1 /* 2 * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileLog.hpp" 33 #include "compiler/disassembler.hpp" 34 #include "compiler/oopMap.hpp" 35 #include "opto/addnode.hpp" 36 #include "opto/block.hpp" 37 #include "opto/c2compiler.hpp" 38 #include "opto/callGenerator.hpp" 39 #include "opto/callnode.hpp" 40 #include "opto/cfgnode.hpp" 41 #include "opto/chaitin.hpp" 42 #include "opto/compile.hpp" 43 #include "opto/connode.hpp" 44 #include "opto/divnode.hpp" 45 #include "opto/escape.hpp" 46 #include "opto/idealGraphPrinter.hpp" 47 #include "opto/loopnode.hpp" 48 #include "opto/machnode.hpp" 49 #include "opto/macro.hpp" 50 #include "opto/matcher.hpp" 51 #include "opto/mathexactnode.hpp" 52 #include "opto/memnode.hpp" 53 #include "opto/mulnode.hpp" 54 #include "opto/narrowptrnode.hpp" 55 #include "opto/node.hpp" 56 #include "opto/opcodes.hpp" 57 #include "opto/output.hpp" 58 #include "opto/parse.hpp" 59 #include "opto/phaseX.hpp" 60 #include "opto/rootnode.hpp" 61 #include "opto/runtime.hpp" 62 #include "opto/stringopts.hpp" 63 #include "opto/type.hpp" 64 #include "opto/vectornode.hpp" 65 #include "runtime/arguments.hpp" 66 #include "runtime/signature.hpp" 67 #include "runtime/stubRoutines.hpp" 68 #include "runtime/timer.hpp" 69 #include "trace/tracing.hpp" 70 #include "utilities/copy.hpp" 71 #ifdef TARGET_ARCH_MODEL_x86_32 72 # include "adfiles/ad_x86_32.hpp" 73 #endif 74 #ifdef TARGET_ARCH_MODEL_x86_64 75 # include "adfiles/ad_x86_64.hpp" 76 #endif 77 #ifdef TARGET_ARCH_MODEL_sparc 78 # include "adfiles/ad_sparc.hpp" 79 #endif 80 #ifdef TARGET_ARCH_MODEL_zero 81 # include "adfiles/ad_zero.hpp" 82 #endif 83 #ifdef TARGET_ARCH_MODEL_arm 84 # include "adfiles/ad_arm.hpp" 85 #endif 86 #ifdef TARGET_ARCH_MODEL_ppc_32 87 # include "adfiles/ad_ppc_32.hpp" 88 #endif 89 #ifdef TARGET_ARCH_MODEL_ppc_64 90 # include "adfiles/ad_ppc_64.hpp" 91 #endif 92 93 94 // -------------------- Compile::mach_constant_base_node ----------------------- 95 // Constant table base node singleton. 96 MachConstantBaseNode* Compile::mach_constant_base_node() { 97 if (_mach_constant_base_node == NULL) { 98 _mach_constant_base_node = new MachConstantBaseNode(); 99 _mach_constant_base_node->add_req(C->root()); 100 } 101 return _mach_constant_base_node; 102 } 103 104 105 /// Support for intrinsics. 106 107 // Return the index at which m must be inserted (or already exists). 108 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 109 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 110 #ifdef ASSERT 111 for (int i = 1; i < _intrinsics->length(); i++) { 112 CallGenerator* cg1 = _intrinsics->at(i-1); 113 CallGenerator* cg2 = _intrinsics->at(i); 114 assert(cg1->method() != cg2->method() 115 ? cg1->method() < cg2->method() 116 : cg1->is_virtual() < cg2->is_virtual(), 117 "compiler intrinsics list must stay sorted"); 118 } 119 #endif 120 // Binary search sorted list, in decreasing intervals [lo, hi]. 121 int lo = 0, hi = _intrinsics->length()-1; 122 while (lo <= hi) { 123 int mid = (uint)(hi + lo) / 2; 124 ciMethod* mid_m = _intrinsics->at(mid)->method(); 125 if (m < mid_m) { 126 hi = mid-1; 127 } else if (m > mid_m) { 128 lo = mid+1; 129 } else { 130 // look at minor sort key 131 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 132 if (is_virtual < mid_virt) { 133 hi = mid-1; 134 } else if (is_virtual > mid_virt) { 135 lo = mid+1; 136 } else { 137 return mid; // exact match 138 } 139 } 140 } 141 return lo; // inexact match 142 } 143 144 void Compile::register_intrinsic(CallGenerator* cg) { 145 if (_intrinsics == NULL) { 146 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 147 } 148 // This code is stolen from ciObjectFactory::insert. 149 // Really, GrowableArray should have methods for 150 // insert_at, remove_at, and binary_search. 151 int len = _intrinsics->length(); 152 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 153 if (index == len) { 154 _intrinsics->append(cg); 155 } else { 156 #ifdef ASSERT 157 CallGenerator* oldcg = _intrinsics->at(index); 158 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 159 #endif 160 _intrinsics->append(_intrinsics->at(len-1)); 161 int pos; 162 for (pos = len-2; pos >= index; pos--) { 163 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 164 } 165 _intrinsics->at_put(index, cg); 166 } 167 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 168 } 169 170 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 171 assert(m->is_loaded(), "don't try this on unloaded methods"); 172 if (_intrinsics != NULL) { 173 int index = intrinsic_insertion_index(m, is_virtual); 174 if (index < _intrinsics->length() 175 && _intrinsics->at(index)->method() == m 176 && _intrinsics->at(index)->is_virtual() == is_virtual) { 177 return _intrinsics->at(index); 178 } 179 } 180 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 181 if (m->intrinsic_id() != vmIntrinsics::_none && 182 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 183 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 184 if (cg != NULL) { 185 // Save it for next time: 186 register_intrinsic(cg); 187 return cg; 188 } else { 189 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 190 } 191 } 192 return NULL; 193 } 194 195 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 196 // in library_call.cpp. 197 198 199 #ifndef PRODUCT 200 // statistics gathering... 201 202 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 203 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 204 205 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 206 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 207 int oflags = _intrinsic_hist_flags[id]; 208 assert(flags != 0, "what happened?"); 209 if (is_virtual) { 210 flags |= _intrinsic_virtual; 211 } 212 bool changed = (flags != oflags); 213 if ((flags & _intrinsic_worked) != 0) { 214 juint count = (_intrinsic_hist_count[id] += 1); 215 if (count == 1) { 216 changed = true; // first time 217 } 218 // increment the overall count also: 219 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 220 } 221 if (changed) { 222 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 223 // Something changed about the intrinsic's virtuality. 224 if ((flags & _intrinsic_virtual) != 0) { 225 // This is the first use of this intrinsic as a virtual call. 226 if (oflags != 0) { 227 // We already saw it as a non-virtual, so note both cases. 228 flags |= _intrinsic_both; 229 } 230 } else if ((oflags & _intrinsic_both) == 0) { 231 // This is the first use of this intrinsic as a non-virtual 232 flags |= _intrinsic_both; 233 } 234 } 235 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 236 } 237 // update the overall flags also: 238 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 239 return changed; 240 } 241 242 static char* format_flags(int flags, char* buf) { 243 buf[0] = 0; 244 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 245 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 246 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 247 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 248 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 249 if (buf[0] == 0) strcat(buf, ","); 250 assert(buf[0] == ',', "must be"); 251 return &buf[1]; 252 } 253 254 void Compile::print_intrinsic_statistics() { 255 char flagsbuf[100]; 256 ttyLocker ttyl; 257 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 258 tty->print_cr("Compiler intrinsic usage:"); 259 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 260 if (total == 0) total = 1; // avoid div0 in case of no successes 261 #define PRINT_STAT_LINE(name, c, f) \ 262 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 263 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 264 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 265 int flags = _intrinsic_hist_flags[id]; 266 juint count = _intrinsic_hist_count[id]; 267 if ((flags | count) != 0) { 268 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 269 } 270 } 271 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 272 if (xtty != NULL) xtty->tail("statistics"); 273 } 274 275 void Compile::print_statistics() { 276 { ttyLocker ttyl; 277 if (xtty != NULL) xtty->head("statistics type='opto'"); 278 Parse::print_statistics(); 279 PhaseCCP::print_statistics(); 280 PhaseRegAlloc::print_statistics(); 281 Scheduling::print_statistics(); 282 PhasePeephole::print_statistics(); 283 PhaseIdealLoop::print_statistics(); 284 if (xtty != NULL) xtty->tail("statistics"); 285 } 286 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 287 // put this under its own <statistics> element. 288 print_intrinsic_statistics(); 289 } 290 } 291 #endif //PRODUCT 292 293 // Support for bundling info 294 Bundle* Compile::node_bundling(const Node *n) { 295 assert(valid_bundle_info(n), "oob"); 296 return &_node_bundling_base[n->_idx]; 297 } 298 299 bool Compile::valid_bundle_info(const Node *n) { 300 return (_node_bundling_limit > n->_idx); 301 } 302 303 304 void Compile::gvn_replace_by(Node* n, Node* nn) { 305 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 306 Node* use = n->last_out(i); 307 bool is_in_table = initial_gvn()->hash_delete(use); 308 uint uses_found = 0; 309 for (uint j = 0; j < use->len(); j++) { 310 if (use->in(j) == n) { 311 if (j < use->req()) 312 use->set_req(j, nn); 313 else 314 use->set_prec(j, nn); 315 uses_found++; 316 } 317 } 318 if (is_in_table) { 319 // reinsert into table 320 initial_gvn()->hash_find_insert(use); 321 } 322 record_for_igvn(use); 323 i -= uses_found; // we deleted 1 or more copies of this edge 324 } 325 } 326 327 328 static inline bool not_a_node(const Node* n) { 329 if (n == NULL) return true; 330 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 331 if (*(address*)n == badAddress) return true; // kill by Node::destruct 332 return false; 333 } 334 335 // Identify all nodes that are reachable from below, useful. 336 // Use breadth-first pass that records state in a Unique_Node_List, 337 // recursive traversal is slower. 338 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 339 int estimated_worklist_size = unique(); 340 useful.map( estimated_worklist_size, NULL ); // preallocate space 341 342 // Initialize worklist 343 if (root() != NULL) { useful.push(root()); } 344 // If 'top' is cached, declare it useful to preserve cached node 345 if( cached_top_node() ) { useful.push(cached_top_node()); } 346 347 // Push all useful nodes onto the list, breadthfirst 348 for( uint next = 0; next < useful.size(); ++next ) { 349 assert( next < unique(), "Unique useful nodes < total nodes"); 350 Node *n = useful.at(next); 351 uint max = n->len(); 352 for( uint i = 0; i < max; ++i ) { 353 Node *m = n->in(i); 354 if (not_a_node(m)) continue; 355 useful.push(m); 356 } 357 } 358 } 359 360 // Update dead_node_list with any missing dead nodes using useful 361 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 362 void Compile::update_dead_node_list(Unique_Node_List &useful) { 363 uint max_idx = unique(); 364 VectorSet& useful_node_set = useful.member_set(); 365 366 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 367 // If node with index node_idx is not in useful set, 368 // mark it as dead in dead node list. 369 if (! useful_node_set.test(node_idx) ) { 370 record_dead_node(node_idx); 371 } 372 } 373 } 374 375 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 376 int shift = 0; 377 for (int i = 0; i < inlines->length(); i++) { 378 CallGenerator* cg = inlines->at(i); 379 CallNode* call = cg->call_node(); 380 if (shift > 0) { 381 inlines->at_put(i-shift, cg); 382 } 383 if (!useful.member(call)) { 384 shift++; 385 } 386 } 387 inlines->trunc_to(inlines->length()-shift); 388 } 389 390 // Disconnect all useless nodes by disconnecting those at the boundary. 391 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 392 uint next = 0; 393 while (next < useful.size()) { 394 Node *n = useful.at(next++); 395 // Use raw traversal of out edges since this code removes out edges 396 int max = n->outcnt(); 397 for (int j = 0; j < max; ++j) { 398 Node* child = n->raw_out(j); 399 if (! useful.member(child)) { 400 assert(!child->is_top() || child != top(), 401 "If top is cached in Compile object it is in useful list"); 402 // Only need to remove this out-edge to the useless node 403 n->raw_del_out(j); 404 --j; 405 --max; 406 } 407 } 408 if (n->outcnt() == 1 && n->has_special_unique_user()) { 409 record_for_igvn(n->unique_out()); 410 } 411 } 412 // Remove useless macro and predicate opaq nodes 413 for (int i = C->macro_count()-1; i >= 0; i--) { 414 Node* n = C->macro_node(i); 415 if (!useful.member(n)) { 416 remove_macro_node(n); 417 } 418 } 419 // Remove useless expensive node 420 for (int i = C->expensive_count()-1; i >= 0; i--) { 421 Node* n = C->expensive_node(i); 422 if (!useful.member(n)) { 423 remove_expensive_node(n); 424 } 425 } 426 // clean up the late inline lists 427 remove_useless_late_inlines(&_string_late_inlines, useful); 428 remove_useless_late_inlines(&_boxing_late_inlines, useful); 429 remove_useless_late_inlines(&_late_inlines, useful); 430 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 431 } 432 433 //------------------------------frame_size_in_words----------------------------- 434 // frame_slots in units of words 435 int Compile::frame_size_in_words() const { 436 // shift is 0 in LP32 and 1 in LP64 437 const int shift = (LogBytesPerWord - LogBytesPerInt); 438 int words = _frame_slots >> shift; 439 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 440 return words; 441 } 442 443 // To bang the stack of this compiled method we use the stack size 444 // that the interpreter would need in case of a deoptimization. This 445 // removes the need to bang the stack in the deoptimization blob which 446 // in turn simplifies stack overflow handling. 447 int Compile::bang_size_in_bytes() const { 448 return MAX2(_interpreter_frame_size, frame_size_in_bytes()); 449 } 450 451 // ============================================================================ 452 //------------------------------CompileWrapper--------------------------------- 453 class CompileWrapper : public StackObj { 454 Compile *const _compile; 455 public: 456 CompileWrapper(Compile* compile); 457 458 ~CompileWrapper(); 459 }; 460 461 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 462 // the Compile* pointer is stored in the current ciEnv: 463 ciEnv* env = compile->env(); 464 assert(env == ciEnv::current(), "must already be a ciEnv active"); 465 assert(env->compiler_data() == NULL, "compile already active?"); 466 env->set_compiler_data(compile); 467 assert(compile == Compile::current(), "sanity"); 468 469 compile->set_type_dict(NULL); 470 compile->set_type_hwm(NULL); 471 compile->set_type_last_size(0); 472 compile->set_last_tf(NULL, NULL); 473 compile->set_indexSet_arena(NULL); 474 compile->set_indexSet_free_block_list(NULL); 475 compile->init_type_arena(); 476 Type::Initialize(compile); 477 _compile->set_scratch_buffer_blob(NULL); 478 _compile->begin_method(); 479 } 480 CompileWrapper::~CompileWrapper() { 481 _compile->end_method(); 482 if (_compile->scratch_buffer_blob() != NULL) 483 BufferBlob::free(_compile->scratch_buffer_blob()); 484 _compile->env()->set_compiler_data(NULL); 485 } 486 487 488 //----------------------------print_compile_messages--------------------------- 489 void Compile::print_compile_messages() { 490 #ifndef PRODUCT 491 // Check if recompiling 492 if (_subsume_loads == false && PrintOpto) { 493 // Recompiling without allowing machine instructions to subsume loads 494 tty->print_cr("*********************************************************"); 495 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 496 tty->print_cr("*********************************************************"); 497 } 498 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 499 // Recompiling without escape analysis 500 tty->print_cr("*********************************************************"); 501 tty->print_cr("** Bailout: Recompile without escape analysis **"); 502 tty->print_cr("*********************************************************"); 503 } 504 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 505 // Recompiling without boxing elimination 506 tty->print_cr("*********************************************************"); 507 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 508 tty->print_cr("*********************************************************"); 509 } 510 if (env()->break_at_compile()) { 511 // Open the debugger when compiling this method. 512 tty->print("### Breaking when compiling: "); 513 method()->print_short_name(); 514 tty->cr(); 515 BREAKPOINT; 516 } 517 518 if( PrintOpto ) { 519 if (is_osr_compilation()) { 520 tty->print("[OSR]%3d", _compile_id); 521 } else { 522 tty->print("%3d", _compile_id); 523 } 524 } 525 #endif 526 } 527 528 529 //-----------------------init_scratch_buffer_blob------------------------------ 530 // Construct a temporary BufferBlob and cache it for this compile. 531 void Compile::init_scratch_buffer_blob(int const_size) { 532 // If there is already a scratch buffer blob allocated and the 533 // constant section is big enough, use it. Otherwise free the 534 // current and allocate a new one. 535 BufferBlob* blob = scratch_buffer_blob(); 536 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 537 // Use the current blob. 538 } else { 539 if (blob != NULL) { 540 BufferBlob::free(blob); 541 } 542 543 ResourceMark rm; 544 _scratch_const_size = const_size; 545 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 546 blob = BufferBlob::create("Compile::scratch_buffer", size); 547 // Record the buffer blob for next time. 548 set_scratch_buffer_blob(blob); 549 // Have we run out of code space? 