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