2); 3571 } 3572 3573 void Compile::sort_expensive_nodes() { 3574 if (!expensive_nodes_sorted()) { 3575 _expensive_nodes->sort(cmp_expensive_nodes); 3576 } 3577 } 3578 3579 bool Compile::expensive_nodes_sorted() const { 3580 for (int i = 1; i < _expensive_nodes->length(); i++) { 3581 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 3582 return false; 3583 } 3584 } 3585 return true; 3586 } 3587 3588 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 3589 if (_expensive_nodes->length() == 0) { 3590 return false; 3591 } 3592 3593 assert(OptimizeExpensiveOps, "optimization off?"); 3594 3595 // Take this opportunity to remove dead nodes from the list 3596 int j = 0; 3597 for (int i = 0; i < _expensive_nodes->length(); i++) { 3598 Node* n = _expensive_nodes->at(i); 3599 if (!n->is_unreachable(igvn)) { 3600 assert(n->is_expensive(), "should be expensive"); 3601 _expensive_nodes->at_put(j, n); 3602 j++; 3603 } 3604 } 3605 _expensive_nodes->trunc_to(j); 3606 3607 // Then sort the list so that similar nodes are next to each other 3608 // and check for at least two nodes of identical kind with same data 3609 // inputs. 3610 sort_expensive_nodes(); 3611 3612 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 3613 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 3614 return true; 3615 } 3616 } 3617 3618 return false; 3619 } 3620 3621 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 3622 if (_expensive_nodes->length() == 0) { 3623 return; 3624 } 3625 3626 assert(OptimizeExpensiveOps, "optimization off?"); 3627 3628 // Sort to bring similar nodes next to each other and clear the 3629 // control input of nodes for which there's only a single copy. 3630 sort_expensive_nodes(); 3631 3632 int j = 0; 3633 int identical = 0; 3634 int i = 0; 3635 for (; i < _expensive_nodes->length()-1; i++) { 3636 assert(j <= i, "can't write beyond current index"); 3637 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 3638 identical++; 3639 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3640 continue; 3641 } 3642 if (identical > 0) { 3643 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3644 identical = 0; 3645 } else { 3646 Node* n = _expensive_nodes->at(i); 3647 igvn.hash_delete(n); 3648 n->set_req(0, NULL); 3649 igvn.hash_insert(n); 3650 } 3651 } 3652 if (identical > 0) { 3653 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3654 } else if (_expensive_nodes->length() >= 1) { 3655 Node* n = _expensive_nodes->at(i); 3656 igvn.hash_delete(n); 3657 n->set_req(0, NULL); 3658 igvn.hash_insert(n); 3659 } 3660 _expensive_nodes->trunc_to(j); 3661 } 3662 3663 void Compile::add_expensive_node(Node * n) { 3664 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 3665 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 3666 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 3667 if (OptimizeExpensiveOps) { 3668 _expensive_nodes->append(n); 3669 } else { 3670 // Clear control input and let IGVN optimize expensive nodes if 3671 // OptimizeExpensiveOps is off. 3672 n->set_req(0, NULL); 3673 } 3674 } 3675 3676 // Auxiliary method to support randomized stressing/fuzzing. 3677 // 3678 // This method can be called the arbitrary number of times, with current count 3679 // as the argument. The logic allows selecting a single candidate from the 3680 // running list of candidates as follows: 3681 // int count = 0; 3682 // Cand* selected = null; 3683 // while(cand = cand->next()) { 3684 // if (randomized_select(++count)) { 3685 // selected = cand; 3686 // } 3687 // } 3688 // 3689 // Including count equalizes the chances any candidate is "selected". 3690 // This is useful when we don't have the complete list of candidates to choose 3691 // from uniformly. In this case, we need to adjust the randomicity of the 3692 // selection, or else we will end up biasing the selection towards the latter 3693 // candidates. 3694 // 3695 // Quick back-envelope calculation shows that for the list of n candidates 3696 // the equal probability for the candidate to persist as "best" can be 3697 // achieved by replacing it with "next" k-th candidate with the probability 3698 // of 1/k. It can be easily shown that by the end of the run, the 3699 // probability for any candidate is converged to 1/n, thus giving the 3700 // uniform distribution among all the candidates. 3701 // 3702 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 3703 #define RANDOMIZED_DOMAIN_POW 29 3704 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 3705 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 3706 bool Compile::randomized_select(int count) { 3707 assert(count > 0, "only positive"); 3708 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 3709 }