550 if (scratch_buffer_blob() == NULL) { 551 // Let CompilerBroker disable further compilations. 552 record_failure("Not enough space for scratch buffer in CodeCache"); 553 return; 554 } 555 } 556 557 // Initialize the relocation buffers 558 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 559 set_scratch_locs_memory(locs_buf); 560 } 561 562 563 //-----------------------scratch_emit_size------------------------------------- 564 // Helper function that computes size by emitting code 565 uint Compile::scratch_emit_size(const Node* n) { 566 // Start scratch_emit_size section. 567 set_in_scratch_emit_size(true); 568 569 // Emit into a trash buffer and count bytes emitted. 570 // This is a pretty expensive way to compute a size, 571 // but it works well enough if seldom used. 572 // All common fixed-size instructions are given a size 573 // method by the AD file. 574 // Note that the scratch buffer blob and locs memory are 575 // allocated at the beginning of the compile task, and 576 // may be shared by several calls to scratch_emit_size. 577 // The allocation of the scratch buffer blob is particularly 578 // expensive, since it has to grab the code cache lock. 579 BufferBlob* blob = this->scratch_buffer_blob(); 580 assert(blob != NULL, "Initialize BufferBlob at start"); 581 assert(blob->size() > MAX_inst_size, "sanity"); 582 relocInfo* locs_buf = scratch_locs_memory(); 583 address blob_begin = blob->content_begin(); 584 address blob_end = (address)locs_buf; 585 assert(blob->content_contains(blob_end), "sanity"); 586 CodeBuffer buf(blob_begin, blob_end - blob_begin); 587 buf.initialize_consts_size(_scratch_const_size); 588 buf.initialize_stubs_size(MAX_stubs_size); 589 assert(locs_buf != NULL, "sanity"); 590 int lsize = MAX_locs_size / 3; 591 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 592 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 593 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 594 595 // Do the emission. 596 597 Label fakeL; // Fake label for branch instructions. 598 Label* saveL = NULL; 599 uint save_bnum = 0; 600 bool is_branch = n->is_MachBranch(); 601 if (is_branch) { 602 MacroAssembler masm(&buf); 603 masm.bind(fakeL); 604 n->as_MachBranch()->save_label(&saveL, &save_bnum); 605 n->as_MachBranch()->label_set(&fakeL, 0); 606 } 607 n->emit(buf, this->regalloc()); 608 if (is_branch) // Restore label. 609 n->as_MachBranch()->label_set(saveL, save_bnum); 610 611 // End scratch_emit_size section. 612 set_in_scratch_emit_size(false); 613 614 return buf.insts_size(); 615 } 616 617 618 // ============================================================================ 619 //------------------------------Compile standard------------------------------- 620 debug_only( int Compile::_debug_idx = 100000; ) 621 622 // Compile a method. entry_bci is -1 for normal compilations and indicates 623 // the continuation bci for on stack replacement. 624 625 626 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 627 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 628 : Phase(Compiler), 629 _env(ci_env), 630 _log(ci_env->log()), 631 _compile_id(ci_env->compile_id()), 632 _save_argument_registers(false), 633 _stub_name(NULL), 634 _stub_function(NULL), 635 _stub_entry_point(NULL), 636 _method(target), 637 _entry_bci(osr_bci), 638 _initial_gvn(NULL), 639 _for_igvn(NULL), 640 _warm_calls(NULL), 641 _subsume_loads(subsume_loads), 642 _do_escape_analysis(do_escape_analysis), 643 _eliminate_boxing(eliminate_boxing), 644 _failure_reason(NULL), 645 _code_buffer("Compile::Fill_buffer"), 646 _orig_pc_slot(0), 647 _orig_pc_slot_offset_in_bytes(0), 648 _has_method_handle_invokes(false), 649 _mach_constant_base_node(NULL), 650 _node_bundling_limit(0), 651 _node_bundling_base(NULL), 652 _java_calls(0), 653 _inner_loops(0), 654 _scratch_const_size(-1), 655 _in_scratch_emit_size(false), 656 _dead_node_list(comp_arena()), 657 _dead_node_count(0), 658 #ifndef PRODUCT 659 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 660 _in_dump_cnt(0), 661 _printer(IdealGraphPrinter::printer()), 662 #endif 663 _congraph(NULL), 664 _replay_inline_data(NULL), 665 _late_inlines(comp_arena(), 2, 0, NULL), 666 _string_late_inlines(comp_arena(), 2, 0, NULL), 667 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 668 _late_inlines_pos(0), 669 _number_of_mh_late_inlines(0), 670 _inlining_progress(false), 671 _inlining_incrementally(false), 672 _print_inlining_list(NULL), 673 _print_inlining_stream(NULL), 674 _print_inlining_idx(0), 675 _print_inlining_output(NULL), 676 _preserve_jvm_state(0), 677 _interpreter_frame_size(0) { 678 C = this; 679 680 CompileWrapper cw(this); 681 #ifndef PRODUCT 682 if (TimeCompiler2) { 683 tty->print(" "); 684 target->holder()->name()->print(); 685 tty->print("."); 686 target->print_short_name(); 687 tty->print(" "); 688 } 689 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 690 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 691 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 692 if (!print_opto_assembly) { 693 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 694 if (print_assembly && !Disassembler::can_decode()) { 695 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 696 print_opto_assembly = true; 697 } 698 } 699 set_print_assembly(print_opto_assembly); 700 set_parsed_irreducible_loop(false); 701 702 if (method()->has_option("ReplayInline")) { 703 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 704 } 705 #endif 706 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 707 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 708 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 709 710 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 711 // Make sure the method being compiled gets its own MDO, 712 // so we can at least track the decompile_count(). 713 // Need MDO to record RTM code generation state. 714 method()->ensure_method_data(); 715 } 716 717 Init(::AliasLevel); 718 719 720 print_compile_messages(); 721 722 _ilt = InlineTree::build_inline_tree_root(); 723 724 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 725 assert(num_alias_types() >= AliasIdxRaw, ""); 726 727 #define MINIMUM_NODE_HASH 1023 728 // Node list that Iterative GVN will start with 729 Unique_Node_List for_igvn(comp_arena()); 730 set_for_igvn(&for_igvn); 731 732 // GVN that will be run immediately on new nodes 733 uint estimated_size = method()->code_size()*4+64; 734 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 735 PhaseGVN gvn(node_arena(), estimated_size); 736 set_initial_gvn(&gvn); 737 738 print_inlining_init(); 739 { // Scope for timing the parser 740 TracePhase t3("parse", &_t_parser, true); 741 742 // Put top into the hash table ASAP. 743 initial_gvn()->transform_no_reclaim(top()); 744 745 // Set up tf(), start(), and find a CallGenerator. 746 CallGenerator* cg = NULL; 747 if (is_osr_compilation()) { 748 const TypeTuple *domain = StartOSRNode::osr_domain(); 749 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 750 init_tf(TypeFunc::make(domain, range)); 751 StartNode* s = new StartOSRNode(root(), domain); 752 initial_gvn()->set_type_bottom(s); 753 init_start(s); 754 cg = CallGenerator::for_osr(method(), entry_bci()); 755 } else { 756 // Normal case. 757 init_tf(TypeFunc::make(method())); 758 StartNode* s = new StartNode(root(), tf()->domain()); 759 initial_gvn()->set_type_bottom(s); 760 init_start(s); 761 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 762 // With java.lang.ref.reference.get() we must go through the 763 // intrinsic when G1 is enabled - even when get() is the root 764 // method of the compile - so that, if necessary, the value in 765 // the referent field of the reference object gets recorded by 766 // the pre-barrier code. 767 // Specifically, if G1 is enabled, the value in the referent 768 // field is recorded by the G1 SATB pre barrier. This will 769 // result in the referent being marked live and the reference 770 // object removed from the list of discovered references during 771 // reference processing. 772 cg = find_intrinsic(method(), false); 773 } 774 if (cg == NULL) { 775 float past_uses = method()->interpreter_invocation_count(); 776 float expected_uses = past_uses; 777 cg = CallGenerator::for_inline(method(), expected_uses); 778 } 779 } 780 if (failing()) return; 781 if (cg == NULL) { 782 record_method_not_compilable_all_tiers("cannot parse method"); 783 return; 784 } 785 JVMState* jvms = build_start_state(start(), tf()); 786 if ((jvms = cg->generate(jvms, NULL)) == NULL) { 787 record_method_not_compilable("method parse failed"); 788 return; 789 } 790 GraphKit kit(jvms); 791 792 if (!kit.stopped()) { 793 // Accept return values, and transfer control we know not where. 794 // This is done by a special, unique ReturnNode bound to root. 795 return_values(kit.jvms()); 796 } 797 798 if (kit.has_exceptions()) { 799 // Any exceptions that escape from this call must be rethrown 800 // to whatever caller is dynamically above us on the stack. 801 // This is done by a special, unique RethrowNode bound to root. 802 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 803 } 804 805 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 806 807 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 808 inline_string_calls(true); 809 } 810 811 if (failing()) return; 812 813 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 814 815 // Remove clutter produced by parsing. 816 if (!failing()) { 817 ResourceMark rm; 818 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 819 } 820 } 821 822 // Note: Large methods are capped off in do_one_bytecode(). 823 if (failing()) return; 824 825 // After parsing, node notes are no longer automagic. 826 // They must be propagated by register_new_node_with_optimizer(), 827 // clone(), or the like. 828 set_default_node_notes(NULL); 829 830 for (;;) { 831 int successes = Inline_Warm(); 832 if (failing()) return; 833 if (successes == 0) break; 834 } 835 836 // Drain the list. 837 Finish_Warm(); 838 #ifndef PRODUCT 839 if (_printer) { 840 _printer->print_inlining(this); 841 } 842 #endif 843 844 if (failing()) return; 845 NOT_PRODUCT( verify_graph_edges(); ) 846 847 // Now optimize 848 Optimize(); 849 if (failing()) return; 850 NOT_PRODUCT( verify_graph_edges(); ) 851 852 #ifndef PRODUCT 853 if (PrintIdeal) { 854 ttyLocker ttyl; // keep the following output all in one block 855 // This output goes directly to the tty, not the compiler log. 856 // To enable tools to match it up with the compilation activity, 857 // be sure to tag this tty output with the compile ID. 858 if (xtty != NULL) { 859 xtty->head("ideal compile_id='%d'%s", compile_id(), 860 is_osr_compilation() ? " compile_kind='osr'" : 861 ""); 862 } 863 root()->dump(9999); 864 if (xtty != NULL) { 865 xtty->tail("ideal"); 866 } 867 } 868 #endif 869 870 NOT_PRODUCT( verify_barriers(); ) 871 872 // Dump compilation data to replay it. 873 if (method()->has_option("DumpReplay")) { 874 env()->dump_replay_data(_compile_id); 875 } 876 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 877 env()->dump_inline_data(_compile_id); 878 } 879 880 // Now that we know the size of all the monitors we can add a fixed slot 881 // for the original deopt pc. 882 883 _orig_pc_slot = fixed_slots(); 884 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 885 set_fixed_slots(next_slot); 886 887 // Compute when to use implicit null checks. Used by matching trap based 888 // nodes and NullCheck optimization. 889 set_allowed_deopt_reasons(); 890 891 // Now generate code 892 Code_Gen(); 893 if (failing()) return; 894 895 // Check if we want to skip execution of all compiled code. 896 { 897 #ifndef PRODUCT 898 if (OptoNoExecute) { 899 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 900 return; 901 } 902 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 903 #endif 904 905 if (is_osr_compilation()) { 906 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 907 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 908 } else { 909 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 910 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 911 } 912 913 env()->register_method(_method, _entry_bci, 914 &_code_offsets, 915 _orig_pc_slot_offset_in_bytes, 916 code_buffer(), 917 frame_size_in_words(), _oop_map_set, 918 &_handler_table, &_inc_table, 919 compiler, 920 env()->comp_level(), 921 has_unsafe_access(), 922 SharedRuntime::is_wide_vector(max_vector_size()), 923 rtm_state() 924 ); 925 926 if (log() != NULL) // Print code cache state into compiler log 927 log()->code_cache_state(); 928 } 929 } 930 931 //------------------------------Compile---------------------------------------- 932 // Compile a runtime stub 933 Compile::Compile( ciEnv* ci_env, 934 TypeFunc_generator generator, 935 address stub_function, 936 const char *stub_name, 937 int is_fancy_jump, 938 bool pass_tls, 939 bool save_arg_registers, 940 bool return_pc ) 941 : Phase(Compiler), 942 _env(ci_env), 943 _log(ci_env->log()), 944 _compile_id(0), 945 _save_argument_registers(save_arg_registers), 946 _method(NULL), 947 _stub_name(stub_name), 948 _stub_function(stub_function), 949 _stub_entry_point(NULL), 950 _entry_bci(InvocationEntryBci), 951 _initial_gvn(NULL), 952 _for_igvn(NULL), 953 _warm_calls(NULL), 954 _orig_pc_slot(0), 955 _orig_pc_slot_offset_in_bytes(0), 956 _subsume_loads(true), 957 _do_escape_analysis(false), 958 _eliminate_boxing(false), 959 _failure_reason(NULL), 960 _code_buffer("Compile::Fill_buffer"), 961 _has_method_handle_invokes(false), 962 _mach_constant_base_node(NULL), 963 _node_bundling_limit(0), 964 _node_bundling_base(NULL), 965 _java_calls(0), 966 _inner_loops(0), 967 #ifndef PRODUCT 968 _trace_opto_output(TraceOptoOutput), 969 _in_dump_cnt(0), 970 _printer(NULL), 971 #endif 972 _dead_node_list(comp_arena()), 973 _dead_node_count(0), 974 _congraph(NULL), 975 _replay_inline_data(NULL), 976 _number_of_mh_late_inlines(0), 977 _inlining_progress(false), 978 _inlining_incrementally(false), 979 _print_inlining_list(NULL), 980 _print_inlining_stream(NULL), 981 _print_inlining_idx(0), 982 _print_inlining_output(NULL), 983 _preserve_jvm_state(0), 984 _allowed_reasons(0), 985 _interpreter_frame_size(0) { 986 C = this; 987 988 #ifndef PRODUCT 989 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 990 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 991 set_print_assembly(PrintFrameConverterAssembly); 992 set_parsed_irreducible_loop(false); 993 #endif 994 set_has_irreducible_loop(false); // no loops 995 996 CompileWrapper cw(this); 997 Init(/*AliasLevel=*/ 0); 998 init_tf((*generator)()); 999 1000 { 1001 // The following is a dummy for the sake of GraphKit::gen_stub 1002 Unique_Node_List for_igvn(comp_arena()); 1003 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1004 PhaseGVN gvn(Thread::current()->resource_area(),255); 1005 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1006 gvn.transform_no_reclaim(top()); 1007 1008 GraphKit kit; 1009 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1010 } 1011 1012 NOT_PRODUCT( verify_graph_edges(); ) 1013 Code_Gen(); 1014 if (failing()) return; 1015 1016 1017 // Entry point will be accessed using compile->stub_entry_point(); 1018 if (code_buffer() == NULL) { 1019 Matcher::soft_match_failure(); 1020 } else { 1021 if (PrintAssembly && (WizardMode || Verbose)) 1022 tty->print_cr("### Stub::%s", stub_name); 1023 1024 if (!failing()) { 1025 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1026 1027 // Make the NMethod 1028 // For now we mark the frame as never safe for profile stackwalking 1029 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1030 code_buffer(), 1031 CodeOffsets::frame_never_safe, 1032 // _code_offsets.value(CodeOffsets::Frame_Complete), 1033 frame_size_in_words(), 1034 _oop_map_set, 1035 save_arg_registers); 1036 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1037 1038 _stub_entry_point = rs->entry_point(); 1039 } 1040 } 1041 } 1042 1043 //------------------------------Init------------------------------------------- 1044 // Prepare for a single compilation 1045 void Compile::Init(int aliaslevel) { 1046 _unique = 0; 1047 _regalloc = NULL; 1048 1049 _tf = NULL; // filled in later 1050 _top = NULL; // cached later 1051 _matcher = NULL; // filled in later 1052 _cfg = NULL; // filled in later 1053 1054 set_24_bit_selection_and_mode(Use24BitFP, false); 1055 1056 _node_note_array = NULL; 1057 _default_node_notes = NULL; 1058 1059 _immutable_memory = NULL; // filled in at first inquiry 1060 1061 // Globally visible Nodes 1062 // First set TOP to NULL to give safe behavior during creation of RootNode 1063 set_cached_top_node(NULL); 1064 set_root(new RootNode()); 1065 // Now that you have a Root to point to, create the real TOP 1066 set_cached_top_node( new ConNode(Type::TOP) ); 1067 set_recent_alloc(NULL, NULL); 1068 1069 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1070 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1071 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1072 env()->set_dependencies(new Dependencies(env())); 1073 1074 _fixed_slots = 0; 1075 set_has_split_ifs(false); 1076 set_has_loops(has_method() && method()->has_loops()); // first approximation 1077 set_has_stringbuilder(false); 1078 set_has_boxed_value(false); 1079 _trap_can_recompile = false; // no traps emitted yet 1080 _major_progress = true; // start out assuming good things will happen 1081 set_has_unsafe_access(false); 1082 set_max_vector_size(0); 1083 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1084 set_decompile_count(0); 1085 1086 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1087 set_num_loop_opts(LoopOptsCount); 1088 set_do_inlining(Inline); 1089 set_max_inline_size(MaxInlineSize); 1090 set_freq_inline_size(FreqInlineSize); 1091 set_do_scheduling(OptoScheduling); 1092 set_do_count_invocations(false); 1093 set_do_method_data_update(false); 1094 set_age_code(has_method() && method()->profile_aging()); 1095 set_rtm_state(NoRTM); // No RTM lock eliding by default 1096 #if INCLUDE_RTM_OPT 1097 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1098 int rtm_state = method()->method_data()->rtm_state(); 1099 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1100 // Don't generate RTM lock eliding code. 1101 set_rtm_state(NoRTM); 1102 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1103 // Generate RTM lock eliding code without abort ratio calculation code. 1104 set_rtm_state(UseRTM); 1105 } else if (UseRTMDeopt) { 1106 // Generate RTM lock eliding code and include abort ratio calculation 1107 // code if UseRTMDeopt is on. 1108 set_rtm_state(ProfileRTM); 1109 } 1110 } 1111 #endif 1112 if (debug_info()->recording_non_safepoints()) { 1113 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1114 (comp_arena(), 8, 0, NULL)); 1115 set_default_node_notes(Node_Notes::make(this)); 1116 } 1117 1118 // // -- Initialize types before each compile -- 1119 // // Update cached type information 1120 // if( _method && _method->constants() ) 1121 // Type::update_loaded_types(_method, _method->constants()); 1122 1123 // Init alias_type map. 1124 if (!