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