_do_escape_analysis && aliaslevel == 3) 1125 aliaslevel = 2; // No unique types without escape analysis 1126 _AliasLevel = aliaslevel; 1127 const int grow_ats = 16; 1128 _max_alias_types = grow_ats; 1129 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1130 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1131 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1132 { 1133 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1134 } 1135 // Initialize the first few types. 1136 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1137 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1138 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1139 _num_alias_types = AliasIdxRaw+1; 1140 // Zero out the alias type cache. 1141 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1142 // A NULL adr_type hits in the cache right away. Preload the right answer. 1143 probe_alias_cache(NULL)->_index = AliasIdxTop; 1144 1145 _intrinsics = NULL; 1146 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1147 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1148 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1149 register_library_intrinsics(); 1150 } 1151 1152 //---------------------------init_start---------------------------------------- 1153 // Install the StartNode on this compile object. 1154 void Compile::init_start(StartNode* s) { 1155 if (failing()) 1156 return; // already failing 1157 assert(s == start(), ""); 1158 } 1159 1160 StartNode* Compile::start() const { 1161 assert(!failing(), ""); 1162 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1163 Node* start = root()->fast_out(i); 1164 if( start->is_Start() ) 1165 return start->as_Start(); 1166 } 1167 fatal("Did not find Start node!"); 1168 return NULL; 1169 } 1170 1171 //-------------------------------immutable_memory------------------------------------- 1172 // Access immutable memory 1173 Node* Compile::immutable_memory() { 1174 if (_immutable_memory != NULL) { 1175 return _immutable_memory; 1176 } 1177 StartNode* s = start(); 1178 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1179 Node *p = s->fast_out(i); 1180 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1181 _immutable_memory = p; 1182 return _immutable_memory; 1183 } 1184 } 1185 ShouldNotReachHere(); 1186 return NULL; 1187 } 1188 1189 //----------------------set_cached_top_node------------------------------------ 1190 // Install the cached top node, and make sure Node::is_top works correctly. 1191 void Compile::set_cached_top_node(Node* tn) { 1192 if (tn != NULL) verify_top(tn); 1193 Node* old_top = _top; 1194 _top = tn; 1195 // Calling Node::setup_is_top allows the nodes the chance to adjust 1196 // their _out arrays. 1197 if (_top != NULL) _top->setup_is_top(); 1198 if (old_top != NULL) old_top->setup_is_top(); 1199 assert(_top == NULL || top()->is_top(), ""); 1200 } 1201 1202 #ifdef ASSERT 1203 uint Compile::count_live_nodes_by_graph_walk() { 1204 Unique_Node_List useful(comp_arena()); 1205 // Get useful node list by walking the graph. 1206 identify_useful_nodes(useful); 1207 return useful.size(); 1208 } 1209 1210 void Compile::print_missing_nodes() { 1211 1212 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1213 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1214 return; 1215 } 1216 1217 // This is an expensive function. It is executed only when the user 1218 // specifies VerifyIdealNodeCount option or otherwise knows the 1219 // additional work that needs to be done to identify reachable nodes 1220 // by walking the flow graph and find the missing ones using 1221 // _dead_node_list. 1222 1223 Unique_Node_List useful(comp_arena()); 1224 // Get useful node list by walking the graph. 1225 identify_useful_nodes(useful); 1226 1227 uint l_nodes = C->live_nodes(); 1228 uint l_nodes_by_walk = useful.size(); 1229 1230 if (l_nodes != l_nodes_by_walk) { 1231 if (_log != NULL) { 1232 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1233 _log->stamp(); 1234 _log->end_head(); 1235 } 1236 VectorSet& useful_member_set = useful.member_set(); 1237 int last_idx = l_nodes_by_walk; 1238 for (int i = 0; i < last_idx; i++) { 1239 if (useful_member_set.test(i)) { 1240 if (_dead_node_list.test(i)) { 1241 if (_log != NULL) { 1242 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1243 } 1244 if (PrintIdealNodeCount) { 1245 // Print the log message to tty 1246 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1247 useful.at(i)->dump(); 1248 } 1249 } 1250 } 1251 else if (! _dead_node_list.test(i)) { 1252 if (_log != NULL) { 1253 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1254 } 1255 if (PrintIdealNodeCount) { 1256 // Print the log message to tty 1257 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1258 } 1259 } 1260 } 1261 if (_log != NULL) { 1262 _log->tail("mismatched_nodes"); 1263 } 1264 } 1265 } 1266 #endif 1267 1268 #ifndef PRODUCT 1269 void Compile::verify_top(Node* tn) const { 1270 if (tn != NULL) { 1271 assert(tn->is_Con(), "top node must be a constant"); 1272 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1273 assert(tn->in(0) != NULL, "must have live top node"); 1274 } 1275 } 1276 #endif 1277 1278 1279 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1280 1281 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1282 guarantee(arr != NULL, ""); 1283 int num_blocks = arr->length(); 1284 if (grow_by < num_blocks) grow_by = num_blocks; 1285 int num_notes = grow_by * _node_notes_block_size; 1286 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1287 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1288 while (num_notes > 0) { 1289 arr->append(notes); 1290 notes += _node_notes_block_size; 1291 num_notes -= _node_notes_block_size; 1292 } 1293 assert(num_notes == 0, "exact multiple, please"); 1294 } 1295 1296 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1297 if (source == NULL || dest == NULL) return false; 1298 1299 if (dest->is_Con()) 1300 return false; // Do not push debug info onto constants. 1301 1302 #ifdef ASSERT 1303 // Leave a bread crumb trail pointing to the original node: 1304 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1305 dest->set_debug_orig(source); 1306 } 1307 #endif 1308 1309 if (node_note_array() == NULL) 1310 return false; // Not collecting any notes now. 1311 1312 // This is a copy onto a pre-existing node, which may already have notes. 1313 // If both nodes have notes, do not overwrite any pre-existing notes. 1314 Node_Notes* source_notes = node_notes_at(source->_idx); 1315 if (source_notes == NULL || source_notes->is_clear()) return false; 1316 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1317 if (dest_notes == NULL || dest_notes->is_clear()) { 1318 return set_node_notes_at(dest->_idx, source_notes); 1319 } 1320 1321 Node_Notes merged_notes = (*source_notes); 1322 // The order of operations here ensures that dest notes will win... 1323 merged_notes.update_from(dest_notes); 1324 return set_node_notes_at(dest->_idx, &merged_notes); 1325 } 1326 1327 1328 //--------------------------allow_range_check_smearing------------------------- 1329 // Gating condition for coalescing similar range checks. 1330 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1331 // single covering check that is at least as strong as any of them. 1332 // If the optimization succeeds, the simplified (strengthened) range check 1333 // will always succeed. If it fails, we will deopt, and then give up 1334 // on the optimization. 1335 bool Compile::allow_range_check_smearing() const { 1336 // If this method has already thrown a range-check, 1337 // assume it was because we already tried range smearing 1338 // and it failed. 1339 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1340 return !already_trapped; 1341 } 1342 1343 1344 //------------------------------flatten_alias_type----------------------------- 1345 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1346 int offset = tj->offset(); 1347 TypePtr::PTR ptr = tj->ptr(); 1348 1349 // Known instance (scalarizable allocation) alias only with itself. 1350 bool is_known_inst = tj->isa_oopptr() != NULL && 1351 tj->is_oopptr()->is_known_instance(); 1352 1353 // Process weird unsafe references. 1354 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1355 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1356 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1357 tj = TypeOopPtr::BOTTOM; 1358 ptr = tj->ptr(); 1359 offset = tj->offset(); 1360 } 1361 1362 // Array pointers need some flattening 1363 const TypeAryPtr *ta = tj->isa_aryptr(); 1364 if (ta && ta->is_stable()) { 1365 // Erase stability property for alias analysis. 1366 tj = ta = ta->cast_to_stable(false); 1367 } 1368 if( ta && is_known_inst ) { 1369 if ( offset != Type::OffsetBot && 1370 offset > arrayOopDesc::length_offset_in_bytes() ) { 1371 offset = Type::OffsetBot; // Flatten constant access into array body only 1372 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1373 } 1374 } else if( ta && _AliasLevel >= 2 ) { 1375 // For arrays indexed by constant indices, we flatten the alias 1376 // space to include all of the array body. Only the header, klass 1377 // and array length can be accessed un-aliased. 1378 if( offset != Type::OffsetBot ) { 1379 if( ta->const_oop() ) { // MethodData* or Method* 1380 offset = Type::OffsetBot; // Flatten constant access into array body 1381 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1382 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1383 // range is OK as-is. 1384 tj = ta = TypeAryPtr::RANGE; 1385 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1386 tj = TypeInstPtr::KLASS; // all klass loads look alike 1387 ta = TypeAryPtr::RANGE; // generic ignored junk 1388 ptr = TypePtr::BotPTR; 1389 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1390 tj = TypeInstPtr::MARK; 1391 ta = TypeAryPtr::RANGE; // generic ignored junk 1392 ptr = TypePtr::BotPTR; 1393 } else { // Random constant offset into array body 1394 offset = Type::OffsetBot; // Flatten constant access into array body 1395 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1396 } 1397 } 1398 // Arrays of fixed size alias with arrays of unknown size. 1399 if (ta->size() != TypeInt::POS) { 1400 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1401 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1402 } 1403 // Arrays of known objects become arrays of unknown objects. 1404 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1405 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1406 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1407 } 1408 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1409 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1410 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1411 } 1412 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1413 // cannot be distinguished by bytecode alone. 1414 if (ta->elem() == TypeInt::BOOL) { 1415 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1416 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1417 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1418 } 1419 // During the 2nd round of IterGVN, NotNull castings are removed. 1420 // Make sure the Bottom and NotNull variants alias the same. 1421 // Also, make sure exact and non-exact variants alias the same. 1422 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1423 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1424 } 1425 } 1426 1427 // Oop pointers need some flattening 1428 const TypeInstPtr *to = tj->isa_instptr(); 1429 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1430 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1431 if( ptr == TypePtr::Constant ) { 1432 if (to->klass() != ciEnv::current()->Class_klass() || 1433 offset < k->size_helper() * wordSize) { 1434 // No constant oop pointers (such as Strings); they alias with 1435 // unknown strings. 1436 assert(!is_known_inst, "not scalarizable allocation"); 1437 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1438 } 1439 } else if( is_known_inst ) { 1440 tj = to; // Keep NotNull and klass_is_exact for instance type 1441 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1442 // During the 2nd round of IterGVN, NotNull castings are removed. 1443 // Make sure the Bottom and NotNull variants alias the same. 1444 // Also, make sure exact and non-exact variants alias the same. 1445 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1446 } 1447 if (to->speculative() != NULL) { 1448 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1449 } 1450 // Canonicalize the holder of this field 1451 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1452 // First handle header references such as a LoadKlassNode, even if the 1453 // object's klass is unloaded at compile time (4965979). 1454 if (!is_known_inst) { // Do it only for non-instance types 1455 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1456 } 1457 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1458 // Static fields are in the space above the normal instance 1459 // fields in the java.lang.Class instance. 1460 if (to->klass() != ciEnv::current()->Class_klass()) { 1461 to = NULL; 1462 tj = TypeOopPtr::BOTTOM; 1463 offset = tj->offset(); 1464 } 1465 } else { 1466 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1467 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1468 if( is_known_inst ) { 1469 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1470 } else { 1471 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1472 } 1473 } 1474 } 1475 } 1476 1477 // Klass pointers to object array klasses need some flattening 1478 const TypeKlassPtr *tk = tj->isa_klassptr(); 1479 if( tk ) { 1480 // If we are referencing a field within a Klass, we need 1481 // to assume the worst case of an Object. Both exact and 1482 // inexact types must flatten to the same alias class so 1483 // use NotNull as the PTR. 1484 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1485 1486 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1487 TypeKlassPtr::OBJECT->klass(), 1488 offset); 1489 } 1490 1491 ciKlass* klass = tk->klass(); 1492 if( klass->is_obj_array_klass() ) { 1493 ciKlass* k = TypeAryPtr::OOPS->klass(); 1494 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1495 k = TypeInstPtr::BOTTOM->klass(); 1496 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1497 } 1498 1499 // Check for precise loads from the primary supertype array and force them 1500 // to the supertype cache alias index. Check for generic array loads from 1501 // the primary supertype array and also force them to the supertype cache 1502 // alias index. Since the same load can reach both, we need to merge 1503 // these 2 disparate memories into the same alias class. Since the 1504 // primary supertype array is read-only, there's no chance of confusion 1505 // where we bypass an array load and an array store. 1506 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1507 if (offset == Type::OffsetBot || 1508 (offset >= primary_supers_offset && 1509 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1510 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1511 offset = in_bytes(Klass::secondary_super_cache_offset()); 1512 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1513 } 1514 } 1515 1516 // Flatten all Raw pointers together. 1517 if (tj->base() == Type::RawPtr) 1518 tj = TypeRawPtr::BOTTOM; 1519 1520 if (tj->base() == Type::AnyPtr) 1521 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1522 1523 // Flatten all to bottom for now 1524 switch( _AliasLevel ) { 1525 case 0: 1526 tj = TypePtr::BOTTOM; 1527 break; 1528 case 1: // Flatten to: oop, static, field or array 1529 switch (tj->base()) { 1530 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1531 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1532 case Type::AryPtr: // do not distinguish arrays at all 1533 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1534 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1535 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1536 default: ShouldNotReachHere(); 1537 } 1538 break; 1539 case 2: // No collapsing at level 2; keep all splits 1540 case 3: // No collapsing at level 3; keep all splits 1541 break; 1542 default: 1543 Unimplemented(); 1544 } 1545 1546 offset = tj->offset(); 1547 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1548 1549 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1550 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1551 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1552 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1553 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1554 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1555 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1556 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1557 assert( tj->ptr() != TypePtr::TopPTR && 1558 tj->ptr() != TypePtr::AnyNull && 1559 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1560 // assert( tj->ptr() != TypePtr::Constant || 1561 // tj->base() == Type::RawPtr || 1562 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1563 1564 return tj; 1565 } 1566 1567 void Compile::AliasType::Init(int i, const TypePtr* at) { 1568 _index = i; 1569 _adr_type = at; 1570 _field = NULL; 1571 _element = NULL; 1572 _is_rewritable = true; // default 1573 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1574 if (atoop != NULL && atoop->is_known_instance()) { 1575 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1576 _general_index = Compile::current()->get_alias_index(gt); 1577 } else { 1578 _general_index = 0; 1579 } 1580 } 1581 1582 //---------------------------------print_on------------------------------------ 1583 #ifndef PRODUCT 1584 void Compile::AliasType::print_on(outputStream* st) { 1585 if (index() < 10) 1586 st->print("@ <%d> ", index()); 1587 else st->print("@ <%d>", index()); 1588 st->print(is_rewritable() ? " " : " RO"); 1589 int offset = adr_type()->offset(); 1590 if (offset == Type::OffsetBot) 1591 st->print(" +any"); 1592 else st->print(" +%-3d", offset); 1593 st->print(" in "); 1594 adr_type()->dump_on(st); 1595 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1596 if (field() != NULL && tjp) { 1597 if (tjp->klass() != field()->holder() || 1598 tjp->offset() != field()->offset_in_bytes()) { 1599 st->print(" != "); 1600 field()->print(); 1601 st->print(" ***"); 1602 } 1603 } 1604 } 1605 1606 void print_alias_types() { 1607 Compile* C = Compile::current(); 1608 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1609 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1610 C->alias_type(idx)->print_on(tty); 1611 tty->cr(); 1612 } 1613 } 1614 #endif 1615 1616 1617 //----------------------------probe_alias_cache-------------------------------- 1618 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1619 intptr_t key = (intptr_t) adr_type; 1620 key ^= key >> logAliasCacheSize; 1621 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1622 } 1623 1624 1625 //-----------------------------grow_alias_types-------------------------------- 1626 void Compile::grow_alias_types() { 1627 const int old_ats = _max_alias_types; // how many before? 1628 const int new_ats = old_ats; // how many more? 1629 const int grow_ats = old_ats+new_ats; // how many now? 1630 _max_alias_types = grow_ats; 1631 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1632 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1633 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1634 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1635 } 1636 1637 1638 //--------------------------------find_alias_type------------------------------ 1639 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1640 if (_AliasLevel == 0) 1641 return alias_type(AliasIdxBot); 1642 1643 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1644 if (ace->_adr_type == adr_type) { 1645 return alias_type(ace->_index); 1646 } 1647 1648 // Handle special cases. 1649 if (adr_type == NULL) return alias_type(AliasIdxTop); 1650 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1651 1652 // Do it the slow way. 1653 const TypePtr* flat = flatten_alias_type(adr_type); 1654 1655 #ifdef ASSERT 1656 assert(flat == flatten_alias_type(flat), "idempotent"); 1657 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1658 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1659 const TypeOopPtr* foop = flat->is_oopptr(); 1660 // Scalarizable allocations have exact klass always. 1661 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1662 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1663 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1664 } 1665 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1666 #endif 1667 1668 int idx = AliasIdxTop; 1669 for (int i = 0; i < num_alias_types(); i++) { 1670 if (alias_type(i)->adr_type() == flat) { 1671 idx = i; 1672 break; 1673 } 1674 } 1675 1676 if (idx == AliasIdxTop) { 1677 if (no_create) return NULL; 1678 // Grow the array if necessary. 1679 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1680 // Add a new alias type. 1681 idx = _num_alias_types++; 1682 _alias_types[idx]->Init(idx, flat); 1683 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1684 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1685 if (flat->isa_instptr()) { 1686 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1687 && flat->is_instptr()->klass() == env()->Class_klass()) 1688 alias_type(idx)->set_rewritable(false); 1689 } 1690 if (flat->isa_aryptr()) { 1691 #ifdef ASSERT 1692 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1693 // (T_BYTE has the weakest alignment and size restrictions...) 1694 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1695 #endif 1696 if (flat->offset() == TypePtr::OffsetBot) { 1697 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1698 } 1699 } 1700 if (flat->isa_klassptr()) { 1701 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1702 alias_type(idx)->set_rewritable(false); 1703 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1704 alias_type(idx)->set_rewritable(false); 1705 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1706 alias_type(idx)->set_rewritable(false); 1707 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1708 alias_type(idx)->set_rewritable(false); 1709 } 1710 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1711 // but the base pointer type is not distinctive enough to identify 1712 // references into JavaThread.) 1713 1714 // Check for final fields. 1715 const TypeInstPtr* tinst = flat->isa_instptr(); 1716 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1717 ciField* field; 1718 if (tinst->const_oop() != NULL && 1719 tinst->klass() == ciEnv::current()->Class_klass() && 1720 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1721 // static field 1722 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1723 field = k->get_field_by_offset(tinst->offset(), true); 1724 } else { 1725 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1726 field = k->get_field_by_offset(tinst->offset(), false); 1727 } 1728 assert(field == NULL || 1729 original_field == NULL || 1730 (field->holder() == original_field->holder() && 1731 field->offset() == original_field->offset() && 1732 field->is_static() == original_field->is_static()), "wrong field?"); 1733 // Set field() and is_rewritable() attributes. 1734 if (field != NULL) alias_type(idx)->set_field(field); 1735 } 1736 } 1737 1738 // Fill the cache for next time. 1739 ace->_adr_type = adr_type; 1740 ace->_index = idx; 1741 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1742 1743 // Might as well try to fill the cache for the flattened version, too. 1744 AliasCacheEntry* face = probe_alias_cache(flat); 1745 if (face->_adr_type == NULL) { 1746 face->_adr_type = flat; 1747 face->_index = idx; 1748 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1749 } 1750 1751 return alias_type(idx); 1752 } 1753 1754 1755 Compile::AliasType* Compile::alias_type(ciField* field) { 1756 const TypeOopPtr* t; 1757 if (field->is_static()) 1758 t = TypeInstPtr::make(field->holder()->java_mirror()); 1759 else 1760 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1761 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1762 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1763 return atp; 1764 } 1765 1766 1767 //------------------------------have_alias_type-------------------------------- 1768 bool Compile::have_alias_type(const TypePtr* adr_type) { 1769 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1770 if (ace->_adr_type == adr_type) { 1771 return true; 1772 } 1773 1774 // Handle special cases. 1775 if (adr_type == NULL) return true; 1776 if (adr_type == TypePtr::BOTTOM) return true; 1777 1778 return find_alias_type(adr_type, true, NULL) != NULL; 1779 } 1780 1781 //-----------------------------must_alias-------------------------------------- 1782 // True if all values of the given address type are in the given alias category. 1783 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1784 if (alias_idx == AliasIdxBot) return true; // the universal category 1785 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1786 if (alias_idx == AliasIdxTop) return false; // the empty category 1787 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1788 1789 // the only remaining possible overlap is identity 1790 int adr_idx = get_alias_index(adr_type); 1791 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1792 assert(adr_idx == alias_idx || 1793 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1794 && adr_type != TypeOopPtr::BOTTOM), 1795 "should not be testing for overlap with an unsafe pointer"); 1796 return adr_idx == alias_idx; 1797 } 1798 1799 //------------------------------can_alias-------------------------------------- 1800 // True if any values of the given address type are in the given alias category. 1801 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1802 if (alias_idx == AliasIdxTop) return false; // the empty category 1803 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1804 if (alias_idx == AliasIdxBot) return true; // the universal category 1805 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1806 1807 // the only remaining possible overlap is identity 1808 int adr_idx = get_alias_index(adr_type); 1809 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1810 return adr_idx == alias_idx; 1811 } 1812 1813 1814 1815 //---------------------------pop_warm_call------------------------------------- 1816 WarmCallInfo* Compile::pop_warm_call() { 1817 WarmCallInfo* wci = _warm_calls; 1818 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1819 return wci; 1820 } 1821 1822 //----------------------------Inline_Warm-------------------------------------- 1823 int Compile::Inline_Warm() { 1824 // If there is room, try to inline some more warm call sites. 1825 // %%% Do a graph index compaction pass when we think we're out of space? 1826 if (!InlineWarmCalls) return 0; 1827 1828 int calls_made_hot = 0; 1829 int room_to_grow = NodeCountInliningCutoff - unique(); 1830 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1831 int amount_grown = 0; 1832 WarmCallInfo* call; 1833 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1834 int est_size = (int)call->size(); 1835 if (est_size > (room_to_grow - amount_grown)) { 1836 // This one won't fit anyway. Get rid of it. 1837 call->make_cold(); 1838 continue; 1839 } 1840 call->make_hot(); 1841 calls_made_hot++; 1842 amount_grown += est_size; 1843 amount_to_grow -= est_size; 1844 } 1845 1846 if (calls_made_hot > 0) set_major_progress(); 1847 return calls_made_hot; 1848 } 1849 1850 1851 //----------------------------Finish_Warm-------------------------------------- 1852 void Compile::Finish_Warm() { 1853 if (!InlineWarmCalls) return; 1854 if (failing()) return; 1855 if (warm_calls() == NULL) return; 1856 1857 // Clean up loose ends, if we are out of space for inlining. 1858 WarmCallInfo* call; 1859 while ((call = pop_warm_call()) != NULL) { 1860 call->make_cold(); 1861 } 1862 } 1863 1864 //---------------------cleanup_loop_predicates----------------------- 1865 // Remove the opaque nodes that protect the predicates so that all unused 1866 // checks and uncommon_traps will be eliminated from the ideal graph 1867 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1868 if (predicate_count()==0) return; 1869 for (int i = predicate_count(); i > 0; i--) { 1870 Node * n = predicate_opaque1_node(i-1); 1871 assert(n->Opcode() == Op_Opaque1, "must be"); 1872 igvn.replace_node(n, n->in(1)); 1873 } 1874 assert(predicate_count()==0, "should be clean!"); 1875 } 1876 1877 // StringOpts and late inlining of string methods 1878 void Compile::inline_string_calls(bool parse_time) { 1879 { 1880 // remove useless nodes to make the usage analysis simpler 1881 ResourceMark rm; 1882 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1883 } 1884 1885 { 1886 ResourceMark rm; 1887 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1888 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1889 print_method(PHASE_AFTER_STRINGOPTS, 3); 1890 } 1891 1892 // now inline anything that we skipped the first time around 1893 if (!parse_time) { 1894 _late_inlines_pos = _late_inlines.length(); 1895 } 1896 1897 while (_string_late_inlines.length() > 0) { 1898 CallGenerator* cg = _string_late_inlines.pop(); 1899 cg->do_late_inline(); 1900 if (failing()) return; 1901 } 1902 _string_late_inlines.trunc_to(0); 1903 } 1904 1905 // Late inlining of boxing methods 1906 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1907 if (_boxing_late_inlines.length() > 0) { 1908 assert(has_boxed_value(), "inconsistent"); 1909 1910 PhaseGVN* gvn = initial_gvn(); 1911 set_inlining_incrementally(true); 1912 1913 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1914 for_igvn()->clear(); 1915 gvn->replace_with(&igvn); 1916 1917 while (_boxing_late_inlines.length() > 0) { 1918 CallGenerator* cg = _boxing_late_inlines.pop(); 1919 cg->do_late_inline(); 1920 if (failing()) return; 1921 } 1922 _boxing_late_inlines.trunc_to(0); 1923 1924 { 1925 ResourceMark rm; 1926 PhaseRemoveUseless pru(gvn, for_igvn()); 1927 } 1928 1929 igvn = PhaseIterGVN(gvn); 1930 igvn.optimize(); 1931 1932 set_inlining_progress(false); 1933 set_inlining_incrementally(false); 1934 } 1935 } 1936 1937 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1938 assert(IncrementalInline, "incremental inlining should be on"); 1939 PhaseGVN* gvn = initial_gvn(); 1940 1941 set_inlining_progress(false); 1942 for_igvn()->clear(); 1943 gvn->replace_with(&igvn); 1944 1945 int i = 0; 1946 1947 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1948 CallGenerator* cg = _late_inlines.at(i); 1949 _late_inlines_pos = i+1; 1950 cg->do_late_inline(); 1951 if (failing()) return; 1952 } 1953 int j = 0; 1954 for (; i < _late_inlines.length(); i++, j++) { 1955 _late_inlines.at_put(j, _late_inlines.at(i)); 1956 } 1957 _late_inlines.trunc_to(j); 1958 1959 { 1960 ResourceMark rm; 1961 PhaseRemoveUseless pru(gvn, for_igvn()); 1962 } 1963 1964 igvn = PhaseIterGVN(gvn); 1965 } 1966 1967 // Perform incremental inlining until bound on number of live nodes is reached 1968 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1969 PhaseGVN* gvn = initial_gvn(); 1970 1971 set_inlining_incrementally(true); 1972 set_inlining_progress(true); 1973 uint low_live_nodes = 0; 1974 1975 while(inlining_progress() && _late_inlines.length() > 0) { 1976 1977 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1978 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1979 // PhaseIdealLoop is expensive so we only try it once we are 1980 // out of loop and we only try it again if the previous helped 1981 // got the number of nodes down significantly 1982 PhaseIdealLoop ideal_loop( igvn, false, true ); 1983 if (failing()) return; 1984 low_live_nodes = live_nodes(); 1985 _major_progress = true; 1986 } 1987 1988 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1989 break; 1990 } 1991 } 1992 1993 inline_incrementally_one(igvn); 1994 1995 if (failing()) return; 1996 1997 igvn.optimize(); 1998 1999 if (failing()) return; 2000 } 2001 2002 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2003 2004 if (_string_late_inlines.length() > 0) { 2005 assert(has_stringbuilder(), "inconsistent"); 2006 for_igvn()->clear(); 2007 initial_gvn()->replace_with(&igvn); 2008 2009 inline_string_calls(false); 2010 2011 if (failing()) return; 2012 2013 { 2014 ResourceMark rm; 2015 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2016 } 2017 2018 igvn = PhaseIterGVN(gvn); 2019 2020 igvn.optimize(); 2021 } 2022 2023 set_inlining_incrementally(false); 2024 } 2025 2026 2027 //------------------------------Optimize--------------------------------------- 2028 // Given a graph, optimize it. 2029 void Compile::Optimize() { 2030 TracePhase t1("optimizer", &_t_optimizer, true); 2031 2032 #ifndef PRODUCT 2033 if (env()->break_at_compile()) { 2034 BREAKPOINT; 2035 } 2036 2037 #endif 2038 2039 ResourceMark rm; 2040 int loop_opts_cnt; 2041 2042 print_inlining_reinit(); 2043 2044 NOT_PRODUCT( verify_graph_edges(); ) 2045 2046 print_method(PHASE_AFTER_PARSING); 2047 2048 { 2049 // Iterative Global Value Numbering, including ideal transforms 2050 // Initialize IterGVN with types and values from parse-time GVN 2051 PhaseIterGVN igvn(initial_gvn()); 2052 { 2053 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 2054 igvn.optimize(); 2055 } 2056 2057 print_method(PHASE_ITER_GVN1, 2); 2058 2059 if (failing()) return; 2060 2061 { 2062 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2063 inline_incrementally(igvn); 2064 } 2065 2066 print_method(PHASE_INCREMENTAL_INLINE, 2); 2067 2068 if (failing()) return; 2069 2070 if (eliminate_boxing()) { 2071 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2072 // Inline valueOf() methods now. 2073 inline_boxing_calls(igvn); 2074 2075 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2076 2077 if (failing()) return; 2078 } 2079 2080 // Remove the speculative part of types and clean up the graph from 2081 // the extra CastPP nodes whose only purpose is to carry them. Do 2082 // that early so that optimizations are not disrupted by the extra 2083 // CastPP nodes. 2084 remove_speculative_types(igvn); 2085 2086 // No more new expensive nodes will be added to the list from here 2087 // so keep only the actual candidates for optimizations. 2088 cleanup_expensive_nodes(igvn); 2089 2090 // Perform escape analysis 2091 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2092 if (has_loops()) { 2093 // Cleanup graph (remove dead nodes). 2094 TracePhase t2("idealLoop", &_t_idealLoop, true); 2095 PhaseIdealLoop ideal_loop( igvn, false, true ); 2096 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2097 if (failing()) return; 2098 } 2099 ConnectionGraph::do_analysis(this, &igvn); 2100 2101 if (failing()) return; 2102 2103 // Optimize out fields loads from scalar replaceable allocations. 2104 igvn.optimize(); 2105 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2106 2107 if (failing()) return; 2108 2109 if (congraph() != NULL && macro_count() > 0) { 2110 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2111 PhaseMacroExpand mexp(igvn); 2112 mexp.eliminate_macro_nodes(); 2113 igvn.set_delay_transform(false); 2114 2115 igvn.optimize(); 2116 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2117 2118 if (failing()) return; 2119 } 2120 } 2121 2122 // Loop transforms on the ideal graph. Range Check Elimination, 2123 // peeling, unrolling, etc. 2124 2125 // Set loop opts counter 2126 loop_opts_cnt = num_loop_opts(); 2127 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2128 { 2129 TracePhase t2("idealLoop", &_t_idealLoop, true); 2130 PhaseIdealLoop ideal_loop( igvn, true ); 2131 loop_opts_cnt--; 2132 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2133 if (failing()) return; 2134 } 2135 // Loop opts pass if partial peeling occurred in previous pass 2136 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2137 TracePhase t3("idealLoop", &_t_idealLoop, true); 2138 PhaseIdealLoop ideal_loop( igvn, false ); 2139 loop_opts_cnt--; 2140 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2141 if (failing()) return; 2142 } 2143 // Loop opts pass for loop-unrolling before CCP 2144 if(major_progress() && (loop_opts_cnt > 0)) { 2145 TracePhase t4("idealLoop", &_t_idealLoop, true); 2146 PhaseIdealLoop ideal_loop( igvn, false ); 2147 loop_opts_cnt--; 2148 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2149 } 2150 if (!failing()) { 2151 // Verify that last round of loop opts produced a valid graph 2152 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2153 PhaseIdealLoop::verify(igvn); 2154 } 2155 } 2156 if (failing()) return; 2157 2158 // Conditional Constant Propagation; 2159 PhaseCCP ccp( &igvn ); 2160 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2161 { 2162 TracePhase t2("ccp", &_t_ccp, true); 2163 ccp.do_transform(); 2164 } 2165 print_method(PHASE_CPP1, 2); 2166 2167 assert( true, "Break here to ccp.dump_old2new_map()"); 2168 2169 // Iterative Global Value Numbering, including ideal transforms 2170 { 2171 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2172 igvn = ccp; 2173 igvn.optimize(); 2174 } 2175 2176 print_method(PHASE_ITER_GVN2, 2); 2177 2178 if (failing()) return; 2179 2180 // Loop transforms on the ideal graph. Range Check Elimination, 2181 // peeling, unrolling, etc. 2182 if(loop_opts_cnt > 0) { 2183 debug_only( int cnt = 0; ); 2184 while(major_progress() && (loop_opts_cnt > 0)) { 2185 TracePhase t2("idealLoop", &_t_idealLoop, true); 2186 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2187 PhaseIdealLoop ideal_loop( igvn, true); 2188 loop_opts_cnt--; 2189 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2190 if (failing()) return; 2191 } 2192 } 2193 2194 { 2195 // Verify that all previous optimizations produced a valid graph 2196 // at least to this point, even if no loop optimizations were done. 2197 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2198 PhaseIdealLoop::verify(igvn); 2199 } 2200 2201 { 2202 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2203 PhaseMacroExpand mex(igvn); 2204 if (mex.expand_macro_nodes()) { 2205 assert(failing(), "must bail out w/ explicit message"); 2206 return; 2207 } 2208 } 2209 2210 } // (End scope of igvn; run destructor if necessary for asserts.) 2211 2212 process_print_inlining(); 2213 // A method with only infinite loops has no edges entering loops from root 2214 { 2215 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2216 if (final_graph_reshaping()) { 2217 assert(failing(), "must bail out w/ explicit message"); 2218 return; 2219 } 2220 } 2221 2222 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2223 } 2224 2225 2226 //------------------------------Code_Gen--------------------------------------- 2227 // Given a graph, generate code for it 2228 void Compile::Code_Gen() { 2229 if (failing()) { 2230 return; 2231 } 2232 2233 // Perform instruction selection. You might think we could reclaim Matcher 2234 // memory PDQ, but actually the Matcher is used in generating spill code. 2235 // Internals of the Matcher (including some VectorSets) must remain live 2236 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2237 // set a bit in reclaimed memory. 2238 2239 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2240 // nodes. Mapping is only valid at the root of each matched subtree. 2241 NOT_PRODUCT( verify_graph_edges(); ) 2242 2243 Matcher matcher; 2244 _matcher = &matcher; 2245 { 2246 TracePhase t2("matcher", &_t_matcher, true); 2247 matcher.match(); 2248 } 2249 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2250 // nodes. Mapping is only valid at the root of each matched subtree. 2251 NOT_PRODUCT( verify_graph_edges(); ) 2252 2253 // If you have too many nodes, or if matching has failed, bail out 2254 check_node_count(0, "out of nodes matching instructions"); 2255 if (failing()) { 2256 return; 2257 } 2258 2259 // Build a proper-looking CFG 2260 PhaseCFG cfg(node_arena(), root(), matcher); 2261 _cfg = &cfg; 2262 { 2263 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2264 bool success = cfg.do_global_code_motion(); 2265 if (!success) { 2266 return; 2267 } 2268 2269 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2270 NOT_PRODUCT( verify_graph_edges(); ) 2271 debug_only( cfg.verify(); ) 2272 } 2273 2274 PhaseChaitin regalloc(unique(), cfg, matcher); 2275 _regalloc = ®alloc; 2276 { 2277 TracePhase t2("regalloc", &_t_registerAllocation, true); 2278 // Perform register allocation. After Chaitin, use-def chains are 2279 // no longer accurate (at spill code) and so must be ignored. 2280 // Node->LRG->reg mappings are still accurate. 2281 _regalloc->Register_Allocate(); 2282 2283 // Bail out if the allocator builds too many nodes 2284 if (failing()) { 2285 return; 2286 } 2287 } 2288 2289 // Prior to register allocation we kept empty basic blocks in case the 2290 // the allocator needed a place to spill. After register allocation we 2291 // are not adding any new instructions. If any basic block is empty, we 2292 // can now safely remove it. 2293 { 2294 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2295 cfg.remove_empty_blocks(); 2296 if (do_freq_based_layout()) { 2297 PhaseBlockLayout layout(cfg); 2298 } else { 2299 cfg.set_loop_alignment(); 2300 } 2301 cfg.fixup_flow(); 2302 } 2303 2304 // Apply peephole optimizations 2305 if( OptoPeephole ) { 2306 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2307 PhasePeephole peep( _regalloc, cfg); 2308 peep.do_transform(); 2309 } 2310 2311 // Do late expand if CPU requires this. 2312 if (Matcher::require_postalloc_expand) { 2313 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true)); 2314 cfg.postalloc_expand(_regalloc); 2315 } 2316 2317 // Convert Nodes to instruction bits in a buffer 2318 { 2319 // %%%% workspace merge brought two timers together for one job 2320 TracePhase t2a("output", &_t_output, true); 2321 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2322 Output(); 2323 } 2324 2325 print_method(PHASE_FINAL_CODE); 2326 2327 // He's dead, Jim. 2328 _cfg = (PhaseCFG*)0xdeadbeef; 2329 _regalloc = (PhaseChaitin*)0xdeadbeef; 2330 } 2331 2332 2333 //------------------------------dump_asm--------------------------------------- 2334 // Dump formatted assembly 2335 #ifndef PRODUCT 2336 void Compile::dump_asm(int *pcs, uint pc_limit) { 2337 bool cut_short = false; 2338 tty->print_cr("#"); 2339 tty->print("# "); _tf->dump(); tty->cr(); 2340 tty->print_cr("#"); 2341 2342 // For all blocks 2343 int pc = 0x0; // Program counter 2344 char starts_bundle = ' '; 2345 _regalloc->dump_frame(); 2346 2347 Node *n = NULL; 2348 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2349 if (VMThread::should_terminate()) { 2350 cut_short = true; 2351 break; 2352 } 2353 Block* block = _cfg->get_block(i); 2354 if (block->is_connector() && !Verbose) { 2355 continue; 2356 } 2357 n = block->head(); 2358 if (pcs && n->_idx < pc_limit) { 2359 tty->print("%3.3x ", pcs[n->_idx]); 2360 } else { 2361 tty->print(" "); 2362 } 2363 block->dump_head(_cfg); 2364 if (block->is_connector()) { 2365 tty->print_cr(" # Empty connector block"); 2366 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2367 tty->print_cr(" # Block is sole successor of call"); 2368 } 2369 2370 // For all instructions 2371 Node *delay = NULL; 2372 for (uint j = 0; j < block->number_of_nodes(); j++) { 2373 if (VMThread::should_terminate()) { 2374 cut_short = true; 2375 break; 2376 } 2377 n = block->get_node(j); 2378 if (valid_bundle_info(n)) { 2379 Bundle* bundle = node_bundling(n); 2380 if (bundle->used_in_unconditional_delay()) { 2381 delay = n; 2382 continue; 2383 } 2384 if (bundle->starts_bundle()) { 2385 starts_bundle = '+'; 2386 } 2387 } 2388 2389 if (WizardMode) { 2390 n->dump(); 2391 } 2392 2393 if( !n->is_Region() && // Dont print in the Assembly 2394 !n->is_Phi() && // a few noisely useless nodes 2395 !n->is_Proj() && 2396 !n->is_MachTemp() && 2397 !n->is_SafePointScalarObject() && 2398 !n->is_Catch() && // Would be nice to print exception table targets 2399 !n->is_MergeMem() && // Not very interesting 2400 !n->is_top() && // Debug info table constants 2401 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2402 ) { 2403 if (pcs && n->_idx < pc_limit) 2404 tty->print("%3.3x", pcs[n->_idx]); 2405 else 2406 tty->print(" "); 2407 tty->print(" %c ", starts_bundle); 2408 starts_bundle = ' '; 2409 tty->print("\t"); 2410 n->format(_regalloc, tty); 2411 tty->cr(); 2412 } 2413 2414 // If we have an instruction with a delay slot, and have seen a delay, 2415 // then back up and print it 2416 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2417 assert(delay != NULL, "no unconditional delay instruction"); 2418 if (WizardMode) delay->dump(); 2419 2420 if (node_bundling(delay)->starts_bundle()) 2421 starts_bundle = '+'; 2422 if (pcs && n->_idx < pc_limit) 2423 tty->print("%3.3x", pcs[n->_idx]); 2424 else 2425 tty->print(" "); 2426 tty->print(" %c ", starts_bundle); 2427 starts_bundle = ' '; 2428 tty->print("\t"); 2429 delay->format(_regalloc, tty); 2430 tty->cr(); 2431 delay = NULL; 2432 } 2433 2434 // Dump the exception table as well 2435 if( n->is_Catch() && (Verbose || WizardMode) ) { 2436 // Print the exception table for this offset 2437 _handler_table.print_subtable_for(pc); 2438 } 2439 } 2440 2441 if (pcs && n->_idx < pc_limit) 2442 tty->print_cr("%3.3x", pcs[n->_idx]); 2443 else 2444 tty->cr(); 2445 2446 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2447 2448 } // End of per-block dump 2449 tty->cr(); 2450 2451 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2452 } 2453 #endif 2454 2455 //------------------------------Final_Reshape_Counts--------------------------- 2456 // This class defines counters to help identify when a method 2457 // may/must be executed using hardware with only 24-bit precision. 2458 struct Final_Reshape_Counts : public StackObj { 2459 int _call_count; // count non-inlined 'common' calls 2460 int _float_count; // count float ops requiring 24-bit precision 2461 int _double_count; // count double ops requiring more precision 2462 int _java_call_count; // count non-inlined 'java' calls 2463 int _inner_loop_count; // count loops which need alignment 2464 VectorSet _visited; // Visitation flags 2465 Node_List _tests; // Set of IfNodes & PCTableNodes 2466 2467 Final_Reshape_Counts() : 2468 _call_count(0), _float_count(0), _double_count(0), 2469 _java_call_count(0), _inner_loop_count(0), 2470 _visited( Thread::current()->resource_area() ) { } 2471 2472 void inc_call_count () { _call_count ++; } 2473 void inc_float_count () { _float_count ++; } 2474 void inc_double_count() { _double_count++; } 2475 void inc_java_call_count() { _java_call_count++; } 2476 void inc_inner_loop_count() { _inner_loop_count++; } 2477 2478 int get_call_count () const { return _call_count ; } 2479 int get_float_count () const { return _float_count ; } 2480 int get_double_count() const { return _double_count; } 2481 int get_java_call_count() const { return _java_call_count; } 2482 int get_inner_loop_count() const { return _inner_loop_count; } 2483 }; 2484 2485 #ifdef ASSERT 2486 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2487 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2488 // Make sure the offset goes inside the instance layout. 2489 return k->contains_field_offset(tp->offset()); 2490 // Note that OffsetBot and OffsetTop are very negative. 2491 } 2492 #endif 2493 2494 // Eliminate trivially redundant StoreCMs and accumulate their 2495 // precedence edges. 2496 void Compile::eliminate_redundant_card_marks(Node* n) { 2497 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2498 if (n->in(MemNode::Address)->outcnt() > 1) { 2499 // There are multiple users of the same address so it might be 2500 // possible to eliminate some of the StoreCMs 2501 Node* mem = n->in(MemNode::Memory); 2502 Node* adr = n->in(MemNode::Address); 2503 Node* val = n->in(MemNode::ValueIn); 2504 Node* prev = n; 2505 bool done = false; 2506 // Walk the chain of StoreCMs eliminating ones that match. As 2507 // long as it's a chain of single users then the optimization is 2508 // safe. Eliminating partially redundant StoreCMs would require 2509 // cloning copies down the other paths. 2510 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2511 if (adr == mem->in(MemNode::Address) && 2512 val == mem->in(MemNode::ValueIn)) { 2513 // redundant StoreCM 2514 if (mem->req() > MemNode::OopStore) { 2515 // Hasn't been processed by this code yet. 2516 n->add_prec(mem->in(MemNode::OopStore)); 2517 } else { 2518 // Already converted to precedence edge 2519 for (uint i = mem->req(); i < mem->len(); i++) { 2520 // Accumulate any precedence edges 2521 if (mem->in(i) != NULL) { 2522 n->add_prec(mem->in(i)); 2523 } 2524 } 2525 // Everything above this point has been processed. 2526 done = true; 2527 } 2528 // Eliminate the previous StoreCM 2529 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2530 assert(mem->outcnt() == 0, "should be dead"); 2531 mem->disconnect_inputs(NULL, this); 2532 } else { 2533 prev = mem; 2534 } 2535 mem = prev->in(MemNode::Memory); 2536 } 2537 } 2538 } 2539 2540 //------------------------------final_graph_reshaping_impl---------------------- 2541 // Implement items 1-5 from final_graph_reshaping below. 2542 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2543 2544 if ( n->outcnt() == 0 ) return; // dead node 2545 uint nop = n->Opcode(); 2546 2547 // Check for 2-input instruction with "last use" on right input. 2548 // Swap to left input. Implements item (2). 2549 if( n->req() == 3 && // two-input instruction 2550 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2551 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2552 n->in(2)->outcnt() == 1 &&// right use IS a last use 2553 !n->in(2)->is_Con() ) { // right use is not a constant 2554 // Check for commutative opcode 2555 switch( nop ) { 2556 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2557 case Op_MaxI: case Op_MinI: 2558 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2559 case Op_AndL: case Op_XorL: case Op_OrL: 2560 case Op_AndI: case Op_XorI: case Op_OrI: { 2561 // Move "last use" input to left by swapping inputs 2562 n->swap_edges(1, 2); 2563 break; 2564 } 2565 default: 2566 break; 2567 } 2568 } 2569 2570 #ifdef ASSERT 2571 if( n->is_Mem() ) { 2572 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2573 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2574 // oop will be recorded in oop map if load crosses safepoint 2575 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2576 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2577 "raw memory operations should have control edge"); 2578 } 2579 #endif 2580 // Count FPU ops and common calls, implements item (3) 2581 switch( nop ) { 2582 // Count all float operations that may use FPU 2583 case Op_AddF: 2584 case Op_SubF: 2585 case Op_MulF: 2586 case Op_DivF: 2587 case Op_NegF: 2588 case Op_ModF: 2589 case Op_ConvI2F: 2590 case Op_ConF: 2591 case Op_CmpF: 2592 case Op_CmpF3: 2593 // case Op_ConvL2F: // longs are split into 32-bit halves 2594 frc.inc_float_count(); 2595 break; 2596 2597 case Op_ConvF2D: 2598 case Op_ConvD2F: 2599 frc.inc_float_count(); 2600 frc.inc_double_count(); 2601 break; 2602 2603 // Count all double operations that may use FPU 2604 case Op_AddD: 2605 case Op_SubD: 2606 case Op_MulD: 2607 case Op_DivD: 2608 case Op_NegD: 2609 case Op_ModD: 2610 case Op_ConvI2D: 2611 case Op_ConvD2I: 2612 // case Op_ConvL2D: // handled by leaf call 2613 // case Op_ConvD2L: // handled by leaf call 2614 case Op_ConD: 2615 case Op_CmpD: 2616 case Op_CmpD3: 2617 frc.inc_double_count(); 2618 break; 2619 case Op_Opaque1: // Remove Opaque Nodes before matching 2620 case Op_Opaque2: // Remove Opaque Nodes before matching 2621 case Op_Opaque3: 2622 n->subsume_by(n->in(1), this); 2623 break; 2624 case Op_CallStaticJava: 2625 case Op_CallJava: 2626 case Op_CallDynamicJava: 2627 frc.inc_java_call_count(); // Count java call site; 2628 case Op_CallRuntime: 2629 case Op_CallLeaf: 2630 case Op_CallLeafNoFP: { 2631 assert( n->is_Call(), "" ); 2632 CallNode *call = n->as_Call(); 2633 // Count call sites where the FP mode bit would have to be flipped. 2634 // Do not count uncommon runtime calls: 2635 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2636 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2637 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2638 frc.inc_call_count(); // Count the call site 2639 } else { // See if uncommon argument is shared 2640 Node *n = call->in(TypeFunc::Parms); 2641 int nop = n->Opcode(); 2642 // Clone shared simple arguments to uncommon calls, item (1). 2643 if( n->outcnt() > 1 && 2644 !n->is_Proj() && 2645 nop != Op_CreateEx && 2646 nop != Op_CheckCastPP && 2647 nop != Op_DecodeN && 2648 nop != Op_DecodeNKlass && 2649 !n->is_Mem() ) { 2650 Node *x = n->clone(); 2651 call->set_req( TypeFunc::Parms, x ); 2652 } 2653 } 2654 break; 2655 } 2656 2657 case Op_StoreD: 2658 case Op_LoadD: 2659 case Op_LoadD_unaligned: 2660 frc.inc_double_count(); 2661 goto handle_mem; 2662 case Op_StoreF: 2663 case Op_LoadF: 2664 frc.inc_float_count(); 2665 goto handle_mem; 2666 2667 case Op_StoreCM: 2668 { 2669 // Convert OopStore dependence into precedence edge 2670 Node* prec = n->in(MemNode::OopStore); 2671 n->del_req(MemNode::OopStore); 2672 n->add_prec(prec); 2673 eliminate_redundant_card_marks(n); 2674 } 2675 2676 // fall through 2677 2678 case Op_StoreB: 2679 case Op_StoreC: 2680 case Op_StorePConditional: 2681 case Op_StoreI: 2682 case Op_StoreL: 2683 case Op_StoreIConditional: 2684 case Op_StoreLConditional: 2685 case Op_CompareAndSwapI: 2686 case Op_CompareAndSwapL: 2687 case Op_CompareAndSwapP: 2688 case Op_CompareAndSwapN: 2689 case Op_GetAndAddI: 2690 case Op_GetAndAddL: 2691 case Op_GetAndSetI: 2692 case Op_GetAndSetL: 2693 case Op_GetAndSetP: 2694 case Op_GetAndSetN: 2695 case Op_StoreP: 2696 case Op_StoreN: 2697 case Op_StoreNKlass: 2698 case Op_LoadB: 2699 case Op_LoadUB: 2700 case Op_LoadUS: 2701 case Op_LoadI: 2702 case Op_LoadKlass: 2703 case Op_LoadNKlass: 2704 case Op_LoadL: 2705 case Op_LoadL_unaligned: 2706 case Op_LoadPLocked: 2707 case Op_LoadP: 2708 case Op_LoadN: 2709 case Op_LoadRange: 2710 case Op_LoadS: { 2711 handle_mem: 2712 #ifdef ASSERT 2713 if( VerifyOptoOopOffsets ) { 2714 assert( n->is_Mem(), "" ); 2715 MemNode *mem = (MemNode*)n; 2716 // Check to see if address types have grounded out somehow. 2717 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2718 assert( !tp || oop_offset_is_sane(tp), "" ); 2719 } 2720 #endif 2721 break; 2722 } 2723 2724 case Op_AddP: { // Assert sane base pointers 2725 Node *addp = n->in(AddPNode::Address); 2726 assert( !addp->is_AddP() || 2727 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2728 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2729 "Base pointers must match" ); 2730 #ifdef _LP64 2731 if ((UseCompressedOops || UseCompressedClassPointers) && 2732 addp->Opcode() == Op_ConP && 2733 addp == n->in(AddPNode::Base) && 2734 n->in(AddPNode::Offset)->is_Con()) { 2735 // Use addressing with narrow klass to load with offset on x86. 2736 // On sparc loading 32-bits constant and decoding it have less 2737 // instructions (4) then load 64-bits constant (7). 2738 // Do this transformation here since IGVN will convert ConN back to ConP. 2739 const Type* t = addp->bottom_type(); 2740 if (t->isa_oopptr() || t->isa_klassptr()) { 2741 Node* nn = NULL; 2742 2743 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2744 2745 // Look for existing ConN node of the same exact type. 2746 Node* r = root(); 2747 uint cnt = r->outcnt(); 2748 for (uint i = 0; i < cnt; i++) { 2749 Node* m = r->raw_out(i); 2750 if (m!= NULL && m->Opcode() == op && 2751 m->bottom_type()->make_ptr() == t) { 2752 nn = m; 2753 break; 2754 } 2755 } 2756 if (nn != NULL) { 2757 // Decode a narrow oop to match address 2758 // [R12 + narrow_oop_reg<<3 + offset] 2759 if (t->isa_oopptr()) { 2760 nn = new DecodeNNode(nn, t); 2761 } else { 2762 nn = new DecodeNKlassNode(nn, t); 2763 } 2764 n->set_req(AddPNode::Base, nn); 2765 n->set_req(AddPNode::Address, nn); 2766 if (addp->outcnt() == 0) { 2767 addp->disconnect_inputs(NULL, this); 2768 } 2769 } 2770 } 2771 } 2772 #endif 2773 break; 2774 } 2775 2776 #ifdef _LP64 2777 case Op_CastPP: 2778 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2779 Node* in1 = n->in(1); 2780 const Type* t = n->bottom_type(); 2781 Node* new_in1 = in1->clone(); 2782 new_in1->as_DecodeN()->set_type(t); 2783 2784 if (!Matcher::narrow_oop_use_complex_address()) { 2785 // 2786 // x86, ARM and friends can handle 2 adds in addressing mode 2787 // and Matcher can fold a DecodeN node into address by using 2788 // a narrow oop directly and do implicit NULL check in address: 2789 // 2790 // [R12 + narrow_oop_reg<<3 + offset] 2791 // NullCheck narrow_oop_reg 2792 // 2793 // On other platforms (Sparc) we have to keep new DecodeN node and 2794 // use it to do implicit NULL check in address: 2795 // 2796 // decode_not_null narrow_oop_reg, base_reg 2797 // [base_reg + offset] 2798 // NullCheck base_reg 2799 // 2800 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2801 // to keep the information to which NULL check the new DecodeN node 2802 // corresponds to use it as value in implicit_null_check(). 2803 // 2804 new_in1->set_req(0, n->in(0)); 2805 } 2806 2807 n->subsume_by(new_in1, this); 2808 if (in1->outcnt() == 0) { 2809 in1->disconnect_inputs(NULL, this); 2810 } 2811 } 2812 break; 2813 2814 case Op_CmpP: 2815 // Do this transformation here to preserve CmpPNode::sub() and 2816 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2817 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2818 Node* in1 = n->in(1); 2819 Node* in2 = n->in(2); 2820 if (!in1->is_DecodeNarrowPtr()) { 2821 in2 = in1; 2822 in1 = n->in(2); 2823 } 2824 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2825 2826 Node* new_in2 = NULL; 2827 if (in2->is_DecodeNarrowPtr()) { 2828 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2829 new_in2 = in2->in(1); 2830 } else if (in2->Opcode() == Op_ConP) { 2831 const Type* t = in2->bottom_type(); 2832 if (t == TypePtr::NULL_PTR) { 2833 assert(in1->is_DecodeN(), "compare klass to null?"); 2834 // Don't convert CmpP null check into CmpN if compressed 2835 // oops implicit null check is not generated. 2836 // This will allow to generate normal oop implicit null check. 2837 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2838 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2839 // 2840 // This transformation together with CastPP transformation above 2841 // will generated code for implicit NULL checks for compressed oops. 2842 // 2843 // The original code after Optimize() 2844 // 2845 // LoadN memory, narrow_oop_reg 2846 // decode narrow_oop_reg, base_reg 2847 // CmpP base_reg, NULL 2848 // CastPP base_reg // NotNull 2849 // Load [base_reg + offset], val_reg 2850 // 2851 // after these transformations will be 2852 // 2853 // LoadN memory, narrow_oop_reg 2854 // CmpN narrow_oop_reg, NULL 2855 // decode_not_null narrow_oop_reg, base_reg 2856 // Load [base_reg + offset], val_reg 2857 // 2858 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2859 // since narrow oops can be used in debug info now (see the code in 2860 // final_graph_reshaping_walk()). 2861 // 2862 // At the end the code will be matched to 2863 // on x86: 2864 // 2865 // Load_narrow_oop memory, narrow_oop_reg 2866 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2867 // NullCheck narrow_oop_reg 2868 // 2869 // and on sparc: 2870 // 2871 // Load_narrow_oop memory, narrow_oop_reg 2872 // decode_not_null narrow_oop_reg, base_reg 2873 // Load [base_reg + offset], val_reg 2874 // NullCheck base_reg 2875 // 2876 } else if (t->isa_oopptr()) { 2877 new_in2 = ConNode::make(this, t->make_narrowoop()); 2878 } else if (t->isa_klassptr()) { 2879 new_in2 = ConNode::make(this, t->make_narrowklass()); 2880 } 2881 } 2882 if (new_in2 != NULL) { 2883 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 2884 n->subsume_by(cmpN, this); 2885 if (in1->outcnt() == 0) { 2886 in1->disconnect_inputs(NULL, this); 2887 } 2888 if (in2->outcnt() == 0) { 2889 in2->disconnect_inputs(NULL, this); 2890 } 2891 } 2892 } 2893 break; 2894 2895 case Op_DecodeN: 2896 case Op_DecodeNKlass: 2897 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2898 // DecodeN could be pinned when it can't be fold into 2899 // an address expression, see the code for Op_CastPP above. 2900 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2901 break; 2902 2903 case Op_EncodeP: 2904 case Op_EncodePKlass: { 2905 Node* in1 = n->in(1); 2906 if (in1->is_DecodeNarrowPtr()) { 2907 n->subsume_by(in1->in(1), this); 2908 } else if (in1->Opcode() == Op_ConP) { 2909 const Type* t = in1->bottom_type(); 2910 if (t == TypePtr::NULL_PTR) { 2911 assert(t->isa_oopptr(), "null klass?"); 2912 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2913 } else if (t->isa_oopptr()) { 2914 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2915 } else if (t->isa_klassptr()) { 2916 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2917 } 2918 } 2919 if (in1->outcnt() == 0) { 2920 in1->disconnect_inputs(NULL, this); 2921 } 2922 break; 2923 } 2924 2925 case Op_Proj: { 2926 if (OptimizeStringConcat) { 2927 ProjNode* p = n->as_Proj(); 2928 if (p->_is_io_use) { 2929 // Separate projections were used for the exception path which 2930 // are normally removed by a late inline. If it wasn't inlined 2931 // then they will hang around and should just be replaced with 2932 // the original one. 2933 Node* proj = NULL; 2934 // Replace with just one 2935 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2936 Node *use = i.get(); 2937 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2938 proj = use; 2939 break; 2940 } 2941 } 2942 assert(proj != NULL, "must be found"); 2943 p->subsume_by(proj, this); 2944 } 2945 } 2946 break; 2947 } 2948 2949 case Op_Phi: 2950 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2951 // The EncodeP optimization may create Phi with the same edges 2952 // for all paths. It is not handled well by Register Allocator. 2953 Node* unique_in = n->in(1); 2954 assert(unique_in != NULL, ""); 2955 uint cnt = n->req(); 2956 for (uint i = 2; i < cnt; i++) { 2957 Node* m = n->in(i); 2958 assert(m != NULL, ""); 2959 if (unique_in != m) 2960 unique_in = NULL; 2961 } 2962 if (unique_in != NULL) { 2963 n->subsume_by(unique_in, this); 2964 } 2965 } 2966 break; 2967 2968 #endif 2969 2970 case Op_ModI: 2971 if (UseDivMod) { 2972 // Check if a%b and a/b both exist 2973 Node* d = n->find_similar(Op_DivI); 2974 if (d) { 2975 // Replace them with a fused divmod if supported 2976 if (Matcher::has_match_rule(Op_DivModI)) { 2977 DivModINode* divmod = DivModINode::make(this, n); 2978 d->subsume_by(divmod->div_proj(), this); 2979 n->subsume_by(divmod->mod_proj(), this); 2980 } else { 2981 // replace a%b with a-((a/b)*b) 2982 Node* mult = new MulINode(d, d->in(2)); 2983 Node* sub = new SubINode(d->in(1), mult); 2984 n->subsume_by(sub, this); 2985 } 2986 } 2987 } 2988 break; 2989 2990 case Op_ModL: 2991 if (UseDivMod) { 2992 // Check if a%b and a/b both exist 2993 Node* d = n->find_similar(Op_DivL); 2994 if (d) { 2995 // Replace them with a fused divmod if supported 2996 if (Matcher::has_match_rule(Op_DivModL)) { 2997 DivModLNode* divmod = DivModLNode::make(this, n); 2998 d->subsume_by(divmod->div_proj(), this); 2999 n->subsume_by(divmod->mod_proj(), this); 3000 } else { 3001 // replace a%b with a-((a/b)*b) 3002 Node* mult = new MulLNode(d, d->in(2)); 3003 Node* sub = new SubLNode(d->in(1), mult); 3004 n->subsume_by(sub, this); 3005 } 3006 } 3007 } 3008 break; 3009 3010 case Op_LoadVector: 3011 case Op_StoreVector: 3012 break; 3013 3014 case Op_PackB: 3015 case Op_PackS: 3016 case Op_PackI: 3017 case Op_PackF: 3018 case Op_PackL: 3019 case Op_PackD: 3020 if (n->req()-1 > 2) { 3021 // Replace many operand PackNodes with a binary tree for matching 3022 PackNode* p = (PackNode*) n; 3023 Node* btp = p->binary_tree_pack(this, 1, n->req()); 3024 n->subsume_by(btp, this); 3025 } 3026 break; 3027 case Op_Loop: 3028 case Op_CountedLoop: 3029 if (n->as_Loop()->is_inner_loop()) { 3030 frc.inc_inner_loop_count(); 3031 } 3032 break; 3033 case Op_LShiftI: 3034 case Op_RShiftI: 3035 case Op_URShiftI: 3036 case Op_LShiftL: 3037 case Op_RShiftL: 3038 case Op_URShiftL: 3039 if (Matcher::need_masked_shift_count) { 3040 // The cpu's shift instructions don't restrict the count to the 3041 // lower 5/6 bits. We need to do the masking ourselves. 3042 Node* in2 = n->in(2); 3043 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3044 const TypeInt* t = in2->find_int_type(); 3045 if (t != NULL && t->is_con()) { 3046 juint shift = t->get_con(); 3047 if (shift > mask) { // Unsigned cmp 3048 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 3049 } 3050 } else { 3051 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3052 Node* shift = new AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 3053 n->set_req(2, shift); 3054 } 3055 } 3056 if (in2->outcnt() == 0) { // Remove dead node 3057 in2->disconnect_inputs(NULL, this); 3058 } 3059 } 3060 break; 3061 case Op_MemBarStoreStore: 3062 case Op_MemBarRelease: 3063 // Break the link with AllocateNode: it is no longer useful and 3064 // confuses register allocation. 3065 if (n->req() > MemBarNode::Precedent) { 3066 n->set_req(MemBarNode::Precedent, top()); 3067 } 3068 break; 3069 default: 3070 assert( !n->is_Call(), "" ); 3071 assert( !n->is_Mem(), "" ); 3072 break; 3073 } 3074 3075 // Collect CFG split points 3076 if (n->is_MultiBranch()) 3077 frc._tests.push(n); 3078 } 3079 3080 //------------------------------final_graph_reshaping_walk--------------------- 3081 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3082 // requires that the walk visits a node's inputs before visiting the node. 3083 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3084 ResourceArea *area = Thread::current()->resource_area(); 3085 Unique_Node_List sfpt(area); 3086 3087 frc._visited.set(root->_idx); // first, mark node as visited 3088 uint cnt = root->req(); 3089 Node *n = root; 3090 uint i = 0; 3091 while (true) { 3092 if (i < cnt) { 3093 // Place all non-visited non-null inputs onto stack 3094 Node* m = n->in(i); 3095 ++i; 3096 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3097 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3098 // compute worst case interpreter size in case of a deoptimization 3099 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3100 3101 sfpt.push(m); 3102 } 3103 cnt = m->req(); 3104 nstack.push(n, i); // put on stack parent and next input's index 3105 n = m; 3106 i = 0; 3107 } 3108 } else { 3109 // Now do post-visit work 3110 final_graph_reshaping_impl( n, frc ); 3111 if (nstack.is_empty()) 3112 break; // finished 3113 n = nstack.node(); // Get node from stack 3114 cnt = n->req(); 3115 i = nstack.index(); 3116 nstack.pop(); // Shift to the next node on stack 3117 } 3118 } 3119 3120 // Skip next transformation if compressed oops are not used. 3121 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3122 (!UseCompressedOops && !UseCompressedClassPointers)) 3123 return; 3124 3125 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3126 // It could be done for an uncommon traps or any safepoints/calls 3127 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3128 while (sfpt.size() > 0) { 3129 n = sfpt.pop(); 3130 JVMState *jvms = n->as_SafePoint()->jvms(); 3131 assert(jvms != NULL, "sanity"); 3132 int start = jvms->debug_start(); 3133 int end = n->req(); 3134 bool is_uncommon = (n->is_CallStaticJava() && 3135 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3136 for (int j = start; j < end; j++) { 3137 Node* in = n->in(j); 3138 if (in->is_DecodeNarrowPtr()) { 3139 bool safe_to_skip = true; 3140 if (!is_uncommon ) { 3141 // Is it safe to skip? 3142 for (uint i = 0; i < in->outcnt(); i++) { 3143 Node* u = in->raw_out(i); 3144 if (!u->is_SafePoint() || 3145 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3146 safe_to_skip = false; 3147 } 3148 } 3149 } 3150 if (safe_to_skip) { 3151 n->set_req(j, in->in(1)); 3152 } 3153 if (in->outcnt() == 0) { 3154 in->disconnect_inputs(NULL, this); 3155 } 3156 } 3157 } 3158 } 3159 } 3160 3161 //------------------------------final_graph_reshaping-------------------------- 3162 // Final Graph Reshaping. 3163 // 3164 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3165 // and not commoned up and forced early. Must come after regular 3166 // optimizations to avoid GVN undoing the cloning. Clone constant 3167 // inputs to Loop Phis; these will be split by the allocator anyways. 3168 // Remove Opaque nodes. 3169 // (2) Move last-uses by commutative operations to the left input to encourage 3170 // Intel update-in-place two-address operations and better register usage 3171 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3172 // calls canonicalizing them back. 3173 // (3) Count the number of double-precision FP ops, single-precision FP ops 3174 // and call sites. On Intel, we can get correct rounding either by 3175 // forcing singles to memory (requires extra stores and loads after each 3176 // FP bytecode) or we can set a rounding mode bit (requires setting and 3177 // clearing the mode bit around call sites). The mode bit is only used 3178 // if the relative frequency of single FP ops to calls is low enough. 3179 // This is a key transform for SPEC mpeg_audio. 3180 // (4) Detect infinite loops; blobs of code reachable from above but not 3181 // below. Several of the Code_Gen algorithms fail on such code shapes, 3182 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3183 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3184 // Detection is by looking for IfNodes where only 1 projection is 3185 // reachable from below or CatchNodes missing some targets. 3186 // (5) Assert for insane oop offsets in debug mode. 3187 3188 bool Compile::final_graph_reshaping() { 3189 // an infinite loop may have been eliminated by the optimizer, 3190 // in which case the graph will be empty. 3191 if (root()->req() == 1) { 3192 record_method_not_compilable("trivial infinite loop"); 3193 return true; 3194 } 3195 3196 // Expensive nodes have their control input set to prevent the GVN 3197 // from freely commoning them. There's no GVN beyond this point so 3198 // no need to keep the control input. We want the expensive nodes to 3199 // be freely moved to the least frequent code path by gcm. 3200 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3201 for (int i = 0; i < expensive_count(); i++) { 3202 _expensive_nodes->at(i)->set_req(0, NULL); 3203 } 3204 3205 Final_Reshape_Counts frc; 3206 3207 // Visit everybody reachable! 3208 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3209 Node_Stack nstack(unique() >> 1); 3210 final_graph_reshaping_walk(nstack, root(), frc); 3211 3212 // Check for unreachable (from below) code (i.e., infinite loops). 3213 for( uint i = 0; i < frc._tests.size(); i++ ) { 3214 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3215 // Get number of CFG targets. 3216 // Note that PCTables include exception targets after calls. 3217 uint required_outcnt = n->required_outcnt(); 3218 if (n->outcnt() != required_outcnt) { 3219 // Check for a few special cases. Rethrow Nodes never take the 3220 // 'fall-thru' path, so expected kids is 1 less. 3221 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3222 if (n->in(0)->in(0)->is_Call()) { 3223 CallNode *call = n->in(0)->in(0)->as_Call(); 3224 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3225 required_outcnt--; // Rethrow always has 1 less kid 3226 } else if (call->req() > TypeFunc::Parms && 3227 call->is_CallDynamicJava()) { 3228 // Check for null receiver. In such case, the optimizer has 3229 // detected that the virtual call will always result in a null 3230 // pointer exception. The fall-through projection of this CatchNode 3231 // will not be populated. 3232 Node *arg0 = call->in(TypeFunc::Parms); 3233 if (arg0->is_Type() && 3234 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3235 required_outcnt--; 3236 } 3237 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3238 call->req() > TypeFunc::Parms+1 && 3239 call->is_CallStaticJava()) { 3240 // Check for negative array length. In such case, the optimizer has 3241 // detected that the allocation attempt will always result in an 3242 // exception. There is no fall-through projection of this CatchNode . 3243 Node *arg1 = call->in(TypeFunc::Parms+1); 3244 if (arg1->is_Type() && 3245 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3246 required_outcnt--; 3247 } 3248 } 3249 } 3250 } 3251 // Recheck with a better notion of 'required_outcnt' 3252 if (n->outcnt() != required_outcnt) { 3253 record_method_not_compilable("malformed control flow"); 3254 return true; // Not all targets reachable! 3255 } 3256 } 3257 // Check that I actually visited all kids. Unreached kids 3258 // must be infinite loops. 3259 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3260 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3261 record_method_not_compilable("infinite loop"); 3262 return true; // Found unvisited kid; must be unreach 3263 } 3264 } 3265 3266 // If original bytecodes contained a mixture of floats and doubles 3267 // check if the optimizer has made it homogenous, item (3). 3268 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3269 frc.get_float_count() > 32 && 3270 frc.get_double_count() == 0 && 3271 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3272 set_24_bit_selection_and_mode( false, true ); 3273 } 3274 3275 set_java_calls(frc.get_java_call_count()); 3276 set_inner_loops(frc.get_inner_loop_count()); 3277 3278 // No infinite loops, no reason to bail out. 3279 return false; 3280 } 3281 3282 //-----------------------------too_many_traps---------------------------------- 3283 // Report if there are too many traps at the current method and bci. 3284 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3285 bool Compile::too_many_traps(ciMethod* method, 3286 int bci, 3287 Deoptimization::DeoptReason reason) { 3288 ciMethodData* md = method->method_data(); 3289 if (md->is_empty()) { 3290 // Assume the trap has not occurred, or that it occurred only 3291 // because of a transient condition during start-up in the interpreter. 3292 return false; 3293 } 3294 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3295 if (md->has_trap_at(bci, m, reason) != 0) { 3296 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3297 // Also, if there are multiple reasons, or if there is no per-BCI record, 3298 // assume the worst. 3299 if (log()) 3300 log()->elem("observe trap='%s' count='%d'", 3301 Deoptimization::trap_reason_name(reason), 3302 md->trap_count(reason)); 3303 return true; 3304 } else { 3305 // Ignore method/bci and see if there have been too many globally. 3306 return too_many_traps(reason, md); 3307 } 3308 } 3309 3310 // Less-accurate variant which does not require a method and bci. 3311 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3312 ciMethodData* logmd) { 3313 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3314 // Too many traps globally. 3315 // Note that we use cumulative trap_count, not just md->trap_count. 3316 if (log()) { 3317 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3318 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3319 Deoptimization::trap_reason_name(reason), 3320 mcount, trap_count(reason)); 3321 } 3322 return true; 3323 } else { 3324 // The coast is clear. 3325 return false; 3326 } 3327 } 3328 3329 //--------------------------too_many_recompiles-------------------------------- 3330 // Report if there are too many recompiles at the current method and bci. 3331 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3332 // Is not eager to return true, since this will cause the compiler to use 3333 // Action_none for a trap point, to avoid too many recompilations. 3334 bool Compile::too_many_recompiles(ciMethod* method, 3335 int bci, 3336 Deoptimization::DeoptReason reason) { 3337 ciMethodData* md = method->method_data(); 3338 if (md->is_empty()) { 3339 // Assume the trap has not occurred, or that it occurred only 3340 // because of a transient condition during start-up in the interpreter. 3341 return false; 3342 } 3343 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3344 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3345 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3346 Deoptimization::DeoptReason per_bc_reason 3347 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3348 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3349 if ((per_bc_reason == Deoptimization::Reason_none 3350 || md->has_trap_at(bci, m, reason) != 0) 3351 // The trap frequency measure we care about is the recompile count: 3352 && md->trap_recompiled_at(bci, m) 3353 && md->overflow_recompile_count() >= bc_cutoff) { 3354 // Do not emit a trap here if it has already caused recompilations. 3355 // Also, if there are multiple reasons, or if there is no per-BCI record, 3356 // assume the worst. 3357 if (log()) 3358 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3359 Deoptimization::trap_reason_name(reason), 3360 md->trap_count(reason), 3361 md->overflow_recompile_count()); 3362 return true; 3363 } else if (trap_count(reason) != 0 3364 && decompile_count() >= m_cutoff) { 3365 // Too many recompiles globally, and we have seen this sort of trap. 3366 // Use cumulative decompile_count, not just md->decompile_count. 3367 if (log()) 3368 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3369 Deoptimization::trap_reason_name(reason), 3370 md->trap_count(reason), trap_count(reason), 3371 md->decompile_count(), decompile_count()); 3372 return true; 3373 } else { 3374 // The coast is clear. 3375 return false; 3376 } 3377 } 3378 3379 // Compute when not to trap. Used by matching trap based nodes and 3380 // NullCheck optimization. 3381 void Compile::set_allowed_deopt_reasons() { 3382 _allowed_reasons = 0; 3383 if (is_method_compilation()) { 3384 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3385 assert(rs < BitsPerInt, "recode bit map"); 3386 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3387 _allowed_reasons |= nth_bit(rs); 3388 } 3389 } 3390 } 3391 } 3392 3393 #ifndef PRODUCT 3394 //------------------------------verify_graph_edges--------------------------- 3395 // Walk the Graph and verify that there is a one-to-one correspondence 3396 // between Use-Def edges and Def-Use edges in the graph. 3397 void Compile::verify_graph_edges(bool no_dead_code) { 3398 if (VerifyGraphEdges) { 3399 ResourceArea *area = Thread::current()->resource_area(); 3400 Unique_Node_List visited(area); 3401 // Call recursive graph walk to check edges 3402 _root->verify_edges(visited); 3403 if (no_dead_code) { 3404 // Now make sure that no visited node is used by an unvisited node. 3405 bool dead_nodes = 0; 3406 Unique_Node_List checked(area); 3407 while (visited.size() > 0) { 3408 Node* n = visited.pop(); 3409 checked.push(n); 3410 for (uint i = 0; i < n->outcnt(); i++) { 3411 Node* use = n->raw_out(i); 3412 if (checked.member(use)) continue; // already checked 3413 if (visited.member(use)) continue; // already in the graph 3414 if (use->is_Con()) continue; // a dead ConNode is OK 3415 // At this point, we have found a dead node which is DU-reachable. 3416 if (dead_nodes++ == 0) 3417 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3418 use->dump(2); 3419 tty->print_cr("---"); 3420 checked.push(use); // No repeats; pretend it is now checked. 3421 } 3422 } 3423 assert(dead_nodes == 0, "using nodes must be reachable from root"); 3424 } 3425 } 3426 } 3427 3428 // Verify GC barriers consistency 3429 // Currently supported: 3430 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3431 void Compile::verify_barriers() { 3432 if (UseG1GC) { 3433 // Verify G1 pre-barriers 3434 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3435 3436 ResourceArea *area = Thread::current()->resource_area(); 3437 Unique_Node_List visited(area); 3438 Node_List worklist(area); 3439 // We're going to walk control flow backwards starting from the Root 3440 worklist.push(_root); 3441 while (worklist.size() > 0) { 3442 Node* x = worklist.pop(); 3443 if (x == NULL || x == top()) continue; 3444 if (visited.member(x)) { 3445 continue; 3446 } else { 3447 visited.push(x); 3448 } 3449 3450 if (x->is_Region()) { 3451 for (uint i = 1; i < x->req(); i++) { 3452 worklist.push(x->in(i)); 3453 } 3454 } else { 3455 worklist.push(x->in(0)); 3456 // We are looking for the pattern: 3457 // /->ThreadLocal 3458 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3459 // \->ConI(0) 3460 // We want to verify that the If and the LoadB have the same control 3461 // See GraphKit::g1_write_barrier_pre() 3462 if (x->is_If()) { 3463 IfNode *iff = x->as_If(); 3464 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3465 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3466 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3467 && cmp->in(1)->is_Load()) { 3468 LoadNode* load = cmp->in(1)->as_Load(); 3469 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3470 && load->in(2)->in(3)->is_Con() 3471 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3472 3473 Node* if_ctrl = iff->in(0); 3474 Node* load_ctrl = load->in(0); 3475 3476 if (if_ctrl != load_ctrl) { 3477 // Skip possible CProj->NeverBranch in infinite loops 3478 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3479 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3480 if_ctrl = if_ctrl->in(0)->in(0); 3481 } 3482 } 3483 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3484 } 3485 } 3486 } 3487 } 3488 } 3489 } 3490 } 3491 } 3492 3493 #endif 3494 3495 // The Compile object keeps track of failure reasons separately from the ciEnv. 3496 // This is required because there is not quite a 1-1 relation between the 3497 // ciEnv and its compilation task and the Compile object. Note that one 3498 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3499 // to backtrack and retry without subsuming loads. Other than this backtracking 3500 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3501 // by the logic in C2Compiler. 3502 void Compile::record_failure(const char* reason) { 3503 if (log() != NULL) { 3504 log()->elem("failure reason='%s' phase='compile'", reason); 3505 } 3506 if (_failure_reason == NULL) { 3507 // Record the first failure reason. 3508 _failure_reason = reason; 3509 } 3510 3511 EventCompilerFailure event; 3512 if (event.should_commit()) { 3513 event.set_compileID(Compile::compile_id()); 3514 event.set_failure(reason); 3515 event.commit(); 3516 } 3517 3518 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3519 C->print_method(PHASE_FAILURE); 3520 } 3521 _root = NULL; // flush the graph, too 3522 } 3523 3524 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3525 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3526 _phase_name(name), _dolog(dolog) 3527 { 3528 if (dolog) { 3529 C = Compile::current(); 3530 _log = C->log(); 3531 } else { 3532 C = NULL; 3533 _log = NULL; 3534 } 3535 if (_log != NULL) { 3536 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3537 _log->stamp(); 3538 _log->end_head(); 3539 } 3540 } 3541 3542 Compile::TracePhase::~TracePhase() { 3543 3544 C = Compile::current(); 3545 if (_dolog) { 3546 _log = C->log(); 3547 } else { 3548 _log = NULL; 3549 } 3550 3551 #ifdef ASSERT 3552 if (PrintIdealNodeCount) { 3553 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3554 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3555 } 3556 3557 if (VerifyIdealNodeCount) { 3558 Compile::current()->print_missing_nodes(); 3559 } 3560 #endif 3561 3562 if (_log != NULL) { 3563 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3564 } 3565 } 3566 3567 //============================================================================= 3568 // Two Constant's are equal when the type and the value are equal. 3569 bool Compile::Constant::operator==(const Constant& other) { 3570 if (type() != other.type() ) return false; 3571 if (can_be_reused() != other.can_be_reused()) return false; 3572 // For floating point values we compare the bit pattern. 3573 switch (type()) { 3574 case T_FLOAT: return (_v._value.i == other._v._value.i); 3575 case T_LONG: 3576 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3577 case T_OBJECT: 3578 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3579 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3580 case T_METADATA: return (_v._metadata == other._v._metadata); 3581 default: ShouldNotReachHere(); 3582 } 3583 return false; 3584 } 3585 3586 static int type_to_size_in_bytes(BasicType t) { 3587 switch (t) { 3588 case T_LONG: return sizeof(jlong ); 3589 case T_FLOAT: return sizeof(jfloat ); 3590 case T_DOUBLE: return sizeof(jdouble); 3591 case T_METADATA: return sizeof(Metadata*); 3592 // We use T_VOID as marker for jump-table entries (labels) which 3593 // need an internal word relocation. 3594 case T_VOID: 3595 case T_ADDRESS: 3596 case T_OBJECT: return sizeof(jobject); 3597 } 3598 3599 ShouldNotReachHere(); 3600 return -1; 3601 } 3602 3603 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3604 // sort descending 3605 if (a->freq() > b->freq()) return -1; 3606 if (a->freq() < b->freq()) return 1; 3607 return 0; 3608 } 3609 3610 void Compile::ConstantTable::calculate_offsets_and_size() { 3611 // First, sort the array by frequencies. 3612 _constants.sort(qsort_comparator); 3613 3614 #ifdef ASSERT 3615 // Make sure all jump-table entries were sorted to the end of the 3616 // array (they have a negative frequency). 3617 bool found_void = false; 3618 for (int i = 0; i < _constants.length(); i++) { 3619 Constant con = _constants.at(i); 3620 if (con.type() == T_VOID) 3621 found_void = true; // jump-tables 3622 else 3623 assert(!found_void, "wrong sorting"); 3624 } 3625 #endif 3626 3627 int offset = 0; 3628 for (int i = 0; i < _constants.length(); i++) { 3629 Constant* con = _constants.adr_at(i); 3630 3631 // Align offset for type. 3632 int typesize = type_to_size_in_bytes(con->type()); 3633 offset = align_size_up(offset, typesize); 3634 con->set_offset(offset); // set constant's offset 3635 3636 if (con->type() == T_VOID) { 3637 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3638 offset = offset + typesize * n->outcnt(); // expand jump-table 3639 } else { 3640 offset = offset + typesize; 3641 } 3642 } 3643 3644 // Align size up to the next section start (which is insts; see 3645 // CodeBuffer::align_at_start). 3646 assert(_size == -1, "already set?"); 3647 _size = align_size_up(offset, CodeEntryAlignment); 3648 } 3649 3650 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3651 MacroAssembler _masm(&cb); 3652 for (int i = 0; i < _constants.length(); i++) { 3653 Constant con = _constants.at(i); 3654 address constant_addr; 3655 switch (con.type()) { 3656 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3657 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3658 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3659 case T_OBJECT: { 3660 jobject obj = con.get_jobject(); 3661 int oop_index = _masm.oop_recorder()->find_index(obj); 3662 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3663 break; 3664 } 3665 case T_ADDRESS: { 3666 address addr = (address) con.get_jobject(); 3667 constant_addr = _masm.address_constant(addr); 3668 break; 3669 } 3670 // We use T_VOID as marker for jump-table entries (labels) which 3671 // need an internal word relocation. 3672 case T_VOID: { 3673 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3674 // Fill the jump-table with a dummy word. The real value is 3675 // filled in later in fill_jump_table. 3676 address dummy = (address) n; 3677 constant_addr = _masm.address_constant(dummy); 3678 // Expand jump-table 3679 for (uint i = 1; i < n->outcnt(); i++) { 3680 address temp_addr = _masm.address_constant(dummy + i); 3681 assert(temp_addr, "consts section too small"); 3682 } 3683 break; 3684 } 3685 case T_METADATA: { 3686 Metadata* obj = con.get_metadata(); 3687 int metadata_index = _masm.oop_recorder()->find_index(obj); 3688 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3689 break; 3690 } 3691 default: ShouldNotReachHere(); 3692 } 3693 assert(constant_addr, "consts section too small"); 3694 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3695 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3696 } 3697 } 3698 3699 int Compile::ConstantTable::find_offset(Constant& con) const { 3700 int idx = _constants.find(con); 3701 assert(idx != -1, "constant must be in constant table"); 3702 int offset = _constants.at(idx).offset(); 3703 assert(offset != -1, "constant table not emitted yet?"); 3704 return offset; 3705 } 3706 3707 void Compile::ConstantTable::add(Constant& con) { 3708 if (con.can_be_reused()) { 3709 int idx = _constants.find(con); 3710 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3711 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3712 return; 3713 } 3714 } 3715 (void) _constants.append(con); 3716 } 3717 3718 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3719 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3720 Constant con(type, value, b->_freq); 3721 add(con); 3722 return con; 3723 } 3724 3725 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3726 Constant con(metadata); 3727 add(con); 3728 return con; 3729 } 3730 3731 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3732 jvalue value; 3733 BasicType type = oper->type()->basic_type(); 3734 switch (type) { 3735 case T_LONG: value.j = oper->constantL(); break; 3736 case T_FLOAT: value.f = oper->constantF(); break; 3737 case T_DOUBLE: value.d = oper->constantD(); break; 3738 case T_OBJECT: 3739 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3740 case T_METADATA: return add((Metadata*)oper->constant()); break; 3741 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3742 } 3743 return add(n, type, value); 3744 } 3745 3746 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3747 jvalue value; 3748 // We can use the node pointer here to identify the right jump-table 3749 // as this method is called from Compile::Fill_buffer right before 3750 // the MachNodes are emitted and the jump-table is filled (means the 3751 // MachNode pointers do not change anymore). 3752 value.l = (jobject) n; 3753 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3754 add(con); 3755 return con; 3756 } 3757 3758 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3759 // If called from Compile::scratch_emit_size do nothing. 3760 if (Compile::current()->in_scratch_emit_size()) return; 3761 3762 assert(labels.is_nonempty(), "must be"); 3763 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3764 3765 // Since MachConstantNode::constant_offset() also contains 3766 // table_base_offset() we need to subtract the table_base_offset() 3767 // to get the plain offset into the constant table. 3768 int offset = n->constant_offset() - table_base_offset(); 3769 3770 MacroAssembler _masm(&cb); 3771 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3772 3773 for (uint i = 0; i < n->outcnt(); i++) { 3774 address* constant_addr = &jump_table_base[i]; 3775 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i))); 3776 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3777 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3778 } 3779 } 3780 3781 // The message about the current inlining is accumulated in 3782 // _print_inlining_stream and transfered into the _print_inlining_list 3783 // once we know whether inlining succeeds or not. For regular 3784 // inlining, messages are appended to the buffer pointed by 3785 // _print_inlining_idx in the _print_inlining_list. For late inlining, 3786 // a new buffer is added after _print_inlining_idx in the list. This 3787 // way we can update the inlining message for late inlining call site 3788 // when the inlining is attempted again. 3789 void Compile::print_inlining_init() { 3790 if (print_inlining() || print_intrinsics()) { 3791 _print_inlining_stream = new stringStream(); 3792 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 3793 } 3794 } 3795 3796 void Compile::print_inlining_reinit() { 3797 if (print_inlining() || print_intrinsics()) { 3798 // Re allocate buffer when we change ResourceMark 3799 _print_inlining_stream = new stringStream(); 3800 } 3801 } 3802 3803 void Compile::print_inlining_reset() { 3804 _print_inlining_stream->reset(); 3805 } 3806 3807 void Compile::print_inlining_commit() { 3808 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 3809 // Transfer the message from _print_inlining_stream to the current 3810 // _print_inlining_list buffer and clear _print_inlining_stream. 3811 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 3812 print_inlining_reset(); 3813 } 3814 3815 void Compile::print_inlining_push() { 3816 // Add new buffer to the _print_inlining_list at current position 3817 _print_inlining_idx++; 3818 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 3819 } 3820 3821 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 3822 return _print_inlining_list->at(_print_inlining_idx); 3823 } 3824 3825 void Compile::print_inlining_update(CallGenerator* cg) { 3826 if (print_inlining() || print_intrinsics()) { 3827 if (!cg->is_late_inline()) { 3828 if (print_inlining_current().cg() != NULL) { 3829 print_inlining_push(); 3830 } 3831 print_inlining_commit(); 3832 } else { 3833 if (print_inlining_current().cg() != cg && 3834 (print_inlining_current().cg() != NULL || 3835 print_inlining_current().ss()->size() != 0)) { 3836 print_inlining_push(); 3837 } 3838 print_inlining_commit(); 3839 print_inlining_current().set_cg(cg); 3840 } 3841 } 3842 } 3843 3844 void Compile::print_inlining_move_to(CallGenerator* cg) { 3845 // We resume inlining at a late inlining call site. Locate the 3846 // corresponding inlining buffer so that we can update it. 3847 if (print_inlining()) { 3848 for (int i = 0; i < _print_inlining_list->length(); i++) { 3849 if (_print_inlining_list->adr_at(i)->cg() == cg) { 3850 _print_inlining_idx = i; 3851 return; 3852 } 3853 } 3854 ShouldNotReachHere(); 3855 } 3856 } 3857 3858 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 3859 if (print_inlining()) { 3860 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 3861 assert(print_inlining_current().cg() == cg, "wrong entry"); 3862 // replace message with new message 3863 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 3864 print_inlining_commit(); 3865 print_inlining_current().set_cg(cg); 3866 } 3867 } 3868 3869 void Compile::print_inlining_assert_ready() { 3870 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 3871 } 3872 3873 void Compile::process_print_inlining() { 3874 bool do_print_inlining = print_inlining() || print_intrinsics(); 3875 if (do_print_inlining || log() != NULL) { 3876 // Print inlining message for candidates that we couldn't inline 3877 // for lack of space 3878 for (int i = 0; i < _late_inlines.length(); i++) { 3879 CallGenerator* cg = _late_inlines.at(i); 3880 if (!cg->is_mh_late_inline()) { 3881 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 3882 if (do_print_inlining) { 3883 cg->print_inlining_late(msg); 3884 } 3885 log_late_inline_failure(cg, msg); 3886 } 3887 } 3888 } 3889 if (do_print_inlining) { 3890 ResourceMark rm; 3891 stringStream ss; 3892 for (int i = 0; i < _print_inlining_list->length(); i++) { 3893 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 3894 } 3895 size_t end = ss.size(); 3896 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 3897 strncpy(_print_inlining_output, ss.base(), end+1); 3898 _print_inlining_output[end] = 0; 3899 } 3900 } 3901 3902 void Compile::dump_print_inlining() { 3903 if (_print_inlining_output != NULL) { 3904 tty->print_raw(_print_inlining_output); 3905 } 3906 } 3907 3908 void Compile::log_late_inline(CallGenerator* cg) { 3909 if (log() != NULL) { 3910 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 3911 cg->unique_id()); 3912 JVMState* p = cg->call_node()->jvms(); 3913 while (p != NULL) { 3914 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 3915 p = p->caller(); 3916 } 3917 log()->tail("late_inline"); 3918 } 3919 } 3920 3921 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 3922 log_late_inline(cg); 3923 if (log() != NULL) { 3924 log()->inline_fail(msg); 3925 } 3926 } 3927 3928 void Compile::log_inline_id(CallGenerator* cg) { 3929 if (log() != NULL) { 3930 // The LogCompilation tool needs a unique way to identify late 3931 // inline call sites. This id must be unique for this call site in 3932 // this compilation. Try to have it unique across compilations as 3933 // well because it can be convenient when grepping through the log 3934 // file. 3935 // Distinguish OSR compilations from others in case CICountOSR is 3936 // on. 3937 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 3938 cg->set_unique_id(id); 3939 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 3940 } 3941 } 3942 3943 void Compile::log_inline_failure(const char* msg) { 3944 if (C->log() != NULL) { 3945 C->log()->inline_fail(msg); 3946 } 3947 } 3948 3949 3950 // Dump inlining replay data to the stream. 3951 // Don't change thread state and acquire any locks. 3952 void Compile::dump_inline_data(outputStream* out) { 3953 InlineTree* inl_tree = ilt(); 3954 if (inl_tree != NULL) { 3955 out->print(" inline %d", inl_tree->count()); 3956 inl_tree->dump_replay_data(out); 3957 } 3958 } 3959 3960 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 3961 if (n1->Opcode() < n2->Opcode()) return -1; 3962 else if (n1->Opcode() > n2->Opcode()) return 1; 3963 3964 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())); 3965 for (uint i = 1; i < n1->req(); i++) { 3966 if (n1->in(i) < n2->in(i)) return -1; 3967 else if (n1->in(i) > n2->in(i)) return 1; 3968 } 3969 3970 return 0; 3971 } 3972 3973 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 3974 Node* n1 = *n1p; 3975 Node* n2 = *n2p; 3976 3977 return cmp_expensive_nodes(n1, n2); 3978 } 3979 3980 void Compile::sort_expensive_nodes() { 3981 if (!expensive_nodes_sorted()) { 3982 _expensive_nodes->sort(cmp_expensive_nodes); 3983 } 3984 } 3985 3986 bool Compile::expensive_nodes_sorted() const { 3987 for (int i = 1; i < _expensive_nodes->length(); i++) { 3988 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 3989 return false; 3990 } 3991 } 3992 return true; 3993 } 3994 3995 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 3996 if (_expensive_nodes->length() == 0) { 3997 return false; 3998 } 3999 4000 assert(OptimizeExpensiveOps, "optimization off?"); 4001 4002 // Take this opportunity to remove dead nodes from the list 4003 int j = 0; 4004 for (int i = 0; i < _expensive_nodes->length(); i++) { 4005 Node* n = _expensive_nodes->at(i); 4006 if (!n->is_unreachable(igvn)) { 4007 assert(n->is_expensive(), "should be expensive"); 4008 _expensive_nodes->at_put(j, n); 4009 j++; 4010 } 4011 } 4012 _expensive_nodes->trunc_to(j); 4013 4014 // Then sort the list so that similar nodes are next to each other 4015 // and check for at least two nodes of identical kind with same data 4016 // inputs. 4017 sort_expensive_nodes(); 4018 4019 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4020 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4021 return true; 4022 } 4023 } 4024 4025 return false; 4026 } 4027 4028 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4029 if (_expensive_nodes->length() == 0) { 4030 return; 4031 } 4032 4033 assert(OptimizeExpensiveOps, "optimization off?"); 4034 4035 // Sort to bring similar nodes next to each other and clear the 4036 // control input of nodes for which there's only a single copy. 4037 sort_expensive_nodes(); 4038 4039 int j = 0; 4040 int identical = 0; 4041 int i = 0; 4042 for (; i < _expensive_nodes->length()-1; i++) { 4043 assert(j <= i, "can't write beyond current index"); 4044 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4045 identical++; 4046 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4047 continue; 4048 } 4049 if (identical > 0) { 4050 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4051 identical = 0; 4052 } else { 4053 Node* n = _expensive_nodes->at(i); 4054 igvn.hash_delete(n); 4055 n->set_req(0, NULL); 4056 igvn.hash_insert(n); 4057 } 4058 } 4059 if (identical > 0) { 4060 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4061 } else if (_expensive_nodes->length() >= 1) { 4062 Node* n = _expensive_nodes->at(i); 4063 igvn.hash_delete(n); 4064 n->set_req(0, NULL); 4065 igvn.hash_insert(n); 4066 } 4067 _expensive_nodes->trunc_to(j); 4068 } 4069 4070 void Compile::add_expensive_node(Node * n) { 4071 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4072 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4073 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4074 if (OptimizeExpensiveOps) { 4075 _expensive_nodes->append(n); 4076 } else { 4077 // Clear control input and let IGVN optimize expensive nodes if 4078 // OptimizeExpensiveOps is off. 4079 n->set_req(0, NULL); 4080 } 4081 } 4082 4083 /** 4084 * Remove the speculative part of types and clean up the graph 4085 */ 4086 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4087 if (UseTypeSpeculation) { 4088 Unique_Node_List worklist; 4089 worklist.push(root()); 4090 int modified = 0; 4091 // Go over all type nodes that carry a speculative type, drop the 4092 // speculative part of the type and enqueue the node for an igvn 4093 // which may optimize it out. 4094 for (uint next = 0; next < worklist.size(); ++next) { 4095 Node *n = worklist.at(next); 4096 if (n->is_Type()) { 4097 TypeNode* tn = n->as_Type(); 4098 const Type* t = tn->type(); 4099 const Type* t_no_spec = t->remove_speculative(); 4100 if (t_no_spec != t) { 4101 bool in_hash = igvn.hash_delete(n); 4102 assert(in_hash, "node should be in igvn hash table"); 4103 tn->set_type(t_no_spec); 4104 igvn.hash_insert(n); 4105 igvn._worklist.push(n); // give it a chance to go away 4106 modified++; 4107 } 4108 } 4109 uint max = n->len(); 4110 for( uint i = 0; i < max; ++i ) { 4111 Node *m = n->in(i); 4112 if (not_a_node(m)) continue; 4113 worklist.push(m); 4114 } 4115 } 4116 // Drop the speculative part of all types in the igvn's type table 4117 igvn.remove_speculative_types(); 4118 if (modified > 0) { 4119 igvn.optimize(); 4120 } 4121 #ifdef ASSERT 4122 // Verify that after the IGVN is over no speculative type has resurfaced 4123 worklist.clear(); 4124 worklist.push(root()); 4125 for (uint next = 0; next < worklist.size(); ++next) { 4126 Node *n = worklist.at(next); 4127 const Type* t = igvn.type_or_null(n); 4128 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4129 if (n->is_Type()) { 4130 t = n->as_Type()->type(); 4131 assert(t == t->remove_speculative(), "no more speculative types"); 4132 } 4133 uint max = n->len(); 4134 for( uint i = 0; i < max; ++i ) { 4135 Node *m = n->in(i); 4136 if (not_a_node(m)) continue; 4137 worklist.push(m); 4138 } 4139 } 4140 igvn.check_no_speculative_types(); 4141 #endif 4142 } 4143 } 4144 4145 // Auxiliary method to support randomized stressing/fuzzing. 4146 // 4147 // This method can be called the arbitrary number of times, with current count 4148 // as the argument. The logic allows selecting a single candidate from the 4149 // running list of candidates as follows: 4150 // int count = 0; 4151 // Cand* selected = null; 4152 // while(cand = cand->next()) { 4153 // if (randomized_select(++count)) { 4154 // selected = cand; 4155 // } 4156 // } 4157 // 4158 // Including count equalizes the chances any candidate is "selected". 4159 // This is useful when we don't have the complete list of candidates to choose 4160 // from uniformly. In this case, we need to adjust the randomicity of the 4161 // selection, or else we will end up biasing the selection towards the latter 4162 // candidates. 4163 // 4164 // Quick back-envelope calculation shows that for the list of n candidates 4165 // the equal probability for the candidate to persist as "best" can be 4166 // achieved by replacing it with "next" k-th candidate with the probability 4167 // of 1/k. It can be easily shown that by the end of the run, the 4168 // probability for any candidate is converged to 1/n, thus giving the 4169 // uniform distribution among all the candidates. 4170 // 4171 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4172 #define RANDOMIZED_DOMAIN_POW 29 4173 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4174 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4175 bool Compile::randomized_select(int count) { 4176 assert(count > 0, "only positive"); 4177 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4178 }